userfaultfd: non-cooperative: closing the uffd without triggering SIGBUS
[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(ACCESS_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 && !ACCESS_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 (ACCESS_ONCE(ctx->released) ||
590                     fatal_signal_pending(current)) {
591                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
592                         if (ewq->msg.event == UFFD_EVENT_FORK) {
593                                 struct userfaultfd_ctx *new;
594
595                                 new = (struct userfaultfd_ctx *)
596                                         (unsigned long)
597                                         ewq->msg.arg.reserved.reserved1;
598
599                                 userfaultfd_ctx_put(new);
600                         }
601                         break;
602                 }
603
604                 spin_unlock(&ctx->event_wqh.lock);
605
606                 wake_up_poll(&ctx->fd_wqh, POLLIN);
607                 schedule();
608
609                 spin_lock(&ctx->event_wqh.lock);
610         }
611         __set_current_state(TASK_RUNNING);
612         spin_unlock(&ctx->event_wqh.lock);
613
614         /*
615          * ctx may go away after this if the userfault pseudo fd is
616          * already released.
617          */
618 out:
619         userfaultfd_ctx_put(ctx);
620 }
621
622 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
623                                        struct userfaultfd_wait_queue *ewq)
624 {
625         ewq->msg.event = 0;
626         wake_up_locked(&ctx->event_wqh);
627         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
628 }
629
630 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
631 {
632         struct userfaultfd_ctx *ctx = NULL, *octx;
633         struct userfaultfd_fork_ctx *fctx;
634
635         octx = vma->vm_userfaultfd_ctx.ctx;
636         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
637                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
638                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
639                 return 0;
640         }
641
642         list_for_each_entry(fctx, fcs, list)
643                 if (fctx->orig == octx) {
644                         ctx = fctx->new;
645                         break;
646                 }
647
648         if (!ctx) {
649                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
650                 if (!fctx)
651                         return -ENOMEM;
652
653                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
654                 if (!ctx) {
655                         kfree(fctx);
656                         return -ENOMEM;
657                 }
658
659                 atomic_set(&ctx->refcount, 1);
660                 ctx->flags = octx->flags;
661                 ctx->state = UFFD_STATE_RUNNING;
662                 ctx->features = octx->features;
663                 ctx->released = false;
664                 ctx->mm = vma->vm_mm;
665                 atomic_inc(&ctx->mm->mm_count);
666
667                 userfaultfd_ctx_get(octx);
668                 fctx->orig = octx;
669                 fctx->new = ctx;
670                 list_add_tail(&fctx->list, fcs);
671         }
672
673         vma->vm_userfaultfd_ctx.ctx = ctx;
674         return 0;
675 }
676
677 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
678 {
679         struct userfaultfd_ctx *ctx = fctx->orig;
680         struct userfaultfd_wait_queue ewq;
681
682         msg_init(&ewq.msg);
683
684         ewq.msg.event = UFFD_EVENT_FORK;
685         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
686
687         userfaultfd_event_wait_completion(ctx, &ewq);
688 }
689
690 void dup_userfaultfd_complete(struct list_head *fcs)
691 {
692         struct userfaultfd_fork_ctx *fctx, *n;
693
694         list_for_each_entry_safe(fctx, n, fcs, list) {
695                 dup_fctx(fctx);
696                 list_del(&fctx->list);
697                 kfree(fctx);
698         }
699 }
700
701 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
702                              struct vm_userfaultfd_ctx *vm_ctx)
703 {
704         struct userfaultfd_ctx *ctx;
705
706         ctx = vma->vm_userfaultfd_ctx.ctx;
707         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
708                 vm_ctx->ctx = ctx;
709                 userfaultfd_ctx_get(ctx);
710         }
711 }
712
713 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
714                                  unsigned long from, unsigned long to,
715                                  unsigned long len)
716 {
717         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
718         struct userfaultfd_wait_queue ewq;
719
720         if (!ctx)
721                 return;
722
723         if (to & ~PAGE_MASK) {
724                 userfaultfd_ctx_put(ctx);
725                 return;
726         }
727
728         msg_init(&ewq.msg);
729
730         ewq.msg.event = UFFD_EVENT_REMAP;
731         ewq.msg.arg.remap.from = from;
732         ewq.msg.arg.remap.to = to;
733         ewq.msg.arg.remap.len = len;
734
735         userfaultfd_event_wait_completion(ctx, &ewq);
736 }
737
738 bool userfaultfd_remove(struct vm_area_struct *vma,
739                         unsigned long start, unsigned long end)
740 {
741         struct mm_struct *mm = vma->vm_mm;
742         struct userfaultfd_ctx *ctx;
743         struct userfaultfd_wait_queue ewq;
744
745         ctx = vma->vm_userfaultfd_ctx.ctx;
746         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
747                 return true;
748
749         userfaultfd_ctx_get(ctx);
750         up_read(&mm->mmap_sem);
751
752         msg_init(&ewq.msg);
753
754         ewq.msg.event = UFFD_EVENT_REMOVE;
755         ewq.msg.arg.remove.start = start;
756         ewq.msg.arg.remove.end = end;
757
758         userfaultfd_event_wait_completion(ctx, &ewq);
759
760         return false;
761 }
762
763 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
764                           unsigned long start, unsigned long end)
765 {
766         struct userfaultfd_unmap_ctx *unmap_ctx;
767
768         list_for_each_entry(unmap_ctx, unmaps, list)
769                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
770                     unmap_ctx->end == end)
771                         return true;
772
773         return false;
774 }
775
776 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
777                            unsigned long start, unsigned long end,
778                            struct list_head *unmaps)
779 {
780         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
781                 struct userfaultfd_unmap_ctx *unmap_ctx;
782                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
783
784                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
785                     has_unmap_ctx(ctx, unmaps, start, end))
786                         continue;
787
788                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
789                 if (!unmap_ctx)
790                         return -ENOMEM;
791
792                 userfaultfd_ctx_get(ctx);
793                 unmap_ctx->ctx = ctx;
794                 unmap_ctx->start = start;
795                 unmap_ctx->end = end;
796                 list_add_tail(&unmap_ctx->list, unmaps);
797         }
798
799         return 0;
800 }
801
802 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
803 {
804         struct userfaultfd_unmap_ctx *ctx, *n;
805         struct userfaultfd_wait_queue ewq;
806
807         list_for_each_entry_safe(ctx, n, uf, list) {
808                 msg_init(&ewq.msg);
809
810                 ewq.msg.event = UFFD_EVENT_UNMAP;
811                 ewq.msg.arg.remove.start = ctx->start;
812                 ewq.msg.arg.remove.end = ctx->end;
813
814                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
815
816                 list_del(&ctx->list);
817                 kfree(ctx);
818         }
819 }
820
821 static int userfaultfd_release(struct inode *inode, struct file *file)
822 {
823         struct userfaultfd_ctx *ctx = file->private_data;
824         struct mm_struct *mm = ctx->mm;
825         struct vm_area_struct *vma, *prev;
826         /* len == 0 means wake all */
827         struct userfaultfd_wake_range range = { .len = 0, };
828         unsigned long new_flags;
829
830         ACCESS_ONCE(ctx->released) = true;
831
832         if (!mmget_not_zero(mm))
833                 goto wakeup;
834
835         /*
836          * Flush page faults out of all CPUs. NOTE: all page faults
837          * must be retried without returning VM_FAULT_SIGBUS if
838          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
839          * changes while handle_userfault released the mmap_sem. So
840          * it's critical that released is set to true (above), before
841          * taking the mmap_sem for writing.
842          */
843         down_write(&mm->mmap_sem);
844         prev = NULL;
845         for (vma = mm->mmap; vma; vma = vma->vm_next) {
846                 cond_resched();
847                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
848                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
849                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
850                         prev = vma;
851                         continue;
852                 }
853                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
854                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
855                                  new_flags, vma->anon_vma,
856                                  vma->vm_file, vma->vm_pgoff,
857                                  vma_policy(vma),
858                                  NULL_VM_UFFD_CTX);
859                 if (prev)
860                         vma = prev;
861                 else
862                         prev = vma;
863                 vma->vm_flags = new_flags;
864                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
865         }
866         up_write(&mm->mmap_sem);
867         mmput(mm);
868 wakeup:
869         /*
870          * After no new page faults can wait on this fault_*wqh, flush
871          * the last page faults that may have been already waiting on
872          * the fault_*wqh.
873          */
874         spin_lock(&ctx->fault_pending_wqh.lock);
875         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
876         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
877         spin_unlock(&ctx->fault_pending_wqh.lock);
878
879         /* Flush pending events that may still wait on event_wqh */
880         wake_up_all(&ctx->event_wqh);
881
882         wake_up_poll(&ctx->fd_wqh, POLLHUP);
883         userfaultfd_ctx_put(ctx);
884         return 0;
885 }
886
887 /* fault_pending_wqh.lock must be hold by the caller */
888 static inline struct userfaultfd_wait_queue *find_userfault_in(
889                 wait_queue_head_t *wqh)
890 {
891         wait_queue_entry_t *wq;
892         struct userfaultfd_wait_queue *uwq;
893
894         VM_BUG_ON(!spin_is_locked(&wqh->lock));
895
896         uwq = NULL;
897         if (!waitqueue_active(wqh))
898                 goto out;
899         /* walk in reverse to provide FIFO behavior to read userfaults */
900         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
901         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
902 out:
903         return uwq;
904 }
905
906 static inline struct userfaultfd_wait_queue *find_userfault(
907                 struct userfaultfd_ctx *ctx)
908 {
909         return find_userfault_in(&ctx->fault_pending_wqh);
910 }
911
912 static inline struct userfaultfd_wait_queue *find_userfault_evt(
913                 struct userfaultfd_ctx *ctx)
914 {
915         return find_userfault_in(&ctx->event_wqh);
916 }
917
918 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
919 {
920         struct userfaultfd_ctx *ctx = file->private_data;
921         unsigned int ret;
922
923         poll_wait(file, &ctx->fd_wqh, wait);
924
925         switch (ctx->state) {
926         case UFFD_STATE_WAIT_API:
927                 return POLLERR;
928         case UFFD_STATE_RUNNING:
929                 /*
930                  * poll() never guarantees that read won't block.
931                  * userfaults can be waken before they're read().
932                  */
933                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
934                         return POLLERR;
935                 /*
936                  * lockless access to see if there are pending faults
937                  * __pollwait last action is the add_wait_queue but
938                  * the spin_unlock would allow the waitqueue_active to
939                  * pass above the actual list_add inside
940                  * add_wait_queue critical section. So use a full
941                  * memory barrier to serialize the list_add write of
942                  * add_wait_queue() with the waitqueue_active read
943                  * below.
944                  */
945                 ret = 0;
946                 smp_mb();
947                 if (waitqueue_active(&ctx->fault_pending_wqh))
948                         ret = POLLIN;
949                 else if (waitqueue_active(&ctx->event_wqh))
950                         ret = POLLIN;
951
952                 return ret;
953         default:
954                 WARN_ON_ONCE(1);
955                 return POLLERR;
956         }
957 }
958
959 static const struct file_operations userfaultfd_fops;
960
961 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
962                                   struct userfaultfd_ctx *new,
963                                   struct uffd_msg *msg)
964 {
965         int fd;
966         struct file *file;
967         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
968
969         fd = get_unused_fd_flags(flags);
970         if (fd < 0)
971                 return fd;
972
973         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
974                                   O_RDWR | flags);
975         if (IS_ERR(file)) {
976                 put_unused_fd(fd);
977                 return PTR_ERR(file);
978         }
979
980         fd_install(fd, file);
981         msg->arg.reserved.reserved1 = 0;
982         msg->arg.fork.ufd = fd;
983
984         return 0;
985 }
986
987 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
988                                     struct uffd_msg *msg)
989 {
990         ssize_t ret;
991         DECLARE_WAITQUEUE(wait, current);
992         struct userfaultfd_wait_queue *uwq;
993         /*
994          * Handling fork event requires sleeping operations, so
995          * we drop the event_wqh lock, then do these ops, then
996          * lock it back and wake up the waiter. While the lock is
997          * dropped the ewq may go away so we keep track of it
998          * carefully.
999          */
1000         LIST_HEAD(fork_event);
1001         struct userfaultfd_ctx *fork_nctx = NULL;
1002
1003         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1004         spin_lock(&ctx->fd_wqh.lock);
1005         __add_wait_queue(&ctx->fd_wqh, &wait);
1006         for (;;) {
1007                 set_current_state(TASK_INTERRUPTIBLE);
1008                 spin_lock(&ctx->fault_pending_wqh.lock);
1009                 uwq = find_userfault(ctx);
1010                 if (uwq) {
1011                         /*
1012                          * Use a seqcount to repeat the lockless check
1013                          * in wake_userfault() to avoid missing
1014                          * wakeups because during the refile both
1015                          * waitqueue could become empty if this is the
1016                          * only userfault.
1017                          */
1018                         write_seqcount_begin(&ctx->refile_seq);
1019
1020                         /*
1021                          * The fault_pending_wqh.lock prevents the uwq
1022                          * to disappear from under us.
1023                          *
1024                          * Refile this userfault from
1025                          * fault_pending_wqh to fault_wqh, it's not
1026                          * pending anymore after we read it.
1027                          *
1028                          * Use list_del() by hand (as
1029                          * userfaultfd_wake_function also uses
1030                          * list_del_init() by hand) to be sure nobody
1031                          * changes __remove_wait_queue() to use
1032                          * list_del_init() in turn breaking the
1033                          * !list_empty_careful() check in
1034                          * handle_userfault(). The uwq->wq.head list
1035                          * must never be empty at any time during the
1036                          * refile, or the waitqueue could disappear
1037                          * from under us. The "wait_queue_head_t"
1038                          * parameter of __remove_wait_queue() is unused
1039                          * anyway.
1040                          */
1041                         list_del(&uwq->wq.entry);
1042                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1043
1044                         write_seqcount_end(&ctx->refile_seq);
1045
1046                         /* careful to always initialize msg if ret == 0 */
1047                         *msg = uwq->msg;
1048                         spin_unlock(&ctx->fault_pending_wqh.lock);
1049                         ret = 0;
1050                         break;
1051                 }
1052                 spin_unlock(&ctx->fault_pending_wqh.lock);
1053
1054                 spin_lock(&ctx->event_wqh.lock);
1055                 uwq = find_userfault_evt(ctx);
1056                 if (uwq) {
1057                         *msg = uwq->msg;
1058
1059                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1060                                 fork_nctx = (struct userfaultfd_ctx *)
1061                                         (unsigned long)
1062                                         uwq->msg.arg.reserved.reserved1;
1063                                 list_move(&uwq->wq.entry, &fork_event);
1064                                 spin_unlock(&ctx->event_wqh.lock);
1065                                 ret = 0;
1066                                 break;
1067                         }
1068
1069                         userfaultfd_event_complete(ctx, uwq);
1070                         spin_unlock(&ctx->event_wqh.lock);
1071                         ret = 0;
1072                         break;
1073                 }
1074                 spin_unlock(&ctx->event_wqh.lock);
1075
1076                 if (signal_pending(current)) {
1077                         ret = -ERESTARTSYS;
1078                         break;
1079                 }
1080                 if (no_wait) {
1081                         ret = -EAGAIN;
1082                         break;
1083                 }
1084                 spin_unlock(&ctx->fd_wqh.lock);
1085                 schedule();
1086                 spin_lock(&ctx->fd_wqh.lock);
1087         }
1088         __remove_wait_queue(&ctx->fd_wqh, &wait);
1089         __set_current_state(TASK_RUNNING);
1090         spin_unlock(&ctx->fd_wqh.lock);
1091
1092         if (!ret && msg->event == UFFD_EVENT_FORK) {
1093                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1094
1095                 if (!ret) {
1096                         spin_lock(&ctx->event_wqh.lock);
1097                         if (!list_empty(&fork_event)) {
1098                                 uwq = list_first_entry(&fork_event,
1099                                                        typeof(*uwq),
1100                                                        wq.entry);
1101                                 list_del(&uwq->wq.entry);
1102                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1103                                 userfaultfd_event_complete(ctx, uwq);
1104                         }
1105                         spin_unlock(&ctx->event_wqh.lock);
1106                 }
1107         }
1108
1109         return ret;
1110 }
1111
1112 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1113                                 size_t count, loff_t *ppos)
1114 {
1115         struct userfaultfd_ctx *ctx = file->private_data;
1116         ssize_t _ret, ret = 0;
1117         struct uffd_msg msg;
1118         int no_wait = file->f_flags & O_NONBLOCK;
1119
1120         if (ctx->state == UFFD_STATE_WAIT_API)
1121                 return -EINVAL;
1122
1123         for (;;) {
1124                 if (count < sizeof(msg))
1125                         return ret ? ret : -EINVAL;
1126                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1127                 if (_ret < 0)
1128                         return ret ? ret : _ret;
1129                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1130                         return ret ? ret : -EFAULT;
1131                 ret += sizeof(msg);
1132                 buf += sizeof(msg);
1133                 count -= sizeof(msg);
1134                 /*
1135                  * Allow to read more than one fault at time but only
1136                  * block if waiting for the very first one.
1137                  */
1138                 no_wait = O_NONBLOCK;
1139         }
1140 }
1141
1142 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1143                              struct userfaultfd_wake_range *range)
1144 {
1145         spin_lock(&ctx->fault_pending_wqh.lock);
1146         /* wake all in the range and autoremove */
1147         if (waitqueue_active(&ctx->fault_pending_wqh))
1148                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1149                                      range);
1150         if (waitqueue_active(&ctx->fault_wqh))
1151                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1152         spin_unlock(&ctx->fault_pending_wqh.lock);
1153 }
1154
1155 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1156                                            struct userfaultfd_wake_range *range)
1157 {
1158         unsigned seq;
1159         bool need_wakeup;
1160
1161         /*
1162          * To be sure waitqueue_active() is not reordered by the CPU
1163          * before the pagetable update, use an explicit SMP memory
1164          * barrier here. PT lock release or up_read(mmap_sem) still
1165          * have release semantics that can allow the
1166          * waitqueue_active() to be reordered before the pte update.
1167          */
1168         smp_mb();
1169
1170         /*
1171          * Use waitqueue_active because it's very frequent to
1172          * change the address space atomically even if there are no
1173          * userfaults yet. So we take the spinlock only when we're
1174          * sure we've userfaults to wake.
1175          */
1176         do {
1177                 seq = read_seqcount_begin(&ctx->refile_seq);
1178                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1179                         waitqueue_active(&ctx->fault_wqh);
1180                 cond_resched();
1181         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1182         if (need_wakeup)
1183                 __wake_userfault(ctx, range);
1184 }
1185
1186 static __always_inline int validate_range(struct mm_struct *mm,
1187                                           __u64 start, __u64 len)
1188 {
1189         __u64 task_size = mm->task_size;
1190
1191         if (start & ~PAGE_MASK)
1192                 return -EINVAL;
1193         if (len & ~PAGE_MASK)
1194                 return -EINVAL;
1195         if (!len)
1196                 return -EINVAL;
1197         if (start < mmap_min_addr)
1198                 return -EINVAL;
1199         if (start >= task_size)
1200                 return -EINVAL;
1201         if (len > task_size - start)
1202                 return -EINVAL;
1203         return 0;
1204 }
1205
1206 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1207 {
1208         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1209                 vma_is_shmem(vma);
1210 }
1211
1212 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1213                                 unsigned long arg)
1214 {
1215         struct mm_struct *mm = ctx->mm;
1216         struct vm_area_struct *vma, *prev, *cur;
1217         int ret;
1218         struct uffdio_register uffdio_register;
1219         struct uffdio_register __user *user_uffdio_register;
1220         unsigned long vm_flags, new_flags;
1221         bool found;
1222         bool basic_ioctls;
1223         unsigned long start, end, vma_end;
1224
1225         user_uffdio_register = (struct uffdio_register __user *) arg;
1226
1227         ret = -EFAULT;
1228         if (copy_from_user(&uffdio_register, user_uffdio_register,
1229                            sizeof(uffdio_register)-sizeof(__u64)))
1230                 goto out;
1231
1232         ret = -EINVAL;
1233         if (!uffdio_register.mode)
1234                 goto out;
1235         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1236                                      UFFDIO_REGISTER_MODE_WP))
1237                 goto out;
1238         vm_flags = 0;
1239         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1240                 vm_flags |= VM_UFFD_MISSING;
1241         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1242                 vm_flags |= VM_UFFD_WP;
1243                 /*
1244                  * FIXME: remove the below error constraint by
1245                  * implementing the wprotect tracking mode.
1246                  */
1247                 ret = -EINVAL;
1248                 goto out;
1249         }
1250
1251         ret = validate_range(mm, uffdio_register.range.start,
1252                              uffdio_register.range.len);
1253         if (ret)
1254                 goto out;
1255
1256         start = uffdio_register.range.start;
1257         end = start + uffdio_register.range.len;
1258
1259         ret = -ENOMEM;
1260         if (!mmget_not_zero(mm))
1261                 goto out;
1262
1263         down_write(&mm->mmap_sem);
1264         vma = find_vma_prev(mm, start, &prev);
1265         if (!vma)
1266                 goto out_unlock;
1267
1268         /* check that there's at least one vma in the range */
1269         ret = -EINVAL;
1270         if (vma->vm_start >= end)
1271                 goto out_unlock;
1272
1273         /*
1274          * If the first vma contains huge pages, make sure start address
1275          * is aligned to huge page size.
1276          */
1277         if (is_vm_hugetlb_page(vma)) {
1278                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1279
1280                 if (start & (vma_hpagesize - 1))
1281                         goto out_unlock;
1282         }
1283
1284         /*
1285          * Search for not compatible vmas.
1286          */
1287         found = false;
1288         basic_ioctls = false;
1289         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1290                 cond_resched();
1291
1292                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1293                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1294
1295                 /* check not compatible vmas */
1296                 ret = -EINVAL;
1297                 if (!vma_can_userfault(cur))
1298                         goto out_unlock;
1299                 /*
1300                  * If this vma contains ending address, and huge pages
1301                  * check alignment.
1302                  */
1303                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1304                     end > cur->vm_start) {
1305                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1306
1307                         ret = -EINVAL;
1308
1309                         if (end & (vma_hpagesize - 1))
1310                                 goto out_unlock;
1311                 }
1312
1313                 /*
1314                  * Check that this vma isn't already owned by a
1315                  * different userfaultfd. We can't allow more than one
1316                  * userfaultfd to own a single vma simultaneously or we
1317                  * wouldn't know which one to deliver the userfaults to.
1318                  */
1319                 ret = -EBUSY;
1320                 if (cur->vm_userfaultfd_ctx.ctx &&
1321                     cur->vm_userfaultfd_ctx.ctx != ctx)
1322                         goto out_unlock;
1323
1324                 /*
1325                  * Note vmas containing huge pages
1326                  */
1327                 if (is_vm_hugetlb_page(cur))
1328                         basic_ioctls = true;
1329
1330                 found = true;
1331         }
1332         BUG_ON(!found);
1333
1334         if (vma->vm_start < start)
1335                 prev = vma;
1336
1337         ret = 0;
1338         do {
1339                 cond_resched();
1340
1341                 BUG_ON(!vma_can_userfault(vma));
1342                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1343                        vma->vm_userfaultfd_ctx.ctx != ctx);
1344
1345                 /*
1346                  * Nothing to do: this vma is already registered into this
1347                  * userfaultfd and with the right tracking mode too.
1348                  */
1349                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1350                     (vma->vm_flags & vm_flags) == vm_flags)
1351                         goto skip;
1352
1353                 if (vma->vm_start > start)
1354                         start = vma->vm_start;
1355                 vma_end = min(end, vma->vm_end);
1356
1357                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1358                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1359                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1360                                  vma_policy(vma),
1361                                  ((struct vm_userfaultfd_ctx){ ctx }));
1362                 if (prev) {
1363                         vma = prev;
1364                         goto next;
1365                 }
1366                 if (vma->vm_start < start) {
1367                         ret = split_vma(mm, vma, start, 1);
1368                         if (ret)
1369                                 break;
1370                 }
1371                 if (vma->vm_end > end) {
1372                         ret = split_vma(mm, vma, end, 0);
1373                         if (ret)
1374                                 break;
1375                 }
1376         next:
1377                 /*
1378                  * In the vma_merge() successful mprotect-like case 8:
1379                  * the next vma was merged into the current one and
1380                  * the current one has not been updated yet.
1381                  */
1382                 vma->vm_flags = new_flags;
1383                 vma->vm_userfaultfd_ctx.ctx = ctx;
1384
1385         skip:
1386                 prev = vma;
1387                 start = vma->vm_end;
1388                 vma = vma->vm_next;
1389         } while (vma && vma->vm_start < end);
1390 out_unlock:
1391         up_write(&mm->mmap_sem);
1392         mmput(mm);
1393         if (!ret) {
1394                 /*
1395                  * Now that we scanned all vmas we can already tell
1396                  * userland which ioctls methods are guaranteed to
1397                  * succeed on this range.
1398                  */
1399                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1400                              UFFD_API_RANGE_IOCTLS,
1401                              &user_uffdio_register->ioctls))
1402                         ret = -EFAULT;
1403         }
1404 out:
1405         return ret;
1406 }
1407
1408 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1409                                   unsigned long arg)
1410 {
1411         struct mm_struct *mm = ctx->mm;
1412         struct vm_area_struct *vma, *prev, *cur;
1413         int ret;
1414         struct uffdio_range uffdio_unregister;
1415         unsigned long new_flags;
1416         bool found;
1417         unsigned long start, end, vma_end;
1418         const void __user *buf = (void __user *)arg;
1419
1420         ret = -EFAULT;
1421         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1422                 goto out;
1423
1424         ret = validate_range(mm, uffdio_unregister.start,
1425                              uffdio_unregister.len);
1426         if (ret)
1427                 goto out;
1428
1429         start = uffdio_unregister.start;
1430         end = start + uffdio_unregister.len;
1431
1432         ret = -ENOMEM;
1433         if (!mmget_not_zero(mm))
1434                 goto out;
1435
1436         down_write(&mm->mmap_sem);
1437         vma = find_vma_prev(mm, start, &prev);
1438         if (!vma)
1439                 goto out_unlock;
1440
1441         /* check that there's at least one vma in the range */
1442         ret = -EINVAL;
1443         if (vma->vm_start >= end)
1444                 goto out_unlock;
1445
1446         /*
1447          * If the first vma contains huge pages, make sure start address
1448          * is aligned to huge page size.
1449          */
1450         if (is_vm_hugetlb_page(vma)) {
1451                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1452
1453                 if (start & (vma_hpagesize - 1))
1454                         goto out_unlock;
1455         }
1456
1457         /*
1458          * Search for not compatible vmas.
1459          */
1460         found = false;
1461         ret = -EINVAL;
1462         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1463                 cond_resched();
1464
1465                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1466                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1467
1468                 /*
1469                  * Check not compatible vmas, not strictly required
1470                  * here as not compatible vmas cannot have an
1471                  * userfaultfd_ctx registered on them, but this
1472                  * provides for more strict behavior to notice
1473                  * unregistration errors.
1474                  */
1475                 if (!vma_can_userfault(cur))
1476                         goto out_unlock;
1477
1478                 found = true;
1479         }
1480         BUG_ON(!found);
1481
1482         if (vma->vm_start < start)
1483                 prev = vma;
1484
1485         ret = 0;
1486         do {
1487                 cond_resched();
1488
1489                 BUG_ON(!vma_can_userfault(vma));
1490
1491                 /*
1492                  * Nothing to do: this vma is already registered into this
1493                  * userfaultfd and with the right tracking mode too.
1494                  */
1495                 if (!vma->vm_userfaultfd_ctx.ctx)
1496                         goto skip;
1497
1498                 if (vma->vm_start > start)
1499                         start = vma->vm_start;
1500                 vma_end = min(end, vma->vm_end);
1501
1502                 if (userfaultfd_missing(vma)) {
1503                         /*
1504                          * Wake any concurrent pending userfault while
1505                          * we unregister, so they will not hang
1506                          * permanently and it avoids userland to call
1507                          * UFFDIO_WAKE explicitly.
1508                          */
1509                         struct userfaultfd_wake_range range;
1510                         range.start = start;
1511                         range.len = vma_end - start;
1512                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1513                 }
1514
1515                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1516                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1517                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1518                                  vma_policy(vma),
1519                                  NULL_VM_UFFD_CTX);
1520                 if (prev) {
1521                         vma = prev;
1522                         goto next;
1523                 }
1524                 if (vma->vm_start < start) {
1525                         ret = split_vma(mm, vma, start, 1);
1526                         if (ret)
1527                                 break;
1528                 }
1529                 if (vma->vm_end > end) {
1530                         ret = split_vma(mm, vma, end, 0);
1531                         if (ret)
1532                                 break;
1533                 }
1534         next:
1535                 /*
1536                  * In the vma_merge() successful mprotect-like case 8:
1537                  * the next vma was merged into the current one and
1538                  * the current one has not been updated yet.
1539                  */
1540                 vma->vm_flags = new_flags;
1541                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1542
1543         skip:
1544                 prev = vma;
1545                 start = vma->vm_end;
1546                 vma = vma->vm_next;
1547         } while (vma && vma->vm_start < end);
1548 out_unlock:
1549         up_write(&mm->mmap_sem);
1550         mmput(mm);
1551 out:
1552         return ret;
1553 }
1554
1555 /*
1556  * userfaultfd_wake may be used in combination with the
1557  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1558  */
1559 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1560                             unsigned long arg)
1561 {
1562         int ret;
1563         struct uffdio_range uffdio_wake;
1564         struct userfaultfd_wake_range range;
1565         const void __user *buf = (void __user *)arg;
1566
1567         ret = -EFAULT;
1568         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1569                 goto out;
1570
1571         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1572         if (ret)
1573                 goto out;
1574
1575         range.start = uffdio_wake.start;
1576         range.len = uffdio_wake.len;
1577
1578         /*
1579          * len == 0 means wake all and we don't want to wake all here,
1580          * so check it again to be sure.
1581          */
1582         VM_BUG_ON(!range.len);
1583
1584         wake_userfault(ctx, &range);
1585         ret = 0;
1586
1587 out:
1588         return ret;
1589 }
1590
1591 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1592                             unsigned long arg)
1593 {
1594         __s64 ret;
1595         struct uffdio_copy uffdio_copy;
1596         struct uffdio_copy __user *user_uffdio_copy;
1597         struct userfaultfd_wake_range range;
1598
1599         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1600
1601         ret = -EFAULT;
1602         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1603                            /* don't copy "copy" last field */
1604                            sizeof(uffdio_copy)-sizeof(__s64)))
1605                 goto out;
1606
1607         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1608         if (ret)
1609                 goto out;
1610         /*
1611          * double check for wraparound just in case. copy_from_user()
1612          * will later check uffdio_copy.src + uffdio_copy.len to fit
1613          * in the userland range.
1614          */
1615         ret = -EINVAL;
1616         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1617                 goto out;
1618         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1619                 goto out;
1620         if (mmget_not_zero(ctx->mm)) {
1621                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1622                                    uffdio_copy.len);
1623                 mmput(ctx->mm);
1624         } else {
1625                 return -ESRCH;
1626         }
1627         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1628                 return -EFAULT;
1629         if (ret < 0)
1630                 goto out;
1631         BUG_ON(!ret);
1632         /* len == 0 would wake all */
1633         range.len = ret;
1634         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1635                 range.start = uffdio_copy.dst;
1636                 wake_userfault(ctx, &range);
1637         }
1638         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1639 out:
1640         return ret;
1641 }
1642
1643 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1644                                 unsigned long arg)
1645 {
1646         __s64 ret;
1647         struct uffdio_zeropage uffdio_zeropage;
1648         struct uffdio_zeropage __user *user_uffdio_zeropage;
1649         struct userfaultfd_wake_range range;
1650
1651         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1652
1653         ret = -EFAULT;
1654         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1655                            /* don't copy "zeropage" last field */
1656                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1657                 goto out;
1658
1659         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1660                              uffdio_zeropage.range.len);
1661         if (ret)
1662                 goto out;
1663         ret = -EINVAL;
1664         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1665                 goto out;
1666
1667         if (mmget_not_zero(ctx->mm)) {
1668                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1669                                      uffdio_zeropage.range.len);
1670                 mmput(ctx->mm);
1671         } else {
1672                 return -ESRCH;
1673         }
1674         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1675                 return -EFAULT;
1676         if (ret < 0)
1677                 goto out;
1678         /* len == 0 would wake all */
1679         BUG_ON(!ret);
1680         range.len = ret;
1681         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1682                 range.start = uffdio_zeropage.range.start;
1683                 wake_userfault(ctx, &range);
1684         }
1685         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1686 out:
1687         return ret;
1688 }
1689
1690 static inline unsigned int uffd_ctx_features(__u64 user_features)
1691 {
1692         /*
1693          * For the current set of features the bits just coincide
1694          */
1695         return (unsigned int)user_features;
1696 }
1697
1698 /*
1699  * userland asks for a certain API version and we return which bits
1700  * and ioctl commands are implemented in this kernel for such API
1701  * version or -EINVAL if unknown.
1702  */
1703 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1704                            unsigned long arg)
1705 {
1706         struct uffdio_api uffdio_api;
1707         void __user *buf = (void __user *)arg;
1708         int ret;
1709         __u64 features;
1710
1711         ret = -EINVAL;
1712         if (ctx->state != UFFD_STATE_WAIT_API)
1713                 goto out;
1714         ret = -EFAULT;
1715         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1716                 goto out;
1717         features = uffdio_api.features;
1718         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1719                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1720                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1721                         goto out;
1722                 ret = -EINVAL;
1723                 goto out;
1724         }
1725         /* report all available features and ioctls to userland */
1726         uffdio_api.features = UFFD_API_FEATURES;
1727         uffdio_api.ioctls = UFFD_API_IOCTLS;
1728         ret = -EFAULT;
1729         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1730                 goto out;
1731         ctx->state = UFFD_STATE_RUNNING;
1732         /* only enable the requested features for this uffd context */
1733         ctx->features = uffd_ctx_features(features);
1734         ret = 0;
1735 out:
1736         return ret;
1737 }
1738
1739 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1740                               unsigned long arg)
1741 {
1742         int ret = -EINVAL;
1743         struct userfaultfd_ctx *ctx = file->private_data;
1744
1745         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1746                 return -EINVAL;
1747
1748         switch(cmd) {
1749         case UFFDIO_API:
1750                 ret = userfaultfd_api(ctx, arg);
1751                 break;
1752         case UFFDIO_REGISTER:
1753                 ret = userfaultfd_register(ctx, arg);
1754                 break;
1755         case UFFDIO_UNREGISTER:
1756                 ret = userfaultfd_unregister(ctx, arg);
1757                 break;
1758         case UFFDIO_WAKE:
1759                 ret = userfaultfd_wake(ctx, arg);
1760                 break;
1761         case UFFDIO_COPY:
1762                 ret = userfaultfd_copy(ctx, arg);
1763                 break;
1764         case UFFDIO_ZEROPAGE:
1765                 ret = userfaultfd_zeropage(ctx, arg);
1766                 break;
1767         }
1768         return ret;
1769 }
1770
1771 #ifdef CONFIG_PROC_FS
1772 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1773 {
1774         struct userfaultfd_ctx *ctx = f->private_data;
1775         wait_queue_entry_t *wq;
1776         struct userfaultfd_wait_queue *uwq;
1777         unsigned long pending = 0, total = 0;
1778
1779         spin_lock(&ctx->fault_pending_wqh.lock);
1780         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1781                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1782                 pending++;
1783                 total++;
1784         }
1785         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1786                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1787                 total++;
1788         }
1789         spin_unlock(&ctx->fault_pending_wqh.lock);
1790
1791         /*
1792          * If more protocols will be added, there will be all shown
1793          * separated by a space. Like this:
1794          *      protocols: aa:... bb:...
1795          */
1796         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1797                    pending, total, UFFD_API, ctx->features,
1798                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1799 }
1800 #endif
1801
1802 static const struct file_operations userfaultfd_fops = {
1803 #ifdef CONFIG_PROC_FS
1804         .show_fdinfo    = userfaultfd_show_fdinfo,
1805 #endif
1806         .release        = userfaultfd_release,
1807         .poll           = userfaultfd_poll,
1808         .read           = userfaultfd_read,
1809         .unlocked_ioctl = userfaultfd_ioctl,
1810         .compat_ioctl   = userfaultfd_ioctl,
1811         .llseek         = noop_llseek,
1812 };
1813
1814 static void init_once_userfaultfd_ctx(void *mem)
1815 {
1816         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1817
1818         init_waitqueue_head(&ctx->fault_pending_wqh);
1819         init_waitqueue_head(&ctx->fault_wqh);
1820         init_waitqueue_head(&ctx->event_wqh);
1821         init_waitqueue_head(&ctx->fd_wqh);
1822         seqcount_init(&ctx->refile_seq);
1823 }
1824
1825 /**
1826  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1827  * @flags: Flags for the userfaultfd file.
1828  *
1829  * This function creates a userfaultfd file pointer, w/out installing
1830  * it into the fd table. This is useful when the userfaultfd file is
1831  * used during the initialization of data structures that require
1832  * extra setup after the userfaultfd creation. So the userfaultfd
1833  * creation is split into the file pointer creation phase, and the
1834  * file descriptor installation phase.  In this way races with
1835  * userspace closing the newly installed file descriptor can be
1836  * avoided.  Returns a userfaultfd file pointer, or a proper error
1837  * pointer.
1838  */
1839 static struct file *userfaultfd_file_create(int flags)
1840 {
1841         struct file *file;
1842         struct userfaultfd_ctx *ctx;
1843
1844         BUG_ON(!current->mm);
1845
1846         /* Check the UFFD_* constants for consistency.  */
1847         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1848         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1849
1850         file = ERR_PTR(-EINVAL);
1851         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1852                 goto out;
1853
1854         file = ERR_PTR(-ENOMEM);
1855         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1856         if (!ctx)
1857                 goto out;
1858
1859         atomic_set(&ctx->refcount, 1);
1860         ctx->flags = flags;
1861         ctx->features = 0;
1862         ctx->state = UFFD_STATE_WAIT_API;
1863         ctx->released = false;
1864         ctx->mm = current->mm;
1865         /* prevent the mm struct to be freed */
1866         mmgrab(ctx->mm);
1867
1868         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1869                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1870         if (IS_ERR(file)) {
1871                 mmdrop(ctx->mm);
1872                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1873         }
1874 out:
1875         return file;
1876 }
1877
1878 SYSCALL_DEFINE1(userfaultfd, int, flags)
1879 {
1880         int fd, error;
1881         struct file *file;
1882
1883         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1884         if (error < 0)
1885                 return error;
1886         fd = error;
1887
1888         file = userfaultfd_file_create(flags);
1889         if (IS_ERR(file)) {
1890                 error = PTR_ERR(file);
1891                 goto err_put_unused_fd;
1892         }
1893         fd_install(fd, file);
1894
1895         return fd;
1896
1897 err_put_unused_fd:
1898         put_unused_fd(fd);
1899
1900         return error;
1901 }
1902
1903 static int __init userfaultfd_init(void)
1904 {
1905         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1906                                                 sizeof(struct userfaultfd_ctx),
1907                                                 0,
1908                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1909                                                 init_once_userfaultfd_ctx);
1910         return 0;
1911 }
1912 __initcall(userfaultfd_init);