Merge tag 'pci-v5.11-fixes-1' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaa...
[linux-2.6-microblaze.git] / fs / userfaultfd.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  fs/userfaultfd.c
4  *
5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
6  *  Copyright (C) 2008-2009 Red Hat, Inc.
7  *  Copyright (C) 2015  Red Hat, Inc.
8  *
9  *  Some part derived from fs/eventfd.c (anon inode setup) and
10  *  mm/ksm.c (mm hashing).
11  */
12
13 #include <linux/list.h>
14 #include <linux/hashtable.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/mm.h>
17 #include <linux/mm.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/seq_file.h>
21 #include <linux/file.h>
22 #include <linux/bug.h>
23 #include <linux/anon_inodes.h>
24 #include <linux/syscalls.h>
25 #include <linux/userfaultfd_k.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ioctl.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
30
31 int sysctl_unprivileged_userfaultfd __read_mostly;
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  * Locking order:
45  *      fd_wqh.lock
46  *              fault_pending_wqh.lock
47  *                      fault_wqh.lock
48  *              event_wqh.lock
49  *
50  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
51  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
52  * also taken in IRQ context.
53  */
54 struct userfaultfd_ctx {
55         /* waitqueue head for the pending (i.e. not read) userfaults */
56         wait_queue_head_t fault_pending_wqh;
57         /* waitqueue head for the userfaults */
58         wait_queue_head_t fault_wqh;
59         /* waitqueue head for the pseudo fd to wakeup poll/read */
60         wait_queue_head_t fd_wqh;
61         /* waitqueue head for events */
62         wait_queue_head_t event_wqh;
63         /* a refile sequence protected by fault_pending_wqh lock */
64         seqcount_spinlock_t refile_seq;
65         /* pseudo fd refcounting */
66         refcount_t refcount;
67         /* userfaultfd syscall flags */
68         unsigned int flags;
69         /* features requested from the userspace */
70         unsigned int features;
71         /* state machine */
72         enum userfaultfd_state state;
73         /* released */
74         bool released;
75         /* memory mappings are changing because of non-cooperative event */
76         bool mmap_changing;
77         /* mm with one ore more vmas attached to this userfaultfd_ctx */
78         struct mm_struct *mm;
79 };
80
81 struct userfaultfd_fork_ctx {
82         struct userfaultfd_ctx *orig;
83         struct userfaultfd_ctx *new;
84         struct list_head list;
85 };
86
87 struct userfaultfd_unmap_ctx {
88         struct userfaultfd_ctx *ctx;
89         unsigned long start;
90         unsigned long end;
91         struct list_head list;
92 };
93
94 struct userfaultfd_wait_queue {
95         struct uffd_msg msg;
96         wait_queue_entry_t wq;
97         struct userfaultfd_ctx *ctx;
98         bool waken;
99 };
100
101 struct userfaultfd_wake_range {
102         unsigned long start;
103         unsigned long len;
104 };
105
106 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
107                                      int wake_flags, void *key)
108 {
109         struct userfaultfd_wake_range *range = key;
110         int ret;
111         struct userfaultfd_wait_queue *uwq;
112         unsigned long start, len;
113
114         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
115         ret = 0;
116         /* len == 0 means wake all */
117         start = range->start;
118         len = range->len;
119         if (len && (start > uwq->msg.arg.pagefault.address ||
120                     start + len <= uwq->msg.arg.pagefault.address))
121                 goto out;
122         WRITE_ONCE(uwq->waken, true);
123         /*
124          * The Program-Order guarantees provided by the scheduler
125          * ensure uwq->waken is visible before the task is woken.
126          */
127         ret = wake_up_state(wq->private, mode);
128         if (ret) {
129                 /*
130                  * Wake only once, autoremove behavior.
131                  *
132                  * After the effect of list_del_init is visible to the other
133                  * CPUs, the waitqueue may disappear from under us, see the
134                  * !list_empty_careful() in handle_userfault().
135                  *
136                  * try_to_wake_up() has an implicit smp_mb(), and the
137                  * wq->private is read before calling the extern function
138                  * "wake_up_state" (which in turns calls try_to_wake_up).
139                  */
140                 list_del_init(&wq->entry);
141         }
142 out:
143         return ret;
144 }
145
146 /**
147  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
148  * context.
149  * @ctx: [in] Pointer to the userfaultfd context.
150  */
151 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
152 {
153         refcount_inc(&ctx->refcount);
154 }
155
156 /**
157  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
158  * context.
159  * @ctx: [in] Pointer to userfaultfd context.
160  *
161  * The userfaultfd context reference must have been previously acquired either
162  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
163  */
164 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
165 {
166         if (refcount_dec_and_test(&ctx->refcount)) {
167                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
168                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
169                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
170                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
171                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
172                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
173                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
174                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
175                 mmdrop(ctx->mm);
176                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
177         }
178 }
179
180 static inline void msg_init(struct uffd_msg *msg)
181 {
182         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
183         /*
184          * Must use memset to zero out the paddings or kernel data is
185          * leaked to userland.
186          */
187         memset(msg, 0, sizeof(struct uffd_msg));
188 }
189
190 static inline struct uffd_msg userfault_msg(unsigned long address,
191                                             unsigned int flags,
192                                             unsigned long reason,
193                                             unsigned int features)
194 {
195         struct uffd_msg msg;
196         msg_init(&msg);
197         msg.event = UFFD_EVENT_PAGEFAULT;
198         msg.arg.pagefault.address = address;
199         if (flags & FAULT_FLAG_WRITE)
200                 /*
201                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
202                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
203                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
204                  * was a read fault, otherwise if set it means it's
205                  * a write fault.
206                  */
207                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
208         if (reason & VM_UFFD_WP)
209                 /*
210                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
211                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
212                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
213                  * a missing fault, otherwise if set it means it's a
214                  * write protect fault.
215                  */
216                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
217         if (features & UFFD_FEATURE_THREAD_ID)
218                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
219         return msg;
220 }
221
222 #ifdef CONFIG_HUGETLB_PAGE
223 /*
224  * Same functionality as userfaultfd_must_wait below with modifications for
225  * hugepmd ranges.
226  */
227 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
228                                          struct vm_area_struct *vma,
229                                          unsigned long address,
230                                          unsigned long flags,
231                                          unsigned long reason)
232 {
233         struct mm_struct *mm = ctx->mm;
234         pte_t *ptep, pte;
235         bool ret = true;
236
237         mmap_assert_locked(mm);
238
239         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
240
241         if (!ptep)
242                 goto out;
243
244         ret = false;
245         pte = huge_ptep_get(ptep);
246
247         /*
248          * Lockless access: we're in a wait_event so it's ok if it
249          * changes under us.
250          */
251         if (huge_pte_none(pte))
252                 ret = true;
253         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
254                 ret = true;
255 out:
256         return ret;
257 }
258 #else
259 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
260                                          struct vm_area_struct *vma,
261                                          unsigned long address,
262                                          unsigned long flags,
263                                          unsigned long reason)
264 {
265         return false;   /* should never get here */
266 }
267 #endif /* CONFIG_HUGETLB_PAGE */
268
269 /*
270  * Verify the pagetables are still not ok after having reigstered into
271  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
272  * userfault that has already been resolved, if userfaultfd_read and
273  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
274  * threads.
275  */
276 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
277                                          unsigned long address,
278                                          unsigned long flags,
279                                          unsigned long reason)
280 {
281         struct mm_struct *mm = ctx->mm;
282         pgd_t *pgd;
283         p4d_t *p4d;
284         pud_t *pud;
285         pmd_t *pmd, _pmd;
286         pte_t *pte;
287         bool ret = true;
288
289         mmap_assert_locked(mm);
290
291         pgd = pgd_offset(mm, address);
292         if (!pgd_present(*pgd))
293                 goto out;
294         p4d = p4d_offset(pgd, address);
295         if (!p4d_present(*p4d))
296                 goto out;
297         pud = pud_offset(p4d, address);
298         if (!pud_present(*pud))
299                 goto out;
300         pmd = pmd_offset(pud, address);
301         /*
302          * READ_ONCE must function as a barrier with narrower scope
303          * and it must be equivalent to:
304          *      _pmd = *pmd; barrier();
305          *
306          * This is to deal with the instability (as in
307          * pmd_trans_unstable) of the pmd.
308          */
309         _pmd = READ_ONCE(*pmd);
310         if (pmd_none(_pmd))
311                 goto out;
312
313         ret = false;
314         if (!pmd_present(_pmd))
315                 goto out;
316
317         if (pmd_trans_huge(_pmd)) {
318                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
319                         ret = true;
320                 goto out;
321         }
322
323         /*
324          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
325          * and use the standard pte_offset_map() instead of parsing _pmd.
326          */
327         pte = pte_offset_map(pmd, address);
328         /*
329          * Lockless access: we're in a wait_event so it's ok if it
330          * changes under us.
331          */
332         if (pte_none(*pte))
333                 ret = true;
334         if (!pte_write(*pte) && (reason & VM_UFFD_WP))
335                 ret = true;
336         pte_unmap(pte);
337
338 out:
339         return ret;
340 }
341
342 static inline long userfaultfd_get_blocking_state(unsigned int flags)
343 {
344         if (flags & FAULT_FLAG_INTERRUPTIBLE)
345                 return TASK_INTERRUPTIBLE;
346
347         if (flags & FAULT_FLAG_KILLABLE)
348                 return TASK_KILLABLE;
349
350         return TASK_UNINTERRUPTIBLE;
351 }
352
353 /*
354  * The locking rules involved in returning VM_FAULT_RETRY depending on
355  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
356  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
357  * recommendation in __lock_page_or_retry is not an understatement.
358  *
359  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
360  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
361  * not set.
362  *
363  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
364  * set, VM_FAULT_RETRY can still be returned if and only if there are
365  * fatal_signal_pending()s, and the mmap_lock must be released before
366  * returning it.
367  */
368 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
369 {
370         struct mm_struct *mm = vmf->vma->vm_mm;
371         struct userfaultfd_ctx *ctx;
372         struct userfaultfd_wait_queue uwq;
373         vm_fault_t ret = VM_FAULT_SIGBUS;
374         bool must_wait;
375         long blocking_state;
376
377         /*
378          * We don't do userfault handling for the final child pid update.
379          *
380          * We also don't do userfault handling during
381          * coredumping. hugetlbfs has the special
382          * follow_hugetlb_page() to skip missing pages in the
383          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
384          * the no_page_table() helper in follow_page_mask(), but the
385          * shmem_vm_ops->fault method is invoked even during
386          * coredumping without mmap_lock and it ends up here.
387          */
388         if (current->flags & (PF_EXITING|PF_DUMPCORE))
389                 goto out;
390
391         /*
392          * Coredumping runs without mmap_lock so we can only check that
393          * the mmap_lock is held, if PF_DUMPCORE was not set.
394          */
395         mmap_assert_locked(mm);
396
397         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
398         if (!ctx)
399                 goto out;
400
401         BUG_ON(ctx->mm != mm);
402
403         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
404         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
405
406         if (ctx->features & UFFD_FEATURE_SIGBUS)
407                 goto out;
408         if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
409             ctx->flags & UFFD_USER_MODE_ONLY) {
410                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
411                         "sysctl knob to 1 if kernel faults must be handled "
412                         "without obtaining CAP_SYS_PTRACE capability\n");
413                 goto out;
414         }
415
416         /*
417          * If it's already released don't get it. This avoids to loop
418          * in __get_user_pages if userfaultfd_release waits on the
419          * caller of handle_userfault to release the mmap_lock.
420          */
421         if (unlikely(READ_ONCE(ctx->released))) {
422                 /*
423                  * Don't return VM_FAULT_SIGBUS in this case, so a non
424                  * cooperative manager can close the uffd after the
425                  * last UFFDIO_COPY, without risking to trigger an
426                  * involuntary SIGBUS if the process was starting the
427                  * userfaultfd while the userfaultfd was still armed
428                  * (but after the last UFFDIO_COPY). If the uffd
429                  * wasn't already closed when the userfault reached
430                  * this point, that would normally be solved by
431                  * userfaultfd_must_wait returning 'false'.
432                  *
433                  * If we were to return VM_FAULT_SIGBUS here, the non
434                  * cooperative manager would be instead forced to
435                  * always call UFFDIO_UNREGISTER before it can safely
436                  * close the uffd.
437                  */
438                 ret = VM_FAULT_NOPAGE;
439                 goto out;
440         }
441
442         /*
443          * Check that we can return VM_FAULT_RETRY.
444          *
445          * NOTE: it should become possible to return VM_FAULT_RETRY
446          * even if FAULT_FLAG_TRIED is set without leading to gup()
447          * -EBUSY failures, if the userfaultfd is to be extended for
448          * VM_UFFD_WP tracking and we intend to arm the userfault
449          * without first stopping userland access to the memory. For
450          * VM_UFFD_MISSING userfaults this is enough for now.
451          */
452         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
453                 /*
454                  * Validate the invariant that nowait must allow retry
455                  * to be sure not to return SIGBUS erroneously on
456                  * nowait invocations.
457                  */
458                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
459 #ifdef CONFIG_DEBUG_VM
460                 if (printk_ratelimit()) {
461                         printk(KERN_WARNING
462                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
463                                vmf->flags);
464                         dump_stack();
465                 }
466 #endif
467                 goto out;
468         }
469
470         /*
471          * Handle nowait, not much to do other than tell it to retry
472          * and wait.
473          */
474         ret = VM_FAULT_RETRY;
475         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
476                 goto out;
477
478         /* take the reference before dropping the mmap_lock */
479         userfaultfd_ctx_get(ctx);
480
481         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
482         uwq.wq.private = current;
483         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
484                         ctx->features);
485         uwq.ctx = ctx;
486         uwq.waken = false;
487
488         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
489
490         spin_lock_irq(&ctx->fault_pending_wqh.lock);
491         /*
492          * After the __add_wait_queue the uwq is visible to userland
493          * through poll/read().
494          */
495         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
496         /*
497          * The smp_mb() after __set_current_state prevents the reads
498          * following the spin_unlock to happen before the list_add in
499          * __add_wait_queue.
500          */
501         set_current_state(blocking_state);
502         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
503
504         if (!is_vm_hugetlb_page(vmf->vma))
505                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
506                                                   reason);
507         else
508                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
509                                                        vmf->address,
510                                                        vmf->flags, reason);
511         mmap_read_unlock(mm);
512
513         if (likely(must_wait && !READ_ONCE(ctx->released))) {
514                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
515                 schedule();
516         }
517
518         __set_current_state(TASK_RUNNING);
519
520         /*
521          * Here we race with the list_del; list_add in
522          * userfaultfd_ctx_read(), however because we don't ever run
523          * list_del_init() to refile across the two lists, the prev
524          * and next pointers will never point to self. list_add also
525          * would never let any of the two pointers to point to
526          * self. So list_empty_careful won't risk to see both pointers
527          * pointing to self at any time during the list refile. The
528          * only case where list_del_init() is called is the full
529          * removal in the wake function and there we don't re-list_add
530          * and it's fine not to block on the spinlock. The uwq on this
531          * kernel stack can be released after the list_del_init.
532          */
533         if (!list_empty_careful(&uwq.wq.entry)) {
534                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
535                 /*
536                  * No need of list_del_init(), the uwq on the stack
537                  * will be freed shortly anyway.
538                  */
539                 list_del(&uwq.wq.entry);
540                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
541         }
542
543         /*
544          * ctx may go away after this if the userfault pseudo fd is
545          * already released.
546          */
547         userfaultfd_ctx_put(ctx);
548
549 out:
550         return ret;
551 }
552
553 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
554                                               struct userfaultfd_wait_queue *ewq)
555 {
556         struct userfaultfd_ctx *release_new_ctx;
557
558         if (WARN_ON_ONCE(current->flags & PF_EXITING))
559                 goto out;
560
561         ewq->ctx = ctx;
562         init_waitqueue_entry(&ewq->wq, current);
563         release_new_ctx = NULL;
564
565         spin_lock_irq(&ctx->event_wqh.lock);
566         /*
567          * After the __add_wait_queue the uwq is visible to userland
568          * through poll/read().
569          */
570         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
571         for (;;) {
572                 set_current_state(TASK_KILLABLE);
573                 if (ewq->msg.event == 0)
574                         break;
575                 if (READ_ONCE(ctx->released) ||
576                     fatal_signal_pending(current)) {
577                         /*
578                          * &ewq->wq may be queued in fork_event, but
579                          * __remove_wait_queue ignores the head
580                          * parameter. It would be a problem if it
581                          * didn't.
582                          */
583                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
584                         if (ewq->msg.event == UFFD_EVENT_FORK) {
585                                 struct userfaultfd_ctx *new;
586
587                                 new = (struct userfaultfd_ctx *)
588                                         (unsigned long)
589                                         ewq->msg.arg.reserved.reserved1;
590                                 release_new_ctx = new;
591                         }
592                         break;
593                 }
594
595                 spin_unlock_irq(&ctx->event_wqh.lock);
596
597                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
598                 schedule();
599
600                 spin_lock_irq(&ctx->event_wqh.lock);
601         }
602         __set_current_state(TASK_RUNNING);
603         spin_unlock_irq(&ctx->event_wqh.lock);
604
605         if (release_new_ctx) {
606                 struct vm_area_struct *vma;
607                 struct mm_struct *mm = release_new_ctx->mm;
608
609                 /* the various vma->vm_userfaultfd_ctx still points to it */
610                 mmap_write_lock(mm);
611                 for (vma = mm->mmap; vma; vma = vma->vm_next)
612                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
613                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
614                                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
615                         }
616                 mmap_write_unlock(mm);
617
618                 userfaultfd_ctx_put(release_new_ctx);
619         }
620
621         /*
622          * ctx may go away after this if the userfault pseudo fd is
623          * already released.
624          */
625 out:
626         WRITE_ONCE(ctx->mmap_changing, false);
627         userfaultfd_ctx_put(ctx);
628 }
629
630 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
631                                        struct userfaultfd_wait_queue *ewq)
632 {
633         ewq->msg.event = 0;
634         wake_up_locked(&ctx->event_wqh);
635         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
636 }
637
638 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
639 {
640         struct userfaultfd_ctx *ctx = NULL, *octx;
641         struct userfaultfd_fork_ctx *fctx;
642
643         octx = vma->vm_userfaultfd_ctx.ctx;
644         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
645                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
646                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
647                 return 0;
648         }
649
650         list_for_each_entry(fctx, fcs, list)
651                 if (fctx->orig == octx) {
652                         ctx = fctx->new;
653                         break;
654                 }
655
656         if (!ctx) {
657                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
658                 if (!fctx)
659                         return -ENOMEM;
660
661                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
662                 if (!ctx) {
663                         kfree(fctx);
664                         return -ENOMEM;
665                 }
666
667                 refcount_set(&ctx->refcount, 1);
668                 ctx->flags = octx->flags;
669                 ctx->state = UFFD_STATE_RUNNING;
670                 ctx->features = octx->features;
671                 ctx->released = false;
672                 ctx->mmap_changing = false;
673                 ctx->mm = vma->vm_mm;
674                 mmgrab(ctx->mm);
675
676                 userfaultfd_ctx_get(octx);
677                 WRITE_ONCE(octx->mmap_changing, true);
678                 fctx->orig = octx;
679                 fctx->new = ctx;
680                 list_add_tail(&fctx->list, fcs);
681         }
682
683         vma->vm_userfaultfd_ctx.ctx = ctx;
684         return 0;
685 }
686
687 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
688 {
689         struct userfaultfd_ctx *ctx = fctx->orig;
690         struct userfaultfd_wait_queue ewq;
691
692         msg_init(&ewq.msg);
693
694         ewq.msg.event = UFFD_EVENT_FORK;
695         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
696
697         userfaultfd_event_wait_completion(ctx, &ewq);
698 }
699
700 void dup_userfaultfd_complete(struct list_head *fcs)
701 {
702         struct userfaultfd_fork_ctx *fctx, *n;
703
704         list_for_each_entry_safe(fctx, n, fcs, list) {
705                 dup_fctx(fctx);
706                 list_del(&fctx->list);
707                 kfree(fctx);
708         }
709 }
710
711 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
712                              struct vm_userfaultfd_ctx *vm_ctx)
713 {
714         struct userfaultfd_ctx *ctx;
715
716         ctx = vma->vm_userfaultfd_ctx.ctx;
717
718         if (!ctx)
719                 return;
720
721         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
722                 vm_ctx->ctx = ctx;
723                 userfaultfd_ctx_get(ctx);
724                 WRITE_ONCE(ctx->mmap_changing, true);
725         } else {
726                 /* Drop uffd context if remap feature not enabled */
727                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
728                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
729         }
730 }
731
732 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
733                                  unsigned long from, unsigned long to,
734                                  unsigned long len)
735 {
736         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
737         struct userfaultfd_wait_queue ewq;
738
739         if (!ctx)
740                 return;
741
742         if (to & ~PAGE_MASK) {
743                 userfaultfd_ctx_put(ctx);
744                 return;
745         }
746
747         msg_init(&ewq.msg);
748
749         ewq.msg.event = UFFD_EVENT_REMAP;
750         ewq.msg.arg.remap.from = from;
751         ewq.msg.arg.remap.to = to;
752         ewq.msg.arg.remap.len = len;
753
754         userfaultfd_event_wait_completion(ctx, &ewq);
755 }
756
757 bool userfaultfd_remove(struct vm_area_struct *vma,
758                         unsigned long start, unsigned long end)
759 {
760         struct mm_struct *mm = vma->vm_mm;
761         struct userfaultfd_ctx *ctx;
762         struct userfaultfd_wait_queue ewq;
763
764         ctx = vma->vm_userfaultfd_ctx.ctx;
765         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
766                 return true;
767
768         userfaultfd_ctx_get(ctx);
769         WRITE_ONCE(ctx->mmap_changing, true);
770         mmap_read_unlock(mm);
771
772         msg_init(&ewq.msg);
773
774         ewq.msg.event = UFFD_EVENT_REMOVE;
775         ewq.msg.arg.remove.start = start;
776         ewq.msg.arg.remove.end = end;
777
778         userfaultfd_event_wait_completion(ctx, &ewq);
779
780         return false;
781 }
782
783 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
784                           unsigned long start, unsigned long end)
785 {
786         struct userfaultfd_unmap_ctx *unmap_ctx;
787
788         list_for_each_entry(unmap_ctx, unmaps, list)
789                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
790                     unmap_ctx->end == end)
791                         return true;
792
793         return false;
794 }
795
796 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
797                            unsigned long start, unsigned long end,
798                            struct list_head *unmaps)
799 {
800         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
801                 struct userfaultfd_unmap_ctx *unmap_ctx;
802                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
803
804                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
805                     has_unmap_ctx(ctx, unmaps, start, end))
806                         continue;
807
808                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
809                 if (!unmap_ctx)
810                         return -ENOMEM;
811
812                 userfaultfd_ctx_get(ctx);
813                 WRITE_ONCE(ctx->mmap_changing, true);
814                 unmap_ctx->ctx = ctx;
815                 unmap_ctx->start = start;
816                 unmap_ctx->end = end;
817                 list_add_tail(&unmap_ctx->list, unmaps);
818         }
819
820         return 0;
821 }
822
823 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
824 {
825         struct userfaultfd_unmap_ctx *ctx, *n;
826         struct userfaultfd_wait_queue ewq;
827
828         list_for_each_entry_safe(ctx, n, uf, list) {
829                 msg_init(&ewq.msg);
830
831                 ewq.msg.event = UFFD_EVENT_UNMAP;
832                 ewq.msg.arg.remove.start = ctx->start;
833                 ewq.msg.arg.remove.end = ctx->end;
834
835                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
836
837                 list_del(&ctx->list);
838                 kfree(ctx);
839         }
840 }
841
842 static int userfaultfd_release(struct inode *inode, struct file *file)
843 {
844         struct userfaultfd_ctx *ctx = file->private_data;
845         struct mm_struct *mm = ctx->mm;
846         struct vm_area_struct *vma, *prev;
847         /* len == 0 means wake all */
848         struct userfaultfd_wake_range range = { .len = 0, };
849         unsigned long new_flags;
850
851         WRITE_ONCE(ctx->released, true);
852
853         if (!mmget_not_zero(mm))
854                 goto wakeup;
855
856         /*
857          * Flush page faults out of all CPUs. NOTE: all page faults
858          * must be retried without returning VM_FAULT_SIGBUS if
859          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
860          * changes while handle_userfault released the mmap_lock. So
861          * it's critical that released is set to true (above), before
862          * taking the mmap_lock for writing.
863          */
864         mmap_write_lock(mm);
865         prev = NULL;
866         for (vma = mm->mmap; vma; vma = vma->vm_next) {
867                 cond_resched();
868                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
869                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
870                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
871                         prev = vma;
872                         continue;
873                 }
874                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
875                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
876                                  new_flags, vma->anon_vma,
877                                  vma->vm_file, vma->vm_pgoff,
878                                  vma_policy(vma),
879                                  NULL_VM_UFFD_CTX);
880                 if (prev)
881                         vma = prev;
882                 else
883                         prev = vma;
884                 vma->vm_flags = new_flags;
885                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
886         }
887         mmap_write_unlock(mm);
888         mmput(mm);
889 wakeup:
890         /*
891          * After no new page faults can wait on this fault_*wqh, flush
892          * the last page faults that may have been already waiting on
893          * the fault_*wqh.
894          */
895         spin_lock_irq(&ctx->fault_pending_wqh.lock);
896         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
897         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
898         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
899
900         /* Flush pending events that may still wait on event_wqh */
901         wake_up_all(&ctx->event_wqh);
902
903         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
904         userfaultfd_ctx_put(ctx);
905         return 0;
906 }
907
908 /* fault_pending_wqh.lock must be hold by the caller */
909 static inline struct userfaultfd_wait_queue *find_userfault_in(
910                 wait_queue_head_t *wqh)
911 {
912         wait_queue_entry_t *wq;
913         struct userfaultfd_wait_queue *uwq;
914
915         lockdep_assert_held(&wqh->lock);
916
917         uwq = NULL;
918         if (!waitqueue_active(wqh))
919                 goto out;
920         /* walk in reverse to provide FIFO behavior to read userfaults */
921         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
922         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
923 out:
924         return uwq;
925 }
926
927 static inline struct userfaultfd_wait_queue *find_userfault(
928                 struct userfaultfd_ctx *ctx)
929 {
930         return find_userfault_in(&ctx->fault_pending_wqh);
931 }
932
933 static inline struct userfaultfd_wait_queue *find_userfault_evt(
934                 struct userfaultfd_ctx *ctx)
935 {
936         return find_userfault_in(&ctx->event_wqh);
937 }
938
939 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
940 {
941         struct userfaultfd_ctx *ctx = file->private_data;
942         __poll_t ret;
943
944         poll_wait(file, &ctx->fd_wqh, wait);
945
946         switch (ctx->state) {
947         case UFFD_STATE_WAIT_API:
948                 return EPOLLERR;
949         case UFFD_STATE_RUNNING:
950                 /*
951                  * poll() never guarantees that read won't block.
952                  * userfaults can be waken before they're read().
953                  */
954                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
955                         return EPOLLERR;
956                 /*
957                  * lockless access to see if there are pending faults
958                  * __pollwait last action is the add_wait_queue but
959                  * the spin_unlock would allow the waitqueue_active to
960                  * pass above the actual list_add inside
961                  * add_wait_queue critical section. So use a full
962                  * memory barrier to serialize the list_add write of
963                  * add_wait_queue() with the waitqueue_active read
964                  * below.
965                  */
966                 ret = 0;
967                 smp_mb();
968                 if (waitqueue_active(&ctx->fault_pending_wqh))
969                         ret = EPOLLIN;
970                 else if (waitqueue_active(&ctx->event_wqh))
971                         ret = EPOLLIN;
972
973                 return ret;
974         default:
975                 WARN_ON_ONCE(1);
976                 return EPOLLERR;
977         }
978 }
979
980 static const struct file_operations userfaultfd_fops;
981
982 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
983                                   struct userfaultfd_ctx *new,
984                                   struct uffd_msg *msg)
985 {
986         int fd;
987
988         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
989                               O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
990         if (fd < 0)
991                 return fd;
992
993         msg->arg.reserved.reserved1 = 0;
994         msg->arg.fork.ufd = fd;
995         return 0;
996 }
997
998 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
999                                     struct uffd_msg *msg)
1000 {
1001         ssize_t ret;
1002         DECLARE_WAITQUEUE(wait, current);
1003         struct userfaultfd_wait_queue *uwq;
1004         /*
1005          * Handling fork event requires sleeping operations, so
1006          * we drop the event_wqh lock, then do these ops, then
1007          * lock it back and wake up the waiter. While the lock is
1008          * dropped the ewq may go away so we keep track of it
1009          * carefully.
1010          */
1011         LIST_HEAD(fork_event);
1012         struct userfaultfd_ctx *fork_nctx = NULL;
1013
1014         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1015         spin_lock_irq(&ctx->fd_wqh.lock);
1016         __add_wait_queue(&ctx->fd_wqh, &wait);
1017         for (;;) {
1018                 set_current_state(TASK_INTERRUPTIBLE);
1019                 spin_lock(&ctx->fault_pending_wqh.lock);
1020                 uwq = find_userfault(ctx);
1021                 if (uwq) {
1022                         /*
1023                          * Use a seqcount to repeat the lockless check
1024                          * in wake_userfault() to avoid missing
1025                          * wakeups because during the refile both
1026                          * waitqueue could become empty if this is the
1027                          * only userfault.
1028                          */
1029                         write_seqcount_begin(&ctx->refile_seq);
1030
1031                         /*
1032                          * The fault_pending_wqh.lock prevents the uwq
1033                          * to disappear from under us.
1034                          *
1035                          * Refile this userfault from
1036                          * fault_pending_wqh to fault_wqh, it's not
1037                          * pending anymore after we read it.
1038                          *
1039                          * Use list_del() by hand (as
1040                          * userfaultfd_wake_function also uses
1041                          * list_del_init() by hand) to be sure nobody
1042                          * changes __remove_wait_queue() to use
1043                          * list_del_init() in turn breaking the
1044                          * !list_empty_careful() check in
1045                          * handle_userfault(). The uwq->wq.head list
1046                          * must never be empty at any time during the
1047                          * refile, or the waitqueue could disappear
1048                          * from under us. The "wait_queue_head_t"
1049                          * parameter of __remove_wait_queue() is unused
1050                          * anyway.
1051                          */
1052                         list_del(&uwq->wq.entry);
1053                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1054
1055                         write_seqcount_end(&ctx->refile_seq);
1056
1057                         /* careful to always initialize msg if ret == 0 */
1058                         *msg = uwq->msg;
1059                         spin_unlock(&ctx->fault_pending_wqh.lock);
1060                         ret = 0;
1061                         break;
1062                 }
1063                 spin_unlock(&ctx->fault_pending_wqh.lock);
1064
1065                 spin_lock(&ctx->event_wqh.lock);
1066                 uwq = find_userfault_evt(ctx);
1067                 if (uwq) {
1068                         *msg = uwq->msg;
1069
1070                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1071                                 fork_nctx = (struct userfaultfd_ctx *)
1072                                         (unsigned long)
1073                                         uwq->msg.arg.reserved.reserved1;
1074                                 list_move(&uwq->wq.entry, &fork_event);
1075                                 /*
1076                                  * fork_nctx can be freed as soon as
1077                                  * we drop the lock, unless we take a
1078                                  * reference on it.
1079                                  */
1080                                 userfaultfd_ctx_get(fork_nctx);
1081                                 spin_unlock(&ctx->event_wqh.lock);
1082                                 ret = 0;
1083                                 break;
1084                         }
1085
1086                         userfaultfd_event_complete(ctx, uwq);
1087                         spin_unlock(&ctx->event_wqh.lock);
1088                         ret = 0;
1089                         break;
1090                 }
1091                 spin_unlock(&ctx->event_wqh.lock);
1092
1093                 if (signal_pending(current)) {
1094                         ret = -ERESTARTSYS;
1095                         break;
1096                 }
1097                 if (no_wait) {
1098                         ret = -EAGAIN;
1099                         break;
1100                 }
1101                 spin_unlock_irq(&ctx->fd_wqh.lock);
1102                 schedule();
1103                 spin_lock_irq(&ctx->fd_wqh.lock);
1104         }
1105         __remove_wait_queue(&ctx->fd_wqh, &wait);
1106         __set_current_state(TASK_RUNNING);
1107         spin_unlock_irq(&ctx->fd_wqh.lock);
1108
1109         if (!ret && msg->event == UFFD_EVENT_FORK) {
1110                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1111                 spin_lock_irq(&ctx->event_wqh.lock);
1112                 if (!list_empty(&fork_event)) {
1113                         /*
1114                          * The fork thread didn't abort, so we can
1115                          * drop the temporary refcount.
1116                          */
1117                         userfaultfd_ctx_put(fork_nctx);
1118
1119                         uwq = list_first_entry(&fork_event,
1120                                                typeof(*uwq),
1121                                                wq.entry);
1122                         /*
1123                          * If fork_event list wasn't empty and in turn
1124                          * the event wasn't already released by fork
1125                          * (the event is allocated on fork kernel
1126                          * stack), put the event back to its place in
1127                          * the event_wq. fork_event head will be freed
1128                          * as soon as we return so the event cannot
1129                          * stay queued there no matter the current
1130                          * "ret" value.
1131                          */
1132                         list_del(&uwq->wq.entry);
1133                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1134
1135                         /*
1136                          * Leave the event in the waitqueue and report
1137                          * error to userland if we failed to resolve
1138                          * the userfault fork.
1139                          */
1140                         if (likely(!ret))
1141                                 userfaultfd_event_complete(ctx, uwq);
1142                 } else {
1143                         /*
1144                          * Here the fork thread aborted and the
1145                          * refcount from the fork thread on fork_nctx
1146                          * has already been released. We still hold
1147                          * the reference we took before releasing the
1148                          * lock above. If resolve_userfault_fork
1149                          * failed we've to drop it because the
1150                          * fork_nctx has to be freed in such case. If
1151                          * it succeeded we'll hold it because the new
1152                          * uffd references it.
1153                          */
1154                         if (ret)
1155                                 userfaultfd_ctx_put(fork_nctx);
1156                 }
1157                 spin_unlock_irq(&ctx->event_wqh.lock);
1158         }
1159
1160         return ret;
1161 }
1162
1163 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1164                                 size_t count, loff_t *ppos)
1165 {
1166         struct userfaultfd_ctx *ctx = file->private_data;
1167         ssize_t _ret, ret = 0;
1168         struct uffd_msg msg;
1169         int no_wait = file->f_flags & O_NONBLOCK;
1170
1171         if (ctx->state == UFFD_STATE_WAIT_API)
1172                 return -EINVAL;
1173
1174         for (;;) {
1175                 if (count < sizeof(msg))
1176                         return ret ? ret : -EINVAL;
1177                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1178                 if (_ret < 0)
1179                         return ret ? ret : _ret;
1180                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1181                         return ret ? ret : -EFAULT;
1182                 ret += sizeof(msg);
1183                 buf += sizeof(msg);
1184                 count -= sizeof(msg);
1185                 /*
1186                  * Allow to read more than one fault at time but only
1187                  * block if waiting for the very first one.
1188                  */
1189                 no_wait = O_NONBLOCK;
1190         }
1191 }
1192
1193 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1194                              struct userfaultfd_wake_range *range)
1195 {
1196         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1197         /* wake all in the range and autoremove */
1198         if (waitqueue_active(&ctx->fault_pending_wqh))
1199                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1200                                      range);
1201         if (waitqueue_active(&ctx->fault_wqh))
1202                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1203         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1204 }
1205
1206 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1207                                            struct userfaultfd_wake_range *range)
1208 {
1209         unsigned seq;
1210         bool need_wakeup;
1211
1212         /*
1213          * To be sure waitqueue_active() is not reordered by the CPU
1214          * before the pagetable update, use an explicit SMP memory
1215          * barrier here. PT lock release or mmap_read_unlock(mm) still
1216          * have release semantics that can allow the
1217          * waitqueue_active() to be reordered before the pte update.
1218          */
1219         smp_mb();
1220
1221         /*
1222          * Use waitqueue_active because it's very frequent to
1223          * change the address space atomically even if there are no
1224          * userfaults yet. So we take the spinlock only when we're
1225          * sure we've userfaults to wake.
1226          */
1227         do {
1228                 seq = read_seqcount_begin(&ctx->refile_seq);
1229                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1230                         waitqueue_active(&ctx->fault_wqh);
1231                 cond_resched();
1232         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1233         if (need_wakeup)
1234                 __wake_userfault(ctx, range);
1235 }
1236
1237 static __always_inline int validate_range(struct mm_struct *mm,
1238                                           __u64 *start, __u64 len)
1239 {
1240         __u64 task_size = mm->task_size;
1241
1242         *start = untagged_addr(*start);
1243
1244         if (*start & ~PAGE_MASK)
1245                 return -EINVAL;
1246         if (len & ~PAGE_MASK)
1247                 return -EINVAL;
1248         if (!len)
1249                 return -EINVAL;
1250         if (*start < mmap_min_addr)
1251                 return -EINVAL;
1252         if (*start >= task_size)
1253                 return -EINVAL;
1254         if (len > task_size - *start)
1255                 return -EINVAL;
1256         return 0;
1257 }
1258
1259 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1260                                      unsigned long vm_flags)
1261 {
1262         /* FIXME: add WP support to hugetlbfs and shmem */
1263         return vma_is_anonymous(vma) ||
1264                 ((is_vm_hugetlb_page(vma) || vma_is_shmem(vma)) &&
1265                  !(vm_flags & VM_UFFD_WP));
1266 }
1267
1268 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1269                                 unsigned long arg)
1270 {
1271         struct mm_struct *mm = ctx->mm;
1272         struct vm_area_struct *vma, *prev, *cur;
1273         int ret;
1274         struct uffdio_register uffdio_register;
1275         struct uffdio_register __user *user_uffdio_register;
1276         unsigned long vm_flags, new_flags;
1277         bool found;
1278         bool basic_ioctls;
1279         unsigned long start, end, vma_end;
1280
1281         user_uffdio_register = (struct uffdio_register __user *) arg;
1282
1283         ret = -EFAULT;
1284         if (copy_from_user(&uffdio_register, user_uffdio_register,
1285                            sizeof(uffdio_register)-sizeof(__u64)))
1286                 goto out;
1287
1288         ret = -EINVAL;
1289         if (!uffdio_register.mode)
1290                 goto out;
1291         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1292                                      UFFDIO_REGISTER_MODE_WP))
1293                 goto out;
1294         vm_flags = 0;
1295         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1296                 vm_flags |= VM_UFFD_MISSING;
1297         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)
1298                 vm_flags |= VM_UFFD_WP;
1299
1300         ret = validate_range(mm, &uffdio_register.range.start,
1301                              uffdio_register.range.len);
1302         if (ret)
1303                 goto out;
1304
1305         start = uffdio_register.range.start;
1306         end = start + uffdio_register.range.len;
1307
1308         ret = -ENOMEM;
1309         if (!mmget_not_zero(mm))
1310                 goto out;
1311
1312         mmap_write_lock(mm);
1313         vma = find_vma_prev(mm, start, &prev);
1314         if (!vma)
1315                 goto out_unlock;
1316
1317         /* check that there's at least one vma in the range */
1318         ret = -EINVAL;
1319         if (vma->vm_start >= end)
1320                 goto out_unlock;
1321
1322         /*
1323          * If the first vma contains huge pages, make sure start address
1324          * is aligned to huge page size.
1325          */
1326         if (is_vm_hugetlb_page(vma)) {
1327                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1328
1329                 if (start & (vma_hpagesize - 1))
1330                         goto out_unlock;
1331         }
1332
1333         /*
1334          * Search for not compatible vmas.
1335          */
1336         found = false;
1337         basic_ioctls = false;
1338         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1339                 cond_resched();
1340
1341                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1342                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1343
1344                 /* check not compatible vmas */
1345                 ret = -EINVAL;
1346                 if (!vma_can_userfault(cur, vm_flags))
1347                         goto out_unlock;
1348
1349                 /*
1350                  * UFFDIO_COPY will fill file holes even without
1351                  * PROT_WRITE. This check enforces that if this is a
1352                  * MAP_SHARED, the process has write permission to the backing
1353                  * file. If VM_MAYWRITE is set it also enforces that on a
1354                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1355                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1356                  */
1357                 ret = -EPERM;
1358                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1359                         goto out_unlock;
1360
1361                 /*
1362                  * If this vma contains ending address, and huge pages
1363                  * check alignment.
1364                  */
1365                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1366                     end > cur->vm_start) {
1367                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1368
1369                         ret = -EINVAL;
1370
1371                         if (end & (vma_hpagesize - 1))
1372                                 goto out_unlock;
1373                 }
1374                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1375                         goto out_unlock;
1376
1377                 /*
1378                  * Check that this vma isn't already owned by a
1379                  * different userfaultfd. We can't allow more than one
1380                  * userfaultfd to own a single vma simultaneously or we
1381                  * wouldn't know which one to deliver the userfaults to.
1382                  */
1383                 ret = -EBUSY;
1384                 if (cur->vm_userfaultfd_ctx.ctx &&
1385                     cur->vm_userfaultfd_ctx.ctx != ctx)
1386                         goto out_unlock;
1387
1388                 /*
1389                  * Note vmas containing huge pages
1390                  */
1391                 if (is_vm_hugetlb_page(cur))
1392                         basic_ioctls = true;
1393
1394                 found = true;
1395         }
1396         BUG_ON(!found);
1397
1398         if (vma->vm_start < start)
1399                 prev = vma;
1400
1401         ret = 0;
1402         do {
1403                 cond_resched();
1404
1405                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1406                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1407                        vma->vm_userfaultfd_ctx.ctx != ctx);
1408                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1409
1410                 /*
1411                  * Nothing to do: this vma is already registered into this
1412                  * userfaultfd and with the right tracking mode too.
1413                  */
1414                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1415                     (vma->vm_flags & vm_flags) == vm_flags)
1416                         goto skip;
1417
1418                 if (vma->vm_start > start)
1419                         start = vma->vm_start;
1420                 vma_end = min(end, vma->vm_end);
1421
1422                 new_flags = (vma->vm_flags &
1423                              ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags;
1424                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1425                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1426                                  vma_policy(vma),
1427                                  ((struct vm_userfaultfd_ctx){ ctx }));
1428                 if (prev) {
1429                         vma = prev;
1430                         goto next;
1431                 }
1432                 if (vma->vm_start < start) {
1433                         ret = split_vma(mm, vma, start, 1);
1434                         if (ret)
1435                                 break;
1436                 }
1437                 if (vma->vm_end > end) {
1438                         ret = split_vma(mm, vma, end, 0);
1439                         if (ret)
1440                                 break;
1441                 }
1442         next:
1443                 /*
1444                  * In the vma_merge() successful mprotect-like case 8:
1445                  * the next vma was merged into the current one and
1446                  * the current one has not been updated yet.
1447                  */
1448                 vma->vm_flags = new_flags;
1449                 vma->vm_userfaultfd_ctx.ctx = ctx;
1450
1451         skip:
1452                 prev = vma;
1453                 start = vma->vm_end;
1454                 vma = vma->vm_next;
1455         } while (vma && vma->vm_start < end);
1456 out_unlock:
1457         mmap_write_unlock(mm);
1458         mmput(mm);
1459         if (!ret) {
1460                 __u64 ioctls_out;
1461
1462                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1463                     UFFD_API_RANGE_IOCTLS;
1464
1465                 /*
1466                  * Declare the WP ioctl only if the WP mode is
1467                  * specified and all checks passed with the range
1468                  */
1469                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1470                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1471
1472                 /*
1473                  * Now that we scanned all vmas we can already tell
1474                  * userland which ioctls methods are guaranteed to
1475                  * succeed on this range.
1476                  */
1477                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1478                         ret = -EFAULT;
1479         }
1480 out:
1481         return ret;
1482 }
1483
1484 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1485                                   unsigned long arg)
1486 {
1487         struct mm_struct *mm = ctx->mm;
1488         struct vm_area_struct *vma, *prev, *cur;
1489         int ret;
1490         struct uffdio_range uffdio_unregister;
1491         unsigned long new_flags;
1492         bool found;
1493         unsigned long start, end, vma_end;
1494         const void __user *buf = (void __user *)arg;
1495
1496         ret = -EFAULT;
1497         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1498                 goto out;
1499
1500         ret = validate_range(mm, &uffdio_unregister.start,
1501                              uffdio_unregister.len);
1502         if (ret)
1503                 goto out;
1504
1505         start = uffdio_unregister.start;
1506         end = start + uffdio_unregister.len;
1507
1508         ret = -ENOMEM;
1509         if (!mmget_not_zero(mm))
1510                 goto out;
1511
1512         mmap_write_lock(mm);
1513         vma = find_vma_prev(mm, start, &prev);
1514         if (!vma)
1515                 goto out_unlock;
1516
1517         /* check that there's at least one vma in the range */
1518         ret = -EINVAL;
1519         if (vma->vm_start >= end)
1520                 goto out_unlock;
1521
1522         /*
1523          * If the first vma contains huge pages, make sure start address
1524          * is aligned to huge page size.
1525          */
1526         if (is_vm_hugetlb_page(vma)) {
1527                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1528
1529                 if (start & (vma_hpagesize - 1))
1530                         goto out_unlock;
1531         }
1532
1533         /*
1534          * Search for not compatible vmas.
1535          */
1536         found = false;
1537         ret = -EINVAL;
1538         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1539                 cond_resched();
1540
1541                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1542                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1543
1544                 /*
1545                  * Check not compatible vmas, not strictly required
1546                  * here as not compatible vmas cannot have an
1547                  * userfaultfd_ctx registered on them, but this
1548                  * provides for more strict behavior to notice
1549                  * unregistration errors.
1550                  */
1551                 if (!vma_can_userfault(cur, cur->vm_flags))
1552                         goto out_unlock;
1553
1554                 found = true;
1555         }
1556         BUG_ON(!found);
1557
1558         if (vma->vm_start < start)
1559                 prev = vma;
1560
1561         ret = 0;
1562         do {
1563                 cond_resched();
1564
1565                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1566
1567                 /*
1568                  * Nothing to do: this vma is already registered into this
1569                  * userfaultfd and with the right tracking mode too.
1570                  */
1571                 if (!vma->vm_userfaultfd_ctx.ctx)
1572                         goto skip;
1573
1574                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1575
1576                 if (vma->vm_start > start)
1577                         start = vma->vm_start;
1578                 vma_end = min(end, vma->vm_end);
1579
1580                 if (userfaultfd_missing(vma)) {
1581                         /*
1582                          * Wake any concurrent pending userfault while
1583                          * we unregister, so they will not hang
1584                          * permanently and it avoids userland to call
1585                          * UFFDIO_WAKE explicitly.
1586                          */
1587                         struct userfaultfd_wake_range range;
1588                         range.start = start;
1589                         range.len = vma_end - start;
1590                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1591                 }
1592
1593                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1594                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1595                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1596                                  vma_policy(vma),
1597                                  NULL_VM_UFFD_CTX);
1598                 if (prev) {
1599                         vma = prev;
1600                         goto next;
1601                 }
1602                 if (vma->vm_start < start) {
1603                         ret = split_vma(mm, vma, start, 1);
1604                         if (ret)
1605                                 break;
1606                 }
1607                 if (vma->vm_end > end) {
1608                         ret = split_vma(mm, vma, end, 0);
1609                         if (ret)
1610                                 break;
1611                 }
1612         next:
1613                 /*
1614                  * In the vma_merge() successful mprotect-like case 8:
1615                  * the next vma was merged into the current one and
1616                  * the current one has not been updated yet.
1617                  */
1618                 vma->vm_flags = new_flags;
1619                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1620
1621         skip:
1622                 prev = vma;
1623                 start = vma->vm_end;
1624                 vma = vma->vm_next;
1625         } while (vma && vma->vm_start < end);
1626 out_unlock:
1627         mmap_write_unlock(mm);
1628         mmput(mm);
1629 out:
1630         return ret;
1631 }
1632
1633 /*
1634  * userfaultfd_wake may be used in combination with the
1635  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1636  */
1637 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1638                             unsigned long arg)
1639 {
1640         int ret;
1641         struct uffdio_range uffdio_wake;
1642         struct userfaultfd_wake_range range;
1643         const void __user *buf = (void __user *)arg;
1644
1645         ret = -EFAULT;
1646         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1647                 goto out;
1648
1649         ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1650         if (ret)
1651                 goto out;
1652
1653         range.start = uffdio_wake.start;
1654         range.len = uffdio_wake.len;
1655
1656         /*
1657          * len == 0 means wake all and we don't want to wake all here,
1658          * so check it again to be sure.
1659          */
1660         VM_BUG_ON(!range.len);
1661
1662         wake_userfault(ctx, &range);
1663         ret = 0;
1664
1665 out:
1666         return ret;
1667 }
1668
1669 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1670                             unsigned long arg)
1671 {
1672         __s64 ret;
1673         struct uffdio_copy uffdio_copy;
1674         struct uffdio_copy __user *user_uffdio_copy;
1675         struct userfaultfd_wake_range range;
1676
1677         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1678
1679         ret = -EAGAIN;
1680         if (READ_ONCE(ctx->mmap_changing))
1681                 goto out;
1682
1683         ret = -EFAULT;
1684         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1685                            /* don't copy "copy" last field */
1686                            sizeof(uffdio_copy)-sizeof(__s64)))
1687                 goto out;
1688
1689         ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1690         if (ret)
1691                 goto out;
1692         /*
1693          * double check for wraparound just in case. copy_from_user()
1694          * will later check uffdio_copy.src + uffdio_copy.len to fit
1695          * in the userland range.
1696          */
1697         ret = -EINVAL;
1698         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1699                 goto out;
1700         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1701                 goto out;
1702         if (mmget_not_zero(ctx->mm)) {
1703                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1704                                    uffdio_copy.len, &ctx->mmap_changing,
1705                                    uffdio_copy.mode);
1706                 mmput(ctx->mm);
1707         } else {
1708                 return -ESRCH;
1709         }
1710         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1711                 return -EFAULT;
1712         if (ret < 0)
1713                 goto out;
1714         BUG_ON(!ret);
1715         /* len == 0 would wake all */
1716         range.len = ret;
1717         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1718                 range.start = uffdio_copy.dst;
1719                 wake_userfault(ctx, &range);
1720         }
1721         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1722 out:
1723         return ret;
1724 }
1725
1726 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1727                                 unsigned long arg)
1728 {
1729         __s64 ret;
1730         struct uffdio_zeropage uffdio_zeropage;
1731         struct uffdio_zeropage __user *user_uffdio_zeropage;
1732         struct userfaultfd_wake_range range;
1733
1734         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1735
1736         ret = -EAGAIN;
1737         if (READ_ONCE(ctx->mmap_changing))
1738                 goto out;
1739
1740         ret = -EFAULT;
1741         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1742                            /* don't copy "zeropage" last field */
1743                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1744                 goto out;
1745
1746         ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1747                              uffdio_zeropage.range.len);
1748         if (ret)
1749                 goto out;
1750         ret = -EINVAL;
1751         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1752                 goto out;
1753
1754         if (mmget_not_zero(ctx->mm)) {
1755                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1756                                      uffdio_zeropage.range.len,
1757                                      &ctx->mmap_changing);
1758                 mmput(ctx->mm);
1759         } else {
1760                 return -ESRCH;
1761         }
1762         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1763                 return -EFAULT;
1764         if (ret < 0)
1765                 goto out;
1766         /* len == 0 would wake all */
1767         BUG_ON(!ret);
1768         range.len = ret;
1769         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1770                 range.start = uffdio_zeropage.range.start;
1771                 wake_userfault(ctx, &range);
1772         }
1773         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1774 out:
1775         return ret;
1776 }
1777
1778 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1779                                     unsigned long arg)
1780 {
1781         int ret;
1782         struct uffdio_writeprotect uffdio_wp;
1783         struct uffdio_writeprotect __user *user_uffdio_wp;
1784         struct userfaultfd_wake_range range;
1785         bool mode_wp, mode_dontwake;
1786
1787         if (READ_ONCE(ctx->mmap_changing))
1788                 return -EAGAIN;
1789
1790         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1791
1792         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1793                            sizeof(struct uffdio_writeprotect)))
1794                 return -EFAULT;
1795
1796         ret = validate_range(ctx->mm, &uffdio_wp.range.start,
1797                              uffdio_wp.range.len);
1798         if (ret)
1799                 return ret;
1800
1801         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1802                                UFFDIO_WRITEPROTECT_MODE_WP))
1803                 return -EINVAL;
1804
1805         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1806         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1807
1808         if (mode_wp && mode_dontwake)
1809                 return -EINVAL;
1810
1811         ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1812                                   uffdio_wp.range.len, mode_wp,
1813                                   &ctx->mmap_changing);
1814         if (ret)
1815                 return ret;
1816
1817         if (!mode_wp && !mode_dontwake) {
1818                 range.start = uffdio_wp.range.start;
1819                 range.len = uffdio_wp.range.len;
1820                 wake_userfault(ctx, &range);
1821         }
1822         return ret;
1823 }
1824
1825 static inline unsigned int uffd_ctx_features(__u64 user_features)
1826 {
1827         /*
1828          * For the current set of features the bits just coincide
1829          */
1830         return (unsigned int)user_features;
1831 }
1832
1833 /*
1834  * userland asks for a certain API version and we return which bits
1835  * and ioctl commands are implemented in this kernel for such API
1836  * version or -EINVAL if unknown.
1837  */
1838 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1839                            unsigned long arg)
1840 {
1841         struct uffdio_api uffdio_api;
1842         void __user *buf = (void __user *)arg;
1843         int ret;
1844         __u64 features;
1845
1846         ret = -EINVAL;
1847         if (ctx->state != UFFD_STATE_WAIT_API)
1848                 goto out;
1849         ret = -EFAULT;
1850         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1851                 goto out;
1852         features = uffdio_api.features;
1853         ret = -EINVAL;
1854         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1855                 goto err_out;
1856         ret = -EPERM;
1857         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1858                 goto err_out;
1859         /* report all available features and ioctls to userland */
1860         uffdio_api.features = UFFD_API_FEATURES;
1861         uffdio_api.ioctls = UFFD_API_IOCTLS;
1862         ret = -EFAULT;
1863         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1864                 goto out;
1865         ctx->state = UFFD_STATE_RUNNING;
1866         /* only enable the requested features for this uffd context */
1867         ctx->features = uffd_ctx_features(features);
1868         ret = 0;
1869 out:
1870         return ret;
1871 err_out:
1872         memset(&uffdio_api, 0, sizeof(uffdio_api));
1873         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1874                 ret = -EFAULT;
1875         goto out;
1876 }
1877
1878 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1879                               unsigned long arg)
1880 {
1881         int ret = -EINVAL;
1882         struct userfaultfd_ctx *ctx = file->private_data;
1883
1884         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1885                 return -EINVAL;
1886
1887         switch(cmd) {
1888         case UFFDIO_API:
1889                 ret = userfaultfd_api(ctx, arg);
1890                 break;
1891         case UFFDIO_REGISTER:
1892                 ret = userfaultfd_register(ctx, arg);
1893                 break;
1894         case UFFDIO_UNREGISTER:
1895                 ret = userfaultfd_unregister(ctx, arg);
1896                 break;
1897         case UFFDIO_WAKE:
1898                 ret = userfaultfd_wake(ctx, arg);
1899                 break;
1900         case UFFDIO_COPY:
1901                 ret = userfaultfd_copy(ctx, arg);
1902                 break;
1903         case UFFDIO_ZEROPAGE:
1904                 ret = userfaultfd_zeropage(ctx, arg);
1905                 break;
1906         case UFFDIO_WRITEPROTECT:
1907                 ret = userfaultfd_writeprotect(ctx, arg);
1908                 break;
1909         }
1910         return ret;
1911 }
1912
1913 #ifdef CONFIG_PROC_FS
1914 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1915 {
1916         struct userfaultfd_ctx *ctx = f->private_data;
1917         wait_queue_entry_t *wq;
1918         unsigned long pending = 0, total = 0;
1919
1920         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1921         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1922                 pending++;
1923                 total++;
1924         }
1925         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1926                 total++;
1927         }
1928         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1929
1930         /*
1931          * If more protocols will be added, there will be all shown
1932          * separated by a space. Like this:
1933          *      protocols: aa:... bb:...
1934          */
1935         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1936                    pending, total, UFFD_API, ctx->features,
1937                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1938 }
1939 #endif
1940
1941 static const struct file_operations userfaultfd_fops = {
1942 #ifdef CONFIG_PROC_FS
1943         .show_fdinfo    = userfaultfd_show_fdinfo,
1944 #endif
1945         .release        = userfaultfd_release,
1946         .poll           = userfaultfd_poll,
1947         .read           = userfaultfd_read,
1948         .unlocked_ioctl = userfaultfd_ioctl,
1949         .compat_ioctl   = compat_ptr_ioctl,
1950         .llseek         = noop_llseek,
1951 };
1952
1953 static void init_once_userfaultfd_ctx(void *mem)
1954 {
1955         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1956
1957         init_waitqueue_head(&ctx->fault_pending_wqh);
1958         init_waitqueue_head(&ctx->fault_wqh);
1959         init_waitqueue_head(&ctx->event_wqh);
1960         init_waitqueue_head(&ctx->fd_wqh);
1961         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
1962 }
1963
1964 SYSCALL_DEFINE1(userfaultfd, int, flags)
1965 {
1966         struct userfaultfd_ctx *ctx;
1967         int fd;
1968
1969         if (!sysctl_unprivileged_userfaultfd &&
1970             (flags & UFFD_USER_MODE_ONLY) == 0 &&
1971             !capable(CAP_SYS_PTRACE)) {
1972                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
1973                         "sysctl knob to 1 if kernel faults must be handled "
1974                         "without obtaining CAP_SYS_PTRACE capability\n");
1975                 return -EPERM;
1976         }
1977
1978         BUG_ON(!current->mm);
1979
1980         /* Check the UFFD_* constants for consistency.  */
1981         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
1982         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1983         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1984
1985         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
1986                 return -EINVAL;
1987
1988         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1989         if (!ctx)
1990                 return -ENOMEM;
1991
1992         refcount_set(&ctx->refcount, 1);
1993         ctx->flags = flags;
1994         ctx->features = 0;
1995         ctx->state = UFFD_STATE_WAIT_API;
1996         ctx->released = false;
1997         ctx->mmap_changing = false;
1998         ctx->mm = current->mm;
1999         /* prevent the mm struct to be freed */
2000         mmgrab(ctx->mm);
2001
2002         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
2003                               O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
2004         if (fd < 0) {
2005                 mmdrop(ctx->mm);
2006                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2007         }
2008         return fd;
2009 }
2010
2011 static int __init userfaultfd_init(void)
2012 {
2013         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2014                                                 sizeof(struct userfaultfd_ctx),
2015                                                 0,
2016                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2017                                                 init_once_userfaultfd_ctx);
2018         return 0;
2019 }
2020 __initcall(userfaultfd_init);