~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/userfaultfd.c

Version: ~ [ linux-5.15-rc1 ] ~ [ linux-5.14.5 ] ~ [ linux-5.13.18 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.66 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.147 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.206 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.246 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.282 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.283 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp