Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/fs/userfaultfd.c

  1 /*
  2  *  fs/userfaultfd.c
  3  *
  4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
  5  *  Copyright (C) 2008-2009 Red Hat, Inc.
  6  *  Copyright (C) 2015  Red Hat, Inc.
  7  *
  8  *  This work is licensed under the terms of the GNU GPL, version 2. See
  9  *  the COPYING file in the top-level directory.
 10  *
 11  *  Some part derived from fs/eventfd.c (anon inode setup) and
 12  *  mm/ksm.c (mm hashing).
 13  */
 14 
 15 #include <linux/hashtable.h>
 16 #include <linux/sched.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 
 30 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
 31 
 32 enum userfaultfd_state {
 33         UFFD_STATE_WAIT_API,
 34         UFFD_STATE_RUNNING,
 35 };
 36 
 37 /*
 38  * Start with fault_pending_wqh and fault_wqh so they're more likely
 39  * to be in the same cacheline.
 40  */
 41 struct userfaultfd_ctx {
 42         /* waitqueue head for the pending (i.e. not read) userfaults */
 43         wait_queue_head_t fault_pending_wqh;
 44         /* waitqueue head for the userfaults */
 45         wait_queue_head_t fault_wqh;
 46         /* waitqueue head for the pseudo fd to wakeup poll/read */
 47         wait_queue_head_t fd_wqh;
 48         /* a refile sequence protected by fault_pending_wqh lock */
 49         struct seqcount refile_seq;
 50         /* pseudo fd refcounting */
 51         atomic_t refcount;
 52         /* userfaultfd syscall flags */
 53         unsigned int flags;
 54         /* state machine */
 55         enum userfaultfd_state state;
 56         /* released */
 57         bool released;
 58         /* mm with one ore more vmas attached to this userfaultfd_ctx */
 59         struct mm_struct *mm;
 60 };
 61 
 62 struct userfaultfd_wait_queue {
 63         struct uffd_msg msg;
 64         wait_queue_t wq;
 65         struct userfaultfd_ctx *ctx;
 66         bool waken;
 67 };
 68 
 69 struct userfaultfd_wake_range {
 70         unsigned long start;
 71         unsigned long len;
 72 };
 73 
 74 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
 75                                      int wake_flags, void *key)
 76 {
 77         struct userfaultfd_wake_range *range = key;
 78         int ret;
 79         struct userfaultfd_wait_queue *uwq;
 80         unsigned long start, len;
 81 
 82         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 83         ret = 0;
 84         /* len == 0 means wake all */
 85         start = range->start;
 86         len = range->len;
 87         if (len && (start > uwq->msg.arg.pagefault.address ||
 88                     start + len <= uwq->msg.arg.pagefault.address))
 89                 goto out;
 90         WRITE_ONCE(uwq->waken, true);
 91         /*
 92          * The implicit smp_mb__before_spinlock in try_to_wake_up()
 93          * renders uwq->waken visible to other CPUs before the task is
 94          * waken.
 95          */
 96         ret = wake_up_state(wq->private, mode);
 97         if (ret)
 98                 /*
 99                  * Wake only once, autoremove behavior.
100                  *
101                  * After the effect of list_del_init is visible to the
102                  * other CPUs, the waitqueue may disappear from under
103                  * us, see the !list_empty_careful() in
104                  * handle_userfault(). try_to_wake_up() has an
105                  * implicit smp_mb__before_spinlock, and the
106                  * wq->private is read before calling the extern
107                  * function "wake_up_state" (which in turns calls
108                  * try_to_wake_up). While the spin_lock;spin_unlock;
109                  * wouldn't be enough, the smp_mb__before_spinlock is
110                  * enough to avoid an explicit smp_mb() here.
111                  */
112                 list_del_init(&wq->task_list);
113 out:
114         return ret;
115 }
116 
117 /**
118  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
119  * context.
120  * @ctx: [in] Pointer to the userfaultfd context.
121  *
122  * Returns: In case of success, returns not zero.
123  */
124 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
125 {
126         if (!atomic_inc_not_zero(&ctx->refcount))
127                 BUG();
128 }
129 
130 /**
131  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
132  * context.
133  * @ctx: [in] Pointer to userfaultfd context.
134  *
135  * The userfaultfd context reference must have been previously acquired either
136  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
137  */
138 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
139 {
140         if (atomic_dec_and_test(&ctx->refcount)) {
141                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
142                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
143                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
144                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
145                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
146                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
147                 mmdrop(ctx->mm);
148                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
149         }
150 }
151 
152 static inline void msg_init(struct uffd_msg *msg)
153 {
154         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
155         /*
156          * Must use memset to zero out the paddings or kernel data is
157          * leaked to userland.
158          */
159         memset(msg, 0, sizeof(struct uffd_msg));
160 }
161 
162 static inline struct uffd_msg userfault_msg(unsigned long address,
163                                             unsigned int flags,
164                                             unsigned long reason)
165 {
166         struct uffd_msg msg;
167         msg_init(&msg);
168         msg.event = UFFD_EVENT_PAGEFAULT;
169         msg.arg.pagefault.address = address;
170         if (flags & FAULT_FLAG_WRITE)
171                 /*
172                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WRITE was set in the
173                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
174                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
175                  * was a read fault, otherwise if set it means it's
176                  * a write fault.
177                  */
178                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
179         if (reason & VM_UFFD_WP)
180                 /*
181                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
182                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
183                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
184                  * a missing fault, otherwise if set it means it's a
185                  * write protect fault.
186                  */
187                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
188         return msg;
189 }
190 
191 /*
192  * Verify the pagetables are still not ok after having reigstered into
193  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
194  * userfault that has already been resolved, if userfaultfd_read and
195  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
196  * threads.
197  */
198 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
199                                          unsigned long address,
200                                          unsigned long flags,
201                                          unsigned long reason)
202 {
203         struct mm_struct *mm = ctx->mm;
204         pgd_t *pgd;
205         pud_t *pud;
206         pmd_t *pmd, _pmd;
207         pte_t *pte;
208         bool ret = true;
209 
210         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
211 
212         pgd = pgd_offset(mm, address);
213         if (!pgd_present(*pgd))
214                 goto out;
215         pud = pud_offset(pgd, address);
216         if (!pud_present(*pud))
217                 goto out;
218         pmd = pmd_offset(pud, address);
219         /*
220          * READ_ONCE must function as a barrier with narrower scope
221          * and it must be equivalent to:
222          *      _pmd = *pmd; barrier();
223          *
224          * This is to deal with the instability (as in
225          * pmd_trans_unstable) of the pmd.
226          */
227         _pmd = READ_ONCE(*pmd);
228         if (!pmd_present(_pmd))
229                 goto out;
230 
231         ret = false;
232         if (pmd_trans_huge(_pmd))
233                 goto out;
234 
235         /*
236          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
237          * and use the standard pte_offset_map() instead of parsing _pmd.
238          */
239         pte = pte_offset_map(pmd, address);
240         /*
241          * Lockless access: we're in a wait_event so it's ok if it
242          * changes under us.
243          */
244         if (pte_none(*pte))
245                 ret = true;
246         pte_unmap(pte);
247 
248 out:
249         return ret;
250 }
251 
252 /*
253  * The locking rules involved in returning VM_FAULT_RETRY depending on
254  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
255  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
256  * recommendation in __lock_page_or_retry is not an understatement.
257  *
258  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
259  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
260  * not set.
261  *
262  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
263  * set, VM_FAULT_RETRY can still be returned if and only if there are
264  * fatal_signal_pending()s, and the mmap_sem must be released before
265  * returning it.
266  */
267 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
268 {
269         struct mm_struct *mm = vmf->vma->vm_mm;
270         struct userfaultfd_ctx *ctx;
271         struct userfaultfd_wait_queue uwq;
272         int ret;
273         bool must_wait, return_to_userland;
274         long blocking_state;
275 
276         BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
277 
278         ret = VM_FAULT_SIGBUS;
279         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
280         if (!ctx)
281                 goto out;
282 
283         BUG_ON(ctx->mm != mm);
284 
285         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
286         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
287 
288         /*
289          * If it's already released don't get it. This avoids to loop
290          * in __get_user_pages if userfaultfd_release waits on the
291          * caller of handle_userfault to release the mmap_sem.
292          */
293         if (unlikely(ACCESS_ONCE(ctx->released)))
294                 goto out;
295 
296         /*
297          * We don't do userfault handling for the final child pid update.
298          */
299         if (current->flags & PF_EXITING)
300                 goto out;
301 
302         /*
303          * Check that we can return VM_FAULT_RETRY.
304          *
305          * NOTE: it should become possible to return VM_FAULT_RETRY
306          * even if FAULT_FLAG_TRIED is set without leading to gup()
307          * -EBUSY failures, if the userfaultfd is to be extended for
308          * VM_UFFD_WP tracking and we intend to arm the userfault
309          * without first stopping userland access to the memory. For
310          * VM_UFFD_MISSING userfaults this is enough for now.
311          */
312         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
313                 /*
314                  * Validate the invariant that nowait must allow retry
315                  * to be sure not to return SIGBUS erroneously on
316                  * nowait invocations.
317                  */
318                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
319 #ifdef CONFIG_DEBUG_VM
320                 if (printk_ratelimit()) {
321                         printk(KERN_WARNING
322                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
323                                vmf->flags);
324                         dump_stack();
325                 }
326 #endif
327                 goto out;
328         }
329 
330         /*
331          * Handle nowait, not much to do other than tell it to retry
332          * and wait.
333          */
334         ret = VM_FAULT_RETRY;
335         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
336                 goto out;
337 
338         /* take the reference before dropping the mmap_sem */
339         userfaultfd_ctx_get(ctx);
340 
341         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
342         uwq.wq.private = current;
343         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
344         uwq.ctx = ctx;
345         uwq.waken = false;
346 
347         return_to_userland =
348                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
349                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
350         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
351                          TASK_KILLABLE;
352 
353         spin_lock(&ctx->fault_pending_wqh.lock);
354         /*
355          * After the __add_wait_queue the uwq is visible to userland
356          * through poll/read().
357          */
358         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
359         /*
360          * The smp_mb() after __set_current_state prevents the reads
361          * following the spin_unlock to happen before the list_add in
362          * __add_wait_queue.
363          */
364         set_current_state(blocking_state);
365         spin_unlock(&ctx->fault_pending_wqh.lock);
366 
367         must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
368                                           reason);
369         up_read(&mm->mmap_sem);
370 
371         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
372                    (return_to_userland ? !signal_pending(current) :
373                     !fatal_signal_pending(current)))) {
374                 wake_up_poll(&ctx->fd_wqh, POLLIN);
375                 schedule();
376                 ret |= VM_FAULT_MAJOR;
377 
378                 /*
379                  * False wakeups can orginate even from rwsem before
380                  * up_read() however userfaults will wait either for a
381                  * targeted wakeup on the specific uwq waitqueue from
382                  * wake_userfault() or for signals or for uffd
383                  * release.
384                  */
385                 while (!READ_ONCE(uwq.waken)) {
386                         /*
387                          * This needs the full smp_store_mb()
388                          * guarantee as the state write must be
389                          * visible to other CPUs before reading
390                          * uwq.waken from other CPUs.
391                          */
392                         set_current_state(blocking_state);
393                         if (READ_ONCE(uwq.waken) ||
394                             READ_ONCE(ctx->released) ||
395                             (return_to_userland ? signal_pending(current) :
396                              fatal_signal_pending(current)))
397                                 break;
398                         schedule();
399                 }
400         }
401 
402         __set_current_state(TASK_RUNNING);
403 
404         if (return_to_userland) {
405                 if (signal_pending(current) &&
406                     !fatal_signal_pending(current)) {
407                         /*
408                          * If we got a SIGSTOP or SIGCONT and this is
409                          * a normal userland page fault, just let
410                          * userland return so the signal will be
411                          * handled and gdb debugging works.  The page
412                          * fault code immediately after we return from
413                          * this function is going to release the
414                          * mmap_sem and it's not depending on it
415                          * (unlike gup would if we were not to return
416                          * VM_FAULT_RETRY).
417                          *
418                          * If a fatal signal is pending we still take
419                          * the streamlined VM_FAULT_RETRY failure path
420                          * and there's no need to retake the mmap_sem
421                          * in such case.
422                          */
423                         down_read(&mm->mmap_sem);
424                         ret = 0;
425                 }
426         }
427 
428         /*
429          * Here we race with the list_del; list_add in
430          * userfaultfd_ctx_read(), however because we don't ever run
431          * list_del_init() to refile across the two lists, the prev
432          * and next pointers will never point to self. list_add also
433          * would never let any of the two pointers to point to
434          * self. So list_empty_careful won't risk to see both pointers
435          * pointing to self at any time during the list refile. The
436          * only case where list_del_init() is called is the full
437          * removal in the wake function and there we don't re-list_add
438          * and it's fine not to block on the spinlock. The uwq on this
439          * kernel stack can be released after the list_del_init.
440          */
441         if (!list_empty_careful(&uwq.wq.task_list)) {
442                 spin_lock(&ctx->fault_pending_wqh.lock);
443                 /*
444                  * No need of list_del_init(), the uwq on the stack
445                  * will be freed shortly anyway.
446                  */
447                 list_del(&uwq.wq.task_list);
448                 spin_unlock(&ctx->fault_pending_wqh.lock);
449         }
450 
451         /*
452          * ctx may go away after this if the userfault pseudo fd is
453          * already released.
454          */
455         userfaultfd_ctx_put(ctx);
456 
457 out:
458         return ret;
459 }
460 
461 static int userfaultfd_release(struct inode *inode, struct file *file)
462 {
463         struct userfaultfd_ctx *ctx = file->private_data;
464         struct mm_struct *mm = ctx->mm;
465         struct vm_area_struct *vma, *prev;
466         /* len == 0 means wake all */
467         struct userfaultfd_wake_range range = { .len = 0, };
468         unsigned long new_flags;
469 
470         ACCESS_ONCE(ctx->released) = true;
471 
472         if (!mmget_not_zero(mm))
473                 goto wakeup;
474 
475         /*
476          * Flush page faults out of all CPUs. NOTE: all page faults
477          * must be retried without returning VM_FAULT_SIGBUS if
478          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
479          * changes while handle_userfault released the mmap_sem. So
480          * it's critical that released is set to true (above), before
481          * taking the mmap_sem for writing.
482          */
483         down_write(&mm->mmap_sem);
484         prev = NULL;
485         for (vma = mm->mmap; vma; vma = vma->vm_next) {
486                 cond_resched();
487                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
488                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
489                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
490                         prev = vma;
491                         continue;
492                 }
493                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
494                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
495                                  new_flags, vma->anon_vma,
496                                  vma->vm_file, vma->vm_pgoff,
497                                  vma_policy(vma),
498                                  NULL_VM_UFFD_CTX);
499                 if (prev)
500                         vma = prev;
501                 else
502                         prev = vma;
503                 vma->vm_flags = new_flags;
504                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
505         }
506         up_write(&mm->mmap_sem);
507         mmput(mm);
508 wakeup:
509         /*
510          * After no new page faults can wait on this fault_*wqh, flush
511          * the last page faults that may have been already waiting on
512          * the fault_*wqh.
513          */
514         spin_lock(&ctx->fault_pending_wqh.lock);
515         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
516         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
517         spin_unlock(&ctx->fault_pending_wqh.lock);
518 
519         wake_up_poll(&ctx->fd_wqh, POLLHUP);
520         userfaultfd_ctx_put(ctx);
521         return 0;
522 }
523 
524 /* fault_pending_wqh.lock must be hold by the caller */
525 static inline struct userfaultfd_wait_queue *find_userfault(
526         struct userfaultfd_ctx *ctx)
527 {
528         wait_queue_t *wq;
529         struct userfaultfd_wait_queue *uwq;
530 
531         VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock));
532 
533         uwq = NULL;
534         if (!waitqueue_active(&ctx->fault_pending_wqh))
535                 goto out;
536         /* walk in reverse to provide FIFO behavior to read userfaults */
537         wq = list_last_entry(&ctx->fault_pending_wqh.task_list,
538                              typeof(*wq), task_list);
539         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
540 out:
541         return uwq;
542 }
543 
544 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
545 {
546         struct userfaultfd_ctx *ctx = file->private_data;
547         unsigned int ret;
548 
549         poll_wait(file, &ctx->fd_wqh, wait);
550 
551         switch (ctx->state) {
552         case UFFD_STATE_WAIT_API:
553                 return POLLERR;
554         case UFFD_STATE_RUNNING:
555                 /*
556                  * poll() never guarantees that read won't block.
557                  * userfaults can be waken before they're read().
558                  */
559                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
560                         return POLLERR;
561                 /*
562                  * lockless access to see if there are pending faults
563                  * __pollwait last action is the add_wait_queue but
564                  * the spin_unlock would allow the waitqueue_active to
565                  * pass above the actual list_add inside
566                  * add_wait_queue critical section. So use a full
567                  * memory barrier to serialize the list_add write of
568                  * add_wait_queue() with the waitqueue_active read
569                  * below.
570                  */
571                 ret = 0;
572                 smp_mb();
573                 if (waitqueue_active(&ctx->fault_pending_wqh))
574                         ret = POLLIN;
575                 return ret;
576         default:
577                 BUG();
578         }
579 }
580 
581 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
582                                     struct uffd_msg *msg)
583 {
584         ssize_t ret;
585         DECLARE_WAITQUEUE(wait, current);
586         struct userfaultfd_wait_queue *uwq;
587 
588         /* always take the fd_wqh lock before the fault_pending_wqh lock */
589         spin_lock(&ctx->fd_wqh.lock);
590         __add_wait_queue(&ctx->fd_wqh, &wait);
591         for (;;) {
592                 set_current_state(TASK_INTERRUPTIBLE);
593                 spin_lock(&ctx->fault_pending_wqh.lock);
594                 uwq = find_userfault(ctx);
595                 if (uwq) {
596                         /*
597                          * Use a seqcount to repeat the lockless check
598                          * in wake_userfault() to avoid missing
599                          * wakeups because during the refile both
600                          * waitqueue could become empty if this is the
601                          * only userfault.
602                          */
603                         write_seqcount_begin(&ctx->refile_seq);
604 
605                         /*
606                          * The fault_pending_wqh.lock prevents the uwq
607                          * to disappear from under us.
608                          *
609                          * Refile this userfault from
610                          * fault_pending_wqh to fault_wqh, it's not
611                          * pending anymore after we read it.
612                          *
613                          * Use list_del() by hand (as
614                          * userfaultfd_wake_function also uses
615                          * list_del_init() by hand) to be sure nobody
616                          * changes __remove_wait_queue() to use
617                          * list_del_init() in turn breaking the
618                          * !list_empty_careful() check in
619                          * handle_userfault(). The uwq->wq.task_list
620                          * must never be empty at any time during the
621                          * refile, or the waitqueue could disappear
622                          * from under us. The "wait_queue_head_t"
623                          * parameter of __remove_wait_queue() is unused
624                          * anyway.
625                          */
626                         list_del(&uwq->wq.task_list);
627                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
628 
629                         write_seqcount_end(&ctx->refile_seq);
630 
631                         /* careful to always initialize msg if ret == 0 */
632                         *msg = uwq->msg;
633                         spin_unlock(&ctx->fault_pending_wqh.lock);
634                         ret = 0;
635                         break;
636                 }
637                 spin_unlock(&ctx->fault_pending_wqh.lock);
638                 if (signal_pending(current)) {
639                         ret = -ERESTARTSYS;
640                         break;
641                 }
642                 if (no_wait) {
643                         ret = -EAGAIN;
644                         break;
645                 }
646                 spin_unlock(&ctx->fd_wqh.lock);
647                 schedule();
648                 spin_lock(&ctx->fd_wqh.lock);
649         }
650         __remove_wait_queue(&ctx->fd_wqh, &wait);
651         __set_current_state(TASK_RUNNING);
652         spin_unlock(&ctx->fd_wqh.lock);
653 
654         return ret;
655 }
656 
657 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
658                                 size_t count, loff_t *ppos)
659 {
660         struct userfaultfd_ctx *ctx = file->private_data;
661         ssize_t _ret, ret = 0;
662         struct uffd_msg msg;
663         int no_wait = file->f_flags & O_NONBLOCK;
664 
665         if (ctx->state == UFFD_STATE_WAIT_API)
666                 return -EINVAL;
667 
668         for (;;) {
669                 if (count < sizeof(msg))
670                         return ret ? ret : -EINVAL;
671                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
672                 if (_ret < 0)
673                         return ret ? ret : _ret;
674                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
675                         return ret ? ret : -EFAULT;
676                 ret += sizeof(msg);
677                 buf += sizeof(msg);
678                 count -= sizeof(msg);
679                 /*
680                  * Allow to read more than one fault at time but only
681                  * block if waiting for the very first one.
682                  */
683                 no_wait = O_NONBLOCK;
684         }
685 }
686 
687 static void __wake_userfault(struct userfaultfd_ctx *ctx,
688                              struct userfaultfd_wake_range *range)
689 {
690         unsigned long start, end;
691 
692         start = range->start;
693         end = range->start + range->len;
694 
695         spin_lock(&ctx->fault_pending_wqh.lock);
696         /* wake all in the range and autoremove */
697         if (waitqueue_active(&ctx->fault_pending_wqh))
698                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
699                                      range);
700         if (waitqueue_active(&ctx->fault_wqh))
701                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
702         spin_unlock(&ctx->fault_pending_wqh.lock);
703 }
704 
705 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
706                                            struct userfaultfd_wake_range *range)
707 {
708         unsigned seq;
709         bool need_wakeup;
710 
711         /*
712          * To be sure waitqueue_active() is not reordered by the CPU
713          * before the pagetable update, use an explicit SMP memory
714          * barrier here. PT lock release or up_read(mmap_sem) still
715          * have release semantics that can allow the
716          * waitqueue_active() to be reordered before the pte update.
717          */
718         smp_mb();
719 
720         /*
721          * Use waitqueue_active because it's very frequent to
722          * change the address space atomically even if there are no
723          * userfaults yet. So we take the spinlock only when we're
724          * sure we've userfaults to wake.
725          */
726         do {
727                 seq = read_seqcount_begin(&ctx->refile_seq);
728                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
729                         waitqueue_active(&ctx->fault_wqh);
730                 cond_resched();
731         } while (read_seqcount_retry(&ctx->refile_seq, seq));
732         if (need_wakeup)
733                 __wake_userfault(ctx, range);
734 }
735 
736 static __always_inline int validate_range(struct mm_struct *mm,
737                                           __u64 start, __u64 len)
738 {
739         __u64 task_size = mm->task_size;
740 
741         if (start & ~PAGE_MASK)
742                 return -EINVAL;
743         if (len & ~PAGE_MASK)
744                 return -EINVAL;
745         if (!len)
746                 return -EINVAL;
747         if (start < mmap_min_addr)
748                 return -EINVAL;
749         if (start >= task_size)
750                 return -EINVAL;
751         if (len > task_size - start)
752                 return -EINVAL;
753         return 0;
754 }
755 
756 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
757                                 unsigned long arg)
758 {
759         struct mm_struct *mm = ctx->mm;
760         struct vm_area_struct *vma, *prev, *cur;
761         int ret;
762         struct uffdio_register uffdio_register;
763         struct uffdio_register __user *user_uffdio_register;
764         unsigned long vm_flags, new_flags;
765         bool found;
766         unsigned long start, end, vma_end;
767 
768         user_uffdio_register = (struct uffdio_register __user *) arg;
769 
770         ret = -EFAULT;
771         if (copy_from_user(&uffdio_register, user_uffdio_register,
772                            sizeof(uffdio_register)-sizeof(__u64)))
773                 goto out;
774 
775         ret = -EINVAL;
776         if (!uffdio_register.mode)
777                 goto out;
778         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
779                                      UFFDIO_REGISTER_MODE_WP))
780                 goto out;
781         vm_flags = 0;
782         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
783                 vm_flags |= VM_UFFD_MISSING;
784         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
785                 vm_flags |= VM_UFFD_WP;
786                 /*
787                  * FIXME: remove the below error constraint by
788                  * implementing the wprotect tracking mode.
789                  */
790                 ret = -EINVAL;
791                 goto out;
792         }
793 
794         ret = validate_range(mm, uffdio_register.range.start,
795                              uffdio_register.range.len);
796         if (ret)
797                 goto out;
798 
799         start = uffdio_register.range.start;
800         end = start + uffdio_register.range.len;
801 
802         ret = -ENOMEM;
803         if (!mmget_not_zero(mm))
804                 goto out;
805 
806         down_write(&mm->mmap_sem);
807         vma = find_vma_prev(mm, start, &prev);
808         if (!vma)
809                 goto out_unlock;
810 
811         /* check that there's at least one vma in the range */
812         ret = -EINVAL;
813         if (vma->vm_start >= end)
814                 goto out_unlock;
815 
816         /*
817          * Search for not compatible vmas.
818          *
819          * FIXME: this shall be relaxed later so that it doesn't fail
820          * on tmpfs backed vmas (in addition to the current allowance
821          * on anonymous vmas).
822          */
823         found = false;
824         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
825                 cond_resched();
826 
827                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
828                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
829 
830                 /* check not compatible vmas */
831                 ret = -EINVAL;
832                 if (cur->vm_ops)
833                         goto out_unlock;
834 
835                 /*
836                  * Check that this vma isn't already owned by a
837                  * different userfaultfd. We can't allow more than one
838                  * userfaultfd to own a single vma simultaneously or we
839                  * wouldn't know which one to deliver the userfaults to.
840                  */
841                 ret = -EBUSY;
842                 if (cur->vm_userfaultfd_ctx.ctx &&
843                     cur->vm_userfaultfd_ctx.ctx != ctx)
844                         goto out_unlock;
845 
846                 found = true;
847         }
848         BUG_ON(!found);
849 
850         if (vma->vm_start < start)
851                 prev = vma;
852 
853         ret = 0;
854         do {
855                 cond_resched();
856 
857                 BUG_ON(vma->vm_ops);
858                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
859                        vma->vm_userfaultfd_ctx.ctx != ctx);
860 
861                 /*
862                  * Nothing to do: this vma is already registered into this
863                  * userfaultfd and with the right tracking mode too.
864                  */
865                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
866                     (vma->vm_flags & vm_flags) == vm_flags)
867                         goto skip;
868 
869                 if (vma->vm_start > start)
870                         start = vma->vm_start;
871                 vma_end = min(end, vma->vm_end);
872 
873                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
874                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
875                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
876                                  vma_policy(vma),
877                                  ((struct vm_userfaultfd_ctx){ ctx }));
878                 if (prev) {
879                         vma = prev;
880                         goto next;
881                 }
882                 if (vma->vm_start < start) {
883                         ret = split_vma(mm, vma, start, 1);
884                         if (ret)
885                                 break;
886                 }
887                 if (vma->vm_end > end) {
888                         ret = split_vma(mm, vma, end, 0);
889                         if (ret)
890                                 break;
891                 }
892         next:
893                 /*
894                  * In the vma_merge() successful mprotect-like case 8:
895                  * the next vma was merged into the current one and
896                  * the current one has not been updated yet.
897                  */
898                 vma->vm_flags = new_flags;
899                 vma->vm_userfaultfd_ctx.ctx = ctx;
900 
901         skip:
902                 prev = vma;
903                 start = vma->vm_end;
904                 vma = vma->vm_next;
905         } while (vma && vma->vm_start < end);
906 out_unlock:
907         up_write(&mm->mmap_sem);
908         mmput(mm);
909         if (!ret) {
910                 /*
911                  * Now that we scanned all vmas we can already tell
912                  * userland which ioctls methods are guaranteed to
913                  * succeed on this range.
914                  */
915                 if (put_user(UFFD_API_RANGE_IOCTLS,
916                              &user_uffdio_register->ioctls))
917                         ret = -EFAULT;
918         }
919 out:
920         return ret;
921 }
922 
923 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
924                                   unsigned long arg)
925 {
926         struct mm_struct *mm = ctx->mm;
927         struct vm_area_struct *vma, *prev, *cur;
928         int ret;
929         struct uffdio_range uffdio_unregister;
930         unsigned long new_flags;
931         bool found;
932         unsigned long start, end, vma_end;
933         const void __user *buf = (void __user *)arg;
934 
935         ret = -EFAULT;
936         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
937                 goto out;
938 
939         ret = validate_range(mm, uffdio_unregister.start,
940                              uffdio_unregister.len);
941         if (ret)
942                 goto out;
943 
944         start = uffdio_unregister.start;
945         end = start + uffdio_unregister.len;
946 
947         ret = -ENOMEM;
948         if (!mmget_not_zero(mm))
949                 goto out;
950 
951         down_write(&mm->mmap_sem);
952         vma = find_vma_prev(mm, start, &prev);
953         if (!vma)
954                 goto out_unlock;
955 
956         /* check that there's at least one vma in the range */
957         ret = -EINVAL;
958         if (vma->vm_start >= end)
959                 goto out_unlock;
960 
961         /*
962          * Search for not compatible vmas.
963          *
964          * FIXME: this shall be relaxed later so that it doesn't fail
965          * on tmpfs backed vmas (in addition to the current allowance
966          * on anonymous vmas).
967          */
968         found = false;
969         ret = -EINVAL;
970         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
971                 cond_resched();
972 
973                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
974                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
975 
976                 /*
977                  * Check not compatible vmas, not strictly required
978                  * here as not compatible vmas cannot have an
979                  * userfaultfd_ctx registered on them, but this
980                  * provides for more strict behavior to notice
981                  * unregistration errors.
982                  */
983                 if (cur->vm_ops)
984                         goto out_unlock;
985 
986                 found = true;
987         }
988         BUG_ON(!found);
989 
990         if (vma->vm_start < start)
991                 prev = vma;
992 
993         ret = 0;
994         do {
995                 cond_resched();
996 
997                 BUG_ON(vma->vm_ops);
998 
999                 /*
1000                  * Nothing to do: this vma is already registered into this
1001                  * userfaultfd and with the right tracking mode too.
1002                  */
1003                 if (!vma->vm_userfaultfd_ctx.ctx)
1004                         goto skip;
1005 
1006                 if (vma->vm_start > start)
1007                         start = vma->vm_start;
1008                 vma_end = min(end, vma->vm_end);
1009 
1010                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1011                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1012                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1013                                  vma_policy(vma),
1014                                  NULL_VM_UFFD_CTX);
1015                 if (prev) {
1016                         vma = prev;
1017                         goto next;
1018                 }
1019                 if (vma->vm_start < start) {
1020                         ret = split_vma(mm, vma, start, 1);
1021                         if (ret)
1022                                 break;
1023                 }
1024                 if (vma->vm_end > end) {
1025                         ret = split_vma(mm, vma, end, 0);
1026                         if (ret)
1027                                 break;
1028                 }
1029         next:
1030                 /*
1031                  * In the vma_merge() successful mprotect-like case 8:
1032                  * the next vma was merged into the current one and
1033                  * the current one has not been updated yet.
1034                  */
1035                 vma->vm_flags = new_flags;
1036                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1037 
1038         skip:
1039                 prev = vma;
1040                 start = vma->vm_end;
1041                 vma = vma->vm_next;
1042         } while (vma && vma->vm_start < end);
1043 out_unlock:
1044         up_write(&mm->mmap_sem);
1045         mmput(mm);
1046 out:
1047         return ret;
1048 }
1049 
1050 /*
1051  * userfaultfd_wake may be used in combination with the
1052  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1053  */
1054 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1055                             unsigned long arg)
1056 {
1057         int ret;
1058         struct uffdio_range uffdio_wake;
1059         struct userfaultfd_wake_range range;
1060         const void __user *buf = (void __user *)arg;
1061 
1062         ret = -EFAULT;
1063         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1064                 goto out;
1065 
1066         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1067         if (ret)
1068                 goto out;
1069 
1070         range.start = uffdio_wake.start;
1071         range.len = uffdio_wake.len;
1072 
1073         /*
1074          * len == 0 means wake all and we don't want to wake all here,
1075          * so check it again to be sure.
1076          */
1077         VM_BUG_ON(!range.len);
1078 
1079         wake_userfault(ctx, &range);
1080         ret = 0;
1081 
1082 out:
1083         return ret;
1084 }
1085 
1086 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1087                             unsigned long arg)
1088 {
1089         __s64 ret;
1090         struct uffdio_copy uffdio_copy;
1091         struct uffdio_copy __user *user_uffdio_copy;
1092         struct userfaultfd_wake_range range;
1093 
1094         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1095 
1096         ret = -EFAULT;
1097         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1098                            /* don't copy "copy" last field */
1099                            sizeof(uffdio_copy)-sizeof(__s64)))
1100                 goto out;
1101 
1102         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1103         if (ret)
1104                 goto out;
1105         /*
1106          * double check for wraparound just in case. copy_from_user()
1107          * will later check uffdio_copy.src + uffdio_copy.len to fit
1108          * in the userland range.
1109          */
1110         ret = -EINVAL;
1111         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1112                 goto out;
1113         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1114                 goto out;
1115         if (mmget_not_zero(ctx->mm)) {
1116                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1117                                    uffdio_copy.len);
1118                 mmput(ctx->mm);
1119         }
1120         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1121                 return -EFAULT;
1122         if (ret < 0)
1123                 goto out;
1124         BUG_ON(!ret);
1125         /* len == 0 would wake all */
1126         range.len = ret;
1127         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1128                 range.start = uffdio_copy.dst;
1129                 wake_userfault(ctx, &range);
1130         }
1131         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1132 out:
1133         return ret;
1134 }
1135 
1136 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1137                                 unsigned long arg)
1138 {
1139         __s64 ret;
1140         struct uffdio_zeropage uffdio_zeropage;
1141         struct uffdio_zeropage __user *user_uffdio_zeropage;
1142         struct userfaultfd_wake_range range;
1143 
1144         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1145 
1146         ret = -EFAULT;
1147         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1148                            /* don't copy "zeropage" last field */
1149                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1150                 goto out;
1151 
1152         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1153                              uffdio_zeropage.range.len);
1154         if (ret)
1155                 goto out;
1156         ret = -EINVAL;
1157         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1158                 goto out;
1159 
1160         if (mmget_not_zero(ctx->mm)) {
1161                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1162                                      uffdio_zeropage.range.len);
1163                 mmput(ctx->mm);
1164         }
1165         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1166                 return -EFAULT;
1167         if (ret < 0)
1168                 goto out;
1169         /* len == 0 would wake all */
1170         BUG_ON(!ret);
1171         range.len = ret;
1172         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1173                 range.start = uffdio_zeropage.range.start;
1174                 wake_userfault(ctx, &range);
1175         }
1176         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1177 out:
1178         return ret;
1179 }
1180 
1181 /*
1182  * userland asks for a certain API version and we return which bits
1183  * and ioctl commands are implemented in this kernel for such API
1184  * version or -EINVAL if unknown.
1185  */
1186 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1187                            unsigned long arg)
1188 {
1189         struct uffdio_api uffdio_api;
1190         void __user *buf = (void __user *)arg;
1191         int ret;
1192 
1193         ret = -EINVAL;
1194         if (ctx->state != UFFD_STATE_WAIT_API)
1195                 goto out;
1196         ret = -EFAULT;
1197         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1198                 goto out;
1199         if (uffdio_api.api != UFFD_API || uffdio_api.features) {
1200                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1201                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1202                         goto out;
1203                 ret = -EINVAL;
1204                 goto out;
1205         }
1206         uffdio_api.features = UFFD_API_FEATURES;
1207         uffdio_api.ioctls = UFFD_API_IOCTLS;
1208         ret = -EFAULT;
1209         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1210                 goto out;
1211         ctx->state = UFFD_STATE_RUNNING;
1212         ret = 0;
1213 out:
1214         return ret;
1215 }
1216 
1217 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1218                               unsigned long arg)
1219 {
1220         int ret = -EINVAL;
1221         struct userfaultfd_ctx *ctx = file->private_data;
1222 
1223         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1224                 return -EINVAL;
1225 
1226         switch(cmd) {
1227         case UFFDIO_API:
1228                 ret = userfaultfd_api(ctx, arg);
1229                 break;
1230         case UFFDIO_REGISTER:
1231                 ret = userfaultfd_register(ctx, arg);
1232                 break;
1233         case UFFDIO_UNREGISTER:
1234                 ret = userfaultfd_unregister(ctx, arg);
1235                 break;
1236         case UFFDIO_WAKE:
1237                 ret = userfaultfd_wake(ctx, arg);
1238                 break;
1239         case UFFDIO_COPY:
1240                 ret = userfaultfd_copy(ctx, arg);
1241                 break;
1242         case UFFDIO_ZEROPAGE:
1243                 ret = userfaultfd_zeropage(ctx, arg);
1244                 break;
1245         }
1246         return ret;
1247 }
1248 
1249 #ifdef CONFIG_PROC_FS
1250 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1251 {
1252         struct userfaultfd_ctx *ctx = f->private_data;
1253         wait_queue_t *wq;
1254         struct userfaultfd_wait_queue *uwq;
1255         unsigned long pending = 0, total = 0;
1256 
1257         spin_lock(&ctx->fault_pending_wqh.lock);
1258         list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1259                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1260                 pending++;
1261                 total++;
1262         }
1263         list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1264                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1265                 total++;
1266         }
1267         spin_unlock(&ctx->fault_pending_wqh.lock);
1268 
1269         /*
1270          * If more protocols will be added, there will be all shown
1271          * separated by a space. Like this:
1272          *      protocols: aa:... bb:...
1273          */
1274         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1275                    pending, total, UFFD_API, UFFD_API_FEATURES,
1276                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1277 }
1278 #endif
1279 
1280 static const struct file_operations userfaultfd_fops = {
1281 #ifdef CONFIG_PROC_FS
1282         .show_fdinfo    = userfaultfd_show_fdinfo,
1283 #endif
1284         .release        = userfaultfd_release,
1285         .poll           = userfaultfd_poll,
1286         .read           = userfaultfd_read,
1287         .unlocked_ioctl = userfaultfd_ioctl,
1288         .compat_ioctl   = userfaultfd_ioctl,
1289         .llseek         = noop_llseek,
1290 };
1291 
1292 static void init_once_userfaultfd_ctx(void *mem)
1293 {
1294         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1295 
1296         init_waitqueue_head(&ctx->fault_pending_wqh);
1297         init_waitqueue_head(&ctx->fault_wqh);
1298         init_waitqueue_head(&ctx->fd_wqh);
1299         seqcount_init(&ctx->refile_seq);
1300 }
1301 
1302 /**
1303  * userfaultfd_file_create - Creates an userfaultfd file pointer.
1304  * @flags: Flags for the userfaultfd file.
1305  *
1306  * This function creates an userfaultfd file pointer, w/out installing
1307  * it into the fd table. This is useful when the userfaultfd file is
1308  * used during the initialization of data structures that require
1309  * extra setup after the userfaultfd creation. So the userfaultfd
1310  * creation is split into the file pointer creation phase, and the
1311  * file descriptor installation phase.  In this way races with
1312  * userspace closing the newly installed file descriptor can be
1313  * avoided.  Returns an userfaultfd file pointer, or a proper error
1314  * pointer.
1315  */
1316 static struct file *userfaultfd_file_create(int flags)
1317 {
1318         struct file *file;
1319         struct userfaultfd_ctx *ctx;
1320 
1321         BUG_ON(!current->mm);
1322 
1323         /* Check the UFFD_* constants for consistency.  */
1324         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1325         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1326 
1327         file = ERR_PTR(-EINVAL);
1328         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1329                 goto out;
1330 
1331         file = ERR_PTR(-ENOMEM);
1332         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1333         if (!ctx)
1334                 goto out;
1335 
1336         atomic_set(&ctx->refcount, 1);
1337         ctx->flags = flags;
1338         ctx->state = UFFD_STATE_WAIT_API;
1339         ctx->released = false;
1340         ctx->mm = current->mm;
1341         /* prevent the mm struct to be freed */
1342         atomic_inc(&ctx->mm->mm_count);
1343 
1344         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1345                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1346         if (IS_ERR(file)) {
1347                 mmdrop(ctx->mm);
1348                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1349         }
1350 out:
1351         return file;
1352 }
1353 
1354 SYSCALL_DEFINE1(userfaultfd, int, flags)
1355 {
1356         int fd, error;
1357         struct file *file;
1358 
1359         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1360         if (error < 0)
1361                 return error;
1362         fd = error;
1363 
1364         file = userfaultfd_file_create(flags);
1365         if (IS_ERR(file)) {
1366                 error = PTR_ERR(file);
1367                 goto err_put_unused_fd;
1368         }
1369         fd_install(fd, file);
1370 
1371         return fd;
1372 
1373 err_put_unused_fd:
1374         put_unused_fd(fd);
1375 
1376         return error;
1377 }
1378 
1379 static int __init userfaultfd_init(void)
1380 {
1381         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1382                                                 sizeof(struct userfaultfd_ctx),
1383                                                 0,
1384                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1385                                                 init_once_userfaultfd_ctx);
1386         return 0;
1387 }
1388 __initcall(userfaultfd_init);
1389 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us