fs/userfaultfd.c

   1 /*
   2  *  fs/userfaultfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc.
   6  *  Copyright (C) 2015  Red Hat, Inc.
   7  *
   8  *  This work is licensed under the terms of the GNU GPL, version 2. See
   9  *  the COPYING file in the top-level directory.
  10  *
  11  *  Some part derived from fs/eventfd.c (anon inode setup) and
  12  *  mm/ksm.c (mm hashing).
  13  */
  14
  15 #include <linux/list.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/mm.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  34
  35 enum userfaultfd_state {
  36         UFFD_STATE_WAIT_API,
  37         UFFD_STATE_RUNNING,
  38 };
  39
  40 /*
  41  * Start with fault_pending_wqh and fault_wqh so they're more likely
  42  * to be in the same cacheline.
  43  */
  44 struct userfaultfd_ctx {
  45         /* waitqueue head for the pending (i.e. not read) userfaults */
  46         wait_queue_head_t fault_pending_wqh;
  47         /* waitqueue head for the userfaults */
  48         wait_queue_head_t fault_wqh;
  49         /* waitqueue head for the pseudo fd to wakeup poll/read */
  50         wait_queue_head_t fd_wqh;
  51         /* waitqueue head for events */
  52         wait_queue_head_t event_wqh;
  53         /* a refile sequence protected by fault_pending_wqh lock */
  54         struct seqcount refile_seq;
  55         /* pseudo fd refcounting */
  56         atomic_t refcount;
  57         /* userfaultfd syscall flags */
  58         unsigned int flags;
  59         /* features requested from the userspace */
  60         unsigned int features;
  61         /* state machine */
  62         enum userfaultfd_state state;
  63         /* released */
  64         bool released;
  65         /* 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         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 740                 vm_ctx->ctx = ctx;
 741                 userfaultfd_ctx_get(ctx);
 742                 WRITE_ONCE(ctx->mmap_changing, true);
 743         }
 744 }
 745
 746 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 747                                  unsigned long from, unsigned long to,
 748                                  unsigned long len)
 749 {
 750         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 751         struct userfaultfd_wait_queue ewq;
 752
 753         if (!ctx)
 754                 return;
 755
 756         if (to & ~PAGE_MASK) {
 757                 userfaultfd_ctx_put(ctx);
 758                 return;
 759         }
 760
 761         msg_init(&ewq.msg);
 762
 763         ewq.msg.event = UFFD_EVENT_REMAP;
 764         ewq.msg.arg.remap.from = from;
 765         ewq.msg.arg.remap.to = to;
 766         ewq.msg.arg.remap.len = len;
 767
 768         userfaultfd_event_wait_completion(ctx, &ewq);
 769 }
 770
 771 bool userfaultfd_remove(struct vm_area_struct *vma,
 772                         unsigned long start, unsigned long end)
 773 {
 774         struct mm_struct *mm = vma->vm_mm;
 775         struct userfaultfd_ctx *ctx;
 776         struct userfaultfd_wait_queue ewq;
 777
 778         ctx = vma->vm_userfaultfd_ctx.ctx;
 779         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 780                 return true;
 781
 782         userfaultfd_ctx_get(ctx);
 783         WRITE_ONCE(ctx->mmap_changing, true);
 784         up_read(&mm->mmap_sem);
 785
 786         msg_init(&ewq.msg);
 787
 788         ewq.msg.event = UFFD_EVENT_REMOVE;
 789         ewq.msg.arg.remove.start = start;
 790         ewq.msg.arg.remove.end = end;
 791
 792         userfaultfd_event_wait_completion(ctx, &ewq);
 793
 794         return false;
 795 }
 796
 797 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 798                           unsigned long start, unsigned long end)
 799 {
 800         struct userfaultfd_unmap_ctx *unmap_ctx;
 801
 802         list_for_each_entry(unmap_ctx, unmaps, list)
 803                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 804                     unmap_ctx->end == end)
 805                         return true;
 806
 807         return false;
 808 }
 809
 810 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 811                            unsigned long start, unsigned long end,
 812                            struct list_head *unmaps)
 813 {
 814         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 815                 struct userfaultfd_unmap_ctx *unmap_ctx;
 816                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 817
 818                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 819                     has_unmap_ctx(ctx, unmaps, start, end))
 820                         continue;
 821
 822                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 823                 if (!unmap_ctx)
 824                         return -ENOMEM;
 825
 826                 userfaultfd_ctx_get(ctx);
 827                 WRITE_ONCE(ctx->mmap_changing, true);
 828                 unmap_ctx->ctx = ctx;
 829                 unmap_ctx->start = start;
 830                 unmap_ctx->end = end;
 831                 list_add_tail(&unmap_ctx->list, unmaps);
 832         }
 833
 834         return 0;
 835 }
 836
 837 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 838 {
 839         struct userfaultfd_unmap_ctx *ctx, *n;
 840         struct userfaultfd_wait_queue ewq;
 841
 842         list_for_each_entry_safe(ctx, n, uf, list) {
 843                 msg_init(&ewq.msg);
 844
 845                 ewq.msg.event = UFFD_EVENT_UNMAP;
 846                 ewq.msg.arg.remove.start = ctx->start;
 847                 ewq.msg.arg.remove.end = ctx->end;
 848
 849                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 850
 851                 list_del(&ctx->list);
 852                 kfree(ctx);
 853         }
 854 }
 855
 856 static int userfaultfd_release(struct inode *inode, struct file *file)
 857 {
 858         struct userfaultfd_ctx *ctx = file->private_data;
 859         struct mm_struct *mm = ctx->mm;
 860         struct vm_area_struct *vma, *prev;
 861         /* len == 0 means wake all */
 862         struct userfaultfd_wake_range range = { .len = 0, };
 863         unsigned long new_flags;
 864
 865         WRITE_ONCE(ctx->released, true);
 866
 867         if (!mmget_not_zero(mm))
 868                 goto wakeup;
 869
 870         /*
 871          * Flush page faults out of all CPUs. NOTE: all page faults
 872          * must be retried without returning VM_FAULT_SIGBUS if
 873          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 874          * changes while handle_userfault released the mmap_sem. So
 875          * it's critical that released is set to true (above), before
 876          * taking the mmap_sem for writing.
 877          */
 878         down_write(&mm->mmap_sem);
 879         prev = NULL;
 880         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 881                 cond_resched();
 882                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 883                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 884                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 885                         prev = vma;
 886                         continue;
 887                 }
 888                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 889                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 890                                  new_flags, vma->anon_vma,
 891                                  vma->vm_file, vma->vm_pgoff,
 892                                  vma_policy(vma),
 893                                  NULL_VM_UFFD_CTX);
 894                 if (prev)
 895                         vma = prev;
 896                 else
 897                         prev = vma;
 898                 vma->vm_flags = new_flags;
 899                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 900         }
 901         up_write(&mm->mmap_sem);
 902         mmput(mm);
 903 wakeup:
 904         /*
 905          * After no new page faults can wait on this fault_*wqh, flush
 906          * the last page faults that may have been already waiting on
 907          * the fault_*wqh.
 908          */
 909         spin_lock(&ctx->fault_pending_wqh.lock);
 910         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 911         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 912         spin_unlock(&ctx->fault_pending_wqh.lock);
 913
 914         /* Flush pending events that may still wait on event_wqh */
 915         wake_up_all(&ctx->event_wqh);
 916
 917         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 918         userfaultfd_ctx_put(ctx);
 919         return 0;
 920 }
 921
 922 /* fault_pending_wqh.lock must be hold by the caller */
 923 static inline struct userfaultfd_wait_queue *find_userfault_in(
 924                 wait_queue_head_t *wqh)
 925 {
 926         wait_queue_entry_t *wq;
 927         struct userfaultfd_wait_queue *uwq;
 928
 929         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 930
 931         uwq = NULL;
 932         if (!waitqueue_active(wqh))
 933                 goto out;
 934         /* walk in reverse to provide FIFO behavior to read userfaults */
 935         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 936         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 937 out:
 938         return uwq;
 939 }
 940
 941 static inline struct userfaultfd_wait_queue *find_userfault(
 942                 struct userfaultfd_ctx *ctx)
 943 {
 944         return find_userfault_in(&ctx->fault_pending_wqh);
 945 }
 946
 947 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 948                 struct userfaultfd_ctx *ctx)
 949 {
 950         return find_userfault_in(&ctx->event_wqh);
 951 }
 952
 953 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 954 {
 955         struct userfaultfd_ctx *ctx = file->private_data;
 956         __poll_t ret;
 957
 958         poll_wait(file, &ctx->fd_wqh, wait);
 959
 960         switch (ctx->state) {
 961         case UFFD_STATE_WAIT_API:
 962                 return EPOLLERR;
 963         case UFFD_STATE_RUNNING:
 964                 /*
 965                  * poll() never guarantees that read won't block.
 966                  * userfaults can be waken before they're read().
 967                  */
 968                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 969                         return EPOLLERR;
 970                 /*
 971                  * lockless access to see if there are pending faults
 972                  * __pollwait last action is the add_wait_queue but
 973                  * the spin_unlock would allow the waitqueue_active to
 974                  * pass above the actual list_add inside
 975                  * add_wait_queue critical section. So use a full
 976                  * memory barrier to serialize the list_add write of
 977                  * add_wait_queue() with the waitqueue_active read
 978                  * below.
 979                  */
 980                 ret = 0;
 981                 smp_mb();
 982                 if (waitqueue_active(&ctx->fault_pending_wqh))
 983                         ret = EPOLLIN;
 984                 else if (waitqueue_active(&ctx->event_wqh))
 985                         ret = EPOLLIN;
 986
 987                 return ret;
 988         default:
 989                 WARN_ON_ONCE(1);
 990                 return EPOLLERR;
 991         }
 992 }
 993
 994 static const struct file_operations userfaultfd_fops;
 995
 996 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 997                                   struct userfaultfd_ctx *new,
 998                                   struct uffd_msg *msg)
 999 {
1000         int fd;
1001
1002         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1003                               O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
1004         if (fd < 0)
1005                 return fd;
1006
1007         msg->arg.reserved.reserved1 = 0;
1008         msg->arg.fork.ufd = fd;
1009         return 0;
1010 }
1011
1012 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1013                                     struct uffd_msg *msg)
1014 {
1015         ssize_t ret;
1016         DECLARE_WAITQUEUE(wait, current);
1017         struct userfaultfd_wait_queue *uwq;
1018         /*
1019          * Handling fork event requires sleeping operations, so
1020          * we drop the event_wqh lock, then do these ops, then
1021          * lock it back and wake up the waiter. While the lock is
1022          * dropped the ewq may go away so we keep track of it
1023          * carefully.
1024          */
1025         LIST_HEAD(fork_event);
1026         struct userfaultfd_ctx *fork_nctx = NULL;
1027
1028         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1029         spin_lock_irq(&ctx->fd_wqh.lock);
1030         __add_wait_queue(&ctx->fd_wqh, &wait);
1031         for (;;) {
1032                 set_current_state(TASK_INTERRUPTIBLE);
1033                 spin_lock(&ctx->fault_pending_wqh.lock);
1034                 uwq = find_userfault(ctx);
1035                 if (uwq) {
1036                         /*
1037                          * Use a seqcount to repeat the lockless check
1038                          * in wake_userfault() to avoid missing
1039                          * wakeups because during the refile both
1040                          * waitqueue could become empty if this is the
1041                          * only userfault.
1042                          */
1043                         write_seqcount_begin(&ctx->refile_seq);
1044
1045                         /*
1046                          * The fault_pending_wqh.lock prevents the uwq
1047                          * to disappear from under us.
1048                          *
1049                          * Refile this userfault from
1050                          * fault_pending_wqh to fault_wqh, it's not
1051                          * pending anymore after we read it.
1052                          *
1053                          * Use list_del() by hand (as
1054                          * userfaultfd_wake_function also uses
1055                          * list_del_init() by hand) to be sure nobody
1056                          * changes __remove_wait_queue() to use
1057                          * list_del_init() in turn breaking the
1058                          * !list_empty_careful() check in
1059                          * handle_userfault(). The uwq->wq.head list
1060                          * must never be empty at any time during the
1061                          * refile, or the waitqueue could disappear
1062                          * from under us. The "wait_queue_head_t"
1063                          * parameter of __remove_wait_queue() is unused
1064                          * anyway.
1065                          */
1066                         list_del(&uwq->wq.entry);
1067                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1068
1069                         write_seqcount_end(&ctx->refile_seq);
1070
1071                         /* careful to always initialize msg if ret == 0 */
1072                         *msg = uwq->msg;
1073                         spin_unlock(&ctx->fault_pending_wqh.lock);
1074                         ret = 0;
1075                         break;
1076                 }
1077                 spin_unlock(&ctx->fault_pending_wqh.lock);
1078
1079                 spin_lock(&ctx->event_wqh.lock);
1080                 uwq = find_userfault_evt(ctx);
1081                 if (uwq) {
1082                         *msg = uwq->msg;
1083
1084                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1085                                 fork_nctx = (struct userfaultfd_ctx *)
1086                                         (unsigned long)
1087                                         uwq->msg.arg.reserved.reserved1;
1088                                 list_move(&uwq->wq.entry, &fork_event);
1089                                 /*
1090                                  * fork_nctx can be freed as soon as
1091                                  * we drop the lock, unless we take a
1092                                  * reference on it.
1093                                  */
1094                                 userfaultfd_ctx_get(fork_nctx);
1095                                 spin_unlock(&ctx->event_wqh.lock);
1096                                 ret = 0;
1097                                 break;
1098                         }
1099
1100                         userfaultfd_event_complete(ctx, uwq);
1101                         spin_unlock(&ctx->event_wqh.lock);
1102                         ret = 0;
1103                         break;
1104                 }
1105                 spin_unlock(&ctx->event_wqh.lock);
1106
1107                 if (signal_pending(current)) {
1108                         ret = -ERESTARTSYS;
1109                         break;
1110                 }
1111                 if (no_wait) {
1112                         ret = -EAGAIN;
1113                         break;
1114                 }
1115                 spin_unlock_irq(&ctx->fd_wqh.lock);
1116                 schedule();
1117                 spin_lock_irq(&ctx->fd_wqh.lock);
1118         }
1119         __remove_wait_queue(&ctx->fd_wqh, &wait);
1120         __set_current_state(TASK_RUNNING);
1121         spin_unlock_irq(&ctx->fd_wqh.lock);
1122
1123         if (!ret && msg->event == UFFD_EVENT_FORK) {
1124                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1125                 spin_lock(&ctx->event_wqh.lock);
1126                 if (!list_empty(&fork_event)) {
1127                         /*
1128                          * The fork thread didn't abort, so we can
1129                          * drop the temporary refcount.
1130                          */
1131                         userfaultfd_ctx_put(fork_nctx);
1132
1133                         uwq = list_first_entry(&fork_event,
1134                                                typeof(*uwq),
1135                                                wq.entry);
1136                         /*
1137                          * If fork_event list wasn't empty and in turn
1138                          * the event wasn't already released by fork
1139                          * (the event is allocated on fork kernel
1140                          * stack), put the event back to its place in
1141                          * the event_wq. fork_event head will be freed
1142                          * as soon as we return so the event cannot
1143                          * stay queued there no matter the current
1144                          * "ret" value.
1145                          */
1146                         list_del(&uwq->wq.entry);
1147                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1148
1149                         /*
1150                          * Leave the event in the waitqueue and report
1151                          * error to userland if we failed to resolve
1152                          * the userfault fork.
1153                          */
1154                         if (likely(!ret))
1155                                 userfaultfd_event_complete(ctx, uwq);
1156                 } else {
1157                         /*
1158                          * Here the fork thread aborted and the
1159                          * refcount from the fork thread on fork_nctx
1160                          * has already been released. We still hold
1161                          * the reference we took before releasing the
1162                          * lock above. If resolve_userfault_fork
1163                          * failed we've to drop it because the
1164                          * fork_nctx has to be freed in such case. If
1165                          * it succeeded we'll hold it because the new
1166                          * uffd references it.
1167                          */
1168                         if (ret)
1169                                 userfaultfd_ctx_put(fork_nctx);
1170                 }
1171                 spin_unlock(&ctx->event_wqh.lock);
1172         }
1173
1174         return ret;
1175 }
1176
1177 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1178                                 size_t count, loff_t *ppos)
1179 {
1180         struct userfaultfd_ctx *ctx = file->private_data;
1181         ssize_t _ret, ret = 0;
1182         struct uffd_msg msg;
1183         int no_wait = file->f_flags & O_NONBLOCK;
1184
1185         if (ctx->state == UFFD_STATE_WAIT_API)
1186                 return -EINVAL;
1187
1188         for (;;) {
1189                 if (count < sizeof(msg))
1190                         return ret ? ret : -EINVAL;
1191                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1192                 if (_ret < 0)
1193                         return ret ? ret : _ret;
1194                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1195                         return ret ? ret : -EFAULT;
1196                 ret += sizeof(msg);
1197                 buf += sizeof(msg);
1198                 count -= sizeof(msg);
1199                 /*
1200                  * Allow to read more than one fault at time but only
1201                  * block if waiting for the very first one.
1202                  */
1203                 no_wait = O_NONBLOCK;
1204         }
1205 }
1206
1207 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1208                              struct userfaultfd_wake_range *range)
1209 {
1210         spin_lock(&ctx->fault_pending_wqh.lock);
1211         /* wake all in the range and autoremove */
1212         if (waitqueue_active(&ctx->fault_pending_wqh))
1213                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1214                                      range);
1215         if (waitqueue_active(&ctx->fault_wqh))
1216                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1217         spin_unlock(&ctx->fault_pending_wqh.lock);
1218 }
1219
1220 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1221                                            struct userfaultfd_wake_range *range)
1222 {
1223         unsigned seq;
1224         bool need_wakeup;
1225
1226         /*
1227          * To be sure waitqueue_active() is not reordered by the CPU
1228          * before the pagetable update, use an explicit SMP memory
1229          * barrier here. PT lock release or up_read(mmap_sem) still
1230          * have release semantics that can allow the
1231          * waitqueue_active() to be reordered before the pte update.
1232          */
1233         smp_mb();
1234
1235         /*
1236          * Use waitqueue_active because it's very frequent to
1237          * change the address space atomically even if there are no
1238          * userfaults yet. So we take the spinlock only when we're
1239          * sure we've userfaults to wake.
1240          */
1241         do {
1242                 seq = read_seqcount_begin(&ctx->refile_seq);
1243                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1244                         waitqueue_active(&ctx->fault_wqh);
1245                 cond_resched();
1246         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1247         if (need_wakeup)
1248                 __wake_userfault(ctx, range);
1249 }
1250
1251 static __always_inline int validate_range(struct mm_struct *mm,
1252                                           __u64 start, __u64 len)
1253 {
1254         __u64 task_size = mm->task_size;
1255
1256         if (start & ~PAGE_MASK)
1257                 return -EINVAL;
1258         if (len & ~PAGE_MASK)
1259                 return -EINVAL;
1260         if (!len)
1261                 return -EINVAL;
1262         if (start < mmap_min_addr)
1263                 return -EINVAL;
1264         if (start >= task_size)
1265                 return -EINVAL;
1266         if (len > task_size - start)
1267                 return -EINVAL;
1268         return 0;
1269 }
1270
1271 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1272 {
1273         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1274                 vma_is_shmem(vma);
1275 }
1276
1277 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1278                                 unsigned long arg)
1279 {
1280         struct mm_struct *mm = ctx->mm;
1281         struct vm_area_struct *vma, *prev, *cur;
1282         int ret;
1283         struct uffdio_register uffdio_register;
1284         struct uffdio_register __user *user_uffdio_register;
1285         unsigned long vm_flags, new_flags;
1286         bool found;
1287         bool basic_ioctls;
1288         unsigned long start, end, vma_end;
1289
1290         user_uffdio_register = (struct uffdio_register __user *) arg;
1291
1292         ret = -EFAULT;
1293         if (copy_from_user(&uffdio_register, user_uffdio_register,
1294                            sizeof(uffdio_register)-sizeof(__u64)))
1295                 goto out;
1296
1297         ret = -EINVAL;
1298         if (!uffdio_register.mode)
1299                 goto out;
1300         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1301                                      UFFDIO_REGISTER_MODE_WP))
1302                 goto out;
1303         vm_flags = 0;
1304         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1305                 vm_flags |= VM_UFFD_MISSING;
1306         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1307                 vm_flags |= VM_UFFD_WP;
1308                 /*
1309                  * FIXME: remove the below error constraint by
1310                  * implementing the wprotect tracking mode.
1311                  */
1312                 ret = -EINVAL;
1313                 goto out;
1314         }
1315
1316         ret = validate_range(mm, uffdio_register.range.start,
1317                              uffdio_register.range.len);
1318         if (ret)
1319                 goto out;
1320
1321         start = uffdio_register.range.start;
1322         end = start + uffdio_register.range.len;
1323
1324         ret = -ENOMEM;
1325         if (!mmget_not_zero(mm))
1326                 goto out;
1327
1328         down_write(&mm->mmap_sem);
1329         vma = find_vma_prev(mm, start, &prev);
1330         if (!vma)
1331                 goto out_unlock;
1332
1333         /* check that there's at least one vma in the range */
1334         ret = -EINVAL;
1335         if (vma->vm_start >= end)
1336                 goto out_unlock;
1337
1338         /*
1339          * If the first vma contains huge pages, make sure start address
1340          * is aligned to huge page size.
1341          */
1342         if (is_vm_hugetlb_page(vma)) {
1343                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1344
1345                 if (start & (vma_hpagesize - 1))
1346                         goto out_unlock;
1347         }
1348
1349         /*
1350          * Search for not compatible vmas.
1351          */
1352         found = false;
1353         basic_ioctls = false;
1354         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1355                 cond_resched();
1356
1357                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1358                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1359
1360                 /* check not compatible vmas */
1361                 ret = -EINVAL;
1362                 if (!vma_can_userfault(cur))
1363                         goto out_unlock;
1364
1365                 /*
1366                  * UFFDIO_COPY will fill file holes even without
1367                  * PROT_WRITE. This check enforces that if this is a
1368                  * MAP_SHARED, the process has write permission to the backing
1369                  * file. If VM_MAYWRITE is set it also enforces that on a
1370                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1371                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1372                  */
1373                 ret = -EPERM;
1374                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1375                         goto out_unlock;
1376
1377                 /*
1378                  * If this vma contains ending address, and huge pages
1379                  * check alignment.
1380                  */
1381                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1382                     end > cur->vm_start) {
1383                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1384
1385                         ret = -EINVAL;
1386
1387                         if (end & (vma_hpagesize - 1))
1388                                 goto out_unlock;
1389                 }
1390
1391                 /*
1392                  * Check that this vma isn't already owned by a
1393                  * different userfaultfd. We can't allow more than one
1394                  * userfaultfd to own a single vma simultaneously or we
1395                  * wouldn't know which one to deliver the userfaults to.
1396                  */
1397                 ret = -EBUSY;
1398                 if (cur->vm_userfaultfd_ctx.ctx &&
1399                     cur->vm_userfaultfd_ctx.ctx != ctx)
1400                         goto out_unlock;
1401
1402                 /*
1403                  * Note vmas containing huge pages
1404                  */
1405                 if (is_vm_hugetlb_page(cur))
1406                         basic_ioctls = true;
1407
1408                 found = true;
1409         }
1410         BUG_ON(!found);
1411
1412         if (vma->vm_start < start)
1413                 prev = vma;
1414
1415         ret = 0;
1416         do {
1417                 cond_resched();
1418
1419                 BUG_ON(!vma_can_userfault(vma));
1420                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1421                        vma->vm_userfaultfd_ctx.ctx != ctx);
1422                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1423
1424                 /*
1425                  * Nothing to do: this vma is already registered into this
1426                  * userfaultfd and with the right tracking mode too.
1427                  */
1428                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1429                     (vma->vm_flags & vm_flags) == vm_flags)
1430                         goto skip;
1431
1432                 if (vma->vm_start > start)
1433                         start = vma->vm_start;
1434                 vma_end = min(end, vma->vm_end);
1435
1436                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1437                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1438                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1439                                  vma_policy(vma),
1440                                  ((struct vm_userfaultfd_ctx){ ctx }));
1441                 if (prev) {
1442                         vma = prev;
1443                         goto next;
1444                 }
1445                 if (vma->vm_start < start) {
1446                         ret = split_vma(mm, vma, start, 1);
1447                         if (ret)
1448                                 break;
1449                 }
1450                 if (vma->vm_end > end) {
1451                         ret = split_vma(mm, vma, end, 0);
1452                         if (ret)
1453                                 break;
1454                 }
1455         next:
1456                 /*
1457                  * In the vma_merge() successful mprotect-like case 8:
1458                  * the next vma was merged into the current one and
1459                  * the current one has not been updated yet.
1460                  */
1461                 vma->vm_flags = new_flags;
1462                 vma->vm_userfaultfd_ctx.ctx = ctx;
1463
1464         skip:
1465                 prev = vma;
1466                 start = vma->vm_end;
1467                 vma = vma->vm_next;
1468         } while (vma && vma->vm_start < end);
1469 out_unlock:
1470         up_write(&mm->mmap_sem);
1471         mmput(mm);
1472         if (!ret) {
1473                 /*
1474                  * Now that we scanned all vmas we can already tell
1475                  * userland which ioctls methods are guaranteed to
1476                  * succeed on this range.
1477                  */
1478                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1479                              UFFD_API_RANGE_IOCTLS,
1480                              &user_uffdio_register->ioctls))
1481                         ret = -EFAULT;
1482         }
1483 out:
1484         return ret;
1485 }
1486
1487 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1488                                   unsigned long arg)
1489 {
1490         struct mm_struct *mm = ctx->mm;
1491         struct vm_area_struct *vma, *prev, *cur;
1492         int ret;
1493         struct uffdio_range uffdio_unregister;
1494         unsigned long new_flags;
1495         bool found;
1496         unsigned long start, end, vma_end;
1497         const void __user *buf = (void __user *)arg;
1498
1499         ret = -EFAULT;
1500         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1501                 goto out;
1502
1503         ret = validate_range(mm, uffdio_unregister.start,
1504                              uffdio_unregister.len);
1505         if (ret)
1506                 goto out;
1507
1508         start = uffdio_unregister.start;
1509         end = start + uffdio_unregister.len;
1510
1511         ret = -ENOMEM;
1512         if (!mmget_not_zero(mm))
1513                 goto out;
1514
1515         down_write(&mm->mmap_sem);
1516         vma = find_vma_prev(mm, start, &prev);
1517         if (!vma)
1518                 goto out_unlock;
1519
1520         /* check that there's at least one vma in the range */
1521         ret = -EINVAL;
1522         if (vma->vm_start >= end)
1523                 goto out_unlock;
1524
1525         /*
1526          * If the first vma contains huge pages, make sure start address
1527          * is aligned to huge page size.
1528          */
1529         if (is_vm_hugetlb_page(vma)) {
1530                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1531
1532                 if (start & (vma_hpagesize - 1))
1533                         goto out_unlock;
1534         }
1535
1536         /*
1537          * Search for not compatible vmas.
1538          */
1539         found = false;
1540         ret = -EINVAL;
1541         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1542                 cond_resched();
1543
1544                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1545                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1546
1547                 /*
1548                  * Check not compatible vmas, not strictly required
1549                  * here as not compatible vmas cannot have an
1550                  * userfaultfd_ctx registered on them, but this
1551                  * provides for more strict behavior to notice
1552                  * unregistration errors.
1553                  */
1554                 if (!vma_can_userfault(cur))
1555                         goto out_unlock;
1556
1557                 found = true;
1558         }
1559         BUG_ON(!found);
1560
1561         if (vma->vm_start < start)
1562                 prev = vma;
1563
1564         ret = 0;
1565         do {
1566                 cond_resched();
1567
1568                 BUG_ON(!vma_can_userfault(vma));
1569                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1570
1571                 /*
1572                  * Nothing to do: this vma is already registered into this
1573                  * userfaultfd and with the right tracking mode too.
1574                  */
1575                 if (!vma->vm_userfaultfd_ctx.ctx)
1576                         goto skip;
1577
1578                 if (vma->vm_start > start)
1579                         start = vma->vm_start;
1580                 vma_end = min(end, vma->vm_end);
1581
1582                 if (userfaultfd_missing(vma)) {
1583                         /*
1584                          * Wake any concurrent pending userfault while
1585                          * we unregister, so they will not hang
1586                          * permanently and it avoids userland to call
1587                          * UFFDIO_WAKE explicitly.
1588                          */
1589                         struct userfaultfd_wake_range range;
1590                         range.start = start;
1591                         range.len = vma_end - start;
1592                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1593                 }
1594
1595                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1596                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1597                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1598                                  vma_policy(vma),
1599                                  NULL_VM_UFFD_CTX);
1600                 if (prev) {
1601                         vma = prev;
1602                         goto next;
1603                 }
1604                 if (vma->vm_start < start) {
1605                         ret = split_vma(mm, vma, start, 1);
1606                         if (ret)
1607                                 break;
1608                 }
1609                 if (vma->vm_end > end) {
1610                         ret = split_vma(mm, vma, end, 0);
1611                         if (ret)
1612                                 break;
1613                 }
1614         next:
1615                 /*
1616                  * In the vma_merge() successful mprotect-like case 8:
1617                  * the next vma was merged into the current one and
1618                  * the current one has not been updated yet.
1619                  */
1620                 vma->vm_flags = new_flags;
1621                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1622
1623         skip:
1624                 prev = vma;
1625                 start = vma->vm_end;
1626                 vma = vma->vm_next;
1627         } while (vma && vma->vm_start < end);
1628 out_unlock:
1629         up_write(&mm->mmap_sem);
1630         mmput(mm);
1631 out:
1632         return ret;
1633 }
1634
1635 /*
1636  * userfaultfd_wake may be used in combination with the
1637  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1638  */
1639 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1640                             unsigned long arg)
1641 {
1642         int ret;
1643         struct uffdio_range uffdio_wake;
1644         struct userfaultfd_wake_range range;
1645         const void __user *buf = (void __user *)arg;
1646
1647         ret = -EFAULT;
1648         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1649                 goto out;
1650
1651         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1652         if (ret)
1653                 goto out;
1654
1655         range.start = uffdio_wake.start;
1656         range.len = uffdio_wake.len;
1657
1658         /*
1659          * len == 0 means wake all and we don't want to wake all here,
1660          * so check it again to be sure.
1661          */
1662         VM_BUG_ON(!range.len);
1663
1664         wake_userfault(ctx, &range);
1665         ret = 0;
1666
1667 out:
1668         return ret;
1669 }
1670
1671 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1672                             unsigned long arg)
1673 {
1674         __s64 ret;
1675         struct uffdio_copy uffdio_copy;
1676         struct uffdio_copy __user *user_uffdio_copy;
1677         struct userfaultfd_wake_range range;
1678
1679         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1680
1681         ret = -EAGAIN;
1682         if (READ_ONCE(ctx->mmap_changing))
1683                 goto out;
1684
1685         ret = -EFAULT;
1686         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1687                            /* don't copy "copy" last field */
1688                            sizeof(uffdio_copy)-sizeof(__s64)))
1689                 goto out;
1690
1691         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1692         if (ret)
1693                 goto out;
1694         /*
1695          * double check for wraparound just in case. copy_from_user()
1696          * will later check uffdio_copy.src + uffdio_copy.len to fit
1697          * in the userland range.
1698          */
1699         ret = -EINVAL;
1700         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1701                 goto out;
1702         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1703                 goto out;
1704         if (mmget_not_zero(ctx->mm)) {
1705                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1706                                    uffdio_copy.len, &ctx->mmap_changing);
1707                 mmput(ctx->mm);
1708         } else {
1709                 return -ESRCH;
1710         }
1711         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1712                 return -EFAULT;
1713         if (ret < 0)
1714                 goto out;
1715         BUG_ON(!ret);
1716         /* len == 0 would wake all */
1717         range.len = ret;
1718         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1719                 range.start = uffdio_copy.dst;
1720                 wake_userfault(ctx, &range);
1721         }
1722         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1723 out:
1724         return ret;
1725 }
1726
1727 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1728                                 unsigned long arg)
1729 {
1730         __s64 ret;
1731         struct uffdio_zeropage uffdio_zeropage;
1732         struct uffdio_zeropage __user *user_uffdio_zeropage;
1733         struct userfaultfd_wake_range range;
1734
1735         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1736
1737         ret = -EAGAIN;
1738         if (READ_ONCE(ctx->mmap_changing))
1739                 goto out;
1740
1741         ret = -EFAULT;
1742         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1743                            /* don't copy "zeropage" last field */
1744                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1745                 goto out;
1746
1747         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1748                              uffdio_zeropage.range.len);
1749         if (ret)
1750                 goto out;
1751         ret = -EINVAL;
1752         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1753                 goto out;
1754
1755         if (mmget_not_zero(ctx->mm)) {
1756                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1757                                      uffdio_zeropage.range.len,
1758                                      &ctx->mmap_changing);
1759                 mmput(ctx->mm);
1760         } else {
1761                 return -ESRCH;
1762         }
1763         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1764                 return -EFAULT;
1765         if (ret < 0)
1766                 goto out;
1767         /* len == 0 would wake all */
1768         BUG_ON(!ret);
1769         range.len = ret;
1770         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1771                 range.start = uffdio_zeropage.range.start;
1772                 wake_userfault(ctx, &range);
1773         }
1774         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1775 out:
1776         return ret;
1777 }
1778
1779 static inline unsigned int uffd_ctx_features(__u64 user_features)
1780 {
1781         /*
1782          * For the current set of features the bits just coincide
1783          */
1784         return (unsigned int)user_features;
1785 }
1786
1787 /*
1788  * userland asks for a certain API version and we return which bits
1789  * and ioctl commands are implemented in this kernel for such API
1790  * version or -EINVAL if unknown.
1791  */
1792 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1793                            unsigned long arg)
1794 {
1795         struct uffdio_api uffdio_api;
1796         void __user *buf = (void __user *)arg;
1797         int ret;
1798         __u64 features;
1799
1800         ret = -EINVAL;
1801         if (ctx->state != UFFD_STATE_WAIT_API)
1802                 goto out;
1803         ret = -EFAULT;
1804         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1805                 goto out;
1806         features = uffdio_api.features;
1807         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1808                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1809                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1810                         goto out;
1811                 ret = -EINVAL;
1812                 goto out;
1813         }
1814         /* report all available features and ioctls to userland */
1815         uffdio_api.features = UFFD_API_FEATURES;
1816         uffdio_api.ioctls = UFFD_API_IOCTLS;
1817         ret = -EFAULT;
1818         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1819                 goto out;
1820         ctx->state = UFFD_STATE_RUNNING;
1821         /* only enable the requested features for this uffd context */
1822         ctx->features = uffd_ctx_features(features);
1823         ret = 0;
1824 out:
1825         return ret;
1826 }
1827
1828 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1829                               unsigned long arg)
1830 {
1831         int ret = -EINVAL;
1832         struct userfaultfd_ctx *ctx = file->private_data;
1833
1834         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1835                 return -EINVAL;
1836
1837         switch(cmd) {
1838         case UFFDIO_API:
1839                 ret = userfaultfd_api(ctx, arg);
1840                 break;
1841         case UFFDIO_REGISTER:
1842                 ret = userfaultfd_register(ctx, arg);
1843                 break;
1844         case UFFDIO_UNREGISTER:
1845                 ret = userfaultfd_unregister(ctx, arg);
1846                 break;
1847         case UFFDIO_WAKE:
1848                 ret = userfaultfd_wake(ctx, arg);
1849                 break;
1850         case UFFDIO_COPY:
1851                 ret = userfaultfd_copy(ctx, arg);
1852                 break;
1853         case UFFDIO_ZEROPAGE:
1854                 ret = userfaultfd_zeropage(ctx, arg);
1855                 break;
1856         }
1857         return ret;
1858 }
1859
1860 #ifdef CONFIG_PROC_FS
1861 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1862 {
1863         struct userfaultfd_ctx *ctx = f->private_data;
1864         wait_queue_entry_t *wq;
1865         unsigned long pending = 0, total = 0;
1866
1867         spin_lock(&ctx->fault_pending_wqh.lock);
1868         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1869                 pending++;
1870                 total++;
1871         }
1872         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1873                 total++;
1874         }
1875         spin_unlock(&ctx->fault_pending_wqh.lock);
1876
1877         /*
1878          * If more protocols will be added, there will be all shown
1879          * separated by a space. Like this:
1880          *      protocols: aa:... bb:...
1881          */
1882         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1883                    pending, total, UFFD_API, ctx->features,
1884                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1885 }
1886 #endif
1887
1888 static const struct file_operations userfaultfd_fops = {
1889 #ifdef CONFIG_PROC_FS
1890         .show_fdinfo    = userfaultfd_show_fdinfo,
1891 #endif
1892         .release        = userfaultfd_release,
1893         .poll           = userfaultfd_poll,
1894         .read           = userfaultfd_read,
1895         .unlocked_ioctl = userfaultfd_ioctl,
1896         .compat_ioctl   = userfaultfd_ioctl,
1897         .llseek         = noop_llseek,
1898 };
1899
1900 static void init_once_userfaultfd_ctx(void *mem)
1901 {
1902         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1903
1904         init_waitqueue_head(&ctx->fault_pending_wqh);
1905         init_waitqueue_head(&ctx->fault_wqh);
1906         init_waitqueue_head(&ctx->event_wqh);
1907         init_waitqueue_head(&ctx->fd_wqh);
1908         seqcount_init(&ctx->refile_seq);
1909 }
1910
1911 SYSCALL_DEFINE1(userfaultfd, int, flags)
1912 {
1913         struct userfaultfd_ctx *ctx;
1914         int fd;
1915
1916         BUG_ON(!current->mm);
1917
1918         /* Check the UFFD_* constants for consistency.  */
1919         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1920         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1921
1922         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1923                 return -EINVAL;
1924
1925         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1926         if (!ctx)
1927                 return -ENOMEM;
1928
1929         atomic_set(&ctx->refcount, 1);
1930         ctx->flags = flags;
1931         ctx->features = 0;
1932         ctx->state = UFFD_STATE_WAIT_API;
1933         ctx->released = false;
1934         ctx->mmap_changing = false;
1935         ctx->mm = current->mm;
1936         /* prevent the mm struct to be freed */
1937         mmgrab(ctx->mm);
1938
1939         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
1940                               O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1941         if (fd < 0) {
1942                 mmdrop(ctx->mm);
1943                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1944         }
1945         return fd;
1946 }
1947
1948 static int __init userfaultfd_init(void)
1949 {
1950         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1951                                                 sizeof(struct userfaultfd_ctx),
1952                                                 0,
1953                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1954                                                 init_once_userfaultfd_ctx);
1955         return 0;
1956 }
1957 __initcall(userfaultfd_init);