fs/userfaultfd.c

   1 /*
   2  *  fs/userfaultfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc.
   6  *  Copyright (C) 2015  Red Hat, Inc.
   7  *
   8  *  This work is licensed under the terms of the GNU GPL, version 2. See
   9  *  the COPYING file in the top-level directory.
  10  *
  11  *  Some part derived from fs/eventfd.c (anon inode setup) and
  12  *  mm/ksm.c (mm hashing).
  13  */
  14
  15 #include <linux/list.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/mm.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  34
  35 enum userfaultfd_state {
  36         UFFD_STATE_WAIT_API,
  37         UFFD_STATE_RUNNING,
  38 };
  39
  40 /*
  41  * Start with fault_pending_wqh and fault_wqh so they're more likely
  42  * to be in the same cacheline.
  43  */
  44 struct userfaultfd_ctx {
  45         /* waitqueue head for the pending (i.e. not read) userfaults */
  46         wait_queue_head_t fault_pending_wqh;
  47         /* waitqueue head for the userfaults */
  48         wait_queue_head_t fault_wqh;
  49         /* waitqueue head for the pseudo fd to wakeup poll/read */
  50         wait_queue_head_t fd_wqh;
  51         /* waitqueue head for events */
  52         wait_queue_head_t event_wqh;
  53         /* a refile sequence protected by fault_pending_wqh lock */
  54         struct seqcount refile_seq;
  55         /* pseudo fd refcounting */
  56         atomic_t refcount;
  57         /* userfaultfd syscall flags */
  58         unsigned int flags;
  59         /* features requested from the userspace */
  60         unsigned int features;
  61         /* state machine */
  62         enum userfaultfd_state state;
  63         /* released */
  64         bool released;
  65         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  66         struct mm_struct *mm;
  67 };
  68
  69 struct userfaultfd_fork_ctx {
  70         struct userfaultfd_ctx *orig;
  71         struct userfaultfd_ctx *new;
  72         struct list_head list;
  73 };
  74
  75 struct userfaultfd_unmap_ctx {
  76         struct userfaultfd_ctx *ctx;
  77         unsigned long start;
  78         unsigned long end;
  79         struct list_head list;
  80 };
  81
  82 struct userfaultfd_wait_queue {
  83         struct uffd_msg msg;
  84         wait_queue_entry_t wq;
  85         struct userfaultfd_ctx *ctx;
  86         bool waken;
  87 };
  88
  89 struct userfaultfd_wake_range {
  90         unsigned long start;
  91         unsigned long len;
  92 };
  93
  94 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
  95                                      int wake_flags, void *key)
  96 {
  97         struct userfaultfd_wake_range *range = key;
  98         int ret;
  99         struct userfaultfd_wait_queue *uwq;
 100         unsigned long start, len;
 101
 102         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 103         ret = 0;
 104         /* len == 0 means wake all */
 105         start = range->start;
 106         len = range->len;
 107         if (len && (start > uwq->msg.arg.pagefault.address ||
 108                     start + len <= uwq->msg.arg.pagefault.address))
 109                 goto out;
 110         WRITE_ONCE(uwq->waken, true);
 111         /*
 112          * The Program-Order guarantees provided by the scheduler
 113          * ensure uwq->waken is visible before the task is woken.
 114          */
 115         ret = wake_up_state(wq->private, mode);
 116         if (ret) {
 117                 /*
 118                  * Wake only once, autoremove behavior.
 119                  *
 120                  * After the effect of list_del_init is visible to the other
 121                  * CPUs, the waitqueue may disappear from under us, see the
 122                  * !list_empty_careful() in handle_userfault().
 123                  *
 124                  * try_to_wake_up() has an implicit smp_mb(), and the
 125                  * wq->private is read before calling the extern function
 126                  * "wake_up_state" (which in turns calls try_to_wake_up).
 127                  */
 128                 list_del_init(&wq->entry);
 129         }
 130 out:
 131         return ret;
 132 }
 133
 134 /**
 135  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 136  * context.
 137  * @ctx: [in] Pointer to the userfaultfd context.
 138  */
 139 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 140 {
 141         if (!atomic_inc_not_zero(&ctx->refcount))
 142                 BUG();
 143 }
 144
 145 /**
 146  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 147  * context.
 148  * @ctx: [in] Pointer to userfaultfd context.
 149  *
 150  * The userfaultfd context reference must have been previously acquired either
 151  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 152  */
 153 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 154 {
 155         if (atomic_dec_and_test(&ctx->refcount)) {
 156                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 157                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 158                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 159                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 160                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 161                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 162                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 163                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 164                 mmdrop(ctx->mm);
 165                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 166         }
 167 }
 168
 169 static inline void msg_init(struct uffd_msg *msg)
 170 {
 171         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 172         /*
 173          * Must use memset to zero out the paddings or kernel data is
 174          * leaked to userland.
 175          */
 176         memset(msg, 0, sizeof(struct uffd_msg));
 177 }
 178
 179 static inline struct uffd_msg userfault_msg(unsigned long address,
 180                                             unsigned int flags,
 181                                             unsigned long reason,
 182                                             unsigned int features)
 183 {
 184         struct uffd_msg msg;
 185         msg_init(&msg);
 186         msg.event = UFFD_EVENT_PAGEFAULT;
 187         msg.arg.pagefault.address = address;
 188         if (flags & FAULT_FLAG_WRITE)
 189                 /*
 190                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 191                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 192                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 193                  * was a read fault, otherwise if set it means it's
 194                  * a write fault.
 195                  */
 196                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 197         if (reason & VM_UFFD_WP)
 198                 /*
 199                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 200                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 201                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 202                  * a missing fault, otherwise if set it means it's a
 203                  * write protect fault.
 204                  */
 205                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 206         if (features & UFFD_FEATURE_THREAD_ID)
 207                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 208         return msg;
 209 }
 210
 211 #ifdef CONFIG_HUGETLB_PAGE
 212 /*
 213  * Same functionality as userfaultfd_must_wait below with modifications for
 214  * hugepmd ranges.
 215  */
 216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 217                                          struct vm_area_struct *vma,
 218                                          unsigned long address,
 219                                          unsigned long flags,
 220                                          unsigned long reason)
 221 {
 222         struct mm_struct *mm = ctx->mm;
 223         pte_t *pte;
 224         bool ret = true;
 225
 226         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 227
 228         pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 229         if (!pte)
 230                 goto out;
 231
 232         ret = false;
 233
 234         /*
 235          * Lockless access: we're in a wait_event so it's ok if it
 236          * changes under us.
 237          */
 238         if (huge_pte_none(*pte))
 239                 ret = true;
 240         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
 241                 ret = true;
 242 out:
 243         return ret;
 244 }
 245 #else
 246 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 247                                          struct vm_area_struct *vma,
 248                                          unsigned long address,
 249                                          unsigned long flags,
 250                                          unsigned long reason)
 251 {
 252         return false;   /* should never get here */
 253 }
 254 #endif /* CONFIG_HUGETLB_PAGE */
 255
 256 /*
 257  * Verify the pagetables are still not ok after having reigstered into
 258  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 259  * userfault that has already been resolved, if userfaultfd_read and
 260  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 261  * threads.
 262  */
 263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 264                                          unsigned long address,
 265                                          unsigned long flags,
 266                                          unsigned long reason)
 267 {
 268         struct mm_struct *mm = ctx->mm;
 269         pgd_t *pgd;
 270         p4d_t *p4d;
 271         pud_t *pud;
 272         pmd_t *pmd, _pmd;
 273         pte_t *pte;
 274         bool ret = true;
 275
 276         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 277
 278         pgd = pgd_offset(mm, address);
 279         if (!pgd_present(*pgd))
 280                 goto out;
 281         p4d = p4d_offset(pgd, address);
 282         if (!p4d_present(*p4d))
 283                 goto out;
 284         pud = pud_offset(p4d, address);
 285         if (!pud_present(*pud))
 286                 goto out;
 287         pmd = pmd_offset(pud, address);
 288         /*
 289          * READ_ONCE must function as a barrier with narrower scope
 290          * and it must be equivalent to:
 291          *      _pmd = *pmd; barrier();
 292          *
 293          * This is to deal with the instability (as in
 294          * pmd_trans_unstable) of the pmd.
 295          */
 296         _pmd = READ_ONCE(*pmd);
 297         if (!pmd_present(_pmd))
 298                 goto out;
 299
 300         ret = false;
 301         if (pmd_trans_huge(_pmd))
 302                 goto out;
 303
 304         /*
 305          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 306          * and use the standard pte_offset_map() instead of parsing _pmd.
 307          */
 308         pte = pte_offset_map(pmd, address);
 309         /*
 310          * Lockless access: we're in a wait_event so it's ok if it
 311          * changes under us.
 312          */
 313         if (pte_none(*pte))
 314                 ret = true;
 315         pte_unmap(pte);
 316
 317 out:
 318         return ret;
 319 }
 320
 321 /*
 322  * The locking rules involved in returning VM_FAULT_RETRY depending on
 323  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 324  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 325  * recommendation in __lock_page_or_retry is not an understatement.
 326  *
 327  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 328  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 329  * not set.
 330  *
 331  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 332  * set, VM_FAULT_RETRY can still be returned if and only if there are
 333  * fatal_signal_pending()s, and the mmap_sem must be released before
 334  * returning it.
 335  */
 336 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 337 {
 338         struct mm_struct *mm = vmf->vma->vm_mm;
 339         struct userfaultfd_ctx *ctx;
 340         struct userfaultfd_wait_queue uwq;
 341         int ret;
 342         bool must_wait, return_to_userland;
 343         long blocking_state;
 344
 345         ret = VM_FAULT_SIGBUS;
 346
 347         /*
 348          * We don't do userfault handling for the final child pid update.
 349          *
 350          * We also don't do userfault handling during
 351          * coredumping. hugetlbfs has the special
 352          * follow_hugetlb_page() to skip missing pages in the
 353          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 354          * the no_page_table() helper in follow_page_mask(), but the
 355          * shmem_vm_ops->fault method is invoked even during
 356          * coredumping without mmap_sem and it ends up here.
 357          */
 358         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 359                 goto out;
 360
 361         /*
 362          * Coredumping runs without mmap_sem so we can only check that
 363          * the mmap_sem is held, if PF_DUMPCORE was not set.
 364          */
 365         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
 366
 367         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 368         if (!ctx)
 369                 goto out;
 370
 371         BUG_ON(ctx->mm != mm);
 372
 373         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 374         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 375
 376         if (ctx->features & UFFD_FEATURE_SIGBUS)
 377                 goto out;
 378
 379         /*
 380          * If it's already released don't get it. This avoids to loop
 381          * in __get_user_pages if userfaultfd_release waits on the
 382          * caller of handle_userfault to release the mmap_sem.
 383          */
 384         if (unlikely(READ_ONCE(ctx->released))) {
 385                 /*
 386                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 387                  * cooperative manager can close the uffd after the
 388                  * last UFFDIO_COPY, without risking to trigger an
 389                  * involuntary SIGBUS if the process was starting the
 390                  * userfaultfd while the userfaultfd was still armed
 391                  * (but after the last UFFDIO_COPY). If the uffd
 392                  * wasn't already closed when the userfault reached
 393                  * this point, that would normally be solved by
 394                  * userfaultfd_must_wait returning 'false'.
 395                  *
 396                  * If we were to return VM_FAULT_SIGBUS here, the non
 397                  * cooperative manager would be instead forced to
 398                  * always call UFFDIO_UNREGISTER before it can safely
 399                  * close the uffd.
 400                  */
 401                 ret = VM_FAULT_NOPAGE;
 402                 goto out;
 403         }
 404
 405         /*
 406          * Check that we can return VM_FAULT_RETRY.
 407          *
 408          * NOTE: it should become possible to return VM_FAULT_RETRY
 409          * even if FAULT_FLAG_TRIED is set without leading to gup()
 410          * -EBUSY failures, if the userfaultfd is to be extended for
 411          * VM_UFFD_WP tracking and we intend to arm the userfault
 412          * without first stopping userland access to the memory. For
 413          * VM_UFFD_MISSING userfaults this is enough for now.
 414          */
 415         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 416                 /*
 417                  * Validate the invariant that nowait must allow retry
 418                  * to be sure not to return SIGBUS erroneously on
 419                  * nowait invocations.
 420                  */
 421                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 422 #ifdef CONFIG_DEBUG_VM
 423                 if (printk_ratelimit()) {
 424                         printk(KERN_WARNING
 425                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 426                                vmf->flags);
 427                         dump_stack();
 428                 }
 429 #endif
 430                 goto out;
 431         }
 432
 433         /*
 434          * Handle nowait, not much to do other than tell it to retry
 435          * and wait.
 436          */
 437         ret = VM_FAULT_RETRY;
 438         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 439                 goto out;
 440
 441         /* take the reference before dropping the mmap_sem */
 442         userfaultfd_ctx_get(ctx);
 443
 444         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 445         uwq.wq.private = current;
 446         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 447                         ctx->features);
 448         uwq.ctx = ctx;
 449         uwq.waken = false;
 450
 451         return_to_userland =
 452                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 453                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 454         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 455                          TASK_KILLABLE;
 456
 457         spin_lock(&ctx->fault_pending_wqh.lock);
 458         /*
 459          * After the __add_wait_queue the uwq is visible to userland
 460          * through poll/read().
 461          */
 462         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 463         /*
 464          * The smp_mb() after __set_current_state prevents the reads
 465          * following the spin_unlock to happen before the list_add in
 466          * __add_wait_queue.
 467          */
 468         set_current_state(blocking_state);
 469         spin_unlock(&ctx->fault_pending_wqh.lock);
 470
 471         if (!is_vm_hugetlb_page(vmf->vma))
 472                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 473                                                   reason);
 474         else
 475                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 476                                                        vmf->address,
 477                                                        vmf->flags, reason);
 478         up_read(&mm->mmap_sem);
 479
 480         if (likely(must_wait && !READ_ONCE(ctx->released) &&
 481                    (return_to_userland ? !signal_pending(current) :
 482                     !fatal_signal_pending(current)))) {
 483                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 484                 schedule();
 485                 ret |= VM_FAULT_MAJOR;
 486
 487                 /*
 488                  * False wakeups can orginate even from rwsem before
 489                  * up_read() however userfaults will wait either for a
 490                  * targeted wakeup on the specific uwq waitqueue from
 491                  * wake_userfault() or for signals or for uffd
 492                  * release.
 493                  */
 494                 while (!READ_ONCE(uwq.waken)) {
 495                         /*
 496                          * This needs the full smp_store_mb()
 497                          * guarantee as the state write must be
 498                          * visible to other CPUs before reading
 499                          * uwq.waken from other CPUs.
 500                          */
 501                         set_current_state(blocking_state);
 502                         if (READ_ONCE(uwq.waken) ||
 503                             READ_ONCE(ctx->released) ||
 504                             (return_to_userland ? signal_pending(current) :
 505                              fatal_signal_pending(current)))
 506                                 break;
 507                         schedule();
 508                 }
 509         }
 510
 511         __set_current_state(TASK_RUNNING);
 512
 513         if (return_to_userland) {
 514                 if (signal_pending(current) &&
 515                     !fatal_signal_pending(current)) {
 516                         /*
 517                          * If we got a SIGSTOP or SIGCONT and this is
 518                          * a normal userland page fault, just let
 519                          * userland return so the signal will be
 520                          * handled and gdb debugging works.  The page
 521                          * fault code immediately after we return from
 522                          * this function is going to release the
 523                          * mmap_sem and it's not depending on it
 524                          * (unlike gup would if we were not to return
 525                          * VM_FAULT_RETRY).
 526                          *
 527                          * If a fatal signal is pending we still take
 528                          * the streamlined VM_FAULT_RETRY failure path
 529                          * and there's no need to retake the mmap_sem
 530                          * in such case.
 531                          */
 532                         down_read(&mm->mmap_sem);
 533                         ret = VM_FAULT_NOPAGE;
 534                 }
 535         }
 536
 537         /*
 538          * Here we race with the list_del; list_add in
 539          * userfaultfd_ctx_read(), however because we don't ever run
 540          * list_del_init() to refile across the two lists, the prev
 541          * and next pointers will never point to self. list_add also
 542          * would never let any of the two pointers to point to
 543          * self. So list_empty_careful won't risk to see both pointers
 544          * pointing to self at any time during the list refile. The
 545          * only case where list_del_init() is called is the full
 546          * removal in the wake function and there we don't re-list_add
 547          * and it's fine not to block on the spinlock. The uwq on this
 548          * kernel stack can be released after the list_del_init.
 549          */
 550         if (!list_empty_careful(&uwq.wq.entry)) {
 551                 spin_lock(&ctx->fault_pending_wqh.lock);
 552                 /*
 553                  * No need of list_del_init(), the uwq on the stack
 554                  * will be freed shortly anyway.
 555                  */
 556                 list_del(&uwq.wq.entry);
 557                 spin_unlock(&ctx->fault_pending_wqh.lock);
 558         }
 559
 560         /*
 561          * ctx may go away after this if the userfault pseudo fd is
 562          * already released.
 563          */
 564         userfaultfd_ctx_put(ctx);
 565
 566 out:
 567         return ret;
 568 }
 569
 570 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 571                                               struct userfaultfd_wait_queue *ewq)
 572 {
 573         struct userfaultfd_ctx *release_new_ctx;
 574
 575         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 576                 goto out;
 577
 578         ewq->ctx = ctx;
 579         init_waitqueue_entry(&ewq->wq, current);
 580         release_new_ctx = NULL;
 581
 582         spin_lock(&ctx->event_wqh.lock);
 583         /*
 584          * After the __add_wait_queue the uwq is visible to userland
 585          * through poll/read().
 586          */
 587         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 588         for (;;) {
 589                 set_current_state(TASK_KILLABLE);
 590                 if (ewq->msg.event == 0)
 591                         break;
 592                 if (READ_ONCE(ctx->released) ||
 593                     fatal_signal_pending(current)) {
 594                         /*
 595                          * &ewq->wq may be queued in fork_event, but
 596                          * __remove_wait_queue ignores the head
 597                          * parameter. It would be a problem if it
 598                          * didn't.
 599                          */
 600                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 601                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 602                                 struct userfaultfd_ctx *new;
 603
 604                                 new = (struct userfaultfd_ctx *)
 605                                         (unsigned long)
 606                                         ewq->msg.arg.reserved.reserved1;
 607                                 release_new_ctx = new;
 608                         }
 609                         break;
 610                 }
 611
 612                 spin_unlock(&ctx->event_wqh.lock);
 613
 614                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 615                 schedule();
 616
 617                 spin_lock(&ctx->event_wqh.lock);
 618         }
 619         __set_current_state(TASK_RUNNING);
 620         spin_unlock(&ctx->event_wqh.lock);
 621
 622         if (release_new_ctx) {
 623                 struct vm_area_struct *vma;
 624                 struct mm_struct *mm = release_new_ctx->mm;
 625
 626                 /* the various vma->vm_userfaultfd_ctx still points to it */
 627                 down_write(&mm->mmap_sem);
 628                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 629                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx)
 630                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 631                 up_write(&mm->mmap_sem);
 632
 633                 userfaultfd_ctx_put(release_new_ctx);
 634         }
 635
 636         /*
 637          * ctx may go away after this if the userfault pseudo fd is
 638          * already released.
 639          */
 640 out:
 641         userfaultfd_ctx_put(ctx);
 642 }
 643
 644 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 645                                        struct userfaultfd_wait_queue *ewq)
 646 {
 647         ewq->msg.event = 0;
 648         wake_up_locked(&ctx->event_wqh);
 649         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 650 }
 651
 652 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 653 {
 654         struct userfaultfd_ctx *ctx = NULL, *octx;
 655         struct userfaultfd_fork_ctx *fctx;
 656
 657         octx = vma->vm_userfaultfd_ctx.ctx;
 658         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 659                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 660                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 661                 return 0;
 662         }
 663
 664         list_for_each_entry(fctx, fcs, list)
 665                 if (fctx->orig == octx) {
 666                         ctx = fctx->new;
 667                         break;
 668                 }
 669
 670         if (!ctx) {
 671                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 672                 if (!fctx)
 673                         return -ENOMEM;
 674
 675                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 676                 if (!ctx) {
 677                         kfree(fctx);
 678                         return -ENOMEM;
 679                 }
 680
 681                 atomic_set(&ctx->refcount, 1);
 682                 ctx->flags = octx->flags;
 683                 ctx->state = UFFD_STATE_RUNNING;
 684                 ctx->features = octx->features;
 685                 ctx->released = false;
 686                 ctx->mm = vma->vm_mm;
 687                 mmgrab(ctx->mm);
 688
 689                 userfaultfd_ctx_get(octx);
 690                 fctx->orig = octx;
 691                 fctx->new = ctx;
 692                 list_add_tail(&fctx->list, fcs);
 693         }
 694
 695         vma->vm_userfaultfd_ctx.ctx = ctx;
 696         return 0;
 697 }
 698
 699 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 700 {
 701         struct userfaultfd_ctx *ctx = fctx->orig;
 702         struct userfaultfd_wait_queue ewq;
 703
 704         msg_init(&ewq.msg);
 705
 706         ewq.msg.event = UFFD_EVENT_FORK;
 707         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 708
 709         userfaultfd_event_wait_completion(ctx, &ewq);
 710 }
 711
 712 void dup_userfaultfd_complete(struct list_head *fcs)
 713 {
 714         struct userfaultfd_fork_ctx *fctx, *n;
 715
 716         list_for_each_entry_safe(fctx, n, fcs, list) {
 717                 dup_fctx(fctx);
 718                 list_del(&fctx->list);
 719                 kfree(fctx);
 720         }
 721 }
 722
 723 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 724                              struct vm_userfaultfd_ctx *vm_ctx)
 725 {
 726         struct userfaultfd_ctx *ctx;
 727
 728         ctx = vma->vm_userfaultfd_ctx.ctx;
 729         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 730                 vm_ctx->ctx = ctx;
 731                 userfaultfd_ctx_get(ctx);
 732         }
 733 }
 734
 735 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 736                                  unsigned long from, unsigned long to,
 737                                  unsigned long len)
 738 {
 739         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 740         struct userfaultfd_wait_queue ewq;
 741
 742         if (!ctx)
 743                 return;
 744
 745         if (to & ~PAGE_MASK) {
 746                 userfaultfd_ctx_put(ctx);
 747                 return;
 748         }
 749
 750         msg_init(&ewq.msg);
 751
 752         ewq.msg.event = UFFD_EVENT_REMAP;
 753         ewq.msg.arg.remap.from = from;
 754         ewq.msg.arg.remap.to = to;
 755         ewq.msg.arg.remap.len = len;
 756
 757         userfaultfd_event_wait_completion(ctx, &ewq);
 758 }
 759
 760 bool userfaultfd_remove(struct vm_area_struct *vma,
 761                         unsigned long start, unsigned long end)
 762 {
 763         struct mm_struct *mm = vma->vm_mm;
 764         struct userfaultfd_ctx *ctx;
 765         struct userfaultfd_wait_queue ewq;
 766
 767         ctx = vma->vm_userfaultfd_ctx.ctx;
 768         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 769                 return true;
 770
 771         userfaultfd_ctx_get(ctx);
 772         up_read(&mm->mmap_sem);
 773
 774         msg_init(&ewq.msg);
 775
 776         ewq.msg.event = UFFD_EVENT_REMOVE;
 777         ewq.msg.arg.remove.start = start;
 778         ewq.msg.arg.remove.end = end;
 779
 780         userfaultfd_event_wait_completion(ctx, &ewq);
 781
 782         return false;
 783 }
 784
 785 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 786                           unsigned long start, unsigned long end)
 787 {
 788         struct userfaultfd_unmap_ctx *unmap_ctx;
 789
 790         list_for_each_entry(unmap_ctx, unmaps, list)
 791                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 792                     unmap_ctx->end == end)
 793                         return true;
 794
 795         return false;
 796 }
 797
 798 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 799                            unsigned long start, unsigned long end,
 800                            struct list_head *unmaps)
 801 {
 802         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 803                 struct userfaultfd_unmap_ctx *unmap_ctx;
 804                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 805
 806                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 807                     has_unmap_ctx(ctx, unmaps, start, end))
 808                         continue;
 809
 810                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 811                 if (!unmap_ctx)
 812                         return -ENOMEM;
 813
 814                 userfaultfd_ctx_get(ctx);
 815                 unmap_ctx->ctx = ctx;
 816                 unmap_ctx->start = start;
 817                 unmap_ctx->end = end;
 818                 list_add_tail(&unmap_ctx->list, unmaps);
 819         }
 820
 821         return 0;
 822 }
 823
 824 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 825 {
 826         struct userfaultfd_unmap_ctx *ctx, *n;
 827         struct userfaultfd_wait_queue ewq;
 828
 829         list_for_each_entry_safe(ctx, n, uf, list) {
 830                 msg_init(&ewq.msg);
 831
 832                 ewq.msg.event = UFFD_EVENT_UNMAP;
 833                 ewq.msg.arg.remove.start = ctx->start;
 834                 ewq.msg.arg.remove.end = ctx->end;
 835
 836                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 837
 838                 list_del(&ctx->list);
 839                 kfree(ctx);
 840         }
 841 }
 842
 843 static int userfaultfd_release(struct inode *inode, struct file *file)
 844 {
 845         struct userfaultfd_ctx *ctx = file->private_data;
 846         struct mm_struct *mm = ctx->mm;
 847         struct vm_area_struct *vma, *prev;
 848         /* len == 0 means wake all */
 849         struct userfaultfd_wake_range range = { .len = 0, };
 850         unsigned long new_flags;
 851
 852         WRITE_ONCE(ctx->released, true);
 853
 854         if (!mmget_not_zero(mm))
 855                 goto wakeup;
 856
 857         /*
 858          * Flush page faults out of all CPUs. NOTE: all page faults
 859          * must be retried without returning VM_FAULT_SIGBUS if
 860          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 861          * changes while handle_userfault released the mmap_sem. So
 862          * it's critical that released is set to true (above), before
 863          * taking the mmap_sem for writing.
 864          */
 865         down_write(&mm->mmap_sem);
 866         prev = NULL;
 867         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 868                 cond_resched();
 869                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 870                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 871                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 872                         prev = vma;
 873                         continue;
 874                 }
 875                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 876                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 877                                  new_flags, vma->anon_vma,
 878                                  vma->vm_file, vma->vm_pgoff,
 879                                  vma_policy(vma),
 880                                  NULL_VM_UFFD_CTX);
 881                 if (prev)
 882                         vma = prev;
 883                 else
 884                         prev = vma;
 885                 vma->vm_flags = new_flags;
 886                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 887         }
 888         up_write(&mm->mmap_sem);
 889         mmput(mm);
 890 wakeup:
 891         /*
 892          * After no new page faults can wait on this fault_*wqh, flush
 893          * the last page faults that may have been already waiting on
 894          * the fault_*wqh.
 895          */
 896         spin_lock(&ctx->fault_pending_wqh.lock);
 897         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 898         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 899         spin_unlock(&ctx->fault_pending_wqh.lock);
 900
 901         /* Flush pending events that may still wait on event_wqh */
 902         wake_up_all(&ctx->event_wqh);
 903
 904         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 905         userfaultfd_ctx_put(ctx);
 906         return 0;
 907 }
 908
 909 /* fault_pending_wqh.lock must be hold by the caller */
 910 static inline struct userfaultfd_wait_queue *find_userfault_in(
 911                 wait_queue_head_t *wqh)
 912 {
 913         wait_queue_entry_t *wq;
 914         struct userfaultfd_wait_queue *uwq;
 915
 916         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 917
 918         uwq = NULL;
 919         if (!waitqueue_active(wqh))
 920                 goto out;
 921         /* walk in reverse to provide FIFO behavior to read userfaults */
 922         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 923         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 924 out:
 925         return uwq;
 926 }
 927
 928 static inline struct userfaultfd_wait_queue *find_userfault(
 929                 struct userfaultfd_ctx *ctx)
 930 {
 931         return find_userfault_in(&ctx->fault_pending_wqh);
 932 }
 933
 934 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 935                 struct userfaultfd_ctx *ctx)
 936 {
 937         return find_userfault_in(&ctx->event_wqh);
 938 }
 939
 940 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 941 {
 942         struct userfaultfd_ctx *ctx = file->private_data;
 943         __poll_t ret;
 944
 945         poll_wait(file, &ctx->fd_wqh, wait);
 946
 947         switch (ctx->state) {
 948         case UFFD_STATE_WAIT_API:
 949                 return POLLERR;
 950         case UFFD_STATE_RUNNING:
 951                 /*
 952                  * poll() never guarantees that read won't block.
 953                  * userfaults can be waken before they're read().
 954                  */
 955                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 956                         return POLLERR;
 957                 /*
 958                  * lockless access to see if there are pending faults
 959                  * __pollwait last action is the add_wait_queue but
 960                  * the spin_unlock would allow the waitqueue_active to
 961                  * pass above the actual list_add inside
 962                  * add_wait_queue critical section. So use a full
 963                  * memory barrier to serialize the list_add write of
 964                  * add_wait_queue() with the waitqueue_active read
 965                  * below.
 966                  */
 967                 ret = 0;
 968                 smp_mb();
 969                 if (waitqueue_active(&ctx->fault_pending_wqh))
 970                         ret = POLLIN;
 971                 else if (waitqueue_active(&ctx->event_wqh))
 972                         ret = POLLIN;
 973
 974                 return ret;
 975         default:
 976                 WARN_ON_ONCE(1);
 977                 return POLLERR;
 978         }
 979 }
 980
 981 static const struct file_operations userfaultfd_fops;
 982
 983 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 984                                   struct userfaultfd_ctx *new,
 985                                   struct uffd_msg *msg)
 986 {
 987         int fd;
 988         struct file *file;
 989         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 990
 991         fd = get_unused_fd_flags(flags);
 992         if (fd < 0)
 993                 return fd;
 994
 995         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 996                                   O_RDWR | flags);
 997         if (IS_ERR(file)) {
 998                 put_unused_fd(fd);
 999                 return PTR_ERR(file);
1000         }
1001
1002         fd_install(fd, file);
1003         msg->arg.reserved.reserved1 = 0;
1004         msg->arg.fork.ufd = fd;
1005
1006         return 0;
1007 }
1008
1009 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1010                                     struct uffd_msg *msg)
1011 {
1012         ssize_t ret;
1013         DECLARE_WAITQUEUE(wait, current);
1014         struct userfaultfd_wait_queue *uwq;
1015         /*
1016          * Handling fork event requires sleeping operations, so
1017          * we drop the event_wqh lock, then do these ops, then
1018          * lock it back and wake up the waiter. While the lock is
1019          * dropped the ewq may go away so we keep track of it
1020          * carefully.
1021          */
1022         LIST_HEAD(fork_event);
1023         struct userfaultfd_ctx *fork_nctx = NULL;
1024
1025         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1026         spin_lock(&ctx->fd_wqh.lock);
1027         __add_wait_queue(&ctx->fd_wqh, &wait);
1028         for (;;) {
1029                 set_current_state(TASK_INTERRUPTIBLE);
1030                 spin_lock(&ctx->fault_pending_wqh.lock);
1031                 uwq = find_userfault(ctx);
1032                 if (uwq) {
1033                         /*
1034                          * Use a seqcount to repeat the lockless check
1035                          * in wake_userfault() to avoid missing
1036                          * wakeups because during the refile both
1037                          * waitqueue could become empty if this is the
1038                          * only userfault.
1039                          */
1040                         write_seqcount_begin(&ctx->refile_seq);
1041
1042                         /*
1043                          * The fault_pending_wqh.lock prevents the uwq
1044                          * to disappear from under us.
1045                          *
1046                          * Refile this userfault from
1047                          * fault_pending_wqh to fault_wqh, it's not
1048                          * pending anymore after we read it.
1049                          *
1050                          * Use list_del() by hand (as
1051                          * userfaultfd_wake_function also uses
1052                          * list_del_init() by hand) to be sure nobody
1053                          * changes __remove_wait_queue() to use
1054                          * list_del_init() in turn breaking the
1055                          * !list_empty_careful() check in
1056                          * handle_userfault(). The uwq->wq.head list
1057                          * must never be empty at any time during the
1058                          * refile, or the waitqueue could disappear
1059                          * from under us. The "wait_queue_head_t"
1060                          * parameter of __remove_wait_queue() is unused
1061                          * anyway.
1062                          */
1063                         list_del(&uwq->wq.entry);
1064                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1065
1066                         write_seqcount_end(&ctx->refile_seq);
1067
1068                         /* careful to always initialize msg if ret == 0 */
1069                         *msg = uwq->msg;
1070                         spin_unlock(&ctx->fault_pending_wqh.lock);
1071                         ret = 0;
1072                         break;
1073                 }
1074                 spin_unlock(&ctx->fault_pending_wqh.lock);
1075
1076                 spin_lock(&ctx->event_wqh.lock);
1077                 uwq = find_userfault_evt(ctx);
1078                 if (uwq) {
1079                         *msg = uwq->msg;
1080
1081                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1082                                 fork_nctx = (struct userfaultfd_ctx *)
1083                                         (unsigned long)
1084                                         uwq->msg.arg.reserved.reserved1;
1085                                 list_move(&uwq->wq.entry, &fork_event);
1086                                 /*
1087                                  * fork_nctx can be freed as soon as
1088                                  * we drop the lock, unless we take a
1089                                  * reference on it.
1090                                  */
1091                                 userfaultfd_ctx_get(fork_nctx);
1092                                 spin_unlock(&ctx->event_wqh.lock);
1093                                 ret = 0;
1094                                 break;
1095                         }
1096
1097                         userfaultfd_event_complete(ctx, uwq);
1098                         spin_unlock(&ctx->event_wqh.lock);
1099                         ret = 0;
1100                         break;
1101                 }
1102                 spin_unlock(&ctx->event_wqh.lock);
1103
1104                 if (signal_pending(current)) {
1105                         ret = -ERESTARTSYS;
1106                         break;
1107                 }
1108                 if (no_wait) {
1109                         ret = -EAGAIN;
1110                         break;
1111                 }
1112                 spin_unlock(&ctx->fd_wqh.lock);
1113                 schedule();
1114                 spin_lock(&ctx->fd_wqh.lock);
1115         }
1116         __remove_wait_queue(&ctx->fd_wqh, &wait);
1117         __set_current_state(TASK_RUNNING);
1118         spin_unlock(&ctx->fd_wqh.lock);
1119
1120         if (!ret && msg->event == UFFD_EVENT_FORK) {
1121                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1122                 spin_lock(&ctx->event_wqh.lock);
1123                 if (!list_empty(&fork_event)) {
1124                         /*
1125                          * The fork thread didn't abort, so we can
1126                          * drop the temporary refcount.
1127                          */
1128                         userfaultfd_ctx_put(fork_nctx);
1129
1130                         uwq = list_first_entry(&fork_event,
1131                                                typeof(*uwq),
1132                                                wq.entry);
1133                         /*
1134                          * If fork_event list wasn't empty and in turn
1135                          * the event wasn't already released by fork
1136                          * (the event is allocated on fork kernel
1137                          * stack), put the event back to its place in
1138                          * the event_wq. fork_event head will be freed
1139                          * as soon as we return so the event cannot
1140                          * stay queued there no matter the current
1141                          * "ret" value.
1142                          */
1143                         list_del(&uwq->wq.entry);
1144                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1145
1146                         /*
1147                          * Leave the event in the waitqueue and report
1148                          * error to userland if we failed to resolve
1149                          * the userfault fork.
1150                          */
1151                         if (likely(!ret))
1152                                 userfaultfd_event_complete(ctx, uwq);
1153                 } else {
1154                         /*
1155                          * Here the fork thread aborted and the
1156                          * refcount from the fork thread on fork_nctx
1157                          * has already been released. We still hold
1158                          * the reference we took before releasing the
1159                          * lock above. If resolve_userfault_fork
1160                          * failed we've to drop it because the
1161                          * fork_nctx has to be freed in such case. If
1162                          * it succeeded we'll hold it because the new
1163                          * uffd references it.
1164                          */
1165                         if (ret)
1166                                 userfaultfd_ctx_put(fork_nctx);
1167                 }
1168                 spin_unlock(&ctx->event_wqh.lock);
1169         }
1170
1171         return ret;
1172 }
1173
1174 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1175                                 size_t count, loff_t *ppos)
1176 {
1177         struct userfaultfd_ctx *ctx = file->private_data;
1178         ssize_t _ret, ret = 0;
1179         struct uffd_msg msg;
1180         int no_wait = file->f_flags & O_NONBLOCK;
1181
1182         if (ctx->state == UFFD_STATE_WAIT_API)
1183                 return -EINVAL;
1184
1185         for (;;) {
1186                 if (count < sizeof(msg))
1187                         return ret ? ret : -EINVAL;
1188                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1189                 if (_ret < 0)
1190                         return ret ? ret : _ret;
1191                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1192                         return ret ? ret : -EFAULT;
1193                 ret += sizeof(msg);
1194                 buf += sizeof(msg);
1195                 count -= sizeof(msg);
1196                 /*
1197                  * Allow to read more than one fault at time but only
1198                  * block if waiting for the very first one.
1199                  */
1200                 no_wait = O_NONBLOCK;
1201         }
1202 }
1203
1204 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1205                              struct userfaultfd_wake_range *range)
1206 {
1207         spin_lock(&ctx->fault_pending_wqh.lock);
1208         /* wake all in the range and autoremove */
1209         if (waitqueue_active(&ctx->fault_pending_wqh))
1210                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1211                                      range);
1212         if (waitqueue_active(&ctx->fault_wqh))
1213                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1214         spin_unlock(&ctx->fault_pending_wqh.lock);
1215 }
1216
1217 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1218                                            struct userfaultfd_wake_range *range)
1219 {
1220         unsigned seq;
1221         bool need_wakeup;
1222
1223         /*
1224          * To be sure waitqueue_active() is not reordered by the CPU
1225          * before the pagetable update, use an explicit SMP memory
1226          * barrier here. PT lock release or up_read(mmap_sem) still
1227          * have release semantics that can allow the
1228          * waitqueue_active() to be reordered before the pte update.
1229          */
1230         smp_mb();
1231
1232         /*
1233          * Use waitqueue_active because it's very frequent to
1234          * change the address space atomically even if there are no
1235          * userfaults yet. So we take the spinlock only when we're
1236          * sure we've userfaults to wake.
1237          */
1238         do {
1239                 seq = read_seqcount_begin(&ctx->refile_seq);
1240                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1241                         waitqueue_active(&ctx->fault_wqh);
1242                 cond_resched();
1243         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1244         if (need_wakeup)
1245                 __wake_userfault(ctx, range);
1246 }
1247
1248 static __always_inline int validate_range(struct mm_struct *mm,
1249                                           __u64 start, __u64 len)
1250 {
1251         __u64 task_size = mm->task_size;
1252
1253         if (start & ~PAGE_MASK)
1254                 return -EINVAL;
1255         if (len & ~PAGE_MASK)
1256                 return -EINVAL;
1257         if (!len)
1258                 return -EINVAL;
1259         if (start < mmap_min_addr)
1260                 return -EINVAL;
1261         if (start >= task_size)
1262                 return -EINVAL;
1263         if (len > task_size - start)
1264                 return -EINVAL;
1265         return 0;
1266 }
1267
1268 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1269 {
1270         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1271                 vma_is_shmem(vma);
1272 }
1273
1274 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1275                                 unsigned long arg)
1276 {
1277         struct mm_struct *mm = ctx->mm;
1278         struct vm_area_struct *vma, *prev, *cur;
1279         int ret;
1280         struct uffdio_register uffdio_register;
1281         struct uffdio_register __user *user_uffdio_register;
1282         unsigned long vm_flags, new_flags;
1283         bool found;
1284         bool basic_ioctls;
1285         unsigned long start, end, vma_end;
1286
1287         user_uffdio_register = (struct uffdio_register __user *) arg;
1288
1289         ret = -EFAULT;
1290         if (copy_from_user(&uffdio_register, user_uffdio_register,
1291                            sizeof(uffdio_register)-sizeof(__u64)))
1292                 goto out;
1293
1294         ret = -EINVAL;
1295         if (!uffdio_register.mode)
1296                 goto out;
1297         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1298                                      UFFDIO_REGISTER_MODE_WP))
1299                 goto out;
1300         vm_flags = 0;
1301         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1302                 vm_flags |= VM_UFFD_MISSING;
1303         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1304                 vm_flags |= VM_UFFD_WP;
1305                 /*
1306                  * FIXME: remove the below error constraint by
1307                  * implementing the wprotect tracking mode.
1308                  */
1309                 ret = -EINVAL;
1310                 goto out;
1311         }
1312
1313         ret = validate_range(mm, uffdio_register.range.start,
1314                              uffdio_register.range.len);
1315         if (ret)
1316                 goto out;
1317
1318         start = uffdio_register.range.start;
1319         end = start + uffdio_register.range.len;
1320
1321         ret = -ENOMEM;
1322         if (!mmget_not_zero(mm))
1323                 goto out;
1324
1325         down_write(&mm->mmap_sem);
1326         vma = find_vma_prev(mm, start, &prev);
1327         if (!vma)
1328                 goto out_unlock;
1329
1330         /* check that there's at least one vma in the range */
1331         ret = -EINVAL;
1332         if (vma->vm_start >= end)
1333                 goto out_unlock;
1334
1335         /*
1336          * If the first vma contains huge pages, make sure start address
1337          * is aligned to huge page size.
1338          */
1339         if (is_vm_hugetlb_page(vma)) {
1340                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1341
1342                 if (start & (vma_hpagesize - 1))
1343                         goto out_unlock;
1344         }
1345
1346         /*
1347          * Search for not compatible vmas.
1348          */
1349         found = false;
1350         basic_ioctls = false;
1351         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1352                 cond_resched();
1353
1354                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1355                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1356
1357                 /* check not compatible vmas */
1358                 ret = -EINVAL;
1359                 if (!vma_can_userfault(cur))
1360                         goto out_unlock;
1361                 /*
1362                  * If this vma contains ending address, and huge pages
1363                  * check alignment.
1364                  */
1365                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1366                     end > cur->vm_start) {
1367                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1368
1369                         ret = -EINVAL;
1370
1371                         if (end & (vma_hpagesize - 1))
1372                                 goto out_unlock;
1373                 }
1374
1375                 /*
1376                  * Check that this vma isn't already owned by a
1377                  * different userfaultfd. We can't allow more than one
1378                  * userfaultfd to own a single vma simultaneously or we
1379                  * wouldn't know which one to deliver the userfaults to.
1380                  */
1381                 ret = -EBUSY;
1382                 if (cur->vm_userfaultfd_ctx.ctx &&
1383                     cur->vm_userfaultfd_ctx.ctx != ctx)
1384                         goto out_unlock;
1385
1386                 /*
1387                  * Note vmas containing huge pages
1388                  */
1389                 if (is_vm_hugetlb_page(cur))
1390                         basic_ioctls = true;
1391
1392                 found = true;
1393         }
1394         BUG_ON(!found);
1395
1396         if (vma->vm_start < start)
1397                 prev = vma;
1398
1399         ret = 0;
1400         do {
1401                 cond_resched();
1402
1403                 BUG_ON(!vma_can_userfault(vma));
1404                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1405                        vma->vm_userfaultfd_ctx.ctx != ctx);
1406
1407                 /*
1408                  * Nothing to do: this vma is already registered into this
1409                  * userfaultfd and with the right tracking mode too.
1410                  */
1411                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1412                     (vma->vm_flags & vm_flags) == vm_flags)
1413                         goto skip;
1414
1415                 if (vma->vm_start > start)
1416                         start = vma->vm_start;
1417                 vma_end = min(end, vma->vm_end);
1418
1419                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1420                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1421                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1422                                  vma_policy(vma),
1423                                  ((struct vm_userfaultfd_ctx){ ctx }));
1424                 if (prev) {
1425                         vma = prev;
1426                         goto next;
1427                 }
1428                 if (vma->vm_start < start) {
1429                         ret = split_vma(mm, vma, start, 1);
1430                         if (ret)
1431                                 break;
1432                 }
1433                 if (vma->vm_end > end) {
1434                         ret = split_vma(mm, vma, end, 0);
1435                         if (ret)
1436                                 break;
1437                 }
1438         next:
1439                 /*
1440                  * In the vma_merge() successful mprotect-like case 8:
1441                  * the next vma was merged into the current one and
1442                  * the current one has not been updated yet.
1443                  */
1444                 vma->vm_flags = new_flags;
1445                 vma->vm_userfaultfd_ctx.ctx = ctx;
1446
1447         skip:
1448                 prev = vma;
1449                 start = vma->vm_end;
1450                 vma = vma->vm_next;
1451         } while (vma && vma->vm_start < end);
1452 out_unlock:
1453         up_write(&mm->mmap_sem);
1454         mmput(mm);
1455         if (!ret) {
1456                 /*
1457                  * Now that we scanned all vmas we can already tell
1458                  * userland which ioctls methods are guaranteed to
1459                  * succeed on this range.
1460                  */
1461                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1462                              UFFD_API_RANGE_IOCTLS,
1463                              &user_uffdio_register->ioctls))
1464                         ret = -EFAULT;
1465         }
1466 out:
1467         return ret;
1468 }
1469
1470 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1471                                   unsigned long arg)
1472 {
1473         struct mm_struct *mm = ctx->mm;
1474         struct vm_area_struct *vma, *prev, *cur;
1475         int ret;
1476         struct uffdio_range uffdio_unregister;
1477         unsigned long new_flags;
1478         bool found;
1479         unsigned long start, end, vma_end;
1480         const void __user *buf = (void __user *)arg;
1481
1482         ret = -EFAULT;
1483         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1484                 goto out;
1485
1486         ret = validate_range(mm, uffdio_unregister.start,
1487                              uffdio_unregister.len);
1488         if (ret)
1489                 goto out;
1490
1491         start = uffdio_unregister.start;
1492         end = start + uffdio_unregister.len;
1493
1494         ret = -ENOMEM;
1495         if (!mmget_not_zero(mm))
1496                 goto out;
1497
1498         down_write(&mm->mmap_sem);
1499         vma = find_vma_prev(mm, start, &prev);
1500         if (!vma)
1501                 goto out_unlock;
1502
1503         /* check that there's at least one vma in the range */
1504         ret = -EINVAL;
1505         if (vma->vm_start >= end)
1506                 goto out_unlock;
1507
1508         /*
1509          * If the first vma contains huge pages, make sure start address
1510          * is aligned to huge page size.
1511          */
1512         if (is_vm_hugetlb_page(vma)) {
1513                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1514
1515                 if (start & (vma_hpagesize - 1))
1516                         goto out_unlock;
1517         }
1518
1519         /*
1520          * Search for not compatible vmas.
1521          */
1522         found = false;
1523         ret = -EINVAL;
1524         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1525                 cond_resched();
1526
1527                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1528                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1529
1530                 /*
1531                  * Check not compatible vmas, not strictly required
1532                  * here as not compatible vmas cannot have an
1533                  * userfaultfd_ctx registered on them, but this
1534                  * provides for more strict behavior to notice
1535                  * unregistration errors.
1536                  */
1537                 if (!vma_can_userfault(cur))
1538                         goto out_unlock;
1539
1540                 found = true;
1541         }
1542         BUG_ON(!found);
1543
1544         if (vma->vm_start < start)
1545                 prev = vma;
1546
1547         ret = 0;
1548         do {
1549                 cond_resched();
1550
1551                 BUG_ON(!vma_can_userfault(vma));
1552
1553                 /*
1554                  * Nothing to do: this vma is already registered into this
1555                  * userfaultfd and with the right tracking mode too.
1556                  */
1557                 if (!vma->vm_userfaultfd_ctx.ctx)
1558                         goto skip;
1559
1560                 if (vma->vm_start > start)
1561                         start = vma->vm_start;
1562                 vma_end = min(end, vma->vm_end);
1563
1564                 if (userfaultfd_missing(vma)) {
1565                         /*
1566                          * Wake any concurrent pending userfault while
1567                          * we unregister, so they will not hang
1568                          * permanently and it avoids userland to call
1569                          * UFFDIO_WAKE explicitly.
1570                          */
1571                         struct userfaultfd_wake_range range;
1572                         range.start = start;
1573                         range.len = vma_end - start;
1574                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1575                 }
1576
1577                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1578                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1579                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1580                                  vma_policy(vma),
1581                                  NULL_VM_UFFD_CTX);
1582                 if (prev) {
1583                         vma = prev;
1584                         goto next;
1585                 }
1586                 if (vma->vm_start < start) {
1587                         ret = split_vma(mm, vma, start, 1);
1588                         if (ret)
1589                                 break;
1590                 }
1591                 if (vma->vm_end > end) {
1592                         ret = split_vma(mm, vma, end, 0);
1593                         if (ret)
1594                                 break;
1595                 }
1596         next:
1597                 /*
1598                  * In the vma_merge() successful mprotect-like case 8:
1599                  * the next vma was merged into the current one and
1600                  * the current one has not been updated yet.
1601                  */
1602                 vma->vm_flags = new_flags;
1603                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1604
1605         skip:
1606                 prev = vma;
1607                 start = vma->vm_end;
1608                 vma = vma->vm_next;
1609         } while (vma && vma->vm_start < end);
1610 out_unlock:
1611         up_write(&mm->mmap_sem);
1612         mmput(mm);
1613 out:
1614         return ret;
1615 }
1616
1617 /*
1618  * userfaultfd_wake may be used in combination with the
1619  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1620  */
1621 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1622                             unsigned long arg)
1623 {
1624         int ret;
1625         struct uffdio_range uffdio_wake;
1626         struct userfaultfd_wake_range range;
1627         const void __user *buf = (void __user *)arg;
1628
1629         ret = -EFAULT;
1630         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1631                 goto out;
1632
1633         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1634         if (ret)
1635                 goto out;
1636
1637         range.start = uffdio_wake.start;
1638         range.len = uffdio_wake.len;
1639
1640         /*
1641          * len == 0 means wake all and we don't want to wake all here,
1642          * so check it again to be sure.
1643          */
1644         VM_BUG_ON(!range.len);
1645
1646         wake_userfault(ctx, &range);
1647         ret = 0;
1648
1649 out:
1650         return ret;
1651 }
1652
1653 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1654                             unsigned long arg)
1655 {
1656         __s64 ret;
1657         struct uffdio_copy uffdio_copy;
1658         struct uffdio_copy __user *user_uffdio_copy;
1659         struct userfaultfd_wake_range range;
1660
1661         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1662
1663         ret = -EFAULT;
1664         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1665                            /* don't copy "copy" last field */
1666                            sizeof(uffdio_copy)-sizeof(__s64)))
1667                 goto out;
1668
1669         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1670         if (ret)
1671                 goto out;
1672         /*
1673          * double check for wraparound just in case. copy_from_user()
1674          * will later check uffdio_copy.src + uffdio_copy.len to fit
1675          * in the userland range.
1676          */
1677         ret = -EINVAL;
1678         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1679                 goto out;
1680         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1681                 goto out;
1682         if (mmget_not_zero(ctx->mm)) {
1683                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1684                                    uffdio_copy.len);
1685                 mmput(ctx->mm);
1686         } else {
1687                 return -ESRCH;
1688         }
1689         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1690                 return -EFAULT;
1691         if (ret < 0)
1692                 goto out;
1693         BUG_ON(!ret);
1694         /* len == 0 would wake all */
1695         range.len = ret;
1696         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1697                 range.start = uffdio_copy.dst;
1698                 wake_userfault(ctx, &range);
1699         }
1700         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1701 out:
1702         return ret;
1703 }
1704
1705 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1706                                 unsigned long arg)
1707 {
1708         __s64 ret;
1709         struct uffdio_zeropage uffdio_zeropage;
1710         struct uffdio_zeropage __user *user_uffdio_zeropage;
1711         struct userfaultfd_wake_range range;
1712
1713         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1714
1715         ret = -EFAULT;
1716         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1717                            /* don't copy "zeropage" last field */
1718                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1719                 goto out;
1720
1721         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1722                              uffdio_zeropage.range.len);
1723         if (ret)
1724                 goto out;
1725         ret = -EINVAL;
1726         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1727                 goto out;
1728
1729         if (mmget_not_zero(ctx->mm)) {
1730                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1731                                      uffdio_zeropage.range.len);
1732                 mmput(ctx->mm);
1733         } else {
1734                 return -ESRCH;
1735         }
1736         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1737                 return -EFAULT;
1738         if (ret < 0)
1739                 goto out;
1740         /* len == 0 would wake all */
1741         BUG_ON(!ret);
1742         range.len = ret;
1743         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1744                 range.start = uffdio_zeropage.range.start;
1745                 wake_userfault(ctx, &range);
1746         }
1747         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1748 out:
1749         return ret;
1750 }
1751
1752 static inline unsigned int uffd_ctx_features(__u64 user_features)
1753 {
1754         /*
1755          * For the current set of features the bits just coincide
1756          */
1757         return (unsigned int)user_features;
1758 }
1759
1760 /*
1761  * userland asks for a certain API version and we return which bits
1762  * and ioctl commands are implemented in this kernel for such API
1763  * version or -EINVAL if unknown.
1764  */
1765 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1766                            unsigned long arg)
1767 {
1768         struct uffdio_api uffdio_api;
1769         void __user *buf = (void __user *)arg;
1770         int ret;
1771         __u64 features;
1772
1773         ret = -EINVAL;
1774         if (ctx->state != UFFD_STATE_WAIT_API)
1775                 goto out;
1776         ret = -EFAULT;
1777         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1778                 goto out;
1779         features = uffdio_api.features;
1780         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1781                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1782                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1783                         goto out;
1784                 ret = -EINVAL;
1785                 goto out;
1786         }
1787         /* report all available features and ioctls to userland */
1788         uffdio_api.features = UFFD_API_FEATURES;
1789         uffdio_api.ioctls = UFFD_API_IOCTLS;
1790         ret = -EFAULT;
1791         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1792                 goto out;
1793         ctx->state = UFFD_STATE_RUNNING;
1794         /* only enable the requested features for this uffd context */
1795         ctx->features = uffd_ctx_features(features);
1796         ret = 0;
1797 out:
1798         return ret;
1799 }
1800
1801 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1802                               unsigned long arg)
1803 {
1804         int ret = -EINVAL;
1805         struct userfaultfd_ctx *ctx = file->private_data;
1806
1807         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1808                 return -EINVAL;
1809
1810         switch(cmd) {
1811         case UFFDIO_API:
1812                 ret = userfaultfd_api(ctx, arg);
1813                 break;
1814         case UFFDIO_REGISTER:
1815                 ret = userfaultfd_register(ctx, arg);
1816                 break;
1817         case UFFDIO_UNREGISTER:
1818                 ret = userfaultfd_unregister(ctx, arg);
1819                 break;
1820         case UFFDIO_WAKE:
1821                 ret = userfaultfd_wake(ctx, arg);
1822                 break;
1823         case UFFDIO_COPY:
1824                 ret = userfaultfd_copy(ctx, arg);
1825                 break;
1826         case UFFDIO_ZEROPAGE:
1827                 ret = userfaultfd_zeropage(ctx, arg);
1828                 break;
1829         }
1830         return ret;
1831 }
1832
1833 #ifdef CONFIG_PROC_FS
1834 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1835 {
1836         struct userfaultfd_ctx *ctx = f->private_data;
1837         wait_queue_entry_t *wq;
1838         struct userfaultfd_wait_queue *uwq;
1839         unsigned long pending = 0, total = 0;
1840
1841         spin_lock(&ctx->fault_pending_wqh.lock);
1842         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1843                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1844                 pending++;
1845                 total++;
1846         }
1847         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1848                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1849                 total++;
1850         }
1851         spin_unlock(&ctx->fault_pending_wqh.lock);
1852
1853         /*
1854          * If more protocols will be added, there will be all shown
1855          * separated by a space. Like this:
1856          *      protocols: aa:... bb:...
1857          */
1858         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1859                    pending, total, UFFD_API, ctx->features,
1860                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1861 }
1862 #endif
1863
1864 static const struct file_operations userfaultfd_fops = {
1865 #ifdef CONFIG_PROC_FS
1866         .show_fdinfo    = userfaultfd_show_fdinfo,
1867 #endif
1868         .release        = userfaultfd_release,
1869         .poll           = userfaultfd_poll,
1870         .read           = userfaultfd_read,
1871         .unlocked_ioctl = userfaultfd_ioctl,
1872         .compat_ioctl   = userfaultfd_ioctl,
1873         .llseek         = noop_llseek,
1874 };
1875
1876 static void init_once_userfaultfd_ctx(void *mem)
1877 {
1878         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1879
1880         init_waitqueue_head(&ctx->fault_pending_wqh);
1881         init_waitqueue_head(&ctx->fault_wqh);
1882         init_waitqueue_head(&ctx->event_wqh);
1883         init_waitqueue_head(&ctx->fd_wqh);
1884         seqcount_init(&ctx->refile_seq);
1885 }
1886
1887 /**
1888  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1889  * @flags: Flags for the userfaultfd file.
1890  *
1891  * This function creates a userfaultfd file pointer, w/out installing
1892  * it into the fd table. This is useful when the userfaultfd file is
1893  * used during the initialization of data structures that require
1894  * extra setup after the userfaultfd creation. So the userfaultfd
1895  * creation is split into the file pointer creation phase, and the
1896  * file descriptor installation phase.  In this way races with
1897  * userspace closing the newly installed file descriptor can be
1898  * avoided.  Returns a userfaultfd file pointer, or a proper error
1899  * pointer.
1900  */
1901 static struct file *userfaultfd_file_create(int flags)
1902 {
1903         struct file *file;
1904         struct userfaultfd_ctx *ctx;
1905
1906         BUG_ON(!current->mm);
1907
1908         /* Check the UFFD_* constants for consistency.  */
1909         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1910         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1911
1912         file = ERR_PTR(-EINVAL);
1913         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1914                 goto out;
1915
1916         file = ERR_PTR(-ENOMEM);
1917         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1918         if (!ctx)
1919                 goto out;
1920
1921         atomic_set(&ctx->refcount, 1);
1922         ctx->flags = flags;
1923         ctx->features = 0;
1924         ctx->state = UFFD_STATE_WAIT_API;
1925         ctx->released = false;
1926         ctx->mm = current->mm;
1927         /* prevent the mm struct to be freed */
1928         mmgrab(ctx->mm);
1929
1930         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1931                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1932         if (IS_ERR(file)) {
1933                 mmdrop(ctx->mm);
1934                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1935         }
1936 out:
1937         return file;
1938 }
1939
1940 SYSCALL_DEFINE1(userfaultfd, int, flags)
1941 {
1942         int fd, error;
1943         struct file *file;
1944
1945         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1946         if (error < 0)
1947                 return error;
1948         fd = error;
1949
1950         file = userfaultfd_file_create(flags);
1951         if (IS_ERR(file)) {
1952                 error = PTR_ERR(file);
1953                 goto err_put_unused_fd;
1954         }
1955         fd_install(fd, file);
1956
1957         return fd;
1958
1959 err_put_unused_fd:
1960         put_unused_fd(fd);
1961
1962         return error;
1963 }
1964
1965 static int __init userfaultfd_init(void)
1966 {
1967         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1968                                                 sizeof(struct userfaultfd_ctx),
1969                                                 0,
1970                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1971                                                 init_once_userfaultfd_ctx);
1972         return 0;
1973 }
1974 __initcall(userfaultfd_init);