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