fs/userfaultfd.c

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