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