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