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