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