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 = 0;
 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 void userfaultfd_exit(struct mm_struct *mm)
 779 {
 780         struct vm_area_struct *vma = mm->mmap;
 781
 782         /*
 783          * We can do the vma walk without locking because the caller
 784          * (exit_mm) knows it now has exclusive access
 785          */
 786         while (vma) {
 787                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 788
 789                 if (ctx && (ctx->features & UFFD_FEATURE_EVENT_EXIT)) {
 790                         struct userfaultfd_wait_queue ewq;
 791
 792                         userfaultfd_ctx_get(ctx);
 793
 794                         msg_init(&ewq.msg);
 795                         ewq.msg.event = UFFD_EVENT_EXIT;
 796
 797                         userfaultfd_event_wait_completion(ctx, &ewq);
 798
 799                         ctx->features &= ~UFFD_FEATURE_EVENT_EXIT;
 800                 }
 801
 802                 vma = vma->vm_next;
 803         }
 804 }
 805
 806 static int userfaultfd_release(struct inode *inode, struct file *file)
 807 {
 808         struct userfaultfd_ctx *ctx = file->private_data;
 809         struct mm_struct *mm = ctx->mm;
 810         struct vm_area_struct *vma, *prev;
 811         /* len == 0 means wake all */
 812         struct userfaultfd_wake_range range = { .len = 0, };
 813         unsigned long new_flags;
 814
 815         ACCESS_ONCE(ctx->released) = true;
 816
 817         if (!mmget_not_zero(mm))
 818                 goto wakeup;
 819
 820         /*
 821          * Flush page faults out of all CPUs. NOTE: all page faults
 822          * must be retried without returning VM_FAULT_SIGBUS if
 823          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 824          * changes while handle_userfault released the mmap_sem. So
 825          * it's critical that released is set to true (above), before
 826          * taking the mmap_sem for writing.
 827          */
 828         down_write(&mm->mmap_sem);
 829         prev = NULL;
 830         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 831                 cond_resched();
 832                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 833                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 834                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 835                         prev = vma;
 836                         continue;
 837                 }
 838                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 839                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 840                                  new_flags, vma->anon_vma,
 841                                  vma->vm_file, vma->vm_pgoff,
 842                                  vma_policy(vma),
 843                                  NULL_VM_UFFD_CTX);
 844                 if (prev)
 845                         vma = prev;
 846                 else
 847                         prev = vma;
 848                 vma->vm_flags = new_flags;
 849                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 850         }
 851         up_write(&mm->mmap_sem);
 852         mmput(mm);
 853 wakeup:
 854         /*
 855          * After no new page faults can wait on this fault_*wqh, flush
 856          * the last page faults that may have been already waiting on
 857          * the fault_*wqh.
 858          */
 859         spin_lock(&ctx->fault_pending_wqh.lock);
 860         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 861         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 862         spin_unlock(&ctx->fault_pending_wqh.lock);
 863
 864         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 865         userfaultfd_ctx_put(ctx);
 866         return 0;
 867 }
 868
 869 /* fault_pending_wqh.lock must be hold by the caller */
 870 static inline struct userfaultfd_wait_queue *find_userfault_in(
 871                 wait_queue_head_t *wqh)
 872 {
 873         wait_queue_t *wq;
 874         struct userfaultfd_wait_queue *uwq;
 875
 876         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 877
 878         uwq = NULL;
 879         if (!waitqueue_active(wqh))
 880                 goto out;
 881         /* walk in reverse to provide FIFO behavior to read userfaults */
 882         wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
 883         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 884 out:
 885         return uwq;
 886 }
 887
 888 static inline struct userfaultfd_wait_queue *find_userfault(
 889                 struct userfaultfd_ctx *ctx)
 890 {
 891         return find_userfault_in(&ctx->fault_pending_wqh);
 892 }
 893
 894 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 895                 struct userfaultfd_ctx *ctx)
 896 {
 897         return find_userfault_in(&ctx->event_wqh);
 898 }
 899
 900 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 901 {
 902         struct userfaultfd_ctx *ctx = file->private_data;
 903         unsigned int ret;
 904
 905         poll_wait(file, &ctx->fd_wqh, wait);
 906
 907         switch (ctx->state) {
 908         case UFFD_STATE_WAIT_API:
 909                 return POLLERR;
 910         case UFFD_STATE_RUNNING:
 911                 /*
 912                  * poll() never guarantees that read won't block.
 913                  * userfaults can be waken before they're read().
 914                  */
 915                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 916                         return POLLERR;
 917                 /*
 918                  * lockless access to see if there are pending faults
 919                  * __pollwait last action is the add_wait_queue but
 920                  * the spin_unlock would allow the waitqueue_active to
 921                  * pass above the actual list_add inside
 922                  * add_wait_queue critical section. So use a full
 923                  * memory barrier to serialize the list_add write of
 924                  * add_wait_queue() with the waitqueue_active read
 925                  * below.
 926                  */
 927                 ret = 0;
 928                 smp_mb();
 929                 if (waitqueue_active(&ctx->fault_pending_wqh))
 930                         ret = POLLIN;
 931                 else if (waitqueue_active(&ctx->event_wqh))
 932                         ret = POLLIN;
 933
 934                 return ret;
 935         default:
 936                 WARN_ON_ONCE(1);
 937                 return POLLERR;
 938         }
 939 }
 940
 941 static const struct file_operations userfaultfd_fops;
 942
 943 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 944                                   struct userfaultfd_ctx *new,
 945                                   struct uffd_msg *msg)
 946 {
 947         int fd;
 948         struct file *file;
 949         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 950
 951         fd = get_unused_fd_flags(flags);
 952         if (fd < 0)
 953                 return fd;
 954
 955         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 956                                   O_RDWR | flags);
 957         if (IS_ERR(file)) {
 958                 put_unused_fd(fd);
 959                 return PTR_ERR(file);
 960         }
 961
 962         fd_install(fd, file);
 963         msg->arg.reserved.reserved1 = 0;
 964         msg->arg.fork.ufd = fd;
 965
 966         return 0;
 967 }
 968
 969 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 970                                     struct uffd_msg *msg)
 971 {
 972         ssize_t ret;
 973         DECLARE_WAITQUEUE(wait, current);
 974         struct userfaultfd_wait_queue *uwq;
 975         /*
 976          * Handling fork event requires sleeping operations, so
 977          * we drop the event_wqh lock, then do these ops, then
 978          * lock it back and wake up the waiter. While the lock is
 979          * dropped the ewq may go away so we keep track of it
 980          * carefully.
 981          */
 982         LIST_HEAD(fork_event);
 983         struct userfaultfd_ctx *fork_nctx = NULL;
 984
 985         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 986         spin_lock(&ctx->fd_wqh.lock);
 987         __add_wait_queue(&ctx->fd_wqh, &wait);
 988         for (;;) {
 989                 set_current_state(TASK_INTERRUPTIBLE);
 990                 spin_lock(&ctx->fault_pending_wqh.lock);
 991                 uwq = find_userfault(ctx);
 992                 if (uwq) {
 993                         /*
 994                          * Use a seqcount to repeat the lockless check
 995                          * in wake_userfault() to avoid missing
 996                          * wakeups because during the refile both
 997                          * waitqueue could become empty if this is the
 998                          * only userfault.
 999                          */
1000                         write_seqcount_begin(&ctx->refile_seq);
1001
1002                         /*
1003                          * The fault_pending_wqh.lock prevents the uwq
1004                          * to disappear from under us.
1005                          *
1006                          * Refile this userfault from
1007                          * fault_pending_wqh to fault_wqh, it's not
1008                          * pending anymore after we read it.
1009                          *
1010                          * Use list_del() by hand (as
1011                          * userfaultfd_wake_function also uses
1012                          * list_del_init() by hand) to be sure nobody
1013                          * changes __remove_wait_queue() to use
1014                          * list_del_init() in turn breaking the
1015                          * !list_empty_careful() check in
1016                          * handle_userfault(). The uwq->wq.task_list
1017                          * must never be empty at any time during the
1018                          * refile, or the waitqueue could disappear
1019                          * from under us. The "wait_queue_head_t"
1020                          * parameter of __remove_wait_queue() is unused
1021                          * anyway.
1022                          */
1023                         list_del(&uwq->wq.task_list);
1024                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1025
1026                         write_seqcount_end(&ctx->refile_seq);
1027
1028                         /* careful to always initialize msg if ret == 0 */
1029                         *msg = uwq->msg;
1030                         spin_unlock(&ctx->fault_pending_wqh.lock);
1031                         ret = 0;
1032                         break;
1033                 }
1034                 spin_unlock(&ctx->fault_pending_wqh.lock);
1035
1036                 spin_lock(&ctx->event_wqh.lock);
1037                 uwq = find_userfault_evt(ctx);
1038                 if (uwq) {
1039                         *msg = uwq->msg;
1040
1041                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1042                                 fork_nctx = (struct userfaultfd_ctx *)
1043                                         (unsigned long)
1044                                         uwq->msg.arg.reserved.reserved1;
1045                                 list_move(&uwq->wq.task_list, &fork_event);
1046                                 spin_unlock(&ctx->event_wqh.lock);
1047                                 ret = 0;
1048                                 break;
1049                         }
1050
1051                         userfaultfd_event_complete(ctx, uwq);
1052                         spin_unlock(&ctx->event_wqh.lock);
1053                         ret = 0;
1054                         break;
1055                 }
1056                 spin_unlock(&ctx->event_wqh.lock);
1057
1058                 if (signal_pending(current)) {
1059                         ret = -ERESTARTSYS;
1060                         break;
1061                 }
1062                 if (no_wait) {
1063                         ret = -EAGAIN;
1064                         break;
1065                 }
1066                 spin_unlock(&ctx->fd_wqh.lock);
1067                 schedule();
1068                 spin_lock(&ctx->fd_wqh.lock);
1069         }
1070         __remove_wait_queue(&ctx->fd_wqh, &wait);
1071         __set_current_state(TASK_RUNNING);
1072         spin_unlock(&ctx->fd_wqh.lock);
1073
1074         if (!ret && msg->event == UFFD_EVENT_FORK) {
1075                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1076
1077                 if (!ret) {
1078                         spin_lock(&ctx->event_wqh.lock);
1079                         if (!list_empty(&fork_event)) {
1080                                 uwq = list_first_entry(&fork_event,
1081                                                        typeof(*uwq),
1082                                                        wq.task_list);
1083                                 list_del(&uwq->wq.task_list);
1084                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1085                                 userfaultfd_event_complete(ctx, uwq);
1086                         }
1087                         spin_unlock(&ctx->event_wqh.lock);
1088                 }
1089         }
1090
1091         return ret;
1092 }
1093
1094 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1095                                 size_t count, loff_t *ppos)
1096 {
1097         struct userfaultfd_ctx *ctx = file->private_data;
1098         ssize_t _ret, ret = 0;
1099         struct uffd_msg msg;
1100         int no_wait = file->f_flags & O_NONBLOCK;
1101
1102         if (ctx->state == UFFD_STATE_WAIT_API)
1103                 return -EINVAL;
1104
1105         for (;;) {
1106                 if (count < sizeof(msg))
1107                         return ret ? ret : -EINVAL;
1108                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1109                 if (_ret < 0)
1110                         return ret ? ret : _ret;
1111                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1112                         return ret ? ret : -EFAULT;
1113                 ret += sizeof(msg);
1114                 buf += sizeof(msg);
1115                 count -= sizeof(msg);
1116                 /*
1117                  * Allow to read more than one fault at time but only
1118                  * block if waiting for the very first one.
1119                  */
1120                 no_wait = O_NONBLOCK;
1121         }
1122 }
1123
1124 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1125                              struct userfaultfd_wake_range *range)
1126 {
1127         unsigned long start, end;
1128
1129         start = range->start;
1130         end = range->start + range->len;
1131
1132         spin_lock(&ctx->fault_pending_wqh.lock);
1133         /* wake all in the range and autoremove */
1134         if (waitqueue_active(&ctx->fault_pending_wqh))
1135                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1136                                      range);
1137         if (waitqueue_active(&ctx->fault_wqh))
1138                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1139         spin_unlock(&ctx->fault_pending_wqh.lock);
1140 }
1141
1142 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1143                                            struct userfaultfd_wake_range *range)
1144 {
1145         unsigned seq;
1146         bool need_wakeup;
1147
1148         /*
1149          * To be sure waitqueue_active() is not reordered by the CPU
1150          * before the pagetable update, use an explicit SMP memory
1151          * barrier here. PT lock release or up_read(mmap_sem) still
1152          * have release semantics that can allow the
1153          * waitqueue_active() to be reordered before the pte update.
1154          */
1155         smp_mb();
1156
1157         /*
1158          * Use waitqueue_active because it's very frequent to
1159          * change the address space atomically even if there are no
1160          * userfaults yet. So we take the spinlock only when we're
1161          * sure we've userfaults to wake.
1162          */
1163         do {
1164                 seq = read_seqcount_begin(&ctx->refile_seq);
1165                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1166                         waitqueue_active(&ctx->fault_wqh);
1167                 cond_resched();
1168         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1169         if (need_wakeup)
1170                 __wake_userfault(ctx, range);
1171 }
1172
1173 static __always_inline int validate_range(struct mm_struct *mm,
1174                                           __u64 start, __u64 len)
1175 {
1176         __u64 task_size = mm->task_size;
1177
1178         if (start & ~PAGE_MASK)
1179                 return -EINVAL;
1180         if (len & ~PAGE_MASK)
1181                 return -EINVAL;
1182         if (!len)
1183                 return -EINVAL;
1184         if (start < mmap_min_addr)
1185                 return -EINVAL;
1186         if (start >= task_size)
1187                 return -EINVAL;
1188         if (len > task_size - start)
1189                 return -EINVAL;
1190         return 0;
1191 }
1192
1193 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1194 {
1195         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1196                 vma_is_shmem(vma);
1197 }
1198
1199 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1200                                 unsigned long arg)
1201 {
1202         struct mm_struct *mm = ctx->mm;
1203         struct vm_area_struct *vma, *prev, *cur;
1204         int ret;
1205         struct uffdio_register uffdio_register;
1206         struct uffdio_register __user *user_uffdio_register;
1207         unsigned long vm_flags, new_flags;
1208         bool found;
1209         bool non_anon_pages;
1210         unsigned long start, end, vma_end;
1211
1212         user_uffdio_register = (struct uffdio_register __user *) arg;
1213
1214         ret = -EFAULT;
1215         if (copy_from_user(&uffdio_register, user_uffdio_register,
1216                            sizeof(uffdio_register)-sizeof(__u64)))
1217                 goto out;
1218
1219         ret = -EINVAL;
1220         if (!uffdio_register.mode)
1221                 goto out;
1222         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1223                                      UFFDIO_REGISTER_MODE_WP))
1224                 goto out;
1225         vm_flags = 0;
1226         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1227                 vm_flags |= VM_UFFD_MISSING;
1228         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1229                 vm_flags |= VM_UFFD_WP;
1230                 /*
1231                  * FIXME: remove the below error constraint by
1232                  * implementing the wprotect tracking mode.
1233                  */
1234                 ret = -EINVAL;
1235                 goto out;
1236         }
1237
1238         ret = validate_range(mm, uffdio_register.range.start,
1239                              uffdio_register.range.len);
1240         if (ret)
1241                 goto out;
1242
1243         start = uffdio_register.range.start;
1244         end = start + uffdio_register.range.len;
1245
1246         ret = -ENOMEM;
1247         if (!mmget_not_zero(mm))
1248                 goto out;
1249
1250         down_write(&mm->mmap_sem);
1251         vma = find_vma_prev(mm, start, &prev);
1252         if (!vma)
1253                 goto out_unlock;
1254
1255         /* check that there's at least one vma in the range */
1256         ret = -EINVAL;
1257         if (vma->vm_start >= end)
1258                 goto out_unlock;
1259
1260         /*
1261          * If the first vma contains huge pages, make sure start address
1262          * is aligned to huge page size.
1263          */
1264         if (is_vm_hugetlb_page(vma)) {
1265                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1266
1267                 if (start & (vma_hpagesize - 1))
1268                         goto out_unlock;
1269         }
1270
1271         /*
1272          * Search for not compatible vmas.
1273          */
1274         found = false;
1275         non_anon_pages = false;
1276         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1277                 cond_resched();
1278
1279                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1280                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1281
1282                 /* check not compatible vmas */
1283                 ret = -EINVAL;
1284                 if (!vma_can_userfault(cur))
1285                         goto out_unlock;
1286                 /*
1287                  * If this vma contains ending address, and huge pages
1288                  * check alignment.
1289                  */
1290                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1291                     end > cur->vm_start) {
1292                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1293
1294                         ret = -EINVAL;
1295
1296                         if (end & (vma_hpagesize - 1))
1297                                 goto out_unlock;
1298                 }
1299
1300                 /*
1301                  * Check that this vma isn't already owned by a
1302                  * different userfaultfd. We can't allow more than one
1303                  * userfaultfd to own a single vma simultaneously or we
1304                  * wouldn't know which one to deliver the userfaults to.
1305                  */
1306                 ret = -EBUSY;
1307                 if (cur->vm_userfaultfd_ctx.ctx &&
1308                     cur->vm_userfaultfd_ctx.ctx != ctx)
1309                         goto out_unlock;
1310
1311                 /*
1312                  * Note vmas containing huge pages
1313                  */
1314                 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1315                         non_anon_pages = true;
1316
1317                 found = true;
1318         }
1319         BUG_ON(!found);
1320
1321         if (vma->vm_start < start)
1322                 prev = vma;
1323
1324         ret = 0;
1325         do {
1326                 cond_resched();
1327
1328                 BUG_ON(!vma_can_userfault(vma));
1329                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1330                        vma->vm_userfaultfd_ctx.ctx != ctx);
1331
1332                 /*
1333                  * Nothing to do: this vma is already registered into this
1334                  * userfaultfd and with the right tracking mode too.
1335                  */
1336                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1337                     (vma->vm_flags & vm_flags) == vm_flags)
1338                         goto skip;
1339
1340                 if (vma->vm_start > start)
1341                         start = vma->vm_start;
1342                 vma_end = min(end, vma->vm_end);
1343
1344                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1345                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1346                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1347                                  vma_policy(vma),
1348                                  ((struct vm_userfaultfd_ctx){ ctx }));
1349                 if (prev) {
1350                         vma = prev;
1351                         goto next;
1352                 }
1353                 if (vma->vm_start < start) {
1354                         ret = split_vma(mm, vma, start, 1);
1355                         if (ret)
1356                                 break;
1357                 }
1358                 if (vma->vm_end > end) {
1359                         ret = split_vma(mm, vma, end, 0);
1360                         if (ret)
1361                                 break;
1362                 }
1363         next:
1364                 /*
1365                  * In the vma_merge() successful mprotect-like case 8:
1366                  * the next vma was merged into the current one and
1367                  * the current one has not been updated yet.
1368                  */
1369                 vma->vm_flags = new_flags;
1370                 vma->vm_userfaultfd_ctx.ctx = ctx;
1371
1372         skip:
1373                 prev = vma;
1374                 start = vma->vm_end;
1375                 vma = vma->vm_next;
1376         } while (vma && vma->vm_start < end);
1377 out_unlock:
1378         up_write(&mm->mmap_sem);
1379         mmput(mm);
1380         if (!ret) {
1381                 /*
1382                  * Now that we scanned all vmas we can already tell
1383                  * userland which ioctls methods are guaranteed to
1384                  * succeed on this range.
1385                  */
1386                 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1387                              UFFD_API_RANGE_IOCTLS,
1388                              &user_uffdio_register->ioctls))
1389                         ret = -EFAULT;
1390         }
1391 out:
1392         return ret;
1393 }
1394
1395 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1396                                   unsigned long arg)
1397 {
1398         struct mm_struct *mm = ctx->mm;
1399         struct vm_area_struct *vma, *prev, *cur;
1400         int ret;
1401         struct uffdio_range uffdio_unregister;
1402         unsigned long new_flags;
1403         bool found;
1404         unsigned long start, end, vma_end;
1405         const void __user *buf = (void __user *)arg;
1406
1407         ret = -EFAULT;
1408         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1409                 goto out;
1410
1411         ret = validate_range(mm, uffdio_unregister.start,
1412                              uffdio_unregister.len);
1413         if (ret)
1414                 goto out;
1415
1416         start = uffdio_unregister.start;
1417         end = start + uffdio_unregister.len;
1418
1419         ret = -ENOMEM;
1420         if (!mmget_not_zero(mm))
1421                 goto out;
1422
1423         down_write(&mm->mmap_sem);
1424         vma = find_vma_prev(mm, start, &prev);
1425         if (!vma)
1426                 goto out_unlock;
1427
1428         /* check that there's at least one vma in the range */
1429         ret = -EINVAL;
1430         if (vma->vm_start >= end)
1431                 goto out_unlock;
1432
1433         /*
1434          * If the first vma contains huge pages, make sure start address
1435          * is aligned to huge page size.
1436          */
1437         if (is_vm_hugetlb_page(vma)) {
1438                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1439
1440                 if (start & (vma_hpagesize - 1))
1441                         goto out_unlock;
1442         }
1443
1444         /*
1445          * Search for not compatible vmas.
1446          */
1447         found = false;
1448         ret = -EINVAL;
1449         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1450                 cond_resched();
1451
1452                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1453                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1454
1455                 /*
1456                  * Check not compatible vmas, not strictly required
1457                  * here as not compatible vmas cannot have an
1458                  * userfaultfd_ctx registered on them, but this
1459                  * provides for more strict behavior to notice
1460                  * unregistration errors.
1461                  */
1462                 if (!vma_can_userfault(cur))
1463                         goto out_unlock;
1464
1465                 found = true;
1466         }
1467         BUG_ON(!found);
1468
1469         if (vma->vm_start < start)
1470                 prev = vma;
1471
1472         ret = 0;
1473         do {
1474                 cond_resched();
1475
1476                 BUG_ON(!vma_can_userfault(vma));
1477
1478                 /*
1479                  * Nothing to do: this vma is already registered into this
1480                  * userfaultfd and with the right tracking mode too.
1481                  */
1482                 if (!vma->vm_userfaultfd_ctx.ctx)
1483                         goto skip;
1484
1485                 if (vma->vm_start > start)
1486                         start = vma->vm_start;
1487                 vma_end = min(end, vma->vm_end);
1488
1489                 if (userfaultfd_missing(vma)) {
1490                         /*
1491                          * Wake any concurrent pending userfault while
1492                          * we unregister, so they will not hang
1493                          * permanently and it avoids userland to call
1494                          * UFFDIO_WAKE explicitly.
1495                          */
1496                         struct userfaultfd_wake_range range;
1497                         range.start = start;
1498                         range.len = vma_end - start;
1499                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1500                 }
1501
1502                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1503                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1504                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1505                                  vma_policy(vma),
1506                                  NULL_VM_UFFD_CTX);
1507                 if (prev) {
1508                         vma = prev;
1509                         goto next;
1510                 }
1511                 if (vma->vm_start < start) {
1512                         ret = split_vma(mm, vma, start, 1);
1513                         if (ret)
1514                                 break;
1515                 }
1516                 if (vma->vm_end > end) {
1517                         ret = split_vma(mm, vma, end, 0);
1518                         if (ret)
1519                                 break;
1520                 }
1521         next:
1522                 /*
1523                  * In the vma_merge() successful mprotect-like case 8:
1524                  * the next vma was merged into the current one and
1525                  * the current one has not been updated yet.
1526                  */
1527                 vma->vm_flags = new_flags;
1528                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1529
1530         skip:
1531                 prev = vma;
1532                 start = vma->vm_end;
1533                 vma = vma->vm_next;
1534         } while (vma && vma->vm_start < end);
1535 out_unlock:
1536         up_write(&mm->mmap_sem);
1537         mmput(mm);
1538 out:
1539         return ret;
1540 }
1541
1542 /*
1543  * userfaultfd_wake may be used in combination with the
1544  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1545  */
1546 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1547                             unsigned long arg)
1548 {
1549         int ret;
1550         struct uffdio_range uffdio_wake;
1551         struct userfaultfd_wake_range range;
1552         const void __user *buf = (void __user *)arg;
1553
1554         ret = -EFAULT;
1555         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1556                 goto out;
1557
1558         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1559         if (ret)
1560                 goto out;
1561
1562         range.start = uffdio_wake.start;
1563         range.len = uffdio_wake.len;
1564
1565         /*
1566          * len == 0 means wake all and we don't want to wake all here,
1567          * so check it again to be sure.
1568          */
1569         VM_BUG_ON(!range.len);
1570
1571         wake_userfault(ctx, &range);
1572         ret = 0;
1573
1574 out:
1575         return ret;
1576 }
1577
1578 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1579                             unsigned long arg)
1580 {
1581         __s64 ret;
1582         struct uffdio_copy uffdio_copy;
1583         struct uffdio_copy __user *user_uffdio_copy;
1584         struct userfaultfd_wake_range range;
1585
1586         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1587
1588         ret = -EFAULT;
1589         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1590                            /* don't copy "copy" last field */
1591                            sizeof(uffdio_copy)-sizeof(__s64)))
1592                 goto out;
1593
1594         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1595         if (ret)
1596                 goto out;
1597         /*
1598          * double check for wraparound just in case. copy_from_user()
1599          * will later check uffdio_copy.src + uffdio_copy.len to fit
1600          * in the userland range.
1601          */
1602         ret = -EINVAL;
1603         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1604                 goto out;
1605         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1606                 goto out;
1607         if (mmget_not_zero(ctx->mm)) {
1608                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1609                                    uffdio_copy.len);
1610                 mmput(ctx->mm);
1611         } else {
1612                 return -ENOSPC;
1613         }
1614         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1615                 return -EFAULT;
1616         if (ret < 0)
1617                 goto out;
1618         BUG_ON(!ret);
1619         /* len == 0 would wake all */
1620         range.len = ret;
1621         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1622                 range.start = uffdio_copy.dst;
1623                 wake_userfault(ctx, &range);
1624         }
1625         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1626 out:
1627         return ret;
1628 }
1629
1630 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1631                                 unsigned long arg)
1632 {
1633         __s64 ret;
1634         struct uffdio_zeropage uffdio_zeropage;
1635         struct uffdio_zeropage __user *user_uffdio_zeropage;
1636         struct userfaultfd_wake_range range;
1637
1638         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1639
1640         ret = -EFAULT;
1641         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1642                            /* don't copy "zeropage" last field */
1643                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1644                 goto out;
1645
1646         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1647                              uffdio_zeropage.range.len);
1648         if (ret)
1649                 goto out;
1650         ret = -EINVAL;
1651         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1652                 goto out;
1653
1654         if (mmget_not_zero(ctx->mm)) {
1655                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1656                                      uffdio_zeropage.range.len);
1657                 mmput(ctx->mm);
1658         }
1659         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1660                 return -EFAULT;
1661         if (ret < 0)
1662                 goto out;
1663         /* len == 0 would wake all */
1664         BUG_ON(!ret);
1665         range.len = ret;
1666         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1667                 range.start = uffdio_zeropage.range.start;
1668                 wake_userfault(ctx, &range);
1669         }
1670         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1671 out:
1672         return ret;
1673 }
1674
1675 static inline unsigned int uffd_ctx_features(__u64 user_features)
1676 {
1677         /*
1678          * For the current set of features the bits just coincide
1679          */
1680         return (unsigned int)user_features;
1681 }
1682
1683 /*
1684  * userland asks for a certain API version and we return which bits
1685  * and ioctl commands are implemented in this kernel for such API
1686  * version or -EINVAL if unknown.
1687  */
1688 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1689                            unsigned long arg)
1690 {
1691         struct uffdio_api uffdio_api;
1692         void __user *buf = (void __user *)arg;
1693         int ret;
1694         __u64 features;
1695
1696         ret = -EINVAL;
1697         if (ctx->state != UFFD_STATE_WAIT_API)
1698                 goto out;
1699         ret = -EFAULT;
1700         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1701                 goto out;
1702         features = uffdio_api.features;
1703         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1704                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1705                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1706                         goto out;
1707                 ret = -EINVAL;
1708                 goto out;
1709         }
1710         /* report all available features and ioctls to userland */
1711         uffdio_api.features = UFFD_API_FEATURES;
1712         uffdio_api.ioctls = UFFD_API_IOCTLS;
1713         ret = -EFAULT;
1714         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1715                 goto out;
1716         ctx->state = UFFD_STATE_RUNNING;
1717         /* only enable the requested features for this uffd context */
1718         ctx->features = uffd_ctx_features(features);
1719         ret = 0;
1720 out:
1721         return ret;
1722 }
1723
1724 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1725                               unsigned long arg)
1726 {
1727         int ret = -EINVAL;
1728         struct userfaultfd_ctx *ctx = file->private_data;
1729
1730         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1731                 return -EINVAL;
1732
1733         switch(cmd) {
1734         case UFFDIO_API:
1735                 ret = userfaultfd_api(ctx, arg);
1736                 break;
1737         case UFFDIO_REGISTER:
1738                 ret = userfaultfd_register(ctx, arg);
1739                 break;
1740         case UFFDIO_UNREGISTER:
1741                 ret = userfaultfd_unregister(ctx, arg);
1742                 break;
1743         case UFFDIO_WAKE:
1744                 ret = userfaultfd_wake(ctx, arg);
1745                 break;
1746         case UFFDIO_COPY:
1747                 ret = userfaultfd_copy(ctx, arg);
1748                 break;
1749         case UFFDIO_ZEROPAGE:
1750                 ret = userfaultfd_zeropage(ctx, arg);
1751                 break;
1752         }
1753         return ret;
1754 }
1755
1756 #ifdef CONFIG_PROC_FS
1757 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1758 {
1759         struct userfaultfd_ctx *ctx = f->private_data;
1760         wait_queue_t *wq;
1761         struct userfaultfd_wait_queue *uwq;
1762         unsigned long pending = 0, total = 0;
1763
1764         spin_lock(&ctx->fault_pending_wqh.lock);
1765         list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1766                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1767                 pending++;
1768                 total++;
1769         }
1770         list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1771                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1772                 total++;
1773         }
1774         spin_unlock(&ctx->fault_pending_wqh.lock);
1775
1776         /*
1777          * If more protocols will be added, there will be all shown
1778          * separated by a space. Like this:
1779          *      protocols: aa:... bb:...
1780          */
1781         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1782                    pending, total, UFFD_API, UFFD_API_FEATURES,
1783                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1784 }
1785 #endif
1786
1787 static const struct file_operations userfaultfd_fops = {
1788 #ifdef CONFIG_PROC_FS
1789         .show_fdinfo    = userfaultfd_show_fdinfo,
1790 #endif
1791         .release        = userfaultfd_release,
1792         .poll           = userfaultfd_poll,
1793         .read           = userfaultfd_read,
1794         .unlocked_ioctl = userfaultfd_ioctl,
1795         .compat_ioctl   = userfaultfd_ioctl,
1796         .llseek         = noop_llseek,
1797 };
1798
1799 static void init_once_userfaultfd_ctx(void *mem)
1800 {
1801         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1802
1803         init_waitqueue_head(&ctx->fault_pending_wqh);
1804         init_waitqueue_head(&ctx->fault_wqh);
1805         init_waitqueue_head(&ctx->event_wqh);
1806         init_waitqueue_head(&ctx->fd_wqh);
1807         seqcount_init(&ctx->refile_seq);
1808 }
1809
1810 /**
1811  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1812  * @flags: Flags for the userfaultfd file.
1813  *
1814  * This function creates a userfaultfd file pointer, w/out installing
1815  * it into the fd table. This is useful when the userfaultfd file is
1816  * used during the initialization of data structures that require
1817  * extra setup after the userfaultfd creation. So the userfaultfd
1818  * creation is split into the file pointer creation phase, and the
1819  * file descriptor installation phase.  In this way races with
1820  * userspace closing the newly installed file descriptor can be
1821  * avoided.  Returns a userfaultfd file pointer, or a proper error
1822  * pointer.
1823  */
1824 static struct file *userfaultfd_file_create(int flags)
1825 {
1826         struct file *file;
1827         struct userfaultfd_ctx *ctx;
1828
1829         BUG_ON(!current->mm);
1830
1831         /* Check the UFFD_* constants for consistency.  */
1832         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1833         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1834
1835         file = ERR_PTR(-EINVAL);
1836         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1837                 goto out;
1838
1839         file = ERR_PTR(-ENOMEM);
1840         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1841         if (!ctx)
1842                 goto out;
1843
1844         atomic_set(&ctx->refcount, 1);
1845         ctx->flags = flags;
1846         ctx->features = 0;
1847         ctx->state = UFFD_STATE_WAIT_API;
1848         ctx->released = false;
1849         ctx->mm = current->mm;
1850         /* prevent the mm struct to be freed */
1851         mmgrab(ctx->mm);
1852
1853         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1854                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1855         if (IS_ERR(file)) {
1856                 mmdrop(ctx->mm);
1857                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1858         }
1859 out:
1860         return file;
1861 }
1862
1863 SYSCALL_DEFINE1(userfaultfd, int, flags)
1864 {
1865         int fd, error;
1866         struct file *file;
1867
1868         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1869         if (error < 0)
1870                 return error;
1871         fd = error;
1872
1873         file = userfaultfd_file_create(flags);
1874         if (IS_ERR(file)) {
1875                 error = PTR_ERR(file);
1876                 goto err_put_unused_fd;
1877         }
1878         fd_install(fd, file);
1879
1880         return fd;
1881
1882 err_put_unused_fd:
1883         put_unused_fd(fd);
1884
1885         return error;
1886 }
1887
1888 static int __init userfaultfd_init(void)
1889 {
1890         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1891                                                 sizeof(struct userfaultfd_ctx),
1892                                                 0,
1893                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1894                                                 init_once_userfaultfd_ctx);
1895         return 0;
1896 }
1897 __initcall(userfaultfd_init);