fs/userfaultfd.c

   1 /*
   2  *  fs/userfaultfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc.
   6  *  Copyright (C) 2015  Red Hat, Inc.
   7  *
   8  *  This work is licensed under the terms of the GNU GPL, version 2. See
   9  *  the COPYING file in the top-level directory.
  10  *
  11  *  Some part derived from fs/eventfd.c (anon inode setup) and
  12  *  mm/ksm.c (mm hashing).
  13  */
  14
  15 #include <linux/list.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/mm.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  34
  35 enum userfaultfd_state {
  36         UFFD_STATE_WAIT_API,
  37         UFFD_STATE_RUNNING,
  38 };
  39
  40 /*
  41  * Start with fault_pending_wqh and fault_wqh so they're more likely
  42  * to be in the same cacheline.
  43  */
  44 struct userfaultfd_ctx {
  45         /* waitqueue head for the pending (i.e. not read) userfaults */
  46         wait_queue_head_t fault_pending_wqh;
  47         /* waitqueue head for the userfaults */
  48         wait_queue_head_t fault_wqh;
  49         /* waitqueue head for the pseudo fd to wakeup poll/read */
  50         wait_queue_head_t fd_wqh;
  51         /* waitqueue head for events */
  52         wait_queue_head_t event_wqh;
  53         /* a refile sequence protected by fault_pending_wqh lock */
  54         struct seqcount refile_seq;
  55         /* pseudo fd refcounting */
  56         atomic_t refcount;
  57         /* userfaultfd syscall flags */
  58         unsigned int flags;
  59         /* features requested from the userspace */
  60         unsigned int features;
  61         /* state machine */
  62         enum userfaultfd_state state;
  63         /* released */
  64         bool released;
  65         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  66         struct mm_struct *mm;
  67 };
  68
  69 struct userfaultfd_fork_ctx {
  70         struct userfaultfd_ctx *orig;
  71         struct userfaultfd_ctx *new;
  72         struct list_head list;
  73 };
  74
  75 struct userfaultfd_unmap_ctx {
  76         struct userfaultfd_ctx *ctx;
  77         unsigned long start;
  78         unsigned long end;
  79         struct list_head list;
  80 };
  81
  82 struct userfaultfd_wait_queue {
  83         struct uffd_msg msg;
  84         wait_queue_entry_t wq;
  85         struct userfaultfd_ctx *ctx;
  86         bool waken;
  87 };
  88
  89 struct userfaultfd_wake_range {
  90         unsigned long start;
  91         unsigned long len;
  92 };
  93
  94 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
  95                                      int wake_flags, void *key)
  96 {
  97         struct userfaultfd_wake_range *range = key;
  98         int ret;
  99         struct userfaultfd_wait_queue *uwq;
 100         unsigned long start, len;
 101
 102         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 103         ret = 0;
 104         /* len == 0 means wake all */
 105         start = range->start;
 106         len = range->len;
 107         if (len && (start > uwq->msg.arg.pagefault.address ||
 108                     start + len <= uwq->msg.arg.pagefault.address))
 109                 goto out;
 110         WRITE_ONCE(uwq->waken, true);
 111         /*
 112          * The Program-Order guarantees provided by the scheduler
 113          * ensure uwq->waken is visible before the task is woken.
 114          */
 115         ret = wake_up_state(wq->private, mode);
 116         if (ret) {
 117                 /*
 118                  * Wake only once, autoremove behavior.
 119                  *
 120                  * After the effect of list_del_init is visible to the other
 121                  * CPUs, the waitqueue may disappear from under us, see the
 122                  * !list_empty_careful() in handle_userfault().
 123                  *
 124                  * try_to_wake_up() has an implicit smp_mb(), and the
 125                  * wq->private is read before calling the extern function
 126                  * "wake_up_state" (which in turns calls try_to_wake_up).
 127                  */
 128                 list_del_init(&wq->entry);
 129         }
 130 out:
 131         return ret;
 132 }
 133
 134 /**
 135  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 136  * context.
 137  * @ctx: [in] Pointer to the userfaultfd context.
 138  */
 139 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 140 {
 141         if (!atomic_inc_not_zero(&ctx->refcount))
 142                 BUG();
 143 }
 144
 145 /**
 146  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 147  * context.
 148  * @ctx: [in] Pointer to userfaultfd context.
 149  *
 150  * The userfaultfd context reference must have been previously acquired either
 151  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 152  */
 153 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 154 {
 155         if (atomic_dec_and_test(&ctx->refcount)) {
 156                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 157                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 158                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 159                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 160                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 161                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 162                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 163                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 164                 mmdrop(ctx->mm);
 165                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 166         }
 167 }
 168
 169 static inline void msg_init(struct uffd_msg *msg)
 170 {
 171         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 172         /*
 173          * Must use memset to zero out the paddings or kernel data is
 174          * leaked to userland.
 175          */
 176         memset(msg, 0, sizeof(struct uffd_msg));
 177 }
 178
 179 static inline struct uffd_msg userfault_msg(unsigned long address,
 180                                             unsigned int flags,
 181                                             unsigned long reason,
 182                                             unsigned int features)
 183 {
 184         struct uffd_msg msg;
 185         msg_init(&msg);
 186         msg.event = UFFD_EVENT_PAGEFAULT;
 187         msg.arg.pagefault.address = address;
 188         if (flags & FAULT_FLAG_WRITE)
 189                 /*
 190                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 191                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 192                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 193                  * was a read fault, otherwise if set it means it's
 194                  * a write fault.
 195                  */
 196                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 197         if (reason & VM_UFFD_WP)
 198                 /*
 199                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 200                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 201                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 202                  * a missing fault, otherwise if set it means it's a
 203                  * write protect fault.
 204                  */
 205                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 206         if (features & UFFD_FEATURE_THREAD_ID)
 207                 msg.arg.pagefault.ptid = task_pid_vnr(current);
 208         return msg;
 209 }
 210
 211 #ifdef CONFIG_HUGETLB_PAGE
 212 /*
 213  * Same functionality as userfaultfd_must_wait below with modifications for
 214  * hugepmd ranges.
 215  */
 216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 217                                          struct vm_area_struct *vma,
 218                                          unsigned long address,
 219                                          unsigned long flags,
 220                                          unsigned long reason)
 221 {
 222         struct mm_struct *mm = ctx->mm;
 223         pte_t *pte;
 224         bool ret = true;
 225
 226         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 227
 228         pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 229         if (!pte)
 230                 goto out;
 231
 232         ret = false;
 233
 234         /*
 235          * Lockless access: we're in a wait_event so it's ok if it
 236          * changes under us.
 237          */
 238         if (huge_pte_none(*pte))
 239                 ret = true;
 240         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
 241                 ret = true;
 242 out:
 243         return ret;
 244 }
 245 #else
 246 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 247                                          struct vm_area_struct *vma,
 248                                          unsigned long address,
 249                                          unsigned long flags,
 250                                          unsigned long reason)
 251 {
 252         return false;   /* should never get here */
 253 }
 254 #endif /* CONFIG_HUGETLB_PAGE */
 255
 256 /*
 257  * Verify the pagetables are still not ok after having reigstered into
 258  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 259  * userfault that has already been resolved, if userfaultfd_read and
 260  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 261  * threads.
 262  */
 263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 264                                          unsigned long address,
 265                                          unsigned long flags,
 266                                          unsigned long reason)
 267 {
 268         struct mm_struct *mm = ctx->mm;
 269         pgd_t *pgd;
 270         p4d_t *p4d;
 271         pud_t *pud;
 272         pmd_t *pmd, _pmd;
 273         pte_t *pte;
 274         bool ret = true;
 275
 276         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 277
 278         pgd = pgd_offset(mm, address);
 279         if (!pgd_present(*pgd))
 280                 goto out;
 281         p4d = p4d_offset(pgd, address);
 282         if (!p4d_present(*p4d))
 283                 goto out;
 284         pud = pud_offset(p4d, address);
 285         if (!pud_present(*pud))
 286                 goto out;
 287         pmd = pmd_offset(pud, address);
 288         /*
 289          * READ_ONCE must function as a barrier with narrower scope
 290          * and it must be equivalent to:
 291          *      _pmd = *pmd; barrier();
 292          *
 293          * This is to deal with the instability (as in
 294          * pmd_trans_unstable) of the pmd.
 295          */
 296         _pmd = READ_ONCE(*pmd);
 297         if (!pmd_present(_pmd))
 298                 goto out;
 299
 300         ret = false;
 301         if (pmd_trans_huge(_pmd))
 302                 goto out;
 303
 304         /*
 305          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 306          * and use the standard pte_offset_map() instead of parsing _pmd.
 307          */
 308         pte = pte_offset_map(pmd, address);
 309         /*
 310          * Lockless access: we're in a wait_event so it's ok if it
 311          * changes under us.
 312          */
 313         if (pte_none(*pte))
 314                 ret = true;
 315         pte_unmap(pte);
 316
 317 out:
 318         return ret;
 319 }
 320
 321 /*
 322  * The locking rules involved in returning VM_FAULT_RETRY depending on
 323  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 324  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 325  * recommendation in __lock_page_or_retry is not an understatement.
 326  *
 327  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 328  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 329  * not set.
 330  *
 331  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 332  * set, VM_FAULT_RETRY can still be returned if and only if there are
 333  * fatal_signal_pending()s, and the mmap_sem must be released before
 334  * returning it.
 335  */
 336 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 337 {
 338         struct mm_struct *mm = vmf->vma->vm_mm;
 339         struct userfaultfd_ctx *ctx;
 340         struct userfaultfd_wait_queue uwq;
 341         int ret;
 342         bool must_wait, return_to_userland;
 343         long blocking_state;
 344
 345         ret = VM_FAULT_SIGBUS;
 346
 347         /*
 348          * We don't do userfault handling for the final child pid update.
 349          *
 350          * We also don't do userfault handling during
 351          * coredumping. hugetlbfs has the special
 352          * follow_hugetlb_page() to skip missing pages in the
 353          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 354          * the no_page_table() helper in follow_page_mask(), but the
 355          * shmem_vm_ops->fault method is invoked even during
 356          * coredumping without mmap_sem and it ends up here.
 357          */
 358         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 359                 goto out;
 360
 361         /*
 362          * Coredumping runs without mmap_sem so we can only check that
 363          * the mmap_sem is held, if PF_DUMPCORE was not set.
 364          */
 365         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
 366
 367         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 368         if (!ctx)
 369                 goto out;
 370
 371         BUG_ON(ctx->mm != mm);
 372
 373         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 374         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 375
 376         if (ctx->features & UFFD_FEATURE_SIGBUS)
 377                 goto out;
 378
 379         /*
 380          * If it's already released don't get it. This avoids to loop
 381          * in __get_user_pages if userfaultfd_release waits on the
 382          * caller of handle_userfault to release the mmap_sem.
 383          */
 384         if (unlikely(ACCESS_ONCE(ctx->released)))
 385                 goto out;
 386
 387         /*
 388          * Check that we can return VM_FAULT_RETRY.
 389          *
 390          * NOTE: it should become possible to return VM_FAULT_RETRY
 391          * even if FAULT_FLAG_TRIED is set without leading to gup()
 392          * -EBUSY failures, if the userfaultfd is to be extended for
 393          * VM_UFFD_WP tracking and we intend to arm the userfault
 394          * without first stopping userland access to the memory. For
 395          * VM_UFFD_MISSING userfaults this is enough for now.
 396          */
 397         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 398                 /*
 399                  * Validate the invariant that nowait must allow retry
 400                  * to be sure not to return SIGBUS erroneously on
 401                  * nowait invocations.
 402                  */
 403                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 404 #ifdef CONFIG_DEBUG_VM
 405                 if (printk_ratelimit()) {
 406                         printk(KERN_WARNING
 407                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 408                                vmf->flags);
 409                         dump_stack();
 410                 }
 411 #endif
 412                 goto out;
 413         }
 414
 415         /*
 416          * Handle nowait, not much to do other than tell it to retry
 417          * and wait.
 418          */
 419         ret = VM_FAULT_RETRY;
 420         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 421                 goto out;
 422
 423         /* take the reference before dropping the mmap_sem */
 424         userfaultfd_ctx_get(ctx);
 425
 426         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 427         uwq.wq.private = current;
 428         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 429                         ctx->features);
 430         uwq.ctx = ctx;
 431         uwq.waken = false;
 432
 433         return_to_userland =
 434                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 435                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 436         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 437                          TASK_KILLABLE;
 438
 439         spin_lock(&ctx->fault_pending_wqh.lock);
 440         /*
 441          * After the __add_wait_queue the uwq is visible to userland
 442          * through poll/read().
 443          */
 444         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 445         /*
 446          * The smp_mb() after __set_current_state prevents the reads
 447          * following the spin_unlock to happen before the list_add in
 448          * __add_wait_queue.
 449          */
 450         set_current_state(blocking_state);
 451         spin_unlock(&ctx->fault_pending_wqh.lock);
 452
 453         if (!is_vm_hugetlb_page(vmf->vma))
 454                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 455                                                   reason);
 456         else
 457                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 458                                                        vmf->address,
 459                                                        vmf->flags, reason);
 460         up_read(&mm->mmap_sem);
 461
 462         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
 463                    (return_to_userland ? !signal_pending(current) :
 464                     !fatal_signal_pending(current)))) {
 465                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 466                 schedule();
 467                 ret |= VM_FAULT_MAJOR;
 468
 469                 /*
 470                  * False wakeups can orginate even from rwsem before
 471                  * up_read() however userfaults will wait either for a
 472                  * targeted wakeup on the specific uwq waitqueue from
 473                  * wake_userfault() or for signals or for uffd
 474                  * release.
 475                  */
 476                 while (!READ_ONCE(uwq.waken)) {
 477                         /*
 478                          * This needs the full smp_store_mb()
 479                          * guarantee as the state write must be
 480                          * visible to other CPUs before reading
 481                          * uwq.waken from other CPUs.
 482                          */
 483                         set_current_state(blocking_state);
 484                         if (READ_ONCE(uwq.waken) ||
 485                             READ_ONCE(ctx->released) ||
 486                             (return_to_userland ? signal_pending(current) :
 487                              fatal_signal_pending(current)))
 488                                 break;
 489                         schedule();
 490                 }
 491         }
 492
 493         __set_current_state(TASK_RUNNING);
 494
 495         if (return_to_userland) {
 496                 if (signal_pending(current) &&
 497                     !fatal_signal_pending(current)) {
 498                         /*
 499                          * If we got a SIGSTOP or SIGCONT and this is
 500                          * a normal userland page fault, just let
 501                          * userland return so the signal will be
 502                          * handled and gdb debugging works.  The page
 503                          * fault code immediately after we return from
 504                          * this function is going to release the
 505                          * mmap_sem and it's not depending on it
 506                          * (unlike gup would if we were not to return
 507                          * VM_FAULT_RETRY).
 508                          *
 509                          * If a fatal signal is pending we still take
 510                          * the streamlined VM_FAULT_RETRY failure path
 511                          * and there's no need to retake the mmap_sem
 512                          * in such case.
 513                          */
 514                         down_read(&mm->mmap_sem);
 515                         ret = VM_FAULT_NOPAGE;
 516                 }
 517         }
 518
 519         /*
 520          * Here we race with the list_del; list_add in
 521          * userfaultfd_ctx_read(), however because we don't ever run
 522          * list_del_init() to refile across the two lists, the prev
 523          * and next pointers will never point to self. list_add also
 524          * would never let any of the two pointers to point to
 525          * self. So list_empty_careful won't risk to see both pointers
 526          * pointing to self at any time during the list refile. The
 527          * only case where list_del_init() is called is the full
 528          * removal in the wake function and there we don't re-list_add
 529          * and it's fine not to block on the spinlock. The uwq on this
 530          * kernel stack can be released after the list_del_init.
 531          */
 532         if (!list_empty_careful(&uwq.wq.entry)) {
 533                 spin_lock(&ctx->fault_pending_wqh.lock);
 534                 /*
 535                  * No need of list_del_init(), the uwq on the stack
 536                  * will be freed shortly anyway.
 537                  */
 538                 list_del(&uwq.wq.entry);
 539                 spin_unlock(&ctx->fault_pending_wqh.lock);
 540         }
 541
 542         /*
 543          * ctx may go away after this if the userfault pseudo fd is
 544          * already released.
 545          */
 546         userfaultfd_ctx_put(ctx);
 547
 548 out:
 549         return ret;
 550 }
 551
 552 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 553                                               struct userfaultfd_wait_queue *ewq)
 554 {
 555         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 556                 goto out;
 557
 558         ewq->ctx = ctx;
 559         init_waitqueue_entry(&ewq->wq, current);
 560
 561         spin_lock(&ctx->event_wqh.lock);
 562         /*
 563          * After the __add_wait_queue the uwq is visible to userland
 564          * through poll/read().
 565          */
 566         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 567         for (;;) {
 568                 set_current_state(TASK_KILLABLE);
 569                 if (ewq->msg.event == 0)
 570                         break;
 571                 if (ACCESS_ONCE(ctx->released) ||
 572                     fatal_signal_pending(current)) {
 573                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 574                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 575                                 struct userfaultfd_ctx *new;
 576
 577                                 new = (struct userfaultfd_ctx *)
 578                                         (unsigned long)
 579                                         ewq->msg.arg.reserved.reserved1;
 580
 581                                 userfaultfd_ctx_put(new);
 582                         }
 583                         break;
 584                 }
 585
 586                 spin_unlock(&ctx->event_wqh.lock);
 587
 588                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 589                 schedule();
 590
 591                 spin_lock(&ctx->event_wqh.lock);
 592         }
 593         __set_current_state(TASK_RUNNING);
 594         spin_unlock(&ctx->event_wqh.lock);
 595
 596         /*
 597          * ctx may go away after this if the userfault pseudo fd is
 598          * already released.
 599          */
 600 out:
 601         userfaultfd_ctx_put(ctx);
 602 }
 603
 604 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 605                                        struct userfaultfd_wait_queue *ewq)
 606 {
 607         ewq->msg.event = 0;
 608         wake_up_locked(&ctx->event_wqh);
 609         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 610 }
 611
 612 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 613 {
 614         struct userfaultfd_ctx *ctx = NULL, *octx;
 615         struct userfaultfd_fork_ctx *fctx;
 616
 617         octx = vma->vm_userfaultfd_ctx.ctx;
 618         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 619                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 620                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 621                 return 0;
 622         }
 623
 624         list_for_each_entry(fctx, fcs, list)
 625                 if (fctx->orig == octx) {
 626                         ctx = fctx->new;
 627                         break;
 628                 }
 629
 630         if (!ctx) {
 631                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 632                 if (!fctx)
 633                         return -ENOMEM;
 634
 635                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 636                 if (!ctx) {
 637                         kfree(fctx);
 638                         return -ENOMEM;
 639                 }
 640
 641                 atomic_set(&ctx->refcount, 1);
 642                 ctx->flags = octx->flags;
 643                 ctx->state = UFFD_STATE_RUNNING;
 644                 ctx->features = octx->features;
 645                 ctx->released = false;
 646                 ctx->mm = vma->vm_mm;
 647                 atomic_inc(&ctx->mm->mm_count);
 648
 649                 userfaultfd_ctx_get(octx);
 650                 fctx->orig = octx;
 651                 fctx->new = ctx;
 652                 list_add_tail(&fctx->list, fcs);
 653         }
 654
 655         vma->vm_userfaultfd_ctx.ctx = ctx;
 656         return 0;
 657 }
 658
 659 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 660 {
 661         struct userfaultfd_ctx *ctx = fctx->orig;
 662         struct userfaultfd_wait_queue ewq;
 663
 664         msg_init(&ewq.msg);
 665
 666         ewq.msg.event = UFFD_EVENT_FORK;
 667         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 668
 669         userfaultfd_event_wait_completion(ctx, &ewq);
 670 }
 671
 672 void dup_userfaultfd_complete(struct list_head *fcs)
 673 {
 674         struct userfaultfd_fork_ctx *fctx, *n;
 675
 676         list_for_each_entry_safe(fctx, n, fcs, list) {
 677                 dup_fctx(fctx);
 678                 list_del(&fctx->list);
 679                 kfree(fctx);
 680         }
 681 }
 682
 683 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 684                              struct vm_userfaultfd_ctx *vm_ctx)
 685 {
 686         struct userfaultfd_ctx *ctx;
 687
 688         ctx = vma->vm_userfaultfd_ctx.ctx;
 689         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 690                 vm_ctx->ctx = ctx;
 691                 userfaultfd_ctx_get(ctx);
 692         }
 693 }
 694
 695 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 696                                  unsigned long from, unsigned long to,
 697                                  unsigned long len)
 698 {
 699         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 700         struct userfaultfd_wait_queue ewq;
 701
 702         if (!ctx)
 703                 return;
 704
 705         if (to & ~PAGE_MASK) {
 706                 userfaultfd_ctx_put(ctx);
 707                 return;
 708         }
 709
 710         msg_init(&ewq.msg);
 711
 712         ewq.msg.event = UFFD_EVENT_REMAP;
 713         ewq.msg.arg.remap.from = from;
 714         ewq.msg.arg.remap.to = to;
 715         ewq.msg.arg.remap.len = len;
 716
 717         userfaultfd_event_wait_completion(ctx, &ewq);
 718 }
 719
 720 bool userfaultfd_remove(struct vm_area_struct *vma,
 721                         unsigned long start, unsigned long end)
 722 {
 723         struct mm_struct *mm = vma->vm_mm;
 724         struct userfaultfd_ctx *ctx;
 725         struct userfaultfd_wait_queue ewq;
 726
 727         ctx = vma->vm_userfaultfd_ctx.ctx;
 728         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 729                 return true;
 730
 731         userfaultfd_ctx_get(ctx);
 732         up_read(&mm->mmap_sem);
 733
 734         msg_init(&ewq.msg);
 735
 736         ewq.msg.event = UFFD_EVENT_REMOVE;
 737         ewq.msg.arg.remove.start = start;
 738         ewq.msg.arg.remove.end = end;
 739
 740         userfaultfd_event_wait_completion(ctx, &ewq);
 741
 742         return false;
 743 }
 744
 745 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 746                           unsigned long start, unsigned long end)
 747 {
 748         struct userfaultfd_unmap_ctx *unmap_ctx;
 749
 750         list_for_each_entry(unmap_ctx, unmaps, list)
 751                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 752                     unmap_ctx->end == end)
 753                         return true;
 754
 755         return false;
 756 }
 757
 758 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 759                            unsigned long start, unsigned long end,
 760                            struct list_head *unmaps)
 761 {
 762         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 763                 struct userfaultfd_unmap_ctx *unmap_ctx;
 764                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 765
 766                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 767                     has_unmap_ctx(ctx, unmaps, start, end))
 768                         continue;
 769
 770                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 771                 if (!unmap_ctx)
 772                         return -ENOMEM;
 773
 774                 userfaultfd_ctx_get(ctx);
 775                 unmap_ctx->ctx = ctx;
 776                 unmap_ctx->start = start;
 777                 unmap_ctx->end = end;
 778                 list_add_tail(&unmap_ctx->list, unmaps);
 779         }
 780
 781         return 0;
 782 }
 783
 784 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 785 {
 786         struct userfaultfd_unmap_ctx *ctx, *n;
 787         struct userfaultfd_wait_queue ewq;
 788
 789         list_for_each_entry_safe(ctx, n, uf, list) {
 790                 msg_init(&ewq.msg);
 791
 792                 ewq.msg.event = UFFD_EVENT_UNMAP;
 793                 ewq.msg.arg.remove.start = ctx->start;
 794                 ewq.msg.arg.remove.end = ctx->end;
 795
 796                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 797
 798                 list_del(&ctx->list);
 799                 kfree(ctx);
 800         }
 801 }
 802
 803 static int userfaultfd_release(struct inode *inode, struct file *file)
 804 {
 805         struct userfaultfd_ctx *ctx = file->private_data;
 806         struct mm_struct *mm = ctx->mm;
 807         struct vm_area_struct *vma, *prev;
 808         /* len == 0 means wake all */
 809         struct userfaultfd_wake_range range = { .len = 0, };
 810         unsigned long new_flags;
 811
 812         ACCESS_ONCE(ctx->released) = true;
 813
 814         if (!mmget_not_zero(mm))
 815                 goto wakeup;
 816
 817         /*
 818          * Flush page faults out of all CPUs. NOTE: all page faults
 819          * must be retried without returning VM_FAULT_SIGBUS if
 820          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 821          * changes while handle_userfault released the mmap_sem. So
 822          * it's critical that released is set to true (above), before
 823          * taking the mmap_sem for writing.
 824          */
 825         down_write(&mm->mmap_sem);
 826         prev = NULL;
 827         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 828                 cond_resched();
 829                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 830                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 831                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 832                         prev = vma;
 833                         continue;
 834                 }
 835                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 836                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 837                                  new_flags, vma->anon_vma,
 838                                  vma->vm_file, vma->vm_pgoff,
 839                                  vma_policy(vma),
 840                                  NULL_VM_UFFD_CTX);
 841                 if (prev)
 842                         vma = prev;
 843                 else
 844                         prev = vma;
 845                 vma->vm_flags = new_flags;
 846                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 847         }
 848         up_write(&mm->mmap_sem);
 849         mmput(mm);
 850 wakeup:
 851         /*
 852          * After no new page faults can wait on this fault_*wqh, flush
 853          * the last page faults that may have been already waiting on
 854          * the fault_*wqh.
 855          */
 856         spin_lock(&ctx->fault_pending_wqh.lock);
 857         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 858         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 859         spin_unlock(&ctx->fault_pending_wqh.lock);
 860
 861         /* Flush pending events that may still wait on event_wqh */
 862         wake_up_all(&ctx->event_wqh);
 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_entry_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->head, typeof(*wq), entry);
 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.head 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.entry);
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.entry, &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.entry);
1083                                 list_del(&uwq->wq.entry);
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         spin_lock(&ctx->fault_pending_wqh.lock);
1128         /* wake all in the range and autoremove */
1129         if (waitqueue_active(&ctx->fault_pending_wqh))
1130                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1131                                      range);
1132         if (waitqueue_active(&ctx->fault_wqh))
1133                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1134         spin_unlock(&ctx->fault_pending_wqh.lock);
1135 }
1136
1137 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1138                                            struct userfaultfd_wake_range *range)
1139 {
1140         unsigned seq;
1141         bool need_wakeup;
1142
1143         /*
1144          * To be sure waitqueue_active() is not reordered by the CPU
1145          * before the pagetable update, use an explicit SMP memory
1146          * barrier here. PT lock release or up_read(mmap_sem) still
1147          * have release semantics that can allow the
1148          * waitqueue_active() to be reordered before the pte update.
1149          */
1150         smp_mb();
1151
1152         /*
1153          * Use waitqueue_active because it's very frequent to
1154          * change the address space atomically even if there are no
1155          * userfaults yet. So we take the spinlock only when we're
1156          * sure we've userfaults to wake.
1157          */
1158         do {
1159                 seq = read_seqcount_begin(&ctx->refile_seq);
1160                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1161                         waitqueue_active(&ctx->fault_wqh);
1162                 cond_resched();
1163         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1164         if (need_wakeup)
1165                 __wake_userfault(ctx, range);
1166 }
1167
1168 static __always_inline int validate_range(struct mm_struct *mm,
1169                                           __u64 start, __u64 len)
1170 {
1171         __u64 task_size = mm->task_size;
1172
1173         if (start & ~PAGE_MASK)
1174                 return -EINVAL;
1175         if (len & ~PAGE_MASK)
1176                 return -EINVAL;
1177         if (!len)
1178                 return -EINVAL;
1179         if (start < mmap_min_addr)
1180                 return -EINVAL;
1181         if (start >= task_size)
1182                 return -EINVAL;
1183         if (len > task_size - start)
1184                 return -EINVAL;
1185         return 0;
1186 }
1187
1188 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1189 {
1190         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1191                 vma_is_shmem(vma);
1192 }
1193
1194 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1195                                 unsigned long arg)
1196 {
1197         struct mm_struct *mm = ctx->mm;
1198         struct vm_area_struct *vma, *prev, *cur;
1199         int ret;
1200         struct uffdio_register uffdio_register;
1201         struct uffdio_register __user *user_uffdio_register;
1202         unsigned long vm_flags, new_flags;
1203         bool found;
1204         bool basic_ioctls;
1205         unsigned long start, end, vma_end;
1206
1207         user_uffdio_register = (struct uffdio_register __user *) arg;
1208
1209         ret = -EFAULT;
1210         if (copy_from_user(&uffdio_register, user_uffdio_register,
1211                            sizeof(uffdio_register)-sizeof(__u64)))
1212                 goto out;
1213
1214         ret = -EINVAL;
1215         if (!uffdio_register.mode)
1216                 goto out;
1217         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1218                                      UFFDIO_REGISTER_MODE_WP))
1219                 goto out;
1220         vm_flags = 0;
1221         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1222                 vm_flags |= VM_UFFD_MISSING;
1223         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1224                 vm_flags |= VM_UFFD_WP;
1225                 /*
1226                  * FIXME: remove the below error constraint by
1227                  * implementing the wprotect tracking mode.
1228                  */
1229                 ret = -EINVAL;
1230                 goto out;
1231         }
1232
1233         ret = validate_range(mm, uffdio_register.range.start,
1234                              uffdio_register.range.len);
1235         if (ret)
1236                 goto out;
1237
1238         start = uffdio_register.range.start;
1239         end = start + uffdio_register.range.len;
1240
1241         ret = -ENOMEM;
1242         if (!mmget_not_zero(mm))
1243                 goto out;
1244
1245         down_write(&mm->mmap_sem);
1246         vma = find_vma_prev(mm, start, &prev);
1247         if (!vma)
1248                 goto out_unlock;
1249
1250         /* check that there's at least one vma in the range */
1251         ret = -EINVAL;
1252         if (vma->vm_start >= end)
1253                 goto out_unlock;
1254
1255         /*
1256          * If the first vma contains huge pages, make sure start address
1257          * is aligned to huge page size.
1258          */
1259         if (is_vm_hugetlb_page(vma)) {
1260                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1261
1262                 if (start & (vma_hpagesize - 1))
1263                         goto out_unlock;
1264         }
1265
1266         /*
1267          * Search for not compatible vmas.
1268          */
1269         found = false;
1270         basic_ioctls = false;
1271         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1272                 cond_resched();
1273
1274                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1275                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1276
1277                 /* check not compatible vmas */
1278                 ret = -EINVAL;
1279                 if (!vma_can_userfault(cur))
1280                         goto out_unlock;
1281                 /*
1282                  * If this vma contains ending address, and huge pages
1283                  * check alignment.
1284                  */
1285                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1286                     end > cur->vm_start) {
1287                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1288
1289                         ret = -EINVAL;
1290
1291                         if (end & (vma_hpagesize - 1))
1292                                 goto out_unlock;
1293                 }
1294
1295                 /*
1296                  * Check that this vma isn't already owned by a
1297                  * different userfaultfd. We can't allow more than one
1298                  * userfaultfd to own a single vma simultaneously or we
1299                  * wouldn't know which one to deliver the userfaults to.
1300                  */
1301                 ret = -EBUSY;
1302                 if (cur->vm_userfaultfd_ctx.ctx &&
1303                     cur->vm_userfaultfd_ctx.ctx != ctx)
1304                         goto out_unlock;
1305
1306                 /*
1307                  * Note vmas containing huge pages
1308                  */
1309                 if (is_vm_hugetlb_page(cur))
1310                         basic_ioctls = true;
1311
1312                 found = true;
1313         }
1314         BUG_ON(!found);
1315
1316         if (vma->vm_start < start)
1317                 prev = vma;
1318
1319         ret = 0;
1320         do {
1321                 cond_resched();
1322
1323                 BUG_ON(!vma_can_userfault(vma));
1324                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1325                        vma->vm_userfaultfd_ctx.ctx != ctx);
1326
1327                 /*
1328                  * Nothing to do: this vma is already registered into this
1329                  * userfaultfd and with the right tracking mode too.
1330                  */
1331                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1332                     (vma->vm_flags & vm_flags) == vm_flags)
1333                         goto skip;
1334
1335                 if (vma->vm_start > start)
1336                         start = vma->vm_start;
1337                 vma_end = min(end, vma->vm_end);
1338
1339                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1340                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1341                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1342                                  vma_policy(vma),
1343                                  ((struct vm_userfaultfd_ctx){ ctx }));
1344                 if (prev) {
1345                         vma = prev;
1346                         goto next;
1347                 }
1348                 if (vma->vm_start < start) {
1349                         ret = split_vma(mm, vma, start, 1);
1350                         if (ret)
1351                                 break;
1352                 }
1353                 if (vma->vm_end > end) {
1354                         ret = split_vma(mm, vma, end, 0);
1355                         if (ret)
1356                                 break;
1357                 }
1358         next:
1359                 /*
1360                  * In the vma_merge() successful mprotect-like case 8:
1361                  * the next vma was merged into the current one and
1362                  * the current one has not been updated yet.
1363                  */
1364                 vma->vm_flags = new_flags;
1365                 vma->vm_userfaultfd_ctx.ctx = ctx;
1366
1367         skip:
1368                 prev = vma;
1369                 start = vma->vm_end;
1370                 vma = vma->vm_next;
1371         } while (vma && vma->vm_start < end);
1372 out_unlock:
1373         up_write(&mm->mmap_sem);
1374         mmput(mm);
1375         if (!ret) {
1376                 /*
1377                  * Now that we scanned all vmas we can already tell
1378                  * userland which ioctls methods are guaranteed to
1379                  * succeed on this range.
1380                  */
1381                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1382                              UFFD_API_RANGE_IOCTLS,
1383                              &user_uffdio_register->ioctls))
1384                         ret = -EFAULT;
1385         }
1386 out:
1387         return ret;
1388 }
1389
1390 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1391                                   unsigned long arg)
1392 {
1393         struct mm_struct *mm = ctx->mm;
1394         struct vm_area_struct *vma, *prev, *cur;
1395         int ret;
1396         struct uffdio_range uffdio_unregister;
1397         unsigned long new_flags;
1398         bool found;
1399         unsigned long start, end, vma_end;
1400         const void __user *buf = (void __user *)arg;
1401
1402         ret = -EFAULT;
1403         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1404                 goto out;
1405
1406         ret = validate_range(mm, uffdio_unregister.start,
1407                              uffdio_unregister.len);
1408         if (ret)
1409                 goto out;
1410
1411         start = uffdio_unregister.start;
1412         end = start + uffdio_unregister.len;
1413
1414         ret = -ENOMEM;
1415         if (!mmget_not_zero(mm))
1416                 goto out;
1417
1418         down_write(&mm->mmap_sem);
1419         vma = find_vma_prev(mm, start, &prev);
1420         if (!vma)
1421                 goto out_unlock;
1422
1423         /* check that there's at least one vma in the range */
1424         ret = -EINVAL;
1425         if (vma->vm_start >= end)
1426                 goto out_unlock;
1427
1428         /*
1429          * If the first vma contains huge pages, make sure start address
1430          * is aligned to huge page size.
1431          */
1432         if (is_vm_hugetlb_page(vma)) {
1433                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1434
1435                 if (start & (vma_hpagesize - 1))
1436                         goto out_unlock;
1437         }
1438
1439         /*
1440          * Search for not compatible vmas.
1441          */
1442         found = false;
1443         ret = -EINVAL;
1444         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1445                 cond_resched();
1446
1447                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1448                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1449
1450                 /*
1451                  * Check not compatible vmas, not strictly required
1452                  * here as not compatible vmas cannot have an
1453                  * userfaultfd_ctx registered on them, but this
1454                  * provides for more strict behavior to notice
1455                  * unregistration errors.
1456                  */
1457                 if (!vma_can_userfault(cur))
1458                         goto out_unlock;
1459
1460                 found = true;
1461         }
1462         BUG_ON(!found);
1463
1464         if (vma->vm_start < start)
1465                 prev = vma;
1466
1467         ret = 0;
1468         do {
1469                 cond_resched();
1470
1471                 BUG_ON(!vma_can_userfault(vma));
1472
1473                 /*
1474                  * Nothing to do: this vma is already registered into this
1475                  * userfaultfd and with the right tracking mode too.
1476                  */
1477                 if (!vma->vm_userfaultfd_ctx.ctx)
1478                         goto skip;
1479
1480                 if (vma->vm_start > start)
1481                         start = vma->vm_start;
1482                 vma_end = min(end, vma->vm_end);
1483
1484                 if (userfaultfd_missing(vma)) {
1485                         /*
1486                          * Wake any concurrent pending userfault while
1487                          * we unregister, so they will not hang
1488                          * permanently and it avoids userland to call
1489                          * UFFDIO_WAKE explicitly.
1490                          */
1491                         struct userfaultfd_wake_range range;
1492                         range.start = start;
1493                         range.len = vma_end - start;
1494                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1495                 }
1496
1497                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1498                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1499                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1500                                  vma_policy(vma),
1501                                  NULL_VM_UFFD_CTX);
1502                 if (prev) {
1503                         vma = prev;
1504                         goto next;
1505                 }
1506                 if (vma->vm_start < start) {
1507                         ret = split_vma(mm, vma, start, 1);
1508                         if (ret)
1509                                 break;
1510                 }
1511                 if (vma->vm_end > end) {
1512                         ret = split_vma(mm, vma, end, 0);
1513                         if (ret)
1514                                 break;
1515                 }
1516         next:
1517                 /*
1518                  * In the vma_merge() successful mprotect-like case 8:
1519                  * the next vma was merged into the current one and
1520                  * the current one has not been updated yet.
1521                  */
1522                 vma->vm_flags = new_flags;
1523                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1524
1525         skip:
1526                 prev = vma;
1527                 start = vma->vm_end;
1528                 vma = vma->vm_next;
1529         } while (vma && vma->vm_start < end);
1530 out_unlock:
1531         up_write(&mm->mmap_sem);
1532         mmput(mm);
1533 out:
1534         return ret;
1535 }
1536
1537 /*
1538  * userfaultfd_wake may be used in combination with the
1539  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1540  */
1541 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1542                             unsigned long arg)
1543 {
1544         int ret;
1545         struct uffdio_range uffdio_wake;
1546         struct userfaultfd_wake_range range;
1547         const void __user *buf = (void __user *)arg;
1548
1549         ret = -EFAULT;
1550         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1551                 goto out;
1552
1553         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1554         if (ret)
1555                 goto out;
1556
1557         range.start = uffdio_wake.start;
1558         range.len = uffdio_wake.len;
1559
1560         /*
1561          * len == 0 means wake all and we don't want to wake all here,
1562          * so check it again to be sure.
1563          */
1564         VM_BUG_ON(!range.len);
1565
1566         wake_userfault(ctx, &range);
1567         ret = 0;
1568
1569 out:
1570         return ret;
1571 }
1572
1573 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1574                             unsigned long arg)
1575 {
1576         __s64 ret;
1577         struct uffdio_copy uffdio_copy;
1578         struct uffdio_copy __user *user_uffdio_copy;
1579         struct userfaultfd_wake_range range;
1580
1581         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1582
1583         ret = -EFAULT;
1584         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1585                            /* don't copy "copy" last field */
1586                            sizeof(uffdio_copy)-sizeof(__s64)))
1587                 goto out;
1588
1589         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1590         if (ret)
1591                 goto out;
1592         /*
1593          * double check for wraparound just in case. copy_from_user()
1594          * will later check uffdio_copy.src + uffdio_copy.len to fit
1595          * in the userland range.
1596          */
1597         ret = -EINVAL;
1598         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1599                 goto out;
1600         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1601                 goto out;
1602         if (mmget_not_zero(ctx->mm)) {
1603                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1604                                    uffdio_copy.len);
1605                 mmput(ctx->mm);
1606         } else {
1607                 return -ESRCH;
1608         }
1609         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1610                 return -EFAULT;
1611         if (ret < 0)
1612                 goto out;
1613         BUG_ON(!ret);
1614         /* len == 0 would wake all */
1615         range.len = ret;
1616         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1617                 range.start = uffdio_copy.dst;
1618                 wake_userfault(ctx, &range);
1619         }
1620         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1621 out:
1622         return ret;
1623 }
1624
1625 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1626                                 unsigned long arg)
1627 {
1628         __s64 ret;
1629         struct uffdio_zeropage uffdio_zeropage;
1630         struct uffdio_zeropage __user *user_uffdio_zeropage;
1631         struct userfaultfd_wake_range range;
1632
1633         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1634
1635         ret = -EFAULT;
1636         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1637                            /* don't copy "zeropage" last field */
1638                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1639                 goto out;
1640
1641         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1642                              uffdio_zeropage.range.len);
1643         if (ret)
1644                 goto out;
1645         ret = -EINVAL;
1646         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1647                 goto out;
1648
1649         if (mmget_not_zero(ctx->mm)) {
1650                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1651                                      uffdio_zeropage.range.len);
1652                 mmput(ctx->mm);
1653         } else {
1654                 return -ESRCH;
1655         }
1656         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1657                 return -EFAULT;
1658         if (ret < 0)
1659                 goto out;
1660         /* len == 0 would wake all */
1661         BUG_ON(!ret);
1662         range.len = ret;
1663         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1664                 range.start = uffdio_zeropage.range.start;
1665                 wake_userfault(ctx, &range);
1666         }
1667         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1668 out:
1669         return ret;
1670 }
1671
1672 static inline unsigned int uffd_ctx_features(__u64 user_features)
1673 {
1674         /*
1675          * For the current set of features the bits just coincide
1676          */
1677         return (unsigned int)user_features;
1678 }
1679
1680 /*
1681  * userland asks for a certain API version and we return which bits
1682  * and ioctl commands are implemented in this kernel for such API
1683  * version or -EINVAL if unknown.
1684  */
1685 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1686                            unsigned long arg)
1687 {
1688         struct uffdio_api uffdio_api;
1689         void __user *buf = (void __user *)arg;
1690         int ret;
1691         __u64 features;
1692
1693         ret = -EINVAL;
1694         if (ctx->state != UFFD_STATE_WAIT_API)
1695                 goto out;
1696         ret = -EFAULT;
1697         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1698                 goto out;
1699         features = uffdio_api.features;
1700         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1701                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1702                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1703                         goto out;
1704                 ret = -EINVAL;
1705                 goto out;
1706         }
1707         /* report all available features and ioctls to userland */
1708         uffdio_api.features = UFFD_API_FEATURES;
1709         uffdio_api.ioctls = UFFD_API_IOCTLS;
1710         ret = -EFAULT;
1711         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1712                 goto out;
1713         ctx->state = UFFD_STATE_RUNNING;
1714         /* only enable the requested features for this uffd context */
1715         ctx->features = uffd_ctx_features(features);
1716         ret = 0;
1717 out:
1718         return ret;
1719 }
1720
1721 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1722                               unsigned long arg)
1723 {
1724         int ret = -EINVAL;
1725         struct userfaultfd_ctx *ctx = file->private_data;
1726
1727         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1728                 return -EINVAL;
1729
1730         switch(cmd) {
1731         case UFFDIO_API:
1732                 ret = userfaultfd_api(ctx, arg);
1733                 break;
1734         case UFFDIO_REGISTER:
1735                 ret = userfaultfd_register(ctx, arg);
1736                 break;
1737         case UFFDIO_UNREGISTER:
1738                 ret = userfaultfd_unregister(ctx, arg);
1739                 break;
1740         case UFFDIO_WAKE:
1741                 ret = userfaultfd_wake(ctx, arg);
1742                 break;
1743         case UFFDIO_COPY:
1744                 ret = userfaultfd_copy(ctx, arg);
1745                 break;
1746         case UFFDIO_ZEROPAGE:
1747                 ret = userfaultfd_zeropage(ctx, arg);
1748                 break;
1749         }
1750         return ret;
1751 }
1752
1753 #ifdef CONFIG_PROC_FS
1754 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1755 {
1756         struct userfaultfd_ctx *ctx = f->private_data;
1757         wait_queue_entry_t *wq;
1758         struct userfaultfd_wait_queue *uwq;
1759         unsigned long pending = 0, total = 0;
1760
1761         spin_lock(&ctx->fault_pending_wqh.lock);
1762         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1763                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1764                 pending++;
1765                 total++;
1766         }
1767         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1768                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1769                 total++;
1770         }
1771         spin_unlock(&ctx->fault_pending_wqh.lock);
1772
1773         /*
1774          * If more protocols will be added, there will be all shown
1775          * separated by a space. Like this:
1776          *      protocols: aa:... bb:...
1777          */
1778         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1779                    pending, total, UFFD_API, ctx->features,
1780                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1781 }
1782 #endif
1783
1784 static const struct file_operations userfaultfd_fops = {
1785 #ifdef CONFIG_PROC_FS
1786         .show_fdinfo    = userfaultfd_show_fdinfo,
1787 #endif
1788         .release        = userfaultfd_release,
1789         .poll           = userfaultfd_poll,
1790         .read           = userfaultfd_read,
1791         .unlocked_ioctl = userfaultfd_ioctl,
1792         .compat_ioctl   = userfaultfd_ioctl,
1793         .llseek         = noop_llseek,
1794 };
1795
1796 static void init_once_userfaultfd_ctx(void *mem)
1797 {
1798         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1799
1800         init_waitqueue_head(&ctx->fault_pending_wqh);
1801         init_waitqueue_head(&ctx->fault_wqh);
1802         init_waitqueue_head(&ctx->event_wqh);
1803         init_waitqueue_head(&ctx->fd_wqh);
1804         seqcount_init(&ctx->refile_seq);
1805 }
1806
1807 /**
1808  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1809  * @flags: Flags for the userfaultfd file.
1810  *
1811  * This function creates a userfaultfd file pointer, w/out installing
1812  * it into the fd table. This is useful when the userfaultfd file is
1813  * used during the initialization of data structures that require
1814  * extra setup after the userfaultfd creation. So the userfaultfd
1815  * creation is split into the file pointer creation phase, and the
1816  * file descriptor installation phase.  In this way races with
1817  * userspace closing the newly installed file descriptor can be
1818  * avoided.  Returns a userfaultfd file pointer, or a proper error
1819  * pointer.
1820  */
1821 static struct file *userfaultfd_file_create(int flags)
1822 {
1823         struct file *file;
1824         struct userfaultfd_ctx *ctx;
1825
1826         BUG_ON(!current->mm);
1827
1828         /* Check the UFFD_* constants for consistency.  */
1829         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1830         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1831
1832         file = ERR_PTR(-EINVAL);
1833         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1834                 goto out;
1835
1836         file = ERR_PTR(-ENOMEM);
1837         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1838         if (!ctx)
1839                 goto out;
1840
1841         atomic_set(&ctx->refcount, 1);
1842         ctx->flags = flags;
1843         ctx->features = 0;
1844         ctx->state = UFFD_STATE_WAIT_API;
1845         ctx->released = false;
1846         ctx->mm = current->mm;
1847         /* prevent the mm struct to be freed */
1848         mmgrab(ctx->mm);
1849
1850         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1851                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1852         if (IS_ERR(file)) {
1853                 mmdrop(ctx->mm);
1854                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1855         }
1856 out:
1857         return file;
1858 }
1859
1860 SYSCALL_DEFINE1(userfaultfd, int, flags)
1861 {
1862         int fd, error;
1863         struct file *file;
1864
1865         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1866         if (error < 0)
1867                 return error;
1868         fd = error;
1869
1870         file = userfaultfd_file_create(flags);
1871         if (IS_ERR(file)) {
1872                 error = PTR_ERR(file);
1873                 goto err_put_unused_fd;
1874         }
1875         fd_install(fd, file);
1876
1877         return fd;
1878
1879 err_put_unused_fd:
1880         put_unused_fd(fd);
1881
1882         return error;
1883 }
1884
1885 static int __init userfaultfd_init(void)
1886 {
1887         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1888                                                 sizeof(struct userfaultfd_ctx),
1889                                                 0,
1890                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1891                                                 init_once_userfaultfd_ctx);
1892         return 0;
1893 }
1894 __initcall(userfaultfd_init);