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 {
 183         struct uffd_msg msg;
 184         msg_init(&msg);
 185         msg.event = UFFD_EVENT_PAGEFAULT;
 186         msg.arg.pagefault.address = address;
 187         if (flags & FAULT_FLAG_WRITE)
 188                 /*
 189                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 190                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 191                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 192                  * was a read fault, otherwise if set it means it's
 193                  * a write fault.
 194                  */
 195                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 196         if (reason & VM_UFFD_WP)
 197                 /*
 198                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 199                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 200                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 201                  * a missing fault, otherwise if set it means it's a
 202                  * write protect fault.
 203                  */
 204                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 205         return msg;
 206 }
 207
 208 #ifdef CONFIG_HUGETLB_PAGE
 209 /*
 210  * Same functionality as userfaultfd_must_wait below with modifications for
 211  * hugepmd ranges.
 212  */
 213 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 214                                          struct vm_area_struct *vma,
 215                                          unsigned long address,
 216                                          unsigned long flags,
 217                                          unsigned long reason)
 218 {
 219         struct mm_struct *mm = ctx->mm;
 220         pte_t *pte;
 221         bool ret = true;
 222
 223         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 224
 225         pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 226         if (!pte)
 227                 goto out;
 228
 229         ret = false;
 230
 231         /*
 232          * Lockless access: we're in a wait_event so it's ok if it
 233          * changes under us.
 234          */
 235         if (huge_pte_none(*pte))
 236                 ret = true;
 237         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
 238                 ret = true;
 239 out:
 240         return ret;
 241 }
 242 #else
 243 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 244                                          struct vm_area_struct *vma,
 245                                          unsigned long address,
 246                                          unsigned long flags,
 247                                          unsigned long reason)
 248 {
 249         return false;   /* should never get here */
 250 }
 251 #endif /* CONFIG_HUGETLB_PAGE */
 252
 253 /*
 254  * Verify the pagetables are still not ok after having reigstered into
 255  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 256  * userfault that has already been resolved, if userfaultfd_read and
 257  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 258  * threads.
 259  */
 260 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 261                                          unsigned long address,
 262                                          unsigned long flags,
 263                                          unsigned long reason)
 264 {
 265         struct mm_struct *mm = ctx->mm;
 266         pgd_t *pgd;
 267         p4d_t *p4d;
 268         pud_t *pud;
 269         pmd_t *pmd, _pmd;
 270         pte_t *pte;
 271         bool ret = true;
 272
 273         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 274
 275         pgd = pgd_offset(mm, address);
 276         if (!pgd_present(*pgd))
 277                 goto out;
 278         p4d = p4d_offset(pgd, address);
 279         if (!p4d_present(*p4d))
 280                 goto out;
 281         pud = pud_offset(p4d, address);
 282         if (!pud_present(*pud))
 283                 goto out;
 284         pmd = pmd_offset(pud, address);
 285         /*
 286          * READ_ONCE must function as a barrier with narrower scope
 287          * and it must be equivalent to:
 288          *      _pmd = *pmd; barrier();
 289          *
 290          * This is to deal with the instability (as in
 291          * pmd_trans_unstable) of the pmd.
 292          */
 293         _pmd = READ_ONCE(*pmd);
 294         if (!pmd_present(_pmd))
 295                 goto out;
 296
 297         ret = false;
 298         if (pmd_trans_huge(_pmd))
 299                 goto out;
 300
 301         /*
 302          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 303          * and use the standard pte_offset_map() instead of parsing _pmd.
 304          */
 305         pte = pte_offset_map(pmd, address);
 306         /*
 307          * Lockless access: we're in a wait_event so it's ok if it
 308          * changes under us.
 309          */
 310         if (pte_none(*pte))
 311                 ret = true;
 312         pte_unmap(pte);
 313
 314 out:
 315         return ret;
 316 }
 317
 318 /*
 319  * The locking rules involved in returning VM_FAULT_RETRY depending on
 320  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 321  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 322  * recommendation in __lock_page_or_retry is not an understatement.
 323  *
 324  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 325  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 326  * not set.
 327  *
 328  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 329  * set, VM_FAULT_RETRY can still be returned if and only if there are
 330  * fatal_signal_pending()s, and the mmap_sem must be released before
 331  * returning it.
 332  */
 333 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 334 {
 335         struct mm_struct *mm = vmf->vma->vm_mm;
 336         struct userfaultfd_ctx *ctx;
 337         struct userfaultfd_wait_queue uwq;
 338         int ret;
 339         bool must_wait, return_to_userland;
 340         long blocking_state;
 341
 342         ret = VM_FAULT_SIGBUS;
 343
 344         /*
 345          * We don't do userfault handling for the final child pid update.
 346          *
 347          * We also don't do userfault handling during
 348          * coredumping. hugetlbfs has the special
 349          * follow_hugetlb_page() to skip missing pages in the
 350          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 351          * the no_page_table() helper in follow_page_mask(), but the
 352          * shmem_vm_ops->fault method is invoked even during
 353          * coredumping without mmap_sem and it ends up here.
 354          */
 355         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 356                 goto out;
 357
 358         /*
 359          * Coredumping runs without mmap_sem so we can only check that
 360          * the mmap_sem is held, if PF_DUMPCORE was not set.
 361          */
 362         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
 363
 364         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 365         if (!ctx)
 366                 goto out;
 367
 368         BUG_ON(ctx->mm != mm);
 369
 370         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 371         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 372
 373         /*
 374          * If it's already released don't get it. This avoids to loop
 375          * in __get_user_pages if userfaultfd_release waits on the
 376          * caller of handle_userfault to release the mmap_sem.
 377          */
 378         if (unlikely(ACCESS_ONCE(ctx->released)))
 379                 goto out;
 380
 381         /*
 382          * Check that we can return VM_FAULT_RETRY.
 383          *
 384          * NOTE: it should become possible to return VM_FAULT_RETRY
 385          * even if FAULT_FLAG_TRIED is set without leading to gup()
 386          * -EBUSY failures, if the userfaultfd is to be extended for
 387          * VM_UFFD_WP tracking and we intend to arm the userfault
 388          * without first stopping userland access to the memory. For
 389          * VM_UFFD_MISSING userfaults this is enough for now.
 390          */
 391         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 392                 /*
 393                  * Validate the invariant that nowait must allow retry
 394                  * to be sure not to return SIGBUS erroneously on
 395                  * nowait invocations.
 396                  */
 397                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 398 #ifdef CONFIG_DEBUG_VM
 399                 if (printk_ratelimit()) {
 400                         printk(KERN_WARNING
 401                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 402                                vmf->flags);
 403                         dump_stack();
 404                 }
 405 #endif
 406                 goto out;
 407         }
 408
 409         /*
 410          * Handle nowait, not much to do other than tell it to retry
 411          * and wait.
 412          */
 413         ret = VM_FAULT_RETRY;
 414         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 415                 goto out;
 416
 417         /* take the reference before dropping the mmap_sem */
 418         userfaultfd_ctx_get(ctx);
 419
 420         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 421         uwq.wq.private = current;
 422         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
 423         uwq.ctx = ctx;
 424         uwq.waken = false;
 425
 426         return_to_userland =
 427                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 428                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 429         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 430                          TASK_KILLABLE;
 431
 432         spin_lock(&ctx->fault_pending_wqh.lock);
 433         /*
 434          * After the __add_wait_queue the uwq is visible to userland
 435          * through poll/read().
 436          */
 437         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 438         /*
 439          * The smp_mb() after __set_current_state prevents the reads
 440          * following the spin_unlock to happen before the list_add in
 441          * __add_wait_queue.
 442          */
 443         set_current_state(blocking_state);
 444         spin_unlock(&ctx->fault_pending_wqh.lock);
 445
 446         if (!is_vm_hugetlb_page(vmf->vma))
 447                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 448                                                   reason);
 449         else
 450                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 451                                                        vmf->address,
 452                                                        vmf->flags, reason);
 453         up_read(&mm->mmap_sem);
 454
 455         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
 456                    (return_to_userland ? !signal_pending(current) :
 457                     !fatal_signal_pending(current)))) {
 458                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 459                 schedule();
 460                 ret |= VM_FAULT_MAJOR;
 461
 462                 /*
 463                  * False wakeups can orginate even from rwsem before
 464                  * up_read() however userfaults will wait either for a
 465                  * targeted wakeup on the specific uwq waitqueue from
 466                  * wake_userfault() or for signals or for uffd
 467                  * release.
 468                  */
 469                 while (!READ_ONCE(uwq.waken)) {
 470                         /*
 471                          * This needs the full smp_store_mb()
 472                          * guarantee as the state write must be
 473                          * visible to other CPUs before reading
 474                          * uwq.waken from other CPUs.
 475                          */
 476                         set_current_state(blocking_state);
 477                         if (READ_ONCE(uwq.waken) ||
 478                             READ_ONCE(ctx->released) ||
 479                             (return_to_userland ? signal_pending(current) :
 480                              fatal_signal_pending(current)))
 481                                 break;
 482                         schedule();
 483                 }
 484         }
 485
 486         __set_current_state(TASK_RUNNING);
 487
 488         if (return_to_userland) {
 489                 if (signal_pending(current) &&
 490                     !fatal_signal_pending(current)) {
 491                         /*
 492                          * If we got a SIGSTOP or SIGCONT and this is
 493                          * a normal userland page fault, just let
 494                          * userland return so the signal will be
 495                          * handled and gdb debugging works.  The page
 496                          * fault code immediately after we return from
 497                          * this function is going to release the
 498                          * mmap_sem and it's not depending on it
 499                          * (unlike gup would if we were not to return
 500                          * VM_FAULT_RETRY).
 501                          *
 502                          * If a fatal signal is pending we still take
 503                          * the streamlined VM_FAULT_RETRY failure path
 504                          * and there's no need to retake the mmap_sem
 505                          * in such case.
 506                          */
 507                         down_read(&mm->mmap_sem);
 508                         ret = VM_FAULT_NOPAGE;
 509                 }
 510         }
 511
 512         /*
 513          * Here we race with the list_del; list_add in
 514          * userfaultfd_ctx_read(), however because we don't ever run
 515          * list_del_init() to refile across the two lists, the prev
 516          * and next pointers will never point to self. list_add also
 517          * would never let any of the two pointers to point to
 518          * self. So list_empty_careful won't risk to see both pointers
 519          * pointing to self at any time during the list refile. The
 520          * only case where list_del_init() is called is the full
 521          * removal in the wake function and there we don't re-list_add
 522          * and it's fine not to block on the spinlock. The uwq on this
 523          * kernel stack can be released after the list_del_init.
 524          */
 525         if (!list_empty_careful(&uwq.wq.entry)) {
 526                 spin_lock(&ctx->fault_pending_wqh.lock);
 527                 /*
 528                  * No need of list_del_init(), the uwq on the stack
 529                  * will be freed shortly anyway.
 530                  */
 531                 list_del(&uwq.wq.entry);
 532                 spin_unlock(&ctx->fault_pending_wqh.lock);
 533         }
 534
 535         /*
 536          * ctx may go away after this if the userfault pseudo fd is
 537          * already released.
 538          */
 539         userfaultfd_ctx_put(ctx);
 540
 541 out:
 542         return ret;
 543 }
 544
 545 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 546                                               struct userfaultfd_wait_queue *ewq)
 547 {
 548         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 549                 goto out;
 550
 551         ewq->ctx = ctx;
 552         init_waitqueue_entry(&ewq->wq, current);
 553
 554         spin_lock(&ctx->event_wqh.lock);
 555         /*
 556          * After the __add_wait_queue the uwq is visible to userland
 557          * through poll/read().
 558          */
 559         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 560         for (;;) {
 561                 set_current_state(TASK_KILLABLE);
 562                 if (ewq->msg.event == 0)
 563                         break;
 564                 if (ACCESS_ONCE(ctx->released) ||
 565                     fatal_signal_pending(current)) {
 566                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 567                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 568                                 struct userfaultfd_ctx *new;
 569
 570                                 new = (struct userfaultfd_ctx *)
 571                                         (unsigned long)
 572                                         ewq->msg.arg.reserved.reserved1;
 573
 574                                 userfaultfd_ctx_put(new);
 575                         }
 576                         break;
 577                 }
 578
 579                 spin_unlock(&ctx->event_wqh.lock);
 580
 581                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 582                 schedule();
 583
 584                 spin_lock(&ctx->event_wqh.lock);
 585         }
 586         __set_current_state(TASK_RUNNING);
 587         spin_unlock(&ctx->event_wqh.lock);
 588
 589         /*
 590          * ctx may go away after this if the userfault pseudo fd is
 591          * already released.
 592          */
 593 out:
 594         userfaultfd_ctx_put(ctx);
 595 }
 596
 597 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 598                                        struct userfaultfd_wait_queue *ewq)
 599 {
 600         ewq->msg.event = 0;
 601         wake_up_locked(&ctx->event_wqh);
 602         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 603 }
 604
 605 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 606 {
 607         struct userfaultfd_ctx *ctx = NULL, *octx;
 608         struct userfaultfd_fork_ctx *fctx;
 609
 610         octx = vma->vm_userfaultfd_ctx.ctx;
 611         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 612                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 613                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 614                 return 0;
 615         }
 616
 617         list_for_each_entry(fctx, fcs, list)
 618                 if (fctx->orig == octx) {
 619                         ctx = fctx->new;
 620                         break;
 621                 }
 622
 623         if (!ctx) {
 624                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 625                 if (!fctx)
 626                         return -ENOMEM;
 627
 628                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 629                 if (!ctx) {
 630                         kfree(fctx);
 631                         return -ENOMEM;
 632                 }
 633
 634                 atomic_set(&ctx->refcount, 1);
 635                 ctx->flags = octx->flags;
 636                 ctx->state = UFFD_STATE_RUNNING;
 637                 ctx->features = octx->features;
 638                 ctx->released = false;
 639                 ctx->mm = vma->vm_mm;
 640                 atomic_inc(&ctx->mm->mm_count);
 641
 642                 userfaultfd_ctx_get(octx);
 643                 fctx->orig = octx;
 644                 fctx->new = ctx;
 645                 list_add_tail(&fctx->list, fcs);
 646         }
 647
 648         vma->vm_userfaultfd_ctx.ctx = ctx;
 649         return 0;
 650 }
 651
 652 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 653 {
 654         struct userfaultfd_ctx *ctx = fctx->orig;
 655         struct userfaultfd_wait_queue ewq;
 656
 657         msg_init(&ewq.msg);
 658
 659         ewq.msg.event = UFFD_EVENT_FORK;
 660         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 661
 662         userfaultfd_event_wait_completion(ctx, &ewq);
 663 }
 664
 665 void dup_userfaultfd_complete(struct list_head *fcs)
 666 {
 667         struct userfaultfd_fork_ctx *fctx, *n;
 668
 669         list_for_each_entry_safe(fctx, n, fcs, list) {
 670                 dup_fctx(fctx);
 671                 list_del(&fctx->list);
 672                 kfree(fctx);
 673         }
 674 }
 675
 676 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 677                              struct vm_userfaultfd_ctx *vm_ctx)
 678 {
 679         struct userfaultfd_ctx *ctx;
 680
 681         ctx = vma->vm_userfaultfd_ctx.ctx;
 682         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 683                 vm_ctx->ctx = ctx;
 684                 userfaultfd_ctx_get(ctx);
 685         }
 686 }
 687
 688 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 689                                  unsigned long from, unsigned long to,
 690                                  unsigned long len)
 691 {
 692         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 693         struct userfaultfd_wait_queue ewq;
 694
 695         if (!ctx)
 696                 return;
 697
 698         if (to & ~PAGE_MASK) {
 699                 userfaultfd_ctx_put(ctx);
 700                 return;
 701         }
 702
 703         msg_init(&ewq.msg);
 704
 705         ewq.msg.event = UFFD_EVENT_REMAP;
 706         ewq.msg.arg.remap.from = from;
 707         ewq.msg.arg.remap.to = to;
 708         ewq.msg.arg.remap.len = len;
 709
 710         userfaultfd_event_wait_completion(ctx, &ewq);
 711 }
 712
 713 bool userfaultfd_remove(struct vm_area_struct *vma,
 714                         unsigned long start, unsigned long end)
 715 {
 716         struct mm_struct *mm = vma->vm_mm;
 717         struct userfaultfd_ctx *ctx;
 718         struct userfaultfd_wait_queue ewq;
 719
 720         ctx = vma->vm_userfaultfd_ctx.ctx;
 721         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 722                 return true;
 723
 724         userfaultfd_ctx_get(ctx);
 725         up_read(&mm->mmap_sem);
 726
 727         msg_init(&ewq.msg);
 728
 729         ewq.msg.event = UFFD_EVENT_REMOVE;
 730         ewq.msg.arg.remove.start = start;
 731         ewq.msg.arg.remove.end = end;
 732
 733         userfaultfd_event_wait_completion(ctx, &ewq);
 734
 735         return false;
 736 }
 737
 738 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 739                           unsigned long start, unsigned long end)
 740 {
 741         struct userfaultfd_unmap_ctx *unmap_ctx;
 742
 743         list_for_each_entry(unmap_ctx, unmaps, list)
 744                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 745                     unmap_ctx->end == end)
 746                         return true;
 747
 748         return false;
 749 }
 750
 751 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 752                            unsigned long start, unsigned long end,
 753                            struct list_head *unmaps)
 754 {
 755         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 756                 struct userfaultfd_unmap_ctx *unmap_ctx;
 757                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 758
 759                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 760                     has_unmap_ctx(ctx, unmaps, start, end))
 761                         continue;
 762
 763                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 764                 if (!unmap_ctx)
 765                         return -ENOMEM;
 766
 767                 userfaultfd_ctx_get(ctx);
 768                 unmap_ctx->ctx = ctx;
 769                 unmap_ctx->start = start;
 770                 unmap_ctx->end = end;
 771                 list_add_tail(&unmap_ctx->list, unmaps);
 772         }
 773
 774         return 0;
 775 }
 776
 777 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 778 {
 779         struct userfaultfd_unmap_ctx *ctx, *n;
 780         struct userfaultfd_wait_queue ewq;
 781
 782         list_for_each_entry_safe(ctx, n, uf, list) {
 783                 msg_init(&ewq.msg);
 784
 785                 ewq.msg.event = UFFD_EVENT_UNMAP;
 786                 ewq.msg.arg.remove.start = ctx->start;
 787                 ewq.msg.arg.remove.end = ctx->end;
 788
 789                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 790
 791                 list_del(&ctx->list);
 792                 kfree(ctx);
 793         }
 794 }
 795
 796 static int userfaultfd_release(struct inode *inode, struct file *file)
 797 {
 798         struct userfaultfd_ctx *ctx = file->private_data;
 799         struct mm_struct *mm = ctx->mm;
 800         struct vm_area_struct *vma, *prev;
 801         /* len == 0 means wake all */
 802         struct userfaultfd_wake_range range = { .len = 0, };
 803         unsigned long new_flags;
 804
 805         ACCESS_ONCE(ctx->released) = true;
 806
 807         if (!mmget_not_zero(mm))
 808                 goto wakeup;
 809
 810         /*
 811          * Flush page faults out of all CPUs. NOTE: all page faults
 812          * must be retried without returning VM_FAULT_SIGBUS if
 813          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 814          * changes while handle_userfault released the mmap_sem. So
 815          * it's critical that released is set to true (above), before
 816          * taking the mmap_sem for writing.
 817          */
 818         down_write(&mm->mmap_sem);
 819         prev = NULL;
 820         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 821                 cond_resched();
 822                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 823                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 824                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 825                         prev = vma;
 826                         continue;
 827                 }
 828                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 829                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 830                                  new_flags, vma->anon_vma,
 831                                  vma->vm_file, vma->vm_pgoff,
 832                                  vma_policy(vma),
 833                                  NULL_VM_UFFD_CTX);
 834                 if (prev)
 835                         vma = prev;
 836                 else
 837                         prev = vma;
 838                 vma->vm_flags = new_flags;
 839                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 840         }
 841         up_write(&mm->mmap_sem);
 842         mmput(mm);
 843 wakeup:
 844         /*
 845          * After no new page faults can wait on this fault_*wqh, flush
 846          * the last page faults that may have been already waiting on
 847          * the fault_*wqh.
 848          */
 849         spin_lock(&ctx->fault_pending_wqh.lock);
 850         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 851         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 852         spin_unlock(&ctx->fault_pending_wqh.lock);
 853
 854         /* Flush pending events that may still wait on event_wqh */
 855         wake_up_all(&ctx->event_wqh);
 856
 857         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 858         userfaultfd_ctx_put(ctx);
 859         return 0;
 860 }
 861
 862 /* fault_pending_wqh.lock must be hold by the caller */
 863 static inline struct userfaultfd_wait_queue *find_userfault_in(
 864                 wait_queue_head_t *wqh)
 865 {
 866         wait_queue_entry_t *wq;
 867         struct userfaultfd_wait_queue *uwq;
 868
 869         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 870
 871         uwq = NULL;
 872         if (!waitqueue_active(wqh))
 873                 goto out;
 874         /* walk in reverse to provide FIFO behavior to read userfaults */
 875         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 876         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 877 out:
 878         return uwq;
 879 }
 880
 881 static inline struct userfaultfd_wait_queue *find_userfault(
 882                 struct userfaultfd_ctx *ctx)
 883 {
 884         return find_userfault_in(&ctx->fault_pending_wqh);
 885 }
 886
 887 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 888                 struct userfaultfd_ctx *ctx)
 889 {
 890         return find_userfault_in(&ctx->event_wqh);
 891 }
 892
 893 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 894 {
 895         struct userfaultfd_ctx *ctx = file->private_data;
 896         unsigned int ret;
 897
 898         poll_wait(file, &ctx->fd_wqh, wait);
 899
 900         switch (ctx->state) {
 901         case UFFD_STATE_WAIT_API:
 902                 return POLLERR;
 903         case UFFD_STATE_RUNNING:
 904                 /*
 905                  * poll() never guarantees that read won't block.
 906                  * userfaults can be waken before they're read().
 907                  */
 908                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 909                         return POLLERR;
 910                 /*
 911                  * lockless access to see if there are pending faults
 912                  * __pollwait last action is the add_wait_queue but
 913                  * the spin_unlock would allow the waitqueue_active to
 914                  * pass above the actual list_add inside
 915                  * add_wait_queue critical section. So use a full
 916                  * memory barrier to serialize the list_add write of
 917                  * add_wait_queue() with the waitqueue_active read
 918                  * below.
 919                  */
 920                 ret = 0;
 921                 smp_mb();
 922                 if (waitqueue_active(&ctx->fault_pending_wqh))
 923                         ret = POLLIN;
 924                 else if (waitqueue_active(&ctx->event_wqh))
 925                         ret = POLLIN;
 926
 927                 return ret;
 928         default:
 929                 WARN_ON_ONCE(1);
 930                 return POLLERR;
 931         }
 932 }
 933
 934 static const struct file_operations userfaultfd_fops;
 935
 936 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 937                                   struct userfaultfd_ctx *new,
 938                                   struct uffd_msg *msg)
 939 {
 940         int fd;
 941         struct file *file;
 942         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 943
 944         fd = get_unused_fd_flags(flags);
 945         if (fd < 0)
 946                 return fd;
 947
 948         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 949                                   O_RDWR | flags);
 950         if (IS_ERR(file)) {
 951                 put_unused_fd(fd);
 952                 return PTR_ERR(file);
 953         }
 954
 955         fd_install(fd, file);
 956         msg->arg.reserved.reserved1 = 0;
 957         msg->arg.fork.ufd = fd;
 958
 959         return 0;
 960 }
 961
 962 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 963                                     struct uffd_msg *msg)
 964 {
 965         ssize_t ret;
 966         DECLARE_WAITQUEUE(wait, current);
 967         struct userfaultfd_wait_queue *uwq;
 968         /*
 969          * Handling fork event requires sleeping operations, so
 970          * we drop the event_wqh lock, then do these ops, then
 971          * lock it back and wake up the waiter. While the lock is
 972          * dropped the ewq may go away so we keep track of it
 973          * carefully.
 974          */
 975         LIST_HEAD(fork_event);
 976         struct userfaultfd_ctx *fork_nctx = NULL;
 977
 978         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 979         spin_lock(&ctx->fd_wqh.lock);
 980         __add_wait_queue(&ctx->fd_wqh, &wait);
 981         for (;;) {
 982                 set_current_state(TASK_INTERRUPTIBLE);
 983                 spin_lock(&ctx->fault_pending_wqh.lock);
 984                 uwq = find_userfault(ctx);
 985                 if (uwq) {
 986                         /*
 987                          * Use a seqcount to repeat the lockless check
 988                          * in wake_userfault() to avoid missing
 989                          * wakeups because during the refile both
 990                          * waitqueue could become empty if this is the
 991                          * only userfault.
 992                          */
 993                         write_seqcount_begin(&ctx->refile_seq);
 994
 995                         /*
 996                          * The fault_pending_wqh.lock prevents the uwq
 997                          * to disappear from under us.
 998                          *
 999                          * Refile this userfault from
1000                          * fault_pending_wqh to fault_wqh, it's not
1001                          * pending anymore after we read it.
1002                          *
1003                          * Use list_del() by hand (as
1004                          * userfaultfd_wake_function also uses
1005                          * list_del_init() by hand) to be sure nobody
1006                          * changes __remove_wait_queue() to use
1007                          * list_del_init() in turn breaking the
1008                          * !list_empty_careful() check in
1009                          * handle_userfault(). The uwq->wq.head list
1010                          * must never be empty at any time during the
1011                          * refile, or the waitqueue could disappear
1012                          * from under us. The "wait_queue_head_t"
1013                          * parameter of __remove_wait_queue() is unused
1014                          * anyway.
1015                          */
1016                         list_del(&uwq->wq.entry);
1017                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1018
1019                         write_seqcount_end(&ctx->refile_seq);
1020
1021                         /* careful to always initialize msg if ret == 0 */
1022                         *msg = uwq->msg;
1023                         spin_unlock(&ctx->fault_pending_wqh.lock);
1024                         ret = 0;
1025                         break;
1026                 }
1027                 spin_unlock(&ctx->fault_pending_wqh.lock);
1028
1029                 spin_lock(&ctx->event_wqh.lock);
1030                 uwq = find_userfault_evt(ctx);
1031                 if (uwq) {
1032                         *msg = uwq->msg;
1033
1034                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1035                                 fork_nctx = (struct userfaultfd_ctx *)
1036                                         (unsigned long)
1037                                         uwq->msg.arg.reserved.reserved1;
1038                                 list_move(&uwq->wq.entry, &fork_event);
1039                                 spin_unlock(&ctx->event_wqh.lock);
1040                                 ret = 0;
1041                                 break;
1042                         }
1043
1044                         userfaultfd_event_complete(ctx, uwq);
1045                         spin_unlock(&ctx->event_wqh.lock);
1046                         ret = 0;
1047                         break;
1048                 }
1049                 spin_unlock(&ctx->event_wqh.lock);
1050
1051                 if (signal_pending(current)) {
1052                         ret = -ERESTARTSYS;
1053                         break;
1054                 }
1055                 if (no_wait) {
1056                         ret = -EAGAIN;
1057                         break;
1058                 }
1059                 spin_unlock(&ctx->fd_wqh.lock);
1060                 schedule();
1061                 spin_lock(&ctx->fd_wqh.lock);
1062         }
1063         __remove_wait_queue(&ctx->fd_wqh, &wait);
1064         __set_current_state(TASK_RUNNING);
1065         spin_unlock(&ctx->fd_wqh.lock);
1066
1067         if (!ret && msg->event == UFFD_EVENT_FORK) {
1068                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1069
1070                 if (!ret) {
1071                         spin_lock(&ctx->event_wqh.lock);
1072                         if (!list_empty(&fork_event)) {
1073                                 uwq = list_first_entry(&fork_event,
1074                                                        typeof(*uwq),
1075                                                        wq.entry);
1076                                 list_del(&uwq->wq.entry);
1077                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1078                                 userfaultfd_event_complete(ctx, uwq);
1079                         }
1080                         spin_unlock(&ctx->event_wqh.lock);
1081                 }
1082         }
1083
1084         return ret;
1085 }
1086
1087 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1088                                 size_t count, loff_t *ppos)
1089 {
1090         struct userfaultfd_ctx *ctx = file->private_data;
1091         ssize_t _ret, ret = 0;
1092         struct uffd_msg msg;
1093         int no_wait = file->f_flags & O_NONBLOCK;
1094
1095         if (ctx->state == UFFD_STATE_WAIT_API)
1096                 return -EINVAL;
1097
1098         for (;;) {
1099                 if (count < sizeof(msg))
1100                         return ret ? ret : -EINVAL;
1101                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1102                 if (_ret < 0)
1103                         return ret ? ret : _ret;
1104                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1105                         return ret ? ret : -EFAULT;
1106                 ret += sizeof(msg);
1107                 buf += sizeof(msg);
1108                 count -= sizeof(msg);
1109                 /*
1110                  * Allow to read more than one fault at time but only
1111                  * block if waiting for the very first one.
1112                  */
1113                 no_wait = O_NONBLOCK;
1114         }
1115 }
1116
1117 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1118                              struct userfaultfd_wake_range *range)
1119 {
1120         spin_lock(&ctx->fault_pending_wqh.lock);
1121         /* wake all in the range and autoremove */
1122         if (waitqueue_active(&ctx->fault_pending_wqh))
1123                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1124                                      range);
1125         if (waitqueue_active(&ctx->fault_wqh))
1126                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1127         spin_unlock(&ctx->fault_pending_wqh.lock);
1128 }
1129
1130 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1131                                            struct userfaultfd_wake_range *range)
1132 {
1133         unsigned seq;
1134         bool need_wakeup;
1135
1136         /*
1137          * To be sure waitqueue_active() is not reordered by the CPU
1138          * before the pagetable update, use an explicit SMP memory
1139          * barrier here. PT lock release or up_read(mmap_sem) still
1140          * have release semantics that can allow the
1141          * waitqueue_active() to be reordered before the pte update.
1142          */
1143         smp_mb();
1144
1145         /*
1146          * Use waitqueue_active because it's very frequent to
1147          * change the address space atomically even if there are no
1148          * userfaults yet. So we take the spinlock only when we're
1149          * sure we've userfaults to wake.
1150          */
1151         do {
1152                 seq = read_seqcount_begin(&ctx->refile_seq);
1153                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1154                         waitqueue_active(&ctx->fault_wqh);
1155                 cond_resched();
1156         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1157         if (need_wakeup)
1158                 __wake_userfault(ctx, range);
1159 }
1160
1161 static __always_inline int validate_range(struct mm_struct *mm,
1162                                           __u64 start, __u64 len)
1163 {
1164         __u64 task_size = mm->task_size;
1165
1166         if (start & ~PAGE_MASK)
1167                 return -EINVAL;
1168         if (len & ~PAGE_MASK)
1169                 return -EINVAL;
1170         if (!len)
1171                 return -EINVAL;
1172         if (start < mmap_min_addr)
1173                 return -EINVAL;
1174         if (start >= task_size)
1175                 return -EINVAL;
1176         if (len > task_size - start)
1177                 return -EINVAL;
1178         return 0;
1179 }
1180
1181 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1182 {
1183         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1184                 vma_is_shmem(vma);
1185 }
1186
1187 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1188                                 unsigned long arg)
1189 {
1190         struct mm_struct *mm = ctx->mm;
1191         struct vm_area_struct *vma, *prev, *cur;
1192         int ret;
1193         struct uffdio_register uffdio_register;
1194         struct uffdio_register __user *user_uffdio_register;
1195         unsigned long vm_flags, new_flags;
1196         bool found;
1197         bool non_anon_pages;
1198         unsigned long start, end, vma_end;
1199
1200         user_uffdio_register = (struct uffdio_register __user *) arg;
1201
1202         ret = -EFAULT;
1203         if (copy_from_user(&uffdio_register, user_uffdio_register,
1204                            sizeof(uffdio_register)-sizeof(__u64)))
1205                 goto out;
1206
1207         ret = -EINVAL;
1208         if (!uffdio_register.mode)
1209                 goto out;
1210         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1211                                      UFFDIO_REGISTER_MODE_WP))
1212                 goto out;
1213         vm_flags = 0;
1214         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1215                 vm_flags |= VM_UFFD_MISSING;
1216         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1217                 vm_flags |= VM_UFFD_WP;
1218                 /*
1219                  * FIXME: remove the below error constraint by
1220                  * implementing the wprotect tracking mode.
1221                  */
1222                 ret = -EINVAL;
1223                 goto out;
1224         }
1225
1226         ret = validate_range(mm, uffdio_register.range.start,
1227                              uffdio_register.range.len);
1228         if (ret)
1229                 goto out;
1230
1231         start = uffdio_register.range.start;
1232         end = start + uffdio_register.range.len;
1233
1234         ret = -ENOMEM;
1235         if (!mmget_not_zero(mm))
1236                 goto out;
1237
1238         down_write(&mm->mmap_sem);
1239         vma = find_vma_prev(mm, start, &prev);
1240         if (!vma)
1241                 goto out_unlock;
1242
1243         /* check that there's at least one vma in the range */
1244         ret = -EINVAL;
1245         if (vma->vm_start >= end)
1246                 goto out_unlock;
1247
1248         /*
1249          * If the first vma contains huge pages, make sure start address
1250          * is aligned to huge page size.
1251          */
1252         if (is_vm_hugetlb_page(vma)) {
1253                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1254
1255                 if (start & (vma_hpagesize - 1))
1256                         goto out_unlock;
1257         }
1258
1259         /*
1260          * Search for not compatible vmas.
1261          */
1262         found = false;
1263         non_anon_pages = false;
1264         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1265                 cond_resched();
1266
1267                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1268                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1269
1270                 /* check not compatible vmas */
1271                 ret = -EINVAL;
1272                 if (!vma_can_userfault(cur))
1273                         goto out_unlock;
1274                 /*
1275                  * If this vma contains ending address, and huge pages
1276                  * check alignment.
1277                  */
1278                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1279                     end > cur->vm_start) {
1280                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1281
1282                         ret = -EINVAL;
1283
1284                         if (end & (vma_hpagesize - 1))
1285                                 goto out_unlock;
1286                 }
1287
1288                 /*
1289                  * Check that this vma isn't already owned by a
1290                  * different userfaultfd. We can't allow more than one
1291                  * userfaultfd to own a single vma simultaneously or we
1292                  * wouldn't know which one to deliver the userfaults to.
1293                  */
1294                 ret = -EBUSY;
1295                 if (cur->vm_userfaultfd_ctx.ctx &&
1296                     cur->vm_userfaultfd_ctx.ctx != ctx)
1297                         goto out_unlock;
1298
1299                 /*
1300                  * Note vmas containing huge pages
1301                  */
1302                 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1303                         non_anon_pages = true;
1304
1305                 found = true;
1306         }
1307         BUG_ON(!found);
1308
1309         if (vma->vm_start < start)
1310                 prev = vma;
1311
1312         ret = 0;
1313         do {
1314                 cond_resched();
1315
1316                 BUG_ON(!vma_can_userfault(vma));
1317                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1318                        vma->vm_userfaultfd_ctx.ctx != ctx);
1319
1320                 /*
1321                  * Nothing to do: this vma is already registered into this
1322                  * userfaultfd and with the right tracking mode too.
1323                  */
1324                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1325                     (vma->vm_flags & vm_flags) == vm_flags)
1326                         goto skip;
1327
1328                 if (vma->vm_start > start)
1329                         start = vma->vm_start;
1330                 vma_end = min(end, vma->vm_end);
1331
1332                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1333                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1334                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1335                                  vma_policy(vma),
1336                                  ((struct vm_userfaultfd_ctx){ ctx }));
1337                 if (prev) {
1338                         vma = prev;
1339                         goto next;
1340                 }
1341                 if (vma->vm_start < start) {
1342                         ret = split_vma(mm, vma, start, 1);
1343                         if (ret)
1344                                 break;
1345                 }
1346                 if (vma->vm_end > end) {
1347                         ret = split_vma(mm, vma, end, 0);
1348                         if (ret)
1349                                 break;
1350                 }
1351         next:
1352                 /*
1353                  * In the vma_merge() successful mprotect-like case 8:
1354                  * the next vma was merged into the current one and
1355                  * the current one has not been updated yet.
1356                  */
1357                 vma->vm_flags = new_flags;
1358                 vma->vm_userfaultfd_ctx.ctx = ctx;
1359
1360         skip:
1361                 prev = vma;
1362                 start = vma->vm_end;
1363                 vma = vma->vm_next;
1364         } while (vma && vma->vm_start < end);
1365 out_unlock:
1366         up_write(&mm->mmap_sem);
1367         mmput(mm);
1368         if (!ret) {
1369                 /*
1370                  * Now that we scanned all vmas we can already tell
1371                  * userland which ioctls methods are guaranteed to
1372                  * succeed on this range.
1373                  */
1374                 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1375                              UFFD_API_RANGE_IOCTLS,
1376                              &user_uffdio_register->ioctls))
1377                         ret = -EFAULT;
1378         }
1379 out:
1380         return ret;
1381 }
1382
1383 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1384                                   unsigned long arg)
1385 {
1386         struct mm_struct *mm = ctx->mm;
1387         struct vm_area_struct *vma, *prev, *cur;
1388         int ret;
1389         struct uffdio_range uffdio_unregister;
1390         unsigned long new_flags;
1391         bool found;
1392         unsigned long start, end, vma_end;
1393         const void __user *buf = (void __user *)arg;
1394
1395         ret = -EFAULT;
1396         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1397                 goto out;
1398
1399         ret = validate_range(mm, uffdio_unregister.start,
1400                              uffdio_unregister.len);
1401         if (ret)
1402                 goto out;
1403
1404         start = uffdio_unregister.start;
1405         end = start + uffdio_unregister.len;
1406
1407         ret = -ENOMEM;
1408         if (!mmget_not_zero(mm))
1409                 goto out;
1410
1411         down_write(&mm->mmap_sem);
1412         vma = find_vma_prev(mm, start, &prev);
1413         if (!vma)
1414                 goto out_unlock;
1415
1416         /* check that there's at least one vma in the range */
1417         ret = -EINVAL;
1418         if (vma->vm_start >= end)
1419                 goto out_unlock;
1420
1421         /*
1422          * If the first vma contains huge pages, make sure start address
1423          * is aligned to huge page size.
1424          */
1425         if (is_vm_hugetlb_page(vma)) {
1426                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1427
1428                 if (start & (vma_hpagesize - 1))
1429                         goto out_unlock;
1430         }
1431
1432         /*
1433          * Search for not compatible vmas.
1434          */
1435         found = false;
1436         ret = -EINVAL;
1437         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1438                 cond_resched();
1439
1440                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1441                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1442
1443                 /*
1444                  * Check not compatible vmas, not strictly required
1445                  * here as not compatible vmas cannot have an
1446                  * userfaultfd_ctx registered on them, but this
1447                  * provides for more strict behavior to notice
1448                  * unregistration errors.
1449                  */
1450                 if (!vma_can_userfault(cur))
1451                         goto out_unlock;
1452
1453                 found = true;
1454         }
1455         BUG_ON(!found);
1456
1457         if (vma->vm_start < start)
1458                 prev = vma;
1459
1460         ret = 0;
1461         do {
1462                 cond_resched();
1463
1464                 BUG_ON(!vma_can_userfault(vma));
1465
1466                 /*
1467                  * Nothing to do: this vma is already registered into this
1468                  * userfaultfd and with the right tracking mode too.
1469                  */
1470                 if (!vma->vm_userfaultfd_ctx.ctx)
1471                         goto skip;
1472
1473                 if (vma->vm_start > start)
1474                         start = vma->vm_start;
1475                 vma_end = min(end, vma->vm_end);
1476
1477                 if (userfaultfd_missing(vma)) {
1478                         /*
1479                          * Wake any concurrent pending userfault while
1480                          * we unregister, so they will not hang
1481                          * permanently and it avoids userland to call
1482                          * UFFDIO_WAKE explicitly.
1483                          */
1484                         struct userfaultfd_wake_range range;
1485                         range.start = start;
1486                         range.len = vma_end - start;
1487                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1488                 }
1489
1490                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1491                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1492                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1493                                  vma_policy(vma),
1494                                  NULL_VM_UFFD_CTX);
1495                 if (prev) {
1496                         vma = prev;
1497                         goto next;
1498                 }
1499                 if (vma->vm_start < start) {
1500                         ret = split_vma(mm, vma, start, 1);
1501                         if (ret)
1502                                 break;
1503                 }
1504                 if (vma->vm_end > end) {
1505                         ret = split_vma(mm, vma, end, 0);
1506                         if (ret)
1507                                 break;
1508                 }
1509         next:
1510                 /*
1511                  * In the vma_merge() successful mprotect-like case 8:
1512                  * the next vma was merged into the current one and
1513                  * the current one has not been updated yet.
1514                  */
1515                 vma->vm_flags = new_flags;
1516                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1517
1518         skip:
1519                 prev = vma;
1520                 start = vma->vm_end;
1521                 vma = vma->vm_next;
1522         } while (vma && vma->vm_start < end);
1523 out_unlock:
1524         up_write(&mm->mmap_sem);
1525         mmput(mm);
1526 out:
1527         return ret;
1528 }
1529
1530 /*
1531  * userfaultfd_wake may be used in combination with the
1532  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1533  */
1534 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1535                             unsigned long arg)
1536 {
1537         int ret;
1538         struct uffdio_range uffdio_wake;
1539         struct userfaultfd_wake_range range;
1540         const void __user *buf = (void __user *)arg;
1541
1542         ret = -EFAULT;
1543         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1544                 goto out;
1545
1546         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1547         if (ret)
1548                 goto out;
1549
1550         range.start = uffdio_wake.start;
1551         range.len = uffdio_wake.len;
1552
1553         /*
1554          * len == 0 means wake all and we don't want to wake all here,
1555          * so check it again to be sure.
1556          */
1557         VM_BUG_ON(!range.len);
1558
1559         wake_userfault(ctx, &range);
1560         ret = 0;
1561
1562 out:
1563         return ret;
1564 }
1565
1566 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1567                             unsigned long arg)
1568 {
1569         __s64 ret;
1570         struct uffdio_copy uffdio_copy;
1571         struct uffdio_copy __user *user_uffdio_copy;
1572         struct userfaultfd_wake_range range;
1573
1574         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1575
1576         ret = -EFAULT;
1577         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1578                            /* don't copy "copy" last field */
1579                            sizeof(uffdio_copy)-sizeof(__s64)))
1580                 goto out;
1581
1582         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1583         if (ret)
1584                 goto out;
1585         /*
1586          * double check for wraparound just in case. copy_from_user()
1587          * will later check uffdio_copy.src + uffdio_copy.len to fit
1588          * in the userland range.
1589          */
1590         ret = -EINVAL;
1591         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1592                 goto out;
1593         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1594                 goto out;
1595         if (mmget_not_zero(ctx->mm)) {
1596                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1597                                    uffdio_copy.len);
1598                 mmput(ctx->mm);
1599         } else {
1600                 return -ESRCH;
1601         }
1602         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1603                 return -EFAULT;
1604         if (ret < 0)
1605                 goto out;
1606         BUG_ON(!ret);
1607         /* len == 0 would wake all */
1608         range.len = ret;
1609         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1610                 range.start = uffdio_copy.dst;
1611                 wake_userfault(ctx, &range);
1612         }
1613         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1614 out:
1615         return ret;
1616 }
1617
1618 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1619                                 unsigned long arg)
1620 {
1621         __s64 ret;
1622         struct uffdio_zeropage uffdio_zeropage;
1623         struct uffdio_zeropage __user *user_uffdio_zeropage;
1624         struct userfaultfd_wake_range range;
1625
1626         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1627
1628         ret = -EFAULT;
1629         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1630                            /* don't copy "zeropage" last field */
1631                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1632                 goto out;
1633
1634         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1635                              uffdio_zeropage.range.len);
1636         if (ret)
1637                 goto out;
1638         ret = -EINVAL;
1639         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1640                 goto out;
1641
1642         if (mmget_not_zero(ctx->mm)) {
1643                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1644                                      uffdio_zeropage.range.len);
1645                 mmput(ctx->mm);
1646         } else {
1647                 return -ESRCH;
1648         }
1649         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1650                 return -EFAULT;
1651         if (ret < 0)
1652                 goto out;
1653         /* len == 0 would wake all */
1654         BUG_ON(!ret);
1655         range.len = ret;
1656         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1657                 range.start = uffdio_zeropage.range.start;
1658                 wake_userfault(ctx, &range);
1659         }
1660         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1661 out:
1662         return ret;
1663 }
1664
1665 static inline unsigned int uffd_ctx_features(__u64 user_features)
1666 {
1667         /*
1668          * For the current set of features the bits just coincide
1669          */
1670         return (unsigned int)user_features;
1671 }
1672
1673 /*
1674  * userland asks for a certain API version and we return which bits
1675  * and ioctl commands are implemented in this kernel for such API
1676  * version or -EINVAL if unknown.
1677  */
1678 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1679                            unsigned long arg)
1680 {
1681         struct uffdio_api uffdio_api;
1682         void __user *buf = (void __user *)arg;
1683         int ret;
1684         __u64 features;
1685
1686         ret = -EINVAL;
1687         if (ctx->state != UFFD_STATE_WAIT_API)
1688                 goto out;
1689         ret = -EFAULT;
1690         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1691                 goto out;
1692         features = uffdio_api.features;
1693         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1694                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1695                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1696                         goto out;
1697                 ret = -EINVAL;
1698                 goto out;
1699         }
1700         /* report all available features and ioctls to userland */
1701         uffdio_api.features = UFFD_API_FEATURES;
1702         uffdio_api.ioctls = UFFD_API_IOCTLS;
1703         ret = -EFAULT;
1704         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1705                 goto out;
1706         ctx->state = UFFD_STATE_RUNNING;
1707         /* only enable the requested features for this uffd context */
1708         ctx->features = uffd_ctx_features(features);
1709         ret = 0;
1710 out:
1711         return ret;
1712 }
1713
1714 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1715                               unsigned long arg)
1716 {
1717         int ret = -EINVAL;
1718         struct userfaultfd_ctx *ctx = file->private_data;
1719
1720         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1721                 return -EINVAL;
1722
1723         switch(cmd) {
1724         case UFFDIO_API:
1725                 ret = userfaultfd_api(ctx, arg);
1726                 break;
1727         case UFFDIO_REGISTER:
1728                 ret = userfaultfd_register(ctx, arg);
1729                 break;
1730         case UFFDIO_UNREGISTER:
1731                 ret = userfaultfd_unregister(ctx, arg);
1732                 break;
1733         case UFFDIO_WAKE:
1734                 ret = userfaultfd_wake(ctx, arg);
1735                 break;
1736         case UFFDIO_COPY:
1737                 ret = userfaultfd_copy(ctx, arg);
1738                 break;
1739         case UFFDIO_ZEROPAGE:
1740                 ret = userfaultfd_zeropage(ctx, arg);
1741                 break;
1742         }
1743         return ret;
1744 }
1745
1746 #ifdef CONFIG_PROC_FS
1747 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1748 {
1749         struct userfaultfd_ctx *ctx = f->private_data;
1750         wait_queue_entry_t *wq;
1751         struct userfaultfd_wait_queue *uwq;
1752         unsigned long pending = 0, total = 0;
1753
1754         spin_lock(&ctx->fault_pending_wqh.lock);
1755         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1756                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1757                 pending++;
1758                 total++;
1759         }
1760         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1761                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1762                 total++;
1763         }
1764         spin_unlock(&ctx->fault_pending_wqh.lock);
1765
1766         /*
1767          * If more protocols will be added, there will be all shown
1768          * separated by a space. Like this:
1769          *      protocols: aa:... bb:...
1770          */
1771         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1772                    pending, total, UFFD_API, ctx->features,
1773                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1774 }
1775 #endif
1776
1777 static const struct file_operations userfaultfd_fops = {
1778 #ifdef CONFIG_PROC_FS
1779         .show_fdinfo    = userfaultfd_show_fdinfo,
1780 #endif
1781         .release        = userfaultfd_release,
1782         .poll           = userfaultfd_poll,
1783         .read           = userfaultfd_read,
1784         .unlocked_ioctl = userfaultfd_ioctl,
1785         .compat_ioctl   = userfaultfd_ioctl,
1786         .llseek         = noop_llseek,
1787 };
1788
1789 static void init_once_userfaultfd_ctx(void *mem)
1790 {
1791         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1792
1793         init_waitqueue_head(&ctx->fault_pending_wqh);
1794         init_waitqueue_head(&ctx->fault_wqh);
1795         init_waitqueue_head(&ctx->event_wqh);
1796         init_waitqueue_head(&ctx->fd_wqh);
1797         seqcount_init(&ctx->refile_seq);
1798 }
1799
1800 /**
1801  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1802  * @flags: Flags for the userfaultfd file.
1803  *
1804  * This function creates a userfaultfd file pointer, w/out installing
1805  * it into the fd table. This is useful when the userfaultfd file is
1806  * used during the initialization of data structures that require
1807  * extra setup after the userfaultfd creation. So the userfaultfd
1808  * creation is split into the file pointer creation phase, and the
1809  * file descriptor installation phase.  In this way races with
1810  * userspace closing the newly installed file descriptor can be
1811  * avoided.  Returns a userfaultfd file pointer, or a proper error
1812  * pointer.
1813  */
1814 static struct file *userfaultfd_file_create(int flags)
1815 {
1816         struct file *file;
1817         struct userfaultfd_ctx *ctx;
1818
1819         BUG_ON(!current->mm);
1820
1821         /* Check the UFFD_* constants for consistency.  */
1822         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1823         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1824
1825         file = ERR_PTR(-EINVAL);
1826         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1827                 goto out;
1828
1829         file = ERR_PTR(-ENOMEM);
1830         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1831         if (!ctx)
1832                 goto out;
1833
1834         atomic_set(&ctx->refcount, 1);
1835         ctx->flags = flags;
1836         ctx->features = 0;
1837         ctx->state = UFFD_STATE_WAIT_API;
1838         ctx->released = false;
1839         ctx->mm = current->mm;
1840         /* prevent the mm struct to be freed */
1841         mmgrab(ctx->mm);
1842
1843         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1844                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1845         if (IS_ERR(file)) {
1846                 mmdrop(ctx->mm);
1847                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1848         }
1849 out:
1850         return file;
1851 }
1852
1853 SYSCALL_DEFINE1(userfaultfd, int, flags)
1854 {
1855         int fd, error;
1856         struct file *file;
1857
1858         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1859         if (error < 0)
1860                 return error;
1861         fd = error;
1862
1863         file = userfaultfd_file_create(flags);
1864         if (IS_ERR(file)) {
1865                 error = PTR_ERR(file);
1866                 goto err_put_unused_fd;
1867         }
1868         fd_install(fd, file);
1869
1870         return fd;
1871
1872 err_put_unused_fd:
1873         put_unused_fd(fd);
1874
1875         return error;
1876 }
1877
1878 static int __init userfaultfd_init(void)
1879 {
1880         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1881                                                 sizeof(struct userfaultfd_ctx),
1882                                                 0,
1883                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1884                                                 init_once_userfaultfd_ctx);
1885         return 0;
1886 }
1887 __initcall(userfaultfd_init);