fs/userfaultfd.c

   1 /*
   2  *  fs/userfaultfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc.
   6  *  Copyright (C) 2015  Red Hat, Inc.
   7  *
   8  *  This work is licensed under the terms of the GNU GPL, version 2. See
   9  *  the COPYING file in the top-level directory.
  10  *
  11  *  Some part derived from fs/eventfd.c (anon inode setup) and
  12  *  mm/ksm.c (mm hashing).
  13  */
  14
  15 #include <linux/list.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/mm.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  34
  35 enum userfaultfd_state {
  36         UFFD_STATE_WAIT_API,
  37         UFFD_STATE_RUNNING,
  38 };
  39
  40 /*
  41  * Start with fault_pending_wqh and fault_wqh so they're more likely
  42  * to be in the same cacheline.
  43  */
  44 struct userfaultfd_ctx {
  45         /* waitqueue head for the pending (i.e. not read) userfaults */
  46         wait_queue_head_t fault_pending_wqh;
  47         /* waitqueue head for the userfaults */
  48         wait_queue_head_t fault_wqh;
  49         /* waitqueue head for the pseudo fd to wakeup poll/read */
  50         wait_queue_head_t fd_wqh;
  51         /* waitqueue head for events */
  52         wait_queue_head_t event_wqh;
  53         /* a refile sequence protected by fault_pending_wqh lock */
  54         struct seqcount refile_seq;
  55         /* pseudo fd refcounting */
  56         atomic_t refcount;
  57         /* userfaultfd syscall flags */
  58         unsigned int flags;
  59         /* features requested from the userspace */
  60         unsigned int features;
  61         /* state machine */
  62         enum userfaultfd_state state;
  63         /* released */
  64         bool released;
  65         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  66         struct mm_struct *mm;
  67 };
  68
  69 struct userfaultfd_fork_ctx {
  70         struct userfaultfd_ctx *orig;
  71         struct userfaultfd_ctx *new;
  72         struct list_head list;
  73 };
  74
  75 struct userfaultfd_unmap_ctx {
  76         struct userfaultfd_ctx *ctx;
  77         unsigned long start;
  78         unsigned long end;
  79         struct list_head list;
  80 };
  81
  82 struct userfaultfd_wait_queue {
  83         struct uffd_msg msg;
  84         wait_queue_t wq;
  85         struct userfaultfd_ctx *ctx;
  86         bool waken;
  87 };
  88
  89 struct userfaultfd_wake_range {
  90         unsigned long start;
  91         unsigned long len;
  92 };
  93
  94 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
  95                                      int wake_flags, void *key)
  96 {
  97         struct userfaultfd_wake_range *range = key;
  98         int ret;
  99         struct userfaultfd_wait_queue *uwq;
 100         unsigned long start, len;
 101
 102         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 103         ret = 0;
 104         /* len == 0 means wake all */
 105         start = range->start;
 106         len = range->len;
 107         if (len && (start > uwq->msg.arg.pagefault.address ||
 108                     start + len <= uwq->msg.arg.pagefault.address))
 109                 goto out;
 110         WRITE_ONCE(uwq->waken, true);
 111         /*
 112          * The implicit smp_mb__before_spinlock in try_to_wake_up()
 113          * renders uwq->waken visible to other CPUs before the task is
 114          * waken.
 115          */
 116         ret = wake_up_state(wq->private, mode);
 117         if (ret)
 118                 /*
 119                  * Wake only once, autoremove behavior.
 120                  *
 121                  * After the effect of list_del_init is visible to the
 122                  * other CPUs, the waitqueue may disappear from under
 123                  * us, see the !list_empty_careful() in
 124                  * handle_userfault(). try_to_wake_up() has an
 125                  * implicit smp_mb__before_spinlock, and the
 126                  * wq->private is read before calling the extern
 127                  * function "wake_up_state" (which in turns calls
 128                  * try_to_wake_up). While the spin_lock;spin_unlock;
 129                  * wouldn't be enough, the smp_mb__before_spinlock is
 130                  * enough to avoid an explicit smp_mb() here.
 131                  */
 132                 list_del_init(&wq->task_list);
 133 out:
 134         return ret;
 135 }
 136
 137 /**
 138  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 139  * context.
 140  * @ctx: [in] Pointer to the userfaultfd context.
 141  *
 142  * Returns: In case of success, returns not zero.
 143  */
 144 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 145 {
 146         if (!atomic_inc_not_zero(&ctx->refcount))
 147                 BUG();
 148 }
 149
 150 /**
 151  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 152  * context.
 153  * @ctx: [in] Pointer to userfaultfd context.
 154  *
 155  * The userfaultfd context reference must have been previously acquired either
 156  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 157  */
 158 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 159 {
 160         if (atomic_dec_and_test(&ctx->refcount)) {
 161                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 162                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 163                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 164                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 165                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 166                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 167                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 168                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 169                 mmdrop(ctx->mm);
 170                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 171         }
 172 }
 173
 174 static inline void msg_init(struct uffd_msg *msg)
 175 {
 176         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 177         /*
 178          * Must use memset to zero out the paddings or kernel data is
 179          * leaked to userland.
 180          */
 181         memset(msg, 0, sizeof(struct uffd_msg));
 182 }
 183
 184 static inline struct uffd_msg userfault_msg(unsigned long address,
 185                                             unsigned int flags,
 186                                             unsigned long reason)
 187 {
 188         struct uffd_msg msg;
 189         msg_init(&msg);
 190         msg.event = UFFD_EVENT_PAGEFAULT;
 191         msg.arg.pagefault.address = address;
 192         if (flags & FAULT_FLAG_WRITE)
 193                 /*
 194                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 195                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 196                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 197                  * was a read fault, otherwise if set it means it's
 198                  * a write fault.
 199                  */
 200                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 201         if (reason & VM_UFFD_WP)
 202                 /*
 203                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 204                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 205                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 206                  * a missing fault, otherwise if set it means it's a
 207                  * write protect fault.
 208                  */
 209                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 210         return msg;
 211 }
 212
 213 #ifdef CONFIG_HUGETLB_PAGE
 214 /*
 215  * Same functionality as userfaultfd_must_wait below with modifications for
 216  * hugepmd ranges.
 217  */
 218 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 219                                          unsigned long address,
 220                                          unsigned long flags,
 221                                          unsigned long reason)
 222 {
 223         struct mm_struct *mm = ctx->mm;
 224         pte_t *pte;
 225         bool ret = true;
 226
 227         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 228
 229         pte = huge_pte_offset(mm, address);
 230         if (!pte)
 231                 goto out;
 232
 233         ret = false;
 234
 235         /*
 236          * Lockless access: we're in a wait_event so it's ok if it
 237          * changes under us.
 238          */
 239         if (huge_pte_none(*pte))
 240                 ret = true;
 241         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
 242                 ret = true;
 243 out:
 244         return ret;
 245 }
 246 #else
 247 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 248                                          unsigned long address,
 249                                          unsigned long flags,
 250                                          unsigned long reason)
 251 {
 252         return false;   /* should never get here */
 253 }
 254 #endif /* CONFIG_HUGETLB_PAGE */
 255
 256 /*
 257  * Verify the pagetables are still not ok after having reigstered into
 258  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 259  * userfault that has already been resolved, if userfaultfd_read and
 260  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 261  * threads.
 262  */
 263 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 264                                          unsigned long address,
 265                                          unsigned long flags,
 266                                          unsigned long reason)
 267 {
 268         struct mm_struct *mm = ctx->mm;
 269         pgd_t *pgd;
 270         pud_t *pud;
 271         pmd_t *pmd, _pmd;
 272         pte_t *pte;
 273         bool ret = true;
 274
 275         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 276
 277         pgd = pgd_offset(mm, address);
 278         if (!pgd_present(*pgd))
 279                 goto out;
 280         pud = pud_offset(pgd, address);
 281         if (!pud_present(*pud))
 282                 goto out;
 283         pmd = pmd_offset(pud, address);
 284         /*
 285          * READ_ONCE must function as a barrier with narrower scope
 286          * and it must be equivalent to:
 287          *      _pmd = *pmd; barrier();
 288          *
 289          * This is to deal with the instability (as in
 290          * pmd_trans_unstable) of the pmd.
 291          */
 292         _pmd = READ_ONCE(*pmd);
 293         if (!pmd_present(_pmd))
 294                 goto out;
 295
 296         ret = false;
 297         if (pmd_trans_huge(_pmd))
 298                 goto out;
 299
 300         /*
 301          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 302          * and use the standard pte_offset_map() instead of parsing _pmd.
 303          */
 304         pte = pte_offset_map(pmd, address);
 305         /*
 306          * Lockless access: we're in a wait_event so it's ok if it
 307          * changes under us.
 308          */
 309         if (pte_none(*pte))
 310                 ret = true;
 311         pte_unmap(pte);
 312
 313 out:
 314         return ret;
 315 }
 316
 317 /*
 318  * The locking rules involved in returning VM_FAULT_RETRY depending on
 319  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 320  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 321  * recommendation in __lock_page_or_retry is not an understatement.
 322  *
 323  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 324  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 325  * not set.
 326  *
 327  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 328  * set, VM_FAULT_RETRY can still be returned if and only if there are
 329  * fatal_signal_pending()s, and the mmap_sem must be released before
 330  * returning it.
 331  */
 332 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 333 {
 334         struct mm_struct *mm = vmf->vma->vm_mm;
 335         struct userfaultfd_ctx *ctx;
 336         struct userfaultfd_wait_queue uwq;
 337         int ret;
 338         bool must_wait, return_to_userland;
 339         long blocking_state;
 340
 341         BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 342
 343         ret = VM_FAULT_SIGBUS;
 344         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 345         if (!ctx)
 346                 goto out;
 347
 348         BUG_ON(ctx->mm != mm);
 349
 350         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 351         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 352
 353         /*
 354          * If it's already released don't get it. This avoids to loop
 355          * in __get_user_pages if userfaultfd_release waits on the
 356          * caller of handle_userfault to release the mmap_sem.
 357          */
 358         if (unlikely(ACCESS_ONCE(ctx->released)))
 359                 goto out;
 360
 361         /*
 362          * We don't do userfault handling for the final child pid update.
 363          */
 364         if (current->flags & PF_EXITING)
 365                 goto out;
 366
 367         /*
 368          * Check that we can return VM_FAULT_RETRY.
 369          *
 370          * NOTE: it should become possible to return VM_FAULT_RETRY
 371          * even if FAULT_FLAG_TRIED is set without leading to gup()
 372          * -EBUSY failures, if the userfaultfd is to be extended for
 373          * VM_UFFD_WP tracking and we intend to arm the userfault
 374          * without first stopping userland access to the memory. For
 375          * VM_UFFD_MISSING userfaults this is enough for now.
 376          */
 377         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 378                 /*
 379                  * Validate the invariant that nowait must allow retry
 380                  * to be sure not to return SIGBUS erroneously on
 381                  * nowait invocations.
 382                  */
 383                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 384 #ifdef CONFIG_DEBUG_VM
 385                 if (printk_ratelimit()) {
 386                         printk(KERN_WARNING
 387                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 388                                vmf->flags);
 389                         dump_stack();
 390                 }
 391 #endif
 392                 goto out;
 393         }
 394
 395         /*
 396          * Handle nowait, not much to do other than tell it to retry
 397          * and wait.
 398          */
 399         ret = VM_FAULT_RETRY;
 400         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 401                 goto out;
 402
 403         /* take the reference before dropping the mmap_sem */
 404         userfaultfd_ctx_get(ctx);
 405
 406         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 407         uwq.wq.private = current;
 408         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
 409         uwq.ctx = ctx;
 410         uwq.waken = false;
 411
 412         return_to_userland =
 413                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 414                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 415         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 416                          TASK_KILLABLE;
 417
 418         spin_lock(&ctx->fault_pending_wqh.lock);
 419         /*
 420          * After the __add_wait_queue the uwq is visible to userland
 421          * through poll/read().
 422          */
 423         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 424         /*
 425          * The smp_mb() after __set_current_state prevents the reads
 426          * following the spin_unlock to happen before the list_add in
 427          * __add_wait_queue.
 428          */
 429         set_current_state(blocking_state);
 430         spin_unlock(&ctx->fault_pending_wqh.lock);
 431
 432         if (!is_vm_hugetlb_page(vmf->vma))
 433                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 434                                                   reason);
 435         else
 436                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->address,
 437                                                        vmf->flags, reason);
 438         up_read(&mm->mmap_sem);
 439
 440         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
 441                    (return_to_userland ? !signal_pending(current) :
 442                     !fatal_signal_pending(current)))) {
 443                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 444                 schedule();
 445                 ret |= VM_FAULT_MAJOR;
 446
 447                 /*
 448                  * False wakeups can orginate even from rwsem before
 449                  * up_read() however userfaults will wait either for a
 450                  * targeted wakeup on the specific uwq waitqueue from
 451                  * wake_userfault() or for signals or for uffd
 452                  * release.
 453                  */
 454                 while (!READ_ONCE(uwq.waken)) {
 455                         /*
 456                          * This needs the full smp_store_mb()
 457                          * guarantee as the state write must be
 458                          * visible to other CPUs before reading
 459                          * uwq.waken from other CPUs.
 460                          */
 461                         set_current_state(blocking_state);
 462                         if (READ_ONCE(uwq.waken) ||
 463                             READ_ONCE(ctx->released) ||
 464                             (return_to_userland ? signal_pending(current) :
 465                              fatal_signal_pending(current)))
 466                                 break;
 467                         schedule();
 468                 }
 469         }
 470
 471         __set_current_state(TASK_RUNNING);
 472
 473         if (return_to_userland) {
 474                 if (signal_pending(current) &&
 475                     !fatal_signal_pending(current)) {
 476                         /*
 477                          * If we got a SIGSTOP or SIGCONT and this is
 478                          * a normal userland page fault, just let
 479                          * userland return so the signal will be
 480                          * handled and gdb debugging works.  The page
 481                          * fault code immediately after we return from
 482                          * this function is going to release the
 483                          * mmap_sem and it's not depending on it
 484                          * (unlike gup would if we were not to return
 485                          * VM_FAULT_RETRY).
 486                          *
 487                          * If a fatal signal is pending we still take
 488                          * the streamlined VM_FAULT_RETRY failure path
 489                          * and there's no need to retake the mmap_sem
 490                          * in such case.
 491                          */
 492                         down_read(&mm->mmap_sem);
 493                         ret = VM_FAULT_NOPAGE;
 494                 }
 495         }
 496
 497         /*
 498          * Here we race with the list_del; list_add in
 499          * userfaultfd_ctx_read(), however because we don't ever run
 500          * list_del_init() to refile across the two lists, the prev
 501          * and next pointers will never point to self. list_add also
 502          * would never let any of the two pointers to point to
 503          * self. So list_empty_careful won't risk to see both pointers
 504          * pointing to self at any time during the list refile. The
 505          * only case where list_del_init() is called is the full
 506          * removal in the wake function and there we don't re-list_add
 507          * and it's fine not to block on the spinlock. The uwq on this
 508          * kernel stack can be released after the list_del_init.
 509          */
 510         if (!list_empty_careful(&uwq.wq.task_list)) {
 511                 spin_lock(&ctx->fault_pending_wqh.lock);
 512                 /*
 513                  * No need of list_del_init(), the uwq on the stack
 514                  * will be freed shortly anyway.
 515                  */
 516                 list_del(&uwq.wq.task_list);
 517                 spin_unlock(&ctx->fault_pending_wqh.lock);
 518         }
 519
 520         /*
 521          * ctx may go away after this if the userfault pseudo fd is
 522          * already released.
 523          */
 524         userfaultfd_ctx_put(ctx);
 525
 526 out:
 527         return ret;
 528 }
 529
 530 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 531                                               struct userfaultfd_wait_queue *ewq)
 532 {
 533         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 534                 goto out;
 535
 536         ewq->ctx = ctx;
 537         init_waitqueue_entry(&ewq->wq, current);
 538
 539         spin_lock(&ctx->event_wqh.lock);
 540         /*
 541          * After the __add_wait_queue the uwq is visible to userland
 542          * through poll/read().
 543          */
 544         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 545         for (;;) {
 546                 set_current_state(TASK_KILLABLE);
 547                 if (ewq->msg.event == 0)
 548                         break;
 549                 if (ACCESS_ONCE(ctx->released) ||
 550                     fatal_signal_pending(current)) {
 551                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 552                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 553                                 struct userfaultfd_ctx *new;
 554
 555                                 new = (struct userfaultfd_ctx *)
 556                                         (unsigned long)
 557                                         ewq->msg.arg.reserved.reserved1;
 558
 559                                 userfaultfd_ctx_put(new);
 560                         }
 561                         break;
 562                 }
 563
 564                 spin_unlock(&ctx->event_wqh.lock);
 565
 566                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 567                 schedule();
 568
 569                 spin_lock(&ctx->event_wqh.lock);
 570         }
 571         __set_current_state(TASK_RUNNING);
 572         spin_unlock(&ctx->event_wqh.lock);
 573
 574         /*
 575          * ctx may go away after this if the userfault pseudo fd is
 576          * already released.
 577          */
 578 out:
 579         userfaultfd_ctx_put(ctx);
 580 }
 581
 582 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 583                                        struct userfaultfd_wait_queue *ewq)
 584 {
 585         ewq->msg.event = 0;
 586         wake_up_locked(&ctx->event_wqh);
 587         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 588 }
 589
 590 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 591 {
 592         struct userfaultfd_ctx *ctx = NULL, *octx;
 593         struct userfaultfd_fork_ctx *fctx;
 594
 595         octx = vma->vm_userfaultfd_ctx.ctx;
 596         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 597                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 598                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 599                 return 0;
 600         }
 601
 602         list_for_each_entry(fctx, fcs, list)
 603                 if (fctx->orig == octx) {
 604                         ctx = fctx->new;
 605                         break;
 606                 }
 607
 608         if (!ctx) {
 609                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 610                 if (!fctx)
 611                         return -ENOMEM;
 612
 613                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 614                 if (!ctx) {
 615                         kfree(fctx);
 616                         return -ENOMEM;
 617                 }
 618
 619                 atomic_set(&ctx->refcount, 1);
 620                 ctx->flags = octx->flags;
 621                 ctx->state = UFFD_STATE_RUNNING;
 622                 ctx->features = octx->features;
 623                 ctx->released = false;
 624                 ctx->mm = vma->vm_mm;
 625                 atomic_inc(&ctx->mm->mm_count);
 626
 627                 userfaultfd_ctx_get(octx);
 628                 fctx->orig = octx;
 629                 fctx->new = ctx;
 630                 list_add_tail(&fctx->list, fcs);
 631         }
 632
 633         vma->vm_userfaultfd_ctx.ctx = ctx;
 634         return 0;
 635 }
 636
 637 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 638 {
 639         struct userfaultfd_ctx *ctx = fctx->orig;
 640         struct userfaultfd_wait_queue ewq;
 641
 642         msg_init(&ewq.msg);
 643
 644         ewq.msg.event = UFFD_EVENT_FORK;
 645         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 646
 647         userfaultfd_event_wait_completion(ctx, &ewq);
 648 }
 649
 650 void dup_userfaultfd_complete(struct list_head *fcs)
 651 {
 652         struct userfaultfd_fork_ctx *fctx, *n;
 653
 654         list_for_each_entry_safe(fctx, n, fcs, list) {
 655                 dup_fctx(fctx);
 656                 list_del(&fctx->list);
 657                 kfree(fctx);
 658         }
 659 }
 660
 661 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 662                              struct vm_userfaultfd_ctx *vm_ctx)
 663 {
 664         struct userfaultfd_ctx *ctx;
 665
 666         ctx = vma->vm_userfaultfd_ctx.ctx;
 667         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 668                 vm_ctx->ctx = ctx;
 669                 userfaultfd_ctx_get(ctx);
 670         }
 671 }
 672
 673 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 674                                  unsigned long from, unsigned long to,
 675                                  unsigned long len)
 676 {
 677         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 678         struct userfaultfd_wait_queue ewq;
 679
 680         if (!ctx)
 681                 return;
 682
 683         if (to & ~PAGE_MASK) {
 684                 userfaultfd_ctx_put(ctx);
 685                 return;
 686         }
 687
 688         msg_init(&ewq.msg);
 689
 690         ewq.msg.event = UFFD_EVENT_REMAP;
 691         ewq.msg.arg.remap.from = from;
 692         ewq.msg.arg.remap.to = to;
 693         ewq.msg.arg.remap.len = len;
 694
 695         userfaultfd_event_wait_completion(ctx, &ewq);
 696 }
 697
 698 bool userfaultfd_remove(struct vm_area_struct *vma,
 699                         unsigned long start, unsigned long end)
 700 {
 701         struct mm_struct *mm = vma->vm_mm;
 702         struct userfaultfd_ctx *ctx;
 703         struct userfaultfd_wait_queue ewq;
 704
 705         ctx = vma->vm_userfaultfd_ctx.ctx;
 706         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 707                 return true;
 708
 709         userfaultfd_ctx_get(ctx);
 710         up_read(&mm->mmap_sem);
 711
 712         msg_init(&ewq.msg);
 713
 714         ewq.msg.event = UFFD_EVENT_REMOVE;
 715         ewq.msg.arg.remove.start = start;
 716         ewq.msg.arg.remove.end = end;
 717
 718         userfaultfd_event_wait_completion(ctx, &ewq);
 719
 720         return false;
 721 }
 722
 723 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 724                           unsigned long start, unsigned long end)
 725 {
 726         struct userfaultfd_unmap_ctx *unmap_ctx;
 727
 728         list_for_each_entry(unmap_ctx, unmaps, list)
 729                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 730                     unmap_ctx->end == end)
 731                         return true;
 732
 733         return false;
 734 }
 735
 736 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 737                            unsigned long start, unsigned long end,
 738                            struct list_head *unmaps)
 739 {
 740         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 741                 struct userfaultfd_unmap_ctx *unmap_ctx;
 742                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 743
 744                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 745                     has_unmap_ctx(ctx, unmaps, start, end))
 746                         continue;
 747
 748                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 749                 if (!unmap_ctx)
 750                         return -ENOMEM;
 751
 752                 userfaultfd_ctx_get(ctx);
 753                 unmap_ctx->ctx = ctx;
 754                 unmap_ctx->start = start;
 755                 unmap_ctx->end = end;
 756                 list_add_tail(&unmap_ctx->list, unmaps);
 757         }
 758
 759         return 0;
 760 }
 761
 762 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 763 {
 764         struct userfaultfd_unmap_ctx *ctx, *n;
 765         struct userfaultfd_wait_queue ewq;
 766
 767         list_for_each_entry_safe(ctx, n, uf, list) {
 768                 msg_init(&ewq.msg);
 769
 770                 ewq.msg.event = UFFD_EVENT_UNMAP;
 771                 ewq.msg.arg.remove.start = ctx->start;
 772                 ewq.msg.arg.remove.end = ctx->end;
 773
 774                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 775
 776                 list_del(&ctx->list);
 777                 kfree(ctx);
 778         }
 779 }
 780
 781 static int userfaultfd_release(struct inode *inode, struct file *file)
 782 {
 783         struct userfaultfd_ctx *ctx = file->private_data;
 784         struct mm_struct *mm = ctx->mm;
 785         struct vm_area_struct *vma, *prev;
 786         /* len == 0 means wake all */
 787         struct userfaultfd_wake_range range = { .len = 0, };
 788         unsigned long new_flags;
 789
 790         ACCESS_ONCE(ctx->released) = true;
 791
 792         if (!mmget_not_zero(mm))
 793                 goto wakeup;
 794
 795         /*
 796          * Flush page faults out of all CPUs. NOTE: all page faults
 797          * must be retried without returning VM_FAULT_SIGBUS if
 798          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 799          * changes while handle_userfault released the mmap_sem. So
 800          * it's critical that released is set to true (above), before
 801          * taking the mmap_sem for writing.
 802          */
 803         down_write(&mm->mmap_sem);
 804         prev = NULL;
 805         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 806                 cond_resched();
 807                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 808                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 809                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 810                         prev = vma;
 811                         continue;
 812                 }
 813                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 814                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 815                                  new_flags, vma->anon_vma,
 816                                  vma->vm_file, vma->vm_pgoff,
 817                                  vma_policy(vma),
 818                                  NULL_VM_UFFD_CTX);
 819                 if (prev)
 820                         vma = prev;
 821                 else
 822                         prev = vma;
 823                 vma->vm_flags = new_flags;
 824                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 825         }
 826         up_write(&mm->mmap_sem);
 827         mmput(mm);
 828 wakeup:
 829         /*
 830          * After no new page faults can wait on this fault_*wqh, flush
 831          * the last page faults that may have been already waiting on
 832          * the fault_*wqh.
 833          */
 834         spin_lock(&ctx->fault_pending_wqh.lock);
 835         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 836         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 837         spin_unlock(&ctx->fault_pending_wqh.lock);
 838
 839         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 840         userfaultfd_ctx_put(ctx);
 841         return 0;
 842 }
 843
 844 /* fault_pending_wqh.lock must be hold by the caller */
 845 static inline struct userfaultfd_wait_queue *find_userfault_in(
 846                 wait_queue_head_t *wqh)
 847 {
 848         wait_queue_t *wq;
 849         struct userfaultfd_wait_queue *uwq;
 850
 851         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 852
 853         uwq = NULL;
 854         if (!waitqueue_active(wqh))
 855                 goto out;
 856         /* walk in reverse to provide FIFO behavior to read userfaults */
 857         wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
 858         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 859 out:
 860         return uwq;
 861 }
 862
 863 static inline struct userfaultfd_wait_queue *find_userfault(
 864                 struct userfaultfd_ctx *ctx)
 865 {
 866         return find_userfault_in(&ctx->fault_pending_wqh);
 867 }
 868
 869 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 870                 struct userfaultfd_ctx *ctx)
 871 {
 872         return find_userfault_in(&ctx->event_wqh);
 873 }
 874
 875 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 876 {
 877         struct userfaultfd_ctx *ctx = file->private_data;
 878         unsigned int ret;
 879
 880         poll_wait(file, &ctx->fd_wqh, wait);
 881
 882         switch (ctx->state) {
 883         case UFFD_STATE_WAIT_API:
 884                 return POLLERR;
 885         case UFFD_STATE_RUNNING:
 886                 /*
 887                  * poll() never guarantees that read won't block.
 888                  * userfaults can be waken before they're read().
 889                  */
 890                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 891                         return POLLERR;
 892                 /*
 893                  * lockless access to see if there are pending faults
 894                  * __pollwait last action is the add_wait_queue but
 895                  * the spin_unlock would allow the waitqueue_active to
 896                  * pass above the actual list_add inside
 897                  * add_wait_queue critical section. So use a full
 898                  * memory barrier to serialize the list_add write of
 899                  * add_wait_queue() with the waitqueue_active read
 900                  * below.
 901                  */
 902                 ret = 0;
 903                 smp_mb();
 904                 if (waitqueue_active(&ctx->fault_pending_wqh))
 905                         ret = POLLIN;
 906                 else if (waitqueue_active(&ctx->event_wqh))
 907                         ret = POLLIN;
 908
 909                 return ret;
 910         default:
 911                 WARN_ON_ONCE(1);
 912                 return POLLERR;
 913         }
 914 }
 915
 916 static const struct file_operations userfaultfd_fops;
 917
 918 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 919                                   struct userfaultfd_ctx *new,
 920                                   struct uffd_msg *msg)
 921 {
 922         int fd;
 923         struct file *file;
 924         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 925
 926         fd = get_unused_fd_flags(flags);
 927         if (fd < 0)
 928                 return fd;
 929
 930         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 931                                   O_RDWR | flags);
 932         if (IS_ERR(file)) {
 933                 put_unused_fd(fd);
 934                 return PTR_ERR(file);
 935         }
 936
 937         fd_install(fd, file);
 938         msg->arg.reserved.reserved1 = 0;
 939         msg->arg.fork.ufd = fd;
 940
 941         return 0;
 942 }
 943
 944 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 945                                     struct uffd_msg *msg)
 946 {
 947         ssize_t ret;
 948         DECLARE_WAITQUEUE(wait, current);
 949         struct userfaultfd_wait_queue *uwq;
 950         /*
 951          * Handling fork event requires sleeping operations, so
 952          * we drop the event_wqh lock, then do these ops, then
 953          * lock it back and wake up the waiter. While the lock is
 954          * dropped the ewq may go away so we keep track of it
 955          * carefully.
 956          */
 957         LIST_HEAD(fork_event);
 958         struct userfaultfd_ctx *fork_nctx = NULL;
 959
 960         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 961         spin_lock(&ctx->fd_wqh.lock);
 962         __add_wait_queue(&ctx->fd_wqh, &wait);
 963         for (;;) {
 964                 set_current_state(TASK_INTERRUPTIBLE);
 965                 spin_lock(&ctx->fault_pending_wqh.lock);
 966                 uwq = find_userfault(ctx);
 967                 if (uwq) {
 968                         /*
 969                          * Use a seqcount to repeat the lockless check
 970                          * in wake_userfault() to avoid missing
 971                          * wakeups because during the refile both
 972                          * waitqueue could become empty if this is the
 973                          * only userfault.
 974                          */
 975                         write_seqcount_begin(&ctx->refile_seq);
 976
 977                         /*
 978                          * The fault_pending_wqh.lock prevents the uwq
 979                          * to disappear from under us.
 980                          *
 981                          * Refile this userfault from
 982                          * fault_pending_wqh to fault_wqh, it's not
 983                          * pending anymore after we read it.
 984                          *
 985                          * Use list_del() by hand (as
 986                          * userfaultfd_wake_function also uses
 987                          * list_del_init() by hand) to be sure nobody
 988                          * changes __remove_wait_queue() to use
 989                          * list_del_init() in turn breaking the
 990                          * !list_empty_careful() check in
 991                          * handle_userfault(). The uwq->wq.task_list
 992                          * must never be empty at any time during the
 993                          * refile, or the waitqueue could disappear
 994                          * from under us. The "wait_queue_head_t"
 995                          * parameter of __remove_wait_queue() is unused
 996                          * anyway.
 997                          */
 998                         list_del(&uwq->wq.task_list);
 999                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1000
1001                         write_seqcount_end(&ctx->refile_seq);
1002
1003                         /* careful to always initialize msg if ret == 0 */
1004                         *msg = uwq->msg;
1005                         spin_unlock(&ctx->fault_pending_wqh.lock);
1006                         ret = 0;
1007                         break;
1008                 }
1009                 spin_unlock(&ctx->fault_pending_wqh.lock);
1010
1011                 spin_lock(&ctx->event_wqh.lock);
1012                 uwq = find_userfault_evt(ctx);
1013                 if (uwq) {
1014                         *msg = uwq->msg;
1015
1016                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1017                                 fork_nctx = (struct userfaultfd_ctx *)
1018                                         (unsigned long)
1019                                         uwq->msg.arg.reserved.reserved1;
1020                                 list_move(&uwq->wq.task_list, &fork_event);
1021                                 spin_unlock(&ctx->event_wqh.lock);
1022                                 ret = 0;
1023                                 break;
1024                         }
1025
1026                         userfaultfd_event_complete(ctx, uwq);
1027                         spin_unlock(&ctx->event_wqh.lock);
1028                         ret = 0;
1029                         break;
1030                 }
1031                 spin_unlock(&ctx->event_wqh.lock);
1032
1033                 if (signal_pending(current)) {
1034                         ret = -ERESTARTSYS;
1035                         break;
1036                 }
1037                 if (no_wait) {
1038                         ret = -EAGAIN;
1039                         break;
1040                 }
1041                 spin_unlock(&ctx->fd_wqh.lock);
1042                 schedule();
1043                 spin_lock(&ctx->fd_wqh.lock);
1044         }
1045         __remove_wait_queue(&ctx->fd_wqh, &wait);
1046         __set_current_state(TASK_RUNNING);
1047         spin_unlock(&ctx->fd_wqh.lock);
1048
1049         if (!ret && msg->event == UFFD_EVENT_FORK) {
1050                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1051
1052                 if (!ret) {
1053                         spin_lock(&ctx->event_wqh.lock);
1054                         if (!list_empty(&fork_event)) {
1055                                 uwq = list_first_entry(&fork_event,
1056                                                        typeof(*uwq),
1057                                                        wq.task_list);
1058                                 list_del(&uwq->wq.task_list);
1059                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1060                                 userfaultfd_event_complete(ctx, uwq);
1061                         }
1062                         spin_unlock(&ctx->event_wqh.lock);
1063                 }
1064         }
1065
1066         return ret;
1067 }
1068
1069 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1070                                 size_t count, loff_t *ppos)
1071 {
1072         struct userfaultfd_ctx *ctx = file->private_data;
1073         ssize_t _ret, ret = 0;
1074         struct uffd_msg msg;
1075         int no_wait = file->f_flags & O_NONBLOCK;
1076
1077         if (ctx->state == UFFD_STATE_WAIT_API)
1078                 return -EINVAL;
1079
1080         for (;;) {
1081                 if (count < sizeof(msg))
1082                         return ret ? ret : -EINVAL;
1083                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1084                 if (_ret < 0)
1085                         return ret ? ret : _ret;
1086                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1087                         return ret ? ret : -EFAULT;
1088                 ret += sizeof(msg);
1089                 buf += sizeof(msg);
1090                 count -= sizeof(msg);
1091                 /*
1092                  * Allow to read more than one fault at time but only
1093                  * block if waiting for the very first one.
1094                  */
1095                 no_wait = O_NONBLOCK;
1096         }
1097 }
1098
1099 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1100                              struct userfaultfd_wake_range *range)
1101 {
1102         unsigned long start, end;
1103
1104         start = range->start;
1105         end = range->start + range->len;
1106
1107         spin_lock(&ctx->fault_pending_wqh.lock);
1108         /* wake all in the range and autoremove */
1109         if (waitqueue_active(&ctx->fault_pending_wqh))
1110                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1111                                      range);
1112         if (waitqueue_active(&ctx->fault_wqh))
1113                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1114         spin_unlock(&ctx->fault_pending_wqh.lock);
1115 }
1116
1117 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1118                                            struct userfaultfd_wake_range *range)
1119 {
1120         unsigned seq;
1121         bool need_wakeup;
1122
1123         /*
1124          * To be sure waitqueue_active() is not reordered by the CPU
1125          * before the pagetable update, use an explicit SMP memory
1126          * barrier here. PT lock release or up_read(mmap_sem) still
1127          * have release semantics that can allow the
1128          * waitqueue_active() to be reordered before the pte update.
1129          */
1130         smp_mb();
1131
1132         /*
1133          * Use waitqueue_active because it's very frequent to
1134          * change the address space atomically even if there are no
1135          * userfaults yet. So we take the spinlock only when we're
1136          * sure we've userfaults to wake.
1137          */
1138         do {
1139                 seq = read_seqcount_begin(&ctx->refile_seq);
1140                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1141                         waitqueue_active(&ctx->fault_wqh);
1142                 cond_resched();
1143         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1144         if (need_wakeup)
1145                 __wake_userfault(ctx, range);
1146 }
1147
1148 static __always_inline int validate_range(struct mm_struct *mm,
1149                                           __u64 start, __u64 len)
1150 {
1151         __u64 task_size = mm->task_size;
1152
1153         if (start & ~PAGE_MASK)
1154                 return -EINVAL;
1155         if (len & ~PAGE_MASK)
1156                 return -EINVAL;
1157         if (!len)
1158                 return -EINVAL;
1159         if (start < mmap_min_addr)
1160                 return -EINVAL;
1161         if (start >= task_size)
1162                 return -EINVAL;
1163         if (len > task_size - start)
1164                 return -EINVAL;
1165         return 0;
1166 }
1167
1168 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1169 {
1170         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1171                 vma_is_shmem(vma);
1172 }
1173
1174 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1175                                 unsigned long arg)
1176 {
1177         struct mm_struct *mm = ctx->mm;
1178         struct vm_area_struct *vma, *prev, *cur;
1179         int ret;
1180         struct uffdio_register uffdio_register;
1181         struct uffdio_register __user *user_uffdio_register;
1182         unsigned long vm_flags, new_flags;
1183         bool found;
1184         bool non_anon_pages;
1185         unsigned long start, end, vma_end;
1186
1187         user_uffdio_register = (struct uffdio_register __user *) arg;
1188
1189         ret = -EFAULT;
1190         if (copy_from_user(&uffdio_register, user_uffdio_register,
1191                            sizeof(uffdio_register)-sizeof(__u64)))
1192                 goto out;
1193
1194         ret = -EINVAL;
1195         if (!uffdio_register.mode)
1196                 goto out;
1197         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1198                                      UFFDIO_REGISTER_MODE_WP))
1199                 goto out;
1200         vm_flags = 0;
1201         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1202                 vm_flags |= VM_UFFD_MISSING;
1203         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1204                 vm_flags |= VM_UFFD_WP;
1205                 /*
1206                  * FIXME: remove the below error constraint by
1207                  * implementing the wprotect tracking mode.
1208                  */
1209                 ret = -EINVAL;
1210                 goto out;
1211         }
1212
1213         ret = validate_range(mm, uffdio_register.range.start,
1214                              uffdio_register.range.len);
1215         if (ret)
1216                 goto out;
1217
1218         start = uffdio_register.range.start;
1219         end = start + uffdio_register.range.len;
1220
1221         ret = -ENOMEM;
1222         if (!mmget_not_zero(mm))
1223                 goto out;
1224
1225         down_write(&mm->mmap_sem);
1226         vma = find_vma_prev(mm, start, &prev);
1227         if (!vma)
1228                 goto out_unlock;
1229
1230         /* check that there's at least one vma in the range */
1231         ret = -EINVAL;
1232         if (vma->vm_start >= end)
1233                 goto out_unlock;
1234
1235         /*
1236          * If the first vma contains huge pages, make sure start address
1237          * is aligned to huge page size.
1238          */
1239         if (is_vm_hugetlb_page(vma)) {
1240                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1241
1242                 if (start & (vma_hpagesize - 1))
1243                         goto out_unlock;
1244         }
1245
1246         /*
1247          * Search for not compatible vmas.
1248          */
1249         found = false;
1250         non_anon_pages = false;
1251         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1252                 cond_resched();
1253
1254                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1255                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1256
1257                 /* check not compatible vmas */
1258                 ret = -EINVAL;
1259                 if (!vma_can_userfault(cur))
1260                         goto out_unlock;
1261                 /*
1262                  * If this vma contains ending address, and huge pages
1263                  * check alignment.
1264                  */
1265                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1266                     end > cur->vm_start) {
1267                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1268
1269                         ret = -EINVAL;
1270
1271                         if (end & (vma_hpagesize - 1))
1272                                 goto out_unlock;
1273                 }
1274
1275                 /*
1276                  * Check that this vma isn't already owned by a
1277                  * different userfaultfd. We can't allow more than one
1278                  * userfaultfd to own a single vma simultaneously or we
1279                  * wouldn't know which one to deliver the userfaults to.
1280                  */
1281                 ret = -EBUSY;
1282                 if (cur->vm_userfaultfd_ctx.ctx &&
1283                     cur->vm_userfaultfd_ctx.ctx != ctx)
1284                         goto out_unlock;
1285
1286                 /*
1287                  * Note vmas containing huge pages
1288                  */
1289                 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1290                         non_anon_pages = true;
1291
1292                 found = true;
1293         }
1294         BUG_ON(!found);
1295
1296         if (vma->vm_start < start)
1297                 prev = vma;
1298
1299         ret = 0;
1300         do {
1301                 cond_resched();
1302
1303                 BUG_ON(!vma_can_userfault(vma));
1304                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1305                        vma->vm_userfaultfd_ctx.ctx != ctx);
1306
1307                 /*
1308                  * Nothing to do: this vma is already registered into this
1309                  * userfaultfd and with the right tracking mode too.
1310                  */
1311                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1312                     (vma->vm_flags & vm_flags) == vm_flags)
1313                         goto skip;
1314
1315                 if (vma->vm_start > start)
1316                         start = vma->vm_start;
1317                 vma_end = min(end, vma->vm_end);
1318
1319                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1320                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1321                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1322                                  vma_policy(vma),
1323                                  ((struct vm_userfaultfd_ctx){ ctx }));
1324                 if (prev) {
1325                         vma = prev;
1326                         goto next;
1327                 }
1328                 if (vma->vm_start < start) {
1329                         ret = split_vma(mm, vma, start, 1);
1330                         if (ret)
1331                                 break;
1332                 }
1333                 if (vma->vm_end > end) {
1334                         ret = split_vma(mm, vma, end, 0);
1335                         if (ret)
1336                                 break;
1337                 }
1338         next:
1339                 /*
1340                  * In the vma_merge() successful mprotect-like case 8:
1341                  * the next vma was merged into the current one and
1342                  * the current one has not been updated yet.
1343                  */
1344                 vma->vm_flags = new_flags;
1345                 vma->vm_userfaultfd_ctx.ctx = ctx;
1346
1347         skip:
1348                 prev = vma;
1349                 start = vma->vm_end;
1350                 vma = vma->vm_next;
1351         } while (vma && vma->vm_start < end);
1352 out_unlock:
1353         up_write(&mm->mmap_sem);
1354         mmput(mm);
1355         if (!ret) {
1356                 /*
1357                  * Now that we scanned all vmas we can already tell
1358                  * userland which ioctls methods are guaranteed to
1359                  * succeed on this range.
1360                  */
1361                 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1362                              UFFD_API_RANGE_IOCTLS,
1363                              &user_uffdio_register->ioctls))
1364                         ret = -EFAULT;
1365         }
1366 out:
1367         return ret;
1368 }
1369
1370 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1371                                   unsigned long arg)
1372 {
1373         struct mm_struct *mm = ctx->mm;
1374         struct vm_area_struct *vma, *prev, *cur;
1375         int ret;
1376         struct uffdio_range uffdio_unregister;
1377         unsigned long new_flags;
1378         bool found;
1379         unsigned long start, end, vma_end;
1380         const void __user *buf = (void __user *)arg;
1381
1382         ret = -EFAULT;
1383         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1384                 goto out;
1385
1386         ret = validate_range(mm, uffdio_unregister.start,
1387                              uffdio_unregister.len);
1388         if (ret)
1389                 goto out;
1390
1391         start = uffdio_unregister.start;
1392         end = start + uffdio_unregister.len;
1393
1394         ret = -ENOMEM;
1395         if (!mmget_not_zero(mm))
1396                 goto out;
1397
1398         down_write(&mm->mmap_sem);
1399         vma = find_vma_prev(mm, start, &prev);
1400         if (!vma)
1401                 goto out_unlock;
1402
1403         /* check that there's at least one vma in the range */
1404         ret = -EINVAL;
1405         if (vma->vm_start >= end)
1406                 goto out_unlock;
1407
1408         /*
1409          * If the first vma contains huge pages, make sure start address
1410          * is aligned to huge page size.
1411          */
1412         if (is_vm_hugetlb_page(vma)) {
1413                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1414
1415                 if (start & (vma_hpagesize - 1))
1416                         goto out_unlock;
1417         }
1418
1419         /*
1420          * Search for not compatible vmas.
1421          */
1422         found = false;
1423         ret = -EINVAL;
1424         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1425                 cond_resched();
1426
1427                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1428                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1429
1430                 /*
1431                  * Check not compatible vmas, not strictly required
1432                  * here as not compatible vmas cannot have an
1433                  * userfaultfd_ctx registered on them, but this
1434                  * provides for more strict behavior to notice
1435                  * unregistration errors.
1436                  */
1437                 if (!vma_can_userfault(cur))
1438                         goto out_unlock;
1439
1440                 found = true;
1441         }
1442         BUG_ON(!found);
1443
1444         if (vma->vm_start < start)
1445                 prev = vma;
1446
1447         ret = 0;
1448         do {
1449                 cond_resched();
1450
1451                 BUG_ON(!vma_can_userfault(vma));
1452
1453                 /*
1454                  * Nothing to do: this vma is already registered into this
1455                  * userfaultfd and with the right tracking mode too.
1456                  */
1457                 if (!vma->vm_userfaultfd_ctx.ctx)
1458                         goto skip;
1459
1460                 if (vma->vm_start > start)
1461                         start = vma->vm_start;
1462                 vma_end = min(end, vma->vm_end);
1463
1464                 if (userfaultfd_missing(vma)) {
1465                         /*
1466                          * Wake any concurrent pending userfault while
1467                          * we unregister, so they will not hang
1468                          * permanently and it avoids userland to call
1469                          * UFFDIO_WAKE explicitly.
1470                          */
1471                         struct userfaultfd_wake_range range;
1472                         range.start = start;
1473                         range.len = vma_end - start;
1474                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1475                 }
1476
1477                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1478                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1479                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1480                                  vma_policy(vma),
1481                                  NULL_VM_UFFD_CTX);
1482                 if (prev) {
1483                         vma = prev;
1484                         goto next;
1485                 }
1486                 if (vma->vm_start < start) {
1487                         ret = split_vma(mm, vma, start, 1);
1488                         if (ret)
1489                                 break;
1490                 }
1491                 if (vma->vm_end > end) {
1492                         ret = split_vma(mm, vma, end, 0);
1493                         if (ret)
1494                                 break;
1495                 }
1496         next:
1497                 /*
1498                  * In the vma_merge() successful mprotect-like case 8:
1499                  * the next vma was merged into the current one and
1500                  * the current one has not been updated yet.
1501                  */
1502                 vma->vm_flags = new_flags;
1503                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1504
1505         skip:
1506                 prev = vma;
1507                 start = vma->vm_end;
1508                 vma = vma->vm_next;
1509         } while (vma && vma->vm_start < end);
1510 out_unlock:
1511         up_write(&mm->mmap_sem);
1512         mmput(mm);
1513 out:
1514         return ret;
1515 }
1516
1517 /*
1518  * userfaultfd_wake may be used in combination with the
1519  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1520  */
1521 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1522                             unsigned long arg)
1523 {
1524         int ret;
1525         struct uffdio_range uffdio_wake;
1526         struct userfaultfd_wake_range range;
1527         const void __user *buf = (void __user *)arg;
1528
1529         ret = -EFAULT;
1530         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1531                 goto out;
1532
1533         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1534         if (ret)
1535                 goto out;
1536
1537         range.start = uffdio_wake.start;
1538         range.len = uffdio_wake.len;
1539
1540         /*
1541          * len == 0 means wake all and we don't want to wake all here,
1542          * so check it again to be sure.
1543          */
1544         VM_BUG_ON(!range.len);
1545
1546         wake_userfault(ctx, &range);
1547         ret = 0;
1548
1549 out:
1550         return ret;
1551 }
1552
1553 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1554                             unsigned long arg)
1555 {
1556         __s64 ret;
1557         struct uffdio_copy uffdio_copy;
1558         struct uffdio_copy __user *user_uffdio_copy;
1559         struct userfaultfd_wake_range range;
1560
1561         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1562
1563         ret = -EFAULT;
1564         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1565                            /* don't copy "copy" last field */
1566                            sizeof(uffdio_copy)-sizeof(__s64)))
1567                 goto out;
1568
1569         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1570         if (ret)
1571                 goto out;
1572         /*
1573          * double check for wraparound just in case. copy_from_user()
1574          * will later check uffdio_copy.src + uffdio_copy.len to fit
1575          * in the userland range.
1576          */
1577         ret = -EINVAL;
1578         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1579                 goto out;
1580         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1581                 goto out;
1582         if (mmget_not_zero(ctx->mm)) {
1583                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1584                                    uffdio_copy.len);
1585                 mmput(ctx->mm);
1586         } else {
1587                 return -ENOSPC;
1588         }
1589         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1590                 return -EFAULT;
1591         if (ret < 0)
1592                 goto out;
1593         BUG_ON(!ret);
1594         /* len == 0 would wake all */
1595         range.len = ret;
1596         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1597                 range.start = uffdio_copy.dst;
1598                 wake_userfault(ctx, &range);
1599         }
1600         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1601 out:
1602         return ret;
1603 }
1604
1605 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1606                                 unsigned long arg)
1607 {
1608         __s64 ret;
1609         struct uffdio_zeropage uffdio_zeropage;
1610         struct uffdio_zeropage __user *user_uffdio_zeropage;
1611         struct userfaultfd_wake_range range;
1612
1613         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1614
1615         ret = -EFAULT;
1616         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1617                            /* don't copy "zeropage" last field */
1618                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1619                 goto out;
1620
1621         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1622                              uffdio_zeropage.range.len);
1623         if (ret)
1624                 goto out;
1625         ret = -EINVAL;
1626         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1627                 goto out;
1628
1629         if (mmget_not_zero(ctx->mm)) {
1630                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1631                                      uffdio_zeropage.range.len);
1632                 mmput(ctx->mm);
1633         }
1634         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1635                 return -EFAULT;
1636         if (ret < 0)
1637                 goto out;
1638         /* len == 0 would wake all */
1639         BUG_ON(!ret);
1640         range.len = ret;
1641         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1642                 range.start = uffdio_zeropage.range.start;
1643                 wake_userfault(ctx, &range);
1644         }
1645         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1646 out:
1647         return ret;
1648 }
1649
1650 static inline unsigned int uffd_ctx_features(__u64 user_features)
1651 {
1652         /*
1653          * For the current set of features the bits just coincide
1654          */
1655         return (unsigned int)user_features;
1656 }
1657
1658 /*
1659  * userland asks for a certain API version and we return which bits
1660  * and ioctl commands are implemented in this kernel for such API
1661  * version or -EINVAL if unknown.
1662  */
1663 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1664                            unsigned long arg)
1665 {
1666         struct uffdio_api uffdio_api;
1667         void __user *buf = (void __user *)arg;
1668         int ret;
1669         __u64 features;
1670
1671         ret = -EINVAL;
1672         if (ctx->state != UFFD_STATE_WAIT_API)
1673                 goto out;
1674         ret = -EFAULT;
1675         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1676                 goto out;
1677         features = uffdio_api.features;
1678         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1679                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1680                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1681                         goto out;
1682                 ret = -EINVAL;
1683                 goto out;
1684         }
1685         /* report all available features and ioctls to userland */
1686         uffdio_api.features = UFFD_API_FEATURES;
1687         uffdio_api.ioctls = UFFD_API_IOCTLS;
1688         ret = -EFAULT;
1689         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1690                 goto out;
1691         ctx->state = UFFD_STATE_RUNNING;
1692         /* only enable the requested features for this uffd context */
1693         ctx->features = uffd_ctx_features(features);
1694         ret = 0;
1695 out:
1696         return ret;
1697 }
1698
1699 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1700                               unsigned long arg)
1701 {
1702         int ret = -EINVAL;
1703         struct userfaultfd_ctx *ctx = file->private_data;
1704
1705         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1706                 return -EINVAL;
1707
1708         switch(cmd) {
1709         case UFFDIO_API:
1710                 ret = userfaultfd_api(ctx, arg);
1711                 break;
1712         case UFFDIO_REGISTER:
1713                 ret = userfaultfd_register(ctx, arg);
1714                 break;
1715         case UFFDIO_UNREGISTER:
1716                 ret = userfaultfd_unregister(ctx, arg);
1717                 break;
1718         case UFFDIO_WAKE:
1719                 ret = userfaultfd_wake(ctx, arg);
1720                 break;
1721         case UFFDIO_COPY:
1722                 ret = userfaultfd_copy(ctx, arg);
1723                 break;
1724         case UFFDIO_ZEROPAGE:
1725                 ret = userfaultfd_zeropage(ctx, arg);
1726                 break;
1727         }
1728         return ret;
1729 }
1730
1731 #ifdef CONFIG_PROC_FS
1732 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1733 {
1734         struct userfaultfd_ctx *ctx = f->private_data;
1735         wait_queue_t *wq;
1736         struct userfaultfd_wait_queue *uwq;
1737         unsigned long pending = 0, total = 0;
1738
1739         spin_lock(&ctx->fault_pending_wqh.lock);
1740         list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1741                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1742                 pending++;
1743                 total++;
1744         }
1745         list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1746                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1747                 total++;
1748         }
1749         spin_unlock(&ctx->fault_pending_wqh.lock);
1750
1751         /*
1752          * If more protocols will be added, there will be all shown
1753          * separated by a space. Like this:
1754          *      protocols: aa:... bb:...
1755          */
1756         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1757                    pending, total, UFFD_API, UFFD_API_FEATURES,
1758                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1759 }
1760 #endif
1761
1762 static const struct file_operations userfaultfd_fops = {
1763 #ifdef CONFIG_PROC_FS
1764         .show_fdinfo    = userfaultfd_show_fdinfo,
1765 #endif
1766         .release        = userfaultfd_release,
1767         .poll           = userfaultfd_poll,
1768         .read           = userfaultfd_read,
1769         .unlocked_ioctl = userfaultfd_ioctl,
1770         .compat_ioctl   = userfaultfd_ioctl,
1771         .llseek         = noop_llseek,
1772 };
1773
1774 static void init_once_userfaultfd_ctx(void *mem)
1775 {
1776         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1777
1778         init_waitqueue_head(&ctx->fault_pending_wqh);
1779         init_waitqueue_head(&ctx->fault_wqh);
1780         init_waitqueue_head(&ctx->event_wqh);
1781         init_waitqueue_head(&ctx->fd_wqh);
1782         seqcount_init(&ctx->refile_seq);
1783 }
1784
1785 /**
1786  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1787  * @flags: Flags for the userfaultfd file.
1788  *
1789  * This function creates a userfaultfd file pointer, w/out installing
1790  * it into the fd table. This is useful when the userfaultfd file is
1791  * used during the initialization of data structures that require
1792  * extra setup after the userfaultfd creation. So the userfaultfd
1793  * creation is split into the file pointer creation phase, and the
1794  * file descriptor installation phase.  In this way races with
1795  * userspace closing the newly installed file descriptor can be
1796  * avoided.  Returns a userfaultfd file pointer, or a proper error
1797  * pointer.
1798  */
1799 static struct file *userfaultfd_file_create(int flags)
1800 {
1801         struct file *file;
1802         struct userfaultfd_ctx *ctx;
1803
1804         BUG_ON(!current->mm);
1805
1806         /* Check the UFFD_* constants for consistency.  */
1807         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1808         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1809
1810         file = ERR_PTR(-EINVAL);
1811         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1812                 goto out;
1813
1814         file = ERR_PTR(-ENOMEM);
1815         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1816         if (!ctx)
1817                 goto out;
1818
1819         atomic_set(&ctx->refcount, 1);
1820         ctx->flags = flags;
1821         ctx->features = 0;
1822         ctx->state = UFFD_STATE_WAIT_API;
1823         ctx->released = false;
1824         ctx->mm = current->mm;
1825         /* prevent the mm struct to be freed */
1826         mmgrab(ctx->mm);
1827
1828         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1829                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1830         if (IS_ERR(file)) {
1831                 mmdrop(ctx->mm);
1832                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1833         }
1834 out:
1835         return file;
1836 }
1837
1838 SYSCALL_DEFINE1(userfaultfd, int, flags)
1839 {
1840         int fd, error;
1841         struct file *file;
1842
1843         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1844         if (error < 0)
1845                 return error;
1846         fd = error;
1847
1848         file = userfaultfd_file_create(flags);
1849         if (IS_ERR(file)) {
1850                 error = PTR_ERR(file);
1851                 goto err_put_unused_fd;
1852         }
1853         fd_install(fd, file);
1854
1855         return fd;
1856
1857 err_put_unused_fd:
1858         put_unused_fd(fd);
1859
1860         return error;
1861 }
1862
1863 static int __init userfaultfd_init(void)
1864 {
1865         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1866                                                 sizeof(struct userfaultfd_ctx),
1867                                                 0,
1868                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1869                                                 init_once_userfaultfd_ctx);
1870         return 0;
1871 }
1872 __initcall(userfaultfd_init);