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                         break;
 553                 }
 554
 555                 spin_unlock(&ctx->event_wqh.lock);
 556
 557                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 558                 schedule();
 559
 560                 spin_lock(&ctx->event_wqh.lock);
 561         }
 562         __set_current_state(TASK_RUNNING);
 563         spin_unlock(&ctx->event_wqh.lock);
 564
 565         /*
 566          * ctx may go away after this if the userfault pseudo fd is
 567          * already released.
 568          */
 569 out:
 570         userfaultfd_ctx_put(ctx);
 571 }
 572
 573 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 574                                        struct userfaultfd_wait_queue *ewq)
 575 {
 576         ewq->msg.event = 0;
 577         wake_up_locked(&ctx->event_wqh);
 578         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 579 }
 580
 581 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 582 {
 583         struct userfaultfd_ctx *ctx = NULL, *octx;
 584         struct userfaultfd_fork_ctx *fctx;
 585
 586         octx = vma->vm_userfaultfd_ctx.ctx;
 587         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 588                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 589                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 590                 return 0;
 591         }
 592
 593         list_for_each_entry(fctx, fcs, list)
 594                 if (fctx->orig == octx) {
 595                         ctx = fctx->new;
 596                         break;
 597                 }
 598
 599         if (!ctx) {
 600                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 601                 if (!fctx)
 602                         return -ENOMEM;
 603
 604                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 605                 if (!ctx) {
 606                         kfree(fctx);
 607                         return -ENOMEM;
 608                 }
 609
 610                 atomic_set(&ctx->refcount, 1);
 611                 ctx->flags = octx->flags;
 612                 ctx->state = UFFD_STATE_RUNNING;
 613                 ctx->features = octx->features;
 614                 ctx->released = false;
 615                 ctx->mm = vma->vm_mm;
 616                 atomic_inc(&ctx->mm->mm_count);
 617
 618                 userfaultfd_ctx_get(octx);
 619                 fctx->orig = octx;
 620                 fctx->new = ctx;
 621                 list_add_tail(&fctx->list, fcs);
 622         }
 623
 624         vma->vm_userfaultfd_ctx.ctx = ctx;
 625         return 0;
 626 }
 627
 628 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 629 {
 630         struct userfaultfd_ctx *ctx = fctx->orig;
 631         struct userfaultfd_wait_queue ewq;
 632
 633         msg_init(&ewq.msg);
 634
 635         ewq.msg.event = UFFD_EVENT_FORK;
 636         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 637
 638         userfaultfd_event_wait_completion(ctx, &ewq);
 639 }
 640
 641 void dup_userfaultfd_complete(struct list_head *fcs)
 642 {
 643         struct userfaultfd_fork_ctx *fctx, *n;
 644
 645         list_for_each_entry_safe(fctx, n, fcs, list) {
 646                 dup_fctx(fctx);
 647                 list_del(&fctx->list);
 648                 kfree(fctx);
 649         }
 650 }
 651
 652 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 653                              struct vm_userfaultfd_ctx *vm_ctx)
 654 {
 655         struct userfaultfd_ctx *ctx;
 656
 657         ctx = vma->vm_userfaultfd_ctx.ctx;
 658         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 659                 vm_ctx->ctx = ctx;
 660                 userfaultfd_ctx_get(ctx);
 661         }
 662 }
 663
 664 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 665                                  unsigned long from, unsigned long to,
 666                                  unsigned long len)
 667 {
 668         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 669         struct userfaultfd_wait_queue ewq;
 670
 671         if (!ctx)
 672                 return;
 673
 674         if (to & ~PAGE_MASK) {
 675                 userfaultfd_ctx_put(ctx);
 676                 return;
 677         }
 678
 679         msg_init(&ewq.msg);
 680
 681         ewq.msg.event = UFFD_EVENT_REMAP;
 682         ewq.msg.arg.remap.from = from;
 683         ewq.msg.arg.remap.to = to;
 684         ewq.msg.arg.remap.len = len;
 685
 686         userfaultfd_event_wait_completion(ctx, &ewq);
 687 }
 688
 689 void userfaultfd_remove(struct vm_area_struct *vma,
 690                         struct vm_area_struct **prev,
 691                         unsigned long start, unsigned long end)
 692 {
 693         struct mm_struct *mm = vma->vm_mm;
 694         struct userfaultfd_ctx *ctx;
 695         struct userfaultfd_wait_queue ewq;
 696
 697         ctx = vma->vm_userfaultfd_ctx.ctx;
 698         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 699                 return;
 700
 701         userfaultfd_ctx_get(ctx);
 702         up_read(&mm->mmap_sem);
 703
 704         *prev = NULL; /* We wait for ACK w/o the mmap semaphore */
 705
 706         msg_init(&ewq.msg);
 707
 708         ewq.msg.event = UFFD_EVENT_REMOVE;
 709         ewq.msg.arg.remove.start = start;
 710         ewq.msg.arg.remove.end = end;
 711
 712         userfaultfd_event_wait_completion(ctx, &ewq);
 713
 714         down_read(&mm->mmap_sem);
 715 }
 716
 717 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 718                           unsigned long start, unsigned long end)
 719 {
 720         struct userfaultfd_unmap_ctx *unmap_ctx;
 721
 722         list_for_each_entry(unmap_ctx, unmaps, list)
 723                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 724                     unmap_ctx->end == end)
 725                         return true;
 726
 727         return false;
 728 }
 729
 730 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 731                            unsigned long start, unsigned long end,
 732                            struct list_head *unmaps)
 733 {
 734         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 735                 struct userfaultfd_unmap_ctx *unmap_ctx;
 736                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 737
 738                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 739                     has_unmap_ctx(ctx, unmaps, start, end))
 740                         continue;
 741
 742                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 743                 if (!unmap_ctx)
 744                         return -ENOMEM;
 745
 746                 userfaultfd_ctx_get(ctx);
 747                 unmap_ctx->ctx = ctx;
 748                 unmap_ctx->start = start;
 749                 unmap_ctx->end = end;
 750                 list_add_tail(&unmap_ctx->list, unmaps);
 751         }
 752
 753         return 0;
 754 }
 755
 756 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 757 {
 758         struct userfaultfd_unmap_ctx *ctx, *n;
 759         struct userfaultfd_wait_queue ewq;
 760
 761         list_for_each_entry_safe(ctx, n, uf, list) {
 762                 msg_init(&ewq.msg);
 763
 764                 ewq.msg.event = UFFD_EVENT_UNMAP;
 765                 ewq.msg.arg.remove.start = ctx->start;
 766                 ewq.msg.arg.remove.end = ctx->end;
 767
 768                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 769
 770                 list_del(&ctx->list);
 771                 kfree(ctx);
 772         }
 773 }
 774
 775 static int userfaultfd_release(struct inode *inode, struct file *file)
 776 {
 777         struct userfaultfd_ctx *ctx = file->private_data;
 778         struct mm_struct *mm = ctx->mm;
 779         struct vm_area_struct *vma, *prev;
 780         /* len == 0 means wake all */
 781         struct userfaultfd_wake_range range = { .len = 0, };
 782         unsigned long new_flags;
 783
 784         ACCESS_ONCE(ctx->released) = true;
 785
 786         if (!mmget_not_zero(mm))
 787                 goto wakeup;
 788
 789         /*
 790          * Flush page faults out of all CPUs. NOTE: all page faults
 791          * must be retried without returning VM_FAULT_SIGBUS if
 792          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 793          * changes while handle_userfault released the mmap_sem. So
 794          * it's critical that released is set to true (above), before
 795          * taking the mmap_sem for writing.
 796          */
 797         down_write(&mm->mmap_sem);
 798         prev = NULL;
 799         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 800                 cond_resched();
 801                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 802                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 803                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 804                         prev = vma;
 805                         continue;
 806                 }
 807                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 808                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 809                                  new_flags, vma->anon_vma,
 810                                  vma->vm_file, vma->vm_pgoff,
 811                                  vma_policy(vma),
 812                                  NULL_VM_UFFD_CTX);
 813                 if (prev)
 814                         vma = prev;
 815                 else
 816                         prev = vma;
 817                 vma->vm_flags = new_flags;
 818                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 819         }
 820         up_write(&mm->mmap_sem);
 821         mmput(mm);
 822 wakeup:
 823         /*
 824          * After no new page faults can wait on this fault_*wqh, flush
 825          * the last page faults that may have been already waiting on
 826          * the fault_*wqh.
 827          */
 828         spin_lock(&ctx->fault_pending_wqh.lock);
 829         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 830         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 831         spin_unlock(&ctx->fault_pending_wqh.lock);
 832
 833         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 834         userfaultfd_ctx_put(ctx);
 835         return 0;
 836 }
 837
 838 /* fault_pending_wqh.lock must be hold by the caller */
 839 static inline struct userfaultfd_wait_queue *find_userfault_in(
 840                 wait_queue_head_t *wqh)
 841 {
 842         wait_queue_t *wq;
 843         struct userfaultfd_wait_queue *uwq;
 844
 845         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 846
 847         uwq = NULL;
 848         if (!waitqueue_active(wqh))
 849                 goto out;
 850         /* walk in reverse to provide FIFO behavior to read userfaults */
 851         wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
 852         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 853 out:
 854         return uwq;
 855 }
 856
 857 static inline struct userfaultfd_wait_queue *find_userfault(
 858                 struct userfaultfd_ctx *ctx)
 859 {
 860         return find_userfault_in(&ctx->fault_pending_wqh);
 861 }
 862
 863 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 864                 struct userfaultfd_ctx *ctx)
 865 {
 866         return find_userfault_in(&ctx->event_wqh);
 867 }
 868
 869 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 870 {
 871         struct userfaultfd_ctx *ctx = file->private_data;
 872         unsigned int ret;
 873
 874         poll_wait(file, &ctx->fd_wqh, wait);
 875
 876         switch (ctx->state) {
 877         case UFFD_STATE_WAIT_API:
 878                 return POLLERR;
 879         case UFFD_STATE_RUNNING:
 880                 /*
 881                  * poll() never guarantees that read won't block.
 882                  * userfaults can be waken before they're read().
 883                  */
 884                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 885                         return POLLERR;
 886                 /*
 887                  * lockless access to see if there are pending faults
 888                  * __pollwait last action is the add_wait_queue but
 889                  * the spin_unlock would allow the waitqueue_active to
 890                  * pass above the actual list_add inside
 891                  * add_wait_queue critical section. So use a full
 892                  * memory barrier to serialize the list_add write of
 893                  * add_wait_queue() with the waitqueue_active read
 894                  * below.
 895                  */
 896                 ret = 0;
 897                 smp_mb();
 898                 if (waitqueue_active(&ctx->fault_pending_wqh))
 899                         ret = POLLIN;
 900                 else if (waitqueue_active(&ctx->event_wqh))
 901                         ret = POLLIN;
 902
 903                 return ret;
 904         default:
 905                 WARN_ON_ONCE(1);
 906                 return POLLERR;
 907         }
 908 }
 909
 910 static const struct file_operations userfaultfd_fops;
 911
 912 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 913                                   struct userfaultfd_ctx *new,
 914                                   struct uffd_msg *msg)
 915 {
 916         int fd;
 917         struct file *file;
 918         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 919
 920         fd = get_unused_fd_flags(flags);
 921         if (fd < 0)
 922                 return fd;
 923
 924         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 925                                   O_RDWR | flags);
 926         if (IS_ERR(file)) {
 927                 put_unused_fd(fd);
 928                 return PTR_ERR(file);
 929         }
 930
 931         fd_install(fd, file);
 932         msg->arg.reserved.reserved1 = 0;
 933         msg->arg.fork.ufd = fd;
 934
 935         return 0;
 936 }
 937
 938 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 939                                     struct uffd_msg *msg)
 940 {
 941         ssize_t ret;
 942         DECLARE_WAITQUEUE(wait, current);
 943         struct userfaultfd_wait_queue *uwq;
 944         /*
 945          * Handling fork event requires sleeping operations, so
 946          * we drop the event_wqh lock, then do these ops, then
 947          * lock it back and wake up the waiter. While the lock is
 948          * dropped the ewq may go away so we keep track of it
 949          * carefully.
 950          */
 951         LIST_HEAD(fork_event);
 952         struct userfaultfd_ctx *fork_nctx = NULL;
 953
 954         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 955         spin_lock(&ctx->fd_wqh.lock);
 956         __add_wait_queue(&ctx->fd_wqh, &wait);
 957         for (;;) {
 958                 set_current_state(TASK_INTERRUPTIBLE);
 959                 spin_lock(&ctx->fault_pending_wqh.lock);
 960                 uwq = find_userfault(ctx);
 961                 if (uwq) {
 962                         /*
 963                          * Use a seqcount to repeat the lockless check
 964                          * in wake_userfault() to avoid missing
 965                          * wakeups because during the refile both
 966                          * waitqueue could become empty if this is the
 967                          * only userfault.
 968                          */
 969                         write_seqcount_begin(&ctx->refile_seq);
 970
 971                         /*
 972                          * The fault_pending_wqh.lock prevents the uwq
 973                          * to disappear from under us.
 974                          *
 975                          * Refile this userfault from
 976                          * fault_pending_wqh to fault_wqh, it's not
 977                          * pending anymore after we read it.
 978                          *
 979                          * Use list_del() by hand (as
 980                          * userfaultfd_wake_function also uses
 981                          * list_del_init() by hand) to be sure nobody
 982                          * changes __remove_wait_queue() to use
 983                          * list_del_init() in turn breaking the
 984                          * !list_empty_careful() check in
 985                          * handle_userfault(). The uwq->wq.task_list
 986                          * must never be empty at any time during the
 987                          * refile, or the waitqueue could disappear
 988                          * from under us. The "wait_queue_head_t"
 989                          * parameter of __remove_wait_queue() is unused
 990                          * anyway.
 991                          */
 992                         list_del(&uwq->wq.task_list);
 993                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
 994
 995                         write_seqcount_end(&ctx->refile_seq);
 996
 997                         /* careful to always initialize msg if ret == 0 */
 998                         *msg = uwq->msg;
 999                         spin_unlock(&ctx->fault_pending_wqh.lock);
1000                         ret = 0;
1001                         break;
1002                 }
1003                 spin_unlock(&ctx->fault_pending_wqh.lock);
1004
1005                 spin_lock(&ctx->event_wqh.lock);
1006                 uwq = find_userfault_evt(ctx);
1007                 if (uwq) {
1008                         *msg = uwq->msg;
1009
1010                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1011                                 fork_nctx = (struct userfaultfd_ctx *)
1012                                         (unsigned long)
1013                                         uwq->msg.arg.reserved.reserved1;
1014                                 list_move(&uwq->wq.task_list, &fork_event);
1015                                 spin_unlock(&ctx->event_wqh.lock);
1016                                 ret = 0;
1017                                 break;
1018                         }
1019
1020                         userfaultfd_event_complete(ctx, uwq);
1021                         spin_unlock(&ctx->event_wqh.lock);
1022                         ret = 0;
1023                         break;
1024                 }
1025                 spin_unlock(&ctx->event_wqh.lock);
1026
1027                 if (signal_pending(current)) {
1028                         ret = -ERESTARTSYS;
1029                         break;
1030                 }
1031                 if (no_wait) {
1032                         ret = -EAGAIN;
1033                         break;
1034                 }
1035                 spin_unlock(&ctx->fd_wqh.lock);
1036                 schedule();
1037                 spin_lock(&ctx->fd_wqh.lock);
1038         }
1039         __remove_wait_queue(&ctx->fd_wqh, &wait);
1040         __set_current_state(TASK_RUNNING);
1041         spin_unlock(&ctx->fd_wqh.lock);
1042
1043         if (!ret && msg->event == UFFD_EVENT_FORK) {
1044                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1045
1046                 if (!ret) {
1047                         spin_lock(&ctx->event_wqh.lock);
1048                         if (!list_empty(&fork_event)) {
1049                                 uwq = list_first_entry(&fork_event,
1050                                                        typeof(*uwq),
1051                                                        wq.task_list);
1052                                 list_del(&uwq->wq.task_list);
1053                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1054                                 userfaultfd_event_complete(ctx, uwq);
1055                         }
1056                         spin_unlock(&ctx->event_wqh.lock);
1057                 }
1058         }
1059
1060         return ret;
1061 }
1062
1063 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1064                                 size_t count, loff_t *ppos)
1065 {
1066         struct userfaultfd_ctx *ctx = file->private_data;
1067         ssize_t _ret, ret = 0;
1068         struct uffd_msg msg;
1069         int no_wait = file->f_flags & O_NONBLOCK;
1070
1071         if (ctx->state == UFFD_STATE_WAIT_API)
1072                 return -EINVAL;
1073
1074         for (;;) {
1075                 if (count < sizeof(msg))
1076                         return ret ? ret : -EINVAL;
1077                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1078                 if (_ret < 0)
1079                         return ret ? ret : _ret;
1080                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1081                         return ret ? ret : -EFAULT;
1082                 ret += sizeof(msg);
1083                 buf += sizeof(msg);
1084                 count -= sizeof(msg);
1085                 /*
1086                  * Allow to read more than one fault at time but only
1087                  * block if waiting for the very first one.
1088                  */
1089                 no_wait = O_NONBLOCK;
1090         }
1091 }
1092
1093 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1094                              struct userfaultfd_wake_range *range)
1095 {
1096         unsigned long start, end;
1097
1098         start = range->start;
1099         end = range->start + range->len;
1100
1101         spin_lock(&ctx->fault_pending_wqh.lock);
1102         /* wake all in the range and autoremove */
1103         if (waitqueue_active(&ctx->fault_pending_wqh))
1104                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1105                                      range);
1106         if (waitqueue_active(&ctx->fault_wqh))
1107                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1108         spin_unlock(&ctx->fault_pending_wqh.lock);
1109 }
1110
1111 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1112                                            struct userfaultfd_wake_range *range)
1113 {
1114         unsigned seq;
1115         bool need_wakeup;
1116
1117         /*
1118          * To be sure waitqueue_active() is not reordered by the CPU
1119          * before the pagetable update, use an explicit SMP memory
1120          * barrier here. PT lock release or up_read(mmap_sem) still
1121          * have release semantics that can allow the
1122          * waitqueue_active() to be reordered before the pte update.
1123          */
1124         smp_mb();
1125
1126         /*
1127          * Use waitqueue_active because it's very frequent to
1128          * change the address space atomically even if there are no
1129          * userfaults yet. So we take the spinlock only when we're
1130          * sure we've userfaults to wake.
1131          */
1132         do {
1133                 seq = read_seqcount_begin(&ctx->refile_seq);
1134                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1135                         waitqueue_active(&ctx->fault_wqh);
1136                 cond_resched();
1137         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1138         if (need_wakeup)
1139                 __wake_userfault(ctx, range);
1140 }
1141
1142 static __always_inline int validate_range(struct mm_struct *mm,
1143                                           __u64 start, __u64 len)
1144 {
1145         __u64 task_size = mm->task_size;
1146
1147         if (start & ~PAGE_MASK)
1148                 return -EINVAL;
1149         if (len & ~PAGE_MASK)
1150                 return -EINVAL;
1151         if (!len)
1152                 return -EINVAL;
1153         if (start < mmap_min_addr)
1154                 return -EINVAL;
1155         if (start >= task_size)
1156                 return -EINVAL;
1157         if (len > task_size - start)
1158                 return -EINVAL;
1159         return 0;
1160 }
1161
1162 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1163 {
1164         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1165                 vma_is_shmem(vma);
1166 }
1167
1168 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1169                                 unsigned long arg)
1170 {
1171         struct mm_struct *mm = ctx->mm;
1172         struct vm_area_struct *vma, *prev, *cur;
1173         int ret;
1174         struct uffdio_register uffdio_register;
1175         struct uffdio_register __user *user_uffdio_register;
1176         unsigned long vm_flags, new_flags;
1177         bool found;
1178         bool non_anon_pages;
1179         unsigned long start, end, vma_end;
1180
1181         user_uffdio_register = (struct uffdio_register __user *) arg;
1182
1183         ret = -EFAULT;
1184         if (copy_from_user(&uffdio_register, user_uffdio_register,
1185                            sizeof(uffdio_register)-sizeof(__u64)))
1186                 goto out;
1187
1188         ret = -EINVAL;
1189         if (!uffdio_register.mode)
1190                 goto out;
1191         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1192                                      UFFDIO_REGISTER_MODE_WP))
1193                 goto out;
1194         vm_flags = 0;
1195         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1196                 vm_flags |= VM_UFFD_MISSING;
1197         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1198                 vm_flags |= VM_UFFD_WP;
1199                 /*
1200                  * FIXME: remove the below error constraint by
1201                  * implementing the wprotect tracking mode.
1202                  */
1203                 ret = -EINVAL;
1204                 goto out;
1205         }
1206
1207         ret = validate_range(mm, uffdio_register.range.start,
1208                              uffdio_register.range.len);
1209         if (ret)
1210                 goto out;
1211
1212         start = uffdio_register.range.start;
1213         end = start + uffdio_register.range.len;
1214
1215         ret = -ENOMEM;
1216         if (!mmget_not_zero(mm))
1217                 goto out;
1218
1219         down_write(&mm->mmap_sem);
1220         vma = find_vma_prev(mm, start, &prev);
1221         if (!vma)
1222                 goto out_unlock;
1223
1224         /* check that there's at least one vma in the range */
1225         ret = -EINVAL;
1226         if (vma->vm_start >= end)
1227                 goto out_unlock;
1228
1229         /*
1230          * If the first vma contains huge pages, make sure start address
1231          * is aligned to huge page size.
1232          */
1233         if (is_vm_hugetlb_page(vma)) {
1234                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1235
1236                 if (start & (vma_hpagesize - 1))
1237                         goto out_unlock;
1238         }
1239
1240         /*
1241          * Search for not compatible vmas.
1242          */
1243         found = false;
1244         non_anon_pages = false;
1245         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1246                 cond_resched();
1247
1248                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1249                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1250
1251                 /* check not compatible vmas */
1252                 ret = -EINVAL;
1253                 if (!vma_can_userfault(cur))
1254                         goto out_unlock;
1255                 /*
1256                  * If this vma contains ending address, and huge pages
1257                  * check alignment.
1258                  */
1259                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1260                     end > cur->vm_start) {
1261                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1262
1263                         ret = -EINVAL;
1264
1265                         if (end & (vma_hpagesize - 1))
1266                                 goto out_unlock;
1267                 }
1268
1269                 /*
1270                  * Check that this vma isn't already owned by a
1271                  * different userfaultfd. We can't allow more than one
1272                  * userfaultfd to own a single vma simultaneously or we
1273                  * wouldn't know which one to deliver the userfaults to.
1274                  */
1275                 ret = -EBUSY;
1276                 if (cur->vm_userfaultfd_ctx.ctx &&
1277                     cur->vm_userfaultfd_ctx.ctx != ctx)
1278                         goto out_unlock;
1279
1280                 /*
1281                  * Note vmas containing huge pages
1282                  */
1283                 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1284                         non_anon_pages = true;
1285
1286                 found = true;
1287         }
1288         BUG_ON(!found);
1289
1290         if (vma->vm_start < start)
1291                 prev = vma;
1292
1293         ret = 0;
1294         do {
1295                 cond_resched();
1296
1297                 BUG_ON(!vma_can_userfault(vma));
1298                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1299                        vma->vm_userfaultfd_ctx.ctx != ctx);
1300
1301                 /*
1302                  * Nothing to do: this vma is already registered into this
1303                  * userfaultfd and with the right tracking mode too.
1304                  */
1305                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1306                     (vma->vm_flags & vm_flags) == vm_flags)
1307                         goto skip;
1308
1309                 if (vma->vm_start > start)
1310                         start = vma->vm_start;
1311                 vma_end = min(end, vma->vm_end);
1312
1313                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1314                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1315                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1316                                  vma_policy(vma),
1317                                  ((struct vm_userfaultfd_ctx){ ctx }));
1318                 if (prev) {
1319                         vma = prev;
1320                         goto next;
1321                 }
1322                 if (vma->vm_start < start) {
1323                         ret = split_vma(mm, vma, start, 1);
1324                         if (ret)
1325                                 break;
1326                 }
1327                 if (vma->vm_end > end) {
1328                         ret = split_vma(mm, vma, end, 0);
1329                         if (ret)
1330                                 break;
1331                 }
1332         next:
1333                 /*
1334                  * In the vma_merge() successful mprotect-like case 8:
1335                  * the next vma was merged into the current one and
1336                  * the current one has not been updated yet.
1337                  */
1338                 vma->vm_flags = new_flags;
1339                 vma->vm_userfaultfd_ctx.ctx = ctx;
1340
1341         skip:
1342                 prev = vma;
1343                 start = vma->vm_end;
1344                 vma = vma->vm_next;
1345         } while (vma && vma->vm_start < end);
1346 out_unlock:
1347         up_write(&mm->mmap_sem);
1348         mmput(mm);
1349         if (!ret) {
1350                 /*
1351                  * Now that we scanned all vmas we can already tell
1352                  * userland which ioctls methods are guaranteed to
1353                  * succeed on this range.
1354                  */
1355                 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1356                              UFFD_API_RANGE_IOCTLS,
1357                              &user_uffdio_register->ioctls))
1358                         ret = -EFAULT;
1359         }
1360 out:
1361         return ret;
1362 }
1363
1364 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1365                                   unsigned long arg)
1366 {
1367         struct mm_struct *mm = ctx->mm;
1368         struct vm_area_struct *vma, *prev, *cur;
1369         int ret;
1370         struct uffdio_range uffdio_unregister;
1371         unsigned long new_flags;
1372         bool found;
1373         unsigned long start, end, vma_end;
1374         const void __user *buf = (void __user *)arg;
1375
1376         ret = -EFAULT;
1377         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1378                 goto out;
1379
1380         ret = validate_range(mm, uffdio_unregister.start,
1381                              uffdio_unregister.len);
1382         if (ret)
1383                 goto out;
1384
1385         start = uffdio_unregister.start;
1386         end = start + uffdio_unregister.len;
1387
1388         ret = -ENOMEM;
1389         if (!mmget_not_zero(mm))
1390                 goto out;
1391
1392         down_write(&mm->mmap_sem);
1393         vma = find_vma_prev(mm, start, &prev);
1394         if (!vma)
1395                 goto out_unlock;
1396
1397         /* check that there's at least one vma in the range */
1398         ret = -EINVAL;
1399         if (vma->vm_start >= end)
1400                 goto out_unlock;
1401
1402         /*
1403          * If the first vma contains huge pages, make sure start address
1404          * is aligned to huge page size.
1405          */
1406         if (is_vm_hugetlb_page(vma)) {
1407                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1408
1409                 if (start & (vma_hpagesize - 1))
1410                         goto out_unlock;
1411         }
1412
1413         /*
1414          * Search for not compatible vmas.
1415          */
1416         found = false;
1417         ret = -EINVAL;
1418         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1419                 cond_resched();
1420
1421                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1422                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1423
1424                 /*
1425                  * Check not compatible vmas, not strictly required
1426                  * here as not compatible vmas cannot have an
1427                  * userfaultfd_ctx registered on them, but this
1428                  * provides for more strict behavior to notice
1429                  * unregistration errors.
1430                  */
1431                 if (!vma_can_userfault(cur))
1432                         goto out_unlock;
1433
1434                 found = true;
1435         }
1436         BUG_ON(!found);
1437
1438         if (vma->vm_start < start)
1439                 prev = vma;
1440
1441         ret = 0;
1442         do {
1443                 cond_resched();
1444
1445                 BUG_ON(!vma_can_userfault(vma));
1446
1447                 /*
1448                  * Nothing to do: this vma is already registered into this
1449                  * userfaultfd and with the right tracking mode too.
1450                  */
1451                 if (!vma->vm_userfaultfd_ctx.ctx)
1452                         goto skip;
1453
1454                 if (vma->vm_start > start)
1455                         start = vma->vm_start;
1456                 vma_end = min(end, vma->vm_end);
1457
1458                 if (userfaultfd_missing(vma)) {
1459                         /*
1460                          * Wake any concurrent pending userfault while
1461                          * we unregister, so they will not hang
1462                          * permanently and it avoids userland to call
1463                          * UFFDIO_WAKE explicitly.
1464                          */
1465                         struct userfaultfd_wake_range range;
1466                         range.start = start;
1467                         range.len = vma_end - start;
1468                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1469                 }
1470
1471                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1472                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1473                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1474                                  vma_policy(vma),
1475                                  NULL_VM_UFFD_CTX);
1476                 if (prev) {
1477                         vma = prev;
1478                         goto next;
1479                 }
1480                 if (vma->vm_start < start) {
1481                         ret = split_vma(mm, vma, start, 1);
1482                         if (ret)
1483                                 break;
1484                 }
1485                 if (vma->vm_end > end) {
1486                         ret = split_vma(mm, vma, end, 0);
1487                         if (ret)
1488                                 break;
1489                 }
1490         next:
1491                 /*
1492                  * In the vma_merge() successful mprotect-like case 8:
1493                  * the next vma was merged into the current one and
1494                  * the current one has not been updated yet.
1495                  */
1496                 vma->vm_flags = new_flags;
1497                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1498
1499         skip:
1500                 prev = vma;
1501                 start = vma->vm_end;
1502                 vma = vma->vm_next;
1503         } while (vma && vma->vm_start < end);
1504 out_unlock:
1505         up_write(&mm->mmap_sem);
1506         mmput(mm);
1507 out:
1508         return ret;
1509 }
1510
1511 /*
1512  * userfaultfd_wake may be used in combination with the
1513  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1514  */
1515 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1516                             unsigned long arg)
1517 {
1518         int ret;
1519         struct uffdio_range uffdio_wake;
1520         struct userfaultfd_wake_range range;
1521         const void __user *buf = (void __user *)arg;
1522
1523         ret = -EFAULT;
1524         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1525                 goto out;
1526
1527         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1528         if (ret)
1529                 goto out;
1530
1531         range.start = uffdio_wake.start;
1532         range.len = uffdio_wake.len;
1533
1534         /*
1535          * len == 0 means wake all and we don't want to wake all here,
1536          * so check it again to be sure.
1537          */
1538         VM_BUG_ON(!range.len);
1539
1540         wake_userfault(ctx, &range);
1541         ret = 0;
1542
1543 out:
1544         return ret;
1545 }
1546
1547 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1548                             unsigned long arg)
1549 {
1550         __s64 ret;
1551         struct uffdio_copy uffdio_copy;
1552         struct uffdio_copy __user *user_uffdio_copy;
1553         struct userfaultfd_wake_range range;
1554
1555         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1556
1557         ret = -EFAULT;
1558         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1559                            /* don't copy "copy" last field */
1560                            sizeof(uffdio_copy)-sizeof(__s64)))
1561                 goto out;
1562
1563         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1564         if (ret)
1565                 goto out;
1566         /*
1567          * double check for wraparound just in case. copy_from_user()
1568          * will later check uffdio_copy.src + uffdio_copy.len to fit
1569          * in the userland range.
1570          */
1571         ret = -EINVAL;
1572         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1573                 goto out;
1574         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1575                 goto out;
1576         if (mmget_not_zero(ctx->mm)) {
1577                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1578                                    uffdio_copy.len);
1579                 mmput(ctx->mm);
1580         } else {
1581                 return -ENOSPC;
1582         }
1583         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1584                 return -EFAULT;
1585         if (ret < 0)
1586                 goto out;
1587         BUG_ON(!ret);
1588         /* len == 0 would wake all */
1589         range.len = ret;
1590         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1591                 range.start = uffdio_copy.dst;
1592                 wake_userfault(ctx, &range);
1593         }
1594         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1595 out:
1596         return ret;
1597 }
1598
1599 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1600                                 unsigned long arg)
1601 {
1602         __s64 ret;
1603         struct uffdio_zeropage uffdio_zeropage;
1604         struct uffdio_zeropage __user *user_uffdio_zeropage;
1605         struct userfaultfd_wake_range range;
1606
1607         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1608
1609         ret = -EFAULT;
1610         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1611                            /* don't copy "zeropage" last field */
1612                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1613                 goto out;
1614
1615         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1616                              uffdio_zeropage.range.len);
1617         if (ret)
1618                 goto out;
1619         ret = -EINVAL;
1620         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1621                 goto out;
1622
1623         if (mmget_not_zero(ctx->mm)) {
1624                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1625                                      uffdio_zeropage.range.len);
1626                 mmput(ctx->mm);
1627         }
1628         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1629                 return -EFAULT;
1630         if (ret < 0)
1631                 goto out;
1632         /* len == 0 would wake all */
1633         BUG_ON(!ret);
1634         range.len = ret;
1635         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1636                 range.start = uffdio_zeropage.range.start;
1637                 wake_userfault(ctx, &range);
1638         }
1639         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1640 out:
1641         return ret;
1642 }
1643
1644 static inline unsigned int uffd_ctx_features(__u64 user_features)
1645 {
1646         /*
1647          * For the current set of features the bits just coincide
1648          */
1649         return (unsigned int)user_features;
1650 }
1651
1652 /*
1653  * userland asks for a certain API version and we return which bits
1654  * and ioctl commands are implemented in this kernel for such API
1655  * version or -EINVAL if unknown.
1656  */
1657 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1658                            unsigned long arg)
1659 {
1660         struct uffdio_api uffdio_api;
1661         void __user *buf = (void __user *)arg;
1662         int ret;
1663         __u64 features;
1664
1665         ret = -EINVAL;
1666         if (ctx->state != UFFD_STATE_WAIT_API)
1667                 goto out;
1668         ret = -EFAULT;
1669         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1670                 goto out;
1671         features = uffdio_api.features;
1672         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1673                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1674                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1675                         goto out;
1676                 ret = -EINVAL;
1677                 goto out;
1678         }
1679         /* report all available features and ioctls to userland */
1680         uffdio_api.features = UFFD_API_FEATURES;
1681         uffdio_api.ioctls = UFFD_API_IOCTLS;
1682         ret = -EFAULT;
1683         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1684                 goto out;
1685         ctx->state = UFFD_STATE_RUNNING;
1686         /* only enable the requested features for this uffd context */
1687         ctx->features = uffd_ctx_features(features);
1688         ret = 0;
1689 out:
1690         return ret;
1691 }
1692
1693 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1694                               unsigned long arg)
1695 {
1696         int ret = -EINVAL;
1697         struct userfaultfd_ctx *ctx = file->private_data;
1698
1699         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1700                 return -EINVAL;
1701
1702         switch(cmd) {
1703         case UFFDIO_API:
1704                 ret = userfaultfd_api(ctx, arg);
1705                 break;
1706         case UFFDIO_REGISTER:
1707                 ret = userfaultfd_register(ctx, arg);
1708                 break;
1709         case UFFDIO_UNREGISTER:
1710                 ret = userfaultfd_unregister(ctx, arg);
1711                 break;
1712         case UFFDIO_WAKE:
1713                 ret = userfaultfd_wake(ctx, arg);
1714                 break;
1715         case UFFDIO_COPY:
1716                 ret = userfaultfd_copy(ctx, arg);
1717                 break;
1718         case UFFDIO_ZEROPAGE:
1719                 ret = userfaultfd_zeropage(ctx, arg);
1720                 break;
1721         }
1722         return ret;
1723 }
1724
1725 #ifdef CONFIG_PROC_FS
1726 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1727 {
1728         struct userfaultfd_ctx *ctx = f->private_data;
1729         wait_queue_t *wq;
1730         struct userfaultfd_wait_queue *uwq;
1731         unsigned long pending = 0, total = 0;
1732
1733         spin_lock(&ctx->fault_pending_wqh.lock);
1734         list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1735                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1736                 pending++;
1737                 total++;
1738         }
1739         list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1740                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1741                 total++;
1742         }
1743         spin_unlock(&ctx->fault_pending_wqh.lock);
1744
1745         /*
1746          * If more protocols will be added, there will be all shown
1747          * separated by a space. Like this:
1748          *      protocols: aa:... bb:...
1749          */
1750         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1751                    pending, total, UFFD_API, UFFD_API_FEATURES,
1752                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1753 }
1754 #endif
1755
1756 static const struct file_operations userfaultfd_fops = {
1757 #ifdef CONFIG_PROC_FS
1758         .show_fdinfo    = userfaultfd_show_fdinfo,
1759 #endif
1760         .release        = userfaultfd_release,
1761         .poll           = userfaultfd_poll,
1762         .read           = userfaultfd_read,
1763         .unlocked_ioctl = userfaultfd_ioctl,
1764         .compat_ioctl   = userfaultfd_ioctl,
1765         .llseek         = noop_llseek,
1766 };
1767
1768 static void init_once_userfaultfd_ctx(void *mem)
1769 {
1770         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1771
1772         init_waitqueue_head(&ctx->fault_pending_wqh);
1773         init_waitqueue_head(&ctx->fault_wqh);
1774         init_waitqueue_head(&ctx->event_wqh);
1775         init_waitqueue_head(&ctx->fd_wqh);
1776         seqcount_init(&ctx->refile_seq);
1777 }
1778
1779 /**
1780  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1781  * @flags: Flags for the userfaultfd file.
1782  *
1783  * This function creates a userfaultfd file pointer, w/out installing
1784  * it into the fd table. This is useful when the userfaultfd file is
1785  * used during the initialization of data structures that require
1786  * extra setup after the userfaultfd creation. So the userfaultfd
1787  * creation is split into the file pointer creation phase, and the
1788  * file descriptor installation phase.  In this way races with
1789  * userspace closing the newly installed file descriptor can be
1790  * avoided.  Returns a userfaultfd file pointer, or a proper error
1791  * pointer.
1792  */
1793 static struct file *userfaultfd_file_create(int flags)
1794 {
1795         struct file *file;
1796         struct userfaultfd_ctx *ctx;
1797
1798         BUG_ON(!current->mm);
1799
1800         /* Check the UFFD_* constants for consistency.  */
1801         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1802         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1803
1804         file = ERR_PTR(-EINVAL);
1805         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1806                 goto out;
1807
1808         file = ERR_PTR(-ENOMEM);
1809         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1810         if (!ctx)
1811                 goto out;
1812
1813         atomic_set(&ctx->refcount, 1);
1814         ctx->flags = flags;
1815         ctx->features = 0;
1816         ctx->state = UFFD_STATE_WAIT_API;
1817         ctx->released = false;
1818         ctx->mm = current->mm;
1819         /* prevent the mm struct to be freed */
1820         mmgrab(ctx->mm);
1821
1822         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1823                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1824         if (IS_ERR(file)) {
1825                 mmdrop(ctx->mm);
1826                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1827         }
1828 out:
1829         return file;
1830 }
1831
1832 SYSCALL_DEFINE1(userfaultfd, int, flags)
1833 {
1834         int fd, error;
1835         struct file *file;
1836
1837         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1838         if (error < 0)
1839                 return error;
1840         fd = error;
1841
1842         file = userfaultfd_file_create(flags);
1843         if (IS_ERR(file)) {
1844                 error = PTR_ERR(file);
1845                 goto err_put_unused_fd;
1846         }
1847         fd_install(fd, file);
1848
1849         return fd;
1850
1851 err_put_unused_fd:
1852         put_unused_fd(fd);
1853
1854         return error;
1855 }
1856
1857 static int __init userfaultfd_init(void)
1858 {
1859         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1860                                                 sizeof(struct userfaultfd_ctx),
1861                                                 0,
1862                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1863                                                 init_once_userfaultfd_ctx);
1864         return 0;
1865 }
1866 __initcall(userfaultfd_init);