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.h>
  18 #include <linux/mm.h>
  19 #include <linux/poll.h>
  20 #include <linux/slab.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/file.h>
  23 #include <linux/bug.h>
  24 #include <linux/anon_inodes.h>
  25 #include <linux/syscalls.h>
  26 #include <linux/userfaultfd_k.h>
  27 #include <linux/mempolicy.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/security.h>
  30 #include <linux/hugetlb.h>
  31
  32 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  33
  34 enum userfaultfd_state {
  35         UFFD_STATE_WAIT_API,
  36         UFFD_STATE_RUNNING,
  37 };
  38
  39 /*
  40  * Start with fault_pending_wqh and fault_wqh so they're more likely
  41  * to be in the same cacheline.
  42  */
  43 struct userfaultfd_ctx {
  44         /* waitqueue head for the pending (i.e. not read) userfaults */
  45         wait_queue_head_t fault_pending_wqh;
  46         /* waitqueue head for the userfaults */
  47         wait_queue_head_t fault_wqh;
  48         /* waitqueue head for the pseudo fd to wakeup poll/read */
  49         wait_queue_head_t fd_wqh;
  50         /* waitqueue head for events */
  51         wait_queue_head_t event_wqh;
  52         /* a refile sequence protected by fault_pending_wqh lock */
  53         struct seqcount refile_seq;
  54         /* pseudo fd refcounting */
  55         atomic_t refcount;
  56         /* userfaultfd syscall flags */
  57         unsigned int flags;
  58         /* features requested from the userspace */
  59         unsigned int features;
  60         /* state machine */
  61         enum userfaultfd_state state;
  62         /* released */
  63         bool released;
  64         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  65         struct mm_struct *mm;
  66 };
  67
  68 struct userfaultfd_fork_ctx {
  69         struct userfaultfd_ctx *orig;
  70         struct userfaultfd_ctx *new;
  71         struct list_head list;
  72 };
  73
  74 struct userfaultfd_unmap_ctx {
  75         struct userfaultfd_ctx *ctx;
  76         unsigned long start;
  77         unsigned long end;
  78         struct list_head list;
  79 };
  80
  81 struct userfaultfd_wait_queue {
  82         struct uffd_msg msg;
  83         wait_queue_t wq;
  84         struct userfaultfd_ctx *ctx;
  85         bool waken;
  86 };
  87
  88 struct userfaultfd_wake_range {
  89         unsigned long start;
  90         unsigned long len;
  91 };
  92
  93 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
  94                                      int wake_flags, void *key)
  95 {
  96         struct userfaultfd_wake_range *range = key;
  97         int ret;
  98         struct userfaultfd_wait_queue *uwq;
  99         unsigned long start, len;
 100
 101         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 102         ret = 0;
 103         /* len == 0 means wake all */
 104         start = range->start;
 105         len = range->len;
 106         if (len && (start > uwq->msg.arg.pagefault.address ||
 107                     start + len <= uwq->msg.arg.pagefault.address))
 108                 goto out;
 109         WRITE_ONCE(uwq->waken, true);
 110         /*
 111          * The implicit smp_mb__before_spinlock in try_to_wake_up()
 112          * renders uwq->waken visible to other CPUs before the task is
 113          * waken.
 114          */
 115         ret = wake_up_state(wq->private, mode);
 116         if (ret)
 117                 /*
 118                  * Wake only once, autoremove behavior.
 119                  *
 120                  * After the effect of list_del_init is visible to the
 121                  * other CPUs, the waitqueue may disappear from under
 122                  * us, see the !list_empty_careful() in
 123                  * handle_userfault(). try_to_wake_up() has an
 124                  * implicit smp_mb__before_spinlock, and the
 125                  * wq->private is read before calling the extern
 126                  * function "wake_up_state" (which in turns calls
 127                  * try_to_wake_up). While the spin_lock;spin_unlock;
 128                  * wouldn't be enough, the smp_mb__before_spinlock is
 129                  * enough to avoid an explicit smp_mb() here.
 130                  */
 131                 list_del_init(&wq->task_list);
 132 out:
 133         return ret;
 134 }
 135
 136 /**
 137  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 138  * context.
 139  * @ctx: [in] Pointer to the userfaultfd context.
 140  *
 141  * Returns: In case of success, returns not zero.
 142  */
 143 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 144 {
 145         if (!atomic_inc_not_zero(&ctx->refcount))
 146                 BUG();
 147 }
 148
 149 /**
 150  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 151  * context.
 152  * @ctx: [in] Pointer to userfaultfd context.
 153  *
 154  * The userfaultfd context reference must have been previously acquired either
 155  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 156  */
 157 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 158 {
 159         if (atomic_dec_and_test(&ctx->refcount)) {
 160                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 161                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 162                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 163                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 164                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 165                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 166                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 167                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 168                 mmdrop(ctx->mm);
 169                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 170         }
 171 }
 172
 173 static inline void msg_init(struct uffd_msg *msg)
 174 {
 175         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 176         /*
 177          * Must use memset to zero out the paddings or kernel data is
 178          * leaked to userland.
 179          */
 180         memset(msg, 0, sizeof(struct uffd_msg));
 181 }
 182
 183 static inline struct uffd_msg userfault_msg(unsigned long address,
 184                                             unsigned int flags,
 185                                             unsigned long reason)
 186 {
 187         struct uffd_msg msg;
 188         msg_init(&msg);
 189         msg.event = UFFD_EVENT_PAGEFAULT;
 190         msg.arg.pagefault.address = address;
 191         if (flags & FAULT_FLAG_WRITE)
 192                 /*
 193                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 194                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 195                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 196                  * was a read fault, otherwise if set it means it's
 197                  * a write fault.
 198                  */
 199                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 200         if (reason & VM_UFFD_WP)
 201                 /*
 202                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 203                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 204                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 205                  * a missing fault, otherwise if set it means it's a
 206                  * write protect fault.
 207                  */
 208                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 209         return msg;
 210 }
 211
 212 #ifdef CONFIG_HUGETLB_PAGE
 213 /*
 214  * Same functionality as userfaultfd_must_wait below with modifications for
 215  * hugepmd ranges.
 216  */
 217 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 218                                          unsigned long address,
 219                                          unsigned long flags,
 220                                          unsigned long reason)
 221 {
 222         struct mm_struct *mm = ctx->mm;
 223         pte_t *pte;
 224         bool ret = true;
 225
 226         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 227
 228         pte = huge_pte_offset(mm, address);
 229         if (!pte)
 230                 goto out;
 231
 232         ret = false;
 233
 234         /*
 235          * Lockless access: we're in a wait_event so it's ok if it
 236          * changes under us.
 237          */
 238         if (huge_pte_none(*pte))
 239                 ret = true;
 240         if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP))
 241                 ret = true;
 242 out:
 243         return ret;
 244 }
 245 #else
 246 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 247                                          unsigned long address,
 248                                          unsigned long flags,
 249                                          unsigned long reason)
 250 {
 251         return false;   /* should never get here */
 252 }
 253 #endif /* CONFIG_HUGETLB_PAGE */
 254
 255 /*
 256  * Verify the pagetables are still not ok after having reigstered into
 257  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 258  * userfault that has already been resolved, if userfaultfd_read and
 259  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 260  * threads.
 261  */
 262 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 263                                          unsigned long address,
 264                                          unsigned long flags,
 265                                          unsigned long reason)
 266 {
 267         struct mm_struct *mm = ctx->mm;
 268         pgd_t *pgd;
 269         pud_t *pud;
 270         pmd_t *pmd, _pmd;
 271         pte_t *pte;
 272         bool ret = true;
 273
 274         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 275
 276         pgd = pgd_offset(mm, address);
 277         if (!pgd_present(*pgd))
 278                 goto out;
 279         pud = pud_offset(pgd, address);
 280         if (!pud_present(*pud))
 281                 goto out;
 282         pmd = pmd_offset(pud, address);
 283         /*
 284          * READ_ONCE must function as a barrier with narrower scope
 285          * and it must be equivalent to:
 286          *      _pmd = *pmd; barrier();
 287          *
 288          * This is to deal with the instability (as in
 289          * pmd_trans_unstable) of the pmd.
 290          */
 291         _pmd = READ_ONCE(*pmd);
 292         if (!pmd_present(_pmd))
 293                 goto out;
 294
 295         ret = false;
 296         if (pmd_trans_huge(_pmd))
 297                 goto out;
 298
 299         /*
 300          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 301          * and use the standard pte_offset_map() instead of parsing _pmd.
 302          */
 303         pte = pte_offset_map(pmd, address);
 304         /*
 305          * Lockless access: we're in a wait_event so it's ok if it
 306          * changes under us.
 307          */
 308         if (pte_none(*pte))
 309                 ret = true;
 310         pte_unmap(pte);
 311
 312 out:
 313         return ret;
 314 }
 315
 316 /*
 317  * The locking rules involved in returning VM_FAULT_RETRY depending on
 318  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 319  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 320  * recommendation in __lock_page_or_retry is not an understatement.
 321  *
 322  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 323  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 324  * not set.
 325  *
 326  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 327  * set, VM_FAULT_RETRY can still be returned if and only if there are
 328  * fatal_signal_pending()s, and the mmap_sem must be released before
 329  * returning it.
 330  */
 331 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 332 {
 333         struct mm_struct *mm = vmf->vma->vm_mm;
 334         struct userfaultfd_ctx *ctx;
 335         struct userfaultfd_wait_queue uwq;
 336         int ret;
 337         bool must_wait, return_to_userland;
 338         long blocking_state;
 339
 340         BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 341
 342         ret = VM_FAULT_SIGBUS;
 343         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 344         if (!ctx)
 345                 goto out;
 346
 347         BUG_ON(ctx->mm != mm);
 348
 349         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 350         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 351
 352         /*
 353          * If it's already released don't get it. This avoids to loop
 354          * in __get_user_pages if userfaultfd_release waits on the
 355          * caller of handle_userfault to release the mmap_sem.
 356          */
 357         if (unlikely(ACCESS_ONCE(ctx->released)))
 358                 goto out;
 359
 360         /*
 361          * We don't do userfault handling for the final child pid update.
 362          */
 363         if (current->flags & PF_EXITING)
 364                 goto out;
 365
 366         /*
 367          * Check that we can return VM_FAULT_RETRY.
 368          *
 369          * NOTE: it should become possible to return VM_FAULT_RETRY
 370          * even if FAULT_FLAG_TRIED is set without leading to gup()
 371          * -EBUSY failures, if the userfaultfd is to be extended for
 372          * VM_UFFD_WP tracking and we intend to arm the userfault
 373          * without first stopping userland access to the memory. For
 374          * VM_UFFD_MISSING userfaults this is enough for now.
 375          */
 376         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 377                 /*
 378                  * Validate the invariant that nowait must allow retry
 379                  * to be sure not to return SIGBUS erroneously on
 380                  * nowait invocations.
 381                  */
 382                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 383 #ifdef CONFIG_DEBUG_VM
 384                 if (printk_ratelimit()) {
 385                         printk(KERN_WARNING
 386                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 387                                vmf->flags);
 388                         dump_stack();
 389                 }
 390 #endif
 391                 goto out;
 392         }
 393
 394         /*
 395          * Handle nowait, not much to do other than tell it to retry
 396          * and wait.
 397          */
 398         ret = VM_FAULT_RETRY;
 399         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 400                 goto out;
 401
 402         /* take the reference before dropping the mmap_sem */
 403         userfaultfd_ctx_get(ctx);
 404
 405         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 406         uwq.wq.private = current;
 407         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason);
 408         uwq.ctx = ctx;
 409         uwq.waken = false;
 410
 411         return_to_userland =
 412                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 413                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 414         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 415                          TASK_KILLABLE;
 416
 417         spin_lock(&ctx->fault_pending_wqh.lock);
 418         /*
 419          * After the __add_wait_queue the uwq is visible to userland
 420          * through poll/read().
 421          */
 422         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 423         /*
 424          * The smp_mb() after __set_current_state prevents the reads
 425          * following the spin_unlock to happen before the list_add in
 426          * __add_wait_queue.
 427          */
 428         set_current_state(blocking_state);
 429         spin_unlock(&ctx->fault_pending_wqh.lock);
 430
 431         if (!is_vm_hugetlb_page(vmf->vma))
 432                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 433                                                   reason);
 434         else
 435                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->address,
 436                                                        vmf->flags, reason);
 437         up_read(&mm->mmap_sem);
 438
 439         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
 440                    (return_to_userland ? !signal_pending(current) :
 441                     !fatal_signal_pending(current)))) {
 442                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 443                 schedule();
 444                 ret |= VM_FAULT_MAJOR;
 445
 446                 /*
 447                  * False wakeups can orginate even from rwsem before
 448                  * up_read() however userfaults will wait either for a
 449                  * targeted wakeup on the specific uwq waitqueue from
 450                  * wake_userfault() or for signals or for uffd
 451                  * release.
 452                  */
 453                 while (!READ_ONCE(uwq.waken)) {
 454                         /*
 455                          * This needs the full smp_store_mb()
 456                          * guarantee as the state write must be
 457                          * visible to other CPUs before reading
 458                          * uwq.waken from other CPUs.
 459                          */
 460                         set_current_state(blocking_state);
 461                         if (READ_ONCE(uwq.waken) ||
 462                             READ_ONCE(ctx->released) ||
 463                             (return_to_userland ? signal_pending(current) :
 464                              fatal_signal_pending(current)))
 465                                 break;
 466                         schedule();
 467                 }
 468         }
 469
 470         __set_current_state(TASK_RUNNING);
 471
 472         if (return_to_userland) {
 473                 if (signal_pending(current) &&
 474                     !fatal_signal_pending(current)) {
 475                         /*
 476                          * If we got a SIGSTOP or SIGCONT and this is
 477                          * a normal userland page fault, just let
 478                          * userland return so the signal will be
 479                          * handled and gdb debugging works.  The page
 480                          * fault code immediately after we return from
 481                          * this function is going to release the
 482                          * mmap_sem and it's not depending on it
 483                          * (unlike gup would if we were not to return
 484                          * VM_FAULT_RETRY).
 485                          *
 486                          * If a fatal signal is pending we still take
 487                          * the streamlined VM_FAULT_RETRY failure path
 488                          * and there's no need to retake the mmap_sem
 489                          * in such case.
 490                          */
 491                         down_read(&mm->mmap_sem);
 492                         ret = 0;
 493                 }
 494         }
 495
 496         /*
 497          * Here we race with the list_del; list_add in
 498          * userfaultfd_ctx_read(), however because we don't ever run
 499          * list_del_init() to refile across the two lists, the prev
 500          * and next pointers will never point to self. list_add also
 501          * would never let any of the two pointers to point to
 502          * self. So list_empty_careful won't risk to see both pointers
 503          * pointing to self at any time during the list refile. The
 504          * only case where list_del_init() is called is the full
 505          * removal in the wake function and there we don't re-list_add
 506          * and it's fine not to block on the spinlock. The uwq on this
 507          * kernel stack can be released after the list_del_init.
 508          */
 509         if (!list_empty_careful(&uwq.wq.task_list)) {
 510                 spin_lock(&ctx->fault_pending_wqh.lock);
 511                 /*
 512                  * No need of list_del_init(), the uwq on the stack
 513                  * will be freed shortly anyway.
 514                  */
 515                 list_del(&uwq.wq.task_list);
 516                 spin_unlock(&ctx->fault_pending_wqh.lock);
 517         }
 518
 519         /*
 520          * ctx may go away after this if the userfault pseudo fd is
 521          * already released.
 522          */
 523         userfaultfd_ctx_put(ctx);
 524
 525 out:
 526         return ret;
 527 }
 528
 529 static int userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 530                                              struct userfaultfd_wait_queue *ewq)
 531 {
 532         int ret = 0;
 533
 534         ewq->ctx = ctx;
 535         init_waitqueue_entry(&ewq->wq, current);
 536
 537         spin_lock(&ctx->event_wqh.lock);
 538         /*
 539          * After the __add_wait_queue the uwq is visible to userland
 540          * through poll/read().
 541          */
 542         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 543         for (;;) {
 544                 set_current_state(TASK_KILLABLE);
 545                 if (ewq->msg.event == 0)
 546                         break;
 547                 if (ACCESS_ONCE(ctx->released) ||
 548                     fatal_signal_pending(current)) {
 549                         ret = -1;
 550                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 551                         break;
 552                 }
 553
 554                 spin_unlock(&ctx->event_wqh.lock);
 555
 556                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 557                 schedule();
 558
 559                 spin_lock(&ctx->event_wqh.lock);
 560         }
 561         __set_current_state(TASK_RUNNING);
 562         spin_unlock(&ctx->event_wqh.lock);
 563
 564         /*
 565          * ctx may go away after this if the userfault pseudo fd is
 566          * already released.
 567          */
 568
 569         userfaultfd_ctx_put(ctx);
 570         return ret;
 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 int 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         return userfaultfd_event_wait_completion(ctx, &ewq);
 639 }
 640
 641 void dup_userfaultfd_complete(struct list_head *fcs)
 642 {
 643         int ret = 0;
 644         struct userfaultfd_fork_ctx *fctx, *n;
 645
 646         list_for_each_entry_safe(fctx, n, fcs, list) {
 647                 if (!ret)
 648                         ret = dup_fctx(fctx);
 649                 list_del(&fctx->list);
 650                 kfree(fctx);
 651         }
 652 }
 653
 654 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 655                              struct vm_userfaultfd_ctx *vm_ctx)
 656 {
 657         struct userfaultfd_ctx *ctx;
 658
 659         ctx = vma->vm_userfaultfd_ctx.ctx;
 660         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 661                 vm_ctx->ctx = ctx;
 662                 userfaultfd_ctx_get(ctx);
 663         }
 664 }
 665
 666 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 667                                  unsigned long from, unsigned long to,
 668                                  unsigned long len)
 669 {
 670         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 671         struct userfaultfd_wait_queue ewq;
 672
 673         if (!ctx)
 674                 return;
 675
 676         if (to & ~PAGE_MASK) {
 677                 userfaultfd_ctx_put(ctx);
 678                 return;
 679         }
 680
 681         msg_init(&ewq.msg);
 682
 683         ewq.msg.event = UFFD_EVENT_REMAP;
 684         ewq.msg.arg.remap.from = from;
 685         ewq.msg.arg.remap.to = to;
 686         ewq.msg.arg.remap.len = len;
 687
 688         userfaultfd_event_wait_completion(ctx, &ewq);
 689 }
 690
 691 void userfaultfd_remove(struct vm_area_struct *vma,
 692                         struct vm_area_struct **prev,
 693                         unsigned long start, unsigned long end)
 694 {
 695         struct mm_struct *mm = vma->vm_mm;
 696         struct userfaultfd_ctx *ctx;
 697         struct userfaultfd_wait_queue ewq;
 698
 699         ctx = vma->vm_userfaultfd_ctx.ctx;
 700         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 701                 return;
 702
 703         userfaultfd_ctx_get(ctx);
 704         up_read(&mm->mmap_sem);
 705
 706         *prev = NULL; /* We wait for ACK w/o the mmap semaphore */
 707
 708         msg_init(&ewq.msg);
 709
 710         ewq.msg.event = UFFD_EVENT_REMOVE;
 711         ewq.msg.arg.remove.start = start;
 712         ewq.msg.arg.remove.end = end;
 713
 714         userfaultfd_event_wait_completion(ctx, &ewq);
 715
 716         down_read(&mm->mmap_sem);
 717 }
 718
 719 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 720                           unsigned long start, unsigned long end)
 721 {
 722         struct userfaultfd_unmap_ctx *unmap_ctx;
 723
 724         list_for_each_entry(unmap_ctx, unmaps, list)
 725                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 726                     unmap_ctx->end == end)
 727                         return true;
 728
 729         return false;
 730 }
 731
 732 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 733                            unsigned long start, unsigned long end,
 734                            struct list_head *unmaps)
 735 {
 736         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 737                 struct userfaultfd_unmap_ctx *unmap_ctx;
 738                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 739
 740                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 741                     has_unmap_ctx(ctx, unmaps, start, end))
 742                         continue;
 743
 744                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 745                 if (!unmap_ctx)
 746                         return -ENOMEM;
 747
 748                 userfaultfd_ctx_get(ctx);
 749                 unmap_ctx->ctx = ctx;
 750                 unmap_ctx->start = start;
 751                 unmap_ctx->end = end;
 752                 list_add_tail(&unmap_ctx->list, unmaps);
 753         }
 754
 755         return 0;
 756 }
 757
 758 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 759 {
 760         struct userfaultfd_unmap_ctx *ctx, *n;
 761         struct userfaultfd_wait_queue ewq;
 762
 763         list_for_each_entry_safe(ctx, n, uf, list) {
 764                 msg_init(&ewq.msg);
 765
 766                 ewq.msg.event = UFFD_EVENT_UNMAP;
 767                 ewq.msg.arg.remove.start = ctx->start;
 768                 ewq.msg.arg.remove.end = ctx->end;
 769
 770                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 771
 772                 list_del(&ctx->list);
 773                 kfree(ctx);
 774         }
 775 }
 776
 777 void userfaultfd_exit(struct mm_struct *mm)
 778 {
 779         struct vm_area_struct *vma = mm->mmap;
 780
 781         /*
 782          * We can do the vma walk without locking because the caller
 783          * (exit_mm) knows it now has exclusive access
 784          */
 785         while (vma) {
 786                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 787
 788                 if (ctx && (ctx->features & UFFD_FEATURE_EVENT_EXIT)) {
 789                         struct userfaultfd_wait_queue ewq;
 790
 791                         userfaultfd_ctx_get(ctx);
 792
 793                         msg_init(&ewq.msg);
 794                         ewq.msg.event = UFFD_EVENT_EXIT;
 795
 796                         userfaultfd_event_wait_completion(ctx, &ewq);
 797
 798                         ctx->features &= ~UFFD_FEATURE_EVENT_EXIT;
 799                 }
 800
 801                 vma = vma->vm_next;
 802         }
 803 }
 804
 805 static int userfaultfd_release(struct inode *inode, struct file *file)
 806 {
 807         struct userfaultfd_ctx *ctx = file->private_data;
 808         struct mm_struct *mm = ctx->mm;
 809         struct vm_area_struct *vma, *prev;
 810         /* len == 0 means wake all */
 811         struct userfaultfd_wake_range range = { .len = 0, };
 812         unsigned long new_flags;
 813
 814         ACCESS_ONCE(ctx->released) = true;
 815
 816         if (!mmget_not_zero(mm))
 817                 goto wakeup;
 818
 819         /*
 820          * Flush page faults out of all CPUs. NOTE: all page faults
 821          * must be retried without returning VM_FAULT_SIGBUS if
 822          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 823          * changes while handle_userfault released the mmap_sem. So
 824          * it's critical that released is set to true (above), before
 825          * taking the mmap_sem for writing.
 826          */
 827         down_write(&mm->mmap_sem);
 828         prev = NULL;
 829         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 830                 cond_resched();
 831                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 832                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 833                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 834                         prev = vma;
 835                         continue;
 836                 }
 837                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 838                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 839                                  new_flags, vma->anon_vma,
 840                                  vma->vm_file, vma->vm_pgoff,
 841                                  vma_policy(vma),
 842                                  NULL_VM_UFFD_CTX);
 843                 if (prev)
 844                         vma = prev;
 845                 else
 846                         prev = vma;
 847                 vma->vm_flags = new_flags;
 848                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 849         }
 850         up_write(&mm->mmap_sem);
 851         mmput(mm);
 852 wakeup:
 853         /*
 854          * After no new page faults can wait on this fault_*wqh, flush
 855          * the last page faults that may have been already waiting on
 856          * the fault_*wqh.
 857          */
 858         spin_lock(&ctx->fault_pending_wqh.lock);
 859         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 860         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 861         spin_unlock(&ctx->fault_pending_wqh.lock);
 862
 863         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 864         userfaultfd_ctx_put(ctx);
 865         return 0;
 866 }
 867
 868 /* fault_pending_wqh.lock must be hold by the caller */
 869 static inline struct userfaultfd_wait_queue *find_userfault_in(
 870                 wait_queue_head_t *wqh)
 871 {
 872         wait_queue_t *wq;
 873         struct userfaultfd_wait_queue *uwq;
 874
 875         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 876
 877         uwq = NULL;
 878         if (!waitqueue_active(wqh))
 879                 goto out;
 880         /* walk in reverse to provide FIFO behavior to read userfaults */
 881         wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list);
 882         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 883 out:
 884         return uwq;
 885 }
 886
 887 static inline struct userfaultfd_wait_queue *find_userfault(
 888                 struct userfaultfd_ctx *ctx)
 889 {
 890         return find_userfault_in(&ctx->fault_pending_wqh);
 891 }
 892
 893 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 894                 struct userfaultfd_ctx *ctx)
 895 {
 896         return find_userfault_in(&ctx->event_wqh);
 897 }
 898
 899 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 900 {
 901         struct userfaultfd_ctx *ctx = file->private_data;
 902         unsigned int ret;
 903
 904         poll_wait(file, &ctx->fd_wqh, wait);
 905
 906         switch (ctx->state) {
 907         case UFFD_STATE_WAIT_API:
 908                 return POLLERR;
 909         case UFFD_STATE_RUNNING:
 910                 /*
 911                  * poll() never guarantees that read won't block.
 912                  * userfaults can be waken before they're read().
 913                  */
 914                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 915                         return POLLERR;
 916                 /*
 917                  * lockless access to see if there are pending faults
 918                  * __pollwait last action is the add_wait_queue but
 919                  * the spin_unlock would allow the waitqueue_active to
 920                  * pass above the actual list_add inside
 921                  * add_wait_queue critical section. So use a full
 922                  * memory barrier to serialize the list_add write of
 923                  * add_wait_queue() with the waitqueue_active read
 924                  * below.
 925                  */
 926                 ret = 0;
 927                 smp_mb();
 928                 if (waitqueue_active(&ctx->fault_pending_wqh))
 929                         ret = POLLIN;
 930                 else if (waitqueue_active(&ctx->event_wqh))
 931                         ret = POLLIN;
 932
 933                 return ret;
 934         default:
 935                 WARN_ON_ONCE(1);
 936                 return POLLERR;
 937         }
 938 }
 939
 940 static const struct file_operations userfaultfd_fops;
 941
 942 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 943                                   struct userfaultfd_ctx *new,
 944                                   struct uffd_msg *msg)
 945 {
 946         int fd;
 947         struct file *file;
 948         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 949
 950         fd = get_unused_fd_flags(flags);
 951         if (fd < 0)
 952                 return fd;
 953
 954         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
 955                                   O_RDWR | flags);
 956         if (IS_ERR(file)) {
 957                 put_unused_fd(fd);
 958                 return PTR_ERR(file);
 959         }
 960
 961         fd_install(fd, file);
 962         msg->arg.reserved.reserved1 = 0;
 963         msg->arg.fork.ufd = fd;
 964
 965         return 0;
 966 }
 967
 968 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 969                                     struct uffd_msg *msg)
 970 {
 971         ssize_t ret;
 972         DECLARE_WAITQUEUE(wait, current);
 973         struct userfaultfd_wait_queue *uwq;
 974         /*
 975          * Handling fork event requires sleeping operations, so
 976          * we drop the event_wqh lock, then do these ops, then
 977          * lock it back and wake up the waiter. While the lock is
 978          * dropped the ewq may go away so we keep track of it
 979          * carefully.
 980          */
 981         LIST_HEAD(fork_event);
 982         struct userfaultfd_ctx *fork_nctx = NULL;
 983
 984         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 985         spin_lock(&ctx->fd_wqh.lock);
 986         __add_wait_queue(&ctx->fd_wqh, &wait);
 987         for (;;) {
 988                 set_current_state(TASK_INTERRUPTIBLE);
 989                 spin_lock(&ctx->fault_pending_wqh.lock);
 990                 uwq = find_userfault(ctx);
 991                 if (uwq) {
 992                         /*
 993                          * Use a seqcount to repeat the lockless check
 994                          * in wake_userfault() to avoid missing
 995                          * wakeups because during the refile both
 996                          * waitqueue could become empty if this is the
 997                          * only userfault.
 998                          */
 999                         write_seqcount_begin(&ctx->refile_seq);
1000
1001                         /*
1002                          * The fault_pending_wqh.lock prevents the uwq
1003                          * to disappear from under us.
1004                          *
1005                          * Refile this userfault from
1006                          * fault_pending_wqh to fault_wqh, it's not
1007                          * pending anymore after we read it.
1008                          *
1009                          * Use list_del() by hand (as
1010                          * userfaultfd_wake_function also uses
1011                          * list_del_init() by hand) to be sure nobody
1012                          * changes __remove_wait_queue() to use
1013                          * list_del_init() in turn breaking the
1014                          * !list_empty_careful() check in
1015                          * handle_userfault(). The uwq->wq.task_list
1016                          * must never be empty at any time during the
1017                          * refile, or the waitqueue could disappear
1018                          * from under us. The "wait_queue_head_t"
1019                          * parameter of __remove_wait_queue() is unused
1020                          * anyway.
1021                          */
1022                         list_del(&uwq->wq.task_list);
1023                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1024
1025                         write_seqcount_end(&ctx->refile_seq);
1026
1027                         /* careful to always initialize msg if ret == 0 */
1028                         *msg = uwq->msg;
1029                         spin_unlock(&ctx->fault_pending_wqh.lock);
1030                         ret = 0;
1031                         break;
1032                 }
1033                 spin_unlock(&ctx->fault_pending_wqh.lock);
1034
1035                 spin_lock(&ctx->event_wqh.lock);
1036                 uwq = find_userfault_evt(ctx);
1037                 if (uwq) {
1038                         *msg = uwq->msg;
1039
1040                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1041                                 fork_nctx = (struct userfaultfd_ctx *)
1042                                         (unsigned long)
1043                                         uwq->msg.arg.reserved.reserved1;
1044                                 list_move(&uwq->wq.task_list, &fork_event);
1045                                 spin_unlock(&ctx->event_wqh.lock);
1046                                 ret = 0;
1047                                 break;
1048                         }
1049
1050                         userfaultfd_event_complete(ctx, uwq);
1051                         spin_unlock(&ctx->event_wqh.lock);
1052                         ret = 0;
1053                         break;
1054                 }
1055                 spin_unlock(&ctx->event_wqh.lock);
1056
1057                 if (signal_pending(current)) {
1058                         ret = -ERESTARTSYS;
1059                         break;
1060                 }
1061                 if (no_wait) {
1062                         ret = -EAGAIN;
1063                         break;
1064                 }
1065                 spin_unlock(&ctx->fd_wqh.lock);
1066                 schedule();
1067                 spin_lock(&ctx->fd_wqh.lock);
1068         }
1069         __remove_wait_queue(&ctx->fd_wqh, &wait);
1070         __set_current_state(TASK_RUNNING);
1071         spin_unlock(&ctx->fd_wqh.lock);
1072
1073         if (!ret && msg->event == UFFD_EVENT_FORK) {
1074                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1075
1076                 if (!ret) {
1077                         spin_lock(&ctx->event_wqh.lock);
1078                         if (!list_empty(&fork_event)) {
1079                                 uwq = list_first_entry(&fork_event,
1080                                                        typeof(*uwq),
1081                                                        wq.task_list);
1082                                 list_del(&uwq->wq.task_list);
1083                                 __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1084                                 userfaultfd_event_complete(ctx, uwq);
1085                         }
1086                         spin_unlock(&ctx->event_wqh.lock);
1087                 }
1088         }
1089
1090         return ret;
1091 }
1092
1093 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1094                                 size_t count, loff_t *ppos)
1095 {
1096         struct userfaultfd_ctx *ctx = file->private_data;
1097         ssize_t _ret, ret = 0;
1098         struct uffd_msg msg;
1099         int no_wait = file->f_flags & O_NONBLOCK;
1100
1101         if (ctx->state == UFFD_STATE_WAIT_API)
1102                 return -EINVAL;
1103
1104         for (;;) {
1105                 if (count < sizeof(msg))
1106                         return ret ? ret : -EINVAL;
1107                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1108                 if (_ret < 0)
1109                         return ret ? ret : _ret;
1110                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1111                         return ret ? ret : -EFAULT;
1112                 ret += sizeof(msg);
1113                 buf += sizeof(msg);
1114                 count -= sizeof(msg);
1115                 /*
1116                  * Allow to read more than one fault at time but only
1117                  * block if waiting for the very first one.
1118                  */
1119                 no_wait = O_NONBLOCK;
1120         }
1121 }
1122
1123 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1124                              struct userfaultfd_wake_range *range)
1125 {
1126         unsigned long start, end;
1127
1128         start = range->start;
1129         end = range->start + range->len;
1130
1131         spin_lock(&ctx->fault_pending_wqh.lock);
1132         /* wake all in the range and autoremove */
1133         if (waitqueue_active(&ctx->fault_pending_wqh))
1134                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1135                                      range);
1136         if (waitqueue_active(&ctx->fault_wqh))
1137                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1138         spin_unlock(&ctx->fault_pending_wqh.lock);
1139 }
1140
1141 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1142                                            struct userfaultfd_wake_range *range)
1143 {
1144         unsigned seq;
1145         bool need_wakeup;
1146
1147         /*
1148          * To be sure waitqueue_active() is not reordered by the CPU
1149          * before the pagetable update, use an explicit SMP memory
1150          * barrier here. PT lock release or up_read(mmap_sem) still
1151          * have release semantics that can allow the
1152          * waitqueue_active() to be reordered before the pte update.
1153          */
1154         smp_mb();
1155
1156         /*
1157          * Use waitqueue_active because it's very frequent to
1158          * change the address space atomically even if there are no
1159          * userfaults yet. So we take the spinlock only when we're
1160          * sure we've userfaults to wake.
1161          */
1162         do {
1163                 seq = read_seqcount_begin(&ctx->refile_seq);
1164                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1165                         waitqueue_active(&ctx->fault_wqh);
1166                 cond_resched();
1167         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1168         if (need_wakeup)
1169                 __wake_userfault(ctx, range);
1170 }
1171
1172 static __always_inline int validate_range(struct mm_struct *mm,
1173                                           __u64 start, __u64 len)
1174 {
1175         __u64 task_size = mm->task_size;
1176
1177         if (start & ~PAGE_MASK)
1178                 return -EINVAL;
1179         if (len & ~PAGE_MASK)
1180                 return -EINVAL;
1181         if (!len)
1182                 return -EINVAL;
1183         if (start < mmap_min_addr)
1184                 return -EINVAL;
1185         if (start >= task_size)
1186                 return -EINVAL;
1187         if (len > task_size - start)
1188                 return -EINVAL;
1189         return 0;
1190 }
1191
1192 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1193 {
1194         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1195                 vma_is_shmem(vma);
1196 }
1197
1198 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1199                                 unsigned long arg)
1200 {
1201         struct mm_struct *mm = ctx->mm;
1202         struct vm_area_struct *vma, *prev, *cur;
1203         int ret;
1204         struct uffdio_register uffdio_register;
1205         struct uffdio_register __user *user_uffdio_register;
1206         unsigned long vm_flags, new_flags;
1207         bool found;
1208         bool non_anon_pages;
1209         unsigned long start, end, vma_end;
1210
1211         user_uffdio_register = (struct uffdio_register __user *) arg;
1212
1213         ret = -EFAULT;
1214         if (copy_from_user(&uffdio_register, user_uffdio_register,
1215                            sizeof(uffdio_register)-sizeof(__u64)))
1216                 goto out;
1217
1218         ret = -EINVAL;
1219         if (!uffdio_register.mode)
1220                 goto out;
1221         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1222                                      UFFDIO_REGISTER_MODE_WP))
1223                 goto out;
1224         vm_flags = 0;
1225         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1226                 vm_flags |= VM_UFFD_MISSING;
1227         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1228                 vm_flags |= VM_UFFD_WP;
1229                 /*
1230                  * FIXME: remove the below error constraint by
1231                  * implementing the wprotect tracking mode.
1232                  */
1233                 ret = -EINVAL;
1234                 goto out;
1235         }
1236
1237         ret = validate_range(mm, uffdio_register.range.start,
1238                              uffdio_register.range.len);
1239         if (ret)
1240                 goto out;
1241
1242         start = uffdio_register.range.start;
1243         end = start + uffdio_register.range.len;
1244
1245         ret = -ENOMEM;
1246         if (!mmget_not_zero(mm))
1247                 goto out;
1248
1249         down_write(&mm->mmap_sem);
1250         vma = find_vma_prev(mm, start, &prev);
1251         if (!vma)
1252                 goto out_unlock;
1253
1254         /* check that there's at least one vma in the range */
1255         ret = -EINVAL;
1256         if (vma->vm_start >= end)
1257                 goto out_unlock;
1258
1259         /*
1260          * If the first vma contains huge pages, make sure start address
1261          * is aligned to huge page size.
1262          */
1263         if (is_vm_hugetlb_page(vma)) {
1264                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1265
1266                 if (start & (vma_hpagesize - 1))
1267                         goto out_unlock;
1268         }
1269
1270         /*
1271          * Search for not compatible vmas.
1272          */
1273         found = false;
1274         non_anon_pages = false;
1275         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1276                 cond_resched();
1277
1278                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1279                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1280
1281                 /* check not compatible vmas */
1282                 ret = -EINVAL;
1283                 if (!vma_can_userfault(cur))
1284                         goto out_unlock;
1285                 /*
1286                  * If this vma contains ending address, and huge pages
1287                  * check alignment.
1288                  */
1289                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1290                     end > cur->vm_start) {
1291                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1292
1293                         ret = -EINVAL;
1294
1295                         if (end & (vma_hpagesize - 1))
1296                                 goto out_unlock;
1297                 }
1298
1299                 /*
1300                  * Check that this vma isn't already owned by a
1301                  * different userfaultfd. We can't allow more than one
1302                  * userfaultfd to own a single vma simultaneously or we
1303                  * wouldn't know which one to deliver the userfaults to.
1304                  */
1305                 ret = -EBUSY;
1306                 if (cur->vm_userfaultfd_ctx.ctx &&
1307                     cur->vm_userfaultfd_ctx.ctx != ctx)
1308                         goto out_unlock;
1309
1310                 /*
1311                  * Note vmas containing huge pages
1312                  */
1313                 if (is_vm_hugetlb_page(cur) || vma_is_shmem(cur))
1314                         non_anon_pages = true;
1315
1316                 found = true;
1317         }
1318         BUG_ON(!found);
1319
1320         if (vma->vm_start < start)
1321                 prev = vma;
1322
1323         ret = 0;
1324         do {
1325                 cond_resched();
1326
1327                 BUG_ON(!vma_can_userfault(vma));
1328                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1329                        vma->vm_userfaultfd_ctx.ctx != ctx);
1330
1331                 /*
1332                  * Nothing to do: this vma is already registered into this
1333                  * userfaultfd and with the right tracking mode too.
1334                  */
1335                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1336                     (vma->vm_flags & vm_flags) == vm_flags)
1337                         goto skip;
1338
1339                 if (vma->vm_start > start)
1340                         start = vma->vm_start;
1341                 vma_end = min(end, vma->vm_end);
1342
1343                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1344                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1345                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1346                                  vma_policy(vma),
1347                                  ((struct vm_userfaultfd_ctx){ ctx }));
1348                 if (prev) {
1349                         vma = prev;
1350                         goto next;
1351                 }
1352                 if (vma->vm_start < start) {
1353                         ret = split_vma(mm, vma, start, 1);
1354                         if (ret)
1355                                 break;
1356                 }
1357                 if (vma->vm_end > end) {
1358                         ret = split_vma(mm, vma, end, 0);
1359                         if (ret)
1360                                 break;
1361                 }
1362         next:
1363                 /*
1364                  * In the vma_merge() successful mprotect-like case 8:
1365                  * the next vma was merged into the current one and
1366                  * the current one has not been updated yet.
1367                  */
1368                 vma->vm_flags = new_flags;
1369                 vma->vm_userfaultfd_ctx.ctx = ctx;
1370
1371         skip:
1372                 prev = vma;
1373                 start = vma->vm_end;
1374                 vma = vma->vm_next;
1375         } while (vma && vma->vm_start < end);
1376 out_unlock:
1377         up_write(&mm->mmap_sem);
1378         mmput(mm);
1379         if (!ret) {
1380                 /*
1381                  * Now that we scanned all vmas we can already tell
1382                  * userland which ioctls methods are guaranteed to
1383                  * succeed on this range.
1384                  */
1385                 if (put_user(non_anon_pages ? UFFD_API_RANGE_IOCTLS_BASIC :
1386                              UFFD_API_RANGE_IOCTLS,
1387                              &user_uffdio_register->ioctls))
1388                         ret = -EFAULT;
1389         }
1390 out:
1391         return ret;
1392 }
1393
1394 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1395                                   unsigned long arg)
1396 {
1397         struct mm_struct *mm = ctx->mm;
1398         struct vm_area_struct *vma, *prev, *cur;
1399         int ret;
1400         struct uffdio_range uffdio_unregister;
1401         unsigned long new_flags;
1402         bool found;
1403         unsigned long start, end, vma_end;
1404         const void __user *buf = (void __user *)arg;
1405
1406         ret = -EFAULT;
1407         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1408                 goto out;
1409
1410         ret = validate_range(mm, uffdio_unregister.start,
1411                              uffdio_unregister.len);
1412         if (ret)
1413                 goto out;
1414
1415         start = uffdio_unregister.start;
1416         end = start + uffdio_unregister.len;
1417
1418         ret = -ENOMEM;
1419         if (!mmget_not_zero(mm))
1420                 goto out;
1421
1422         down_write(&mm->mmap_sem);
1423         vma = find_vma_prev(mm, start, &prev);
1424         if (!vma)
1425                 goto out_unlock;
1426
1427         /* check that there's at least one vma in the range */
1428         ret = -EINVAL;
1429         if (vma->vm_start >= end)
1430                 goto out_unlock;
1431
1432         /*
1433          * If the first vma contains huge pages, make sure start address
1434          * is aligned to huge page size.
1435          */
1436         if (is_vm_hugetlb_page(vma)) {
1437                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1438
1439                 if (start & (vma_hpagesize - 1))
1440                         goto out_unlock;
1441         }
1442
1443         /*
1444          * Search for not compatible vmas.
1445          */
1446         found = false;
1447         ret = -EINVAL;
1448         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1449                 cond_resched();
1450
1451                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1452                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1453
1454                 /*
1455                  * Check not compatible vmas, not strictly required
1456                  * here as not compatible vmas cannot have an
1457                  * userfaultfd_ctx registered on them, but this
1458                  * provides for more strict behavior to notice
1459                  * unregistration errors.
1460                  */
1461                 if (!vma_can_userfault(cur))
1462                         goto out_unlock;
1463
1464                 found = true;
1465         }
1466         BUG_ON(!found);
1467
1468         if (vma->vm_start < start)
1469                 prev = vma;
1470
1471         ret = 0;
1472         do {
1473                 cond_resched();
1474
1475                 BUG_ON(!vma_can_userfault(vma));
1476
1477                 /*
1478                  * Nothing to do: this vma is already registered into this
1479                  * userfaultfd and with the right tracking mode too.
1480                  */
1481                 if (!vma->vm_userfaultfd_ctx.ctx)
1482                         goto skip;
1483
1484                 if (vma->vm_start > start)
1485                         start = vma->vm_start;
1486                 vma_end = min(end, vma->vm_end);
1487
1488                 if (userfaultfd_missing(vma)) {
1489                         /*
1490                          * Wake any concurrent pending userfault while
1491                          * we unregister, so they will not hang
1492                          * permanently and it avoids userland to call
1493                          * UFFDIO_WAKE explicitly.
1494                          */
1495                         struct userfaultfd_wake_range range;
1496                         range.start = start;
1497                         range.len = vma_end - start;
1498                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1499                 }
1500
1501                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1502                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1503                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1504                                  vma_policy(vma),
1505                                  NULL_VM_UFFD_CTX);
1506                 if (prev) {
1507                         vma = prev;
1508                         goto next;
1509                 }
1510                 if (vma->vm_start < start) {
1511                         ret = split_vma(mm, vma, start, 1);
1512                         if (ret)
1513                                 break;
1514                 }
1515                 if (vma->vm_end > end) {
1516                         ret = split_vma(mm, vma, end, 0);
1517                         if (ret)
1518                                 break;
1519                 }
1520         next:
1521                 /*
1522                  * In the vma_merge() successful mprotect-like case 8:
1523                  * the next vma was merged into the current one and
1524                  * the current one has not been updated yet.
1525                  */
1526                 vma->vm_flags = new_flags;
1527                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1528
1529         skip:
1530                 prev = vma;
1531                 start = vma->vm_end;
1532                 vma = vma->vm_next;
1533         } while (vma && vma->vm_start < end);
1534 out_unlock:
1535         up_write(&mm->mmap_sem);
1536         mmput(mm);
1537 out:
1538         return ret;
1539 }
1540
1541 /*
1542  * userfaultfd_wake may be used in combination with the
1543  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1544  */
1545 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1546                             unsigned long arg)
1547 {
1548         int ret;
1549         struct uffdio_range uffdio_wake;
1550         struct userfaultfd_wake_range range;
1551         const void __user *buf = (void __user *)arg;
1552
1553         ret = -EFAULT;
1554         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1555                 goto out;
1556
1557         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1558         if (ret)
1559                 goto out;
1560
1561         range.start = uffdio_wake.start;
1562         range.len = uffdio_wake.len;
1563
1564         /*
1565          * len == 0 means wake all and we don't want to wake all here,
1566          * so check it again to be sure.
1567          */
1568         VM_BUG_ON(!range.len);
1569
1570         wake_userfault(ctx, &range);
1571         ret = 0;
1572
1573 out:
1574         return ret;
1575 }
1576
1577 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1578                             unsigned long arg)
1579 {
1580         __s64 ret;
1581         struct uffdio_copy uffdio_copy;
1582         struct uffdio_copy __user *user_uffdio_copy;
1583         struct userfaultfd_wake_range range;
1584
1585         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1586
1587         ret = -EFAULT;
1588         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1589                            /* don't copy "copy" last field */
1590                            sizeof(uffdio_copy)-sizeof(__s64)))
1591                 goto out;
1592
1593         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1594         if (ret)
1595                 goto out;
1596         /*
1597          * double check for wraparound just in case. copy_from_user()
1598          * will later check uffdio_copy.src + uffdio_copy.len to fit
1599          * in the userland range.
1600          */
1601         ret = -EINVAL;
1602         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1603                 goto out;
1604         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1605                 goto out;
1606         if (mmget_not_zero(ctx->mm)) {
1607                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1608                                    uffdio_copy.len);
1609                 mmput(ctx->mm);
1610         } else {
1611                 return -ENOSPC;
1612         }
1613         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1614                 return -EFAULT;
1615         if (ret < 0)
1616                 goto out;
1617         BUG_ON(!ret);
1618         /* len == 0 would wake all */
1619         range.len = ret;
1620         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1621                 range.start = uffdio_copy.dst;
1622                 wake_userfault(ctx, &range);
1623         }
1624         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1625 out:
1626         return ret;
1627 }
1628
1629 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1630                                 unsigned long arg)
1631 {
1632         __s64 ret;
1633         struct uffdio_zeropage uffdio_zeropage;
1634         struct uffdio_zeropage __user *user_uffdio_zeropage;
1635         struct userfaultfd_wake_range range;
1636
1637         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1638
1639         ret = -EFAULT;
1640         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1641                            /* don't copy "zeropage" last field */
1642                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1643                 goto out;
1644
1645         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1646                              uffdio_zeropage.range.len);
1647         if (ret)
1648                 goto out;
1649         ret = -EINVAL;
1650         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1651                 goto out;
1652
1653         if (mmget_not_zero(ctx->mm)) {
1654                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1655                                      uffdio_zeropage.range.len);
1656                 mmput(ctx->mm);
1657         }
1658         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1659                 return -EFAULT;
1660         if (ret < 0)
1661                 goto out;
1662         /* len == 0 would wake all */
1663         BUG_ON(!ret);
1664         range.len = ret;
1665         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1666                 range.start = uffdio_zeropage.range.start;
1667                 wake_userfault(ctx, &range);
1668         }
1669         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1670 out:
1671         return ret;
1672 }
1673
1674 static inline unsigned int uffd_ctx_features(__u64 user_features)
1675 {
1676         /*
1677          * For the current set of features the bits just coincide
1678          */
1679         return (unsigned int)user_features;
1680 }
1681
1682 /*
1683  * userland asks for a certain API version and we return which bits
1684  * and ioctl commands are implemented in this kernel for such API
1685  * version or -EINVAL if unknown.
1686  */
1687 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1688                            unsigned long arg)
1689 {
1690         struct uffdio_api uffdio_api;
1691         void __user *buf = (void __user *)arg;
1692         int ret;
1693         __u64 features;
1694
1695         ret = -EINVAL;
1696         if (ctx->state != UFFD_STATE_WAIT_API)
1697                 goto out;
1698         ret = -EFAULT;
1699         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1700                 goto out;
1701         features = uffdio_api.features;
1702         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1703                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1704                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1705                         goto out;
1706                 ret = -EINVAL;
1707                 goto out;
1708         }
1709         /* report all available features and ioctls to userland */
1710         uffdio_api.features = UFFD_API_FEATURES;
1711         uffdio_api.ioctls = UFFD_API_IOCTLS;
1712         ret = -EFAULT;
1713         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1714                 goto out;
1715         ctx->state = UFFD_STATE_RUNNING;
1716         /* only enable the requested features for this uffd context */
1717         ctx->features = uffd_ctx_features(features);
1718         ret = 0;
1719 out:
1720         return ret;
1721 }
1722
1723 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1724                               unsigned long arg)
1725 {
1726         int ret = -EINVAL;
1727         struct userfaultfd_ctx *ctx = file->private_data;
1728
1729         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1730                 return -EINVAL;
1731
1732         switch(cmd) {
1733         case UFFDIO_API:
1734                 ret = userfaultfd_api(ctx, arg);
1735                 break;
1736         case UFFDIO_REGISTER:
1737                 ret = userfaultfd_register(ctx, arg);
1738                 break;
1739         case UFFDIO_UNREGISTER:
1740                 ret = userfaultfd_unregister(ctx, arg);
1741                 break;
1742         case UFFDIO_WAKE:
1743                 ret = userfaultfd_wake(ctx, arg);
1744                 break;
1745         case UFFDIO_COPY:
1746                 ret = userfaultfd_copy(ctx, arg);
1747                 break;
1748         case UFFDIO_ZEROPAGE:
1749                 ret = userfaultfd_zeropage(ctx, arg);
1750                 break;
1751         }
1752         return ret;
1753 }
1754
1755 #ifdef CONFIG_PROC_FS
1756 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1757 {
1758         struct userfaultfd_ctx *ctx = f->private_data;
1759         wait_queue_t *wq;
1760         struct userfaultfd_wait_queue *uwq;
1761         unsigned long pending = 0, total = 0;
1762
1763         spin_lock(&ctx->fault_pending_wqh.lock);
1764         list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) {
1765                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1766                 pending++;
1767                 total++;
1768         }
1769         list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
1770                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1771                 total++;
1772         }
1773         spin_unlock(&ctx->fault_pending_wqh.lock);
1774
1775         /*
1776          * If more protocols will be added, there will be all shown
1777          * separated by a space. Like this:
1778          *      protocols: aa:... bb:...
1779          */
1780         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1781                    pending, total, UFFD_API, UFFD_API_FEATURES,
1782                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1783 }
1784 #endif
1785
1786 static const struct file_operations userfaultfd_fops = {
1787 #ifdef CONFIG_PROC_FS
1788         .show_fdinfo    = userfaultfd_show_fdinfo,
1789 #endif
1790         .release        = userfaultfd_release,
1791         .poll           = userfaultfd_poll,
1792         .read           = userfaultfd_read,
1793         .unlocked_ioctl = userfaultfd_ioctl,
1794         .compat_ioctl   = userfaultfd_ioctl,
1795         .llseek         = noop_llseek,
1796 };
1797
1798 static void init_once_userfaultfd_ctx(void *mem)
1799 {
1800         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1801
1802         init_waitqueue_head(&ctx->fault_pending_wqh);
1803         init_waitqueue_head(&ctx->fault_wqh);
1804         init_waitqueue_head(&ctx->event_wqh);
1805         init_waitqueue_head(&ctx->fd_wqh);
1806         seqcount_init(&ctx->refile_seq);
1807 }
1808
1809 /**
1810  * userfaultfd_file_create - Creates an userfaultfd file pointer.
1811  * @flags: Flags for the userfaultfd file.
1812  *
1813  * This function creates an userfaultfd file pointer, w/out installing
1814  * it into the fd table. This is useful when the userfaultfd file is
1815  * used during the initialization of data structures that require
1816  * extra setup after the userfaultfd creation. So the userfaultfd
1817  * creation is split into the file pointer creation phase, and the
1818  * file descriptor installation phase.  In this way races with
1819  * userspace closing the newly installed file descriptor can be
1820  * avoided.  Returns an userfaultfd file pointer, or a proper error
1821  * pointer.
1822  */
1823 static struct file *userfaultfd_file_create(int flags)
1824 {
1825         struct file *file;
1826         struct userfaultfd_ctx *ctx;
1827
1828         BUG_ON(!current->mm);
1829
1830         /* Check the UFFD_* constants for consistency.  */
1831         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1832         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1833
1834         file = ERR_PTR(-EINVAL);
1835         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1836                 goto out;
1837
1838         file = ERR_PTR(-ENOMEM);
1839         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1840         if (!ctx)
1841                 goto out;
1842
1843         atomic_set(&ctx->refcount, 1);
1844         ctx->flags = flags;
1845         ctx->features = 0;
1846         ctx->state = UFFD_STATE_WAIT_API;
1847         ctx->released = false;
1848         ctx->mm = current->mm;
1849         /* prevent the mm struct to be freed */
1850         atomic_inc(&ctx->mm->mm_count);
1851
1852         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1853                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1854         if (IS_ERR(file)) {
1855                 mmdrop(ctx->mm);
1856                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1857         }
1858 out:
1859         return file;
1860 }
1861
1862 SYSCALL_DEFINE1(userfaultfd, int, flags)
1863 {
1864         int fd, error;
1865         struct file *file;
1866
1867         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1868         if (error < 0)
1869                 return error;
1870         fd = error;
1871
1872         file = userfaultfd_file_create(flags);
1873         if (IS_ERR(file)) {
1874                 error = PTR_ERR(file);
1875                 goto err_put_unused_fd;
1876         }
1877         fd_install(fd, file);
1878
1879         return fd;
1880
1881 err_put_unused_fd:
1882         put_unused_fd(fd);
1883
1884         return error;
1885 }
1886
1887 static int __init userfaultfd_init(void)
1888 {
1889         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1890                                                 sizeof(struct userfaultfd_ctx),
1891                                                 0,
1892                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1893                                                 init_once_userfaultfd_ctx);
1894         return 0;
1895 }
1896 __initcall(userfaultfd_init);