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