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