fs/userfaultfd.c

   1 /*
   2  *  fs/userfaultfd.c
   3  *
   4  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc.
   6  *  Copyright (C) 2015  Red Hat, Inc.
   7  *
   8  *  This work is licensed under the terms of the GNU GPL, version 2. See
   9  *  the COPYING file in the top-level directory.
  10  *
  11  *  Some part derived from fs/eventfd.c (anon inode setup) and
  12  *  mm/ksm.c (mm hashing).
  13  */
  14
  15 #include <linux/list.h>
  16 #include <linux/hashtable.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/mm.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  34
  35 enum userfaultfd_state {
  36         UFFD_STATE_WAIT_API,
  37         UFFD_STATE_RUNNING,
  38 };
  39
  40 /*
  41  * Start with fault_pending_wqh and fault_wqh so they're more likely
  42  * to be in the same cacheline.
  43  */
  44 struct userfaultfd_ctx {
  45         /* waitqueue head for the pending (i.e. not read) userfaults */
  46         wait_queue_head_t fault_pending_wqh;
  47         /* waitqueue head for the userfaults */
  48         wait_queue_head_t fault_wqh;
  49         /* waitqueue head for the pseudo fd to wakeup poll/read */
  50         wait_queue_head_t fd_wqh;
  51         /* waitqueue head for events */
  52         wait_queue_head_t event_wqh;
  53         /* a refile sequence protected by fault_pending_wqh lock */
  54         struct seqcount refile_seq;
  55         /* pseudo fd refcounting */
  56         atomic_t refcount;
  57         /* userfaultfd syscall flags */
  58         unsigned int flags;
  59         /* features requested from the userspace */
  60         unsigned int features;
  61         /* state machine */
  62         enum userfaultfd_state state;
  63         /* released */
  64         bool released;
  65         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  66         struct mm_struct *mm;
  67 };
  68
  69 struct userfaultfd_fork_ctx {
  70         struct userfaultfd_ctx *orig;
  71         struct userfaultfd_ctx *new;
  72         struct list_head list;
  73 };
  74
  75 struct userfaultfd_unmap_ctx {
  76         struct userfaultfd_ctx *ctx;
  77         unsigned long start;
  78         unsigned long end;
  79         struct list_head list;
  80 };
  81
  82 struct userfaultfd_wait_queue {
  83         struct uffd_msg msg;
  84         wait_queue_entry_t wq;
  85         struct userfaultfd_ctx *ctx;
  86         bool waken;
  87 };
  88
  89 struct userfaultfd_wake_range {
  90         unsigned long start;
  91         unsigned long len;
  92 };
  93
  94 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
  95                                      int wake_flags, void *key)
  96 {
  97         struct userfaultfd_wake_range *range = key;
  98         int ret;
  99         struct userfaultfd_wait_queue *uwq;
 100         unsigned long start, len;
 101
 102         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 103         ret = 0;
 104         /* len == 0 means wake all */
 105         start = range->start;
 106         len = range->len;
 107         if (len && (start > uwq->msg.arg.pagefault.address ||
 108                     start + len <= uwq->msg.arg.pagefault.address))
 109                 goto out;
 110         WRITE_ONCE(uwq->waken, true);
 111         /*
 112          * The Program-Order guarantees provided by the scheduler
 113          * ensure uwq->waken is visible before the task is woken.
 114          */
 115         ret = wake_up_state(wq->private, mode);
 116         if (ret) {
 117                 /*
 118                  * Wake only once, autoremove behavior.
 119                  *
 120                  * After the effect of list_del_init is visible to the other
 121                  * CPUs, the waitqueue may disappear from under us, see the
 122                  * !list_empty_careful() in handle_userfault().
 123                  *
 124                  * try_to_wake_up() has an implicit smp_mb(), and the
 125                  * wq->private is read before calling the extern function
 126                  * "wake_up_state" (which in turns calls try_to_wake_up).
 127                  */
 128                 list_del_init(&wq->entry);
 129         }
 130 out:
 131         return ret;
 132 }
 133
 134 /**
 135  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 136  * context.
 137  * @ctx: [in] Pointer to the userfaultfd context.
 138  */
 139 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 140 {
 141         if (!atomic_inc_not_zero(&ctx->refcount))
 142                 BUG();
 143 }
 144
 145 /**
 146  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 147  * context.
 148  * @ctx: [in] Pointer to userfaultfd context.
 149  *
 150  * The userfaultfd context reference must have been previously acquired either
 151  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 152  */
 153 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 154 {
 155         if (atomic_dec_and_test(&ctx->refcount)) {
 156                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 157                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 158                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 159                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 160                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 161                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 162                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 163                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 164                 mmdrop(ctx->mm);
 165                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 166         }
 167 }
 168
 169 static inline void msg_init(struct uffd_msg *msg)
 170 {
 171         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 172         /*
 173          * Must use memset to zero out the paddings or kernel data is
 174          * leaked to userland.
 175          */
 176         memset(msg, 0, sizeof(struct uffd_msg));
 177 }
 178
 179 static inline struct uffd_msg userfault_msg(unsigned long address,
 180                                             unsigned int flags,
 181                                             unsigned long reason,
 182                                             unsigned int features)
 183 {
 184         struct uffd_msg msg;
 185         msg_init(&msg);
 186         msg.event = UFFD_EVENT_PAGEFAULT;
 187         msg.arg.pagefault.address = address;
 188         if (flags & FAULT_FLAG_WRITE)
 189                 /*
 190                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 191                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 192                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 193                  * was a read fault, otherwise if set it means it's
 194                  * a write fault.
 195                  */
 196                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 197         if (reason & VM_UFFD_WP)
 198                 /*
 199                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 200                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 201                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 202                  * a missing fault, otherwise if set it means it's a
 203                  * write protect fault.
 204                  */
 205                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 206         if (features & UFFD_FEATURE_THREAD_ID)
 207                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 208         return msg;
 209 }
 210
 211 #ifdef CONFIG_HUGETLB_PAGE
 212 /*
 213  * Same functionality as userfaultfd_must_wait below with modifications for
 214  * hugepmd ranges.
 215  */
 216 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 217                                          struct vm_area_struct *vma,
 218                                          unsigned long address,
 219                                          unsigned long flags,
 220                                          unsigned long reason)
 221 {
 222         struct mm_struct *mm = ctx->mm;
 223         pte_t *ptep, pte;
 224         bool ret = true;
 225
 226         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 227
 228         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 229
 230         if (!ptep)
 231                 goto out;
 232
 233         ret = false;
 234         pte = huge_ptep_get(ptep);
 235
 236         /*
 237          * Lockless access: we're in a wait_event so it's ok if it
 238          * changes under us.
 239          */
 240         if (huge_pte_none(pte))
 241                 ret = true;
 242         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 243                 ret = true;
 244 out:
 245         return ret;
 246 }
 247 #else
 248 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 249                                          struct vm_area_struct *vma,
 250                                          unsigned long address,
 251                                          unsigned long flags,
 252                                          unsigned long reason)
 253 {
 254         return false;   /* should never get here */
 255 }
 256 #endif /* CONFIG_HUGETLB_PAGE */
 257
 258 /*
 259  * Verify the pagetables are still not ok after having reigstered into
 260  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 261  * userfault that has already been resolved, if userfaultfd_read and
 262  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 263  * threads.
 264  */
 265 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 266                                          unsigned long address,
 267                                          unsigned long flags,
 268                                          unsigned long reason)
 269 {
 270         struct mm_struct *mm = ctx->mm;
 271         pgd_t *pgd;
 272         p4d_t *p4d;
 273         pud_t *pud;
 274         pmd_t *pmd, _pmd;
 275         pte_t *pte;
 276         bool ret = true;
 277
 278         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 279
 280         pgd = pgd_offset(mm, address);
 281         if (!pgd_present(*pgd))
 282                 goto out;
 283         p4d = p4d_offset(pgd, address);
 284         if (!p4d_present(*p4d))
 285                 goto out;
 286         pud = pud_offset(p4d, address);
 287         if (!pud_present(*pud))
 288                 goto out;
 289         pmd = pmd_offset(pud, address);
 290         /*
 291          * READ_ONCE must function as a barrier with narrower scope
 292          * and it must be equivalent to:
 293          *      _pmd = *pmd; barrier();
 294          *
 295          * This is to deal with the instability (as in
 296          * pmd_trans_unstable) of the pmd.
 297          */
 298         _pmd = READ_ONCE(*pmd);
 299         if (!pmd_present(_pmd))
 300                 goto out;
 301
 302         ret = false;
 303         if (pmd_trans_huge(_pmd))
 304                 goto out;
 305
 306         /*
 307          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 308          * and use the standard pte_offset_map() instead of parsing _pmd.
 309          */
 310         pte = pte_offset_map(pmd, address);
 311         /*
 312          * Lockless access: we're in a wait_event so it's ok if it
 313          * changes under us.
 314          */
 315         if (pte_none(*pte))
 316                 ret = true;
 317         pte_unmap(pte);
 318
 319 out:
 320         return ret;
 321 }
 322
 323 /*
 324  * The locking rules involved in returning VM_FAULT_RETRY depending on
 325  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 326  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 327  * recommendation in __lock_page_or_retry is not an understatement.
 328  *
 329  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
 330  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 331  * not set.
 332  *
 333  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 334  * set, VM_FAULT_RETRY can still be returned if and only if there are
 335  * fatal_signal_pending()s, and the mmap_sem must be released before
 336  * returning it.
 337  */
 338 int handle_userfault(struct vm_fault *vmf, unsigned long reason)
 339 {
 340         struct mm_struct *mm = vmf->vma->vm_mm;
 341         struct userfaultfd_ctx *ctx;
 342         struct userfaultfd_wait_queue uwq;
 343         int ret;
 344         bool must_wait, return_to_userland;
 345         long blocking_state;
 346
 347         ret = VM_FAULT_SIGBUS;
 348
 349         /*
 350          * We don't do userfault handling for the final child pid update.
 351          *
 352          * We also don't do userfault handling during
 353          * coredumping. hugetlbfs has the special
 354          * follow_hugetlb_page() to skip missing pages in the
 355          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 356          * the no_page_table() helper in follow_page_mask(), but the
 357          * shmem_vm_ops->fault method is invoked even during
 358          * coredumping without mmap_sem and it ends up here.
 359          */
 360         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 361                 goto out;
 362
 363         /*
 364          * Coredumping runs without mmap_sem so we can only check that
 365          * the mmap_sem is held, if PF_DUMPCORE was not set.
 366          */
 367         WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
 368
 369         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 370         if (!ctx)
 371                 goto out;
 372
 373         BUG_ON(ctx->mm != mm);
 374
 375         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 376         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 377
 378         if (ctx->features & UFFD_FEATURE_SIGBUS)
 379                 goto out;
 380
 381         /*
 382          * If it's already released don't get it. This avoids to loop
 383          * in __get_user_pages if userfaultfd_release waits on the
 384          * caller of handle_userfault to release the mmap_sem.
 385          */
 386         if (unlikely(ACCESS_ONCE(ctx->released))) {
 387                 /*
 388                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 389                  * cooperative manager can close the uffd after the
 390                  * last UFFDIO_COPY, without risking to trigger an
 391                  * involuntary SIGBUS if the process was starting the
 392                  * userfaultfd while the userfaultfd was still armed
 393                  * (but after the last UFFDIO_COPY). If the uffd
 394                  * wasn't already closed when the userfault reached
 395                  * this point, that would normally be solved by
 396                  * userfaultfd_must_wait returning 'false'.
 397                  *
 398                  * If we were to return VM_FAULT_SIGBUS here, the non
 399                  * cooperative manager would be instead forced to
 400                  * always call UFFDIO_UNREGISTER before it can safely
 401                  * close the uffd.
 402                  */
 403                 ret = VM_FAULT_NOPAGE;
 404                 goto out;
 405         }
 406
 407         /*
 408          * Check that we can return VM_FAULT_RETRY.
 409          *
 410          * NOTE: it should become possible to return VM_FAULT_RETRY
 411          * even if FAULT_FLAG_TRIED is set without leading to gup()
 412          * -EBUSY failures, if the userfaultfd is to be extended for
 413          * VM_UFFD_WP tracking and we intend to arm the userfault
 414          * without first stopping userland access to the memory. For
 415          * VM_UFFD_MISSING userfaults this is enough for now.
 416          */
 417         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 418                 /*
 419                  * Validate the invariant that nowait must allow retry
 420                  * to be sure not to return SIGBUS erroneously on
 421                  * nowait invocations.
 422                  */
 423                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 424 #ifdef CONFIG_DEBUG_VM
 425                 if (printk_ratelimit()) {
 426                         printk(KERN_WARNING
 427                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 428                                vmf->flags);
 429                         dump_stack();
 430                 }
 431 #endif
 432                 goto out;
 433         }
 434
 435         /*
 436          * Handle nowait, not much to do other than tell it to retry
 437          * and wait.
 438          */
 439         ret = VM_FAULT_RETRY;
 440         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 441                 goto out;
 442
 443         /* take the reference before dropping the mmap_sem */
 444         userfaultfd_ctx_get(ctx);
 445
 446         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 447         uwq.wq.private = current;
 448         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 449                         ctx->features);
 450         uwq.ctx = ctx;
 451         uwq.waken = false;
 452
 453         return_to_userland =
 454                 (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
 455                 (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
 456         blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
 457                          TASK_KILLABLE;
 458
 459         spin_lock(&ctx->fault_pending_wqh.lock);
 460         /*
 461          * After the __add_wait_queue the uwq is visible to userland
 462          * through poll/read().
 463          */
 464         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 465         /*
 466          * The smp_mb() after __set_current_state prevents the reads
 467          * following the spin_unlock to happen before the list_add in
 468          * __add_wait_queue.
 469          */
 470         set_current_state(blocking_state);
 471         spin_unlock(&ctx->fault_pending_wqh.lock);
 472
 473         if (!is_vm_hugetlb_page(vmf->vma))
 474                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 475                                                   reason);
 476         else
 477                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 478                                                        vmf->address,
 479                                                        vmf->flags, reason);
 480         up_read(&mm->mmap_sem);
 481
 482         if (likely(must_wait && !ACCESS_ONCE(ctx->released) &&
 483                    (return_to_userland ? !signal_pending(current) :
 484                     !fatal_signal_pending(current)))) {
 485                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 486                 schedule();
 487                 ret |= VM_FAULT_MAJOR;
 488
 489                 /*
 490                  * False wakeups can orginate even from rwsem before
 491                  * up_read() however userfaults will wait either for a
 492                  * targeted wakeup on the specific uwq waitqueue from
 493                  * wake_userfault() or for signals or for uffd
 494                  * release.
 495                  */
 496                 while (!READ_ONCE(uwq.waken)) {
 497                         /*
 498                          * This needs the full smp_store_mb()
 499                          * guarantee as the state write must be
 500                          * visible to other CPUs before reading
 501                          * uwq.waken from other CPUs.
 502                          */
 503                         set_current_state(blocking_state);
 504                         if (READ_ONCE(uwq.waken) ||
 505                             READ_ONCE(ctx->released) ||
 506                             (return_to_userland ? signal_pending(current) :
 507                              fatal_signal_pending(current)))
 508                                 break;
 509                         schedule();
 510                 }
 511         }
 512
 513         __set_current_state(TASK_RUNNING);
 514
 515         if (return_to_userland) {
 516                 if (signal_pending(current) &&
 517                     !fatal_signal_pending(current)) {
 518                         /*
 519                          * If we got a SIGSTOP or SIGCONT and this is
 520                          * a normal userland page fault, just let
 521                          * userland return so the signal will be
 522                          * handled and gdb debugging works.  The page
 523                          * fault code immediately after we return from
 524                          * this function is going to release the
 525                          * mmap_sem and it's not depending on it
 526                          * (unlike gup would if we were not to return
 527                          * VM_FAULT_RETRY).
 528                          *
 529                          * If a fatal signal is pending we still take
 530                          * the streamlined VM_FAULT_RETRY failure path
 531                          * and there's no need to retake the mmap_sem
 532                          * in such case.
 533                          */
 534                         down_read(&mm->mmap_sem);
 535                         ret = VM_FAULT_NOPAGE;
 536                 }
 537         }
 538
 539         /*
 540          * Here we race with the list_del; list_add in
 541          * userfaultfd_ctx_read(), however because we don't ever run
 542          * list_del_init() to refile across the two lists, the prev
 543          * and next pointers will never point to self. list_add also
 544          * would never let any of the two pointers to point to
 545          * self. So list_empty_careful won't risk to see both pointers
 546          * pointing to self at any time during the list refile. The
 547          * only case where list_del_init() is called is the full
 548          * removal in the wake function and there we don't re-list_add
 549          * and it's fine not to block on the spinlock. The uwq on this
 550          * kernel stack can be released after the list_del_init.
 551          */
 552         if (!list_empty_careful(&uwq.wq.entry)) {
 553                 spin_lock(&ctx->fault_pending_wqh.lock);
 554                 /*
 555                  * No need of list_del_init(), the uwq on the stack
 556                  * will be freed shortly anyway.
 557                  */
 558                 list_del(&uwq.wq.entry);
 559                 spin_unlock(&ctx->fault_pending_wqh.lock);
 560         }
 561
 562         /*
 563          * ctx may go away after this if the userfault pseudo fd is
 564          * already released.
 565          */
 566         userfaultfd_ctx_put(ctx);
 567
 568 out:
 569         return ret;
 570 }
 571
 572 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 573                                               struct userfaultfd_wait_queue *ewq)
 574 {
 575         struct userfaultfd_ctx *release_new_ctx;
 576
 577         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 578                 goto out;
 579
 580         ewq->ctx = ctx;
 581         init_waitqueue_entry(&ewq->wq, current);
 582         release_new_ctx = NULL;
 583
 584         spin_lock(&ctx->event_wqh.lock);
 585         /*
 586          * After the __add_wait_queue the uwq is visible to userland
 587          * through poll/read().
 588          */
 589         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 590         for (;;) {
 591                 set_current_state(TASK_KILLABLE);
 592                 if (ewq->msg.event == 0)
 593                         break;
 594                 if (ACCESS_ONCE(ctx->released) ||
 595                     fatal_signal_pending(current)) {
 596                         /*
 597                          * &ewq->wq may be queued in fork_event, but
 598                          * __remove_wait_queue ignores the head
 599                          * parameter. It would be a problem if it
 600                          * didn't.
 601                          */
 602                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 603                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 604                                 struct userfaultfd_ctx *new;
 605
 606                                 new = (struct userfaultfd_ctx *)
 607                                         (unsigned long)
 608                                         ewq->msg.arg.reserved.reserved1;
 609                                 release_new_ctx = new;
 610                         }
 611                         break;
 612                 }
 613
 614                 spin_unlock(&ctx->event_wqh.lock);
 615
 616                 wake_up_poll(&ctx->fd_wqh, POLLIN);
 617                 schedule();
 618
 619                 spin_lock(&ctx->event_wqh.lock);
 620         }
 621         __set_current_state(TASK_RUNNING);
 622         spin_unlock(&ctx->event_wqh.lock);
 623
 624         if (release_new_ctx) {
 625                 struct vm_area_struct *vma;
 626                 struct mm_struct *mm = release_new_ctx->mm;
 627
 628                 /* the various vma->vm_userfaultfd_ctx still points to it */
 629                 down_write(&mm->mmap_sem);
 630                 /* no task can run (and in turn coredump) yet */
 631                 VM_WARN_ON(!mmget_still_valid(mm));
 632                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 633                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 634                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 635                                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 636                         }
 637                 up_write(&mm->mmap_sem);
 638
 639                 userfaultfd_ctx_put(release_new_ctx);
 640         }
 641
 642         /*
 643          * ctx may go away after this if the userfault pseudo fd is
 644          * already released.
 645          */
 646 out:
 647         userfaultfd_ctx_put(ctx);
 648 }
 649
 650 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 651                                        struct userfaultfd_wait_queue *ewq)
 652 {
 653         ewq->msg.event = 0;
 654         wake_up_locked(&ctx->event_wqh);
 655         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 656 }
 657
 658 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 659 {
 660         struct userfaultfd_ctx *ctx = NULL, *octx;
 661         struct userfaultfd_fork_ctx *fctx;
 662
 663         octx = vma->vm_userfaultfd_ctx.ctx;
 664         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 665                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 666                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 667                 return 0;
 668         }
 669
 670         list_for_each_entry(fctx, fcs, list)
 671                 if (fctx->orig == octx) {
 672                         ctx = fctx->new;
 673                         break;
 674                 }
 675
 676         if (!ctx) {
 677                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 678                 if (!fctx)
 679                         return -ENOMEM;
 680
 681                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 682                 if (!ctx) {
 683                         kfree(fctx);
 684                         return -ENOMEM;
 685                 }
 686
 687                 atomic_set(&ctx->refcount, 1);
 688                 ctx->flags = octx->flags;
 689                 ctx->state = UFFD_STATE_RUNNING;
 690                 ctx->features = octx->features;
 691                 ctx->released = false;
 692                 ctx->mm = vma->vm_mm;
 693                 atomic_inc(&ctx->mm->mm_count);
 694
 695                 userfaultfd_ctx_get(octx);
 696                 fctx->orig = octx;
 697                 fctx->new = ctx;
 698                 list_add_tail(&fctx->list, fcs);
 699         }
 700
 701         vma->vm_userfaultfd_ctx.ctx = ctx;
 702         return 0;
 703 }
 704
 705 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 706 {
 707         struct userfaultfd_ctx *ctx = fctx->orig;
 708         struct userfaultfd_wait_queue ewq;
 709
 710         msg_init(&ewq.msg);
 711
 712         ewq.msg.event = UFFD_EVENT_FORK;
 713         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 714
 715         userfaultfd_event_wait_completion(ctx, &ewq);
 716 }
 717
 718 void dup_userfaultfd_complete(struct list_head *fcs)
 719 {
 720         struct userfaultfd_fork_ctx *fctx, *n;
 721
 722         list_for_each_entry_safe(fctx, n, fcs, list) {
 723                 dup_fctx(fctx);
 724                 list_del(&fctx->list);
 725                 kfree(fctx);
 726         }
 727 }
 728
 729 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 730                              struct vm_userfaultfd_ctx *vm_ctx)
 731 {
 732         struct userfaultfd_ctx *ctx;
 733
 734         ctx = vma->vm_userfaultfd_ctx.ctx;
 735         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 736                 vm_ctx->ctx = ctx;
 737                 userfaultfd_ctx_get(ctx);
 738         }
 739 }
 740
 741 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 742                                  unsigned long from, unsigned long to,
 743                                  unsigned long len)
 744 {
 745         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 746         struct userfaultfd_wait_queue ewq;
 747
 748         if (!ctx)
 749                 return;
 750
 751         if (to & ~PAGE_MASK) {
 752                 userfaultfd_ctx_put(ctx);
 753                 return;
 754         }
 755
 756         msg_init(&ewq.msg);
 757
 758         ewq.msg.event = UFFD_EVENT_REMAP;
 759         ewq.msg.arg.remap.from = from;
 760         ewq.msg.arg.remap.to = to;
 761         ewq.msg.arg.remap.len = len;
 762
 763         userfaultfd_event_wait_completion(ctx, &ewq);
 764 }
 765
 766 bool userfaultfd_remove(struct vm_area_struct *vma,
 767                         unsigned long start, unsigned long end)
 768 {
 769         struct mm_struct *mm = vma->vm_mm;
 770         struct userfaultfd_ctx *ctx;
 771         struct userfaultfd_wait_queue ewq;
 772
 773         ctx = vma->vm_userfaultfd_ctx.ctx;
 774         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 775                 return true;
 776
 777         userfaultfd_ctx_get(ctx);
 778         up_read(&mm->mmap_sem);
 779
 780         msg_init(&ewq.msg);
 781
 782         ewq.msg.event = UFFD_EVENT_REMOVE;
 783         ewq.msg.arg.remove.start = start;
 784         ewq.msg.arg.remove.end = end;
 785
 786         userfaultfd_event_wait_completion(ctx, &ewq);
 787
 788         return false;
 789 }
 790
 791 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 792                           unsigned long start, unsigned long end)
 793 {
 794         struct userfaultfd_unmap_ctx *unmap_ctx;
 795
 796         list_for_each_entry(unmap_ctx, unmaps, list)
 797                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 798                     unmap_ctx->end == end)
 799                         return true;
 800
 801         return false;
 802 }
 803
 804 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 805                            unsigned long start, unsigned long end,
 806                            struct list_head *unmaps)
 807 {
 808         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 809                 struct userfaultfd_unmap_ctx *unmap_ctx;
 810                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 811
 812                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 813                     has_unmap_ctx(ctx, unmaps, start, end))
 814                         continue;
 815
 816                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 817                 if (!unmap_ctx)
 818                         return -ENOMEM;
 819
 820                 userfaultfd_ctx_get(ctx);
 821                 unmap_ctx->ctx = ctx;
 822                 unmap_ctx->start = start;
 823                 unmap_ctx->end = end;
 824                 list_add_tail(&unmap_ctx->list, unmaps);
 825         }
 826
 827         return 0;
 828 }
 829
 830 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 831 {
 832         struct userfaultfd_unmap_ctx *ctx, *n;
 833         struct userfaultfd_wait_queue ewq;
 834
 835         list_for_each_entry_safe(ctx, n, uf, list) {
 836                 msg_init(&ewq.msg);
 837
 838                 ewq.msg.event = UFFD_EVENT_UNMAP;
 839                 ewq.msg.arg.remove.start = ctx->start;
 840                 ewq.msg.arg.remove.end = ctx->end;
 841
 842                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 843
 844                 list_del(&ctx->list);
 845                 kfree(ctx);
 846         }
 847 }
 848
 849 static int userfaultfd_release(struct inode *inode, struct file *file)
 850 {
 851         struct userfaultfd_ctx *ctx = file->private_data;
 852         struct mm_struct *mm = ctx->mm;
 853         struct vm_area_struct *vma, *prev;
 854         /* len == 0 means wake all */
 855         struct userfaultfd_wake_range range = { .len = 0, };
 856         unsigned long new_flags;
 857         bool still_valid;
 858
 859         ACCESS_ONCE(ctx->released) = true;
 860
 861         if (!mmget_not_zero(mm))
 862                 goto wakeup;
 863
 864         /*
 865          * Flush page faults out of all CPUs. NOTE: all page faults
 866          * must be retried without returning VM_FAULT_SIGBUS if
 867          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 868          * changes while handle_userfault released the mmap_sem. So
 869          * it's critical that released is set to true (above), before
 870          * taking the mmap_sem for writing.
 871          */
 872         down_write(&mm->mmap_sem);
 873         still_valid = mmget_still_valid(mm);
 874         prev = NULL;
 875         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 876                 cond_resched();
 877                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 878                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 879                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 880                         prev = vma;
 881                         continue;
 882                 }
 883                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 884                 if (still_valid) {
 885                         prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 886                                          new_flags, vma->anon_vma,
 887                                          vma->vm_file, vma->vm_pgoff,
 888                                          vma_policy(vma),
 889                                          NULL_VM_UFFD_CTX);
 890                         if (prev)
 891                                 vma = prev;
 892                         else
 893                                 prev = vma;
 894                 }
 895                 vma->vm_flags = new_flags;
 896                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 897         }
 898         up_write(&mm->mmap_sem);
 899         mmput(mm);
 900 wakeup:
 901         /*
 902          * After no new page faults can wait on this fault_*wqh, flush
 903          * the last page faults that may have been already waiting on
 904          * the fault_*wqh.
 905          */
 906         spin_lock(&ctx->fault_pending_wqh.lock);
 907         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 908         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 909         spin_unlock(&ctx->fault_pending_wqh.lock);
 910
 911         /* Flush pending events that may still wait on event_wqh */
 912         wake_up_all(&ctx->event_wqh);
 913
 914         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 915         userfaultfd_ctx_put(ctx);
 916         return 0;
 917 }
 918
 919 /* fault_pending_wqh.lock must be hold by the caller */
 920 static inline struct userfaultfd_wait_queue *find_userfault_in(
 921                 wait_queue_head_t *wqh)
 922 {
 923         wait_queue_entry_t *wq;
 924         struct userfaultfd_wait_queue *uwq;
 925
 926         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 927
 928         uwq = NULL;
 929         if (!waitqueue_active(wqh))
 930                 goto out;
 931         /* walk in reverse to provide FIFO behavior to read userfaults */
 932         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 933         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 934 out:
 935         return uwq;
 936 }
 937
 938 static inline struct userfaultfd_wait_queue *find_userfault(
 939                 struct userfaultfd_ctx *ctx)
 940 {
 941         return find_userfault_in(&ctx->fault_pending_wqh);
 942 }
 943
 944 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 945                 struct userfaultfd_ctx *ctx)
 946 {
 947         return find_userfault_in(&ctx->event_wqh);
 948 }
 949
 950 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 951 {
 952         struct userfaultfd_ctx *ctx = file->private_data;
 953         unsigned int ret;
 954
 955         poll_wait(file, &ctx->fd_wqh, wait);
 956
 957         switch (ctx->state) {
 958         case UFFD_STATE_WAIT_API:
 959                 return POLLERR;
 960         case UFFD_STATE_RUNNING:
 961                 /*
 962                  * poll() never guarantees that read won't block.
 963                  * userfaults can be waken before they're read().
 964                  */
 965                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 966                         return POLLERR;
 967                 /*
 968                  * lockless access to see if there are pending faults
 969                  * __pollwait last action is the add_wait_queue but
 970                  * the spin_unlock would allow the waitqueue_active to
 971                  * pass above the actual list_add inside
 972                  * add_wait_queue critical section. So use a full
 973                  * memory barrier to serialize the list_add write of
 974                  * add_wait_queue() with the waitqueue_active read
 975                  * below.
 976                  */
 977                 ret = 0;
 978                 smp_mb();
 979                 if (waitqueue_active(&ctx->fault_pending_wqh))
 980                         ret = POLLIN;
 981                 else if (waitqueue_active(&ctx->event_wqh))
 982                         ret = POLLIN;
 983
 984                 return ret;
 985         default:
 986                 WARN_ON_ONCE(1);
 987                 return POLLERR;
 988         }
 989 }
 990
 991 static const struct file_operations userfaultfd_fops;
 992
 993 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 994                                   struct userfaultfd_ctx *new,
 995                                   struct uffd_msg *msg)
 996 {
 997         int fd;
 998         struct file *file;
 999         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
1000
1001         fd = get_unused_fd_flags(flags);
1002         if (fd < 0)
1003                 return fd;
1004
1005         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
1006                                   O_RDWR | flags);
1007         if (IS_ERR(file)) {
1008                 put_unused_fd(fd);
1009                 return PTR_ERR(file);
1010         }
1011
1012         fd_install(fd, file);
1013         msg->arg.reserved.reserved1 = 0;
1014         msg->arg.fork.ufd = fd;
1015
1016         return 0;
1017 }
1018
1019 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1020                                     struct uffd_msg *msg)
1021 {
1022         ssize_t ret;
1023         DECLARE_WAITQUEUE(wait, current);
1024         struct userfaultfd_wait_queue *uwq;
1025         /*
1026          * Handling fork event requires sleeping operations, so
1027          * we drop the event_wqh lock, then do these ops, then
1028          * lock it back and wake up the waiter. While the lock is
1029          * dropped the ewq may go away so we keep track of it
1030          * carefully.
1031          */
1032         LIST_HEAD(fork_event);
1033         struct userfaultfd_ctx *fork_nctx = NULL;
1034
1035         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1036         spin_lock(&ctx->fd_wqh.lock);
1037         __add_wait_queue(&ctx->fd_wqh, &wait);
1038         for (;;) {
1039                 set_current_state(TASK_INTERRUPTIBLE);
1040                 spin_lock(&ctx->fault_pending_wqh.lock);
1041                 uwq = find_userfault(ctx);
1042                 if (uwq) {
1043                         /*
1044                          * Use a seqcount to repeat the lockless check
1045                          * in wake_userfault() to avoid missing
1046                          * wakeups because during the refile both
1047                          * waitqueue could become empty if this is the
1048                          * only userfault.
1049                          */
1050                         write_seqcount_begin(&ctx->refile_seq);
1051
1052                         /*
1053                          * The fault_pending_wqh.lock prevents the uwq
1054                          * to disappear from under us.
1055                          *
1056                          * Refile this userfault from
1057                          * fault_pending_wqh to fault_wqh, it's not
1058                          * pending anymore after we read it.
1059                          *
1060                          * Use list_del() by hand (as
1061                          * userfaultfd_wake_function also uses
1062                          * list_del_init() by hand) to be sure nobody
1063                          * changes __remove_wait_queue() to use
1064                          * list_del_init() in turn breaking the
1065                          * !list_empty_careful() check in
1066                          * handle_userfault(). The uwq->wq.head list
1067                          * must never be empty at any time during the
1068                          * refile, or the waitqueue could disappear
1069                          * from under us. The "wait_queue_head_t"
1070                          * parameter of __remove_wait_queue() is unused
1071                          * anyway.
1072                          */
1073                         list_del(&uwq->wq.entry);
1074                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1075
1076                         write_seqcount_end(&ctx->refile_seq);
1077
1078                         /* careful to always initialize msg if ret == 0 */
1079                         *msg = uwq->msg;
1080                         spin_unlock(&ctx->fault_pending_wqh.lock);
1081                         ret = 0;
1082                         break;
1083                 }
1084                 spin_unlock(&ctx->fault_pending_wqh.lock);
1085
1086                 spin_lock(&ctx->event_wqh.lock);
1087                 uwq = find_userfault_evt(ctx);
1088                 if (uwq) {
1089                         *msg = uwq->msg;
1090
1091                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1092                                 fork_nctx = (struct userfaultfd_ctx *)
1093                                         (unsigned long)
1094                                         uwq->msg.arg.reserved.reserved1;
1095                                 list_move(&uwq->wq.entry, &fork_event);
1096                                 /*
1097                                  * fork_nctx can be freed as soon as
1098                                  * we drop the lock, unless we take a
1099                                  * reference on it.
1100                                  */
1101                                 userfaultfd_ctx_get(fork_nctx);
1102                                 spin_unlock(&ctx->event_wqh.lock);
1103                                 ret = 0;
1104                                 break;
1105                         }
1106
1107                         userfaultfd_event_complete(ctx, uwq);
1108                         spin_unlock(&ctx->event_wqh.lock);
1109                         ret = 0;
1110                         break;
1111                 }
1112                 spin_unlock(&ctx->event_wqh.lock);
1113
1114                 if (signal_pending(current)) {
1115                         ret = -ERESTARTSYS;
1116                         break;
1117                 }
1118                 if (no_wait) {
1119                         ret = -EAGAIN;
1120                         break;
1121                 }
1122                 spin_unlock(&ctx->fd_wqh.lock);
1123                 schedule();
1124                 spin_lock(&ctx->fd_wqh.lock);
1125         }
1126         __remove_wait_queue(&ctx->fd_wqh, &wait);
1127         __set_current_state(TASK_RUNNING);
1128         spin_unlock(&ctx->fd_wqh.lock);
1129
1130         if (!ret && msg->event == UFFD_EVENT_FORK) {
1131                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1132                 spin_lock(&ctx->event_wqh.lock);
1133                 if (!list_empty(&fork_event)) {
1134                         /*
1135                          * The fork thread didn't abort, so we can
1136                          * drop the temporary refcount.
1137                          */
1138                         userfaultfd_ctx_put(fork_nctx);
1139
1140                         uwq = list_first_entry(&fork_event,
1141                                                typeof(*uwq),
1142                                                wq.entry);
1143                         /*
1144                          * If fork_event list wasn't empty and in turn
1145                          * the event wasn't already released by fork
1146                          * (the event is allocated on fork kernel
1147                          * stack), put the event back to its place in
1148                          * the event_wq. fork_event head will be freed
1149                          * as soon as we return so the event cannot
1150                          * stay queued there no matter the current
1151                          * "ret" value.
1152                          */
1153                         list_del(&uwq->wq.entry);
1154                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1155
1156                         /*
1157                          * Leave the event in the waitqueue and report
1158                          * error to userland if we failed to resolve
1159                          * the userfault fork.
1160                          */
1161                         if (likely(!ret))
1162                                 userfaultfd_event_complete(ctx, uwq);
1163                 } else {
1164                         /*
1165                          * Here the fork thread aborted and the
1166                          * refcount from the fork thread on fork_nctx
1167                          * has already been released. We still hold
1168                          * the reference we took before releasing the
1169                          * lock above. If resolve_userfault_fork
1170                          * failed we've to drop it because the
1171                          * fork_nctx has to be freed in such case. If
1172                          * it succeeded we'll hold it because the new
1173                          * uffd references it.
1174                          */
1175                         if (ret)
1176                                 userfaultfd_ctx_put(fork_nctx);
1177                 }
1178                 spin_unlock(&ctx->event_wqh.lock);
1179         }
1180
1181         return ret;
1182 }
1183
1184 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1185                                 size_t count, loff_t *ppos)
1186 {
1187         struct userfaultfd_ctx *ctx = file->private_data;
1188         ssize_t _ret, ret = 0;
1189         struct uffd_msg msg;
1190         int no_wait = file->f_flags & O_NONBLOCK;
1191
1192         if (ctx->state == UFFD_STATE_WAIT_API)
1193                 return -EINVAL;
1194
1195         for (;;) {
1196                 if (count < sizeof(msg))
1197                         return ret ? ret : -EINVAL;
1198                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1199                 if (_ret < 0)
1200                         return ret ? ret : _ret;
1201                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1202                         return ret ? ret : -EFAULT;
1203                 ret += sizeof(msg);
1204                 buf += sizeof(msg);
1205                 count -= sizeof(msg);
1206                 /*
1207                  * Allow to read more than one fault at time but only
1208                  * block if waiting for the very first one.
1209                  */
1210                 no_wait = O_NONBLOCK;
1211         }
1212 }
1213
1214 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1215                              struct userfaultfd_wake_range *range)
1216 {
1217         spin_lock(&ctx->fault_pending_wqh.lock);
1218         /* wake all in the range and autoremove */
1219         if (waitqueue_active(&ctx->fault_pending_wqh))
1220                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1221                                      range);
1222         if (waitqueue_active(&ctx->fault_wqh))
1223                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1224         spin_unlock(&ctx->fault_pending_wqh.lock);
1225 }
1226
1227 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1228                                            struct userfaultfd_wake_range *range)
1229 {
1230         unsigned seq;
1231         bool need_wakeup;
1232
1233         /*
1234          * To be sure waitqueue_active() is not reordered by the CPU
1235          * before the pagetable update, use an explicit SMP memory
1236          * barrier here. PT lock release or up_read(mmap_sem) still
1237          * have release semantics that can allow the
1238          * waitqueue_active() to be reordered before the pte update.
1239          */
1240         smp_mb();
1241
1242         /*
1243          * Use waitqueue_active because it's very frequent to
1244          * change the address space atomically even if there are no
1245          * userfaults yet. So we take the spinlock only when we're
1246          * sure we've userfaults to wake.
1247          */
1248         do {
1249                 seq = read_seqcount_begin(&ctx->refile_seq);
1250                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1251                         waitqueue_active(&ctx->fault_wqh);
1252                 cond_resched();
1253         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1254         if (need_wakeup)
1255                 __wake_userfault(ctx, range);
1256 }
1257
1258 static __always_inline int validate_range(struct mm_struct *mm,
1259                                           __u64 start, __u64 len)
1260 {
1261         __u64 task_size = mm->task_size;
1262
1263         if (start & ~PAGE_MASK)
1264                 return -EINVAL;
1265         if (len & ~PAGE_MASK)
1266                 return -EINVAL;
1267         if (!len)
1268                 return -EINVAL;
1269         if (start < mmap_min_addr)
1270                 return -EINVAL;
1271         if (start >= task_size)
1272                 return -EINVAL;
1273         if (len > task_size - start)
1274                 return -EINVAL;
1275         return 0;
1276 }
1277
1278 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1279 {
1280         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1281                 vma_is_shmem(vma);
1282 }
1283
1284 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1285                                 unsigned long arg)
1286 {
1287         struct mm_struct *mm = ctx->mm;
1288         struct vm_area_struct *vma, *prev, *cur;
1289         int ret;
1290         struct uffdio_register uffdio_register;
1291         struct uffdio_register __user *user_uffdio_register;
1292         unsigned long vm_flags, new_flags;
1293         bool found;
1294         bool basic_ioctls;
1295         unsigned long start, end, vma_end;
1296
1297         user_uffdio_register = (struct uffdio_register __user *) arg;
1298
1299         ret = -EFAULT;
1300         if (copy_from_user(&uffdio_register, user_uffdio_register,
1301                            sizeof(uffdio_register)-sizeof(__u64)))
1302                 goto out;
1303
1304         ret = -EINVAL;
1305         if (!uffdio_register.mode)
1306                 goto out;
1307         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1308                                      UFFDIO_REGISTER_MODE_WP))
1309                 goto out;
1310         vm_flags = 0;
1311         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1312                 vm_flags |= VM_UFFD_MISSING;
1313         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1314                 vm_flags |= VM_UFFD_WP;
1315                 /*
1316                  * FIXME: remove the below error constraint by
1317                  * implementing the wprotect tracking mode.
1318                  */
1319                 ret = -EINVAL;
1320                 goto out;
1321         }
1322
1323         ret = validate_range(mm, uffdio_register.range.start,
1324                              uffdio_register.range.len);
1325         if (ret)
1326                 goto out;
1327
1328         start = uffdio_register.range.start;
1329         end = start + uffdio_register.range.len;
1330
1331         ret = -ENOMEM;
1332         if (!mmget_not_zero(mm))
1333                 goto out;
1334
1335         down_write(&mm->mmap_sem);
1336         if (!mmget_still_valid(mm))
1337                 goto out_unlock;
1338         vma = find_vma_prev(mm, start, &prev);
1339         if (!vma)
1340                 goto out_unlock;
1341
1342         /* check that there's at least one vma in the range */
1343         ret = -EINVAL;
1344         if (vma->vm_start >= end)
1345                 goto out_unlock;
1346
1347         /*
1348          * If the first vma contains huge pages, make sure start address
1349          * is aligned to huge page size.
1350          */
1351         if (is_vm_hugetlb_page(vma)) {
1352                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1353
1354                 if (start & (vma_hpagesize - 1))
1355                         goto out_unlock;
1356         }
1357
1358         /*
1359          * Search for not compatible vmas.
1360          */
1361         found = false;
1362         basic_ioctls = false;
1363         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1364                 cond_resched();
1365
1366                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1367                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1368
1369                 /* check not compatible vmas */
1370                 ret = -EINVAL;
1371                 if (!vma_can_userfault(cur))
1372                         goto out_unlock;
1373
1374                 /*
1375                  * UFFDIO_COPY will fill file holes even without
1376                  * PROT_WRITE. This check enforces that if this is a
1377                  * MAP_SHARED, the process has write permission to the backing
1378                  * file. If VM_MAYWRITE is set it also enforces that on a
1379                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1380                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1381                  */
1382                 ret = -EPERM;
1383                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1384                         goto out_unlock;
1385
1386                 /*
1387                  * If this vma contains ending address, and huge pages
1388                  * check alignment.
1389                  */
1390                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1391                     end > cur->vm_start) {
1392                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1393
1394                         ret = -EINVAL;
1395
1396                         if (end & (vma_hpagesize - 1))
1397                                 goto out_unlock;
1398                 }
1399
1400                 /*
1401                  * Check that this vma isn't already owned by a
1402                  * different userfaultfd. We can't allow more than one
1403                  * userfaultfd to own a single vma simultaneously or we
1404                  * wouldn't know which one to deliver the userfaults to.
1405                  */
1406                 ret = -EBUSY;
1407                 if (cur->vm_userfaultfd_ctx.ctx &&
1408                     cur->vm_userfaultfd_ctx.ctx != ctx)
1409                         goto out_unlock;
1410
1411                 /*
1412                  * Note vmas containing huge pages
1413                  */
1414                 if (is_vm_hugetlb_page(cur))
1415                         basic_ioctls = true;
1416
1417                 found = true;
1418         }
1419         BUG_ON(!found);
1420
1421         if (vma->vm_start < start)
1422                 prev = vma;
1423
1424         ret = 0;
1425         do {
1426                 cond_resched();
1427
1428                 BUG_ON(!vma_can_userfault(vma));
1429                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1430                        vma->vm_userfaultfd_ctx.ctx != ctx);
1431                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1432
1433                 /*
1434                  * Nothing to do: this vma is already registered into this
1435                  * userfaultfd and with the right tracking mode too.
1436                  */
1437                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1438                     (vma->vm_flags & vm_flags) == vm_flags)
1439                         goto skip;
1440
1441                 if (vma->vm_start > start)
1442                         start = vma->vm_start;
1443                 vma_end = min(end, vma->vm_end);
1444
1445                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1446                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1447                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1448                                  vma_policy(vma),
1449                                  ((struct vm_userfaultfd_ctx){ ctx }));
1450                 if (prev) {
1451                         vma = prev;
1452                         goto next;
1453                 }
1454                 if (vma->vm_start < start) {
1455                         ret = split_vma(mm, vma, start, 1);
1456                         if (ret)
1457                                 break;
1458                 }
1459                 if (vma->vm_end > end) {
1460                         ret = split_vma(mm, vma, end, 0);
1461                         if (ret)
1462                                 break;
1463                 }
1464         next:
1465                 /*
1466                  * In the vma_merge() successful mprotect-like case 8:
1467                  * the next vma was merged into the current one and
1468                  * the current one has not been updated yet.
1469                  */
1470                 vma->vm_flags = new_flags;
1471                 vma->vm_userfaultfd_ctx.ctx = ctx;
1472
1473         skip:
1474                 prev = vma;
1475                 start = vma->vm_end;
1476                 vma = vma->vm_next;
1477         } while (vma && vma->vm_start < end);
1478 out_unlock:
1479         up_write(&mm->mmap_sem);
1480         mmput(mm);
1481         if (!ret) {
1482                 /*
1483                  * Now that we scanned all vmas we can already tell
1484                  * userland which ioctls methods are guaranteed to
1485                  * succeed on this range.
1486                  */
1487                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1488                              UFFD_API_RANGE_IOCTLS,
1489                              &user_uffdio_register->ioctls))
1490                         ret = -EFAULT;
1491         }
1492 out:
1493         return ret;
1494 }
1495
1496 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1497                                   unsigned long arg)
1498 {
1499         struct mm_struct *mm = ctx->mm;
1500         struct vm_area_struct *vma, *prev, *cur;
1501         int ret;
1502         struct uffdio_range uffdio_unregister;
1503         unsigned long new_flags;
1504         bool found;
1505         unsigned long start, end, vma_end;
1506         const void __user *buf = (void __user *)arg;
1507
1508         ret = -EFAULT;
1509         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1510                 goto out;
1511
1512         ret = validate_range(mm, uffdio_unregister.start,
1513                              uffdio_unregister.len);
1514         if (ret)
1515                 goto out;
1516
1517         start = uffdio_unregister.start;
1518         end = start + uffdio_unregister.len;
1519
1520         ret = -ENOMEM;
1521         if (!mmget_not_zero(mm))
1522                 goto out;
1523
1524         down_write(&mm->mmap_sem);
1525         if (!mmget_still_valid(mm))
1526                 goto out_unlock;
1527         vma = find_vma_prev(mm, start, &prev);
1528         if (!vma)
1529                 goto out_unlock;
1530
1531         /* check that there's at least one vma in the range */
1532         ret = -EINVAL;
1533         if (vma->vm_start >= end)
1534                 goto out_unlock;
1535
1536         /*
1537          * If the first vma contains huge pages, make sure start address
1538          * is aligned to huge page size.
1539          */
1540         if (is_vm_hugetlb_page(vma)) {
1541                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1542
1543                 if (start & (vma_hpagesize - 1))
1544                         goto out_unlock;
1545         }
1546
1547         /*
1548          * Search for not compatible vmas.
1549          */
1550         found = false;
1551         ret = -EINVAL;
1552         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1553                 cond_resched();
1554
1555                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1556                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1557
1558                 /*
1559                  * Check not compatible vmas, not strictly required
1560                  * here as not compatible vmas cannot have an
1561                  * userfaultfd_ctx registered on them, but this
1562                  * provides for more strict behavior to notice
1563                  * unregistration errors.
1564                  */
1565                 if (!vma_can_userfault(cur))
1566                         goto out_unlock;
1567
1568                 found = true;
1569         }
1570         BUG_ON(!found);
1571
1572         if (vma->vm_start < start)
1573                 prev = vma;
1574
1575         ret = 0;
1576         do {
1577                 cond_resched();
1578
1579                 BUG_ON(!vma_can_userfault(vma));
1580
1581                 /*
1582                  * Nothing to do: this vma is already registered into this
1583                  * userfaultfd and with the right tracking mode too.
1584                  */
1585                 if (!vma->vm_userfaultfd_ctx.ctx)
1586                         goto skip;
1587
1588                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1589
1590                 if (vma->vm_start > start)
1591                         start = vma->vm_start;
1592                 vma_end = min(end, vma->vm_end);
1593
1594                 if (userfaultfd_missing(vma)) {
1595                         /*
1596                          * Wake any concurrent pending userfault while
1597                          * we unregister, so they will not hang
1598                          * permanently and it avoids userland to call
1599                          * UFFDIO_WAKE explicitly.
1600                          */
1601                         struct userfaultfd_wake_range range;
1602                         range.start = start;
1603                         range.len = vma_end - start;
1604                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1605                 }
1606
1607                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1608                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1609                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1610                                  vma_policy(vma),
1611                                  NULL_VM_UFFD_CTX);
1612                 if (prev) {
1613                         vma = prev;
1614                         goto next;
1615                 }
1616                 if (vma->vm_start < start) {
1617                         ret = split_vma(mm, vma, start, 1);
1618                         if (ret)
1619                                 break;
1620                 }
1621                 if (vma->vm_end > end) {
1622                         ret = split_vma(mm, vma, end, 0);
1623                         if (ret)
1624                                 break;
1625                 }
1626         next:
1627                 /*
1628                  * In the vma_merge() successful mprotect-like case 8:
1629                  * the next vma was merged into the current one and
1630                  * the current one has not been updated yet.
1631                  */
1632                 vma->vm_flags = new_flags;
1633                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1634
1635         skip:
1636                 prev = vma;
1637                 start = vma->vm_end;
1638                 vma = vma->vm_next;
1639         } while (vma && vma->vm_start < end);
1640 out_unlock:
1641         up_write(&mm->mmap_sem);
1642         mmput(mm);
1643 out:
1644         return ret;
1645 }
1646
1647 /*
1648  * userfaultfd_wake may be used in combination with the
1649  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1650  */
1651 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1652                             unsigned long arg)
1653 {
1654         int ret;
1655         struct uffdio_range uffdio_wake;
1656         struct userfaultfd_wake_range range;
1657         const void __user *buf = (void __user *)arg;
1658
1659         ret = -EFAULT;
1660         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1661                 goto out;
1662
1663         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1664         if (ret)
1665                 goto out;
1666
1667         range.start = uffdio_wake.start;
1668         range.len = uffdio_wake.len;
1669
1670         /*
1671          * len == 0 means wake all and we don't want to wake all here,
1672          * so check it again to be sure.
1673          */
1674         VM_BUG_ON(!range.len);
1675
1676         wake_userfault(ctx, &range);
1677         ret = 0;
1678
1679 out:
1680         return ret;
1681 }
1682
1683 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1684                             unsigned long arg)
1685 {
1686         __s64 ret;
1687         struct uffdio_copy uffdio_copy;
1688         struct uffdio_copy __user *user_uffdio_copy;
1689         struct userfaultfd_wake_range range;
1690
1691         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1692
1693         ret = -EFAULT;
1694         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1695                            /* don't copy "copy" last field */
1696                            sizeof(uffdio_copy)-sizeof(__s64)))
1697                 goto out;
1698
1699         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1700         if (ret)
1701                 goto out;
1702         /*
1703          * double check for wraparound just in case. copy_from_user()
1704          * will later check uffdio_copy.src + uffdio_copy.len to fit
1705          * in the userland range.
1706          */
1707         ret = -EINVAL;
1708         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1709                 goto out;
1710         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1711                 goto out;
1712         if (mmget_not_zero(ctx->mm)) {
1713                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1714                                    uffdio_copy.len);
1715                 mmput(ctx->mm);
1716         } else {
1717                 return -ESRCH;
1718         }
1719         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1720                 return -EFAULT;
1721         if (ret < 0)
1722                 goto out;
1723         BUG_ON(!ret);
1724         /* len == 0 would wake all */
1725         range.len = ret;
1726         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1727                 range.start = uffdio_copy.dst;
1728                 wake_userfault(ctx, &range);
1729         }
1730         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1731 out:
1732         return ret;
1733 }
1734
1735 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1736                                 unsigned long arg)
1737 {
1738         __s64 ret;
1739         struct uffdio_zeropage uffdio_zeropage;
1740         struct uffdio_zeropage __user *user_uffdio_zeropage;
1741         struct userfaultfd_wake_range range;
1742
1743         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1744
1745         ret = -EFAULT;
1746         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1747                            /* don't copy "zeropage" last field */
1748                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1749                 goto out;
1750
1751         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1752                              uffdio_zeropage.range.len);
1753         if (ret)
1754                 goto out;
1755         ret = -EINVAL;
1756         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1757                 goto out;
1758
1759         if (mmget_not_zero(ctx->mm)) {
1760                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1761                                      uffdio_zeropage.range.len);
1762                 mmput(ctx->mm);
1763         } else {
1764                 return -ESRCH;
1765         }
1766         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1767                 return -EFAULT;
1768         if (ret < 0)
1769                 goto out;
1770         /* len == 0 would wake all */
1771         BUG_ON(!ret);
1772         range.len = ret;
1773         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1774                 range.start = uffdio_zeropage.range.start;
1775                 wake_userfault(ctx, &range);
1776         }
1777         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1778 out:
1779         return ret;
1780 }
1781
1782 static inline unsigned int uffd_ctx_features(__u64 user_features)
1783 {
1784         /*
1785          * For the current set of features the bits just coincide
1786          */
1787         return (unsigned int)user_features;
1788 }
1789
1790 /*
1791  * userland asks for a certain API version and we return which bits
1792  * and ioctl commands are implemented in this kernel for such API
1793  * version or -EINVAL if unknown.
1794  */
1795 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1796                            unsigned long arg)
1797 {
1798         struct uffdio_api uffdio_api;
1799         void __user *buf = (void __user *)arg;
1800         int ret;
1801         __u64 features;
1802
1803         ret = -EINVAL;
1804         if (ctx->state != UFFD_STATE_WAIT_API)
1805                 goto out;
1806         ret = -EFAULT;
1807         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1808                 goto out;
1809         features = uffdio_api.features;
1810         ret = -EINVAL;
1811         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1812                 goto err_out;
1813         ret = -EPERM;
1814         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1815                 goto err_out;
1816         /* report all available features and ioctls to userland */
1817         uffdio_api.features = UFFD_API_FEATURES;
1818         uffdio_api.ioctls = UFFD_API_IOCTLS;
1819         ret = -EFAULT;
1820         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1821                 goto out;
1822         ctx->state = UFFD_STATE_RUNNING;
1823         /* only enable the requested features for this uffd context */
1824         ctx->features = uffd_ctx_features(features);
1825         ret = 0;
1826 out:
1827         return ret;
1828 err_out:
1829         memset(&uffdio_api, 0, sizeof(uffdio_api));
1830         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1831                 ret = -EFAULT;
1832         goto out;
1833 }
1834
1835 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1836                               unsigned long arg)
1837 {
1838         int ret = -EINVAL;
1839         struct userfaultfd_ctx *ctx = file->private_data;
1840
1841         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1842                 return -EINVAL;
1843
1844         switch(cmd) {
1845         case UFFDIO_API:
1846                 ret = userfaultfd_api(ctx, arg);
1847                 break;
1848         case UFFDIO_REGISTER:
1849                 ret = userfaultfd_register(ctx, arg);
1850                 break;
1851         case UFFDIO_UNREGISTER:
1852                 ret = userfaultfd_unregister(ctx, arg);
1853                 break;
1854         case UFFDIO_WAKE:
1855                 ret = userfaultfd_wake(ctx, arg);
1856                 break;
1857         case UFFDIO_COPY:
1858                 ret = userfaultfd_copy(ctx, arg);
1859                 break;
1860         case UFFDIO_ZEROPAGE:
1861                 ret = userfaultfd_zeropage(ctx, arg);
1862                 break;
1863         }
1864         return ret;
1865 }
1866
1867 #ifdef CONFIG_PROC_FS
1868 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1869 {
1870         struct userfaultfd_ctx *ctx = f->private_data;
1871         wait_queue_entry_t *wq;
1872         struct userfaultfd_wait_queue *uwq;
1873         unsigned long pending = 0, total = 0;
1874
1875         spin_lock(&ctx->fault_pending_wqh.lock);
1876         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1877                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1878                 pending++;
1879                 total++;
1880         }
1881         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1882                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1883                 total++;
1884         }
1885         spin_unlock(&ctx->fault_pending_wqh.lock);
1886
1887         /*
1888          * If more protocols will be added, there will be all shown
1889          * separated by a space. Like this:
1890          *      protocols: aa:... bb:...
1891          */
1892         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1893                    pending, total, UFFD_API, ctx->features,
1894                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1895 }
1896 #endif
1897
1898 static const struct file_operations userfaultfd_fops = {
1899 #ifdef CONFIG_PROC_FS
1900         .show_fdinfo    = userfaultfd_show_fdinfo,
1901 #endif
1902         .release        = userfaultfd_release,
1903         .poll           = userfaultfd_poll,
1904         .read           = userfaultfd_read,
1905         .unlocked_ioctl = userfaultfd_ioctl,
1906         .compat_ioctl   = userfaultfd_ioctl,
1907         .llseek         = noop_llseek,
1908 };
1909
1910 static void init_once_userfaultfd_ctx(void *mem)
1911 {
1912         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1913
1914         init_waitqueue_head(&ctx->fault_pending_wqh);
1915         init_waitqueue_head(&ctx->fault_wqh);
1916         init_waitqueue_head(&ctx->event_wqh);
1917         init_waitqueue_head(&ctx->fd_wqh);
1918         seqcount_init(&ctx->refile_seq);
1919 }
1920
1921 /**
1922  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1923  * @flags: Flags for the userfaultfd file.
1924  *
1925  * This function creates a userfaultfd file pointer, w/out installing
1926  * it into the fd table. This is useful when the userfaultfd file is
1927  * used during the initialization of data structures that require
1928  * extra setup after the userfaultfd creation. So the userfaultfd
1929  * creation is split into the file pointer creation phase, and the
1930  * file descriptor installation phase.  In this way races with
1931  * userspace closing the newly installed file descriptor can be
1932  * avoided.  Returns a userfaultfd file pointer, or a proper error
1933  * pointer.
1934  */
1935 static struct file *userfaultfd_file_create(int flags)
1936 {
1937         struct file *file;
1938         struct userfaultfd_ctx *ctx;
1939
1940         BUG_ON(!current->mm);
1941
1942         /* Check the UFFD_* constants for consistency.  */
1943         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1944         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1945
1946         file = ERR_PTR(-EINVAL);
1947         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1948                 goto out;
1949
1950         file = ERR_PTR(-ENOMEM);
1951         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1952         if (!ctx)
1953                 goto out;
1954
1955         atomic_set(&ctx->refcount, 1);
1956         ctx->flags = flags;
1957         ctx->features = 0;
1958         ctx->state = UFFD_STATE_WAIT_API;
1959         ctx->released = false;
1960         ctx->mm = current->mm;
1961         /* prevent the mm struct to be freed */
1962         mmgrab(ctx->mm);
1963
1964         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1965                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1966         if (IS_ERR(file)) {
1967                 mmdrop(ctx->mm);
1968                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1969         }
1970 out:
1971         return file;
1972 }
1973
1974 SYSCALL_DEFINE1(userfaultfd, int, flags)
1975 {
1976         int fd, error;
1977         struct file *file;
1978
1979         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1980         if (error < 0)
1981                 return error;
1982         fd = error;
1983
1984         file = userfaultfd_file_create(flags);
1985         if (IS_ERR(file)) {
1986                 error = PTR_ERR(file);
1987                 goto err_put_unused_fd;
1988         }
1989         fd_install(fd, file);
1990
1991         return fd;
1992
1993 err_put_unused_fd:
1994         put_unused_fd(fd);
1995
1996         return error;
1997 }
1998
1999 static int __init userfaultfd_init(void)
2000 {
2001         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2002                                                 sizeof(struct userfaultfd_ctx),
2003                                                 0,
2004                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2005                                                 init_once_userfaultfd_ctx);
2006         return 0;
2007 }
2008 __initcall(userfaultfd_init);