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                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 631                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 632                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 633                                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 634                         }
 635                 up_write(&mm->mmap_sem);
 636
 637                 userfaultfd_ctx_put(release_new_ctx);
 638         }
 639
 640         /*
 641          * ctx may go away after this if the userfault pseudo fd is
 642          * already released.
 643          */
 644 out:
 645         userfaultfd_ctx_put(ctx);
 646 }
 647
 648 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 649                                        struct userfaultfd_wait_queue *ewq)
 650 {
 651         ewq->msg.event = 0;
 652         wake_up_locked(&ctx->event_wqh);
 653         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 654 }
 655
 656 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 657 {
 658         struct userfaultfd_ctx *ctx = NULL, *octx;
 659         struct userfaultfd_fork_ctx *fctx;
 660
 661         octx = vma->vm_userfaultfd_ctx.ctx;
 662         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 663                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 664                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 665                 return 0;
 666         }
 667
 668         list_for_each_entry(fctx, fcs, list)
 669                 if (fctx->orig == octx) {
 670                         ctx = fctx->new;
 671                         break;
 672                 }
 673
 674         if (!ctx) {
 675                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 676                 if (!fctx)
 677                         return -ENOMEM;
 678
 679                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 680                 if (!ctx) {
 681                         kfree(fctx);
 682                         return -ENOMEM;
 683                 }
 684
 685                 atomic_set(&ctx->refcount, 1);
 686                 ctx->flags = octx->flags;
 687                 ctx->state = UFFD_STATE_RUNNING;
 688                 ctx->features = octx->features;
 689                 ctx->released = false;
 690                 ctx->mm = vma->vm_mm;
 691                 atomic_inc(&ctx->mm->mm_count);
 692
 693                 userfaultfd_ctx_get(octx);
 694                 fctx->orig = octx;
 695                 fctx->new = ctx;
 696                 list_add_tail(&fctx->list, fcs);
 697         }
 698
 699         vma->vm_userfaultfd_ctx.ctx = ctx;
 700         return 0;
 701 }
 702
 703 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 704 {
 705         struct userfaultfd_ctx *ctx = fctx->orig;
 706         struct userfaultfd_wait_queue ewq;
 707
 708         msg_init(&ewq.msg);
 709
 710         ewq.msg.event = UFFD_EVENT_FORK;
 711         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 712
 713         userfaultfd_event_wait_completion(ctx, &ewq);
 714 }
 715
 716 void dup_userfaultfd_complete(struct list_head *fcs)
 717 {
 718         struct userfaultfd_fork_ctx *fctx, *n;
 719
 720         list_for_each_entry_safe(fctx, n, fcs, list) {
 721                 dup_fctx(fctx);
 722                 list_del(&fctx->list);
 723                 kfree(fctx);
 724         }
 725 }
 726
 727 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 728                              struct vm_userfaultfd_ctx *vm_ctx)
 729 {
 730         struct userfaultfd_ctx *ctx;
 731
 732         ctx = vma->vm_userfaultfd_ctx.ctx;
 733         if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) {
 734                 vm_ctx->ctx = ctx;
 735                 userfaultfd_ctx_get(ctx);
 736         }
 737 }
 738
 739 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 740                                  unsigned long from, unsigned long to,
 741                                  unsigned long len)
 742 {
 743         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 744         struct userfaultfd_wait_queue ewq;
 745
 746         if (!ctx)
 747                 return;
 748
 749         if (to & ~PAGE_MASK) {
 750                 userfaultfd_ctx_put(ctx);
 751                 return;
 752         }
 753
 754         msg_init(&ewq.msg);
 755
 756         ewq.msg.event = UFFD_EVENT_REMAP;
 757         ewq.msg.arg.remap.from = from;
 758         ewq.msg.arg.remap.to = to;
 759         ewq.msg.arg.remap.len = len;
 760
 761         userfaultfd_event_wait_completion(ctx, &ewq);
 762 }
 763
 764 bool userfaultfd_remove(struct vm_area_struct *vma,
 765                         unsigned long start, unsigned long end)
 766 {
 767         struct mm_struct *mm = vma->vm_mm;
 768         struct userfaultfd_ctx *ctx;
 769         struct userfaultfd_wait_queue ewq;
 770
 771         ctx = vma->vm_userfaultfd_ctx.ctx;
 772         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 773                 return true;
 774
 775         userfaultfd_ctx_get(ctx);
 776         up_read(&mm->mmap_sem);
 777
 778         msg_init(&ewq.msg);
 779
 780         ewq.msg.event = UFFD_EVENT_REMOVE;
 781         ewq.msg.arg.remove.start = start;
 782         ewq.msg.arg.remove.end = end;
 783
 784         userfaultfd_event_wait_completion(ctx, &ewq);
 785
 786         return false;
 787 }
 788
 789 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 790                           unsigned long start, unsigned long end)
 791 {
 792         struct userfaultfd_unmap_ctx *unmap_ctx;
 793
 794         list_for_each_entry(unmap_ctx, unmaps, list)
 795                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 796                     unmap_ctx->end == end)
 797                         return true;
 798
 799         return false;
 800 }
 801
 802 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 803                            unsigned long start, unsigned long end,
 804                            struct list_head *unmaps)
 805 {
 806         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 807                 struct userfaultfd_unmap_ctx *unmap_ctx;
 808                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 809
 810                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 811                     has_unmap_ctx(ctx, unmaps, start, end))
 812                         continue;
 813
 814                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 815                 if (!unmap_ctx)
 816                         return -ENOMEM;
 817
 818                 userfaultfd_ctx_get(ctx);
 819                 unmap_ctx->ctx = ctx;
 820                 unmap_ctx->start = start;
 821                 unmap_ctx->end = end;
 822                 list_add_tail(&unmap_ctx->list, unmaps);
 823         }
 824
 825         return 0;
 826 }
 827
 828 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 829 {
 830         struct userfaultfd_unmap_ctx *ctx, *n;
 831         struct userfaultfd_wait_queue ewq;
 832
 833         list_for_each_entry_safe(ctx, n, uf, list) {
 834                 msg_init(&ewq.msg);
 835
 836                 ewq.msg.event = UFFD_EVENT_UNMAP;
 837                 ewq.msg.arg.remove.start = ctx->start;
 838                 ewq.msg.arg.remove.end = ctx->end;
 839
 840                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 841
 842                 list_del(&ctx->list);
 843                 kfree(ctx);
 844         }
 845 }
 846
 847 static int userfaultfd_release(struct inode *inode, struct file *file)
 848 {
 849         struct userfaultfd_ctx *ctx = file->private_data;
 850         struct mm_struct *mm = ctx->mm;
 851         struct vm_area_struct *vma, *prev;
 852         /* len == 0 means wake all */
 853         struct userfaultfd_wake_range range = { .len = 0, };
 854         unsigned long new_flags;
 855
 856         ACCESS_ONCE(ctx->released) = true;
 857
 858         if (!mmget_not_zero(mm))
 859                 goto wakeup;
 860
 861         /*
 862          * Flush page faults out of all CPUs. NOTE: all page faults
 863          * must be retried without returning VM_FAULT_SIGBUS if
 864          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 865          * changes while handle_userfault released the mmap_sem. So
 866          * it's critical that released is set to true (above), before
 867          * taking the mmap_sem for writing.
 868          */
 869         down_write(&mm->mmap_sem);
 870         prev = NULL;
 871         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 872                 cond_resched();
 873                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 874                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 875                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 876                         prev = vma;
 877                         continue;
 878                 }
 879                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 880                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 881                                  new_flags, vma->anon_vma,
 882                                  vma->vm_file, vma->vm_pgoff,
 883                                  vma_policy(vma),
 884                                  NULL_VM_UFFD_CTX);
 885                 if (prev)
 886                         vma = prev;
 887                 else
 888                         prev = vma;
 889                 vma->vm_flags = new_flags;
 890                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 891         }
 892         up_write(&mm->mmap_sem);
 893         mmput(mm);
 894 wakeup:
 895         /*
 896          * After no new page faults can wait on this fault_*wqh, flush
 897          * the last page faults that may have been already waiting on
 898          * the fault_*wqh.
 899          */
 900         spin_lock(&ctx->fault_pending_wqh.lock);
 901         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 902         __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range);
 903         spin_unlock(&ctx->fault_pending_wqh.lock);
 904
 905         /* Flush pending events that may still wait on event_wqh */
 906         wake_up_all(&ctx->event_wqh);
 907
 908         wake_up_poll(&ctx->fd_wqh, POLLHUP);
 909         userfaultfd_ctx_put(ctx);
 910         return 0;
 911 }
 912
 913 /* fault_pending_wqh.lock must be hold by the caller */
 914 static inline struct userfaultfd_wait_queue *find_userfault_in(
 915                 wait_queue_head_t *wqh)
 916 {
 917         wait_queue_entry_t *wq;
 918         struct userfaultfd_wait_queue *uwq;
 919
 920         VM_BUG_ON(!spin_is_locked(&wqh->lock));
 921
 922         uwq = NULL;
 923         if (!waitqueue_active(wqh))
 924                 goto out;
 925         /* walk in reverse to provide FIFO behavior to read userfaults */
 926         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 927         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 928 out:
 929         return uwq;
 930 }
 931
 932 static inline struct userfaultfd_wait_queue *find_userfault(
 933                 struct userfaultfd_ctx *ctx)
 934 {
 935         return find_userfault_in(&ctx->fault_pending_wqh);
 936 }
 937
 938 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 939                 struct userfaultfd_ctx *ctx)
 940 {
 941         return find_userfault_in(&ctx->event_wqh);
 942 }
 943
 944 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 945 {
 946         struct userfaultfd_ctx *ctx = file->private_data;
 947         unsigned int ret;
 948
 949         poll_wait(file, &ctx->fd_wqh, wait);
 950
 951         switch (ctx->state) {
 952         case UFFD_STATE_WAIT_API:
 953                 return POLLERR;
 954         case UFFD_STATE_RUNNING:
 955                 /*
 956                  * poll() never guarantees that read won't block.
 957                  * userfaults can be waken before they're read().
 958                  */
 959                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 960                         return POLLERR;
 961                 /*
 962                  * lockless access to see if there are pending faults
 963                  * __pollwait last action is the add_wait_queue but
 964                  * the spin_unlock would allow the waitqueue_active to
 965                  * pass above the actual list_add inside
 966                  * add_wait_queue critical section. So use a full
 967                  * memory barrier to serialize the list_add write of
 968                  * add_wait_queue() with the waitqueue_active read
 969                  * below.
 970                  */
 971                 ret = 0;
 972                 smp_mb();
 973                 if (waitqueue_active(&ctx->fault_pending_wqh))
 974                         ret = POLLIN;
 975                 else if (waitqueue_active(&ctx->event_wqh))
 976                         ret = POLLIN;
 977
 978                 return ret;
 979         default:
 980                 WARN_ON_ONCE(1);
 981                 return POLLERR;
 982         }
 983 }
 984
 985 static const struct file_operations userfaultfd_fops;
 986
 987 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
 988                                   struct userfaultfd_ctx *new,
 989                                   struct uffd_msg *msg)
 990 {
 991         int fd;
 992         struct file *file;
 993         unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS;
 994
 995         fd = get_unused_fd_flags(flags);
 996         if (fd < 0)
 997                 return fd;
 998
 999         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
1000                                   O_RDWR | flags);
1001         if (IS_ERR(file)) {
1002                 put_unused_fd(fd);
1003                 return PTR_ERR(file);
1004         }
1005
1006         fd_install(fd, file);
1007         msg->arg.reserved.reserved1 = 0;
1008         msg->arg.fork.ufd = fd;
1009
1010         return 0;
1011 }
1012
1013 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1014                                     struct uffd_msg *msg)
1015 {
1016         ssize_t ret;
1017         DECLARE_WAITQUEUE(wait, current);
1018         struct userfaultfd_wait_queue *uwq;
1019         /*
1020          * Handling fork event requires sleeping operations, so
1021          * we drop the event_wqh lock, then do these ops, then
1022          * lock it back and wake up the waiter. While the lock is
1023          * dropped the ewq may go away so we keep track of it
1024          * carefully.
1025          */
1026         LIST_HEAD(fork_event);
1027         struct userfaultfd_ctx *fork_nctx = NULL;
1028
1029         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1030         spin_lock(&ctx->fd_wqh.lock);
1031         __add_wait_queue(&ctx->fd_wqh, &wait);
1032         for (;;) {
1033                 set_current_state(TASK_INTERRUPTIBLE);
1034                 spin_lock(&ctx->fault_pending_wqh.lock);
1035                 uwq = find_userfault(ctx);
1036                 if (uwq) {
1037                         /*
1038                          * Use a seqcount to repeat the lockless check
1039                          * in wake_userfault() to avoid missing
1040                          * wakeups because during the refile both
1041                          * waitqueue could become empty if this is the
1042                          * only userfault.
1043                          */
1044                         write_seqcount_begin(&ctx->refile_seq);
1045
1046                         /*
1047                          * The fault_pending_wqh.lock prevents the uwq
1048                          * to disappear from under us.
1049                          *
1050                          * Refile this userfault from
1051                          * fault_pending_wqh to fault_wqh, it's not
1052                          * pending anymore after we read it.
1053                          *
1054                          * Use list_del() by hand (as
1055                          * userfaultfd_wake_function also uses
1056                          * list_del_init() by hand) to be sure nobody
1057                          * changes __remove_wait_queue() to use
1058                          * list_del_init() in turn breaking the
1059                          * !list_empty_careful() check in
1060                          * handle_userfault(). The uwq->wq.head list
1061                          * must never be empty at any time during the
1062                          * refile, or the waitqueue could disappear
1063                          * from under us. The "wait_queue_head_t"
1064                          * parameter of __remove_wait_queue() is unused
1065                          * anyway.
1066                          */
1067                         list_del(&uwq->wq.entry);
1068                         __add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1069
1070                         write_seqcount_end(&ctx->refile_seq);
1071
1072                         /* careful to always initialize msg if ret == 0 */
1073                         *msg = uwq->msg;
1074                         spin_unlock(&ctx->fault_pending_wqh.lock);
1075                         ret = 0;
1076                         break;
1077                 }
1078                 spin_unlock(&ctx->fault_pending_wqh.lock);
1079
1080                 spin_lock(&ctx->event_wqh.lock);
1081                 uwq = find_userfault_evt(ctx);
1082                 if (uwq) {
1083                         *msg = uwq->msg;
1084
1085                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1086                                 fork_nctx = (struct userfaultfd_ctx *)
1087                                         (unsigned long)
1088                                         uwq->msg.arg.reserved.reserved1;
1089                                 list_move(&uwq->wq.entry, &fork_event);
1090                                 /*
1091                                  * fork_nctx can be freed as soon as
1092                                  * we drop the lock, unless we take a
1093                                  * reference on it.
1094                                  */
1095                                 userfaultfd_ctx_get(fork_nctx);
1096                                 spin_unlock(&ctx->event_wqh.lock);
1097                                 ret = 0;
1098                                 break;
1099                         }
1100
1101                         userfaultfd_event_complete(ctx, uwq);
1102                         spin_unlock(&ctx->event_wqh.lock);
1103                         ret = 0;
1104                         break;
1105                 }
1106                 spin_unlock(&ctx->event_wqh.lock);
1107
1108                 if (signal_pending(current)) {
1109                         ret = -ERESTARTSYS;
1110                         break;
1111                 }
1112                 if (no_wait) {
1113                         ret = -EAGAIN;
1114                         break;
1115                 }
1116                 spin_unlock(&ctx->fd_wqh.lock);
1117                 schedule();
1118                 spin_lock(&ctx->fd_wqh.lock);
1119         }
1120         __remove_wait_queue(&ctx->fd_wqh, &wait);
1121         __set_current_state(TASK_RUNNING);
1122         spin_unlock(&ctx->fd_wqh.lock);
1123
1124         if (!ret && msg->event == UFFD_EVENT_FORK) {
1125                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1126                 spin_lock(&ctx->event_wqh.lock);
1127                 if (!list_empty(&fork_event)) {
1128                         /*
1129                          * The fork thread didn't abort, so we can
1130                          * drop the temporary refcount.
1131                          */
1132                         userfaultfd_ctx_put(fork_nctx);
1133
1134                         uwq = list_first_entry(&fork_event,
1135                                                typeof(*uwq),
1136                                                wq.entry);
1137                         /*
1138                          * If fork_event list wasn't empty and in turn
1139                          * the event wasn't already released by fork
1140                          * (the event is allocated on fork kernel
1141                          * stack), put the event back to its place in
1142                          * the event_wq. fork_event head will be freed
1143                          * as soon as we return so the event cannot
1144                          * stay queued there no matter the current
1145                          * "ret" value.
1146                          */
1147                         list_del(&uwq->wq.entry);
1148                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1149
1150                         /*
1151                          * Leave the event in the waitqueue and report
1152                          * error to userland if we failed to resolve
1153                          * the userfault fork.
1154                          */
1155                         if (likely(!ret))
1156                                 userfaultfd_event_complete(ctx, uwq);
1157                 } else {
1158                         /*
1159                          * Here the fork thread aborted and the
1160                          * refcount from the fork thread on fork_nctx
1161                          * has already been released. We still hold
1162                          * the reference we took before releasing the
1163                          * lock above. If resolve_userfault_fork
1164                          * failed we've to drop it because the
1165                          * fork_nctx has to be freed in such case. If
1166                          * it succeeded we'll hold it because the new
1167                          * uffd references it.
1168                          */
1169                         if (ret)
1170                                 userfaultfd_ctx_put(fork_nctx);
1171                 }
1172                 spin_unlock(&ctx->event_wqh.lock);
1173         }
1174
1175         return ret;
1176 }
1177
1178 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1179                                 size_t count, loff_t *ppos)
1180 {
1181         struct userfaultfd_ctx *ctx = file->private_data;
1182         ssize_t _ret, ret = 0;
1183         struct uffd_msg msg;
1184         int no_wait = file->f_flags & O_NONBLOCK;
1185
1186         if (ctx->state == UFFD_STATE_WAIT_API)
1187                 return -EINVAL;
1188
1189         for (;;) {
1190                 if (count < sizeof(msg))
1191                         return ret ? ret : -EINVAL;
1192                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1193                 if (_ret < 0)
1194                         return ret ? ret : _ret;
1195                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1196                         return ret ? ret : -EFAULT;
1197                 ret += sizeof(msg);
1198                 buf += sizeof(msg);
1199                 count -= sizeof(msg);
1200                 /*
1201                  * Allow to read more than one fault at time but only
1202                  * block if waiting for the very first one.
1203                  */
1204                 no_wait = O_NONBLOCK;
1205         }
1206 }
1207
1208 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1209                              struct userfaultfd_wake_range *range)
1210 {
1211         spin_lock(&ctx->fault_pending_wqh.lock);
1212         /* wake all in the range and autoremove */
1213         if (waitqueue_active(&ctx->fault_pending_wqh))
1214                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1215                                      range);
1216         if (waitqueue_active(&ctx->fault_wqh))
1217                 __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range);
1218         spin_unlock(&ctx->fault_pending_wqh.lock);
1219 }
1220
1221 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1222                                            struct userfaultfd_wake_range *range)
1223 {
1224         unsigned seq;
1225         bool need_wakeup;
1226
1227         /*
1228          * To be sure waitqueue_active() is not reordered by the CPU
1229          * before the pagetable update, use an explicit SMP memory
1230          * barrier here. PT lock release or up_read(mmap_sem) still
1231          * have release semantics that can allow the
1232          * waitqueue_active() to be reordered before the pte update.
1233          */
1234         smp_mb();
1235
1236         /*
1237          * Use waitqueue_active because it's very frequent to
1238          * change the address space atomically even if there are no
1239          * userfaults yet. So we take the spinlock only when we're
1240          * sure we've userfaults to wake.
1241          */
1242         do {
1243                 seq = read_seqcount_begin(&ctx->refile_seq);
1244                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1245                         waitqueue_active(&ctx->fault_wqh);
1246                 cond_resched();
1247         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1248         if (need_wakeup)
1249                 __wake_userfault(ctx, range);
1250 }
1251
1252 static __always_inline int validate_range(struct mm_struct *mm,
1253                                           __u64 start, __u64 len)
1254 {
1255         __u64 task_size = mm->task_size;
1256
1257         if (start & ~PAGE_MASK)
1258                 return -EINVAL;
1259         if (len & ~PAGE_MASK)
1260                 return -EINVAL;
1261         if (!len)
1262                 return -EINVAL;
1263         if (start < mmap_min_addr)
1264                 return -EINVAL;
1265         if (start >= task_size)
1266                 return -EINVAL;
1267         if (len > task_size - start)
1268                 return -EINVAL;
1269         return 0;
1270 }
1271
1272 static inline bool vma_can_userfault(struct vm_area_struct *vma)
1273 {
1274         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1275                 vma_is_shmem(vma);
1276 }
1277
1278 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1279                                 unsigned long arg)
1280 {
1281         struct mm_struct *mm = ctx->mm;
1282         struct vm_area_struct *vma, *prev, *cur;
1283         int ret;
1284         struct uffdio_register uffdio_register;
1285         struct uffdio_register __user *user_uffdio_register;
1286         unsigned long vm_flags, new_flags;
1287         bool found;
1288         bool basic_ioctls;
1289         unsigned long start, end, vma_end;
1290
1291         user_uffdio_register = (struct uffdio_register __user *) arg;
1292
1293         ret = -EFAULT;
1294         if (copy_from_user(&uffdio_register, user_uffdio_register,
1295                            sizeof(uffdio_register)-sizeof(__u64)))
1296                 goto out;
1297
1298         ret = -EINVAL;
1299         if (!uffdio_register.mode)
1300                 goto out;
1301         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1302                                      UFFDIO_REGISTER_MODE_WP))
1303                 goto out;
1304         vm_flags = 0;
1305         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1306                 vm_flags |= VM_UFFD_MISSING;
1307         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1308                 vm_flags |= VM_UFFD_WP;
1309                 /*
1310                  * FIXME: remove the below error constraint by
1311                  * implementing the wprotect tracking mode.
1312                  */
1313                 ret = -EINVAL;
1314                 goto out;
1315         }
1316
1317         ret = validate_range(mm, uffdio_register.range.start,
1318                              uffdio_register.range.len);
1319         if (ret)
1320                 goto out;
1321
1322         start = uffdio_register.range.start;
1323         end = start + uffdio_register.range.len;
1324
1325         ret = -ENOMEM;
1326         if (!mmget_not_zero(mm))
1327                 goto out;
1328
1329         down_write(&mm->mmap_sem);
1330         vma = find_vma_prev(mm, start, &prev);
1331         if (!vma)
1332                 goto out_unlock;
1333
1334         /* check that there's at least one vma in the range */
1335         ret = -EINVAL;
1336         if (vma->vm_start >= end)
1337                 goto out_unlock;
1338
1339         /*
1340          * If the first vma contains huge pages, make sure start address
1341          * is aligned to huge page size.
1342          */
1343         if (is_vm_hugetlb_page(vma)) {
1344                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1345
1346                 if (start & (vma_hpagesize - 1))
1347                         goto out_unlock;
1348         }
1349
1350         /*
1351          * Search for not compatible vmas.
1352          */
1353         found = false;
1354         basic_ioctls = false;
1355         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1356                 cond_resched();
1357
1358                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1359                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1360
1361                 /* check not compatible vmas */
1362                 ret = -EINVAL;
1363                 if (!vma_can_userfault(cur))
1364                         goto out_unlock;
1365
1366                 /*
1367                  * UFFDIO_COPY will fill file holes even without
1368                  * PROT_WRITE. This check enforces that if this is a
1369                  * MAP_SHARED, the process has write permission to the backing
1370                  * file. If VM_MAYWRITE is set it also enforces that on a
1371                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1372                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1373                  */
1374                 ret = -EPERM;
1375                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1376                         goto out_unlock;
1377
1378                 /*
1379                  * If this vma contains ending address, and huge pages
1380                  * check alignment.
1381                  */
1382                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1383                     end > cur->vm_start) {
1384                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1385
1386                         ret = -EINVAL;
1387
1388                         if (end & (vma_hpagesize - 1))
1389                                 goto out_unlock;
1390                 }
1391
1392                 /*
1393                  * Check that this vma isn't already owned by a
1394                  * different userfaultfd. We can't allow more than one
1395                  * userfaultfd to own a single vma simultaneously or we
1396                  * wouldn't know which one to deliver the userfaults to.
1397                  */
1398                 ret = -EBUSY;
1399                 if (cur->vm_userfaultfd_ctx.ctx &&
1400                     cur->vm_userfaultfd_ctx.ctx != ctx)
1401                         goto out_unlock;
1402
1403                 /*
1404                  * Note vmas containing huge pages
1405                  */
1406                 if (is_vm_hugetlb_page(cur))
1407                         basic_ioctls = true;
1408
1409                 found = true;
1410         }
1411         BUG_ON(!found);
1412
1413         if (vma->vm_start < start)
1414                 prev = vma;
1415
1416         ret = 0;
1417         do {
1418                 cond_resched();
1419
1420                 BUG_ON(!vma_can_userfault(vma));
1421                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1422                        vma->vm_userfaultfd_ctx.ctx != ctx);
1423                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1424
1425                 /*
1426                  * Nothing to do: this vma is already registered into this
1427                  * userfaultfd and with the right tracking mode too.
1428                  */
1429                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1430                     (vma->vm_flags & vm_flags) == vm_flags)
1431                         goto skip;
1432
1433                 if (vma->vm_start > start)
1434                         start = vma->vm_start;
1435                 vma_end = min(end, vma->vm_end);
1436
1437                 new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1438                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1439                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1440                                  vma_policy(vma),
1441                                  ((struct vm_userfaultfd_ctx){ ctx }));
1442                 if (prev) {
1443                         vma = prev;
1444                         goto next;
1445                 }
1446                 if (vma->vm_start < start) {
1447                         ret = split_vma(mm, vma, start, 1);
1448                         if (ret)
1449                                 break;
1450                 }
1451                 if (vma->vm_end > end) {
1452                         ret = split_vma(mm, vma, end, 0);
1453                         if (ret)
1454                                 break;
1455                 }
1456         next:
1457                 /*
1458                  * In the vma_merge() successful mprotect-like case 8:
1459                  * the next vma was merged into the current one and
1460                  * the current one has not been updated yet.
1461                  */
1462                 vma->vm_flags = new_flags;
1463                 vma->vm_userfaultfd_ctx.ctx = ctx;
1464
1465         skip:
1466                 prev = vma;
1467                 start = vma->vm_end;
1468                 vma = vma->vm_next;
1469         } while (vma && vma->vm_start < end);
1470 out_unlock:
1471         up_write(&mm->mmap_sem);
1472         mmput(mm);
1473         if (!ret) {
1474                 /*
1475                  * Now that we scanned all vmas we can already tell
1476                  * userland which ioctls methods are guaranteed to
1477                  * succeed on this range.
1478                  */
1479                 if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1480                              UFFD_API_RANGE_IOCTLS,
1481                              &user_uffdio_register->ioctls))
1482                         ret = -EFAULT;
1483         }
1484 out:
1485         return ret;
1486 }
1487
1488 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1489                                   unsigned long arg)
1490 {
1491         struct mm_struct *mm = ctx->mm;
1492         struct vm_area_struct *vma, *prev, *cur;
1493         int ret;
1494         struct uffdio_range uffdio_unregister;
1495         unsigned long new_flags;
1496         bool found;
1497         unsigned long start, end, vma_end;
1498         const void __user *buf = (void __user *)arg;
1499
1500         ret = -EFAULT;
1501         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1502                 goto out;
1503
1504         ret = validate_range(mm, uffdio_unregister.start,
1505                              uffdio_unregister.len);
1506         if (ret)
1507                 goto out;
1508
1509         start = uffdio_unregister.start;
1510         end = start + uffdio_unregister.len;
1511
1512         ret = -ENOMEM;
1513         if (!mmget_not_zero(mm))
1514                 goto out;
1515
1516         down_write(&mm->mmap_sem);
1517         vma = find_vma_prev(mm, start, &prev);
1518         if (!vma)
1519                 goto out_unlock;
1520
1521         /* check that there's at least one vma in the range */
1522         ret = -EINVAL;
1523         if (vma->vm_start >= end)
1524                 goto out_unlock;
1525
1526         /*
1527          * If the first vma contains huge pages, make sure start address
1528          * is aligned to huge page size.
1529          */
1530         if (is_vm_hugetlb_page(vma)) {
1531                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1532
1533                 if (start & (vma_hpagesize - 1))
1534                         goto out_unlock;
1535         }
1536
1537         /*
1538          * Search for not compatible vmas.
1539          */
1540         found = false;
1541         ret = -EINVAL;
1542         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1543                 cond_resched();
1544
1545                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1546                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1547
1548                 /*
1549                  * Check not compatible vmas, not strictly required
1550                  * here as not compatible vmas cannot have an
1551                  * userfaultfd_ctx registered on them, but this
1552                  * provides for more strict behavior to notice
1553                  * unregistration errors.
1554                  */
1555                 if (!vma_can_userfault(cur))
1556                         goto out_unlock;
1557
1558                 found = true;
1559         }
1560         BUG_ON(!found);
1561
1562         if (vma->vm_start < start)
1563                 prev = vma;
1564
1565         ret = 0;
1566         do {
1567                 cond_resched();
1568
1569                 BUG_ON(!vma_can_userfault(vma));
1570
1571                 /*
1572                  * Nothing to do: this vma is already registered into this
1573                  * userfaultfd and with the right tracking mode too.
1574                  */
1575                 if (!vma->vm_userfaultfd_ctx.ctx)
1576                         goto skip;
1577
1578                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1579
1580                 if (vma->vm_start > start)
1581                         start = vma->vm_start;
1582                 vma_end = min(end, vma->vm_end);
1583
1584                 if (userfaultfd_missing(vma)) {
1585                         /*
1586                          * Wake any concurrent pending userfault while
1587                          * we unregister, so they will not hang
1588                          * permanently and it avoids userland to call
1589                          * UFFDIO_WAKE explicitly.
1590                          */
1591                         struct userfaultfd_wake_range range;
1592                         range.start = start;
1593                         range.len = vma_end - start;
1594                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1595                 }
1596
1597                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1598                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1599                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1600                                  vma_policy(vma),
1601                                  NULL_VM_UFFD_CTX);
1602                 if (prev) {
1603                         vma = prev;
1604                         goto next;
1605                 }
1606                 if (vma->vm_start < start) {
1607                         ret = split_vma(mm, vma, start, 1);
1608                         if (ret)
1609                                 break;
1610                 }
1611                 if (vma->vm_end > end) {
1612                         ret = split_vma(mm, vma, end, 0);
1613                         if (ret)
1614                                 break;
1615                 }
1616         next:
1617                 /*
1618                  * In the vma_merge() successful mprotect-like case 8:
1619                  * the next vma was merged into the current one and
1620                  * the current one has not been updated yet.
1621                  */
1622                 vma->vm_flags = new_flags;
1623                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1624
1625         skip:
1626                 prev = vma;
1627                 start = vma->vm_end;
1628                 vma = vma->vm_next;
1629         } while (vma && vma->vm_start < end);
1630 out_unlock:
1631         up_write(&mm->mmap_sem);
1632         mmput(mm);
1633 out:
1634         return ret;
1635 }
1636
1637 /*
1638  * userfaultfd_wake may be used in combination with the
1639  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1640  */
1641 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1642                             unsigned long arg)
1643 {
1644         int ret;
1645         struct uffdio_range uffdio_wake;
1646         struct userfaultfd_wake_range range;
1647         const void __user *buf = (void __user *)arg;
1648
1649         ret = -EFAULT;
1650         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1651                 goto out;
1652
1653         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1654         if (ret)
1655                 goto out;
1656
1657         range.start = uffdio_wake.start;
1658         range.len = uffdio_wake.len;
1659
1660         /*
1661          * len == 0 means wake all and we don't want to wake all here,
1662          * so check it again to be sure.
1663          */
1664         VM_BUG_ON(!range.len);
1665
1666         wake_userfault(ctx, &range);
1667         ret = 0;
1668
1669 out:
1670         return ret;
1671 }
1672
1673 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1674                             unsigned long arg)
1675 {
1676         __s64 ret;
1677         struct uffdio_copy uffdio_copy;
1678         struct uffdio_copy __user *user_uffdio_copy;
1679         struct userfaultfd_wake_range range;
1680
1681         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1682
1683         ret = -EFAULT;
1684         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1685                            /* don't copy "copy" last field */
1686                            sizeof(uffdio_copy)-sizeof(__s64)))
1687                 goto out;
1688
1689         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1690         if (ret)
1691                 goto out;
1692         /*
1693          * double check for wraparound just in case. copy_from_user()
1694          * will later check uffdio_copy.src + uffdio_copy.len to fit
1695          * in the userland range.
1696          */
1697         ret = -EINVAL;
1698         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1699                 goto out;
1700         if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1701                 goto out;
1702         if (mmget_not_zero(ctx->mm)) {
1703                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1704                                    uffdio_copy.len);
1705                 mmput(ctx->mm);
1706         } else {
1707                 return -ESRCH;
1708         }
1709         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1710                 return -EFAULT;
1711         if (ret < 0)
1712                 goto out;
1713         BUG_ON(!ret);
1714         /* len == 0 would wake all */
1715         range.len = ret;
1716         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1717                 range.start = uffdio_copy.dst;
1718                 wake_userfault(ctx, &range);
1719         }
1720         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1721 out:
1722         return ret;
1723 }
1724
1725 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1726                                 unsigned long arg)
1727 {
1728         __s64 ret;
1729         struct uffdio_zeropage uffdio_zeropage;
1730         struct uffdio_zeropage __user *user_uffdio_zeropage;
1731         struct userfaultfd_wake_range range;
1732
1733         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1734
1735         ret = -EFAULT;
1736         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1737                            /* don't copy "zeropage" last field */
1738                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1739                 goto out;
1740
1741         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1742                              uffdio_zeropage.range.len);
1743         if (ret)
1744                 goto out;
1745         ret = -EINVAL;
1746         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1747                 goto out;
1748
1749         if (mmget_not_zero(ctx->mm)) {
1750                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1751                                      uffdio_zeropage.range.len);
1752                 mmput(ctx->mm);
1753         } else {
1754                 return -ESRCH;
1755         }
1756         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1757                 return -EFAULT;
1758         if (ret < 0)
1759                 goto out;
1760         /* len == 0 would wake all */
1761         BUG_ON(!ret);
1762         range.len = ret;
1763         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1764                 range.start = uffdio_zeropage.range.start;
1765                 wake_userfault(ctx, &range);
1766         }
1767         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1768 out:
1769         return ret;
1770 }
1771
1772 static inline unsigned int uffd_ctx_features(__u64 user_features)
1773 {
1774         /*
1775          * For the current set of features the bits just coincide
1776          */
1777         return (unsigned int)user_features;
1778 }
1779
1780 /*
1781  * userland asks for a certain API version and we return which bits
1782  * and ioctl commands are implemented in this kernel for such API
1783  * version or -EINVAL if unknown.
1784  */
1785 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1786                            unsigned long arg)
1787 {
1788         struct uffdio_api uffdio_api;
1789         void __user *buf = (void __user *)arg;
1790         int ret;
1791         __u64 features;
1792
1793         ret = -EINVAL;
1794         if (ctx->state != UFFD_STATE_WAIT_API)
1795                 goto out;
1796         ret = -EFAULT;
1797         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1798                 goto out;
1799         features = uffdio_api.features;
1800         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) {
1801                 memset(&uffdio_api, 0, sizeof(uffdio_api));
1802                 if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1803                         goto out;
1804                 ret = -EINVAL;
1805                 goto out;
1806         }
1807         /* report all available features and ioctls to userland */
1808         uffdio_api.features = UFFD_API_FEATURES;
1809         uffdio_api.ioctls = UFFD_API_IOCTLS;
1810         ret = -EFAULT;
1811         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1812                 goto out;
1813         ctx->state = UFFD_STATE_RUNNING;
1814         /* only enable the requested features for this uffd context */
1815         ctx->features = uffd_ctx_features(features);
1816         ret = 0;
1817 out:
1818         return ret;
1819 }
1820
1821 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1822                               unsigned long arg)
1823 {
1824         int ret = -EINVAL;
1825         struct userfaultfd_ctx *ctx = file->private_data;
1826
1827         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1828                 return -EINVAL;
1829
1830         switch(cmd) {
1831         case UFFDIO_API:
1832                 ret = userfaultfd_api(ctx, arg);
1833                 break;
1834         case UFFDIO_REGISTER:
1835                 ret = userfaultfd_register(ctx, arg);
1836                 break;
1837         case UFFDIO_UNREGISTER:
1838                 ret = userfaultfd_unregister(ctx, arg);
1839                 break;
1840         case UFFDIO_WAKE:
1841                 ret = userfaultfd_wake(ctx, arg);
1842                 break;
1843         case UFFDIO_COPY:
1844                 ret = userfaultfd_copy(ctx, arg);
1845                 break;
1846         case UFFDIO_ZEROPAGE:
1847                 ret = userfaultfd_zeropage(ctx, arg);
1848                 break;
1849         }
1850         return ret;
1851 }
1852
1853 #ifdef CONFIG_PROC_FS
1854 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1855 {
1856         struct userfaultfd_ctx *ctx = f->private_data;
1857         wait_queue_entry_t *wq;
1858         struct userfaultfd_wait_queue *uwq;
1859         unsigned long pending = 0, total = 0;
1860
1861         spin_lock(&ctx->fault_pending_wqh.lock);
1862         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1863                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1864                 pending++;
1865                 total++;
1866         }
1867         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1868                 uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
1869                 total++;
1870         }
1871         spin_unlock(&ctx->fault_pending_wqh.lock);
1872
1873         /*
1874          * If more protocols will be added, there will be all shown
1875          * separated by a space. Like this:
1876          *      protocols: aa:... bb:...
1877          */
1878         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1879                    pending, total, UFFD_API, ctx->features,
1880                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1881 }
1882 #endif
1883
1884 static const struct file_operations userfaultfd_fops = {
1885 #ifdef CONFIG_PROC_FS
1886         .show_fdinfo    = userfaultfd_show_fdinfo,
1887 #endif
1888         .release        = userfaultfd_release,
1889         .poll           = userfaultfd_poll,
1890         .read           = userfaultfd_read,
1891         .unlocked_ioctl = userfaultfd_ioctl,
1892         .compat_ioctl   = userfaultfd_ioctl,
1893         .llseek         = noop_llseek,
1894 };
1895
1896 static void init_once_userfaultfd_ctx(void *mem)
1897 {
1898         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1899
1900         init_waitqueue_head(&ctx->fault_pending_wqh);
1901         init_waitqueue_head(&ctx->fault_wqh);
1902         init_waitqueue_head(&ctx->event_wqh);
1903         init_waitqueue_head(&ctx->fd_wqh);
1904         seqcount_init(&ctx->refile_seq);
1905 }
1906
1907 /**
1908  * userfaultfd_file_create - Creates a userfaultfd file pointer.
1909  * @flags: Flags for the userfaultfd file.
1910  *
1911  * This function creates a userfaultfd file pointer, w/out installing
1912  * it into the fd table. This is useful when the userfaultfd file is
1913  * used during the initialization of data structures that require
1914  * extra setup after the userfaultfd creation. So the userfaultfd
1915  * creation is split into the file pointer creation phase, and the
1916  * file descriptor installation phase.  In this way races with
1917  * userspace closing the newly installed file descriptor can be
1918  * avoided.  Returns a userfaultfd file pointer, or a proper error
1919  * pointer.
1920  */
1921 static struct file *userfaultfd_file_create(int flags)
1922 {
1923         struct file *file;
1924         struct userfaultfd_ctx *ctx;
1925
1926         BUG_ON(!current->mm);
1927
1928         /* Check the UFFD_* constants for consistency.  */
1929         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
1930         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
1931
1932         file = ERR_PTR(-EINVAL);
1933         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
1934                 goto out;
1935
1936         file = ERR_PTR(-ENOMEM);
1937         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
1938         if (!ctx)
1939                 goto out;
1940
1941         atomic_set(&ctx->refcount, 1);
1942         ctx->flags = flags;
1943         ctx->features = 0;
1944         ctx->state = UFFD_STATE_WAIT_API;
1945         ctx->released = false;
1946         ctx->mm = current->mm;
1947         /* prevent the mm struct to be freed */
1948         mmgrab(ctx->mm);
1949
1950         file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
1951                                   O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
1952         if (IS_ERR(file)) {
1953                 mmdrop(ctx->mm);
1954                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
1955         }
1956 out:
1957         return file;
1958 }
1959
1960 SYSCALL_DEFINE1(userfaultfd, int, flags)
1961 {
1962         int fd, error;
1963         struct file *file;
1964
1965         error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
1966         if (error < 0)
1967                 return error;
1968         fd = error;
1969
1970         file = userfaultfd_file_create(flags);
1971         if (IS_ERR(file)) {
1972                 error = PTR_ERR(file);
1973                 goto err_put_unused_fd;
1974         }
1975         fd_install(fd, file);
1976
1977         return fd;
1978
1979 err_put_unused_fd:
1980         put_unused_fd(fd);
1981
1982         return error;
1983 }
1984
1985 static int __init userfaultfd_init(void)
1986 {
1987         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
1988                                                 sizeof(struct userfaultfd_ctx),
1989                                                 0,
1990                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
1991                                                 init_once_userfaultfd_ctx);
1992         return 0;
1993 }
1994 __initcall(userfaultfd_init);