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