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