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