fs/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  fs/userfaultfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   6  *  Copyright (C) 2008-2009 Red Hat, Inc.
   7  *  Copyright (C) 2015  Red Hat, Inc.
   8  *
   9  *  Some part derived from fs/eventfd.c (anon inode setup) and
  10  *  mm/ksm.c (mm hashing).
  11  */
  12
  13 #include <linux/list.h>
  14 #include <linux/hashtable.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/mm.h>
  18 #include <linux/poll.h>
  19 #include <linux/slab.h>
  20 #include <linux/seq_file.h>
  21 #include <linux/file.h>
  22 #include <linux/bug.h>
  23 #include <linux/anon_inodes.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/userfaultfd_k.h>
  26 #include <linux/mempolicy.h>
  27 #include <linux/ioctl.h>
  28 #include <linux/security.h>
  29 #include <linux/hugetlb.h>
  30
  31 int sysctl_unprivileged_userfaultfd __read_mostly = 1;
  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  * Locking order:
  45  *      fd_wqh.lock
  46  *              fault_pending_wqh.lock
  47  *                      fault_wqh.lock
  48  *              event_wqh.lock
  49  *
  50  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
  51  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
  52  * also taken in IRQ context.
  53  */
  54 struct userfaultfd_ctx {
  55         /* waitqueue head for the pending (i.e. not read) userfaults */
  56         wait_queue_head_t fault_pending_wqh;
  57         /* waitqueue head for the userfaults */
  58         wait_queue_head_t fault_wqh;
  59         /* waitqueue head for the pseudo fd to wakeup poll/read */
  60         wait_queue_head_t fd_wqh;
  61         /* waitqueue head for events */
  62         wait_queue_head_t event_wqh;
  63         /* a refile sequence protected by fault_pending_wqh lock */
  64         struct seqcount refile_seq;
  65         /* pseudo fd refcounting */
  66         refcount_t refcount;
  67         /* userfaultfd syscall flags */
  68         unsigned int flags;
  69         /* features requested from the userspace */
  70         unsigned int features;
  71         /* state machine */
  72         enum userfaultfd_state state;
  73         /* released */
  74         bool released;
  75         /* memory mappings are changing because of non-cooperative event */
  76         bool mmap_changing;
  77         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  78         struct mm_struct *mm;
  79 };
  80
  81 struct userfaultfd_fork_ctx {
  82         struct userfaultfd_ctx *orig;
  83         struct userfaultfd_ctx *new;
  84         struct list_head list;
  85 };
  86
  87 struct userfaultfd_unmap_ctx {
  88         struct userfaultfd_ctx *ctx;
  89         unsigned long start;
  90         unsigned long end;
  91         struct list_head list;
  92 };
  93
  94 struct userfaultfd_wait_queue {
  95         struct uffd_msg msg;
  96         wait_queue_entry_t wq;
  97         struct userfaultfd_ctx *ctx;
  98         bool waken;
  99 };
 100
 101 struct userfaultfd_wake_range {
 102         unsigned long start;
 103         unsigned long len;
 104 };
 105
 106 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 107                                      int wake_flags, void *key)
 108 {
 109         struct userfaultfd_wake_range *range = key;
 110         int ret;
 111         struct userfaultfd_wait_queue *uwq;
 112         unsigned long start, len;
 113
 114         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 115         ret = 0;
 116         /* len == 0 means wake all */
 117         start = range->start;
 118         len = range->len;
 119         if (len && (start > uwq->msg.arg.pagefault.address ||
 120                     start + len <= uwq->msg.arg.pagefault.address))
 121                 goto out;
 122         WRITE_ONCE(uwq->waken, true);
 123         /*
 124          * The Program-Order guarantees provided by the scheduler
 125          * ensure uwq->waken is visible before the task is woken.
 126          */
 127         ret = wake_up_state(wq->private, mode);
 128         if (ret) {
 129                 /*
 130                  * Wake only once, autoremove behavior.
 131                  *
 132                  * After the effect of list_del_init is visible to the other
 133                  * CPUs, the waitqueue may disappear from under us, see the
 134                  * !list_empty_careful() in handle_userfault().
 135                  *
 136                  * try_to_wake_up() has an implicit smp_mb(), and the
 137                  * wq->private is read before calling the extern function
 138                  * "wake_up_state" (which in turns calls try_to_wake_up).
 139                  */
 140                 list_del_init(&wq->entry);
 141         }
 142 out:
 143         return ret;
 144 }
 145
 146 /**
 147  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 148  * context.
 149  * @ctx: [in] Pointer to the userfaultfd context.
 150  */
 151 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 152 {
 153         refcount_inc(&ctx->refcount);
 154 }
 155
 156 /**
 157  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 158  * context.
 159  * @ctx: [in] Pointer to userfaultfd context.
 160  *
 161  * The userfaultfd context reference must have been previously acquired either
 162  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 163  */
 164 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 165 {
 166         if (refcount_dec_and_test(&ctx->refcount)) {
 167                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 168                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 169                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 170                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 171                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 172                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 173                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 174                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 175                 mmdrop(ctx->mm);
 176                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 177         }
 178 }
 179
 180 static inline void msg_init(struct uffd_msg *msg)
 181 {
 182         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 183         /*
 184          * Must use memset to zero out the paddings or kernel data is
 185          * leaked to userland.
 186          */
 187         memset(msg, 0, sizeof(struct uffd_msg));
 188 }
 189
 190 static inline struct uffd_msg userfault_msg(unsigned long address,
 191                                             unsigned int flags,
 192                                             unsigned long reason,
 193                                             unsigned int features)
 194 {
 195         struct uffd_msg msg;
 196         msg_init(&msg);
 197         msg.event = UFFD_EVENT_PAGEFAULT;
 198         msg.arg.pagefault.address = address;
 199         if (flags & FAULT_FLAG_WRITE)
 200                 /*
 201                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 202                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
 203                  * was not set in a UFFD_EVENT_PAGEFAULT, it means it
 204                  * was a read fault, otherwise if set it means it's
 205                  * a write fault.
 206                  */
 207                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 208         if (reason & VM_UFFD_WP)
 209                 /*
 210                  * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
 211                  * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
 212                  * not set in a UFFD_EVENT_PAGEFAULT, it means it was
 213                  * a missing fault, otherwise if set it means it's a
 214                  * write protect fault.
 215                  */
 216                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 217         if (features & UFFD_FEATURE_THREAD_ID)
 218                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 219         return msg;
 220 }
 221
 222 #ifdef CONFIG_HUGETLB_PAGE
 223 /*
 224  * Same functionality as userfaultfd_must_wait below with modifications for
 225  * hugepmd ranges.
 226  */
 227 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 228                                          struct vm_area_struct *vma,
 229                                          unsigned long address,
 230                                          unsigned long flags,
 231                                          unsigned long reason)
 232 {
 233         struct mm_struct *mm = ctx->mm;
 234         pte_t *ptep, pte;
 235         bool ret = true;
 236
 237         mmap_assert_locked(mm);
 238
 239         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 240
 241         if (!ptep)
 242                 goto out;
 243
 244         ret = false;
 245         pte = huge_ptep_get(ptep);
 246
 247         /*
 248          * Lockless access: we're in a wait_event so it's ok if it
 249          * changes under us.
 250          */
 251         if (huge_pte_none(pte))
 252                 ret = true;
 253         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 254                 ret = true;
 255 out:
 256         return ret;
 257 }
 258 #else
 259 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 260                                          struct vm_area_struct *vma,
 261                                          unsigned long address,
 262                                          unsigned long flags,
 263                                          unsigned long reason)
 264 {
 265         return false;   /* should never get here */
 266 }
 267 #endif /* CONFIG_HUGETLB_PAGE */
 268
 269 /*
 270  * Verify the pagetables are still not ok after having reigstered into
 271  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 272  * userfault that has already been resolved, if userfaultfd_read and
 273  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 274  * threads.
 275  */
 276 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 277                                          unsigned long address,
 278                                          unsigned long flags,
 279                                          unsigned long reason)
 280 {
 281         struct mm_struct *mm = ctx->mm;
 282         pgd_t *pgd;
 283         p4d_t *p4d;
 284         pud_t *pud;
 285         pmd_t *pmd, _pmd;
 286         pte_t *pte;
 287         bool ret = true;
 288
 289         mmap_assert_locked(mm);
 290
 291         pgd = pgd_offset(mm, address);
 292         if (!pgd_present(*pgd))
 293                 goto out;
 294         p4d = p4d_offset(pgd, address);
 295         if (!p4d_present(*p4d))
 296                 goto out;
 297         pud = pud_offset(p4d, address);
 298         if (!pud_present(*pud))
 299                 goto out;
 300         pmd = pmd_offset(pud, address);
 301         /*
 302          * READ_ONCE must function as a barrier with narrower scope
 303          * and it must be equivalent to:
 304          *      _pmd = *pmd; barrier();
 305          *
 306          * This is to deal with the instability (as in
 307          * pmd_trans_unstable) of the pmd.
 308          */
 309         _pmd = READ_ONCE(*pmd);
 310         if (pmd_none(_pmd))
 311                 goto out;
 312
 313         ret = false;
 314         if (!pmd_present(_pmd))
 315                 goto out;
 316
 317         if (pmd_trans_huge(_pmd)) {
 318                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 319                         ret = true;
 320                 goto out;
 321         }
 322
 323         /*
 324          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 325          * and use the standard pte_offset_map() instead of parsing _pmd.
 326          */
 327         pte = pte_offset_map(pmd, address);
 328         /*
 329          * Lockless access: we're in a wait_event so it's ok if it
 330          * changes under us.
 331          */
 332         if (pte_none(*pte))
 333                 ret = true;
 334         if (!pte_write(*pte) && (reason & VM_UFFD_WP))
 335                 ret = true;
 336         pte_unmap(pte);
 337
 338 out:
 339         return ret;
 340 }
 341
 342 /* Should pair with userfaultfd_signal_pending() */
 343 static inline long userfaultfd_get_blocking_state(unsigned int flags)
 344 {
 345         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 346                 return TASK_INTERRUPTIBLE;
 347
 348         if (flags & FAULT_FLAG_KILLABLE)
 349                 return TASK_KILLABLE;
 350
 351         return TASK_UNINTERRUPTIBLE;
 352 }
 353
 354 /* Should pair with userfaultfd_get_blocking_state() */
 355 static inline bool userfaultfd_signal_pending(unsigned int flags)
 356 {
 357         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 358                 return signal_pending(current);
 359
 360         if (flags & FAULT_FLAG_KILLABLE)
 361                 return fatal_signal_pending(current);
 362
 363         return false;
 364 }
 365
 366 /*
 367  * The locking rules involved in returning VM_FAULT_RETRY depending on
 368  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 369  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 370  * recommendation in __lock_page_or_retry is not an understatement.
 371  *
 372  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
 373  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 374  * not set.
 375  *
 376  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 377  * set, VM_FAULT_RETRY can still be returned if and only if there are
 378  * fatal_signal_pending()s, and the mmap_lock must be released before
 379  * returning it.
 380  */
 381 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 382 {
 383         struct mm_struct *mm = vmf->vma->vm_mm;
 384         struct userfaultfd_ctx *ctx;
 385         struct userfaultfd_wait_queue uwq;
 386         vm_fault_t ret = VM_FAULT_SIGBUS;
 387         bool must_wait;
 388         long blocking_state;
 389
 390         /*
 391          * We don't do userfault handling for the final child pid update.
 392          *
 393          * We also don't do userfault handling during
 394          * coredumping. hugetlbfs has the special
 395          * follow_hugetlb_page() to skip missing pages in the
 396          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 397          * the no_page_table() helper in follow_page_mask(), but the
 398          * shmem_vm_ops->fault method is invoked even during
 399          * coredumping without mmap_lock and it ends up here.
 400          */
 401         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 402                 goto out;
 403
 404         /*
 405          * Coredumping runs without mmap_lock so we can only check that
 406          * the mmap_lock is held, if PF_DUMPCORE was not set.
 407          */
 408         mmap_assert_locked(mm);
 409
 410         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 411         if (!ctx)
 412                 goto out;
 413
 414         BUG_ON(ctx->mm != mm);
 415
 416         VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
 417         VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
 418
 419         if (ctx->features & UFFD_FEATURE_SIGBUS)
 420                 goto out;
 421
 422         /*
 423          * If it's already released don't get it. This avoids to loop
 424          * in __get_user_pages if userfaultfd_release waits on the
 425          * caller of handle_userfault to release the mmap_lock.
 426          */
 427         if (unlikely(READ_ONCE(ctx->released))) {
 428                 /*
 429                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 430                  * cooperative manager can close the uffd after the
 431                  * last UFFDIO_COPY, without risking to trigger an
 432                  * involuntary SIGBUS if the process was starting the
 433                  * userfaultfd while the userfaultfd was still armed
 434                  * (but after the last UFFDIO_COPY). If the uffd
 435                  * wasn't already closed when the userfault reached
 436                  * this point, that would normally be solved by
 437                  * userfaultfd_must_wait returning 'false'.
 438                  *
 439                  * If we were to return VM_FAULT_SIGBUS here, the non
 440                  * cooperative manager would be instead forced to
 441                  * always call UFFDIO_UNREGISTER before it can safely
 442                  * close the uffd.
 443                  */
 444                 ret = VM_FAULT_NOPAGE;
 445                 goto out;
 446         }
 447
 448         /*
 449          * Check that we can return VM_FAULT_RETRY.
 450          *
 451          * NOTE: it should become possible to return VM_FAULT_RETRY
 452          * even if FAULT_FLAG_TRIED is set without leading to gup()
 453          * -EBUSY failures, if the userfaultfd is to be extended for
 454          * VM_UFFD_WP tracking and we intend to arm the userfault
 455          * without first stopping userland access to the memory. For
 456          * VM_UFFD_MISSING userfaults this is enough for now.
 457          */
 458         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 459                 /*
 460                  * Validate the invariant that nowait must allow retry
 461                  * to be sure not to return SIGBUS erroneously on
 462                  * nowait invocations.
 463                  */
 464                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 465 #ifdef CONFIG_DEBUG_VM
 466                 if (printk_ratelimit()) {
 467                         printk(KERN_WARNING
 468                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 469                                vmf->flags);
 470                         dump_stack();
 471                 }
 472 #endif
 473                 goto out;
 474         }
 475
 476         /*
 477          * Handle nowait, not much to do other than tell it to retry
 478          * and wait.
 479          */
 480         ret = VM_FAULT_RETRY;
 481         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 482                 goto out;
 483
 484         /* take the reference before dropping the mmap_lock */
 485         userfaultfd_ctx_get(ctx);
 486
 487         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 488         uwq.wq.private = current;
 489         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 490                         ctx->features);
 491         uwq.ctx = ctx;
 492         uwq.waken = false;
 493
 494         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 495
 496         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 497         /*
 498          * After the __add_wait_queue the uwq is visible to userland
 499          * through poll/read().
 500          */
 501         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 502         /*
 503          * The smp_mb() after __set_current_state prevents the reads
 504          * following the spin_unlock to happen before the list_add in
 505          * __add_wait_queue.
 506          */
 507         set_current_state(blocking_state);
 508         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 509
 510         if (!is_vm_hugetlb_page(vmf->vma))
 511                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 512                                                   reason);
 513         else
 514                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 515                                                        vmf->address,
 516                                                        vmf->flags, reason);
 517         mmap_read_unlock(mm);
 518
 519         if (likely(must_wait && !READ_ONCE(ctx->released) &&
 520                    !userfaultfd_signal_pending(vmf->flags))) {
 521                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 522                 schedule();
 523                 ret |= VM_FAULT_MAJOR;
 524
 525                 /*
 526                  * False wakeups can orginate even from rwsem before
 527                  * up_read() however userfaults will wait either for a
 528                  * targeted wakeup on the specific uwq waitqueue from
 529                  * wake_userfault() or for signals or for uffd
 530                  * release.
 531                  */
 532                 while (!READ_ONCE(uwq.waken)) {
 533                         /*
 534                          * This needs the full smp_store_mb()
 535                          * guarantee as the state write must be
 536                          * visible to other CPUs before reading
 537                          * uwq.waken from other CPUs.
 538                          */
 539                         set_current_state(blocking_state);
 540                         if (READ_ONCE(uwq.waken) ||
 541                             READ_ONCE(ctx->released) ||
 542                             userfaultfd_signal_pending(vmf->flags))
 543                                 break;
 544                         schedule();
 545                 }
 546         }
 547
 548         __set_current_state(TASK_RUNNING);
 549
 550         /*
 551          * Here we race with the list_del; list_add in
 552          * userfaultfd_ctx_read(), however because we don't ever run
 553          * list_del_init() to refile across the two lists, the prev
 554          * and next pointers will never point to self. list_add also
 555          * would never let any of the two pointers to point to
 556          * self. So list_empty_careful won't risk to see both pointers
 557          * pointing to self at any time during the list refile. The
 558          * only case where list_del_init() is called is the full
 559          * removal in the wake function and there we don't re-list_add
 560          * and it's fine not to block on the spinlock. The uwq on this
 561          * kernel stack can be released after the list_del_init.
 562          */
 563         if (!list_empty_careful(&uwq.wq.entry)) {
 564                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 565                 /*
 566                  * No need of list_del_init(), the uwq on the stack
 567                  * will be freed shortly anyway.
 568                  */
 569                 list_del(&uwq.wq.entry);
 570                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 571         }
 572
 573         /*
 574          * ctx may go away after this if the userfault pseudo fd is
 575          * already released.
 576          */
 577         userfaultfd_ctx_put(ctx);
 578
 579 out:
 580         return ret;
 581 }
 582
 583 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 584                                               struct userfaultfd_wait_queue *ewq)
 585 {
 586         struct userfaultfd_ctx *release_new_ctx;
 587
 588         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 589                 goto out;
 590
 591         ewq->ctx = ctx;
 592         init_waitqueue_entry(&ewq->wq, current);
 593         release_new_ctx = NULL;
 594
 595         spin_lock_irq(&ctx->event_wqh.lock);
 596         /*
 597          * After the __add_wait_queue the uwq is visible to userland
 598          * through poll/read().
 599          */
 600         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 601         for (;;) {
 602                 set_current_state(TASK_KILLABLE);
 603                 if (ewq->msg.event == 0)
 604                         break;
 605                 if (READ_ONCE(ctx->released) ||
 606                     fatal_signal_pending(current)) {
 607                         /*
 608                          * &ewq->wq may be queued in fork_event, but
 609                          * __remove_wait_queue ignores the head
 610                          * parameter. It would be a problem if it
 611                          * didn't.
 612                          */
 613                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 614                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 615                                 struct userfaultfd_ctx *new;
 616
 617                                 new = (struct userfaultfd_ctx *)
 618                                         (unsigned long)
 619                                         ewq->msg.arg.reserved.reserved1;
 620                                 release_new_ctx = new;
 621                         }
 622                         break;
 623                 }
 624
 625                 spin_unlock_irq(&ctx->event_wqh.lock);
 626
 627                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 628                 schedule();
 629
 630                 spin_lock_irq(&ctx->event_wqh.lock);
 631         }
 632         __set_current_state(TASK_RUNNING);
 633         spin_unlock_irq(&ctx->event_wqh.lock);
 634
 635         if (release_new_ctx) {
 636                 struct vm_area_struct *vma;
 637                 struct mm_struct *mm = release_new_ctx->mm;
 638
 639                 /* the various vma->vm_userfaultfd_ctx still points to it */
 640                 mmap_write_lock(mm);
 641                 /* no task can run (and in turn coredump) yet */
 642                 VM_WARN_ON(!mmget_still_valid(mm));
 643                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 644                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 645                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 646                                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 647                         }
 648                 mmap_write_unlock(mm);
 649
 650                 userfaultfd_ctx_put(release_new_ctx);
 651         }
 652
 653         /*
 654          * ctx may go away after this if the userfault pseudo fd is
 655          * already released.
 656          */
 657 out:
 658         WRITE_ONCE(ctx->mmap_changing, false);
 659         userfaultfd_ctx_put(ctx);
 660 }
 661
 662 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 663                                        struct userfaultfd_wait_queue *ewq)
 664 {
 665         ewq->msg.event = 0;
 666         wake_up_locked(&ctx->event_wqh);
 667         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 668 }
 669
 670 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 671 {
 672         struct userfaultfd_ctx *ctx = NULL, *octx;
 673         struct userfaultfd_fork_ctx *fctx;
 674
 675         octx = vma->vm_userfaultfd_ctx.ctx;
 676         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 677                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 678                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 679                 return 0;
 680         }
 681
 682         list_for_each_entry(fctx, fcs, list)
 683                 if (fctx->orig == octx) {
 684                         ctx = fctx->new;
 685                         break;
 686                 }
 687
 688         if (!ctx) {
 689                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 690                 if (!fctx)
 691                         return -ENOMEM;
 692
 693                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 694                 if (!ctx) {
 695                         kfree(fctx);
 696                         return -ENOMEM;
 697                 }
 698
 699                 refcount_set(&ctx->refcount, 1);
 700                 ctx->flags = octx->flags;
 701                 ctx->state = UFFD_STATE_RUNNING;
 702                 ctx->features = octx->features;
 703                 ctx->released = false;
 704                 ctx->mmap_changing = false;
 705                 ctx->mm = vma->vm_mm;
 706                 mmgrab(ctx->mm);
 707
 708                 userfaultfd_ctx_get(octx);
 709                 WRITE_ONCE(octx->mmap_changing, true);
 710                 fctx->orig = octx;
 711                 fctx->new = ctx;
 712                 list_add_tail(&fctx->list, fcs);
 713         }
 714
 715         vma->vm_userfaultfd_ctx.ctx = ctx;
 716         return 0;
 717 }
 718
 719 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 720 {
 721         struct userfaultfd_ctx *ctx = fctx->orig;
 722         struct userfaultfd_wait_queue ewq;
 723
 724         msg_init(&ewq.msg);
 725
 726         ewq.msg.event = UFFD_EVENT_FORK;
 727         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 728
 729         userfaultfd_event_wait_completion(ctx, &ewq);
 730 }
 731
 732 void dup_userfaultfd_complete(struct list_head *fcs)
 733 {
 734         struct userfaultfd_fork_ctx *fctx, *n;
 735
 736         list_for_each_entry_safe(fctx, n, fcs, list) {
 737                 dup_fctx(fctx);
 738                 list_del(&fctx->list);
 739                 kfree(fctx);
 740         }
 741 }
 742
 743 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 744                              struct vm_userfaultfd_ctx *vm_ctx)
 745 {
 746         struct userfaultfd_ctx *ctx;
 747
 748         ctx = vma->vm_userfaultfd_ctx.ctx;
 749
 750         if (!ctx)
 751                 return;
 752
 753         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 754                 vm_ctx->ctx = ctx;
 755                 userfaultfd_ctx_get(ctx);
 756                 WRITE_ONCE(ctx->mmap_changing, true);
 757         } else {
 758                 /* Drop uffd context if remap feature not enabled */
 759                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 760                 vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
 761         }
 762 }
 763
 764 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 765                                  unsigned long from, unsigned long to,
 766                                  unsigned long len)
 767 {
 768         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 769         struct userfaultfd_wait_queue ewq;
 770
 771         if (!ctx)
 772                 return;
 773
 774         if (to & ~PAGE_MASK) {
 775                 userfaultfd_ctx_put(ctx);
 776                 return;
 777         }
 778
 779         msg_init(&ewq.msg);
 780
 781         ewq.msg.event = UFFD_EVENT_REMAP;
 782         ewq.msg.arg.remap.from = from;
 783         ewq.msg.arg.remap.to = to;
 784         ewq.msg.arg.remap.len = len;
 785
 786         userfaultfd_event_wait_completion(ctx, &ewq);
 787 }
 788
 789 bool userfaultfd_remove(struct vm_area_struct *vma,
 790                         unsigned long start, unsigned long end)
 791 {
 792         struct mm_struct *mm = vma->vm_mm;
 793         struct userfaultfd_ctx *ctx;
 794         struct userfaultfd_wait_queue ewq;
 795
 796         ctx = vma->vm_userfaultfd_ctx.ctx;
 797         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 798                 return true;
 799
 800         userfaultfd_ctx_get(ctx);
 801         WRITE_ONCE(ctx->mmap_changing, true);
 802         mmap_read_unlock(mm);
 803
 804         msg_init(&ewq.msg);
 805
 806         ewq.msg.event = UFFD_EVENT_REMOVE;
 807         ewq.msg.arg.remove.start = start;
 808         ewq.msg.arg.remove.end = end;
 809
 810         userfaultfd_event_wait_completion(ctx, &ewq);
 811
 812         return false;
 813 }
 814
 815 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 816                           unsigned long start, unsigned long end)
 817 {
 818         struct userfaultfd_unmap_ctx *unmap_ctx;
 819
 820         list_for_each_entry(unmap_ctx, unmaps, list)
 821                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 822                     unmap_ctx->end == end)
 823                         return true;
 824
 825         return false;
 826 }
 827
 828 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 829                            unsigned long start, unsigned long end,
 830                            struct list_head *unmaps)
 831 {
 832         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 833                 struct userfaultfd_unmap_ctx *unmap_ctx;
 834                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 835
 836                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 837                     has_unmap_ctx(ctx, unmaps, start, end))
 838                         continue;
 839
 840                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 841                 if (!unmap_ctx)
 842                         return -ENOMEM;
 843
 844                 userfaultfd_ctx_get(ctx);
 845                 WRITE_ONCE(ctx->mmap_changing, true);
 846                 unmap_ctx->ctx = ctx;
 847                 unmap_ctx->start = start;
 848                 unmap_ctx->end = end;
 849                 list_add_tail(&unmap_ctx->list, unmaps);
 850         }
 851
 852         return 0;
 853 }
 854
 855 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 856 {
 857         struct userfaultfd_unmap_ctx *ctx, *n;
 858         struct userfaultfd_wait_queue ewq;
 859
 860         list_for_each_entry_safe(ctx, n, uf, list) {
 861                 msg_init(&ewq.msg);
 862
 863                 ewq.msg.event = UFFD_EVENT_UNMAP;
 864                 ewq.msg.arg.remove.start = ctx->start;
 865                 ewq.msg.arg.remove.end = ctx->end;
 866
 867                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 868
 869                 list_del(&ctx->list);
 870                 kfree(ctx);
 871         }
 872 }
 873
 874 static int userfaultfd_release(struct inode *inode, struct file *file)
 875 {
 876         struct userfaultfd_ctx *ctx = file->private_data;
 877         struct mm_struct *mm = ctx->mm;
 878         struct vm_area_struct *vma, *prev;
 879         /* len == 0 means wake all */
 880         struct userfaultfd_wake_range range = { .len = 0, };
 881         unsigned long new_flags;
 882         bool still_valid;
 883
 884         WRITE_ONCE(ctx->released, true);
 885
 886         if (!mmget_not_zero(mm))
 887                 goto wakeup;
 888
 889         /*
 890          * Flush page faults out of all CPUs. NOTE: all page faults
 891          * must be retried without returning VM_FAULT_SIGBUS if
 892          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 893          * changes while handle_userfault released the mmap_lock. So
 894          * it's critical that released is set to true (above), before
 895          * taking the mmap_lock for writing.
 896          */
 897         mmap_write_lock(mm);
 898         still_valid = mmget_still_valid(mm);
 899         prev = NULL;
 900         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 901                 cond_resched();
 902                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 903                        !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
 904                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 905                         prev = vma;
 906                         continue;
 907                 }
 908                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
 909                 if (still_valid) {
 910                         prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 911                                          new_flags, vma->anon_vma,
 912                                          vma->vm_file, vma->vm_pgoff,
 913                                          vma_policy(vma),
 914                                          NULL_VM_UFFD_CTX);
 915                         if (prev)
 916                                 vma = prev;
 917                         else
 918                                 prev = vma;
 919                 }
 920                 vma->vm_flags = new_flags;
 921                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 922         }
 923         mmap_write_unlock(mm);
 924         mmput(mm);
 925 wakeup:
 926         /*
 927          * After no new page faults can wait on this fault_*wqh, flush
 928          * the last page faults that may have been already waiting on
 929          * the fault_*wqh.
 930          */
 931         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 932         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 933         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 934         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 935
 936         /* Flush pending events that may still wait on event_wqh */
 937         wake_up_all(&ctx->event_wqh);
 938
 939         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 940         userfaultfd_ctx_put(ctx);
 941         return 0;
 942 }
 943
 944 /* fault_pending_wqh.lock must be hold by the caller */
 945 static inline struct userfaultfd_wait_queue *find_userfault_in(
 946                 wait_queue_head_t *wqh)
 947 {
 948         wait_queue_entry_t *wq;
 949         struct userfaultfd_wait_queue *uwq;
 950
 951         lockdep_assert_held(&wqh->lock);
 952
 953         uwq = NULL;
 954         if (!waitqueue_active(wqh))
 955                 goto out;
 956         /* walk in reverse to provide FIFO behavior to read userfaults */
 957         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 958         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 959 out:
 960         return uwq;
 961 }
 962
 963 static inline struct userfaultfd_wait_queue *find_userfault(
 964                 struct userfaultfd_ctx *ctx)
 965 {
 966         return find_userfault_in(&ctx->fault_pending_wqh);
 967 }
 968
 969 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 970                 struct userfaultfd_ctx *ctx)
 971 {
 972         return find_userfault_in(&ctx->event_wqh);
 973 }
 974
 975 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 976 {
 977         struct userfaultfd_ctx *ctx = file->private_data;
 978         __poll_t ret;
 979
 980         poll_wait(file, &ctx->fd_wqh, wait);
 981
 982         switch (ctx->state) {
 983         case UFFD_STATE_WAIT_API:
 984                 return EPOLLERR;
 985         case UFFD_STATE_RUNNING:
 986                 /*
 987                  * poll() never guarantees that read won't block.
 988                  * userfaults can be waken before they're read().
 989                  */
 990                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 991                         return EPOLLERR;
 992                 /*
 993                  * lockless access to see if there are pending faults
 994                  * __pollwait last action is the add_wait_queue but
 995                  * the spin_unlock would allow the waitqueue_active to
 996                  * pass above the actual list_add inside
 997                  * add_wait_queue critical section. So use a full
 998                  * memory barrier to serialize the list_add write of
 999                  * add_wait_queue() with the waitqueue_active read
1000                  * below.
1001                  */
1002                 ret = 0;
1003                 smp_mb();
1004                 if (waitqueue_active(&ctx->fault_pending_wqh))
1005                         ret = EPOLLIN;
1006                 else if (waitqueue_active(&ctx->event_wqh))
1007                         ret = EPOLLIN;
1008
1009                 return ret;
1010         default:
1011                 WARN_ON_ONCE(1);
1012                 return EPOLLERR;
1013         }
1014 }
1015
1016 static const struct file_operations userfaultfd_fops;
1017
1018 static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
1019                                   struct userfaultfd_ctx *new,
1020                                   struct uffd_msg *msg)
1021 {
1022         int fd;
1023
1024         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1025                               O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
1026         if (fd < 0)
1027                 return fd;
1028
1029         msg->arg.reserved.reserved1 = 0;
1030         msg->arg.fork.ufd = fd;
1031         return 0;
1032 }
1033
1034 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1035                                     struct uffd_msg *msg)
1036 {
1037         ssize_t ret;
1038         DECLARE_WAITQUEUE(wait, current);
1039         struct userfaultfd_wait_queue *uwq;
1040         /*
1041          * Handling fork event requires sleeping operations, so
1042          * we drop the event_wqh lock, then do these ops, then
1043          * lock it back and wake up the waiter. While the lock is
1044          * dropped the ewq may go away so we keep track of it
1045          * carefully.
1046          */
1047         LIST_HEAD(fork_event);
1048         struct userfaultfd_ctx *fork_nctx = NULL;
1049
1050         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1051         spin_lock_irq(&ctx->fd_wqh.lock);
1052         __add_wait_queue(&ctx->fd_wqh, &wait);
1053         for (;;) {
1054                 set_current_state(TASK_INTERRUPTIBLE);
1055                 spin_lock(&ctx->fault_pending_wqh.lock);
1056                 uwq = find_userfault(ctx);
1057                 if (uwq) {
1058                         /*
1059                          * Use a seqcount to repeat the lockless check
1060                          * in wake_userfault() to avoid missing
1061                          * wakeups because during the refile both
1062                          * waitqueue could become empty if this is the
1063                          * only userfault.
1064                          */
1065                         write_seqcount_begin(&ctx->refile_seq);
1066
1067                         /*
1068                          * The fault_pending_wqh.lock prevents the uwq
1069                          * to disappear from under us.
1070                          *
1071                          * Refile this userfault from
1072                          * fault_pending_wqh to fault_wqh, it's not
1073                          * pending anymore after we read it.
1074                          *
1075                          * Use list_del() by hand (as
1076                          * userfaultfd_wake_function also uses
1077                          * list_del_init() by hand) to be sure nobody
1078                          * changes __remove_wait_queue() to use
1079                          * list_del_init() in turn breaking the
1080                          * !list_empty_careful() check in
1081                          * handle_userfault(). The uwq->wq.head list
1082                          * must never be empty at any time during the
1083                          * refile, or the waitqueue could disappear
1084                          * from under us. The "wait_queue_head_t"
1085                          * parameter of __remove_wait_queue() is unused
1086                          * anyway.
1087                          */
1088                         list_del(&uwq->wq.entry);
1089                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1090
1091                         write_seqcount_end(&ctx->refile_seq);
1092
1093                         /* careful to always initialize msg if ret == 0 */
1094                         *msg = uwq->msg;
1095                         spin_unlock(&ctx->fault_pending_wqh.lock);
1096                         ret = 0;
1097                         break;
1098                 }
1099                 spin_unlock(&ctx->fault_pending_wqh.lock);
1100
1101                 spin_lock(&ctx->event_wqh.lock);
1102                 uwq = find_userfault_evt(ctx);
1103                 if (uwq) {
1104                         *msg = uwq->msg;
1105
1106                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1107                                 fork_nctx = (struct userfaultfd_ctx *)
1108                                         (unsigned long)
1109                                         uwq->msg.arg.reserved.reserved1;
1110                                 list_move(&uwq->wq.entry, &fork_event);
1111                                 /*
1112                                  * fork_nctx can be freed as soon as
1113                                  * we drop the lock, unless we take a
1114                                  * reference on it.
1115                                  */
1116                                 userfaultfd_ctx_get(fork_nctx);
1117                                 spin_unlock(&ctx->event_wqh.lock);
1118                                 ret = 0;
1119                                 break;
1120                         }
1121
1122                         userfaultfd_event_complete(ctx, uwq);
1123                         spin_unlock(&ctx->event_wqh.lock);
1124                         ret = 0;
1125                         break;
1126                 }
1127                 spin_unlock(&ctx->event_wqh.lock);
1128
1129                 if (signal_pending(current)) {
1130                         ret = -ERESTARTSYS;
1131                         break;
1132                 }
1133                 if (no_wait) {
1134                         ret = -EAGAIN;
1135                         break;
1136                 }
1137                 spin_unlock_irq(&ctx->fd_wqh.lock);
1138                 schedule();
1139                 spin_lock_irq(&ctx->fd_wqh.lock);
1140         }
1141         __remove_wait_queue(&ctx->fd_wqh, &wait);
1142         __set_current_state(TASK_RUNNING);
1143         spin_unlock_irq(&ctx->fd_wqh.lock);
1144
1145         if (!ret && msg->event == UFFD_EVENT_FORK) {
1146                 ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1147                 spin_lock_irq(&ctx->event_wqh.lock);
1148                 if (!list_empty(&fork_event)) {
1149                         /*
1150                          * The fork thread didn't abort, so we can
1151                          * drop the temporary refcount.
1152                          */
1153                         userfaultfd_ctx_put(fork_nctx);
1154
1155                         uwq = list_first_entry(&fork_event,
1156                                                typeof(*uwq),
1157                                                wq.entry);
1158                         /*
1159                          * If fork_event list wasn't empty and in turn
1160                          * the event wasn't already released by fork
1161                          * (the event is allocated on fork kernel
1162                          * stack), put the event back to its place in
1163                          * the event_wq. fork_event head will be freed
1164                          * as soon as we return so the event cannot
1165                          * stay queued there no matter the current
1166                          * "ret" value.
1167                          */
1168                         list_del(&uwq->wq.entry);
1169                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1170
1171                         /*
1172                          * Leave the event in the waitqueue and report
1173                          * error to userland if we failed to resolve
1174                          * the userfault fork.
1175                          */
1176                         if (likely(!ret))
1177                                 userfaultfd_event_complete(ctx, uwq);
1178                 } else {
1179                         /*
1180                          * Here the fork thread aborted and the
1181                          * refcount from the fork thread on fork_nctx
1182                          * has already been released. We still hold
1183                          * the reference we took before releasing the
1184                          * lock above. If resolve_userfault_fork
1185                          * failed we've to drop it because the
1186                          * fork_nctx has to be freed in such case. If
1187                          * it succeeded we'll hold it because the new
1188                          * uffd references it.
1189                          */
1190                         if (ret)
1191                                 userfaultfd_ctx_put(fork_nctx);
1192                 }
1193                 spin_unlock_irq(&ctx->event_wqh.lock);
1194         }
1195
1196         return ret;
1197 }
1198
1199 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1200                                 size_t count, loff_t *ppos)
1201 {
1202         struct userfaultfd_ctx *ctx = file->private_data;
1203         ssize_t _ret, ret = 0;
1204         struct uffd_msg msg;
1205         int no_wait = file->f_flags & O_NONBLOCK;
1206
1207         if (ctx->state == UFFD_STATE_WAIT_API)
1208                 return -EINVAL;
1209
1210         for (;;) {
1211                 if (count < sizeof(msg))
1212                         return ret ? ret : -EINVAL;
1213                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1214                 if (_ret < 0)
1215                         return ret ? ret : _ret;
1216                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1217                         return ret ? ret : -EFAULT;
1218                 ret += sizeof(msg);
1219                 buf += sizeof(msg);
1220                 count -= sizeof(msg);
1221                 /*
1222                  * Allow to read more than one fault at time but only
1223                  * block if waiting for the very first one.
1224                  */
1225                 no_wait = O_NONBLOCK;
1226         }
1227 }
1228
1229 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1230                              struct userfaultfd_wake_range *range)
1231 {
1232         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1233         /* wake all in the range and autoremove */
1234         if (waitqueue_active(&ctx->fault_pending_wqh))
1235                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1236                                      range);
1237         if (waitqueue_active(&ctx->fault_wqh))
1238                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1239         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1240 }
1241
1242 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1243                                            struct userfaultfd_wake_range *range)
1244 {
1245         unsigned seq;
1246         bool need_wakeup;
1247
1248         /*
1249          * To be sure waitqueue_active() is not reordered by the CPU
1250          * before the pagetable update, use an explicit SMP memory
1251          * barrier here. PT lock release or mmap_read_unlock(mm) still
1252          * have release semantics that can allow the
1253          * waitqueue_active() to be reordered before the pte update.
1254          */
1255         smp_mb();
1256
1257         /*
1258          * Use waitqueue_active because it's very frequent to
1259          * change the address space atomically even if there are no
1260          * userfaults yet. So we take the spinlock only when we're
1261          * sure we've userfaults to wake.
1262          */
1263         do {
1264                 seq = read_seqcount_begin(&ctx->refile_seq);
1265                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1266                         waitqueue_active(&ctx->fault_wqh);
1267                 cond_resched();
1268         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1269         if (need_wakeup)
1270                 __wake_userfault(ctx, range);
1271 }
1272
1273 static __always_inline int validate_range(struct mm_struct *mm,
1274                                           __u64 *start, __u64 len)
1275 {
1276         __u64 task_size = mm->task_size;
1277
1278         *start = untagged_addr(*start);
1279
1280         if (*start & ~PAGE_MASK)
1281                 return -EINVAL;
1282         if (len & ~PAGE_MASK)
1283                 return -EINVAL;
1284         if (!len)
1285                 return -EINVAL;
1286         if (*start < mmap_min_addr)
1287                 return -EINVAL;
1288         if (*start >= task_size)
1289                 return -EINVAL;
1290         if (len > task_size - *start)
1291                 return -EINVAL;
1292         return 0;
1293 }
1294
1295 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1296                                      unsigned long vm_flags)
1297 {
1298         /* FIXME: add WP support to hugetlbfs and shmem */
1299         return vma_is_anonymous(vma) ||
1300                 ((is_vm_hugetlb_page(vma) || vma_is_shmem(vma)) &&
1301                  !(vm_flags & VM_UFFD_WP));
1302 }
1303
1304 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1305                                 unsigned long arg)
1306 {
1307         struct mm_struct *mm = ctx->mm;
1308         struct vm_area_struct *vma, *prev, *cur;
1309         int ret;
1310         struct uffdio_register uffdio_register;
1311         struct uffdio_register __user *user_uffdio_register;
1312         unsigned long vm_flags, new_flags;
1313         bool found;
1314         bool basic_ioctls;
1315         unsigned long start, end, vma_end;
1316
1317         user_uffdio_register = (struct uffdio_register __user *) arg;
1318
1319         ret = -EFAULT;
1320         if (copy_from_user(&uffdio_register, user_uffdio_register,
1321                            sizeof(uffdio_register)-sizeof(__u64)))
1322                 goto out;
1323
1324         ret = -EINVAL;
1325         if (!uffdio_register.mode)
1326                 goto out;
1327         if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1328                                      UFFDIO_REGISTER_MODE_WP))
1329                 goto out;
1330         vm_flags = 0;
1331         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1332                 vm_flags |= VM_UFFD_MISSING;
1333         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)
1334                 vm_flags |= VM_UFFD_WP;
1335
1336         ret = validate_range(mm, &uffdio_register.range.start,
1337                              uffdio_register.range.len);
1338         if (ret)
1339                 goto out;
1340
1341         start = uffdio_register.range.start;
1342         end = start + uffdio_register.range.len;
1343
1344         ret = -ENOMEM;
1345         if (!mmget_not_zero(mm))
1346                 goto out;
1347
1348         mmap_write_lock(mm);
1349         if (!mmget_still_valid(mm))
1350                 goto out_unlock;
1351         vma = find_vma_prev(mm, start, &prev);
1352         if (!vma)
1353                 goto out_unlock;
1354
1355         /* check that there's at least one vma in the range */
1356         ret = -EINVAL;
1357         if (vma->vm_start >= end)
1358                 goto out_unlock;
1359
1360         /*
1361          * If the first vma contains huge pages, make sure start address
1362          * is aligned to huge page size.
1363          */
1364         if (is_vm_hugetlb_page(vma)) {
1365                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1366
1367                 if (start & (vma_hpagesize - 1))
1368                         goto out_unlock;
1369         }
1370
1371         /*
1372          * Search for not compatible vmas.
1373          */
1374         found = false;
1375         basic_ioctls = false;
1376         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1377                 cond_resched();
1378
1379                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1380                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1381
1382                 /* check not compatible vmas */
1383                 ret = -EINVAL;
1384                 if (!vma_can_userfault(cur, vm_flags))
1385                         goto out_unlock;
1386
1387                 /*
1388                  * UFFDIO_COPY will fill file holes even without
1389                  * PROT_WRITE. This check enforces that if this is a
1390                  * MAP_SHARED, the process has write permission to the backing
1391                  * file. If VM_MAYWRITE is set it also enforces that on a
1392                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1393                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1394                  */
1395                 ret = -EPERM;
1396                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1397                         goto out_unlock;
1398
1399                 /*
1400                  * If this vma contains ending address, and huge pages
1401                  * check alignment.
1402                  */
1403                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1404                     end > cur->vm_start) {
1405                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1406
1407                         ret = -EINVAL;
1408
1409                         if (end & (vma_hpagesize - 1))
1410                                 goto out_unlock;
1411                 }
1412                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1413                         goto out_unlock;
1414
1415                 /*
1416                  * Check that this vma isn't already owned by a
1417                  * different userfaultfd. We can't allow more than one
1418                  * userfaultfd to own a single vma simultaneously or we
1419                  * wouldn't know which one to deliver the userfaults to.
1420                  */
1421                 ret = -EBUSY;
1422                 if (cur->vm_userfaultfd_ctx.ctx &&
1423                     cur->vm_userfaultfd_ctx.ctx != ctx)
1424                         goto out_unlock;
1425
1426                 /*
1427                  * Note vmas containing huge pages
1428                  */
1429                 if (is_vm_hugetlb_page(cur))
1430                         basic_ioctls = true;
1431
1432                 found = true;
1433         }
1434         BUG_ON(!found);
1435
1436         if (vma->vm_start < start)
1437                 prev = vma;
1438
1439         ret = 0;
1440         do {
1441                 cond_resched();
1442
1443                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1444                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1445                        vma->vm_userfaultfd_ctx.ctx != ctx);
1446                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1447
1448                 /*
1449                  * Nothing to do: this vma is already registered into this
1450                  * userfaultfd and with the right tracking mode too.
1451                  */
1452                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1453                     (vma->vm_flags & vm_flags) == vm_flags)
1454                         goto skip;
1455
1456                 if (vma->vm_start > start)
1457                         start = vma->vm_start;
1458                 vma_end = min(end, vma->vm_end);
1459
1460                 new_flags = (vma->vm_flags &
1461                              ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags;
1462                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1463                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1464                                  vma_policy(vma),
1465                                  ((struct vm_userfaultfd_ctx){ ctx }));
1466                 if (prev) {
1467                         vma = prev;
1468                         goto next;
1469                 }
1470                 if (vma->vm_start < start) {
1471                         ret = split_vma(mm, vma, start, 1);
1472                         if (ret)
1473                                 break;
1474                 }
1475                 if (vma->vm_end > end) {
1476                         ret = split_vma(mm, vma, end, 0);
1477                         if (ret)
1478                                 break;
1479                 }
1480         next:
1481                 /*
1482                  * In the vma_merge() successful mprotect-like case 8:
1483                  * the next vma was merged into the current one and
1484                  * the current one has not been updated yet.
1485                  */
1486                 vma->vm_flags = new_flags;
1487                 vma->vm_userfaultfd_ctx.ctx = ctx;
1488
1489         skip:
1490                 prev = vma;
1491                 start = vma->vm_end;
1492                 vma = vma->vm_next;
1493         } while (vma && vma->vm_start < end);
1494 out_unlock:
1495         mmap_write_unlock(mm);
1496         mmput(mm);
1497         if (!ret) {
1498                 __u64 ioctls_out;
1499
1500                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1501                     UFFD_API_RANGE_IOCTLS;
1502
1503                 /*
1504                  * Declare the WP ioctl only if the WP mode is
1505                  * specified and all checks passed with the range
1506                  */
1507                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1508                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1509
1510                 /*
1511                  * Now that we scanned all vmas we can already tell
1512                  * userland which ioctls methods are guaranteed to
1513                  * succeed on this range.
1514                  */
1515                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1516                         ret = -EFAULT;
1517         }
1518 out:
1519         return ret;
1520 }
1521
1522 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1523                                   unsigned long arg)
1524 {
1525         struct mm_struct *mm = ctx->mm;
1526         struct vm_area_struct *vma, *prev, *cur;
1527         int ret;
1528         struct uffdio_range uffdio_unregister;
1529         unsigned long new_flags;
1530         bool found;
1531         unsigned long start, end, vma_end;
1532         const void __user *buf = (void __user *)arg;
1533
1534         ret = -EFAULT;
1535         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1536                 goto out;
1537
1538         ret = validate_range(mm, &uffdio_unregister.start,
1539                              uffdio_unregister.len);
1540         if (ret)
1541                 goto out;
1542
1543         start = uffdio_unregister.start;
1544         end = start + uffdio_unregister.len;
1545
1546         ret = -ENOMEM;
1547         if (!mmget_not_zero(mm))
1548                 goto out;
1549
1550         mmap_write_lock(mm);
1551         if (!mmget_still_valid(mm))
1552                 goto out_unlock;
1553         vma = find_vma_prev(mm, start, &prev);
1554         if (!vma)
1555                 goto out_unlock;
1556
1557         /* check that there's at least one vma in the range */
1558         ret = -EINVAL;
1559         if (vma->vm_start >= end)
1560                 goto out_unlock;
1561
1562         /*
1563          * If the first vma contains huge pages, make sure start address
1564          * is aligned to huge page size.
1565          */
1566         if (is_vm_hugetlb_page(vma)) {
1567                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1568
1569                 if (start & (vma_hpagesize - 1))
1570                         goto out_unlock;
1571         }
1572
1573         /*
1574          * Search for not compatible vmas.
1575          */
1576         found = false;
1577         ret = -EINVAL;
1578         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1579                 cond_resched();
1580
1581                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1582                        !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1583
1584                 /*
1585                  * Check not compatible vmas, not strictly required
1586                  * here as not compatible vmas cannot have an
1587                  * userfaultfd_ctx registered on them, but this
1588                  * provides for more strict behavior to notice
1589                  * unregistration errors.
1590                  */
1591                 if (!vma_can_userfault(cur, cur->vm_flags))
1592                         goto out_unlock;
1593
1594                 found = true;
1595         }
1596         BUG_ON(!found);
1597
1598         if (vma->vm_start < start)
1599                 prev = vma;
1600
1601         ret = 0;
1602         do {
1603                 cond_resched();
1604
1605                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1606
1607                 /*
1608                  * Nothing to do: this vma is already registered into this
1609                  * userfaultfd and with the right tracking mode too.
1610                  */
1611                 if (!vma->vm_userfaultfd_ctx.ctx)
1612                         goto skip;
1613
1614                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1615
1616                 if (vma->vm_start > start)
1617                         start = vma->vm_start;
1618                 vma_end = min(end, vma->vm_end);
1619
1620                 if (userfaultfd_missing(vma)) {
1621                         /*
1622                          * Wake any concurrent pending userfault while
1623                          * we unregister, so they will not hang
1624                          * permanently and it avoids userland to call
1625                          * UFFDIO_WAKE explicitly.
1626                          */
1627                         struct userfaultfd_wake_range range;
1628                         range.start = start;
1629                         range.len = vma_end - start;
1630                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1631                 }
1632
1633                 new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1634                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1635                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1636                                  vma_policy(vma),
1637                                  NULL_VM_UFFD_CTX);
1638                 if (prev) {
1639                         vma = prev;
1640                         goto next;
1641                 }
1642                 if (vma->vm_start < start) {
1643                         ret = split_vma(mm, vma, start, 1);
1644                         if (ret)
1645                                 break;
1646                 }
1647                 if (vma->vm_end > end) {
1648                         ret = split_vma(mm, vma, end, 0);
1649                         if (ret)
1650                                 break;
1651                 }
1652         next:
1653                 /*
1654                  * In the vma_merge() successful mprotect-like case 8:
1655                  * the next vma was merged into the current one and
1656                  * the current one has not been updated yet.
1657                  */
1658                 vma->vm_flags = new_flags;
1659                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1660
1661         skip:
1662                 prev = vma;
1663                 start = vma->vm_end;
1664                 vma = vma->vm_next;
1665         } while (vma && vma->vm_start < end);
1666 out_unlock:
1667         mmap_write_unlock(mm);
1668         mmput(mm);
1669 out:
1670         return ret;
1671 }
1672
1673 /*
1674  * userfaultfd_wake may be used in combination with the
1675  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1676  */
1677 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1678                             unsigned long arg)
1679 {
1680         int ret;
1681         struct uffdio_range uffdio_wake;
1682         struct userfaultfd_wake_range range;
1683         const void __user *buf = (void __user *)arg;
1684
1685         ret = -EFAULT;
1686         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1687                 goto out;
1688
1689         ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1690         if (ret)
1691                 goto out;
1692
1693         range.start = uffdio_wake.start;
1694         range.len = uffdio_wake.len;
1695
1696         /*
1697          * len == 0 means wake all and we don't want to wake all here,
1698          * so check it again to be sure.
1699          */
1700         VM_BUG_ON(!range.len);
1701
1702         wake_userfault(ctx, &range);
1703         ret = 0;
1704
1705 out:
1706         return ret;
1707 }
1708
1709 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1710                             unsigned long arg)
1711 {
1712         __s64 ret;
1713         struct uffdio_copy uffdio_copy;
1714         struct uffdio_copy __user *user_uffdio_copy;
1715         struct userfaultfd_wake_range range;
1716
1717         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1718
1719         ret = -EAGAIN;
1720         if (READ_ONCE(ctx->mmap_changing))
1721                 goto out;
1722
1723         ret = -EFAULT;
1724         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1725                            /* don't copy "copy" last field */
1726                            sizeof(uffdio_copy)-sizeof(__s64)))
1727                 goto out;
1728
1729         ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1730         if (ret)
1731                 goto out;
1732         /*
1733          * double check for wraparound just in case. copy_from_user()
1734          * will later check uffdio_copy.src + uffdio_copy.len to fit
1735          * in the userland range.
1736          */
1737         ret = -EINVAL;
1738         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1739                 goto out;
1740         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1741                 goto out;
1742         if (mmget_not_zero(ctx->mm)) {
1743                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1744                                    uffdio_copy.len, &ctx->mmap_changing,
1745                                    uffdio_copy.mode);
1746                 mmput(ctx->mm);
1747         } else {
1748                 return -ESRCH;
1749         }
1750         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1751                 return -EFAULT;
1752         if (ret < 0)
1753                 goto out;
1754         BUG_ON(!ret);
1755         /* len == 0 would wake all */
1756         range.len = ret;
1757         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1758                 range.start = uffdio_copy.dst;
1759                 wake_userfault(ctx, &range);
1760         }
1761         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1762 out:
1763         return ret;
1764 }
1765
1766 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1767                                 unsigned long arg)
1768 {
1769         __s64 ret;
1770         struct uffdio_zeropage uffdio_zeropage;
1771         struct uffdio_zeropage __user *user_uffdio_zeropage;
1772         struct userfaultfd_wake_range range;
1773
1774         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1775
1776         ret = -EAGAIN;
1777         if (READ_ONCE(ctx->mmap_changing))
1778                 goto out;
1779
1780         ret = -EFAULT;
1781         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1782                            /* don't copy "zeropage" last field */
1783                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1784                 goto out;
1785
1786         ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1787                              uffdio_zeropage.range.len);
1788         if (ret)
1789                 goto out;
1790         ret = -EINVAL;
1791         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1792                 goto out;
1793
1794         if (mmget_not_zero(ctx->mm)) {
1795                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1796                                      uffdio_zeropage.range.len,
1797                                      &ctx->mmap_changing);
1798                 mmput(ctx->mm);
1799         } else {
1800                 return -ESRCH;
1801         }
1802         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1803                 return -EFAULT;
1804         if (ret < 0)
1805                 goto out;
1806         /* len == 0 would wake all */
1807         BUG_ON(!ret);
1808         range.len = ret;
1809         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1810                 range.start = uffdio_zeropage.range.start;
1811                 wake_userfault(ctx, &range);
1812         }
1813         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1814 out:
1815         return ret;
1816 }
1817
1818 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1819                                     unsigned long arg)
1820 {
1821         int ret;
1822         struct uffdio_writeprotect uffdio_wp;
1823         struct uffdio_writeprotect __user *user_uffdio_wp;
1824         struct userfaultfd_wake_range range;
1825         bool mode_wp, mode_dontwake;
1826
1827         if (READ_ONCE(ctx->mmap_changing))
1828                 return -EAGAIN;
1829
1830         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1831
1832         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1833                            sizeof(struct uffdio_writeprotect)))
1834                 return -EFAULT;
1835
1836         ret = validate_range(ctx->mm, &uffdio_wp.range.start,
1837                              uffdio_wp.range.len);
1838         if (ret)
1839                 return ret;
1840
1841         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1842                                UFFDIO_WRITEPROTECT_MODE_WP))
1843                 return -EINVAL;
1844
1845         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1846         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1847
1848         if (mode_wp && mode_dontwake)
1849                 return -EINVAL;
1850
1851         ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1852                                   uffdio_wp.range.len, mode_wp,
1853                                   &ctx->mmap_changing);
1854         if (ret)
1855                 return ret;
1856
1857         if (!mode_wp && !mode_dontwake) {
1858                 range.start = uffdio_wp.range.start;
1859                 range.len = uffdio_wp.range.len;
1860                 wake_userfault(ctx, &range);
1861         }
1862         return ret;
1863 }
1864
1865 static inline unsigned int uffd_ctx_features(__u64 user_features)
1866 {
1867         /*
1868          * For the current set of features the bits just coincide
1869          */
1870         return (unsigned int)user_features;
1871 }
1872
1873 /*
1874  * userland asks for a certain API version and we return which bits
1875  * and ioctl commands are implemented in this kernel for such API
1876  * version or -EINVAL if unknown.
1877  */
1878 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1879                            unsigned long arg)
1880 {
1881         struct uffdio_api uffdio_api;
1882         void __user *buf = (void __user *)arg;
1883         int ret;
1884         __u64 features;
1885
1886         ret = -EINVAL;
1887         if (ctx->state != UFFD_STATE_WAIT_API)
1888                 goto out;
1889         ret = -EFAULT;
1890         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1891                 goto out;
1892         features = uffdio_api.features;
1893         ret = -EINVAL;
1894         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1895                 goto err_out;
1896         ret = -EPERM;
1897         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1898                 goto err_out;
1899         /* report all available features and ioctls to userland */
1900         uffdio_api.features = UFFD_API_FEATURES;
1901         uffdio_api.ioctls = UFFD_API_IOCTLS;
1902         ret = -EFAULT;
1903         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1904                 goto out;
1905         ctx->state = UFFD_STATE_RUNNING;
1906         /* only enable the requested features for this uffd context */
1907         ctx->features = uffd_ctx_features(features);
1908         ret = 0;
1909 out:
1910         return ret;
1911 err_out:
1912         memset(&uffdio_api, 0, sizeof(uffdio_api));
1913         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1914                 ret = -EFAULT;
1915         goto out;
1916 }
1917
1918 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1919                               unsigned long arg)
1920 {
1921         int ret = -EINVAL;
1922         struct userfaultfd_ctx *ctx = file->private_data;
1923
1924         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1925                 return -EINVAL;
1926
1927         switch(cmd) {
1928         case UFFDIO_API:
1929                 ret = userfaultfd_api(ctx, arg);
1930                 break;
1931         case UFFDIO_REGISTER:
1932                 ret = userfaultfd_register(ctx, arg);
1933                 break;
1934         case UFFDIO_UNREGISTER:
1935                 ret = userfaultfd_unregister(ctx, arg);
1936                 break;
1937         case UFFDIO_WAKE:
1938                 ret = userfaultfd_wake(ctx, arg);
1939                 break;
1940         case UFFDIO_COPY:
1941                 ret = userfaultfd_copy(ctx, arg);
1942                 break;
1943         case UFFDIO_ZEROPAGE:
1944                 ret = userfaultfd_zeropage(ctx, arg);
1945                 break;
1946         case UFFDIO_WRITEPROTECT:
1947                 ret = userfaultfd_writeprotect(ctx, arg);
1948                 break;
1949         }
1950         return ret;
1951 }
1952
1953 #ifdef CONFIG_PROC_FS
1954 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
1955 {
1956         struct userfaultfd_ctx *ctx = f->private_data;
1957         wait_queue_entry_t *wq;
1958         unsigned long pending = 0, total = 0;
1959
1960         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1961         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
1962                 pending++;
1963                 total++;
1964         }
1965         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
1966                 total++;
1967         }
1968         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1969
1970         /*
1971          * If more protocols will be added, there will be all shown
1972          * separated by a space. Like this:
1973          *      protocols: aa:... bb:...
1974          */
1975         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
1976                    pending, total, UFFD_API, ctx->features,
1977                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
1978 }
1979 #endif
1980
1981 static const struct file_operations userfaultfd_fops = {
1982 #ifdef CONFIG_PROC_FS
1983         .show_fdinfo    = userfaultfd_show_fdinfo,
1984 #endif
1985         .release        = userfaultfd_release,
1986         .poll           = userfaultfd_poll,
1987         .read           = userfaultfd_read,
1988         .unlocked_ioctl = userfaultfd_ioctl,
1989         .compat_ioctl   = compat_ptr_ioctl,
1990         .llseek         = noop_llseek,
1991 };
1992
1993 static void init_once_userfaultfd_ctx(void *mem)
1994 {
1995         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
1996
1997         init_waitqueue_head(&ctx->fault_pending_wqh);
1998         init_waitqueue_head(&ctx->fault_wqh);
1999         init_waitqueue_head(&ctx->event_wqh);
2000         init_waitqueue_head(&ctx->fd_wqh);
2001         seqcount_init(&ctx->refile_seq);
2002 }
2003
2004 SYSCALL_DEFINE1(userfaultfd, int, flags)
2005 {
2006         struct userfaultfd_ctx *ctx;
2007         int fd;
2008
2009         if (!sysctl_unprivileged_userfaultfd && !capable(CAP_SYS_PTRACE))
2010                 return -EPERM;
2011
2012         BUG_ON(!current->mm);
2013
2014         /* Check the UFFD_* constants for consistency.  */
2015         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2016         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2017
2018         if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
2019                 return -EINVAL;
2020
2021         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2022         if (!ctx)
2023                 return -ENOMEM;
2024
2025         refcount_set(&ctx->refcount, 1);
2026         ctx->flags = flags;
2027         ctx->features = 0;
2028         ctx->state = UFFD_STATE_WAIT_API;
2029         ctx->released = false;
2030         ctx->mmap_changing = false;
2031         ctx->mm = current->mm;
2032         /* prevent the mm struct to be freed */
2033         mmgrab(ctx->mm);
2034
2035         fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
2036                               O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
2037         if (fd < 0) {
2038                 mmdrop(ctx->mm);
2039                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2040         }
2041         return fd;
2042 }
2043
2044 static int __init userfaultfd_init(void)
2045 {
2046         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2047                                                 sizeof(struct userfaultfd_ctx),
2048                                                 0,
2049                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2050                                                 init_once_userfaultfd_ctx);
2051         return 0;
2052 }
2053 __initcall(userfaultfd_init);