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/mmu_notifier.h>
  19 #include <linux/poll.h>
  20 #include <linux/slab.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/file.h>
  23 #include <linux/bug.h>
  24 #include <linux/anon_inodes.h>
  25 #include <linux/syscalls.h>
  26 #include <linux/userfaultfd_k.h>
  27 #include <linux/mempolicy.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/security.h>
  30 #include <linux/hugetlb.h>
  31
  32 int sysctl_unprivileged_userfaultfd __read_mostly;
  33
  34 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  35
  36 enum userfaultfd_state {
  37         UFFD_STATE_WAIT_API,
  38         UFFD_STATE_RUNNING,
  39 };
  40
  41 /*
  42  * Start with fault_pending_wqh and fault_wqh so they're more likely
  43  * to be in the same cacheline.
  44  *
  45  * Locking order:
  46  *      fd_wqh.lock
  47  *              fault_pending_wqh.lock
  48  *                      fault_wqh.lock
  49  *              event_wqh.lock
  50  *
  51  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
  52  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
  53  * also taken in IRQ context.
  54  */
  55 struct userfaultfd_ctx {
  56         /* waitqueue head for the pending (i.e. not read) userfaults */
  57         wait_queue_head_t fault_pending_wqh;
  58         /* waitqueue head for the userfaults */
  59         wait_queue_head_t fault_wqh;
  60         /* waitqueue head for the pseudo fd to wakeup poll/read */
  61         wait_queue_head_t fd_wqh;
  62         /* waitqueue head for events */
  63         wait_queue_head_t event_wqh;
  64         /* a refile sequence protected by fault_pending_wqh lock */
  65         seqcount_spinlock_t refile_seq;
  66         /* pseudo fd refcounting */
  67         refcount_t refcount;
  68         /* userfaultfd syscall flags */
  69         unsigned int flags;
  70         /* features requested from the userspace */
  71         unsigned int features;
  72         /* state machine */
  73         enum userfaultfd_state state;
  74         /* released */
  75         bool released;
  76         /* memory mappings are changing because of non-cooperative event */
  77         atomic_t mmap_changing;
  78         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  79         struct mm_struct *mm;
  80 };
  81
  82 struct userfaultfd_fork_ctx {
  83         struct userfaultfd_ctx *orig;
  84         struct userfaultfd_ctx *new;
  85         struct list_head list;
  86 };
  87
  88 struct userfaultfd_unmap_ctx {
  89         struct userfaultfd_ctx *ctx;
  90         unsigned long start;
  91         unsigned long end;
  92         struct list_head list;
  93 };
  94
  95 struct userfaultfd_wait_queue {
  96         struct uffd_msg msg;
  97         wait_queue_entry_t wq;
  98         struct userfaultfd_ctx *ctx;
  99         bool waken;
 100 };
 101
 102 struct userfaultfd_wake_range {
 103         unsigned long start;
 104         unsigned long len;
 105 };
 106
 107 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 108                                      int wake_flags, void *key)
 109 {
 110         struct userfaultfd_wake_range *range = key;
 111         int ret;
 112         struct userfaultfd_wait_queue *uwq;
 113         unsigned long start, len;
 114
 115         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 116         ret = 0;
 117         /* len == 0 means wake all */
 118         start = range->start;
 119         len = range->len;
 120         if (len && (start > uwq->msg.arg.pagefault.address ||
 121                     start + len <= uwq->msg.arg.pagefault.address))
 122                 goto out;
 123         WRITE_ONCE(uwq->waken, true);
 124         /*
 125          * The Program-Order guarantees provided by the scheduler
 126          * ensure uwq->waken is visible before the task is woken.
 127          */
 128         ret = wake_up_state(wq->private, mode);
 129         if (ret) {
 130                 /*
 131                  * Wake only once, autoremove behavior.
 132                  *
 133                  * After the effect of list_del_init is visible to the other
 134                  * CPUs, the waitqueue may disappear from under us, see the
 135                  * !list_empty_careful() in handle_userfault().
 136                  *
 137                  * try_to_wake_up() has an implicit smp_mb(), and the
 138                  * wq->private is read before calling the extern function
 139                  * "wake_up_state" (which in turns calls try_to_wake_up).
 140                  */
 141                 list_del_init(&wq->entry);
 142         }
 143 out:
 144         return ret;
 145 }
 146
 147 /**
 148  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 149  * context.
 150  * @ctx: [in] Pointer to the userfaultfd context.
 151  */
 152 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 153 {
 154         refcount_inc(&ctx->refcount);
 155 }
 156
 157 /**
 158  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 159  * context.
 160  * @ctx: [in] Pointer to userfaultfd context.
 161  *
 162  * The userfaultfd context reference must have been previously acquired either
 163  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 164  */
 165 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 166 {
 167         if (refcount_dec_and_test(&ctx->refcount)) {
 168                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 169                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 170                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 171                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 172                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 173                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 174                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 175                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 176                 mmdrop(ctx->mm);
 177                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 178         }
 179 }
 180
 181 static inline void msg_init(struct uffd_msg *msg)
 182 {
 183         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 184         /*
 185          * Must use memset to zero out the paddings or kernel data is
 186          * leaked to userland.
 187          */
 188         memset(msg, 0, sizeof(struct uffd_msg));
 189 }
 190
 191 static inline struct uffd_msg userfault_msg(unsigned long address,
 192                                             unsigned int flags,
 193                                             unsigned long reason,
 194                                             unsigned int features)
 195 {
 196         struct uffd_msg msg;
 197         msg_init(&msg);
 198         msg.event = UFFD_EVENT_PAGEFAULT;
 199         msg.arg.pagefault.address = address;
 200         /*
 201          * These flags indicate why the userfault occurred:
 202          * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
 203          * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
 204          * - Neither of these flags being set indicates a MISSING fault.
 205          *
 206          * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
 207          * fault. Otherwise, it was a read fault.
 208          */
 209         if (flags & FAULT_FLAG_WRITE)
 210                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 211         if (reason & VM_UFFD_WP)
 212                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 213         if (reason & VM_UFFD_MINOR)
 214                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
 215         if (features & UFFD_FEATURE_THREAD_ID)
 216                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 217         return msg;
 218 }
 219
 220 #ifdef CONFIG_HUGETLB_PAGE
 221 /*
 222  * Same functionality as userfaultfd_must_wait below with modifications for
 223  * hugepmd ranges.
 224  */
 225 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 226                                          struct vm_area_struct *vma,
 227                                          unsigned long address,
 228                                          unsigned long flags,
 229                                          unsigned long reason)
 230 {
 231         struct mm_struct *mm = ctx->mm;
 232         pte_t *ptep, pte;
 233         bool ret = true;
 234
 235         mmap_assert_locked(mm);
 236
 237         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 238
 239         if (!ptep)
 240                 goto out;
 241
 242         ret = false;
 243         pte = huge_ptep_get(ptep);
 244
 245         /*
 246          * Lockless access: we're in a wait_event so it's ok if it
 247          * changes under us.
 248          */
 249         if (huge_pte_none(pte))
 250                 ret = true;
 251         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 252                 ret = true;
 253 out:
 254         return ret;
 255 }
 256 #else
 257 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 258                                          struct vm_area_struct *vma,
 259                                          unsigned long address,
 260                                          unsigned long flags,
 261                                          unsigned long reason)
 262 {
 263         return false;   /* should never get here */
 264 }
 265 #endif /* CONFIG_HUGETLB_PAGE */
 266
 267 /*
 268  * Verify the pagetables are still not ok after having reigstered into
 269  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 270  * userfault that has already been resolved, if userfaultfd_read and
 271  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 272  * threads.
 273  */
 274 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 275                                          unsigned long address,
 276                                          unsigned long flags,
 277                                          unsigned long reason)
 278 {
 279         struct mm_struct *mm = ctx->mm;
 280         pgd_t *pgd;
 281         p4d_t *p4d;
 282         pud_t *pud;
 283         pmd_t *pmd, _pmd;
 284         pte_t *pte;
 285         bool ret = true;
 286
 287         mmap_assert_locked(mm);
 288
 289         pgd = pgd_offset(mm, address);
 290         if (!pgd_present(*pgd))
 291                 goto out;
 292         p4d = p4d_offset(pgd, address);
 293         if (!p4d_present(*p4d))
 294                 goto out;
 295         pud = pud_offset(p4d, address);
 296         if (!pud_present(*pud))
 297                 goto out;
 298         pmd = pmd_offset(pud, address);
 299         /*
 300          * READ_ONCE must function as a barrier with narrower scope
 301          * and it must be equivalent to:
 302          *      _pmd = *pmd; barrier();
 303          *
 304          * This is to deal with the instability (as in
 305          * pmd_trans_unstable) of the pmd.
 306          */
 307         _pmd = READ_ONCE(*pmd);
 308         if (pmd_none(_pmd))
 309                 goto out;
 310
 311         ret = false;
 312         if (!pmd_present(_pmd))
 313                 goto out;
 314
 315         if (pmd_trans_huge(_pmd)) {
 316                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 317                         ret = true;
 318                 goto out;
 319         }
 320
 321         /*
 322          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 323          * and use the standard pte_offset_map() instead of parsing _pmd.
 324          */
 325         pte = pte_offset_map(pmd, address);
 326         /*
 327          * Lockless access: we're in a wait_event so it's ok if it
 328          * changes under us.
 329          */
 330         if (pte_none(*pte))
 331                 ret = true;
 332         if (!pte_write(*pte) && (reason & VM_UFFD_WP))
 333                 ret = true;
 334         pte_unmap(pte);
 335
 336 out:
 337         return ret;
 338 }
 339
 340 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
 341 {
 342         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 343                 return TASK_INTERRUPTIBLE;
 344
 345         if (flags & FAULT_FLAG_KILLABLE)
 346                 return TASK_KILLABLE;
 347
 348         return TASK_UNINTERRUPTIBLE;
 349 }
 350
 351 /*
 352  * The locking rules involved in returning VM_FAULT_RETRY depending on
 353  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 354  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 355  * recommendation in __lock_page_or_retry is not an understatement.
 356  *
 357  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
 358  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 359  * not set.
 360  *
 361  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 362  * set, VM_FAULT_RETRY can still be returned if and only if there are
 363  * fatal_signal_pending()s, and the mmap_lock must be released before
 364  * returning it.
 365  */
 366 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 367 {
 368         struct mm_struct *mm = vmf->vma->vm_mm;
 369         struct userfaultfd_ctx *ctx;
 370         struct userfaultfd_wait_queue uwq;
 371         vm_fault_t ret = VM_FAULT_SIGBUS;
 372         bool must_wait;
 373         unsigned int blocking_state;
 374
 375         /*
 376          * We don't do userfault handling for the final child pid update.
 377          *
 378          * We also don't do userfault handling during
 379          * coredumping. hugetlbfs has the special
 380          * follow_hugetlb_page() to skip missing pages in the
 381          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 382          * the no_page_table() helper in follow_page_mask(), but the
 383          * shmem_vm_ops->fault method is invoked even during
 384          * coredumping without mmap_lock and it ends up here.
 385          */
 386         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 387                 goto out;
 388
 389         /*
 390          * Coredumping runs without mmap_lock so we can only check that
 391          * the mmap_lock is held, if PF_DUMPCORE was not set.
 392          */
 393         mmap_assert_locked(mm);
 394
 395         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 396         if (!ctx)
 397                 goto out;
 398
 399         BUG_ON(ctx->mm != mm);
 400
 401         /* Any unrecognized flag is a bug. */
 402         VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
 403         /* 0 or > 1 flags set is a bug; we expect exactly 1. */
 404         VM_BUG_ON(!reason || (reason & (reason - 1)));
 405
 406         if (ctx->features & UFFD_FEATURE_SIGBUS)
 407                 goto out;
 408         if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
 409             ctx->flags & UFFD_USER_MODE_ONLY) {
 410                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
 411                         "sysctl knob to 1 if kernel faults must be handled "
 412                         "without obtaining CAP_SYS_PTRACE capability\n");
 413                 goto out;
 414         }
 415
 416         /*
 417          * If it's already released don't get it. This avoids to loop
 418          * in __get_user_pages if userfaultfd_release waits on the
 419          * caller of handle_userfault to release the mmap_lock.
 420          */
 421         if (unlikely(READ_ONCE(ctx->released))) {
 422                 /*
 423                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 424                  * cooperative manager can close the uffd after the
 425                  * last UFFDIO_COPY, without risking to trigger an
 426                  * involuntary SIGBUS if the process was starting the
 427                  * userfaultfd while the userfaultfd was still armed
 428                  * (but after the last UFFDIO_COPY). If the uffd
 429                  * wasn't already closed when the userfault reached
 430                  * this point, that would normally be solved by
 431                  * userfaultfd_must_wait returning 'false'.
 432                  *
 433                  * If we were to return VM_FAULT_SIGBUS here, the non
 434                  * cooperative manager would be instead forced to
 435                  * always call UFFDIO_UNREGISTER before it can safely
 436                  * close the uffd.
 437                  */
 438                 ret = VM_FAULT_NOPAGE;
 439                 goto out;
 440         }
 441
 442         /*
 443          * Check that we can return VM_FAULT_RETRY.
 444          *
 445          * NOTE: it should become possible to return VM_FAULT_RETRY
 446          * even if FAULT_FLAG_TRIED is set without leading to gup()
 447          * -EBUSY failures, if the userfaultfd is to be extended for
 448          * VM_UFFD_WP tracking and we intend to arm the userfault
 449          * without first stopping userland access to the memory. For
 450          * VM_UFFD_MISSING userfaults this is enough for now.
 451          */
 452         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 453                 /*
 454                  * Validate the invariant that nowait must allow retry
 455                  * to be sure not to return SIGBUS erroneously on
 456                  * nowait invocations.
 457                  */
 458                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 459 #ifdef CONFIG_DEBUG_VM
 460                 if (printk_ratelimit()) {
 461                         printk(KERN_WARNING
 462                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 463                                vmf->flags);
 464                         dump_stack();
 465                 }
 466 #endif
 467                 goto out;
 468         }
 469
 470         /*
 471          * Handle nowait, not much to do other than tell it to retry
 472          * and wait.
 473          */
 474         ret = VM_FAULT_RETRY;
 475         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 476                 goto out;
 477
 478         /* take the reference before dropping the mmap_lock */
 479         userfaultfd_ctx_get(ctx);
 480
 481         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 482         uwq.wq.private = current;
 483         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 484                         ctx->features);
 485         uwq.ctx = ctx;
 486         uwq.waken = false;
 487
 488         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 489
 490         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 491         /*
 492          * After the __add_wait_queue the uwq is visible to userland
 493          * through poll/read().
 494          */
 495         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 496         /*
 497          * The smp_mb() after __set_current_state prevents the reads
 498          * following the spin_unlock to happen before the list_add in
 499          * __add_wait_queue.
 500          */
 501         set_current_state(blocking_state);
 502         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 503
 504         if (!is_vm_hugetlb_page(vmf->vma))
 505                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 506                                                   reason);
 507         else
 508                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 509                                                        vmf->address,
 510                                                        vmf->flags, reason);
 511         mmap_read_unlock(mm);
 512
 513         if (likely(must_wait && !READ_ONCE(ctx->released))) {
 514                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 515                 schedule();
 516         }
 517
 518         __set_current_state(TASK_RUNNING);
 519
 520         /*
 521          * Here we race with the list_del; list_add in
 522          * userfaultfd_ctx_read(), however because we don't ever run
 523          * list_del_init() to refile across the two lists, the prev
 524          * and next pointers will never point to self. list_add also
 525          * would never let any of the two pointers to point to
 526          * self. So list_empty_careful won't risk to see both pointers
 527          * pointing to self at any time during the list refile. The
 528          * only case where list_del_init() is called is the full
 529          * removal in the wake function and there we don't re-list_add
 530          * and it's fine not to block on the spinlock. The uwq on this
 531          * kernel stack can be released after the list_del_init.
 532          */
 533         if (!list_empty_careful(&uwq.wq.entry)) {
 534                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 535                 /*
 536                  * No need of list_del_init(), the uwq on the stack
 537                  * will be freed shortly anyway.
 538                  */
 539                 list_del(&uwq.wq.entry);
 540                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 541         }
 542
 543         /*
 544          * ctx may go away after this if the userfault pseudo fd is
 545          * already released.
 546          */
 547         userfaultfd_ctx_put(ctx);
 548
 549 out:
 550         return ret;
 551 }
 552
 553 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 554                                               struct userfaultfd_wait_queue *ewq)
 555 {
 556         struct userfaultfd_ctx *release_new_ctx;
 557
 558         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 559                 goto out;
 560
 561         ewq->ctx = ctx;
 562         init_waitqueue_entry(&ewq->wq, current);
 563         release_new_ctx = NULL;
 564
 565         spin_lock_irq(&ctx->event_wqh.lock);
 566         /*
 567          * After the __add_wait_queue the uwq is visible to userland
 568          * through poll/read().
 569          */
 570         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 571         for (;;) {
 572                 set_current_state(TASK_KILLABLE);
 573                 if (ewq->msg.event == 0)
 574                         break;
 575                 if (READ_ONCE(ctx->released) ||
 576                     fatal_signal_pending(current)) {
 577                         /*
 578                          * &ewq->wq may be queued in fork_event, but
 579                          * __remove_wait_queue ignores the head
 580                          * parameter. It would be a problem if it
 581                          * didn't.
 582                          */
 583                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 584                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 585                                 struct userfaultfd_ctx *new;
 586
 587                                 new = (struct userfaultfd_ctx *)
 588                                         (unsigned long)
 589                                         ewq->msg.arg.reserved.reserved1;
 590                                 release_new_ctx = new;
 591                         }
 592                         break;
 593                 }
 594
 595                 spin_unlock_irq(&ctx->event_wqh.lock);
 596
 597                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 598                 schedule();
 599
 600                 spin_lock_irq(&ctx->event_wqh.lock);
 601         }
 602         __set_current_state(TASK_RUNNING);
 603         spin_unlock_irq(&ctx->event_wqh.lock);
 604
 605         if (release_new_ctx) {
 606                 struct vm_area_struct *vma;
 607                 struct mm_struct *mm = release_new_ctx->mm;
 608
 609                 /* the various vma->vm_userfaultfd_ctx still points to it */
 610                 mmap_write_lock(mm);
 611                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 612                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 613                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 614                                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 615                         }
 616                 mmap_write_unlock(mm);
 617
 618                 userfaultfd_ctx_put(release_new_ctx);
 619         }
 620
 621         /*
 622          * ctx may go away after this if the userfault pseudo fd is
 623          * already released.
 624          */
 625 out:
 626         atomic_dec(&ctx->mmap_changing);
 627         VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
 628         userfaultfd_ctx_put(ctx);
 629 }
 630
 631 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 632                                        struct userfaultfd_wait_queue *ewq)
 633 {
 634         ewq->msg.event = 0;
 635         wake_up_locked(&ctx->event_wqh);
 636         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 637 }
 638
 639 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 640 {
 641         struct userfaultfd_ctx *ctx = NULL, *octx;
 642         struct userfaultfd_fork_ctx *fctx;
 643
 644         octx = vma->vm_userfaultfd_ctx.ctx;
 645         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 646                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 647                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 648                 return 0;
 649         }
 650
 651         list_for_each_entry(fctx, fcs, list)
 652                 if (fctx->orig == octx) {
 653                         ctx = fctx->new;
 654                         break;
 655                 }
 656
 657         if (!ctx) {
 658                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 659                 if (!fctx)
 660                         return -ENOMEM;
 661
 662                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 663                 if (!ctx) {
 664                         kfree(fctx);
 665                         return -ENOMEM;
 666                 }
 667
 668                 refcount_set(&ctx->refcount, 1);
 669                 ctx->flags = octx->flags;
 670                 ctx->state = UFFD_STATE_RUNNING;
 671                 ctx->features = octx->features;
 672                 ctx->released = false;
 673                 atomic_set(&ctx->mmap_changing, 0);
 674                 ctx->mm = vma->vm_mm;
 675                 mmgrab(ctx->mm);
 676
 677                 userfaultfd_ctx_get(octx);
 678                 atomic_inc(&octx->mmap_changing);
 679                 fctx->orig = octx;
 680                 fctx->new = ctx;
 681                 list_add_tail(&fctx->list, fcs);
 682         }
 683
 684         vma->vm_userfaultfd_ctx.ctx = ctx;
 685         return 0;
 686 }
 687
 688 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 689 {
 690         struct userfaultfd_ctx *ctx = fctx->orig;
 691         struct userfaultfd_wait_queue ewq;
 692
 693         msg_init(&ewq.msg);
 694
 695         ewq.msg.event = UFFD_EVENT_FORK;
 696         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 697
 698         userfaultfd_event_wait_completion(ctx, &ewq);
 699 }
 700
 701 void dup_userfaultfd_complete(struct list_head *fcs)
 702 {
 703         struct userfaultfd_fork_ctx *fctx, *n;
 704
 705         list_for_each_entry_safe(fctx, n, fcs, list) {
 706                 dup_fctx(fctx);
 707                 list_del(&fctx->list);
 708                 kfree(fctx);
 709         }
 710 }
 711
 712 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 713                              struct vm_userfaultfd_ctx *vm_ctx)
 714 {
 715         struct userfaultfd_ctx *ctx;
 716
 717         ctx = vma->vm_userfaultfd_ctx.ctx;
 718
 719         if (!ctx)
 720                 return;
 721
 722         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 723                 vm_ctx->ctx = ctx;
 724                 userfaultfd_ctx_get(ctx);
 725                 atomic_inc(&ctx->mmap_changing);
 726         } else {
 727                 /* Drop uffd context if remap feature not enabled */
 728                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 729                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 730         }
 731 }
 732
 733 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 734                                  unsigned long from, unsigned long to,
 735                                  unsigned long len)
 736 {
 737         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 738         struct userfaultfd_wait_queue ewq;
 739
 740         if (!ctx)
 741                 return;
 742
 743         if (to & ~PAGE_MASK) {
 744                 userfaultfd_ctx_put(ctx);
 745                 return;
 746         }
 747
 748         msg_init(&ewq.msg);
 749
 750         ewq.msg.event = UFFD_EVENT_REMAP;
 751         ewq.msg.arg.remap.from = from;
 752         ewq.msg.arg.remap.to = to;
 753         ewq.msg.arg.remap.len = len;
 754
 755         userfaultfd_event_wait_completion(ctx, &ewq);
 756 }
 757
 758 bool userfaultfd_remove(struct vm_area_struct *vma,
 759                         unsigned long start, unsigned long end)
 760 {
 761         struct mm_struct *mm = vma->vm_mm;
 762         struct userfaultfd_ctx *ctx;
 763         struct userfaultfd_wait_queue ewq;
 764
 765         ctx = vma->vm_userfaultfd_ctx.ctx;
 766         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 767                 return true;
 768
 769         userfaultfd_ctx_get(ctx);
 770         atomic_inc(&ctx->mmap_changing);
 771         mmap_read_unlock(mm);
 772
 773         msg_init(&ewq.msg);
 774
 775         ewq.msg.event = UFFD_EVENT_REMOVE;
 776         ewq.msg.arg.remove.start = start;
 777         ewq.msg.arg.remove.end = end;
 778
 779         userfaultfd_event_wait_completion(ctx, &ewq);
 780
 781         return false;
 782 }
 783
 784 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 785                           unsigned long start, unsigned long end)
 786 {
 787         struct userfaultfd_unmap_ctx *unmap_ctx;
 788
 789         list_for_each_entry(unmap_ctx, unmaps, list)
 790                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 791                     unmap_ctx->end == end)
 792                         return true;
 793
 794         return false;
 795 }
 796
 797 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 798                            unsigned long start, unsigned long end,
 799                            struct list_head *unmaps)
 800 {
 801         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 802                 struct userfaultfd_unmap_ctx *unmap_ctx;
 803                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 804
 805                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 806                     has_unmap_ctx(ctx, unmaps, start, end))
 807                         continue;
 808
 809                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 810                 if (!unmap_ctx)
 811                         return -ENOMEM;
 812
 813                 userfaultfd_ctx_get(ctx);
 814                 atomic_inc(&ctx->mmap_changing);
 815                 unmap_ctx->ctx = ctx;
 816                 unmap_ctx->start = start;
 817                 unmap_ctx->end = end;
 818                 list_add_tail(&unmap_ctx->list, unmaps);
 819         }
 820
 821         return 0;
 822 }
 823
 824 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 825 {
 826         struct userfaultfd_unmap_ctx *ctx, *n;
 827         struct userfaultfd_wait_queue ewq;
 828
 829         list_for_each_entry_safe(ctx, n, uf, list) {
 830                 msg_init(&ewq.msg);
 831
 832                 ewq.msg.event = UFFD_EVENT_UNMAP;
 833                 ewq.msg.arg.remove.start = ctx->start;
 834                 ewq.msg.arg.remove.end = ctx->end;
 835
 836                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 837
 838                 list_del(&ctx->list);
 839                 kfree(ctx);
 840         }
 841 }
 842
 843 static int userfaultfd_release(struct inode *inode, struct file *file)
 844 {
 845         struct userfaultfd_ctx *ctx = file->private_data;
 846         struct mm_struct *mm = ctx->mm;
 847         struct vm_area_struct *vma, *prev;
 848         /* len == 0 means wake all */
 849         struct userfaultfd_wake_range range = { .len = 0, };
 850         unsigned long new_flags;
 851
 852         WRITE_ONCE(ctx->released, true);
 853
 854         if (!mmget_not_zero(mm))
 855                 goto wakeup;
 856
 857         /*
 858          * Flush page faults out of all CPUs. NOTE: all page faults
 859          * must be retried without returning VM_FAULT_SIGBUS if
 860          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 861          * changes while handle_userfault released the mmap_lock. So
 862          * it's critical that released is set to true (above), before
 863          * taking the mmap_lock for writing.
 864          */
 865         mmap_write_lock(mm);
 866         prev = NULL;
 867         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 868                 cond_resched();
 869                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 870                        !!(vma->vm_flags & __VM_UFFD_FLAGS));
 871                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 872                         prev = vma;
 873                         continue;
 874                 }
 875                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
 876                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 877                                  new_flags, vma->anon_vma,
 878                                  vma->vm_file, vma->vm_pgoff,
 879                                  vma_policy(vma),
 880                                  NULL_VM_UFFD_CTX);
 881                 if (prev)
 882                         vma = prev;
 883                 else
 884                         prev = vma;
 885                 vma->vm_flags = new_flags;
 886                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 887         }
 888         mmap_write_unlock(mm);
 889         mmput(mm);
 890 wakeup:
 891         /*
 892          * After no new page faults can wait on this fault_*wqh, flush
 893          * the last page faults that may have been already waiting on
 894          * the fault_*wqh.
 895          */
 896         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 897         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 898         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 899         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 900
 901         /* Flush pending events that may still wait on event_wqh */
 902         wake_up_all(&ctx->event_wqh);
 903
 904         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 905         userfaultfd_ctx_put(ctx);
 906         return 0;
 907 }
 908
 909 /* fault_pending_wqh.lock must be hold by the caller */
 910 static inline struct userfaultfd_wait_queue *find_userfault_in(
 911                 wait_queue_head_t *wqh)
 912 {
 913         wait_queue_entry_t *wq;
 914         struct userfaultfd_wait_queue *uwq;
 915
 916         lockdep_assert_held(&wqh->lock);
 917
 918         uwq = NULL;
 919         if (!waitqueue_active(wqh))
 920                 goto out;
 921         /* walk in reverse to provide FIFO behavior to read userfaults */
 922         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 923         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 924 out:
 925         return uwq;
 926 }
 927
 928 static inline struct userfaultfd_wait_queue *find_userfault(
 929                 struct userfaultfd_ctx *ctx)
 930 {
 931         return find_userfault_in(&ctx->fault_pending_wqh);
 932 }
 933
 934 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 935                 struct userfaultfd_ctx *ctx)
 936 {
 937         return find_userfault_in(&ctx->event_wqh);
 938 }
 939
 940 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 941 {
 942         struct userfaultfd_ctx *ctx = file->private_data;
 943         __poll_t ret;
 944
 945         poll_wait(file, &ctx->fd_wqh, wait);
 946
 947         switch (ctx->state) {
 948         case UFFD_STATE_WAIT_API:
 949                 return EPOLLERR;
 950         case UFFD_STATE_RUNNING:
 951                 /*
 952                  * poll() never guarantees that read won't block.
 953                  * userfaults can be waken before they're read().
 954                  */
 955                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 956                         return EPOLLERR;
 957                 /*
 958                  * lockless access to see if there are pending faults
 959                  * __pollwait last action is the add_wait_queue but
 960                  * the spin_unlock would allow the waitqueue_active to
 961                  * pass above the actual list_add inside
 962                  * add_wait_queue critical section. So use a full
 963                  * memory barrier to serialize the list_add write of
 964                  * add_wait_queue() with the waitqueue_active read
 965                  * below.
 966                  */
 967                 ret = 0;
 968                 smp_mb();
 969                 if (waitqueue_active(&ctx->fault_pending_wqh))
 970                         ret = EPOLLIN;
 971                 else if (waitqueue_active(&ctx->event_wqh))
 972                         ret = EPOLLIN;
 973
 974                 return ret;
 975         default:
 976                 WARN_ON_ONCE(1);
 977                 return EPOLLERR;
 978         }
 979 }
 980
 981 static const struct file_operations userfaultfd_fops;
 982
 983 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
 984                                   struct inode *inode,
 985                                   struct uffd_msg *msg)
 986 {
 987         int fd;
 988
 989         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
 990                         O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
 991         if (fd < 0)
 992                 return fd;
 993
 994         msg->arg.reserved.reserved1 = 0;
 995         msg->arg.fork.ufd = fd;
 996         return 0;
 997 }
 998
 999 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1000                                     struct uffd_msg *msg, struct inode *inode)
1001 {
1002         ssize_t ret;
1003         DECLARE_WAITQUEUE(wait, current);
1004         struct userfaultfd_wait_queue *uwq;
1005         /*
1006          * Handling fork event requires sleeping operations, so
1007          * we drop the event_wqh lock, then do these ops, then
1008          * lock it back and wake up the waiter. While the lock is
1009          * dropped the ewq may go away so we keep track of it
1010          * carefully.
1011          */
1012         LIST_HEAD(fork_event);
1013         struct userfaultfd_ctx *fork_nctx = NULL;
1014
1015         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1016         spin_lock_irq(&ctx->fd_wqh.lock);
1017         __add_wait_queue(&ctx->fd_wqh, &wait);
1018         for (;;) {
1019                 set_current_state(TASK_INTERRUPTIBLE);
1020                 spin_lock(&ctx->fault_pending_wqh.lock);
1021                 uwq = find_userfault(ctx);
1022                 if (uwq) {
1023                         /*
1024                          * Use a seqcount to repeat the lockless check
1025                          * in wake_userfault() to avoid missing
1026                          * wakeups because during the refile both
1027                          * waitqueue could become empty if this is the
1028                          * only userfault.
1029                          */
1030                         write_seqcount_begin(&ctx->refile_seq);
1031
1032                         /*
1033                          * The fault_pending_wqh.lock prevents the uwq
1034                          * to disappear from under us.
1035                          *
1036                          * Refile this userfault from
1037                          * fault_pending_wqh to fault_wqh, it's not
1038                          * pending anymore after we read it.
1039                          *
1040                          * Use list_del() by hand (as
1041                          * userfaultfd_wake_function also uses
1042                          * list_del_init() by hand) to be sure nobody
1043                          * changes __remove_wait_queue() to use
1044                          * list_del_init() in turn breaking the
1045                          * !list_empty_careful() check in
1046                          * handle_userfault(). The uwq->wq.head list
1047                          * must never be empty at any time during the
1048                          * refile, or the waitqueue could disappear
1049                          * from under us. The "wait_queue_head_t"
1050                          * parameter of __remove_wait_queue() is unused
1051                          * anyway.
1052                          */
1053                         list_del(&uwq->wq.entry);
1054                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1055
1056                         write_seqcount_end(&ctx->refile_seq);
1057
1058                         /* careful to always initialize msg if ret == 0 */
1059                         *msg = uwq->msg;
1060                         spin_unlock(&ctx->fault_pending_wqh.lock);
1061                         ret = 0;
1062                         break;
1063                 }
1064                 spin_unlock(&ctx->fault_pending_wqh.lock);
1065
1066                 spin_lock(&ctx->event_wqh.lock);
1067                 uwq = find_userfault_evt(ctx);
1068                 if (uwq) {
1069                         *msg = uwq->msg;
1070
1071                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1072                                 fork_nctx = (struct userfaultfd_ctx *)
1073                                         (unsigned long)
1074                                         uwq->msg.arg.reserved.reserved1;
1075                                 list_move(&uwq->wq.entry, &fork_event);
1076                                 /*
1077                                  * fork_nctx can be freed as soon as
1078                                  * we drop the lock, unless we take a
1079                                  * reference on it.
1080                                  */
1081                                 userfaultfd_ctx_get(fork_nctx);
1082                                 spin_unlock(&ctx->event_wqh.lock);
1083                                 ret = 0;
1084                                 break;
1085                         }
1086
1087                         userfaultfd_event_complete(ctx, uwq);
1088                         spin_unlock(&ctx->event_wqh.lock);
1089                         ret = 0;
1090                         break;
1091                 }
1092                 spin_unlock(&ctx->event_wqh.lock);
1093
1094                 if (signal_pending(current)) {
1095                         ret = -ERESTARTSYS;
1096                         break;
1097                 }
1098                 if (no_wait) {
1099                         ret = -EAGAIN;
1100                         break;
1101                 }
1102                 spin_unlock_irq(&ctx->fd_wqh.lock);
1103                 schedule();
1104                 spin_lock_irq(&ctx->fd_wqh.lock);
1105         }
1106         __remove_wait_queue(&ctx->fd_wqh, &wait);
1107         __set_current_state(TASK_RUNNING);
1108         spin_unlock_irq(&ctx->fd_wqh.lock);
1109
1110         if (!ret && msg->event == UFFD_EVENT_FORK) {
1111                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1112                 spin_lock_irq(&ctx->event_wqh.lock);
1113                 if (!list_empty(&fork_event)) {
1114                         /*
1115                          * The fork thread didn't abort, so we can
1116                          * drop the temporary refcount.
1117                          */
1118                         userfaultfd_ctx_put(fork_nctx);
1119
1120                         uwq = list_first_entry(&fork_event,
1121                                                typeof(*uwq),
1122                                                wq.entry);
1123                         /*
1124                          * If fork_event list wasn't empty and in turn
1125                          * the event wasn't already released by fork
1126                          * (the event is allocated on fork kernel
1127                          * stack), put the event back to its place in
1128                          * the event_wq. fork_event head will be freed
1129                          * as soon as we return so the event cannot
1130                          * stay queued there no matter the current
1131                          * "ret" value.
1132                          */
1133                         list_del(&uwq->wq.entry);
1134                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1135
1136                         /*
1137                          * Leave the event in the waitqueue and report
1138                          * error to userland if we failed to resolve
1139                          * the userfault fork.
1140                          */
1141                         if (likely(!ret))
1142                                 userfaultfd_event_complete(ctx, uwq);
1143                 } else {
1144                         /*
1145                          * Here the fork thread aborted and the
1146                          * refcount from the fork thread on fork_nctx
1147                          * has already been released. We still hold
1148                          * the reference we took before releasing the
1149                          * lock above. If resolve_userfault_fork
1150                          * failed we've to drop it because the
1151                          * fork_nctx has to be freed in such case. If
1152                          * it succeeded we'll hold it because the new
1153                          * uffd references it.
1154                          */
1155                         if (ret)
1156                                 userfaultfd_ctx_put(fork_nctx);
1157                 }
1158                 spin_unlock_irq(&ctx->event_wqh.lock);
1159         }
1160
1161         return ret;
1162 }
1163
1164 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1165                                 size_t count, loff_t *ppos)
1166 {
1167         struct userfaultfd_ctx *ctx = file->private_data;
1168         ssize_t _ret, ret = 0;
1169         struct uffd_msg msg;
1170         int no_wait = file->f_flags & O_NONBLOCK;
1171         struct inode *inode = file_inode(file);
1172
1173         if (ctx->state == UFFD_STATE_WAIT_API)
1174                 return -EINVAL;
1175
1176         for (;;) {
1177                 if (count < sizeof(msg))
1178                         return ret ? ret : -EINVAL;
1179                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1180                 if (_ret < 0)
1181                         return ret ? ret : _ret;
1182                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1183                         return ret ? ret : -EFAULT;
1184                 ret += sizeof(msg);
1185                 buf += sizeof(msg);
1186                 count -= sizeof(msg);
1187                 /*
1188                  * Allow to read more than one fault at time but only
1189                  * block if waiting for the very first one.
1190                  */
1191                 no_wait = O_NONBLOCK;
1192         }
1193 }
1194
1195 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1196                              struct userfaultfd_wake_range *range)
1197 {
1198         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1199         /* wake all in the range and autoremove */
1200         if (waitqueue_active(&ctx->fault_pending_wqh))
1201                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1202                                      range);
1203         if (waitqueue_active(&ctx->fault_wqh))
1204                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1205         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1206 }
1207
1208 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1209                                            struct userfaultfd_wake_range *range)
1210 {
1211         unsigned seq;
1212         bool need_wakeup;
1213
1214         /*
1215          * To be sure waitqueue_active() is not reordered by the CPU
1216          * before the pagetable update, use an explicit SMP memory
1217          * barrier here. PT lock release or mmap_read_unlock(mm) still
1218          * have release semantics that can allow the
1219          * waitqueue_active() to be reordered before the pte update.
1220          */
1221         smp_mb();
1222
1223         /*
1224          * Use waitqueue_active because it's very frequent to
1225          * change the address space atomically even if there are no
1226          * userfaults yet. So we take the spinlock only when we're
1227          * sure we've userfaults to wake.
1228          */
1229         do {
1230                 seq = read_seqcount_begin(&ctx->refile_seq);
1231                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1232                         waitqueue_active(&ctx->fault_wqh);
1233                 cond_resched();
1234         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1235         if (need_wakeup)
1236                 __wake_userfault(ctx, range);
1237 }
1238
1239 static __always_inline int validate_range(struct mm_struct *mm,
1240                                           __u64 start, __u64 len)
1241 {
1242         __u64 task_size = mm->task_size;
1243
1244         if (start & ~PAGE_MASK)
1245                 return -EINVAL;
1246         if (len & ~PAGE_MASK)
1247                 return -EINVAL;
1248         if (!len)
1249                 return -EINVAL;
1250         if (start < mmap_min_addr)
1251                 return -EINVAL;
1252         if (start >= task_size)
1253                 return -EINVAL;
1254         if (len > task_size - start)
1255                 return -EINVAL;
1256         return 0;
1257 }
1258
1259 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1260                                      unsigned long vm_flags)
1261 {
1262         /* FIXME: add WP support to hugetlbfs and shmem */
1263         if (vm_flags & VM_UFFD_WP) {
1264                 if (is_vm_hugetlb_page(vma) || vma_is_shmem(vma))
1265                         return false;
1266         }
1267
1268         if (vm_flags & VM_UFFD_MINOR) {
1269                 if (!(is_vm_hugetlb_page(vma) || vma_is_shmem(vma)))
1270                         return false;
1271         }
1272
1273         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1274                vma_is_shmem(vma);
1275 }
1276
1277 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1278                                 unsigned long arg)
1279 {
1280         struct mm_struct *mm = ctx->mm;
1281         struct vm_area_struct *vma, *prev, *cur;
1282         int ret;
1283         struct uffdio_register uffdio_register;
1284         struct uffdio_register __user *user_uffdio_register;
1285         unsigned long vm_flags, new_flags;
1286         bool found;
1287         bool basic_ioctls;
1288         unsigned long start, end, vma_end;
1289
1290         user_uffdio_register = (struct uffdio_register __user *) arg;
1291
1292         ret = -EFAULT;
1293         if (copy_from_user(&uffdio_register, user_uffdio_register,
1294                            sizeof(uffdio_register)-sizeof(__u64)))
1295                 goto out;
1296
1297         ret = -EINVAL;
1298         if (!uffdio_register.mode)
1299                 goto out;
1300         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1301                 goto out;
1302         vm_flags = 0;
1303         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1304                 vm_flags |= VM_UFFD_MISSING;
1305         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1306 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1307                 goto out;
1308 #endif
1309                 vm_flags |= VM_UFFD_WP;
1310         }
1311         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1312 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1313                 goto out;
1314 #endif
1315                 vm_flags |= VM_UFFD_MINOR;
1316         }
1317
1318         ret = validate_range(mm, uffdio_register.range.start,
1319                              uffdio_register.range.len);
1320         if (ret)
1321                 goto out;
1322
1323         start = uffdio_register.range.start;
1324         end = start + uffdio_register.range.len;
1325
1326         ret = -ENOMEM;
1327         if (!mmget_not_zero(mm))
1328                 goto out;
1329
1330         mmap_write_lock(mm);
1331         vma = find_vma_prev(mm, start, &prev);
1332         if (!vma)
1333                 goto out_unlock;
1334
1335         /* check that there's at least one vma in the range */
1336         ret = -EINVAL;
1337         if (vma->vm_start >= end)
1338                 goto out_unlock;
1339
1340         /*
1341          * If the first vma contains huge pages, make sure start address
1342          * is aligned to huge page size.
1343          */
1344         if (is_vm_hugetlb_page(vma)) {
1345                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1346
1347                 if (start & (vma_hpagesize - 1))
1348                         goto out_unlock;
1349         }
1350
1351         /*
1352          * Search for not compatible vmas.
1353          */
1354         found = false;
1355         basic_ioctls = false;
1356         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1357                 cond_resched();
1358
1359                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1360                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1361
1362                 /* check not compatible vmas */
1363                 ret = -EINVAL;
1364                 if (!vma_can_userfault(cur, vm_flags))
1365                         goto out_unlock;
1366
1367                 /*
1368                  * UFFDIO_COPY will fill file holes even without
1369                  * PROT_WRITE. This check enforces that if this is a
1370                  * MAP_SHARED, the process has write permission to the backing
1371                  * file. If VM_MAYWRITE is set it also enforces that on a
1372                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1373                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1374                  */
1375                 ret = -EPERM;
1376                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1377                         goto out_unlock;
1378
1379                 /*
1380                  * If this vma contains ending address, and huge pages
1381                  * check alignment.
1382                  */
1383                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1384                     end > cur->vm_start) {
1385                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1386
1387                         ret = -EINVAL;
1388
1389                         if (end & (vma_hpagesize - 1))
1390                                 goto out_unlock;
1391                 }
1392                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1393                         goto out_unlock;
1394
1395                 /*
1396                  * Check that this vma isn't already owned by a
1397                  * different userfaultfd. We can't allow more than one
1398                  * userfaultfd to own a single vma simultaneously or we
1399                  * wouldn't know which one to deliver the userfaults to.
1400                  */
1401                 ret = -EBUSY;
1402                 if (cur->vm_userfaultfd_ctx.ctx &&
1403                     cur->vm_userfaultfd_ctx.ctx != ctx)
1404                         goto out_unlock;
1405
1406                 /*
1407                  * Note vmas containing huge pages
1408                  */
1409                 if (is_vm_hugetlb_page(cur))
1410                         basic_ioctls = true;
1411
1412                 found = true;
1413         }
1414         BUG_ON(!found);
1415
1416         if (vma->vm_start < start)
1417                 prev = vma;
1418
1419         ret = 0;
1420         do {
1421                 cond_resched();
1422
1423                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1424                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1425                        vma->vm_userfaultfd_ctx.ctx != ctx);
1426                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1427
1428                 /*
1429                  * Nothing to do: this vma is already registered into this
1430                  * userfaultfd and with the right tracking mode too.
1431                  */
1432                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1433                     (vma->vm_flags & vm_flags) == vm_flags)
1434                         goto skip;
1435
1436                 if (vma->vm_start > start)
1437                         start = vma->vm_start;
1438                 vma_end = min(end, vma->vm_end);
1439
1440                 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1441                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1442                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1443                                  vma_policy(vma),
1444                                  ((struct vm_userfaultfd_ctx){ ctx }));
1445                 if (prev) {
1446                         vma = prev;
1447                         goto next;
1448                 }
1449                 if (vma->vm_start < start) {
1450                         ret = split_vma(mm, vma, start, 1);
1451                         if (ret)
1452                                 break;
1453                 }
1454                 if (vma->vm_end > end) {
1455                         ret = split_vma(mm, vma, end, 0);
1456                         if (ret)
1457                                 break;
1458                 }
1459         next:
1460                 /*
1461                  * In the vma_merge() successful mprotect-like case 8:
1462                  * the next vma was merged into the current one and
1463                  * the current one has not been updated yet.
1464                  */
1465                 vma->vm_flags = new_flags;
1466                 vma->vm_userfaultfd_ctx.ctx = ctx;
1467
1468                 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1469                         hugetlb_unshare_all_pmds(vma);
1470
1471         skip:
1472                 prev = vma;
1473                 start = vma->vm_end;
1474                 vma = vma->vm_next;
1475         } while (vma && vma->vm_start < end);
1476 out_unlock:
1477         mmap_write_unlock(mm);
1478         mmput(mm);
1479         if (!ret) {
1480                 __u64 ioctls_out;
1481
1482                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1483                     UFFD_API_RANGE_IOCTLS;
1484
1485                 /*
1486                  * Declare the WP ioctl only if the WP mode is
1487                  * specified and all checks passed with the range
1488                  */
1489                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1490                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1491
1492                 /* CONTINUE ioctl is only supported for MINOR ranges. */
1493                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1494                         ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1495
1496                 /*
1497                  * Now that we scanned all vmas we can already tell
1498                  * userland which ioctls methods are guaranteed to
1499                  * succeed on this range.
1500                  */
1501                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1502                         ret = -EFAULT;
1503         }
1504 out:
1505         return ret;
1506 }
1507
1508 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1509                                   unsigned long arg)
1510 {
1511         struct mm_struct *mm = ctx->mm;
1512         struct vm_area_struct *vma, *prev, *cur;
1513         int ret;
1514         struct uffdio_range uffdio_unregister;
1515         unsigned long new_flags;
1516         bool found;
1517         unsigned long start, end, vma_end;
1518         const void __user *buf = (void __user *)arg;
1519
1520         ret = -EFAULT;
1521         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1522                 goto out;
1523
1524         ret = validate_range(mm, uffdio_unregister.start,
1525                              uffdio_unregister.len);
1526         if (ret)
1527                 goto out;
1528
1529         start = uffdio_unregister.start;
1530         end = start + uffdio_unregister.len;
1531
1532         ret = -ENOMEM;
1533         if (!mmget_not_zero(mm))
1534                 goto out;
1535
1536         mmap_write_lock(mm);
1537         vma = find_vma_prev(mm, start, &prev);
1538         if (!vma)
1539                 goto out_unlock;
1540
1541         /* check that there's at least one vma in the range */
1542         ret = -EINVAL;
1543         if (vma->vm_start >= end)
1544                 goto out_unlock;
1545
1546         /*
1547          * If the first vma contains huge pages, make sure start address
1548          * is aligned to huge page size.
1549          */
1550         if (is_vm_hugetlb_page(vma)) {
1551                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1552
1553                 if (start & (vma_hpagesize - 1))
1554                         goto out_unlock;
1555         }
1556
1557         /*
1558          * Search for not compatible vmas.
1559          */
1560         found = false;
1561         ret = -EINVAL;
1562         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1563                 cond_resched();
1564
1565                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1566                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1567
1568                 /*
1569                  * Check not compatible vmas, not strictly required
1570                  * here as not compatible vmas cannot have an
1571                  * userfaultfd_ctx registered on them, but this
1572                  * provides for more strict behavior to notice
1573                  * unregistration errors.
1574                  */
1575                 if (!vma_can_userfault(cur, cur->vm_flags))
1576                         goto out_unlock;
1577
1578                 found = true;
1579         }
1580         BUG_ON(!found);
1581
1582         if (vma->vm_start < start)
1583                 prev = vma;
1584
1585         ret = 0;
1586         do {
1587                 cond_resched();
1588
1589                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1590
1591                 /*
1592                  * Nothing to do: this vma is already registered into this
1593                  * userfaultfd and with the right tracking mode too.
1594                  */
1595                 if (!vma->vm_userfaultfd_ctx.ctx)
1596                         goto skip;
1597
1598                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1599
1600                 if (vma->vm_start > start)
1601                         start = vma->vm_start;
1602                 vma_end = min(end, vma->vm_end);
1603
1604                 if (userfaultfd_missing(vma)) {
1605                         /*
1606                          * Wake any concurrent pending userfault while
1607                          * we unregister, so they will not hang
1608                          * permanently and it avoids userland to call
1609                          * UFFDIO_WAKE explicitly.
1610                          */
1611                         struct userfaultfd_wake_range range;
1612                         range.start = start;
1613                         range.len = vma_end - start;
1614                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1615                 }
1616
1617                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1618                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1619                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1620                                  vma_policy(vma),
1621                                  NULL_VM_UFFD_CTX);
1622                 if (prev) {
1623                         vma = prev;
1624                         goto next;
1625                 }
1626                 if (vma->vm_start < start) {
1627                         ret = split_vma(mm, vma, start, 1);
1628                         if (ret)
1629                                 break;
1630                 }
1631                 if (vma->vm_end > end) {
1632                         ret = split_vma(mm, vma, end, 0);
1633                         if (ret)
1634                                 break;
1635                 }
1636         next:
1637                 /*
1638                  * In the vma_merge() successful mprotect-like case 8:
1639                  * the next vma was merged into the current one and
1640                  * the current one has not been updated yet.
1641                  */
1642                 vma->vm_flags = new_flags;
1643                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1644
1645         skip:
1646                 prev = vma;
1647                 start = vma->vm_end;
1648                 vma = vma->vm_next;
1649         } while (vma && vma->vm_start < end);
1650 out_unlock:
1651         mmap_write_unlock(mm);
1652         mmput(mm);
1653 out:
1654         return ret;
1655 }
1656
1657 /*
1658  * userfaultfd_wake may be used in combination with the
1659  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1660  */
1661 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1662                             unsigned long arg)
1663 {
1664         int ret;
1665         struct uffdio_range uffdio_wake;
1666         struct userfaultfd_wake_range range;
1667         const void __user *buf = (void __user *)arg;
1668
1669         ret = -EFAULT;
1670         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1671                 goto out;
1672
1673         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1674         if (ret)
1675                 goto out;
1676
1677         range.start = uffdio_wake.start;
1678         range.len = uffdio_wake.len;
1679
1680         /*
1681          * len == 0 means wake all and we don't want to wake all here,
1682          * so check it again to be sure.
1683          */
1684         VM_BUG_ON(!range.len);
1685
1686         wake_userfault(ctx, &range);
1687         ret = 0;
1688
1689 out:
1690         return ret;
1691 }
1692
1693 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1694                             unsigned long arg)
1695 {
1696         __s64 ret;
1697         struct uffdio_copy uffdio_copy;
1698         struct uffdio_copy __user *user_uffdio_copy;
1699         struct userfaultfd_wake_range range;
1700
1701         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1702
1703         ret = -EAGAIN;
1704         if (atomic_read(&ctx->mmap_changing))
1705                 goto out;
1706
1707         ret = -EFAULT;
1708         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1709                            /* don't copy "copy" last field */
1710                            sizeof(uffdio_copy)-sizeof(__s64)))
1711                 goto out;
1712
1713         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1714         if (ret)
1715                 goto out;
1716         /*
1717          * double check for wraparound just in case. copy_from_user()
1718          * will later check uffdio_copy.src + uffdio_copy.len to fit
1719          * in the userland range.
1720          */
1721         ret = -EINVAL;
1722         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1723                 goto out;
1724         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1725                 goto out;
1726         if (mmget_not_zero(ctx->mm)) {
1727                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1728                                    uffdio_copy.len, &ctx->mmap_changing,
1729                                    uffdio_copy.mode);
1730                 mmput(ctx->mm);
1731         } else {
1732                 return -ESRCH;
1733         }
1734         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1735                 return -EFAULT;
1736         if (ret < 0)
1737                 goto out;
1738         BUG_ON(!ret);
1739         /* len == 0 would wake all */
1740         range.len = ret;
1741         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1742                 range.start = uffdio_copy.dst;
1743                 wake_userfault(ctx, &range);
1744         }
1745         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1746 out:
1747         return ret;
1748 }
1749
1750 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1751                                 unsigned long arg)
1752 {
1753         __s64 ret;
1754         struct uffdio_zeropage uffdio_zeropage;
1755         struct uffdio_zeropage __user *user_uffdio_zeropage;
1756         struct userfaultfd_wake_range range;
1757
1758         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1759
1760         ret = -EAGAIN;
1761         if (atomic_read(&ctx->mmap_changing))
1762                 goto out;
1763
1764         ret = -EFAULT;
1765         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1766                            /* don't copy "zeropage" last field */
1767                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1768                 goto out;
1769
1770         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1771                              uffdio_zeropage.range.len);
1772         if (ret)
1773                 goto out;
1774         ret = -EINVAL;
1775         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1776                 goto out;
1777
1778         if (mmget_not_zero(ctx->mm)) {
1779                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1780                                      uffdio_zeropage.range.len,
1781                                      &ctx->mmap_changing);
1782                 mmput(ctx->mm);
1783         } else {
1784                 return -ESRCH;
1785         }
1786         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1787                 return -EFAULT;
1788         if (ret < 0)
1789                 goto out;
1790         /* len == 0 would wake all */
1791         BUG_ON(!ret);
1792         range.len = ret;
1793         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1794                 range.start = uffdio_zeropage.range.start;
1795                 wake_userfault(ctx, &range);
1796         }
1797         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1798 out:
1799         return ret;
1800 }
1801
1802 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1803                                     unsigned long arg)
1804 {
1805         int ret;
1806         struct uffdio_writeprotect uffdio_wp;
1807         struct uffdio_writeprotect __user *user_uffdio_wp;
1808         struct userfaultfd_wake_range range;
1809         bool mode_wp, mode_dontwake;
1810
1811         if (atomic_read(&ctx->mmap_changing))
1812                 return -EAGAIN;
1813
1814         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1815
1816         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1817                            sizeof(struct uffdio_writeprotect)))
1818                 return -EFAULT;
1819
1820         ret = validate_range(ctx->mm, uffdio_wp.range.start,
1821                              uffdio_wp.range.len);
1822         if (ret)
1823                 return ret;
1824
1825         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1826                                UFFDIO_WRITEPROTECT_MODE_WP))
1827                 return -EINVAL;
1828
1829         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1830         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1831
1832         if (mode_wp && mode_dontwake)
1833                 return -EINVAL;
1834
1835         ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1836                                   uffdio_wp.range.len, mode_wp,
1837                                   &ctx->mmap_changing);
1838         if (ret)
1839                 return ret;
1840
1841         if (!mode_wp && !mode_dontwake) {
1842                 range.start = uffdio_wp.range.start;
1843                 range.len = uffdio_wp.range.len;
1844                 wake_userfault(ctx, &range);
1845         }
1846         return ret;
1847 }
1848
1849 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1850 {
1851         __s64 ret;
1852         struct uffdio_continue uffdio_continue;
1853         struct uffdio_continue __user *user_uffdio_continue;
1854         struct userfaultfd_wake_range range;
1855
1856         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1857
1858         ret = -EAGAIN;
1859         if (atomic_read(&ctx->mmap_changing))
1860                 goto out;
1861
1862         ret = -EFAULT;
1863         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1864                            /* don't copy the output fields */
1865                            sizeof(uffdio_continue) - (sizeof(__s64))))
1866                 goto out;
1867
1868         ret = validate_range(ctx->mm, uffdio_continue.range.start,
1869                              uffdio_continue.range.len);
1870         if (ret)
1871                 goto out;
1872
1873         ret = -EINVAL;
1874         /* double check for wraparound just in case. */
1875         if (uffdio_continue.range.start + uffdio_continue.range.len <=
1876             uffdio_continue.range.start) {
1877                 goto out;
1878         }
1879         if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1880                 goto out;
1881
1882         if (mmget_not_zero(ctx->mm)) {
1883                 ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1884                                      uffdio_continue.range.len,
1885                                      &ctx->mmap_changing);
1886                 mmput(ctx->mm);
1887         } else {
1888                 return -ESRCH;
1889         }
1890
1891         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1892                 return -EFAULT;
1893         if (ret < 0)
1894                 goto out;
1895
1896         /* len == 0 would wake all */
1897         BUG_ON(!ret);
1898         range.len = ret;
1899         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1900                 range.start = uffdio_continue.range.start;
1901                 wake_userfault(ctx, &range);
1902         }
1903         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1904
1905 out:
1906         return ret;
1907 }
1908
1909 static inline unsigned int uffd_ctx_features(__u64 user_features)
1910 {
1911         /*
1912          * For the current set of features the bits just coincide
1913          */
1914         return (unsigned int)user_features;
1915 }
1916
1917 /*
1918  * userland asks for a certain API version and we return which bits
1919  * and ioctl commands are implemented in this kernel for such API
1920  * version or -EINVAL if unknown.
1921  */
1922 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1923                            unsigned long arg)
1924 {
1925         struct uffdio_api uffdio_api;
1926         void __user *buf = (void __user *)arg;
1927         int ret;
1928         __u64 features;
1929
1930         ret = -EINVAL;
1931         if (ctx->state != UFFD_STATE_WAIT_API)
1932                 goto out;
1933         ret = -EFAULT;
1934         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1935                 goto out;
1936         features = uffdio_api.features;
1937         ret = -EINVAL;
1938         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1939                 goto err_out;
1940         ret = -EPERM;
1941         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1942                 goto err_out;
1943         /* report all available features and ioctls to userland */
1944         uffdio_api.features = UFFD_API_FEATURES;
1945 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1946         uffdio_api.features &=
1947                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1948 #endif
1949 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1950         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1951 #endif
1952         uffdio_api.ioctls = UFFD_API_IOCTLS;
1953         ret = -EFAULT;
1954         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1955                 goto out;
1956         ctx->state = UFFD_STATE_RUNNING;
1957         /* only enable the requested features for this uffd context */
1958         ctx->features = uffd_ctx_features(features);
1959         ret = 0;
1960 out:
1961         return ret;
1962 err_out:
1963         memset(&uffdio_api, 0, sizeof(uffdio_api));
1964         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1965                 ret = -EFAULT;
1966         goto out;
1967 }
1968
1969 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1970                               unsigned long arg)
1971 {
1972         int ret = -EINVAL;
1973         struct userfaultfd_ctx *ctx = file->private_data;
1974
1975         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1976                 return -EINVAL;
1977
1978         switch(cmd) {
1979         case UFFDIO_API:
1980                 ret = userfaultfd_api(ctx, arg);
1981                 break;
1982         case UFFDIO_REGISTER:
1983                 ret = userfaultfd_register(ctx, arg);
1984                 break;
1985         case UFFDIO_UNREGISTER:
1986                 ret = userfaultfd_unregister(ctx, arg);
1987                 break;
1988         case UFFDIO_WAKE:
1989                 ret = userfaultfd_wake(ctx, arg);
1990                 break;
1991         case UFFDIO_COPY:
1992                 ret = userfaultfd_copy(ctx, arg);
1993                 break;
1994         case UFFDIO_ZEROPAGE:
1995                 ret = userfaultfd_zeropage(ctx, arg);
1996                 break;
1997         case UFFDIO_WRITEPROTECT:
1998                 ret = userfaultfd_writeprotect(ctx, arg);
1999                 break;
2000         case UFFDIO_CONTINUE:
2001                 ret = userfaultfd_continue(ctx, arg);
2002                 break;
2003         }
2004         return ret;
2005 }
2006
2007 #ifdef CONFIG_PROC_FS
2008 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2009 {
2010         struct userfaultfd_ctx *ctx = f->private_data;
2011         wait_queue_entry_t *wq;
2012         unsigned long pending = 0, total = 0;
2013
2014         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2015         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2016                 pending++;
2017                 total++;
2018         }
2019         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2020                 total++;
2021         }
2022         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2023
2024         /*
2025          * If more protocols will be added, there will be all shown
2026          * separated by a space. Like this:
2027          *      protocols: aa:... bb:...
2028          */
2029         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2030                    pending, total, UFFD_API, ctx->features,
2031                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2032 }
2033 #endif
2034
2035 static const struct file_operations userfaultfd_fops = {
2036 #ifdef CONFIG_PROC_FS
2037         .show_fdinfo    = userfaultfd_show_fdinfo,
2038 #endif
2039         .release        = userfaultfd_release,
2040         .poll           = userfaultfd_poll,
2041         .read           = userfaultfd_read,
2042         .unlocked_ioctl = userfaultfd_ioctl,
2043         .compat_ioctl   = compat_ptr_ioctl,
2044         .llseek         = noop_llseek,
2045 };
2046
2047 static void init_once_userfaultfd_ctx(void *mem)
2048 {
2049         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2050
2051         init_waitqueue_head(&ctx->fault_pending_wqh);
2052         init_waitqueue_head(&ctx->fault_wqh);
2053         init_waitqueue_head(&ctx->event_wqh);
2054         init_waitqueue_head(&ctx->fd_wqh);
2055         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2056 }
2057
2058 SYSCALL_DEFINE1(userfaultfd, int, flags)
2059 {
2060         struct userfaultfd_ctx *ctx;
2061         int fd;
2062
2063         if (!sysctl_unprivileged_userfaultfd &&
2064             (flags & UFFD_USER_MODE_ONLY) == 0 &&
2065             !capable(CAP_SYS_PTRACE)) {
2066                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2067                         "sysctl knob to 1 if kernel faults must be handled "
2068                         "without obtaining CAP_SYS_PTRACE capability\n");
2069                 return -EPERM;
2070         }
2071
2072         BUG_ON(!current->mm);
2073
2074         /* Check the UFFD_* constants for consistency.  */
2075         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2076         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2077         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2078
2079         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2080                 return -EINVAL;
2081
2082         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2083         if (!ctx)
2084                 return -ENOMEM;
2085
2086         refcount_set(&ctx->refcount, 1);
2087         ctx->flags = flags;
2088         ctx->features = 0;
2089         ctx->state = UFFD_STATE_WAIT_API;
2090         ctx->released = false;
2091         atomic_set(&ctx->mmap_changing, 0);
2092         ctx->mm = current->mm;
2093         /* prevent the mm struct to be freed */
2094         mmgrab(ctx->mm);
2095
2096         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2097                         O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2098         if (fd < 0) {
2099                 mmdrop(ctx->mm);
2100                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2101         }
2102         return fd;
2103 }
2104
2105 static int __init userfaultfd_init(void)
2106 {
2107         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2108                                                 sizeof(struct userfaultfd_ctx),
2109                                                 0,
2110                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2111                                                 init_once_userfaultfd_ctx);
2112         return 0;
2113 }
2114 __initcall(userfaultfd_init);