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         bool 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         WRITE_ONCE(ctx->mmap_changing, false);
 627         userfaultfd_ctx_put(ctx);
 628 }
 629
 630 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 631                                        struct userfaultfd_wait_queue *ewq)
 632 {
 633         ewq->msg.event = 0;
 634         wake_up_locked(&ctx->event_wqh);
 635         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 636 }
 637
 638 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 639 {
 640         struct userfaultfd_ctx *ctx = NULL, *octx;
 641         struct userfaultfd_fork_ctx *fctx;
 642
 643         octx = vma->vm_userfaultfd_ctx.ctx;
 644         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 645                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 646                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 647                 return 0;
 648         }
 649
 650         list_for_each_entry(fctx, fcs, list)
 651                 if (fctx->orig == octx) {
 652                         ctx = fctx->new;
 653                         break;
 654                 }
 655
 656         if (!ctx) {
 657                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 658                 if (!fctx)
 659                         return -ENOMEM;
 660
 661                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 662                 if (!ctx) {
 663                         kfree(fctx);
 664                         return -ENOMEM;
 665                 }
 666
 667                 refcount_set(&ctx->refcount, 1);
 668                 ctx->flags = octx->flags;
 669                 ctx->state = UFFD_STATE_RUNNING;
 670                 ctx->features = octx->features;
 671                 ctx->released = false;
 672                 ctx->mmap_changing = false;
 673                 ctx->mm = vma->vm_mm;
 674                 mmgrab(ctx->mm);
 675
 676                 userfaultfd_ctx_get(octx);
 677                 WRITE_ONCE(octx->mmap_changing, true);
 678                 fctx->orig = octx;
 679                 fctx->new = ctx;
 680                 list_add_tail(&fctx->list, fcs);
 681         }
 682
 683         vma->vm_userfaultfd_ctx.ctx = ctx;
 684         return 0;
 685 }
 686
 687 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 688 {
 689         struct userfaultfd_ctx *ctx = fctx->orig;
 690         struct userfaultfd_wait_queue ewq;
 691
 692         msg_init(&ewq.msg);
 693
 694         ewq.msg.event = UFFD_EVENT_FORK;
 695         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 696
 697         userfaultfd_event_wait_completion(ctx, &ewq);
 698 }
 699
 700 void dup_userfaultfd_complete(struct list_head *fcs)
 701 {
 702         struct userfaultfd_fork_ctx *fctx, *n;
 703
 704         list_for_each_entry_safe(fctx, n, fcs, list) {
 705                 dup_fctx(fctx);
 706                 list_del(&fctx->list);
 707                 kfree(fctx);
 708         }
 709 }
 710
 711 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 712                              struct vm_userfaultfd_ctx *vm_ctx)
 713 {
 714         struct userfaultfd_ctx *ctx;
 715
 716         ctx = vma->vm_userfaultfd_ctx.ctx;
 717
 718         if (!ctx)
 719                 return;
 720
 721         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 722                 vm_ctx->ctx = ctx;
 723                 userfaultfd_ctx_get(ctx);
 724                 WRITE_ONCE(ctx->mmap_changing, true);
 725         } else {
 726                 /* Drop uffd context if remap feature not enabled */
 727                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 728                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 729         }
 730 }
 731
 732 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 733                                  unsigned long from, unsigned long to,
 734                                  unsigned long len)
 735 {
 736         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 737         struct userfaultfd_wait_queue ewq;
 738
 739         if (!ctx)
 740                 return;
 741
 742         if (to & ~PAGE_MASK) {
 743                 userfaultfd_ctx_put(ctx);
 744                 return;
 745         }
 746
 747         msg_init(&ewq.msg);
 748
 749         ewq.msg.event = UFFD_EVENT_REMAP;
 750         ewq.msg.arg.remap.from = from;
 751         ewq.msg.arg.remap.to = to;
 752         ewq.msg.arg.remap.len = len;
 753
 754         userfaultfd_event_wait_completion(ctx, &ewq);
 755 }
 756
 757 bool userfaultfd_remove(struct vm_area_struct *vma,
 758                         unsigned long start, unsigned long end)
 759 {
 760         struct mm_struct *mm = vma->vm_mm;
 761         struct userfaultfd_ctx *ctx;
 762         struct userfaultfd_wait_queue ewq;
 763
 764         ctx = vma->vm_userfaultfd_ctx.ctx;
 765         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 766                 return true;
 767
 768         userfaultfd_ctx_get(ctx);
 769         WRITE_ONCE(ctx->mmap_changing, true);
 770         mmap_read_unlock(mm);
 771
 772         msg_init(&ewq.msg);
 773
 774         ewq.msg.event = UFFD_EVENT_REMOVE;
 775         ewq.msg.arg.remove.start = start;
 776         ewq.msg.arg.remove.end = end;
 777
 778         userfaultfd_event_wait_completion(ctx, &ewq);
 779
 780         return false;
 781 }
 782
 783 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 784                           unsigned long start, unsigned long end)
 785 {
 786         struct userfaultfd_unmap_ctx *unmap_ctx;
 787
 788         list_for_each_entry(unmap_ctx, unmaps, list)
 789                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 790                     unmap_ctx->end == end)
 791                         return true;
 792
 793         return false;
 794 }
 795
 796 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 797                            unsigned long start, unsigned long end,
 798                            struct list_head *unmaps)
 799 {
 800         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 801                 struct userfaultfd_unmap_ctx *unmap_ctx;
 802                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 803
 804                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 805                     has_unmap_ctx(ctx, unmaps, start, end))
 806                         continue;
 807
 808                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 809                 if (!unmap_ctx)
 810                         return -ENOMEM;
 811
 812                 userfaultfd_ctx_get(ctx);
 813                 WRITE_ONCE(ctx->mmap_changing, true);
 814                 unmap_ctx->ctx = ctx;
 815                 unmap_ctx->start = start;
 816                 unmap_ctx->end = end;
 817                 list_add_tail(&unmap_ctx->list, unmaps);
 818         }
 819
 820         return 0;
 821 }
 822
 823 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 824 {
 825         struct userfaultfd_unmap_ctx *ctx, *n;
 826         struct userfaultfd_wait_queue ewq;
 827
 828         list_for_each_entry_safe(ctx, n, uf, list) {
 829                 msg_init(&ewq.msg);
 830
 831                 ewq.msg.event = UFFD_EVENT_UNMAP;
 832                 ewq.msg.arg.remove.start = ctx->start;
 833                 ewq.msg.arg.remove.end = ctx->end;
 834
 835                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 836
 837                 list_del(&ctx->list);
 838                 kfree(ctx);
 839         }
 840 }
 841
 842 static int userfaultfd_release(struct inode *inode, struct file *file)
 843 {
 844         struct userfaultfd_ctx *ctx = file->private_data;
 845         struct mm_struct *mm = ctx->mm;
 846         struct vm_area_struct *vma, *prev;
 847         /* len == 0 means wake all */
 848         struct userfaultfd_wake_range range = { .len = 0, };
 849         unsigned long new_flags;
 850
 851         WRITE_ONCE(ctx->released, true);
 852
 853         if (!mmget_not_zero(mm))
 854                 goto wakeup;
 855
 856         /*
 857          * Flush page faults out of all CPUs. NOTE: all page faults
 858          * must be retried without returning VM_FAULT_SIGBUS if
 859          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 860          * changes while handle_userfault released the mmap_lock. So
 861          * it's critical that released is set to true (above), before
 862          * taking the mmap_lock for writing.
 863          */
 864         mmap_write_lock(mm);
 865         prev = NULL;
 866         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 867                 cond_resched();
 868                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 869                        !!(vma->vm_flags & __VM_UFFD_FLAGS));
 870                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 871                         prev = vma;
 872                         continue;
 873                 }
 874                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
 875                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 876                                  new_flags, vma->anon_vma,
 877                                  vma->vm_file, vma->vm_pgoff,
 878                                  vma_policy(vma),
 879                                  NULL_VM_UFFD_CTX);
 880                 if (prev)
 881                         vma = prev;
 882                 else
 883                         prev = vma;
 884                 vma->vm_flags = new_flags;
 885                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 886         }
 887         mmap_write_unlock(mm);
 888         mmput(mm);
 889 wakeup:
 890         /*
 891          * After no new page faults can wait on this fault_*wqh, flush
 892          * the last page faults that may have been already waiting on
 893          * the fault_*wqh.
 894          */
 895         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 896         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 897         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 898         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 899
 900         /* Flush pending events that may still wait on event_wqh */
 901         wake_up_all(&ctx->event_wqh);
 902
 903         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 904         userfaultfd_ctx_put(ctx);
 905         return 0;
 906 }
 907
 908 /* fault_pending_wqh.lock must be hold by the caller */
 909 static inline struct userfaultfd_wait_queue *find_userfault_in(
 910                 wait_queue_head_t *wqh)
 911 {
 912         wait_queue_entry_t *wq;
 913         struct userfaultfd_wait_queue *uwq;
 914
 915         lockdep_assert_held(&wqh->lock);
 916
 917         uwq = NULL;
 918         if (!waitqueue_active(wqh))
 919                 goto out;
 920         /* walk in reverse to provide FIFO behavior to read userfaults */
 921         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 922         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 923 out:
 924         return uwq;
 925 }
 926
 927 static inline struct userfaultfd_wait_queue *find_userfault(
 928                 struct userfaultfd_ctx *ctx)
 929 {
 930         return find_userfault_in(&ctx->fault_pending_wqh);
 931 }
 932
 933 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 934                 struct userfaultfd_ctx *ctx)
 935 {
 936         return find_userfault_in(&ctx->event_wqh);
 937 }
 938
 939 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 940 {
 941         struct userfaultfd_ctx *ctx = file->private_data;
 942         __poll_t ret;
 943
 944         poll_wait(file, &ctx->fd_wqh, wait);
 945
 946         switch (ctx->state) {
 947         case UFFD_STATE_WAIT_API:
 948                 return EPOLLERR;
 949         case UFFD_STATE_RUNNING:
 950                 /*
 951                  * poll() never guarantees that read won't block.
 952                  * userfaults can be waken before they're read().
 953                  */
 954                 if (unlikely(!(file->f_flags & O_NONBLOCK)))
 955                         return EPOLLERR;
 956                 /*
 957                  * lockless access to see if there are pending faults
 958                  * __pollwait last action is the add_wait_queue but
 959                  * the spin_unlock would allow the waitqueue_active to
 960                  * pass above the actual list_add inside
 961                  * add_wait_queue critical section. So use a full
 962                  * memory barrier to serialize the list_add write of
 963                  * add_wait_queue() with the waitqueue_active read
 964                  * below.
 965                  */
 966                 ret = 0;
 967                 smp_mb();
 968                 if (waitqueue_active(&ctx->fault_pending_wqh))
 969                         ret = EPOLLIN;
 970                 else if (waitqueue_active(&ctx->event_wqh))
 971                         ret = EPOLLIN;
 972
 973                 return ret;
 974         default:
 975                 WARN_ON_ONCE(1);
 976                 return EPOLLERR;
 977         }
 978 }
 979
 980 static const struct file_operations userfaultfd_fops;
 981
 982 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
 983                                   struct inode *inode,
 984                                   struct uffd_msg *msg)
 985 {
 986         int fd;
 987
 988         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
 989                         O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
 990         if (fd < 0)
 991                 return fd;
 992
 993         msg->arg.reserved.reserved1 = 0;
 994         msg->arg.fork.ufd = fd;
 995         return 0;
 996 }
 997
 998 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 999                                     struct uffd_msg *msg, struct inode *inode)
1000 {
1001         ssize_t ret;
1002         DECLARE_WAITQUEUE(wait, current);
1003         struct userfaultfd_wait_queue *uwq;
1004         /*
1005          * Handling fork event requires sleeping operations, so
1006          * we drop the event_wqh lock, then do these ops, then
1007          * lock it back and wake up the waiter. While the lock is
1008          * dropped the ewq may go away so we keep track of it
1009          * carefully.
1010          */
1011         LIST_HEAD(fork_event);
1012         struct userfaultfd_ctx *fork_nctx = NULL;
1013
1014         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1015         spin_lock_irq(&ctx->fd_wqh.lock);
1016         __add_wait_queue(&ctx->fd_wqh, &wait);
1017         for (;;) {
1018                 set_current_state(TASK_INTERRUPTIBLE);
1019                 spin_lock(&ctx->fault_pending_wqh.lock);
1020                 uwq = find_userfault(ctx);
1021                 if (uwq) {
1022                         /*
1023                          * Use a seqcount to repeat the lockless check
1024                          * in wake_userfault() to avoid missing
1025                          * wakeups because during the refile both
1026                          * waitqueue could become empty if this is the
1027                          * only userfault.
1028                          */
1029                         write_seqcount_begin(&ctx->refile_seq);
1030
1031                         /*
1032                          * The fault_pending_wqh.lock prevents the uwq
1033                          * to disappear from under us.
1034                          *
1035                          * Refile this userfault from
1036                          * fault_pending_wqh to fault_wqh, it's not
1037                          * pending anymore after we read it.
1038                          *
1039                          * Use list_del() by hand (as
1040                          * userfaultfd_wake_function also uses
1041                          * list_del_init() by hand) to be sure nobody
1042                          * changes __remove_wait_queue() to use
1043                          * list_del_init() in turn breaking the
1044                          * !list_empty_careful() check in
1045                          * handle_userfault(). The uwq->wq.head list
1046                          * must never be empty at any time during the
1047                          * refile, or the waitqueue could disappear
1048                          * from under us. The "wait_queue_head_t"
1049                          * parameter of __remove_wait_queue() is unused
1050                          * anyway.
1051                          */
1052                         list_del(&uwq->wq.entry);
1053                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1054
1055                         write_seqcount_end(&ctx->refile_seq);
1056
1057                         /* careful to always initialize msg if ret == 0 */
1058                         *msg = uwq->msg;
1059                         spin_unlock(&ctx->fault_pending_wqh.lock);
1060                         ret = 0;
1061                         break;
1062                 }
1063                 spin_unlock(&ctx->fault_pending_wqh.lock);
1064
1065                 spin_lock(&ctx->event_wqh.lock);
1066                 uwq = find_userfault_evt(ctx);
1067                 if (uwq) {
1068                         *msg = uwq->msg;
1069
1070                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1071                                 fork_nctx = (struct userfaultfd_ctx *)
1072                                         (unsigned long)
1073                                         uwq->msg.arg.reserved.reserved1;
1074                                 list_move(&uwq->wq.entry, &fork_event);
1075                                 /*
1076                                  * fork_nctx can be freed as soon as
1077                                  * we drop the lock, unless we take a
1078                                  * reference on it.
1079                                  */
1080                                 userfaultfd_ctx_get(fork_nctx);
1081                                 spin_unlock(&ctx->event_wqh.lock);
1082                                 ret = 0;
1083                                 break;
1084                         }
1085
1086                         userfaultfd_event_complete(ctx, uwq);
1087                         spin_unlock(&ctx->event_wqh.lock);
1088                         ret = 0;
1089                         break;
1090                 }
1091                 spin_unlock(&ctx->event_wqh.lock);
1092
1093                 if (signal_pending(current)) {
1094                         ret = -ERESTARTSYS;
1095                         break;
1096                 }
1097                 if (no_wait) {
1098                         ret = -EAGAIN;
1099                         break;
1100                 }
1101                 spin_unlock_irq(&ctx->fd_wqh.lock);
1102                 schedule();
1103                 spin_lock_irq(&ctx->fd_wqh.lock);
1104         }
1105         __remove_wait_queue(&ctx->fd_wqh, &wait);
1106         __set_current_state(TASK_RUNNING);
1107         spin_unlock_irq(&ctx->fd_wqh.lock);
1108
1109         if (!ret && msg->event == UFFD_EVENT_FORK) {
1110                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1111                 spin_lock_irq(&ctx->event_wqh.lock);
1112                 if (!list_empty(&fork_event)) {
1113                         /*
1114                          * The fork thread didn't abort, so we can
1115                          * drop the temporary refcount.
1116                          */
1117                         userfaultfd_ctx_put(fork_nctx);
1118
1119                         uwq = list_first_entry(&fork_event,
1120                                                typeof(*uwq),
1121                                                wq.entry);
1122                         /*
1123                          * If fork_event list wasn't empty and in turn
1124                          * the event wasn't already released by fork
1125                          * (the event is allocated on fork kernel
1126                          * stack), put the event back to its place in
1127                          * the event_wq. fork_event head will be freed
1128                          * as soon as we return so the event cannot
1129                          * stay queued there no matter the current
1130                          * "ret" value.
1131                          */
1132                         list_del(&uwq->wq.entry);
1133                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1134
1135                         /*
1136                          * Leave the event in the waitqueue and report
1137                          * error to userland if we failed to resolve
1138                          * the userfault fork.
1139                          */
1140                         if (likely(!ret))
1141                                 userfaultfd_event_complete(ctx, uwq);
1142                 } else {
1143                         /*
1144                          * Here the fork thread aborted and the
1145                          * refcount from the fork thread on fork_nctx
1146                          * has already been released. We still hold
1147                          * the reference we took before releasing the
1148                          * lock above. If resolve_userfault_fork
1149                          * failed we've to drop it because the
1150                          * fork_nctx has to be freed in such case. If
1151                          * it succeeded we'll hold it because the new
1152                          * uffd references it.
1153                          */
1154                         if (ret)
1155                                 userfaultfd_ctx_put(fork_nctx);
1156                 }
1157                 spin_unlock_irq(&ctx->event_wqh.lock);
1158         }
1159
1160         return ret;
1161 }
1162
1163 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1164                                 size_t count, loff_t *ppos)
1165 {
1166         struct userfaultfd_ctx *ctx = file->private_data;
1167         ssize_t _ret, ret = 0;
1168         struct uffd_msg msg;
1169         int no_wait = file->f_flags & O_NONBLOCK;
1170         struct inode *inode = file_inode(file);
1171
1172         if (ctx->state == UFFD_STATE_WAIT_API)
1173                 return -EINVAL;
1174
1175         for (;;) {
1176                 if (count < sizeof(msg))
1177                         return ret ? ret : -EINVAL;
1178                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1179                 if (_ret < 0)
1180                         return ret ? ret : _ret;
1181                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1182                         return ret ? ret : -EFAULT;
1183                 ret += sizeof(msg);
1184                 buf += sizeof(msg);
1185                 count -= sizeof(msg);
1186                 /*
1187                  * Allow to read more than one fault at time but only
1188                  * block if waiting for the very first one.
1189                  */
1190                 no_wait = O_NONBLOCK;
1191         }
1192 }
1193
1194 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1195                              struct userfaultfd_wake_range *range)
1196 {
1197         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1198         /* wake all in the range and autoremove */
1199         if (waitqueue_active(&ctx->fault_pending_wqh))
1200                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1201                                      range);
1202         if (waitqueue_active(&ctx->fault_wqh))
1203                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1204         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1205 }
1206
1207 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1208                                            struct userfaultfd_wake_range *range)
1209 {
1210         unsigned seq;
1211         bool need_wakeup;
1212
1213         /*
1214          * To be sure waitqueue_active() is not reordered by the CPU
1215          * before the pagetable update, use an explicit SMP memory
1216          * barrier here. PT lock release or mmap_read_unlock(mm) still
1217          * have release semantics that can allow the
1218          * waitqueue_active() to be reordered before the pte update.
1219          */
1220         smp_mb();
1221
1222         /*
1223          * Use waitqueue_active because it's very frequent to
1224          * change the address space atomically even if there are no
1225          * userfaults yet. So we take the spinlock only when we're
1226          * sure we've userfaults to wake.
1227          */
1228         do {
1229                 seq = read_seqcount_begin(&ctx->refile_seq);
1230                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1231                         waitqueue_active(&ctx->fault_wqh);
1232                 cond_resched();
1233         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1234         if (need_wakeup)
1235                 __wake_userfault(ctx, range);
1236 }
1237
1238 static __always_inline int validate_range(struct mm_struct *mm,
1239                                           __u64 *start, __u64 len)
1240 {
1241         __u64 task_size = mm->task_size;
1242
1243         *start = untagged_addr(*start);
1244
1245         if (*start & ~PAGE_MASK)
1246                 return -EINVAL;
1247         if (len & ~PAGE_MASK)
1248                 return -EINVAL;
1249         if (!len)
1250                 return -EINVAL;
1251         if (*start < mmap_min_addr)
1252                 return -EINVAL;
1253         if (*start >= task_size)
1254                 return -EINVAL;
1255         if (len > task_size - *start)
1256                 return -EINVAL;
1257         return 0;
1258 }
1259
1260 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1261                                      unsigned long vm_flags)
1262 {
1263         /* FIXME: add WP support to hugetlbfs and shmem */
1264         if (vm_flags & VM_UFFD_WP) {
1265                 if (is_vm_hugetlb_page(vma) || vma_is_shmem(vma))
1266                         return false;
1267         }
1268
1269         if (vm_flags & VM_UFFD_MINOR) {
1270                 if (!(is_vm_hugetlb_page(vma) || vma_is_shmem(vma)))
1271                         return false;
1272         }
1273
1274         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1275                vma_is_shmem(vma);
1276 }
1277
1278 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1279                                 unsigned long arg)
1280 {
1281         struct mm_struct *mm = ctx->mm;
1282         struct vm_area_struct *vma, *prev, *cur;
1283         int ret;
1284         struct uffdio_register uffdio_register;
1285         struct uffdio_register __user *user_uffdio_register;
1286         unsigned long vm_flags, new_flags;
1287         bool found;
1288         bool basic_ioctls;
1289         unsigned long start, end, vma_end;
1290
1291         user_uffdio_register = (struct uffdio_register __user *) arg;
1292
1293         ret = -EFAULT;
1294         if (copy_from_user(&uffdio_register, user_uffdio_register,
1295                            sizeof(uffdio_register)-sizeof(__u64)))
1296                 goto out;
1297
1298         ret = -EINVAL;
1299         if (!uffdio_register.mode)
1300                 goto out;
1301         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1302                 goto out;
1303         vm_flags = 0;
1304         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1305                 vm_flags |= VM_UFFD_MISSING;
1306         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1307 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1308                 goto out;
1309 #endif
1310                 vm_flags |= VM_UFFD_WP;
1311         }
1312         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1313 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1314                 goto out;
1315 #endif
1316                 vm_flags |= VM_UFFD_MINOR;
1317         }
1318
1319         ret = validate_range(mm, &uffdio_register.range.start,
1320                              uffdio_register.range.len);
1321         if (ret)
1322                 goto out;
1323
1324         start = uffdio_register.range.start;
1325         end = start + uffdio_register.range.len;
1326
1327         ret = -ENOMEM;
1328         if (!mmget_not_zero(mm))
1329                 goto out;
1330
1331         mmap_write_lock(mm);
1332         vma = find_vma_prev(mm, start, &prev);
1333         if (!vma)
1334                 goto out_unlock;
1335
1336         /* check that there's at least one vma in the range */
1337         ret = -EINVAL;
1338         if (vma->vm_start >= end)
1339                 goto out_unlock;
1340
1341         /*
1342          * If the first vma contains huge pages, make sure start address
1343          * is aligned to huge page size.
1344          */
1345         if (is_vm_hugetlb_page(vma)) {
1346                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1347
1348                 if (start & (vma_hpagesize - 1))
1349                         goto out_unlock;
1350         }
1351
1352         /*
1353          * Search for not compatible vmas.
1354          */
1355         found = false;
1356         basic_ioctls = false;
1357         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1358                 cond_resched();
1359
1360                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1361                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1362
1363                 /* check not compatible vmas */
1364                 ret = -EINVAL;
1365                 if (!vma_can_userfault(cur, vm_flags))
1366                         goto out_unlock;
1367
1368                 /*
1369                  * UFFDIO_COPY will fill file holes even without
1370                  * PROT_WRITE. This check enforces that if this is a
1371                  * MAP_SHARED, the process has write permission to the backing
1372                  * file. If VM_MAYWRITE is set it also enforces that on a
1373                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1374                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1375                  */
1376                 ret = -EPERM;
1377                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1378                         goto out_unlock;
1379
1380                 /*
1381                  * If this vma contains ending address, and huge pages
1382                  * check alignment.
1383                  */
1384                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1385                     end > cur->vm_start) {
1386                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1387
1388                         ret = -EINVAL;
1389
1390                         if (end & (vma_hpagesize - 1))
1391                                 goto out_unlock;
1392                 }
1393                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1394                         goto out_unlock;
1395
1396                 /*
1397                  * Check that this vma isn't already owned by a
1398                  * different userfaultfd. We can't allow more than one
1399                  * userfaultfd to own a single vma simultaneously or we
1400                  * wouldn't know which one to deliver the userfaults to.
1401                  */
1402                 ret = -EBUSY;
1403                 if (cur->vm_userfaultfd_ctx.ctx &&
1404                     cur->vm_userfaultfd_ctx.ctx != ctx)
1405                         goto out_unlock;
1406
1407                 /*
1408                  * Note vmas containing huge pages
1409                  */
1410                 if (is_vm_hugetlb_page(cur))
1411                         basic_ioctls = true;
1412
1413                 found = true;
1414         }
1415         BUG_ON(!found);
1416
1417         if (vma->vm_start < start)
1418                 prev = vma;
1419
1420         ret = 0;
1421         do {
1422                 cond_resched();
1423
1424                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1425                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1426                        vma->vm_userfaultfd_ctx.ctx != ctx);
1427                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1428
1429                 /*
1430                  * Nothing to do: this vma is already registered into this
1431                  * userfaultfd and with the right tracking mode too.
1432                  */
1433                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1434                     (vma->vm_flags & vm_flags) == vm_flags)
1435                         goto skip;
1436
1437                 if (vma->vm_start > start)
1438                         start = vma->vm_start;
1439                 vma_end = min(end, vma->vm_end);
1440
1441                 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1442                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1443                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1444                                  vma_policy(vma),
1445                                  ((struct vm_userfaultfd_ctx){ ctx }));
1446                 if (prev) {
1447                         vma = prev;
1448                         goto next;
1449                 }
1450                 if (vma->vm_start < start) {
1451                         ret = split_vma(mm, vma, start, 1);
1452                         if (ret)
1453                                 break;
1454                 }
1455                 if (vma->vm_end > end) {
1456                         ret = split_vma(mm, vma, end, 0);
1457                         if (ret)
1458                                 break;
1459                 }
1460         next:
1461                 /*
1462                  * In the vma_merge() successful mprotect-like case 8:
1463                  * the next vma was merged into the current one and
1464                  * the current one has not been updated yet.
1465                  */
1466                 vma->vm_flags = new_flags;
1467                 vma->vm_userfaultfd_ctx.ctx = ctx;
1468
1469                 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1470                         hugetlb_unshare_all_pmds(vma);
1471
1472         skip:
1473                 prev = vma;
1474                 start = vma->vm_end;
1475                 vma = vma->vm_next;
1476         } while (vma && vma->vm_start < end);
1477 out_unlock:
1478         mmap_write_unlock(mm);
1479         mmput(mm);
1480         if (!ret) {
1481                 __u64 ioctls_out;
1482
1483                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1484                     UFFD_API_RANGE_IOCTLS;
1485
1486                 /*
1487                  * Declare the WP ioctl only if the WP mode is
1488                  * specified and all checks passed with the range
1489                  */
1490                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1491                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1492
1493                 /* CONTINUE ioctl is only supported for MINOR ranges. */
1494                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1495                         ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1496
1497                 /*
1498                  * Now that we scanned all vmas we can already tell
1499                  * userland which ioctls methods are guaranteed to
1500                  * succeed on this range.
1501                  */
1502                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1503                         ret = -EFAULT;
1504         }
1505 out:
1506         return ret;
1507 }
1508
1509 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1510                                   unsigned long arg)
1511 {
1512         struct mm_struct *mm = ctx->mm;
1513         struct vm_area_struct *vma, *prev, *cur;
1514         int ret;
1515         struct uffdio_range uffdio_unregister;
1516         unsigned long new_flags;
1517         bool found;
1518         unsigned long start, end, vma_end;
1519         const void __user *buf = (void __user *)arg;
1520
1521         ret = -EFAULT;
1522         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1523                 goto out;
1524
1525         ret = validate_range(mm, &uffdio_unregister.start,
1526                              uffdio_unregister.len);
1527         if (ret)
1528                 goto out;
1529
1530         start = uffdio_unregister.start;
1531         end = start + uffdio_unregister.len;
1532
1533         ret = -ENOMEM;
1534         if (!mmget_not_zero(mm))
1535                 goto out;
1536
1537         mmap_write_lock(mm);
1538         vma = find_vma_prev(mm, start, &prev);
1539         if (!vma)
1540                 goto out_unlock;
1541
1542         /* check that there's at least one vma in the range */
1543         ret = -EINVAL;
1544         if (vma->vm_start >= end)
1545                 goto out_unlock;
1546
1547         /*
1548          * If the first vma contains huge pages, make sure start address
1549          * is aligned to huge page size.
1550          */
1551         if (is_vm_hugetlb_page(vma)) {
1552                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1553
1554                 if (start & (vma_hpagesize - 1))
1555                         goto out_unlock;
1556         }
1557
1558         /*
1559          * Search for not compatible vmas.
1560          */
1561         found = false;
1562         ret = -EINVAL;
1563         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1564                 cond_resched();
1565
1566                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1567                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1568
1569                 /*
1570                  * Check not compatible vmas, not strictly required
1571                  * here as not compatible vmas cannot have an
1572                  * userfaultfd_ctx registered on them, but this
1573                  * provides for more strict behavior to notice
1574                  * unregistration errors.
1575                  */
1576                 if (!vma_can_userfault(cur, cur->vm_flags))
1577                         goto out_unlock;
1578
1579                 found = true;
1580         }
1581         BUG_ON(!found);
1582
1583         if (vma->vm_start < start)
1584                 prev = vma;
1585
1586         ret = 0;
1587         do {
1588                 cond_resched();
1589
1590                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1591
1592                 /*
1593                  * Nothing to do: this vma is already registered into this
1594                  * userfaultfd and with the right tracking mode too.
1595                  */
1596                 if (!vma->vm_userfaultfd_ctx.ctx)
1597                         goto skip;
1598
1599                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1600
1601                 if (vma->vm_start > start)
1602                         start = vma->vm_start;
1603                 vma_end = min(end, vma->vm_end);
1604
1605                 if (userfaultfd_missing(vma)) {
1606                         /*
1607                          * Wake any concurrent pending userfault while
1608                          * we unregister, so they will not hang
1609                          * permanently and it avoids userland to call
1610                          * UFFDIO_WAKE explicitly.
1611                          */
1612                         struct userfaultfd_wake_range range;
1613                         range.start = start;
1614                         range.len = vma_end - start;
1615                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1616                 }
1617
1618                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1619                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1620                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1621                                  vma_policy(vma),
1622                                  NULL_VM_UFFD_CTX);
1623                 if (prev) {
1624                         vma = prev;
1625                         goto next;
1626                 }
1627                 if (vma->vm_start < start) {
1628                         ret = split_vma(mm, vma, start, 1);
1629                         if (ret)
1630                                 break;
1631                 }
1632                 if (vma->vm_end > end) {
1633                         ret = split_vma(mm, vma, end, 0);
1634                         if (ret)
1635                                 break;
1636                 }
1637         next:
1638                 /*
1639                  * In the vma_merge() successful mprotect-like case 8:
1640                  * the next vma was merged into the current one and
1641                  * the current one has not been updated yet.
1642                  */
1643                 vma->vm_flags = new_flags;
1644                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645
1646         skip:
1647                 prev = vma;
1648                 start = vma->vm_end;
1649                 vma = vma->vm_next;
1650         } while (vma && vma->vm_start < end);
1651 out_unlock:
1652         mmap_write_unlock(mm);
1653         mmput(mm);
1654 out:
1655         return ret;
1656 }
1657
1658 /*
1659  * userfaultfd_wake may be used in combination with the
1660  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1661  */
1662 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1663                             unsigned long arg)
1664 {
1665         int ret;
1666         struct uffdio_range uffdio_wake;
1667         struct userfaultfd_wake_range range;
1668         const void __user *buf = (void __user *)arg;
1669
1670         ret = -EFAULT;
1671         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1672                 goto out;
1673
1674         ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1675         if (ret)
1676                 goto out;
1677
1678         range.start = uffdio_wake.start;
1679         range.len = uffdio_wake.len;
1680
1681         /*
1682          * len == 0 means wake all and we don't want to wake all here,
1683          * so check it again to be sure.
1684          */
1685         VM_BUG_ON(!range.len);
1686
1687         wake_userfault(ctx, &range);
1688         ret = 0;
1689
1690 out:
1691         return ret;
1692 }
1693
1694 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1695                             unsigned long arg)
1696 {
1697         __s64 ret;
1698         struct uffdio_copy uffdio_copy;
1699         struct uffdio_copy __user *user_uffdio_copy;
1700         struct userfaultfd_wake_range range;
1701
1702         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1703
1704         ret = -EAGAIN;
1705         if (READ_ONCE(ctx->mmap_changing))
1706                 goto out;
1707
1708         ret = -EFAULT;
1709         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1710                            /* don't copy "copy" last field */
1711                            sizeof(uffdio_copy)-sizeof(__s64)))
1712                 goto out;
1713
1714         ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1715         if (ret)
1716                 goto out;
1717         /*
1718          * double check for wraparound just in case. copy_from_user()
1719          * will later check uffdio_copy.src + uffdio_copy.len to fit
1720          * in the userland range.
1721          */
1722         ret = -EINVAL;
1723         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1724                 goto out;
1725         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1726                 goto out;
1727         if (mmget_not_zero(ctx->mm)) {
1728                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1729                                    uffdio_copy.len, &ctx->mmap_changing,
1730                                    uffdio_copy.mode);
1731                 mmput(ctx->mm);
1732         } else {
1733                 return -ESRCH;
1734         }
1735         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1736                 return -EFAULT;
1737         if (ret < 0)
1738                 goto out;
1739         BUG_ON(!ret);
1740         /* len == 0 would wake all */
1741         range.len = ret;
1742         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1743                 range.start = uffdio_copy.dst;
1744                 wake_userfault(ctx, &range);
1745         }
1746         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1747 out:
1748         return ret;
1749 }
1750
1751 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1752                                 unsigned long arg)
1753 {
1754         __s64 ret;
1755         struct uffdio_zeropage uffdio_zeropage;
1756         struct uffdio_zeropage __user *user_uffdio_zeropage;
1757         struct userfaultfd_wake_range range;
1758
1759         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1760
1761         ret = -EAGAIN;
1762         if (READ_ONCE(ctx->mmap_changing))
1763                 goto out;
1764
1765         ret = -EFAULT;
1766         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1767                            /* don't copy "zeropage" last field */
1768                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1769                 goto out;
1770
1771         ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1772                              uffdio_zeropage.range.len);
1773         if (ret)
1774                 goto out;
1775         ret = -EINVAL;
1776         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1777                 goto out;
1778
1779         if (mmget_not_zero(ctx->mm)) {
1780                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1781                                      uffdio_zeropage.range.len,
1782                                      &ctx->mmap_changing);
1783                 mmput(ctx->mm);
1784         } else {
1785                 return -ESRCH;
1786         }
1787         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1788                 return -EFAULT;
1789         if (ret < 0)
1790                 goto out;
1791         /* len == 0 would wake all */
1792         BUG_ON(!ret);
1793         range.len = ret;
1794         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1795                 range.start = uffdio_zeropage.range.start;
1796                 wake_userfault(ctx, &range);
1797         }
1798         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1799 out:
1800         return ret;
1801 }
1802
1803 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1804                                     unsigned long arg)
1805 {
1806         int ret;
1807         struct uffdio_writeprotect uffdio_wp;
1808         struct uffdio_writeprotect __user *user_uffdio_wp;
1809         struct userfaultfd_wake_range range;
1810         bool mode_wp, mode_dontwake;
1811
1812         if (READ_ONCE(ctx->mmap_changing))
1813                 return -EAGAIN;
1814
1815         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1816
1817         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1818                            sizeof(struct uffdio_writeprotect)))
1819                 return -EFAULT;
1820
1821         ret = validate_range(ctx->mm, &uffdio_wp.range.start,
1822                              uffdio_wp.range.len);
1823         if (ret)
1824                 return ret;
1825
1826         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1827                                UFFDIO_WRITEPROTECT_MODE_WP))
1828                 return -EINVAL;
1829
1830         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1831         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1832
1833         if (mode_wp && mode_dontwake)
1834                 return -EINVAL;
1835
1836         ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1837                                   uffdio_wp.range.len, mode_wp,
1838                                   &ctx->mmap_changing);
1839         if (ret)
1840                 return ret;
1841
1842         if (!mode_wp && !mode_dontwake) {
1843                 range.start = uffdio_wp.range.start;
1844                 range.len = uffdio_wp.range.len;
1845                 wake_userfault(ctx, &range);
1846         }
1847         return ret;
1848 }
1849
1850 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1851 {
1852         __s64 ret;
1853         struct uffdio_continue uffdio_continue;
1854         struct uffdio_continue __user *user_uffdio_continue;
1855         struct userfaultfd_wake_range range;
1856
1857         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1858
1859         ret = -EAGAIN;
1860         if (READ_ONCE(ctx->mmap_changing))
1861                 goto out;
1862
1863         ret = -EFAULT;
1864         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1865                            /* don't copy the output fields */
1866                            sizeof(uffdio_continue) - (sizeof(__s64))))
1867                 goto out;
1868
1869         ret = validate_range(ctx->mm, &uffdio_continue.range.start,
1870                              uffdio_continue.range.len);
1871         if (ret)
1872                 goto out;
1873
1874         ret = -EINVAL;
1875         /* double check for wraparound just in case. */
1876         if (uffdio_continue.range.start + uffdio_continue.range.len <=
1877             uffdio_continue.range.start) {
1878                 goto out;
1879         }
1880         if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1881                 goto out;
1882
1883         if (mmget_not_zero(ctx->mm)) {
1884                 ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1885                                      uffdio_continue.range.len,
1886                                      &ctx->mmap_changing);
1887                 mmput(ctx->mm);
1888         } else {
1889                 return -ESRCH;
1890         }
1891
1892         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1893                 return -EFAULT;
1894         if (ret < 0)
1895                 goto out;
1896
1897         /* len == 0 would wake all */
1898         BUG_ON(!ret);
1899         range.len = ret;
1900         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1901                 range.start = uffdio_continue.range.start;
1902                 wake_userfault(ctx, &range);
1903         }
1904         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1905
1906 out:
1907         return ret;
1908 }
1909
1910 static inline unsigned int uffd_ctx_features(__u64 user_features)
1911 {
1912         /*
1913          * For the current set of features the bits just coincide
1914          */
1915         return (unsigned int)user_features;
1916 }
1917
1918 /*
1919  * userland asks for a certain API version and we return which bits
1920  * and ioctl commands are implemented in this kernel for such API
1921  * version or -EINVAL if unknown.
1922  */
1923 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1924                            unsigned long arg)
1925 {
1926         struct uffdio_api uffdio_api;
1927         void __user *buf = (void __user *)arg;
1928         int ret;
1929         __u64 features;
1930
1931         ret = -EINVAL;
1932         if (ctx->state != UFFD_STATE_WAIT_API)
1933                 goto out;
1934         ret = -EFAULT;
1935         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1936                 goto out;
1937         features = uffdio_api.features;
1938         ret = -EINVAL;
1939         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1940                 goto err_out;
1941         ret = -EPERM;
1942         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1943                 goto err_out;
1944         /* report all available features and ioctls to userland */
1945         uffdio_api.features = UFFD_API_FEATURES;
1946 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1947         uffdio_api.features &=
1948                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1949 #endif
1950 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1951         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1952 #endif
1953         uffdio_api.ioctls = UFFD_API_IOCTLS;
1954         ret = -EFAULT;
1955         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1956                 goto out;
1957         ctx->state = UFFD_STATE_RUNNING;
1958         /* only enable the requested features for this uffd context */
1959         ctx->features = uffd_ctx_features(features);
1960         ret = 0;
1961 out:
1962         return ret;
1963 err_out:
1964         memset(&uffdio_api, 0, sizeof(uffdio_api));
1965         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1966                 ret = -EFAULT;
1967         goto out;
1968 }
1969
1970 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1971                               unsigned long arg)
1972 {
1973         int ret = -EINVAL;
1974         struct userfaultfd_ctx *ctx = file->private_data;
1975
1976         if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
1977                 return -EINVAL;
1978
1979         switch(cmd) {
1980         case UFFDIO_API:
1981                 ret = userfaultfd_api(ctx, arg);
1982                 break;
1983         case UFFDIO_REGISTER:
1984                 ret = userfaultfd_register(ctx, arg);
1985                 break;
1986         case UFFDIO_UNREGISTER:
1987                 ret = userfaultfd_unregister(ctx, arg);
1988                 break;
1989         case UFFDIO_WAKE:
1990                 ret = userfaultfd_wake(ctx, arg);
1991                 break;
1992         case UFFDIO_COPY:
1993                 ret = userfaultfd_copy(ctx, arg);
1994                 break;
1995         case UFFDIO_ZEROPAGE:
1996                 ret = userfaultfd_zeropage(ctx, arg);
1997                 break;
1998         case UFFDIO_WRITEPROTECT:
1999                 ret = userfaultfd_writeprotect(ctx, arg);
2000                 break;
2001         case UFFDIO_CONTINUE:
2002                 ret = userfaultfd_continue(ctx, arg);
2003                 break;
2004         }
2005         return ret;
2006 }
2007
2008 #ifdef CONFIG_PROC_FS
2009 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2010 {
2011         struct userfaultfd_ctx *ctx = f->private_data;
2012         wait_queue_entry_t *wq;
2013         unsigned long pending = 0, total = 0;
2014
2015         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2016         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2017                 pending++;
2018                 total++;
2019         }
2020         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2021                 total++;
2022         }
2023         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2024
2025         /*
2026          * If more protocols will be added, there will be all shown
2027          * separated by a space. Like this:
2028          *      protocols: aa:... bb:...
2029          */
2030         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2031                    pending, total, UFFD_API, ctx->features,
2032                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2033 }
2034 #endif
2035
2036 static const struct file_operations userfaultfd_fops = {
2037 #ifdef CONFIG_PROC_FS
2038         .show_fdinfo    = userfaultfd_show_fdinfo,
2039 #endif
2040         .release        = userfaultfd_release,
2041         .poll           = userfaultfd_poll,
2042         .read           = userfaultfd_read,
2043         .unlocked_ioctl = userfaultfd_ioctl,
2044         .compat_ioctl   = compat_ptr_ioctl,
2045         .llseek         = noop_llseek,
2046 };
2047
2048 static void init_once_userfaultfd_ctx(void *mem)
2049 {
2050         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2051
2052         init_waitqueue_head(&ctx->fault_pending_wqh);
2053         init_waitqueue_head(&ctx->fault_wqh);
2054         init_waitqueue_head(&ctx->event_wqh);
2055         init_waitqueue_head(&ctx->fd_wqh);
2056         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2057 }
2058
2059 SYSCALL_DEFINE1(userfaultfd, int, flags)
2060 {
2061         struct userfaultfd_ctx *ctx;
2062         int fd;
2063
2064         if (!sysctl_unprivileged_userfaultfd &&
2065             (flags & UFFD_USER_MODE_ONLY) == 0 &&
2066             !capable(CAP_SYS_PTRACE)) {
2067                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2068                         "sysctl knob to 1 if kernel faults must be handled "
2069                         "without obtaining CAP_SYS_PTRACE capability\n");
2070                 return -EPERM;
2071         }
2072
2073         BUG_ON(!current->mm);
2074
2075         /* Check the UFFD_* constants for consistency.  */
2076         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2077         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2078         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2079
2080         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2081                 return -EINVAL;
2082
2083         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2084         if (!ctx)
2085                 return -ENOMEM;
2086
2087         refcount_set(&ctx->refcount, 1);
2088         ctx->flags = flags;
2089         ctx->features = 0;
2090         ctx->state = UFFD_STATE_WAIT_API;
2091         ctx->released = false;
2092         ctx->mmap_changing = false;
2093         ctx->mm = current->mm;
2094         /* prevent the mm struct to be freed */
2095         mmgrab(ctx->mm);
2096
2097         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2098                         O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2099         if (fd < 0) {
2100                 mmdrop(ctx->mm);
2101                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2102         }
2103         return fd;
2104 }
2105
2106 static int __init userfaultfd_init(void)
2107 {
2108         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2109                                                 sizeof(struct userfaultfd_ctx),
2110                                                 0,
2111                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2112                                                 init_once_userfaultfd_ctx);
2113         return 0;
2114 }
2115 __initcall(userfaultfd_init);