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