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/mm_inline.h>
  19 #include <linux/mmu_notifier.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32
  33 int sysctl_unprivileged_userfaultfd __read_mostly;
  34
  35 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  36
  37 /*
  38  * Start with fault_pending_wqh and fault_wqh so they're more likely
  39  * to be in the same cacheline.
  40  *
  41  * Locking order:
  42  *      fd_wqh.lock
  43  *              fault_pending_wqh.lock
  44  *                      fault_wqh.lock
  45  *              event_wqh.lock
  46  *
  47  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
  48  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
  49  * also taken in IRQ context.
  50  */
  51 struct userfaultfd_ctx {
  52         /* waitqueue head for the pending (i.e. not read) userfaults */
  53         wait_queue_head_t fault_pending_wqh;
  54         /* waitqueue head for the userfaults */
  55         wait_queue_head_t fault_wqh;
  56         /* waitqueue head for the pseudo fd to wakeup poll/read */
  57         wait_queue_head_t fd_wqh;
  58         /* waitqueue head for events */
  59         wait_queue_head_t event_wqh;
  60         /* a refile sequence protected by fault_pending_wqh lock */
  61         seqcount_spinlock_t refile_seq;
  62         /* pseudo fd refcounting */
  63         refcount_t refcount;
  64         /* userfaultfd syscall flags */
  65         unsigned int flags;
  66         /* features requested from the userspace */
  67         unsigned int features;
  68         /* released */
  69         bool released;
  70         /* memory mappings are changing because of non-cooperative event */
  71         atomic_t mmap_changing;
  72         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  73         struct mm_struct *mm;
  74 };
  75
  76 struct userfaultfd_fork_ctx {
  77         struct userfaultfd_ctx *orig;
  78         struct userfaultfd_ctx *new;
  79         struct list_head list;
  80 };
  81
  82 struct userfaultfd_unmap_ctx {
  83         struct userfaultfd_ctx *ctx;
  84         unsigned long start;
  85         unsigned long end;
  86         struct list_head list;
  87 };
  88
  89 struct userfaultfd_wait_queue {
  90         struct uffd_msg msg;
  91         wait_queue_entry_t wq;
  92         struct userfaultfd_ctx *ctx;
  93         bool waken;
  94 };
  95
  96 struct userfaultfd_wake_range {
  97         unsigned long start;
  98         unsigned long len;
  99 };
 100
 101 /* internal indication that UFFD_API ioctl was successfully executed */
 102 #define UFFD_FEATURE_INITIALIZED                (1u << 31)
 103
 104 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
 105 {
 106         return ctx->features & UFFD_FEATURE_INITIALIZED;
 107 }
 108
 109 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 110                                      int wake_flags, void *key)
 111 {
 112         struct userfaultfd_wake_range *range = key;
 113         int ret;
 114         struct userfaultfd_wait_queue *uwq;
 115         unsigned long start, len;
 116
 117         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 118         ret = 0;
 119         /* len == 0 means wake all */
 120         start = range->start;
 121         len = range->len;
 122         if (len && (start > uwq->msg.arg.pagefault.address ||
 123                     start + len <= uwq->msg.arg.pagefault.address))
 124                 goto out;
 125         WRITE_ONCE(uwq->waken, true);
 126         /*
 127          * The Program-Order guarantees provided by the scheduler
 128          * ensure uwq->waken is visible before the task is woken.
 129          */
 130         ret = wake_up_state(wq->private, mode);
 131         if (ret) {
 132                 /*
 133                  * Wake only once, autoremove behavior.
 134                  *
 135                  * After the effect of list_del_init is visible to the other
 136                  * CPUs, the waitqueue may disappear from under us, see the
 137                  * !list_empty_careful() in handle_userfault().
 138                  *
 139                  * try_to_wake_up() has an implicit smp_mb(), and the
 140                  * wq->private is read before calling the extern function
 141                  * "wake_up_state" (which in turns calls try_to_wake_up).
 142                  */
 143                 list_del_init(&wq->entry);
 144         }
 145 out:
 146         return ret;
 147 }
 148
 149 /**
 150  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 151  * context.
 152  * @ctx: [in] Pointer to the userfaultfd context.
 153  */
 154 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 155 {
 156         refcount_inc(&ctx->refcount);
 157 }
 158
 159 /**
 160  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 161  * context.
 162  * @ctx: [in] Pointer to userfaultfd context.
 163  *
 164  * The userfaultfd context reference must have been previously acquired either
 165  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 166  */
 167 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 168 {
 169         if (refcount_dec_and_test(&ctx->refcount)) {
 170                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 171                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 172                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 173                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 174                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 175                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 176                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 177                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 178                 mmdrop(ctx->mm);
 179                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 180         }
 181 }
 182
 183 static inline void msg_init(struct uffd_msg *msg)
 184 {
 185         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 186         /*
 187          * Must use memset to zero out the paddings or kernel data is
 188          * leaked to userland.
 189          */
 190         memset(msg, 0, sizeof(struct uffd_msg));
 191 }
 192
 193 static inline struct uffd_msg userfault_msg(unsigned long address,
 194                                             unsigned int flags,
 195                                             unsigned long reason,
 196                                             unsigned int features)
 197 {
 198         struct uffd_msg msg;
 199         msg_init(&msg);
 200         msg.event = UFFD_EVENT_PAGEFAULT;
 201         msg.arg.pagefault.address = address;
 202         /*
 203          * These flags indicate why the userfault occurred:
 204          * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
 205          * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
 206          * - Neither of these flags being set indicates a MISSING fault.
 207          *
 208          * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
 209          * fault. Otherwise, it was a read fault.
 210          */
 211         if (flags & FAULT_FLAG_WRITE)
 212                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 213         if (reason & VM_UFFD_WP)
 214                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 215         if (reason & VM_UFFD_MINOR)
 216                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
 217         if (features & UFFD_FEATURE_THREAD_ID)
 218                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 219         return msg;
 220 }
 221
 222 #ifdef CONFIG_HUGETLB_PAGE
 223 /*
 224  * Same functionality as userfaultfd_must_wait below with modifications for
 225  * hugepmd ranges.
 226  */
 227 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 228                                          struct vm_area_struct *vma,
 229                                          unsigned long address,
 230                                          unsigned long flags,
 231                                          unsigned long reason)
 232 {
 233         struct mm_struct *mm = ctx->mm;
 234         pte_t *ptep, pte;
 235         bool ret = true;
 236
 237         mmap_assert_locked(mm);
 238
 239         ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
 240
 241         if (!ptep)
 242                 goto out;
 243
 244         ret = false;
 245         pte = huge_ptep_get(ptep);
 246
 247         /*
 248          * Lockless access: we're in a wait_event so it's ok if it
 249          * changes under us.
 250          */
 251         if (huge_pte_none(pte))
 252                 ret = true;
 253         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 254                 ret = true;
 255 out:
 256         return ret;
 257 }
 258 #else
 259 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 260                                          struct vm_area_struct *vma,
 261                                          unsigned long address,
 262                                          unsigned long flags,
 263                                          unsigned long reason)
 264 {
 265         return false;   /* should never get here */
 266 }
 267 #endif /* CONFIG_HUGETLB_PAGE */
 268
 269 /*
 270  * Verify the pagetables are still not ok after having reigstered into
 271  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 272  * userfault that has already been resolved, if userfaultfd_read and
 273  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 274  * threads.
 275  */
 276 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 277                                          unsigned long address,
 278                                          unsigned long flags,
 279                                          unsigned long reason)
 280 {
 281         struct mm_struct *mm = ctx->mm;
 282         pgd_t *pgd;
 283         p4d_t *p4d;
 284         pud_t *pud;
 285         pmd_t *pmd, _pmd;
 286         pte_t *pte;
 287         bool ret = true;
 288
 289         mmap_assert_locked(mm);
 290
 291         pgd = pgd_offset(mm, address);
 292         if (!pgd_present(*pgd))
 293                 goto out;
 294         p4d = p4d_offset(pgd, address);
 295         if (!p4d_present(*p4d))
 296                 goto out;
 297         pud = pud_offset(p4d, address);
 298         if (!pud_present(*pud))
 299                 goto out;
 300         pmd = pmd_offset(pud, address);
 301         /*
 302          * READ_ONCE must function as a barrier with narrower scope
 303          * and it must be equivalent to:
 304          *      _pmd = *pmd; barrier();
 305          *
 306          * This is to deal with the instability (as in
 307          * pmd_trans_unstable) of the pmd.
 308          */
 309         _pmd = READ_ONCE(*pmd);
 310         if (pmd_none(_pmd))
 311                 goto out;
 312
 313         ret = false;
 314         if (!pmd_present(_pmd))
 315                 goto out;
 316
 317         if (pmd_trans_huge(_pmd)) {
 318                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 319                         ret = true;
 320                 goto out;
 321         }
 322
 323         /*
 324          * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
 325          * and use the standard pte_offset_map() instead of parsing _pmd.
 326          */
 327         pte = pte_offset_map(pmd, address);
 328         /*
 329          * Lockless access: we're in a wait_event so it's ok if it
 330          * changes under us.
 331          */
 332         if (pte_none(*pte))
 333                 ret = true;
 334         if (!pte_write(*pte) && (reason & VM_UFFD_WP))
 335                 ret = true;
 336         pte_unmap(pte);
 337
 338 out:
 339         return ret;
 340 }
 341
 342 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
 343 {
 344         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 345                 return TASK_INTERRUPTIBLE;
 346
 347         if (flags & FAULT_FLAG_KILLABLE)
 348                 return TASK_KILLABLE;
 349
 350         return TASK_UNINTERRUPTIBLE;
 351 }
 352
 353 /*
 354  * The locking rules involved in returning VM_FAULT_RETRY depending on
 355  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 356  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 357  * recommendation in __lock_page_or_retry is not an understatement.
 358  *
 359  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
 360  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 361  * not set.
 362  *
 363  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 364  * set, VM_FAULT_RETRY can still be returned if and only if there are
 365  * fatal_signal_pending()s, and the mmap_lock must be released before
 366  * returning it.
 367  */
 368 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 369 {
 370         struct mm_struct *mm = vmf->vma->vm_mm;
 371         struct userfaultfd_ctx *ctx;
 372         struct userfaultfd_wait_queue uwq;
 373         vm_fault_t ret = VM_FAULT_SIGBUS;
 374         bool must_wait;
 375         unsigned int blocking_state;
 376
 377         /*
 378          * We don't do userfault handling for the final child pid update.
 379          *
 380          * We also don't do userfault handling during
 381          * coredumping. hugetlbfs has the special
 382          * follow_hugetlb_page() to skip missing pages in the
 383          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 384          * the no_page_table() helper in follow_page_mask(), but the
 385          * shmem_vm_ops->fault method is invoked even during
 386          * coredumping without mmap_lock and it ends up here.
 387          */
 388         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 389                 goto out;
 390
 391         /*
 392          * Coredumping runs without mmap_lock so we can only check that
 393          * the mmap_lock is held, if PF_DUMPCORE was not set.
 394          */
 395         mmap_assert_locked(mm);
 396
 397         ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
 398         if (!ctx)
 399                 goto out;
 400
 401         BUG_ON(ctx->mm != mm);
 402
 403         /* Any unrecognized flag is a bug. */
 404         VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
 405         /* 0 or > 1 flags set is a bug; we expect exactly 1. */
 406         VM_BUG_ON(!reason || (reason & (reason - 1)));
 407
 408         if (ctx->features & UFFD_FEATURE_SIGBUS)
 409                 goto out;
 410         if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
 411             ctx->flags & UFFD_USER_MODE_ONLY) {
 412                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
 413                         "sysctl knob to 1 if kernel faults must be handled "
 414                         "without obtaining CAP_SYS_PTRACE capability\n");
 415                 goto out;
 416         }
 417
 418         /*
 419          * If it's already released don't get it. This avoids to loop
 420          * in __get_user_pages if userfaultfd_release waits on the
 421          * caller of handle_userfault to release the mmap_lock.
 422          */
 423         if (unlikely(READ_ONCE(ctx->released))) {
 424                 /*
 425                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 426                  * cooperative manager can close the uffd after the
 427                  * last UFFDIO_COPY, without risking to trigger an
 428                  * involuntary SIGBUS if the process was starting the
 429                  * userfaultfd while the userfaultfd was still armed
 430                  * (but after the last UFFDIO_COPY). If the uffd
 431                  * wasn't already closed when the userfault reached
 432                  * this point, that would normally be solved by
 433                  * userfaultfd_must_wait returning 'false'.
 434                  *
 435                  * If we were to return VM_FAULT_SIGBUS here, the non
 436                  * cooperative manager would be instead forced to
 437                  * always call UFFDIO_UNREGISTER before it can safely
 438                  * close the uffd.
 439                  */
 440                 ret = VM_FAULT_NOPAGE;
 441                 goto out;
 442         }
 443
 444         /*
 445          * Check that we can return VM_FAULT_RETRY.
 446          *
 447          * NOTE: it should become possible to return VM_FAULT_RETRY
 448          * even if FAULT_FLAG_TRIED is set without leading to gup()
 449          * -EBUSY failures, if the userfaultfd is to be extended for
 450          * VM_UFFD_WP tracking and we intend to arm the userfault
 451          * without first stopping userland access to the memory. For
 452          * VM_UFFD_MISSING userfaults this is enough for now.
 453          */
 454         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 455                 /*
 456                  * Validate the invariant that nowait must allow retry
 457                  * to be sure not to return SIGBUS erroneously on
 458                  * nowait invocations.
 459                  */
 460                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 461 #ifdef CONFIG_DEBUG_VM
 462                 if (printk_ratelimit()) {
 463                         printk(KERN_WARNING
 464                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 465                                vmf->flags);
 466                         dump_stack();
 467                 }
 468 #endif
 469                 goto out;
 470         }
 471
 472         /*
 473          * Handle nowait, not much to do other than tell it to retry
 474          * and wait.
 475          */
 476         ret = VM_FAULT_RETRY;
 477         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 478                 goto out;
 479
 480         /* take the reference before dropping the mmap_lock */
 481         userfaultfd_ctx_get(ctx);
 482
 483         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 484         uwq.wq.private = current;
 485         uwq.msg = userfault_msg(vmf->address, vmf->flags, reason,
 486                         ctx->features);
 487         uwq.ctx = ctx;
 488         uwq.waken = false;
 489
 490         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 491
 492         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 493         /*
 494          * After the __add_wait_queue the uwq is visible to userland
 495          * through poll/read().
 496          */
 497         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 498         /*
 499          * The smp_mb() after __set_current_state prevents the reads
 500          * following the spin_unlock to happen before the list_add in
 501          * __add_wait_queue.
 502          */
 503         set_current_state(blocking_state);
 504         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 505
 506         if (!is_vm_hugetlb_page(vmf->vma))
 507                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 508                                                   reason);
 509         else
 510                 must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
 511                                                        vmf->address,
 512                                                        vmf->flags, reason);
 513         mmap_read_unlock(mm);
 514
 515         if (likely(must_wait && !READ_ONCE(ctx->released))) {
 516                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 517                 schedule();
 518         }
 519
 520         __set_current_state(TASK_RUNNING);
 521
 522         /*
 523          * Here we race with the list_del; list_add in
 524          * userfaultfd_ctx_read(), however because we don't ever run
 525          * list_del_init() to refile across the two lists, the prev
 526          * and next pointers will never point to self. list_add also
 527          * would never let any of the two pointers to point to
 528          * self. So list_empty_careful won't risk to see both pointers
 529          * pointing to self at any time during the list refile. The
 530          * only case where list_del_init() is called is the full
 531          * removal in the wake function and there we don't re-list_add
 532          * and it's fine not to block on the spinlock. The uwq on this
 533          * kernel stack can be released after the list_del_init.
 534          */
 535         if (!list_empty_careful(&uwq.wq.entry)) {
 536                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 537                 /*
 538                  * No need of list_del_init(), the uwq on the stack
 539                  * will be freed shortly anyway.
 540                  */
 541                 list_del(&uwq.wq.entry);
 542                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 543         }
 544
 545         /*
 546          * ctx may go away after this if the userfault pseudo fd is
 547          * already released.
 548          */
 549         userfaultfd_ctx_put(ctx);
 550
 551 out:
 552         return ret;
 553 }
 554
 555 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 556                                               struct userfaultfd_wait_queue *ewq)
 557 {
 558         struct userfaultfd_ctx *release_new_ctx;
 559
 560         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 561                 goto out;
 562
 563         ewq->ctx = ctx;
 564         init_waitqueue_entry(&ewq->wq, current);
 565         release_new_ctx = NULL;
 566
 567         spin_lock_irq(&ctx->event_wqh.lock);
 568         /*
 569          * After the __add_wait_queue the uwq is visible to userland
 570          * through poll/read().
 571          */
 572         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 573         for (;;) {
 574                 set_current_state(TASK_KILLABLE);
 575                 if (ewq->msg.event == 0)
 576                         break;
 577                 if (READ_ONCE(ctx->released) ||
 578                     fatal_signal_pending(current)) {
 579                         /*
 580                          * &ewq->wq may be queued in fork_event, but
 581                          * __remove_wait_queue ignores the head
 582                          * parameter. It would be a problem if it
 583                          * didn't.
 584                          */
 585                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 586                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 587                                 struct userfaultfd_ctx *new;
 588
 589                                 new = (struct userfaultfd_ctx *)
 590                                         (unsigned long)
 591                                         ewq->msg.arg.reserved.reserved1;
 592                                 release_new_ctx = new;
 593                         }
 594                         break;
 595                 }
 596
 597                 spin_unlock_irq(&ctx->event_wqh.lock);
 598
 599                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 600                 schedule();
 601
 602                 spin_lock_irq(&ctx->event_wqh.lock);
 603         }
 604         __set_current_state(TASK_RUNNING);
 605         spin_unlock_irq(&ctx->event_wqh.lock);
 606
 607         if (release_new_ctx) {
 608                 struct vm_area_struct *vma;
 609                 struct mm_struct *mm = release_new_ctx->mm;
 610
 611                 /* the various vma->vm_userfaultfd_ctx still points to it */
 612                 mmap_write_lock(mm);
 613                 for (vma = mm->mmap; vma; vma = vma->vm_next)
 614                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 615                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 616                                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 617                         }
 618                 mmap_write_unlock(mm);
 619
 620                 userfaultfd_ctx_put(release_new_ctx);
 621         }
 622
 623         /*
 624          * ctx may go away after this if the userfault pseudo fd is
 625          * already released.
 626          */
 627 out:
 628         atomic_dec(&ctx->mmap_changing);
 629         VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
 630         userfaultfd_ctx_put(ctx);
 631 }
 632
 633 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 634                                        struct userfaultfd_wait_queue *ewq)
 635 {
 636         ewq->msg.event = 0;
 637         wake_up_locked(&ctx->event_wqh);
 638         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 639 }
 640
 641 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 642 {
 643         struct userfaultfd_ctx *ctx = NULL, *octx;
 644         struct userfaultfd_fork_ctx *fctx;
 645
 646         octx = vma->vm_userfaultfd_ctx.ctx;
 647         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 648                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 649                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 650                 return 0;
 651         }
 652
 653         list_for_each_entry(fctx, fcs, list)
 654                 if (fctx->orig == octx) {
 655                         ctx = fctx->new;
 656                         break;
 657                 }
 658
 659         if (!ctx) {
 660                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 661                 if (!fctx)
 662                         return -ENOMEM;
 663
 664                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 665                 if (!ctx) {
 666                         kfree(fctx);
 667                         return -ENOMEM;
 668                 }
 669
 670                 refcount_set(&ctx->refcount, 1);
 671                 ctx->flags = octx->flags;
 672                 ctx->features = octx->features;
 673                 ctx->released = false;
 674                 atomic_set(&ctx->mmap_changing, 0);
 675                 ctx->mm = vma->vm_mm;
 676                 mmgrab(ctx->mm);
 677
 678                 userfaultfd_ctx_get(octx);
 679                 atomic_inc(&octx->mmap_changing);
 680                 fctx->orig = octx;
 681                 fctx->new = ctx;
 682                 list_add_tail(&fctx->list, fcs);
 683         }
 684
 685         vma->vm_userfaultfd_ctx.ctx = ctx;
 686         return 0;
 687 }
 688
 689 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 690 {
 691         struct userfaultfd_ctx *ctx = fctx->orig;
 692         struct userfaultfd_wait_queue ewq;
 693
 694         msg_init(&ewq.msg);
 695
 696         ewq.msg.event = UFFD_EVENT_FORK;
 697         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 698
 699         userfaultfd_event_wait_completion(ctx, &ewq);
 700 }
 701
 702 void dup_userfaultfd_complete(struct list_head *fcs)
 703 {
 704         struct userfaultfd_fork_ctx *fctx, *n;
 705
 706         list_for_each_entry_safe(fctx, n, fcs, list) {
 707                 dup_fctx(fctx);
 708                 list_del(&fctx->list);
 709                 kfree(fctx);
 710         }
 711 }
 712
 713 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 714                              struct vm_userfaultfd_ctx *vm_ctx)
 715 {
 716         struct userfaultfd_ctx *ctx;
 717
 718         ctx = vma->vm_userfaultfd_ctx.ctx;
 719
 720         if (!ctx)
 721                 return;
 722
 723         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 724                 vm_ctx->ctx = ctx;
 725                 userfaultfd_ctx_get(ctx);
 726                 atomic_inc(&ctx->mmap_changing);
 727         } else {
 728                 /* Drop uffd context if remap feature not enabled */
 729                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 730                 vma->vm_flags &= ~__VM_UFFD_FLAGS;
 731         }
 732 }
 733
 734 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 735                                  unsigned long from, unsigned long to,
 736                                  unsigned long len)
 737 {
 738         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 739         struct userfaultfd_wait_queue ewq;
 740
 741         if (!ctx)
 742                 return;
 743
 744         if (to & ~PAGE_MASK) {
 745                 userfaultfd_ctx_put(ctx);
 746                 return;
 747         }
 748
 749         msg_init(&ewq.msg);
 750
 751         ewq.msg.event = UFFD_EVENT_REMAP;
 752         ewq.msg.arg.remap.from = from;
 753         ewq.msg.arg.remap.to = to;
 754         ewq.msg.arg.remap.len = len;
 755
 756         userfaultfd_event_wait_completion(ctx, &ewq);
 757 }
 758
 759 bool userfaultfd_remove(struct vm_area_struct *vma,
 760                         unsigned long start, unsigned long end)
 761 {
 762         struct mm_struct *mm = vma->vm_mm;
 763         struct userfaultfd_ctx *ctx;
 764         struct userfaultfd_wait_queue ewq;
 765
 766         ctx = vma->vm_userfaultfd_ctx.ctx;
 767         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 768                 return true;
 769
 770         userfaultfd_ctx_get(ctx);
 771         atomic_inc(&ctx->mmap_changing);
 772         mmap_read_unlock(mm);
 773
 774         msg_init(&ewq.msg);
 775
 776         ewq.msg.event = UFFD_EVENT_REMOVE;
 777         ewq.msg.arg.remove.start = start;
 778         ewq.msg.arg.remove.end = end;
 779
 780         userfaultfd_event_wait_completion(ctx, &ewq);
 781
 782         return false;
 783 }
 784
 785 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 786                           unsigned long start, unsigned long end)
 787 {
 788         struct userfaultfd_unmap_ctx *unmap_ctx;
 789
 790         list_for_each_entry(unmap_ctx, unmaps, list)
 791                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 792                     unmap_ctx->end == end)
 793                         return true;
 794
 795         return false;
 796 }
 797
 798 int userfaultfd_unmap_prep(struct vm_area_struct *vma,
 799                            unsigned long start, unsigned long end,
 800                            struct list_head *unmaps)
 801 {
 802         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
 803                 struct userfaultfd_unmap_ctx *unmap_ctx;
 804                 struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 805
 806                 if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 807                     has_unmap_ctx(ctx, unmaps, start, end))
 808                         continue;
 809
 810                 unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 811                 if (!unmap_ctx)
 812                         return -ENOMEM;
 813
 814                 userfaultfd_ctx_get(ctx);
 815                 atomic_inc(&ctx->mmap_changing);
 816                 unmap_ctx->ctx = ctx;
 817                 unmap_ctx->start = start;
 818                 unmap_ctx->end = end;
 819                 list_add_tail(&unmap_ctx->list, unmaps);
 820         }
 821
 822         return 0;
 823 }
 824
 825 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 826 {
 827         struct userfaultfd_unmap_ctx *ctx, *n;
 828         struct userfaultfd_wait_queue ewq;
 829
 830         list_for_each_entry_safe(ctx, n, uf, list) {
 831                 msg_init(&ewq.msg);
 832
 833                 ewq.msg.event = UFFD_EVENT_UNMAP;
 834                 ewq.msg.arg.remove.start = ctx->start;
 835                 ewq.msg.arg.remove.end = ctx->end;
 836
 837                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 838
 839                 list_del(&ctx->list);
 840                 kfree(ctx);
 841         }
 842 }
 843
 844 static int userfaultfd_release(struct inode *inode, struct file *file)
 845 {
 846         struct userfaultfd_ctx *ctx = file->private_data;
 847         struct mm_struct *mm = ctx->mm;
 848         struct vm_area_struct *vma, *prev;
 849         /* len == 0 means wake all */
 850         struct userfaultfd_wake_range range = { .len = 0, };
 851         unsigned long new_flags;
 852
 853         WRITE_ONCE(ctx->released, true);
 854
 855         if (!mmget_not_zero(mm))
 856                 goto wakeup;
 857
 858         /*
 859          * Flush page faults out of all CPUs. NOTE: all page faults
 860          * must be retried without returning VM_FAULT_SIGBUS if
 861          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 862          * changes while handle_userfault released the mmap_lock. So
 863          * it's critical that released is set to true (above), before
 864          * taking the mmap_lock for writing.
 865          */
 866         mmap_write_lock(mm);
 867         prev = NULL;
 868         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 869                 cond_resched();
 870                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 871                        !!(vma->vm_flags & __VM_UFFD_FLAGS));
 872                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 873                         prev = vma;
 874                         continue;
 875                 }
 876                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
 877                 prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
 878                                  new_flags, vma->anon_vma,
 879                                  vma->vm_file, vma->vm_pgoff,
 880                                  vma_policy(vma),
 881                                  NULL_VM_UFFD_CTX, anon_vma_name(vma));
 882                 if (prev)
 883                         vma = prev;
 884                 else
 885                         prev = vma;
 886                 vma->vm_flags = new_flags;
 887                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 888         }
 889         mmap_write_unlock(mm);
 890         mmput(mm);
 891 wakeup:
 892         /*
 893          * After no new page faults can wait on this fault_*wqh, flush
 894          * the last page faults that may have been already waiting on
 895          * the fault_*wqh.
 896          */
 897         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 898         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 899         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 900         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 901
 902         /* Flush pending events that may still wait on event_wqh */
 903         wake_up_all(&ctx->event_wqh);
 904
 905         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 906         userfaultfd_ctx_put(ctx);
 907         return 0;
 908 }
 909
 910 /* fault_pending_wqh.lock must be hold by the caller */
 911 static inline struct userfaultfd_wait_queue *find_userfault_in(
 912                 wait_queue_head_t *wqh)
 913 {
 914         wait_queue_entry_t *wq;
 915         struct userfaultfd_wait_queue *uwq;
 916
 917         lockdep_assert_held(&wqh->lock);
 918
 919         uwq = NULL;
 920         if (!waitqueue_active(wqh))
 921                 goto out;
 922         /* walk in reverse to provide FIFO behavior to read userfaults */
 923         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 924         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 925 out:
 926         return uwq;
 927 }
 928
 929 static inline struct userfaultfd_wait_queue *find_userfault(
 930                 struct userfaultfd_ctx *ctx)
 931 {
 932         return find_userfault_in(&ctx->fault_pending_wqh);
 933 }
 934
 935 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 936                 struct userfaultfd_ctx *ctx)
 937 {
 938         return find_userfault_in(&ctx->event_wqh);
 939 }
 940
 941 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 942 {
 943         struct userfaultfd_ctx *ctx = file->private_data;
 944         __poll_t ret;
 945
 946         poll_wait(file, &ctx->fd_wqh, wait);
 947
 948         if (!userfaultfd_is_initialized(ctx))
 949                 return EPOLLERR;
 950
 951         /*
 952          * poll() never guarantees that read won't block.
 953          * userfaults can be waken before they're read().
 954          */
 955         if (unlikely(!(file->f_flags & O_NONBLOCK)))
 956                 return EPOLLERR;
 957         /*
 958          * lockless access to see if there are pending faults
 959          * __pollwait last action is the add_wait_queue but
 960          * the spin_unlock would allow the waitqueue_active to
 961          * pass above the actual list_add inside
 962          * add_wait_queue critical section. So use a full
 963          * memory barrier to serialize the list_add write of
 964          * add_wait_queue() with the waitqueue_active read
 965          * below.
 966          */
 967         ret = 0;
 968         smp_mb();
 969         if (waitqueue_active(&ctx->fault_pending_wqh))
 970                 ret = EPOLLIN;
 971         else if (waitqueue_active(&ctx->event_wqh))
 972                 ret = EPOLLIN;
 973
 974         return ret;
 975 }
 976
 977 static const struct file_operations userfaultfd_fops;
 978
 979 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
 980                                   struct inode *inode,
 981                                   struct uffd_msg *msg)
 982 {
 983         int fd;
 984
 985         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
 986                         O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
 987         if (fd < 0)
 988                 return fd;
 989
 990         msg->arg.reserved.reserved1 = 0;
 991         msg->arg.fork.ufd = fd;
 992         return 0;
 993 }
 994
 995 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 996                                     struct uffd_msg *msg, struct inode *inode)
 997 {
 998         ssize_t ret;
 999         DECLARE_WAITQUEUE(wait, current);
1000         struct userfaultfd_wait_queue *uwq;
1001         /*
1002          * Handling fork event requires sleeping operations, so
1003          * we drop the event_wqh lock, then do these ops, then
1004          * lock it back and wake up the waiter. While the lock is
1005          * dropped the ewq may go away so we keep track of it
1006          * carefully.
1007          */
1008         LIST_HEAD(fork_event);
1009         struct userfaultfd_ctx *fork_nctx = NULL;
1010
1011         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1012         spin_lock_irq(&ctx->fd_wqh.lock);
1013         __add_wait_queue(&ctx->fd_wqh, &wait);
1014         for (;;) {
1015                 set_current_state(TASK_INTERRUPTIBLE);
1016                 spin_lock(&ctx->fault_pending_wqh.lock);
1017                 uwq = find_userfault(ctx);
1018                 if (uwq) {
1019                         /*
1020                          * Use a seqcount to repeat the lockless check
1021                          * in wake_userfault() to avoid missing
1022                          * wakeups because during the refile both
1023                          * waitqueue could become empty if this is the
1024                          * only userfault.
1025                          */
1026                         write_seqcount_begin(&ctx->refile_seq);
1027
1028                         /*
1029                          * The fault_pending_wqh.lock prevents the uwq
1030                          * to disappear from under us.
1031                          *
1032                          * Refile this userfault from
1033                          * fault_pending_wqh to fault_wqh, it's not
1034                          * pending anymore after we read it.
1035                          *
1036                          * Use list_del() by hand (as
1037                          * userfaultfd_wake_function also uses
1038                          * list_del_init() by hand) to be sure nobody
1039                          * changes __remove_wait_queue() to use
1040                          * list_del_init() in turn breaking the
1041                          * !list_empty_careful() check in
1042                          * handle_userfault(). The uwq->wq.head list
1043                          * must never be empty at any time during the
1044                          * refile, or the waitqueue could disappear
1045                          * from under us. The "wait_queue_head_t"
1046                          * parameter of __remove_wait_queue() is unused
1047                          * anyway.
1048                          */
1049                         list_del(&uwq->wq.entry);
1050                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1051
1052                         write_seqcount_end(&ctx->refile_seq);
1053
1054                         /* careful to always initialize msg if ret == 0 */
1055                         *msg = uwq->msg;
1056                         spin_unlock(&ctx->fault_pending_wqh.lock);
1057                         ret = 0;
1058                         break;
1059                 }
1060                 spin_unlock(&ctx->fault_pending_wqh.lock);
1061
1062                 spin_lock(&ctx->event_wqh.lock);
1063                 uwq = find_userfault_evt(ctx);
1064                 if (uwq) {
1065                         *msg = uwq->msg;
1066
1067                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1068                                 fork_nctx = (struct userfaultfd_ctx *)
1069                                         (unsigned long)
1070                                         uwq->msg.arg.reserved.reserved1;
1071                                 list_move(&uwq->wq.entry, &fork_event);
1072                                 /*
1073                                  * fork_nctx can be freed as soon as
1074                                  * we drop the lock, unless we take a
1075                                  * reference on it.
1076                                  */
1077                                 userfaultfd_ctx_get(fork_nctx);
1078                                 spin_unlock(&ctx->event_wqh.lock);
1079                                 ret = 0;
1080                                 break;
1081                         }
1082
1083                         userfaultfd_event_complete(ctx, uwq);
1084                         spin_unlock(&ctx->event_wqh.lock);
1085                         ret = 0;
1086                         break;
1087                 }
1088                 spin_unlock(&ctx->event_wqh.lock);
1089
1090                 if (signal_pending(current)) {
1091                         ret = -ERESTARTSYS;
1092                         break;
1093                 }
1094                 if (no_wait) {
1095                         ret = -EAGAIN;
1096                         break;
1097                 }
1098                 spin_unlock_irq(&ctx->fd_wqh.lock);
1099                 schedule();
1100                 spin_lock_irq(&ctx->fd_wqh.lock);
1101         }
1102         __remove_wait_queue(&ctx->fd_wqh, &wait);
1103         __set_current_state(TASK_RUNNING);
1104         spin_unlock_irq(&ctx->fd_wqh.lock);
1105
1106         if (!ret && msg->event == UFFD_EVENT_FORK) {
1107                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1108                 spin_lock_irq(&ctx->event_wqh.lock);
1109                 if (!list_empty(&fork_event)) {
1110                         /*
1111                          * The fork thread didn't abort, so we can
1112                          * drop the temporary refcount.
1113                          */
1114                         userfaultfd_ctx_put(fork_nctx);
1115
1116                         uwq = list_first_entry(&fork_event,
1117                                                typeof(*uwq),
1118                                                wq.entry);
1119                         /*
1120                          * If fork_event list wasn't empty and in turn
1121                          * the event wasn't already released by fork
1122                          * (the event is allocated on fork kernel
1123                          * stack), put the event back to its place in
1124                          * the event_wq. fork_event head will be freed
1125                          * as soon as we return so the event cannot
1126                          * stay queued there no matter the current
1127                          * "ret" value.
1128                          */
1129                         list_del(&uwq->wq.entry);
1130                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1131
1132                         /*
1133                          * Leave the event in the waitqueue and report
1134                          * error to userland if we failed to resolve
1135                          * the userfault fork.
1136                          */
1137                         if (likely(!ret))
1138                                 userfaultfd_event_complete(ctx, uwq);
1139                 } else {
1140                         /*
1141                          * Here the fork thread aborted and the
1142                          * refcount from the fork thread on fork_nctx
1143                          * has already been released. We still hold
1144                          * the reference we took before releasing the
1145                          * lock above. If resolve_userfault_fork
1146                          * failed we've to drop it because the
1147                          * fork_nctx has to be freed in such case. If
1148                          * it succeeded we'll hold it because the new
1149                          * uffd references it.
1150                          */
1151                         if (ret)
1152                                 userfaultfd_ctx_put(fork_nctx);
1153                 }
1154                 spin_unlock_irq(&ctx->event_wqh.lock);
1155         }
1156
1157         return ret;
1158 }
1159
1160 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1161                                 size_t count, loff_t *ppos)
1162 {
1163         struct userfaultfd_ctx *ctx = file->private_data;
1164         ssize_t _ret, ret = 0;
1165         struct uffd_msg msg;
1166         int no_wait = file->f_flags & O_NONBLOCK;
1167         struct inode *inode = file_inode(file);
1168
1169         if (!userfaultfd_is_initialized(ctx))
1170                 return -EINVAL;
1171
1172         for (;;) {
1173                 if (count < sizeof(msg))
1174                         return ret ? ret : -EINVAL;
1175                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1176                 if (_ret < 0)
1177                         return ret ? ret : _ret;
1178                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1179                         return ret ? ret : -EFAULT;
1180                 ret += sizeof(msg);
1181                 buf += sizeof(msg);
1182                 count -= sizeof(msg);
1183                 /*
1184                  * Allow to read more than one fault at time but only
1185                  * block if waiting for the very first one.
1186                  */
1187                 no_wait = O_NONBLOCK;
1188         }
1189 }
1190
1191 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1192                              struct userfaultfd_wake_range *range)
1193 {
1194         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1195         /* wake all in the range and autoremove */
1196         if (waitqueue_active(&ctx->fault_pending_wqh))
1197                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1198                                      range);
1199         if (waitqueue_active(&ctx->fault_wqh))
1200                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1201         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1202 }
1203
1204 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1205                                            struct userfaultfd_wake_range *range)
1206 {
1207         unsigned seq;
1208         bool need_wakeup;
1209
1210         /*
1211          * To be sure waitqueue_active() is not reordered by the CPU
1212          * before the pagetable update, use an explicit SMP memory
1213          * barrier here. PT lock release or mmap_read_unlock(mm) still
1214          * have release semantics that can allow the
1215          * waitqueue_active() to be reordered before the pte update.
1216          */
1217         smp_mb();
1218
1219         /*
1220          * Use waitqueue_active because it's very frequent to
1221          * change the address space atomically even if there are no
1222          * userfaults yet. So we take the spinlock only when we're
1223          * sure we've userfaults to wake.
1224          */
1225         do {
1226                 seq = read_seqcount_begin(&ctx->refile_seq);
1227                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1228                         waitqueue_active(&ctx->fault_wqh);
1229                 cond_resched();
1230         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1231         if (need_wakeup)
1232                 __wake_userfault(ctx, range);
1233 }
1234
1235 static __always_inline int validate_range(struct mm_struct *mm,
1236                                           __u64 start, __u64 len)
1237 {
1238         __u64 task_size = mm->task_size;
1239
1240         if (start & ~PAGE_MASK)
1241                 return -EINVAL;
1242         if (len & ~PAGE_MASK)
1243                 return -EINVAL;
1244         if (!len)
1245                 return -EINVAL;
1246         if (start < mmap_min_addr)
1247                 return -EINVAL;
1248         if (start >= task_size)
1249                 return -EINVAL;
1250         if (len > task_size - start)
1251                 return -EINVAL;
1252         return 0;
1253 }
1254
1255 static inline bool vma_can_userfault(struct vm_area_struct *vma,
1256                                      unsigned long vm_flags)
1257 {
1258         /* FIXME: add WP support to hugetlbfs and shmem */
1259         if (vm_flags & VM_UFFD_WP) {
1260                 if (is_vm_hugetlb_page(vma) || vma_is_shmem(vma))
1261                         return false;
1262         }
1263
1264         if (vm_flags & VM_UFFD_MINOR) {
1265                 if (!(is_vm_hugetlb_page(vma) || vma_is_shmem(vma)))
1266                         return false;
1267         }
1268
1269         return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1270                vma_is_shmem(vma);
1271 }
1272
1273 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1274                                 unsigned long arg)
1275 {
1276         struct mm_struct *mm = ctx->mm;
1277         struct vm_area_struct *vma, *prev, *cur;
1278         int ret;
1279         struct uffdio_register uffdio_register;
1280         struct uffdio_register __user *user_uffdio_register;
1281         unsigned long vm_flags, new_flags;
1282         bool found;
1283         bool basic_ioctls;
1284         unsigned long start, end, vma_end;
1285
1286         user_uffdio_register = (struct uffdio_register __user *) arg;
1287
1288         ret = -EFAULT;
1289         if (copy_from_user(&uffdio_register, user_uffdio_register,
1290                            sizeof(uffdio_register)-sizeof(__u64)))
1291                 goto out;
1292
1293         ret = -EINVAL;
1294         if (!uffdio_register.mode)
1295                 goto out;
1296         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1297                 goto out;
1298         vm_flags = 0;
1299         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1300                 vm_flags |= VM_UFFD_MISSING;
1301         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1302 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1303                 goto out;
1304 #endif
1305                 vm_flags |= VM_UFFD_WP;
1306         }
1307         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1308 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1309                 goto out;
1310 #endif
1311                 vm_flags |= VM_UFFD_MINOR;
1312         }
1313
1314         ret = validate_range(mm, uffdio_register.range.start,
1315                              uffdio_register.range.len);
1316         if (ret)
1317                 goto out;
1318
1319         start = uffdio_register.range.start;
1320         end = start + uffdio_register.range.len;
1321
1322         ret = -ENOMEM;
1323         if (!mmget_not_zero(mm))
1324                 goto out;
1325
1326         mmap_write_lock(mm);
1327         vma = find_vma_prev(mm, start, &prev);
1328         if (!vma)
1329                 goto out_unlock;
1330
1331         /* check that there's at least one vma in the range */
1332         ret = -EINVAL;
1333         if (vma->vm_start >= end)
1334                 goto out_unlock;
1335
1336         /*
1337          * If the first vma contains huge pages, make sure start address
1338          * is aligned to huge page size.
1339          */
1340         if (is_vm_hugetlb_page(vma)) {
1341                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1342
1343                 if (start & (vma_hpagesize - 1))
1344                         goto out_unlock;
1345         }
1346
1347         /*
1348          * Search for not compatible vmas.
1349          */
1350         found = false;
1351         basic_ioctls = false;
1352         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1353                 cond_resched();
1354
1355                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1356                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1357
1358                 /* check not compatible vmas */
1359                 ret = -EINVAL;
1360                 if (!vma_can_userfault(cur, vm_flags))
1361                         goto out_unlock;
1362
1363                 /*
1364                  * UFFDIO_COPY will fill file holes even without
1365                  * PROT_WRITE. This check enforces that if this is a
1366                  * MAP_SHARED, the process has write permission to the backing
1367                  * file. If VM_MAYWRITE is set it also enforces that on a
1368                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1369                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1370                  */
1371                 ret = -EPERM;
1372                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1373                         goto out_unlock;
1374
1375                 /*
1376                  * If this vma contains ending address, and huge pages
1377                  * check alignment.
1378                  */
1379                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1380                     end > cur->vm_start) {
1381                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1382
1383                         ret = -EINVAL;
1384
1385                         if (end & (vma_hpagesize - 1))
1386                                 goto out_unlock;
1387                 }
1388                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1389                         goto out_unlock;
1390
1391                 /*
1392                  * Check that this vma isn't already owned by a
1393                  * different userfaultfd. We can't allow more than one
1394                  * userfaultfd to own a single vma simultaneously or we
1395                  * wouldn't know which one to deliver the userfaults to.
1396                  */
1397                 ret = -EBUSY;
1398                 if (cur->vm_userfaultfd_ctx.ctx &&
1399                     cur->vm_userfaultfd_ctx.ctx != ctx)
1400                         goto out_unlock;
1401
1402                 /*
1403                  * Note vmas containing huge pages
1404                  */
1405                 if (is_vm_hugetlb_page(cur))
1406                         basic_ioctls = true;
1407
1408                 found = true;
1409         }
1410         BUG_ON(!found);
1411
1412         if (vma->vm_start < start)
1413                 prev = vma;
1414
1415         ret = 0;
1416         do {
1417                 cond_resched();
1418
1419                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1420                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1421                        vma->vm_userfaultfd_ctx.ctx != ctx);
1422                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1423
1424                 /*
1425                  * Nothing to do: this vma is already registered into this
1426                  * userfaultfd and with the right tracking mode too.
1427                  */
1428                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1429                     (vma->vm_flags & vm_flags) == vm_flags)
1430                         goto skip;
1431
1432                 if (vma->vm_start > start)
1433                         start = vma->vm_start;
1434                 vma_end = min(end, vma->vm_end);
1435
1436                 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1437                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1438                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1439                                  vma_policy(vma),
1440                                  ((struct vm_userfaultfd_ctx){ ctx }),
1441                                  anon_vma_name(vma));
1442                 if (prev) {
1443                         vma = prev;
1444                         goto next;
1445                 }
1446                 if (vma->vm_start < start) {
1447                         ret = split_vma(mm, vma, start, 1);
1448                         if (ret)
1449                                 break;
1450                 }
1451                 if (vma->vm_end > end) {
1452                         ret = split_vma(mm, vma, end, 0);
1453                         if (ret)
1454                                 break;
1455                 }
1456         next:
1457                 /*
1458                  * In the vma_merge() successful mprotect-like case 8:
1459                  * the next vma was merged into the current one and
1460                  * the current one has not been updated yet.
1461                  */
1462                 vma->vm_flags = new_flags;
1463                 vma->vm_userfaultfd_ctx.ctx = ctx;
1464
1465                 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1466                         hugetlb_unshare_all_pmds(vma);
1467
1468         skip:
1469                 prev = vma;
1470                 start = vma->vm_end;
1471                 vma = vma->vm_next;
1472         } while (vma && vma->vm_start < end);
1473 out_unlock:
1474         mmap_write_unlock(mm);
1475         mmput(mm);
1476         if (!ret) {
1477                 __u64 ioctls_out;
1478
1479                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1480                     UFFD_API_RANGE_IOCTLS;
1481
1482                 /*
1483                  * Declare the WP ioctl only if the WP mode is
1484                  * specified and all checks passed with the range
1485                  */
1486                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1487                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1488
1489                 /* CONTINUE ioctl is only supported for MINOR ranges. */
1490                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1491                         ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1492
1493                 /*
1494                  * Now that we scanned all vmas we can already tell
1495                  * userland which ioctls methods are guaranteed to
1496                  * succeed on this range.
1497                  */
1498                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1499                         ret = -EFAULT;
1500         }
1501 out:
1502         return ret;
1503 }
1504
1505 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1506                                   unsigned long arg)
1507 {
1508         struct mm_struct *mm = ctx->mm;
1509         struct vm_area_struct *vma, *prev, *cur;
1510         int ret;
1511         struct uffdio_range uffdio_unregister;
1512         unsigned long new_flags;
1513         bool found;
1514         unsigned long start, end, vma_end;
1515         const void __user *buf = (void __user *)arg;
1516
1517         ret = -EFAULT;
1518         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1519                 goto out;
1520
1521         ret = validate_range(mm, uffdio_unregister.start,
1522                              uffdio_unregister.len);
1523         if (ret)
1524                 goto out;
1525
1526         start = uffdio_unregister.start;
1527         end = start + uffdio_unregister.len;
1528
1529         ret = -ENOMEM;
1530         if (!mmget_not_zero(mm))
1531                 goto out;
1532
1533         mmap_write_lock(mm);
1534         vma = find_vma_prev(mm, start, &prev);
1535         if (!vma)
1536                 goto out_unlock;
1537
1538         /* check that there's at least one vma in the range */
1539         ret = -EINVAL;
1540         if (vma->vm_start >= end)
1541                 goto out_unlock;
1542
1543         /*
1544          * If the first vma contains huge pages, make sure start address
1545          * is aligned to huge page size.
1546          */
1547         if (is_vm_hugetlb_page(vma)) {
1548                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1549
1550                 if (start & (vma_hpagesize - 1))
1551                         goto out_unlock;
1552         }
1553
1554         /*
1555          * Search for not compatible vmas.
1556          */
1557         found = false;
1558         ret = -EINVAL;
1559         for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) {
1560                 cond_resched();
1561
1562                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1563                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1564
1565                 /*
1566                  * Check not compatible vmas, not strictly required
1567                  * here as not compatible vmas cannot have an
1568                  * userfaultfd_ctx registered on them, but this
1569                  * provides for more strict behavior to notice
1570                  * unregistration errors.
1571                  */
1572                 if (!vma_can_userfault(cur, cur->vm_flags))
1573                         goto out_unlock;
1574
1575                 found = true;
1576         }
1577         BUG_ON(!found);
1578
1579         if (vma->vm_start < start)
1580                 prev = vma;
1581
1582         ret = 0;
1583         do {
1584                 cond_resched();
1585
1586                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1587
1588                 /*
1589                  * Nothing to do: this vma is already registered into this
1590                  * userfaultfd and with the right tracking mode too.
1591                  */
1592                 if (!vma->vm_userfaultfd_ctx.ctx)
1593                         goto skip;
1594
1595                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1596
1597                 if (vma->vm_start > start)
1598                         start = vma->vm_start;
1599                 vma_end = min(end, vma->vm_end);
1600
1601                 if (userfaultfd_missing(vma)) {
1602                         /*
1603                          * Wake any concurrent pending userfault while
1604                          * we unregister, so they will not hang
1605                          * permanently and it avoids userland to call
1606                          * UFFDIO_WAKE explicitly.
1607                          */
1608                         struct userfaultfd_wake_range range;
1609                         range.start = start;
1610                         range.len = vma_end - start;
1611                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1612                 }
1613
1614                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1615                 prev = vma_merge(mm, prev, start, vma_end, new_flags,
1616                                  vma->anon_vma, vma->vm_file, vma->vm_pgoff,
1617                                  vma_policy(vma),
1618                                  NULL_VM_UFFD_CTX, anon_vma_name(vma));
1619                 if (prev) {
1620                         vma = prev;
1621                         goto next;
1622                 }
1623                 if (vma->vm_start < start) {
1624                         ret = split_vma(mm, vma, start, 1);
1625                         if (ret)
1626                                 break;
1627                 }
1628                 if (vma->vm_end > end) {
1629                         ret = split_vma(mm, vma, end, 0);
1630                         if (ret)
1631                                 break;
1632                 }
1633         next:
1634                 /*
1635                  * In the vma_merge() successful mprotect-like case 8:
1636                  * the next vma was merged into the current one and
1637                  * the current one has not been updated yet.
1638                  */
1639                 vma->vm_flags = new_flags;
1640                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1641
1642         skip:
1643                 prev = vma;
1644                 start = vma->vm_end;
1645                 vma = vma->vm_next;
1646         } while (vma && vma->vm_start < end);
1647 out_unlock:
1648         mmap_write_unlock(mm);
1649         mmput(mm);
1650 out:
1651         return ret;
1652 }
1653
1654 /*
1655  * userfaultfd_wake may be used in combination with the
1656  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1657  */
1658 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1659                             unsigned long arg)
1660 {
1661         int ret;
1662         struct uffdio_range uffdio_wake;
1663         struct userfaultfd_wake_range range;
1664         const void __user *buf = (void __user *)arg;
1665
1666         ret = -EFAULT;
1667         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1668                 goto out;
1669
1670         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1671         if (ret)
1672                 goto out;
1673
1674         range.start = uffdio_wake.start;
1675         range.len = uffdio_wake.len;
1676
1677         /*
1678          * len == 0 means wake all and we don't want to wake all here,
1679          * so check it again to be sure.
1680          */
1681         VM_BUG_ON(!range.len);
1682
1683         wake_userfault(ctx, &range);
1684         ret = 0;
1685
1686 out:
1687         return ret;
1688 }
1689
1690 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1691                             unsigned long arg)
1692 {
1693         __s64 ret;
1694         struct uffdio_copy uffdio_copy;
1695         struct uffdio_copy __user *user_uffdio_copy;
1696         struct userfaultfd_wake_range range;
1697
1698         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1699
1700         ret = -EAGAIN;
1701         if (atomic_read(&ctx->mmap_changing))
1702                 goto out;
1703
1704         ret = -EFAULT;
1705         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1706                            /* don't copy "copy" last field */
1707                            sizeof(uffdio_copy)-sizeof(__s64)))
1708                 goto out;
1709
1710         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1711         if (ret)
1712                 goto out;
1713         /*
1714          * double check for wraparound just in case. copy_from_user()
1715          * will later check uffdio_copy.src + uffdio_copy.len to fit
1716          * in the userland range.
1717          */
1718         ret = -EINVAL;
1719         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1720                 goto out;
1721         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1722                 goto out;
1723         if (mmget_not_zero(ctx->mm)) {
1724                 ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1725                                    uffdio_copy.len, &ctx->mmap_changing,
1726                                    uffdio_copy.mode);
1727                 mmput(ctx->mm);
1728         } else {
1729                 return -ESRCH;
1730         }
1731         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1732                 return -EFAULT;
1733         if (ret < 0)
1734                 goto out;
1735         BUG_ON(!ret);
1736         /* len == 0 would wake all */
1737         range.len = ret;
1738         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1739                 range.start = uffdio_copy.dst;
1740                 wake_userfault(ctx, &range);
1741         }
1742         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1743 out:
1744         return ret;
1745 }
1746
1747 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1748                                 unsigned long arg)
1749 {
1750         __s64 ret;
1751         struct uffdio_zeropage uffdio_zeropage;
1752         struct uffdio_zeropage __user *user_uffdio_zeropage;
1753         struct userfaultfd_wake_range range;
1754
1755         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1756
1757         ret = -EAGAIN;
1758         if (atomic_read(&ctx->mmap_changing))
1759                 goto out;
1760
1761         ret = -EFAULT;
1762         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1763                            /* don't copy "zeropage" last field */
1764                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1765                 goto out;
1766
1767         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1768                              uffdio_zeropage.range.len);
1769         if (ret)
1770                 goto out;
1771         ret = -EINVAL;
1772         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1773                 goto out;
1774
1775         if (mmget_not_zero(ctx->mm)) {
1776                 ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start,
1777                                      uffdio_zeropage.range.len,
1778                                      &ctx->mmap_changing);
1779                 mmput(ctx->mm);
1780         } else {
1781                 return -ESRCH;
1782         }
1783         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1784                 return -EFAULT;
1785         if (ret < 0)
1786                 goto out;
1787         /* len == 0 would wake all */
1788         BUG_ON(!ret);
1789         range.len = ret;
1790         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1791                 range.start = uffdio_zeropage.range.start;
1792                 wake_userfault(ctx, &range);
1793         }
1794         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1795 out:
1796         return ret;
1797 }
1798
1799 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1800                                     unsigned long arg)
1801 {
1802         int ret;
1803         struct uffdio_writeprotect uffdio_wp;
1804         struct uffdio_writeprotect __user *user_uffdio_wp;
1805         struct userfaultfd_wake_range range;
1806         bool mode_wp, mode_dontwake;
1807
1808         if (atomic_read(&ctx->mmap_changing))
1809                 return -EAGAIN;
1810
1811         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1812
1813         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1814                            sizeof(struct uffdio_writeprotect)))
1815                 return -EFAULT;
1816
1817         ret = validate_range(ctx->mm, uffdio_wp.range.start,
1818                              uffdio_wp.range.len);
1819         if (ret)
1820                 return ret;
1821
1822         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1823                                UFFDIO_WRITEPROTECT_MODE_WP))
1824                 return -EINVAL;
1825
1826         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1827         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1828
1829         if (mode_wp && mode_dontwake)
1830                 return -EINVAL;
1831
1832         if (mmget_not_zero(ctx->mm)) {
1833                 ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1834                                           uffdio_wp.range.len, mode_wp,
1835                                           &ctx->mmap_changing);
1836                 mmput(ctx->mm);
1837         } else {
1838                 return -ESRCH;
1839         }
1840
1841         if (ret)
1842                 return ret;
1843
1844         if (!mode_wp && !mode_dontwake) {
1845                 range.start = uffdio_wp.range.start;
1846                 range.len = uffdio_wp.range.len;
1847                 wake_userfault(ctx, &range);
1848         }
1849         return ret;
1850 }
1851
1852 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1853 {
1854         __s64 ret;
1855         struct uffdio_continue uffdio_continue;
1856         struct uffdio_continue __user *user_uffdio_continue;
1857         struct userfaultfd_wake_range range;
1858
1859         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1860
1861         ret = -EAGAIN;
1862         if (atomic_read(&ctx->mmap_changing))
1863                 goto out;
1864
1865         ret = -EFAULT;
1866         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1867                            /* don't copy the output fields */
1868                            sizeof(uffdio_continue) - (sizeof(__s64))))
1869                 goto out;
1870
1871         ret = validate_range(ctx->mm, uffdio_continue.range.start,
1872                              uffdio_continue.range.len);
1873         if (ret)
1874                 goto out;
1875
1876         ret = -EINVAL;
1877         /* double check for wraparound just in case. */
1878         if (uffdio_continue.range.start + uffdio_continue.range.len <=
1879             uffdio_continue.range.start) {
1880                 goto out;
1881         }
1882         if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1883                 goto out;
1884
1885         if (mmget_not_zero(ctx->mm)) {
1886                 ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1887                                      uffdio_continue.range.len,
1888                                      &ctx->mmap_changing);
1889                 mmput(ctx->mm);
1890         } else {
1891                 return -ESRCH;
1892         }
1893
1894         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1895                 return -EFAULT;
1896         if (ret < 0)
1897                 goto out;
1898
1899         /* len == 0 would wake all */
1900         BUG_ON(!ret);
1901         range.len = ret;
1902         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1903                 range.start = uffdio_continue.range.start;
1904                 wake_userfault(ctx, &range);
1905         }
1906         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1907
1908 out:
1909         return ret;
1910 }
1911
1912 static inline unsigned int uffd_ctx_features(__u64 user_features)
1913 {
1914         /*
1915          * For the current set of features the bits just coincide. Set
1916          * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1917          */
1918         return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1919 }
1920
1921 /*
1922  * userland asks for a certain API version and we return which bits
1923  * and ioctl commands are implemented in this kernel for such API
1924  * version or -EINVAL if unknown.
1925  */
1926 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1927                            unsigned long arg)
1928 {
1929         struct uffdio_api uffdio_api;
1930         void __user *buf = (void __user *)arg;
1931         unsigned int ctx_features;
1932         int ret;
1933         __u64 features;
1934
1935         ret = -EFAULT;
1936         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1937                 goto out;
1938         features = uffdio_api.features;
1939         ret = -EINVAL;
1940         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1941                 goto err_out;
1942         ret = -EPERM;
1943         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1944                 goto err_out;
1945         /* report all available features and ioctls to userland */
1946         uffdio_api.features = UFFD_API_FEATURES;
1947 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1948         uffdio_api.features &=
1949                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1950 #endif
1951 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1952         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1953 #endif
1954         uffdio_api.ioctls = UFFD_API_IOCTLS;
1955         ret = -EFAULT;
1956         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1957                 goto out;
1958
1959         /* only enable the requested features for this uffd context */
1960         ctx_features = uffd_ctx_features(features);
1961         ret = -EINVAL;
1962         if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
1963                 goto err_out;
1964
1965         ret = 0;
1966 out:
1967         return ret;
1968 err_out:
1969         memset(&uffdio_api, 0, sizeof(uffdio_api));
1970         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1971                 ret = -EFAULT;
1972         goto out;
1973 }
1974
1975 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1976                               unsigned long arg)
1977 {
1978         int ret = -EINVAL;
1979         struct userfaultfd_ctx *ctx = file->private_data;
1980
1981         if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
1982                 return -EINVAL;
1983
1984         switch(cmd) {
1985         case UFFDIO_API:
1986                 ret = userfaultfd_api(ctx, arg);
1987                 break;
1988         case UFFDIO_REGISTER:
1989                 ret = userfaultfd_register(ctx, arg);
1990                 break;
1991         case UFFDIO_UNREGISTER:
1992                 ret = userfaultfd_unregister(ctx, arg);
1993                 break;
1994         case UFFDIO_WAKE:
1995                 ret = userfaultfd_wake(ctx, arg);
1996                 break;
1997         case UFFDIO_COPY:
1998                 ret = userfaultfd_copy(ctx, arg);
1999                 break;
2000         case UFFDIO_ZEROPAGE:
2001                 ret = userfaultfd_zeropage(ctx, arg);
2002                 break;
2003         case UFFDIO_WRITEPROTECT:
2004                 ret = userfaultfd_writeprotect(ctx, arg);
2005                 break;
2006         case UFFDIO_CONTINUE:
2007                 ret = userfaultfd_continue(ctx, arg);
2008                 break;
2009         }
2010         return ret;
2011 }
2012
2013 #ifdef CONFIG_PROC_FS
2014 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2015 {
2016         struct userfaultfd_ctx *ctx = f->private_data;
2017         wait_queue_entry_t *wq;
2018         unsigned long pending = 0, total = 0;
2019
2020         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2021         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2022                 pending++;
2023                 total++;
2024         }
2025         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2026                 total++;
2027         }
2028         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2029
2030         /*
2031          * If more protocols will be added, there will be all shown
2032          * separated by a space. Like this:
2033          *      protocols: aa:... bb:...
2034          */
2035         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2036                    pending, total, UFFD_API, ctx->features,
2037                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2038 }
2039 #endif
2040
2041 static const struct file_operations userfaultfd_fops = {
2042 #ifdef CONFIG_PROC_FS
2043         .show_fdinfo    = userfaultfd_show_fdinfo,
2044 #endif
2045         .release        = userfaultfd_release,
2046         .poll           = userfaultfd_poll,
2047         .read           = userfaultfd_read,
2048         .unlocked_ioctl = userfaultfd_ioctl,
2049         .compat_ioctl   = compat_ptr_ioctl,
2050         .llseek         = noop_llseek,
2051 };
2052
2053 static void init_once_userfaultfd_ctx(void *mem)
2054 {
2055         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2056
2057         init_waitqueue_head(&ctx->fault_pending_wqh);
2058         init_waitqueue_head(&ctx->fault_wqh);
2059         init_waitqueue_head(&ctx->event_wqh);
2060         init_waitqueue_head(&ctx->fd_wqh);
2061         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2062 }
2063
2064 SYSCALL_DEFINE1(userfaultfd, int, flags)
2065 {
2066         struct userfaultfd_ctx *ctx;
2067         int fd;
2068
2069         if (!sysctl_unprivileged_userfaultfd &&
2070             (flags & UFFD_USER_MODE_ONLY) == 0 &&
2071             !capable(CAP_SYS_PTRACE)) {
2072                 printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2073                         "sysctl knob to 1 if kernel faults must be handled "
2074                         "without obtaining CAP_SYS_PTRACE capability\n");
2075                 return -EPERM;
2076         }
2077
2078         BUG_ON(!current->mm);
2079
2080         /* Check the UFFD_* constants for consistency.  */
2081         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2082         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2083         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2084
2085         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2086                 return -EINVAL;
2087
2088         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2089         if (!ctx)
2090                 return -ENOMEM;
2091
2092         refcount_set(&ctx->refcount, 1);
2093         ctx->flags = flags;
2094         ctx->features = 0;
2095         ctx->released = false;
2096         atomic_set(&ctx->mmap_changing, 0);
2097         ctx->mm = current->mm;
2098         /* prevent the mm struct to be freed */
2099         mmgrab(ctx->mm);
2100
2101         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2102                         O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2103         if (fd < 0) {
2104                 mmdrop(ctx->mm);
2105                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2106         }
2107         return fd;
2108 }
2109
2110 static int __init userfaultfd_init(void)
2111 {
2112         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2113                                                 sizeof(struct userfaultfd_ctx),
2114                                                 0,
2115                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2116                                                 init_once_userfaultfd_ctx);
2117         return 0;
2118 }
2119 __initcall(userfaultfd_init);