fs/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  fs/userfaultfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <[email protected]>
   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 #include <linux/swapops.h>
  33 #include <linux/miscdevice.h>
  34 #include <linux/uio.h>
  35
  36 static int sysctl_unprivileged_userfaultfd __read_mostly;
  37
  38 #ifdef CONFIG_SYSCTL
  39 static struct ctl_table vm_userfaultfd_table[] = {
  40         {
  41                 .procname       = "unprivileged_userfaultfd",
  42                 .data           = &sysctl_unprivileged_userfaultfd,
  43                 .maxlen         = sizeof(sysctl_unprivileged_userfaultfd),
  44                 .mode           = 0644,
  45                 .proc_handler   = proc_dointvec_minmax,
  46                 .extra1         = SYSCTL_ZERO,
  47                 .extra2         = SYSCTL_ONE,
  48         },
  49 };
  50 #endif
  51
  52 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
  53
  54 struct userfaultfd_fork_ctx {
  55         struct userfaultfd_ctx *orig;
  56         struct userfaultfd_ctx *new;
  57         struct list_head list;
  58 };
  59
  60 struct userfaultfd_unmap_ctx {
  61         struct userfaultfd_ctx *ctx;
  62         unsigned long start;
  63         unsigned long end;
  64         struct list_head list;
  65 };
  66
  67 struct userfaultfd_wait_queue {
  68         struct uffd_msg msg;
  69         wait_queue_entry_t wq;
  70         struct userfaultfd_ctx *ctx;
  71         bool waken;
  72 };
  73
  74 struct userfaultfd_wake_range {
  75         unsigned long start;
  76         unsigned long len;
  77 };
  78
  79 /* internal indication that UFFD_API ioctl was successfully executed */
  80 #define UFFD_FEATURE_INITIALIZED                (1u << 31)
  81
  82 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
  83 {
  84         return ctx->features & UFFD_FEATURE_INITIALIZED;
  85 }
  86
  87 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
  88 {
  89         return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
  90 }
  91
  92 /*
  93  * Whether WP_UNPOPULATED is enabled on the uffd context.  It is only
  94  * meaningful when userfaultfd_wp()==true on the vma and when it's
  95  * anonymous.
  96  */
  97 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
  98 {
  99         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 100
 101         if (!ctx)
 102                 return false;
 103
 104         return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
 105 }
 106
 107 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 108                                      int wake_flags, void *key)
 109 {
 110         struct userfaultfd_wake_range *range = key;
 111         int ret;
 112         struct userfaultfd_wait_queue *uwq;
 113         unsigned long start, len;
 114
 115         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 116         ret = 0;
 117         /* len == 0 means wake all */
 118         start = range->start;
 119         len = range->len;
 120         if (len && (start > uwq->msg.arg.pagefault.address ||
 121                     start + len <= uwq->msg.arg.pagefault.address))
 122                 goto out;
 123         WRITE_ONCE(uwq->waken, true);
 124         /*
 125          * The Program-Order guarantees provided by the scheduler
 126          * ensure uwq->waken is visible before the task is woken.
 127          */
 128         ret = wake_up_state(wq->private, mode);
 129         if (ret) {
 130                 /*
 131                  * Wake only once, autoremove behavior.
 132                  *
 133                  * After the effect of list_del_init is visible to the other
 134                  * CPUs, the waitqueue may disappear from under us, see the
 135                  * !list_empty_careful() in handle_userfault().
 136                  *
 137                  * try_to_wake_up() has an implicit smp_mb(), and the
 138                  * wq->private is read before calling the extern function
 139                  * "wake_up_state" (which in turns calls try_to_wake_up).
 140                  */
 141                 list_del_init(&wq->entry);
 142         }
 143 out:
 144         return ret;
 145 }
 146
 147 /**
 148  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 149  * context.
 150  * @ctx: [in] Pointer to the userfaultfd context.
 151  */
 152 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 153 {
 154         refcount_inc(&ctx->refcount);
 155 }
 156
 157 /**
 158  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 159  * context.
 160  * @ctx: [in] Pointer to userfaultfd context.
 161  *
 162  * The userfaultfd context reference must have been previously acquired either
 163  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 164  */
 165 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 166 {
 167         if (refcount_dec_and_test(&ctx->refcount)) {
 168                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 169                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 170                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 171                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 172                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 173                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 174                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 175                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 176                 mmdrop(ctx->mm);
 177                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 178         }
 179 }
 180
 181 static inline void msg_init(struct uffd_msg *msg)
 182 {
 183         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 184         /*
 185          * Must use memset to zero out the paddings or kernel data is
 186          * leaked to userland.
 187          */
 188         memset(msg, 0, sizeof(struct uffd_msg));
 189 }
 190
 191 static inline struct uffd_msg userfault_msg(unsigned long address,
 192                                             unsigned long real_address,
 193                                             unsigned int flags,
 194                                             unsigned long reason,
 195                                             unsigned int features)
 196 {
 197         struct uffd_msg msg;
 198
 199         msg_init(&msg);
 200         msg.event = UFFD_EVENT_PAGEFAULT;
 201
 202         msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
 203                                     real_address : address;
 204
 205         /*
 206          * These flags indicate why the userfault occurred:
 207          * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
 208          * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
 209          * - Neither of these flags being set indicates a MISSING fault.
 210          *
 211          * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
 212          * fault. Otherwise, it was a read fault.
 213          */
 214         if (flags & FAULT_FLAG_WRITE)
 215                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 216         if (reason & VM_UFFD_WP)
 217                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 218         if (reason & VM_UFFD_MINOR)
 219                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
 220         if (features & UFFD_FEATURE_THREAD_ID)
 221                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 222         return msg;
 223 }
 224
 225 #ifdef CONFIG_HUGETLB_PAGE
 226 /*
 227  * Same functionality as userfaultfd_must_wait below with modifications for
 228  * hugepmd ranges.
 229  */
 230 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 231                                               struct vm_fault *vmf,
 232                                               unsigned long reason)
 233 {
 234         struct vm_area_struct *vma = vmf->vma;
 235         pte_t *ptep, pte;
 236         bool ret = true;
 237
 238         assert_fault_locked(vmf);
 239
 240         ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
 241         if (!ptep)
 242                 goto out;
 243
 244         ret = false;
 245         pte = huge_ptep_get(vma->vm_mm, vmf->address, ptep);
 246
 247         /*
 248          * Lockless access: we're in a wait_event so it's ok if it
 249          * changes under us.  PTE markers should be handled the same as none
 250          * ptes here.
 251          */
 252         if (huge_pte_none_mostly(pte))
 253                 ret = true;
 254         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 255                 ret = true;
 256 out:
 257         return ret;
 258 }
 259 #else
 260 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 261                                               struct vm_fault *vmf,
 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_iter 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                                          struct vm_fault *vmf,
 277                                          unsigned long reason)
 278 {
 279         struct mm_struct *mm = ctx->mm;
 280         unsigned long address = vmf->address;
 281         pgd_t *pgd;
 282         p4d_t *p4d;
 283         pud_t *pud;
 284         pmd_t *pmd, _pmd;
 285         pte_t *pte;
 286         pte_t ptent;
 287         bool ret = true;
 288
 289         assert_fault_locked(vmf);
 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 again:
 302         _pmd = pmdp_get_lockless(pmd);
 303         if (pmd_none(_pmd))
 304                 goto out;
 305
 306         ret = false;
 307         if (!pmd_present(_pmd) || pmd_devmap(_pmd))
 308                 goto out;
 309
 310         if (pmd_trans_huge(_pmd)) {
 311                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 312                         ret = true;
 313                 goto out;
 314         }
 315
 316         pte = pte_offset_map(pmd, address);
 317         if (!pte) {
 318                 ret = true;
 319                 goto again;
 320         }
 321         /*
 322          * Lockless access: we're in a wait_event so it's ok if it
 323          * changes under us.  PTE markers should be handled the same as none
 324          * ptes here.
 325          */
 326         ptent = ptep_get(pte);
 327         if (pte_none_mostly(ptent))
 328                 ret = true;
 329         if (!pte_write(ptent) && (reason & VM_UFFD_WP))
 330                 ret = true;
 331         pte_unmap(pte);
 332
 333 out:
 334         return ret;
 335 }
 336
 337 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
 338 {
 339         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 340                 return TASK_INTERRUPTIBLE;
 341
 342         if (flags & FAULT_FLAG_KILLABLE)
 343                 return TASK_KILLABLE;
 344
 345         return TASK_UNINTERRUPTIBLE;
 346 }
 347
 348 /*
 349  * The locking rules involved in returning VM_FAULT_RETRY depending on
 350  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 351  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 352  * recommendation in __lock_page_or_retry is not an understatement.
 353  *
 354  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
 355  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 356  * not set.
 357  *
 358  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 359  * set, VM_FAULT_RETRY can still be returned if and only if there are
 360  * fatal_signal_pending()s, and the mmap_lock must be released before
 361  * returning it.
 362  */
 363 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 364 {
 365         struct vm_area_struct *vma = vmf->vma;
 366         struct mm_struct *mm = vma->vm_mm;
 367         struct userfaultfd_ctx *ctx;
 368         struct userfaultfd_wait_queue uwq;
 369         vm_fault_t ret = VM_FAULT_SIGBUS;
 370         bool must_wait;
 371         unsigned int blocking_state;
 372
 373         /*
 374          * We don't do userfault handling for the final child pid update
 375          * and when coredumping (faults triggered by get_dump_page()).
 376          */
 377         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 378                 goto out;
 379
 380         assert_fault_locked(vmf);
 381
 382         ctx = vma->vm_userfaultfd_ctx.ctx;
 383         if (!ctx)
 384                 goto out;
 385
 386         BUG_ON(ctx->mm != mm);
 387
 388         /* Any unrecognized flag is a bug. */
 389         VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
 390         /* 0 or > 1 flags set is a bug; we expect exactly 1. */
 391         VM_BUG_ON(!reason || (reason & (reason - 1)));
 392
 393         if (ctx->features & UFFD_FEATURE_SIGBUS)
 394                 goto out;
 395         if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
 396                 goto out;
 397
 398         /*
 399          * If it's already released don't get it. This avoids to loop
 400          * in __get_user_pages if userfaultfd_release waits on the
 401          * caller of handle_userfault to release the mmap_lock.
 402          */
 403         if (unlikely(READ_ONCE(ctx->released))) {
 404                 /*
 405                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 406                  * cooperative manager can close the uffd after the
 407                  * last UFFDIO_COPY, without risking to trigger an
 408                  * involuntary SIGBUS if the process was starting the
 409                  * userfaultfd while the userfaultfd was still armed
 410                  * (but after the last UFFDIO_COPY). If the uffd
 411                  * wasn't already closed when the userfault reached
 412                  * this point, that would normally be solved by
 413                  * userfaultfd_must_wait returning 'false'.
 414                  *
 415                  * If we were to return VM_FAULT_SIGBUS here, the non
 416                  * cooperative manager would be instead forced to
 417                  * always call UFFDIO_UNREGISTER before it can safely
 418                  * close the uffd.
 419                  */
 420                 ret = VM_FAULT_NOPAGE;
 421                 goto out;
 422         }
 423
 424         /*
 425          * Check that we can return VM_FAULT_RETRY.
 426          *
 427          * NOTE: it should become possible to return VM_FAULT_RETRY
 428          * even if FAULT_FLAG_TRIED is set without leading to gup()
 429          * -EBUSY failures, if the userfaultfd is to be extended for
 430          * VM_UFFD_WP tracking and we intend to arm the userfault
 431          * without first stopping userland access to the memory. For
 432          * VM_UFFD_MISSING userfaults this is enough for now.
 433          */
 434         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 435                 /*
 436                  * Validate the invariant that nowait must allow retry
 437                  * to be sure not to return SIGBUS erroneously on
 438                  * nowait invocations.
 439                  */
 440                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 441 #ifdef CONFIG_DEBUG_VM
 442                 if (printk_ratelimit()) {
 443                         printk(KERN_WARNING
 444                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 445                                vmf->flags);
 446                         dump_stack();
 447                 }
 448 #endif
 449                 goto out;
 450         }
 451
 452         /*
 453          * Handle nowait, not much to do other than tell it to retry
 454          * and wait.
 455          */
 456         ret = VM_FAULT_RETRY;
 457         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 458                 goto out;
 459
 460         /* take the reference before dropping the mmap_lock */
 461         userfaultfd_ctx_get(ctx);
 462
 463         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 464         uwq.wq.private = current;
 465         uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
 466                                 reason, ctx->features);
 467         uwq.ctx = ctx;
 468         uwq.waken = false;
 469
 470         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 471
 472         /*
 473          * Take the vma lock now, in order to safely call
 474          * userfaultfd_huge_must_wait() later. Since acquiring the
 475          * (sleepable) vma lock can modify the current task state, that
 476          * must be before explicitly calling set_current_state().
 477          */
 478         if (is_vm_hugetlb_page(vma))
 479                 hugetlb_vma_lock_read(vma);
 480
 481         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 482         /*
 483          * After the __add_wait_queue the uwq is visible to userland
 484          * through poll/read().
 485          */
 486         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 487         /*
 488          * The smp_mb() after __set_current_state prevents the reads
 489          * following the spin_unlock to happen before the list_add in
 490          * __add_wait_queue.
 491          */
 492         set_current_state(blocking_state);
 493         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 494
 495         if (!is_vm_hugetlb_page(vma))
 496                 must_wait = userfaultfd_must_wait(ctx, vmf, reason);
 497         else
 498                 must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
 499         if (is_vm_hugetlb_page(vma))
 500                 hugetlb_vma_unlock_read(vma);
 501         release_fault_lock(vmf);
 502
 503         if (likely(must_wait && !READ_ONCE(ctx->released))) {
 504                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 505                 schedule();
 506         }
 507
 508         __set_current_state(TASK_RUNNING);
 509
 510         /*
 511          * Here we race with the list_del; list_add in
 512          * userfaultfd_ctx_read(), however because we don't ever run
 513          * list_del_init() to refile across the two lists, the prev
 514          * and next pointers will never point to self. list_add also
 515          * would never let any of the two pointers to point to
 516          * self. So list_empty_careful won't risk to see both pointers
 517          * pointing to self at any time during the list refile. The
 518          * only case where list_del_init() is called is the full
 519          * removal in the wake function and there we don't re-list_add
 520          * and it's fine not to block on the spinlock. The uwq on this
 521          * kernel stack can be released after the list_del_init.
 522          */
 523         if (!list_empty_careful(&uwq.wq.entry)) {
 524                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 525                 /*
 526                  * No need of list_del_init(), the uwq on the stack
 527                  * will be freed shortly anyway.
 528                  */
 529                 list_del(&uwq.wq.entry);
 530                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 531         }
 532
 533         /*
 534          * ctx may go away after this if the userfault pseudo fd is
 535          * already released.
 536          */
 537         userfaultfd_ctx_put(ctx);
 538
 539 out:
 540         return ret;
 541 }
 542
 543 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 544                                               struct userfaultfd_wait_queue *ewq)
 545 {
 546         struct userfaultfd_ctx *release_new_ctx;
 547
 548         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 549                 goto out;
 550
 551         ewq->ctx = ctx;
 552         init_waitqueue_entry(&ewq->wq, current);
 553         release_new_ctx = NULL;
 554
 555         spin_lock_irq(&ctx->event_wqh.lock);
 556         /*
 557          * After the __add_wait_queue the uwq is visible to userland
 558          * through poll/read().
 559          */
 560         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 561         for (;;) {
 562                 set_current_state(TASK_KILLABLE);
 563                 if (ewq->msg.event == 0)
 564                         break;
 565                 if (READ_ONCE(ctx->released) ||
 566                     fatal_signal_pending(current)) {
 567                         /*
 568                          * &ewq->wq may be queued in fork_event, but
 569                          * __remove_wait_queue ignores the head
 570                          * parameter. It would be a problem if it
 571                          * didn't.
 572                          */
 573                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 574                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 575                                 struct userfaultfd_ctx *new;
 576
 577                                 new = (struct userfaultfd_ctx *)
 578                                         (unsigned long)
 579                                         ewq->msg.arg.reserved.reserved1;
 580                                 release_new_ctx = new;
 581                         }
 582                         break;
 583                 }
 584
 585                 spin_unlock_irq(&ctx->event_wqh.lock);
 586
 587                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 588                 schedule();
 589
 590                 spin_lock_irq(&ctx->event_wqh.lock);
 591         }
 592         __set_current_state(TASK_RUNNING);
 593         spin_unlock_irq(&ctx->event_wqh.lock);
 594
 595         if (release_new_ctx) {
 596                 userfaultfd_release_new(release_new_ctx);
 597                 userfaultfd_ctx_put(release_new_ctx);
 598         }
 599
 600         /*
 601          * ctx may go away after this if the userfault pseudo fd is
 602          * already released.
 603          */
 604 out:
 605         atomic_dec(&ctx->mmap_changing);
 606         VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
 607         userfaultfd_ctx_put(ctx);
 608 }
 609
 610 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 611                                        struct userfaultfd_wait_queue *ewq)
 612 {
 613         ewq->msg.event = 0;
 614         wake_up_locked(&ctx->event_wqh);
 615         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 616 }
 617
 618 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 619 {
 620         struct userfaultfd_ctx *ctx = NULL, *octx;
 621         struct userfaultfd_fork_ctx *fctx;
 622
 623         octx = vma->vm_userfaultfd_ctx.ctx;
 624         if (!octx)
 625                 return 0;
 626
 627         if (!(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 628                 userfaultfd_reset_ctx(vma);
 629                 return 0;
 630         }
 631
 632         list_for_each_entry(fctx, fcs, list)
 633                 if (fctx->orig == octx) {
 634                         ctx = fctx->new;
 635                         break;
 636                 }
 637
 638         if (!ctx) {
 639                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 640                 if (!fctx)
 641                         return -ENOMEM;
 642
 643                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 644                 if (!ctx) {
 645                         kfree(fctx);
 646                         return -ENOMEM;
 647                 }
 648
 649                 refcount_set(&ctx->refcount, 1);
 650                 ctx->flags = octx->flags;
 651                 ctx->features = octx->features;
 652                 ctx->released = false;
 653                 init_rwsem(&ctx->map_changing_lock);
 654                 atomic_set(&ctx->mmap_changing, 0);
 655                 ctx->mm = vma->vm_mm;
 656                 mmgrab(ctx->mm);
 657
 658                 userfaultfd_ctx_get(octx);
 659                 down_write(&octx->map_changing_lock);
 660                 atomic_inc(&octx->mmap_changing);
 661                 up_write(&octx->map_changing_lock);
 662                 fctx->orig = octx;
 663                 fctx->new = ctx;
 664                 list_add_tail(&fctx->list, fcs);
 665         }
 666
 667         vma->vm_userfaultfd_ctx.ctx = ctx;
 668         return 0;
 669 }
 670
 671 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 672 {
 673         struct userfaultfd_ctx *ctx = fctx->orig;
 674         struct userfaultfd_wait_queue ewq;
 675
 676         msg_init(&ewq.msg);
 677
 678         ewq.msg.event = UFFD_EVENT_FORK;
 679         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 680
 681         userfaultfd_event_wait_completion(ctx, &ewq);
 682 }
 683
 684 void dup_userfaultfd_complete(struct list_head *fcs)
 685 {
 686         struct userfaultfd_fork_ctx *fctx, *n;
 687
 688         list_for_each_entry_safe(fctx, n, fcs, list) {
 689                 dup_fctx(fctx);
 690                 list_del(&fctx->list);
 691                 kfree(fctx);
 692         }
 693 }
 694
 695 void dup_userfaultfd_fail(struct list_head *fcs)
 696 {
 697         struct userfaultfd_fork_ctx *fctx, *n;
 698
 699         /*
 700          * An error has occurred on fork, we will tear memory down, but have
 701          * allocated memory for fctx's and raised reference counts for both the
 702          * original and child contexts (and on the mm for each as a result).
 703          *
 704          * These would ordinarily be taken care of by a user handling the event,
 705          * but we are no longer doing so, so manually clean up here.
 706          *
 707          * mm tear down will take care of cleaning up VMA contexts.
 708          */
 709         list_for_each_entry_safe(fctx, n, fcs, list) {
 710                 struct userfaultfd_ctx *octx = fctx->orig;
 711                 struct userfaultfd_ctx *ctx = fctx->new;
 712
 713                 atomic_dec(&octx->mmap_changing);
 714                 VM_BUG_ON(atomic_read(&octx->mmap_changing) < 0);
 715                 userfaultfd_ctx_put(octx);
 716                 userfaultfd_ctx_put(ctx);
 717
 718                 list_del(&fctx->list);
 719                 kfree(fctx);
 720         }
 721 }
 722
 723 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 724                              struct vm_userfaultfd_ctx *vm_ctx)
 725 {
 726         struct userfaultfd_ctx *ctx;
 727
 728         ctx = vma->vm_userfaultfd_ctx.ctx;
 729
 730         if (!ctx)
 731                 return;
 732
 733         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 734                 vm_ctx->ctx = ctx;
 735                 userfaultfd_ctx_get(ctx);
 736                 down_write(&ctx->map_changing_lock);
 737                 atomic_inc(&ctx->mmap_changing);
 738                 up_write(&ctx->map_changing_lock);
 739         } else {
 740                 /* Drop uffd context if remap feature not enabled */
 741                 userfaultfd_reset_ctx(vma);
 742         }
 743 }
 744
 745 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 746                                  unsigned long from, unsigned long to,
 747                                  unsigned long len)
 748 {
 749         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 750         struct userfaultfd_wait_queue ewq;
 751
 752         if (!ctx)
 753                 return;
 754
 755         if (to & ~PAGE_MASK) {
 756                 userfaultfd_ctx_put(ctx);
 757                 return;
 758         }
 759
 760         msg_init(&ewq.msg);
 761
 762         ewq.msg.event = UFFD_EVENT_REMAP;
 763         ewq.msg.arg.remap.from = from;
 764         ewq.msg.arg.remap.to = to;
 765         ewq.msg.arg.remap.len = len;
 766
 767         userfaultfd_event_wait_completion(ctx, &ewq);
 768 }
 769
 770 bool userfaultfd_remove(struct vm_area_struct *vma,
 771                         unsigned long start, unsigned long end)
 772 {
 773         struct mm_struct *mm = vma->vm_mm;
 774         struct userfaultfd_ctx *ctx;
 775         struct userfaultfd_wait_queue ewq;
 776
 777         ctx = vma->vm_userfaultfd_ctx.ctx;
 778         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 779                 return true;
 780
 781         userfaultfd_ctx_get(ctx);
 782         down_write(&ctx->map_changing_lock);
 783         atomic_inc(&ctx->mmap_changing);
 784         up_write(&ctx->map_changing_lock);
 785         mmap_read_unlock(mm);
 786
 787         msg_init(&ewq.msg);
 788
 789         ewq.msg.event = UFFD_EVENT_REMOVE;
 790         ewq.msg.arg.remove.start = start;
 791         ewq.msg.arg.remove.end = end;
 792
 793         userfaultfd_event_wait_completion(ctx, &ewq);
 794
 795         return false;
 796 }
 797
 798 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 799                           unsigned long start, unsigned long end)
 800 {
 801         struct userfaultfd_unmap_ctx *unmap_ctx;
 802
 803         list_for_each_entry(unmap_ctx, unmaps, list)
 804                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 805                     unmap_ctx->end == end)
 806                         return true;
 807
 808         return false;
 809 }
 810
 811 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
 812                            unsigned long end, struct list_head *unmaps)
 813 {
 814         struct userfaultfd_unmap_ctx *unmap_ctx;
 815         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 816
 817         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 818             has_unmap_ctx(ctx, unmaps, start, end))
 819                 return 0;
 820
 821         unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 822         if (!unmap_ctx)
 823                 return -ENOMEM;
 824
 825         userfaultfd_ctx_get(ctx);
 826         down_write(&ctx->map_changing_lock);
 827         atomic_inc(&ctx->mmap_changing);
 828         up_write(&ctx->map_changing_lock);
 829         unmap_ctx->ctx = ctx;
 830         unmap_ctx->start = start;
 831         unmap_ctx->end = end;
 832         list_add_tail(&unmap_ctx->list, unmaps);
 833
 834         return 0;
 835 }
 836
 837 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 838 {
 839         struct userfaultfd_unmap_ctx *ctx, *n;
 840         struct userfaultfd_wait_queue ewq;
 841
 842         list_for_each_entry_safe(ctx, n, uf, list) {
 843                 msg_init(&ewq.msg);
 844
 845                 ewq.msg.event = UFFD_EVENT_UNMAP;
 846                 ewq.msg.arg.remove.start = ctx->start;
 847                 ewq.msg.arg.remove.end = ctx->end;
 848
 849                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 850
 851                 list_del(&ctx->list);
 852                 kfree(ctx);
 853         }
 854 }
 855
 856 static int userfaultfd_release(struct inode *inode, struct file *file)
 857 {
 858         struct userfaultfd_ctx *ctx = file->private_data;
 859         struct mm_struct *mm = ctx->mm;
 860         /* len == 0 means wake all */
 861         struct userfaultfd_wake_range range = { .len = 0, };
 862
 863         WRITE_ONCE(ctx->released, true);
 864
 865         userfaultfd_release_all(mm, ctx);
 866
 867         /*
 868          * After no new page faults can wait on this fault_*wqh, flush
 869          * the last page faults that may have been already waiting on
 870          * the fault_*wqh.
 871          */
 872         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 873         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 874         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 875         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 876
 877         /* Flush pending events that may still wait on event_wqh */
 878         wake_up_all(&ctx->event_wqh);
 879
 880         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 881         userfaultfd_ctx_put(ctx);
 882         return 0;
 883 }
 884
 885 /* fault_pending_wqh.lock must be hold by the caller */
 886 static inline struct userfaultfd_wait_queue *find_userfault_in(
 887                 wait_queue_head_t *wqh)
 888 {
 889         wait_queue_entry_t *wq;
 890         struct userfaultfd_wait_queue *uwq;
 891
 892         lockdep_assert_held(&wqh->lock);
 893
 894         uwq = NULL;
 895         if (!waitqueue_active(wqh))
 896                 goto out;
 897         /* walk in reverse to provide FIFO behavior to read userfaults */
 898         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 899         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 900 out:
 901         return uwq;
 902 }
 903
 904 static inline struct userfaultfd_wait_queue *find_userfault(
 905                 struct userfaultfd_ctx *ctx)
 906 {
 907         return find_userfault_in(&ctx->fault_pending_wqh);
 908 }
 909
 910 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 911                 struct userfaultfd_ctx *ctx)
 912 {
 913         return find_userfault_in(&ctx->event_wqh);
 914 }
 915
 916 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 917 {
 918         struct userfaultfd_ctx *ctx = file->private_data;
 919         __poll_t ret;
 920
 921         poll_wait(file, &ctx->fd_wqh, wait);
 922
 923         if (!userfaultfd_is_initialized(ctx))
 924                 return EPOLLERR;
 925
 926         /*
 927          * poll() never guarantees that read won't block.
 928          * userfaults can be waken before they're read().
 929          */
 930         if (unlikely(!(file->f_flags & O_NONBLOCK)))
 931                 return EPOLLERR;
 932         /*
 933          * lockless access to see if there are pending faults
 934          * __pollwait last action is the add_wait_queue but
 935          * the spin_unlock would allow the waitqueue_active to
 936          * pass above the actual list_add inside
 937          * add_wait_queue critical section. So use a full
 938          * memory barrier to serialize the list_add write of
 939          * add_wait_queue() with the waitqueue_active read
 940          * below.
 941          */
 942         ret = 0;
 943         smp_mb();
 944         if (waitqueue_active(&ctx->fault_pending_wqh))
 945                 ret = EPOLLIN;
 946         else if (waitqueue_active(&ctx->event_wqh))
 947                 ret = EPOLLIN;
 948
 949         return ret;
 950 }
 951
 952 static const struct file_operations userfaultfd_fops;
 953
 954 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
 955                                   struct inode *inode,
 956                                   struct uffd_msg *msg)
 957 {
 958         int fd;
 959
 960         fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, new,
 961                         O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
 962         if (fd < 0)
 963                 return fd;
 964
 965         msg->arg.reserved.reserved1 = 0;
 966         msg->arg.fork.ufd = fd;
 967         return 0;
 968 }
 969
 970 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 971                                     struct uffd_msg *msg, struct inode *inode)
 972 {
 973         ssize_t ret;
 974         DECLARE_WAITQUEUE(wait, current);
 975         struct userfaultfd_wait_queue *uwq;
 976         /*
 977          * Handling fork event requires sleeping operations, so
 978          * we drop the event_wqh lock, then do these ops, then
 979          * lock it back and wake up the waiter. While the lock is
 980          * dropped the ewq may go away so we keep track of it
 981          * carefully.
 982          */
 983         LIST_HEAD(fork_event);
 984         struct userfaultfd_ctx *fork_nctx = NULL;
 985
 986         /* always take the fd_wqh lock before the fault_pending_wqh lock */
 987         spin_lock_irq(&ctx->fd_wqh.lock);
 988         __add_wait_queue(&ctx->fd_wqh, &wait);
 989         for (;;) {
 990                 set_current_state(TASK_INTERRUPTIBLE);
 991                 spin_lock(&ctx->fault_pending_wqh.lock);
 992                 uwq = find_userfault(ctx);
 993                 if (uwq) {
 994                         /*
 995                          * Use a seqcount to repeat the lockless check
 996                          * in wake_userfault() to avoid missing
 997                          * wakeups because during the refile both
 998                          * waitqueue could become empty if this is the
 999                          * only userfault.
1000                          */
1001                         write_seqcount_begin(&ctx->refile_seq);
1002
1003                         /*
1004                          * The fault_pending_wqh.lock prevents the uwq
1005                          * to disappear from under us.
1006                          *
1007                          * Refile this userfault from
1008                          * fault_pending_wqh to fault_wqh, it's not
1009                          * pending anymore after we read it.
1010                          *
1011                          * Use list_del() by hand (as
1012                          * userfaultfd_wake_function also uses
1013                          * list_del_init() by hand) to be sure nobody
1014                          * changes __remove_wait_queue() to use
1015                          * list_del_init() in turn breaking the
1016                          * !list_empty_careful() check in
1017                          * handle_userfault(). The uwq->wq.head list
1018                          * must never be empty at any time during the
1019                          * refile, or the waitqueue could disappear
1020                          * from under us. The "wait_queue_head_t"
1021                          * parameter of __remove_wait_queue() is unused
1022                          * anyway.
1023                          */
1024                         list_del(&uwq->wq.entry);
1025                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1026
1027                         write_seqcount_end(&ctx->refile_seq);
1028
1029                         /* careful to always initialize msg if ret == 0 */
1030                         *msg = uwq->msg;
1031                         spin_unlock(&ctx->fault_pending_wqh.lock);
1032                         ret = 0;
1033                         break;
1034                 }
1035                 spin_unlock(&ctx->fault_pending_wqh.lock);
1036
1037                 spin_lock(&ctx->event_wqh.lock);
1038                 uwq = find_userfault_evt(ctx);
1039                 if (uwq) {
1040                         *msg = uwq->msg;
1041
1042                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1043                                 fork_nctx = (struct userfaultfd_ctx *)
1044                                         (unsigned long)
1045                                         uwq->msg.arg.reserved.reserved1;
1046                                 list_move(&uwq->wq.entry, &fork_event);
1047                                 /*
1048                                  * fork_nctx can be freed as soon as
1049                                  * we drop the lock, unless we take a
1050                                  * reference on it.
1051                                  */
1052                                 userfaultfd_ctx_get(fork_nctx);
1053                                 spin_unlock(&ctx->event_wqh.lock);
1054                                 ret = 0;
1055                                 break;
1056                         }
1057
1058                         userfaultfd_event_complete(ctx, uwq);
1059                         spin_unlock(&ctx->event_wqh.lock);
1060                         ret = 0;
1061                         break;
1062                 }
1063                 spin_unlock(&ctx->event_wqh.lock);
1064
1065                 if (signal_pending(current)) {
1066                         ret = -ERESTARTSYS;
1067                         break;
1068                 }
1069                 if (no_wait) {
1070                         ret = -EAGAIN;
1071                         break;
1072                 }
1073                 spin_unlock_irq(&ctx->fd_wqh.lock);
1074                 schedule();
1075                 spin_lock_irq(&ctx->fd_wqh.lock);
1076         }
1077         __remove_wait_queue(&ctx->fd_wqh, &wait);
1078         __set_current_state(TASK_RUNNING);
1079         spin_unlock_irq(&ctx->fd_wqh.lock);
1080
1081         if (!ret && msg->event == UFFD_EVENT_FORK) {
1082                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1083                 spin_lock_irq(&ctx->event_wqh.lock);
1084                 if (!list_empty(&fork_event)) {
1085                         /*
1086                          * The fork thread didn't abort, so we can
1087                          * drop the temporary refcount.
1088                          */
1089                         userfaultfd_ctx_put(fork_nctx);
1090
1091                         uwq = list_first_entry(&fork_event,
1092                                                typeof(*uwq),
1093                                                wq.entry);
1094                         /*
1095                          * If fork_event list wasn't empty and in turn
1096                          * the event wasn't already released by fork
1097                          * (the event is allocated on fork kernel
1098                          * stack), put the event back to its place in
1099                          * the event_wq. fork_event head will be freed
1100                          * as soon as we return so the event cannot
1101                          * stay queued there no matter the current
1102                          * "ret" value.
1103                          */
1104                         list_del(&uwq->wq.entry);
1105                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1106
1107                         /*
1108                          * Leave the event in the waitqueue and report
1109                          * error to userland if we failed to resolve
1110                          * the userfault fork.
1111                          */
1112                         if (likely(!ret))
1113                                 userfaultfd_event_complete(ctx, uwq);
1114                 } else {
1115                         /*
1116                          * Here the fork thread aborted and the
1117                          * refcount from the fork thread on fork_nctx
1118                          * has already been released. We still hold
1119                          * the reference we took before releasing the
1120                          * lock above. If resolve_userfault_fork
1121                          * failed we've to drop it because the
1122                          * fork_nctx has to be freed in such case. If
1123                          * it succeeded we'll hold it because the new
1124                          * uffd references it.
1125                          */
1126                         if (ret)
1127                                 userfaultfd_ctx_put(fork_nctx);
1128                 }
1129                 spin_unlock_irq(&ctx->event_wqh.lock);
1130         }
1131
1132         return ret;
1133 }
1134
1135 static ssize_t userfaultfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
1136 {
1137         struct file *file = iocb->ki_filp;
1138         struct userfaultfd_ctx *ctx = file->private_data;
1139         ssize_t _ret, ret = 0;
1140         struct uffd_msg msg;
1141         struct inode *inode = file_inode(file);
1142         bool no_wait;
1143
1144         if (!userfaultfd_is_initialized(ctx))
1145                 return -EINVAL;
1146
1147         no_wait = file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT;
1148         for (;;) {
1149                 if (iov_iter_count(to) < sizeof(msg))
1150                         return ret ? ret : -EINVAL;
1151                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1152                 if (_ret < 0)
1153                         return ret ? ret : _ret;
1154                 _ret = !copy_to_iter_full(&msg, sizeof(msg), to);
1155                 if (_ret)
1156                         return ret ? ret : -EFAULT;
1157                 ret += sizeof(msg);
1158                 /*
1159                  * Allow to read more than one fault at time but only
1160                  * block if waiting for the very first one.
1161                  */
1162                 no_wait = true;
1163         }
1164 }
1165
1166 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1167                              struct userfaultfd_wake_range *range)
1168 {
1169         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1170         /* wake all in the range and autoremove */
1171         if (waitqueue_active(&ctx->fault_pending_wqh))
1172                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1173                                      range);
1174         if (waitqueue_active(&ctx->fault_wqh))
1175                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1176         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1177 }
1178
1179 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1180                                            struct userfaultfd_wake_range *range)
1181 {
1182         unsigned seq;
1183         bool need_wakeup;
1184
1185         /*
1186          * To be sure waitqueue_active() is not reordered by the CPU
1187          * before the pagetable update, use an explicit SMP memory
1188          * barrier here. PT lock release or mmap_read_unlock(mm) still
1189          * have release semantics that can allow the
1190          * waitqueue_active() to be reordered before the pte update.
1191          */
1192         smp_mb();
1193
1194         /*
1195          * Use waitqueue_active because it's very frequent to
1196          * change the address space atomically even if there are no
1197          * userfaults yet. So we take the spinlock only when we're
1198          * sure we've userfaults to wake.
1199          */
1200         do {
1201                 seq = read_seqcount_begin(&ctx->refile_seq);
1202                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1203                         waitqueue_active(&ctx->fault_wqh);
1204                 cond_resched();
1205         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1206         if (need_wakeup)
1207                 __wake_userfault(ctx, range);
1208 }
1209
1210 static __always_inline int validate_unaligned_range(
1211         struct mm_struct *mm, __u64 start, __u64 len)
1212 {
1213         __u64 task_size = mm->task_size;
1214
1215         if (len & ~PAGE_MASK)
1216                 return -EINVAL;
1217         if (!len)
1218                 return -EINVAL;
1219         if (start < mmap_min_addr)
1220                 return -EINVAL;
1221         if (start >= task_size)
1222                 return -EINVAL;
1223         if (len > task_size - start)
1224                 return -EINVAL;
1225         if (start + len <= start)
1226                 return -EINVAL;
1227         return 0;
1228 }
1229
1230 static __always_inline int validate_range(struct mm_struct *mm,
1231                                           __u64 start, __u64 len)
1232 {
1233         if (start & ~PAGE_MASK)
1234                 return -EINVAL;
1235
1236         return validate_unaligned_range(mm, start, len);
1237 }
1238
1239 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1240                                 unsigned long arg)
1241 {
1242         struct mm_struct *mm = ctx->mm;
1243         struct vm_area_struct *vma, *cur;
1244         int ret;
1245         struct uffdio_register uffdio_register;
1246         struct uffdio_register __user *user_uffdio_register;
1247         unsigned long vm_flags;
1248         bool found;
1249         bool basic_ioctls;
1250         unsigned long start, end;
1251         struct vma_iterator vmi;
1252         bool wp_async = userfaultfd_wp_async_ctx(ctx);
1253
1254         user_uffdio_register = (struct uffdio_register __user *) arg;
1255
1256         ret = -EFAULT;
1257         if (copy_from_user(&uffdio_register, user_uffdio_register,
1258                            sizeof(uffdio_register)-sizeof(__u64)))
1259                 goto out;
1260
1261         ret = -EINVAL;
1262         if (!uffdio_register.mode)
1263                 goto out;
1264         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1265                 goto out;
1266         vm_flags = 0;
1267         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1268                 vm_flags |= VM_UFFD_MISSING;
1269         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1270 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1271                 goto out;
1272 #endif
1273                 vm_flags |= VM_UFFD_WP;
1274         }
1275         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1276 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1277                 goto out;
1278 #endif
1279                 vm_flags |= VM_UFFD_MINOR;
1280         }
1281
1282         ret = validate_range(mm, uffdio_register.range.start,
1283                              uffdio_register.range.len);
1284         if (ret)
1285                 goto out;
1286
1287         start = uffdio_register.range.start;
1288         end = start + uffdio_register.range.len;
1289
1290         ret = -ENOMEM;
1291         if (!mmget_not_zero(mm))
1292                 goto out;
1293
1294         ret = -EINVAL;
1295         mmap_write_lock(mm);
1296         vma_iter_init(&vmi, mm, start);
1297         vma = vma_find(&vmi, end);
1298         if (!vma)
1299                 goto out_unlock;
1300
1301         /*
1302          * If the first vma contains huge pages, make sure start address
1303          * is aligned to huge page size.
1304          */
1305         if (is_vm_hugetlb_page(vma)) {
1306                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1307
1308                 if (start & (vma_hpagesize - 1))
1309                         goto out_unlock;
1310         }
1311
1312         /*
1313          * Search for not compatible vmas.
1314          */
1315         found = false;
1316         basic_ioctls = false;
1317         cur = vma;
1318         do {
1319                 cond_resched();
1320
1321                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1322                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1323
1324                 /* check not compatible vmas */
1325                 ret = -EINVAL;
1326                 if (!vma_can_userfault(cur, vm_flags, wp_async))
1327                         goto out_unlock;
1328
1329                 /*
1330                  * UFFDIO_COPY will fill file holes even without
1331                  * PROT_WRITE. This check enforces that if this is a
1332                  * MAP_SHARED, the process has write permission to the backing
1333                  * file. If VM_MAYWRITE is set it also enforces that on a
1334                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1335                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1336                  */
1337                 ret = -EPERM;
1338                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1339                         goto out_unlock;
1340
1341                 /*
1342                  * If this vma contains ending address, and huge pages
1343                  * check alignment.
1344                  */
1345                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1346                     end > cur->vm_start) {
1347                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1348
1349                         ret = -EINVAL;
1350
1351                         if (end & (vma_hpagesize - 1))
1352                                 goto out_unlock;
1353                 }
1354                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1355                         goto out_unlock;
1356
1357                 /*
1358                  * Check that this vma isn't already owned by a
1359                  * different userfaultfd. We can't allow more than one
1360                  * userfaultfd to own a single vma simultaneously or we
1361                  * wouldn't know which one to deliver the userfaults to.
1362                  */
1363                 ret = -EBUSY;
1364                 if (cur->vm_userfaultfd_ctx.ctx &&
1365                     cur->vm_userfaultfd_ctx.ctx != ctx)
1366                         goto out_unlock;
1367
1368                 /*
1369                  * Note vmas containing huge pages
1370                  */
1371                 if (is_vm_hugetlb_page(cur))
1372                         basic_ioctls = true;
1373
1374                 found = true;
1375         } for_each_vma_range(vmi, cur, end);
1376         BUG_ON(!found);
1377
1378         ret = userfaultfd_register_range(ctx, vma, vm_flags, start, end,
1379                                          wp_async);
1380
1381 out_unlock:
1382         mmap_write_unlock(mm);
1383         mmput(mm);
1384         if (!ret) {
1385                 __u64 ioctls_out;
1386
1387                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1388                     UFFD_API_RANGE_IOCTLS;
1389
1390                 /*
1391                  * Declare the WP ioctl only if the WP mode is
1392                  * specified and all checks passed with the range
1393                  */
1394                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1395                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1396
1397                 /* CONTINUE ioctl is only supported for MINOR ranges. */
1398                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1399                         ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1400
1401                 /*
1402                  * Now that we scanned all vmas we can already tell
1403                  * userland which ioctls methods are guaranteed to
1404                  * succeed on this range.
1405                  */
1406                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1407                         ret = -EFAULT;
1408         }
1409 out:
1410         return ret;
1411 }
1412
1413 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1414                                   unsigned long arg)
1415 {
1416         struct mm_struct *mm = ctx->mm;
1417         struct vm_area_struct *vma, *prev, *cur;
1418         int ret;
1419         struct uffdio_range uffdio_unregister;
1420         bool found;
1421         unsigned long start, end, vma_end;
1422         const void __user *buf = (void __user *)arg;
1423         struct vma_iterator vmi;
1424         bool wp_async = userfaultfd_wp_async_ctx(ctx);
1425
1426         ret = -EFAULT;
1427         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1428                 goto out;
1429
1430         ret = validate_range(mm, uffdio_unregister.start,
1431                              uffdio_unregister.len);
1432         if (ret)
1433                 goto out;
1434
1435         start = uffdio_unregister.start;
1436         end = start + uffdio_unregister.len;
1437
1438         ret = -ENOMEM;
1439         if (!mmget_not_zero(mm))
1440                 goto out;
1441
1442         mmap_write_lock(mm);
1443         ret = -EINVAL;
1444         vma_iter_init(&vmi, mm, start);
1445         vma = vma_find(&vmi, end);
1446         if (!vma)
1447                 goto out_unlock;
1448
1449         /*
1450          * If the first vma contains huge pages, make sure start address
1451          * is aligned to huge page size.
1452          */
1453         if (is_vm_hugetlb_page(vma)) {
1454                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1455
1456                 if (start & (vma_hpagesize - 1))
1457                         goto out_unlock;
1458         }
1459
1460         /*
1461          * Search for not compatible vmas.
1462          */
1463         found = false;
1464         cur = vma;
1465         do {
1466                 cond_resched();
1467
1468                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1469                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1470
1471                 /*
1472                  * Check not compatible vmas, not strictly required
1473                  * here as not compatible vmas cannot have an
1474                  * userfaultfd_ctx registered on them, but this
1475                  * provides for more strict behavior to notice
1476                  * unregistration errors.
1477                  */
1478                 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1479                         goto out_unlock;
1480
1481                 found = true;
1482         } for_each_vma_range(vmi, cur, end);
1483         BUG_ON(!found);
1484
1485         vma_iter_set(&vmi, start);
1486         prev = vma_prev(&vmi);
1487         if (vma->vm_start < start)
1488                 prev = vma;
1489
1490         ret = 0;
1491         for_each_vma_range(vmi, vma, end) {
1492                 cond_resched();
1493
1494                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1495
1496                 /*
1497                  * Nothing to do: this vma is already registered into this
1498                  * userfaultfd and with the right tracking mode too.
1499                  */
1500                 if (!vma->vm_userfaultfd_ctx.ctx)
1501                         goto skip;
1502
1503                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1504
1505                 if (vma->vm_start > start)
1506                         start = vma->vm_start;
1507                 vma_end = min(end, vma->vm_end);
1508
1509                 if (userfaultfd_missing(vma)) {
1510                         /*
1511                          * Wake any concurrent pending userfault while
1512                          * we unregister, so they will not hang
1513                          * permanently and it avoids userland to call
1514                          * UFFDIO_WAKE explicitly.
1515                          */
1516                         struct userfaultfd_wake_range range;
1517                         range.start = start;
1518                         range.len = vma_end - start;
1519                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1520                 }
1521
1522                 vma = userfaultfd_clear_vma(&vmi, prev, vma,
1523                                             start, vma_end);
1524                 if (IS_ERR(vma)) {
1525                         ret = PTR_ERR(vma);
1526                         break;
1527                 }
1528
1529         skip:
1530                 prev = vma;
1531                 start = vma->vm_end;
1532         }
1533
1534 out_unlock:
1535         mmap_write_unlock(mm);
1536         mmput(mm);
1537 out:
1538         return ret;
1539 }
1540
1541 /*
1542  * userfaultfd_wake may be used in combination with the
1543  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1544  */
1545 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1546                             unsigned long arg)
1547 {
1548         int ret;
1549         struct uffdio_range uffdio_wake;
1550         struct userfaultfd_wake_range range;
1551         const void __user *buf = (void __user *)arg;
1552
1553         ret = -EFAULT;
1554         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1555                 goto out;
1556
1557         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1558         if (ret)
1559                 goto out;
1560
1561         range.start = uffdio_wake.start;
1562         range.len = uffdio_wake.len;
1563
1564         /*
1565          * len == 0 means wake all and we don't want to wake all here,
1566          * so check it again to be sure.
1567          */
1568         VM_BUG_ON(!range.len);
1569
1570         wake_userfault(ctx, &range);
1571         ret = 0;
1572
1573 out:
1574         return ret;
1575 }
1576
1577 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1578                             unsigned long arg)
1579 {
1580         __s64 ret;
1581         struct uffdio_copy uffdio_copy;
1582         struct uffdio_copy __user *user_uffdio_copy;
1583         struct userfaultfd_wake_range range;
1584         uffd_flags_t flags = 0;
1585
1586         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1587
1588         ret = -EAGAIN;
1589         if (atomic_read(&ctx->mmap_changing))
1590                 goto out;
1591
1592         ret = -EFAULT;
1593         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1594                            /* don't copy "copy" last field */
1595                            sizeof(uffdio_copy)-sizeof(__s64)))
1596                 goto out;
1597
1598         ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1599                                        uffdio_copy.len);
1600         if (ret)
1601                 goto out;
1602         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1603         if (ret)
1604                 goto out;
1605
1606         ret = -EINVAL;
1607         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1608                 goto out;
1609         if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1610                 flags |= MFILL_ATOMIC_WP;
1611         if (mmget_not_zero(ctx->mm)) {
1612                 ret = mfill_atomic_copy(ctx, uffdio_copy.dst, uffdio_copy.src,
1613                                         uffdio_copy.len, flags);
1614                 mmput(ctx->mm);
1615         } else {
1616                 return -ESRCH;
1617         }
1618         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1619                 return -EFAULT;
1620         if (ret < 0)
1621                 goto out;
1622         BUG_ON(!ret);
1623         /* len == 0 would wake all */
1624         range.len = ret;
1625         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1626                 range.start = uffdio_copy.dst;
1627                 wake_userfault(ctx, &range);
1628         }
1629         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1630 out:
1631         return ret;
1632 }
1633
1634 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1635                                 unsigned long arg)
1636 {
1637         __s64 ret;
1638         struct uffdio_zeropage uffdio_zeropage;
1639         struct uffdio_zeropage __user *user_uffdio_zeropage;
1640         struct userfaultfd_wake_range range;
1641
1642         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1643
1644         ret = -EAGAIN;
1645         if (atomic_read(&ctx->mmap_changing))
1646                 goto out;
1647
1648         ret = -EFAULT;
1649         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1650                            /* don't copy "zeropage" last field */
1651                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1652                 goto out;
1653
1654         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1655                              uffdio_zeropage.range.len);
1656         if (ret)
1657                 goto out;
1658         ret = -EINVAL;
1659         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1660                 goto out;
1661
1662         if (mmget_not_zero(ctx->mm)) {
1663                 ret = mfill_atomic_zeropage(ctx, uffdio_zeropage.range.start,
1664                                            uffdio_zeropage.range.len);
1665                 mmput(ctx->mm);
1666         } else {
1667                 return -ESRCH;
1668         }
1669         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1670                 return -EFAULT;
1671         if (ret < 0)
1672                 goto out;
1673         /* len == 0 would wake all */
1674         BUG_ON(!ret);
1675         range.len = ret;
1676         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1677                 range.start = uffdio_zeropage.range.start;
1678                 wake_userfault(ctx, &range);
1679         }
1680         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1681 out:
1682         return ret;
1683 }
1684
1685 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1686                                     unsigned long arg)
1687 {
1688         int ret;
1689         struct uffdio_writeprotect uffdio_wp;
1690         struct uffdio_writeprotect __user *user_uffdio_wp;
1691         struct userfaultfd_wake_range range;
1692         bool mode_wp, mode_dontwake;
1693
1694         if (atomic_read(&ctx->mmap_changing))
1695                 return -EAGAIN;
1696
1697         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1698
1699         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1700                            sizeof(struct uffdio_writeprotect)))
1701                 return -EFAULT;
1702
1703         ret = validate_range(ctx->mm, uffdio_wp.range.start,
1704                              uffdio_wp.range.len);
1705         if (ret)
1706                 return ret;
1707
1708         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1709                                UFFDIO_WRITEPROTECT_MODE_WP))
1710                 return -EINVAL;
1711
1712         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1713         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1714
1715         if (mode_wp && mode_dontwake)
1716                 return -EINVAL;
1717
1718         if (mmget_not_zero(ctx->mm)) {
1719                 ret = mwriteprotect_range(ctx, uffdio_wp.range.start,
1720                                           uffdio_wp.range.len, mode_wp);
1721                 mmput(ctx->mm);
1722         } else {
1723                 return -ESRCH;
1724         }
1725
1726         if (ret)
1727                 return ret;
1728
1729         if (!mode_wp && !mode_dontwake) {
1730                 range.start = uffdio_wp.range.start;
1731                 range.len = uffdio_wp.range.len;
1732                 wake_userfault(ctx, &range);
1733         }
1734         return ret;
1735 }
1736
1737 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1738 {
1739         __s64 ret;
1740         struct uffdio_continue uffdio_continue;
1741         struct uffdio_continue __user *user_uffdio_continue;
1742         struct userfaultfd_wake_range range;
1743         uffd_flags_t flags = 0;
1744
1745         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1746
1747         ret = -EAGAIN;
1748         if (atomic_read(&ctx->mmap_changing))
1749                 goto out;
1750
1751         ret = -EFAULT;
1752         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1753                            /* don't copy the output fields */
1754                            sizeof(uffdio_continue) - (sizeof(__s64))))
1755                 goto out;
1756
1757         ret = validate_range(ctx->mm, uffdio_continue.range.start,
1758                              uffdio_continue.range.len);
1759         if (ret)
1760                 goto out;
1761
1762         ret = -EINVAL;
1763         if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1764                                      UFFDIO_CONTINUE_MODE_WP))
1765                 goto out;
1766         if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1767                 flags |= MFILL_ATOMIC_WP;
1768
1769         if (mmget_not_zero(ctx->mm)) {
1770                 ret = mfill_atomic_continue(ctx, uffdio_continue.range.start,
1771                                             uffdio_continue.range.len, flags);
1772                 mmput(ctx->mm);
1773         } else {
1774                 return -ESRCH;
1775         }
1776
1777         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1778                 return -EFAULT;
1779         if (ret < 0)
1780                 goto out;
1781
1782         /* len == 0 would wake all */
1783         BUG_ON(!ret);
1784         range.len = ret;
1785         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1786                 range.start = uffdio_continue.range.start;
1787                 wake_userfault(ctx, &range);
1788         }
1789         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1790
1791 out:
1792         return ret;
1793 }
1794
1795 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1796 {
1797         __s64 ret;
1798         struct uffdio_poison uffdio_poison;
1799         struct uffdio_poison __user *user_uffdio_poison;
1800         struct userfaultfd_wake_range range;
1801
1802         user_uffdio_poison = (struct uffdio_poison __user *)arg;
1803
1804         ret = -EAGAIN;
1805         if (atomic_read(&ctx->mmap_changing))
1806                 goto out;
1807
1808         ret = -EFAULT;
1809         if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1810                            /* don't copy the output fields */
1811                            sizeof(uffdio_poison) - (sizeof(__s64))))
1812                 goto out;
1813
1814         ret = validate_range(ctx->mm, uffdio_poison.range.start,
1815                              uffdio_poison.range.len);
1816         if (ret)
1817                 goto out;
1818
1819         ret = -EINVAL;
1820         if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1821                 goto out;
1822
1823         if (mmget_not_zero(ctx->mm)) {
1824                 ret = mfill_atomic_poison(ctx, uffdio_poison.range.start,
1825                                           uffdio_poison.range.len, 0);
1826                 mmput(ctx->mm);
1827         } else {
1828                 return -ESRCH;
1829         }
1830
1831         if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1832                 return -EFAULT;
1833         if (ret < 0)
1834                 goto out;
1835
1836         /* len == 0 would wake all */
1837         BUG_ON(!ret);
1838         range.len = ret;
1839         if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1840                 range.start = uffdio_poison.range.start;
1841                 wake_userfault(ctx, &range);
1842         }
1843         ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1844
1845 out:
1846         return ret;
1847 }
1848
1849 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1850 {
1851         return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1852 }
1853
1854 static inline unsigned int uffd_ctx_features(__u64 user_features)
1855 {
1856         /*
1857          * For the current set of features the bits just coincide. Set
1858          * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1859          */
1860         return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1861 }
1862
1863 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
1864                             unsigned long arg)
1865 {
1866         __s64 ret;
1867         struct uffdio_move uffdio_move;
1868         struct uffdio_move __user *user_uffdio_move;
1869         struct userfaultfd_wake_range range;
1870         struct mm_struct *mm = ctx->mm;
1871
1872         user_uffdio_move = (struct uffdio_move __user *) arg;
1873
1874         if (atomic_read(&ctx->mmap_changing))
1875                 return -EAGAIN;
1876
1877         if (copy_from_user(&uffdio_move, user_uffdio_move,
1878                            /* don't copy "move" last field */
1879                            sizeof(uffdio_move)-sizeof(__s64)))
1880                 return -EFAULT;
1881
1882         /* Do not allow cross-mm moves. */
1883         if (mm != current->mm)
1884                 return -EINVAL;
1885
1886         ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
1887         if (ret)
1888                 return ret;
1889
1890         ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
1891         if (ret)
1892                 return ret;
1893
1894         if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
1895                                   UFFDIO_MOVE_MODE_DONTWAKE))
1896                 return -EINVAL;
1897
1898         if (mmget_not_zero(mm)) {
1899                 ret = move_pages(ctx, uffdio_move.dst, uffdio_move.src,
1900                                  uffdio_move.len, uffdio_move.mode);
1901                 mmput(mm);
1902         } else {
1903                 return -ESRCH;
1904         }
1905
1906         if (unlikely(put_user(ret, &user_uffdio_move->move)))
1907                 return -EFAULT;
1908         if (ret < 0)
1909                 goto out;
1910
1911         /* len == 0 would wake all */
1912         VM_WARN_ON(!ret);
1913         range.len = ret;
1914         if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
1915                 range.start = uffdio_move.dst;
1916                 wake_userfault(ctx, &range);
1917         }
1918         ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
1919
1920 out:
1921         return ret;
1922 }
1923
1924 /*
1925  * userland asks for a certain API version and we return which bits
1926  * and ioctl commands are implemented in this kernel for such API
1927  * version or -EINVAL if unknown.
1928  */
1929 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1930                            unsigned long arg)
1931 {
1932         struct uffdio_api uffdio_api;
1933         void __user *buf = (void __user *)arg;
1934         unsigned int ctx_features;
1935         int ret;
1936         __u64 features;
1937
1938         ret = -EFAULT;
1939         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1940                 goto out;
1941         features = uffdio_api.features;
1942         ret = -EINVAL;
1943         if (uffdio_api.api != UFFD_API)
1944                 goto err_out;
1945         ret = -EPERM;
1946         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
1947                 goto err_out;
1948
1949         /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
1950         if (features & UFFD_FEATURE_WP_ASYNC)
1951                 features |= UFFD_FEATURE_WP_UNPOPULATED;
1952
1953         /* report all available features and ioctls to userland */
1954         uffdio_api.features = UFFD_API_FEATURES;
1955 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1956         uffdio_api.features &=
1957                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1958 #endif
1959 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1960         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
1961 #endif
1962 #ifndef CONFIG_PTE_MARKER_UFFD_WP
1963         uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
1964         uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
1965         uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
1966 #endif
1967
1968         ret = -EINVAL;
1969         if (features & ~uffdio_api.features)
1970                 goto err_out;
1971
1972         uffdio_api.ioctls = UFFD_API_IOCTLS;
1973         ret = -EFAULT;
1974         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1975                 goto out;
1976
1977         /* only enable the requested features for this uffd context */
1978         ctx_features = uffd_ctx_features(features);
1979         ret = -EINVAL;
1980         if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
1981                 goto err_out;
1982
1983         ret = 0;
1984 out:
1985         return ret;
1986 err_out:
1987         memset(&uffdio_api, 0, sizeof(uffdio_api));
1988         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
1989                 ret = -EFAULT;
1990         goto out;
1991 }
1992
1993 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
1994                               unsigned long arg)
1995 {
1996         int ret = -EINVAL;
1997         struct userfaultfd_ctx *ctx = file->private_data;
1998
1999         if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2000                 return -EINVAL;
2001
2002         switch(cmd) {
2003         case UFFDIO_API:
2004                 ret = userfaultfd_api(ctx, arg);
2005                 break;
2006         case UFFDIO_REGISTER:
2007                 ret = userfaultfd_register(ctx, arg);
2008                 break;
2009         case UFFDIO_UNREGISTER:
2010                 ret = userfaultfd_unregister(ctx, arg);
2011                 break;
2012         case UFFDIO_WAKE:
2013                 ret = userfaultfd_wake(ctx, arg);
2014                 break;
2015         case UFFDIO_COPY:
2016                 ret = userfaultfd_copy(ctx, arg);
2017                 break;
2018         case UFFDIO_ZEROPAGE:
2019                 ret = userfaultfd_zeropage(ctx, arg);
2020                 break;
2021         case UFFDIO_MOVE:
2022                 ret = userfaultfd_move(ctx, arg);
2023                 break;
2024         case UFFDIO_WRITEPROTECT:
2025                 ret = userfaultfd_writeprotect(ctx, arg);
2026                 break;
2027         case UFFDIO_CONTINUE:
2028                 ret = userfaultfd_continue(ctx, arg);
2029                 break;
2030         case UFFDIO_POISON:
2031                 ret = userfaultfd_poison(ctx, arg);
2032                 break;
2033         }
2034         return ret;
2035 }
2036
2037 #ifdef CONFIG_PROC_FS
2038 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2039 {
2040         struct userfaultfd_ctx *ctx = f->private_data;
2041         wait_queue_entry_t *wq;
2042         unsigned long pending = 0, total = 0;
2043
2044         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2045         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2046                 pending++;
2047                 total++;
2048         }
2049         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2050                 total++;
2051         }
2052         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2053
2054         /*
2055          * If more protocols will be added, there will be all shown
2056          * separated by a space. Like this:
2057          *      protocols: aa:... bb:...
2058          */
2059         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2060                    pending, total, UFFD_API, ctx->features,
2061                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2062 }
2063 #endif
2064
2065 static const struct file_operations userfaultfd_fops = {
2066 #ifdef CONFIG_PROC_FS
2067         .show_fdinfo    = userfaultfd_show_fdinfo,
2068 #endif
2069         .release        = userfaultfd_release,
2070         .poll           = userfaultfd_poll,
2071         .read_iter      = userfaultfd_read_iter,
2072         .unlocked_ioctl = userfaultfd_ioctl,
2073         .compat_ioctl   = compat_ptr_ioctl,
2074         .llseek         = noop_llseek,
2075 };
2076
2077 static void init_once_userfaultfd_ctx(void *mem)
2078 {
2079         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2080
2081         init_waitqueue_head(&ctx->fault_pending_wqh);
2082         init_waitqueue_head(&ctx->fault_wqh);
2083         init_waitqueue_head(&ctx->event_wqh);
2084         init_waitqueue_head(&ctx->fd_wqh);
2085         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2086 }
2087
2088 static int new_userfaultfd(int flags)
2089 {
2090         struct userfaultfd_ctx *ctx;
2091         struct file *file;
2092         int fd;
2093
2094         BUG_ON(!current->mm);
2095
2096         /* Check the UFFD_* constants for consistency.  */
2097         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2098         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2099         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2100
2101         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2102                 return -EINVAL;
2103
2104         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2105         if (!ctx)
2106                 return -ENOMEM;
2107
2108         refcount_set(&ctx->refcount, 1);
2109         ctx->flags = flags;
2110         ctx->features = 0;
2111         ctx->released = false;
2112         init_rwsem(&ctx->map_changing_lock);
2113         atomic_set(&ctx->mmap_changing, 0);
2114         ctx->mm = current->mm;
2115
2116         fd = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
2117         if (fd < 0)
2118                 goto err_out;
2119
2120         /* Create a new inode so that the LSM can block the creation.  */
2121         file = anon_inode_create_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
2122                         O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2123         if (IS_ERR(file)) {
2124                 put_unused_fd(fd);
2125                 fd = PTR_ERR(file);
2126                 goto err_out;
2127         }
2128         /* prevent the mm struct to be freed */
2129         mmgrab(ctx->mm);
2130         file->f_mode |= FMODE_NOWAIT;
2131         fd_install(fd, file);
2132         return fd;
2133 err_out:
2134         kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2135         return fd;
2136 }
2137
2138 static inline bool userfaultfd_syscall_allowed(int flags)
2139 {
2140         /* Userspace-only page faults are always allowed */
2141         if (flags & UFFD_USER_MODE_ONLY)
2142                 return true;
2143
2144         /*
2145          * The user is requesting a userfaultfd which can handle kernel faults.
2146          * Privileged users are always allowed to do this.
2147          */
2148         if (capable(CAP_SYS_PTRACE))
2149                 return true;
2150
2151         /* Otherwise, access to kernel fault handling is sysctl controlled. */
2152         return sysctl_unprivileged_userfaultfd;
2153 }
2154
2155 SYSCALL_DEFINE1(userfaultfd, int, flags)
2156 {
2157         if (!userfaultfd_syscall_allowed(flags))
2158                 return -EPERM;
2159
2160         return new_userfaultfd(flags);
2161 }
2162
2163 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2164 {
2165         if (cmd != USERFAULTFD_IOC_NEW)
2166                 return -EINVAL;
2167
2168         return new_userfaultfd(flags);
2169 }
2170
2171 static const struct file_operations userfaultfd_dev_fops = {
2172         .unlocked_ioctl = userfaultfd_dev_ioctl,
2173         .compat_ioctl = userfaultfd_dev_ioctl,
2174         .owner = THIS_MODULE,
2175         .llseek = noop_llseek,
2176 };
2177
2178 static struct miscdevice userfaultfd_misc = {
2179         .minor = MISC_DYNAMIC_MINOR,
2180         .name = "userfaultfd",
2181         .fops = &userfaultfd_dev_fops
2182 };
2183
2184 static int __init userfaultfd_init(void)
2185 {
2186         int ret;
2187
2188         ret = misc_register(&userfaultfd_misc);
2189         if (ret)
2190                 return ret;
2191
2192         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2193                                                 sizeof(struct userfaultfd_ctx),
2194                                                 0,
2195                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2196                                                 init_once_userfaultfd_ctx);
2197 #ifdef CONFIG_SYSCTL
2198         register_sysctl_init("vm", vm_userfaultfd_table);
2199 #endif
2200         return 0;
2201 }
2202 __initcall(userfaultfd_init);