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 #include <linux/swapops.h>
  33 #include <linux/miscdevice.h>
  34
  35 static int sysctl_unprivileged_userfaultfd __read_mostly;
  36
  37 #ifdef CONFIG_SYSCTL
  38 static struct ctl_table vm_userfaultfd_table[] = {
  39         {
  40                 .procname       = "unprivileged_userfaultfd",
  41                 .data           = &sysctl_unprivileged_userfaultfd,
  42                 .maxlen         = sizeof(sysctl_unprivileged_userfaultfd),
  43                 .mode           = 0644,
  44                 .proc_handler   = proc_dointvec_minmax,
  45                 .extra1         = SYSCTL_ZERO,
  46                 .extra2         = SYSCTL_ONE,
  47         },
  48         { }
  49 };
  50 #endif
  51
  52 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
  53
  54 /*
  55  * Start with fault_pending_wqh and fault_wqh so they're more likely
  56  * to be in the same cacheline.
  57  *
  58  * Locking order:
  59  *      fd_wqh.lock
  60  *              fault_pending_wqh.lock
  61  *                      fault_wqh.lock
  62  *              event_wqh.lock
  63  *
  64  * To avoid deadlocks, IRQs must be disabled when taking any of the above locks,
  65  * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's
  66  * also taken in IRQ context.
  67  */
  68 struct userfaultfd_ctx {
  69         /* waitqueue head for the pending (i.e. not read) userfaults */
  70         wait_queue_head_t fault_pending_wqh;
  71         /* waitqueue head for the userfaults */
  72         wait_queue_head_t fault_wqh;
  73         /* waitqueue head for the pseudo fd to wakeup poll/read */
  74         wait_queue_head_t fd_wqh;
  75         /* waitqueue head for events */
  76         wait_queue_head_t event_wqh;
  77         /* a refile sequence protected by fault_pending_wqh lock */
  78         seqcount_spinlock_t refile_seq;
  79         /* pseudo fd refcounting */
  80         refcount_t refcount;
  81         /* userfaultfd syscall flags */
  82         unsigned int flags;
  83         /* features requested from the userspace */
  84         unsigned int features;
  85         /* released */
  86         bool released;
  87         /* memory mappings are changing because of non-cooperative event */
  88         atomic_t mmap_changing;
  89         /* mm with one ore more vmas attached to this userfaultfd_ctx */
  90         struct mm_struct *mm;
  91 };
  92
  93 struct userfaultfd_fork_ctx {
  94         struct userfaultfd_ctx *orig;
  95         struct userfaultfd_ctx *new;
  96         struct list_head list;
  97 };
  98
  99 struct userfaultfd_unmap_ctx {
 100         struct userfaultfd_ctx *ctx;
 101         unsigned long start;
 102         unsigned long end;
 103         struct list_head list;
 104 };
 105
 106 struct userfaultfd_wait_queue {
 107         struct uffd_msg msg;
 108         wait_queue_entry_t wq;
 109         struct userfaultfd_ctx *ctx;
 110         bool waken;
 111 };
 112
 113 struct userfaultfd_wake_range {
 114         unsigned long start;
 115         unsigned long len;
 116 };
 117
 118 /* internal indication that UFFD_API ioctl was successfully executed */
 119 #define UFFD_FEATURE_INITIALIZED                (1u << 31)
 120
 121 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
 122 {
 123         return ctx->features & UFFD_FEATURE_INITIALIZED;
 124 }
 125
 126 /*
 127  * Whether WP_UNPOPULATED is enabled on the uffd context.  It is only
 128  * meaningful when userfaultfd_wp()==true on the vma and when it's
 129  * anonymous.
 130  */
 131 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
 132 {
 133         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 134
 135         if (!ctx)
 136                 return false;
 137
 138         return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
 139 }
 140
 141 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
 142                                      vm_flags_t flags)
 143 {
 144         const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
 145
 146         vm_flags_reset(vma, flags);
 147         /*
 148          * For shared mappings, we want to enable writenotify while
 149          * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
 150          * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
 151          */
 152         if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
 153                 vma_set_page_prot(vma);
 154 }
 155
 156 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 157                                      int wake_flags, void *key)
 158 {
 159         struct userfaultfd_wake_range *range = key;
 160         int ret;
 161         struct userfaultfd_wait_queue *uwq;
 162         unsigned long start, len;
 163
 164         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 165         ret = 0;
 166         /* len == 0 means wake all */
 167         start = range->start;
 168         len = range->len;
 169         if (len && (start > uwq->msg.arg.pagefault.address ||
 170                     start + len <= uwq->msg.arg.pagefault.address))
 171                 goto out;
 172         WRITE_ONCE(uwq->waken, true);
 173         /*
 174          * The Program-Order guarantees provided by the scheduler
 175          * ensure uwq->waken is visible before the task is woken.
 176          */
 177         ret = wake_up_state(wq->private, mode);
 178         if (ret) {
 179                 /*
 180                  * Wake only once, autoremove behavior.
 181                  *
 182                  * After the effect of list_del_init is visible to the other
 183                  * CPUs, the waitqueue may disappear from under us, see the
 184                  * !list_empty_careful() in handle_userfault().
 185                  *
 186                  * try_to_wake_up() has an implicit smp_mb(), and the
 187                  * wq->private is read before calling the extern function
 188                  * "wake_up_state" (which in turns calls try_to_wake_up).
 189                  */
 190                 list_del_init(&wq->entry);
 191         }
 192 out:
 193         return ret;
 194 }
 195
 196 /**
 197  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 198  * context.
 199  * @ctx: [in] Pointer to the userfaultfd context.
 200  */
 201 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 202 {
 203         refcount_inc(&ctx->refcount);
 204 }
 205
 206 /**
 207  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 208  * context.
 209  * @ctx: [in] Pointer to userfaultfd context.
 210  *
 211  * The userfaultfd context reference must have been previously acquired either
 212  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 213  */
 214 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 215 {
 216         if (refcount_dec_and_test(&ctx->refcount)) {
 217                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 218                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 219                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 220                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 221                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 222                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 223                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 224                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 225                 mmdrop(ctx->mm);
 226                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 227         }
 228 }
 229
 230 static inline void msg_init(struct uffd_msg *msg)
 231 {
 232         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 233         /*
 234          * Must use memset to zero out the paddings or kernel data is
 235          * leaked to userland.
 236          */
 237         memset(msg, 0, sizeof(struct uffd_msg));
 238 }
 239
 240 static inline struct uffd_msg userfault_msg(unsigned long address,
 241                                             unsigned long real_address,
 242                                             unsigned int flags,
 243                                             unsigned long reason,
 244                                             unsigned int features)
 245 {
 246         struct uffd_msg msg;
 247
 248         msg_init(&msg);
 249         msg.event = UFFD_EVENT_PAGEFAULT;
 250
 251         msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
 252                                     real_address : address;
 253
 254         /*
 255          * These flags indicate why the userfault occurred:
 256          * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
 257          * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
 258          * - Neither of these flags being set indicates a MISSING fault.
 259          *
 260          * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
 261          * fault. Otherwise, it was a read fault.
 262          */
 263         if (flags & FAULT_FLAG_WRITE)
 264                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 265         if (reason & VM_UFFD_WP)
 266                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 267         if (reason & VM_UFFD_MINOR)
 268                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
 269         if (features & UFFD_FEATURE_THREAD_ID)
 270                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 271         return msg;
 272 }
 273
 274 #ifdef CONFIG_HUGETLB_PAGE
 275 /*
 276  * Same functionality as userfaultfd_must_wait below with modifications for
 277  * hugepmd ranges.
 278  */
 279 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 280                                          struct vm_area_struct *vma,
 281                                          unsigned long address,
 282                                          unsigned long flags,
 283                                          unsigned long reason)
 284 {
 285         pte_t *ptep, pte;
 286         bool ret = true;
 287
 288         mmap_assert_locked(ctx->mm);
 289
 290         ptep = hugetlb_walk(vma, address, vma_mmu_pagesize(vma));
 291         if (!ptep)
 292                 goto out;
 293
 294         ret = false;
 295         pte = huge_ptep_get(ptep);
 296
 297         /*
 298          * Lockless access: we're in a wait_event so it's ok if it
 299          * changes under us.  PTE markers should be handled the same as none
 300          * ptes here.
 301          */
 302         if (huge_pte_none_mostly(pte))
 303                 ret = true;
 304         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 305                 ret = true;
 306 out:
 307         return ret;
 308 }
 309 #else
 310 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 311                                          struct vm_area_struct *vma,
 312                                          unsigned long address,
 313                                          unsigned long flags,
 314                                          unsigned long reason)
 315 {
 316         return false;   /* should never get here */
 317 }
 318 #endif /* CONFIG_HUGETLB_PAGE */
 319
 320 /*
 321  * Verify the pagetables are still not ok after having reigstered into
 322  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 323  * userfault that has already been resolved, if userfaultfd_read and
 324  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 325  * threads.
 326  */
 327 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 328                                          unsigned long address,
 329                                          unsigned long flags,
 330                                          unsigned long reason)
 331 {
 332         struct mm_struct *mm = ctx->mm;
 333         pgd_t *pgd;
 334         p4d_t *p4d;
 335         pud_t *pud;
 336         pmd_t *pmd, _pmd;
 337         pte_t *pte;
 338         pte_t ptent;
 339         bool ret = true;
 340
 341         mmap_assert_locked(mm);
 342
 343         pgd = pgd_offset(mm, address);
 344         if (!pgd_present(*pgd))
 345                 goto out;
 346         p4d = p4d_offset(pgd, address);
 347         if (!p4d_present(*p4d))
 348                 goto out;
 349         pud = pud_offset(p4d, address);
 350         if (!pud_present(*pud))
 351                 goto out;
 352         pmd = pmd_offset(pud, address);
 353 again:
 354         _pmd = pmdp_get_lockless(pmd);
 355         if (pmd_none(_pmd))
 356                 goto out;
 357
 358         ret = false;
 359         if (!pmd_present(_pmd) || pmd_devmap(_pmd))
 360                 goto out;
 361
 362         if (pmd_trans_huge(_pmd)) {
 363                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 364                         ret = true;
 365                 goto out;
 366         }
 367
 368         pte = pte_offset_map(pmd, address);
 369         if (!pte) {
 370                 ret = true;
 371                 goto again;
 372         }
 373         /*
 374          * Lockless access: we're in a wait_event so it's ok if it
 375          * changes under us.  PTE markers should be handled the same as none
 376          * ptes here.
 377          */
 378         ptent = ptep_get(pte);
 379         if (pte_none_mostly(ptent))
 380                 ret = true;
 381         if (!pte_write(ptent) && (reason & VM_UFFD_WP))
 382                 ret = true;
 383         pte_unmap(pte);
 384
 385 out:
 386         return ret;
 387 }
 388
 389 static inline unsigned int userfaultfd_get_blocking_state(unsigned int flags)
 390 {
 391         if (flags & FAULT_FLAG_INTERRUPTIBLE)
 392                 return TASK_INTERRUPTIBLE;
 393
 394         if (flags & FAULT_FLAG_KILLABLE)
 395                 return TASK_KILLABLE;
 396
 397         return TASK_UNINTERRUPTIBLE;
 398 }
 399
 400 /*
 401  * The locking rules involved in returning VM_FAULT_RETRY depending on
 402  * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and
 403  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
 404  * recommendation in __lock_page_or_retry is not an understatement.
 405  *
 406  * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
 407  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
 408  * not set.
 409  *
 410  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
 411  * set, VM_FAULT_RETRY can still be returned if and only if there are
 412  * fatal_signal_pending()s, and the mmap_lock must be released before
 413  * returning it.
 414  */
 415 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 416 {
 417         struct vm_area_struct *vma = vmf->vma;
 418         struct mm_struct *mm = vma->vm_mm;
 419         struct userfaultfd_ctx *ctx;
 420         struct userfaultfd_wait_queue uwq;
 421         vm_fault_t ret = VM_FAULT_SIGBUS;
 422         bool must_wait;
 423         unsigned int blocking_state;
 424
 425         /*
 426          * We don't do userfault handling for the final child pid update.
 427          *
 428          * We also don't do userfault handling during
 429          * coredumping. hugetlbfs has the special
 430          * follow_hugetlb_page() to skip missing pages in the
 431          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 432          * the no_page_table() helper in follow_page_mask(), but the
 433          * shmem_vm_ops->fault method is invoked even during
 434          * coredumping without mmap_lock and it ends up here.
 435          */
 436         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 437                 goto out;
 438
 439         /*
 440          * Coredumping runs without mmap_lock so we can only check that
 441          * the mmap_lock is held, if PF_DUMPCORE was not set.
 442          */
 443         mmap_assert_locked(mm);
 444
 445         ctx = vma->vm_userfaultfd_ctx.ctx;
 446         if (!ctx)
 447                 goto out;
 448
 449         BUG_ON(ctx->mm != mm);
 450
 451         /* Any unrecognized flag is a bug. */
 452         VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
 453         /* 0 or > 1 flags set is a bug; we expect exactly 1. */
 454         VM_BUG_ON(!reason || (reason & (reason - 1)));
 455
 456         if (ctx->features & UFFD_FEATURE_SIGBUS)
 457                 goto out;
 458         if (!(vmf->flags & FAULT_FLAG_USER) && (ctx->flags & UFFD_USER_MODE_ONLY))
 459                 goto out;
 460
 461         /*
 462          * If it's already released don't get it. This avoids to loop
 463          * in __get_user_pages if userfaultfd_release waits on the
 464          * caller of handle_userfault to release the mmap_lock.
 465          */
 466         if (unlikely(READ_ONCE(ctx->released))) {
 467                 /*
 468                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 469                  * cooperative manager can close the uffd after the
 470                  * last UFFDIO_COPY, without risking to trigger an
 471                  * involuntary SIGBUS if the process was starting the
 472                  * userfaultfd while the userfaultfd was still armed
 473                  * (but after the last UFFDIO_COPY). If the uffd
 474                  * wasn't already closed when the userfault reached
 475                  * this point, that would normally be solved by
 476                  * userfaultfd_must_wait returning 'false'.
 477                  *
 478                  * If we were to return VM_FAULT_SIGBUS here, the non
 479                  * cooperative manager would be instead forced to
 480                  * always call UFFDIO_UNREGISTER before it can safely
 481                  * close the uffd.
 482                  */
 483                 ret = VM_FAULT_NOPAGE;
 484                 goto out;
 485         }
 486
 487         /*
 488          * Check that we can return VM_FAULT_RETRY.
 489          *
 490          * NOTE: it should become possible to return VM_FAULT_RETRY
 491          * even if FAULT_FLAG_TRIED is set without leading to gup()
 492          * -EBUSY failures, if the userfaultfd is to be extended for
 493          * VM_UFFD_WP tracking and we intend to arm the userfault
 494          * without first stopping userland access to the memory. For
 495          * VM_UFFD_MISSING userfaults this is enough for now.
 496          */
 497         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 498                 /*
 499                  * Validate the invariant that nowait must allow retry
 500                  * to be sure not to return SIGBUS erroneously on
 501                  * nowait invocations.
 502                  */
 503                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 504 #ifdef CONFIG_DEBUG_VM
 505                 if (printk_ratelimit()) {
 506                         printk(KERN_WARNING
 507                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 508                                vmf->flags);
 509                         dump_stack();
 510                 }
 511 #endif
 512                 goto out;
 513         }
 514
 515         /*
 516          * Handle nowait, not much to do other than tell it to retry
 517          * and wait.
 518          */
 519         ret = VM_FAULT_RETRY;
 520         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 521                 goto out;
 522
 523         /* take the reference before dropping the mmap_lock */
 524         userfaultfd_ctx_get(ctx);
 525
 526         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 527         uwq.wq.private = current;
 528         uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
 529                                 reason, ctx->features);
 530         uwq.ctx = ctx;
 531         uwq.waken = false;
 532
 533         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 534
 535         /*
 536          * Take the vma lock now, in order to safely call
 537          * userfaultfd_huge_must_wait() later. Since acquiring the
 538          * (sleepable) vma lock can modify the current task state, that
 539          * must be before explicitly calling set_current_state().
 540          */
 541         if (is_vm_hugetlb_page(vma))
 542                 hugetlb_vma_lock_read(vma);
 543
 544         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 545         /*
 546          * After the __add_wait_queue the uwq is visible to userland
 547          * through poll/read().
 548          */
 549         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 550         /*
 551          * The smp_mb() after __set_current_state prevents the reads
 552          * following the spin_unlock to happen before the list_add in
 553          * __add_wait_queue.
 554          */
 555         set_current_state(blocking_state);
 556         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 557
 558         if (!is_vm_hugetlb_page(vma))
 559                 must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
 560                                                   reason);
 561         else
 562                 must_wait = userfaultfd_huge_must_wait(ctx, vma,
 563                                                        vmf->address,
 564                                                        vmf->flags, reason);
 565         if (is_vm_hugetlb_page(vma))
 566                 hugetlb_vma_unlock_read(vma);
 567         mmap_read_unlock(mm);
 568
 569         if (likely(must_wait && !READ_ONCE(ctx->released))) {
 570                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 571                 schedule();
 572         }
 573
 574         __set_current_state(TASK_RUNNING);
 575
 576         /*
 577          * Here we race with the list_del; list_add in
 578          * userfaultfd_ctx_read(), however because we don't ever run
 579          * list_del_init() to refile across the two lists, the prev
 580          * and next pointers will never point to self. list_add also
 581          * would never let any of the two pointers to point to
 582          * self. So list_empty_careful won't risk to see both pointers
 583          * pointing to self at any time during the list refile. The
 584          * only case where list_del_init() is called is the full
 585          * removal in the wake function and there we don't re-list_add
 586          * and it's fine not to block on the spinlock. The uwq on this
 587          * kernel stack can be released after the list_del_init.
 588          */
 589         if (!list_empty_careful(&uwq.wq.entry)) {
 590                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 591                 /*
 592                  * No need of list_del_init(), the uwq on the stack
 593                  * will be freed shortly anyway.
 594                  */
 595                 list_del(&uwq.wq.entry);
 596                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 597         }
 598
 599         /*
 600          * ctx may go away after this if the userfault pseudo fd is
 601          * already released.
 602          */
 603         userfaultfd_ctx_put(ctx);
 604
 605 out:
 606         return ret;
 607 }
 608
 609 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 610                                               struct userfaultfd_wait_queue *ewq)
 611 {
 612         struct userfaultfd_ctx *release_new_ctx;
 613
 614         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 615                 goto out;
 616
 617         ewq->ctx = ctx;
 618         init_waitqueue_entry(&ewq->wq, current);
 619         release_new_ctx = NULL;
 620
 621         spin_lock_irq(&ctx->event_wqh.lock);
 622         /*
 623          * After the __add_wait_queue the uwq is visible to userland
 624          * through poll/read().
 625          */
 626         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 627         for (;;) {
 628                 set_current_state(TASK_KILLABLE);
 629                 if (ewq->msg.event == 0)
 630                         break;
 631                 if (READ_ONCE(ctx->released) ||
 632                     fatal_signal_pending(current)) {
 633                         /*
 634                          * &ewq->wq may be queued in fork_event, but
 635                          * __remove_wait_queue ignores the head
 636                          * parameter. It would be a problem if it
 637                          * didn't.
 638                          */
 639                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 640                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 641                                 struct userfaultfd_ctx *new;
 642
 643                                 new = (struct userfaultfd_ctx *)
 644                                         (unsigned long)
 645                                         ewq->msg.arg.reserved.reserved1;
 646                                 release_new_ctx = new;
 647                         }
 648                         break;
 649                 }
 650
 651                 spin_unlock_irq(&ctx->event_wqh.lock);
 652
 653                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 654                 schedule();
 655
 656                 spin_lock_irq(&ctx->event_wqh.lock);
 657         }
 658         __set_current_state(TASK_RUNNING);
 659         spin_unlock_irq(&ctx->event_wqh.lock);
 660
 661         if (release_new_ctx) {
 662                 struct vm_area_struct *vma;
 663                 struct mm_struct *mm = release_new_ctx->mm;
 664                 VMA_ITERATOR(vmi, mm, 0);
 665
 666                 /* the various vma->vm_userfaultfd_ctx still points to it */
 667                 mmap_write_lock(mm);
 668                 for_each_vma(vmi, vma) {
 669                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 670                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 671                                 userfaultfd_set_vm_flags(vma,
 672                                                          vma->vm_flags & ~__VM_UFFD_FLAGS);
 673                         }
 674                 }
 675                 mmap_write_unlock(mm);
 676
 677                 userfaultfd_ctx_put(release_new_ctx);
 678         }
 679
 680         /*
 681          * ctx may go away after this if the userfault pseudo fd is
 682          * already released.
 683          */
 684 out:
 685         atomic_dec(&ctx->mmap_changing);
 686         VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
 687         userfaultfd_ctx_put(ctx);
 688 }
 689
 690 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 691                                        struct userfaultfd_wait_queue *ewq)
 692 {
 693         ewq->msg.event = 0;
 694         wake_up_locked(&ctx->event_wqh);
 695         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 696 }
 697
 698 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 699 {
 700         struct userfaultfd_ctx *ctx = NULL, *octx;
 701         struct userfaultfd_fork_ctx *fctx;
 702
 703         octx = vma->vm_userfaultfd_ctx.ctx;
 704         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 705                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 706                 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
 707                 return 0;
 708         }
 709
 710         list_for_each_entry(fctx, fcs, list)
 711                 if (fctx->orig == octx) {
 712                         ctx = fctx->new;
 713                         break;
 714                 }
 715
 716         if (!ctx) {
 717                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 718                 if (!fctx)
 719                         return -ENOMEM;
 720
 721                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 722                 if (!ctx) {
 723                         kfree(fctx);
 724                         return -ENOMEM;
 725                 }
 726
 727                 refcount_set(&ctx->refcount, 1);
 728                 ctx->flags = octx->flags;
 729                 ctx->features = octx->features;
 730                 ctx->released = false;
 731                 atomic_set(&ctx->mmap_changing, 0);
 732                 ctx->mm = vma->vm_mm;
 733                 mmgrab(ctx->mm);
 734
 735                 userfaultfd_ctx_get(octx);
 736                 atomic_inc(&octx->mmap_changing);
 737                 fctx->orig = octx;
 738                 fctx->new = ctx;
 739                 list_add_tail(&fctx->list, fcs);
 740         }
 741
 742         vma->vm_userfaultfd_ctx.ctx = ctx;
 743         return 0;
 744 }
 745
 746 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 747 {
 748         struct userfaultfd_ctx *ctx = fctx->orig;
 749         struct userfaultfd_wait_queue ewq;
 750
 751         msg_init(&ewq.msg);
 752
 753         ewq.msg.event = UFFD_EVENT_FORK;
 754         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 755
 756         userfaultfd_event_wait_completion(ctx, &ewq);
 757 }
 758
 759 void dup_userfaultfd_complete(struct list_head *fcs)
 760 {
 761         struct userfaultfd_fork_ctx *fctx, *n;
 762
 763         list_for_each_entry_safe(fctx, n, fcs, list) {
 764                 dup_fctx(fctx);
 765                 list_del(&fctx->list);
 766                 kfree(fctx);
 767         }
 768 }
 769
 770 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 771                              struct vm_userfaultfd_ctx *vm_ctx)
 772 {
 773         struct userfaultfd_ctx *ctx;
 774
 775         ctx = vma->vm_userfaultfd_ctx.ctx;
 776
 777         if (!ctx)
 778                 return;
 779
 780         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 781                 vm_ctx->ctx = ctx;
 782                 userfaultfd_ctx_get(ctx);
 783                 atomic_inc(&ctx->mmap_changing);
 784         } else {
 785                 /* Drop uffd context if remap feature not enabled */
 786                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 787                 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
 788         }
 789 }
 790
 791 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 792                                  unsigned long from, unsigned long to,
 793                                  unsigned long len)
 794 {
 795         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 796         struct userfaultfd_wait_queue ewq;
 797
 798         if (!ctx)
 799                 return;
 800
 801         if (to & ~PAGE_MASK) {
 802                 userfaultfd_ctx_put(ctx);
 803                 return;
 804         }
 805
 806         msg_init(&ewq.msg);
 807
 808         ewq.msg.event = UFFD_EVENT_REMAP;
 809         ewq.msg.arg.remap.from = from;
 810         ewq.msg.arg.remap.to = to;
 811         ewq.msg.arg.remap.len = len;
 812
 813         userfaultfd_event_wait_completion(ctx, &ewq);
 814 }
 815
 816 bool userfaultfd_remove(struct vm_area_struct *vma,
 817                         unsigned long start, unsigned long end)
 818 {
 819         struct mm_struct *mm = vma->vm_mm;
 820         struct userfaultfd_ctx *ctx;
 821         struct userfaultfd_wait_queue ewq;
 822
 823         ctx = vma->vm_userfaultfd_ctx.ctx;
 824         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 825                 return true;
 826
 827         userfaultfd_ctx_get(ctx);
 828         atomic_inc(&ctx->mmap_changing);
 829         mmap_read_unlock(mm);
 830
 831         msg_init(&ewq.msg);
 832
 833         ewq.msg.event = UFFD_EVENT_REMOVE;
 834         ewq.msg.arg.remove.start = start;
 835         ewq.msg.arg.remove.end = end;
 836
 837         userfaultfd_event_wait_completion(ctx, &ewq);
 838
 839         return false;
 840 }
 841
 842 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 843                           unsigned long start, unsigned long end)
 844 {
 845         struct userfaultfd_unmap_ctx *unmap_ctx;
 846
 847         list_for_each_entry(unmap_ctx, unmaps, list)
 848                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 849                     unmap_ctx->end == end)
 850                         return true;
 851
 852         return false;
 853 }
 854
 855 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
 856                            unsigned long end, struct list_head *unmaps)
 857 {
 858         struct userfaultfd_unmap_ctx *unmap_ctx;
 859         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 860
 861         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 862             has_unmap_ctx(ctx, unmaps, start, end))
 863                 return 0;
 864
 865         unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 866         if (!unmap_ctx)
 867                 return -ENOMEM;
 868
 869         userfaultfd_ctx_get(ctx);
 870         atomic_inc(&ctx->mmap_changing);
 871         unmap_ctx->ctx = ctx;
 872         unmap_ctx->start = start;
 873         unmap_ctx->end = end;
 874         list_add_tail(&unmap_ctx->list, unmaps);
 875
 876         return 0;
 877 }
 878
 879 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 880 {
 881         struct userfaultfd_unmap_ctx *ctx, *n;
 882         struct userfaultfd_wait_queue ewq;
 883
 884         list_for_each_entry_safe(ctx, n, uf, list) {
 885                 msg_init(&ewq.msg);
 886
 887                 ewq.msg.event = UFFD_EVENT_UNMAP;
 888                 ewq.msg.arg.remove.start = ctx->start;
 889                 ewq.msg.arg.remove.end = ctx->end;
 890
 891                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 892
 893                 list_del(&ctx->list);
 894                 kfree(ctx);
 895         }
 896 }
 897
 898 static int userfaultfd_release(struct inode *inode, struct file *file)
 899 {
 900         struct userfaultfd_ctx *ctx = file->private_data;
 901         struct mm_struct *mm = ctx->mm;
 902         struct vm_area_struct *vma, *prev;
 903         /* len == 0 means wake all */
 904         struct userfaultfd_wake_range range = { .len = 0, };
 905         unsigned long new_flags;
 906         VMA_ITERATOR(vmi, mm, 0);
 907
 908         WRITE_ONCE(ctx->released, true);
 909
 910         if (!mmget_not_zero(mm))
 911                 goto wakeup;
 912
 913         /*
 914          * Flush page faults out of all CPUs. NOTE: all page faults
 915          * must be retried without returning VM_FAULT_SIGBUS if
 916          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 917          * changes while handle_userfault released the mmap_lock. So
 918          * it's critical that released is set to true (above), before
 919          * taking the mmap_lock for writing.
 920          */
 921         mmap_write_lock(mm);
 922         prev = NULL;
 923         for_each_vma(vmi, vma) {
 924                 cond_resched();
 925                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 926                        !!(vma->vm_flags & __VM_UFFD_FLAGS));
 927                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 928                         prev = vma;
 929                         continue;
 930                 }
 931                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
 932                 prev = vma_merge(&vmi, mm, prev, vma->vm_start, vma->vm_end,
 933                                  new_flags, vma->anon_vma,
 934                                  vma->vm_file, vma->vm_pgoff,
 935                                  vma_policy(vma),
 936                                  NULL_VM_UFFD_CTX, anon_vma_name(vma));
 937                 if (prev) {
 938                         vma = prev;
 939                 } else {
 940                         prev = vma;
 941                 }
 942
 943                 userfaultfd_set_vm_flags(vma, new_flags);
 944                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 945         }
 946         mmap_write_unlock(mm);
 947         mmput(mm);
 948 wakeup:
 949         /*
 950          * After no new page faults can wait on this fault_*wqh, flush
 951          * the last page faults that may have been already waiting on
 952          * the fault_*wqh.
 953          */
 954         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 955         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 956         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 957         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 958
 959         /* Flush pending events that may still wait on event_wqh */
 960         wake_up_all(&ctx->event_wqh);
 961
 962         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 963         userfaultfd_ctx_put(ctx);
 964         return 0;
 965 }
 966
 967 /* fault_pending_wqh.lock must be hold by the caller */
 968 static inline struct userfaultfd_wait_queue *find_userfault_in(
 969                 wait_queue_head_t *wqh)
 970 {
 971         wait_queue_entry_t *wq;
 972         struct userfaultfd_wait_queue *uwq;
 973
 974         lockdep_assert_held(&wqh->lock);
 975
 976         uwq = NULL;
 977         if (!waitqueue_active(wqh))
 978                 goto out;
 979         /* walk in reverse to provide FIFO behavior to read userfaults */
 980         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 981         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 982 out:
 983         return uwq;
 984 }
 985
 986 static inline struct userfaultfd_wait_queue *find_userfault(
 987                 struct userfaultfd_ctx *ctx)
 988 {
 989         return find_userfault_in(&ctx->fault_pending_wqh);
 990 }
 991
 992 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 993                 struct userfaultfd_ctx *ctx)
 994 {
 995         return find_userfault_in(&ctx->event_wqh);
 996 }
 997
 998 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 999 {
1000         struct userfaultfd_ctx *ctx = file->private_data;
1001         __poll_t ret;
1002
1003         poll_wait(file, &ctx->fd_wqh, wait);
1004
1005         if (!userfaultfd_is_initialized(ctx))
1006                 return EPOLLERR;
1007
1008         /*
1009          * poll() never guarantees that read won't block.
1010          * userfaults can be waken before they're read().
1011          */
1012         if (unlikely(!(file->f_flags & O_NONBLOCK)))
1013                 return EPOLLERR;
1014         /*
1015          * lockless access to see if there are pending faults
1016          * __pollwait last action is the add_wait_queue but
1017          * the spin_unlock would allow the waitqueue_active to
1018          * pass above the actual list_add inside
1019          * add_wait_queue critical section. So use a full
1020          * memory barrier to serialize the list_add write of
1021          * add_wait_queue() with the waitqueue_active read
1022          * below.
1023          */
1024         ret = 0;
1025         smp_mb();
1026         if (waitqueue_active(&ctx->fault_pending_wqh))
1027                 ret = EPOLLIN;
1028         else if (waitqueue_active(&ctx->event_wqh))
1029                 ret = EPOLLIN;
1030
1031         return ret;
1032 }
1033
1034 static const struct file_operations userfaultfd_fops;
1035
1036 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1037                                   struct inode *inode,
1038                                   struct uffd_msg *msg)
1039 {
1040         int fd;
1041
1042         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
1043                         O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1044         if (fd < 0)
1045                 return fd;
1046
1047         msg->arg.reserved.reserved1 = 0;
1048         msg->arg.fork.ufd = fd;
1049         return 0;
1050 }
1051
1052 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1053                                     struct uffd_msg *msg, struct inode *inode)
1054 {
1055         ssize_t ret;
1056         DECLARE_WAITQUEUE(wait, current);
1057         struct userfaultfd_wait_queue *uwq;
1058         /*
1059          * Handling fork event requires sleeping operations, so
1060          * we drop the event_wqh lock, then do these ops, then
1061          * lock it back and wake up the waiter. While the lock is
1062          * dropped the ewq may go away so we keep track of it
1063          * carefully.
1064          */
1065         LIST_HEAD(fork_event);
1066         struct userfaultfd_ctx *fork_nctx = NULL;
1067
1068         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1069         spin_lock_irq(&ctx->fd_wqh.lock);
1070         __add_wait_queue(&ctx->fd_wqh, &wait);
1071         for (;;) {
1072                 set_current_state(TASK_INTERRUPTIBLE);
1073                 spin_lock(&ctx->fault_pending_wqh.lock);
1074                 uwq = find_userfault(ctx);
1075                 if (uwq) {
1076                         /*
1077                          * Use a seqcount to repeat the lockless check
1078                          * in wake_userfault() to avoid missing
1079                          * wakeups because during the refile both
1080                          * waitqueue could become empty if this is the
1081                          * only userfault.
1082                          */
1083                         write_seqcount_begin(&ctx->refile_seq);
1084
1085                         /*
1086                          * The fault_pending_wqh.lock prevents the uwq
1087                          * to disappear from under us.
1088                          *
1089                          * Refile this userfault from
1090                          * fault_pending_wqh to fault_wqh, it's not
1091                          * pending anymore after we read it.
1092                          *
1093                          * Use list_del() by hand (as
1094                          * userfaultfd_wake_function also uses
1095                          * list_del_init() by hand) to be sure nobody
1096                          * changes __remove_wait_queue() to use
1097                          * list_del_init() in turn breaking the
1098                          * !list_empty_careful() check in
1099                          * handle_userfault(). The uwq->wq.head list
1100                          * must never be empty at any time during the
1101                          * refile, or the waitqueue could disappear
1102                          * from under us. The "wait_queue_head_t"
1103                          * parameter of __remove_wait_queue() is unused
1104                          * anyway.
1105                          */
1106                         list_del(&uwq->wq.entry);
1107                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1108
1109                         write_seqcount_end(&ctx->refile_seq);
1110
1111                         /* careful to always initialize msg if ret == 0 */
1112                         *msg = uwq->msg;
1113                         spin_unlock(&ctx->fault_pending_wqh.lock);
1114                         ret = 0;
1115                         break;
1116                 }
1117                 spin_unlock(&ctx->fault_pending_wqh.lock);
1118
1119                 spin_lock(&ctx->event_wqh.lock);
1120                 uwq = find_userfault_evt(ctx);
1121                 if (uwq) {
1122                         *msg = uwq->msg;
1123
1124                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1125                                 fork_nctx = (struct userfaultfd_ctx *)
1126                                         (unsigned long)
1127                                         uwq->msg.arg.reserved.reserved1;
1128                                 list_move(&uwq->wq.entry, &fork_event);
1129                                 /*
1130                                  * fork_nctx can be freed as soon as
1131                                  * we drop the lock, unless we take a
1132                                  * reference on it.
1133                                  */
1134                                 userfaultfd_ctx_get(fork_nctx);
1135                                 spin_unlock(&ctx->event_wqh.lock);
1136                                 ret = 0;
1137                                 break;
1138                         }
1139
1140                         userfaultfd_event_complete(ctx, uwq);
1141                         spin_unlock(&ctx->event_wqh.lock);
1142                         ret = 0;
1143                         break;
1144                 }
1145                 spin_unlock(&ctx->event_wqh.lock);
1146
1147                 if (signal_pending(current)) {
1148                         ret = -ERESTARTSYS;
1149                         break;
1150                 }
1151                 if (no_wait) {
1152                         ret = -EAGAIN;
1153                         break;
1154                 }
1155                 spin_unlock_irq(&ctx->fd_wqh.lock);
1156                 schedule();
1157                 spin_lock_irq(&ctx->fd_wqh.lock);
1158         }
1159         __remove_wait_queue(&ctx->fd_wqh, &wait);
1160         __set_current_state(TASK_RUNNING);
1161         spin_unlock_irq(&ctx->fd_wqh.lock);
1162
1163         if (!ret && msg->event == UFFD_EVENT_FORK) {
1164                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1165                 spin_lock_irq(&ctx->event_wqh.lock);
1166                 if (!list_empty(&fork_event)) {
1167                         /*
1168                          * The fork thread didn't abort, so we can
1169                          * drop the temporary refcount.
1170                          */
1171                         userfaultfd_ctx_put(fork_nctx);
1172
1173                         uwq = list_first_entry(&fork_event,
1174                                                typeof(*uwq),
1175                                                wq.entry);
1176                         /*
1177                          * If fork_event list wasn't empty and in turn
1178                          * the event wasn't already released by fork
1179                          * (the event is allocated on fork kernel
1180                          * stack), put the event back to its place in
1181                          * the event_wq. fork_event head will be freed
1182                          * as soon as we return so the event cannot
1183                          * stay queued there no matter the current
1184                          * "ret" value.
1185                          */
1186                         list_del(&uwq->wq.entry);
1187                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1188
1189                         /*
1190                          * Leave the event in the waitqueue and report
1191                          * error to userland if we failed to resolve
1192                          * the userfault fork.
1193                          */
1194                         if (likely(!ret))
1195                                 userfaultfd_event_complete(ctx, uwq);
1196                 } else {
1197                         /*
1198                          * Here the fork thread aborted and the
1199                          * refcount from the fork thread on fork_nctx
1200                          * has already been released. We still hold
1201                          * the reference we took before releasing the
1202                          * lock above. If resolve_userfault_fork
1203                          * failed we've to drop it because the
1204                          * fork_nctx has to be freed in such case. If
1205                          * it succeeded we'll hold it because the new
1206                          * uffd references it.
1207                          */
1208                         if (ret)
1209                                 userfaultfd_ctx_put(fork_nctx);
1210                 }
1211                 spin_unlock_irq(&ctx->event_wqh.lock);
1212         }
1213
1214         return ret;
1215 }
1216
1217 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1218                                 size_t count, loff_t *ppos)
1219 {
1220         struct userfaultfd_ctx *ctx = file->private_data;
1221         ssize_t _ret, ret = 0;
1222         struct uffd_msg msg;
1223         int no_wait = file->f_flags & O_NONBLOCK;
1224         struct inode *inode = file_inode(file);
1225
1226         if (!userfaultfd_is_initialized(ctx))
1227                 return -EINVAL;
1228
1229         for (;;) {
1230                 if (count < sizeof(msg))
1231                         return ret ? ret : -EINVAL;
1232                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1233                 if (_ret < 0)
1234                         return ret ? ret : _ret;
1235                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1236                         return ret ? ret : -EFAULT;
1237                 ret += sizeof(msg);
1238                 buf += sizeof(msg);
1239                 count -= sizeof(msg);
1240                 /*
1241                  * Allow to read more than one fault at time but only
1242                  * block if waiting for the very first one.
1243                  */
1244                 no_wait = O_NONBLOCK;
1245         }
1246 }
1247
1248 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1249                              struct userfaultfd_wake_range *range)
1250 {
1251         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1252         /* wake all in the range and autoremove */
1253         if (waitqueue_active(&ctx->fault_pending_wqh))
1254                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1255                                      range);
1256         if (waitqueue_active(&ctx->fault_wqh))
1257                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1258         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1259 }
1260
1261 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1262                                            struct userfaultfd_wake_range *range)
1263 {
1264         unsigned seq;
1265         bool need_wakeup;
1266
1267         /*
1268          * To be sure waitqueue_active() is not reordered by the CPU
1269          * before the pagetable update, use an explicit SMP memory
1270          * barrier here. PT lock release or mmap_read_unlock(mm) still
1271          * have release semantics that can allow the
1272          * waitqueue_active() to be reordered before the pte update.
1273          */
1274         smp_mb();
1275
1276         /*
1277          * Use waitqueue_active because it's very frequent to
1278          * change the address space atomically even if there are no
1279          * userfaults yet. So we take the spinlock only when we're
1280          * sure we've userfaults to wake.
1281          */
1282         do {
1283                 seq = read_seqcount_begin(&ctx->refile_seq);
1284                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1285                         waitqueue_active(&ctx->fault_wqh);
1286                 cond_resched();
1287         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1288         if (need_wakeup)
1289                 __wake_userfault(ctx, range);
1290 }
1291
1292 static __always_inline int validate_range(struct mm_struct *mm,
1293                                           __u64 start, __u64 len)
1294 {
1295         __u64 task_size = mm->task_size;
1296
1297         if (start & ~PAGE_MASK)
1298                 return -EINVAL;
1299         if (len & ~PAGE_MASK)
1300                 return -EINVAL;
1301         if (!len)
1302                 return -EINVAL;
1303         if (start < mmap_min_addr)
1304                 return -EINVAL;
1305         if (start >= task_size)
1306                 return -EINVAL;
1307         if (len > task_size - start)
1308                 return -EINVAL;
1309         return 0;
1310 }
1311
1312 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1313                                 unsigned long arg)
1314 {
1315         struct mm_struct *mm = ctx->mm;
1316         struct vm_area_struct *vma, *prev, *cur;
1317         int ret;
1318         struct uffdio_register uffdio_register;
1319         struct uffdio_register __user *user_uffdio_register;
1320         unsigned long vm_flags, new_flags;
1321         bool found;
1322         bool basic_ioctls;
1323         unsigned long start, end, vma_end;
1324         struct vma_iterator vmi;
1325         pgoff_t pgoff;
1326
1327         user_uffdio_register = (struct uffdio_register __user *) arg;
1328
1329         ret = -EFAULT;
1330         if (copy_from_user(&uffdio_register, user_uffdio_register,
1331                            sizeof(uffdio_register)-sizeof(__u64)))
1332                 goto out;
1333
1334         ret = -EINVAL;
1335         if (!uffdio_register.mode)
1336                 goto out;
1337         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1338                 goto out;
1339         vm_flags = 0;
1340         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1341                 vm_flags |= VM_UFFD_MISSING;
1342         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1343 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1344                 goto out;
1345 #endif
1346                 vm_flags |= VM_UFFD_WP;
1347         }
1348         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1349 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1350                 goto out;
1351 #endif
1352                 vm_flags |= VM_UFFD_MINOR;
1353         }
1354
1355         ret = validate_range(mm, uffdio_register.range.start,
1356                              uffdio_register.range.len);
1357         if (ret)
1358                 goto out;
1359
1360         start = uffdio_register.range.start;
1361         end = start + uffdio_register.range.len;
1362
1363         ret = -ENOMEM;
1364         if (!mmget_not_zero(mm))
1365                 goto out;
1366
1367         ret = -EINVAL;
1368         mmap_write_lock(mm);
1369         vma_iter_init(&vmi, mm, start);
1370         vma = vma_find(&vmi, end);
1371         if (!vma)
1372                 goto out_unlock;
1373
1374         /*
1375          * If the first vma contains huge pages, make sure start address
1376          * is aligned to huge page size.
1377          */
1378         if (is_vm_hugetlb_page(vma)) {
1379                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1380
1381                 if (start & (vma_hpagesize - 1))
1382                         goto out_unlock;
1383         }
1384
1385         /*
1386          * Search for not compatible vmas.
1387          */
1388         found = false;
1389         basic_ioctls = false;
1390         cur = vma;
1391         do {
1392                 cond_resched();
1393
1394                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1395                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1396
1397                 /* check not compatible vmas */
1398                 ret = -EINVAL;
1399                 if (!vma_can_userfault(cur, vm_flags))
1400                         goto out_unlock;
1401
1402                 /*
1403                  * UFFDIO_COPY will fill file holes even without
1404                  * PROT_WRITE. This check enforces that if this is a
1405                  * MAP_SHARED, the process has write permission to the backing
1406                  * file. If VM_MAYWRITE is set it also enforces that on a
1407                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1408                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1409                  */
1410                 ret = -EPERM;
1411                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1412                         goto out_unlock;
1413
1414                 /*
1415                  * If this vma contains ending address, and huge pages
1416                  * check alignment.
1417                  */
1418                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1419                     end > cur->vm_start) {
1420                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1421
1422                         ret = -EINVAL;
1423
1424                         if (end & (vma_hpagesize - 1))
1425                                 goto out_unlock;
1426                 }
1427                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1428                         goto out_unlock;
1429
1430                 /*
1431                  * Check that this vma isn't already owned by a
1432                  * different userfaultfd. We can't allow more than one
1433                  * userfaultfd to own a single vma simultaneously or we
1434                  * wouldn't know which one to deliver the userfaults to.
1435                  */
1436                 ret = -EBUSY;
1437                 if (cur->vm_userfaultfd_ctx.ctx &&
1438                     cur->vm_userfaultfd_ctx.ctx != ctx)
1439                         goto out_unlock;
1440
1441                 /*
1442                  * Note vmas containing huge pages
1443                  */
1444                 if (is_vm_hugetlb_page(cur))
1445                         basic_ioctls = true;
1446
1447                 found = true;
1448         } for_each_vma_range(vmi, cur, end);
1449         BUG_ON(!found);
1450
1451         vma_iter_set(&vmi, start);
1452         prev = vma_prev(&vmi);
1453         if (vma->vm_start < start)
1454                 prev = vma;
1455
1456         ret = 0;
1457         for_each_vma_range(vmi, vma, end) {
1458                 cond_resched();
1459
1460                 BUG_ON(!vma_can_userfault(vma, vm_flags));
1461                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1462                        vma->vm_userfaultfd_ctx.ctx != ctx);
1463                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1464
1465                 /*
1466                  * Nothing to do: this vma is already registered into this
1467                  * userfaultfd and with the right tracking mode too.
1468                  */
1469                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1470                     (vma->vm_flags & vm_flags) == vm_flags)
1471                         goto skip;
1472
1473                 if (vma->vm_start > start)
1474                         start = vma->vm_start;
1475                 vma_end = min(end, vma->vm_end);
1476
1477                 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1478                 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
1479                 prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags,
1480                                  vma->anon_vma, vma->vm_file, pgoff,
1481                                  vma_policy(vma),
1482                                  ((struct vm_userfaultfd_ctx){ ctx }),
1483                                  anon_vma_name(vma));
1484                 if (prev) {
1485                         /* vma_merge() invalidated the mas */
1486                         vma = prev;
1487                         goto next;
1488                 }
1489                 if (vma->vm_start < start) {
1490                         ret = split_vma(&vmi, vma, start, 1);
1491                         if (ret)
1492                                 break;
1493                 }
1494                 if (vma->vm_end > end) {
1495                         ret = split_vma(&vmi, vma, end, 0);
1496                         if (ret)
1497                                 break;
1498                 }
1499         next:
1500                 /*
1501                  * In the vma_merge() successful mprotect-like case 8:
1502                  * the next vma was merged into the current one and
1503                  * the current one has not been updated yet.
1504                  */
1505                 userfaultfd_set_vm_flags(vma, new_flags);
1506                 vma->vm_userfaultfd_ctx.ctx = ctx;
1507
1508                 if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1509                         hugetlb_unshare_all_pmds(vma);
1510
1511         skip:
1512                 prev = vma;
1513                 start = vma->vm_end;
1514         }
1515
1516 out_unlock:
1517         mmap_write_unlock(mm);
1518         mmput(mm);
1519         if (!ret) {
1520                 __u64 ioctls_out;
1521
1522                 ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1523                     UFFD_API_RANGE_IOCTLS;
1524
1525                 /*
1526                  * Declare the WP ioctl only if the WP mode is
1527                  * specified and all checks passed with the range
1528                  */
1529                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1530                         ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1531
1532                 /* CONTINUE ioctl is only supported for MINOR ranges. */
1533                 if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1534                         ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1535
1536                 /*
1537                  * Now that we scanned all vmas we can already tell
1538                  * userland which ioctls methods are guaranteed to
1539                  * succeed on this range.
1540                  */
1541                 if (put_user(ioctls_out, &user_uffdio_register->ioctls))
1542                         ret = -EFAULT;
1543         }
1544 out:
1545         return ret;
1546 }
1547
1548 static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
1549                                   unsigned long arg)
1550 {
1551         struct mm_struct *mm = ctx->mm;
1552         struct vm_area_struct *vma, *prev, *cur;
1553         int ret;
1554         struct uffdio_range uffdio_unregister;
1555         unsigned long new_flags;
1556         bool found;
1557         unsigned long start, end, vma_end;
1558         const void __user *buf = (void __user *)arg;
1559         struct vma_iterator vmi;
1560         pgoff_t pgoff;
1561
1562         ret = -EFAULT;
1563         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1564                 goto out;
1565
1566         ret = validate_range(mm, uffdio_unregister.start,
1567                              uffdio_unregister.len);
1568         if (ret)
1569                 goto out;
1570
1571         start = uffdio_unregister.start;
1572         end = start + uffdio_unregister.len;
1573
1574         ret = -ENOMEM;
1575         if (!mmget_not_zero(mm))
1576                 goto out;
1577
1578         mmap_write_lock(mm);
1579         ret = -EINVAL;
1580         vma_iter_init(&vmi, mm, start);
1581         vma = vma_find(&vmi, end);
1582         if (!vma)
1583                 goto out_unlock;
1584
1585         /*
1586          * If the first vma contains huge pages, make sure start address
1587          * is aligned to huge page size.
1588          */
1589         if (is_vm_hugetlb_page(vma)) {
1590                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1591
1592                 if (start & (vma_hpagesize - 1))
1593                         goto out_unlock;
1594         }
1595
1596         /*
1597          * Search for not compatible vmas.
1598          */
1599         found = false;
1600         cur = vma;
1601         do {
1602                 cond_resched();
1603
1604                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1605                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1606
1607                 /*
1608                  * Check not compatible vmas, not strictly required
1609                  * here as not compatible vmas cannot have an
1610                  * userfaultfd_ctx registered on them, but this
1611                  * provides for more strict behavior to notice
1612                  * unregistration errors.
1613                  */
1614                 if (!vma_can_userfault(cur, cur->vm_flags))
1615                         goto out_unlock;
1616
1617                 found = true;
1618         } for_each_vma_range(vmi, cur, end);
1619         BUG_ON(!found);
1620
1621         vma_iter_set(&vmi, start);
1622         prev = vma_prev(&vmi);
1623         if (vma->vm_start < start)
1624                 prev = vma;
1625
1626         ret = 0;
1627         for_each_vma_range(vmi, vma, end) {
1628                 cond_resched();
1629
1630                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
1631
1632                 /*
1633                  * Nothing to do: this vma is already registered into this
1634                  * userfaultfd and with the right tracking mode too.
1635                  */
1636                 if (!vma->vm_userfaultfd_ctx.ctx)
1637                         goto skip;
1638
1639                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1640
1641                 if (vma->vm_start > start)
1642                         start = vma->vm_start;
1643                 vma_end = min(end, vma->vm_end);
1644
1645                 if (userfaultfd_missing(vma)) {
1646                         /*
1647                          * Wake any concurrent pending userfault while
1648                          * we unregister, so they will not hang
1649                          * permanently and it avoids userland to call
1650                          * UFFDIO_WAKE explicitly.
1651                          */
1652                         struct userfaultfd_wake_range range;
1653                         range.start = start;
1654                         range.len = vma_end - start;
1655                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1656                 }
1657
1658                 /* Reset ptes for the whole vma range if wr-protected */
1659                 if (userfaultfd_wp(vma))
1660                         uffd_wp_range(vma, start, vma_end - start, false);
1661
1662                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1663                 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
1664                 prev = vma_merge(&vmi, mm, prev, start, vma_end, new_flags,
1665                                  vma->anon_vma, vma->vm_file, pgoff,
1666                                  vma_policy(vma),
1667                                  NULL_VM_UFFD_CTX, anon_vma_name(vma));
1668                 if (prev) {
1669                         vma = prev;
1670                         goto next;
1671                 }
1672                 if (vma->vm_start < start) {
1673                         ret = split_vma(&vmi, vma, start, 1);
1674                         if (ret)
1675                                 break;
1676                 }
1677                 if (vma->vm_end > end) {
1678                         ret = split_vma(&vmi, vma, end, 0);
1679                         if (ret)
1680                                 break;
1681                 }
1682         next:
1683                 /*
1684                  * In the vma_merge() successful mprotect-like case 8:
1685                  * the next vma was merged into the current one and
1686                  * the current one has not been updated yet.
1687                  */
1688                 userfaultfd_set_vm_flags(vma, new_flags);
1689                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1690
1691         skip:
1692                 prev = vma;
1693                 start = vma->vm_end;
1694         }
1695
1696 out_unlock:
1697         mmap_write_unlock(mm);
1698         mmput(mm);
1699 out:
1700         return ret;
1701 }
1702
1703 /*
1704  * userfaultfd_wake may be used in combination with the
1705  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1706  */
1707 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1708                             unsigned long arg)
1709 {
1710         int ret;
1711         struct uffdio_range uffdio_wake;
1712         struct userfaultfd_wake_range range;
1713         const void __user *buf = (void __user *)arg;
1714
1715         ret = -EFAULT;
1716         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1717                 goto out;
1718
1719         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1720         if (ret)
1721                 goto out;
1722
1723         range.start = uffdio_wake.start;
1724         range.len = uffdio_wake.len;
1725
1726         /*
1727          * len == 0 means wake all and we don't want to wake all here,
1728          * so check it again to be sure.
1729          */
1730         VM_BUG_ON(!range.len);
1731
1732         wake_userfault(ctx, &range);
1733         ret = 0;
1734
1735 out:
1736         return ret;
1737 }
1738
1739 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1740                             unsigned long arg)
1741 {
1742         __s64 ret;
1743         struct uffdio_copy uffdio_copy;
1744         struct uffdio_copy __user *user_uffdio_copy;
1745         struct userfaultfd_wake_range range;
1746         uffd_flags_t flags = 0;
1747
1748         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1749
1750         ret = -EAGAIN;
1751         if (atomic_read(&ctx->mmap_changing))
1752                 goto out;
1753
1754         ret = -EFAULT;
1755         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1756                            /* don't copy "copy" last field */
1757                            sizeof(uffdio_copy)-sizeof(__s64)))
1758                 goto out;
1759
1760         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1761         if (ret)
1762                 goto out;
1763         /*
1764          * double check for wraparound just in case. copy_from_user()
1765          * will later check uffdio_copy.src + uffdio_copy.len to fit
1766          * in the userland range.
1767          */
1768         ret = -EINVAL;
1769         if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
1770                 goto out;
1771         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1772                 goto out;
1773         if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1774                 flags |= MFILL_ATOMIC_WP;
1775         if (mmget_not_zero(ctx->mm)) {
1776                 ret = mfill_atomic_copy(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1777                                         uffdio_copy.len, &ctx->mmap_changing,
1778                                         flags);
1779                 mmput(ctx->mm);
1780         } else {
1781                 return -ESRCH;
1782         }
1783         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1784                 return -EFAULT;
1785         if (ret < 0)
1786                 goto out;
1787         BUG_ON(!ret);
1788         /* len == 0 would wake all */
1789         range.len = ret;
1790         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1791                 range.start = uffdio_copy.dst;
1792                 wake_userfault(ctx, &range);
1793         }
1794         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1795 out:
1796         return ret;
1797 }
1798
1799 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1800                                 unsigned long arg)
1801 {
1802         __s64 ret;
1803         struct uffdio_zeropage uffdio_zeropage;
1804         struct uffdio_zeropage __user *user_uffdio_zeropage;
1805         struct userfaultfd_wake_range range;
1806
1807         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1808
1809         ret = -EAGAIN;
1810         if (atomic_read(&ctx->mmap_changing))
1811                 goto out;
1812
1813         ret = -EFAULT;
1814         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1815                            /* don't copy "zeropage" last field */
1816                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1817                 goto out;
1818
1819         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1820                              uffdio_zeropage.range.len);
1821         if (ret)
1822                 goto out;
1823         ret = -EINVAL;
1824         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1825                 goto out;
1826
1827         if (mmget_not_zero(ctx->mm)) {
1828                 ret = mfill_atomic_zeropage(ctx->mm, uffdio_zeropage.range.start,
1829                                            uffdio_zeropage.range.len,
1830                                            &ctx->mmap_changing);
1831                 mmput(ctx->mm);
1832         } else {
1833                 return -ESRCH;
1834         }
1835         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1836                 return -EFAULT;
1837         if (ret < 0)
1838                 goto out;
1839         /* len == 0 would wake all */
1840         BUG_ON(!ret);
1841         range.len = ret;
1842         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1843                 range.start = uffdio_zeropage.range.start;
1844                 wake_userfault(ctx, &range);
1845         }
1846         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1847 out:
1848         return ret;
1849 }
1850
1851 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1852                                     unsigned long arg)
1853 {
1854         int ret;
1855         struct uffdio_writeprotect uffdio_wp;
1856         struct uffdio_writeprotect __user *user_uffdio_wp;
1857         struct userfaultfd_wake_range range;
1858         bool mode_wp, mode_dontwake;
1859
1860         if (atomic_read(&ctx->mmap_changing))
1861                 return -EAGAIN;
1862
1863         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1864
1865         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1866                            sizeof(struct uffdio_writeprotect)))
1867                 return -EFAULT;
1868
1869         ret = validate_range(ctx->mm, uffdio_wp.range.start,
1870                              uffdio_wp.range.len);
1871         if (ret)
1872                 return ret;
1873
1874         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1875                                UFFDIO_WRITEPROTECT_MODE_WP))
1876                 return -EINVAL;
1877
1878         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1879         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1880
1881         if (mode_wp && mode_dontwake)
1882                 return -EINVAL;
1883
1884         if (mmget_not_zero(ctx->mm)) {
1885                 ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1886                                           uffdio_wp.range.len, mode_wp,
1887                                           &ctx->mmap_changing);
1888                 mmput(ctx->mm);
1889         } else {
1890                 return -ESRCH;
1891         }
1892
1893         if (ret)
1894                 return ret;
1895
1896         if (!mode_wp && !mode_dontwake) {
1897                 range.start = uffdio_wp.range.start;
1898                 range.len = uffdio_wp.range.len;
1899                 wake_userfault(ctx, &range);
1900         }
1901         return ret;
1902 }
1903
1904 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1905 {
1906         __s64 ret;
1907         struct uffdio_continue uffdio_continue;
1908         struct uffdio_continue __user *user_uffdio_continue;
1909         struct userfaultfd_wake_range range;
1910         uffd_flags_t flags = 0;
1911
1912         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1913
1914         ret = -EAGAIN;
1915         if (atomic_read(&ctx->mmap_changing))
1916                 goto out;
1917
1918         ret = -EFAULT;
1919         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1920                            /* don't copy the output fields */
1921                            sizeof(uffdio_continue) - (sizeof(__s64))))
1922                 goto out;
1923
1924         ret = validate_range(ctx->mm, uffdio_continue.range.start,
1925                              uffdio_continue.range.len);
1926         if (ret)
1927                 goto out;
1928
1929         ret = -EINVAL;
1930         /* double check for wraparound just in case. */
1931         if (uffdio_continue.range.start + uffdio_continue.range.len <=
1932             uffdio_continue.range.start) {
1933                 goto out;
1934         }
1935         if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1936                                      UFFDIO_CONTINUE_MODE_WP))
1937                 goto out;
1938         if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1939                 flags |= MFILL_ATOMIC_WP;
1940
1941         if (mmget_not_zero(ctx->mm)) {
1942                 ret = mfill_atomic_continue(ctx->mm, uffdio_continue.range.start,
1943                                             uffdio_continue.range.len,
1944                                             &ctx->mmap_changing, flags);
1945                 mmput(ctx->mm);
1946         } else {
1947                 return -ESRCH;
1948         }
1949
1950         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1951                 return -EFAULT;
1952         if (ret < 0)
1953                 goto out;
1954
1955         /* len == 0 would wake all */
1956         BUG_ON(!ret);
1957         range.len = ret;
1958         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1959                 range.start = uffdio_continue.range.start;
1960                 wake_userfault(ctx, &range);
1961         }
1962         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1963
1964 out:
1965         return ret;
1966 }
1967
1968 static inline unsigned int uffd_ctx_features(__u64 user_features)
1969 {
1970         /*
1971          * For the current set of features the bits just coincide. Set
1972          * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1973          */
1974         return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1975 }
1976
1977 /*
1978  * userland asks for a certain API version and we return which bits
1979  * and ioctl commands are implemented in this kernel for such API
1980  * version or -EINVAL if unknown.
1981  */
1982 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
1983                            unsigned long arg)
1984 {
1985         struct uffdio_api uffdio_api;
1986         void __user *buf = (void __user *)arg;
1987         unsigned int ctx_features;
1988         int ret;
1989         __u64 features;
1990
1991         ret = -EFAULT;
1992         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
1993                 goto out;
1994         features = uffdio_api.features;
1995         ret = -EINVAL;
1996         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
1997                 goto err_out;
1998         ret = -EPERM;
1999         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2000                 goto err_out;
2001         /* report all available features and ioctls to userland */
2002         uffdio_api.features = UFFD_API_FEATURES;
2003 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2004         uffdio_api.features &=
2005                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2006 #endif
2007 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2008         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2009 #endif
2010 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2011         uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2012         uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2013 #endif
2014         uffdio_api.ioctls = UFFD_API_IOCTLS;
2015         ret = -EFAULT;
2016         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2017                 goto out;
2018
2019         /* only enable the requested features for this uffd context */
2020         ctx_features = uffd_ctx_features(features);
2021         ret = -EINVAL;
2022         if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2023                 goto err_out;
2024
2025         ret = 0;
2026 out:
2027         return ret;
2028 err_out:
2029         memset(&uffdio_api, 0, sizeof(uffdio_api));
2030         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2031                 ret = -EFAULT;
2032         goto out;
2033 }
2034
2035 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2036                               unsigned long arg)
2037 {
2038         int ret = -EINVAL;
2039         struct userfaultfd_ctx *ctx = file->private_data;
2040
2041         if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2042                 return -EINVAL;
2043
2044         switch(cmd) {
2045         case UFFDIO_API:
2046                 ret = userfaultfd_api(ctx, arg);
2047                 break;
2048         case UFFDIO_REGISTER:
2049                 ret = userfaultfd_register(ctx, arg);
2050                 break;
2051         case UFFDIO_UNREGISTER:
2052                 ret = userfaultfd_unregister(ctx, arg);
2053                 break;
2054         case UFFDIO_WAKE:
2055                 ret = userfaultfd_wake(ctx, arg);
2056                 break;
2057         case UFFDIO_COPY:
2058                 ret = userfaultfd_copy(ctx, arg);
2059                 break;
2060         case UFFDIO_ZEROPAGE:
2061                 ret = userfaultfd_zeropage(ctx, arg);
2062                 break;
2063         case UFFDIO_WRITEPROTECT:
2064                 ret = userfaultfd_writeprotect(ctx, arg);
2065                 break;
2066         case UFFDIO_CONTINUE:
2067                 ret = userfaultfd_continue(ctx, arg);
2068                 break;
2069         }
2070         return ret;
2071 }
2072
2073 #ifdef CONFIG_PROC_FS
2074 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2075 {
2076         struct userfaultfd_ctx *ctx = f->private_data;
2077         wait_queue_entry_t *wq;
2078         unsigned long pending = 0, total = 0;
2079
2080         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2081         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2082                 pending++;
2083                 total++;
2084         }
2085         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2086                 total++;
2087         }
2088         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2089
2090         /*
2091          * If more protocols will be added, there will be all shown
2092          * separated by a space. Like this:
2093          *      protocols: aa:... bb:...
2094          */
2095         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2096                    pending, total, UFFD_API, ctx->features,
2097                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2098 }
2099 #endif
2100
2101 static const struct file_operations userfaultfd_fops = {
2102 #ifdef CONFIG_PROC_FS
2103         .show_fdinfo    = userfaultfd_show_fdinfo,
2104 #endif
2105         .release        = userfaultfd_release,
2106         .poll           = userfaultfd_poll,
2107         .read           = userfaultfd_read,
2108         .unlocked_ioctl = userfaultfd_ioctl,
2109         .compat_ioctl   = compat_ptr_ioctl,
2110         .llseek         = noop_llseek,
2111 };
2112
2113 static void init_once_userfaultfd_ctx(void *mem)
2114 {
2115         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2116
2117         init_waitqueue_head(&ctx->fault_pending_wqh);
2118         init_waitqueue_head(&ctx->fault_wqh);
2119         init_waitqueue_head(&ctx->event_wqh);
2120         init_waitqueue_head(&ctx->fd_wqh);
2121         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2122 }
2123
2124 static int new_userfaultfd(int flags)
2125 {
2126         struct userfaultfd_ctx *ctx;
2127         int fd;
2128
2129         BUG_ON(!current->mm);
2130
2131         /* Check the UFFD_* constants for consistency.  */
2132         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2133         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2134         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2135
2136         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2137                 return -EINVAL;
2138
2139         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2140         if (!ctx)
2141                 return -ENOMEM;
2142
2143         refcount_set(&ctx->refcount, 1);
2144         ctx->flags = flags;
2145         ctx->features = 0;
2146         ctx->released = false;
2147         atomic_set(&ctx->mmap_changing, 0);
2148         ctx->mm = current->mm;
2149         /* prevent the mm struct to be freed */
2150         mmgrab(ctx->mm);
2151
2152         fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2153                         O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2154         if (fd < 0) {
2155                 mmdrop(ctx->mm);
2156                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2157         }
2158         return fd;
2159 }
2160
2161 static inline bool userfaultfd_syscall_allowed(int flags)
2162 {
2163         /* Userspace-only page faults are always allowed */
2164         if (flags & UFFD_USER_MODE_ONLY)
2165                 return true;
2166
2167         /*
2168          * The user is requesting a userfaultfd which can handle kernel faults.
2169          * Privileged users are always allowed to do this.
2170          */
2171         if (capable(CAP_SYS_PTRACE))
2172                 return true;
2173
2174         /* Otherwise, access to kernel fault handling is sysctl controlled. */
2175         return sysctl_unprivileged_userfaultfd;
2176 }
2177
2178 SYSCALL_DEFINE1(userfaultfd, int, flags)
2179 {
2180         if (!userfaultfd_syscall_allowed(flags))
2181                 return -EPERM;
2182
2183         return new_userfaultfd(flags);
2184 }
2185
2186 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2187 {
2188         if (cmd != USERFAULTFD_IOC_NEW)
2189                 return -EINVAL;
2190
2191         return new_userfaultfd(flags);
2192 }
2193
2194 static const struct file_operations userfaultfd_dev_fops = {
2195         .unlocked_ioctl = userfaultfd_dev_ioctl,
2196         .compat_ioctl = userfaultfd_dev_ioctl,
2197         .owner = THIS_MODULE,
2198         .llseek = noop_llseek,
2199 };
2200
2201 static struct miscdevice userfaultfd_misc = {
2202         .minor = MISC_DYNAMIC_MINOR,
2203         .name = "userfaultfd",
2204         .fops = &userfaultfd_dev_fops
2205 };
2206
2207 static int __init userfaultfd_init(void)
2208 {
2209         int ret;
2210
2211         ret = misc_register(&userfaultfd_misc);
2212         if (ret)
2213                 return ret;
2214
2215         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2216                                                 sizeof(struct userfaultfd_ctx),
2217                                                 0,
2218                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2219                                                 init_once_userfaultfd_ctx);
2220 #ifdef CONFIG_SYSCTL
2221         register_sysctl_init("vm", vm_userfaultfd_table);
2222 #endif
2223         return 0;
2224 }
2225 __initcall(userfaultfd_init);