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