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