fs/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  fs/userfaultfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   6  *  Copyright (C) 2008-2009 Red Hat, Inc.
   7  *  Copyright (C) 2015  Red Hat, Inc.
   8  *
   9  *  Some part derived from fs/eventfd.c (anon inode setup) and
  10  *  mm/ksm.c (mm hashing).
  11  */
  12
  13 #include <linux/list.h>
  14 #include <linux/hashtable.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/mm.h>
  18 #include <linux/mm_inline.h>
  19 #include <linux/mmu_notifier.h>
  20 #include <linux/poll.h>
  21 #include <linux/slab.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/file.h>
  24 #include <linux/bug.h>
  25 #include <linux/anon_inodes.h>
  26 #include <linux/syscalls.h>
  27 #include <linux/userfaultfd_k.h>
  28 #include <linux/mempolicy.h>
  29 #include <linux/ioctl.h>
  30 #include <linux/security.h>
  31 #include <linux/hugetlb.h>
  32 #include <linux/swapops.h>
  33 #include <linux/miscdevice.h>
  34
  35 static int sysctl_unprivileged_userfaultfd __read_mostly;
  36
  37 #ifdef CONFIG_SYSCTL
  38 static struct ctl_table vm_userfaultfd_table[] = {
  39         {
  40                 .procname       = "unprivileged_userfaultfd",
  41                 .data           = &sysctl_unprivileged_userfaultfd,
  42                 .maxlen         = sizeof(sysctl_unprivileged_userfaultfd),
  43                 .mode           = 0644,
  44                 .proc_handler   = proc_dointvec_minmax,
  45                 .extra1         = SYSCTL_ZERO,
  46                 .extra2         = SYSCTL_ONE,
  47         },
  48 };
  49 #endif
  50
  51 static struct kmem_cache *userfaultfd_ctx_cachep __ro_after_init;
  52
  53 struct userfaultfd_fork_ctx {
  54         struct userfaultfd_ctx *orig;
  55         struct userfaultfd_ctx *new;
  56         struct list_head list;
  57 };
  58
  59 struct userfaultfd_unmap_ctx {
  60         struct userfaultfd_ctx *ctx;
  61         unsigned long start;
  62         unsigned long end;
  63         struct list_head list;
  64 };
  65
  66 struct userfaultfd_wait_queue {
  67         struct uffd_msg msg;
  68         wait_queue_entry_t wq;
  69         struct userfaultfd_ctx *ctx;
  70         bool waken;
  71 };
  72
  73 struct userfaultfd_wake_range {
  74         unsigned long start;
  75         unsigned long len;
  76 };
  77
  78 /* internal indication that UFFD_API ioctl was successfully executed */
  79 #define UFFD_FEATURE_INITIALIZED                (1u << 31)
  80
  81 static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
  82 {
  83         return ctx->features & UFFD_FEATURE_INITIALIZED;
  84 }
  85
  86 static bool userfaultfd_wp_async_ctx(struct userfaultfd_ctx *ctx)
  87 {
  88         return ctx && (ctx->features & UFFD_FEATURE_WP_ASYNC);
  89 }
  90
  91 /*
  92  * Whether WP_UNPOPULATED is enabled on the uffd context.  It is only
  93  * meaningful when userfaultfd_wp()==true on the vma and when it's
  94  * anonymous.
  95  */
  96 bool userfaultfd_wp_unpopulated(struct vm_area_struct *vma)
  97 {
  98         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
  99
 100         if (!ctx)
 101                 return false;
 102
 103         return ctx->features & UFFD_FEATURE_WP_UNPOPULATED;
 104 }
 105
 106 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
 107                                      vm_flags_t flags)
 108 {
 109         const bool uffd_wp_changed = (vma->vm_flags ^ flags) & VM_UFFD_WP;
 110
 111         vm_flags_reset(vma, flags);
 112         /*
 113          * For shared mappings, we want to enable writenotify while
 114          * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
 115          * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
 116          */
 117         if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
 118                 vma_set_page_prot(vma);
 119 }
 120
 121 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 122                                      int wake_flags, void *key)
 123 {
 124         struct userfaultfd_wake_range *range = key;
 125         int ret;
 126         struct userfaultfd_wait_queue *uwq;
 127         unsigned long start, len;
 128
 129         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 130         ret = 0;
 131         /* len == 0 means wake all */
 132         start = range->start;
 133         len = range->len;
 134         if (len && (start > uwq->msg.arg.pagefault.address ||
 135                     start + len <= uwq->msg.arg.pagefault.address))
 136                 goto out;
 137         WRITE_ONCE(uwq->waken, true);
 138         /*
 139          * The Program-Order guarantees provided by the scheduler
 140          * ensure uwq->waken is visible before the task is woken.
 141          */
 142         ret = wake_up_state(wq->private, mode);
 143         if (ret) {
 144                 /*
 145                  * Wake only once, autoremove behavior.
 146                  *
 147                  * After the effect of list_del_init is visible to the other
 148                  * CPUs, the waitqueue may disappear from under us, see the
 149                  * !list_empty_careful() in handle_userfault().
 150                  *
 151                  * try_to_wake_up() has an implicit smp_mb(), and the
 152                  * wq->private is read before calling the extern function
 153                  * "wake_up_state" (which in turns calls try_to_wake_up).
 154                  */
 155                 list_del_init(&wq->entry);
 156         }
 157 out:
 158         return ret;
 159 }
 160
 161 /**
 162  * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd
 163  * context.
 164  * @ctx: [in] Pointer to the userfaultfd context.
 165  */
 166 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
 167 {
 168         refcount_inc(&ctx->refcount);
 169 }
 170
 171 /**
 172  * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd
 173  * context.
 174  * @ctx: [in] Pointer to userfaultfd context.
 175  *
 176  * The userfaultfd context reference must have been previously acquired either
 177  * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget().
 178  */
 179 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
 180 {
 181         if (refcount_dec_and_test(&ctx->refcount)) {
 182                 VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
 183                 VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
 184                 VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
 185                 VM_BUG_ON(waitqueue_active(&ctx->fault_wqh));
 186                 VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock));
 187                 VM_BUG_ON(waitqueue_active(&ctx->event_wqh));
 188                 VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock));
 189                 VM_BUG_ON(waitqueue_active(&ctx->fd_wqh));
 190                 mmdrop(ctx->mm);
 191                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 192         }
 193 }
 194
 195 static inline void msg_init(struct uffd_msg *msg)
 196 {
 197         BUILD_BUG_ON(sizeof(struct uffd_msg) != 32);
 198         /*
 199          * Must use memset to zero out the paddings or kernel data is
 200          * leaked to userland.
 201          */
 202         memset(msg, 0, sizeof(struct uffd_msg));
 203 }
 204
 205 static inline struct uffd_msg userfault_msg(unsigned long address,
 206                                             unsigned long real_address,
 207                                             unsigned int flags,
 208                                             unsigned long reason,
 209                                             unsigned int features)
 210 {
 211         struct uffd_msg msg;
 212
 213         msg_init(&msg);
 214         msg.event = UFFD_EVENT_PAGEFAULT;
 215
 216         msg.arg.pagefault.address = (features & UFFD_FEATURE_EXACT_ADDRESS) ?
 217                                     real_address : address;
 218
 219         /*
 220          * These flags indicate why the userfault occurred:
 221          * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
 222          * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
 223          * - Neither of these flags being set indicates a MISSING fault.
 224          *
 225          * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
 226          * fault. Otherwise, it was a read fault.
 227          */
 228         if (flags & FAULT_FLAG_WRITE)
 229                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
 230         if (reason & VM_UFFD_WP)
 231                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
 232         if (reason & VM_UFFD_MINOR)
 233                 msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
 234         if (features & UFFD_FEATURE_THREAD_ID)
 235                 msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
 236         return msg;
 237 }
 238
 239 #ifdef CONFIG_HUGETLB_PAGE
 240 /*
 241  * Same functionality as userfaultfd_must_wait below with modifications for
 242  * hugepmd ranges.
 243  */
 244 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 245                                               struct vm_fault *vmf,
 246                                               unsigned long reason)
 247 {
 248         struct vm_area_struct *vma = vmf->vma;
 249         pte_t *ptep, pte;
 250         bool ret = true;
 251
 252         assert_fault_locked(vmf);
 253
 254         ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
 255         if (!ptep)
 256                 goto out;
 257
 258         ret = false;
 259         pte = huge_ptep_get(ptep);
 260
 261         /*
 262          * Lockless access: we're in a wait_event so it's ok if it
 263          * changes under us.  PTE markers should be handled the same as none
 264          * ptes here.
 265          */
 266         if (huge_pte_none_mostly(pte))
 267                 ret = true;
 268         if (!huge_pte_write(pte) && (reason & VM_UFFD_WP))
 269                 ret = true;
 270 out:
 271         return ret;
 272 }
 273 #else
 274 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 275                                               struct vm_fault *vmf,
 276                                               unsigned long reason)
 277 {
 278         return false;   /* should never get here */
 279 }
 280 #endif /* CONFIG_HUGETLB_PAGE */
 281
 282 /*
 283  * Verify the pagetables are still not ok after having reigstered into
 284  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
 285  * userfault that has already been resolved, if userfaultfd_read and
 286  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
 287  * threads.
 288  */
 289 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 290                                          struct vm_fault *vmf,
 291                                          unsigned long reason)
 292 {
 293         struct mm_struct *mm = ctx->mm;
 294         unsigned long address = vmf->address;
 295         pgd_t *pgd;
 296         p4d_t *p4d;
 297         pud_t *pud;
 298         pmd_t *pmd, _pmd;
 299         pte_t *pte;
 300         pte_t ptent;
 301         bool ret = true;
 302
 303         assert_fault_locked(vmf);
 304
 305         pgd = pgd_offset(mm, address);
 306         if (!pgd_present(*pgd))
 307                 goto out;
 308         p4d = p4d_offset(pgd, address);
 309         if (!p4d_present(*p4d))
 310                 goto out;
 311         pud = pud_offset(p4d, address);
 312         if (!pud_present(*pud))
 313                 goto out;
 314         pmd = pmd_offset(pud, address);
 315 again:
 316         _pmd = pmdp_get_lockless(pmd);
 317         if (pmd_none(_pmd))
 318                 goto out;
 319
 320         ret = false;
 321         if (!pmd_present(_pmd) || pmd_devmap(_pmd))
 322                 goto out;
 323
 324         if (pmd_trans_huge(_pmd)) {
 325                 if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
 326                         ret = true;
 327                 goto out;
 328         }
 329
 330         pte = pte_offset_map(pmd, address);
 331         if (!pte) {
 332                 ret = true;
 333                 goto again;
 334         }
 335         /*
 336          * Lockless access: we're in a wait_event so it's ok if it
 337          * changes under us.  PTE markers should be handled the same as none
 338          * ptes here.
 339          */
 340         ptent = ptep_get(pte);
 341         if (pte_none_mostly(ptent))
 342                 ret = true;
 343         if (!pte_write(ptent) && (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 vm_area_struct *vma = vmf->vma;
 380         struct mm_struct *mm = vma->vm_mm;
 381         struct userfaultfd_ctx *ctx;
 382         struct userfaultfd_wait_queue uwq;
 383         vm_fault_t ret = VM_FAULT_SIGBUS;
 384         bool must_wait;
 385         unsigned int blocking_state;
 386
 387         /*
 388          * We don't do userfault handling for the final child pid update.
 389          *
 390          * We also don't do userfault handling during
 391          * coredumping. hugetlbfs has the special
 392          * hugetlb_follow_page_mask() to skip missing pages in the
 393          * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
 394          * the no_page_table() helper in follow_page_mask(), but the
 395          * shmem_vm_ops->fault method is invoked even during
 396          * coredumping and it ends up here.
 397          */
 398         if (current->flags & (PF_EXITING|PF_DUMPCORE))
 399                 goto out;
 400
 401         assert_fault_locked(vmf);
 402
 403         ctx = 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) && (ctx->flags & UFFD_USER_MODE_ONLY))
 417                 goto out;
 418
 419         /*
 420          * If it's already released don't get it. This avoids to loop
 421          * in __get_user_pages if userfaultfd_release waits on the
 422          * caller of handle_userfault to release the mmap_lock.
 423          */
 424         if (unlikely(READ_ONCE(ctx->released))) {
 425                 /*
 426                  * Don't return VM_FAULT_SIGBUS in this case, so a non
 427                  * cooperative manager can close the uffd after the
 428                  * last UFFDIO_COPY, without risking to trigger an
 429                  * involuntary SIGBUS if the process was starting the
 430                  * userfaultfd while the userfaultfd was still armed
 431                  * (but after the last UFFDIO_COPY). If the uffd
 432                  * wasn't already closed when the userfault reached
 433                  * this point, that would normally be solved by
 434                  * userfaultfd_must_wait returning 'false'.
 435                  *
 436                  * If we were to return VM_FAULT_SIGBUS here, the non
 437                  * cooperative manager would be instead forced to
 438                  * always call UFFDIO_UNREGISTER before it can safely
 439                  * close the uffd.
 440                  */
 441                 ret = VM_FAULT_NOPAGE;
 442                 goto out;
 443         }
 444
 445         /*
 446          * Check that we can return VM_FAULT_RETRY.
 447          *
 448          * NOTE: it should become possible to return VM_FAULT_RETRY
 449          * even if FAULT_FLAG_TRIED is set without leading to gup()
 450          * -EBUSY failures, if the userfaultfd is to be extended for
 451          * VM_UFFD_WP tracking and we intend to arm the userfault
 452          * without first stopping userland access to the memory. For
 453          * VM_UFFD_MISSING userfaults this is enough for now.
 454          */
 455         if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) {
 456                 /*
 457                  * Validate the invariant that nowait must allow retry
 458                  * to be sure not to return SIGBUS erroneously on
 459                  * nowait invocations.
 460                  */
 461                 BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT);
 462 #ifdef CONFIG_DEBUG_VM
 463                 if (printk_ratelimit()) {
 464                         printk(KERN_WARNING
 465                                "FAULT_FLAG_ALLOW_RETRY missing %x\n",
 466                                vmf->flags);
 467                         dump_stack();
 468                 }
 469 #endif
 470                 goto out;
 471         }
 472
 473         /*
 474          * Handle nowait, not much to do other than tell it to retry
 475          * and wait.
 476          */
 477         ret = VM_FAULT_RETRY;
 478         if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 479                 goto out;
 480
 481         /* take the reference before dropping the mmap_lock */
 482         userfaultfd_ctx_get(ctx);
 483
 484         init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
 485         uwq.wq.private = current;
 486         uwq.msg = userfault_msg(vmf->address, vmf->real_address, vmf->flags,
 487                                 reason, ctx->features);
 488         uwq.ctx = ctx;
 489         uwq.waken = false;
 490
 491         blocking_state = userfaultfd_get_blocking_state(vmf->flags);
 492
 493         /*
 494          * Take the vma lock now, in order to safely call
 495          * userfaultfd_huge_must_wait() later. Since acquiring the
 496          * (sleepable) vma lock can modify the current task state, that
 497          * must be before explicitly calling set_current_state().
 498          */
 499         if (is_vm_hugetlb_page(vma))
 500                 hugetlb_vma_lock_read(vma);
 501
 502         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 503         /*
 504          * After the __add_wait_queue the uwq is visible to userland
 505          * through poll/read().
 506          */
 507         __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq);
 508         /*
 509          * The smp_mb() after __set_current_state prevents the reads
 510          * following the spin_unlock to happen before the list_add in
 511          * __add_wait_queue.
 512          */
 513         set_current_state(blocking_state);
 514         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 515
 516         if (!is_vm_hugetlb_page(vma))
 517                 must_wait = userfaultfd_must_wait(ctx, vmf, reason);
 518         else
 519                 must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
 520         if (is_vm_hugetlb_page(vma))
 521                 hugetlb_vma_unlock_read(vma);
 522         release_fault_lock(vmf);
 523
 524         if (likely(must_wait && !READ_ONCE(ctx->released))) {
 525                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 526                 schedule();
 527         }
 528
 529         __set_current_state(TASK_RUNNING);
 530
 531         /*
 532          * Here we race with the list_del; list_add in
 533          * userfaultfd_ctx_read(), however because we don't ever run
 534          * list_del_init() to refile across the two lists, the prev
 535          * and next pointers will never point to self. list_add also
 536          * would never let any of the two pointers to point to
 537          * self. So list_empty_careful won't risk to see both pointers
 538          * pointing to self at any time during the list refile. The
 539          * only case where list_del_init() is called is the full
 540          * removal in the wake function and there we don't re-list_add
 541          * and it's fine not to block on the spinlock. The uwq on this
 542          * kernel stack can be released after the list_del_init.
 543          */
 544         if (!list_empty_careful(&uwq.wq.entry)) {
 545                 spin_lock_irq(&ctx->fault_pending_wqh.lock);
 546                 /*
 547                  * No need of list_del_init(), the uwq on the stack
 548                  * will be freed shortly anyway.
 549                  */
 550                 list_del(&uwq.wq.entry);
 551                 spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 552         }
 553
 554         /*
 555          * ctx may go away after this if the userfault pseudo fd is
 556          * already released.
 557          */
 558         userfaultfd_ctx_put(ctx);
 559
 560 out:
 561         return ret;
 562 }
 563
 564 static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx,
 565                                               struct userfaultfd_wait_queue *ewq)
 566 {
 567         struct userfaultfd_ctx *release_new_ctx;
 568
 569         if (WARN_ON_ONCE(current->flags & PF_EXITING))
 570                 goto out;
 571
 572         ewq->ctx = ctx;
 573         init_waitqueue_entry(&ewq->wq, current);
 574         release_new_ctx = NULL;
 575
 576         spin_lock_irq(&ctx->event_wqh.lock);
 577         /*
 578          * After the __add_wait_queue the uwq is visible to userland
 579          * through poll/read().
 580          */
 581         __add_wait_queue(&ctx->event_wqh, &ewq->wq);
 582         for (;;) {
 583                 set_current_state(TASK_KILLABLE);
 584                 if (ewq->msg.event == 0)
 585                         break;
 586                 if (READ_ONCE(ctx->released) ||
 587                     fatal_signal_pending(current)) {
 588                         /*
 589                          * &ewq->wq may be queued in fork_event, but
 590                          * __remove_wait_queue ignores the head
 591                          * parameter. It would be a problem if it
 592                          * didn't.
 593                          */
 594                         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 595                         if (ewq->msg.event == UFFD_EVENT_FORK) {
 596                                 struct userfaultfd_ctx *new;
 597
 598                                 new = (struct userfaultfd_ctx *)
 599                                         (unsigned long)
 600                                         ewq->msg.arg.reserved.reserved1;
 601                                 release_new_ctx = new;
 602                         }
 603                         break;
 604                 }
 605
 606                 spin_unlock_irq(&ctx->event_wqh.lock);
 607
 608                 wake_up_poll(&ctx->fd_wqh, EPOLLIN);
 609                 schedule();
 610
 611                 spin_lock_irq(&ctx->event_wqh.lock);
 612         }
 613         __set_current_state(TASK_RUNNING);
 614         spin_unlock_irq(&ctx->event_wqh.lock);
 615
 616         if (release_new_ctx) {
 617                 struct vm_area_struct *vma;
 618                 struct mm_struct *mm = release_new_ctx->mm;
 619                 VMA_ITERATOR(vmi, mm, 0);
 620
 621                 /* the various vma->vm_userfaultfd_ctx still points to it */
 622                 mmap_write_lock(mm);
 623                 for_each_vma(vmi, vma) {
 624                         if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
 625                                 vma_start_write(vma);
 626                                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 627                                 userfaultfd_set_vm_flags(vma,
 628                                                          vma->vm_flags & ~__VM_UFFD_FLAGS);
 629                         }
 630                 }
 631                 mmap_write_unlock(mm);
 632
 633                 userfaultfd_ctx_put(release_new_ctx);
 634         }
 635
 636         /*
 637          * ctx may go away after this if the userfault pseudo fd is
 638          * already released.
 639          */
 640 out:
 641         atomic_dec(&ctx->mmap_changing);
 642         VM_BUG_ON(atomic_read(&ctx->mmap_changing) < 0);
 643         userfaultfd_ctx_put(ctx);
 644 }
 645
 646 static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx,
 647                                        struct userfaultfd_wait_queue *ewq)
 648 {
 649         ewq->msg.event = 0;
 650         wake_up_locked(&ctx->event_wqh);
 651         __remove_wait_queue(&ctx->event_wqh, &ewq->wq);
 652 }
 653
 654 int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 655 {
 656         struct userfaultfd_ctx *ctx = NULL, *octx;
 657         struct userfaultfd_fork_ctx *fctx;
 658
 659         octx = vma->vm_userfaultfd_ctx.ctx;
 660         if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
 661                 vma_start_write(vma);
 662                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 663                 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
 664                 return 0;
 665         }
 666
 667         list_for_each_entry(fctx, fcs, list)
 668                 if (fctx->orig == octx) {
 669                         ctx = fctx->new;
 670                         break;
 671                 }
 672
 673         if (!ctx) {
 674                 fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
 675                 if (!fctx)
 676                         return -ENOMEM;
 677
 678                 ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
 679                 if (!ctx) {
 680                         kfree(fctx);
 681                         return -ENOMEM;
 682                 }
 683
 684                 refcount_set(&ctx->refcount, 1);
 685                 ctx->flags = octx->flags;
 686                 ctx->features = octx->features;
 687                 ctx->released = false;
 688                 init_rwsem(&ctx->map_changing_lock);
 689                 atomic_set(&ctx->mmap_changing, 0);
 690                 ctx->mm = vma->vm_mm;
 691                 mmgrab(ctx->mm);
 692
 693                 userfaultfd_ctx_get(octx);
 694                 down_write(&octx->map_changing_lock);
 695                 atomic_inc(&octx->mmap_changing);
 696                 up_write(&octx->map_changing_lock);
 697                 fctx->orig = octx;
 698                 fctx->new = ctx;
 699                 list_add_tail(&fctx->list, fcs);
 700         }
 701
 702         vma->vm_userfaultfd_ctx.ctx = ctx;
 703         return 0;
 704 }
 705
 706 static void dup_fctx(struct userfaultfd_fork_ctx *fctx)
 707 {
 708         struct userfaultfd_ctx *ctx = fctx->orig;
 709         struct userfaultfd_wait_queue ewq;
 710
 711         msg_init(&ewq.msg);
 712
 713         ewq.msg.event = UFFD_EVENT_FORK;
 714         ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new;
 715
 716         userfaultfd_event_wait_completion(ctx, &ewq);
 717 }
 718
 719 void dup_userfaultfd_complete(struct list_head *fcs)
 720 {
 721         struct userfaultfd_fork_ctx *fctx, *n;
 722
 723         list_for_each_entry_safe(fctx, n, fcs, list) {
 724                 dup_fctx(fctx);
 725                 list_del(&fctx->list);
 726                 kfree(fctx);
 727         }
 728 }
 729
 730 void mremap_userfaultfd_prep(struct vm_area_struct *vma,
 731                              struct vm_userfaultfd_ctx *vm_ctx)
 732 {
 733         struct userfaultfd_ctx *ctx;
 734
 735         ctx = vma->vm_userfaultfd_ctx.ctx;
 736
 737         if (!ctx)
 738                 return;
 739
 740         if (ctx->features & UFFD_FEATURE_EVENT_REMAP) {
 741                 vm_ctx->ctx = ctx;
 742                 userfaultfd_ctx_get(ctx);
 743                 down_write(&ctx->map_changing_lock);
 744                 atomic_inc(&ctx->mmap_changing);
 745                 up_write(&ctx->map_changing_lock);
 746         } else {
 747                 /* Drop uffd context if remap feature not enabled */
 748                 vma_start_write(vma);
 749                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 750                 userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
 751         }
 752 }
 753
 754 void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx,
 755                                  unsigned long from, unsigned long to,
 756                                  unsigned long len)
 757 {
 758         struct userfaultfd_ctx *ctx = vm_ctx->ctx;
 759         struct userfaultfd_wait_queue ewq;
 760
 761         if (!ctx)
 762                 return;
 763
 764         if (to & ~PAGE_MASK) {
 765                 userfaultfd_ctx_put(ctx);
 766                 return;
 767         }
 768
 769         msg_init(&ewq.msg);
 770
 771         ewq.msg.event = UFFD_EVENT_REMAP;
 772         ewq.msg.arg.remap.from = from;
 773         ewq.msg.arg.remap.to = to;
 774         ewq.msg.arg.remap.len = len;
 775
 776         userfaultfd_event_wait_completion(ctx, &ewq);
 777 }
 778
 779 bool userfaultfd_remove(struct vm_area_struct *vma,
 780                         unsigned long start, unsigned long end)
 781 {
 782         struct mm_struct *mm = vma->vm_mm;
 783         struct userfaultfd_ctx *ctx;
 784         struct userfaultfd_wait_queue ewq;
 785
 786         ctx = vma->vm_userfaultfd_ctx.ctx;
 787         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE))
 788                 return true;
 789
 790         userfaultfd_ctx_get(ctx);
 791         down_write(&ctx->map_changing_lock);
 792         atomic_inc(&ctx->mmap_changing);
 793         up_write(&ctx->map_changing_lock);
 794         mmap_read_unlock(mm);
 795
 796         msg_init(&ewq.msg);
 797
 798         ewq.msg.event = UFFD_EVENT_REMOVE;
 799         ewq.msg.arg.remove.start = start;
 800         ewq.msg.arg.remove.end = end;
 801
 802         userfaultfd_event_wait_completion(ctx, &ewq);
 803
 804         return false;
 805 }
 806
 807 static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps,
 808                           unsigned long start, unsigned long end)
 809 {
 810         struct userfaultfd_unmap_ctx *unmap_ctx;
 811
 812         list_for_each_entry(unmap_ctx, unmaps, list)
 813                 if (unmap_ctx->ctx == ctx && unmap_ctx->start == start &&
 814                     unmap_ctx->end == end)
 815                         return true;
 816
 817         return false;
 818 }
 819
 820 int userfaultfd_unmap_prep(struct vm_area_struct *vma, unsigned long start,
 821                            unsigned long end, struct list_head *unmaps)
 822 {
 823         struct userfaultfd_unmap_ctx *unmap_ctx;
 824         struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx;
 825
 826         if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) ||
 827             has_unmap_ctx(ctx, unmaps, start, end))
 828                 return 0;
 829
 830         unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL);
 831         if (!unmap_ctx)
 832                 return -ENOMEM;
 833
 834         userfaultfd_ctx_get(ctx);
 835         down_write(&ctx->map_changing_lock);
 836         atomic_inc(&ctx->mmap_changing);
 837         up_write(&ctx->map_changing_lock);
 838         unmap_ctx->ctx = ctx;
 839         unmap_ctx->start = start;
 840         unmap_ctx->end = end;
 841         list_add_tail(&unmap_ctx->list, unmaps);
 842
 843         return 0;
 844 }
 845
 846 void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf)
 847 {
 848         struct userfaultfd_unmap_ctx *ctx, *n;
 849         struct userfaultfd_wait_queue ewq;
 850
 851         list_for_each_entry_safe(ctx, n, uf, list) {
 852                 msg_init(&ewq.msg);
 853
 854                 ewq.msg.event = UFFD_EVENT_UNMAP;
 855                 ewq.msg.arg.remove.start = ctx->start;
 856                 ewq.msg.arg.remove.end = ctx->end;
 857
 858                 userfaultfd_event_wait_completion(ctx->ctx, &ewq);
 859
 860                 list_del(&ctx->list);
 861                 kfree(ctx);
 862         }
 863 }
 864
 865 static int userfaultfd_release(struct inode *inode, struct file *file)
 866 {
 867         struct userfaultfd_ctx *ctx = file->private_data;
 868         struct mm_struct *mm = ctx->mm;
 869         struct vm_area_struct *vma, *prev;
 870         /* len == 0 means wake all */
 871         struct userfaultfd_wake_range range = { .len = 0, };
 872         unsigned long new_flags;
 873         VMA_ITERATOR(vmi, mm, 0);
 874
 875         WRITE_ONCE(ctx->released, true);
 876
 877         if (!mmget_not_zero(mm))
 878                 goto wakeup;
 879
 880         /*
 881          * Flush page faults out of all CPUs. NOTE: all page faults
 882          * must be retried without returning VM_FAULT_SIGBUS if
 883          * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
 884          * changes while handle_userfault released the mmap_lock. So
 885          * it's critical that released is set to true (above), before
 886          * taking the mmap_lock for writing.
 887          */
 888         mmap_write_lock(mm);
 889         prev = NULL;
 890         for_each_vma(vmi, vma) {
 891                 cond_resched();
 892                 BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
 893                        !!(vma->vm_flags & __VM_UFFD_FLAGS));
 894                 if (vma->vm_userfaultfd_ctx.ctx != ctx) {
 895                         prev = vma;
 896                         continue;
 897                 }
 898                 /* Reset ptes for the whole vma range if wr-protected */
 899                 if (userfaultfd_wp(vma))
 900                         uffd_wp_range(vma, vma->vm_start,
 901                                       vma->vm_end - vma->vm_start, false);
 902                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
 903                 vma = vma_modify_flags_uffd(&vmi, prev, vma, vma->vm_start,
 904                                             vma->vm_end, new_flags,
 905                                             NULL_VM_UFFD_CTX);
 906
 907                 vma_start_write(vma);
 908                 userfaultfd_set_vm_flags(vma, new_flags);
 909                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 910
 911                 prev = vma;
 912         }
 913         mmap_write_unlock(mm);
 914         mmput(mm);
 915 wakeup:
 916         /*
 917          * After no new page faults can wait on this fault_*wqh, flush
 918          * the last page faults that may have been already waiting on
 919          * the fault_*wqh.
 920          */
 921         spin_lock_irq(&ctx->fault_pending_wqh.lock);
 922         __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range);
 923         __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range);
 924         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 925
 926         /* Flush pending events that may still wait on event_wqh */
 927         wake_up_all(&ctx->event_wqh);
 928
 929         wake_up_poll(&ctx->fd_wqh, EPOLLHUP);
 930         userfaultfd_ctx_put(ctx);
 931         return 0;
 932 }
 933
 934 /* fault_pending_wqh.lock must be hold by the caller */
 935 static inline struct userfaultfd_wait_queue *find_userfault_in(
 936                 wait_queue_head_t *wqh)
 937 {
 938         wait_queue_entry_t *wq;
 939         struct userfaultfd_wait_queue *uwq;
 940
 941         lockdep_assert_held(&wqh->lock);
 942
 943         uwq = NULL;
 944         if (!waitqueue_active(wqh))
 945                 goto out;
 946         /* walk in reverse to provide FIFO behavior to read userfaults */
 947         wq = list_last_entry(&wqh->head, typeof(*wq), entry);
 948         uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 949 out:
 950         return uwq;
 951 }
 952
 953 static inline struct userfaultfd_wait_queue *find_userfault(
 954                 struct userfaultfd_ctx *ctx)
 955 {
 956         return find_userfault_in(&ctx->fault_pending_wqh);
 957 }
 958
 959 static inline struct userfaultfd_wait_queue *find_userfault_evt(
 960                 struct userfaultfd_ctx *ctx)
 961 {
 962         return find_userfault_in(&ctx->event_wqh);
 963 }
 964
 965 static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 966 {
 967         struct userfaultfd_ctx *ctx = file->private_data;
 968         __poll_t ret;
 969
 970         poll_wait(file, &ctx->fd_wqh, wait);
 971
 972         if (!userfaultfd_is_initialized(ctx))
 973                 return EPOLLERR;
 974
 975         /*
 976          * poll() never guarantees that read won't block.
 977          * userfaults can be waken before they're read().
 978          */
 979         if (unlikely(!(file->f_flags & O_NONBLOCK)))
 980                 return EPOLLERR;
 981         /*
 982          * lockless access to see if there are pending faults
 983          * __pollwait last action is the add_wait_queue but
 984          * the spin_unlock would allow the waitqueue_active to
 985          * pass above the actual list_add inside
 986          * add_wait_queue critical section. So use a full
 987          * memory barrier to serialize the list_add write of
 988          * add_wait_queue() with the waitqueue_active read
 989          * below.
 990          */
 991         ret = 0;
 992         smp_mb();
 993         if (waitqueue_active(&ctx->fault_pending_wqh))
 994                 ret = EPOLLIN;
 995         else if (waitqueue_active(&ctx->event_wqh))
 996                 ret = EPOLLIN;
 997
 998         return ret;
 999 }
1000
1001 static const struct file_operations userfaultfd_fops;
1002
1003 static int resolve_userfault_fork(struct userfaultfd_ctx *new,
1004                                   struct inode *inode,
1005                                   struct uffd_msg *msg)
1006 {
1007         int fd;
1008
1009         fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, new,
1010                         O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1011         if (fd < 0)
1012                 return fd;
1013
1014         msg->arg.reserved.reserved1 = 0;
1015         msg->arg.fork.ufd = fd;
1016         return 0;
1017 }
1018
1019 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1020                                     struct uffd_msg *msg, struct inode *inode)
1021 {
1022         ssize_t ret;
1023         DECLARE_WAITQUEUE(wait, current);
1024         struct userfaultfd_wait_queue *uwq;
1025         /*
1026          * Handling fork event requires sleeping operations, so
1027          * we drop the event_wqh lock, then do these ops, then
1028          * lock it back and wake up the waiter. While the lock is
1029          * dropped the ewq may go away so we keep track of it
1030          * carefully.
1031          */
1032         LIST_HEAD(fork_event);
1033         struct userfaultfd_ctx *fork_nctx = NULL;
1034
1035         /* always take the fd_wqh lock before the fault_pending_wqh lock */
1036         spin_lock_irq(&ctx->fd_wqh.lock);
1037         __add_wait_queue(&ctx->fd_wqh, &wait);
1038         for (;;) {
1039                 set_current_state(TASK_INTERRUPTIBLE);
1040                 spin_lock(&ctx->fault_pending_wqh.lock);
1041                 uwq = find_userfault(ctx);
1042                 if (uwq) {
1043                         /*
1044                          * Use a seqcount to repeat the lockless check
1045                          * in wake_userfault() to avoid missing
1046                          * wakeups because during the refile both
1047                          * waitqueue could become empty if this is the
1048                          * only userfault.
1049                          */
1050                         write_seqcount_begin(&ctx->refile_seq);
1051
1052                         /*
1053                          * The fault_pending_wqh.lock prevents the uwq
1054                          * to disappear from under us.
1055                          *
1056                          * Refile this userfault from
1057                          * fault_pending_wqh to fault_wqh, it's not
1058                          * pending anymore after we read it.
1059                          *
1060                          * Use list_del() by hand (as
1061                          * userfaultfd_wake_function also uses
1062                          * list_del_init() by hand) to be sure nobody
1063                          * changes __remove_wait_queue() to use
1064                          * list_del_init() in turn breaking the
1065                          * !list_empty_careful() check in
1066                          * handle_userfault(). The uwq->wq.head list
1067                          * must never be empty at any time during the
1068                          * refile, or the waitqueue could disappear
1069                          * from under us. The "wait_queue_head_t"
1070                          * parameter of __remove_wait_queue() is unused
1071                          * anyway.
1072                          */
1073                         list_del(&uwq->wq.entry);
1074                         add_wait_queue(&ctx->fault_wqh, &uwq->wq);
1075
1076                         write_seqcount_end(&ctx->refile_seq);
1077
1078                         /* careful to always initialize msg if ret == 0 */
1079                         *msg = uwq->msg;
1080                         spin_unlock(&ctx->fault_pending_wqh.lock);
1081                         ret = 0;
1082                         break;
1083                 }
1084                 spin_unlock(&ctx->fault_pending_wqh.lock);
1085
1086                 spin_lock(&ctx->event_wqh.lock);
1087                 uwq = find_userfault_evt(ctx);
1088                 if (uwq) {
1089                         *msg = uwq->msg;
1090
1091                         if (uwq->msg.event == UFFD_EVENT_FORK) {
1092                                 fork_nctx = (struct userfaultfd_ctx *)
1093                                         (unsigned long)
1094                                         uwq->msg.arg.reserved.reserved1;
1095                                 list_move(&uwq->wq.entry, &fork_event);
1096                                 /*
1097                                  * fork_nctx can be freed as soon as
1098                                  * we drop the lock, unless we take a
1099                                  * reference on it.
1100                                  */
1101                                 userfaultfd_ctx_get(fork_nctx);
1102                                 spin_unlock(&ctx->event_wqh.lock);
1103                                 ret = 0;
1104                                 break;
1105                         }
1106
1107                         userfaultfd_event_complete(ctx, uwq);
1108                         spin_unlock(&ctx->event_wqh.lock);
1109                         ret = 0;
1110                         break;
1111                 }
1112                 spin_unlock(&ctx->event_wqh.lock);
1113
1114                 if (signal_pending(current)) {
1115                         ret = -ERESTARTSYS;
1116                         break;
1117                 }
1118                 if (no_wait) {
1119                         ret = -EAGAIN;
1120                         break;
1121                 }
1122                 spin_unlock_irq(&ctx->fd_wqh.lock);
1123                 schedule();
1124                 spin_lock_irq(&ctx->fd_wqh.lock);
1125         }
1126         __remove_wait_queue(&ctx->fd_wqh, &wait);
1127         __set_current_state(TASK_RUNNING);
1128         spin_unlock_irq(&ctx->fd_wqh.lock);
1129
1130         if (!ret && msg->event == UFFD_EVENT_FORK) {
1131                 ret = resolve_userfault_fork(fork_nctx, inode, msg);
1132                 spin_lock_irq(&ctx->event_wqh.lock);
1133                 if (!list_empty(&fork_event)) {
1134                         /*
1135                          * The fork thread didn't abort, so we can
1136                          * drop the temporary refcount.
1137                          */
1138                         userfaultfd_ctx_put(fork_nctx);
1139
1140                         uwq = list_first_entry(&fork_event,
1141                                                typeof(*uwq),
1142                                                wq.entry);
1143                         /*
1144                          * If fork_event list wasn't empty and in turn
1145                          * the event wasn't already released by fork
1146                          * (the event is allocated on fork kernel
1147                          * stack), put the event back to its place in
1148                          * the event_wq. fork_event head will be freed
1149                          * as soon as we return so the event cannot
1150                          * stay queued there no matter the current
1151                          * "ret" value.
1152                          */
1153                         list_del(&uwq->wq.entry);
1154                         __add_wait_queue(&ctx->event_wqh, &uwq->wq);
1155
1156                         /*
1157                          * Leave the event in the waitqueue and report
1158                          * error to userland if we failed to resolve
1159                          * the userfault fork.
1160                          */
1161                         if (likely(!ret))
1162                                 userfaultfd_event_complete(ctx, uwq);
1163                 } else {
1164                         /*
1165                          * Here the fork thread aborted and the
1166                          * refcount from the fork thread on fork_nctx
1167                          * has already been released. We still hold
1168                          * the reference we took before releasing the
1169                          * lock above. If resolve_userfault_fork
1170                          * failed we've to drop it because the
1171                          * fork_nctx has to be freed in such case. If
1172                          * it succeeded we'll hold it because the new
1173                          * uffd references it.
1174                          */
1175                         if (ret)
1176                                 userfaultfd_ctx_put(fork_nctx);
1177                 }
1178                 spin_unlock_irq(&ctx->event_wqh.lock);
1179         }
1180
1181         return ret;
1182 }
1183
1184 static ssize_t userfaultfd_read(struct file *file, char __user *buf,
1185                                 size_t count, loff_t *ppos)
1186 {
1187         struct userfaultfd_ctx *ctx = file->private_data;
1188         ssize_t _ret, ret = 0;
1189         struct uffd_msg msg;
1190         int no_wait = file->f_flags & O_NONBLOCK;
1191         struct inode *inode = file_inode(file);
1192
1193         if (!userfaultfd_is_initialized(ctx))
1194                 return -EINVAL;
1195
1196         for (;;) {
1197                 if (count < sizeof(msg))
1198                         return ret ? ret : -EINVAL;
1199                 _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
1200                 if (_ret < 0)
1201                         return ret ? ret : _ret;
1202                 if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
1203                         return ret ? ret : -EFAULT;
1204                 ret += sizeof(msg);
1205                 buf += sizeof(msg);
1206                 count -= sizeof(msg);
1207                 /*
1208                  * Allow to read more than one fault at time but only
1209                  * block if waiting for the very first one.
1210                  */
1211                 no_wait = O_NONBLOCK;
1212         }
1213 }
1214
1215 static void __wake_userfault(struct userfaultfd_ctx *ctx,
1216                              struct userfaultfd_wake_range *range)
1217 {
1218         spin_lock_irq(&ctx->fault_pending_wqh.lock);
1219         /* wake all in the range and autoremove */
1220         if (waitqueue_active(&ctx->fault_pending_wqh))
1221                 __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL,
1222                                      range);
1223         if (waitqueue_active(&ctx->fault_wqh))
1224                 __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range);
1225         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
1226 }
1227
1228 static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx,
1229                                            struct userfaultfd_wake_range *range)
1230 {
1231         unsigned seq;
1232         bool need_wakeup;
1233
1234         /*
1235          * To be sure waitqueue_active() is not reordered by the CPU
1236          * before the pagetable update, use an explicit SMP memory
1237          * barrier here. PT lock release or mmap_read_unlock(mm) still
1238          * have release semantics that can allow the
1239          * waitqueue_active() to be reordered before the pte update.
1240          */
1241         smp_mb();
1242
1243         /*
1244          * Use waitqueue_active because it's very frequent to
1245          * change the address space atomically even if there are no
1246          * userfaults yet. So we take the spinlock only when we're
1247          * sure we've userfaults to wake.
1248          */
1249         do {
1250                 seq = read_seqcount_begin(&ctx->refile_seq);
1251                 need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) ||
1252                         waitqueue_active(&ctx->fault_wqh);
1253                 cond_resched();
1254         } while (read_seqcount_retry(&ctx->refile_seq, seq));
1255         if (need_wakeup)
1256                 __wake_userfault(ctx, range);
1257 }
1258
1259 static __always_inline int validate_unaligned_range(
1260         struct mm_struct *mm, __u64 start, __u64 len)
1261 {
1262         __u64 task_size = mm->task_size;
1263
1264         if (len & ~PAGE_MASK)
1265                 return -EINVAL;
1266         if (!len)
1267                 return -EINVAL;
1268         if (start < mmap_min_addr)
1269                 return -EINVAL;
1270         if (start >= task_size)
1271                 return -EINVAL;
1272         if (len > task_size - start)
1273                 return -EINVAL;
1274         if (start + len <= start)
1275                 return -EINVAL;
1276         return 0;
1277 }
1278
1279 static __always_inline int validate_range(struct mm_struct *mm,
1280                                           __u64 start, __u64 len)
1281 {
1282         if (start & ~PAGE_MASK)
1283                 return -EINVAL;
1284
1285         return validate_unaligned_range(mm, start, len);
1286 }
1287
1288 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
1289                                 unsigned long arg)
1290 {
1291         struct mm_struct *mm = ctx->mm;
1292         struct vm_area_struct *vma, *prev, *cur;
1293         int ret;
1294         struct uffdio_register uffdio_register;
1295         struct uffdio_register __user *user_uffdio_register;
1296         unsigned long vm_flags, new_flags;
1297         bool found;
1298         bool basic_ioctls;
1299         unsigned long start, end, vma_end;
1300         struct vma_iterator vmi;
1301         bool wp_async = userfaultfd_wp_async_ctx(ctx);
1302
1303         user_uffdio_register = (struct uffdio_register __user *) arg;
1304
1305         ret = -EFAULT;
1306         if (copy_from_user(&uffdio_register, user_uffdio_register,
1307                            sizeof(uffdio_register)-sizeof(__u64)))
1308                 goto out;
1309
1310         ret = -EINVAL;
1311         if (!uffdio_register.mode)
1312                 goto out;
1313         if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
1314                 goto out;
1315         vm_flags = 0;
1316         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
1317                 vm_flags |= VM_UFFD_MISSING;
1318         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1319 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
1320                 goto out;
1321 #endif
1322                 vm_flags |= VM_UFFD_WP;
1323         }
1324         if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1325 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1326                 goto out;
1327 #endif
1328                 vm_flags |= VM_UFFD_MINOR;
1329         }
1330
1331         ret = validate_range(mm, uffdio_register.range.start,
1332                              uffdio_register.range.len);
1333         if (ret)
1334                 goto out;
1335
1336         start = uffdio_register.range.start;
1337         end = start + uffdio_register.range.len;
1338
1339         ret = -ENOMEM;
1340         if (!mmget_not_zero(mm))
1341                 goto out;
1342
1343         ret = -EINVAL;
1344         mmap_write_lock(mm);
1345         vma_iter_init(&vmi, mm, start);
1346         vma = vma_find(&vmi, end);
1347         if (!vma)
1348                 goto out_unlock;
1349
1350         /*
1351          * If the first vma contains huge pages, make sure start address
1352          * is aligned to huge page size.
1353          */
1354         if (is_vm_hugetlb_page(vma)) {
1355                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1356
1357                 if (start & (vma_hpagesize - 1))
1358                         goto out_unlock;
1359         }
1360
1361         /*
1362          * Search for not compatible vmas.
1363          */
1364         found = false;
1365         basic_ioctls = false;
1366         cur = vma;
1367         do {
1368                 cond_resched();
1369
1370                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1371                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1372
1373                 /* check not compatible vmas */
1374                 ret = -EINVAL;
1375                 if (!vma_can_userfault(cur, vm_flags, wp_async))
1376                         goto out_unlock;
1377
1378                 /*
1379                  * UFFDIO_COPY will fill file holes even without
1380                  * PROT_WRITE. This check enforces that if this is a
1381                  * MAP_SHARED, the process has write permission to the backing
1382                  * file. If VM_MAYWRITE is set it also enforces that on a
1383                  * MAP_SHARED vma: there is no F_WRITE_SEAL and no further
1384                  * F_WRITE_SEAL can be taken until the vma is destroyed.
1385                  */
1386                 ret = -EPERM;
1387                 if (unlikely(!(cur->vm_flags & VM_MAYWRITE)))
1388                         goto out_unlock;
1389
1390                 /*
1391                  * If this vma contains ending address, and huge pages
1392                  * check alignment.
1393                  */
1394                 if (is_vm_hugetlb_page(cur) && end <= cur->vm_end &&
1395                     end > cur->vm_start) {
1396                         unsigned long vma_hpagesize = vma_kernel_pagesize(cur);
1397
1398                         ret = -EINVAL;
1399
1400                         if (end & (vma_hpagesize - 1))
1401                                 goto out_unlock;
1402                 }
1403                 if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1404                         goto out_unlock;
1405
1406                 /*
1407                  * Check that this vma isn't already owned by a
1408                  * different userfaultfd. We can't allow more than one
1409                  * userfaultfd to own a single vma simultaneously or we
1410                  * wouldn't know which one to deliver the userfaults to.
1411                  */
1412                 ret = -EBUSY;
1413                 if (cur->vm_userfaultfd_ctx.ctx &&
1414                     cur->vm_userfaultfd_ctx.ctx != ctx)
1415                         goto out_unlock;
1416
1417                 /*
1418                  * Note vmas containing huge pages
1419                  */
1420                 if (is_vm_hugetlb_page(cur))
1421                         basic_ioctls = true;
1422
1423                 found = true;
1424         } for_each_vma_range(vmi, cur, end);
1425         BUG_ON(!found);
1426
1427         vma_iter_set(&vmi, start);
1428         prev = vma_prev(&vmi);
1429         if (vma->vm_start < start)
1430                 prev = vma;
1431
1432         ret = 0;
1433         for_each_vma_range(vmi, vma, end) {
1434                 cond_resched();
1435
1436                 BUG_ON(!vma_can_userfault(vma, vm_flags, wp_async));
1437                 BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
1438                        vma->vm_userfaultfd_ctx.ctx != ctx);
1439                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1440
1441                 /*
1442                  * Nothing to do: this vma is already registered into this
1443                  * userfaultfd and with the right tracking mode too.
1444                  */
1445                 if (vma->vm_userfaultfd_ctx.ctx == ctx &&
1446                     (vma->vm_flags & vm_flags) == vm_flags)
1447                         goto skip;
1448
1449                 if (vma->vm_start > start)
1450                         start = vma->vm_start;
1451                 vma_end = min(end, vma->vm_end);
1452
1453                 new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
1454                 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1455                                             new_flags,
1456                                             (struct vm_userfaultfd_ctx){ctx});
1457                 if (IS_ERR(vma)) {
1458                         ret = PTR_ERR(vma);
1459                         break;
1460                 }
1461
1462                 /*
1463                  * In the vma_merge() successful mprotect-like case 8:
1464                  * the next vma was merged into the current one and
1465                  * the current one has not been updated yet.
1466                  */
1467                 vma_start_write(vma);
1468                 userfaultfd_set_vm_flags(vma, 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         }
1478
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         struct vma_iterator vmi;
1523         bool wp_async = userfaultfd_wp_async_ctx(ctx);
1524
1525         ret = -EFAULT;
1526         if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
1527                 goto out;
1528
1529         ret = validate_range(mm, uffdio_unregister.start,
1530                              uffdio_unregister.len);
1531         if (ret)
1532                 goto out;
1533
1534         start = uffdio_unregister.start;
1535         end = start + uffdio_unregister.len;
1536
1537         ret = -ENOMEM;
1538         if (!mmget_not_zero(mm))
1539                 goto out;
1540
1541         mmap_write_lock(mm);
1542         ret = -EINVAL;
1543         vma_iter_init(&vmi, mm, start);
1544         vma = vma_find(&vmi, end);
1545         if (!vma)
1546                 goto out_unlock;
1547
1548         /*
1549          * If the first vma contains huge pages, make sure start address
1550          * is aligned to huge page size.
1551          */
1552         if (is_vm_hugetlb_page(vma)) {
1553                 unsigned long vma_hpagesize = vma_kernel_pagesize(vma);
1554
1555                 if (start & (vma_hpagesize - 1))
1556                         goto out_unlock;
1557         }
1558
1559         /*
1560          * Search for not compatible vmas.
1561          */
1562         found = false;
1563         cur = vma;
1564         do {
1565                 cond_resched();
1566
1567                 BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1568                        !!(cur->vm_flags & __VM_UFFD_FLAGS));
1569
1570                 /*
1571                  * Check not compatible vmas, not strictly required
1572                  * here as not compatible vmas cannot have an
1573                  * userfaultfd_ctx registered on them, but this
1574                  * provides for more strict behavior to notice
1575                  * unregistration errors.
1576                  */
1577                 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
1578                         goto out_unlock;
1579
1580                 found = true;
1581         } for_each_vma_range(vmi, cur, end);
1582         BUG_ON(!found);
1583
1584         vma_iter_set(&vmi, start);
1585         prev = vma_prev(&vmi);
1586         if (vma->vm_start < start)
1587                 prev = vma;
1588
1589         ret = 0;
1590         for_each_vma_range(vmi, vma, end) {
1591                 cond_resched();
1592
1593                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
1594
1595                 /*
1596                  * Nothing to do: this vma is already registered into this
1597                  * userfaultfd and with the right tracking mode too.
1598                  */
1599                 if (!vma->vm_userfaultfd_ctx.ctx)
1600                         goto skip;
1601
1602                 WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
1603
1604                 if (vma->vm_start > start)
1605                         start = vma->vm_start;
1606                 vma_end = min(end, vma->vm_end);
1607
1608                 if (userfaultfd_missing(vma)) {
1609                         /*
1610                          * Wake any concurrent pending userfault while
1611                          * we unregister, so they will not hang
1612                          * permanently and it avoids userland to call
1613                          * UFFDIO_WAKE explicitly.
1614                          */
1615                         struct userfaultfd_wake_range range;
1616                         range.start = start;
1617                         range.len = vma_end - start;
1618                         wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
1619                 }
1620
1621                 /* Reset ptes for the whole vma range if wr-protected */
1622                 if (userfaultfd_wp(vma))
1623                         uffd_wp_range(vma, start, vma_end - start, false);
1624
1625                 new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
1626                 vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
1627                                             new_flags, NULL_VM_UFFD_CTX);
1628                 if (IS_ERR(vma)) {
1629                         ret = PTR_ERR(vma);
1630                         break;
1631                 }
1632
1633                 /*
1634                  * In the vma_merge() successful mprotect-like case 8:
1635                  * the next vma was merged into the current one and
1636                  * the current one has not been updated yet.
1637                  */
1638                 vma_start_write(vma);
1639                 userfaultfd_set_vm_flags(vma, new_flags);
1640                 vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1641
1642         skip:
1643                 prev = vma;
1644                 start = vma->vm_end;
1645         }
1646
1647 out_unlock:
1648         mmap_write_unlock(mm);
1649         mmput(mm);
1650 out:
1651         return ret;
1652 }
1653
1654 /*
1655  * userfaultfd_wake may be used in combination with the
1656  * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches.
1657  */
1658 static int userfaultfd_wake(struct userfaultfd_ctx *ctx,
1659                             unsigned long arg)
1660 {
1661         int ret;
1662         struct uffdio_range uffdio_wake;
1663         struct userfaultfd_wake_range range;
1664         const void __user *buf = (void __user *)arg;
1665
1666         ret = -EFAULT;
1667         if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
1668                 goto out;
1669
1670         ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
1671         if (ret)
1672                 goto out;
1673
1674         range.start = uffdio_wake.start;
1675         range.len = uffdio_wake.len;
1676
1677         /*
1678          * len == 0 means wake all and we don't want to wake all here,
1679          * so check it again to be sure.
1680          */
1681         VM_BUG_ON(!range.len);
1682
1683         wake_userfault(ctx, &range);
1684         ret = 0;
1685
1686 out:
1687         return ret;
1688 }
1689
1690 static int userfaultfd_copy(struct userfaultfd_ctx *ctx,
1691                             unsigned long arg)
1692 {
1693         __s64 ret;
1694         struct uffdio_copy uffdio_copy;
1695         struct uffdio_copy __user *user_uffdio_copy;
1696         struct userfaultfd_wake_range range;
1697         uffd_flags_t flags = 0;
1698
1699         user_uffdio_copy = (struct uffdio_copy __user *) arg;
1700
1701         ret = -EAGAIN;
1702         if (atomic_read(&ctx->mmap_changing))
1703                 goto out;
1704
1705         ret = -EFAULT;
1706         if (copy_from_user(&uffdio_copy, user_uffdio_copy,
1707                            /* don't copy "copy" last field */
1708                            sizeof(uffdio_copy)-sizeof(__s64)))
1709                 goto out;
1710
1711         ret = validate_unaligned_range(ctx->mm, uffdio_copy.src,
1712                                        uffdio_copy.len);
1713         if (ret)
1714                 goto out;
1715         ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
1716         if (ret)
1717                 goto out;
1718
1719         ret = -EINVAL;
1720         if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
1721                 goto out;
1722         if (uffdio_copy.mode & UFFDIO_COPY_MODE_WP)
1723                 flags |= MFILL_ATOMIC_WP;
1724         if (mmget_not_zero(ctx->mm)) {
1725                 ret = mfill_atomic_copy(ctx, uffdio_copy.dst, uffdio_copy.src,
1726                                         uffdio_copy.len, flags);
1727                 mmput(ctx->mm);
1728         } else {
1729                 return -ESRCH;
1730         }
1731         if (unlikely(put_user(ret, &user_uffdio_copy->copy)))
1732                 return -EFAULT;
1733         if (ret < 0)
1734                 goto out;
1735         BUG_ON(!ret);
1736         /* len == 0 would wake all */
1737         range.len = ret;
1738         if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) {
1739                 range.start = uffdio_copy.dst;
1740                 wake_userfault(ctx, &range);
1741         }
1742         ret = range.len == uffdio_copy.len ? 0 : -EAGAIN;
1743 out:
1744         return ret;
1745 }
1746
1747 static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx,
1748                                 unsigned long arg)
1749 {
1750         __s64 ret;
1751         struct uffdio_zeropage uffdio_zeropage;
1752         struct uffdio_zeropage __user *user_uffdio_zeropage;
1753         struct userfaultfd_wake_range range;
1754
1755         user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg;
1756
1757         ret = -EAGAIN;
1758         if (atomic_read(&ctx->mmap_changing))
1759                 goto out;
1760
1761         ret = -EFAULT;
1762         if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage,
1763                            /* don't copy "zeropage" last field */
1764                            sizeof(uffdio_zeropage)-sizeof(__s64)))
1765                 goto out;
1766
1767         ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
1768                              uffdio_zeropage.range.len);
1769         if (ret)
1770                 goto out;
1771         ret = -EINVAL;
1772         if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE)
1773                 goto out;
1774
1775         if (mmget_not_zero(ctx->mm)) {
1776                 ret = mfill_atomic_zeropage(ctx, uffdio_zeropage.range.start,
1777                                            uffdio_zeropage.range.len);
1778                 mmput(ctx->mm);
1779         } else {
1780                 return -ESRCH;
1781         }
1782         if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage)))
1783                 return -EFAULT;
1784         if (ret < 0)
1785                 goto out;
1786         /* len == 0 would wake all */
1787         BUG_ON(!ret);
1788         range.len = ret;
1789         if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) {
1790                 range.start = uffdio_zeropage.range.start;
1791                 wake_userfault(ctx, &range);
1792         }
1793         ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1794 out:
1795         return ret;
1796 }
1797
1798 static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1799                                     unsigned long arg)
1800 {
1801         int ret;
1802         struct uffdio_writeprotect uffdio_wp;
1803         struct uffdio_writeprotect __user *user_uffdio_wp;
1804         struct userfaultfd_wake_range range;
1805         bool mode_wp, mode_dontwake;
1806
1807         if (atomic_read(&ctx->mmap_changing))
1808                 return -EAGAIN;
1809
1810         user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1811
1812         if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1813                            sizeof(struct uffdio_writeprotect)))
1814                 return -EFAULT;
1815
1816         ret = validate_range(ctx->mm, uffdio_wp.range.start,
1817                              uffdio_wp.range.len);
1818         if (ret)
1819                 return ret;
1820
1821         if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1822                                UFFDIO_WRITEPROTECT_MODE_WP))
1823                 return -EINVAL;
1824
1825         mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1826         mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1827
1828         if (mode_wp && mode_dontwake)
1829                 return -EINVAL;
1830
1831         if (mmget_not_zero(ctx->mm)) {
1832                 ret = mwriteprotect_range(ctx, uffdio_wp.range.start,
1833                                           uffdio_wp.range.len, mode_wp);
1834                 mmput(ctx->mm);
1835         } else {
1836                 return -ESRCH;
1837         }
1838
1839         if (ret)
1840                 return ret;
1841
1842         if (!mode_wp && !mode_dontwake) {
1843                 range.start = uffdio_wp.range.start;
1844                 range.len = uffdio_wp.range.len;
1845                 wake_userfault(ctx, &range);
1846         }
1847         return ret;
1848 }
1849
1850 static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1851 {
1852         __s64 ret;
1853         struct uffdio_continue uffdio_continue;
1854         struct uffdio_continue __user *user_uffdio_continue;
1855         struct userfaultfd_wake_range range;
1856         uffd_flags_t flags = 0;
1857
1858         user_uffdio_continue = (struct uffdio_continue __user *)arg;
1859
1860         ret = -EAGAIN;
1861         if (atomic_read(&ctx->mmap_changing))
1862                 goto out;
1863
1864         ret = -EFAULT;
1865         if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1866                            /* don't copy the output fields */
1867                            sizeof(uffdio_continue) - (sizeof(__s64))))
1868                 goto out;
1869
1870         ret = validate_range(ctx->mm, uffdio_continue.range.start,
1871                              uffdio_continue.range.len);
1872         if (ret)
1873                 goto out;
1874
1875         ret = -EINVAL;
1876         if (uffdio_continue.mode & ~(UFFDIO_CONTINUE_MODE_DONTWAKE |
1877                                      UFFDIO_CONTINUE_MODE_WP))
1878                 goto out;
1879         if (uffdio_continue.mode & UFFDIO_CONTINUE_MODE_WP)
1880                 flags |= MFILL_ATOMIC_WP;
1881
1882         if (mmget_not_zero(ctx->mm)) {
1883                 ret = mfill_atomic_continue(ctx, uffdio_continue.range.start,
1884                                             uffdio_continue.range.len, flags);
1885                 mmput(ctx->mm);
1886         } else {
1887                 return -ESRCH;
1888         }
1889
1890         if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1891                 return -EFAULT;
1892         if (ret < 0)
1893                 goto out;
1894
1895         /* len == 0 would wake all */
1896         BUG_ON(!ret);
1897         range.len = ret;
1898         if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1899                 range.start = uffdio_continue.range.start;
1900                 wake_userfault(ctx, &range);
1901         }
1902         ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1903
1904 out:
1905         return ret;
1906 }
1907
1908 static inline int userfaultfd_poison(struct userfaultfd_ctx *ctx, unsigned long arg)
1909 {
1910         __s64 ret;
1911         struct uffdio_poison uffdio_poison;
1912         struct uffdio_poison __user *user_uffdio_poison;
1913         struct userfaultfd_wake_range range;
1914
1915         user_uffdio_poison = (struct uffdio_poison __user *)arg;
1916
1917         ret = -EAGAIN;
1918         if (atomic_read(&ctx->mmap_changing))
1919                 goto out;
1920
1921         ret = -EFAULT;
1922         if (copy_from_user(&uffdio_poison, user_uffdio_poison,
1923                            /* don't copy the output fields */
1924                            sizeof(uffdio_poison) - (sizeof(__s64))))
1925                 goto out;
1926
1927         ret = validate_range(ctx->mm, uffdio_poison.range.start,
1928                              uffdio_poison.range.len);
1929         if (ret)
1930                 goto out;
1931
1932         ret = -EINVAL;
1933         if (uffdio_poison.mode & ~UFFDIO_POISON_MODE_DONTWAKE)
1934                 goto out;
1935
1936         if (mmget_not_zero(ctx->mm)) {
1937                 ret = mfill_atomic_poison(ctx, uffdio_poison.range.start,
1938                                           uffdio_poison.range.len, 0);
1939                 mmput(ctx->mm);
1940         } else {
1941                 return -ESRCH;
1942         }
1943
1944         if (unlikely(put_user(ret, &user_uffdio_poison->updated)))
1945                 return -EFAULT;
1946         if (ret < 0)
1947                 goto out;
1948
1949         /* len == 0 would wake all */
1950         BUG_ON(!ret);
1951         range.len = ret;
1952         if (!(uffdio_poison.mode & UFFDIO_POISON_MODE_DONTWAKE)) {
1953                 range.start = uffdio_poison.range.start;
1954                 wake_userfault(ctx, &range);
1955         }
1956         ret = range.len == uffdio_poison.range.len ? 0 : -EAGAIN;
1957
1958 out:
1959         return ret;
1960 }
1961
1962 bool userfaultfd_wp_async(struct vm_area_struct *vma)
1963 {
1964         return userfaultfd_wp_async_ctx(vma->vm_userfaultfd_ctx.ctx);
1965 }
1966
1967 static inline unsigned int uffd_ctx_features(__u64 user_features)
1968 {
1969         /*
1970          * For the current set of features the bits just coincide. Set
1971          * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
1972          */
1973         return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
1974 }
1975
1976 static int userfaultfd_move(struct userfaultfd_ctx *ctx,
1977                             unsigned long arg)
1978 {
1979         __s64 ret;
1980         struct uffdio_move uffdio_move;
1981         struct uffdio_move __user *user_uffdio_move;
1982         struct userfaultfd_wake_range range;
1983         struct mm_struct *mm = ctx->mm;
1984
1985         user_uffdio_move = (struct uffdio_move __user *) arg;
1986
1987         if (atomic_read(&ctx->mmap_changing))
1988                 return -EAGAIN;
1989
1990         if (copy_from_user(&uffdio_move, user_uffdio_move,
1991                            /* don't copy "move" last field */
1992                            sizeof(uffdio_move)-sizeof(__s64)))
1993                 return -EFAULT;
1994
1995         /* Do not allow cross-mm moves. */
1996         if (mm != current->mm)
1997                 return -EINVAL;
1998
1999         ret = validate_range(mm, uffdio_move.dst, uffdio_move.len);
2000         if (ret)
2001                 return ret;
2002
2003         ret = validate_range(mm, uffdio_move.src, uffdio_move.len);
2004         if (ret)
2005                 return ret;
2006
2007         if (uffdio_move.mode & ~(UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES|
2008                                   UFFDIO_MOVE_MODE_DONTWAKE))
2009                 return -EINVAL;
2010
2011         if (mmget_not_zero(mm)) {
2012                 ret = move_pages(ctx, uffdio_move.dst, uffdio_move.src,
2013                                  uffdio_move.len, uffdio_move.mode);
2014                 mmput(mm);
2015         } else {
2016                 return -ESRCH;
2017         }
2018
2019         if (unlikely(put_user(ret, &user_uffdio_move->move)))
2020                 return -EFAULT;
2021         if (ret < 0)
2022                 goto out;
2023
2024         /* len == 0 would wake all */
2025         VM_WARN_ON(!ret);
2026         range.len = ret;
2027         if (!(uffdio_move.mode & UFFDIO_MOVE_MODE_DONTWAKE)) {
2028                 range.start = uffdio_move.dst;
2029                 wake_userfault(ctx, &range);
2030         }
2031         ret = range.len == uffdio_move.len ? 0 : -EAGAIN;
2032
2033 out:
2034         return ret;
2035 }
2036
2037 /*
2038  * userland asks for a certain API version and we return which bits
2039  * and ioctl commands are implemented in this kernel for such API
2040  * version or -EINVAL if unknown.
2041  */
2042 static int userfaultfd_api(struct userfaultfd_ctx *ctx,
2043                            unsigned long arg)
2044 {
2045         struct uffdio_api uffdio_api;
2046         void __user *buf = (void __user *)arg;
2047         unsigned int ctx_features;
2048         int ret;
2049         __u64 features;
2050
2051         ret = -EFAULT;
2052         if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
2053                 goto out;
2054         features = uffdio_api.features;
2055         ret = -EINVAL;
2056         if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES))
2057                 goto err_out;
2058         ret = -EPERM;
2059         if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE))
2060                 goto err_out;
2061
2062         /* WP_ASYNC relies on WP_UNPOPULATED, choose it unconditionally */
2063         if (features & UFFD_FEATURE_WP_ASYNC)
2064                 features |= UFFD_FEATURE_WP_UNPOPULATED;
2065
2066         /* report all available features and ioctls to userland */
2067         uffdio_api.features = UFFD_API_FEATURES;
2068 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
2069         uffdio_api.features &=
2070                 ~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
2071 #endif
2072 #ifndef CONFIG_HAVE_ARCH_USERFAULTFD_WP
2073         uffdio_api.features &= ~UFFD_FEATURE_PAGEFAULT_FLAG_WP;
2074 #endif
2075 #ifndef CONFIG_PTE_MARKER_UFFD_WP
2076         uffdio_api.features &= ~UFFD_FEATURE_WP_HUGETLBFS_SHMEM;
2077         uffdio_api.features &= ~UFFD_FEATURE_WP_UNPOPULATED;
2078         uffdio_api.features &= ~UFFD_FEATURE_WP_ASYNC;
2079 #endif
2080         uffdio_api.ioctls = UFFD_API_IOCTLS;
2081         ret = -EFAULT;
2082         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2083                 goto out;
2084
2085         /* only enable the requested features for this uffd context */
2086         ctx_features = uffd_ctx_features(features);
2087         ret = -EINVAL;
2088         if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
2089                 goto err_out;
2090
2091         ret = 0;
2092 out:
2093         return ret;
2094 err_out:
2095         memset(&uffdio_api, 0, sizeof(uffdio_api));
2096         if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
2097                 ret = -EFAULT;
2098         goto out;
2099 }
2100
2101 static long userfaultfd_ioctl(struct file *file, unsigned cmd,
2102                               unsigned long arg)
2103 {
2104         int ret = -EINVAL;
2105         struct userfaultfd_ctx *ctx = file->private_data;
2106
2107         if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
2108                 return -EINVAL;
2109
2110         switch(cmd) {
2111         case UFFDIO_API:
2112                 ret = userfaultfd_api(ctx, arg);
2113                 break;
2114         case UFFDIO_REGISTER:
2115                 ret = userfaultfd_register(ctx, arg);
2116                 break;
2117         case UFFDIO_UNREGISTER:
2118                 ret = userfaultfd_unregister(ctx, arg);
2119                 break;
2120         case UFFDIO_WAKE:
2121                 ret = userfaultfd_wake(ctx, arg);
2122                 break;
2123         case UFFDIO_COPY:
2124                 ret = userfaultfd_copy(ctx, arg);
2125                 break;
2126         case UFFDIO_ZEROPAGE:
2127                 ret = userfaultfd_zeropage(ctx, arg);
2128                 break;
2129         case UFFDIO_MOVE:
2130                 ret = userfaultfd_move(ctx, arg);
2131                 break;
2132         case UFFDIO_WRITEPROTECT:
2133                 ret = userfaultfd_writeprotect(ctx, arg);
2134                 break;
2135         case UFFDIO_CONTINUE:
2136                 ret = userfaultfd_continue(ctx, arg);
2137                 break;
2138         case UFFDIO_POISON:
2139                 ret = userfaultfd_poison(ctx, arg);
2140                 break;
2141         }
2142         return ret;
2143 }
2144
2145 #ifdef CONFIG_PROC_FS
2146 static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
2147 {
2148         struct userfaultfd_ctx *ctx = f->private_data;
2149         wait_queue_entry_t *wq;
2150         unsigned long pending = 0, total = 0;
2151
2152         spin_lock_irq(&ctx->fault_pending_wqh.lock);
2153         list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) {
2154                 pending++;
2155                 total++;
2156         }
2157         list_for_each_entry(wq, &ctx->fault_wqh.head, entry) {
2158                 total++;
2159         }
2160         spin_unlock_irq(&ctx->fault_pending_wqh.lock);
2161
2162         /*
2163          * If more protocols will be added, there will be all shown
2164          * separated by a space. Like this:
2165          *      protocols: aa:... bb:...
2166          */
2167         seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
2168                    pending, total, UFFD_API, ctx->features,
2169                    UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
2170 }
2171 #endif
2172
2173 static const struct file_operations userfaultfd_fops = {
2174 #ifdef CONFIG_PROC_FS
2175         .show_fdinfo    = userfaultfd_show_fdinfo,
2176 #endif
2177         .release        = userfaultfd_release,
2178         .poll           = userfaultfd_poll,
2179         .read           = userfaultfd_read,
2180         .unlocked_ioctl = userfaultfd_ioctl,
2181         .compat_ioctl   = compat_ptr_ioctl,
2182         .llseek         = noop_llseek,
2183 };
2184
2185 static void init_once_userfaultfd_ctx(void *mem)
2186 {
2187         struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem;
2188
2189         init_waitqueue_head(&ctx->fault_pending_wqh);
2190         init_waitqueue_head(&ctx->fault_wqh);
2191         init_waitqueue_head(&ctx->event_wqh);
2192         init_waitqueue_head(&ctx->fd_wqh);
2193         seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
2194 }
2195
2196 static int new_userfaultfd(int flags)
2197 {
2198         struct userfaultfd_ctx *ctx;
2199         int fd;
2200
2201         BUG_ON(!current->mm);
2202
2203         /* Check the UFFD_* constants for consistency.  */
2204         BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
2205         BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
2206         BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
2207
2208         if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
2209                 return -EINVAL;
2210
2211         ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
2212         if (!ctx)
2213                 return -ENOMEM;
2214
2215         refcount_set(&ctx->refcount, 1);
2216         ctx->flags = flags;
2217         ctx->features = 0;
2218         ctx->released = false;
2219         init_rwsem(&ctx->map_changing_lock);
2220         atomic_set(&ctx->mmap_changing, 0);
2221         ctx->mm = current->mm;
2222         /* prevent the mm struct to be freed */
2223         mmgrab(ctx->mm);
2224
2225         /* Create a new inode so that the LSM can block the creation.  */
2226         fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
2227                         O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
2228         if (fd < 0) {
2229                 mmdrop(ctx->mm);
2230                 kmem_cache_free(userfaultfd_ctx_cachep, ctx);
2231         }
2232         return fd;
2233 }
2234
2235 static inline bool userfaultfd_syscall_allowed(int flags)
2236 {
2237         /* Userspace-only page faults are always allowed */
2238         if (flags & UFFD_USER_MODE_ONLY)
2239                 return true;
2240
2241         /*
2242          * The user is requesting a userfaultfd which can handle kernel faults.
2243          * Privileged users are always allowed to do this.
2244          */
2245         if (capable(CAP_SYS_PTRACE))
2246                 return true;
2247
2248         /* Otherwise, access to kernel fault handling is sysctl controlled. */
2249         return sysctl_unprivileged_userfaultfd;
2250 }
2251
2252 SYSCALL_DEFINE1(userfaultfd, int, flags)
2253 {
2254         if (!userfaultfd_syscall_allowed(flags))
2255                 return -EPERM;
2256
2257         return new_userfaultfd(flags);
2258 }
2259
2260 static long userfaultfd_dev_ioctl(struct file *file, unsigned int cmd, unsigned long flags)
2261 {
2262         if (cmd != USERFAULTFD_IOC_NEW)
2263                 return -EINVAL;
2264
2265         return new_userfaultfd(flags);
2266 }
2267
2268 static const struct file_operations userfaultfd_dev_fops = {
2269         .unlocked_ioctl = userfaultfd_dev_ioctl,
2270         .compat_ioctl = userfaultfd_dev_ioctl,
2271         .owner = THIS_MODULE,
2272         .llseek = noop_llseek,
2273 };
2274
2275 static struct miscdevice userfaultfd_misc = {
2276         .minor = MISC_DYNAMIC_MINOR,
2277         .name = "userfaultfd",
2278         .fops = &userfaultfd_dev_fops
2279 };
2280
2281 static int __init userfaultfd_init(void)
2282 {
2283         int ret;
2284
2285         ret = misc_register(&userfaultfd_misc);
2286         if (ret)
2287                 return ret;
2288
2289         userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache",
2290                                                 sizeof(struct userfaultfd_ctx),
2291                                                 0,
2292                                                 SLAB_HWCACHE_ALIGN|SLAB_PANIC,
2293                                                 init_once_userfaultfd_ctx);
2294 #ifdef CONFIG_SYSCTL
2295         register_sysctl_init("vm", vm_userfaultfd_table);
2296 #endif
2297         return 0;
2298 }
2299 __initcall(userfaultfd_init);