sys/vm/vm_fault.c

   1 /*-
   2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
   3  *
   4  * Copyright (c) 1991, 1993
   5  *      The Regents of the University of California.  All rights reserved.
   6  * Copyright (c) 1994 John S. Dyson
   7  * All rights reserved.
   8  * Copyright (c) 1994 David Greenman
   9  * All rights reserved.
  10  *
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * The Mach Operating System project at Carnegie-Mellon University.
  14  *
  15  * Redistribution and use in source and binary forms, with or without
  16  * modification, are permitted provided that the following conditions
  17  * are met:
  18  * 1. Redistributions of source code must retain the above copyright
  19  *    notice, this list of conditions and the following disclaimer.
  20  * 2. Redistributions in binary form must reproduce the above copyright
  21  *    notice, this list of conditions and the following disclaimer in the
  22  *    documentation and/or other materials provided with the distribution.
  23  * 3. All advertising materials mentioning features or use of this software
  24  *    must display the following acknowledgement:
  25  *      This product includes software developed by the University of
  26  *      California, Berkeley and its contributors.
  27  * 4. Neither the name of the University nor the names of its contributors
  28  *    may be used to endorse or promote products derived from this software
  29  *    without specific prior written permission.
  30  *
  31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  41  * SUCH DAMAGE.
  42  *
  43  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  44  *
  45  *
  46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  47  * All rights reserved.
  48  *
  49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  50  *
  51  * Permission to use, copy, modify and distribute this software and
  52  * its documentation is hereby granted, provided that both the copyright
  53  * notice and this permission notice appear in all copies of the
  54  * software, derivative works or modified versions, and any portions
  55  * thereof, and that both notices appear in supporting documentation.
  56  *
  57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  60  *
  61  * Carnegie Mellon requests users of this software to return to
  62  *
  63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  64  *  School of Computer Science
  65  *  Carnegie Mellon University
  66  *  Pittsburgh PA 15213-3890
  67  *
  68  * any improvements or extensions that they make and grant Carnegie the
  69  * rights to redistribute these changes.
  70  */
  71
  72 /*
  73  *      Page fault handling module.
  74  */
  75
  76 #include <sys/cdefs.h>
  77 __FBSDID("$FreeBSD$");
  78
  79 #include "opt_ktrace.h"
  80 #include "opt_vm.h"
  81
  82 #include <sys/param.h>
  83 #include <sys/systm.h>
  84 #include <sys/kernel.h>
  85 #include <sys/lock.h>
  86 #include <sys/mman.h>
  87 #include <sys/proc.h>
  88 #include <sys/racct.h>
  89 #include <sys/resourcevar.h>
  90 #include <sys/rwlock.h>
  91 #include <sys/sysctl.h>
  92 #include <sys/vmmeter.h>
  93 #include <sys/vnode.h>
  94 #ifdef KTRACE
  95 #include <sys/ktrace.h>
  96 #endif
  97
  98 #include <vm/vm.h>
  99 #include <vm/vm_param.h>
 100 #include <vm/pmap.h>
 101 #include <vm/vm_map.h>
 102 #include <vm/vm_object.h>
 103 #include <vm/vm_page.h>
 104 #include <vm/vm_pageout.h>
 105 #include <vm/vm_kern.h>
 106 #include <vm/vm_pager.h>
 107 #include <vm/vm_extern.h>
 108 #include <vm/vm_reserv.h>
 109
 110 #define PFBAK 4
 111 #define PFFOR 4
 112
 113 #define VM_FAULT_READ_DEFAULT   (1 + VM_FAULT_READ_AHEAD_INIT)
 114 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 115
 116 #define VM_FAULT_DONTNEED_MIN   1048576
 117
 118 struct faultstate {
 119         vm_page_t m;
 120         vm_object_t object;
 121         vm_pindex_t pindex;
 122         vm_page_t first_m;
 123         vm_object_t     first_object;
 124         vm_pindex_t first_pindex;
 125         vm_map_t map;
 126         vm_map_entry_t entry;
 127         int map_generation;
 128         bool lookup_still_valid;
 129         struct vnode *vp;
 130 };
 131
 132 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 133             int ahead);
 134 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 135             int backward, int forward, bool obj_locked);
 136
 137 static inline void
 138 release_page(struct faultstate *fs)
 139 {
 140
 141         vm_page_xunbusy(fs->m);
 142         vm_page_lock(fs->m);
 143         vm_page_deactivate(fs->m);
 144         vm_page_unlock(fs->m);
 145         fs->m = NULL;
 146 }
 147
 148 static inline void
 149 unlock_map(struct faultstate *fs)
 150 {
 151
 152         if (fs->lookup_still_valid) {
 153                 vm_map_lookup_done(fs->map, fs->entry);
 154                 fs->lookup_still_valid = false;
 155         }
 156 }
 157
 158 static void
 159 unlock_vp(struct faultstate *fs)
 160 {
 161
 162         if (fs->vp != NULL) {
 163                 vput(fs->vp);
 164                 fs->vp = NULL;
 165         }
 166 }
 167
 168 static void
 169 unlock_and_deallocate(struct faultstate *fs)
 170 {
 171
 172         vm_object_pip_wakeup(fs->object);
 173         VM_OBJECT_WUNLOCK(fs->object);
 174         if (fs->object != fs->first_object) {
 175                 VM_OBJECT_WLOCK(fs->first_object);
 176                 vm_page_lock(fs->first_m);
 177                 vm_page_free(fs->first_m);
 178                 vm_page_unlock(fs->first_m);
 179                 vm_object_pip_wakeup(fs->first_object);
 180                 VM_OBJECT_WUNLOCK(fs->first_object);
 181                 fs->first_m = NULL;
 182         }
 183         vm_object_deallocate(fs->first_object);
 184         unlock_map(fs);
 185         unlock_vp(fs);
 186 }
 187
 188 static void
 189 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 190     vm_prot_t fault_type, int fault_flags, bool set_wd)
 191 {
 192         bool need_dirty;
 193
 194         if (((prot & VM_PROT_WRITE) == 0 &&
 195             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 196             (m->oflags & VPO_UNMANAGED) != 0)
 197                 return;
 198
 199         VM_OBJECT_ASSERT_LOCKED(m->object);
 200
 201         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 202             (fault_flags & VM_FAULT_WIRE) == 0) ||
 203             (fault_flags & VM_FAULT_DIRTY) != 0;
 204
 205         if (set_wd)
 206                 vm_object_set_writeable_dirty(m->object);
 207         else
 208                 /*
 209                  * If two callers of vm_fault_dirty() with set_wd ==
 210                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 211                  * flag set, other with flag clear, race, it is
 212                  * possible for the no-NOSYNC thread to see m->dirty
 213                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 214                  * around manipulation of VPO_NOSYNC and
 215                  * vm_page_dirty() call, to avoid the race and keep
 216                  * m->oflags consistent.
 217                  */
 218                 vm_page_lock(m);
 219
 220         /*
 221          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 222          * if the page is already dirty to prevent data written with
 223          * the expectation of being synced from not being synced.
 224          * Likewise if this entry does not request NOSYNC then make
 225          * sure the page isn't marked NOSYNC.  Applications sharing
 226          * data should use the same flags to avoid ping ponging.
 227          */
 228         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 229                 if (m->dirty == 0) {
 230                         m->oflags |= VPO_NOSYNC;
 231                 }
 232         } else {
 233                 m->oflags &= ~VPO_NOSYNC;
 234         }
 235
 236         /*
 237          * If the fault is a write, we know that this page is being
 238          * written NOW so dirty it explicitly to save on
 239          * pmap_is_modified() calls later.
 240          *
 241          * Also, since the page is now dirty, we can possibly tell
 242          * the pager to release any swap backing the page.  Calling
 243          * the pager requires a write lock on the object.
 244          */
 245         if (need_dirty)
 246                 vm_page_dirty(m);
 247         if (!set_wd)
 248                 vm_page_unlock(m);
 249         else if (need_dirty)
 250                 vm_pager_page_unswapped(m);
 251 }
 252
 253 static void
 254 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m)
 255 {
 256
 257         if (m_hold != NULL) {
 258                 *m_hold = m;
 259                 vm_page_lock(m);
 260                 vm_page_hold(m);
 261                 vm_page_unlock(m);
 262         }
 263 }
 264
 265 /*
 266  * Unlocks fs.first_object and fs.map on success.
 267  */
 268 static int
 269 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 270     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 271 {
 272         vm_page_t m, m_map;
 273 #if defined(__amd64__) && VM_NRESERVLEVEL > 0
 274         vm_page_t m_super;
 275         int flags;
 276 #endif
 277         int psind, rv;
 278
 279         MPASS(fs->vp == NULL);
 280         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 281         /* A busy page can be mapped for read|execute access. */
 282         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 283             vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 284                 return (KERN_FAILURE);
 285         m_map = m;
 286         psind = 0;
 287 #if defined(__amd64__) && VM_NRESERVLEVEL > 0
 288         if ((m->flags & PG_FICTITIOUS) == 0 &&
 289             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 290             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 291             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 292             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 293             (pagesizes[m_super->psind] - 1)) &&
 294             pmap_ps_enabled(fs->map->pmap)) {
 295                 flags = PS_ALL_VALID;
 296                 if ((prot & VM_PROT_WRITE) != 0) {
 297                         /*
 298                          * Create a superpage mapping allowing write access
 299                          * only if none of the constituent pages are busy and
 300                          * all of them are already dirty (except possibly for
 301                          * the page that was faulted on).
 302                          */
 303                         flags |= PS_NONE_BUSY;
 304                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 305                                 flags |= PS_ALL_DIRTY;
 306                 }
 307                 if (vm_page_ps_test(m_super, flags, m)) {
 308                         m_map = m_super;
 309                         psind = m_super->psind;
 310                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 311                         /* Preset the modified bit for dirty superpages. */
 312                         if ((flags & PS_ALL_DIRTY) != 0)
 313                                 fault_type |= VM_PROT_WRITE;
 314                 }
 315         }
 316 #endif
 317         rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
 318             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
 319         if (rv != KERN_SUCCESS)
 320                 return (rv);
 321         vm_fault_fill_hold(m_hold, m);
 322         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
 323         if (psind == 0 && !wired)
 324                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 325         VM_OBJECT_RUNLOCK(fs->first_object);
 326         vm_map_lookup_done(fs->map, fs->entry);
 327         curthread->td_ru.ru_minflt++;
 328         return (KERN_SUCCESS);
 329 }
 330
 331 static void
 332 vm_fault_restore_map_lock(struct faultstate *fs)
 333 {
 334
 335         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 336         MPASS(fs->first_object->paging_in_progress > 0);
 337
 338         if (!vm_map_trylock_read(fs->map)) {
 339                 VM_OBJECT_WUNLOCK(fs->first_object);
 340                 vm_map_lock_read(fs->map);
 341                 VM_OBJECT_WLOCK(fs->first_object);
 342         }
 343         fs->lookup_still_valid = true;
 344 }
 345
 346 static void
 347 vm_fault_populate_check_page(vm_page_t m)
 348 {
 349
 350         /*
 351          * Check each page to ensure that the pager is obeying the
 352          * interface: the page must be installed in the object, fully
 353          * valid, and exclusively busied.
 354          */
 355         MPASS(m != NULL);
 356         MPASS(m->valid == VM_PAGE_BITS_ALL);
 357         MPASS(vm_page_xbusied(m));
 358 }
 359
 360 static void
 361 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 362     vm_pindex_t last)
 363 {
 364         vm_page_t m;
 365         vm_pindex_t pidx;
 366
 367         VM_OBJECT_ASSERT_WLOCKED(object);
 368         MPASS(first <= last);
 369         for (pidx = first, m = vm_page_lookup(object, pidx);
 370             pidx <= last; pidx++, m = vm_page_next(m)) {
 371                 vm_fault_populate_check_page(m);
 372                 vm_page_lock(m);
 373                 vm_page_deactivate(m);
 374                 vm_page_unlock(m);
 375                 vm_page_xunbusy(m);
 376         }
 377 }
 378
 379 static int
 380 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 381     int fault_flags, boolean_t wired, vm_page_t *m_hold)
 382 {
 383         struct mtx *m_mtx;
 384         vm_offset_t vaddr;
 385         vm_page_t m;
 386         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 387         int i, npages, psind, rv;
 388
 389         MPASS(fs->object == fs->first_object);
 390         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 391         MPASS(fs->first_object->paging_in_progress > 0);
 392         MPASS(fs->first_object->backing_object == NULL);
 393         MPASS(fs->lookup_still_valid);
 394
 395         pager_first = OFF_TO_IDX(fs->entry->offset);
 396         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 397         unlock_map(fs);
 398         unlock_vp(fs);
 399
 400         /*
 401          * Call the pager (driver) populate() method.
 402          *
 403          * There is no guarantee that the method will be called again
 404          * if the current fault is for read, and a future fault is
 405          * for write.  Report the entry's maximum allowed protection
 406          * to the driver.
 407          */
 408         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 409             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 410
 411         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 412         if (rv == VM_PAGER_BAD) {
 413                 /*
 414                  * VM_PAGER_BAD is the backdoor for a pager to request
 415                  * normal fault handling.
 416                  */
 417                 vm_fault_restore_map_lock(fs);
 418                 if (fs->map->timestamp != fs->map_generation)
 419                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 420                 return (KERN_NOT_RECEIVER);
 421         }
 422         if (rv != VM_PAGER_OK)
 423                 return (KERN_FAILURE); /* AKA SIGSEGV */
 424
 425         /* Ensure that the driver is obeying the interface. */
 426         MPASS(pager_first <= pager_last);
 427         MPASS(fs->first_pindex <= pager_last);
 428         MPASS(fs->first_pindex >= pager_first);
 429         MPASS(pager_last < fs->first_object->size);
 430
 431         vm_fault_restore_map_lock(fs);
 432         if (fs->map->timestamp != fs->map_generation) {
 433                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 434                     pager_last);
 435                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 436         }
 437
 438         /*
 439          * The map is unchanged after our last unlock.  Process the fault.
 440          *
 441          * The range [pager_first, pager_last] that is given to the
 442          * pager is only a hint.  The pager may populate any range
 443          * within the object that includes the requested page index.
 444          * In case the pager expanded the range, clip it to fit into
 445          * the map entry.
 446          */
 447         map_first = OFF_TO_IDX(fs->entry->offset);
 448         if (map_first > pager_first) {
 449                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 450                     map_first - 1);
 451                 pager_first = map_first;
 452         }
 453         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 454         if (map_last < pager_last) {
 455                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 456                     pager_last);
 457                 pager_last = map_last;
 458         }
 459         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 460             pidx <= pager_last;
 461             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 462                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 463 #if defined(__amd64__)
 464                 psind = m->psind;
 465                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 466                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 467                     !pmap_ps_enabled(fs->map->pmap)))
 468                         psind = 0;
 469 #else
 470                 psind = 0;
 471 #endif
 472                 npages = atop(pagesizes[psind]);
 473                 for (i = 0; i < npages; i++) {
 474                         vm_fault_populate_check_page(&m[i]);
 475                         vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
 476                             fault_flags, true);
 477                 }
 478                 VM_OBJECT_WUNLOCK(fs->first_object);
 479                 pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ?
 480                     PMAP_ENTER_WIRED : 0), psind);
 481                 VM_OBJECT_WLOCK(fs->first_object);
 482                 m_mtx = NULL;
 483                 for (i = 0; i < npages; i++) {
 484                         vm_page_change_lock(&m[i], &m_mtx);
 485                         if ((fault_flags & VM_FAULT_WIRE) != 0)
 486                                 vm_page_wire(&m[i]);
 487                         else
 488                                 vm_page_activate(&m[i]);
 489                         if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
 490                                 *m_hold = &m[i];
 491                                 vm_page_hold(&m[i]);
 492                         }
 493                         vm_page_xunbusy_maybelocked(&m[i]);
 494                 }
 495                 if (m_mtx != NULL)
 496                         mtx_unlock(m_mtx);
 497         }
 498         curthread->td_ru.ru_majflt++;
 499         return (KERN_SUCCESS);
 500 }
 501
 502 /*
 503  *      vm_fault:
 504  *
 505  *      Handle a page fault occurring at the given address,
 506  *      requiring the given permissions, in the map specified.
 507  *      If successful, the page is inserted into the
 508  *      associated physical map.
 509  *
 510  *      NOTE: the given address should be truncated to the
 511  *      proper page address.
 512  *
 513  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 514  *      a standard error specifying why the fault is fatal is returned.
 515  *
 516  *      The map in question must be referenced, and remains so.
 517  *      Caller may hold no locks.
 518  */
 519 int
 520 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 521     int fault_flags)
 522 {
 523         struct thread *td;
 524         int result;
 525
 526         td = curthread;
 527         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 528                 return (KERN_PROTECTION_FAILURE);
 529 #ifdef KTRACE
 530         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 531                 ktrfault(vaddr, fault_type);
 532 #endif
 533         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 534             NULL);
 535 #ifdef KTRACE
 536         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 537                 ktrfaultend(result);
 538 #endif
 539         return (result);
 540 }
 541
 542 int
 543 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 544     int fault_flags, vm_page_t *m_hold)
 545 {
 546         struct faultstate fs;
 547         struct vnode *vp;
 548         vm_object_t next_object, retry_object;
 549         vm_offset_t e_end, e_start;
 550         vm_pindex_t retry_pindex;
 551         vm_prot_t prot, retry_prot;
 552         int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
 553         int locked, nera, result, rv;
 554         u_char behavior;
 555         boolean_t wired;        /* Passed by reference. */
 556         bool dead, hardfault, is_first_object_locked;
 557
 558         VM_CNT_INC(v_vm_faults);
 559         fs.vp = NULL;
 560         faultcount = 0;
 561         nera = -1;
 562         hardfault = false;
 563
 564 RetryFault:;
 565
 566         /*
 567          * Find the backing store object and offset into it to begin the
 568          * search.
 569          */
 570         fs.map = map;
 571         result = vm_map_lookup(&fs.map, vaddr, fault_type |
 572             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 573             &fs.first_pindex, &prot, &wired);
 574         if (result != KERN_SUCCESS) {
 575                 unlock_vp(&fs);
 576                 return (result);
 577         }
 578
 579         fs.map_generation = fs.map->timestamp;
 580
 581         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 582                 panic("%s: fault on nofault entry, addr: %#lx",
 583                     __func__, (u_long)vaddr);
 584         }
 585
 586         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 587             fs.entry->wiring_thread != curthread) {
 588                 vm_map_unlock_read(fs.map);
 589                 vm_map_lock(fs.map);
 590                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 591                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 592                         unlock_vp(&fs);
 593                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 594                         vm_map_unlock_and_wait(fs.map, 0);
 595                 } else
 596                         vm_map_unlock(fs.map);
 597                 goto RetryFault;
 598         }
 599
 600         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 601
 602         if (wired)
 603                 fault_type = prot | (fault_type & VM_PROT_COPY);
 604         else
 605                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 606                     ("!wired && VM_FAULT_WIRE"));
 607
 608         /*
 609          * Try to avoid lock contention on the top-level object through
 610          * special-case handling of some types of page faults, specifically,
 611          * those that are both (1) mapping an existing page from the top-
 612          * level object and (2) not having to mark that object as containing
 613          * dirty pages.  Under these conditions, a read lock on the top-level
 614          * object suffices, allowing multiple page faults of a similar type to
 615          * run in parallel on the same top-level object.
 616          */
 617         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 618             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
 619             /* avoid calling vm_object_set_writeable_dirty() */
 620             ((prot & VM_PROT_WRITE) == 0 ||
 621             (fs.first_object->type != OBJT_VNODE &&
 622             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 623             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 624                 VM_OBJECT_RLOCK(fs.first_object);
 625                 if ((prot & VM_PROT_WRITE) == 0 ||
 626                     (fs.first_object->type != OBJT_VNODE &&
 627                     (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 628                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 629                         rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 630                             fault_flags, wired, m_hold);
 631                         if (rv == KERN_SUCCESS)
 632                                 return (rv);
 633                 }
 634                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 635                         VM_OBJECT_RUNLOCK(fs.first_object);
 636                         VM_OBJECT_WLOCK(fs.first_object);
 637                 }
 638         } else {
 639                 VM_OBJECT_WLOCK(fs.first_object);
 640         }
 641
 642         /*
 643          * Make a reference to this object to prevent its disposal while we
 644          * are messing with it.  Once we have the reference, the map is free
 645          * to be diddled.  Since objects reference their shadows (and copies),
 646          * they will stay around as well.
 647          *
 648          * Bump the paging-in-progress count to prevent size changes (e.g.
 649          * truncation operations) during I/O.
 650          */
 651         vm_object_reference_locked(fs.first_object);
 652         vm_object_pip_add(fs.first_object, 1);
 653
 654         fs.lookup_still_valid = true;
 655
 656         fs.first_m = NULL;
 657
 658         /*
 659          * Search for the page at object/offset.
 660          */
 661         fs.object = fs.first_object;
 662         fs.pindex = fs.first_pindex;
 663         while (TRUE) {
 664                 /*
 665                  * If the object is marked for imminent termination,
 666                  * we retry here, since the collapse pass has raced
 667                  * with us.  Otherwise, if we see terminally dead
 668                  * object, return fail.
 669                  */
 670                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 671                         dead = fs.object->type == OBJT_DEAD;
 672                         unlock_and_deallocate(&fs);
 673                         if (dead)
 674                                 return (KERN_PROTECTION_FAILURE);
 675                         pause("vmf_de", 1);
 676                         goto RetryFault;
 677                 }
 678
 679                 /*
 680                  * See if page is resident
 681                  */
 682                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 683                 if (fs.m != NULL) {
 684                         /*
 685                          * Wait/Retry if the page is busy.  We have to do this
 686                          * if the page is either exclusive or shared busy
 687                          * because the vm_pager may be using read busy for
 688                          * pageouts (and even pageins if it is the vnode
 689                          * pager), and we could end up trying to pagein and
 690                          * pageout the same page simultaneously.
 691                          *
 692                          * We can theoretically allow the busy case on a read
 693                          * fault if the page is marked valid, but since such
 694                          * pages are typically already pmap'd, putting that
 695                          * special case in might be more effort then it is
 696                          * worth.  We cannot under any circumstances mess
 697                          * around with a shared busied page except, perhaps,
 698                          * to pmap it.
 699                          */
 700                         if (vm_page_busied(fs.m)) {
 701                                 /*
 702                                  * Reference the page before unlocking and
 703                                  * sleeping so that the page daemon is less
 704                                  * likely to reclaim it.
 705                                  */
 706                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 707                                 if (fs.object != fs.first_object) {
 708                                         if (!VM_OBJECT_TRYWLOCK(
 709                                             fs.first_object)) {
 710                                                 VM_OBJECT_WUNLOCK(fs.object);
 711                                                 VM_OBJECT_WLOCK(fs.first_object);
 712                                                 VM_OBJECT_WLOCK(fs.object);
 713                                         }
 714                                         vm_page_lock(fs.first_m);
 715                                         vm_page_free(fs.first_m);
 716                                         vm_page_unlock(fs.first_m);
 717                                         vm_object_pip_wakeup(fs.first_object);
 718                                         VM_OBJECT_WUNLOCK(fs.first_object);
 719                                         fs.first_m = NULL;
 720                                 }
 721                                 unlock_map(&fs);
 722                                 if (fs.m == vm_page_lookup(fs.object,
 723                                     fs.pindex)) {
 724                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 725                                 }
 726                                 vm_object_pip_wakeup(fs.object);
 727                                 VM_OBJECT_WUNLOCK(fs.object);
 728                                 VM_CNT_INC(v_intrans);
 729                                 vm_object_deallocate(fs.first_object);
 730                                 goto RetryFault;
 731                         }
 732
 733                         /*
 734                          * Mark page busy for other processes, and the
 735                          * pagedaemon.  If it still isn't completely valid
 736                          * (readable), jump to readrest, else break-out ( we
 737                          * found the page ).
 738                          */
 739                         vm_page_xbusy(fs.m);
 740                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 741                                 goto readrest;
 742                         break; /* break to PAGE HAS BEEN FOUND */
 743                 }
 744                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 745
 746                 /*
 747                  * Page is not resident.  If the pager might contain the page
 748                  * or this is the beginning of the search, allocate a new
 749                  * page.  (Default objects are zero-fill, so there is no real
 750                  * pager for them.)
 751                  */
 752                 if (fs.object->type != OBJT_DEFAULT ||
 753                     fs.object == fs.first_object) {
 754                         if (fs.pindex >= fs.object->size) {
 755                                 unlock_and_deallocate(&fs);
 756                                 return (KERN_PROTECTION_FAILURE);
 757                         }
 758
 759                         if (fs.object == fs.first_object &&
 760                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 761                             fs.first_object->shadow_count == 0) {
 762                                 rv = vm_fault_populate(&fs, prot, fault_type,
 763                                     fault_flags, wired, m_hold);
 764                                 switch (rv) {
 765                                 case KERN_SUCCESS:
 766                                 case KERN_FAILURE:
 767                                         unlock_and_deallocate(&fs);
 768                                         return (rv);
 769                                 case KERN_RESOURCE_SHORTAGE:
 770                                         unlock_and_deallocate(&fs);
 771                                         goto RetryFault;
 772                                 case KERN_NOT_RECEIVER:
 773                                         /*
 774                                          * Pager's populate() method
 775                                          * returned VM_PAGER_BAD.
 776                                          */
 777                                         break;
 778                                 default:
 779                                         panic("inconsistent return codes");
 780                                 }
 781                         }
 782
 783                         /*
 784                          * Allocate a new page for this object/offset pair.
 785                          *
 786                          * Unlocked read of the p_flag is harmless. At
 787                          * worst, the P_KILLED might be not observed
 788                          * there, and allocation can fail, causing
 789                          * restart and new reading of the p_flag.
 790                          */
 791                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 792 #if VM_NRESERVLEVEL > 0
 793                                 vm_object_color(fs.object, atop(vaddr) -
 794                                     fs.pindex);
 795 #endif
 796                                 alloc_req = P_KILLED(curproc) ?
 797                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 798                                 if (fs.object->type != OBJT_VNODE &&
 799                                     fs.object->backing_object == NULL)
 800                                         alloc_req |= VM_ALLOC_ZERO;
 801                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 802                                     alloc_req);
 803                         }
 804                         if (fs.m == NULL) {
 805                                 unlock_and_deallocate(&fs);
 806                                 vm_waitpfault();
 807                                 goto RetryFault;
 808                         }
 809                 }
 810
 811 readrest:
 812                 /*
 813                  * At this point, we have either allocated a new page or found
 814                  * an existing page that is only partially valid.
 815                  *
 816                  * We hold a reference on the current object and the page is
 817                  * exclusive busied.
 818                  */
 819
 820                 /*
 821                  * If the pager for the current object might have the page,
 822                  * then determine the number of additional pages to read and
 823                  * potentially reprioritize previously read pages for earlier
 824                  * reclamation.  These operations should only be performed
 825                  * once per page fault.  Even if the current pager doesn't
 826                  * have the page, the number of additional pages to read will
 827                  * apply to subsequent objects in the shadow chain.
 828                  */
 829                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 830                     !P_KILLED(curproc)) {
 831                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 832                         era = fs.entry->read_ahead;
 833                         behavior = vm_map_entry_behavior(fs.entry);
 834                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 835                                 nera = 0;
 836                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 837                                 nera = VM_FAULT_READ_AHEAD_MAX;
 838                                 if (vaddr == fs.entry->next_read)
 839                                         vm_fault_dontneed(&fs, vaddr, nera);
 840                         } else if (vaddr == fs.entry->next_read) {
 841                                 /*
 842                                  * This is a sequential fault.  Arithmetically
 843                                  * increase the requested number of pages in
 844                                  * the read-ahead window.  The requested
 845                                  * number of pages is "# of sequential faults
 846                                  * x (read ahead min + 1) + read ahead min"
 847                                  */
 848                                 nera = VM_FAULT_READ_AHEAD_MIN;
 849                                 if (era > 0) {
 850                                         nera += era + 1;
 851                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 852                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 853                                 }
 854                                 if (era == VM_FAULT_READ_AHEAD_MAX)
 855                                         vm_fault_dontneed(&fs, vaddr, nera);
 856                         } else {
 857                                 /*
 858                                  * This is a non-sequential fault.
 859                                  */
 860                                 nera = 0;
 861                         }
 862                         if (era != nera) {
 863                                 /*
 864                                  * A read lock on the map suffices to update
 865                                  * the read ahead count safely.
 866                                  */
 867                                 fs.entry->read_ahead = nera;
 868                         }
 869
 870                         /*
 871                          * Prepare for unlocking the map.  Save the map
 872                          * entry's start and end addresses, which are used to
 873                          * optimize the size of the pager operation below.
 874                          * Even if the map entry's addresses change after
 875                          * unlocking the map, using the saved addresses is
 876                          * safe.
 877                          */
 878                         e_start = fs.entry->start;
 879                         e_end = fs.entry->end;
 880                 }
 881
 882                 /*
 883                  * Call the pager to retrieve the page if there is a chance
 884                  * that the pager has it, and potentially retrieve additional
 885                  * pages at the same time.
 886                  */
 887                 if (fs.object->type != OBJT_DEFAULT) {
 888                         /*
 889                          * Release the map lock before locking the vnode or
 890                          * sleeping in the pager.  (If the current object has
 891                          * a shadow, then an earlier iteration of this loop
 892                          * may have already unlocked the map.)
 893                          */
 894                         unlock_map(&fs);
 895
 896                         if (fs.object->type == OBJT_VNODE &&
 897                             (vp = fs.object->handle) != fs.vp) {
 898                                 /*
 899                                  * Perform an unlock in case the desired vnode
 900                                  * changed while the map was unlocked during a
 901                                  * retry.
 902                                  */
 903                                 unlock_vp(&fs);
 904
 905                                 locked = VOP_ISLOCKED(vp);
 906                                 if (locked != LK_EXCLUSIVE)
 907                                         locked = LK_SHARED;
 908
 909                                 /*
 910                                  * We must not sleep acquiring the vnode lock
 911                                  * while we have the page exclusive busied or
 912                                  * the object's paging-in-progress count
 913                                  * incremented.  Otherwise, we could deadlock.
 914                                  */
 915                                 error = vget(vp, locked | LK_CANRECURSE |
 916                                     LK_NOWAIT, curthread);
 917                                 if (error != 0) {
 918                                         vhold(vp);
 919                                         release_page(&fs);
 920                                         unlock_and_deallocate(&fs);
 921                                         error = vget(vp, locked | LK_RETRY |
 922                                             LK_CANRECURSE, curthread);
 923                                         vdrop(vp);
 924                                         fs.vp = vp;
 925                                         KASSERT(error == 0,
 926                                             ("vm_fault: vget failed"));
 927                                         goto RetryFault;
 928                                 }
 929                                 fs.vp = vp;
 930                         }
 931                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 932                             ("vm_fault: vnode-backed object mapped by system map"));
 933
 934                         /*
 935                          * Page in the requested page and hint the pager,
 936                          * that it may bring up surrounding pages.
 937                          */
 938                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
 939                             P_KILLED(curproc)) {
 940                                 behind = 0;
 941                                 ahead = 0;
 942                         } else {
 943                                 /* Is this a sequential fault? */
 944                                 if (nera > 0) {
 945                                         behind = 0;
 946                                         ahead = nera;
 947                                 } else {
 948                                         /*
 949                                          * Request a cluster of pages that is
 950                                          * aligned to a VM_FAULT_READ_DEFAULT
 951                                          * page offset boundary within the
 952                                          * object.  Alignment to a page offset
 953                                          * boundary is more likely to coincide
 954                                          * with the underlying file system
 955                                          * block than alignment to a virtual
 956                                          * address boundary.
 957                                          */
 958                                         cluster_offset = fs.pindex %
 959                                             VM_FAULT_READ_DEFAULT;
 960                                         behind = ulmin(cluster_offset,
 961                                             atop(vaddr - e_start));
 962                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
 963                                             cluster_offset;
 964                                 }
 965                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
 966                         }
 967                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
 968                             &behind, &ahead);
 969                         if (rv == VM_PAGER_OK) {
 970                                 faultcount = behind + 1 + ahead;
 971                                 hardfault = true;
 972                                 break; /* break to PAGE HAS BEEN FOUND */
 973                         }
 974                         if (rv == VM_PAGER_ERROR)
 975                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 976                                     curproc->p_pid, curproc->p_comm);
 977
 978                         /*
 979                          * If an I/O error occurred or the requested page was
 980                          * outside the range of the pager, clean up and return
 981                          * an error.
 982                          */
 983                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
 984                                 vm_page_lock(fs.m);
 985                                 if (fs.m->wire_count == 0)
 986                                         vm_page_free(fs.m);
 987                                 else
 988                                         vm_page_xunbusy_maybelocked(fs.m);
 989                                 vm_page_unlock(fs.m);
 990                                 fs.m = NULL;
 991                                 unlock_and_deallocate(&fs);
 992                                 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
 993                                     KERN_PROTECTION_FAILURE);
 994                         }
 995
 996                         /*
 997                          * The requested page does not exist at this object/
 998                          * offset.  Remove the invalid page from the object,
 999                          * waking up anyone waiting for it, and continue on to
1000                          * the next object.  However, if this is the top-level
1001                          * object, we must leave the busy page in place to
1002                          * prevent another process from rushing past us, and
1003                          * inserting the page in that object at the same time
1004                          * that we are.
1005                          */
1006                         if (fs.object != fs.first_object) {
1007                                 vm_page_lock(fs.m);
1008                                 if (fs.m->wire_count == 0)
1009                                         vm_page_free(fs.m);
1010                                 else
1011                                         vm_page_xunbusy_maybelocked(fs.m);
1012                                 vm_page_unlock(fs.m);
1013                                 fs.m = NULL;
1014                         }
1015                 }
1016
1017                 /*
1018                  * We get here if the object has default pager (or unwiring)
1019                  * or the pager doesn't have the page.
1020                  */
1021                 if (fs.object == fs.first_object)
1022                         fs.first_m = fs.m;
1023
1024                 /*
1025                  * Move on to the next object.  Lock the next object before
1026                  * unlocking the current one.
1027                  */
1028                 next_object = fs.object->backing_object;
1029                 if (next_object == NULL) {
1030                         /*
1031                          * If there's no object left, fill the page in the top
1032                          * object with zeros.
1033                          */
1034                         if (fs.object != fs.first_object) {
1035                                 vm_object_pip_wakeup(fs.object);
1036                                 VM_OBJECT_WUNLOCK(fs.object);
1037
1038                                 fs.object = fs.first_object;
1039                                 fs.pindex = fs.first_pindex;
1040                                 fs.m = fs.first_m;
1041                                 VM_OBJECT_WLOCK(fs.object);
1042                         }
1043                         fs.first_m = NULL;
1044
1045                         /*
1046                          * Zero the page if necessary and mark it valid.
1047                          */
1048                         if ((fs.m->flags & PG_ZERO) == 0) {
1049                                 pmap_zero_page(fs.m);
1050                         } else {
1051                                 VM_CNT_INC(v_ozfod);
1052                         }
1053                         VM_CNT_INC(v_zfod);
1054                         fs.m->valid = VM_PAGE_BITS_ALL;
1055                         /* Don't try to prefault neighboring pages. */
1056                         faultcount = 1;
1057                         break;  /* break to PAGE HAS BEEN FOUND */
1058                 } else {
1059                         KASSERT(fs.object != next_object,
1060                             ("object loop %p", next_object));
1061                         VM_OBJECT_WLOCK(next_object);
1062                         vm_object_pip_add(next_object, 1);
1063                         if (fs.object != fs.first_object)
1064                                 vm_object_pip_wakeup(fs.object);
1065                         fs.pindex +=
1066                             OFF_TO_IDX(fs.object->backing_object_offset);
1067                         VM_OBJECT_WUNLOCK(fs.object);
1068                         fs.object = next_object;
1069                 }
1070         }
1071
1072         vm_page_assert_xbusied(fs.m);
1073
1074         /*
1075          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1076          * is held.]
1077          */
1078
1079         /*
1080          * If the page is being written, but isn't already owned by the
1081          * top-level object, we have to copy it into a new page owned by the
1082          * top-level object.
1083          */
1084         if (fs.object != fs.first_object) {
1085                 /*
1086                  * We only really need to copy if we want to write it.
1087                  */
1088                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1089                         /*
1090                          * This allows pages to be virtually copied from a
1091                          * backing_object into the first_object, where the
1092                          * backing object has no other refs to it, and cannot
1093                          * gain any more refs.  Instead of a bcopy, we just
1094                          * move the page from the backing object to the
1095                          * first object.  Note that we must mark the page
1096                          * dirty in the first object so that it will go out
1097                          * to swap when needed.
1098                          */
1099                         is_first_object_locked = false;
1100                         if (
1101                                 /*
1102                                  * Only one shadow object
1103                                  */
1104                                 (fs.object->shadow_count == 1) &&
1105                                 /*
1106                                  * No COW refs, except us
1107                                  */
1108                                 (fs.object->ref_count == 1) &&
1109                                 /*
1110                                  * No one else can look this object up
1111                                  */
1112                                 (fs.object->handle == NULL) &&
1113                                 /*
1114                                  * No other ways to look the object up
1115                                  */
1116                                 ((fs.object->type == OBJT_DEFAULT) ||
1117                                  (fs.object->type == OBJT_SWAP)) &&
1118                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1119                                 /*
1120                                  * We don't chase down the shadow chain
1121                                  */
1122                             fs.object == fs.first_object->backing_object) {
1123                                 vm_page_lock(fs.m);
1124                                 vm_page_remque(fs.m);
1125                                 vm_page_remove(fs.m);
1126                                 vm_page_unlock(fs.m);
1127                                 vm_page_lock(fs.first_m);
1128                                 vm_page_replace_checked(fs.m, fs.first_object,
1129                                     fs.first_pindex, fs.first_m);
1130                                 vm_page_free(fs.first_m);
1131                                 vm_page_unlock(fs.first_m);
1132                                 vm_page_dirty(fs.m);
1133 #if VM_NRESERVLEVEL > 0
1134                                 /*
1135                                  * Rename the reservation.
1136                                  */
1137                                 vm_reserv_rename(fs.m, fs.first_object,
1138                                     fs.object, OFF_TO_IDX(
1139                                     fs.first_object->backing_object_offset));
1140 #endif
1141                                 /*
1142                                  * Removing the page from the backing object
1143                                  * unbusied it.
1144                                  */
1145                                 vm_page_xbusy(fs.m);
1146                                 fs.first_m = fs.m;
1147                                 fs.m = NULL;
1148                                 VM_CNT_INC(v_cow_optim);
1149                         } else {
1150                                 /*
1151                                  * Oh, well, lets copy it.
1152                                  */
1153                                 pmap_copy_page(fs.m, fs.first_m);
1154                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
1155                                 if ((fault_flags & VM_FAULT_WIRE) == 0) {
1156                                         prot &= ~VM_PROT_WRITE;
1157                                         fault_type &= ~VM_PROT_WRITE;
1158                                 }
1159                                 if (wired && (fault_flags &
1160                                     VM_FAULT_WIRE) == 0) {
1161                                         vm_page_lock(fs.first_m);
1162                                         vm_page_wire(fs.first_m);
1163                                         vm_page_unlock(fs.first_m);
1164
1165                                         vm_page_lock(fs.m);
1166                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1167                                         vm_page_unlock(fs.m);
1168                                 }
1169                                 /*
1170                                  * We no longer need the old page or object.
1171                                  */
1172                                 release_page(&fs);
1173                         }
1174                         /*
1175                          * fs.object != fs.first_object due to above
1176                          * conditional
1177                          */
1178                         vm_object_pip_wakeup(fs.object);
1179                         VM_OBJECT_WUNLOCK(fs.object);
1180                         /*
1181                          * Only use the new page below...
1182                          */
1183                         fs.object = fs.first_object;
1184                         fs.pindex = fs.first_pindex;
1185                         fs.m = fs.first_m;
1186                         if (!is_first_object_locked)
1187                                 VM_OBJECT_WLOCK(fs.object);
1188                         VM_CNT_INC(v_cow_faults);
1189                         curthread->td_cow++;
1190                 } else {
1191                         prot &= ~VM_PROT_WRITE;
1192                 }
1193         }
1194
1195         /*
1196          * We must verify that the maps have not changed since our last
1197          * lookup.
1198          */
1199         if (!fs.lookup_still_valid) {
1200                 if (!vm_map_trylock_read(fs.map)) {
1201                         release_page(&fs);
1202                         unlock_and_deallocate(&fs);
1203                         goto RetryFault;
1204                 }
1205                 fs.lookup_still_valid = true;
1206                 if (fs.map->timestamp != fs.map_generation) {
1207                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1208                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1209
1210                         /*
1211                          * If we don't need the page any longer, put it on the inactive
1212                          * list (the easiest thing to do here).  If no one needs it,
1213                          * pageout will grab it eventually.
1214                          */
1215                         if (result != KERN_SUCCESS) {
1216                                 release_page(&fs);
1217                                 unlock_and_deallocate(&fs);
1218
1219                                 /*
1220                                  * If retry of map lookup would have blocked then
1221                                  * retry fault from start.
1222                                  */
1223                                 if (result == KERN_FAILURE)
1224                                         goto RetryFault;
1225                                 return (result);
1226                         }
1227                         if ((retry_object != fs.first_object) ||
1228                             (retry_pindex != fs.first_pindex)) {
1229                                 release_page(&fs);
1230                                 unlock_and_deallocate(&fs);
1231                                 goto RetryFault;
1232                         }
1233
1234                         /*
1235                          * Check whether the protection has changed or the object has
1236                          * been copied while we left the map unlocked. Changing from
1237                          * read to write permission is OK - we leave the page
1238                          * write-protected, and catch the write fault. Changing from
1239                          * write to read permission means that we can't mark the page
1240                          * write-enabled after all.
1241                          */
1242                         prot &= retry_prot;
1243                         fault_type &= retry_prot;
1244                         if (prot == 0) {
1245                                 release_page(&fs);
1246                                 unlock_and_deallocate(&fs);
1247                                 goto RetryFault;
1248                         }
1249
1250                         /* Reassert because wired may have changed. */
1251                         KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1252                             ("!wired && VM_FAULT_WIRE"));
1253                 }
1254         }
1255
1256         /*
1257          * If the page was filled by a pager, save the virtual address that
1258          * should be faulted on next under a sequential access pattern to the
1259          * map entry.  A read lock on the map suffices to update this address
1260          * safely.
1261          */
1262         if (hardfault)
1263                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1264
1265         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1266         vm_page_assert_xbusied(fs.m);
1267
1268         /*
1269          * Page must be completely valid or it is not fit to
1270          * map into user space.  vm_pager_get_pages() ensures this.
1271          */
1272         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1273             ("vm_fault: page %p partially invalid", fs.m));
1274         VM_OBJECT_WUNLOCK(fs.object);
1275
1276         /*
1277          * Put this page into the physical map.  We had to do the unlock above
1278          * because pmap_enter() may sleep.  We don't put the page
1279          * back on the active queue until later so that the pageout daemon
1280          * won't find it (yet).
1281          */
1282         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1283             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1284         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1285             wired == 0)
1286                 vm_fault_prefault(&fs, vaddr,
1287                     faultcount > 0 ? behind : PFBAK,
1288                     faultcount > 0 ? ahead : PFFOR, false);
1289         VM_OBJECT_WLOCK(fs.object);
1290         vm_page_lock(fs.m);
1291
1292         /*
1293          * If the page is not wired down, then put it where the pageout daemon
1294          * can find it.
1295          */
1296         if ((fault_flags & VM_FAULT_WIRE) != 0)
1297                 vm_page_wire(fs.m);
1298         else
1299                 vm_page_activate(fs.m);
1300         if (m_hold != NULL) {
1301                 *m_hold = fs.m;
1302                 vm_page_hold(fs.m);
1303         }
1304         vm_page_unlock(fs.m);
1305         vm_page_xunbusy(fs.m);
1306
1307         /*
1308          * Unlock everything, and return
1309          */
1310         unlock_and_deallocate(&fs);
1311         if (hardfault) {
1312                 VM_CNT_INC(v_io_faults);
1313                 curthread->td_ru.ru_majflt++;
1314 #ifdef RACCT
1315                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1316                         PROC_LOCK(curproc);
1317                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1318                                 racct_add_force(curproc, RACCT_WRITEBPS,
1319                                     PAGE_SIZE + behind * PAGE_SIZE);
1320                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1321                         } else {
1322                                 racct_add_force(curproc, RACCT_READBPS,
1323                                     PAGE_SIZE + ahead * PAGE_SIZE);
1324                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1325                         }
1326                         PROC_UNLOCK(curproc);
1327                 }
1328 #endif
1329         } else
1330                 curthread->td_ru.ru_minflt++;
1331
1332         return (KERN_SUCCESS);
1333 }
1334
1335 /*
1336  * Speed up the reclamation of pages that precede the faulting pindex within
1337  * the first object of the shadow chain.  Essentially, perform the equivalent
1338  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1339  * the faulting pindex by the cluster size when the pages read by vm_fault()
1340  * cross a cluster-size boundary.  The cluster size is the greater of the
1341  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1342  *
1343  * When "fs->first_object" is a shadow object, the pages in the backing object
1344  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1345  * function must only be concerned with pages in the first object.
1346  */
1347 static void
1348 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1349 {
1350         vm_map_entry_t entry;
1351         vm_object_t first_object, object;
1352         vm_offset_t end, start;
1353         vm_page_t m, m_next;
1354         vm_pindex_t pend, pstart;
1355         vm_size_t size;
1356
1357         object = fs->object;
1358         VM_OBJECT_ASSERT_WLOCKED(object);
1359         first_object = fs->first_object;
1360         if (first_object != object) {
1361                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1362                         VM_OBJECT_WUNLOCK(object);
1363                         VM_OBJECT_WLOCK(first_object);
1364                         VM_OBJECT_WLOCK(object);
1365                 }
1366         }
1367         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1368         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1369                 size = VM_FAULT_DONTNEED_MIN;
1370                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1371                         size = pagesizes[1];
1372                 end = rounddown2(vaddr, size);
1373                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1374                     (entry = fs->entry)->start < end) {
1375                         if (end - entry->start < size)
1376                                 start = entry->start;
1377                         else
1378                                 start = end - size;
1379                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1380                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1381                             entry->start);
1382                         m_next = vm_page_find_least(first_object, pstart);
1383                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1384                             entry->start);
1385                         while ((m = m_next) != NULL && m->pindex < pend) {
1386                                 m_next = TAILQ_NEXT(m, listq);
1387                                 if (m->valid != VM_PAGE_BITS_ALL ||
1388                                     vm_page_busied(m))
1389                                         continue;
1390
1391                                 /*
1392                                  * Don't clear PGA_REFERENCED, since it would
1393                                  * likely represent a reference by a different
1394                                  * process.
1395                                  *
1396                                  * Typically, at this point, prefetched pages
1397                                  * are still in the inactive queue.  Only
1398                                  * pages that triggered page faults are in the
1399                                  * active queue.
1400                                  */
1401                                 vm_page_lock(m);
1402                                 if (!vm_page_inactive(m))
1403                                         vm_page_deactivate(m);
1404                                 vm_page_unlock(m);
1405                         }
1406                 }
1407         }
1408         if (first_object != object)
1409                 VM_OBJECT_WUNLOCK(first_object);
1410 }
1411
1412 /*
1413  * vm_fault_prefault provides a quick way of clustering
1414  * pagefaults into a processes address space.  It is a "cousin"
1415  * of vm_map_pmap_enter, except it runs at page fault time instead
1416  * of mmap time.
1417  */
1418 static void
1419 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1420     int backward, int forward, bool obj_locked)
1421 {
1422         pmap_t pmap;
1423         vm_map_entry_t entry;
1424         vm_object_t backing_object, lobject;
1425         vm_offset_t addr, starta;
1426         vm_pindex_t pindex;
1427         vm_page_t m;
1428         int i;
1429
1430         pmap = fs->map->pmap;
1431         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1432                 return;
1433
1434         entry = fs->entry;
1435
1436         if (addra < backward * PAGE_SIZE) {
1437                 starta = entry->start;
1438         } else {
1439                 starta = addra - backward * PAGE_SIZE;
1440                 if (starta < entry->start)
1441                         starta = entry->start;
1442         }
1443
1444         /*
1445          * Generate the sequence of virtual addresses that are candidates for
1446          * prefaulting in an outward spiral from the faulting virtual address,
1447          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1448          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1449          * If the candidate address doesn't have a backing physical page, then
1450          * the loop immediately terminates.
1451          */
1452         for (i = 0; i < 2 * imax(backward, forward); i++) {
1453                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1454                     PAGE_SIZE);
1455                 if (addr > addra + forward * PAGE_SIZE)
1456                         addr = 0;
1457
1458                 if (addr < starta || addr >= entry->end)
1459                         continue;
1460
1461                 if (!pmap_is_prefaultable(pmap, addr))
1462                         continue;
1463
1464                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1465                 lobject = entry->object.vm_object;
1466                 if (!obj_locked)
1467                         VM_OBJECT_RLOCK(lobject);
1468                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1469                     lobject->type == OBJT_DEFAULT &&
1470                     (backing_object = lobject->backing_object) != NULL) {
1471                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1472                             0, ("vm_fault_prefault: unaligned object offset"));
1473                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1474                         VM_OBJECT_RLOCK(backing_object);
1475                         if (!obj_locked || lobject != entry->object.vm_object)
1476                                 VM_OBJECT_RUNLOCK(lobject);
1477                         lobject = backing_object;
1478                 }
1479                 if (m == NULL) {
1480                         if (!obj_locked || lobject != entry->object.vm_object)
1481                                 VM_OBJECT_RUNLOCK(lobject);
1482                         break;
1483                 }
1484                 if (m->valid == VM_PAGE_BITS_ALL &&
1485                     (m->flags & PG_FICTITIOUS) == 0)
1486                         pmap_enter_quick(pmap, addr, m, entry->protection);
1487                 if (!obj_locked || lobject != entry->object.vm_object)
1488                         VM_OBJECT_RUNLOCK(lobject);
1489         }
1490 }
1491
1492 /*
1493  * Hold each of the physical pages that are mapped by the specified range of
1494  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1495  * and allow the specified types of access, "prot".  If all of the implied
1496  * pages are successfully held, then the number of held pages is returned
1497  * together with pointers to those pages in the array "ma".  However, if any
1498  * of the pages cannot be held, -1 is returned.
1499  */
1500 int
1501 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1502     vm_prot_t prot, vm_page_t *ma, int max_count)
1503 {
1504         vm_offset_t end, va;
1505         vm_page_t *mp;
1506         int count;
1507         boolean_t pmap_failed;
1508
1509         if (len == 0)
1510                 return (0);
1511         end = round_page(addr + len);
1512         addr = trunc_page(addr);
1513
1514         /*
1515          * Check for illegal addresses.
1516          */
1517         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1518                 return (-1);
1519
1520         if (atop(end - addr) > max_count)
1521                 panic("vm_fault_quick_hold_pages: count > max_count");
1522         count = atop(end - addr);
1523
1524         /*
1525          * Most likely, the physical pages are resident in the pmap, so it is
1526          * faster to try pmap_extract_and_hold() first.
1527          */
1528         pmap_failed = FALSE;
1529         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1530                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1531                 if (*mp == NULL)
1532                         pmap_failed = TRUE;
1533                 else if ((prot & VM_PROT_WRITE) != 0 &&
1534                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1535                         /*
1536                          * Explicitly dirty the physical page.  Otherwise, the
1537                          * caller's changes may go unnoticed because they are
1538                          * performed through an unmanaged mapping or by a DMA
1539                          * operation.
1540                          *
1541                          * The object lock is not held here.
1542                          * See vm_page_clear_dirty_mask().
1543                          */
1544                         vm_page_dirty(*mp);
1545                 }
1546         }
1547         if (pmap_failed) {
1548                 /*
1549                  * One or more pages could not be held by the pmap.  Either no
1550                  * page was mapped at the specified virtual address or that
1551                  * mapping had insufficient permissions.  Attempt to fault in
1552                  * and hold these pages.
1553                  *
1554                  * If vm_fault_disable_pagefaults() was called,
1555                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1556                  * acquire MD VM locks, which means we must not call
1557                  * vm_fault_hold().  Some (out of tree) callers mark
1558                  * too wide a code area with vm_fault_disable_pagefaults()
1559                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1560                  * the proper behaviour explicitly.
1561                  */
1562                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1563                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1564                         goto error;
1565                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1566                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1567                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1568                                 goto error;
1569         }
1570         return (count);
1571 error:
1572         for (mp = ma; mp < ma + count; mp++)
1573                 if (*mp != NULL) {
1574                         vm_page_lock(*mp);
1575                         vm_page_unhold(*mp);
1576                         vm_page_unlock(*mp);
1577                 }
1578         return (-1);
1579 }
1580
1581 /*
1582  *      Routine:
1583  *              vm_fault_copy_entry
1584  *      Function:
1585  *              Create new shadow object backing dst_entry with private copy of
1586  *              all underlying pages. When src_entry is equal to dst_entry,
1587  *              function implements COW for wired-down map entry. Otherwise,
1588  *              it forks wired entry into dst_map.
1589  *
1590  *      In/out conditions:
1591  *              The source and destination maps must be locked for write.
1592  *              The source map entry must be wired down (or be a sharing map
1593  *              entry corresponding to a main map entry that is wired down).
1594  */
1595 void
1596 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1597     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1598     vm_ooffset_t *fork_charge)
1599 {
1600         vm_object_t backing_object, dst_object, object, src_object;
1601         vm_pindex_t dst_pindex, pindex, src_pindex;
1602         vm_prot_t access, prot;
1603         vm_offset_t vaddr;
1604         vm_page_t dst_m;
1605         vm_page_t src_m;
1606         boolean_t upgrade;
1607
1608 #ifdef  lint
1609         src_map++;
1610 #endif  /* lint */
1611
1612         upgrade = src_entry == dst_entry;
1613         access = prot = dst_entry->protection;
1614
1615         src_object = src_entry->object.vm_object;
1616         src_pindex = OFF_TO_IDX(src_entry->offset);
1617
1618         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1619                 dst_object = src_object;
1620                 vm_object_reference(dst_object);
1621         } else {
1622                 /*
1623                  * Create the top-level object for the destination entry. (Doesn't
1624                  * actually shadow anything - we copy the pages directly.)
1625                  */
1626                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1627                     atop(dst_entry->end - dst_entry->start));
1628 #if VM_NRESERVLEVEL > 0
1629                 dst_object->flags |= OBJ_COLORED;
1630                 dst_object->pg_color = atop(dst_entry->start);
1631 #endif
1632         }
1633
1634         VM_OBJECT_WLOCK(dst_object);
1635         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1636             ("vm_fault_copy_entry: vm_object not NULL"));
1637         if (src_object != dst_object) {
1638                 dst_object->domain = src_object->domain;
1639                 dst_entry->object.vm_object = dst_object;
1640                 dst_entry->offset = 0;
1641                 dst_object->charge = dst_entry->end - dst_entry->start;
1642         }
1643         if (fork_charge != NULL) {
1644                 KASSERT(dst_entry->cred == NULL,
1645                     ("vm_fault_copy_entry: leaked swp charge"));
1646                 dst_object->cred = curthread->td_ucred;
1647                 crhold(dst_object->cred);
1648                 *fork_charge += dst_object->charge;
1649         } else if (dst_object->cred == NULL) {
1650                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1651                     dst_entry));
1652                 dst_object->cred = dst_entry->cred;
1653                 dst_entry->cred = NULL;
1654         }
1655
1656         /*
1657          * If not an upgrade, then enter the mappings in the pmap as
1658          * read and/or execute accesses.  Otherwise, enter them as
1659          * write accesses.
1660          *
1661          * A writeable large page mapping is only created if all of
1662          * the constituent small page mappings are modified. Marking
1663          * PTEs as modified on inception allows promotion to happen
1664          * without taking potentially large number of soft faults.
1665          */
1666         if (!upgrade)
1667                 access &= ~VM_PROT_WRITE;
1668
1669         /*
1670          * Loop through all of the virtual pages within the entry's
1671          * range, copying each page from the source object to the
1672          * destination object.  Since the source is wired, those pages
1673          * must exist.  In contrast, the destination is pageable.
1674          * Since the destination object doesn't share any backing storage
1675          * with the source object, all of its pages must be dirtied,
1676          * regardless of whether they can be written.
1677          */
1678         for (vaddr = dst_entry->start, dst_pindex = 0;
1679             vaddr < dst_entry->end;
1680             vaddr += PAGE_SIZE, dst_pindex++) {
1681 again:
1682                 /*
1683                  * Find the page in the source object, and copy it in.
1684                  * Because the source is wired down, the page will be
1685                  * in memory.
1686                  */
1687                 if (src_object != dst_object)
1688                         VM_OBJECT_RLOCK(src_object);
1689                 object = src_object;
1690                 pindex = src_pindex + dst_pindex;
1691                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1692                     (backing_object = object->backing_object) != NULL) {
1693                         /*
1694                          * Unless the source mapping is read-only or
1695                          * it is presently being upgraded from
1696                          * read-only, the first object in the shadow
1697                          * chain should provide all of the pages.  In
1698                          * other words, this loop body should never be
1699                          * executed when the source mapping is already
1700                          * read/write.
1701                          */
1702                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1703                             upgrade,
1704                             ("vm_fault_copy_entry: main object missing page"));
1705
1706                         VM_OBJECT_RLOCK(backing_object);
1707                         pindex += OFF_TO_IDX(object->backing_object_offset);
1708                         if (object != dst_object)
1709                                 VM_OBJECT_RUNLOCK(object);
1710                         object = backing_object;
1711                 }
1712                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1713
1714                 if (object != dst_object) {
1715                         /*
1716                          * Allocate a page in the destination object.
1717                          */
1718                         dst_m = vm_page_alloc(dst_object, (src_object ==
1719                             dst_object ? src_pindex : 0) + dst_pindex,
1720                             VM_ALLOC_NORMAL);
1721                         if (dst_m == NULL) {
1722                                 VM_OBJECT_WUNLOCK(dst_object);
1723                                 VM_OBJECT_RUNLOCK(object);
1724                                 vm_wait(dst_object);
1725                                 VM_OBJECT_WLOCK(dst_object);
1726                                 goto again;
1727                         }
1728                         pmap_copy_page(src_m, dst_m);
1729                         VM_OBJECT_RUNLOCK(object);
1730                         dst_m->valid = VM_PAGE_BITS_ALL;
1731                         dst_m->dirty = VM_PAGE_BITS_ALL;
1732                 } else {
1733                         dst_m = src_m;
1734                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1735                                 goto again;
1736                         vm_page_xbusy(dst_m);
1737                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1738                             ("invalid dst page %p", dst_m));
1739                 }
1740                 VM_OBJECT_WUNLOCK(dst_object);
1741
1742                 /*
1743                  * Enter it in the pmap. If a wired, copy-on-write
1744                  * mapping is being replaced by a write-enabled
1745                  * mapping, then wire that new mapping.
1746                  */
1747                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1748                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1749
1750                 /*
1751                  * Mark it no longer busy, and put it on the active list.
1752                  */
1753                 VM_OBJECT_WLOCK(dst_object);
1754
1755                 if (upgrade) {
1756                         if (src_m != dst_m) {
1757                                 vm_page_lock(src_m);
1758                                 vm_page_unwire(src_m, PQ_INACTIVE);
1759                                 vm_page_unlock(src_m);
1760                                 vm_page_lock(dst_m);
1761                                 vm_page_wire(dst_m);
1762                                 vm_page_unlock(dst_m);
1763                         } else {
1764                                 KASSERT(dst_m->wire_count > 0,
1765                                     ("dst_m %p is not wired", dst_m));
1766                         }
1767                 } else {
1768                         vm_page_lock(dst_m);
1769                         vm_page_activate(dst_m);
1770                         vm_page_unlock(dst_m);
1771                 }
1772                 vm_page_xunbusy(dst_m);
1773         }
1774         VM_OBJECT_WUNLOCK(dst_object);
1775         if (upgrade) {
1776                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1777                 vm_object_deallocate(src_object);
1778         }
1779 }
1780
1781 /*
1782  * Block entry into the machine-independent layer's page fault handler by
1783  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1784  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1785  * spurious page faults.
1786  */
1787 int
1788 vm_fault_disable_pagefaults(void)
1789 {
1790
1791         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1792 }
1793
1794 void
1795 vm_fault_enable_pagefaults(int save)
1796 {
1797
1798         curthread_pflags_restore(save);
1799 }