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/mutex.h>
  88 #include <sys/proc.h>
  89 #include <sys/racct.h>
  90 #include <sys/resourcevar.h>
  91 #include <sys/rwlock.h>
  92 #include <sys/sysctl.h>
  93 #include <sys/vmmeter.h>
  94 #include <sys/vnode.h>
  95 #ifdef KTRACE
  96 #include <sys/ktrace.h>
  97 #endif
  98
  99 #include <vm/vm.h>
 100 #include <vm/vm_param.h>
 101 #include <vm/pmap.h>
 102 #include <vm/vm_map.h>
 103 #include <vm/vm_object.h>
 104 #include <vm/vm_page.h>
 105 #include <vm/vm_pageout.h>
 106 #include <vm/vm_kern.h>
 107 #include <vm/vm_pager.h>
 108 #include <vm/vm_extern.h>
 109 #include <vm/vm_reserv.h>
 110
 111 #define PFBAK 4
 112 #define PFFOR 4
 113
 114 #define VM_FAULT_READ_DEFAULT   (1 + VM_FAULT_READ_AHEAD_INIT)
 115 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 116
 117 #define VM_FAULT_DONTNEED_MIN   1048576
 118
 119 struct faultstate {
 120         vm_page_t m;
 121         vm_object_t object;
 122         vm_pindex_t pindex;
 123         vm_page_t first_m;
 124         vm_object_t     first_object;
 125         vm_pindex_t first_pindex;
 126         vm_map_t map;
 127         vm_map_entry_t entry;
 128         int map_generation;
 129         bool lookup_still_valid;
 130         struct vnode *vp;
 131 };
 132
 133 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 134             int ahead);
 135 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 136             int backward, int forward, bool obj_locked);
 137
 138 static inline void
 139 release_page(struct faultstate *fs)
 140 {
 141
 142         vm_page_xunbusy(fs->m);
 143         vm_page_lock(fs->m);
 144         vm_page_deactivate(fs->m);
 145         vm_page_unlock(fs->m);
 146         fs->m = NULL;
 147 }
 148
 149 static inline void
 150 unlock_map(struct faultstate *fs)
 151 {
 152
 153         if (fs->lookup_still_valid) {
 154                 vm_map_lookup_done(fs->map, fs->entry);
 155                 fs->lookup_still_valid = false;
 156         }
 157 }
 158
 159 static void
 160 unlock_vp(struct faultstate *fs)
 161 {
 162
 163         if (fs->vp != NULL) {
 164                 vput(fs->vp);
 165                 fs->vp = NULL;
 166         }
 167 }
 168
 169 static void
 170 unlock_and_deallocate(struct faultstate *fs)
 171 {
 172
 173         vm_object_pip_wakeup(fs->object);
 174         VM_OBJECT_WUNLOCK(fs->object);
 175         if (fs->object != fs->first_object) {
 176                 VM_OBJECT_WLOCK(fs->first_object);
 177                 vm_page_lock(fs->first_m);
 178                 vm_page_free(fs->first_m);
 179                 vm_page_unlock(fs->first_m);
 180                 vm_object_pip_wakeup(fs->first_object);
 181                 VM_OBJECT_WUNLOCK(fs->first_object);
 182                 fs->first_m = NULL;
 183         }
 184         vm_object_deallocate(fs->first_object);
 185         unlock_map(fs);
 186         unlock_vp(fs);
 187 }
 188
 189 static void
 190 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 191     vm_prot_t fault_type, int fault_flags, bool set_wd)
 192 {
 193         bool need_dirty;
 194
 195         if (((prot & VM_PROT_WRITE) == 0 &&
 196             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 197             (m->oflags & VPO_UNMANAGED) != 0)
 198                 return;
 199
 200         VM_OBJECT_ASSERT_LOCKED(m->object);
 201
 202         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 203             (fault_flags & VM_FAULT_WIRE) == 0) ||
 204             (fault_flags & VM_FAULT_DIRTY) != 0;
 205
 206         if (set_wd)
 207                 vm_object_set_writeable_dirty(m->object);
 208         else
 209                 /*
 210                  * If two callers of vm_fault_dirty() with set_wd ==
 211                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 212                  * flag set, other with flag clear, race, it is
 213                  * possible for the no-NOSYNC thread to see m->dirty
 214                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 215                  * around manipulation of VPO_NOSYNC and
 216                  * vm_page_dirty() call, to avoid the race and keep
 217                  * m->oflags consistent.
 218                  */
 219                 vm_page_lock(m);
 220
 221         /*
 222          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 223          * if the page is already dirty to prevent data written with
 224          * the expectation of being synced from not being synced.
 225          * Likewise if this entry does not request NOSYNC then make
 226          * sure the page isn't marked NOSYNC.  Applications sharing
 227          * data should use the same flags to avoid ping ponging.
 228          */
 229         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 230                 if (m->dirty == 0) {
 231                         m->oflags |= VPO_NOSYNC;
 232                 }
 233         } else {
 234                 m->oflags &= ~VPO_NOSYNC;
 235         }
 236
 237         /*
 238          * If the fault is a write, we know that this page is being
 239          * written NOW so dirty it explicitly to save on
 240          * pmap_is_modified() calls later.
 241          *
 242          * Also, since the page is now dirty, we can possibly tell
 243          * the pager to release any swap backing the page.  Calling
 244          * the pager requires a write lock on the object.
 245          */
 246         if (need_dirty)
 247                 vm_page_dirty(m);
 248         if (!set_wd)
 249                 vm_page_unlock(m);
 250         else if (need_dirty)
 251                 vm_pager_page_unswapped(m);
 252 }
 253
 254 static void
 255 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m)
 256 {
 257
 258         if (m_hold != NULL) {
 259                 *m_hold = m;
 260                 vm_page_lock(m);
 261                 vm_page_wire(m);
 262                 vm_page_unlock(m);
 263         }
 264 }
 265
 266 /*
 267  * Unlocks fs.first_object and fs.map on success.
 268  */
 269 static int
 270 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 271     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 272 {
 273         vm_page_t m, m_map;
 274 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 275     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 276     VM_NRESERVLEVEL > 0
 277         vm_page_t m_super;
 278         int flags;
 279 #endif
 280         int psind, rv;
 281
 282         MPASS(fs->vp == NULL);
 283         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 284         /* A busy page can be mapped for read|execute access. */
 285         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 286             vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 287                 return (KERN_FAILURE);
 288         m_map = m;
 289         psind = 0;
 290 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 291     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 292     VM_NRESERVLEVEL > 0
 293         if ((m->flags & PG_FICTITIOUS) == 0 &&
 294             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 295             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 296             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 297             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 298             (pagesizes[m_super->psind] - 1)) && !wired &&
 299             pmap_ps_enabled(fs->map->pmap)) {
 300                 flags = PS_ALL_VALID;
 301                 if ((prot & VM_PROT_WRITE) != 0) {
 302                         /*
 303                          * Create a superpage mapping allowing write access
 304                          * only if none of the constituent pages are busy and
 305                          * all of them are already dirty (except possibly for
 306                          * the page that was faulted on).
 307                          */
 308                         flags |= PS_NONE_BUSY;
 309                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 310                                 flags |= PS_ALL_DIRTY;
 311                 }
 312                 if (vm_page_ps_test(m_super, flags, m)) {
 313                         m_map = m_super;
 314                         psind = m_super->psind;
 315                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 316                         /* Preset the modified bit for dirty superpages. */
 317                         if ((flags & PS_ALL_DIRTY) != 0)
 318                                 fault_type |= VM_PROT_WRITE;
 319                 }
 320         }
 321 #endif
 322         rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
 323             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
 324         if (rv != KERN_SUCCESS)
 325                 return (rv);
 326         vm_fault_fill_hold(m_hold, m);
 327         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
 328         if (psind == 0 && !wired)
 329                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 330         VM_OBJECT_RUNLOCK(fs->first_object);
 331         vm_map_lookup_done(fs->map, fs->entry);
 332         curthread->td_ru.ru_minflt++;
 333         return (KERN_SUCCESS);
 334 }
 335
 336 static void
 337 vm_fault_restore_map_lock(struct faultstate *fs)
 338 {
 339
 340         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 341         MPASS(fs->first_object->paging_in_progress > 0);
 342
 343         if (!vm_map_trylock_read(fs->map)) {
 344                 VM_OBJECT_WUNLOCK(fs->first_object);
 345                 vm_map_lock_read(fs->map);
 346                 VM_OBJECT_WLOCK(fs->first_object);
 347         }
 348         fs->lookup_still_valid = true;
 349 }
 350
 351 static void
 352 vm_fault_populate_check_page(vm_page_t m)
 353 {
 354
 355         /*
 356          * Check each page to ensure that the pager is obeying the
 357          * interface: the page must be installed in the object, fully
 358          * valid, and exclusively busied.
 359          */
 360         MPASS(m != NULL);
 361         MPASS(m->valid == VM_PAGE_BITS_ALL);
 362         MPASS(vm_page_xbusied(m));
 363 }
 364
 365 static void
 366 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 367     vm_pindex_t last)
 368 {
 369         vm_page_t m;
 370         vm_pindex_t pidx;
 371
 372         VM_OBJECT_ASSERT_WLOCKED(object);
 373         MPASS(first <= last);
 374         for (pidx = first, m = vm_page_lookup(object, pidx);
 375             pidx <= last; pidx++, m = vm_page_next(m)) {
 376                 vm_fault_populate_check_page(m);
 377                 vm_page_lock(m);
 378                 vm_page_deactivate(m);
 379                 vm_page_unlock(m);
 380                 vm_page_xunbusy(m);
 381         }
 382 }
 383
 384 static int
 385 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 386     int fault_flags, boolean_t wired, vm_page_t *m_hold)
 387 {
 388         struct mtx *m_mtx;
 389         vm_offset_t vaddr;
 390         vm_page_t m;
 391         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 392         int i, npages, psind, rv;
 393
 394         MPASS(fs->object == fs->first_object);
 395         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 396         MPASS(fs->first_object->paging_in_progress > 0);
 397         MPASS(fs->first_object->backing_object == NULL);
 398         MPASS(fs->lookup_still_valid);
 399
 400         pager_first = OFF_TO_IDX(fs->entry->offset);
 401         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 402         unlock_map(fs);
 403         unlock_vp(fs);
 404
 405         /*
 406          * Call the pager (driver) populate() method.
 407          *
 408          * There is no guarantee that the method will be called again
 409          * if the current fault is for read, and a future fault is
 410          * for write.  Report the entry's maximum allowed protection
 411          * to the driver.
 412          */
 413         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 414             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 415
 416         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 417         if (rv == VM_PAGER_BAD) {
 418                 /*
 419                  * VM_PAGER_BAD is the backdoor for a pager to request
 420                  * normal fault handling.
 421                  */
 422                 vm_fault_restore_map_lock(fs);
 423                 if (fs->map->timestamp != fs->map_generation)
 424                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 425                 return (KERN_NOT_RECEIVER);
 426         }
 427         if (rv != VM_PAGER_OK)
 428                 return (KERN_FAILURE); /* AKA SIGSEGV */
 429
 430         /* Ensure that the driver is obeying the interface. */
 431         MPASS(pager_first <= pager_last);
 432         MPASS(fs->first_pindex <= pager_last);
 433         MPASS(fs->first_pindex >= pager_first);
 434         MPASS(pager_last < fs->first_object->size);
 435
 436         vm_fault_restore_map_lock(fs);
 437         if (fs->map->timestamp != fs->map_generation) {
 438                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 439                     pager_last);
 440                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 441         }
 442
 443         /*
 444          * The map is unchanged after our last unlock.  Process the fault.
 445          *
 446          * The range [pager_first, pager_last] that is given to the
 447          * pager is only a hint.  The pager may populate any range
 448          * within the object that includes the requested page index.
 449          * In case the pager expanded the range, clip it to fit into
 450          * the map entry.
 451          */
 452         map_first = OFF_TO_IDX(fs->entry->offset);
 453         if (map_first > pager_first) {
 454                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 455                     map_first - 1);
 456                 pager_first = map_first;
 457         }
 458         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 459         if (map_last < pager_last) {
 460                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 461                     pager_last);
 462                 pager_last = map_last;
 463         }
 464         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 465             pidx <= pager_last;
 466             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 467                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 468 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 469     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
 470                 psind = m->psind;
 471                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 472                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 473                     !pmap_ps_enabled(fs->map->pmap) || wired))
 474                         psind = 0;
 475 #else
 476                 psind = 0;
 477 #endif
 478                 npages = atop(pagesizes[psind]);
 479                 for (i = 0; i < npages; i++) {
 480                         vm_fault_populate_check_page(&m[i]);
 481                         vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
 482                             fault_flags, true);
 483                 }
 484                 VM_OBJECT_WUNLOCK(fs->first_object);
 485                 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
 486                     (wired ? PMAP_ENTER_WIRED : 0), psind);
 487 #if defined(__amd64__)
 488                 if (psind > 0 && rv == KERN_FAILURE) {
 489                         for (i = 0; i < npages; i++) {
 490                                 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
 491                                     &m[i], prot, fault_type |
 492                                     (wired ? PMAP_ENTER_WIRED : 0), 0);
 493                                 MPASS(rv == KERN_SUCCESS);
 494                         }
 495                 }
 496 #else
 497                 MPASS(rv == KERN_SUCCESS);
 498 #endif
 499                 VM_OBJECT_WLOCK(fs->first_object);
 500                 m_mtx = NULL;
 501                 for (i = 0; i < npages; i++) {
 502                         vm_page_change_lock(&m[i], &m_mtx);
 503                         if ((fault_flags & VM_FAULT_WIRE) != 0)
 504                                 vm_page_wire(&m[i]);
 505                         else
 506                                 vm_page_activate(&m[i]);
 507                         if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
 508                                 *m_hold = &m[i];
 509                                 vm_page_wire(&m[i]);
 510                         }
 511                         vm_page_xunbusy_maybelocked(&m[i]);
 512                 }
 513                 if (m_mtx != NULL)
 514                         mtx_unlock(m_mtx);
 515         }
 516         curthread->td_ru.ru_majflt++;
 517         return (KERN_SUCCESS);
 518 }
 519
 520 /*
 521  *      vm_fault:
 522  *
 523  *      Handle a page fault occurring at the given address,
 524  *      requiring the given permissions, in the map specified.
 525  *      If successful, the page is inserted into the
 526  *      associated physical map.
 527  *
 528  *      NOTE: the given address should be truncated to the
 529  *      proper page address.
 530  *
 531  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 532  *      a standard error specifying why the fault is fatal is returned.
 533  *
 534  *      The map in question must be referenced, and remains so.
 535  *      Caller may hold no locks.
 536  */
 537 int
 538 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 539     int fault_flags)
 540 {
 541         struct thread *td;
 542         int result;
 543
 544         td = curthread;
 545         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 546                 return (KERN_PROTECTION_FAILURE);
 547 #ifdef KTRACE
 548         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 549                 ktrfault(vaddr, fault_type);
 550 #endif
 551         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 552             NULL);
 553 #ifdef KTRACE
 554         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 555                 ktrfaultend(result);
 556 #endif
 557         return (result);
 558 }
 559
 560 int
 561 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 562     int fault_flags, vm_page_t *m_hold)
 563 {
 564         struct faultstate fs;
 565         struct vnode *vp;
 566         struct domainset *dset;
 567         struct mtx *mtx;
 568         vm_object_t next_object, retry_object;
 569         vm_offset_t e_end, e_start;
 570         vm_pindex_t retry_pindex;
 571         vm_prot_t prot, retry_prot;
 572         int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
 573         int locked, nera, result, rv;
 574         u_char behavior;
 575         boolean_t wired;        /* Passed by reference. */
 576         bool dead, hardfault, is_first_object_locked;
 577
 578         VM_CNT_INC(v_vm_faults);
 579         fs.vp = NULL;
 580         faultcount = 0;
 581         nera = -1;
 582         hardfault = false;
 583
 584 RetryFault:;
 585
 586         /*
 587          * Find the backing store object and offset into it to begin the
 588          * search.
 589          */
 590         fs.map = map;
 591         result = vm_map_lookup(&fs.map, vaddr, fault_type |
 592             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 593             &fs.first_pindex, &prot, &wired);
 594         if (result != KERN_SUCCESS) {
 595                 unlock_vp(&fs);
 596                 return (result);
 597         }
 598
 599         fs.map_generation = fs.map->timestamp;
 600
 601         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 602                 panic("%s: fault on nofault entry, addr: %#lx",
 603                     __func__, (u_long)vaddr);
 604         }
 605
 606         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 607             fs.entry->wiring_thread != curthread) {
 608                 vm_map_unlock_read(fs.map);
 609                 vm_map_lock(fs.map);
 610                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 611                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 612                         unlock_vp(&fs);
 613                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 614                         vm_map_unlock_and_wait(fs.map, 0);
 615                 } else
 616                         vm_map_unlock(fs.map);
 617                 goto RetryFault;
 618         }
 619
 620         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 621
 622         if (wired)
 623                 fault_type = prot | (fault_type & VM_PROT_COPY);
 624         else
 625                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 626                     ("!wired && VM_FAULT_WIRE"));
 627
 628         /*
 629          * Try to avoid lock contention on the top-level object through
 630          * special-case handling of some types of page faults, specifically,
 631          * those that are both (1) mapping an existing page from the top-
 632          * level object and (2) not having to mark that object as containing
 633          * dirty pages.  Under these conditions, a read lock on the top-level
 634          * object suffices, allowing multiple page faults of a similar type to
 635          * run in parallel on the same top-level object.
 636          */
 637         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 638             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
 639             /* avoid calling vm_object_set_writeable_dirty() */
 640             ((prot & VM_PROT_WRITE) == 0 ||
 641             (fs.first_object->type != OBJT_VNODE &&
 642             (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 643             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 644                 VM_OBJECT_RLOCK(fs.first_object);
 645                 if ((prot & VM_PROT_WRITE) == 0 ||
 646                     (fs.first_object->type != OBJT_VNODE &&
 647                     (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
 648                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 649                         rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 650                             fault_flags, wired, m_hold);
 651                         if (rv == KERN_SUCCESS)
 652                                 return (rv);
 653                 }
 654                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 655                         VM_OBJECT_RUNLOCK(fs.first_object);
 656                         VM_OBJECT_WLOCK(fs.first_object);
 657                 }
 658         } else {
 659                 VM_OBJECT_WLOCK(fs.first_object);
 660         }
 661
 662         /*
 663          * Make a reference to this object to prevent its disposal while we
 664          * are messing with it.  Once we have the reference, the map is free
 665          * to be diddled.  Since objects reference their shadows (and copies),
 666          * they will stay around as well.
 667          *
 668          * Bump the paging-in-progress count to prevent size changes (e.g.
 669          * truncation operations) during I/O.
 670          */
 671         vm_object_reference_locked(fs.first_object);
 672         vm_object_pip_add(fs.first_object, 1);
 673
 674         fs.lookup_still_valid = true;
 675
 676         fs.first_m = NULL;
 677
 678         /*
 679          * Search for the page at object/offset.
 680          */
 681         fs.object = fs.first_object;
 682         fs.pindex = fs.first_pindex;
 683         while (TRUE) {
 684                 /*
 685                  * If the object is marked for imminent termination,
 686                  * we retry here, since the collapse pass has raced
 687                  * with us.  Otherwise, if we see terminally dead
 688                  * object, return fail.
 689                  */
 690                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 691                         dead = fs.object->type == OBJT_DEAD;
 692                         unlock_and_deallocate(&fs);
 693                         if (dead)
 694                                 return (KERN_PROTECTION_FAILURE);
 695                         pause("vmf_de", 1);
 696                         goto RetryFault;
 697                 }
 698
 699                 /*
 700                  * See if page is resident
 701                  */
 702                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 703                 if (fs.m != NULL) {
 704                         /*
 705                          * Wait/Retry if the page is busy.  We have to do this
 706                          * if the page is either exclusive or shared busy
 707                          * because the vm_pager may be using read busy for
 708                          * pageouts (and even pageins if it is the vnode
 709                          * pager), and we could end up trying to pagein and
 710                          * pageout the same page simultaneously.
 711                          *
 712                          * We can theoretically allow the busy case on a read
 713                          * fault if the page is marked valid, but since such
 714                          * pages are typically already pmap'd, putting that
 715                          * special case in might be more effort then it is
 716                          * worth.  We cannot under any circumstances mess
 717                          * around with a shared busied page except, perhaps,
 718                          * to pmap it.
 719                          */
 720                         if (vm_page_busied(fs.m)) {
 721                                 /*
 722                                  * Reference the page before unlocking and
 723                                  * sleeping so that the page daemon is less
 724                                  * likely to reclaim it.
 725                                  */
 726                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 727                                 if (fs.object != fs.first_object) {
 728                                         if (!VM_OBJECT_TRYWLOCK(
 729                                             fs.first_object)) {
 730                                                 VM_OBJECT_WUNLOCK(fs.object);
 731                                                 VM_OBJECT_WLOCK(fs.first_object);
 732                                                 VM_OBJECT_WLOCK(fs.object);
 733                                         }
 734                                         vm_page_lock(fs.first_m);
 735                                         vm_page_free(fs.first_m);
 736                                         vm_page_unlock(fs.first_m);
 737                                         vm_object_pip_wakeup(fs.first_object);
 738                                         VM_OBJECT_WUNLOCK(fs.first_object);
 739                                         fs.first_m = NULL;
 740                                 }
 741                                 unlock_map(&fs);
 742                                 if (fs.m == vm_page_lookup(fs.object,
 743                                     fs.pindex)) {
 744                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 745                                 }
 746                                 vm_object_pip_wakeup(fs.object);
 747                                 VM_OBJECT_WUNLOCK(fs.object);
 748                                 VM_CNT_INC(v_intrans);
 749                                 vm_object_deallocate(fs.first_object);
 750                                 goto RetryFault;
 751                         }
 752
 753                         /*
 754                          * Mark page busy for other processes, and the
 755                          * pagedaemon.  If it still isn't completely valid
 756                          * (readable), jump to readrest, else break-out ( we
 757                          * found the page ).
 758                          */
 759                         vm_page_xbusy(fs.m);
 760                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 761                                 goto readrest;
 762                         break; /* break to PAGE HAS BEEN FOUND */
 763                 }
 764                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 765
 766                 /*
 767                  * Page is not resident.  If the pager might contain the page
 768                  * or this is the beginning of the search, allocate a new
 769                  * page.  (Default objects are zero-fill, so there is no real
 770                  * pager for them.)
 771                  */
 772                 if (fs.object->type != OBJT_DEFAULT ||
 773                     fs.object == fs.first_object) {
 774                         if (fs.pindex >= fs.object->size) {
 775                                 unlock_and_deallocate(&fs);
 776                                 return (KERN_PROTECTION_FAILURE);
 777                         }
 778
 779                         if (fs.object == fs.first_object &&
 780                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 781                             fs.first_object->shadow_count == 0) {
 782                                 rv = vm_fault_populate(&fs, prot, fault_type,
 783                                     fault_flags, wired, m_hold);
 784                                 switch (rv) {
 785                                 case KERN_SUCCESS:
 786                                 case KERN_FAILURE:
 787                                         unlock_and_deallocate(&fs);
 788                                         return (rv);
 789                                 case KERN_RESOURCE_SHORTAGE:
 790                                         unlock_and_deallocate(&fs);
 791                                         goto RetryFault;
 792                                 case KERN_NOT_RECEIVER:
 793                                         /*
 794                                          * Pager's populate() method
 795                                          * returned VM_PAGER_BAD.
 796                                          */
 797                                         break;
 798                                 default:
 799                                         panic("inconsistent return codes");
 800                                 }
 801                         }
 802
 803                         /*
 804                          * Allocate a new page for this object/offset pair.
 805                          *
 806                          * Unlocked read of the p_flag is harmless. At
 807                          * worst, the P_KILLED might be not observed
 808                          * there, and allocation can fail, causing
 809                          * restart and new reading of the p_flag.
 810                          */
 811                         dset = fs.object->domain.dr_policy;
 812                         if (dset == NULL)
 813                                 dset = curthread->td_domain.dr_policy;
 814                         if (!vm_page_count_severe_set(&dset->ds_mask) ||
 815                             P_KILLED(curproc)) {
 816 #if VM_NRESERVLEVEL > 0
 817                                 vm_object_color(fs.object, atop(vaddr) -
 818                                     fs.pindex);
 819 #endif
 820                                 alloc_req = P_KILLED(curproc) ?
 821                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 822                                 if (fs.object->type != OBJT_VNODE &&
 823                                     fs.object->backing_object == NULL)
 824                                         alloc_req |= VM_ALLOC_ZERO;
 825                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 826                                     alloc_req);
 827                         }
 828                         if (fs.m == NULL) {
 829                                 unlock_and_deallocate(&fs);
 830                                 vm_waitpfault(dset);
 831                                 goto RetryFault;
 832                         }
 833                 }
 834
 835 readrest:
 836                 /*
 837                  * At this point, we have either allocated a new page or found
 838                  * an existing page that is only partially valid.
 839                  *
 840                  * We hold a reference on the current object and the page is
 841                  * exclusive busied.
 842                  */
 843
 844                 /*
 845                  * If the pager for the current object might have the page,
 846                  * then determine the number of additional pages to read and
 847                  * potentially reprioritize previously read pages for earlier
 848                  * reclamation.  These operations should only be performed
 849                  * once per page fault.  Even if the current pager doesn't
 850                  * have the page, the number of additional pages to read will
 851                  * apply to subsequent objects in the shadow chain.
 852                  */
 853                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 854                     !P_KILLED(curproc)) {
 855                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 856                         era = fs.entry->read_ahead;
 857                         behavior = vm_map_entry_behavior(fs.entry);
 858                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 859                                 nera = 0;
 860                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 861                                 nera = VM_FAULT_READ_AHEAD_MAX;
 862                                 if (vaddr == fs.entry->next_read)
 863                                         vm_fault_dontneed(&fs, vaddr, nera);
 864                         } else if (vaddr == fs.entry->next_read) {
 865                                 /*
 866                                  * This is a sequential fault.  Arithmetically
 867                                  * increase the requested number of pages in
 868                                  * the read-ahead window.  The requested
 869                                  * number of pages is "# of sequential faults
 870                                  * x (read ahead min + 1) + read ahead min"
 871                                  */
 872                                 nera = VM_FAULT_READ_AHEAD_MIN;
 873                                 if (era > 0) {
 874                                         nera += era + 1;
 875                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 876                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 877                                 }
 878                                 if (era == VM_FAULT_READ_AHEAD_MAX)
 879                                         vm_fault_dontneed(&fs, vaddr, nera);
 880                         } else {
 881                                 /*
 882                                  * This is a non-sequential fault.
 883                                  */
 884                                 nera = 0;
 885                         }
 886                         if (era != nera) {
 887                                 /*
 888                                  * A read lock on the map suffices to update
 889                                  * the read ahead count safely.
 890                                  */
 891                                 fs.entry->read_ahead = nera;
 892                         }
 893
 894                         /*
 895                          * Prepare for unlocking the map.  Save the map
 896                          * entry's start and end addresses, which are used to
 897                          * optimize the size of the pager operation below.
 898                          * Even if the map entry's addresses change after
 899                          * unlocking the map, using the saved addresses is
 900                          * safe.
 901                          */
 902                         e_start = fs.entry->start;
 903                         e_end = fs.entry->end;
 904                 }
 905
 906                 /*
 907                  * Call the pager to retrieve the page if there is a chance
 908                  * that the pager has it, and potentially retrieve additional
 909                  * pages at the same time.
 910                  */
 911                 if (fs.object->type != OBJT_DEFAULT) {
 912                         /*
 913                          * Release the map lock before locking the vnode or
 914                          * sleeping in the pager.  (If the current object has
 915                          * a shadow, then an earlier iteration of this loop
 916                          * may have already unlocked the map.)
 917                          */
 918                         unlock_map(&fs);
 919
 920                         if (fs.object->type == OBJT_VNODE &&
 921                             (vp = fs.object->handle) != fs.vp) {
 922                                 /*
 923                                  * Perform an unlock in case the desired vnode
 924                                  * changed while the map was unlocked during a
 925                                  * retry.
 926                                  */
 927                                 unlock_vp(&fs);
 928
 929                                 locked = VOP_ISLOCKED(vp);
 930                                 if (locked != LK_EXCLUSIVE)
 931                                         locked = LK_SHARED;
 932
 933                                 /*
 934                                  * We must not sleep acquiring the vnode lock
 935                                  * while we have the page exclusive busied or
 936                                  * the object's paging-in-progress count
 937                                  * incremented.  Otherwise, we could deadlock.
 938                                  */
 939                                 error = vget(vp, locked | LK_CANRECURSE |
 940                                     LK_NOWAIT, curthread);
 941                                 if (error != 0) {
 942                                         vhold(vp);
 943                                         release_page(&fs);
 944                                         unlock_and_deallocate(&fs);
 945                                         error = vget(vp, locked | LK_RETRY |
 946                                             LK_CANRECURSE, curthread);
 947                                         vdrop(vp);
 948                                         fs.vp = vp;
 949                                         KASSERT(error == 0,
 950                                             ("vm_fault: vget failed"));
 951                                         goto RetryFault;
 952                                 }
 953                                 fs.vp = vp;
 954                         }
 955                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 956                             ("vm_fault: vnode-backed object mapped by system map"));
 957
 958                         /*
 959                          * Page in the requested page and hint the pager,
 960                          * that it may bring up surrounding pages.
 961                          */
 962                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
 963                             P_KILLED(curproc)) {
 964                                 behind = 0;
 965                                 ahead = 0;
 966                         } else {
 967                                 /* Is this a sequential fault? */
 968                                 if (nera > 0) {
 969                                         behind = 0;
 970                                         ahead = nera;
 971                                 } else {
 972                                         /*
 973                                          * Request a cluster of pages that is
 974                                          * aligned to a VM_FAULT_READ_DEFAULT
 975                                          * page offset boundary within the
 976                                          * object.  Alignment to a page offset
 977                                          * boundary is more likely to coincide
 978                                          * with the underlying file system
 979                                          * block than alignment to a virtual
 980                                          * address boundary.
 981                                          */
 982                                         cluster_offset = fs.pindex %
 983                                             VM_FAULT_READ_DEFAULT;
 984                                         behind = ulmin(cluster_offset,
 985                                             atop(vaddr - e_start));
 986                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
 987                                             cluster_offset;
 988                                 }
 989                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
 990                         }
 991                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
 992                             &behind, &ahead);
 993                         if (rv == VM_PAGER_OK) {
 994                                 faultcount = behind + 1 + ahead;
 995                                 hardfault = true;
 996                                 break; /* break to PAGE HAS BEEN FOUND */
 997                         }
 998                         if (rv == VM_PAGER_ERROR)
 999                                 printf("vm_fault: pager read error, pid %d (%s)\n",
1000                                     curproc->p_pid, curproc->p_comm);
1001
1002                         /*
1003                          * If an I/O error occurred or the requested page was
1004                          * outside the range of the pager, clean up and return
1005                          * an error.
1006                          */
1007                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1008                                 vm_page_lock(fs.m);
1009                                 if (!vm_page_wired(fs.m))
1010                                         vm_page_free(fs.m);
1011                                 else
1012                                         vm_page_xunbusy_maybelocked(fs.m);
1013                                 vm_page_unlock(fs.m);
1014                                 fs.m = NULL;
1015                                 unlock_and_deallocate(&fs);
1016                                 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
1017                                     KERN_PROTECTION_FAILURE);
1018                         }
1019
1020                         /*
1021                          * The requested page does not exist at this object/
1022                          * offset.  Remove the invalid page from the object,
1023                          * waking up anyone waiting for it, and continue on to
1024                          * the next object.  However, if this is the top-level
1025                          * object, we must leave the busy page in place to
1026                          * prevent another process from rushing past us, and
1027                          * inserting the page in that object at the same time
1028                          * that we are.
1029                          */
1030                         if (fs.object != fs.first_object) {
1031                                 vm_page_lock(fs.m);
1032                                 if (!vm_page_wired(fs.m))
1033                                         vm_page_free(fs.m);
1034                                 else
1035                                         vm_page_xunbusy_maybelocked(fs.m);
1036                                 vm_page_unlock(fs.m);
1037                                 fs.m = NULL;
1038                         }
1039                 }
1040
1041                 /*
1042                  * We get here if the object has default pager (or unwiring)
1043                  * or the pager doesn't have the page.
1044                  */
1045                 if (fs.object == fs.first_object)
1046                         fs.first_m = fs.m;
1047
1048                 /*
1049                  * Move on to the next object.  Lock the next object before
1050                  * unlocking the current one.
1051                  */
1052                 next_object = fs.object->backing_object;
1053                 if (next_object == NULL) {
1054                         /*
1055                          * If there's no object left, fill the page in the top
1056                          * object with zeros.
1057                          */
1058                         if (fs.object != fs.first_object) {
1059                                 vm_object_pip_wakeup(fs.object);
1060                                 VM_OBJECT_WUNLOCK(fs.object);
1061
1062                                 fs.object = fs.first_object;
1063                                 fs.pindex = fs.first_pindex;
1064                                 fs.m = fs.first_m;
1065                                 VM_OBJECT_WLOCK(fs.object);
1066                         }
1067                         fs.first_m = NULL;
1068
1069                         /*
1070                          * Zero the page if necessary and mark it valid.
1071                          */
1072                         if ((fs.m->flags & PG_ZERO) == 0) {
1073                                 pmap_zero_page(fs.m);
1074                         } else {
1075                                 VM_CNT_INC(v_ozfod);
1076                         }
1077                         VM_CNT_INC(v_zfod);
1078                         fs.m->valid = VM_PAGE_BITS_ALL;
1079                         /* Don't try to prefault neighboring pages. */
1080                         faultcount = 1;
1081                         break;  /* break to PAGE HAS BEEN FOUND */
1082                 } else {
1083                         KASSERT(fs.object != next_object,
1084                             ("object loop %p", next_object));
1085                         VM_OBJECT_WLOCK(next_object);
1086                         vm_object_pip_add(next_object, 1);
1087                         if (fs.object != fs.first_object)
1088                                 vm_object_pip_wakeup(fs.object);
1089                         fs.pindex +=
1090                             OFF_TO_IDX(fs.object->backing_object_offset);
1091                         VM_OBJECT_WUNLOCK(fs.object);
1092                         fs.object = next_object;
1093                 }
1094         }
1095
1096         vm_page_assert_xbusied(fs.m);
1097
1098         /*
1099          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1100          * is held.]
1101          */
1102
1103         /*
1104          * If the page is being written, but isn't already owned by the
1105          * top-level object, we have to copy it into a new page owned by the
1106          * top-level object.
1107          */
1108         if (fs.object != fs.first_object) {
1109                 /*
1110                  * We only really need to copy if we want to write it.
1111                  */
1112                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1113                         /*
1114                          * This allows pages to be virtually copied from a
1115                          * backing_object into the first_object, where the
1116                          * backing object has no other refs to it, and cannot
1117                          * gain any more refs.  Instead of a bcopy, we just
1118                          * move the page from the backing object to the
1119                          * first object.  Note that we must mark the page
1120                          * dirty in the first object so that it will go out
1121                          * to swap when needed.
1122                          */
1123                         is_first_object_locked = false;
1124                         if (
1125                                 /*
1126                                  * Only one shadow object
1127                                  */
1128                                 (fs.object->shadow_count == 1) &&
1129                                 /*
1130                                  * No COW refs, except us
1131                                  */
1132                                 (fs.object->ref_count == 1) &&
1133                                 /*
1134                                  * No one else can look this object up
1135                                  */
1136                                 (fs.object->handle == NULL) &&
1137                                 /*
1138                                  * No other ways to look the object up
1139                                  */
1140                                 ((fs.object->type == OBJT_DEFAULT) ||
1141                                  (fs.object->type == OBJT_SWAP)) &&
1142                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1143                                 /*
1144                                  * We don't chase down the shadow chain
1145                                  */
1146                             fs.object == fs.first_object->backing_object) {
1147                                 /*
1148                                  * Keep the page wired to ensure that it is not
1149                                  * freed by another thread, such as the page
1150                                  * daemon, while it is disassociated from an
1151                                  * object.
1152                                  */
1153                                 mtx = NULL;
1154                                 vm_page_change_lock(fs.m, &mtx);
1155                                 vm_page_wire(fs.m);
1156                                 (void)vm_page_remove(fs.m);
1157                                 vm_page_change_lock(fs.first_m, &mtx);
1158                                 vm_page_replace_checked(fs.m, fs.first_object,
1159                                     fs.first_pindex, fs.first_m);
1160                                 vm_page_free(fs.first_m);
1161                                 vm_page_change_lock(fs.m, &mtx);
1162                                 vm_page_unwire(fs.m, PQ_ACTIVE);
1163                                 mtx_unlock(mtx);
1164                                 vm_page_dirty(fs.m);
1165 #if VM_NRESERVLEVEL > 0
1166                                 /*
1167                                  * Rename the reservation.
1168                                  */
1169                                 vm_reserv_rename(fs.m, fs.first_object,
1170                                     fs.object, OFF_TO_IDX(
1171                                     fs.first_object->backing_object_offset));
1172 #endif
1173                                 /*
1174                                  * Removing the page from the backing object
1175                                  * unbusied it.
1176                                  */
1177                                 vm_page_xbusy(fs.m);
1178                                 fs.first_m = fs.m;
1179                                 fs.m = NULL;
1180                                 VM_CNT_INC(v_cow_optim);
1181                         } else {
1182                                 /*
1183                                  * Oh, well, lets copy it.
1184                                  */
1185                                 pmap_copy_page(fs.m, fs.first_m);
1186                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
1187                                 if (wired && (fault_flags &
1188                                     VM_FAULT_WIRE) == 0) {
1189                                         vm_page_lock(fs.first_m);
1190                                         vm_page_wire(fs.first_m);
1191                                         vm_page_unlock(fs.first_m);
1192
1193                                         vm_page_lock(fs.m);
1194                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1195                                         vm_page_unlock(fs.m);
1196                                 }
1197                                 /*
1198                                  * We no longer need the old page or object.
1199                                  */
1200                                 release_page(&fs);
1201                         }
1202                         /*
1203                          * fs.object != fs.first_object due to above
1204                          * conditional
1205                          */
1206                         vm_object_pip_wakeup(fs.object);
1207                         VM_OBJECT_WUNLOCK(fs.object);
1208
1209                         /*
1210                          * We only try to prefault read-only mappings to the
1211                          * neighboring pages when this copy-on-write fault is
1212                          * a hard fault.  In other cases, trying to prefault
1213                          * is typically wasted effort.
1214                          */
1215                         if (faultcount == 0)
1216                                 faultcount = 1;
1217
1218                         /*
1219                          * Only use the new page below...
1220                          */
1221                         fs.object = fs.first_object;
1222                         fs.pindex = fs.first_pindex;
1223                         fs.m = fs.first_m;
1224                         if (!is_first_object_locked)
1225                                 VM_OBJECT_WLOCK(fs.object);
1226                         VM_CNT_INC(v_cow_faults);
1227                         curthread->td_cow++;
1228                 } else {
1229                         prot &= ~VM_PROT_WRITE;
1230                 }
1231         }
1232
1233         /*
1234          * We must verify that the maps have not changed since our last
1235          * lookup.
1236          */
1237         if (!fs.lookup_still_valid) {
1238                 if (!vm_map_trylock_read(fs.map)) {
1239                         release_page(&fs);
1240                         unlock_and_deallocate(&fs);
1241                         goto RetryFault;
1242                 }
1243                 fs.lookup_still_valid = true;
1244                 if (fs.map->timestamp != fs.map_generation) {
1245                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1246                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1247
1248                         /*
1249                          * If we don't need the page any longer, put it on the inactive
1250                          * list (the easiest thing to do here).  If no one needs it,
1251                          * pageout will grab it eventually.
1252                          */
1253                         if (result != KERN_SUCCESS) {
1254                                 release_page(&fs);
1255                                 unlock_and_deallocate(&fs);
1256
1257                                 /*
1258                                  * If retry of map lookup would have blocked then
1259                                  * retry fault from start.
1260                                  */
1261                                 if (result == KERN_FAILURE)
1262                                         goto RetryFault;
1263                                 return (result);
1264                         }
1265                         if ((retry_object != fs.first_object) ||
1266                             (retry_pindex != fs.first_pindex)) {
1267                                 release_page(&fs);
1268                                 unlock_and_deallocate(&fs);
1269                                 goto RetryFault;
1270                         }
1271
1272                         /*
1273                          * Check whether the protection has changed or the object has
1274                          * been copied while we left the map unlocked. Changing from
1275                          * read to write permission is OK - we leave the page
1276                          * write-protected, and catch the write fault. Changing from
1277                          * write to read permission means that we can't mark the page
1278                          * write-enabled after all.
1279                          */
1280                         prot &= retry_prot;
1281                         fault_type &= retry_prot;
1282                         if (prot == 0) {
1283                                 release_page(&fs);
1284                                 unlock_and_deallocate(&fs);
1285                                 goto RetryFault;
1286                         }
1287
1288                         /* Reassert because wired may have changed. */
1289                         KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1290                             ("!wired && VM_FAULT_WIRE"));
1291                 }
1292         }
1293
1294         /*
1295          * If the page was filled by a pager, save the virtual address that
1296          * should be faulted on next under a sequential access pattern to the
1297          * map entry.  A read lock on the map suffices to update this address
1298          * safely.
1299          */
1300         if (hardfault)
1301                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1302
1303         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1304         vm_page_assert_xbusied(fs.m);
1305
1306         /*
1307          * Page must be completely valid or it is not fit to
1308          * map into user space.  vm_pager_get_pages() ensures this.
1309          */
1310         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1311             ("vm_fault: page %p partially invalid", fs.m));
1312         VM_OBJECT_WUNLOCK(fs.object);
1313
1314         /*
1315          * Put this page into the physical map.  We had to do the unlock above
1316          * because pmap_enter() may sleep.  We don't put the page
1317          * back on the active queue until later so that the pageout daemon
1318          * won't find it (yet).
1319          */
1320         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1321             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1322         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1323             wired == 0)
1324                 vm_fault_prefault(&fs, vaddr,
1325                     faultcount > 0 ? behind : PFBAK,
1326                     faultcount > 0 ? ahead : PFFOR, false);
1327         VM_OBJECT_WLOCK(fs.object);
1328         vm_page_lock(fs.m);
1329
1330         /*
1331          * If the page is not wired down, then put it where the pageout daemon
1332          * can find it.
1333          */
1334         if ((fault_flags & VM_FAULT_WIRE) != 0)
1335                 vm_page_wire(fs.m);
1336         else
1337                 vm_page_activate(fs.m);
1338         if (m_hold != NULL) {
1339                 *m_hold = fs.m;
1340                 vm_page_wire(fs.m);
1341         }
1342         vm_page_unlock(fs.m);
1343         vm_page_xunbusy(fs.m);
1344
1345         /*
1346          * Unlock everything, and return
1347          */
1348         unlock_and_deallocate(&fs);
1349         if (hardfault) {
1350                 VM_CNT_INC(v_io_faults);
1351                 curthread->td_ru.ru_majflt++;
1352 #ifdef RACCT
1353                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1354                         PROC_LOCK(curproc);
1355                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1356                                 racct_add_force(curproc, RACCT_WRITEBPS,
1357                                     PAGE_SIZE + behind * PAGE_SIZE);
1358                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1359                         } else {
1360                                 racct_add_force(curproc, RACCT_READBPS,
1361                                     PAGE_SIZE + ahead * PAGE_SIZE);
1362                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1363                         }
1364                         PROC_UNLOCK(curproc);
1365                 }
1366 #endif
1367         } else
1368                 curthread->td_ru.ru_minflt++;
1369
1370         return (KERN_SUCCESS);
1371 }
1372
1373 /*
1374  * Speed up the reclamation of pages that precede the faulting pindex within
1375  * the first object of the shadow chain.  Essentially, perform the equivalent
1376  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1377  * the faulting pindex by the cluster size when the pages read by vm_fault()
1378  * cross a cluster-size boundary.  The cluster size is the greater of the
1379  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1380  *
1381  * When "fs->first_object" is a shadow object, the pages in the backing object
1382  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1383  * function must only be concerned with pages in the first object.
1384  */
1385 static void
1386 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1387 {
1388         vm_map_entry_t entry;
1389         vm_object_t first_object, object;
1390         vm_offset_t end, start;
1391         vm_page_t m, m_next;
1392         vm_pindex_t pend, pstart;
1393         vm_size_t size;
1394
1395         object = fs->object;
1396         VM_OBJECT_ASSERT_WLOCKED(object);
1397         first_object = fs->first_object;
1398         if (first_object != object) {
1399                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1400                         VM_OBJECT_WUNLOCK(object);
1401                         VM_OBJECT_WLOCK(first_object);
1402                         VM_OBJECT_WLOCK(object);
1403                 }
1404         }
1405         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1406         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1407                 size = VM_FAULT_DONTNEED_MIN;
1408                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1409                         size = pagesizes[1];
1410                 end = rounddown2(vaddr, size);
1411                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1412                     (entry = fs->entry)->start < end) {
1413                         if (end - entry->start < size)
1414                                 start = entry->start;
1415                         else
1416                                 start = end - size;
1417                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1418                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1419                             entry->start);
1420                         m_next = vm_page_find_least(first_object, pstart);
1421                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1422                             entry->start);
1423                         while ((m = m_next) != NULL && m->pindex < pend) {
1424                                 m_next = TAILQ_NEXT(m, listq);
1425                                 if (m->valid != VM_PAGE_BITS_ALL ||
1426                                     vm_page_busied(m))
1427                                         continue;
1428
1429                                 /*
1430                                  * Don't clear PGA_REFERENCED, since it would
1431                                  * likely represent a reference by a different
1432                                  * process.
1433                                  *
1434                                  * Typically, at this point, prefetched pages
1435                                  * are still in the inactive queue.  Only
1436                                  * pages that triggered page faults are in the
1437                                  * active queue.
1438                                  */
1439                                 vm_page_lock(m);
1440                                 if (!vm_page_inactive(m))
1441                                         vm_page_deactivate(m);
1442                                 vm_page_unlock(m);
1443                         }
1444                 }
1445         }
1446         if (first_object != object)
1447                 VM_OBJECT_WUNLOCK(first_object);
1448 }
1449
1450 /*
1451  * vm_fault_prefault provides a quick way of clustering
1452  * pagefaults into a processes address space.  It is a "cousin"
1453  * of vm_map_pmap_enter, except it runs at page fault time instead
1454  * of mmap time.
1455  */
1456 static void
1457 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1458     int backward, int forward, bool obj_locked)
1459 {
1460         pmap_t pmap;
1461         vm_map_entry_t entry;
1462         vm_object_t backing_object, lobject;
1463         vm_offset_t addr, starta;
1464         vm_pindex_t pindex;
1465         vm_page_t m;
1466         int i;
1467
1468         pmap = fs->map->pmap;
1469         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1470                 return;
1471
1472         entry = fs->entry;
1473
1474         if (addra < backward * PAGE_SIZE) {
1475                 starta = entry->start;
1476         } else {
1477                 starta = addra - backward * PAGE_SIZE;
1478                 if (starta < entry->start)
1479                         starta = entry->start;
1480         }
1481
1482         /*
1483          * Generate the sequence of virtual addresses that are candidates for
1484          * prefaulting in an outward spiral from the faulting virtual address,
1485          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1486          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1487          * If the candidate address doesn't have a backing physical page, then
1488          * the loop immediately terminates.
1489          */
1490         for (i = 0; i < 2 * imax(backward, forward); i++) {
1491                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1492                     PAGE_SIZE);
1493                 if (addr > addra + forward * PAGE_SIZE)
1494                         addr = 0;
1495
1496                 if (addr < starta || addr >= entry->end)
1497                         continue;
1498
1499                 if (!pmap_is_prefaultable(pmap, addr))
1500                         continue;
1501
1502                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1503                 lobject = entry->object.vm_object;
1504                 if (!obj_locked)
1505                         VM_OBJECT_RLOCK(lobject);
1506                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1507                     lobject->type == OBJT_DEFAULT &&
1508                     (backing_object = lobject->backing_object) != NULL) {
1509                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1510                             0, ("vm_fault_prefault: unaligned object offset"));
1511                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1512                         VM_OBJECT_RLOCK(backing_object);
1513                         if (!obj_locked || lobject != entry->object.vm_object)
1514                                 VM_OBJECT_RUNLOCK(lobject);
1515                         lobject = backing_object;
1516                 }
1517                 if (m == NULL) {
1518                         if (!obj_locked || lobject != entry->object.vm_object)
1519                                 VM_OBJECT_RUNLOCK(lobject);
1520                         break;
1521                 }
1522                 if (m->valid == VM_PAGE_BITS_ALL &&
1523                     (m->flags & PG_FICTITIOUS) == 0)
1524                         pmap_enter_quick(pmap, addr, m, entry->protection);
1525                 if (!obj_locked || lobject != entry->object.vm_object)
1526                         VM_OBJECT_RUNLOCK(lobject);
1527         }
1528 }
1529
1530 /*
1531  * Hold each of the physical pages that are mapped by the specified range of
1532  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1533  * and allow the specified types of access, "prot".  If all of the implied
1534  * pages are successfully held, then the number of held pages is returned
1535  * together with pointers to those pages in the array "ma".  However, if any
1536  * of the pages cannot be held, -1 is returned.
1537  */
1538 int
1539 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1540     vm_prot_t prot, vm_page_t *ma, int max_count)
1541 {
1542         vm_offset_t end, va;
1543         vm_page_t *mp;
1544         int count;
1545         boolean_t pmap_failed;
1546
1547         if (len == 0)
1548                 return (0);
1549         end = round_page(addr + len);
1550         addr = trunc_page(addr);
1551
1552         /*
1553          * Check for illegal addresses.
1554          */
1555         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1556                 return (-1);
1557
1558         if (atop(end - addr) > max_count)
1559                 panic("vm_fault_quick_hold_pages: count > max_count");
1560         count = atop(end - addr);
1561
1562         /*
1563          * Most likely, the physical pages are resident in the pmap, so it is
1564          * faster to try pmap_extract_and_hold() first.
1565          */
1566         pmap_failed = FALSE;
1567         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1568                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1569                 if (*mp == NULL)
1570                         pmap_failed = TRUE;
1571                 else if ((prot & VM_PROT_WRITE) != 0 &&
1572                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1573                         /*
1574                          * Explicitly dirty the physical page.  Otherwise, the
1575                          * caller's changes may go unnoticed because they are
1576                          * performed through an unmanaged mapping or by a DMA
1577                          * operation.
1578                          *
1579                          * The object lock is not held here.
1580                          * See vm_page_clear_dirty_mask().
1581                          */
1582                         vm_page_dirty(*mp);
1583                 }
1584         }
1585         if (pmap_failed) {
1586                 /*
1587                  * One or more pages could not be held by the pmap.  Either no
1588                  * page was mapped at the specified virtual address or that
1589                  * mapping had insufficient permissions.  Attempt to fault in
1590                  * and hold these pages.
1591                  *
1592                  * If vm_fault_disable_pagefaults() was called,
1593                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1594                  * acquire MD VM locks, which means we must not call
1595                  * vm_fault_hold().  Some (out of tree) callers mark
1596                  * too wide a code area with vm_fault_disable_pagefaults()
1597                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1598                  * the proper behaviour explicitly.
1599                  */
1600                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1601                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1602                         goto error;
1603                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1604                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1605                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1606                                 goto error;
1607         }
1608         return (count);
1609 error:
1610         for (mp = ma; mp < ma + count; mp++)
1611                 if (*mp != NULL) {
1612                         vm_page_lock(*mp);
1613                         if (vm_page_unwire(*mp, PQ_INACTIVE) &&
1614                             (*mp)->object == NULL)
1615                                 vm_page_free(*mp);
1616                         vm_page_unlock(*mp);
1617                 }
1618         return (-1);
1619 }
1620
1621 /*
1622  *      Routine:
1623  *              vm_fault_copy_entry
1624  *      Function:
1625  *              Create new shadow object backing dst_entry with private copy of
1626  *              all underlying pages. When src_entry is equal to dst_entry,
1627  *              function implements COW for wired-down map entry. Otherwise,
1628  *              it forks wired entry into dst_map.
1629  *
1630  *      In/out conditions:
1631  *              The source and destination maps must be locked for write.
1632  *              The source map entry must be wired down (or be a sharing map
1633  *              entry corresponding to a main map entry that is wired down).
1634  */
1635 void
1636 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1637     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1638     vm_ooffset_t *fork_charge)
1639 {
1640         vm_object_t backing_object, dst_object, object, src_object;
1641         vm_pindex_t dst_pindex, pindex, src_pindex;
1642         vm_prot_t access, prot;
1643         vm_offset_t vaddr;
1644         vm_page_t dst_m;
1645         vm_page_t src_m;
1646         boolean_t upgrade;
1647
1648 #ifdef  lint
1649         src_map++;
1650 #endif  /* lint */
1651
1652         upgrade = src_entry == dst_entry;
1653         access = prot = dst_entry->protection;
1654
1655         src_object = src_entry->object.vm_object;
1656         src_pindex = OFF_TO_IDX(src_entry->offset);
1657
1658         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1659                 dst_object = src_object;
1660                 vm_object_reference(dst_object);
1661         } else {
1662                 /*
1663                  * Create the top-level object for the destination entry. (Doesn't
1664                  * actually shadow anything - we copy the pages directly.)
1665                  */
1666                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1667                     atop(dst_entry->end - dst_entry->start));
1668 #if VM_NRESERVLEVEL > 0
1669                 dst_object->flags |= OBJ_COLORED;
1670                 dst_object->pg_color = atop(dst_entry->start);
1671 #endif
1672                 dst_object->domain = src_object->domain;
1673                 dst_object->charge = dst_entry->end - dst_entry->start;
1674         }
1675
1676         VM_OBJECT_WLOCK(dst_object);
1677         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1678             ("vm_fault_copy_entry: vm_object not NULL"));
1679         if (src_object != dst_object) {
1680                 dst_entry->object.vm_object = dst_object;
1681                 dst_entry->offset = 0;
1682                 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1683         }
1684         if (fork_charge != NULL) {
1685                 KASSERT(dst_entry->cred == NULL,
1686                     ("vm_fault_copy_entry: leaked swp charge"));
1687                 dst_object->cred = curthread->td_ucred;
1688                 crhold(dst_object->cred);
1689                 *fork_charge += dst_object->charge;
1690         } else if ((dst_object->type == OBJT_DEFAULT ||
1691             dst_object->type == OBJT_SWAP) &&
1692             dst_object->cred == NULL) {
1693                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1694                     dst_entry));
1695                 dst_object->cred = dst_entry->cred;
1696                 dst_entry->cred = NULL;
1697         }
1698
1699         /*
1700          * If not an upgrade, then enter the mappings in the pmap as
1701          * read and/or execute accesses.  Otherwise, enter them as
1702          * write accesses.
1703          *
1704          * A writeable large page mapping is only created if all of
1705          * the constituent small page mappings are modified. Marking
1706          * PTEs as modified on inception allows promotion to happen
1707          * without taking potentially large number of soft faults.
1708          */
1709         if (!upgrade)
1710                 access &= ~VM_PROT_WRITE;
1711
1712         /*
1713          * Loop through all of the virtual pages within the entry's
1714          * range, copying each page from the source object to the
1715          * destination object.  Since the source is wired, those pages
1716          * must exist.  In contrast, the destination is pageable.
1717          * Since the destination object doesn't share any backing storage
1718          * with the source object, all of its pages must be dirtied,
1719          * regardless of whether they can be written.
1720          */
1721         for (vaddr = dst_entry->start, dst_pindex = 0;
1722             vaddr < dst_entry->end;
1723             vaddr += PAGE_SIZE, dst_pindex++) {
1724 again:
1725                 /*
1726                  * Find the page in the source object, and copy it in.
1727                  * Because the source is wired down, the page will be
1728                  * in memory.
1729                  */
1730                 if (src_object != dst_object)
1731                         VM_OBJECT_RLOCK(src_object);
1732                 object = src_object;
1733                 pindex = src_pindex + dst_pindex;
1734                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1735                     (backing_object = object->backing_object) != NULL) {
1736                         /*
1737                          * Unless the source mapping is read-only or
1738                          * it is presently being upgraded from
1739                          * read-only, the first object in the shadow
1740                          * chain should provide all of the pages.  In
1741                          * other words, this loop body should never be
1742                          * executed when the source mapping is already
1743                          * read/write.
1744                          */
1745                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1746                             upgrade,
1747                             ("vm_fault_copy_entry: main object missing page"));
1748
1749                         VM_OBJECT_RLOCK(backing_object);
1750                         pindex += OFF_TO_IDX(object->backing_object_offset);
1751                         if (object != dst_object)
1752                                 VM_OBJECT_RUNLOCK(object);
1753                         object = backing_object;
1754                 }
1755                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1756
1757                 if (object != dst_object) {
1758                         /*
1759                          * Allocate a page in the destination object.
1760                          */
1761                         dst_m = vm_page_alloc(dst_object, (src_object ==
1762                             dst_object ? src_pindex : 0) + dst_pindex,
1763                             VM_ALLOC_NORMAL);
1764                         if (dst_m == NULL) {
1765                                 VM_OBJECT_WUNLOCK(dst_object);
1766                                 VM_OBJECT_RUNLOCK(object);
1767                                 vm_wait(dst_object);
1768                                 VM_OBJECT_WLOCK(dst_object);
1769                                 goto again;
1770                         }
1771                         pmap_copy_page(src_m, dst_m);
1772                         VM_OBJECT_RUNLOCK(object);
1773                         dst_m->dirty = dst_m->valid = src_m->valid;
1774                 } else {
1775                         dst_m = src_m;
1776                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1777                                 goto again;
1778                         if (dst_m->pindex >= dst_object->size)
1779                                 /*
1780                                  * We are upgrading.  Index can occur
1781                                  * out of bounds if the object type is
1782                                  * vnode and the file was truncated.
1783                                  */
1784                                 break;
1785                         vm_page_xbusy(dst_m);
1786                 }
1787                 VM_OBJECT_WUNLOCK(dst_object);
1788
1789                 /*
1790                  * Enter it in the pmap. If a wired, copy-on-write
1791                  * mapping is being replaced by a write-enabled
1792                  * mapping, then wire that new mapping.
1793                  *
1794                  * The page can be invalid if the user called
1795                  * msync(MS_INVALIDATE) or truncated the backing vnode
1796                  * or shared memory object.  In this case, do not
1797                  * insert it into pmap, but still do the copy so that
1798                  * all copies of the wired map entry have similar
1799                  * backing pages.
1800                  */
1801                 if (dst_m->valid == VM_PAGE_BITS_ALL) {
1802                         pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1803                             access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1804                 }
1805
1806                 /*
1807                  * Mark it no longer busy, and put it on the active list.
1808                  */
1809                 VM_OBJECT_WLOCK(dst_object);
1810
1811                 if (upgrade) {
1812                         if (src_m != dst_m) {
1813                                 vm_page_lock(src_m);
1814                                 vm_page_unwire(src_m, PQ_INACTIVE);
1815                                 vm_page_unlock(src_m);
1816                                 vm_page_lock(dst_m);
1817                                 vm_page_wire(dst_m);
1818                                 vm_page_unlock(dst_m);
1819                         } else {
1820                                 KASSERT(vm_page_wired(dst_m),
1821                                     ("dst_m %p is not wired", dst_m));
1822                         }
1823                 } else {
1824                         vm_page_lock(dst_m);
1825                         vm_page_activate(dst_m);
1826                         vm_page_unlock(dst_m);
1827                 }
1828                 vm_page_xunbusy(dst_m);
1829         }
1830         VM_OBJECT_WUNLOCK(dst_object);
1831         if (upgrade) {
1832                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1833                 vm_object_deallocate(src_object);
1834         }
1835 }
1836
1837 /*
1838  * Block entry into the machine-independent layer's page fault handler by
1839  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1840  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1841  * spurious page faults.
1842  */
1843 int
1844 vm_fault_disable_pagefaults(void)
1845 {
1846
1847         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1848 }
1849
1850 void
1851 vm_fault_enable_pagefaults(int save)
1852 {
1853
1854         curthread_pflags_restore(save);
1855 }