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/refcount.h>
  91 #include <sys/resourcevar.h>
  92 #include <sys/rwlock.h>
  93 #include <sys/signalvar.h>
  94 #include <sys/sysctl.h>
  95 #include <sys/sysent.h>
  96 #include <sys/vmmeter.h>
  97 #include <sys/vnode.h>
  98 #ifdef KTRACE
  99 #include <sys/ktrace.h>
 100 #endif
 101
 102 #include <vm/vm.h>
 103 #include <vm/vm_param.h>
 104 #include <vm/pmap.h>
 105 #include <vm/vm_map.h>
 106 #include <vm/vm_object.h>
 107 #include <vm/vm_page.h>
 108 #include <vm/vm_pageout.h>
 109 #include <vm/vm_kern.h>
 110 #include <vm/vm_pager.h>
 111 #include <vm/vm_extern.h>
 112 #include <vm/vm_reserv.h>
 113
 114 #define PFBAK 4
 115 #define PFFOR 4
 116
 117 #define VM_FAULT_READ_DEFAULT   (1 + VM_FAULT_READ_AHEAD_INIT)
 118 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 119
 120 #define VM_FAULT_DONTNEED_MIN   1048576
 121
 122 struct faultstate {
 123         vm_page_t m;
 124         vm_object_t object;
 125         vm_pindex_t pindex;
 126         vm_page_t first_m;
 127         vm_object_t     first_object;
 128         vm_pindex_t first_pindex;
 129         vm_map_t map;
 130         vm_map_entry_t entry;
 131         int map_generation;
 132         bool lookup_still_valid;
 133         struct vnode *vp;
 134 };
 135
 136 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 137             int ahead);
 138 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 139             int backward, int forward, bool obj_locked);
 140
 141 static int vm_pfault_oom_attempts = 3;
 142 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
 143     &vm_pfault_oom_attempts, 0,
 144     "Number of page allocation attempts in page fault handler before it "
 145     "triggers OOM handling");
 146
 147 static int vm_pfault_oom_wait = 10;
 148 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
 149     &vm_pfault_oom_wait, 0,
 150     "Number of seconds to wait for free pages before retrying "
 151     "the page fault handler");
 152
 153 static inline void
 154 fault_page_release(vm_page_t *mp)
 155 {
 156         vm_page_t m;
 157
 158         m = *mp;
 159         if (m != NULL) {
 160                 /*
 161                  * We are likely to loop around again and attempt to busy
 162                  * this page.  Deactivating it leaves it available for
 163                  * pageout while optimizing fault restarts.
 164                  */
 165                 vm_page_deactivate(m);
 166                 vm_page_xunbusy(m);
 167                 *mp = NULL;
 168         }
 169 }
 170
 171 static inline void
 172 fault_page_free(vm_page_t *mp)
 173 {
 174         vm_page_t m;
 175
 176         m = *mp;
 177         if (m != NULL) {
 178                 VM_OBJECT_ASSERT_WLOCKED(m->object);
 179                 if (!vm_page_wired(m))
 180                         vm_page_free(m);
 181                 else
 182                         vm_page_xunbusy(m);
 183                 *mp = NULL;
 184         }
 185 }
 186
 187 static inline void
 188 unlock_map(struct faultstate *fs)
 189 {
 190
 191         if (fs->lookup_still_valid) {
 192                 vm_map_lookup_done(fs->map, fs->entry);
 193                 fs->lookup_still_valid = false;
 194         }
 195 }
 196
 197 static void
 198 unlock_vp(struct faultstate *fs)
 199 {
 200
 201         if (fs->vp != NULL) {
 202                 vput(fs->vp);
 203                 fs->vp = NULL;
 204         }
 205 }
 206
 207 static void
 208 fault_deallocate(struct faultstate *fs)
 209 {
 210
 211         fault_page_release(&fs->m);
 212         vm_object_pip_wakeup(fs->object);
 213         if (fs->object != fs->first_object) {
 214                 VM_OBJECT_WLOCK(fs->first_object);
 215                 fault_page_free(&fs->first_m);
 216                 VM_OBJECT_WUNLOCK(fs->first_object);
 217                 vm_object_pip_wakeup(fs->first_object);
 218         }
 219         vm_object_deallocate(fs->first_object);
 220         unlock_map(fs);
 221         unlock_vp(fs);
 222 }
 223
 224 static void
 225 unlock_and_deallocate(struct faultstate *fs)
 226 {
 227
 228         VM_OBJECT_WUNLOCK(fs->object);
 229         fault_deallocate(fs);
 230 }
 231
 232 static void
 233 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 234     vm_prot_t fault_type, int fault_flags)
 235 {
 236         bool need_dirty;
 237
 238         if (((prot & VM_PROT_WRITE) == 0 &&
 239             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 240             (m->oflags & VPO_UNMANAGED) != 0)
 241                 return;
 242
 243         VM_PAGE_OBJECT_BUSY_ASSERT(m);
 244
 245         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 246             (fault_flags & VM_FAULT_WIRE) == 0) ||
 247             (fault_flags & VM_FAULT_DIRTY) != 0;
 248
 249         vm_object_set_writeable_dirty(m->object);
 250
 251         /*
 252          * If the fault is a write, we know that this page is being
 253          * written NOW so dirty it explicitly to save on
 254          * pmap_is_modified() calls later.
 255          *
 256          * Also, since the page is now dirty, we can possibly tell
 257          * the pager to release any swap backing the page.
 258          */
 259         if (need_dirty && vm_page_set_dirty(m) == 0) {
 260                 /*
 261                  * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
 262                  * if the page is already dirty to prevent data written with
 263                  * the expectation of being synced from not being synced.
 264                  * Likewise if this entry does not request NOSYNC then make
 265                  * sure the page isn't marked NOSYNC.  Applications sharing
 266                  * data should use the same flags to avoid ping ponging.
 267                  */
 268                 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0)
 269                         vm_page_aflag_set(m, PGA_NOSYNC);
 270                 else
 271                         vm_page_aflag_clear(m, PGA_NOSYNC);
 272         }
 273
 274 }
 275
 276 /*
 277  * Unlocks fs.first_object and fs.map on success.
 278  */
 279 static int
 280 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 281     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 282 {
 283         vm_page_t m, m_map;
 284 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 285     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 286     VM_NRESERVLEVEL > 0
 287         vm_page_t m_super;
 288         int flags;
 289 #endif
 290         int psind, rv;
 291
 292         MPASS(fs->vp == NULL);
 293         vm_object_busy(fs->first_object);
 294         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 295         /* A busy page can be mapped for read|execute access. */
 296         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 297             vm_page_busied(m)) || !vm_page_all_valid(m)) {
 298                 rv = KERN_FAILURE;
 299                 goto out;
 300         }
 301         m_map = m;
 302         psind = 0;
 303 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 304     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 305     VM_NRESERVLEVEL > 0
 306         if ((m->flags & PG_FICTITIOUS) == 0 &&
 307             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 308             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 309             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 310             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 311             (pagesizes[m_super->psind] - 1)) && !wired &&
 312             pmap_ps_enabled(fs->map->pmap)) {
 313                 flags = PS_ALL_VALID;
 314                 if ((prot & VM_PROT_WRITE) != 0) {
 315                         /*
 316                          * Create a superpage mapping allowing write access
 317                          * only if none of the constituent pages are busy and
 318                          * all of them are already dirty (except possibly for
 319                          * the page that was faulted on).
 320                          */
 321                         flags |= PS_NONE_BUSY;
 322                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 323                                 flags |= PS_ALL_DIRTY;
 324                 }
 325                 if (vm_page_ps_test(m_super, flags, m)) {
 326                         m_map = m_super;
 327                         psind = m_super->psind;
 328                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 329                         /* Preset the modified bit for dirty superpages. */
 330                         if ((flags & PS_ALL_DIRTY) != 0)
 331                                 fault_type |= VM_PROT_WRITE;
 332                 }
 333         }
 334 #endif
 335         rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
 336             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
 337         if (rv != KERN_SUCCESS)
 338                 goto out;
 339         if (m_hold != NULL) {
 340                 *m_hold = m;
 341                 vm_page_wire(m);
 342         }
 343         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags);
 344         if (psind == 0 && !wired)
 345                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 346         VM_OBJECT_RUNLOCK(fs->first_object);
 347         vm_map_lookup_done(fs->map, fs->entry);
 348         curthread->td_ru.ru_minflt++;
 349
 350 out:
 351         vm_object_unbusy(fs->first_object);
 352         return (rv);
 353 }
 354
 355 static void
 356 vm_fault_restore_map_lock(struct faultstate *fs)
 357 {
 358
 359         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 360         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 361
 362         if (!vm_map_trylock_read(fs->map)) {
 363                 VM_OBJECT_WUNLOCK(fs->first_object);
 364                 vm_map_lock_read(fs->map);
 365                 VM_OBJECT_WLOCK(fs->first_object);
 366         }
 367         fs->lookup_still_valid = true;
 368 }
 369
 370 static void
 371 vm_fault_populate_check_page(vm_page_t m)
 372 {
 373
 374         /*
 375          * Check each page to ensure that the pager is obeying the
 376          * interface: the page must be installed in the object, fully
 377          * valid, and exclusively busied.
 378          */
 379         MPASS(m != NULL);
 380         MPASS(vm_page_all_valid(m));
 381         MPASS(vm_page_xbusied(m));
 382 }
 383
 384 static void
 385 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 386     vm_pindex_t last)
 387 {
 388         vm_page_t m;
 389         vm_pindex_t pidx;
 390
 391         VM_OBJECT_ASSERT_WLOCKED(object);
 392         MPASS(first <= last);
 393         for (pidx = first, m = vm_page_lookup(object, pidx);
 394             pidx <= last; pidx++, m = vm_page_next(m)) {
 395                 vm_fault_populate_check_page(m);
 396                 vm_page_deactivate(m);
 397                 vm_page_xunbusy(m);
 398         }
 399 }
 400
 401 static int
 402 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 403     int fault_flags, boolean_t wired, vm_page_t *m_hold)
 404 {
 405         vm_offset_t vaddr;
 406         vm_page_t m;
 407         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 408         int i, npages, psind, rv;
 409
 410         MPASS(fs->object == fs->first_object);
 411         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 412         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 413         MPASS(fs->first_object->backing_object == NULL);
 414         MPASS(fs->lookup_still_valid);
 415
 416         pager_first = OFF_TO_IDX(fs->entry->offset);
 417         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 418         unlock_map(fs);
 419         unlock_vp(fs);
 420
 421         /*
 422          * Call the pager (driver) populate() method.
 423          *
 424          * There is no guarantee that the method will be called again
 425          * if the current fault is for read, and a future fault is
 426          * for write.  Report the entry's maximum allowed protection
 427          * to the driver.
 428          */
 429         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 430             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 431
 432         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 433         if (rv == VM_PAGER_BAD) {
 434                 /*
 435                  * VM_PAGER_BAD is the backdoor for a pager to request
 436                  * normal fault handling.
 437                  */
 438                 vm_fault_restore_map_lock(fs);
 439                 if (fs->map->timestamp != fs->map_generation)
 440                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 441                 return (KERN_NOT_RECEIVER);
 442         }
 443         if (rv != VM_PAGER_OK)
 444                 return (KERN_FAILURE); /* AKA SIGSEGV */
 445
 446         /* Ensure that the driver is obeying the interface. */
 447         MPASS(pager_first <= pager_last);
 448         MPASS(fs->first_pindex <= pager_last);
 449         MPASS(fs->first_pindex >= pager_first);
 450         MPASS(pager_last < fs->first_object->size);
 451
 452         vm_fault_restore_map_lock(fs);
 453         if (fs->map->timestamp != fs->map_generation) {
 454                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 455                     pager_last);
 456                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 457         }
 458
 459         /*
 460          * The map is unchanged after our last unlock.  Process the fault.
 461          *
 462          * The range [pager_first, pager_last] that is given to the
 463          * pager is only a hint.  The pager may populate any range
 464          * within the object that includes the requested page index.
 465          * In case the pager expanded the range, clip it to fit into
 466          * the map entry.
 467          */
 468         map_first = OFF_TO_IDX(fs->entry->offset);
 469         if (map_first > pager_first) {
 470                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 471                     map_first - 1);
 472                 pager_first = map_first;
 473         }
 474         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 475         if (map_last < pager_last) {
 476                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 477                     pager_last);
 478                 pager_last = map_last;
 479         }
 480         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 481             pidx <= pager_last;
 482             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 483                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 484 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 485     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
 486                 psind = m->psind;
 487                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 488                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 489                     !pmap_ps_enabled(fs->map->pmap) || wired))
 490                         psind = 0;
 491 #else
 492                 psind = 0;
 493 #endif
 494                 npages = atop(pagesizes[psind]);
 495                 for (i = 0; i < npages; i++) {
 496                         vm_fault_populate_check_page(&m[i]);
 497                         vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
 498                             fault_flags);
 499                 }
 500                 VM_OBJECT_WUNLOCK(fs->first_object);
 501                 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
 502                     (wired ? PMAP_ENTER_WIRED : 0), psind);
 503 #if defined(__amd64__)
 504                 if (psind > 0 && rv == KERN_FAILURE) {
 505                         for (i = 0; i < npages; i++) {
 506                                 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
 507                                     &m[i], prot, fault_type |
 508                                     (wired ? PMAP_ENTER_WIRED : 0), 0);
 509                                 MPASS(rv == KERN_SUCCESS);
 510                         }
 511                 }
 512 #else
 513                 MPASS(rv == KERN_SUCCESS);
 514 #endif
 515                 VM_OBJECT_WLOCK(fs->first_object);
 516                 for (i = 0; i < npages; i++) {
 517                         if ((fault_flags & VM_FAULT_WIRE) != 0)
 518                                 vm_page_wire(&m[i]);
 519                         else
 520                                 vm_page_activate(&m[i]);
 521                         if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
 522                                 *m_hold = &m[i];
 523                                 vm_page_wire(&m[i]);
 524                         }
 525                         vm_page_xunbusy(&m[i]);
 526                 }
 527         }
 528         curthread->td_ru.ru_majflt++;
 529         return (KERN_SUCCESS);
 530 }
 531
 532 static int prot_fault_translation;
 533 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
 534     &prot_fault_translation, 0,
 535     "Control signal to deliver on protection fault");
 536
 537 /* compat definition to keep common code for signal translation */
 538 #define UCODE_PAGEFLT   12
 539 #ifdef T_PAGEFLT
 540 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
 541 #endif
 542
 543 /*
 544  *      vm_fault_trap:
 545  *
 546  *      Handle a page fault occurring at the given address,
 547  *      requiring the given permissions, in the map specified.
 548  *      If successful, the page is inserted into the
 549  *      associated physical map.
 550  *
 551  *      NOTE: the given address should be truncated to the
 552  *      proper page address.
 553  *
 554  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 555  *      a standard error specifying why the fault is fatal is returned.
 556  *
 557  *      The map in question must be referenced, and remains so.
 558  *      Caller may hold no locks.
 559  */
 560 int
 561 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 562     int fault_flags, int *signo, int *ucode)
 563 {
 564         int result;
 565
 566         MPASS(signo == NULL || ucode != NULL);
 567 #ifdef KTRACE
 568         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
 569                 ktrfault(vaddr, fault_type);
 570 #endif
 571         result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
 572             NULL);
 573         KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
 574             result == KERN_INVALID_ADDRESS ||
 575             result == KERN_RESOURCE_SHORTAGE ||
 576             result == KERN_PROTECTION_FAILURE ||
 577             result == KERN_OUT_OF_BOUNDS,
 578             ("Unexpected Mach error %d from vm_fault()", result));
 579 #ifdef KTRACE
 580         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
 581                 ktrfaultend(result);
 582 #endif
 583         if (result != KERN_SUCCESS && signo != NULL) {
 584                 switch (result) {
 585                 case KERN_FAILURE:
 586                 case KERN_INVALID_ADDRESS:
 587                         *signo = SIGSEGV;
 588                         *ucode = SEGV_MAPERR;
 589                         break;
 590                 case KERN_RESOURCE_SHORTAGE:
 591                         *signo = SIGBUS;
 592                         *ucode = BUS_OOMERR;
 593                         break;
 594                 case KERN_OUT_OF_BOUNDS:
 595                         *signo = SIGBUS;
 596                         *ucode = BUS_OBJERR;
 597                         break;
 598                 case KERN_PROTECTION_FAILURE:
 599                         if (prot_fault_translation == 0) {
 600                                 /*
 601                                  * Autodetect.  This check also covers
 602                                  * the images without the ABI-tag ELF
 603                                  * note.
 604                                  */
 605                                 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
 606                                     curproc->p_osrel >= P_OSREL_SIGSEGV) {
 607                                         *signo = SIGSEGV;
 608                                         *ucode = SEGV_ACCERR;
 609                                 } else {
 610                                         *signo = SIGBUS;
 611                                         *ucode = UCODE_PAGEFLT;
 612                                 }
 613                         } else if (prot_fault_translation == 1) {
 614                                 /* Always compat mode. */
 615                                 *signo = SIGBUS;
 616                                 *ucode = UCODE_PAGEFLT;
 617                         } else {
 618                                 /* Always SIGSEGV mode. */
 619                                 *signo = SIGSEGV;
 620                                 *ucode = SEGV_ACCERR;
 621                         }
 622                         break;
 623                 default:
 624                         KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
 625                             result));
 626                         break;
 627                 }
 628         }
 629         return (result);
 630 }
 631
 632 static int
 633 vm_fault_lock_vnode(struct faultstate *fs)
 634 {
 635         struct vnode *vp;
 636         int error, locked;
 637
 638         if (fs->object->type != OBJT_VNODE)
 639                 return (KERN_SUCCESS);
 640         vp = fs->object->handle;
 641         if (vp == fs->vp) {
 642                 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
 643                 return (KERN_SUCCESS);
 644         }
 645
 646         /*
 647          * Perform an unlock in case the desired vnode changed while
 648          * the map was unlocked during a retry.
 649          */
 650         unlock_vp(fs);
 651
 652         locked = VOP_ISLOCKED(vp);
 653         if (locked != LK_EXCLUSIVE)
 654                 locked = LK_SHARED;
 655
 656         /*
 657          * We must not sleep acquiring the vnode lock while we have
 658          * the page exclusive busied or the object's
 659          * paging-in-progress count incremented.  Otherwise, we could
 660          * deadlock.
 661          */
 662         error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
 663         if (error == 0) {
 664                 fs->vp = vp;
 665                 return (KERN_SUCCESS);
 666         }
 667
 668         vhold(vp);
 669         unlock_and_deallocate(fs);
 670         error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
 671         vdrop(vp);
 672         fs->vp = vp;
 673         KASSERT(error == 0, ("vm_fault: vget failed %d", error));
 674         return (KERN_RESOURCE_SHORTAGE);
 675 }
 676
 677 /*
 678  * Wait/Retry if the page is busy.  We have to do this if the page is
 679  * either exclusive or shared busy because the vm_pager may be using
 680  * read busy for pageouts (and even pageins if it is the vnode pager),
 681  * and we could end up trying to pagein and pageout the same page
 682  * simultaneously.
 683  *
 684  * We can theoretically allow the busy case on a read fault if the page
 685  * is marked valid, but since such pages are typically already pmap'd,
 686  * putting that special case in might be more effort then it is worth.
 687  * We cannot under any circumstances mess around with a shared busied
 688  * page except, perhaps, to pmap it.
 689  */
 690 static void
 691 vm_fault_busy_sleep(struct faultstate *fs)
 692 {
 693         /*
 694          * Reference the page before unlocking and
 695          * sleeping so that the page daemon is less
 696          * likely to reclaim it.
 697          */
 698         vm_page_aflag_set(fs->m, PGA_REFERENCED);
 699         if (fs->object != fs->first_object) {
 700                 fault_page_release(&fs->first_m);
 701                 vm_object_pip_wakeup(fs->first_object);
 702         }
 703         vm_object_pip_wakeup(fs->object);
 704         unlock_map(fs);
 705         if (fs->m == vm_page_lookup(fs->object, fs->pindex))
 706                 vm_page_busy_sleep(fs->m, "vmpfw", false);
 707         else
 708                 VM_OBJECT_WUNLOCK(fs->object);
 709         VM_CNT_INC(v_intrans);
 710         vm_object_deallocate(fs->first_object);
 711 }
 712
 713 int
 714 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 715     int fault_flags, vm_page_t *m_hold)
 716 {
 717         struct faultstate fs;
 718         struct domainset *dset;
 719         vm_object_t next_object, retry_object;
 720         vm_offset_t e_end, e_start;
 721         vm_pindex_t retry_pindex;
 722         vm_prot_t prot, retry_prot;
 723         int ahead, alloc_req, behind, cluster_offset, era, faultcount;
 724         int nera, oom, result, rv;
 725         u_char behavior;
 726         boolean_t wired;        /* Passed by reference. */
 727         bool dead, hardfault, is_first_object_locked;
 728
 729         VM_CNT_INC(v_vm_faults);
 730
 731         if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
 732                 return (KERN_PROTECTION_FAILURE);
 733
 734         fs.vp = NULL;
 735         faultcount = 0;
 736         nera = -1;
 737         hardfault = false;
 738
 739 RetryFault:
 740         oom = 0;
 741 RetryFault_oom:
 742
 743         /*
 744          * Find the backing store object and offset into it to begin the
 745          * search.
 746          */
 747         fs.map = map;
 748         result = vm_map_lookup(&fs.map, vaddr, fault_type |
 749             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 750             &fs.first_pindex, &prot, &wired);
 751         if (result != KERN_SUCCESS) {
 752                 unlock_vp(&fs);
 753                 return (result);
 754         }
 755
 756         fs.map_generation = fs.map->timestamp;
 757
 758         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 759                 panic("%s: fault on nofault entry, addr: %#lx",
 760                     __func__, (u_long)vaddr);
 761         }
 762
 763         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 764             fs.entry->wiring_thread != curthread) {
 765                 vm_map_unlock_read(fs.map);
 766                 vm_map_lock(fs.map);
 767                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 768                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 769                         unlock_vp(&fs);
 770                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 771                         vm_map_unlock_and_wait(fs.map, 0);
 772                 } else
 773                         vm_map_unlock(fs.map);
 774                 goto RetryFault;
 775         }
 776
 777         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 778
 779         if (wired)
 780                 fault_type = prot | (fault_type & VM_PROT_COPY);
 781         else
 782                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 783                     ("!wired && VM_FAULT_WIRE"));
 784
 785         /*
 786          * Try to avoid lock contention on the top-level object through
 787          * special-case handling of some types of page faults, specifically,
 788          * those that are mapping an existing page from the top-level object.
 789          * Under this condition, a read lock on the object suffices, allowing
 790          * multiple page faults of a similar type to run in parallel.
 791          */
 792         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 793             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
 794                 VM_OBJECT_RLOCK(fs.first_object);
 795                 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 796                     fault_flags, wired, m_hold);
 797                 if (rv == KERN_SUCCESS)
 798                         return (rv);
 799                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 800                         VM_OBJECT_RUNLOCK(fs.first_object);
 801                         VM_OBJECT_WLOCK(fs.first_object);
 802                 }
 803         } else {
 804                 VM_OBJECT_WLOCK(fs.first_object);
 805         }
 806
 807         /*
 808          * Make a reference to this object to prevent its disposal while we
 809          * are messing with it.  Once we have the reference, the map is free
 810          * to be diddled.  Since objects reference their shadows (and copies),
 811          * they will stay around as well.
 812          *
 813          * Bump the paging-in-progress count to prevent size changes (e.g.
 814          * truncation operations) during I/O.
 815          */
 816         vm_object_reference_locked(fs.first_object);
 817         vm_object_pip_add(fs.first_object, 1);
 818
 819         fs.lookup_still_valid = true;
 820
 821         fs.m = fs.first_m = NULL;
 822
 823         /*
 824          * Search for the page at object/offset.
 825          */
 826         fs.object = fs.first_object;
 827         fs.pindex = fs.first_pindex;
 828         while (TRUE) {
 829                 KASSERT(fs.m == NULL,
 830                     ("page still set %p at loop start", fs.m));
 831                 /*
 832                  * If the object is marked for imminent termination,
 833                  * we retry here, since the collapse pass has raced
 834                  * with us.  Otherwise, if we see terminally dead
 835                  * object, return fail.
 836                  */
 837                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 838                         dead = fs.object->type == OBJT_DEAD;
 839                         unlock_and_deallocate(&fs);
 840                         if (dead)
 841                                 return (KERN_PROTECTION_FAILURE);
 842                         pause("vmf_de", 1);
 843                         goto RetryFault;
 844                 }
 845
 846                 /*
 847                  * See if page is resident
 848                  */
 849                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 850                 if (fs.m != NULL) {
 851                         if (vm_page_tryxbusy(fs.m) == 0) {
 852                                 vm_fault_busy_sleep(&fs);
 853                                 goto RetryFault;
 854                         }
 855
 856                         /*
 857                          * The page is marked busy for other processes and the
 858                          * pagedaemon.  If it still isn't completely valid
 859                          * (readable), jump to readrest, else break-out ( we
 860                          * found the page ).
 861                          */
 862                         if (!vm_page_all_valid(fs.m))
 863                                 goto readrest;
 864                         break; /* break to PAGE HAS BEEN FOUND */
 865                 }
 866                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 867
 868                 /*
 869                  * Page is not resident.  If the pager might contain the page
 870                  * or this is the beginning of the search, allocate a new
 871                  * page.  (Default objects are zero-fill, so there is no real
 872                  * pager for them.)
 873                  */
 874                 if (fs.object->type != OBJT_DEFAULT ||
 875                     fs.object == fs.first_object) {
 876                         if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
 877                                 rv = vm_fault_lock_vnode(&fs);
 878                                 MPASS(rv == KERN_SUCCESS ||
 879                                     rv == KERN_RESOURCE_SHORTAGE);
 880                                 if (rv == KERN_RESOURCE_SHORTAGE)
 881                                         goto RetryFault;
 882                         }
 883                         if (fs.pindex >= fs.object->size) {
 884                                 unlock_and_deallocate(&fs);
 885                                 return (KERN_OUT_OF_BOUNDS);
 886                         }
 887
 888                         if (fs.object == fs.first_object &&
 889                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 890                             fs.first_object->shadow_count == 0) {
 891                                 rv = vm_fault_populate(&fs, prot, fault_type,
 892                                     fault_flags, wired, m_hold);
 893                                 switch (rv) {
 894                                 case KERN_SUCCESS:
 895                                 case KERN_FAILURE:
 896                                         unlock_and_deallocate(&fs);
 897                                         return (rv);
 898                                 case KERN_RESOURCE_SHORTAGE:
 899                                         unlock_and_deallocate(&fs);
 900                                         goto RetryFault;
 901                                 case KERN_NOT_RECEIVER:
 902                                         /*
 903                                          * Pager's populate() method
 904                                          * returned VM_PAGER_BAD.
 905                                          */
 906                                         break;
 907                                 default:
 908                                         panic("inconsistent return codes");
 909                                 }
 910                         }
 911
 912                         /*
 913                          * Allocate a new page for this object/offset pair.
 914                          *
 915                          * Unlocked read of the p_flag is harmless. At
 916                          * worst, the P_KILLED might be not observed
 917                          * there, and allocation can fail, causing
 918                          * restart and new reading of the p_flag.
 919                          */
 920                         dset = fs.object->domain.dr_policy;
 921                         if (dset == NULL)
 922                                 dset = curthread->td_domain.dr_policy;
 923                         if (!vm_page_count_severe_set(&dset->ds_mask) ||
 924                             P_KILLED(curproc)) {
 925 #if VM_NRESERVLEVEL > 0
 926                                 vm_object_color(fs.object, atop(vaddr) -
 927                                     fs.pindex);
 928 #endif
 929                                 alloc_req = P_KILLED(curproc) ?
 930                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 931                                 if (fs.object->type != OBJT_VNODE &&
 932                                     fs.object->backing_object == NULL)
 933                                         alloc_req |= VM_ALLOC_ZERO;
 934                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 935                                     alloc_req);
 936                         }
 937                         if (fs.m == NULL) {
 938                                 unlock_and_deallocate(&fs);
 939                                 if (vm_pfault_oom_attempts < 0 ||
 940                                     oom < vm_pfault_oom_attempts) {
 941                                         oom++;
 942                                         vm_waitpfault(dset,
 943                                             vm_pfault_oom_wait * hz);
 944                                         goto RetryFault_oom;
 945                                 }
 946                                 if (bootverbose)
 947                                         printf(
 948         "proc %d (%s) failed to alloc page on fault, starting OOM\n",
 949                                             curproc->p_pid, curproc->p_comm);
 950                                 vm_pageout_oom(VM_OOM_MEM_PF);
 951                                 goto RetryFault;
 952                         }
 953                 }
 954
 955 readrest:
 956                 /*
 957                  * At this point, we have either allocated a new page or found
 958                  * an existing page that is only partially valid.
 959                  *
 960                  * We hold a reference on the current object and the page is
 961                  * exclusive busied.
 962                  */
 963
 964                 /*
 965                  * If the pager for the current object might have the page,
 966                  * then determine the number of additional pages to read and
 967                  * potentially reprioritize previously read pages for earlier
 968                  * reclamation.  These operations should only be performed
 969                  * once per page fault.  Even if the current pager doesn't
 970                  * have the page, the number of additional pages to read will
 971                  * apply to subsequent objects in the shadow chain.
 972                  */
 973                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 974                     !P_KILLED(curproc)) {
 975                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 976                         era = fs.entry->read_ahead;
 977                         behavior = vm_map_entry_behavior(fs.entry);
 978                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 979                                 nera = 0;
 980                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 981                                 nera = VM_FAULT_READ_AHEAD_MAX;
 982                                 if (vaddr == fs.entry->next_read)
 983                                         vm_fault_dontneed(&fs, vaddr, nera);
 984                         } else if (vaddr == fs.entry->next_read) {
 985                                 /*
 986                                  * This is a sequential fault.  Arithmetically
 987                                  * increase the requested number of pages in
 988                                  * the read-ahead window.  The requested
 989                                  * number of pages is "# of sequential faults
 990                                  * x (read ahead min + 1) + read ahead min"
 991                                  */
 992                                 nera = VM_FAULT_READ_AHEAD_MIN;
 993                                 if (era > 0) {
 994                                         nera += era + 1;
 995                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 996                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 997                                 }
 998                                 if (era == VM_FAULT_READ_AHEAD_MAX)
 999                                         vm_fault_dontneed(&fs, vaddr, nera);
1000                         } else {
1001                                 /*
1002                                  * This is a non-sequential fault.
1003                                  */
1004                                 nera = 0;
1005                         }
1006                         if (era != nera) {
1007                                 /*
1008                                  * A read lock on the map suffices to update
1009                                  * the read ahead count safely.
1010                                  */
1011                                 fs.entry->read_ahead = nera;
1012                         }
1013
1014                         /*
1015                          * Prepare for unlocking the map.  Save the map
1016                          * entry's start and end addresses, which are used to
1017                          * optimize the size of the pager operation below.
1018                          * Even if the map entry's addresses change after
1019                          * unlocking the map, using the saved addresses is
1020                          * safe.
1021                          */
1022                         e_start = fs.entry->start;
1023                         e_end = fs.entry->end;
1024                 }
1025
1026                 /*
1027                  * Call the pager to retrieve the page if there is a chance
1028                  * that the pager has it, and potentially retrieve additional
1029                  * pages at the same time.
1030                  */
1031                 if (fs.object->type != OBJT_DEFAULT) {
1032                         /*
1033                          * Release the map lock before locking the vnode or
1034                          * sleeping in the pager.  (If the current object has
1035                          * a shadow, then an earlier iteration of this loop
1036                          * may have already unlocked the map.)
1037                          */
1038                         unlock_map(&fs);
1039
1040                         rv = vm_fault_lock_vnode(&fs);
1041                         MPASS(rv == KERN_SUCCESS ||
1042                             rv == KERN_RESOURCE_SHORTAGE);
1043                         if (rv == KERN_RESOURCE_SHORTAGE)
1044                                 goto RetryFault;
1045                         KASSERT(fs.vp == NULL || !fs.map->system_map,
1046                             ("vm_fault: vnode-backed object mapped by system map"));
1047
1048                         /*
1049                          * Page in the requested page and hint the pager,
1050                          * that it may bring up surrounding pages.
1051                          */
1052                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1053                             P_KILLED(curproc)) {
1054                                 behind = 0;
1055                                 ahead = 0;
1056                         } else {
1057                                 /* Is this a sequential fault? */
1058                                 if (nera > 0) {
1059                                         behind = 0;
1060                                         ahead = nera;
1061                                 } else {
1062                                         /*
1063                                          * Request a cluster of pages that is
1064                                          * aligned to a VM_FAULT_READ_DEFAULT
1065                                          * page offset boundary within the
1066                                          * object.  Alignment to a page offset
1067                                          * boundary is more likely to coincide
1068                                          * with the underlying file system
1069                                          * block than alignment to a virtual
1070                                          * address boundary.
1071                                          */
1072                                         cluster_offset = fs.pindex %
1073                                             VM_FAULT_READ_DEFAULT;
1074                                         behind = ulmin(cluster_offset,
1075                                             atop(vaddr - e_start));
1076                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
1077                                             cluster_offset;
1078                                 }
1079                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1080                         }
1081                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1082                             &behind, &ahead);
1083                         if (rv == VM_PAGER_OK) {
1084                                 faultcount = behind + 1 + ahead;
1085                                 hardfault = true;
1086                                 break; /* break to PAGE HAS BEEN FOUND */
1087                         }
1088                         if (rv == VM_PAGER_ERROR)
1089                                 printf("vm_fault: pager read error, pid %d (%s)\n",
1090                                     curproc->p_pid, curproc->p_comm);
1091
1092                         /*
1093                          * If an I/O error occurred or the requested page was
1094                          * outside the range of the pager, clean up and return
1095                          * an error.
1096                          */
1097                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1098                                 fault_page_free(&fs.m);
1099                                 unlock_and_deallocate(&fs);
1100                                 return (KERN_OUT_OF_BOUNDS);
1101                         }
1102
1103                 }
1104
1105                 /*
1106                  * The requested page does not exist at this object/
1107                  * offset.  Remove the invalid page from the object,
1108                  * waking up anyone waiting for it, and continue on to
1109                  * the next object.  However, if this is the top-level
1110                  * object, we must leave the busy page in place to
1111                  * prevent another process from rushing past us, and
1112                  * inserting the page in that object at the same time
1113                  * that we are.
1114                  */
1115                 if (fs.object == fs.first_object) {
1116                         fs.first_m = fs.m;
1117                         fs.m = NULL;
1118                 } else
1119                         fault_page_free(&fs.m);
1120
1121                 /*
1122                  * Move on to the next object.  Lock the next object before
1123                  * unlocking the current one.
1124                  */
1125                 next_object = fs.object->backing_object;
1126                 if (next_object == NULL) {
1127                         /*
1128                          * If there's no object left, fill the page in the top
1129                          * object with zeros.
1130                          */
1131                         if (fs.object != fs.first_object) {
1132                                 vm_object_pip_wakeup(fs.object);
1133                                 VM_OBJECT_WUNLOCK(fs.object);
1134
1135                                 fs.object = fs.first_object;
1136                                 fs.pindex = fs.first_pindex;
1137                                 VM_OBJECT_WLOCK(fs.object);
1138                         }
1139                         MPASS(fs.first_m != NULL);
1140                         MPASS(fs.m == NULL);
1141                         fs.m = fs.first_m;
1142                         fs.first_m = NULL;
1143
1144                         /*
1145                          * Zero the page if necessary and mark it valid.
1146                          */
1147                         if ((fs.m->flags & PG_ZERO) == 0) {
1148                                 pmap_zero_page(fs.m);
1149                         } else {
1150                                 VM_CNT_INC(v_ozfod);
1151                         }
1152                         VM_CNT_INC(v_zfod);
1153                         vm_page_valid(fs.m);
1154                         /* Don't try to prefault neighboring pages. */
1155                         faultcount = 1;
1156                         break;  /* break to PAGE HAS BEEN FOUND */
1157                 } else {
1158                         MPASS(fs.first_m != NULL);
1159                         KASSERT(fs.object != next_object,
1160                             ("object loop %p", next_object));
1161                         VM_OBJECT_WLOCK(next_object);
1162                         vm_object_pip_add(next_object, 1);
1163                         if (fs.object != fs.first_object)
1164                                 vm_object_pip_wakeup(fs.object);
1165                         fs.pindex +=
1166                             OFF_TO_IDX(fs.object->backing_object_offset);
1167                         VM_OBJECT_WUNLOCK(fs.object);
1168                         fs.object = next_object;
1169                 }
1170         }
1171
1172         vm_page_assert_xbusied(fs.m);
1173
1174         /*
1175          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1176          * is held.]
1177          */
1178
1179         /*
1180          * If the page is being written, but isn't already owned by the
1181          * top-level object, we have to copy it into a new page owned by the
1182          * top-level object.
1183          */
1184         if (fs.object != fs.first_object) {
1185                 /*
1186                  * We only really need to copy if we want to write it.
1187                  */
1188                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1189                         /*
1190                          * This allows pages to be virtually copied from a
1191                          * backing_object into the first_object, where the
1192                          * backing object has no other refs to it, and cannot
1193                          * gain any more refs.  Instead of a bcopy, we just
1194                          * move the page from the backing object to the
1195                          * first object.  Note that we must mark the page
1196                          * dirty in the first object so that it will go out
1197                          * to swap when needed.
1198                          */
1199                         is_first_object_locked = false;
1200                         if (
1201                                 /*
1202                                  * Only one shadow object
1203                                  */
1204                                 (fs.object->shadow_count == 1) &&
1205                                 /*
1206                                  * No COW refs, except us
1207                                  */
1208                                 (fs.object->ref_count == 1) &&
1209                                 /*
1210                                  * No one else can look this object up
1211                                  */
1212                                 (fs.object->handle == NULL) &&
1213                                 /*
1214                                  * No other ways to look the object up
1215                                  */
1216                                 ((fs.object->flags & OBJ_ANON) != 0) &&
1217                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1218                                 /*
1219                                  * We don't chase down the shadow chain
1220                                  */
1221                             fs.object == fs.first_object->backing_object) {
1222
1223                                 /*
1224                                  * Remove but keep xbusy for replace.  fs.m is
1225                                  * moved into fs.first_object and left busy
1226                                  * while fs.first_m is conditionally freed.
1227                                  */
1228                                 vm_page_remove_xbusy(fs.m);
1229                                 vm_page_replace(fs.m, fs.first_object,
1230                                     fs.first_pindex, fs.first_m);
1231                                 vm_page_dirty(fs.m);
1232 #if VM_NRESERVLEVEL > 0
1233                                 /*
1234                                  * Rename the reservation.
1235                                  */
1236                                 vm_reserv_rename(fs.m, fs.first_object,
1237                                     fs.object, OFF_TO_IDX(
1238                                     fs.first_object->backing_object_offset));
1239 #endif
1240                                 VM_OBJECT_WUNLOCK(fs.object);
1241                                 fs.first_m = fs.m;
1242                                 fs.m = NULL;
1243                                 VM_CNT_INC(v_cow_optim);
1244                         } else {
1245                                 VM_OBJECT_WUNLOCK(fs.object);
1246                                 /*
1247                                  * Oh, well, lets copy it.
1248                                  */
1249                                 pmap_copy_page(fs.m, fs.first_m);
1250                                 vm_page_valid(fs.first_m);
1251                                 if (wired && (fault_flags &
1252                                     VM_FAULT_WIRE) == 0) {
1253                                         vm_page_wire(fs.first_m);
1254                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1255                                 }
1256                                 /*
1257                                  * We no longer need the old page or object.
1258                                  */
1259                                 fault_page_release(&fs.m);
1260                         }
1261                         /*
1262                          * fs.object != fs.first_object due to above
1263                          * conditional
1264                          */
1265                         vm_object_pip_wakeup(fs.object);
1266
1267                         /*
1268                          * We only try to prefault read-only mappings to the
1269                          * neighboring pages when this copy-on-write fault is
1270                          * a hard fault.  In other cases, trying to prefault
1271                          * is typically wasted effort.
1272                          */
1273                         if (faultcount == 0)
1274                                 faultcount = 1;
1275
1276                         /*
1277                          * Only use the new page below...
1278                          */
1279                         fs.object = fs.first_object;
1280                         fs.pindex = fs.first_pindex;
1281                         fs.m = fs.first_m;
1282                         if (!is_first_object_locked)
1283                                 VM_OBJECT_WLOCK(fs.object);
1284                         VM_CNT_INC(v_cow_faults);
1285                         curthread->td_cow++;
1286                 } else {
1287                         prot &= ~VM_PROT_WRITE;
1288                 }
1289         }
1290
1291         /*
1292          * We must verify that the maps have not changed since our last
1293          * lookup.
1294          */
1295         if (!fs.lookup_still_valid) {
1296                 if (!vm_map_trylock_read(fs.map)) {
1297                         unlock_and_deallocate(&fs);
1298                         goto RetryFault;
1299                 }
1300                 fs.lookup_still_valid = true;
1301                 if (fs.map->timestamp != fs.map_generation) {
1302                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1303                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1304
1305                         /*
1306                          * If we don't need the page any longer, put it on the inactive
1307                          * list (the easiest thing to do here).  If no one needs it,
1308                          * pageout will grab it eventually.
1309                          */
1310                         if (result != KERN_SUCCESS) {
1311                                 unlock_and_deallocate(&fs);
1312
1313                                 /*
1314                                  * If retry of map lookup would have blocked then
1315                                  * retry fault from start.
1316                                  */
1317                                 if (result == KERN_FAILURE)
1318                                         goto RetryFault;
1319                                 return (result);
1320                         }
1321                         if ((retry_object != fs.first_object) ||
1322                             (retry_pindex != fs.first_pindex)) {
1323                                 unlock_and_deallocate(&fs);
1324                                 goto RetryFault;
1325                         }
1326
1327                         /*
1328                          * Check whether the protection has changed or the object has
1329                          * been copied while we left the map unlocked. Changing from
1330                          * read to write permission is OK - we leave the page
1331                          * write-protected, and catch the write fault. Changing from
1332                          * write to read permission means that we can't mark the page
1333                          * write-enabled after all.
1334                          */
1335                         prot &= retry_prot;
1336                         fault_type &= retry_prot;
1337                         if (prot == 0) {
1338                                 unlock_and_deallocate(&fs);
1339                                 goto RetryFault;
1340                         }
1341
1342                         /* Reassert because wired may have changed. */
1343                         KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1344                             ("!wired && VM_FAULT_WIRE"));
1345                 }
1346         }
1347
1348         /*
1349          * If the page was filled by a pager, save the virtual address that
1350          * should be faulted on next under a sequential access pattern to the
1351          * map entry.  A read lock on the map suffices to update this address
1352          * safely.
1353          */
1354         if (hardfault)
1355                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1356
1357         vm_page_assert_xbusied(fs.m);
1358         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags);
1359
1360         /*
1361          * Page must be completely valid or it is not fit to
1362          * map into user space.  vm_pager_get_pages() ensures this.
1363          */
1364         KASSERT(vm_page_all_valid(fs.m),
1365             ("vm_fault: page %p partially invalid", fs.m));
1366         VM_OBJECT_WUNLOCK(fs.object);
1367
1368         /*
1369          * Put this page into the physical map.  We had to do the unlock above
1370          * because pmap_enter() may sleep.  We don't put the page
1371          * back on the active queue until later so that the pageout daemon
1372          * won't find it (yet).
1373          */
1374         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1375             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1376         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1377             wired == 0)
1378                 vm_fault_prefault(&fs, vaddr,
1379                     faultcount > 0 ? behind : PFBAK,
1380                     faultcount > 0 ? ahead : PFFOR, false);
1381
1382         /*
1383          * If the page is not wired down, then put it where the pageout daemon
1384          * can find it.
1385          */
1386         if ((fault_flags & VM_FAULT_WIRE) != 0)
1387                 vm_page_wire(fs.m);
1388         else
1389                 vm_page_activate(fs.m);
1390         if (m_hold != NULL) {
1391                 *m_hold = fs.m;
1392                 vm_page_wire(fs.m);
1393         }
1394         vm_page_xunbusy(fs.m);
1395         fs.m = NULL;
1396
1397         /*
1398          * Unlock everything, and return
1399          */
1400         fault_deallocate(&fs);
1401         if (hardfault) {
1402                 VM_CNT_INC(v_io_faults);
1403                 curthread->td_ru.ru_majflt++;
1404 #ifdef RACCT
1405                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1406                         PROC_LOCK(curproc);
1407                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1408                                 racct_add_force(curproc, RACCT_WRITEBPS,
1409                                     PAGE_SIZE + behind * PAGE_SIZE);
1410                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1411                         } else {
1412                                 racct_add_force(curproc, RACCT_READBPS,
1413                                     PAGE_SIZE + ahead * PAGE_SIZE);
1414                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1415                         }
1416                         PROC_UNLOCK(curproc);
1417                 }
1418 #endif
1419         } else
1420                 curthread->td_ru.ru_minflt++;
1421
1422         return (KERN_SUCCESS);
1423 }
1424
1425 /*
1426  * Speed up the reclamation of pages that precede the faulting pindex within
1427  * the first object of the shadow chain.  Essentially, perform the equivalent
1428  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1429  * the faulting pindex by the cluster size when the pages read by vm_fault()
1430  * cross a cluster-size boundary.  The cluster size is the greater of the
1431  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1432  *
1433  * When "fs->first_object" is a shadow object, the pages in the backing object
1434  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1435  * function must only be concerned with pages in the first object.
1436  */
1437 static void
1438 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1439 {
1440         vm_map_entry_t entry;
1441         vm_object_t first_object, object;
1442         vm_offset_t end, start;
1443         vm_page_t m, m_next;
1444         vm_pindex_t pend, pstart;
1445         vm_size_t size;
1446
1447         object = fs->object;
1448         VM_OBJECT_ASSERT_WLOCKED(object);
1449         first_object = fs->first_object;
1450         if (first_object != object) {
1451                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1452                         VM_OBJECT_WUNLOCK(object);
1453                         VM_OBJECT_WLOCK(first_object);
1454                         VM_OBJECT_WLOCK(object);
1455                 }
1456         }
1457         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1458         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1459                 size = VM_FAULT_DONTNEED_MIN;
1460                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1461                         size = pagesizes[1];
1462                 end = rounddown2(vaddr, size);
1463                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1464                     (entry = fs->entry)->start < end) {
1465                         if (end - entry->start < size)
1466                                 start = entry->start;
1467                         else
1468                                 start = end - size;
1469                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1470                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1471                             entry->start);
1472                         m_next = vm_page_find_least(first_object, pstart);
1473                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1474                             entry->start);
1475                         while ((m = m_next) != NULL && m->pindex < pend) {
1476                                 m_next = TAILQ_NEXT(m, listq);
1477                                 if (!vm_page_all_valid(m) ||
1478                                     vm_page_busied(m))
1479                                         continue;
1480
1481                                 /*
1482                                  * Don't clear PGA_REFERENCED, since it would
1483                                  * likely represent a reference by a different
1484                                  * process.
1485                                  *
1486                                  * Typically, at this point, prefetched pages
1487                                  * are still in the inactive queue.  Only
1488                                  * pages that triggered page faults are in the
1489                                  * active queue.  The test for whether the page
1490                                  * is in the inactive queue is racy; in the
1491                                  * worst case we will requeue the page
1492                                  * unnecessarily.
1493                                  */
1494                                 if (!vm_page_inactive(m))
1495                                         vm_page_deactivate(m);
1496                         }
1497                 }
1498         }
1499         if (first_object != object)
1500                 VM_OBJECT_WUNLOCK(first_object);
1501 }
1502
1503 /*
1504  * vm_fault_prefault provides a quick way of clustering
1505  * pagefaults into a processes address space.  It is a "cousin"
1506  * of vm_map_pmap_enter, except it runs at page fault time instead
1507  * of mmap time.
1508  */
1509 static void
1510 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1511     int backward, int forward, bool obj_locked)
1512 {
1513         pmap_t pmap;
1514         vm_map_entry_t entry;
1515         vm_object_t backing_object, lobject;
1516         vm_offset_t addr, starta;
1517         vm_pindex_t pindex;
1518         vm_page_t m;
1519         int i;
1520
1521         pmap = fs->map->pmap;
1522         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1523                 return;
1524
1525         entry = fs->entry;
1526
1527         if (addra < backward * PAGE_SIZE) {
1528                 starta = entry->start;
1529         } else {
1530                 starta = addra - backward * PAGE_SIZE;
1531                 if (starta < entry->start)
1532                         starta = entry->start;
1533         }
1534
1535         /*
1536          * Generate the sequence of virtual addresses that are candidates for
1537          * prefaulting in an outward spiral from the faulting virtual address,
1538          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1539          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1540          * If the candidate address doesn't have a backing physical page, then
1541          * the loop immediately terminates.
1542          */
1543         for (i = 0; i < 2 * imax(backward, forward); i++) {
1544                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1545                     PAGE_SIZE);
1546                 if (addr > addra + forward * PAGE_SIZE)
1547                         addr = 0;
1548
1549                 if (addr < starta || addr >= entry->end)
1550                         continue;
1551
1552                 if (!pmap_is_prefaultable(pmap, addr))
1553                         continue;
1554
1555                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1556                 lobject = entry->object.vm_object;
1557                 if (!obj_locked)
1558                         VM_OBJECT_RLOCK(lobject);
1559                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1560                     lobject->type == OBJT_DEFAULT &&
1561                     (backing_object = lobject->backing_object) != NULL) {
1562                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1563                             0, ("vm_fault_prefault: unaligned object offset"));
1564                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1565                         VM_OBJECT_RLOCK(backing_object);
1566                         if (!obj_locked || lobject != entry->object.vm_object)
1567                                 VM_OBJECT_RUNLOCK(lobject);
1568                         lobject = backing_object;
1569                 }
1570                 if (m == NULL) {
1571                         if (!obj_locked || lobject != entry->object.vm_object)
1572                                 VM_OBJECT_RUNLOCK(lobject);
1573                         break;
1574                 }
1575                 if (vm_page_all_valid(m) &&
1576                     (m->flags & PG_FICTITIOUS) == 0)
1577                         pmap_enter_quick(pmap, addr, m, entry->protection);
1578                 if (!obj_locked || lobject != entry->object.vm_object)
1579                         VM_OBJECT_RUNLOCK(lobject);
1580         }
1581 }
1582
1583 /*
1584  * Hold each of the physical pages that are mapped by the specified range of
1585  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1586  * and allow the specified types of access, "prot".  If all of the implied
1587  * pages are successfully held, then the number of held pages is returned
1588  * together with pointers to those pages in the array "ma".  However, if any
1589  * of the pages cannot be held, -1 is returned.
1590  */
1591 int
1592 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1593     vm_prot_t prot, vm_page_t *ma, int max_count)
1594 {
1595         vm_offset_t end, va;
1596         vm_page_t *mp;
1597         int count;
1598         boolean_t pmap_failed;
1599
1600         if (len == 0)
1601                 return (0);
1602         end = round_page(addr + len);
1603         addr = trunc_page(addr);
1604
1605         /*
1606          * Check for illegal addresses.
1607          */
1608         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1609                 return (-1);
1610
1611         if (atop(end - addr) > max_count)
1612                 panic("vm_fault_quick_hold_pages: count > max_count");
1613         count = atop(end - addr);
1614
1615         /*
1616          * Most likely, the physical pages are resident in the pmap, so it is
1617          * faster to try pmap_extract_and_hold() first.
1618          */
1619         pmap_failed = FALSE;
1620         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1621                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1622                 if (*mp == NULL)
1623                         pmap_failed = TRUE;
1624                 else if ((prot & VM_PROT_WRITE) != 0 &&
1625                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1626                         /*
1627                          * Explicitly dirty the physical page.  Otherwise, the
1628                          * caller's changes may go unnoticed because they are
1629                          * performed through an unmanaged mapping or by a DMA
1630                          * operation.
1631                          *
1632                          * The object lock is not held here.
1633                          * See vm_page_clear_dirty_mask().
1634                          */
1635                         vm_page_dirty(*mp);
1636                 }
1637         }
1638         if (pmap_failed) {
1639                 /*
1640                  * One or more pages could not be held by the pmap.  Either no
1641                  * page was mapped at the specified virtual address or that
1642                  * mapping had insufficient permissions.  Attempt to fault in
1643                  * and hold these pages.
1644                  *
1645                  * If vm_fault_disable_pagefaults() was called,
1646                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1647                  * acquire MD VM locks, which means we must not call
1648                  * vm_fault().  Some (out of tree) callers mark
1649                  * too wide a code area with vm_fault_disable_pagefaults()
1650                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1651                  * the proper behaviour explicitly.
1652                  */
1653                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1654                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1655                         goto error;
1656                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1657                         if (*mp == NULL && vm_fault(map, va, prot,
1658                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1659                                 goto error;
1660         }
1661         return (count);
1662 error:
1663         for (mp = ma; mp < ma + count; mp++)
1664                 if (*mp != NULL)
1665                         vm_page_unwire(*mp, PQ_INACTIVE);
1666         return (-1);
1667 }
1668
1669 /*
1670  *      Routine:
1671  *              vm_fault_copy_entry
1672  *      Function:
1673  *              Create new shadow object backing dst_entry with private copy of
1674  *              all underlying pages. When src_entry is equal to dst_entry,
1675  *              function implements COW for wired-down map entry. Otherwise,
1676  *              it forks wired entry into dst_map.
1677  *
1678  *      In/out conditions:
1679  *              The source and destination maps must be locked for write.
1680  *              The source map entry must be wired down (or be a sharing map
1681  *              entry corresponding to a main map entry that is wired down).
1682  */
1683 void
1684 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1685     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1686     vm_ooffset_t *fork_charge)
1687 {
1688         vm_object_t backing_object, dst_object, object, src_object;
1689         vm_pindex_t dst_pindex, pindex, src_pindex;
1690         vm_prot_t access, prot;
1691         vm_offset_t vaddr;
1692         vm_page_t dst_m;
1693         vm_page_t src_m;
1694         boolean_t upgrade;
1695
1696 #ifdef  lint
1697         src_map++;
1698 #endif  /* lint */
1699
1700         upgrade = src_entry == dst_entry;
1701         access = prot = dst_entry->protection;
1702
1703         src_object = src_entry->object.vm_object;
1704         src_pindex = OFF_TO_IDX(src_entry->offset);
1705
1706         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1707                 dst_object = src_object;
1708                 vm_object_reference(dst_object);
1709         } else {
1710                 /*
1711                  * Create the top-level object for the destination entry.
1712                  * Doesn't actually shadow anything - we copy the pages
1713                  * directly.
1714                  */
1715                 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1716                     dst_entry->start), NULL, NULL, 0);
1717 #if VM_NRESERVLEVEL > 0
1718                 dst_object->flags |= OBJ_COLORED;
1719                 dst_object->pg_color = atop(dst_entry->start);
1720 #endif
1721                 dst_object->domain = src_object->domain;
1722                 dst_object->charge = dst_entry->end - dst_entry->start;
1723         }
1724
1725         VM_OBJECT_WLOCK(dst_object);
1726         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1727             ("vm_fault_copy_entry: vm_object not NULL"));
1728         if (src_object != dst_object) {
1729                 dst_entry->object.vm_object = dst_object;
1730                 dst_entry->offset = 0;
1731                 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1732         }
1733         if (fork_charge != NULL) {
1734                 KASSERT(dst_entry->cred == NULL,
1735                     ("vm_fault_copy_entry: leaked swp charge"));
1736                 dst_object->cred = curthread->td_ucred;
1737                 crhold(dst_object->cred);
1738                 *fork_charge += dst_object->charge;
1739         } else if ((dst_object->type == OBJT_DEFAULT ||
1740             dst_object->type == OBJT_SWAP) &&
1741             dst_object->cred == NULL) {
1742                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1743                     dst_entry));
1744                 dst_object->cred = dst_entry->cred;
1745                 dst_entry->cred = NULL;
1746         }
1747
1748         /*
1749          * If not an upgrade, then enter the mappings in the pmap as
1750          * read and/or execute accesses.  Otherwise, enter them as
1751          * write accesses.
1752          *
1753          * A writeable large page mapping is only created if all of
1754          * the constituent small page mappings are modified. Marking
1755          * PTEs as modified on inception allows promotion to happen
1756          * without taking potentially large number of soft faults.
1757          */
1758         if (!upgrade)
1759                 access &= ~VM_PROT_WRITE;
1760
1761         /*
1762          * Loop through all of the virtual pages within the entry's
1763          * range, copying each page from the source object to the
1764          * destination object.  Since the source is wired, those pages
1765          * must exist.  In contrast, the destination is pageable.
1766          * Since the destination object doesn't share any backing storage
1767          * with the source object, all of its pages must be dirtied,
1768          * regardless of whether they can be written.
1769          */
1770         for (vaddr = dst_entry->start, dst_pindex = 0;
1771             vaddr < dst_entry->end;
1772             vaddr += PAGE_SIZE, dst_pindex++) {
1773 again:
1774                 /*
1775                  * Find the page in the source object, and copy it in.
1776                  * Because the source is wired down, the page will be
1777                  * in memory.
1778                  */
1779                 if (src_object != dst_object)
1780                         VM_OBJECT_RLOCK(src_object);
1781                 object = src_object;
1782                 pindex = src_pindex + dst_pindex;
1783                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1784                     (backing_object = object->backing_object) != NULL) {
1785                         /*
1786                          * Unless the source mapping is read-only or
1787                          * it is presently being upgraded from
1788                          * read-only, the first object in the shadow
1789                          * chain should provide all of the pages.  In
1790                          * other words, this loop body should never be
1791                          * executed when the source mapping is already
1792                          * read/write.
1793                          */
1794                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1795                             upgrade,
1796                             ("vm_fault_copy_entry: main object missing page"));
1797
1798                         VM_OBJECT_RLOCK(backing_object);
1799                         pindex += OFF_TO_IDX(object->backing_object_offset);
1800                         if (object != dst_object)
1801                                 VM_OBJECT_RUNLOCK(object);
1802                         object = backing_object;
1803                 }
1804                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1805
1806                 if (object != dst_object) {
1807                         /*
1808                          * Allocate a page in the destination object.
1809                          */
1810                         dst_m = vm_page_alloc(dst_object, (src_object ==
1811                             dst_object ? src_pindex : 0) + dst_pindex,
1812                             VM_ALLOC_NORMAL);
1813                         if (dst_m == NULL) {
1814                                 VM_OBJECT_WUNLOCK(dst_object);
1815                                 VM_OBJECT_RUNLOCK(object);
1816                                 vm_wait(dst_object);
1817                                 VM_OBJECT_WLOCK(dst_object);
1818                                 goto again;
1819                         }
1820                         pmap_copy_page(src_m, dst_m);
1821                         VM_OBJECT_RUNLOCK(object);
1822                         dst_m->dirty = dst_m->valid = src_m->valid;
1823                 } else {
1824                         dst_m = src_m;
1825                         if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1826                                 goto again;
1827                         if (dst_m->pindex >= dst_object->size) {
1828                                 /*
1829                                  * We are upgrading.  Index can occur
1830                                  * out of bounds if the object type is
1831                                  * vnode and the file was truncated.
1832                                  */
1833                                 vm_page_xunbusy(dst_m);
1834                                 break;
1835                         }
1836                 }
1837                 VM_OBJECT_WUNLOCK(dst_object);
1838
1839                 /*
1840                  * Enter it in the pmap. If a wired, copy-on-write
1841                  * mapping is being replaced by a write-enabled
1842                  * mapping, then wire that new mapping.
1843                  *
1844                  * The page can be invalid if the user called
1845                  * msync(MS_INVALIDATE) or truncated the backing vnode
1846                  * or shared memory object.  In this case, do not
1847                  * insert it into pmap, but still do the copy so that
1848                  * all copies of the wired map entry have similar
1849                  * backing pages.
1850                  */
1851                 if (vm_page_all_valid(dst_m)) {
1852                         pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1853                             access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1854                 }
1855
1856                 /*
1857                  * Mark it no longer busy, and put it on the active list.
1858                  */
1859                 VM_OBJECT_WLOCK(dst_object);
1860
1861                 if (upgrade) {
1862                         if (src_m != dst_m) {
1863                                 vm_page_unwire(src_m, PQ_INACTIVE);
1864                                 vm_page_wire(dst_m);
1865                         } else {
1866                                 KASSERT(vm_page_wired(dst_m),
1867                                     ("dst_m %p is not wired", dst_m));
1868                         }
1869                 } else {
1870                         vm_page_activate(dst_m);
1871                 }
1872                 vm_page_xunbusy(dst_m);
1873         }
1874         VM_OBJECT_WUNLOCK(dst_object);
1875         if (upgrade) {
1876                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1877                 vm_object_deallocate(src_object);
1878         }
1879 }
1880
1881 /*
1882  * Block entry into the machine-independent layer's page fault handler by
1883  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1884  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1885  * spurious page faults.
1886  */
1887 int
1888 vm_fault_disable_pagefaults(void)
1889 {
1890
1891         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1892 }
1893
1894 void
1895 vm_fault_enable_pagefaults(int save)
1896 {
1897
1898         curthread_pflags_restore(save);
1899 }