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_offset_t vaddr;
 124         vm_page_t m;
 125         vm_page_t m_cow;
 126         vm_object_t object;
 127         vm_pindex_t pindex;
 128         vm_page_t first_m;
 129         vm_object_t     first_object;
 130         vm_pindex_t first_pindex;
 131         vm_map_t map;
 132         vm_map_entry_t entry;
 133         int map_generation;
 134         bool lookup_still_valid;
 135         struct vnode *vp;
 136 };
 137
 138 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 139             int ahead);
 140 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 141             int backward, int forward, bool obj_locked);
 142
 143 static int vm_pfault_oom_attempts = 3;
 144 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
 145     &vm_pfault_oom_attempts, 0,
 146     "Number of page allocation attempts in page fault handler before it "
 147     "triggers OOM handling");
 148
 149 static int vm_pfault_oom_wait = 10;
 150 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
 151     &vm_pfault_oom_wait, 0,
 152     "Number of seconds to wait for free pages before retrying "
 153     "the page fault handler");
 154
 155 static inline void
 156 fault_page_release(vm_page_t *mp)
 157 {
 158         vm_page_t m;
 159
 160         m = *mp;
 161         if (m != NULL) {
 162                 /*
 163                  * We are likely to loop around again and attempt to busy
 164                  * this page.  Deactivating it leaves it available for
 165                  * pageout while optimizing fault restarts.
 166                  */
 167                 vm_page_deactivate(m);
 168                 vm_page_xunbusy(m);
 169                 *mp = NULL;
 170         }
 171 }
 172
 173 static inline void
 174 fault_page_free(vm_page_t *mp)
 175 {
 176         vm_page_t m;
 177
 178         m = *mp;
 179         if (m != NULL) {
 180                 VM_OBJECT_ASSERT_WLOCKED(m->object);
 181                 if (!vm_page_wired(m))
 182                         vm_page_free(m);
 183                 else
 184                         vm_page_xunbusy(m);
 185                 *mp = NULL;
 186         }
 187 }
 188
 189 static inline void
 190 unlock_map(struct faultstate *fs)
 191 {
 192
 193         if (fs->lookup_still_valid) {
 194                 vm_map_lookup_done(fs->map, fs->entry);
 195                 fs->lookup_still_valid = false;
 196         }
 197 }
 198
 199 static void
 200 unlock_vp(struct faultstate *fs)
 201 {
 202
 203         if (fs->vp != NULL) {
 204                 vput(fs->vp);
 205                 fs->vp = NULL;
 206         }
 207 }
 208
 209 static void
 210 fault_deallocate(struct faultstate *fs)
 211 {
 212
 213         fault_page_release(&fs->m_cow);
 214         fault_page_release(&fs->m);
 215         vm_object_pip_wakeup(fs->object);
 216         if (fs->object != fs->first_object) {
 217                 VM_OBJECT_WLOCK(fs->first_object);
 218                 fault_page_free(&fs->first_m);
 219                 VM_OBJECT_WUNLOCK(fs->first_object);
 220                 vm_object_pip_wakeup(fs->first_object);
 221         }
 222         vm_object_deallocate(fs->first_object);
 223         unlock_map(fs);
 224         unlock_vp(fs);
 225 }
 226
 227 static void
 228 unlock_and_deallocate(struct faultstate *fs)
 229 {
 230
 231         VM_OBJECT_WUNLOCK(fs->object);
 232         fault_deallocate(fs);
 233 }
 234
 235 static void
 236 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 237     vm_prot_t fault_type, int fault_flags)
 238 {
 239         bool need_dirty;
 240
 241         if (((prot & VM_PROT_WRITE) == 0 &&
 242             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 243             (m->oflags & VPO_UNMANAGED) != 0)
 244                 return;
 245
 246         VM_PAGE_OBJECT_BUSY_ASSERT(m);
 247
 248         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 249             (fault_flags & VM_FAULT_WIRE) == 0) ||
 250             (fault_flags & VM_FAULT_DIRTY) != 0;
 251
 252         vm_object_set_writeable_dirty(m->object);
 253
 254         /*
 255          * If the fault is a write, we know that this page is being
 256          * written NOW so dirty it explicitly to save on
 257          * pmap_is_modified() calls later.
 258          *
 259          * Also, since the page is now dirty, we can possibly tell
 260          * the pager to release any swap backing the page.
 261          */
 262         if (need_dirty && vm_page_set_dirty(m) == 0) {
 263                 /*
 264                  * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
 265                  * if the page is already dirty to prevent data written with
 266                  * the expectation of being synced from not being synced.
 267                  * Likewise if this entry does not request NOSYNC then make
 268                  * sure the page isn't marked NOSYNC.  Applications sharing
 269                  * data should use the same flags to avoid ping ponging.
 270                  */
 271                 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0)
 272                         vm_page_aflag_set(m, PGA_NOSYNC);
 273                 else
 274                         vm_page_aflag_clear(m, PGA_NOSYNC);
 275         }
 276
 277 }
 278
 279 /*
 280  * Unlocks fs.first_object and fs.map on success.
 281  */
 282 static int
 283 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 284     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 285 {
 286         vm_page_t m, m_map;
 287 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 288     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 289     VM_NRESERVLEVEL > 0
 290         vm_page_t m_super;
 291         int flags;
 292 #endif
 293         int psind, rv;
 294
 295         MPASS(fs->vp == NULL);
 296         vm_object_busy(fs->first_object);
 297         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 298         /* A busy page can be mapped for read|execute access. */
 299         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 300             vm_page_busied(m)) || !vm_page_all_valid(m)) {
 301                 rv = KERN_FAILURE;
 302                 goto out;
 303         }
 304         m_map = m;
 305         psind = 0;
 306 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 307     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 308     VM_NRESERVLEVEL > 0
 309         if ((m->flags & PG_FICTITIOUS) == 0 &&
 310             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 311             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 312             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 313             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 314             (pagesizes[m_super->psind] - 1)) && !wired &&
 315             pmap_ps_enabled(fs->map->pmap)) {
 316                 flags = PS_ALL_VALID;
 317                 if ((prot & VM_PROT_WRITE) != 0) {
 318                         /*
 319                          * Create a superpage mapping allowing write access
 320                          * only if none of the constituent pages are busy and
 321                          * all of them are already dirty (except possibly for
 322                          * the page that was faulted on).
 323                          */
 324                         flags |= PS_NONE_BUSY;
 325                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 326                                 flags |= PS_ALL_DIRTY;
 327                 }
 328                 if (vm_page_ps_test(m_super, flags, m)) {
 329                         m_map = m_super;
 330                         psind = m_super->psind;
 331                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 332                         /* Preset the modified bit for dirty superpages. */
 333                         if ((flags & PS_ALL_DIRTY) != 0)
 334                                 fault_type |= VM_PROT_WRITE;
 335                 }
 336         }
 337 #endif
 338         rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
 339             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
 340         if (rv != KERN_SUCCESS)
 341                 goto out;
 342         if (m_hold != NULL) {
 343                 *m_hold = m;
 344                 vm_page_wire(m);
 345         }
 346         if (psind == 0 && !wired)
 347                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 348         VM_OBJECT_RUNLOCK(fs->first_object);
 349         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags);
 350         vm_map_lookup_done(fs->map, fs->entry);
 351         curthread->td_ru.ru_minflt++;
 352
 353 out:
 354         vm_object_unbusy(fs->first_object);
 355         return (rv);
 356 }
 357
 358 static void
 359 vm_fault_restore_map_lock(struct faultstate *fs)
 360 {
 361
 362         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 363         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 364
 365         if (!vm_map_trylock_read(fs->map)) {
 366                 VM_OBJECT_WUNLOCK(fs->first_object);
 367                 vm_map_lock_read(fs->map);
 368                 VM_OBJECT_WLOCK(fs->first_object);
 369         }
 370         fs->lookup_still_valid = true;
 371 }
 372
 373 static void
 374 vm_fault_populate_check_page(vm_page_t m)
 375 {
 376
 377         /*
 378          * Check each page to ensure that the pager is obeying the
 379          * interface: the page must be installed in the object, fully
 380          * valid, and exclusively busied.
 381          */
 382         MPASS(m != NULL);
 383         MPASS(vm_page_all_valid(m));
 384         MPASS(vm_page_xbusied(m));
 385 }
 386
 387 static void
 388 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 389     vm_pindex_t last)
 390 {
 391         vm_page_t m;
 392         vm_pindex_t pidx;
 393
 394         VM_OBJECT_ASSERT_WLOCKED(object);
 395         MPASS(first <= last);
 396         for (pidx = first, m = vm_page_lookup(object, pidx);
 397             pidx <= last; pidx++, m = vm_page_next(m)) {
 398                 vm_fault_populate_check_page(m);
 399                 vm_page_deactivate(m);
 400                 vm_page_xunbusy(m);
 401         }
 402 }
 403
 404 static int
 405 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 406     int fault_flags, boolean_t wired, vm_page_t *m_hold)
 407 {
 408         vm_offset_t vaddr;
 409         vm_page_t m;
 410         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 411         int i, npages, psind, rv;
 412
 413         MPASS(fs->object == fs->first_object);
 414         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 415         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 416         MPASS(fs->first_object->backing_object == NULL);
 417         MPASS(fs->lookup_still_valid);
 418
 419         pager_first = OFF_TO_IDX(fs->entry->offset);
 420         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 421         unlock_map(fs);
 422         unlock_vp(fs);
 423
 424         /*
 425          * Call the pager (driver) populate() method.
 426          *
 427          * There is no guarantee that the method will be called again
 428          * if the current fault is for read, and a future fault is
 429          * for write.  Report the entry's maximum allowed protection
 430          * to the driver.
 431          */
 432         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 433             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 434
 435         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 436         if (rv == VM_PAGER_BAD) {
 437                 /*
 438                  * VM_PAGER_BAD is the backdoor for a pager to request
 439                  * normal fault handling.
 440                  */
 441                 vm_fault_restore_map_lock(fs);
 442                 if (fs->map->timestamp != fs->map_generation)
 443                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 444                 return (KERN_NOT_RECEIVER);
 445         }
 446         if (rv != VM_PAGER_OK)
 447                 return (KERN_FAILURE); /* AKA SIGSEGV */
 448
 449         /* Ensure that the driver is obeying the interface. */
 450         MPASS(pager_first <= pager_last);
 451         MPASS(fs->first_pindex <= pager_last);
 452         MPASS(fs->first_pindex >= pager_first);
 453         MPASS(pager_last < fs->first_object->size);
 454
 455         vm_fault_restore_map_lock(fs);
 456         if (fs->map->timestamp != fs->map_generation) {
 457                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 458                     pager_last);
 459                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 460         }
 461
 462         /*
 463          * The map is unchanged after our last unlock.  Process the fault.
 464          *
 465          * The range [pager_first, pager_last] that is given to the
 466          * pager is only a hint.  The pager may populate any range
 467          * within the object that includes the requested page index.
 468          * In case the pager expanded the range, clip it to fit into
 469          * the map entry.
 470          */
 471         map_first = OFF_TO_IDX(fs->entry->offset);
 472         if (map_first > pager_first) {
 473                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 474                     map_first - 1);
 475                 pager_first = map_first;
 476         }
 477         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 478         if (map_last < pager_last) {
 479                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 480                     pager_last);
 481                 pager_last = map_last;
 482         }
 483         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 484             pidx <= pager_last;
 485             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 486                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 487 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 488     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
 489                 psind = m->psind;
 490                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 491                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 492                     !pmap_ps_enabled(fs->map->pmap) || wired))
 493                         psind = 0;
 494 #else
 495                 psind = 0;
 496 #endif
 497                 npages = atop(pagesizes[psind]);
 498                 for (i = 0; i < npages; i++) {
 499                         vm_fault_populate_check_page(&m[i]);
 500                         vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
 501                             fault_flags);
 502                 }
 503                 VM_OBJECT_WUNLOCK(fs->first_object);
 504                 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
 505                     (wired ? PMAP_ENTER_WIRED : 0), psind);
 506 #if defined(__amd64__)
 507                 if (psind > 0 && rv == KERN_FAILURE) {
 508                         for (i = 0; i < npages; i++) {
 509                                 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
 510                                     &m[i], prot, fault_type |
 511                                     (wired ? PMAP_ENTER_WIRED : 0), 0);
 512                                 MPASS(rv == KERN_SUCCESS);
 513                         }
 514                 }
 515 #else
 516                 MPASS(rv == KERN_SUCCESS);
 517 #endif
 518                 VM_OBJECT_WLOCK(fs->first_object);
 519                 for (i = 0; i < npages; i++) {
 520                         if ((fault_flags & VM_FAULT_WIRE) != 0)
 521                                 vm_page_wire(&m[i]);
 522                         else
 523                                 vm_page_activate(&m[i]);
 524                         if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
 525                                 *m_hold = &m[i];
 526                                 vm_page_wire(&m[i]);
 527                         }
 528                         vm_page_xunbusy(&m[i]);
 529                 }
 530         }
 531         curthread->td_ru.ru_majflt++;
 532         return (KERN_SUCCESS);
 533 }
 534
 535 static int prot_fault_translation;
 536 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
 537     &prot_fault_translation, 0,
 538     "Control signal to deliver on protection fault");
 539
 540 /* compat definition to keep common code for signal translation */
 541 #define UCODE_PAGEFLT   12
 542 #ifdef T_PAGEFLT
 543 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
 544 #endif
 545
 546 /*
 547  *      vm_fault_trap:
 548  *
 549  *      Handle a page fault occurring at the given address,
 550  *      requiring the given permissions, in the map specified.
 551  *      If successful, the page is inserted into the
 552  *      associated physical map.
 553  *
 554  *      NOTE: the given address should be truncated to the
 555  *      proper page address.
 556  *
 557  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 558  *      a standard error specifying why the fault is fatal is returned.
 559  *
 560  *      The map in question must be referenced, and remains so.
 561  *      Caller may hold no locks.
 562  */
 563 int
 564 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 565     int fault_flags, int *signo, int *ucode)
 566 {
 567         int result;
 568
 569         MPASS(signo == NULL || ucode != NULL);
 570 #ifdef KTRACE
 571         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
 572                 ktrfault(vaddr, fault_type);
 573 #endif
 574         result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
 575             NULL);
 576         KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
 577             result == KERN_INVALID_ADDRESS ||
 578             result == KERN_RESOURCE_SHORTAGE ||
 579             result == KERN_PROTECTION_FAILURE ||
 580             result == KERN_OUT_OF_BOUNDS,
 581             ("Unexpected Mach error %d from vm_fault()", result));
 582 #ifdef KTRACE
 583         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
 584                 ktrfaultend(result);
 585 #endif
 586         if (result != KERN_SUCCESS && signo != NULL) {
 587                 switch (result) {
 588                 case KERN_FAILURE:
 589                 case KERN_INVALID_ADDRESS:
 590                         *signo = SIGSEGV;
 591                         *ucode = SEGV_MAPERR;
 592                         break;
 593                 case KERN_RESOURCE_SHORTAGE:
 594                         *signo = SIGBUS;
 595                         *ucode = BUS_OOMERR;
 596                         break;
 597                 case KERN_OUT_OF_BOUNDS:
 598                         *signo = SIGBUS;
 599                         *ucode = BUS_OBJERR;
 600                         break;
 601                 case KERN_PROTECTION_FAILURE:
 602                         if (prot_fault_translation == 0) {
 603                                 /*
 604                                  * Autodetect.  This check also covers
 605                                  * the images without the ABI-tag ELF
 606                                  * note.
 607                                  */
 608                                 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
 609                                     curproc->p_osrel >= P_OSREL_SIGSEGV) {
 610                                         *signo = SIGSEGV;
 611                                         *ucode = SEGV_ACCERR;
 612                                 } else {
 613                                         *signo = SIGBUS;
 614                                         *ucode = UCODE_PAGEFLT;
 615                                 }
 616                         } else if (prot_fault_translation == 1) {
 617                                 /* Always compat mode. */
 618                                 *signo = SIGBUS;
 619                                 *ucode = UCODE_PAGEFLT;
 620                         } else {
 621                                 /* Always SIGSEGV mode. */
 622                                 *signo = SIGSEGV;
 623                                 *ucode = SEGV_ACCERR;
 624                         }
 625                         break;
 626                 default:
 627                         KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
 628                             result));
 629                         break;
 630                 }
 631         }
 632         return (result);
 633 }
 634
 635 static int
 636 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
 637 {
 638         struct vnode *vp;
 639         int error, locked;
 640
 641         if (fs->object->type != OBJT_VNODE)
 642                 return (KERN_SUCCESS);
 643         vp = fs->object->handle;
 644         if (vp == fs->vp) {
 645                 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
 646                 return (KERN_SUCCESS);
 647         }
 648
 649         /*
 650          * Perform an unlock in case the desired vnode changed while
 651          * the map was unlocked during a retry.
 652          */
 653         unlock_vp(fs);
 654
 655         locked = VOP_ISLOCKED(vp);
 656         if (locked != LK_EXCLUSIVE)
 657                 locked = LK_SHARED;
 658
 659         /*
 660          * We must not sleep acquiring the vnode lock while we have
 661          * the page exclusive busied or the object's
 662          * paging-in-progress count incremented.  Otherwise, we could
 663          * deadlock.
 664          */
 665         error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
 666         if (error == 0) {
 667                 fs->vp = vp;
 668                 return (KERN_SUCCESS);
 669         }
 670
 671         vhold(vp);
 672         if (objlocked)
 673                 unlock_and_deallocate(fs);
 674         else
 675                 fault_deallocate(fs);
 676         error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
 677         vdrop(vp);
 678         fs->vp = vp;
 679         KASSERT(error == 0, ("vm_fault: vget failed %d", error));
 680         return (KERN_RESOURCE_SHORTAGE);
 681 }
 682
 683 /*
 684  * Calculate the desired readahead.  Handle drop-behind.
 685  *
 686  * Returns the number of readahead blocks to pass to the pager.
 687  */
 688 static int
 689 vm_fault_readahead(struct faultstate *fs)
 690 {
 691         int era, nera;
 692         u_char behavior;
 693
 694         KASSERT(fs->lookup_still_valid, ("map unlocked"));
 695         era = fs->entry->read_ahead;
 696         behavior = vm_map_entry_behavior(fs->entry);
 697         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 698                 nera = 0;
 699         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 700                 nera = VM_FAULT_READ_AHEAD_MAX;
 701                 if (fs->vaddr == fs->entry->next_read)
 702                         vm_fault_dontneed(fs, fs->vaddr, nera);
 703         } else if (fs->vaddr == fs->entry->next_read) {
 704                 /*
 705                  * This is a sequential fault.  Arithmetically
 706                  * increase the requested number of pages in
 707                  * the read-ahead window.  The requested
 708                  * number of pages is "# of sequential faults
 709                  * x (read ahead min + 1) + read ahead min"
 710                  */
 711                 nera = VM_FAULT_READ_AHEAD_MIN;
 712                 if (era > 0) {
 713                         nera += era + 1;
 714                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 715                                 nera = VM_FAULT_READ_AHEAD_MAX;
 716                 }
 717                 if (era == VM_FAULT_READ_AHEAD_MAX)
 718                         vm_fault_dontneed(fs, fs->vaddr, nera);
 719         } else {
 720                 /*
 721                  * This is a non-sequential fault.
 722                  */
 723                 nera = 0;
 724         }
 725         if (era != nera) {
 726                 /*
 727                  * A read lock on the map suffices to update
 728                  * the read ahead count safely.
 729                  */
 730                 fs->entry->read_ahead = nera;
 731         }
 732
 733         return (nera);
 734 }
 735
 736 /*
 737  * Wait/Retry if the page is busy.  We have to do this if the page is
 738  * either exclusive or shared busy because the vm_pager may be using
 739  * read busy for pageouts (and even pageins if it is the vnode pager),
 740  * and we could end up trying to pagein and pageout the same page
 741  * simultaneously.
 742  *
 743  * We can theoretically allow the busy case on a read fault if the page
 744  * is marked valid, but since such pages are typically already pmap'd,
 745  * putting that special case in might be more effort then it is worth.
 746  * We cannot under any circumstances mess around with a shared busied
 747  * page except, perhaps, to pmap it.
 748  */
 749 static void
 750 vm_fault_busy_sleep(struct faultstate *fs)
 751 {
 752         /*
 753          * Reference the page before unlocking and
 754          * sleeping so that the page daemon is less
 755          * likely to reclaim it.
 756          */
 757         vm_page_aflag_set(fs->m, PGA_REFERENCED);
 758         if (fs->object != fs->first_object) {
 759                 fault_page_release(&fs->first_m);
 760                 vm_object_pip_wakeup(fs->first_object);
 761         }
 762         vm_object_pip_wakeup(fs->object);
 763         unlock_map(fs);
 764         if (fs->m == vm_page_lookup(fs->object, fs->pindex))
 765                 vm_page_busy_sleep(fs->m, "vmpfw", false);
 766         else
 767                 VM_OBJECT_WUNLOCK(fs->object);
 768         VM_CNT_INC(v_intrans);
 769         vm_object_deallocate(fs->first_object);
 770 }
 771
 772 int
 773 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 774     int fault_flags, vm_page_t *m_hold)
 775 {
 776         struct faultstate fs;
 777         struct domainset *dset;
 778         vm_object_t next_object, retry_object;
 779         vm_offset_t e_end, e_start;
 780         vm_pindex_t retry_pindex;
 781         vm_prot_t prot, retry_prot;
 782         int ahead, alloc_req, behind, cluster_offset, faultcount;
 783         int nera, oom, result, rv;
 784         u_char behavior;
 785         boolean_t wired;        /* Passed by reference. */
 786         bool dead, hardfault, is_first_object_locked;
 787
 788         VM_CNT_INC(v_vm_faults);
 789
 790         if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
 791                 return (KERN_PROTECTION_FAILURE);
 792
 793         fs.vp = NULL;
 794         fs.vaddr = vaddr;
 795         faultcount = 0;
 796         nera = -1;
 797         hardfault = false;
 798
 799 RetryFault:
 800         oom = 0;
 801 RetryFault_oom:
 802
 803         /*
 804          * Find the backing store object and offset into it to begin the
 805          * search.
 806          */
 807         fs.map = map;
 808         result = vm_map_lookup(&fs.map, vaddr, fault_type |
 809             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 810             &fs.first_pindex, &prot, &wired);
 811         if (result != KERN_SUCCESS) {
 812                 unlock_vp(&fs);
 813                 return (result);
 814         }
 815
 816         fs.map_generation = fs.map->timestamp;
 817
 818         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 819                 panic("%s: fault on nofault entry, addr: %#lx",
 820                     __func__, (u_long)vaddr);
 821         }
 822
 823         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 824             fs.entry->wiring_thread != curthread) {
 825                 vm_map_unlock_read(fs.map);
 826                 vm_map_lock(fs.map);
 827                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 828                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 829                         unlock_vp(&fs);
 830                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 831                         vm_map_unlock_and_wait(fs.map, 0);
 832                 } else
 833                         vm_map_unlock(fs.map);
 834                 goto RetryFault;
 835         }
 836
 837         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 838
 839         if (wired)
 840                 fault_type = prot | (fault_type & VM_PROT_COPY);
 841         else
 842                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 843                     ("!wired && VM_FAULT_WIRE"));
 844
 845         /*
 846          * Try to avoid lock contention on the top-level object through
 847          * special-case handling of some types of page faults, specifically,
 848          * those that are mapping an existing page from the top-level object.
 849          * Under this condition, a read lock on the object suffices, allowing
 850          * multiple page faults of a similar type to run in parallel.
 851          */
 852         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 853             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
 854                 VM_OBJECT_RLOCK(fs.first_object);
 855                 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 856                     fault_flags, wired, m_hold);
 857                 if (rv == KERN_SUCCESS)
 858                         return (rv);
 859                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 860                         VM_OBJECT_RUNLOCK(fs.first_object);
 861                         VM_OBJECT_WLOCK(fs.first_object);
 862                 }
 863         } else {
 864                 VM_OBJECT_WLOCK(fs.first_object);
 865         }
 866
 867         /*
 868          * Make a reference to this object to prevent its disposal while we
 869          * are messing with it.  Once we have the reference, the map is free
 870          * to be diddled.  Since objects reference their shadows (and copies),
 871          * they will stay around as well.
 872          *
 873          * Bump the paging-in-progress count to prevent size changes (e.g.
 874          * truncation operations) during I/O.
 875          */
 876         vm_object_reference_locked(fs.first_object);
 877         vm_object_pip_add(fs.first_object, 1);
 878
 879         fs.lookup_still_valid = true;
 880
 881         fs.m_cow = fs.m = fs.first_m = NULL;
 882
 883         /*
 884          * Search for the page at object/offset.
 885          */
 886         fs.object = fs.first_object;
 887         fs.pindex = fs.first_pindex;
 888         while (TRUE) {
 889                 KASSERT(fs.m == NULL,
 890                     ("page still set %p at loop start", fs.m));
 891                 /*
 892                  * If the object is marked for imminent termination,
 893                  * we retry here, since the collapse pass has raced
 894                  * with us.  Otherwise, if we see terminally dead
 895                  * object, return fail.
 896                  */
 897                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 898                         dead = fs.object->type == OBJT_DEAD;
 899                         unlock_and_deallocate(&fs);
 900                         if (dead)
 901                                 return (KERN_PROTECTION_FAILURE);
 902                         pause("vmf_de", 1);
 903                         goto RetryFault;
 904                 }
 905
 906                 /*
 907                  * See if page is resident
 908                  */
 909                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 910                 if (fs.m != NULL) {
 911                         if (vm_page_tryxbusy(fs.m) == 0) {
 912                                 vm_fault_busy_sleep(&fs);
 913                                 goto RetryFault;
 914                         }
 915
 916                         /*
 917                          * The page is marked busy for other processes and the
 918                          * pagedaemon.  If it still isn't completely valid
 919                          * (readable), jump to readrest, else break-out ( we
 920                          * found the page ).
 921                          */
 922                         if (!vm_page_all_valid(fs.m))
 923                                 goto readrest;
 924                         VM_OBJECT_WUNLOCK(fs.object);
 925                         break; /* break to PAGE HAS BEEN FOUND. */
 926                 }
 927                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 928                 VM_OBJECT_ASSERT_WLOCKED(fs.object);
 929
 930                 /*
 931                  * Page is not resident.  If the pager might contain the page
 932                  * or this is the beginning of the search, allocate a new
 933                  * page.  (Default objects are zero-fill, so there is no real
 934                  * pager for them.)
 935                  */
 936                 if (fs.object->type != OBJT_DEFAULT ||
 937                     fs.object == fs.first_object) {
 938                         if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
 939                                 rv = vm_fault_lock_vnode(&fs, true);
 940                                 MPASS(rv == KERN_SUCCESS ||
 941                                     rv == KERN_RESOURCE_SHORTAGE);
 942                                 if (rv == KERN_RESOURCE_SHORTAGE)
 943                                         goto RetryFault;
 944                         }
 945                         if (fs.pindex >= fs.object->size) {
 946                                 unlock_and_deallocate(&fs);
 947                                 return (KERN_OUT_OF_BOUNDS);
 948                         }
 949
 950                         if (fs.object == fs.first_object &&
 951                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 952                             fs.first_object->shadow_count == 0) {
 953                                 rv = vm_fault_populate(&fs, prot, fault_type,
 954                                     fault_flags, wired, m_hold);
 955                                 switch (rv) {
 956                                 case KERN_SUCCESS:
 957                                 case KERN_FAILURE:
 958                                         unlock_and_deallocate(&fs);
 959                                         return (rv);
 960                                 case KERN_RESOURCE_SHORTAGE:
 961                                         unlock_and_deallocate(&fs);
 962                                         goto RetryFault;
 963                                 case KERN_NOT_RECEIVER:
 964                                         /*
 965                                          * Pager's populate() method
 966                                          * returned VM_PAGER_BAD.
 967                                          */
 968                                         break;
 969                                 default:
 970                                         panic("inconsistent return codes");
 971                                 }
 972                         }
 973
 974                         /*
 975                          * Allocate a new page for this object/offset pair.
 976                          *
 977                          * Unlocked read of the p_flag is harmless. At
 978                          * worst, the P_KILLED might be not observed
 979                          * there, and allocation can fail, causing
 980                          * restart and new reading of the p_flag.
 981                          */
 982                         dset = fs.object->domain.dr_policy;
 983                         if (dset == NULL)
 984                                 dset = curthread->td_domain.dr_policy;
 985                         if (!vm_page_count_severe_set(&dset->ds_mask) ||
 986                             P_KILLED(curproc)) {
 987 #if VM_NRESERVLEVEL > 0
 988                                 vm_object_color(fs.object, atop(vaddr) -
 989                                     fs.pindex);
 990 #endif
 991                                 alloc_req = P_KILLED(curproc) ?
 992                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 993                                 if (fs.object->type != OBJT_VNODE &&
 994                                     fs.object->backing_object == NULL)
 995                                         alloc_req |= VM_ALLOC_ZERO;
 996                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 997                                     alloc_req);
 998                         }
 999                         if (fs.m == NULL) {
1000                                 unlock_and_deallocate(&fs);
1001                                 if (vm_pfault_oom_attempts < 0 ||
1002                                     oom < vm_pfault_oom_attempts) {
1003                                         oom++;
1004                                         vm_waitpfault(dset,
1005                                             vm_pfault_oom_wait * hz);
1006                                         goto RetryFault_oom;
1007                                 }
1008                                 if (bootverbose)
1009                                         printf(
1010         "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1011                                             curproc->p_pid, curproc->p_comm);
1012                                 vm_pageout_oom(VM_OOM_MEM_PF);
1013                                 goto RetryFault;
1014                         }
1015                 }
1016
1017 readrest:
1018                 /*
1019                  * Default objects have no pager so no exclusive busy exists
1020                  * to protect this page in the chain.  Skip to the next
1021                  * object without dropping the lock to preserve atomicity of
1022                  * shadow faults.
1023                  */
1024                 if (fs.object->type == OBJT_DEFAULT)
1025                         goto next;
1026
1027                 /*
1028                  * At this point, we have either allocated a new page or found
1029                  * an existing page that is only partially valid.
1030                  *
1031                  * We hold a reference on the current object and the page is
1032                  * exclusive busied.  The exclusive busy prevents simultaneous
1033                  * faults and collapses while the object lock is dropped.
1034                  */
1035                 VM_OBJECT_WUNLOCK(fs.object);
1036
1037                 /*
1038                  * If the pager for the current object might have the page,
1039                  * then determine the number of additional pages to read and
1040                  * potentially reprioritize previously read pages for earlier
1041                  * reclamation.  These operations should only be performed
1042                  * once per page fault.  Even if the current pager doesn't
1043                  * have the page, the number of additional pages to read will
1044                  * apply to subsequent objects in the shadow chain.
1045                  */
1046                 if (nera == -1 && !P_KILLED(curproc)) {
1047                         nera = vm_fault_readahead(&fs);
1048                         /*
1049                          * Prepare for unlocking the map.  Save the map
1050                          * entry's start and end addresses, which are used to
1051                          * optimize the size of the pager operation below.
1052                          * Even if the map entry's addresses change after
1053                          * unlocking the map, using the saved addresses is
1054                          * safe.
1055                          */
1056                         e_start = fs.entry->start;
1057                         e_end = fs.entry->end;
1058                         behavior = vm_map_entry_behavior(fs.entry);
1059                 }
1060
1061                 /*
1062                  * Call the pager to retrieve the page if there is a chance
1063                  * that the pager has it, and potentially retrieve additional
1064                  * pages at the same time.
1065                  */
1066                 if (fs.object->type != OBJT_DEFAULT) {
1067                         /*
1068                          * Release the map lock before locking the vnode or
1069                          * sleeping in the pager.  (If the current object has
1070                          * a shadow, then an earlier iteration of this loop
1071                          * may have already unlocked the map.)
1072                          */
1073                         unlock_map(&fs);
1074
1075                         rv = vm_fault_lock_vnode(&fs, false);
1076                         MPASS(rv == KERN_SUCCESS ||
1077                             rv == KERN_RESOURCE_SHORTAGE);
1078                         if (rv == KERN_RESOURCE_SHORTAGE)
1079                                 goto RetryFault;
1080                         KASSERT(fs.vp == NULL || !fs.map->system_map,
1081                             ("vm_fault: vnode-backed object mapped by system map"));
1082
1083                         /*
1084                          * Page in the requested page and hint the pager,
1085                          * that it may bring up surrounding pages.
1086                          */
1087                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1088                             P_KILLED(curproc)) {
1089                                 behind = 0;
1090                                 ahead = 0;
1091                         } else {
1092                                 /* Is this a sequential fault? */
1093                                 if (nera > 0) {
1094                                         behind = 0;
1095                                         ahead = nera;
1096                                 } else {
1097                                         /*
1098                                          * Request a cluster of pages that is
1099                                          * aligned to a VM_FAULT_READ_DEFAULT
1100                                          * page offset boundary within the
1101                                          * object.  Alignment to a page offset
1102                                          * boundary is more likely to coincide
1103                                          * with the underlying file system
1104                                          * block than alignment to a virtual
1105                                          * address boundary.
1106                                          */
1107                                         cluster_offset = fs.pindex %
1108                                             VM_FAULT_READ_DEFAULT;
1109                                         behind = ulmin(cluster_offset,
1110                                             atop(vaddr - e_start));
1111                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
1112                                             cluster_offset;
1113                                 }
1114                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1115                         }
1116                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1117                             &behind, &ahead);
1118                         if (rv == VM_PAGER_OK) {
1119                                 faultcount = behind + 1 + ahead;
1120                                 hardfault = true;
1121                                 break; /* break to PAGE HAS BEEN FOUND. */
1122                         }
1123                         VM_OBJECT_WLOCK(fs.object);
1124                         if (rv == VM_PAGER_ERROR)
1125                                 printf("vm_fault: pager read error, pid %d (%s)\n",
1126                                     curproc->p_pid, curproc->p_comm);
1127
1128                         /*
1129                          * If an I/O error occurred or the requested page was
1130                          * outside the range of the pager, clean up and return
1131                          * an error.
1132                          */
1133                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1134                                 fault_page_free(&fs.m);
1135                                 unlock_and_deallocate(&fs);
1136                                 return (KERN_OUT_OF_BOUNDS);
1137                         }
1138
1139                 }
1140
1141 next:
1142                 /*
1143                  * The requested page does not exist at this object/
1144                  * offset.  Remove the invalid page from the object,
1145                  * waking up anyone waiting for it, and continue on to
1146                  * the next object.  However, if this is the top-level
1147                  * object, we must leave the busy page in place to
1148                  * prevent another process from rushing past us, and
1149                  * inserting the page in that object at the same time
1150                  * that we are.
1151                  */
1152                 if (fs.object == fs.first_object) {
1153                         fs.first_m = fs.m;
1154                         fs.m = NULL;
1155                 } else
1156                         fault_page_free(&fs.m);
1157
1158                 /*
1159                  * Move on to the next object.  Lock the next object before
1160                  * unlocking the current one.
1161                  */
1162                 VM_OBJECT_ASSERT_WLOCKED(fs.object);
1163                 next_object = fs.object->backing_object;
1164                 if (next_object == NULL) {
1165                         /*
1166                          * If there's no object left, fill the page in the top
1167                          * object with zeros.
1168                          */
1169                         VM_OBJECT_WUNLOCK(fs.object);
1170                         if (fs.object != fs.first_object) {
1171                                 vm_object_pip_wakeup(fs.object);
1172                                 fs.object = fs.first_object;
1173                                 fs.pindex = fs.first_pindex;
1174                         }
1175                         MPASS(fs.first_m != NULL);
1176                         MPASS(fs.m == NULL);
1177                         fs.m = fs.first_m;
1178                         fs.first_m = NULL;
1179
1180                         /*
1181                          * Zero the page if necessary and mark it valid.
1182                          */
1183                         if ((fs.m->flags & PG_ZERO) == 0) {
1184                                 pmap_zero_page(fs.m);
1185                         } else {
1186                                 VM_CNT_INC(v_ozfod);
1187                         }
1188                         VM_CNT_INC(v_zfod);
1189                         vm_page_valid(fs.m);
1190                         /* Don't try to prefault neighboring pages. */
1191                         faultcount = 1;
1192                         break;  /* break to PAGE HAS BEEN FOUND. */
1193                 } else {
1194                         MPASS(fs.first_m != NULL);
1195                         KASSERT(fs.object != next_object,
1196                             ("object loop %p", next_object));
1197                         VM_OBJECT_WLOCK(next_object);
1198                         vm_object_pip_add(next_object, 1);
1199                         if (fs.object != fs.first_object)
1200                                 vm_object_pip_wakeup(fs.object);
1201                         fs.pindex +=
1202                             OFF_TO_IDX(fs.object->backing_object_offset);
1203                         VM_OBJECT_WUNLOCK(fs.object);
1204                         fs.object = next_object;
1205                 }
1206         }
1207
1208         /*
1209          * PAGE HAS BEEN FOUND.  A valid page has been found and exclusively
1210          * busied.  The object lock must no longer be held.
1211          */
1212         vm_page_assert_xbusied(fs.m);
1213         VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1214
1215         /*
1216          * If the page is being written, but isn't already owned by the
1217          * top-level object, we have to copy it into a new page owned by the
1218          * top-level object.
1219          */
1220         if (fs.object != fs.first_object) {
1221                 /*
1222                  * We only really need to copy if we want to write it.
1223                  */
1224                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1225                         /*
1226                          * This allows pages to be virtually copied from a
1227                          * backing_object into the first_object, where the
1228                          * backing object has no other refs to it, and cannot
1229                          * gain any more refs.  Instead of a bcopy, we just
1230                          * move the page from the backing object to the
1231                          * first object.  Note that we must mark the page
1232                          * dirty in the first object so that it will go out
1233                          * to swap when needed.
1234                          */
1235                         is_first_object_locked = false;
1236                         if (
1237                             /*
1238                              * Only one shadow object
1239                              */
1240                             fs.object->shadow_count == 1 &&
1241                             /*
1242                              * No COW refs, except us
1243                              */
1244                             fs.object->ref_count == 1 &&
1245                             /*
1246                              * No one else can look this object up
1247                              */
1248                             fs.object->handle == NULL &&
1249                             /*
1250                              * No other ways to look the object up
1251                              */
1252                             (fs.object->flags & OBJ_ANON) != 0 &&
1253                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1254                             /*
1255                              * We don't chase down the shadow chain
1256                              */
1257                             fs.object == fs.first_object->backing_object &&
1258                             VM_OBJECT_TRYWLOCK(fs.object)) {
1259
1260                                 /*
1261                                  * Remove but keep xbusy for replace.  fs.m is
1262                                  * moved into fs.first_object and left busy
1263                                  * while fs.first_m is conditionally freed.
1264                                  */
1265                                 vm_page_remove_xbusy(fs.m);
1266                                 vm_page_replace(fs.m, fs.first_object,
1267                                     fs.first_pindex, fs.first_m);
1268                                 vm_page_dirty(fs.m);
1269 #if VM_NRESERVLEVEL > 0
1270                                 /*
1271                                  * Rename the reservation.
1272                                  */
1273                                 vm_reserv_rename(fs.m, fs.first_object,
1274                                     fs.object, OFF_TO_IDX(
1275                                     fs.first_object->backing_object_offset));
1276 #endif
1277                                 VM_OBJECT_WUNLOCK(fs.object);
1278                                 VM_OBJECT_WUNLOCK(fs.first_object);
1279                                 fs.first_m = fs.m;
1280                                 fs.m = NULL;
1281                                 VM_CNT_INC(v_cow_optim);
1282                         } else {
1283                                 if (is_first_object_locked)
1284                                         VM_OBJECT_WUNLOCK(fs.first_object);
1285                                 /*
1286                                  * Oh, well, lets copy it.
1287                                  */
1288                                 pmap_copy_page(fs.m, fs.first_m);
1289                                 vm_page_valid(fs.first_m);
1290                                 if (wired && (fault_flags &
1291                                     VM_FAULT_WIRE) == 0) {
1292                                         vm_page_wire(fs.first_m);
1293                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1294                                 }
1295                                 /*
1296                                  * Save the cow page to be released after
1297                                  * pmap_enter is complete.
1298                                  */
1299                                 fs.m_cow = fs.m;
1300                                 fs.m = NULL;
1301                         }
1302                         /*
1303                          * fs.object != fs.first_object due to above
1304                          * conditional
1305                          */
1306                         vm_object_pip_wakeup(fs.object);
1307
1308                         /*
1309                          * We only try to prefault read-only mappings to the
1310                          * neighboring pages when this copy-on-write fault is
1311                          * a hard fault.  In other cases, trying to prefault
1312                          * is typically wasted effort.
1313                          */
1314                         if (faultcount == 0)
1315                                 faultcount = 1;
1316
1317                         /*
1318                          * Only use the new page below...
1319                          */
1320                         fs.object = fs.first_object;
1321                         fs.pindex = fs.first_pindex;
1322                         fs.m = fs.first_m;
1323                         VM_CNT_INC(v_cow_faults);
1324                         curthread->td_cow++;
1325                 } else {
1326                         prot &= ~VM_PROT_WRITE;
1327                 }
1328         }
1329
1330         /*
1331          * We must verify that the maps have not changed since our last
1332          * lookup.
1333          */
1334         if (!fs.lookup_still_valid) {
1335                 if (!vm_map_trylock_read(fs.map)) {
1336                         fault_deallocate(&fs);
1337                         goto RetryFault;
1338                 }
1339                 fs.lookup_still_valid = true;
1340                 if (fs.map->timestamp != fs.map_generation) {
1341                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1342                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1343
1344                         /*
1345                          * If we don't need the page any longer, put it on the inactive
1346                          * list (the easiest thing to do here).  If no one needs it,
1347                          * pageout will grab it eventually.
1348                          */
1349                         if (result != KERN_SUCCESS) {
1350                                 fault_deallocate(&fs);
1351
1352                                 /*
1353                                  * If retry of map lookup would have blocked then
1354                                  * retry fault from start.
1355                                  */
1356                                 if (result == KERN_FAILURE)
1357                                         goto RetryFault;
1358                                 return (result);
1359                         }
1360                         if ((retry_object != fs.first_object) ||
1361                             (retry_pindex != fs.first_pindex)) {
1362                                 fault_deallocate(&fs);
1363                                 goto RetryFault;
1364                         }
1365
1366                         /*
1367                          * Check whether the protection has changed or the object has
1368                          * been copied while we left the map unlocked. Changing from
1369                          * read to write permission is OK - we leave the page
1370                          * write-protected, and catch the write fault. Changing from
1371                          * write to read permission means that we can't mark the page
1372                          * write-enabled after all.
1373                          */
1374                         prot &= retry_prot;
1375                         fault_type &= retry_prot;
1376                         if (prot == 0) {
1377                                 fault_deallocate(&fs);
1378                                 goto RetryFault;
1379                         }
1380
1381                         /* Reassert because wired may have changed. */
1382                         KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1383                             ("!wired && VM_FAULT_WIRE"));
1384                 }
1385         }
1386         VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1387
1388         /*
1389          * If the page was filled by a pager, save the virtual address that
1390          * should be faulted on next under a sequential access pattern to the
1391          * map entry.  A read lock on the map suffices to update this address
1392          * safely.
1393          */
1394         if (hardfault)
1395                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1396
1397         /*
1398          * Page must be completely valid or it is not fit to
1399          * map into user space.  vm_pager_get_pages() ensures this.
1400          */
1401         vm_page_assert_xbusied(fs.m);
1402         KASSERT(vm_page_all_valid(fs.m),
1403             ("vm_fault: page %p partially invalid", fs.m));
1404
1405         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags);
1406
1407         /*
1408          * Put this page into the physical map.  We had to do the unlock above
1409          * because pmap_enter() may sleep.  We don't put the page
1410          * back on the active queue until later so that the pageout daemon
1411          * won't find it (yet).
1412          */
1413         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1414             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1415         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1416             wired == 0)
1417                 vm_fault_prefault(&fs, vaddr,
1418                     faultcount > 0 ? behind : PFBAK,
1419                     faultcount > 0 ? ahead : PFFOR, false);
1420
1421         /*
1422          * If the page is not wired down, then put it where the pageout daemon
1423          * can find it.
1424          */
1425         if ((fault_flags & VM_FAULT_WIRE) != 0)
1426                 vm_page_wire(fs.m);
1427         else
1428                 vm_page_activate(fs.m);
1429         if (m_hold != NULL) {
1430                 *m_hold = fs.m;
1431                 vm_page_wire(fs.m);
1432         }
1433         vm_page_xunbusy(fs.m);
1434         fs.m = NULL;
1435
1436         /*
1437          * Unlock everything, and return
1438          */
1439         fault_deallocate(&fs);
1440         if (hardfault) {
1441                 VM_CNT_INC(v_io_faults);
1442                 curthread->td_ru.ru_majflt++;
1443 #ifdef RACCT
1444                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1445                         PROC_LOCK(curproc);
1446                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1447                                 racct_add_force(curproc, RACCT_WRITEBPS,
1448                                     PAGE_SIZE + behind * PAGE_SIZE);
1449                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1450                         } else {
1451                                 racct_add_force(curproc, RACCT_READBPS,
1452                                     PAGE_SIZE + ahead * PAGE_SIZE);
1453                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1454                         }
1455                         PROC_UNLOCK(curproc);
1456                 }
1457 #endif
1458         } else
1459                 curthread->td_ru.ru_minflt++;
1460
1461         return (KERN_SUCCESS);
1462 }
1463
1464 /*
1465  * Speed up the reclamation of pages that precede the faulting pindex within
1466  * the first object of the shadow chain.  Essentially, perform the equivalent
1467  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1468  * the faulting pindex by the cluster size when the pages read by vm_fault()
1469  * cross a cluster-size boundary.  The cluster size is the greater of the
1470  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1471  *
1472  * When "fs->first_object" is a shadow object, the pages in the backing object
1473  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1474  * function must only be concerned with pages in the first object.
1475  */
1476 static void
1477 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1478 {
1479         vm_map_entry_t entry;
1480         vm_object_t first_object, object;
1481         vm_offset_t end, start;
1482         vm_page_t m, m_next;
1483         vm_pindex_t pend, pstart;
1484         vm_size_t size;
1485
1486         object = fs->object;
1487         VM_OBJECT_ASSERT_UNLOCKED(object);
1488         first_object = fs->first_object;
1489         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1490         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1491                 VM_OBJECT_RLOCK(first_object);
1492                 size = VM_FAULT_DONTNEED_MIN;
1493                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1494                         size = pagesizes[1];
1495                 end = rounddown2(vaddr, size);
1496                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1497                     (entry = fs->entry)->start < end) {
1498                         if (end - entry->start < size)
1499                                 start = entry->start;
1500                         else
1501                                 start = end - size;
1502                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1503                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1504                             entry->start);
1505                         m_next = vm_page_find_least(first_object, pstart);
1506                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1507                             entry->start);
1508                         while ((m = m_next) != NULL && m->pindex < pend) {
1509                                 m_next = TAILQ_NEXT(m, listq);
1510                                 if (!vm_page_all_valid(m) ||
1511                                     vm_page_busied(m))
1512                                         continue;
1513
1514                                 /*
1515                                  * Don't clear PGA_REFERENCED, since it would
1516                                  * likely represent a reference by a different
1517                                  * process.
1518                                  *
1519                                  * Typically, at this point, prefetched pages
1520                                  * are still in the inactive queue.  Only
1521                                  * pages that triggered page faults are in the
1522                                  * active queue.  The test for whether the page
1523                                  * is in the inactive queue is racy; in the
1524                                  * worst case we will requeue the page
1525                                  * unnecessarily.
1526                                  */
1527                                 if (!vm_page_inactive(m))
1528                                         vm_page_deactivate(m);
1529                         }
1530                 }
1531                 VM_OBJECT_RUNLOCK(first_object);
1532         }
1533 }
1534
1535 /*
1536  * vm_fault_prefault provides a quick way of clustering
1537  * pagefaults into a processes address space.  It is a "cousin"
1538  * of vm_map_pmap_enter, except it runs at page fault time instead
1539  * of mmap time.
1540  */
1541 static void
1542 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1543     int backward, int forward, bool obj_locked)
1544 {
1545         pmap_t pmap;
1546         vm_map_entry_t entry;
1547         vm_object_t backing_object, lobject;
1548         vm_offset_t addr, starta;
1549         vm_pindex_t pindex;
1550         vm_page_t m;
1551         int i;
1552
1553         pmap = fs->map->pmap;
1554         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1555                 return;
1556
1557         entry = fs->entry;
1558
1559         if (addra < backward * PAGE_SIZE) {
1560                 starta = entry->start;
1561         } else {
1562                 starta = addra - backward * PAGE_SIZE;
1563                 if (starta < entry->start)
1564                         starta = entry->start;
1565         }
1566
1567         /*
1568          * Generate the sequence of virtual addresses that are candidates for
1569          * prefaulting in an outward spiral from the faulting virtual address,
1570          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1571          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1572          * If the candidate address doesn't have a backing physical page, then
1573          * the loop immediately terminates.
1574          */
1575         for (i = 0; i < 2 * imax(backward, forward); i++) {
1576                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1577                     PAGE_SIZE);
1578                 if (addr > addra + forward * PAGE_SIZE)
1579                         addr = 0;
1580
1581                 if (addr < starta || addr >= entry->end)
1582                         continue;
1583
1584                 if (!pmap_is_prefaultable(pmap, addr))
1585                         continue;
1586
1587                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1588                 lobject = entry->object.vm_object;
1589                 if (!obj_locked)
1590                         VM_OBJECT_RLOCK(lobject);
1591                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1592                     lobject->type == OBJT_DEFAULT &&
1593                     (backing_object = lobject->backing_object) != NULL) {
1594                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1595                             0, ("vm_fault_prefault: unaligned object offset"));
1596                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1597                         VM_OBJECT_RLOCK(backing_object);
1598                         if (!obj_locked || lobject != entry->object.vm_object)
1599                                 VM_OBJECT_RUNLOCK(lobject);
1600                         lobject = backing_object;
1601                 }
1602                 if (m == NULL) {
1603                         if (!obj_locked || lobject != entry->object.vm_object)
1604                                 VM_OBJECT_RUNLOCK(lobject);
1605                         break;
1606                 }
1607                 if (vm_page_all_valid(m) &&
1608                     (m->flags & PG_FICTITIOUS) == 0)
1609                         pmap_enter_quick(pmap, addr, m, entry->protection);
1610                 if (!obj_locked || lobject != entry->object.vm_object)
1611                         VM_OBJECT_RUNLOCK(lobject);
1612         }
1613 }
1614
1615 /*
1616  * Hold each of the physical pages that are mapped by the specified range of
1617  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1618  * and allow the specified types of access, "prot".  If all of the implied
1619  * pages are successfully held, then the number of held pages is returned
1620  * together with pointers to those pages in the array "ma".  However, if any
1621  * of the pages cannot be held, -1 is returned.
1622  */
1623 int
1624 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1625     vm_prot_t prot, vm_page_t *ma, int max_count)
1626 {
1627         vm_offset_t end, va;
1628         vm_page_t *mp;
1629         int count;
1630         boolean_t pmap_failed;
1631
1632         if (len == 0)
1633                 return (0);
1634         end = round_page(addr + len);
1635         addr = trunc_page(addr);
1636
1637         /*
1638          * Check for illegal addresses.
1639          */
1640         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1641                 return (-1);
1642
1643         if (atop(end - addr) > max_count)
1644                 panic("vm_fault_quick_hold_pages: count > max_count");
1645         count = atop(end - addr);
1646
1647         /*
1648          * Most likely, the physical pages are resident in the pmap, so it is
1649          * faster to try pmap_extract_and_hold() first.
1650          */
1651         pmap_failed = FALSE;
1652         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1653                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1654                 if (*mp == NULL)
1655                         pmap_failed = TRUE;
1656                 else if ((prot & VM_PROT_WRITE) != 0 &&
1657                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1658                         /*
1659                          * Explicitly dirty the physical page.  Otherwise, the
1660                          * caller's changes may go unnoticed because they are
1661                          * performed through an unmanaged mapping or by a DMA
1662                          * operation.
1663                          *
1664                          * The object lock is not held here.
1665                          * See vm_page_clear_dirty_mask().
1666                          */
1667                         vm_page_dirty(*mp);
1668                 }
1669         }
1670         if (pmap_failed) {
1671                 /*
1672                  * One or more pages could not be held by the pmap.  Either no
1673                  * page was mapped at the specified virtual address or that
1674                  * mapping had insufficient permissions.  Attempt to fault in
1675                  * and hold these pages.
1676                  *
1677                  * If vm_fault_disable_pagefaults() was called,
1678                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1679                  * acquire MD VM locks, which means we must not call
1680                  * vm_fault().  Some (out of tree) callers mark
1681                  * too wide a code area with vm_fault_disable_pagefaults()
1682                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1683                  * the proper behaviour explicitly.
1684                  */
1685                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1686                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1687                         goto error;
1688                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1689                         if (*mp == NULL && vm_fault(map, va, prot,
1690                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1691                                 goto error;
1692         }
1693         return (count);
1694 error:
1695         for (mp = ma; mp < ma + count; mp++)
1696                 if (*mp != NULL)
1697                         vm_page_unwire(*mp, PQ_INACTIVE);
1698         return (-1);
1699 }
1700
1701 /*
1702  *      Routine:
1703  *              vm_fault_copy_entry
1704  *      Function:
1705  *              Create new shadow object backing dst_entry with private copy of
1706  *              all underlying pages. When src_entry is equal to dst_entry,
1707  *              function implements COW for wired-down map entry. Otherwise,
1708  *              it forks wired entry into dst_map.
1709  *
1710  *      In/out conditions:
1711  *              The source and destination maps must be locked for write.
1712  *              The source map entry must be wired down (or be a sharing map
1713  *              entry corresponding to a main map entry that is wired down).
1714  */
1715 void
1716 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1717     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1718     vm_ooffset_t *fork_charge)
1719 {
1720         vm_object_t backing_object, dst_object, object, src_object;
1721         vm_pindex_t dst_pindex, pindex, src_pindex;
1722         vm_prot_t access, prot;
1723         vm_offset_t vaddr;
1724         vm_page_t dst_m;
1725         vm_page_t src_m;
1726         boolean_t upgrade;
1727
1728 #ifdef  lint
1729         src_map++;
1730 #endif  /* lint */
1731
1732         upgrade = src_entry == dst_entry;
1733         access = prot = dst_entry->protection;
1734
1735         src_object = src_entry->object.vm_object;
1736         src_pindex = OFF_TO_IDX(src_entry->offset);
1737
1738         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1739                 dst_object = src_object;
1740                 vm_object_reference(dst_object);
1741         } else {
1742                 /*
1743                  * Create the top-level object for the destination entry.
1744                  * Doesn't actually shadow anything - we copy the pages
1745                  * directly.
1746                  */
1747                 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1748                     dst_entry->start), NULL, NULL, 0);
1749 #if VM_NRESERVLEVEL > 0
1750                 dst_object->flags |= OBJ_COLORED;
1751                 dst_object->pg_color = atop(dst_entry->start);
1752 #endif
1753                 dst_object->domain = src_object->domain;
1754                 dst_object->charge = dst_entry->end - dst_entry->start;
1755         }
1756
1757         VM_OBJECT_WLOCK(dst_object);
1758         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1759             ("vm_fault_copy_entry: vm_object not NULL"));
1760         if (src_object != dst_object) {
1761                 dst_entry->object.vm_object = dst_object;
1762                 dst_entry->offset = 0;
1763                 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1764         }
1765         if (fork_charge != NULL) {
1766                 KASSERT(dst_entry->cred == NULL,
1767                     ("vm_fault_copy_entry: leaked swp charge"));
1768                 dst_object->cred = curthread->td_ucred;
1769                 crhold(dst_object->cred);
1770                 *fork_charge += dst_object->charge;
1771         } else if ((dst_object->type == OBJT_DEFAULT ||
1772             dst_object->type == OBJT_SWAP) &&
1773             dst_object->cred == NULL) {
1774                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1775                     dst_entry));
1776                 dst_object->cred = dst_entry->cred;
1777                 dst_entry->cred = NULL;
1778         }
1779
1780         /*
1781          * If not an upgrade, then enter the mappings in the pmap as
1782          * read and/or execute accesses.  Otherwise, enter them as
1783          * write accesses.
1784          *
1785          * A writeable large page mapping is only created if all of
1786          * the constituent small page mappings are modified. Marking
1787          * PTEs as modified on inception allows promotion to happen
1788          * without taking potentially large number of soft faults.
1789          */
1790         if (!upgrade)
1791                 access &= ~VM_PROT_WRITE;
1792
1793         /*
1794          * Loop through all of the virtual pages within the entry's
1795          * range, copying each page from the source object to the
1796          * destination object.  Since the source is wired, those pages
1797          * must exist.  In contrast, the destination is pageable.
1798          * Since the destination object doesn't share any backing storage
1799          * with the source object, all of its pages must be dirtied,
1800          * regardless of whether they can be written.
1801          */
1802         for (vaddr = dst_entry->start, dst_pindex = 0;
1803             vaddr < dst_entry->end;
1804             vaddr += PAGE_SIZE, dst_pindex++) {
1805 again:
1806                 /*
1807                  * Find the page in the source object, and copy it in.
1808                  * Because the source is wired down, the page will be
1809                  * in memory.
1810                  */
1811                 if (src_object != dst_object)
1812                         VM_OBJECT_RLOCK(src_object);
1813                 object = src_object;
1814                 pindex = src_pindex + dst_pindex;
1815                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1816                     (backing_object = object->backing_object) != NULL) {
1817                         /*
1818                          * Unless the source mapping is read-only or
1819                          * it is presently being upgraded from
1820                          * read-only, the first object in the shadow
1821                          * chain should provide all of the pages.  In
1822                          * other words, this loop body should never be
1823                          * executed when the source mapping is already
1824                          * read/write.
1825                          */
1826                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1827                             upgrade,
1828                             ("vm_fault_copy_entry: main object missing page"));
1829
1830                         VM_OBJECT_RLOCK(backing_object);
1831                         pindex += OFF_TO_IDX(object->backing_object_offset);
1832                         if (object != dst_object)
1833                                 VM_OBJECT_RUNLOCK(object);
1834                         object = backing_object;
1835                 }
1836                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1837
1838                 if (object != dst_object) {
1839                         /*
1840                          * Allocate a page in the destination object.
1841                          */
1842                         dst_m = vm_page_alloc(dst_object, (src_object ==
1843                             dst_object ? src_pindex : 0) + dst_pindex,
1844                             VM_ALLOC_NORMAL);
1845                         if (dst_m == NULL) {
1846                                 VM_OBJECT_WUNLOCK(dst_object);
1847                                 VM_OBJECT_RUNLOCK(object);
1848                                 vm_wait(dst_object);
1849                                 VM_OBJECT_WLOCK(dst_object);
1850                                 goto again;
1851                         }
1852                         pmap_copy_page(src_m, dst_m);
1853                         VM_OBJECT_RUNLOCK(object);
1854                         dst_m->dirty = dst_m->valid = src_m->valid;
1855                 } else {
1856                         dst_m = src_m;
1857                         if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1858                                 goto again;
1859                         if (dst_m->pindex >= dst_object->size) {
1860                                 /*
1861                                  * We are upgrading.  Index can occur
1862                                  * out of bounds if the object type is
1863                                  * vnode and the file was truncated.
1864                                  */
1865                                 vm_page_xunbusy(dst_m);
1866                                 break;
1867                         }
1868                 }
1869                 VM_OBJECT_WUNLOCK(dst_object);
1870
1871                 /*
1872                  * Enter it in the pmap. If a wired, copy-on-write
1873                  * mapping is being replaced by a write-enabled
1874                  * mapping, then wire that new mapping.
1875                  *
1876                  * The page can be invalid if the user called
1877                  * msync(MS_INVALIDATE) or truncated the backing vnode
1878                  * or shared memory object.  In this case, do not
1879                  * insert it into pmap, but still do the copy so that
1880                  * all copies of the wired map entry have similar
1881                  * backing pages.
1882                  */
1883                 if (vm_page_all_valid(dst_m)) {
1884                         pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1885                             access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1886                 }
1887
1888                 /*
1889                  * Mark it no longer busy, and put it on the active list.
1890                  */
1891                 VM_OBJECT_WLOCK(dst_object);
1892
1893                 if (upgrade) {
1894                         if (src_m != dst_m) {
1895                                 vm_page_unwire(src_m, PQ_INACTIVE);
1896                                 vm_page_wire(dst_m);
1897                         } else {
1898                                 KASSERT(vm_page_wired(dst_m),
1899                                     ("dst_m %p is not wired", dst_m));
1900                         }
1901                 } else {
1902                         vm_page_activate(dst_m);
1903                 }
1904                 vm_page_xunbusy(dst_m);
1905         }
1906         VM_OBJECT_WUNLOCK(dst_object);
1907         if (upgrade) {
1908                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1909                 vm_object_deallocate(src_object);
1910         }
1911 }
1912
1913 /*
1914  * Block entry into the machine-independent layer's page fault handler by
1915  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1916  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1917  * spurious page faults.
1918  */
1919 int
1920 vm_fault_disable_pagefaults(void)
1921 {
1922
1923         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1924 }
1925
1926 void
1927 vm_fault_enable_pagefaults(int save)
1928 {
1929
1930         curthread_pflags_restore(save);
1931 }