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         /* Fault parameters. */
 124         vm_offset_t     vaddr;
 125         vm_page_t       *m_hold;
 126         vm_prot_t       fault_type;
 127         vm_prot_t       prot;
 128         int             fault_flags;
 129         boolean_t       wired;
 130
 131         /* Page reference for cow. */
 132         vm_page_t m_cow;
 133
 134         /* Current object. */
 135         vm_object_t     object;
 136         vm_pindex_t     pindex;
 137         vm_page_t       m;
 138
 139         /* Top-level map object. */
 140         vm_object_t     first_object;
 141         vm_pindex_t     first_pindex;
 142         vm_page_t       first_m;
 143
 144         /* Map state. */
 145         vm_map_t        map;
 146         vm_map_entry_t  entry;
 147         int             map_generation;
 148         bool            lookup_still_valid;
 149
 150         /* Vnode if locked. */
 151         struct vnode    *vp;
 152 };
 153
 154 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 155             int ahead);
 156 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 157             int backward, int forward, bool obj_locked);
 158
 159 static int vm_pfault_oom_attempts = 3;
 160 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
 161     &vm_pfault_oom_attempts, 0,
 162     "Number of page allocation attempts in page fault handler before it "
 163     "triggers OOM handling");
 164
 165 static int vm_pfault_oom_wait = 10;
 166 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
 167     &vm_pfault_oom_wait, 0,
 168     "Number of seconds to wait for free pages before retrying "
 169     "the page fault handler");
 170
 171 static inline void
 172 fault_page_release(vm_page_t *mp)
 173 {
 174         vm_page_t m;
 175
 176         m = *mp;
 177         if (m != NULL) {
 178                 /*
 179                  * We are likely to loop around again and attempt to busy
 180                  * this page.  Deactivating it leaves it available for
 181                  * pageout while optimizing fault restarts.
 182                  */
 183                 vm_page_deactivate(m);
 184                 vm_page_xunbusy(m);
 185                 *mp = NULL;
 186         }
 187 }
 188
 189 static inline void
 190 fault_page_free(vm_page_t *mp)
 191 {
 192         vm_page_t m;
 193
 194         m = *mp;
 195         if (m != NULL) {
 196                 VM_OBJECT_ASSERT_WLOCKED(m->object);
 197                 if (!vm_page_wired(m))
 198                         vm_page_free(m);
 199                 else
 200                         vm_page_xunbusy(m);
 201                 *mp = NULL;
 202         }
 203 }
 204
 205 static inline void
 206 unlock_map(struct faultstate *fs)
 207 {
 208
 209         if (fs->lookup_still_valid) {
 210                 vm_map_lookup_done(fs->map, fs->entry);
 211                 fs->lookup_still_valid = false;
 212         }
 213 }
 214
 215 static void
 216 unlock_vp(struct faultstate *fs)
 217 {
 218
 219         if (fs->vp != NULL) {
 220                 vput(fs->vp);
 221                 fs->vp = NULL;
 222         }
 223 }
 224
 225 static void
 226 fault_deallocate(struct faultstate *fs)
 227 {
 228
 229         fault_page_release(&fs->m_cow);
 230         fault_page_release(&fs->m);
 231         vm_object_pip_wakeup(fs->object);
 232         if (fs->object != fs->first_object) {
 233                 VM_OBJECT_WLOCK(fs->first_object);
 234                 fault_page_free(&fs->first_m);
 235                 VM_OBJECT_WUNLOCK(fs->first_object);
 236                 vm_object_pip_wakeup(fs->first_object);
 237         }
 238         vm_object_deallocate(fs->first_object);
 239         unlock_map(fs);
 240         unlock_vp(fs);
 241 }
 242
 243 static void
 244 unlock_and_deallocate(struct faultstate *fs)
 245 {
 246
 247         VM_OBJECT_WUNLOCK(fs->object);
 248         fault_deallocate(fs);
 249 }
 250
 251 static void
 252 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
 253 {
 254         bool need_dirty;
 255
 256         if (((fs->prot & VM_PROT_WRITE) == 0 &&
 257             (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
 258             (m->oflags & VPO_UNMANAGED) != 0)
 259                 return;
 260
 261         VM_PAGE_OBJECT_BUSY_ASSERT(m);
 262
 263         need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
 264             (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
 265             (fs->fault_flags & VM_FAULT_DIRTY) != 0;
 266
 267         vm_object_set_writeable_dirty(m->object);
 268
 269         /*
 270          * If the fault is a write, we know that this page is being
 271          * written NOW so dirty it explicitly to save on
 272          * pmap_is_modified() calls later.
 273          *
 274          * Also, since the page is now dirty, we can possibly tell
 275          * the pager to release any swap backing the page.
 276          */
 277         if (need_dirty && vm_page_set_dirty(m) == 0) {
 278                 /*
 279                  * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
 280                  * if the page is already dirty to prevent data written with
 281                  * the expectation of being synced from not being synced.
 282                  * Likewise if this entry does not request NOSYNC then make
 283                  * sure the page isn't marked NOSYNC.  Applications sharing
 284                  * data should use the same flags to avoid ping ponging.
 285                  */
 286                 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
 287                         vm_page_aflag_set(m, PGA_NOSYNC);
 288                 else
 289                         vm_page_aflag_clear(m, PGA_NOSYNC);
 290         }
 291
 292 }
 293
 294 /*
 295  * Unlocks fs.first_object and fs.map on success.
 296  */
 297 static int
 298 vm_fault_soft_fast(struct faultstate *fs)
 299 {
 300         vm_page_t m, m_map;
 301 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 302     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 303     VM_NRESERVLEVEL > 0
 304         vm_page_t m_super;
 305         int flags;
 306 #endif
 307         int psind, rv;
 308         vm_offset_t vaddr;
 309
 310         MPASS(fs->vp == NULL);
 311         vaddr = fs->vaddr;
 312         vm_object_busy(fs->first_object);
 313         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 314         /* A busy page can be mapped for read|execute access. */
 315         if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
 316             vm_page_busied(m)) || !vm_page_all_valid(m)) {
 317                 rv = KERN_FAILURE;
 318                 goto out;
 319         }
 320         m_map = m;
 321         psind = 0;
 322 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 323     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 324     VM_NRESERVLEVEL > 0
 325         if ((m->flags & PG_FICTITIOUS) == 0 &&
 326             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 327             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 328             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 329             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 330             (pagesizes[m_super->psind] - 1)) && !fs->wired &&
 331             pmap_ps_enabled(fs->map->pmap)) {
 332                 flags = PS_ALL_VALID;
 333                 if ((fs->prot & VM_PROT_WRITE) != 0) {
 334                         /*
 335                          * Create a superpage mapping allowing write access
 336                          * only if none of the constituent pages are busy and
 337                          * all of them are already dirty (except possibly for
 338                          * the page that was faulted on).
 339                          */
 340                         flags |= PS_NONE_BUSY;
 341                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 342                                 flags |= PS_ALL_DIRTY;
 343                 }
 344                 if (vm_page_ps_test(m_super, flags, m)) {
 345                         m_map = m_super;
 346                         psind = m_super->psind;
 347                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 348                         /* Preset the modified bit for dirty superpages. */
 349                         if ((flags & PS_ALL_DIRTY) != 0)
 350                                 fs->fault_type |= VM_PROT_WRITE;
 351                 }
 352         }
 353 #endif
 354         rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
 355             PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
 356         if (rv != KERN_SUCCESS)
 357                 goto out;
 358         if (fs->m_hold != NULL) {
 359                 (*fs->m_hold) = m;
 360                 vm_page_wire(m);
 361         }
 362         if (psind == 0 && !fs->wired)
 363                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 364         VM_OBJECT_RUNLOCK(fs->first_object);
 365         vm_fault_dirty(fs, m);
 366         vm_map_lookup_done(fs->map, fs->entry);
 367         curthread->td_ru.ru_minflt++;
 368
 369 out:
 370         vm_object_unbusy(fs->first_object);
 371         return (rv);
 372 }
 373
 374 static void
 375 vm_fault_restore_map_lock(struct faultstate *fs)
 376 {
 377
 378         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 379         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 380
 381         if (!vm_map_trylock_read(fs->map)) {
 382                 VM_OBJECT_WUNLOCK(fs->first_object);
 383                 vm_map_lock_read(fs->map);
 384                 VM_OBJECT_WLOCK(fs->first_object);
 385         }
 386         fs->lookup_still_valid = true;
 387 }
 388
 389 static void
 390 vm_fault_populate_check_page(vm_page_t m)
 391 {
 392
 393         /*
 394          * Check each page to ensure that the pager is obeying the
 395          * interface: the page must be installed in the object, fully
 396          * valid, and exclusively busied.
 397          */
 398         MPASS(m != NULL);
 399         MPASS(vm_page_all_valid(m));
 400         MPASS(vm_page_xbusied(m));
 401 }
 402
 403 static void
 404 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 405     vm_pindex_t last)
 406 {
 407         vm_page_t m;
 408         vm_pindex_t pidx;
 409
 410         VM_OBJECT_ASSERT_WLOCKED(object);
 411         MPASS(first <= last);
 412         for (pidx = first, m = vm_page_lookup(object, pidx);
 413             pidx <= last; pidx++, m = vm_page_next(m)) {
 414                 vm_fault_populate_check_page(m);
 415                 vm_page_deactivate(m);
 416                 vm_page_xunbusy(m);
 417         }
 418 }
 419
 420 static int
 421 vm_fault_populate(struct faultstate *fs)
 422 {
 423         vm_offset_t vaddr;
 424         vm_page_t m;
 425         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 426         int i, npages, psind, rv;
 427
 428         MPASS(fs->object == fs->first_object);
 429         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 430         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 431         MPASS(fs->first_object->backing_object == NULL);
 432         MPASS(fs->lookup_still_valid);
 433
 434         pager_first = OFF_TO_IDX(fs->entry->offset);
 435         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 436         unlock_map(fs);
 437         unlock_vp(fs);
 438
 439         /*
 440          * Call the pager (driver) populate() method.
 441          *
 442          * There is no guarantee that the method will be called again
 443          * if the current fault is for read, and a future fault is
 444          * for write.  Report the entry's maximum allowed protection
 445          * to the driver.
 446          */
 447         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 448             fs->fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 449
 450         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 451         if (rv == VM_PAGER_BAD) {
 452                 /*
 453                  * VM_PAGER_BAD is the backdoor for a pager to request
 454                  * normal fault handling.
 455                  */
 456                 vm_fault_restore_map_lock(fs);
 457                 if (fs->map->timestamp != fs->map_generation)
 458                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 459                 return (KERN_NOT_RECEIVER);
 460         }
 461         if (rv != VM_PAGER_OK)
 462                 return (KERN_FAILURE); /* AKA SIGSEGV */
 463
 464         /* Ensure that the driver is obeying the interface. */
 465         MPASS(pager_first <= pager_last);
 466         MPASS(fs->first_pindex <= pager_last);
 467         MPASS(fs->first_pindex >= pager_first);
 468         MPASS(pager_last < fs->first_object->size);
 469
 470         vm_fault_restore_map_lock(fs);
 471         if (fs->map->timestamp != fs->map_generation) {
 472                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 473                     pager_last);
 474                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 475         }
 476
 477         /*
 478          * The map is unchanged after our last unlock.  Process the fault.
 479          *
 480          * The range [pager_first, pager_last] that is given to the
 481          * pager is only a hint.  The pager may populate any range
 482          * within the object that includes the requested page index.
 483          * In case the pager expanded the range, clip it to fit into
 484          * the map entry.
 485          */
 486         map_first = OFF_TO_IDX(fs->entry->offset);
 487         if (map_first > pager_first) {
 488                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 489                     map_first - 1);
 490                 pager_first = map_first;
 491         }
 492         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 493         if (map_last < pager_last) {
 494                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 495                     pager_last);
 496                 pager_last = map_last;
 497         }
 498         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 499             pidx <= pager_last;
 500             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 501                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 502 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 503     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
 504                 psind = m->psind;
 505                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 506                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 507                     !pmap_ps_enabled(fs->map->pmap) || fs->wired))
 508                         psind = 0;
 509 #else
 510                 psind = 0;
 511 #endif
 512                 npages = atop(pagesizes[psind]);
 513                 for (i = 0; i < npages; i++) {
 514                         vm_fault_populate_check_page(&m[i]);
 515                         vm_fault_dirty(fs, &m[i]);
 516                 }
 517                 VM_OBJECT_WUNLOCK(fs->first_object);
 518                 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
 519                     (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
 520 #if defined(__amd64__)
 521                 if (psind > 0 && rv == KERN_FAILURE) {
 522                         for (i = 0; i < npages; i++) {
 523                                 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
 524                                     &m[i], fs->prot, fs->fault_type |
 525                                     (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
 526                                 MPASS(rv == KERN_SUCCESS);
 527                         }
 528                 }
 529 #else
 530                 MPASS(rv == KERN_SUCCESS);
 531 #endif
 532                 VM_OBJECT_WLOCK(fs->first_object);
 533                 for (i = 0; i < npages; i++) {
 534                         if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
 535                                 vm_page_wire(&m[i]);
 536                         else
 537                                 vm_page_activate(&m[i]);
 538                         if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
 539                                 (*fs->m_hold) = &m[i];
 540                                 vm_page_wire(&m[i]);
 541                         }
 542                         vm_page_xunbusy(&m[i]);
 543                 }
 544         }
 545         curthread->td_ru.ru_majflt++;
 546         return (KERN_SUCCESS);
 547 }
 548
 549 static int prot_fault_translation;
 550 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
 551     &prot_fault_translation, 0,
 552     "Control signal to deliver on protection fault");
 553
 554 /* compat definition to keep common code for signal translation */
 555 #define UCODE_PAGEFLT   12
 556 #ifdef T_PAGEFLT
 557 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
 558 #endif
 559
 560 /*
 561  *      vm_fault_trap:
 562  *
 563  *      Handle a page fault occurring at the given address,
 564  *      requiring the given permissions, in the map specified.
 565  *      If successful, the page is inserted into the
 566  *      associated physical map.
 567  *
 568  *      NOTE: the given address should be truncated to the
 569  *      proper page address.
 570  *
 571  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 572  *      a standard error specifying why the fault is fatal is returned.
 573  *
 574  *      The map in question must be referenced, and remains so.
 575  *      Caller may hold no locks.
 576  */
 577 int
 578 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 579     int fault_flags, int *signo, int *ucode)
 580 {
 581         int result;
 582
 583         MPASS(signo == NULL || ucode != NULL);
 584 #ifdef KTRACE
 585         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
 586                 ktrfault(vaddr, fault_type);
 587 #endif
 588         result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
 589             NULL);
 590         KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
 591             result == KERN_INVALID_ADDRESS ||
 592             result == KERN_RESOURCE_SHORTAGE ||
 593             result == KERN_PROTECTION_FAILURE ||
 594             result == KERN_OUT_OF_BOUNDS,
 595             ("Unexpected Mach error %d from vm_fault()", result));
 596 #ifdef KTRACE
 597         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
 598                 ktrfaultend(result);
 599 #endif
 600         if (result != KERN_SUCCESS && signo != NULL) {
 601                 switch (result) {
 602                 case KERN_FAILURE:
 603                 case KERN_INVALID_ADDRESS:
 604                         *signo = SIGSEGV;
 605                         *ucode = SEGV_MAPERR;
 606                         break;
 607                 case KERN_RESOURCE_SHORTAGE:
 608                         *signo = SIGBUS;
 609                         *ucode = BUS_OOMERR;
 610                         break;
 611                 case KERN_OUT_OF_BOUNDS:
 612                         *signo = SIGBUS;
 613                         *ucode = BUS_OBJERR;
 614                         break;
 615                 case KERN_PROTECTION_FAILURE:
 616                         if (prot_fault_translation == 0) {
 617                                 /*
 618                                  * Autodetect.  This check also covers
 619                                  * the images without the ABI-tag ELF
 620                                  * note.
 621                                  */
 622                                 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
 623                                     curproc->p_osrel >= P_OSREL_SIGSEGV) {
 624                                         *signo = SIGSEGV;
 625                                         *ucode = SEGV_ACCERR;
 626                                 } else {
 627                                         *signo = SIGBUS;
 628                                         *ucode = UCODE_PAGEFLT;
 629                                 }
 630                         } else if (prot_fault_translation == 1) {
 631                                 /* Always compat mode. */
 632                                 *signo = SIGBUS;
 633                                 *ucode = UCODE_PAGEFLT;
 634                         } else {
 635                                 /* Always SIGSEGV mode. */
 636                                 *signo = SIGSEGV;
 637                                 *ucode = SEGV_ACCERR;
 638                         }
 639                         break;
 640                 default:
 641                         KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
 642                             result));
 643                         break;
 644                 }
 645         }
 646         return (result);
 647 }
 648
 649 static int
 650 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
 651 {
 652         struct vnode *vp;
 653         int error, locked;
 654
 655         if (fs->object->type != OBJT_VNODE)
 656                 return (KERN_SUCCESS);
 657         vp = fs->object->handle;
 658         if (vp == fs->vp) {
 659                 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
 660                 return (KERN_SUCCESS);
 661         }
 662
 663         /*
 664          * Perform an unlock in case the desired vnode changed while
 665          * the map was unlocked during a retry.
 666          */
 667         unlock_vp(fs);
 668
 669         locked = VOP_ISLOCKED(vp);
 670         if (locked != LK_EXCLUSIVE)
 671                 locked = LK_SHARED;
 672
 673         /*
 674          * We must not sleep acquiring the vnode lock while we have
 675          * the page exclusive busied or the object's
 676          * paging-in-progress count incremented.  Otherwise, we could
 677          * deadlock.
 678          */
 679         error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
 680         if (error == 0) {
 681                 fs->vp = vp;
 682                 return (KERN_SUCCESS);
 683         }
 684
 685         vhold(vp);
 686         if (objlocked)
 687                 unlock_and_deallocate(fs);
 688         else
 689                 fault_deallocate(fs);
 690         error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
 691         vdrop(vp);
 692         fs->vp = vp;
 693         KASSERT(error == 0, ("vm_fault: vget failed %d", error));
 694         return (KERN_RESOURCE_SHORTAGE);
 695 }
 696
 697 /*
 698  * Calculate the desired readahead.  Handle drop-behind.
 699  *
 700  * Returns the number of readahead blocks to pass to the pager.
 701  */
 702 static int
 703 vm_fault_readahead(struct faultstate *fs)
 704 {
 705         int era, nera;
 706         u_char behavior;
 707
 708         KASSERT(fs->lookup_still_valid, ("map unlocked"));
 709         era = fs->entry->read_ahead;
 710         behavior = vm_map_entry_behavior(fs->entry);
 711         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 712                 nera = 0;
 713         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 714                 nera = VM_FAULT_READ_AHEAD_MAX;
 715                 if (fs->vaddr == fs->entry->next_read)
 716                         vm_fault_dontneed(fs, fs->vaddr, nera);
 717         } else if (fs->vaddr == fs->entry->next_read) {
 718                 /*
 719                  * This is a sequential fault.  Arithmetically
 720                  * increase the requested number of pages in
 721                  * the read-ahead window.  The requested
 722                  * number of pages is "# of sequential faults
 723                  * x (read ahead min + 1) + read ahead min"
 724                  */
 725                 nera = VM_FAULT_READ_AHEAD_MIN;
 726                 if (era > 0) {
 727                         nera += era + 1;
 728                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 729                                 nera = VM_FAULT_READ_AHEAD_MAX;
 730                 }
 731                 if (era == VM_FAULT_READ_AHEAD_MAX)
 732                         vm_fault_dontneed(fs, fs->vaddr, nera);
 733         } else {
 734                 /*
 735                  * This is a non-sequential fault.
 736                  */
 737                 nera = 0;
 738         }
 739         if (era != nera) {
 740                 /*
 741                  * A read lock on the map suffices to update
 742                  * the read ahead count safely.
 743                  */
 744                 fs->entry->read_ahead = nera;
 745         }
 746
 747         return (nera);
 748 }
 749
 750 /*
 751  * Wait/Retry if the page is busy.  We have to do this if the page is
 752  * either exclusive or shared busy because the vm_pager may be using
 753  * read busy for pageouts (and even pageins if it is the vnode pager),
 754  * and we could end up trying to pagein and pageout the same page
 755  * simultaneously.
 756  *
 757  * We can theoretically allow the busy case on a read fault if the page
 758  * is marked valid, but since such pages are typically already pmap'd,
 759  * putting that special case in might be more effort then it is worth.
 760  * We cannot under any circumstances mess around with a shared busied
 761  * page except, perhaps, to pmap it.
 762  */
 763 static void
 764 vm_fault_busy_sleep(struct faultstate *fs)
 765 {
 766         /*
 767          * Reference the page before unlocking and
 768          * sleeping so that the page daemon is less
 769          * likely to reclaim it.
 770          */
 771         vm_page_aflag_set(fs->m, PGA_REFERENCED);
 772         if (fs->object != fs->first_object) {
 773                 fault_page_release(&fs->first_m);
 774                 vm_object_pip_wakeup(fs->first_object);
 775         }
 776         vm_object_pip_wakeup(fs->object);
 777         unlock_map(fs);
 778         if (fs->m == vm_page_lookup(fs->object, fs->pindex))
 779                 vm_page_busy_sleep(fs->m, "vmpfw", false);
 780         else
 781                 VM_OBJECT_WUNLOCK(fs->object);
 782         VM_CNT_INC(v_intrans);
 783         vm_object_deallocate(fs->first_object);
 784 }
 785
 786 int
 787 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 788     int fault_flags, vm_page_t *m_hold)
 789 {
 790         struct faultstate fs;
 791         struct domainset *dset;
 792         vm_object_t next_object, retry_object;
 793         vm_offset_t e_end, e_start;
 794         vm_pindex_t retry_pindex;
 795         vm_prot_t retry_prot;
 796         int ahead, alloc_req, behind, cluster_offset, faultcount;
 797         int nera, oom, result, rv;
 798         u_char behavior;
 799         bool dead, hardfault, is_first_object_locked;
 800
 801         VM_CNT_INC(v_vm_faults);
 802
 803         if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
 804                 return (KERN_PROTECTION_FAILURE);
 805
 806         fs.vp = NULL;
 807         fs.vaddr = vaddr;
 808         fs.m_hold = m_hold;
 809         fs.fault_flags = fault_flags;
 810         faultcount = 0;
 811         nera = -1;
 812         hardfault = false;
 813
 814 RetryFault:
 815         oom = 0;
 816 RetryFault_oom:
 817         fs.fault_type = fault_type;
 818
 819         /*
 820          * Find the backing store object and offset into it to begin the
 821          * search.
 822          */
 823         fs.map = map;
 824         result = vm_map_lookup(&fs.map, fs.vaddr, fs.fault_type |
 825             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 826             &fs.first_pindex, &fs.prot, &fs.wired);
 827         if (result != KERN_SUCCESS) {
 828                 unlock_vp(&fs);
 829                 return (result);
 830         }
 831
 832         fs.map_generation = fs.map->timestamp;
 833
 834         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 835                 panic("%s: fault on nofault entry, addr: %#lx",
 836                     __func__, (u_long)fs.vaddr);
 837         }
 838
 839         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 840             fs.entry->wiring_thread != curthread) {
 841                 vm_map_unlock_read(fs.map);
 842                 vm_map_lock(fs.map);
 843                 if (vm_map_lookup_entry(fs.map, fs.vaddr, &fs.entry) &&
 844                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 845                         unlock_vp(&fs);
 846                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 847                         vm_map_unlock_and_wait(fs.map, 0);
 848                 } else
 849                         vm_map_unlock(fs.map);
 850                 goto RetryFault;
 851         }
 852
 853         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 854
 855         if (fs.wired)
 856                 fs.fault_type = fs.prot | (fs.fault_type & VM_PROT_COPY);
 857         else
 858                 KASSERT((fs.fault_flags & VM_FAULT_WIRE) == 0,
 859                     ("!fs.wired && VM_FAULT_WIRE"));
 860
 861         /*
 862          * Try to avoid lock contention on the top-level object through
 863          * special-case handling of some types of page faults, specifically,
 864          * those that are mapping an existing page from the top-level object.
 865          * Under this condition, a read lock on the object suffices, allowing
 866          * multiple page faults of a similar type to run in parallel.
 867          */
 868         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 869             (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
 870                 VM_OBJECT_RLOCK(fs.first_object);
 871                 rv = vm_fault_soft_fast(&fs);
 872                 if (rv == KERN_SUCCESS)
 873                         return (rv);
 874                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 875                         VM_OBJECT_RUNLOCK(fs.first_object);
 876                         VM_OBJECT_WLOCK(fs.first_object);
 877                 }
 878         } else {
 879                 VM_OBJECT_WLOCK(fs.first_object);
 880         }
 881
 882         /*
 883          * Make a reference to this object to prevent its disposal while we
 884          * are messing with it.  Once we have the reference, the map is free
 885          * to be diddled.  Since objects reference their shadows (and copies),
 886          * they will stay around as well.
 887          *
 888          * Bump the paging-in-progress count to prevent size changes (e.g.
 889          * truncation operations) during I/O.
 890          */
 891         vm_object_reference_locked(fs.first_object);
 892         vm_object_pip_add(fs.first_object, 1);
 893
 894         fs.lookup_still_valid = true;
 895
 896         fs.m_cow = fs.m = fs.first_m = NULL;
 897
 898         /*
 899          * Search for the page at object/offset.
 900          */
 901         fs.object = fs.first_object;
 902         fs.pindex = fs.first_pindex;
 903         while (TRUE) {
 904                 KASSERT(fs.m == NULL,
 905                     ("page still set %p at loop start", fs.m));
 906                 /*
 907                  * If the object is marked for imminent termination,
 908                  * we retry here, since the collapse pass has raced
 909                  * with us.  Otherwise, if we see terminally dead
 910                  * object, return fail.
 911                  */
 912                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 913                         dead = fs.object->type == OBJT_DEAD;
 914                         unlock_and_deallocate(&fs);
 915                         if (dead)
 916                                 return (KERN_PROTECTION_FAILURE);
 917                         pause("vmf_de", 1);
 918                         goto RetryFault;
 919                 }
 920
 921                 /*
 922                  * See if page is resident
 923                  */
 924                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 925                 if (fs.m != NULL) {
 926                         if (vm_page_tryxbusy(fs.m) == 0) {
 927                                 vm_fault_busy_sleep(&fs);
 928                                 goto RetryFault;
 929                         }
 930
 931                         /*
 932                          * The page is marked busy for other processes and the
 933                          * pagedaemon.  If it still isn't completely valid
 934                          * (readable), jump to readrest, else break-out ( we
 935                          * found the page ).
 936                          */
 937                         if (!vm_page_all_valid(fs.m))
 938                                 goto readrest;
 939                         VM_OBJECT_WUNLOCK(fs.object);
 940                         break; /* break to PAGE HAS BEEN FOUND. */
 941                 }
 942                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 943                 VM_OBJECT_ASSERT_WLOCKED(fs.object);
 944
 945                 /*
 946                  * Page is not resident.  If the pager might contain the page
 947                  * or this is the beginning of the search, allocate a new
 948                  * page.  (Default objects are zero-fill, so there is no real
 949                  * pager for them.)
 950                  */
 951                 if (fs.object->type != OBJT_DEFAULT ||
 952                     fs.object == fs.first_object) {
 953                         if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
 954                                 rv = vm_fault_lock_vnode(&fs, true);
 955                                 MPASS(rv == KERN_SUCCESS ||
 956                                     rv == KERN_RESOURCE_SHORTAGE);
 957                                 if (rv == KERN_RESOURCE_SHORTAGE)
 958                                         goto RetryFault;
 959                         }
 960                         if (fs.pindex >= fs.object->size) {
 961                                 unlock_and_deallocate(&fs);
 962                                 return (KERN_OUT_OF_BOUNDS);
 963                         }
 964
 965                         if (fs.object == fs.first_object &&
 966                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 967                             fs.first_object->shadow_count == 0) {
 968                                 rv = vm_fault_populate(&fs);
 969                                 switch (rv) {
 970                                 case KERN_SUCCESS:
 971                                 case KERN_FAILURE:
 972                                         unlock_and_deallocate(&fs);
 973                                         return (rv);
 974                                 case KERN_RESOURCE_SHORTAGE:
 975                                         unlock_and_deallocate(&fs);
 976                                         goto RetryFault;
 977                                 case KERN_NOT_RECEIVER:
 978                                         /*
 979                                          * Pager's populate() method
 980                                          * returned VM_PAGER_BAD.
 981                                          */
 982                                         break;
 983                                 default:
 984                                         panic("inconsistent return codes");
 985                                 }
 986                         }
 987
 988                         /*
 989                          * Allocate a new page for this object/offset pair.
 990                          *
 991                          * Unlocked read of the p_flag is harmless. At
 992                          * worst, the P_KILLED might be not observed
 993                          * there, and allocation can fail, causing
 994                          * restart and new reading of the p_flag.
 995                          */
 996                         dset = fs.object->domain.dr_policy;
 997                         if (dset == NULL)
 998                                 dset = curthread->td_domain.dr_policy;
 999                         if (!vm_page_count_severe_set(&dset->ds_mask) ||
1000                             P_KILLED(curproc)) {
1001 #if VM_NRESERVLEVEL > 0
1002                                 vm_object_color(fs.object, atop(vaddr) -
1003                                     fs.pindex);
1004 #endif
1005                                 alloc_req = P_KILLED(curproc) ?
1006                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
1007                                 if (fs.object->type != OBJT_VNODE &&
1008                                     fs.object->backing_object == NULL)
1009                                         alloc_req |= VM_ALLOC_ZERO;
1010                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
1011                                     alloc_req);
1012                         }
1013                         if (fs.m == NULL) {
1014                                 unlock_and_deallocate(&fs);
1015                                 if (vm_pfault_oom_attempts < 0 ||
1016                                     oom < vm_pfault_oom_attempts) {
1017                                         oom++;
1018                                         vm_waitpfault(dset,
1019                                             vm_pfault_oom_wait * hz);
1020                                         goto RetryFault_oom;
1021                                 }
1022                                 if (bootverbose)
1023                                         printf(
1024         "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1025                                             curproc->p_pid, curproc->p_comm);
1026                                 vm_pageout_oom(VM_OOM_MEM_PF);
1027                                 goto RetryFault;
1028                         }
1029                 }
1030
1031 readrest:
1032                 /*
1033                  * Default objects have no pager so no exclusive busy exists
1034                  * to protect this page in the chain.  Skip to the next
1035                  * object without dropping the lock to preserve atomicity of
1036                  * shadow faults.
1037                  */
1038                 if (fs.object->type == OBJT_DEFAULT)
1039                         goto next;
1040
1041                 /*
1042                  * At this point, we have either allocated a new page or found
1043                  * an existing page that is only partially valid.
1044                  *
1045                  * We hold a reference on the current object and the page is
1046                  * exclusive busied.  The exclusive busy prevents simultaneous
1047                  * faults and collapses while the object lock is dropped.
1048                  */
1049                 VM_OBJECT_WUNLOCK(fs.object);
1050
1051                 /*
1052                  * If the pager for the current object might have the page,
1053                  * then determine the number of additional pages to read and
1054                  * potentially reprioritize previously read pages for earlier
1055                  * reclamation.  These operations should only be performed
1056                  * once per page fault.  Even if the current pager doesn't
1057                  * have the page, the number of additional pages to read will
1058                  * apply to subsequent objects in the shadow chain.
1059                  */
1060                 if (nera == -1 && !P_KILLED(curproc)) {
1061                         nera = vm_fault_readahead(&fs);
1062                         /*
1063                          * Prepare for unlocking the map.  Save the map
1064                          * entry's start and end addresses, which are used to
1065                          * optimize the size of the pager operation below.
1066                          * Even if the map entry's addresses change after
1067                          * unlocking the map, using the saved addresses is
1068                          * safe.
1069                          */
1070                         e_start = fs.entry->start;
1071                         e_end = fs.entry->end;
1072                         behavior = vm_map_entry_behavior(fs.entry);
1073                 }
1074
1075                 /*
1076                  * Call the pager to retrieve the page if there is a chance
1077                  * that the pager has it, and potentially retrieve additional
1078                  * pages at the same time.
1079                  */
1080                 if (fs.object->type != OBJT_DEFAULT) {
1081                         /*
1082                          * Release the map lock before locking the vnode or
1083                          * sleeping in the pager.  (If the current object has
1084                          * a shadow, then an earlier iteration of this loop
1085                          * may have already unlocked the map.)
1086                          */
1087                         unlock_map(&fs);
1088
1089                         rv = vm_fault_lock_vnode(&fs, false);
1090                         MPASS(rv == KERN_SUCCESS ||
1091                             rv == KERN_RESOURCE_SHORTAGE);
1092                         if (rv == KERN_RESOURCE_SHORTAGE)
1093                                 goto RetryFault;
1094                         KASSERT(fs.vp == NULL || !fs.map->system_map,
1095                             ("vm_fault: vnode-backed object mapped by system map"));
1096
1097                         /*
1098                          * Page in the requested page and hint the pager,
1099                          * that it may bring up surrounding pages.
1100                          */
1101                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1102                             P_KILLED(curproc)) {
1103                                 behind = 0;
1104                                 ahead = 0;
1105                         } else {
1106                                 /* Is this a sequential fault? */
1107                                 if (nera > 0) {
1108                                         behind = 0;
1109                                         ahead = nera;
1110                                 } else {
1111                                         /*
1112                                          * Request a cluster of pages that is
1113                                          * aligned to a VM_FAULT_READ_DEFAULT
1114                                          * page offset boundary within the
1115                                          * object.  Alignment to a page offset
1116                                          * boundary is more likely to coincide
1117                                          * with the underlying file system
1118                                          * block than alignment to a virtual
1119                                          * address boundary.
1120                                          */
1121                                         cluster_offset = fs.pindex %
1122                                             VM_FAULT_READ_DEFAULT;
1123                                         behind = ulmin(cluster_offset,
1124                                             atop(vaddr - e_start));
1125                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
1126                                             cluster_offset;
1127                                 }
1128                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1129                         }
1130                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1131                             &behind, &ahead);
1132                         if (rv == VM_PAGER_OK) {
1133                                 faultcount = behind + 1 + ahead;
1134                                 hardfault = true;
1135                                 break; /* break to PAGE HAS BEEN FOUND. */
1136                         }
1137                         VM_OBJECT_WLOCK(fs.object);
1138                         if (rv == VM_PAGER_ERROR)
1139                                 printf("vm_fault: pager read error, pid %d (%s)\n",
1140                                     curproc->p_pid, curproc->p_comm);
1141
1142                         /*
1143                          * If an I/O error occurred or the requested page was
1144                          * outside the range of the pager, clean up and return
1145                          * an error.
1146                          */
1147                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1148                                 fault_page_free(&fs.m);
1149                                 unlock_and_deallocate(&fs);
1150                                 return (KERN_OUT_OF_BOUNDS);
1151                         }
1152
1153                 }
1154
1155 next:
1156                 /*
1157                  * The requested page does not exist at this object/
1158                  * offset.  Remove the invalid page from the object,
1159                  * waking up anyone waiting for it, and continue on to
1160                  * the next object.  However, if this is the top-level
1161                  * object, we must leave the busy page in place to
1162                  * prevent another process from rushing past us, and
1163                  * inserting the page in that object at the same time
1164                  * that we are.
1165                  */
1166                 if (fs.object == fs.first_object) {
1167                         fs.first_m = fs.m;
1168                         fs.m = NULL;
1169                 } else
1170                         fault_page_free(&fs.m);
1171
1172                 /*
1173                  * Move on to the next object.  Lock the next object before
1174                  * unlocking the current one.
1175                  */
1176                 VM_OBJECT_ASSERT_WLOCKED(fs.object);
1177                 next_object = fs.object->backing_object;
1178                 if (next_object == NULL) {
1179                         /*
1180                          * If there's no object left, fill the page in the top
1181                          * object with zeros.
1182                          */
1183                         VM_OBJECT_WUNLOCK(fs.object);
1184                         if (fs.object != fs.first_object) {
1185                                 vm_object_pip_wakeup(fs.object);
1186                                 fs.object = fs.first_object;
1187                                 fs.pindex = fs.first_pindex;
1188                         }
1189                         MPASS(fs.first_m != NULL);
1190                         MPASS(fs.m == NULL);
1191                         fs.m = fs.first_m;
1192                         fs.first_m = NULL;
1193
1194                         /*
1195                          * Zero the page if necessary and mark it valid.
1196                          */
1197                         if ((fs.m->flags & PG_ZERO) == 0) {
1198                                 pmap_zero_page(fs.m);
1199                         } else {
1200                                 VM_CNT_INC(v_ozfod);
1201                         }
1202                         VM_CNT_INC(v_zfod);
1203                         vm_page_valid(fs.m);
1204                         /* Don't try to prefault neighboring pages. */
1205                         faultcount = 1;
1206                         break;  /* break to PAGE HAS BEEN FOUND. */
1207                 } else {
1208                         MPASS(fs.first_m != NULL);
1209                         KASSERT(fs.object != next_object,
1210                             ("object loop %p", next_object));
1211                         VM_OBJECT_WLOCK(next_object);
1212                         vm_object_pip_add(next_object, 1);
1213                         if (fs.object != fs.first_object)
1214                                 vm_object_pip_wakeup(fs.object);
1215                         fs.pindex +=
1216                             OFF_TO_IDX(fs.object->backing_object_offset);
1217                         VM_OBJECT_WUNLOCK(fs.object);
1218                         fs.object = next_object;
1219                 }
1220         }
1221
1222         /*
1223          * PAGE HAS BEEN FOUND.  A valid page has been found and exclusively
1224          * busied.  The object lock must no longer be held.
1225          */
1226         vm_page_assert_xbusied(fs.m);
1227         VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1228
1229         /*
1230          * If the page is being written, but isn't already owned by the
1231          * top-level object, we have to copy it into a new page owned by the
1232          * top-level object.
1233          */
1234         if (fs.object != fs.first_object) {
1235                 /*
1236                  * We only really need to copy if we want to write it.
1237                  */
1238                 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1239                         /*
1240                          * This allows pages to be virtually copied from a
1241                          * backing_object into the first_object, where the
1242                          * backing object has no other refs to it, and cannot
1243                          * gain any more refs.  Instead of a bcopy, we just
1244                          * move the page from the backing object to the
1245                          * first object.  Note that we must mark the page
1246                          * dirty in the first object so that it will go out
1247                          * to swap when needed.
1248                          */
1249                         is_first_object_locked = false;
1250                         if (
1251                             /*
1252                              * Only one shadow object
1253                              */
1254                             fs.object->shadow_count == 1 &&
1255                             /*
1256                              * No COW refs, except us
1257                              */
1258                             fs.object->ref_count == 1 &&
1259                             /*
1260                              * No one else can look this object up
1261                              */
1262                             fs.object->handle == NULL &&
1263                             /*
1264                              * No other ways to look the object up
1265                              */
1266                             (fs.object->flags & OBJ_ANON) != 0 &&
1267                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1268                             /*
1269                              * We don't chase down the shadow chain
1270                              */
1271                             fs.object == fs.first_object->backing_object &&
1272                             VM_OBJECT_TRYWLOCK(fs.object)) {
1273
1274                                 /*
1275                                  * Remove but keep xbusy for replace.  fs.m is
1276                                  * moved into fs.first_object and left busy
1277                                  * while fs.first_m is conditionally freed.
1278                                  */
1279                                 vm_page_remove_xbusy(fs.m);
1280                                 vm_page_replace(fs.m, fs.first_object,
1281                                     fs.first_pindex, fs.first_m);
1282                                 vm_page_dirty(fs.m);
1283 #if VM_NRESERVLEVEL > 0
1284                                 /*
1285                                  * Rename the reservation.
1286                                  */
1287                                 vm_reserv_rename(fs.m, fs.first_object,
1288                                     fs.object, OFF_TO_IDX(
1289                                     fs.first_object->backing_object_offset));
1290 #endif
1291                                 VM_OBJECT_WUNLOCK(fs.object);
1292                                 VM_OBJECT_WUNLOCK(fs.first_object);
1293                                 fs.first_m = fs.m;
1294                                 fs.m = NULL;
1295                                 VM_CNT_INC(v_cow_optim);
1296                         } else {
1297                                 if (is_first_object_locked)
1298                                         VM_OBJECT_WUNLOCK(fs.first_object);
1299                                 /*
1300                                  * Oh, well, lets copy it.
1301                                  */
1302                                 pmap_copy_page(fs.m, fs.first_m);
1303                                 vm_page_valid(fs.first_m);
1304                                 if (fs.wired && (fs.fault_flags &
1305                                     VM_FAULT_WIRE) == 0) {
1306                                         vm_page_wire(fs.first_m);
1307                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1308                                 }
1309                                 /*
1310                                  * Save the cow page to be released after
1311                                  * pmap_enter is complete.
1312                                  */
1313                                 fs.m_cow = fs.m;
1314                                 fs.m = NULL;
1315                         }
1316                         /*
1317                          * fs.object != fs.first_object due to above
1318                          * conditional
1319                          */
1320                         vm_object_pip_wakeup(fs.object);
1321
1322                         /*
1323                          * We only try to prefault read-only mappings to the
1324                          * neighboring pages when this copy-on-write fault is
1325                          * a hard fault.  In other cases, trying to prefault
1326                          * is typically wasted effort.
1327                          */
1328                         if (faultcount == 0)
1329                                 faultcount = 1;
1330
1331                         /*
1332                          * Only use the new page below...
1333                          */
1334                         fs.object = fs.first_object;
1335                         fs.pindex = fs.first_pindex;
1336                         fs.m = fs.first_m;
1337                         VM_CNT_INC(v_cow_faults);
1338                         curthread->td_cow++;
1339                 } else {
1340                         fs.prot &= ~VM_PROT_WRITE;
1341                 }
1342         }
1343
1344         /*
1345          * We must verify that the maps have not changed since our last
1346          * lookup.
1347          */
1348         if (!fs.lookup_still_valid) {
1349                 if (!vm_map_trylock_read(fs.map)) {
1350                         fault_deallocate(&fs);
1351                         goto RetryFault;
1352                 }
1353                 fs.lookup_still_valid = true;
1354                 if (fs.map->timestamp != fs.map_generation) {
1355                         result = vm_map_lookup_locked(&fs.map, vaddr, fs.fault_type,
1356                             &fs.entry, &retry_object, &retry_pindex, &retry_prot,
1357                             &fs.wired);
1358
1359                         /*
1360                          * If we don't need the page any longer, put it on the inactive
1361                          * list (the easiest thing to do here).  If no one needs it,
1362                          * pageout will grab it eventually.
1363                          */
1364                         if (result != KERN_SUCCESS) {
1365                                 fault_deallocate(&fs);
1366
1367                                 /*
1368                                  * If retry of map lookup would have blocked then
1369                                  * retry fault from start.
1370                                  */
1371                                 if (result == KERN_FAILURE)
1372                                         goto RetryFault;
1373                                 return (result);
1374                         }
1375                         if ((retry_object != fs.first_object) ||
1376                             (retry_pindex != fs.first_pindex)) {
1377                                 fault_deallocate(&fs);
1378                                 goto RetryFault;
1379                         }
1380
1381                         /*
1382                          * Check whether the protection has changed or the object has
1383                          * been copied while we left the map unlocked. Changing from
1384                          * read to write permission is OK - we leave the page
1385                          * write-protected, and catch the write fault. Changing from
1386                          * write to read permission means that we can't mark the page
1387                          * write-enabled after all.
1388                          */
1389                         fs.prot &= retry_prot;
1390                         fs.fault_type &= retry_prot;
1391                         if (fs.prot == 0) {
1392                                 fault_deallocate(&fs);
1393                                 goto RetryFault;
1394                         }
1395
1396                         /* Reassert because wired may have changed. */
1397                         KASSERT(fs.wired || (fs.fault_flags & VM_FAULT_WIRE) == 0,
1398                             ("!wired && VM_FAULT_WIRE"));
1399                 }
1400         }
1401         VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1402
1403         /*
1404          * If the page was filled by a pager, save the virtual address that
1405          * should be faulted on next under a sequential access pattern to the
1406          * map entry.  A read lock on the map suffices to update this address
1407          * safely.
1408          */
1409         if (hardfault)
1410                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1411
1412         /*
1413          * Page must be completely valid or it is not fit to
1414          * map into user space.  vm_pager_get_pages() ensures this.
1415          */
1416         vm_page_assert_xbusied(fs.m);
1417         KASSERT(vm_page_all_valid(fs.m),
1418             ("vm_fault: page %p partially invalid", fs.m));
1419
1420         vm_fault_dirty(&fs, fs.m);
1421
1422         /*
1423          * Put this page into the physical map.  We had to do the unlock above
1424          * because pmap_enter() may sleep.  We don't put the page
1425          * back on the active queue until later so that the pageout daemon
1426          * won't find it (yet).
1427          */
1428         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1429             fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1430         if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1431             fs.wired == 0)
1432                 vm_fault_prefault(&fs, vaddr,
1433                     faultcount > 0 ? behind : PFBAK,
1434                     faultcount > 0 ? ahead : PFFOR, false);
1435
1436         /*
1437          * If the page is not wired down, then put it where the pageout daemon
1438          * can find it.
1439          */
1440         if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1441                 vm_page_wire(fs.m);
1442         else
1443                 vm_page_activate(fs.m);
1444         if (fs.m_hold != NULL) {
1445                 (*fs.m_hold) = fs.m;
1446                 vm_page_wire(fs.m);
1447         }
1448         vm_page_xunbusy(fs.m);
1449         fs.m = NULL;
1450
1451         /*
1452          * Unlock everything, and return
1453          */
1454         fault_deallocate(&fs);
1455         if (hardfault) {
1456                 VM_CNT_INC(v_io_faults);
1457                 curthread->td_ru.ru_majflt++;
1458 #ifdef RACCT
1459                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1460                         PROC_LOCK(curproc);
1461                         if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1462                                 racct_add_force(curproc, RACCT_WRITEBPS,
1463                                     PAGE_SIZE + behind * PAGE_SIZE);
1464                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1465                         } else {
1466                                 racct_add_force(curproc, RACCT_READBPS,
1467                                     PAGE_SIZE + ahead * PAGE_SIZE);
1468                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1469                         }
1470                         PROC_UNLOCK(curproc);
1471                 }
1472 #endif
1473         } else
1474                 curthread->td_ru.ru_minflt++;
1475
1476         return (KERN_SUCCESS);
1477 }
1478
1479 /*
1480  * Speed up the reclamation of pages that precede the faulting pindex within
1481  * the first object of the shadow chain.  Essentially, perform the equivalent
1482  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1483  * the faulting pindex by the cluster size when the pages read by vm_fault()
1484  * cross a cluster-size boundary.  The cluster size is the greater of the
1485  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1486  *
1487  * When "fs->first_object" is a shadow object, the pages in the backing object
1488  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1489  * function must only be concerned with pages in the first object.
1490  */
1491 static void
1492 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1493 {
1494         vm_map_entry_t entry;
1495         vm_object_t first_object, object;
1496         vm_offset_t end, start;
1497         vm_page_t m, m_next;
1498         vm_pindex_t pend, pstart;
1499         vm_size_t size;
1500
1501         object = fs->object;
1502         VM_OBJECT_ASSERT_UNLOCKED(object);
1503         first_object = fs->first_object;
1504         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1505         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1506                 VM_OBJECT_RLOCK(first_object);
1507                 size = VM_FAULT_DONTNEED_MIN;
1508                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1509                         size = pagesizes[1];
1510                 end = rounddown2(vaddr, size);
1511                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1512                     (entry = fs->entry)->start < end) {
1513                         if (end - entry->start < size)
1514                                 start = entry->start;
1515                         else
1516                                 start = end - size;
1517                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1518                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1519                             entry->start);
1520                         m_next = vm_page_find_least(first_object, pstart);
1521                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1522                             entry->start);
1523                         while ((m = m_next) != NULL && m->pindex < pend) {
1524                                 m_next = TAILQ_NEXT(m, listq);
1525                                 if (!vm_page_all_valid(m) ||
1526                                     vm_page_busied(m))
1527                                         continue;
1528
1529                                 /*
1530                                  * Don't clear PGA_REFERENCED, since it would
1531                                  * likely represent a reference by a different
1532                                  * process.
1533                                  *
1534                                  * Typically, at this point, prefetched pages
1535                                  * are still in the inactive queue.  Only
1536                                  * pages that triggered page faults are in the
1537                                  * active queue.  The test for whether the page
1538                                  * is in the inactive queue is racy; in the
1539                                  * worst case we will requeue the page
1540                                  * unnecessarily.
1541                                  */
1542                                 if (!vm_page_inactive(m))
1543                                         vm_page_deactivate(m);
1544                         }
1545                 }
1546                 VM_OBJECT_RUNLOCK(first_object);
1547         }
1548 }
1549
1550 /*
1551  * vm_fault_prefault provides a quick way of clustering
1552  * pagefaults into a processes address space.  It is a "cousin"
1553  * of vm_map_pmap_enter, except it runs at page fault time instead
1554  * of mmap time.
1555  */
1556 static void
1557 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1558     int backward, int forward, bool obj_locked)
1559 {
1560         pmap_t pmap;
1561         vm_map_entry_t entry;
1562         vm_object_t backing_object, lobject;
1563         vm_offset_t addr, starta;
1564         vm_pindex_t pindex;
1565         vm_page_t m;
1566         int i;
1567
1568         pmap = fs->map->pmap;
1569         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1570                 return;
1571
1572         entry = fs->entry;
1573
1574         if (addra < backward * PAGE_SIZE) {
1575                 starta = entry->start;
1576         } else {
1577                 starta = addra - backward * PAGE_SIZE;
1578                 if (starta < entry->start)
1579                         starta = entry->start;
1580         }
1581
1582         /*
1583          * Generate the sequence of virtual addresses that are candidates for
1584          * prefaulting in an outward spiral from the faulting virtual address,
1585          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1586          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1587          * If the candidate address doesn't have a backing physical page, then
1588          * the loop immediately terminates.
1589          */
1590         for (i = 0; i < 2 * imax(backward, forward); i++) {
1591                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1592                     PAGE_SIZE);
1593                 if (addr > addra + forward * PAGE_SIZE)
1594                         addr = 0;
1595
1596                 if (addr < starta || addr >= entry->end)
1597                         continue;
1598
1599                 if (!pmap_is_prefaultable(pmap, addr))
1600                         continue;
1601
1602                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1603                 lobject = entry->object.vm_object;
1604                 if (!obj_locked)
1605                         VM_OBJECT_RLOCK(lobject);
1606                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1607                     lobject->type == OBJT_DEFAULT &&
1608                     (backing_object = lobject->backing_object) != NULL) {
1609                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1610                             0, ("vm_fault_prefault: unaligned object offset"));
1611                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1612                         VM_OBJECT_RLOCK(backing_object);
1613                         if (!obj_locked || lobject != entry->object.vm_object)
1614                                 VM_OBJECT_RUNLOCK(lobject);
1615                         lobject = backing_object;
1616                 }
1617                 if (m == NULL) {
1618                         if (!obj_locked || lobject != entry->object.vm_object)
1619                                 VM_OBJECT_RUNLOCK(lobject);
1620                         break;
1621                 }
1622                 if (vm_page_all_valid(m) &&
1623                     (m->flags & PG_FICTITIOUS) == 0)
1624                         pmap_enter_quick(pmap, addr, m, entry->protection);
1625                 if (!obj_locked || lobject != entry->object.vm_object)
1626                         VM_OBJECT_RUNLOCK(lobject);
1627         }
1628 }
1629
1630 /*
1631  * Hold each of the physical pages that are mapped by the specified range of
1632  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1633  * and allow the specified types of access, "prot".  If all of the implied
1634  * pages are successfully held, then the number of held pages is returned
1635  * together with pointers to those pages in the array "ma".  However, if any
1636  * of the pages cannot be held, -1 is returned.
1637  */
1638 int
1639 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1640     vm_prot_t prot, vm_page_t *ma, int max_count)
1641 {
1642         vm_offset_t end, va;
1643         vm_page_t *mp;
1644         int count;
1645         boolean_t pmap_failed;
1646
1647         if (len == 0)
1648                 return (0);
1649         end = round_page(addr + len);
1650         addr = trunc_page(addr);
1651
1652         /*
1653          * Check for illegal addresses.
1654          */
1655         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1656                 return (-1);
1657
1658         if (atop(end - addr) > max_count)
1659                 panic("vm_fault_quick_hold_pages: count > max_count");
1660         count = atop(end - addr);
1661
1662         /*
1663          * Most likely, the physical pages are resident in the pmap, so it is
1664          * faster to try pmap_extract_and_hold() first.
1665          */
1666         pmap_failed = FALSE;
1667         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1668                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1669                 if (*mp == NULL)
1670                         pmap_failed = TRUE;
1671                 else if ((prot & VM_PROT_WRITE) != 0 &&
1672                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1673                         /*
1674                          * Explicitly dirty the physical page.  Otherwise, the
1675                          * caller's changes may go unnoticed because they are
1676                          * performed through an unmanaged mapping or by a DMA
1677                          * operation.
1678                          *
1679                          * The object lock is not held here.
1680                          * See vm_page_clear_dirty_mask().
1681                          */
1682                         vm_page_dirty(*mp);
1683                 }
1684         }
1685         if (pmap_failed) {
1686                 /*
1687                  * One or more pages could not be held by the pmap.  Either no
1688                  * page was mapped at the specified virtual address or that
1689                  * mapping had insufficient permissions.  Attempt to fault in
1690                  * and hold these pages.
1691                  *
1692                  * If vm_fault_disable_pagefaults() was called,
1693                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1694                  * acquire MD VM locks, which means we must not call
1695                  * vm_fault().  Some (out of tree) callers mark
1696                  * too wide a code area with vm_fault_disable_pagefaults()
1697                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1698                  * the proper behaviour explicitly.
1699                  */
1700                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1701                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1702                         goto error;
1703                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1704                         if (*mp == NULL && vm_fault(map, va, prot,
1705                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1706                                 goto error;
1707         }
1708         return (count);
1709 error:
1710         for (mp = ma; mp < ma + count; mp++)
1711                 if (*mp != NULL)
1712                         vm_page_unwire(*mp, PQ_INACTIVE);
1713         return (-1);
1714 }
1715
1716 /*
1717  *      Routine:
1718  *              vm_fault_copy_entry
1719  *      Function:
1720  *              Create new shadow object backing dst_entry with private copy of
1721  *              all underlying pages. When src_entry is equal to dst_entry,
1722  *              function implements COW for wired-down map entry. Otherwise,
1723  *              it forks wired entry into dst_map.
1724  *
1725  *      In/out conditions:
1726  *              The source and destination maps must be locked for write.
1727  *              The source map entry must be wired down (or be a sharing map
1728  *              entry corresponding to a main map entry that is wired down).
1729  */
1730 void
1731 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1732     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1733     vm_ooffset_t *fork_charge)
1734 {
1735         vm_object_t backing_object, dst_object, object, src_object;
1736         vm_pindex_t dst_pindex, pindex, src_pindex;
1737         vm_prot_t access, prot;
1738         vm_offset_t vaddr;
1739         vm_page_t dst_m;
1740         vm_page_t src_m;
1741         boolean_t upgrade;
1742
1743 #ifdef  lint
1744         src_map++;
1745 #endif  /* lint */
1746
1747         upgrade = src_entry == dst_entry;
1748         access = prot = dst_entry->protection;
1749
1750         src_object = src_entry->object.vm_object;
1751         src_pindex = OFF_TO_IDX(src_entry->offset);
1752
1753         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1754                 dst_object = src_object;
1755                 vm_object_reference(dst_object);
1756         } else {
1757                 /*
1758                  * Create the top-level object for the destination entry.
1759                  * Doesn't actually shadow anything - we copy the pages
1760                  * directly.
1761                  */
1762                 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1763                     dst_entry->start), NULL, NULL, 0);
1764 #if VM_NRESERVLEVEL > 0
1765                 dst_object->flags |= OBJ_COLORED;
1766                 dst_object->pg_color = atop(dst_entry->start);
1767 #endif
1768                 dst_object->domain = src_object->domain;
1769                 dst_object->charge = dst_entry->end - dst_entry->start;
1770         }
1771
1772         VM_OBJECT_WLOCK(dst_object);
1773         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1774             ("vm_fault_copy_entry: vm_object not NULL"));
1775         if (src_object != dst_object) {
1776                 dst_entry->object.vm_object = dst_object;
1777                 dst_entry->offset = 0;
1778                 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1779         }
1780         if (fork_charge != NULL) {
1781                 KASSERT(dst_entry->cred == NULL,
1782                     ("vm_fault_copy_entry: leaked swp charge"));
1783                 dst_object->cred = curthread->td_ucred;
1784                 crhold(dst_object->cred);
1785                 *fork_charge += dst_object->charge;
1786         } else if ((dst_object->type == OBJT_DEFAULT ||
1787             dst_object->type == OBJT_SWAP) &&
1788             dst_object->cred == NULL) {
1789                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1790                     dst_entry));
1791                 dst_object->cred = dst_entry->cred;
1792                 dst_entry->cred = NULL;
1793         }
1794
1795         /*
1796          * If not an upgrade, then enter the mappings in the pmap as
1797          * read and/or execute accesses.  Otherwise, enter them as
1798          * write accesses.
1799          *
1800          * A writeable large page mapping is only created if all of
1801          * the constituent small page mappings are modified. Marking
1802          * PTEs as modified on inception allows promotion to happen
1803          * without taking potentially large number of soft faults.
1804          */
1805         if (!upgrade)
1806                 access &= ~VM_PROT_WRITE;
1807
1808         /*
1809          * Loop through all of the virtual pages within the entry's
1810          * range, copying each page from the source object to the
1811          * destination object.  Since the source is wired, those pages
1812          * must exist.  In contrast, the destination is pageable.
1813          * Since the destination object doesn't share any backing storage
1814          * with the source object, all of its pages must be dirtied,
1815          * regardless of whether they can be written.
1816          */
1817         for (vaddr = dst_entry->start, dst_pindex = 0;
1818             vaddr < dst_entry->end;
1819             vaddr += PAGE_SIZE, dst_pindex++) {
1820 again:
1821                 /*
1822                  * Find the page in the source object, and copy it in.
1823                  * Because the source is wired down, the page will be
1824                  * in memory.
1825                  */
1826                 if (src_object != dst_object)
1827                         VM_OBJECT_RLOCK(src_object);
1828                 object = src_object;
1829                 pindex = src_pindex + dst_pindex;
1830                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1831                     (backing_object = object->backing_object) != NULL) {
1832                         /*
1833                          * Unless the source mapping is read-only or
1834                          * it is presently being upgraded from
1835                          * read-only, the first object in the shadow
1836                          * chain should provide all of the pages.  In
1837                          * other words, this loop body should never be
1838                          * executed when the source mapping is already
1839                          * read/write.
1840                          */
1841                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1842                             upgrade,
1843                             ("vm_fault_copy_entry: main object missing page"));
1844
1845                         VM_OBJECT_RLOCK(backing_object);
1846                         pindex += OFF_TO_IDX(object->backing_object_offset);
1847                         if (object != dst_object)
1848                                 VM_OBJECT_RUNLOCK(object);
1849                         object = backing_object;
1850                 }
1851                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1852
1853                 if (object != dst_object) {
1854                         /*
1855                          * Allocate a page in the destination object.
1856                          */
1857                         dst_m = vm_page_alloc(dst_object, (src_object ==
1858                             dst_object ? src_pindex : 0) + dst_pindex,
1859                             VM_ALLOC_NORMAL);
1860                         if (dst_m == NULL) {
1861                                 VM_OBJECT_WUNLOCK(dst_object);
1862                                 VM_OBJECT_RUNLOCK(object);
1863                                 vm_wait(dst_object);
1864                                 VM_OBJECT_WLOCK(dst_object);
1865                                 goto again;
1866                         }
1867                         pmap_copy_page(src_m, dst_m);
1868                         VM_OBJECT_RUNLOCK(object);
1869                         dst_m->dirty = dst_m->valid = src_m->valid;
1870                 } else {
1871                         dst_m = src_m;
1872                         if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1873                                 goto again;
1874                         if (dst_m->pindex >= dst_object->size) {
1875                                 /*
1876                                  * We are upgrading.  Index can occur
1877                                  * out of bounds if the object type is
1878                                  * vnode and the file was truncated.
1879                                  */
1880                                 vm_page_xunbusy(dst_m);
1881                                 break;
1882                         }
1883                 }
1884                 VM_OBJECT_WUNLOCK(dst_object);
1885
1886                 /*
1887                  * Enter it in the pmap. If a wired, copy-on-write
1888                  * mapping is being replaced by a write-enabled
1889                  * mapping, then wire that new mapping.
1890                  *
1891                  * The page can be invalid if the user called
1892                  * msync(MS_INVALIDATE) or truncated the backing vnode
1893                  * or shared memory object.  In this case, do not
1894                  * insert it into pmap, but still do the copy so that
1895                  * all copies of the wired map entry have similar
1896                  * backing pages.
1897                  */
1898                 if (vm_page_all_valid(dst_m)) {
1899                         pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1900                             access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1901                 }
1902
1903                 /*
1904                  * Mark it no longer busy, and put it on the active list.
1905                  */
1906                 VM_OBJECT_WLOCK(dst_object);
1907
1908                 if (upgrade) {
1909                         if (src_m != dst_m) {
1910                                 vm_page_unwire(src_m, PQ_INACTIVE);
1911                                 vm_page_wire(dst_m);
1912                         } else {
1913                                 KASSERT(vm_page_wired(dst_m),
1914                                     ("dst_m %p is not wired", dst_m));
1915                         }
1916                 } else {
1917                         vm_page_activate(dst_m);
1918                 }
1919                 vm_page_xunbusy(dst_m);
1920         }
1921         VM_OBJECT_WUNLOCK(dst_object);
1922         if (upgrade) {
1923                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1924                 vm_object_deallocate(src_object);
1925         }
1926 }
1927
1928 /*
1929  * Block entry into the machine-independent layer's page fault handler by
1930  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1931  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1932  * spurious page faults.
1933  */
1934 int
1935 vm_fault_disable_pagefaults(void)
1936 {
1937
1938         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1939 }
1940
1941 void
1942 vm_fault_enable_pagefaults(int save)
1943 {
1944
1945         curthread_pflags_restore(save);
1946 }