sys/vm/vm_fault.c

   1 /*-
   2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
   3  *
   4  * Copyright (c) 1991, 1993
   5  *      The Regents of the University of California.  All rights reserved.
   6  * Copyright (c) 1994 John S. Dyson
   7  * All rights reserved.
   8  * Copyright (c) 1994 David Greenman
   9  * All rights reserved.
  10  *
  11  *
  12  * This code is derived from software contributed to Berkeley by
  13  * The Mach Operating System project at Carnegie-Mellon University.
  14  *
  15  * Redistribution and use in source and binary forms, with or without
  16  * modification, are permitted provided that the following conditions
  17  * are met:
  18  * 1. Redistributions of source code must retain the above copyright
  19  *    notice, this list of conditions and the following disclaimer.
  20  * 2. Redistributions in binary form must reproduce the above copyright
  21  *    notice, this list of conditions and the following disclaimer in the
  22  *    documentation and/or other materials provided with the distribution.
  23  * 3. All advertising materials mentioning features or use of this software
  24  *    must display the following acknowledgement:
  25  *      This product includes software developed by the University of
  26  *      California, Berkeley and its contributors.
  27  * 4. Neither the name of the University nor the names of its contributors
  28  *    may be used to endorse or promote products derived from this software
  29  *    without specific prior written permission.
  30  *
  31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  41  * SUCH DAMAGE.
  42  *
  43  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  44  *
  45  *
  46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  47  * All rights reserved.
  48  *
  49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  50  *
  51  * Permission to use, copy, modify and distribute this software and
  52  * its documentation is hereby granted, provided that both the copyright
  53  * notice and this permission notice appear in all copies of the
  54  * software, derivative works or modified versions, and any portions
  55  * thereof, and that both notices appear in supporting documentation.
  56  *
  57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  60  *
  61  * Carnegie Mellon requests users of this software to return to
  62  *
  63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  64  *  School of Computer Science
  65  *  Carnegie Mellon University
  66  *  Pittsburgh PA 15213-3890
  67  *
  68  * any improvements or extensions that they make and grant Carnegie the
  69  * rights to redistribute these changes.
  70  */
  71
  72 /*
  73  *      Page fault handling module.
  74  */
  75
  76 #include <sys/cdefs.h>
  77 __FBSDID("$FreeBSD$");
  78
  79 #include "opt_ktrace.h"
  80 #include "opt_vm.h"
  81
  82 #include <sys/param.h>
  83 #include <sys/systm.h>
  84 #include <sys/kernel.h>
  85 #include <sys/lock.h>
  86 #include <sys/mman.h>
  87 #include <sys/mutex.h>
  88 #include <sys/proc.h>
  89 #include <sys/racct.h>
  90 #include <sys/refcount.h>
  91 #include <sys/resourcevar.h>
  92 #include <sys/rwlock.h>
  93 #include <sys/signalvar.h>
  94 #include <sys/sysctl.h>
  95 #include <sys/sysent.h>
  96 #include <sys/vmmeter.h>
  97 #include <sys/vnode.h>
  98 #ifdef KTRACE
  99 #include <sys/ktrace.h>
 100 #endif
 101
 102 #include <vm/vm.h>
 103 #include <vm/vm_param.h>
 104 #include <vm/pmap.h>
 105 #include <vm/vm_map.h>
 106 #include <vm/vm_object.h>
 107 #include <vm/vm_page.h>
 108 #include <vm/vm_pageout.h>
 109 #include <vm/vm_kern.h>
 110 #include <vm/vm_pager.h>
 111 #include <vm/vm_extern.h>
 112 #include <vm/vm_reserv.h>
 113
 114 #define PFBAK 4
 115 #define PFFOR 4
 116
 117 #define VM_FAULT_READ_DEFAULT   (1 + VM_FAULT_READ_AHEAD_INIT)
 118 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 119
 120 #define VM_FAULT_DONTNEED_MIN   1048576
 121
 122 struct faultstate {
 123         vm_page_t m;
 124         vm_page_t m_cow;
 125         vm_object_t object;
 126         vm_pindex_t pindex;
 127         vm_page_t first_m;
 128         vm_object_t     first_object;
 129         vm_pindex_t first_pindex;
 130         vm_map_t map;
 131         vm_map_entry_t entry;
 132         int map_generation;
 133         bool lookup_still_valid;
 134         struct vnode *vp;
 135 };
 136
 137 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
 138             int ahead);
 139 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 140             int backward, int forward, bool obj_locked);
 141
 142 static int vm_pfault_oom_attempts = 3;
 143 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
 144     &vm_pfault_oom_attempts, 0,
 145     "Number of page allocation attempts in page fault handler before it "
 146     "triggers OOM handling");
 147
 148 static int vm_pfault_oom_wait = 10;
 149 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
 150     &vm_pfault_oom_wait, 0,
 151     "Number of seconds to wait for free pages before retrying "
 152     "the page fault handler");
 153
 154 static inline void
 155 fault_page_release(vm_page_t *mp)
 156 {
 157         vm_page_t m;
 158
 159         m = *mp;
 160         if (m != NULL) {
 161                 /*
 162                  * We are likely to loop around again and attempt to busy
 163                  * this page.  Deactivating it leaves it available for
 164                  * pageout while optimizing fault restarts.
 165                  */
 166                 vm_page_deactivate(m);
 167                 vm_page_xunbusy(m);
 168                 *mp = NULL;
 169         }
 170 }
 171
 172 static inline void
 173 fault_page_free(vm_page_t *mp)
 174 {
 175         vm_page_t m;
 176
 177         m = *mp;
 178         if (m != NULL) {
 179                 VM_OBJECT_ASSERT_WLOCKED(m->object);
 180                 if (!vm_page_wired(m))
 181                         vm_page_free(m);
 182                 else
 183                         vm_page_xunbusy(m);
 184                 *mp = NULL;
 185         }
 186 }
 187
 188 static inline void
 189 unlock_map(struct faultstate *fs)
 190 {
 191
 192         if (fs->lookup_still_valid) {
 193                 vm_map_lookup_done(fs->map, fs->entry);
 194                 fs->lookup_still_valid = false;
 195         }
 196 }
 197
 198 static void
 199 unlock_vp(struct faultstate *fs)
 200 {
 201
 202         if (fs->vp != NULL) {
 203                 vput(fs->vp);
 204                 fs->vp = NULL;
 205         }
 206 }
 207
 208 static void
 209 fault_deallocate(struct faultstate *fs)
 210 {
 211
 212         fault_page_release(&fs->m_cow);
 213         fault_page_release(&fs->m);
 214         vm_object_pip_wakeup(fs->object);
 215         if (fs->object != fs->first_object) {
 216                 VM_OBJECT_WLOCK(fs->first_object);
 217                 fault_page_free(&fs->first_m);
 218                 VM_OBJECT_WUNLOCK(fs->first_object);
 219                 vm_object_pip_wakeup(fs->first_object);
 220         }
 221         vm_object_deallocate(fs->first_object);
 222         unlock_map(fs);
 223         unlock_vp(fs);
 224 }
 225
 226 static void
 227 unlock_and_deallocate(struct faultstate *fs)
 228 {
 229
 230         VM_OBJECT_WUNLOCK(fs->object);
 231         fault_deallocate(fs);
 232 }
 233
 234 static void
 235 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 236     vm_prot_t fault_type, int fault_flags)
 237 {
 238         bool need_dirty;
 239
 240         if (((prot & VM_PROT_WRITE) == 0 &&
 241             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 242             (m->oflags & VPO_UNMANAGED) != 0)
 243                 return;
 244
 245         VM_PAGE_OBJECT_BUSY_ASSERT(m);
 246
 247         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 248             (fault_flags & VM_FAULT_WIRE) == 0) ||
 249             (fault_flags & VM_FAULT_DIRTY) != 0;
 250
 251         vm_object_set_writeable_dirty(m->object);
 252
 253         /*
 254          * If the fault is a write, we know that this page is being
 255          * written NOW so dirty it explicitly to save on
 256          * pmap_is_modified() calls later.
 257          *
 258          * Also, since the page is now dirty, we can possibly tell
 259          * the pager to release any swap backing the page.
 260          */
 261         if (need_dirty && vm_page_set_dirty(m) == 0) {
 262                 /*
 263                  * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
 264                  * if the page is already dirty to prevent data written with
 265                  * the expectation of being synced from not being synced.
 266                  * Likewise if this entry does not request NOSYNC then make
 267                  * sure the page isn't marked NOSYNC.  Applications sharing
 268                  * data should use the same flags to avoid ping ponging.
 269                  */
 270                 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0)
 271                         vm_page_aflag_set(m, PGA_NOSYNC);
 272                 else
 273                         vm_page_aflag_clear(m, PGA_NOSYNC);
 274         }
 275
 276 }
 277
 278 /*
 279  * Unlocks fs.first_object and fs.map on success.
 280  */
 281 static int
 282 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
 283     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
 284 {
 285         vm_page_t m, m_map;
 286 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 287     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 288     VM_NRESERVLEVEL > 0
 289         vm_page_t m_super;
 290         int flags;
 291 #endif
 292         int psind, rv;
 293
 294         MPASS(fs->vp == NULL);
 295         vm_object_busy(fs->first_object);
 296         m = vm_page_lookup(fs->first_object, fs->first_pindex);
 297         /* A busy page can be mapped for read|execute access. */
 298         if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 299             vm_page_busied(m)) || !vm_page_all_valid(m)) {
 300                 rv = KERN_FAILURE;
 301                 goto out;
 302         }
 303         m_map = m;
 304         psind = 0;
 305 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 306     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
 307     VM_NRESERVLEVEL > 0
 308         if ((m->flags & PG_FICTITIOUS) == 0 &&
 309             (m_super = vm_reserv_to_superpage(m)) != NULL &&
 310             rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
 311             roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
 312             (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
 313             (pagesizes[m_super->psind] - 1)) && !wired &&
 314             pmap_ps_enabled(fs->map->pmap)) {
 315                 flags = PS_ALL_VALID;
 316                 if ((prot & VM_PROT_WRITE) != 0) {
 317                         /*
 318                          * Create a superpage mapping allowing write access
 319                          * only if none of the constituent pages are busy and
 320                          * all of them are already dirty (except possibly for
 321                          * the page that was faulted on).
 322                          */
 323                         flags |= PS_NONE_BUSY;
 324                         if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
 325                                 flags |= PS_ALL_DIRTY;
 326                 }
 327                 if (vm_page_ps_test(m_super, flags, m)) {
 328                         m_map = m_super;
 329                         psind = m_super->psind;
 330                         vaddr = rounddown2(vaddr, pagesizes[psind]);
 331                         /* Preset the modified bit for dirty superpages. */
 332                         if ((flags & PS_ALL_DIRTY) != 0)
 333                                 fault_type |= VM_PROT_WRITE;
 334                 }
 335         }
 336 #endif
 337         rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
 338             PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
 339         if (rv != KERN_SUCCESS)
 340                 goto out;
 341         if (m_hold != NULL) {
 342                 *m_hold = m;
 343                 vm_page_wire(m);
 344         }
 345         vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags);
 346         if (psind == 0 && !wired)
 347                 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
 348         VM_OBJECT_RUNLOCK(fs->first_object);
 349         vm_map_lookup_done(fs->map, fs->entry);
 350         curthread->td_ru.ru_minflt++;
 351
 352 out:
 353         vm_object_unbusy(fs->first_object);
 354         return (rv);
 355 }
 356
 357 static void
 358 vm_fault_restore_map_lock(struct faultstate *fs)
 359 {
 360
 361         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 362         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 363
 364         if (!vm_map_trylock_read(fs->map)) {
 365                 VM_OBJECT_WUNLOCK(fs->first_object);
 366                 vm_map_lock_read(fs->map);
 367                 VM_OBJECT_WLOCK(fs->first_object);
 368         }
 369         fs->lookup_still_valid = true;
 370 }
 371
 372 static void
 373 vm_fault_populate_check_page(vm_page_t m)
 374 {
 375
 376         /*
 377          * Check each page to ensure that the pager is obeying the
 378          * interface: the page must be installed in the object, fully
 379          * valid, and exclusively busied.
 380          */
 381         MPASS(m != NULL);
 382         MPASS(vm_page_all_valid(m));
 383         MPASS(vm_page_xbusied(m));
 384 }
 385
 386 static void
 387 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
 388     vm_pindex_t last)
 389 {
 390         vm_page_t m;
 391         vm_pindex_t pidx;
 392
 393         VM_OBJECT_ASSERT_WLOCKED(object);
 394         MPASS(first <= last);
 395         for (pidx = first, m = vm_page_lookup(object, pidx);
 396             pidx <= last; pidx++, m = vm_page_next(m)) {
 397                 vm_fault_populate_check_page(m);
 398                 vm_page_deactivate(m);
 399                 vm_page_xunbusy(m);
 400         }
 401 }
 402
 403 static int
 404 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
 405     int fault_flags, boolean_t wired, vm_page_t *m_hold)
 406 {
 407         vm_offset_t vaddr;
 408         vm_page_t m;
 409         vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
 410         int i, npages, psind, rv;
 411
 412         MPASS(fs->object == fs->first_object);
 413         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 414         MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
 415         MPASS(fs->first_object->backing_object == NULL);
 416         MPASS(fs->lookup_still_valid);
 417
 418         pager_first = OFF_TO_IDX(fs->entry->offset);
 419         pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
 420         unlock_map(fs);
 421         unlock_vp(fs);
 422
 423         /*
 424          * Call the pager (driver) populate() method.
 425          *
 426          * There is no guarantee that the method will be called again
 427          * if the current fault is for read, and a future fault is
 428          * for write.  Report the entry's maximum allowed protection
 429          * to the driver.
 430          */
 431         rv = vm_pager_populate(fs->first_object, fs->first_pindex,
 432             fault_type, fs->entry->max_protection, &pager_first, &pager_last);
 433
 434         VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
 435         if (rv == VM_PAGER_BAD) {
 436                 /*
 437                  * VM_PAGER_BAD is the backdoor for a pager to request
 438                  * normal fault handling.
 439                  */
 440                 vm_fault_restore_map_lock(fs);
 441                 if (fs->map->timestamp != fs->map_generation)
 442                         return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 443                 return (KERN_NOT_RECEIVER);
 444         }
 445         if (rv != VM_PAGER_OK)
 446                 return (KERN_FAILURE); /* AKA SIGSEGV */
 447
 448         /* Ensure that the driver is obeying the interface. */
 449         MPASS(pager_first <= pager_last);
 450         MPASS(fs->first_pindex <= pager_last);
 451         MPASS(fs->first_pindex >= pager_first);
 452         MPASS(pager_last < fs->first_object->size);
 453
 454         vm_fault_restore_map_lock(fs);
 455         if (fs->map->timestamp != fs->map_generation) {
 456                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 457                     pager_last);
 458                 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
 459         }
 460
 461         /*
 462          * The map is unchanged after our last unlock.  Process the fault.
 463          *
 464          * The range [pager_first, pager_last] that is given to the
 465          * pager is only a hint.  The pager may populate any range
 466          * within the object that includes the requested page index.
 467          * In case the pager expanded the range, clip it to fit into
 468          * the map entry.
 469          */
 470         map_first = OFF_TO_IDX(fs->entry->offset);
 471         if (map_first > pager_first) {
 472                 vm_fault_populate_cleanup(fs->first_object, pager_first,
 473                     map_first - 1);
 474                 pager_first = map_first;
 475         }
 476         map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
 477         if (map_last < pager_last) {
 478                 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
 479                     pager_last);
 480                 pager_last = map_last;
 481         }
 482         for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
 483             pidx <= pager_last;
 484             pidx += npages, m = vm_page_next(&m[npages - 1])) {
 485                 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
 486 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
 487     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
 488                 psind = m->psind;
 489                 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
 490                     pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
 491                     !pmap_ps_enabled(fs->map->pmap) || wired))
 492                         psind = 0;
 493 #else
 494                 psind = 0;
 495 #endif
 496                 npages = atop(pagesizes[psind]);
 497                 for (i = 0; i < npages; i++) {
 498                         vm_fault_populate_check_page(&m[i]);
 499                         vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
 500                             fault_flags);
 501                 }
 502                 VM_OBJECT_WUNLOCK(fs->first_object);
 503                 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
 504                     (wired ? PMAP_ENTER_WIRED : 0), psind);
 505 #if defined(__amd64__)
 506                 if (psind > 0 && rv == KERN_FAILURE) {
 507                         for (i = 0; i < npages; i++) {
 508                                 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
 509                                     &m[i], prot, fault_type |
 510                                     (wired ? PMAP_ENTER_WIRED : 0), 0);
 511                                 MPASS(rv == KERN_SUCCESS);
 512                         }
 513                 }
 514 #else
 515                 MPASS(rv == KERN_SUCCESS);
 516 #endif
 517                 VM_OBJECT_WLOCK(fs->first_object);
 518                 for (i = 0; i < npages; i++) {
 519                         if ((fault_flags & VM_FAULT_WIRE) != 0)
 520                                 vm_page_wire(&m[i]);
 521                         else
 522                                 vm_page_activate(&m[i]);
 523                         if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
 524                                 *m_hold = &m[i];
 525                                 vm_page_wire(&m[i]);
 526                         }
 527                         vm_page_xunbusy(&m[i]);
 528                 }
 529         }
 530         curthread->td_ru.ru_majflt++;
 531         return (KERN_SUCCESS);
 532 }
 533
 534 static int prot_fault_translation;
 535 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
 536     &prot_fault_translation, 0,
 537     "Control signal to deliver on protection fault");
 538
 539 /* compat definition to keep common code for signal translation */
 540 #define UCODE_PAGEFLT   12
 541 #ifdef T_PAGEFLT
 542 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
 543 #endif
 544
 545 /*
 546  *      vm_fault_trap:
 547  *
 548  *      Handle a page fault occurring at the given address,
 549  *      requiring the given permissions, in the map specified.
 550  *      If successful, the page is inserted into the
 551  *      associated physical map.
 552  *
 553  *      NOTE: the given address should be truncated to the
 554  *      proper page address.
 555  *
 556  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 557  *      a standard error specifying why the fault is fatal is returned.
 558  *
 559  *      The map in question must be referenced, and remains so.
 560  *      Caller may hold no locks.
 561  */
 562 int
 563 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 564     int fault_flags, int *signo, int *ucode)
 565 {
 566         int result;
 567
 568         MPASS(signo == NULL || ucode != NULL);
 569 #ifdef KTRACE
 570         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
 571                 ktrfault(vaddr, fault_type);
 572 #endif
 573         result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
 574             NULL);
 575         KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
 576             result == KERN_INVALID_ADDRESS ||
 577             result == KERN_RESOURCE_SHORTAGE ||
 578             result == KERN_PROTECTION_FAILURE ||
 579             result == KERN_OUT_OF_BOUNDS,
 580             ("Unexpected Mach error %d from vm_fault()", result));
 581 #ifdef KTRACE
 582         if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
 583                 ktrfaultend(result);
 584 #endif
 585         if (result != KERN_SUCCESS && signo != NULL) {
 586                 switch (result) {
 587                 case KERN_FAILURE:
 588                 case KERN_INVALID_ADDRESS:
 589                         *signo = SIGSEGV;
 590                         *ucode = SEGV_MAPERR;
 591                         break;
 592                 case KERN_RESOURCE_SHORTAGE:
 593                         *signo = SIGBUS;
 594                         *ucode = BUS_OOMERR;
 595                         break;
 596                 case KERN_OUT_OF_BOUNDS:
 597                         *signo = SIGBUS;
 598                         *ucode = BUS_OBJERR;
 599                         break;
 600                 case KERN_PROTECTION_FAILURE:
 601                         if (prot_fault_translation == 0) {
 602                                 /*
 603                                  * Autodetect.  This check also covers
 604                                  * the images without the ABI-tag ELF
 605                                  * note.
 606                                  */
 607                                 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
 608                                     curproc->p_osrel >= P_OSREL_SIGSEGV) {
 609                                         *signo = SIGSEGV;
 610                                         *ucode = SEGV_ACCERR;
 611                                 } else {
 612                                         *signo = SIGBUS;
 613                                         *ucode = UCODE_PAGEFLT;
 614                                 }
 615                         } else if (prot_fault_translation == 1) {
 616                                 /* Always compat mode. */
 617                                 *signo = SIGBUS;
 618                                 *ucode = UCODE_PAGEFLT;
 619                         } else {
 620                                 /* Always SIGSEGV mode. */
 621                                 *signo = SIGSEGV;
 622                                 *ucode = SEGV_ACCERR;
 623                         }
 624                         break;
 625                 default:
 626                         KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
 627                             result));
 628                         break;
 629                 }
 630         }
 631         return (result);
 632 }
 633
 634 static int
 635 vm_fault_lock_vnode(struct faultstate *fs)
 636 {
 637         struct vnode *vp;
 638         int error, locked;
 639
 640         if (fs->object->type != OBJT_VNODE)
 641                 return (KERN_SUCCESS);
 642         vp = fs->object->handle;
 643         if (vp == fs->vp) {
 644                 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
 645                 return (KERN_SUCCESS);
 646         }
 647
 648         /*
 649          * Perform an unlock in case the desired vnode changed while
 650          * the map was unlocked during a retry.
 651          */
 652         unlock_vp(fs);
 653
 654         locked = VOP_ISLOCKED(vp);
 655         if (locked != LK_EXCLUSIVE)
 656                 locked = LK_SHARED;
 657
 658         /*
 659          * We must not sleep acquiring the vnode lock while we have
 660          * the page exclusive busied or the object's
 661          * paging-in-progress count incremented.  Otherwise, we could
 662          * deadlock.
 663          */
 664         error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
 665         if (error == 0) {
 666                 fs->vp = vp;
 667                 return (KERN_SUCCESS);
 668         }
 669
 670         vhold(vp);
 671         unlock_and_deallocate(fs);
 672         error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
 673         vdrop(vp);
 674         fs->vp = vp;
 675         KASSERT(error == 0, ("vm_fault: vget failed %d", error));
 676         return (KERN_RESOURCE_SHORTAGE);
 677 }
 678
 679 /*
 680  * Wait/Retry if the page is busy.  We have to do this if the page is
 681  * either exclusive or shared busy because the vm_pager may be using
 682  * read busy for pageouts (and even pageins if it is the vnode pager),
 683  * and we could end up trying to pagein and pageout the same page
 684  * simultaneously.
 685  *
 686  * We can theoretically allow the busy case on a read fault if the page
 687  * is marked valid, but since such pages are typically already pmap'd,
 688  * putting that special case in might be more effort then it is worth.
 689  * We cannot under any circumstances mess around with a shared busied
 690  * page except, perhaps, to pmap it.
 691  */
 692 static void
 693 vm_fault_busy_sleep(struct faultstate *fs)
 694 {
 695         /*
 696          * Reference the page before unlocking and
 697          * sleeping so that the page daemon is less
 698          * likely to reclaim it.
 699          */
 700         vm_page_aflag_set(fs->m, PGA_REFERENCED);
 701         if (fs->object != fs->first_object) {
 702                 fault_page_release(&fs->first_m);
 703                 vm_object_pip_wakeup(fs->first_object);
 704         }
 705         vm_object_pip_wakeup(fs->object);
 706         unlock_map(fs);
 707         if (fs->m == vm_page_lookup(fs->object, fs->pindex))
 708                 vm_page_busy_sleep(fs->m, "vmpfw", false);
 709         else
 710                 VM_OBJECT_WUNLOCK(fs->object);
 711         VM_CNT_INC(v_intrans);
 712         vm_object_deallocate(fs->first_object);
 713 }
 714
 715 int
 716 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 717     int fault_flags, vm_page_t *m_hold)
 718 {
 719         struct faultstate fs;
 720         struct domainset *dset;
 721         vm_object_t next_object, retry_object;
 722         vm_offset_t e_end, e_start;
 723         vm_pindex_t retry_pindex;
 724         vm_prot_t prot, retry_prot;
 725         int ahead, alloc_req, behind, cluster_offset, era, faultcount;
 726         int nera, oom, result, rv;
 727         u_char behavior;
 728         boolean_t wired;        /* Passed by reference. */
 729         bool dead, hardfault, is_first_object_locked;
 730
 731         VM_CNT_INC(v_vm_faults);
 732
 733         if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
 734                 return (KERN_PROTECTION_FAILURE);
 735
 736         fs.vp = NULL;
 737         faultcount = 0;
 738         nera = -1;
 739         hardfault = false;
 740
 741 RetryFault:
 742         oom = 0;
 743 RetryFault_oom:
 744
 745         /*
 746          * Find the backing store object and offset into it to begin the
 747          * search.
 748          */
 749         fs.map = map;
 750         result = vm_map_lookup(&fs.map, vaddr, fault_type |
 751             VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
 752             &fs.first_pindex, &prot, &wired);
 753         if (result != KERN_SUCCESS) {
 754                 unlock_vp(&fs);
 755                 return (result);
 756         }
 757
 758         fs.map_generation = fs.map->timestamp;
 759
 760         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 761                 panic("%s: fault on nofault entry, addr: %#lx",
 762                     __func__, (u_long)vaddr);
 763         }
 764
 765         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 766             fs.entry->wiring_thread != curthread) {
 767                 vm_map_unlock_read(fs.map);
 768                 vm_map_lock(fs.map);
 769                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 770                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 771                         unlock_vp(&fs);
 772                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 773                         vm_map_unlock_and_wait(fs.map, 0);
 774                 } else
 775                         vm_map_unlock(fs.map);
 776                 goto RetryFault;
 777         }
 778
 779         MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
 780
 781         if (wired)
 782                 fault_type = prot | (fault_type & VM_PROT_COPY);
 783         else
 784                 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
 785                     ("!wired && VM_FAULT_WIRE"));
 786
 787         /*
 788          * Try to avoid lock contention on the top-level object through
 789          * special-case handling of some types of page faults, specifically,
 790          * those that are mapping an existing page from the top-level object.
 791          * Under this condition, a read lock on the object suffices, allowing
 792          * multiple page faults of a similar type to run in parallel.
 793          */
 794         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 795             (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
 796                 VM_OBJECT_RLOCK(fs.first_object);
 797                 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
 798                     fault_flags, wired, m_hold);
 799                 if (rv == KERN_SUCCESS)
 800                         return (rv);
 801                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 802                         VM_OBJECT_RUNLOCK(fs.first_object);
 803                         VM_OBJECT_WLOCK(fs.first_object);
 804                 }
 805         } else {
 806                 VM_OBJECT_WLOCK(fs.first_object);
 807         }
 808
 809         /*
 810          * Make a reference to this object to prevent its disposal while we
 811          * are messing with it.  Once we have the reference, the map is free
 812          * to be diddled.  Since objects reference their shadows (and copies),
 813          * they will stay around as well.
 814          *
 815          * Bump the paging-in-progress count to prevent size changes (e.g.
 816          * truncation operations) during I/O.
 817          */
 818         vm_object_reference_locked(fs.first_object);
 819         vm_object_pip_add(fs.first_object, 1);
 820
 821         fs.lookup_still_valid = true;
 822
 823         fs.m_cow = fs.m = fs.first_m = NULL;
 824
 825         /*
 826          * Search for the page at object/offset.
 827          */
 828         fs.object = fs.first_object;
 829         fs.pindex = fs.first_pindex;
 830         while (TRUE) {
 831                 KASSERT(fs.m == NULL,
 832                     ("page still set %p at loop start", fs.m));
 833                 /*
 834                  * If the object is marked for imminent termination,
 835                  * we retry here, since the collapse pass has raced
 836                  * with us.  Otherwise, if we see terminally dead
 837                  * object, return fail.
 838                  */
 839                 if ((fs.object->flags & OBJ_DEAD) != 0) {
 840                         dead = fs.object->type == OBJT_DEAD;
 841                         unlock_and_deallocate(&fs);
 842                         if (dead)
 843                                 return (KERN_PROTECTION_FAILURE);
 844                         pause("vmf_de", 1);
 845                         goto RetryFault;
 846                 }
 847
 848                 /*
 849                  * See if page is resident
 850                  */
 851                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 852                 if (fs.m != NULL) {
 853                         if (vm_page_tryxbusy(fs.m) == 0) {
 854                                 vm_fault_busy_sleep(&fs);
 855                                 goto RetryFault;
 856                         }
 857
 858                         /*
 859                          * The page is marked busy for other processes and the
 860                          * pagedaemon.  If it still isn't completely valid
 861                          * (readable), jump to readrest, else break-out ( we
 862                          * found the page ).
 863                          */
 864                         if (!vm_page_all_valid(fs.m))
 865                                 goto readrest;
 866                         break; /* break to PAGE HAS BEEN FOUND */
 867                 }
 868                 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
 869
 870                 /*
 871                  * Page is not resident.  If the pager might contain the page
 872                  * or this is the beginning of the search, allocate a new
 873                  * page.  (Default objects are zero-fill, so there is no real
 874                  * pager for them.)
 875                  */
 876                 if (fs.object->type != OBJT_DEFAULT ||
 877                     fs.object == fs.first_object) {
 878                         if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
 879                                 rv = vm_fault_lock_vnode(&fs);
 880                                 MPASS(rv == KERN_SUCCESS ||
 881                                     rv == KERN_RESOURCE_SHORTAGE);
 882                                 if (rv == KERN_RESOURCE_SHORTAGE)
 883                                         goto RetryFault;
 884                         }
 885                         if (fs.pindex >= fs.object->size) {
 886                                 unlock_and_deallocate(&fs);
 887                                 return (KERN_OUT_OF_BOUNDS);
 888                         }
 889
 890                         if (fs.object == fs.first_object &&
 891                             (fs.first_object->flags & OBJ_POPULATE) != 0 &&
 892                             fs.first_object->shadow_count == 0) {
 893                                 rv = vm_fault_populate(&fs, prot, fault_type,
 894                                     fault_flags, wired, m_hold);
 895                                 switch (rv) {
 896                                 case KERN_SUCCESS:
 897                                 case KERN_FAILURE:
 898                                         unlock_and_deallocate(&fs);
 899                                         return (rv);
 900                                 case KERN_RESOURCE_SHORTAGE:
 901                                         unlock_and_deallocate(&fs);
 902                                         goto RetryFault;
 903                                 case KERN_NOT_RECEIVER:
 904                                         /*
 905                                          * Pager's populate() method
 906                                          * returned VM_PAGER_BAD.
 907                                          */
 908                                         break;
 909                                 default:
 910                                         panic("inconsistent return codes");
 911                                 }
 912                         }
 913
 914                         /*
 915                          * Allocate a new page for this object/offset pair.
 916                          *
 917                          * Unlocked read of the p_flag is harmless. At
 918                          * worst, the P_KILLED might be not observed
 919                          * there, and allocation can fail, causing
 920                          * restart and new reading of the p_flag.
 921                          */
 922                         dset = fs.object->domain.dr_policy;
 923                         if (dset == NULL)
 924                                 dset = curthread->td_domain.dr_policy;
 925                         if (!vm_page_count_severe_set(&dset->ds_mask) ||
 926                             P_KILLED(curproc)) {
 927 #if VM_NRESERVLEVEL > 0
 928                                 vm_object_color(fs.object, atop(vaddr) -
 929                                     fs.pindex);
 930 #endif
 931                                 alloc_req = P_KILLED(curproc) ?
 932                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 933                                 if (fs.object->type != OBJT_VNODE &&
 934                                     fs.object->backing_object == NULL)
 935                                         alloc_req |= VM_ALLOC_ZERO;
 936                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 937                                     alloc_req);
 938                         }
 939                         if (fs.m == NULL) {
 940                                 unlock_and_deallocate(&fs);
 941                                 if (vm_pfault_oom_attempts < 0 ||
 942                                     oom < vm_pfault_oom_attempts) {
 943                                         oom++;
 944                                         vm_waitpfault(dset,
 945                                             vm_pfault_oom_wait * hz);
 946                                         goto RetryFault_oom;
 947                                 }
 948                                 if (bootverbose)
 949                                         printf(
 950         "proc %d (%s) failed to alloc page on fault, starting OOM\n",
 951                                             curproc->p_pid, curproc->p_comm);
 952                                 vm_pageout_oom(VM_OOM_MEM_PF);
 953                                 goto RetryFault;
 954                         }
 955                 }
 956
 957 readrest:
 958                 /*
 959                  * At this point, we have either allocated a new page or found
 960                  * an existing page that is only partially valid.
 961                  *
 962                  * We hold a reference on the current object and the page is
 963                  * exclusive busied.
 964                  */
 965
 966                 /*
 967                  * If the pager for the current object might have the page,
 968                  * then determine the number of additional pages to read and
 969                  * potentially reprioritize previously read pages for earlier
 970                  * reclamation.  These operations should only be performed
 971                  * once per page fault.  Even if the current pager doesn't
 972                  * have the page, the number of additional pages to read will
 973                  * apply to subsequent objects in the shadow chain.
 974                  */
 975                 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
 976                     !P_KILLED(curproc)) {
 977                         KASSERT(fs.lookup_still_valid, ("map unlocked"));
 978                         era = fs.entry->read_ahead;
 979                         behavior = vm_map_entry_behavior(fs.entry);
 980                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 981                                 nera = 0;
 982                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 983                                 nera = VM_FAULT_READ_AHEAD_MAX;
 984                                 if (vaddr == fs.entry->next_read)
 985                                         vm_fault_dontneed(&fs, vaddr, nera);
 986                         } else if (vaddr == fs.entry->next_read) {
 987                                 /*
 988                                  * This is a sequential fault.  Arithmetically
 989                                  * increase the requested number of pages in
 990                                  * the read-ahead window.  The requested
 991                                  * number of pages is "# of sequential faults
 992                                  * x (read ahead min + 1) + read ahead min"
 993                                  */
 994                                 nera = VM_FAULT_READ_AHEAD_MIN;
 995                                 if (era > 0) {
 996                                         nera += era + 1;
 997                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 998                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 999                                 }
1000                                 if (era == VM_FAULT_READ_AHEAD_MAX)
1001                                         vm_fault_dontneed(&fs, vaddr, nera);
1002                         } else {
1003                                 /*
1004                                  * This is a non-sequential fault.
1005                                  */
1006                                 nera = 0;
1007                         }
1008                         if (era != nera) {
1009                                 /*
1010                                  * A read lock on the map suffices to update
1011                                  * the read ahead count safely.
1012                                  */
1013                                 fs.entry->read_ahead = nera;
1014                         }
1015
1016                         /*
1017                          * Prepare for unlocking the map.  Save the map
1018                          * entry's start and end addresses, which are used to
1019                          * optimize the size of the pager operation below.
1020                          * Even if the map entry's addresses change after
1021                          * unlocking the map, using the saved addresses is
1022                          * safe.
1023                          */
1024                         e_start = fs.entry->start;
1025                         e_end = fs.entry->end;
1026                 }
1027
1028                 /*
1029                  * Call the pager to retrieve the page if there is a chance
1030                  * that the pager has it, and potentially retrieve additional
1031                  * pages at the same time.
1032                  */
1033                 if (fs.object->type != OBJT_DEFAULT) {
1034                         /*
1035                          * Release the map lock before locking the vnode or
1036                          * sleeping in the pager.  (If the current object has
1037                          * a shadow, then an earlier iteration of this loop
1038                          * may have already unlocked the map.)
1039                          */
1040                         unlock_map(&fs);
1041
1042                         rv = vm_fault_lock_vnode(&fs);
1043                         MPASS(rv == KERN_SUCCESS ||
1044                             rv == KERN_RESOURCE_SHORTAGE);
1045                         if (rv == KERN_RESOURCE_SHORTAGE)
1046                                 goto RetryFault;
1047                         KASSERT(fs.vp == NULL || !fs.map->system_map,
1048                             ("vm_fault: vnode-backed object mapped by system map"));
1049
1050                         /*
1051                          * Page in the requested page and hint the pager,
1052                          * that it may bring up surrounding pages.
1053                          */
1054                         if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1055                             P_KILLED(curproc)) {
1056                                 behind = 0;
1057                                 ahead = 0;
1058                         } else {
1059                                 /* Is this a sequential fault? */
1060                                 if (nera > 0) {
1061                                         behind = 0;
1062                                         ahead = nera;
1063                                 } else {
1064                                         /*
1065                                          * Request a cluster of pages that is
1066                                          * aligned to a VM_FAULT_READ_DEFAULT
1067                                          * page offset boundary within the
1068                                          * object.  Alignment to a page offset
1069                                          * boundary is more likely to coincide
1070                                          * with the underlying file system
1071                                          * block than alignment to a virtual
1072                                          * address boundary.
1073                                          */
1074                                         cluster_offset = fs.pindex %
1075                                             VM_FAULT_READ_DEFAULT;
1076                                         behind = ulmin(cluster_offset,
1077                                             atop(vaddr - e_start));
1078                                         ahead = VM_FAULT_READ_DEFAULT - 1 -
1079                                             cluster_offset;
1080                                 }
1081                                 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1082                         }
1083                         VM_OBJECT_WUNLOCK(fs.object);
1084                         rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1085                             &behind, &ahead);
1086                         VM_OBJECT_WLOCK(fs.object);
1087                         if (rv == VM_PAGER_OK) {
1088                                 faultcount = behind + 1 + ahead;
1089                                 hardfault = true;
1090                                 break; /* break to PAGE HAS BEEN FOUND */
1091                         }
1092                         if (rv == VM_PAGER_ERROR)
1093                                 printf("vm_fault: pager read error, pid %d (%s)\n",
1094                                     curproc->p_pid, curproc->p_comm);
1095
1096                         /*
1097                          * If an I/O error occurred or the requested page was
1098                          * outside the range of the pager, clean up and return
1099                          * an error.
1100                          */
1101                         if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1102                                 fault_page_free(&fs.m);
1103                                 unlock_and_deallocate(&fs);
1104                                 return (KERN_OUT_OF_BOUNDS);
1105                         }
1106
1107                 }
1108
1109                 /*
1110                  * The requested page does not exist at this object/
1111                  * offset.  Remove the invalid page from the object,
1112                  * waking up anyone waiting for it, and continue on to
1113                  * the next object.  However, if this is the top-level
1114                  * object, we must leave the busy page in place to
1115                  * prevent another process from rushing past us, and
1116                  * inserting the page in that object at the same time
1117                  * that we are.
1118                  */
1119                 if (fs.object == fs.first_object) {
1120                         fs.first_m = fs.m;
1121                         fs.m = NULL;
1122                 } else
1123                         fault_page_free(&fs.m);
1124
1125                 /*
1126                  * Move on to the next object.  Lock the next object before
1127                  * unlocking the current one.
1128                  */
1129                 next_object = fs.object->backing_object;
1130                 if (next_object == NULL) {
1131                         /*
1132                          * If there's no object left, fill the page in the top
1133                          * object with zeros.
1134                          */
1135                         if (fs.object != fs.first_object) {
1136                                 vm_object_pip_wakeup(fs.object);
1137                                 VM_OBJECT_WUNLOCK(fs.object);
1138
1139                                 fs.object = fs.first_object;
1140                                 fs.pindex = fs.first_pindex;
1141                                 VM_OBJECT_WLOCK(fs.object);
1142                         }
1143                         MPASS(fs.first_m != NULL);
1144                         MPASS(fs.m == NULL);
1145                         fs.m = fs.first_m;
1146                         fs.first_m = NULL;
1147
1148                         /*
1149                          * Zero the page if necessary and mark it valid.
1150                          */
1151                         if ((fs.m->flags & PG_ZERO) == 0) {
1152                                 pmap_zero_page(fs.m);
1153                         } else {
1154                                 VM_CNT_INC(v_ozfod);
1155                         }
1156                         VM_CNT_INC(v_zfod);
1157                         vm_page_valid(fs.m);
1158                         /* Don't try to prefault neighboring pages. */
1159                         faultcount = 1;
1160                         break;  /* break to PAGE HAS BEEN FOUND */
1161                 } else {
1162                         MPASS(fs.first_m != NULL);
1163                         KASSERT(fs.object != next_object,
1164                             ("object loop %p", next_object));
1165                         VM_OBJECT_WLOCK(next_object);
1166                         vm_object_pip_add(next_object, 1);
1167                         if (fs.object != fs.first_object)
1168                                 vm_object_pip_wakeup(fs.object);
1169                         fs.pindex +=
1170                             OFF_TO_IDX(fs.object->backing_object_offset);
1171                         VM_OBJECT_WUNLOCK(fs.object);
1172                         fs.object = next_object;
1173                 }
1174         }
1175
1176         vm_page_assert_xbusied(fs.m);
1177
1178         /*
1179          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1180          * is held.]
1181          */
1182
1183         /*
1184          * If the page is being written, but isn't already owned by the
1185          * top-level object, we have to copy it into a new page owned by the
1186          * top-level object.
1187          */
1188         if (fs.object != fs.first_object) {
1189                 /*
1190                  * We only really need to copy if we want to write it.
1191                  */
1192                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1193                         /*
1194                          * This allows pages to be virtually copied from a
1195                          * backing_object into the first_object, where the
1196                          * backing object has no other refs to it, and cannot
1197                          * gain any more refs.  Instead of a bcopy, we just
1198                          * move the page from the backing object to the
1199                          * first object.  Note that we must mark the page
1200                          * dirty in the first object so that it will go out
1201                          * to swap when needed.
1202                          */
1203                         is_first_object_locked = false;
1204                         if (
1205                                 /*
1206                                  * Only one shadow object
1207                                  */
1208                                 (fs.object->shadow_count == 1) &&
1209                                 /*
1210                                  * No COW refs, except us
1211                                  */
1212                                 (fs.object->ref_count == 1) &&
1213                                 /*
1214                                  * No one else can look this object up
1215                                  */
1216                                 (fs.object->handle == NULL) &&
1217                                 /*
1218                                  * No other ways to look the object up
1219                                  */
1220                                 ((fs.object->flags & OBJ_ANON) != 0) &&
1221                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1222                                 /*
1223                                  * We don't chase down the shadow chain
1224                                  */
1225                             fs.object == fs.first_object->backing_object) {
1226
1227                                 /*
1228                                  * Remove but keep xbusy for replace.  fs.m is
1229                                  * moved into fs.first_object and left busy
1230                                  * while fs.first_m is conditionally freed.
1231                                  */
1232                                 vm_page_remove_xbusy(fs.m);
1233                                 vm_page_replace(fs.m, fs.first_object,
1234                                     fs.first_pindex, fs.first_m);
1235                                 vm_page_dirty(fs.m);
1236 #if VM_NRESERVLEVEL > 0
1237                                 /*
1238                                  * Rename the reservation.
1239                                  */
1240                                 vm_reserv_rename(fs.m, fs.first_object,
1241                                     fs.object, OFF_TO_IDX(
1242                                     fs.first_object->backing_object_offset));
1243 #endif
1244                                 VM_OBJECT_WUNLOCK(fs.object);
1245                                 fs.first_m = fs.m;
1246                                 fs.m = NULL;
1247                                 VM_CNT_INC(v_cow_optim);
1248                         } else {
1249                                 VM_OBJECT_WUNLOCK(fs.object);
1250                                 /*
1251                                  * Oh, well, lets copy it.
1252                                  */
1253                                 pmap_copy_page(fs.m, fs.first_m);
1254                                 vm_page_valid(fs.first_m);
1255                                 if (wired && (fault_flags &
1256                                     VM_FAULT_WIRE) == 0) {
1257                                         vm_page_wire(fs.first_m);
1258                                         vm_page_unwire(fs.m, PQ_INACTIVE);
1259                                 }
1260                                 /*
1261                                  * Save the cow page to be released after
1262                                  * pmap_enter is complete.
1263                                  */
1264                                 fs.m_cow = fs.m;
1265                                 fs.m = NULL;
1266                         }
1267                         /*
1268                          * fs.object != fs.first_object due to above
1269                          * conditional
1270                          */
1271                         vm_object_pip_wakeup(fs.object);
1272
1273                         /*
1274                          * We only try to prefault read-only mappings to the
1275                          * neighboring pages when this copy-on-write fault is
1276                          * a hard fault.  In other cases, trying to prefault
1277                          * is typically wasted effort.
1278                          */
1279                         if (faultcount == 0)
1280                                 faultcount = 1;
1281
1282                         /*
1283                          * Only use the new page below...
1284                          */
1285                         fs.object = fs.first_object;
1286                         fs.pindex = fs.first_pindex;
1287                         fs.m = fs.first_m;
1288                         if (!is_first_object_locked)
1289                                 VM_OBJECT_WLOCK(fs.object);
1290                         VM_CNT_INC(v_cow_faults);
1291                         curthread->td_cow++;
1292                 } else {
1293                         prot &= ~VM_PROT_WRITE;
1294                 }
1295         }
1296
1297         /*
1298          * We must verify that the maps have not changed since our last
1299          * lookup.
1300          */
1301         if (!fs.lookup_still_valid) {
1302                 if (!vm_map_trylock_read(fs.map)) {
1303                         unlock_and_deallocate(&fs);
1304                         goto RetryFault;
1305                 }
1306                 fs.lookup_still_valid = true;
1307                 if (fs.map->timestamp != fs.map_generation) {
1308                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1309                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1310
1311                         /*
1312                          * If we don't need the page any longer, put it on the inactive
1313                          * list (the easiest thing to do here).  If no one needs it,
1314                          * pageout will grab it eventually.
1315                          */
1316                         if (result != KERN_SUCCESS) {
1317                                 unlock_and_deallocate(&fs);
1318
1319                                 /*
1320                                  * If retry of map lookup would have blocked then
1321                                  * retry fault from start.
1322                                  */
1323                                 if (result == KERN_FAILURE)
1324                                         goto RetryFault;
1325                                 return (result);
1326                         }
1327                         if ((retry_object != fs.first_object) ||
1328                             (retry_pindex != fs.first_pindex)) {
1329                                 unlock_and_deallocate(&fs);
1330                                 goto RetryFault;
1331                         }
1332
1333                         /*
1334                          * Check whether the protection has changed or the object has
1335                          * been copied while we left the map unlocked. Changing from
1336                          * read to write permission is OK - we leave the page
1337                          * write-protected, and catch the write fault. Changing from
1338                          * write to read permission means that we can't mark the page
1339                          * write-enabled after all.
1340                          */
1341                         prot &= retry_prot;
1342                         fault_type &= retry_prot;
1343                         if (prot == 0) {
1344                                 unlock_and_deallocate(&fs);
1345                                 goto RetryFault;
1346                         }
1347
1348                         /* Reassert because wired may have changed. */
1349                         KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1350                             ("!wired && VM_FAULT_WIRE"));
1351                 }
1352         }
1353
1354         /*
1355          * If the page was filled by a pager, save the virtual address that
1356          * should be faulted on next under a sequential access pattern to the
1357          * map entry.  A read lock on the map suffices to update this address
1358          * safely.
1359          */
1360         if (hardfault)
1361                 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1362
1363         vm_page_assert_xbusied(fs.m);
1364         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags);
1365
1366         /*
1367          * Page must be completely valid or it is not fit to
1368          * map into user space.  vm_pager_get_pages() ensures this.
1369          */
1370         KASSERT(vm_page_all_valid(fs.m),
1371             ("vm_fault: page %p partially invalid", fs.m));
1372         VM_OBJECT_WUNLOCK(fs.object);
1373
1374         /*
1375          * Put this page into the physical map.  We had to do the unlock above
1376          * because pmap_enter() may sleep.  We don't put the page
1377          * back on the active queue until later so that the pageout daemon
1378          * won't find it (yet).
1379          */
1380         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1381             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1382         if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1383             wired == 0)
1384                 vm_fault_prefault(&fs, vaddr,
1385                     faultcount > 0 ? behind : PFBAK,
1386                     faultcount > 0 ? ahead : PFFOR, false);
1387
1388         /*
1389          * If the page is not wired down, then put it where the pageout daemon
1390          * can find it.
1391          */
1392         if ((fault_flags & VM_FAULT_WIRE) != 0)
1393                 vm_page_wire(fs.m);
1394         else
1395                 vm_page_activate(fs.m);
1396         if (m_hold != NULL) {
1397                 *m_hold = fs.m;
1398                 vm_page_wire(fs.m);
1399         }
1400         vm_page_xunbusy(fs.m);
1401         fs.m = NULL;
1402
1403         /*
1404          * Unlock everything, and return
1405          */
1406         fault_deallocate(&fs);
1407         if (hardfault) {
1408                 VM_CNT_INC(v_io_faults);
1409                 curthread->td_ru.ru_majflt++;
1410 #ifdef RACCT
1411                 if (racct_enable && fs.object->type == OBJT_VNODE) {
1412                         PROC_LOCK(curproc);
1413                         if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1414                                 racct_add_force(curproc, RACCT_WRITEBPS,
1415                                     PAGE_SIZE + behind * PAGE_SIZE);
1416                                 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1417                         } else {
1418                                 racct_add_force(curproc, RACCT_READBPS,
1419                                     PAGE_SIZE + ahead * PAGE_SIZE);
1420                                 racct_add_force(curproc, RACCT_READIOPS, 1);
1421                         }
1422                         PROC_UNLOCK(curproc);
1423                 }
1424 #endif
1425         } else
1426                 curthread->td_ru.ru_minflt++;
1427
1428         return (KERN_SUCCESS);
1429 }
1430
1431 /*
1432  * Speed up the reclamation of pages that precede the faulting pindex within
1433  * the first object of the shadow chain.  Essentially, perform the equivalent
1434  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1435  * the faulting pindex by the cluster size when the pages read by vm_fault()
1436  * cross a cluster-size boundary.  The cluster size is the greater of the
1437  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1438  *
1439  * When "fs->first_object" is a shadow object, the pages in the backing object
1440  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1441  * function must only be concerned with pages in the first object.
1442  */
1443 static void
1444 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1445 {
1446         vm_map_entry_t entry;
1447         vm_object_t first_object, object;
1448         vm_offset_t end, start;
1449         vm_page_t m, m_next;
1450         vm_pindex_t pend, pstart;
1451         vm_size_t size;
1452
1453         object = fs->object;
1454         VM_OBJECT_ASSERT_WLOCKED(object);
1455         first_object = fs->first_object;
1456         if (first_object != object) {
1457                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1458                         VM_OBJECT_WUNLOCK(object);
1459                         VM_OBJECT_WLOCK(first_object);
1460                         VM_OBJECT_WLOCK(object);
1461                 }
1462         }
1463         /* Neither fictitious nor unmanaged pages can be reclaimed. */
1464         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1465                 size = VM_FAULT_DONTNEED_MIN;
1466                 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1467                         size = pagesizes[1];
1468                 end = rounddown2(vaddr, size);
1469                 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1470                     (entry = fs->entry)->start < end) {
1471                         if (end - entry->start < size)
1472                                 start = entry->start;
1473                         else
1474                                 start = end - size;
1475                         pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1476                         pstart = OFF_TO_IDX(entry->offset) + atop(start -
1477                             entry->start);
1478                         m_next = vm_page_find_least(first_object, pstart);
1479                         pend = OFF_TO_IDX(entry->offset) + atop(end -
1480                             entry->start);
1481                         while ((m = m_next) != NULL && m->pindex < pend) {
1482                                 m_next = TAILQ_NEXT(m, listq);
1483                                 if (!vm_page_all_valid(m) ||
1484                                     vm_page_busied(m))
1485                                         continue;
1486
1487                                 /*
1488                                  * Don't clear PGA_REFERENCED, since it would
1489                                  * likely represent a reference by a different
1490                                  * process.
1491                                  *
1492                                  * Typically, at this point, prefetched pages
1493                                  * are still in the inactive queue.  Only
1494                                  * pages that triggered page faults are in the
1495                                  * active queue.  The test for whether the page
1496                                  * is in the inactive queue is racy; in the
1497                                  * worst case we will requeue the page
1498                                  * unnecessarily.
1499                                  */
1500                                 if (!vm_page_inactive(m))
1501                                         vm_page_deactivate(m);
1502                         }
1503                 }
1504         }
1505         if (first_object != object)
1506                 VM_OBJECT_WUNLOCK(first_object);
1507 }
1508
1509 /*
1510  * vm_fault_prefault provides a quick way of clustering
1511  * pagefaults into a processes address space.  It is a "cousin"
1512  * of vm_map_pmap_enter, except it runs at page fault time instead
1513  * of mmap time.
1514  */
1515 static void
1516 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1517     int backward, int forward, bool obj_locked)
1518 {
1519         pmap_t pmap;
1520         vm_map_entry_t entry;
1521         vm_object_t backing_object, lobject;
1522         vm_offset_t addr, starta;
1523         vm_pindex_t pindex;
1524         vm_page_t m;
1525         int i;
1526
1527         pmap = fs->map->pmap;
1528         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1529                 return;
1530
1531         entry = fs->entry;
1532
1533         if (addra < backward * PAGE_SIZE) {
1534                 starta = entry->start;
1535         } else {
1536                 starta = addra - backward * PAGE_SIZE;
1537                 if (starta < entry->start)
1538                         starta = entry->start;
1539         }
1540
1541         /*
1542          * Generate the sequence of virtual addresses that are candidates for
1543          * prefaulting in an outward spiral from the faulting virtual address,
1544          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1545          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1546          * If the candidate address doesn't have a backing physical page, then
1547          * the loop immediately terminates.
1548          */
1549         for (i = 0; i < 2 * imax(backward, forward); i++) {
1550                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1551                     PAGE_SIZE);
1552                 if (addr > addra + forward * PAGE_SIZE)
1553                         addr = 0;
1554
1555                 if (addr < starta || addr >= entry->end)
1556                         continue;
1557
1558                 if (!pmap_is_prefaultable(pmap, addr))
1559                         continue;
1560
1561                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1562                 lobject = entry->object.vm_object;
1563                 if (!obj_locked)
1564                         VM_OBJECT_RLOCK(lobject);
1565                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1566                     lobject->type == OBJT_DEFAULT &&
1567                     (backing_object = lobject->backing_object) != NULL) {
1568                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1569                             0, ("vm_fault_prefault: unaligned object offset"));
1570                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1571                         VM_OBJECT_RLOCK(backing_object);
1572                         if (!obj_locked || lobject != entry->object.vm_object)
1573                                 VM_OBJECT_RUNLOCK(lobject);
1574                         lobject = backing_object;
1575                 }
1576                 if (m == NULL) {
1577                         if (!obj_locked || lobject != entry->object.vm_object)
1578                                 VM_OBJECT_RUNLOCK(lobject);
1579                         break;
1580                 }
1581                 if (vm_page_all_valid(m) &&
1582                     (m->flags & PG_FICTITIOUS) == 0)
1583                         pmap_enter_quick(pmap, addr, m, entry->protection);
1584                 if (!obj_locked || lobject != entry->object.vm_object)
1585                         VM_OBJECT_RUNLOCK(lobject);
1586         }
1587 }
1588
1589 /*
1590  * Hold each of the physical pages that are mapped by the specified range of
1591  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1592  * and allow the specified types of access, "prot".  If all of the implied
1593  * pages are successfully held, then the number of held pages is returned
1594  * together with pointers to those pages in the array "ma".  However, if any
1595  * of the pages cannot be held, -1 is returned.
1596  */
1597 int
1598 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1599     vm_prot_t prot, vm_page_t *ma, int max_count)
1600 {
1601         vm_offset_t end, va;
1602         vm_page_t *mp;
1603         int count;
1604         boolean_t pmap_failed;
1605
1606         if (len == 0)
1607                 return (0);
1608         end = round_page(addr + len);
1609         addr = trunc_page(addr);
1610
1611         /*
1612          * Check for illegal addresses.
1613          */
1614         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1615                 return (-1);
1616
1617         if (atop(end - addr) > max_count)
1618                 panic("vm_fault_quick_hold_pages: count > max_count");
1619         count = atop(end - addr);
1620
1621         /*
1622          * Most likely, the physical pages are resident in the pmap, so it is
1623          * faster to try pmap_extract_and_hold() first.
1624          */
1625         pmap_failed = FALSE;
1626         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1627                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1628                 if (*mp == NULL)
1629                         pmap_failed = TRUE;
1630                 else if ((prot & VM_PROT_WRITE) != 0 &&
1631                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1632                         /*
1633                          * Explicitly dirty the physical page.  Otherwise, the
1634                          * caller's changes may go unnoticed because they are
1635                          * performed through an unmanaged mapping or by a DMA
1636                          * operation.
1637                          *
1638                          * The object lock is not held here.
1639                          * See vm_page_clear_dirty_mask().
1640                          */
1641                         vm_page_dirty(*mp);
1642                 }
1643         }
1644         if (pmap_failed) {
1645                 /*
1646                  * One or more pages could not be held by the pmap.  Either no
1647                  * page was mapped at the specified virtual address or that
1648                  * mapping had insufficient permissions.  Attempt to fault in
1649                  * and hold these pages.
1650                  *
1651                  * If vm_fault_disable_pagefaults() was called,
1652                  * i.e., TDP_NOFAULTING is set, we must not sleep nor
1653                  * acquire MD VM locks, which means we must not call
1654                  * vm_fault().  Some (out of tree) callers mark
1655                  * too wide a code area with vm_fault_disable_pagefaults()
1656                  * already, use the VM_PROT_QUICK_NOFAULT flag to request
1657                  * the proper behaviour explicitly.
1658                  */
1659                 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1660                     (curthread->td_pflags & TDP_NOFAULTING) != 0)
1661                         goto error;
1662                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1663                         if (*mp == NULL && vm_fault(map, va, prot,
1664                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1665                                 goto error;
1666         }
1667         return (count);
1668 error:
1669         for (mp = ma; mp < ma + count; mp++)
1670                 if (*mp != NULL)
1671                         vm_page_unwire(*mp, PQ_INACTIVE);
1672         return (-1);
1673 }
1674
1675 /*
1676  *      Routine:
1677  *              vm_fault_copy_entry
1678  *      Function:
1679  *              Create new shadow object backing dst_entry with private copy of
1680  *              all underlying pages. When src_entry is equal to dst_entry,
1681  *              function implements COW for wired-down map entry. Otherwise,
1682  *              it forks wired entry into dst_map.
1683  *
1684  *      In/out conditions:
1685  *              The source and destination maps must be locked for write.
1686  *              The source map entry must be wired down (or be a sharing map
1687  *              entry corresponding to a main map entry that is wired down).
1688  */
1689 void
1690 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1691     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1692     vm_ooffset_t *fork_charge)
1693 {
1694         vm_object_t backing_object, dst_object, object, src_object;
1695         vm_pindex_t dst_pindex, pindex, src_pindex;
1696         vm_prot_t access, prot;
1697         vm_offset_t vaddr;
1698         vm_page_t dst_m;
1699         vm_page_t src_m;
1700         boolean_t upgrade;
1701
1702 #ifdef  lint
1703         src_map++;
1704 #endif  /* lint */
1705
1706         upgrade = src_entry == dst_entry;
1707         access = prot = dst_entry->protection;
1708
1709         src_object = src_entry->object.vm_object;
1710         src_pindex = OFF_TO_IDX(src_entry->offset);
1711
1712         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1713                 dst_object = src_object;
1714                 vm_object_reference(dst_object);
1715         } else {
1716                 /*
1717                  * Create the top-level object for the destination entry.
1718                  * Doesn't actually shadow anything - we copy the pages
1719                  * directly.
1720                  */
1721                 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1722                     dst_entry->start), NULL, NULL, 0);
1723 #if VM_NRESERVLEVEL > 0
1724                 dst_object->flags |= OBJ_COLORED;
1725                 dst_object->pg_color = atop(dst_entry->start);
1726 #endif
1727                 dst_object->domain = src_object->domain;
1728                 dst_object->charge = dst_entry->end - dst_entry->start;
1729         }
1730
1731         VM_OBJECT_WLOCK(dst_object);
1732         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1733             ("vm_fault_copy_entry: vm_object not NULL"));
1734         if (src_object != dst_object) {
1735                 dst_entry->object.vm_object = dst_object;
1736                 dst_entry->offset = 0;
1737                 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1738         }
1739         if (fork_charge != NULL) {
1740                 KASSERT(dst_entry->cred == NULL,
1741                     ("vm_fault_copy_entry: leaked swp charge"));
1742                 dst_object->cred = curthread->td_ucred;
1743                 crhold(dst_object->cred);
1744                 *fork_charge += dst_object->charge;
1745         } else if ((dst_object->type == OBJT_DEFAULT ||
1746             dst_object->type == OBJT_SWAP) &&
1747             dst_object->cred == NULL) {
1748                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1749                     dst_entry));
1750                 dst_object->cred = dst_entry->cred;
1751                 dst_entry->cred = NULL;
1752         }
1753
1754         /*
1755          * If not an upgrade, then enter the mappings in the pmap as
1756          * read and/or execute accesses.  Otherwise, enter them as
1757          * write accesses.
1758          *
1759          * A writeable large page mapping is only created if all of
1760          * the constituent small page mappings are modified. Marking
1761          * PTEs as modified on inception allows promotion to happen
1762          * without taking potentially large number of soft faults.
1763          */
1764         if (!upgrade)
1765                 access &= ~VM_PROT_WRITE;
1766
1767         /*
1768          * Loop through all of the virtual pages within the entry's
1769          * range, copying each page from the source object to the
1770          * destination object.  Since the source is wired, those pages
1771          * must exist.  In contrast, the destination is pageable.
1772          * Since the destination object doesn't share any backing storage
1773          * with the source object, all of its pages must be dirtied,
1774          * regardless of whether they can be written.
1775          */
1776         for (vaddr = dst_entry->start, dst_pindex = 0;
1777             vaddr < dst_entry->end;
1778             vaddr += PAGE_SIZE, dst_pindex++) {
1779 again:
1780                 /*
1781                  * Find the page in the source object, and copy it in.
1782                  * Because the source is wired down, the page will be
1783                  * in memory.
1784                  */
1785                 if (src_object != dst_object)
1786                         VM_OBJECT_RLOCK(src_object);
1787                 object = src_object;
1788                 pindex = src_pindex + dst_pindex;
1789                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1790                     (backing_object = object->backing_object) != NULL) {
1791                         /*
1792                          * Unless the source mapping is read-only or
1793                          * it is presently being upgraded from
1794                          * read-only, the first object in the shadow
1795                          * chain should provide all of the pages.  In
1796                          * other words, this loop body should never be
1797                          * executed when the source mapping is already
1798                          * read/write.
1799                          */
1800                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1801                             upgrade,
1802                             ("vm_fault_copy_entry: main object missing page"));
1803
1804                         VM_OBJECT_RLOCK(backing_object);
1805                         pindex += OFF_TO_IDX(object->backing_object_offset);
1806                         if (object != dst_object)
1807                                 VM_OBJECT_RUNLOCK(object);
1808                         object = backing_object;
1809                 }
1810                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1811
1812                 if (object != dst_object) {
1813                         /*
1814                          * Allocate a page in the destination object.
1815                          */
1816                         dst_m = vm_page_alloc(dst_object, (src_object ==
1817                             dst_object ? src_pindex : 0) + dst_pindex,
1818                             VM_ALLOC_NORMAL);
1819                         if (dst_m == NULL) {
1820                                 VM_OBJECT_WUNLOCK(dst_object);
1821                                 VM_OBJECT_RUNLOCK(object);
1822                                 vm_wait(dst_object);
1823                                 VM_OBJECT_WLOCK(dst_object);
1824                                 goto again;
1825                         }
1826                         pmap_copy_page(src_m, dst_m);
1827                         VM_OBJECT_RUNLOCK(object);
1828                         dst_m->dirty = dst_m->valid = src_m->valid;
1829                 } else {
1830                         dst_m = src_m;
1831                         if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1832                                 goto again;
1833                         if (dst_m->pindex >= dst_object->size) {
1834                                 /*
1835                                  * We are upgrading.  Index can occur
1836                                  * out of bounds if the object type is
1837                                  * vnode and the file was truncated.
1838                                  */
1839                                 vm_page_xunbusy(dst_m);
1840                                 break;
1841                         }
1842                 }
1843                 VM_OBJECT_WUNLOCK(dst_object);
1844
1845                 /*
1846                  * Enter it in the pmap. If a wired, copy-on-write
1847                  * mapping is being replaced by a write-enabled
1848                  * mapping, then wire that new mapping.
1849                  *
1850                  * The page can be invalid if the user called
1851                  * msync(MS_INVALIDATE) or truncated the backing vnode
1852                  * or shared memory object.  In this case, do not
1853                  * insert it into pmap, but still do the copy so that
1854                  * all copies of the wired map entry have similar
1855                  * backing pages.
1856                  */
1857                 if (vm_page_all_valid(dst_m)) {
1858                         pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1859                             access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1860                 }
1861
1862                 /*
1863                  * Mark it no longer busy, and put it on the active list.
1864                  */
1865                 VM_OBJECT_WLOCK(dst_object);
1866
1867                 if (upgrade) {
1868                         if (src_m != dst_m) {
1869                                 vm_page_unwire(src_m, PQ_INACTIVE);
1870                                 vm_page_wire(dst_m);
1871                         } else {
1872                                 KASSERT(vm_page_wired(dst_m),
1873                                     ("dst_m %p is not wired", dst_m));
1874                         }
1875                 } else {
1876                         vm_page_activate(dst_m);
1877                 }
1878                 vm_page_xunbusy(dst_m);
1879         }
1880         VM_OBJECT_WUNLOCK(dst_object);
1881         if (upgrade) {
1882                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1883                 vm_object_deallocate(src_object);
1884         }
1885 }
1886
1887 /*
1888  * Block entry into the machine-independent layer's page fault handler by
1889  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1890  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1891  * spurious page faults.
1892  */
1893 int
1894 vm_fault_disable_pagefaults(void)
1895 {
1896
1897         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1898 }
1899
1900 void
1901 vm_fault_enable_pagefaults(int save)
1902 {
1903
1904         curthread_pflags_restore(save);
1905 }