sys/vm/vm_fault.c

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