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/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/rwlock.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vm_extern.h>
 104
 105 #define PFBAK 4
 106 #define PFFOR 4
 107
 108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 109
 110 #define VM_FAULT_READ_BEHIND    8
 111 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 112 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 113 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 114 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 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 lookup_still_valid;
 126         struct vnode *vp;
 127 };
 128
 129 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 130 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 131             int faultcount, int reqpage);
 132
 133 static inline void
 134 release_page(struct faultstate *fs)
 135 {
 136
 137         vm_page_xunbusy(fs->m);
 138         vm_page_lock(fs->m);
 139         vm_page_deactivate(fs->m);
 140         vm_page_unlock(fs->m);
 141         fs->m = NULL;
 142 }
 143
 144 static inline void
 145 unlock_map(struct faultstate *fs)
 146 {
 147
 148         if (fs->lookup_still_valid) {
 149                 vm_map_lookup_done(fs->map, fs->entry);
 150                 fs->lookup_still_valid = FALSE;
 151         }
 152 }
 153
 154 static void
 155 unlock_and_deallocate(struct faultstate *fs)
 156 {
 157
 158         vm_object_pip_wakeup(fs->object);
 159         VM_OBJECT_WUNLOCK(fs->object);
 160         if (fs->object != fs->first_object) {
 161                 VM_OBJECT_WLOCK(fs->first_object);
 162                 vm_page_lock(fs->first_m);
 163                 vm_page_free(fs->first_m);
 164                 vm_page_unlock(fs->first_m);
 165                 vm_object_pip_wakeup(fs->first_object);
 166                 VM_OBJECT_WUNLOCK(fs->first_object);
 167                 fs->first_m = NULL;
 168         }
 169         vm_object_deallocate(fs->first_object);
 170         unlock_map(fs);
 171         if (fs->vp != NULL) {
 172                 vput(fs->vp);
 173                 fs->vp = NULL;
 174         }
 175 }
 176
 177 static void
 178 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
 179     vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
 180 {
 181         boolean_t need_dirty;
 182
 183         if (((prot & VM_PROT_WRITE) == 0 &&
 184             (fault_flags & VM_FAULT_DIRTY) == 0) ||
 185             (m->oflags & VPO_UNMANAGED) != 0)
 186                 return;
 187
 188         VM_OBJECT_ASSERT_LOCKED(m->object);
 189
 190         need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
 191             (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 192             (fault_flags & VM_FAULT_DIRTY) != 0;
 193
 194         if (set_wd)
 195                 vm_object_set_writeable_dirty(m->object);
 196         else
 197                 /*
 198                  * If two callers of vm_fault_dirty() with set_wd ==
 199                  * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
 200                  * flag set, other with flag clear, race, it is
 201                  * possible for the no-NOSYNC thread to see m->dirty
 202                  * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
 203                  * around manipulation of VPO_NOSYNC and
 204                  * vm_page_dirty() call, to avoid the race and keep
 205                  * m->oflags consistent.
 206                  */
 207                 vm_page_lock(m);
 208
 209         /*
 210          * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 211          * if the page is already dirty to prevent data written with
 212          * the expectation of being synced from not being synced.
 213          * Likewise if this entry does not request NOSYNC then make
 214          * sure the page isn't marked NOSYNC.  Applications sharing
 215          * data should use the same flags to avoid ping ponging.
 216          */
 217         if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
 218                 if (m->dirty == 0) {
 219                         m->oflags |= VPO_NOSYNC;
 220                 }
 221         } else {
 222                 m->oflags &= ~VPO_NOSYNC;
 223         }
 224
 225         /*
 226          * If the fault is a write, we know that this page is being
 227          * written NOW so dirty it explicitly to save on
 228          * pmap_is_modified() calls later.
 229          *
 230          * Also tell the backing pager, if any, that it should remove
 231          * any swap backing since the page is now dirty.
 232          */
 233         if (need_dirty)
 234                 vm_page_dirty(m);
 235         if (!set_wd)
 236                 vm_page_unlock(m);
 237         if (need_dirty)
 238                 vm_pager_page_unswapped(m);
 239 }
 240
 241 /*
 242  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 243  *            current object *might* contain the page.
 244  *
 245  *            default objects are zero-fill, there is no real pager.
 246  */
 247 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 248                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 249
 250 /*
 251  *      vm_fault:
 252  *
 253  *      Handle a page fault occurring at the given address,
 254  *      requiring the given permissions, in the map specified.
 255  *      If successful, the page is inserted into the
 256  *      associated physical map.
 257  *
 258  *      NOTE: the given address should be truncated to the
 259  *      proper page address.
 260  *
 261  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 262  *      a standard error specifying why the fault is fatal is returned.
 263  *
 264  *      The map in question must be referenced, and remains so.
 265  *      Caller may hold no locks.
 266  */
 267 int
 268 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 269     int fault_flags)
 270 {
 271         struct thread *td;
 272         int result;
 273
 274         td = curthread;
 275         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 276                 return (KERN_PROTECTION_FAILURE);
 277 #ifdef KTRACE
 278         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 279                 ktrfault(vaddr, fault_type);
 280 #endif
 281         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 282             NULL);
 283 #ifdef KTRACE
 284         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 285                 ktrfaultend(result);
 286 #endif
 287         return (result);
 288 }
 289
 290 int
 291 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 292     int fault_flags, vm_page_t *m_hold)
 293 {
 294         vm_prot_t prot;
 295         long ahead, behind;
 296         int alloc_req, era, faultcount, nera, reqpage, result;
 297         boolean_t growstack, is_first_object_locked, wired;
 298         int map_generation;
 299         vm_object_t next_object;
 300         vm_page_t marray[VM_FAULT_READ_MAX];
 301         int hardfault;
 302         struct faultstate fs;
 303         struct vnode *vp;
 304         vm_page_t m;
 305         int locked, error;
 306
 307         hardfault = 0;
 308         growstack = TRUE;
 309         PCPU_INC(cnt.v_vm_faults);
 310         fs.vp = NULL;
 311         faultcount = reqpage = 0;
 312
 313 RetryFault:;
 314
 315         /*
 316          * Find the backing store object and offset into it to begin the
 317          * search.
 318          */
 319         fs.map = map;
 320         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 321             &fs.first_object, &fs.first_pindex, &prot, &wired);
 322         if (result != KERN_SUCCESS) {
 323                 if (growstack && result == KERN_INVALID_ADDRESS &&
 324                     map != kernel_map) {
 325                         result = vm_map_growstack(curproc, vaddr);
 326                         if (result != KERN_SUCCESS)
 327                                 return (KERN_FAILURE);
 328                         growstack = FALSE;
 329                         goto RetryFault;
 330                 }
 331                 return (result);
 332         }
 333
 334         map_generation = fs.map->timestamp;
 335
 336         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 337                 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
 338                         vm_map_unlock_read(fs.map);
 339                         return (KERN_FAILURE);
 340                 }
 341                 panic("vm_fault: fault on nofault entry, addr: %lx",
 342                     (u_long)vaddr);
 343         }
 344
 345         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 346             fs.entry->wiring_thread != curthread) {
 347                 vm_map_unlock_read(fs.map);
 348                 vm_map_lock(fs.map);
 349                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 350                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 351                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 352                         vm_map_unlock_and_wait(fs.map, 0);
 353                 } else
 354                         vm_map_unlock(fs.map);
 355                 goto RetryFault;
 356         }
 357
 358         if (wired)
 359                 fault_type = prot | (fault_type & VM_PROT_COPY);
 360
 361         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 362             (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
 363             /* avoid calling vm_object_set_writeable_dirty() */
 364             ((prot & VM_PROT_WRITE) == 0 ||
 365             fs.first_object->type != OBJT_VNODE ||
 366             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 367                 VM_OBJECT_RLOCK(fs.first_object);
 368                 if ((prot & VM_PROT_WRITE) != 0 &&
 369                     fs.first_object->type == OBJT_VNODE &&
 370                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
 371                         goto fast_failed;
 372                 m = vm_page_lookup(fs.first_object, fs.first_pindex);
 373                 /* A busy page can be mapped for read|execute access. */
 374                 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 375                     vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 376                         goto fast_failed;
 377                 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
 378                    fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
 379                    0), 0);
 380                 if (result != KERN_SUCCESS)
 381                         goto fast_failed;
 382                 if (m_hold != NULL) {
 383                         *m_hold = m;
 384                         vm_page_lock(m);
 385                         vm_page_hold(m);
 386                         vm_page_unlock(m);
 387                 }
 388                 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
 389                     FALSE);
 390                 VM_OBJECT_RUNLOCK(fs.first_object);
 391                 if (!wired)
 392                         vm_fault_prefault(&fs, vaddr, 0, 0);
 393                 vm_map_lookup_done(fs.map, fs.entry);
 394                 curthread->td_ru.ru_minflt++;
 395                 return (KERN_SUCCESS);
 396 fast_failed:
 397                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 398                         VM_OBJECT_RUNLOCK(fs.first_object);
 399                         VM_OBJECT_WLOCK(fs.first_object);
 400                 }
 401         } else {
 402                 VM_OBJECT_WLOCK(fs.first_object);
 403         }
 404
 405         /*
 406          * Make a reference to this object to prevent its disposal while we
 407          * are messing with it.  Once we have the reference, the map is free
 408          * to be diddled.  Since objects reference their shadows (and copies),
 409          * they will stay around as well.
 410          *
 411          * Bump the paging-in-progress count to prevent size changes (e.g.
 412          * truncation operations) during I/O.  This must be done after
 413          * obtaining the vnode lock in order to avoid possible deadlocks.
 414          */
 415         vm_object_reference_locked(fs.first_object);
 416         vm_object_pip_add(fs.first_object, 1);
 417
 418         fs.lookup_still_valid = TRUE;
 419
 420         fs.first_m = NULL;
 421
 422         /*
 423          * Search for the page at object/offset.
 424          */
 425         fs.object = fs.first_object;
 426         fs.pindex = fs.first_pindex;
 427         while (TRUE) {
 428                 /*
 429                  * If the object is dead, we stop here
 430                  */
 431                 if (fs.object->flags & OBJ_DEAD) {
 432                         unlock_and_deallocate(&fs);
 433                         return (KERN_PROTECTION_FAILURE);
 434                 }
 435
 436                 /*
 437                  * See if page is resident
 438                  */
 439                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 440                 if (fs.m != NULL) {
 441                         /*
 442                          * Wait/Retry if the page is busy.  We have to do this
 443                          * if the page is either exclusive or shared busy
 444                          * because the vm_pager may be using read busy for
 445                          * pageouts (and even pageins if it is the vnode
 446                          * pager), and we could end up trying to pagein and
 447                          * pageout the same page simultaneously.
 448                          *
 449                          * We can theoretically allow the busy case on a read
 450                          * fault if the page is marked valid, but since such
 451                          * pages are typically already pmap'd, putting that
 452                          * special case in might be more effort then it is
 453                          * worth.  We cannot under any circumstances mess
 454                          * around with a shared busied page except, perhaps,
 455                          * to pmap it.
 456                          */
 457                         if (vm_page_busied(fs.m)) {
 458                                 /*
 459                                  * Reference the page before unlocking and
 460                                  * sleeping so that the page daemon is less
 461                                  * likely to reclaim it.
 462                                  */
 463                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 464                                 if (fs.object != fs.first_object) {
 465                                         if (!VM_OBJECT_TRYWLOCK(
 466                                             fs.first_object)) {
 467                                                 VM_OBJECT_WUNLOCK(fs.object);
 468                                                 VM_OBJECT_WLOCK(fs.first_object);
 469                                                 VM_OBJECT_WLOCK(fs.object);
 470                                         }
 471                                         vm_page_lock(fs.first_m);
 472                                         vm_page_free(fs.first_m);
 473                                         vm_page_unlock(fs.first_m);
 474                                         vm_object_pip_wakeup(fs.first_object);
 475                                         VM_OBJECT_WUNLOCK(fs.first_object);
 476                                         fs.first_m = NULL;
 477                                 }
 478                                 unlock_map(&fs);
 479                                 if (fs.m == vm_page_lookup(fs.object,
 480                                     fs.pindex)) {
 481                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 482                                 }
 483                                 vm_object_pip_wakeup(fs.object);
 484                                 VM_OBJECT_WUNLOCK(fs.object);
 485                                 PCPU_INC(cnt.v_intrans);
 486                                 vm_object_deallocate(fs.first_object);
 487                                 goto RetryFault;
 488                         }
 489                         vm_page_lock(fs.m);
 490                         vm_page_remque(fs.m);
 491                         vm_page_unlock(fs.m);
 492
 493                         /*
 494                          * Mark page busy for other processes, and the
 495                          * pagedaemon.  If it still isn't completely valid
 496                          * (readable), jump to readrest, else break-out ( we
 497                          * found the page ).
 498                          */
 499                         vm_page_xbusy(fs.m);
 500                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 501                                 goto readrest;
 502                         break;
 503                 }
 504
 505                 /*
 506                  * Page is not resident, If this is the search termination
 507                  * or the pager might contain the page, allocate a new page.
 508                  */
 509                 if (TRYPAGER || fs.object == fs.first_object) {
 510                         if (fs.pindex >= fs.object->size) {
 511                                 unlock_and_deallocate(&fs);
 512                                 return (KERN_PROTECTION_FAILURE);
 513                         }
 514
 515                         /*
 516                          * Allocate a new page for this object/offset pair.
 517                          *
 518                          * Unlocked read of the p_flag is harmless. At
 519                          * worst, the P_KILLED might be not observed
 520                          * there, and allocation can fail, causing
 521                          * restart and new reading of the p_flag.
 522                          */
 523                         fs.m = NULL;
 524                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 525 #if VM_NRESERVLEVEL > 0
 526                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 527                                         fs.object->flags |= OBJ_COLORED;
 528                                         fs.object->pg_color = atop(vaddr) -
 529                                             fs.pindex;
 530                                 }
 531 #endif
 532                                 alloc_req = P_KILLED(curproc) ?
 533                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 534                                 if (fs.object->type != OBJT_VNODE &&
 535                                     fs.object->backing_object == NULL)
 536                                         alloc_req |= VM_ALLOC_ZERO;
 537                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 538                                     alloc_req);
 539                         }
 540                         if (fs.m == NULL) {
 541                                 unlock_and_deallocate(&fs);
 542                                 VM_WAITPFAULT;
 543                                 goto RetryFault;
 544                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 545                                 break;
 546                 }
 547
 548 readrest:
 549                 /*
 550                  * We have found a valid page or we have allocated a new page.
 551                  * The page thus may not be valid or may not be entirely
 552                  * valid.
 553                  *
 554                  * Attempt to fault-in the page if there is a chance that the
 555                  * pager has it, and potentially fault in additional pages
 556                  * at the same time.
 557                  */
 558                 if (TRYPAGER) {
 559                         int rv;
 560                         u_char behavior = vm_map_entry_behavior(fs.entry);
 561
 562                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 563                             P_KILLED(curproc)) {
 564                                 behind = 0;
 565                                 ahead = 0;
 566                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 567                                 behind = 0;
 568                                 ahead = atop(fs.entry->end - vaddr) - 1;
 569                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 570                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 571                                 if (fs.pindex == fs.entry->next_read)
 572                                         vm_fault_cache_behind(&fs,
 573                                             VM_FAULT_READ_MAX);
 574                         } else {
 575                                 /*
 576                                  * If this is a sequential page fault, then
 577                                  * arithmetically increase the number of pages
 578                                  * in the read-ahead window.  Otherwise, reset
 579                                  * the read-ahead window to its smallest size.
 580                                  */
 581                                 behind = atop(vaddr - fs.entry->start);
 582                                 if (behind > VM_FAULT_READ_BEHIND)
 583                                         behind = VM_FAULT_READ_BEHIND;
 584                                 ahead = atop(fs.entry->end - vaddr) - 1;
 585                                 era = fs.entry->read_ahead;
 586                                 if (fs.pindex == fs.entry->next_read) {
 587                                         nera = era + behind;
 588                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 589                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 590                                         behind = 0;
 591                                         if (ahead > nera)
 592                                                 ahead = nera;
 593                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 594                                                 vm_fault_cache_behind(&fs,
 595                                                     VM_FAULT_CACHE_BEHIND);
 596                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 597                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 598                                 if (era != ahead)
 599                                         fs.entry->read_ahead = ahead;
 600                         }
 601
 602                         /*
 603                          * Call the pager to retrieve the data, if any, after
 604                          * releasing the lock on the map.  We hold a ref on
 605                          * fs.object and the pages are exclusive busied.
 606                          */
 607                         unlock_map(&fs);
 608
 609                         if (fs.object->type == OBJT_VNODE) {
 610                                 vp = fs.object->handle;
 611                                 if (vp == fs.vp)
 612                                         goto vnode_locked;
 613                                 else if (fs.vp != NULL) {
 614                                         vput(fs.vp);
 615                                         fs.vp = NULL;
 616                                 }
 617                                 locked = VOP_ISLOCKED(vp);
 618
 619                                 if (locked != LK_EXCLUSIVE)
 620                                         locked = LK_SHARED;
 621                                 /* Do not sleep for vnode lock while fs.m is busy */
 622                                 error = vget(vp, locked | LK_CANRECURSE |
 623                                     LK_NOWAIT, curthread);
 624                                 if (error != 0) {
 625                                         vhold(vp);
 626                                         release_page(&fs);
 627                                         unlock_and_deallocate(&fs);
 628                                         error = vget(vp, locked | LK_RETRY |
 629                                             LK_CANRECURSE, curthread);
 630                                         vdrop(vp);
 631                                         fs.vp = vp;
 632                                         KASSERT(error == 0,
 633                                             ("vm_fault: vget failed"));
 634                                         goto RetryFault;
 635                                 }
 636                                 fs.vp = vp;
 637                         }
 638 vnode_locked:
 639                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 640                             ("vm_fault: vnode-backed object mapped by system map"));
 641
 642                         /*
 643                          * now we find out if any other pages should be paged
 644                          * in at this time this routine checks to see if the
 645                          * pages surrounding this fault reside in the same
 646                          * object as the page for this fault.  If they do,
 647                          * then they are faulted in also into the object.  The
 648                          * array "marray" returned contains an array of
 649                          * vm_page_t structs where one of them is the
 650                          * vm_page_t passed to the routine.  The reqpage
 651                          * return value is the index into the marray for the
 652                          * vm_page_t passed to the routine.
 653                          *
 654                          * fs.m plus the additional pages are exclusive busied.
 655                          */
 656                         faultcount = vm_fault_additional_pages(
 657                             fs.m, behind, ahead, marray, &reqpage);
 658
 659                         rv = faultcount ?
 660                             vm_pager_get_pages(fs.object, marray, faultcount,
 661                                 reqpage) : VM_PAGER_FAIL;
 662
 663                         if (rv == VM_PAGER_OK) {
 664                                 /*
 665                                  * Found the page. Leave it busy while we play
 666                                  * with it.
 667                                  */
 668
 669                                 /*
 670                                  * Relookup in case pager changed page. Pager
 671                                  * is responsible for disposition of old page
 672                                  * if moved.
 673                                  */
 674                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 675                                 if (!fs.m) {
 676                                         unlock_and_deallocate(&fs);
 677                                         goto RetryFault;
 678                                 }
 679
 680                                 hardfault++;
 681                                 break; /* break to PAGE HAS BEEN FOUND */
 682                         }
 683                         /*
 684                          * Remove the bogus page (which does not exist at this
 685                          * object/offset); before doing so, we must get back
 686                          * our object lock to preserve our invariant.
 687                          *
 688                          * Also wake up any other process that may want to bring
 689                          * in this page.
 690                          *
 691                          * If this is the top-level object, we must leave the
 692                          * busy page to prevent another process from rushing
 693                          * past us, and inserting the page in that object at
 694                          * the same time that we are.
 695                          */
 696                         if (rv == VM_PAGER_ERROR)
 697                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 698                                     curproc->p_pid, curproc->p_comm);
 699                         /*
 700                          * Data outside the range of the pager or an I/O error
 701                          */
 702                         /*
 703                          * XXX - the check for kernel_map is a kludge to work
 704                          * around having the machine panic on a kernel space
 705                          * fault w/ I/O error.
 706                          */
 707                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 708                                 (rv == VM_PAGER_BAD)) {
 709                                 vm_page_lock(fs.m);
 710                                 vm_page_free(fs.m);
 711                                 vm_page_unlock(fs.m);
 712                                 fs.m = NULL;
 713                                 unlock_and_deallocate(&fs);
 714                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 715                         }
 716                         if (fs.object != fs.first_object) {
 717                                 vm_page_lock(fs.m);
 718                                 vm_page_free(fs.m);
 719                                 vm_page_unlock(fs.m);
 720                                 fs.m = NULL;
 721                                 /*
 722                                  * XXX - we cannot just fall out at this
 723                                  * point, m has been freed and is invalid!
 724                                  */
 725                         }
 726                 }
 727
 728                 /*
 729                  * We get here if the object has default pager (or unwiring)
 730                  * or the pager doesn't have the page.
 731                  */
 732                 if (fs.object == fs.first_object)
 733                         fs.first_m = fs.m;
 734
 735                 /*
 736                  * Move on to the next object.  Lock the next object before
 737                  * unlocking the current one.
 738                  */
 739                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 740                 next_object = fs.object->backing_object;
 741                 if (next_object == NULL) {
 742                         /*
 743                          * If there's no object left, fill the page in the top
 744                          * object with zeros.
 745                          */
 746                         if (fs.object != fs.first_object) {
 747                                 vm_object_pip_wakeup(fs.object);
 748                                 VM_OBJECT_WUNLOCK(fs.object);
 749
 750                                 fs.object = fs.first_object;
 751                                 fs.pindex = fs.first_pindex;
 752                                 fs.m = fs.first_m;
 753                                 VM_OBJECT_WLOCK(fs.object);
 754                         }
 755                         fs.first_m = NULL;
 756
 757                         /*
 758                          * Zero the page if necessary and mark it valid.
 759                          */
 760                         if ((fs.m->flags & PG_ZERO) == 0) {
 761                                 pmap_zero_page(fs.m);
 762                         } else {
 763                                 PCPU_INC(cnt.v_ozfod);
 764                         }
 765                         PCPU_INC(cnt.v_zfod);
 766                         fs.m->valid = VM_PAGE_BITS_ALL;
 767                         /* Don't try to prefault neighboring pages. */
 768                         faultcount = 1;
 769                         break;  /* break to PAGE HAS BEEN FOUND */
 770                 } else {
 771                         KASSERT(fs.object != next_object,
 772                             ("object loop %p", next_object));
 773                         VM_OBJECT_WLOCK(next_object);
 774                         vm_object_pip_add(next_object, 1);
 775                         if (fs.object != fs.first_object)
 776                                 vm_object_pip_wakeup(fs.object);
 777                         VM_OBJECT_WUNLOCK(fs.object);
 778                         fs.object = next_object;
 779                 }
 780         }
 781
 782         vm_page_assert_xbusied(fs.m);
 783
 784         /*
 785          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 786          * is held.]
 787          */
 788
 789         /*
 790          * If the page is being written, but isn't already owned by the
 791          * top-level object, we have to copy it into a new page owned by the
 792          * top-level object.
 793          */
 794         if (fs.object != fs.first_object) {
 795                 /*
 796                  * We only really need to copy if we want to write it.
 797                  */
 798                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 799                         /*
 800                          * This allows pages to be virtually copied from a
 801                          * backing_object into the first_object, where the
 802                          * backing object has no other refs to it, and cannot
 803                          * gain any more refs.  Instead of a bcopy, we just
 804                          * move the page from the backing object to the
 805                          * first object.  Note that we must mark the page
 806                          * dirty in the first object so that it will go out
 807                          * to swap when needed.
 808                          */
 809                         is_first_object_locked = FALSE;
 810                         if (
 811                                 /*
 812                                  * Only one shadow object
 813                                  */
 814                                 (fs.object->shadow_count == 1) &&
 815                                 /*
 816                                  * No COW refs, except us
 817                                  */
 818                                 (fs.object->ref_count == 1) &&
 819                                 /*
 820                                  * No one else can look this object up
 821                                  */
 822                                 (fs.object->handle == NULL) &&
 823                                 /*
 824                                  * No other ways to look the object up
 825                                  */
 826                                 ((fs.object->type == OBJT_DEFAULT) ||
 827                                  (fs.object->type == OBJT_SWAP)) &&
 828                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 829                                 /*
 830                                  * We don't chase down the shadow chain
 831                                  */
 832                             fs.object == fs.first_object->backing_object) {
 833                                 /*
 834                                  * get rid of the unnecessary page
 835                                  */
 836                                 vm_page_lock(fs.first_m);
 837                                 vm_page_free(fs.first_m);
 838                                 vm_page_unlock(fs.first_m);
 839                                 /*
 840                                  * grab the page and put it into the
 841                                  * process'es object.  The page is
 842                                  * automatically made dirty.
 843                                  */
 844                                 if (vm_page_rename(fs.m, fs.first_object,
 845                                     fs.first_pindex)) {
 846                                         unlock_and_deallocate(&fs);
 847                                         goto RetryFault;
 848                                 }
 849                                 vm_page_xbusy(fs.m);
 850                                 fs.first_m = fs.m;
 851                                 fs.m = NULL;
 852                                 PCPU_INC(cnt.v_cow_optim);
 853                         } else {
 854                                 /*
 855                                  * Oh, well, lets copy it.
 856                                  */
 857                                 pmap_copy_page(fs.m, fs.first_m);
 858                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 859                                 if (wired && (fault_flags &
 860                                     VM_FAULT_CHANGE_WIRING) == 0) {
 861                                         vm_page_lock(fs.first_m);
 862                                         vm_page_wire(fs.first_m);
 863                                         vm_page_unlock(fs.first_m);
 864
 865                                         vm_page_lock(fs.m);
 866                                         vm_page_unwire(fs.m, FALSE);
 867                                         vm_page_unlock(fs.m);
 868                                 }
 869                                 /*
 870                                  * We no longer need the old page or object.
 871                                  */
 872                                 release_page(&fs);
 873                         }
 874                         /*
 875                          * fs.object != fs.first_object due to above
 876                          * conditional
 877                          */
 878                         vm_object_pip_wakeup(fs.object);
 879                         VM_OBJECT_WUNLOCK(fs.object);
 880                         /*
 881                          * Only use the new page below...
 882                          */
 883                         fs.object = fs.first_object;
 884                         fs.pindex = fs.first_pindex;
 885                         fs.m = fs.first_m;
 886                         if (!is_first_object_locked)
 887                                 VM_OBJECT_WLOCK(fs.object);
 888                         PCPU_INC(cnt.v_cow_faults);
 889                         curthread->td_cow++;
 890                 } else {
 891                         prot &= ~VM_PROT_WRITE;
 892                 }
 893         }
 894
 895         /*
 896          * We must verify that the maps have not changed since our last
 897          * lookup.
 898          */
 899         if (!fs.lookup_still_valid) {
 900                 vm_object_t retry_object;
 901                 vm_pindex_t retry_pindex;
 902                 vm_prot_t retry_prot;
 903
 904                 if (!vm_map_trylock_read(fs.map)) {
 905                         release_page(&fs);
 906                         unlock_and_deallocate(&fs);
 907                         goto RetryFault;
 908                 }
 909                 fs.lookup_still_valid = TRUE;
 910                 if (fs.map->timestamp != map_generation) {
 911                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 912                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 913
 914                         /*
 915                          * If we don't need the page any longer, put it on the inactive
 916                          * list (the easiest thing to do here).  If no one needs it,
 917                          * pageout will grab it eventually.
 918                          */
 919                         if (result != KERN_SUCCESS) {
 920                                 release_page(&fs);
 921                                 unlock_and_deallocate(&fs);
 922
 923                                 /*
 924                                  * If retry of map lookup would have blocked then
 925                                  * retry fault from start.
 926                                  */
 927                                 if (result == KERN_FAILURE)
 928                                         goto RetryFault;
 929                                 return (result);
 930                         }
 931                         if ((retry_object != fs.first_object) ||
 932                             (retry_pindex != fs.first_pindex)) {
 933                                 release_page(&fs);
 934                                 unlock_and_deallocate(&fs);
 935                                 goto RetryFault;
 936                         }
 937
 938                         /*
 939                          * Check whether the protection has changed or the object has
 940                          * been copied while we left the map unlocked. Changing from
 941                          * read to write permission is OK - we leave the page
 942                          * write-protected, and catch the write fault. Changing from
 943                          * write to read permission means that we can't mark the page
 944                          * write-enabled after all.
 945                          */
 946                         prot &= retry_prot;
 947                 }
 948         }
 949         /*
 950          * If the page was filled by a pager, update the map entry's
 951          * last read offset.  Since the pager does not return the
 952          * actual set of pages that it read, this update is based on
 953          * the requested set.  Typically, the requested and actual
 954          * sets are the same.
 955          *
 956          * XXX The following assignment modifies the map
 957          * without holding a write lock on it.
 958          */
 959         if (hardfault)
 960                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 961
 962         vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
 963         vm_page_assert_xbusied(fs.m);
 964
 965         /*
 966          * Page must be completely valid or it is not fit to
 967          * map into user space.  vm_pager_get_pages() ensures this.
 968          */
 969         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 970             ("vm_fault: page %p partially invalid", fs.m));
 971         VM_OBJECT_WUNLOCK(fs.object);
 972
 973         /*
 974          * Put this page into the physical map.  We had to do the unlock above
 975          * because pmap_enter() may sleep.  We don't put the page
 976          * back on the active queue until later so that the pageout daemon
 977          * won't find it (yet).
 978          */
 979         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
 980             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
 981         if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
 982             wired == 0)
 983                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
 984         VM_OBJECT_WLOCK(fs.object);
 985         vm_page_lock(fs.m);
 986
 987         /*
 988          * If the page is not wired down, then put it where the pageout daemon
 989          * can find it.
 990          */
 991         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 992                 if (wired)
 993                         vm_page_wire(fs.m);
 994                 else
 995                         vm_page_unwire(fs.m, 1);
 996         } else
 997                 vm_page_activate(fs.m);
 998         if (m_hold != NULL) {
 999                 *m_hold = fs.m;
1000                 vm_page_hold(fs.m);
1001         }
1002         vm_page_unlock(fs.m);
1003         vm_page_xunbusy(fs.m);
1004
1005         /*
1006          * Unlock everything, and return
1007          */
1008         unlock_and_deallocate(&fs);
1009         if (hardfault) {
1010                 PCPU_INC(cnt.v_io_faults);
1011                 curthread->td_ru.ru_majflt++;
1012         } else
1013                 curthread->td_ru.ru_minflt++;
1014
1015         return (KERN_SUCCESS);
1016 }
1017
1018 /*
1019  * Speed up the reclamation of up to "distance" pages that precede the
1020  * faulting pindex within the first object of the shadow chain.
1021  */
1022 static void
1023 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1024 {
1025         vm_object_t first_object, object;
1026         vm_page_t m, m_prev;
1027         vm_pindex_t pindex;
1028
1029         object = fs->object;
1030         VM_OBJECT_ASSERT_WLOCKED(object);
1031         first_object = fs->first_object;
1032         if (first_object != object) {
1033                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1034                         VM_OBJECT_WUNLOCK(object);
1035                         VM_OBJECT_WLOCK(first_object);
1036                         VM_OBJECT_WLOCK(object);
1037                 }
1038         }
1039         /* Neither fictitious nor unmanaged pages can be cached. */
1040         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1041                 if (fs->first_pindex < distance)
1042                         pindex = 0;
1043                 else
1044                         pindex = fs->first_pindex - distance;
1045                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1046                         pindex = OFF_TO_IDX(fs->entry->offset);
1047                 m = first_object != object ? fs->first_m : fs->m;
1048                 vm_page_assert_xbusied(m);
1049                 m_prev = vm_page_prev(m);
1050                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1051                     m->valid == VM_PAGE_BITS_ALL) {
1052                         m_prev = vm_page_prev(m);
1053                         if (vm_page_busied(m))
1054                                 continue;
1055                         vm_page_lock(m);
1056                         if (m->hold_count == 0 && m->wire_count == 0) {
1057                                 pmap_remove_all(m);
1058                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1059                                 if (m->dirty != 0)
1060                                         vm_page_deactivate(m);
1061                                 else
1062                                         vm_page_cache(m);
1063                         }
1064                         vm_page_unlock(m);
1065                 }
1066         }
1067         if (first_object != object)
1068                 VM_OBJECT_WUNLOCK(first_object);
1069 }
1070
1071 /*
1072  * vm_fault_prefault provides a quick way of clustering
1073  * pagefaults into a processes address space.  It is a "cousin"
1074  * of vm_map_pmap_enter, except it runs at page fault time instead
1075  * of mmap time.
1076  */
1077 static void
1078 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1079     int faultcount, int reqpage)
1080 {
1081         pmap_t pmap;
1082         vm_map_entry_t entry;
1083         vm_object_t backing_object, lobject;
1084         vm_offset_t addr, starta;
1085         vm_pindex_t pindex;
1086         vm_page_t m;
1087         int backward, forward, i;
1088
1089         pmap = fs->map->pmap;
1090         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1091                 return;
1092
1093         if (faultcount > 0) {
1094                 backward = reqpage;
1095                 forward = faultcount - reqpage - 1;
1096         } else {
1097                 backward = PFBAK;
1098                 forward = PFFOR;
1099         }
1100         entry = fs->entry;
1101
1102         starta = addra - backward * PAGE_SIZE;
1103         if (starta < entry->start) {
1104                 starta = entry->start;
1105         } else if (starta > addra) {
1106                 starta = 0;
1107         }
1108
1109         /*
1110          * Generate the sequence of virtual addresses that are candidates for
1111          * prefaulting in an outward spiral from the faulting virtual address,
1112          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1113          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1114          * If the candidate address doesn't have a backing physical page, then
1115          * the loop immediately terminates.
1116          */
1117         for (i = 0; i < 2 * imax(backward, forward); i++) {
1118                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1119                     PAGE_SIZE);
1120                 if (addr > addra + forward * PAGE_SIZE)
1121                         addr = 0;
1122
1123                 if (addr < starta || addr >= entry->end)
1124                         continue;
1125
1126                 if (!pmap_is_prefaultable(pmap, addr))
1127                         continue;
1128
1129                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1130                 lobject = entry->object.vm_object;
1131                 VM_OBJECT_RLOCK(lobject);
1132                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1133                     lobject->type == OBJT_DEFAULT &&
1134                     (backing_object = lobject->backing_object) != NULL) {
1135                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1136                             0, ("vm_fault_prefault: unaligned object offset"));
1137                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1138                         VM_OBJECT_RLOCK(backing_object);
1139                         VM_OBJECT_RUNLOCK(lobject);
1140                         lobject = backing_object;
1141                 }
1142                 if (m == NULL) {
1143                         VM_OBJECT_RUNLOCK(lobject);
1144                         break;
1145                 }
1146                 if (m->valid == VM_PAGE_BITS_ALL &&
1147                     (m->flags & PG_FICTITIOUS) == 0)
1148                         pmap_enter_quick(pmap, addr, m, entry->protection);
1149                 VM_OBJECT_RUNLOCK(lobject);
1150         }
1151 }
1152
1153 /*
1154  * Hold each of the physical pages that are mapped by the specified range of
1155  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1156  * and allow the specified types of access, "prot".  If all of the implied
1157  * pages are successfully held, then the number of held pages is returned
1158  * together with pointers to those pages in the array "ma".  However, if any
1159  * of the pages cannot be held, -1 is returned.
1160  */
1161 int
1162 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1163     vm_prot_t prot, vm_page_t *ma, int max_count)
1164 {
1165         vm_offset_t end, va;
1166         vm_page_t *mp;
1167         int count;
1168         boolean_t pmap_failed;
1169
1170         if (len == 0)
1171                 return (0);
1172         end = round_page(addr + len);
1173         addr = trunc_page(addr);
1174
1175         /*
1176          * Check for illegal addresses.
1177          */
1178         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1179                 return (-1);
1180
1181         if (atop(end - addr) > max_count)
1182                 panic("vm_fault_quick_hold_pages: count > max_count");
1183         count = atop(end - addr);
1184
1185         /*
1186          * Most likely, the physical pages are resident in the pmap, so it is
1187          * faster to try pmap_extract_and_hold() first.
1188          */
1189         pmap_failed = FALSE;
1190         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1191                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1192                 if (*mp == NULL)
1193                         pmap_failed = TRUE;
1194                 else if ((prot & VM_PROT_WRITE) != 0 &&
1195                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1196                         /*
1197                          * Explicitly dirty the physical page.  Otherwise, the
1198                          * caller's changes may go unnoticed because they are
1199                          * performed through an unmanaged mapping or by a DMA
1200                          * operation.
1201                          *
1202                          * The object lock is not held here.
1203                          * See vm_page_clear_dirty_mask().
1204                          */
1205                         vm_page_dirty(*mp);
1206                 }
1207         }
1208         if (pmap_failed) {
1209                 /*
1210                  * One or more pages could not be held by the pmap.  Either no
1211                  * page was mapped at the specified virtual address or that
1212                  * mapping had insufficient permissions.  Attempt to fault in
1213                  * and hold these pages.
1214                  */
1215                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1216                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1217                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1218                                 goto error;
1219         }
1220         return (count);
1221 error:
1222         for (mp = ma; mp < ma + count; mp++)
1223                 if (*mp != NULL) {
1224                         vm_page_lock(*mp);
1225                         vm_page_unhold(*mp);
1226                         vm_page_unlock(*mp);
1227                 }
1228         return (-1);
1229 }
1230
1231 /*
1232  *      Routine:
1233  *              vm_fault_copy_entry
1234  *      Function:
1235  *              Create new shadow object backing dst_entry with private copy of
1236  *              all underlying pages. When src_entry is equal to dst_entry,
1237  *              function implements COW for wired-down map entry. Otherwise,
1238  *              it forks wired entry into dst_map.
1239  *
1240  *      In/out conditions:
1241  *              The source and destination maps must be locked for write.
1242  *              The source map entry must be wired down (or be a sharing map
1243  *              entry corresponding to a main map entry that is wired down).
1244  */
1245 void
1246 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1247     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1248     vm_ooffset_t *fork_charge)
1249 {
1250         vm_object_t backing_object, dst_object, object, src_object;
1251         vm_pindex_t dst_pindex, pindex, src_pindex;
1252         vm_prot_t access, prot;
1253         vm_offset_t vaddr;
1254         vm_page_t dst_m;
1255         vm_page_t src_m;
1256         boolean_t upgrade;
1257
1258 #ifdef  lint
1259         src_map++;
1260 #endif  /* lint */
1261
1262         upgrade = src_entry == dst_entry;
1263         access = prot = dst_entry->protection;
1264
1265         src_object = src_entry->object.vm_object;
1266         src_pindex = OFF_TO_IDX(src_entry->offset);
1267
1268         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1269                 dst_object = src_object;
1270                 vm_object_reference(dst_object);
1271         } else {
1272                 /*
1273                  * Create the top-level object for the destination entry. (Doesn't
1274                  * actually shadow anything - we copy the pages directly.)
1275                  */
1276                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1277                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1278 #if VM_NRESERVLEVEL > 0
1279                 dst_object->flags |= OBJ_COLORED;
1280                 dst_object->pg_color = atop(dst_entry->start);
1281 #endif
1282         }
1283
1284         VM_OBJECT_WLOCK(dst_object);
1285         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1286             ("vm_fault_copy_entry: vm_object not NULL"));
1287         if (src_object != dst_object) {
1288                 dst_entry->object.vm_object = dst_object;
1289                 dst_entry->offset = 0;
1290                 dst_object->charge = dst_entry->end - dst_entry->start;
1291         }
1292         if (fork_charge != NULL) {
1293                 KASSERT(dst_entry->cred == NULL,
1294                     ("vm_fault_copy_entry: leaked swp charge"));
1295                 dst_object->cred = curthread->td_ucred;
1296                 crhold(dst_object->cred);
1297                 *fork_charge += dst_object->charge;
1298         } else if (dst_object->cred == NULL) {
1299                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1300                     dst_entry));
1301                 dst_object->cred = dst_entry->cred;
1302                 dst_entry->cred = NULL;
1303         }
1304
1305         /*
1306          * If not an upgrade, then enter the mappings in the pmap as
1307          * read and/or execute accesses.  Otherwise, enter them as
1308          * write accesses.
1309          *
1310          * A writeable large page mapping is only created if all of
1311          * the constituent small page mappings are modified. Marking
1312          * PTEs as modified on inception allows promotion to happen
1313          * without taking potentially large number of soft faults.
1314          */
1315         if (!upgrade)
1316                 access &= ~VM_PROT_WRITE;
1317
1318         /*
1319          * Loop through all of the virtual pages within the entry's
1320          * range, copying each page from the source object to the
1321          * destination object.  Since the source is wired, those pages
1322          * must exist.  In contrast, the destination is pageable.
1323          * Since the destination object does share any backing storage
1324          * with the source object, all of its pages must be dirtied,
1325          * regardless of whether they can be written.
1326          */
1327         for (vaddr = dst_entry->start, dst_pindex = 0;
1328             vaddr < dst_entry->end;
1329             vaddr += PAGE_SIZE, dst_pindex++) {
1330 again:
1331                 /*
1332                  * Find the page in the source object, and copy it in.
1333                  * Because the source is wired down, the page will be
1334                  * in memory.
1335                  */
1336                 if (src_object != dst_object)
1337                         VM_OBJECT_RLOCK(src_object);
1338                 object = src_object;
1339                 pindex = src_pindex + dst_pindex;
1340                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1341                     (backing_object = object->backing_object) != NULL) {
1342                         /*
1343                          * Unless the source mapping is read-only or
1344                          * it is presently being upgraded from
1345                          * read-only, the first object in the shadow
1346                          * chain should provide all of the pages.  In
1347                          * other words, this loop body should never be
1348                          * executed when the source mapping is already
1349                          * read/write.
1350                          */
1351                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1352                             upgrade,
1353                             ("vm_fault_copy_entry: main object missing page"));
1354
1355                         VM_OBJECT_RLOCK(backing_object);
1356                         pindex += OFF_TO_IDX(object->backing_object_offset);
1357                         if (object != dst_object)
1358                                 VM_OBJECT_RUNLOCK(object);
1359                         object = backing_object;
1360                 }
1361                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1362
1363                 if (object != dst_object) {
1364                         /*
1365                          * Allocate a page in the destination object.
1366                          */
1367                         dst_m = vm_page_alloc(dst_object, (src_object ==
1368                             dst_object ? src_pindex : 0) + dst_pindex,
1369                             VM_ALLOC_NORMAL);
1370                         if (dst_m == NULL) {
1371                                 VM_OBJECT_WUNLOCK(dst_object);
1372                                 VM_OBJECT_RUNLOCK(object);
1373                                 VM_WAIT;
1374                                 VM_OBJECT_WLOCK(dst_object);
1375                                 goto again;
1376                         }
1377                         pmap_copy_page(src_m, dst_m);
1378                         VM_OBJECT_RUNLOCK(object);
1379                         dst_m->valid = VM_PAGE_BITS_ALL;
1380                         dst_m->dirty = VM_PAGE_BITS_ALL;
1381                 } else {
1382                         dst_m = src_m;
1383                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1384                                 goto again;
1385                         vm_page_xbusy(dst_m);
1386                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1387                             ("invalid dst page %p", dst_m));
1388                 }
1389                 VM_OBJECT_WUNLOCK(dst_object);
1390
1391                 /*
1392                  * Enter it in the pmap. If a wired, copy-on-write
1393                  * mapping is being replaced by a write-enabled
1394                  * mapping, then wire that new mapping.
1395                  */
1396                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1397                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1398
1399                 /*
1400                  * Mark it no longer busy, and put it on the active list.
1401                  */
1402                 VM_OBJECT_WLOCK(dst_object);
1403
1404                 if (upgrade) {
1405                         if (src_m != dst_m) {
1406                                 vm_page_lock(src_m);
1407                                 vm_page_unwire(src_m, 0);
1408                                 vm_page_unlock(src_m);
1409                                 vm_page_lock(dst_m);
1410                                 vm_page_wire(dst_m);
1411                                 vm_page_unlock(dst_m);
1412                         } else {
1413                                 KASSERT(dst_m->wire_count > 0,
1414                                     ("dst_m %p is not wired", dst_m));
1415                         }
1416                 } else {
1417                         vm_page_lock(dst_m);
1418                         vm_page_activate(dst_m);
1419                         vm_page_unlock(dst_m);
1420                 }
1421                 vm_page_xunbusy(dst_m);
1422         }
1423         VM_OBJECT_WUNLOCK(dst_object);
1424         if (upgrade) {
1425                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1426                 vm_object_deallocate(src_object);
1427         }
1428 }
1429
1430
1431 /*
1432  * This routine checks around the requested page for other pages that
1433  * might be able to be faulted in.  This routine brackets the viable
1434  * pages for the pages to be paged in.
1435  *
1436  * Inputs:
1437  *      m, rbehind, rahead
1438  *
1439  * Outputs:
1440  *  marray (array of vm_page_t), reqpage (index of requested page)
1441  *
1442  * Return value:
1443  *  number of pages in marray
1444  */
1445 static int
1446 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1447         vm_page_t m;
1448         int rbehind;
1449         int rahead;
1450         vm_page_t *marray;
1451         int *reqpage;
1452 {
1453         int i,j;
1454         vm_object_t object;
1455         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1456         vm_page_t rtm;
1457         int cbehind, cahead;
1458
1459         VM_OBJECT_ASSERT_WLOCKED(m->object);
1460
1461         object = m->object;
1462         pindex = m->pindex;
1463         cbehind = cahead = 0;
1464
1465         /*
1466          * if the requested page is not available, then give up now
1467          */
1468         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1469                 return 0;
1470         }
1471
1472         if ((cbehind == 0) && (cahead == 0)) {
1473                 *reqpage = 0;
1474                 marray[0] = m;
1475                 return 1;
1476         }
1477
1478         if (rahead > cahead) {
1479                 rahead = cahead;
1480         }
1481
1482         if (rbehind > cbehind) {
1483                 rbehind = cbehind;
1484         }
1485
1486         /*
1487          * scan backward for the read behind pages -- in memory
1488          */
1489         if (pindex > 0) {
1490                 if (rbehind > pindex) {
1491                         rbehind = pindex;
1492                         startpindex = 0;
1493                 } else {
1494                         startpindex = pindex - rbehind;
1495                 }
1496
1497                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1498                     rtm->pindex >= startpindex)
1499                         startpindex = rtm->pindex + 1;
1500
1501                 /* tpindex is unsigned; beware of numeric underflow. */
1502                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1503                     tpindex < pindex; i++, tpindex--) {
1504
1505                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1506                             VM_ALLOC_IFNOTCACHED);
1507                         if (rtm == NULL) {
1508                                 /*
1509                                  * Shift the allocated pages to the
1510                                  * beginning of the array.
1511                                  */
1512                                 for (j = 0; j < i; j++) {
1513                                         marray[j] = marray[j + tpindex + 1 -
1514                                             startpindex];
1515                                 }
1516                                 break;
1517                         }
1518
1519                         marray[tpindex - startpindex] = rtm;
1520                 }
1521         } else {
1522                 startpindex = 0;
1523                 i = 0;
1524         }
1525
1526         marray[i] = m;
1527         /* page offset of the required page */
1528         *reqpage = i;
1529
1530         tpindex = pindex + 1;
1531         i++;
1532
1533         /*
1534          * scan forward for the read ahead pages
1535          */
1536         endpindex = tpindex + rahead;
1537         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1538                 endpindex = rtm->pindex;
1539         if (endpindex > object->size)
1540                 endpindex = object->size;
1541
1542         for (; tpindex < endpindex; i++, tpindex++) {
1543
1544                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1545                     VM_ALLOC_IFNOTCACHED);
1546                 if (rtm == NULL) {
1547                         break;
1548                 }
1549
1550                 marray[i] = rtm;
1551         }
1552
1553         /* return number of pages */
1554         return i;
1555 }
1556
1557 /*
1558  * Block entry into the machine-independent layer's page fault handler by
1559  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1560  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1561  * spurious page faults.
1562  */
1563 int
1564 vm_fault_disable_pagefaults(void)
1565 {
1566
1567         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1568 }
1569
1570 void
1571 vm_fault_enable_pagefaults(int save)
1572 {
1573
1574         curthread_pflags_restore(save);
1575 }