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_vm.h"
  78
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/kernel.h>
  82 #include <sys/lock.h>
  83 #include <sys/mutex.h>
  84 #include <sys/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/sysctl.h>
  87 #include <sys/vmmeter.h>
  88 #include <sys/vnode.h>
  89
  90 #include <vm/vm.h>
  91 #include <vm/vm_param.h>
  92 #include <vm/pmap.h>
  93 #include <vm/vm_map.h>
  94 #include <vm/vm_object.h>
  95 #include <vm/vm_page.h>
  96 #include <vm/vm_pageout.h>
  97 #include <vm/vm_kern.h>
  98 #include <vm/vm_pager.h>
  99 #include <vm/vnode_pager.h>
 100 #include <vm/vm_extern.h>
 101
 102 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 103
 104 #define PFBAK 4
 105 #define PFFOR 4
 106 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 107
 108 static int prefault_pageorder[] = {
 109         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 110         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 111         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 112         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 113 };
 114
 115 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 116 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 117
 118 #define VM_FAULT_READ_AHEAD 8
 119 #define VM_FAULT_READ_BEHIND 7
 120 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 121
 122 struct faultstate {
 123         vm_page_t m;
 124         vm_object_t object;
 125         vm_pindex_t pindex;
 126         vm_page_t first_m;
 127         vm_object_t     first_object;
 128         vm_pindex_t first_pindex;
 129         vm_map_t map;
 130         vm_map_entry_t entry;
 131         int lookup_still_valid;
 132         struct vnode *vp;
 133         int vfslocked;
 134 };
 135
 136 static inline void
 137 release_page(struct faultstate *fs)
 138 {
 139
 140         vm_page_wakeup(fs->m);
 141         vm_page_lock_queues();
 142         vm_page_deactivate(fs->m);
 143         vm_page_unlock_queues();
 144         fs->m = NULL;
 145 }
 146
 147 static inline void
 148 unlock_map(struct faultstate *fs)
 149 {
 150
 151         if (fs->lookup_still_valid) {
 152                 vm_map_lookup_done(fs->map, fs->entry);
 153                 fs->lookup_still_valid = FALSE;
 154         }
 155 }
 156
 157 static void
 158 unlock_and_deallocate(struct faultstate *fs)
 159 {
 160
 161         vm_object_pip_wakeup(fs->object);
 162         VM_OBJECT_UNLOCK(fs->object);
 163         if (fs->object != fs->first_object) {
 164                 VM_OBJECT_LOCK(fs->first_object);
 165                 vm_page_lock_queues();
 166                 vm_page_free(fs->first_m);
 167                 vm_page_unlock_queues();
 168                 vm_object_pip_wakeup(fs->first_object);
 169                 VM_OBJECT_UNLOCK(fs->first_object);
 170                 fs->first_m = NULL;
 171         }
 172         vm_object_deallocate(fs->first_object);
 173         unlock_map(fs);
 174         if (fs->vp != NULL) {
 175                 vput(fs->vp);
 176                 fs->vp = NULL;
 177         }
 178         VFS_UNLOCK_GIANT(fs->vfslocked);
 179         fs->vfslocked = 0;
 180 }
 181
 182 /*
 183  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 184  *            current object *might* contain the page.
 185  *
 186  *            default objects are zero-fill, there is no real pager.
 187  */
 188 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 189                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 190
 191 /*
 192  *      vm_fault:
 193  *
 194  *      Handle a page fault occurring at the given address,
 195  *      requiring the given permissions, in the map specified.
 196  *      If successful, the page is inserted into the
 197  *      associated physical map.
 198  *
 199  *      NOTE: the given address should be truncated to the
 200  *      proper page address.
 201  *
 202  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 203  *      a standard error specifying why the fault is fatal is returned.
 204  *
 205  *
 206  *      The map in question must be referenced, and remains so.
 207  *      Caller may hold no locks.
 208  */
 209 int
 210 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 211          int fault_flags)
 212 {
 213         vm_prot_t prot;
 214         int is_first_object_locked, result;
 215         boolean_t are_queues_locked, growstack, wired;
 216         int map_generation;
 217         vm_object_t next_object;
 218         vm_page_t marray[VM_FAULT_READ];
 219         int hardfault;
 220         int faultcount, ahead, behind, alloc_req;
 221         struct faultstate fs;
 222         struct vnode *vp;
 223         int locked, error;
 224
 225         hardfault = 0;
 226         growstack = TRUE;
 227         PCPU_INC(cnt.v_vm_faults);
 228         fs.vp = NULL;
 229         fs.vfslocked = 0;
 230         faultcount = behind = 0;
 231
 232 RetryFault:;
 233
 234         /*
 235          * Find the backing store object and offset into it to begin the
 236          * search.
 237          */
 238         fs.map = map;
 239         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 240             &fs.first_object, &fs.first_pindex, &prot, &wired);
 241         if (result != KERN_SUCCESS) {
 242                 if (result != KERN_PROTECTION_FAILURE ||
 243                     (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
 244                         if (growstack && result == KERN_INVALID_ADDRESS &&
 245                             map != kernel_map && curproc != NULL) {
 246                                 result = vm_map_growstack(curproc, vaddr);
 247                                 if (result != KERN_SUCCESS)
 248                                         return (KERN_FAILURE);
 249                                 growstack = FALSE;
 250                                 goto RetryFault;
 251                         }
 252                         return (result);
 253                 }
 254
 255                 /*
 256                  * If we are user-wiring a r/w segment, and it is COW, then
 257                  * we need to do the COW operation.  Note that we don't COW
 258                  * currently RO sections now, because it is NOT desirable
 259                  * to COW .text.  We simply keep .text from ever being COW'ed
 260                  * and take the heat that one cannot debug wired .text sections.
 261                  */
 262                 result = vm_map_lookup(&fs.map, vaddr,
 263                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 264                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 265                 if (result != KERN_SUCCESS)
 266                         return (result);
 267
 268                 /*
 269                  * If we don't COW now, on a user wire, the user will never
 270                  * be able to write to the mapping.  If we don't make this
 271                  * restriction, the bookkeeping would be nearly impossible.
 272                  *
 273                  * XXX The following assignment modifies the map without
 274                  * holding a write lock on it.
 275                  */
 276                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 277                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 278         }
 279
 280         map_generation = fs.map->timestamp;
 281
 282         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 283                 panic("vm_fault: fault on nofault entry, addr: %lx",
 284                     (u_long)vaddr);
 285         }
 286
 287         /*
 288          * Make a reference to this object to prevent its disposal while we
 289          * are messing with it.  Once we have the reference, the map is free
 290          * to be diddled.  Since objects reference their shadows (and copies),
 291          * they will stay around as well.
 292          *
 293          * Bump the paging-in-progress count to prevent size changes (e.g.
 294          * truncation operations) during I/O.  This must be done after
 295          * obtaining the vnode lock in order to avoid possible deadlocks.
 296          */
 297         VM_OBJECT_LOCK(fs.first_object);
 298         vm_object_reference_locked(fs.first_object);
 299         vm_object_pip_add(fs.first_object, 1);
 300
 301         fs.lookup_still_valid = TRUE;
 302
 303         if (wired)
 304                 fault_type = prot;
 305
 306         fs.first_m = NULL;
 307
 308         /*
 309          * Search for the page at object/offset.
 310          */
 311         fs.object = fs.first_object;
 312         fs.pindex = fs.first_pindex;
 313         while (TRUE) {
 314                 /*
 315                  * If the object is dead, we stop here
 316                  */
 317                 if (fs.object->flags & OBJ_DEAD) {
 318                         unlock_and_deallocate(&fs);
 319                         return (KERN_PROTECTION_FAILURE);
 320                 }
 321
 322                 /*
 323                  * See if page is resident
 324                  */
 325                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 326                 if (fs.m != NULL) {
 327                         /*
 328                          * check for page-based copy on write.
 329                          * We check fs.object == fs.first_object so
 330                          * as to ensure the legacy COW mechanism is
 331                          * used when the page in question is part of
 332                          * a shadow object.  Otherwise, vm_page_cowfault()
 333                          * removes the page from the backing object,
 334                          * which is not what we want.
 335                          */
 336                         vm_page_lock_queues();
 337                         if ((fs.m->cow) &&
 338                             (fault_type & VM_PROT_WRITE) &&
 339                             (fs.object == fs.first_object)) {
 340                                 vm_page_cowfault(fs.m);
 341                                 vm_page_unlock_queues();
 342                                 unlock_and_deallocate(&fs);
 343                                 goto RetryFault;
 344                         }
 345
 346                         /*
 347                          * Wait/Retry if the page is busy.  We have to do this
 348                          * if the page is busy via either VPO_BUSY or
 349                          * vm_page_t->busy because the vm_pager may be using
 350                          * vm_page_t->busy for pageouts ( and even pageins if
 351                          * it is the vnode pager ), and we could end up trying
 352                          * to pagein and pageout the same page simultaneously.
 353                          *
 354                          * We can theoretically allow the busy case on a read
 355                          * fault if the page is marked valid, but since such
 356                          * pages are typically already pmap'd, putting that
 357                          * special case in might be more effort then it is
 358                          * worth.  We cannot under any circumstances mess
 359                          * around with a vm_page_t->busy page except, perhaps,
 360                          * to pmap it.
 361                          */
 362                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 363                                 vm_page_unlock_queues();
 364                                 VM_OBJECT_UNLOCK(fs.object);
 365                                 if (fs.object != fs.first_object) {
 366                                         VM_OBJECT_LOCK(fs.first_object);
 367                                         vm_page_lock_queues();
 368                                         vm_page_free(fs.first_m);
 369                                         vm_page_unlock_queues();
 370                                         vm_object_pip_wakeup(fs.first_object);
 371                                         VM_OBJECT_UNLOCK(fs.first_object);
 372                                         fs.first_m = NULL;
 373                                 }
 374                                 unlock_map(&fs);
 375                                 VM_OBJECT_LOCK(fs.object);
 376                                 if (fs.m == vm_page_lookup(fs.object,
 377                                     fs.pindex)) {
 378                                         vm_page_sleep_if_busy(fs.m, TRUE,
 379                                             "vmpfw");
 380                                 }
 381                                 vm_object_pip_wakeup(fs.object);
 382                                 VM_OBJECT_UNLOCK(fs.object);
 383                                 PCPU_INC(cnt.v_intrans);
 384                                 vm_object_deallocate(fs.first_object);
 385                                 goto RetryFault;
 386                         }
 387                         vm_pageq_remove(fs.m);
 388                         vm_page_unlock_queues();
 389
 390                         /*
 391                          * Mark page busy for other processes, and the
 392                          * pagedaemon.  If it still isn't completely valid
 393                          * (readable), jump to readrest, else break-out ( we
 394                          * found the page ).
 395                          */
 396                         vm_page_busy(fs.m);
 397                         if (fs.m->valid != VM_PAGE_BITS_ALL &&
 398                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 399                                 goto readrest;
 400                         }
 401
 402                         break;
 403                 }
 404
 405                 /*
 406                  * Page is not resident, If this is the search termination
 407                  * or the pager might contain the page, allocate a new page.
 408                  */
 409                 if (TRYPAGER || fs.object == fs.first_object) {
 410                         if (fs.pindex >= fs.object->size) {
 411                                 unlock_and_deallocate(&fs);
 412                                 return (KERN_PROTECTION_FAILURE);
 413                         }
 414
 415                         /*
 416                          * Allocate a new page for this object/offset pair.
 417                          *
 418                          * Unlocked read of the p_flag is harmless. At
 419                          * worst, the P_KILLED might be not observed
 420                          * there, and allocation can fail, causing
 421                          * restart and new reading of the p_flag.
 422                          */
 423                         fs.m = NULL;
 424                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 425 #if VM_NRESERVLEVEL > 0
 426                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 427                                         fs.object->flags |= OBJ_COLORED;
 428                                         fs.object->pg_color = atop(vaddr) -
 429                                             fs.pindex;
 430                                 }
 431 #endif
 432                                 alloc_req = P_KILLED(curproc) ?
 433                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 434                                 if (fs.object->type != OBJT_VNODE &&
 435                                     fs.object->backing_object == NULL)
 436                                         alloc_req |= VM_ALLOC_ZERO;
 437                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 438                                     alloc_req);
 439                         }
 440                         if (fs.m == NULL) {
 441                                 unlock_and_deallocate(&fs);
 442                                 VM_WAITPFAULT;
 443                                 goto RetryFault;
 444                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 445                                 break;
 446                 }
 447
 448 readrest:
 449                 /*
 450                  * We have found a valid page or we have allocated a new page.
 451                  * The page thus may not be valid or may not be entirely
 452                  * valid.
 453                  *
 454                  * Attempt to fault-in the page if there is a chance that the
 455                  * pager has it, and potentially fault in additional pages
 456                  * at the same time.
 457                  */
 458                 if (TRYPAGER) {
 459                         int rv;
 460                         int reqpage = 0;
 461                         u_char behavior = vm_map_entry_behavior(fs.entry);
 462
 463                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 464                             P_KILLED(curproc)) {
 465                                 ahead = 0;
 466                                 behind = 0;
 467                         } else {
 468                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 469                                 if (behind > VM_FAULT_READ_BEHIND)
 470                                         behind = VM_FAULT_READ_BEHIND;
 471
 472                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 473                                 if (ahead > VM_FAULT_READ_AHEAD)
 474                                         ahead = VM_FAULT_READ_AHEAD;
 475                         }
 476                         is_first_object_locked = FALSE;
 477                         if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 478                              (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 479                               fs.pindex >= fs.entry->lastr &&
 480                               fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
 481                             (fs.first_object == fs.object ||
 482                              (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
 483                             fs.first_object->type != OBJT_DEVICE &&
 484                             fs.first_object->type != OBJT_PHYS &&
 485                             fs.first_object->type != OBJT_SG) {
 486                                 vm_pindex_t firstpindex, tmppindex;
 487
 488                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 489                                         firstpindex = 0;
 490                                 else
 491                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 492
 493                                 are_queues_locked = FALSE;
 494                                 /*
 495                                  * note: partially valid pages cannot be
 496                                  * included in the lookahead - NFS piecemeal
 497                                  * writes will barf on it badly.
 498                                  */
 499                                 for (tmppindex = fs.first_pindex - 1;
 500                                         tmppindex >= firstpindex;
 501                                         --tmppindex) {
 502                                         vm_page_t mt;
 503
 504                                         mt = vm_page_lookup(fs.first_object, tmppindex);
 505                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 506                                                 break;
 507                                         if (mt->busy ||
 508                                             (mt->oflags & VPO_BUSY))
 509                                                 continue;
 510                                         if (!are_queues_locked) {
 511                                                 are_queues_locked = TRUE;
 512                                                 vm_page_lock_queues();
 513                                         }
 514                                         if (mt->hold_count ||
 515                                                 mt->wire_count)
 516                                                 continue;
 517                                         pmap_remove_all(mt);
 518                                         if (mt->dirty) {
 519                                                 vm_page_deactivate(mt);
 520                                         } else {
 521                                                 vm_page_cache(mt);
 522                                         }
 523                                 }
 524                                 if (are_queues_locked)
 525                                         vm_page_unlock_queues();
 526                                 ahead += behind;
 527                                 behind = 0;
 528                         }
 529                         if (is_first_object_locked)
 530                                 VM_OBJECT_UNLOCK(fs.first_object);
 531
 532                         /*
 533                          * Call the pager to retrieve the data, if any, after
 534                          * releasing the lock on the map.  We hold a ref on
 535                          * fs.object and the pages are VPO_BUSY'd.
 536                          */
 537                         unlock_map(&fs);
 538
 539 vnode_lock:
 540                         if (fs.object->type == OBJT_VNODE) {
 541                                 vp = fs.object->handle;
 542                                 if (vp == fs.vp)
 543                                         goto vnode_locked;
 544                                 else if (fs.vp != NULL) {
 545                                         vput(fs.vp);
 546                                         fs.vp = NULL;
 547                                 }
 548                                 locked = VOP_ISLOCKED(vp);
 549
 550                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 551                                         fs.vfslocked = 1;
 552                                         if (!mtx_trylock(&Giant)) {
 553                                                 VM_OBJECT_UNLOCK(fs.object);
 554                                                 mtx_lock(&Giant);
 555                                                 VM_OBJECT_LOCK(fs.object);
 556                                                 goto vnode_lock;
 557                                         }
 558                                 }
 559                                 if (locked != LK_EXCLUSIVE)
 560                                         locked = LK_SHARED;
 561                                 /* Do not sleep for vnode lock while fs.m is busy */
 562                                 error = vget(vp, locked | LK_CANRECURSE |
 563                                     LK_NOWAIT, curthread);
 564                                 if (error != 0) {
 565                                         int vfslocked;
 566
 567                                         vfslocked = fs.vfslocked;
 568                                         fs.vfslocked = 0; /* Keep Giant */
 569                                         vhold(vp);
 570                                         release_page(&fs);
 571                                         unlock_and_deallocate(&fs);
 572                                         error = vget(vp, locked | LK_RETRY |
 573                                             LK_CANRECURSE, curthread);
 574                                         vdrop(vp);
 575                                         fs.vp = vp;
 576                                         fs.vfslocked = vfslocked;
 577                                         KASSERT(error == 0,
 578                                             ("vm_fault: vget failed"));
 579                                         goto RetryFault;
 580                                 }
 581                                 fs.vp = vp;
 582                         }
 583 vnode_locked:
 584                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 585                             ("vm_fault: vnode-backed object mapped by system map"));
 586
 587                         /*
 588                          * now we find out if any other pages should be paged
 589                          * in at this time this routine checks to see if the
 590                          * pages surrounding this fault reside in the same
 591                          * object as the page for this fault.  If they do,
 592                          * then they are faulted in also into the object.  The
 593                          * array "marray" returned contains an array of
 594                          * vm_page_t structs where one of them is the
 595                          * vm_page_t passed to the routine.  The reqpage
 596                          * return value is the index into the marray for the
 597                          * vm_page_t passed to the routine.
 598                          *
 599                          * fs.m plus the additional pages are VPO_BUSY'd.
 600                          */
 601                         faultcount = vm_fault_additional_pages(
 602                             fs.m, behind, ahead, marray, &reqpage);
 603
 604                         rv = faultcount ?
 605                             vm_pager_get_pages(fs.object, marray, faultcount,
 606                                 reqpage) : VM_PAGER_FAIL;
 607
 608                         if (rv == VM_PAGER_OK) {
 609                                 /*
 610                                  * Found the page. Leave it busy while we play
 611                                  * with it.
 612                                  */
 613
 614                                 /*
 615                                  * Relookup in case pager changed page. Pager
 616                                  * is responsible for disposition of old page
 617                                  * if moved.
 618                                  */
 619                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 620                                 if (!fs.m) {
 621                                         unlock_and_deallocate(&fs);
 622                                         goto RetryFault;
 623                                 }
 624
 625                                 hardfault++;
 626                                 break; /* break to PAGE HAS BEEN FOUND */
 627                         }
 628                         /*
 629                          * Remove the bogus page (which does not exist at this
 630                          * object/offset); before doing so, we must get back
 631                          * our object lock to preserve our invariant.
 632                          *
 633                          * Also wake up any other process that may want to bring
 634                          * in this page.
 635                          *
 636                          * If this is the top-level object, we must leave the
 637                          * busy page to prevent another process from rushing
 638                          * past us, and inserting the page in that object at
 639                          * the same time that we are.
 640                          */
 641                         if (rv == VM_PAGER_ERROR)
 642                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 643                                     curproc->p_pid, curproc->p_comm);
 644                         /*
 645                          * Data outside the range of the pager or an I/O error
 646                          */
 647                         /*
 648                          * XXX - the check for kernel_map is a kludge to work
 649                          * around having the machine panic on a kernel space
 650                          * fault w/ I/O error.
 651                          */
 652                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 653                                 (rv == VM_PAGER_BAD)) {
 654                                 vm_page_lock_queues();
 655                                 vm_page_free(fs.m);
 656                                 vm_page_unlock_queues();
 657                                 fs.m = NULL;
 658                                 unlock_and_deallocate(&fs);
 659                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 660                         }
 661                         if (fs.object != fs.first_object) {
 662                                 vm_page_lock_queues();
 663                                 vm_page_free(fs.m);
 664                                 vm_page_unlock_queues();
 665                                 fs.m = NULL;
 666                                 /*
 667                                  * XXX - we cannot just fall out at this
 668                                  * point, m has been freed and is invalid!
 669                                  */
 670                         }
 671                 }
 672
 673                 /*
 674                  * We get here if the object has default pager (or unwiring)
 675                  * or the pager doesn't have the page.
 676                  */
 677                 if (fs.object == fs.first_object)
 678                         fs.first_m = fs.m;
 679
 680                 /*
 681                  * Move on to the next object.  Lock the next object before
 682                  * unlocking the current one.
 683                  */
 684                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 685                 next_object = fs.object->backing_object;
 686                 if (next_object == NULL) {
 687                         /*
 688                          * If there's no object left, fill the page in the top
 689                          * object with zeros.
 690                          */
 691                         if (fs.object != fs.first_object) {
 692                                 vm_object_pip_wakeup(fs.object);
 693                                 VM_OBJECT_UNLOCK(fs.object);
 694
 695                                 fs.object = fs.first_object;
 696                                 fs.pindex = fs.first_pindex;
 697                                 fs.m = fs.first_m;
 698                                 VM_OBJECT_LOCK(fs.object);
 699                         }
 700                         fs.first_m = NULL;
 701
 702                         /*
 703                          * Zero the page if necessary and mark it valid.
 704                          */
 705                         if ((fs.m->flags & PG_ZERO) == 0) {
 706                                 pmap_zero_page(fs.m);
 707                         } else {
 708                                 PCPU_INC(cnt.v_ozfod);
 709                         }
 710                         PCPU_INC(cnt.v_zfod);
 711                         fs.m->valid = VM_PAGE_BITS_ALL;
 712                         break;  /* break to PAGE HAS BEEN FOUND */
 713                 } else {
 714                         KASSERT(fs.object != next_object,
 715                             ("object loop %p", next_object));
 716                         VM_OBJECT_LOCK(next_object);
 717                         vm_object_pip_add(next_object, 1);
 718                         if (fs.object != fs.first_object)
 719                                 vm_object_pip_wakeup(fs.object);
 720                         VM_OBJECT_UNLOCK(fs.object);
 721                         fs.object = next_object;
 722                 }
 723         }
 724
 725         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 726             ("vm_fault: not busy after main loop"));
 727
 728         /*
 729          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 730          * is held.]
 731          */
 732
 733         /*
 734          * If the page is being written, but isn't already owned by the
 735          * top-level object, we have to copy it into a new page owned by the
 736          * top-level object.
 737          */
 738         if (fs.object != fs.first_object) {
 739                 /*
 740                  * We only really need to copy if we want to write it.
 741                  */
 742                 if (fault_type & VM_PROT_WRITE) {
 743                         /*
 744                          * This allows pages to be virtually copied from a
 745                          * backing_object into the first_object, where the
 746                          * backing object has no other refs to it, and cannot
 747                          * gain any more refs.  Instead of a bcopy, we just
 748                          * move the page from the backing object to the
 749                          * first object.  Note that we must mark the page
 750                          * dirty in the first object so that it will go out
 751                          * to swap when needed.
 752                          */
 753                         is_first_object_locked = FALSE;
 754                         if (
 755                                 /*
 756                                  * Only one shadow object
 757                                  */
 758                                 (fs.object->shadow_count == 1) &&
 759                                 /*
 760                                  * No COW refs, except us
 761                                  */
 762                                 (fs.object->ref_count == 1) &&
 763                                 /*
 764                                  * No one else can look this object up
 765                                  */
 766                                 (fs.object->handle == NULL) &&
 767                                 /*
 768                                  * No other ways to look the object up
 769                                  */
 770                                 ((fs.object->type == OBJT_DEFAULT) ||
 771                                  (fs.object->type == OBJT_SWAP)) &&
 772                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 773                                 /*
 774                                  * We don't chase down the shadow chain
 775                                  */
 776                             fs.object == fs.first_object->backing_object) {
 777                                 vm_page_lock_queues();
 778                                 /*
 779                                  * get rid of the unnecessary page
 780                                  */
 781                                 vm_page_free(fs.first_m);
 782                                 /*
 783                                  * grab the page and put it into the
 784                                  * process'es object.  The page is
 785                                  * automatically made dirty.
 786                                  */
 787                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 788                                 vm_page_unlock_queues();
 789                                 vm_page_busy(fs.m);
 790                                 fs.first_m = fs.m;
 791                                 fs.m = NULL;
 792                                 PCPU_INC(cnt.v_cow_optim);
 793                         } else {
 794                                 /*
 795                                  * Oh, well, lets copy it.
 796                                  */
 797                                 pmap_copy_page(fs.m, fs.first_m);
 798                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 799                         }
 800                         if (fs.m) {
 801                                 /*
 802                                  * We no longer need the old page or object.
 803                                  */
 804                                 release_page(&fs);
 805                         }
 806                         /*
 807                          * fs.object != fs.first_object due to above
 808                          * conditional
 809                          */
 810                         vm_object_pip_wakeup(fs.object);
 811                         VM_OBJECT_UNLOCK(fs.object);
 812                         /*
 813                          * Only use the new page below...
 814                          */
 815                         fs.object = fs.first_object;
 816                         fs.pindex = fs.first_pindex;
 817                         fs.m = fs.first_m;
 818                         if (!is_first_object_locked)
 819                                 VM_OBJECT_LOCK(fs.object);
 820                         PCPU_INC(cnt.v_cow_faults);
 821                 } else {
 822                         prot &= ~VM_PROT_WRITE;
 823                 }
 824         }
 825
 826         /*
 827          * We must verify that the maps have not changed since our last
 828          * lookup.
 829          */
 830         if (!fs.lookup_still_valid) {
 831                 vm_object_t retry_object;
 832                 vm_pindex_t retry_pindex;
 833                 vm_prot_t retry_prot;
 834
 835                 if (!vm_map_trylock_read(fs.map)) {
 836                         release_page(&fs);
 837                         unlock_and_deallocate(&fs);
 838                         goto RetryFault;
 839                 }
 840                 fs.lookup_still_valid = TRUE;
 841                 if (fs.map->timestamp != map_generation) {
 842                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 843                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 844
 845                         /*
 846                          * If we don't need the page any longer, put it on the inactive
 847                          * list (the easiest thing to do here).  If no one needs it,
 848                          * pageout will grab it eventually.
 849                          */
 850                         if (result != KERN_SUCCESS) {
 851                                 release_page(&fs);
 852                                 unlock_and_deallocate(&fs);
 853
 854                                 /*
 855                                  * If retry of map lookup would have blocked then
 856                                  * retry fault from start.
 857                                  */
 858                                 if (result == KERN_FAILURE)
 859                                         goto RetryFault;
 860                                 return (result);
 861                         }
 862                         if ((retry_object != fs.first_object) ||
 863                             (retry_pindex != fs.first_pindex)) {
 864                                 release_page(&fs);
 865                                 unlock_and_deallocate(&fs);
 866                                 goto RetryFault;
 867                         }
 868
 869                         /*
 870                          * Check whether the protection has changed or the object has
 871                          * been copied while we left the map unlocked. Changing from
 872                          * read to write permission is OK - we leave the page
 873                          * write-protected, and catch the write fault. Changing from
 874                          * write to read permission means that we can't mark the page
 875                          * write-enabled after all.
 876                          */
 877                         prot &= retry_prot;
 878                 }
 879         }
 880         /*
 881          * If the page was filled by a pager, update the map entry's
 882          * last read offset.  Since the pager does not return the
 883          * actual set of pages that it read, this update is based on
 884          * the requested set.  Typically, the requested and actual
 885          * sets are the same.
 886          *
 887          * XXX The following assignment modifies the map
 888          * without holding a write lock on it.
 889          */
 890         if (hardfault)
 891                 fs.entry->lastr = fs.pindex + faultcount - behind;
 892
 893         if (prot & VM_PROT_WRITE) {
 894                 vm_object_set_writeable_dirty(fs.object);
 895
 896                 /*
 897                  * If the fault is a write, we know that this page is being
 898                  * written NOW so dirty it explicitly to save on
 899                  * pmap_is_modified() calls later.
 900                  *
 901                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 902                  * if the page is already dirty to prevent data written with
 903                  * the expectation of being synced from not being synced.
 904                  * Likewise if this entry does not request NOSYNC then make
 905                  * sure the page isn't marked NOSYNC.  Applications sharing
 906                  * data should use the same flags to avoid ping ponging.
 907                  *
 908                  * Also tell the backing pager, if any, that it should remove
 909                  * any swap backing since the page is now dirty.
 910                  */
 911                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 912                         if (fs.m->dirty == 0)
 913                                 fs.m->oflags |= VPO_NOSYNC;
 914                 } else {
 915                         fs.m->oflags &= ~VPO_NOSYNC;
 916                 }
 917                 if (fault_flags & VM_FAULT_DIRTY) {
 918                         vm_page_dirty(fs.m);
 919                         vm_pager_page_unswapped(fs.m);
 920                 }
 921         }
 922
 923         /*
 924          * Page had better still be busy
 925          */
 926         KASSERT(fs.m->oflags & VPO_BUSY,
 927                 ("vm_fault: page %p not busy!", fs.m));
 928         /*
 929          * Page must be completely valid or it is not fit to
 930          * map into user space.  vm_pager_get_pages() ensures this.
 931          */
 932         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 933             ("vm_fault: page %p partially invalid", fs.m));
 934         VM_OBJECT_UNLOCK(fs.object);
 935
 936         /*
 937          * Put this page into the physical map.  We had to do the unlock above
 938          * because pmap_enter() may sleep.  We don't put the page
 939          * back on the active queue until later so that the pageout daemon
 940          * won't find it (yet).
 941          */
 942         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 943         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 944                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 945         }
 946         VM_OBJECT_LOCK(fs.object);
 947         vm_page_lock_queues();
 948         vm_page_flag_set(fs.m, PG_REFERENCED);
 949
 950         /*
 951          * If the page is not wired down, then put it where the pageout daemon
 952          * can find it.
 953          */
 954         if (fault_flags & VM_FAULT_WIRE_MASK) {
 955                 if (wired)
 956                         vm_page_wire(fs.m);
 957                 else
 958                         vm_page_unwire(fs.m, 1);
 959         } else {
 960                 vm_page_activate(fs.m);
 961         }
 962         vm_page_unlock_queues();
 963         vm_page_wakeup(fs.m);
 964
 965         /*
 966          * Unlock everything, and return
 967          */
 968         unlock_and_deallocate(&fs);
 969         if (hardfault)
 970                 curthread->td_ru.ru_majflt++;
 971         else
 972                 curthread->td_ru.ru_minflt++;
 973
 974         return (KERN_SUCCESS);
 975 }
 976
 977 /*
 978  * vm_fault_prefault provides a quick way of clustering
 979  * pagefaults into a processes address space.  It is a "cousin"
 980  * of vm_map_pmap_enter, except it runs at page fault time instead
 981  * of mmap time.
 982  */
 983 static void
 984 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
 985 {
 986         int i;
 987         vm_offset_t addr, starta;
 988         vm_pindex_t pindex;
 989         vm_page_t m;
 990         vm_object_t object;
 991
 992         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
 993                 return;
 994
 995         object = entry->object.vm_object;
 996
 997         starta = addra - PFBAK * PAGE_SIZE;
 998         if (starta < entry->start) {
 999                 starta = entry->start;
1000         } else if (starta > addra) {
1001                 starta = 0;
1002         }
1003
1004         for (i = 0; i < PAGEORDER_SIZE; i++) {
1005                 vm_object_t backing_object, lobject;
1006
1007                 addr = addra + prefault_pageorder[i];
1008                 if (addr > addra + (PFFOR * PAGE_SIZE))
1009                         addr = 0;
1010
1011                 if (addr < starta || addr >= entry->end)
1012                         continue;
1013
1014                 if (!pmap_is_prefaultable(pmap, addr))
1015                         continue;
1016
1017                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1018                 lobject = object;
1019                 VM_OBJECT_LOCK(lobject);
1020                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1021                     lobject->type == OBJT_DEFAULT &&
1022                     (backing_object = lobject->backing_object) != NULL) {
1023                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1024                             0, ("vm_fault_prefault: unaligned object offset"));
1025                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1026                         VM_OBJECT_LOCK(backing_object);
1027                         VM_OBJECT_UNLOCK(lobject);
1028                         lobject = backing_object;
1029                 }
1030                 /*
1031                  * give-up when a page is not in memory
1032                  */
1033                 if (m == NULL) {
1034                         VM_OBJECT_UNLOCK(lobject);
1035                         break;
1036                 }
1037                 if (m->valid == VM_PAGE_BITS_ALL &&
1038                     (m->flags & PG_FICTITIOUS) == 0) {
1039                         vm_page_lock_queues();
1040                         pmap_enter_quick(pmap, addr, m, entry->protection);
1041                         vm_page_unlock_queues();
1042                 }
1043                 VM_OBJECT_UNLOCK(lobject);
1044         }
1045 }
1046
1047 /*
1048  *      vm_fault_quick:
1049  *
1050  *      Ensure that the requested virtual address, which may be in userland,
1051  *      is valid.  Fault-in the page if necessary.  Return -1 on failure.
1052  */
1053 int
1054 vm_fault_quick(caddr_t v, int prot)
1055 {
1056         int r;
1057
1058         if (prot & VM_PROT_WRITE)
1059                 r = subyte(v, fubyte(v));
1060         else
1061                 r = fubyte(v);
1062         return(r);
1063 }
1064
1065 /*
1066  *      vm_fault_wire:
1067  *
1068  *      Wire down a range of virtual addresses in a map.
1069  */
1070 int
1071 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1072     boolean_t user_wire, boolean_t fictitious)
1073 {
1074         vm_offset_t va;
1075         int rv;
1076
1077         /*
1078          * We simulate a fault to get the page and enter it in the physical
1079          * map.  For user wiring, we only ask for read access on currently
1080          * read-only sections.
1081          */
1082         for (va = start; va < end; va += PAGE_SIZE) {
1083                 rv = vm_fault(map, va,
1084                     user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1085                     user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1086                 if (rv) {
1087                         if (va != start)
1088                                 vm_fault_unwire(map, start, va, fictitious);
1089                         return (rv);
1090                 }
1091         }
1092         return (KERN_SUCCESS);
1093 }
1094
1095 /*
1096  *      vm_fault_unwire:
1097  *
1098  *      Unwire a range of virtual addresses in a map.
1099  */
1100 void
1101 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1102     boolean_t fictitious)
1103 {
1104         vm_paddr_t pa;
1105         vm_offset_t va;
1106         pmap_t pmap;
1107
1108         pmap = vm_map_pmap(map);
1109
1110         /*
1111          * Since the pages are wired down, we must be able to get their
1112          * mappings from the physical map system.
1113          */
1114         for (va = start; va < end; va += PAGE_SIZE) {
1115                 pa = pmap_extract(pmap, va);
1116                 if (pa != 0) {
1117                         pmap_change_wiring(pmap, va, FALSE);
1118                         if (!fictitious) {
1119                                 vm_page_lock_queues();
1120                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1121                                 vm_page_unlock_queues();
1122                         }
1123                 }
1124         }
1125 }
1126
1127 /*
1128  *      Routine:
1129  *              vm_fault_copy_entry
1130  *      Function:
1131  *              Create new shadow object backing dst_entry with private copy of
1132  *              all underlying pages. When src_entry is equal to dst_entry,
1133  *              function implements COW for wired-down map entry. Otherwise,
1134  *              it forks wired entry into dst_map.
1135  *
1136  *      In/out conditions:
1137  *              The source and destination maps must be locked for write.
1138  *              The source map entry must be wired down (or be a sharing map
1139  *              entry corresponding to a main map entry that is wired down).
1140  */
1141 void
1142 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1143     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1144     vm_ooffset_t *fork_charge)
1145 {
1146         vm_object_t backing_object, dst_object, object, src_object;
1147         vm_pindex_t dst_pindex, pindex, src_pindex;
1148         vm_prot_t access, prot;
1149         vm_offset_t vaddr;
1150         vm_page_t dst_m;
1151         vm_page_t src_m;
1152         boolean_t src_readonly, upgrade;
1153
1154 #ifdef  lint
1155         src_map++;
1156 #endif  /* lint */
1157
1158         upgrade = src_entry == dst_entry;
1159
1160         src_object = src_entry->object.vm_object;
1161         src_pindex = OFF_TO_IDX(src_entry->offset);
1162         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1163
1164         /*
1165          * Create the top-level object for the destination entry. (Doesn't
1166          * actually shadow anything - we copy the pages directly.)
1167          */
1168         dst_object = vm_object_allocate(OBJT_DEFAULT,
1169             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1170 #if VM_NRESERVLEVEL > 0
1171         dst_object->flags |= OBJ_COLORED;
1172         dst_object->pg_color = atop(dst_entry->start);
1173 #endif
1174
1175         VM_OBJECT_LOCK(dst_object);
1176         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1177             ("vm_fault_copy_entry: vm_object not NULL"));
1178         dst_entry->object.vm_object = dst_object;
1179         dst_entry->offset = 0;
1180         dst_object->charge = dst_entry->end - dst_entry->start;
1181         if (fork_charge != NULL) {
1182                 KASSERT(dst_entry->uip == NULL,
1183                     ("vm_fault_copy_entry: leaked swp charge"));
1184                 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1185                 uihold(dst_object->uip);
1186                 *fork_charge += dst_object->charge;
1187         } else {
1188                 dst_object->uip = dst_entry->uip;
1189                 dst_entry->uip = NULL;
1190         }
1191         access = prot = dst_entry->max_protection;
1192         /*
1193          * If not an upgrade, then enter the mappings in the pmap as
1194          * read and/or execute accesses.  Otherwise, enter them as
1195          * write accesses.
1196          *
1197          * A writeable large page mapping is only created if all of
1198          * the constituent small page mappings are modified. Marking
1199          * PTEs as modified on inception allows promotion to happen
1200          * without taking potentially large number of soft faults.
1201          */
1202         if (!upgrade)
1203                 access &= ~VM_PROT_WRITE;
1204
1205         /*
1206          * Loop through all of the pages in the entry's range, copying each
1207          * one from the source object (it should be there) to the destination
1208          * object.
1209          */
1210         for (vaddr = dst_entry->start, dst_pindex = 0;
1211             vaddr < dst_entry->end;
1212             vaddr += PAGE_SIZE, dst_pindex++) {
1213
1214                 /*
1215                  * Allocate a page in the destination object.
1216                  */
1217                 do {
1218                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1219                             VM_ALLOC_NORMAL);
1220                         if (dst_m == NULL) {
1221                                 VM_OBJECT_UNLOCK(dst_object);
1222                                 VM_WAIT;
1223                                 VM_OBJECT_LOCK(dst_object);
1224                         }
1225                 } while (dst_m == NULL);
1226
1227                 /*
1228                  * Find the page in the source object, and copy it in.
1229                  * (Because the source is wired down, the page will be in
1230                  * memory.)
1231                  */
1232                 VM_OBJECT_LOCK(src_object);
1233                 object = src_object;
1234                 pindex = src_pindex + dst_pindex;
1235                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1236                     src_readonly &&
1237                     (backing_object = object->backing_object) != NULL) {
1238                         /*
1239                          * Allow fallback to backing objects if we are reading.
1240                          */
1241                         VM_OBJECT_LOCK(backing_object);
1242                         pindex += OFF_TO_IDX(object->backing_object_offset);
1243                         VM_OBJECT_UNLOCK(object);
1244                         object = backing_object;
1245                 }
1246                 if (src_m == NULL)
1247                         panic("vm_fault_copy_wired: page missing");
1248                 pmap_copy_page(src_m, dst_m);
1249                 VM_OBJECT_UNLOCK(object);
1250                 dst_m->valid = VM_PAGE_BITS_ALL;
1251                 VM_OBJECT_UNLOCK(dst_object);
1252
1253                 /*
1254                  * Enter it in the pmap. If a wired, copy-on-write
1255                  * mapping is being replaced by a write-enabled
1256                  * mapping, then wire that new mapping.
1257                  */
1258                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1259
1260                 /*
1261                  * Mark it no longer busy, and put it on the active list.
1262                  */
1263                 VM_OBJECT_LOCK(dst_object);
1264                 vm_page_lock_queues();
1265                 if (upgrade) {
1266                         vm_page_unwire(src_m, 0);
1267                         vm_page_wire(dst_m);
1268                 } else
1269                         vm_page_activate(dst_m);
1270                 vm_page_unlock_queues();
1271                 vm_page_wakeup(dst_m);
1272         }
1273         VM_OBJECT_UNLOCK(dst_object);
1274         if (upgrade) {
1275                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1276                 vm_object_deallocate(src_object);
1277         }
1278 }
1279
1280
1281 /*
1282  * This routine checks around the requested page for other pages that
1283  * might be able to be faulted in.  This routine brackets the viable
1284  * pages for the pages to be paged in.
1285  *
1286  * Inputs:
1287  *      m, rbehind, rahead
1288  *
1289  * Outputs:
1290  *  marray (array of vm_page_t), reqpage (index of requested page)
1291  *
1292  * Return value:
1293  *  number of pages in marray
1294  */
1295 static int
1296 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1297         vm_page_t m;
1298         int rbehind;
1299         int rahead;
1300         vm_page_t *marray;
1301         int *reqpage;
1302 {
1303         int i,j;
1304         vm_object_t object;
1305         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1306         vm_page_t rtm;
1307         int cbehind, cahead;
1308
1309         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1310
1311         object = m->object;
1312         pindex = m->pindex;
1313         cbehind = cahead = 0;
1314
1315         /*
1316          * if the requested page is not available, then give up now
1317          */
1318         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1319                 return 0;
1320         }
1321
1322         if ((cbehind == 0) && (cahead == 0)) {
1323                 *reqpage = 0;
1324                 marray[0] = m;
1325                 return 1;
1326         }
1327
1328         if (rahead > cahead) {
1329                 rahead = cahead;
1330         }
1331
1332         if (rbehind > cbehind) {
1333                 rbehind = cbehind;
1334         }
1335
1336         /*
1337          * scan backward for the read behind pages -- in memory
1338          */
1339         if (pindex > 0) {
1340                 if (rbehind > pindex) {
1341                         rbehind = pindex;
1342                         startpindex = 0;
1343                 } else {
1344                         startpindex = pindex - rbehind;
1345                 }
1346
1347                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1348                     rtm->pindex >= startpindex)
1349                         startpindex = rtm->pindex + 1;
1350
1351                 /* tpindex is unsigned; beware of numeric underflow. */
1352                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1353                     tpindex < pindex; i++, tpindex--) {
1354
1355                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1356                             VM_ALLOC_IFNOTCACHED);
1357                         if (rtm == NULL) {
1358                                 /*
1359                                  * Shift the allocated pages to the
1360                                  * beginning of the array.
1361                                  */
1362                                 for (j = 0; j < i; j++) {
1363                                         marray[j] = marray[j + tpindex + 1 -
1364                                             startpindex];
1365                                 }
1366                                 break;
1367                         }
1368
1369                         marray[tpindex - startpindex] = rtm;
1370                 }
1371         } else {
1372                 startpindex = 0;
1373                 i = 0;
1374         }
1375
1376         marray[i] = m;
1377         /* page offset of the required page */
1378         *reqpage = i;
1379
1380         tpindex = pindex + 1;
1381         i++;
1382
1383         /*
1384          * scan forward for the read ahead pages
1385          */
1386         endpindex = tpindex + rahead;
1387         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1388                 endpindex = rtm->pindex;
1389         if (endpindex > object->size)
1390                 endpindex = object->size;
1391
1392         for (; tpindex < endpindex; i++, tpindex++) {
1393
1394                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1395                     VM_ALLOC_IFNOTCACHED);
1396                 if (rtm == NULL) {
1397                         break;
1398                 }
1399
1400                 marray[i] = rtm;
1401         }
1402
1403         /* return number of pages */
1404         return i;
1405 }