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], mt, mt_prev;
 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                                 if (fs.object != fs.first_object) {
 365                                         if (!VM_OBJECT_TRYLOCK(
 366                                             fs.first_object)) {
 367                                                 VM_OBJECT_UNLOCK(fs.object);
 368                                                 VM_OBJECT_LOCK(fs.first_object);
 369                                                 VM_OBJECT_LOCK(fs.object);
 370                                         }
 371                                         vm_page_lock_queues();
 372                                         vm_page_free(fs.first_m);
 373                                         vm_page_unlock_queues();
 374                                         vm_object_pip_wakeup(fs.first_object);
 375                                         VM_OBJECT_UNLOCK(fs.first_object);
 376                                         fs.first_m = NULL;
 377                                 }
 378                                 unlock_map(&fs);
 379                                 if (fs.m == vm_page_lookup(fs.object,
 380                                     fs.pindex)) {
 381                                         vm_page_sleep_if_busy(fs.m, TRUE,
 382                                             "vmpfw");
 383                                 }
 384                                 vm_object_pip_wakeup(fs.object);
 385                                 VM_OBJECT_UNLOCK(fs.object);
 386                                 PCPU_INC(cnt.v_intrans);
 387                                 vm_object_deallocate(fs.first_object);
 388                                 goto RetryFault;
 389                         }
 390                         vm_pageq_remove(fs.m);
 391                         vm_page_unlock_queues();
 392
 393                         /*
 394                          * Mark page busy for other processes, and the
 395                          * pagedaemon.  If it still isn't completely valid
 396                          * (readable), jump to readrest, else break-out ( we
 397                          * found the page ).
 398                          */
 399                         vm_page_busy(fs.m);
 400                         if (fs.m->valid != VM_PAGE_BITS_ALL &&
 401                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 402                                 goto readrest;
 403                         }
 404
 405                         break;
 406                 }
 407
 408                 /*
 409                  * Page is not resident, If this is the search termination
 410                  * or the pager might contain the page, allocate a new page.
 411                  */
 412                 if (TRYPAGER || fs.object == fs.first_object) {
 413                         if (fs.pindex >= fs.object->size) {
 414                                 unlock_and_deallocate(&fs);
 415                                 return (KERN_PROTECTION_FAILURE);
 416                         }
 417
 418                         /*
 419                          * Allocate a new page for this object/offset pair.
 420                          *
 421                          * Unlocked read of the p_flag is harmless. At
 422                          * worst, the P_KILLED might be not observed
 423                          * there, and allocation can fail, causing
 424                          * restart and new reading of the p_flag.
 425                          */
 426                         fs.m = NULL;
 427                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 428 #if VM_NRESERVLEVEL > 0
 429                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 430                                         fs.object->flags |= OBJ_COLORED;
 431                                         fs.object->pg_color = atop(vaddr) -
 432                                             fs.pindex;
 433                                 }
 434 #endif
 435                                 alloc_req = P_KILLED(curproc) ?
 436                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 437                                 if (fs.object->type != OBJT_VNODE &&
 438                                     fs.object->backing_object == NULL)
 439                                         alloc_req |= VM_ALLOC_ZERO;
 440                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 441                                     alloc_req);
 442                         }
 443                         if (fs.m == NULL) {
 444                                 unlock_and_deallocate(&fs);
 445                                 VM_WAITPFAULT;
 446                                 goto RetryFault;
 447                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 448                                 break;
 449                 }
 450
 451 readrest:
 452                 /*
 453                  * We have found a valid page or we have allocated a new page.
 454                  * The page thus may not be valid or may not be entirely
 455                  * valid.
 456                  *
 457                  * Attempt to fault-in the page if there is a chance that the
 458                  * pager has it, and potentially fault in additional pages
 459                  * at the same time.
 460                  */
 461                 if (TRYPAGER) {
 462                         int rv;
 463                         int reqpage = 0;
 464                         u_char behavior = vm_map_entry_behavior(fs.entry);
 465
 466                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 467                             P_KILLED(curproc)) {
 468                                 ahead = 0;
 469                                 behind = 0;
 470                         } else {
 471                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 472                                 if (behind > VM_FAULT_READ_BEHIND)
 473                                         behind = VM_FAULT_READ_BEHIND;
 474
 475                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 476                                 if (ahead > VM_FAULT_READ_AHEAD)
 477                                         ahead = VM_FAULT_READ_AHEAD;
 478                         }
 479                         is_first_object_locked = FALSE;
 480                         if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 481                              (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 482                               fs.pindex >= fs.entry->lastr &&
 483                               fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
 484                             (fs.first_object == fs.object ||
 485                              (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
 486                             fs.first_object->type != OBJT_DEVICE &&
 487                             fs.first_object->type != OBJT_PHYS &&
 488                             fs.first_object->type != OBJT_SG) {
 489                                 vm_pindex_t firstpindex;
 490
 491                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 492                                         firstpindex = 0;
 493                                 else
 494                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 495                                 mt = fs.first_object != fs.object ?
 496                                     fs.first_m : fs.m;
 497                                 KASSERT(mt != NULL, ("vm_fault: missing mt"));
 498                                 KASSERT((mt->oflags & VPO_BUSY) != 0,
 499                                     ("vm_fault: mt %p not busy", mt));
 500                                 mt_prev = vm_page_prev(mt);
 501
 502                                 are_queues_locked = FALSE;
 503                                 /*
 504                                  * note: partially valid pages cannot be
 505                                  * included in the lookahead - NFS piecemeal
 506                                  * writes will barf on it badly.
 507                                  */
 508                                 while ((mt = mt_prev) != NULL &&
 509                                     mt->pindex >= firstpindex &&
 510                                     mt->valid == VM_PAGE_BITS_ALL) {
 511                                         mt_prev = vm_page_prev(mt);
 512                                         if (mt->busy ||
 513                                             (mt->oflags & VPO_BUSY))
 514                                                 continue;
 515                                         if (!are_queues_locked) {
 516                                                 are_queues_locked = TRUE;
 517                                                 vm_page_lock_queues();
 518                                         }
 519                                         if (mt->hold_count ||
 520                                                 mt->wire_count)
 521                                                 continue;
 522                                         pmap_remove_all(mt);
 523                                         if (mt->dirty) {
 524                                                 vm_page_deactivate(mt);
 525                                         } else {
 526                                                 vm_page_cache(mt);
 527                                         }
 528                                 }
 529                                 if (are_queues_locked)
 530                                         vm_page_unlock_queues();
 531                                 ahead += behind;
 532                                 behind = 0;
 533                         }
 534                         if (is_first_object_locked)
 535                                 VM_OBJECT_UNLOCK(fs.first_object);
 536
 537                         /*
 538                          * Call the pager to retrieve the data, if any, after
 539                          * releasing the lock on the map.  We hold a ref on
 540                          * fs.object and the pages are VPO_BUSY'd.
 541                          */
 542                         unlock_map(&fs);
 543
 544 vnode_lock:
 545                         if (fs.object->type == OBJT_VNODE) {
 546                                 vp = fs.object->handle;
 547                                 if (vp == fs.vp)
 548                                         goto vnode_locked;
 549                                 else if (fs.vp != NULL) {
 550                                         vput(fs.vp);
 551                                         fs.vp = NULL;
 552                                 }
 553                                 locked = VOP_ISLOCKED(vp);
 554
 555                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 556                                         fs.vfslocked = 1;
 557                                         if (!mtx_trylock(&Giant)) {
 558                                                 VM_OBJECT_UNLOCK(fs.object);
 559                                                 mtx_lock(&Giant);
 560                                                 VM_OBJECT_LOCK(fs.object);
 561                                                 goto vnode_lock;
 562                                         }
 563                                 }
 564                                 if (locked != LK_EXCLUSIVE)
 565                                         locked = LK_SHARED;
 566                                 /* Do not sleep for vnode lock while fs.m is busy */
 567                                 error = vget(vp, locked | LK_CANRECURSE |
 568                                     LK_NOWAIT, curthread);
 569                                 if (error != 0) {
 570                                         int vfslocked;
 571
 572                                         vfslocked = fs.vfslocked;
 573                                         fs.vfslocked = 0; /* Keep Giant */
 574                                         vhold(vp);
 575                                         release_page(&fs);
 576                                         unlock_and_deallocate(&fs);
 577                                         error = vget(vp, locked | LK_RETRY |
 578                                             LK_CANRECURSE, curthread);
 579                                         vdrop(vp);
 580                                         fs.vp = vp;
 581                                         fs.vfslocked = vfslocked;
 582                                         KASSERT(error == 0,
 583                                             ("vm_fault: vget failed"));
 584                                         goto RetryFault;
 585                                 }
 586                                 fs.vp = vp;
 587                         }
 588 vnode_locked:
 589                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 590                             ("vm_fault: vnode-backed object mapped by system map"));
 591
 592                         /*
 593                          * now we find out if any other pages should be paged
 594                          * in at this time this routine checks to see if the
 595                          * pages surrounding this fault reside in the same
 596                          * object as the page for this fault.  If they do,
 597                          * then they are faulted in also into the object.  The
 598                          * array "marray" returned contains an array of
 599                          * vm_page_t structs where one of them is the
 600                          * vm_page_t passed to the routine.  The reqpage
 601                          * return value is the index into the marray for the
 602                          * vm_page_t passed to the routine.
 603                          *
 604                          * fs.m plus the additional pages are VPO_BUSY'd.
 605                          */
 606                         faultcount = vm_fault_additional_pages(
 607                             fs.m, behind, ahead, marray, &reqpage);
 608
 609                         rv = faultcount ?
 610                             vm_pager_get_pages(fs.object, marray, faultcount,
 611                                 reqpage) : VM_PAGER_FAIL;
 612
 613                         if (rv == VM_PAGER_OK) {
 614                                 /*
 615                                  * Found the page. Leave it busy while we play
 616                                  * with it.
 617                                  */
 618
 619                                 /*
 620                                  * Relookup in case pager changed page. Pager
 621                                  * is responsible for disposition of old page
 622                                  * if moved.
 623                                  */
 624                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 625                                 if (!fs.m) {
 626                                         unlock_and_deallocate(&fs);
 627                                         goto RetryFault;
 628                                 }
 629
 630                                 hardfault++;
 631                                 break; /* break to PAGE HAS BEEN FOUND */
 632                         }
 633                         /*
 634                          * Remove the bogus page (which does not exist at this
 635                          * object/offset); before doing so, we must get back
 636                          * our object lock to preserve our invariant.
 637                          *
 638                          * Also wake up any other process that may want to bring
 639                          * in this page.
 640                          *
 641                          * If this is the top-level object, we must leave the
 642                          * busy page to prevent another process from rushing
 643                          * past us, and inserting the page in that object at
 644                          * the same time that we are.
 645                          */
 646                         if (rv == VM_PAGER_ERROR)
 647                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 648                                     curproc->p_pid, curproc->p_comm);
 649                         /*
 650                          * Data outside the range of the pager or an I/O error
 651                          */
 652                         /*
 653                          * XXX - the check for kernel_map is a kludge to work
 654                          * around having the machine panic on a kernel space
 655                          * fault w/ I/O error.
 656                          */
 657                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 658                                 (rv == VM_PAGER_BAD)) {
 659                                 vm_page_lock_queues();
 660                                 vm_page_free(fs.m);
 661                                 vm_page_unlock_queues();
 662                                 fs.m = NULL;
 663                                 unlock_and_deallocate(&fs);
 664                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 665                         }
 666                         if (fs.object != fs.first_object) {
 667                                 vm_page_lock_queues();
 668                                 vm_page_free(fs.m);
 669                                 vm_page_unlock_queues();
 670                                 fs.m = NULL;
 671                                 /*
 672                                  * XXX - we cannot just fall out at this
 673                                  * point, m has been freed and is invalid!
 674                                  */
 675                         }
 676                 }
 677
 678                 /*
 679                  * We get here if the object has default pager (or unwiring)
 680                  * or the pager doesn't have the page.
 681                  */
 682                 if (fs.object == fs.first_object)
 683                         fs.first_m = fs.m;
 684
 685                 /*
 686                  * Move on to the next object.  Lock the next object before
 687                  * unlocking the current one.
 688                  */
 689                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 690                 next_object = fs.object->backing_object;
 691                 if (next_object == NULL) {
 692                         /*
 693                          * If there's no object left, fill the page in the top
 694                          * object with zeros.
 695                          */
 696                         if (fs.object != fs.first_object) {
 697                                 vm_object_pip_wakeup(fs.object);
 698                                 VM_OBJECT_UNLOCK(fs.object);
 699
 700                                 fs.object = fs.first_object;
 701                                 fs.pindex = fs.first_pindex;
 702                                 fs.m = fs.first_m;
 703                                 VM_OBJECT_LOCK(fs.object);
 704                         }
 705                         fs.first_m = NULL;
 706
 707                         /*
 708                          * Zero the page if necessary and mark it valid.
 709                          */
 710                         if ((fs.m->flags & PG_ZERO) == 0) {
 711                                 pmap_zero_page(fs.m);
 712                         } else {
 713                                 PCPU_INC(cnt.v_ozfod);
 714                         }
 715                         PCPU_INC(cnt.v_zfod);
 716                         fs.m->valid = VM_PAGE_BITS_ALL;
 717                         break;  /* break to PAGE HAS BEEN FOUND */
 718                 } else {
 719                         KASSERT(fs.object != next_object,
 720                             ("object loop %p", next_object));
 721                         VM_OBJECT_LOCK(next_object);
 722                         vm_object_pip_add(next_object, 1);
 723                         if (fs.object != fs.first_object)
 724                                 vm_object_pip_wakeup(fs.object);
 725                         VM_OBJECT_UNLOCK(fs.object);
 726                         fs.object = next_object;
 727                 }
 728         }
 729
 730         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 731             ("vm_fault: not busy after main loop"));
 732
 733         /*
 734          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 735          * is held.]
 736          */
 737
 738         /*
 739          * If the page is being written, but isn't already owned by the
 740          * top-level object, we have to copy it into a new page owned by the
 741          * top-level object.
 742          */
 743         if (fs.object != fs.first_object) {
 744                 /*
 745                  * We only really need to copy if we want to write it.
 746                  */
 747                 if (fault_type & VM_PROT_WRITE) {
 748                         /*
 749                          * This allows pages to be virtually copied from a
 750                          * backing_object into the first_object, where the
 751                          * backing object has no other refs to it, and cannot
 752                          * gain any more refs.  Instead of a bcopy, we just
 753                          * move the page from the backing object to the
 754                          * first object.  Note that we must mark the page
 755                          * dirty in the first object so that it will go out
 756                          * to swap when needed.
 757                          */
 758                         is_first_object_locked = FALSE;
 759                         if (
 760                                 /*
 761                                  * Only one shadow object
 762                                  */
 763                                 (fs.object->shadow_count == 1) &&
 764                                 /*
 765                                  * No COW refs, except us
 766                                  */
 767                                 (fs.object->ref_count == 1) &&
 768                                 /*
 769                                  * No one else can look this object up
 770                                  */
 771                                 (fs.object->handle == NULL) &&
 772                                 /*
 773                                  * No other ways to look the object up
 774                                  */
 775                                 ((fs.object->type == OBJT_DEFAULT) ||
 776                                  (fs.object->type == OBJT_SWAP)) &&
 777                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 778                                 /*
 779                                  * We don't chase down the shadow chain
 780                                  */
 781                             fs.object == fs.first_object->backing_object) {
 782                                 vm_page_lock_queues();
 783                                 /*
 784                                  * get rid of the unnecessary page
 785                                  */
 786                                 vm_page_free(fs.first_m);
 787                                 /*
 788                                  * grab the page and put it into the
 789                                  * process'es object.  The page is
 790                                  * automatically made dirty.
 791                                  */
 792                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 793                                 vm_page_unlock_queues();
 794                                 vm_page_busy(fs.m);
 795                                 fs.first_m = fs.m;
 796                                 fs.m = NULL;
 797                                 PCPU_INC(cnt.v_cow_optim);
 798                         } else {
 799                                 /*
 800                                  * Oh, well, lets copy it.
 801                                  */
 802                                 pmap_copy_page(fs.m, fs.first_m);
 803                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 804                         }
 805                         if (fs.m) {
 806                                 /*
 807                                  * We no longer need the old page or object.
 808                                  */
 809                                 release_page(&fs);
 810                         }
 811                         /*
 812                          * fs.object != fs.first_object due to above
 813                          * conditional
 814                          */
 815                         vm_object_pip_wakeup(fs.object);
 816                         VM_OBJECT_UNLOCK(fs.object);
 817                         /*
 818                          * Only use the new page below...
 819                          */
 820                         fs.object = fs.first_object;
 821                         fs.pindex = fs.first_pindex;
 822                         fs.m = fs.first_m;
 823                         if (!is_first_object_locked)
 824                                 VM_OBJECT_LOCK(fs.object);
 825                         PCPU_INC(cnt.v_cow_faults);
 826                 } else {
 827                         prot &= ~VM_PROT_WRITE;
 828                 }
 829         }
 830
 831         /*
 832          * We must verify that the maps have not changed since our last
 833          * lookup.
 834          */
 835         if (!fs.lookup_still_valid) {
 836                 vm_object_t retry_object;
 837                 vm_pindex_t retry_pindex;
 838                 vm_prot_t retry_prot;
 839
 840                 if (!vm_map_trylock_read(fs.map)) {
 841                         release_page(&fs);
 842                         unlock_and_deallocate(&fs);
 843                         goto RetryFault;
 844                 }
 845                 fs.lookup_still_valid = TRUE;
 846                 if (fs.map->timestamp != map_generation) {
 847                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 848                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 849
 850                         /*
 851                          * If we don't need the page any longer, put it on the inactive
 852                          * list (the easiest thing to do here).  If no one needs it,
 853                          * pageout will grab it eventually.
 854                          */
 855                         if (result != KERN_SUCCESS) {
 856                                 release_page(&fs);
 857                                 unlock_and_deallocate(&fs);
 858
 859                                 /*
 860                                  * If retry of map lookup would have blocked then
 861                                  * retry fault from start.
 862                                  */
 863                                 if (result == KERN_FAILURE)
 864                                         goto RetryFault;
 865                                 return (result);
 866                         }
 867                         if ((retry_object != fs.first_object) ||
 868                             (retry_pindex != fs.first_pindex)) {
 869                                 release_page(&fs);
 870                                 unlock_and_deallocate(&fs);
 871                                 goto RetryFault;
 872                         }
 873
 874                         /*
 875                          * Check whether the protection has changed or the object has
 876                          * been copied while we left the map unlocked. Changing from
 877                          * read to write permission is OK - we leave the page
 878                          * write-protected, and catch the write fault. Changing from
 879                          * write to read permission means that we can't mark the page
 880                          * write-enabled after all.
 881                          */
 882                         prot &= retry_prot;
 883                 }
 884         }
 885         /*
 886          * If the page was filled by a pager, update the map entry's
 887          * last read offset.  Since the pager does not return the
 888          * actual set of pages that it read, this update is based on
 889          * the requested set.  Typically, the requested and actual
 890          * sets are the same.
 891          *
 892          * XXX The following assignment modifies the map
 893          * without holding a write lock on it.
 894          */
 895         if (hardfault)
 896                 fs.entry->lastr = fs.pindex + faultcount - behind;
 897
 898         if (prot & VM_PROT_WRITE) {
 899                 vm_object_set_writeable_dirty(fs.object);
 900
 901                 /*
 902                  * If the fault is a write, we know that this page is being
 903                  * written NOW so dirty it explicitly to save on
 904                  * pmap_is_modified() calls later.
 905                  *
 906                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 907                  * if the page is already dirty to prevent data written with
 908                  * the expectation of being synced from not being synced.
 909                  * Likewise if this entry does not request NOSYNC then make
 910                  * sure the page isn't marked NOSYNC.  Applications sharing
 911                  * data should use the same flags to avoid ping ponging.
 912                  *
 913                  * Also tell the backing pager, if any, that it should remove
 914                  * any swap backing since the page is now dirty.
 915                  */
 916                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 917                         if (fs.m->dirty == 0)
 918                                 fs.m->oflags |= VPO_NOSYNC;
 919                 } else {
 920                         fs.m->oflags &= ~VPO_NOSYNC;
 921                 }
 922                 if (fault_flags & VM_FAULT_DIRTY) {
 923                         vm_page_dirty(fs.m);
 924                         vm_pager_page_unswapped(fs.m);
 925                 }
 926         }
 927
 928         /*
 929          * Page had better still be busy
 930          */
 931         KASSERT(fs.m->oflags & VPO_BUSY,
 932                 ("vm_fault: page %p not busy!", fs.m));
 933         /*
 934          * Page must be completely valid or it is not fit to
 935          * map into user space.  vm_pager_get_pages() ensures this.
 936          */
 937         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 938             ("vm_fault: page %p partially invalid", fs.m));
 939         VM_OBJECT_UNLOCK(fs.object);
 940
 941         /*
 942          * Put this page into the physical map.  We had to do the unlock above
 943          * because pmap_enter() may sleep.  We don't put the page
 944          * back on the active queue until later so that the pageout daemon
 945          * won't find it (yet).
 946          */
 947         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 948         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 949                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 950         }
 951         VM_OBJECT_LOCK(fs.object);
 952         vm_page_lock_queues();
 953         vm_page_flag_set(fs.m, PG_REFERENCED);
 954
 955         /*
 956          * If the page is not wired down, then put it where the pageout daemon
 957          * can find it.
 958          */
 959         if (fault_flags & VM_FAULT_WIRE_MASK) {
 960                 if (wired)
 961                         vm_page_wire(fs.m);
 962                 else
 963                         vm_page_unwire(fs.m, 1);
 964         } else {
 965                 vm_page_activate(fs.m);
 966         }
 967         vm_page_unlock_queues();
 968         vm_page_wakeup(fs.m);
 969
 970         /*
 971          * Unlock everything, and return
 972          */
 973         unlock_and_deallocate(&fs);
 974         if (hardfault)
 975                 curthread->td_ru.ru_majflt++;
 976         else
 977                 curthread->td_ru.ru_minflt++;
 978
 979         return (KERN_SUCCESS);
 980 }
 981
 982 /*
 983  * vm_fault_prefault provides a quick way of clustering
 984  * pagefaults into a processes address space.  It is a "cousin"
 985  * of vm_map_pmap_enter, except it runs at page fault time instead
 986  * of mmap time.
 987  */
 988 static void
 989 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
 990 {
 991         int i;
 992         vm_offset_t addr, starta;
 993         vm_pindex_t pindex;
 994         vm_page_t m;
 995         vm_object_t object;
 996
 997         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
 998                 return;
 999
1000         object = entry->object.vm_object;
1001
1002         starta = addra - PFBAK * PAGE_SIZE;
1003         if (starta < entry->start) {
1004                 starta = entry->start;
1005         } else if (starta > addra) {
1006                 starta = 0;
1007         }
1008
1009         for (i = 0; i < PAGEORDER_SIZE; i++) {
1010                 vm_object_t backing_object, lobject;
1011
1012                 addr = addra + prefault_pageorder[i];
1013                 if (addr > addra + (PFFOR * PAGE_SIZE))
1014                         addr = 0;
1015
1016                 if (addr < starta || addr >= entry->end)
1017                         continue;
1018
1019                 if (!pmap_is_prefaultable(pmap, addr))
1020                         continue;
1021
1022                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1023                 lobject = object;
1024                 VM_OBJECT_LOCK(lobject);
1025                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1026                     lobject->type == OBJT_DEFAULT &&
1027                     (backing_object = lobject->backing_object) != NULL) {
1028                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1029                             0, ("vm_fault_prefault: unaligned object offset"));
1030                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1031                         VM_OBJECT_LOCK(backing_object);
1032                         VM_OBJECT_UNLOCK(lobject);
1033                         lobject = backing_object;
1034                 }
1035                 /*
1036                  * give-up when a page is not in memory
1037                  */
1038                 if (m == NULL) {
1039                         VM_OBJECT_UNLOCK(lobject);
1040                         break;
1041                 }
1042                 if (m->valid == VM_PAGE_BITS_ALL &&
1043                     (m->flags & PG_FICTITIOUS) == 0) {
1044                         vm_page_lock_queues();
1045                         pmap_enter_quick(pmap, addr, m, entry->protection);
1046                         vm_page_unlock_queues();
1047                 }
1048                 VM_OBJECT_UNLOCK(lobject);
1049         }
1050 }
1051
1052 /*
1053  *      vm_fault_quick:
1054  *
1055  *      Ensure that the requested virtual address, which may be in userland,
1056  *      is valid.  Fault-in the page if necessary.  Return -1 on failure.
1057  */
1058 int
1059 vm_fault_quick(caddr_t v, int prot)
1060 {
1061         int r;
1062
1063         if (prot & VM_PROT_WRITE)
1064                 r = subyte(v, fubyte(v));
1065         else
1066                 r = fubyte(v);
1067         return(r);
1068 }
1069
1070 /*
1071  *      vm_fault_wire:
1072  *
1073  *      Wire down a range of virtual addresses in a map.
1074  */
1075 int
1076 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1077     boolean_t user_wire, boolean_t fictitious)
1078 {
1079         vm_offset_t va;
1080         int rv;
1081
1082         /*
1083          * We simulate a fault to get the page and enter it in the physical
1084          * map.  For user wiring, we only ask for read access on currently
1085          * read-only sections.
1086          */
1087         for (va = start; va < end; va += PAGE_SIZE) {
1088                 rv = vm_fault(map, va,
1089                     user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1090                     user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1091                 if (rv) {
1092                         if (va != start)
1093                                 vm_fault_unwire(map, start, va, fictitious);
1094                         return (rv);
1095                 }
1096         }
1097         return (KERN_SUCCESS);
1098 }
1099
1100 /*
1101  *      vm_fault_unwire:
1102  *
1103  *      Unwire a range of virtual addresses in a map.
1104  */
1105 void
1106 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1107     boolean_t fictitious)
1108 {
1109         vm_paddr_t pa;
1110         vm_offset_t va;
1111         pmap_t pmap;
1112
1113         pmap = vm_map_pmap(map);
1114
1115         /*
1116          * Since the pages are wired down, we must be able to get their
1117          * mappings from the physical map system.
1118          */
1119         for (va = start; va < end; va += PAGE_SIZE) {
1120                 pa = pmap_extract(pmap, va);
1121                 if (pa != 0) {
1122                         pmap_change_wiring(pmap, va, FALSE);
1123                         if (!fictitious) {
1124                                 vm_page_lock_queues();
1125                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1126                                 vm_page_unlock_queues();
1127                         }
1128                 }
1129         }
1130 }
1131
1132 /*
1133  *      Routine:
1134  *              vm_fault_copy_entry
1135  *      Function:
1136  *              Create new shadow object backing dst_entry with private copy of
1137  *              all underlying pages. When src_entry is equal to dst_entry,
1138  *              function implements COW for wired-down map entry. Otherwise,
1139  *              it forks wired entry into dst_map.
1140  *
1141  *      In/out conditions:
1142  *              The source and destination maps must be locked for write.
1143  *              The source map entry must be wired down (or be a sharing map
1144  *              entry corresponding to a main map entry that is wired down).
1145  */
1146 void
1147 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1148     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1149     vm_ooffset_t *fork_charge)
1150 {
1151         vm_object_t backing_object, dst_object, object, src_object;
1152         vm_pindex_t dst_pindex, pindex, src_pindex;
1153         vm_prot_t access, prot;
1154         vm_offset_t vaddr;
1155         vm_page_t dst_m;
1156         vm_page_t src_m;
1157         boolean_t src_readonly, upgrade;
1158
1159 #ifdef  lint
1160         src_map++;
1161 #endif  /* lint */
1162
1163         upgrade = src_entry == dst_entry;
1164
1165         src_object = src_entry->object.vm_object;
1166         src_pindex = OFF_TO_IDX(src_entry->offset);
1167         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1168
1169         /*
1170          * Create the top-level object for the destination entry. (Doesn't
1171          * actually shadow anything - we copy the pages directly.)
1172          */
1173         dst_object = vm_object_allocate(OBJT_DEFAULT,
1174             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1175 #if VM_NRESERVLEVEL > 0
1176         dst_object->flags |= OBJ_COLORED;
1177         dst_object->pg_color = atop(dst_entry->start);
1178 #endif
1179
1180         VM_OBJECT_LOCK(dst_object);
1181         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1182             ("vm_fault_copy_entry: vm_object not NULL"));
1183         dst_entry->object.vm_object = dst_object;
1184         dst_entry->offset = 0;
1185         dst_object->charge = dst_entry->end - dst_entry->start;
1186         if (fork_charge != NULL) {
1187                 KASSERT(dst_entry->uip == NULL,
1188                     ("vm_fault_copy_entry: leaked swp charge"));
1189                 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1190                 uihold(dst_object->uip);
1191                 *fork_charge += dst_object->charge;
1192         } else {
1193                 dst_object->uip = dst_entry->uip;
1194                 dst_entry->uip = NULL;
1195         }
1196         access = prot = dst_entry->max_protection;
1197         /*
1198          * If not an upgrade, then enter the mappings in the pmap as
1199          * read and/or execute accesses.  Otherwise, enter them as
1200          * write accesses.
1201          *
1202          * A writeable large page mapping is only created if all of
1203          * the constituent small page mappings are modified. Marking
1204          * PTEs as modified on inception allows promotion to happen
1205          * without taking potentially large number of soft faults.
1206          */
1207         if (!upgrade)
1208                 access &= ~VM_PROT_WRITE;
1209
1210         /*
1211          * Loop through all of the pages in the entry's range, copying each
1212          * one from the source object (it should be there) to the destination
1213          * object.
1214          */
1215         for (vaddr = dst_entry->start, dst_pindex = 0;
1216             vaddr < dst_entry->end;
1217             vaddr += PAGE_SIZE, dst_pindex++) {
1218
1219                 /*
1220                  * Allocate a page in the destination object.
1221                  */
1222                 do {
1223                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1224                             VM_ALLOC_NORMAL);
1225                         if (dst_m == NULL) {
1226                                 VM_OBJECT_UNLOCK(dst_object);
1227                                 VM_WAIT;
1228                                 VM_OBJECT_LOCK(dst_object);
1229                         }
1230                 } while (dst_m == NULL);
1231
1232                 /*
1233                  * Find the page in the source object, and copy it in.
1234                  * (Because the source is wired down, the page will be in
1235                  * memory.)
1236                  */
1237                 VM_OBJECT_LOCK(src_object);
1238                 object = src_object;
1239                 pindex = src_pindex + dst_pindex;
1240                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1241                     src_readonly &&
1242                     (backing_object = object->backing_object) != NULL) {
1243                         /*
1244                          * Allow fallback to backing objects if we are reading.
1245                          */
1246                         VM_OBJECT_LOCK(backing_object);
1247                         pindex += OFF_TO_IDX(object->backing_object_offset);
1248                         VM_OBJECT_UNLOCK(object);
1249                         object = backing_object;
1250                 }
1251                 if (src_m == NULL)
1252                         panic("vm_fault_copy_wired: page missing");
1253                 pmap_copy_page(src_m, dst_m);
1254                 VM_OBJECT_UNLOCK(object);
1255                 dst_m->valid = VM_PAGE_BITS_ALL;
1256                 VM_OBJECT_UNLOCK(dst_object);
1257
1258                 /*
1259                  * Enter it in the pmap. If a wired, copy-on-write
1260                  * mapping is being replaced by a write-enabled
1261                  * mapping, then wire that new mapping.
1262                  */
1263                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1264
1265                 /*
1266                  * Mark it no longer busy, and put it on the active list.
1267                  */
1268                 VM_OBJECT_LOCK(dst_object);
1269                 vm_page_lock_queues();
1270                 if (upgrade) {
1271                         vm_page_unwire(src_m, 0);
1272                         vm_page_wire(dst_m);
1273                 } else
1274                         vm_page_activate(dst_m);
1275                 vm_page_unlock_queues();
1276                 vm_page_wakeup(dst_m);
1277         }
1278         VM_OBJECT_UNLOCK(dst_object);
1279         if (upgrade) {
1280                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1281                 vm_object_deallocate(src_object);
1282         }
1283 }
1284
1285
1286 /*
1287  * This routine checks around the requested page for other pages that
1288  * might be able to be faulted in.  This routine brackets the viable
1289  * pages for the pages to be paged in.
1290  *
1291  * Inputs:
1292  *      m, rbehind, rahead
1293  *
1294  * Outputs:
1295  *  marray (array of vm_page_t), reqpage (index of requested page)
1296  *
1297  * Return value:
1298  *  number of pages in marray
1299  */
1300 static int
1301 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1302         vm_page_t m;
1303         int rbehind;
1304         int rahead;
1305         vm_page_t *marray;
1306         int *reqpage;
1307 {
1308         int i,j;
1309         vm_object_t object;
1310         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1311         vm_page_t rtm;
1312         int cbehind, cahead;
1313
1314         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1315
1316         object = m->object;
1317         pindex = m->pindex;
1318         cbehind = cahead = 0;
1319
1320         /*
1321          * if the requested page is not available, then give up now
1322          */
1323         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1324                 return 0;
1325         }
1326
1327         if ((cbehind == 0) && (cahead == 0)) {
1328                 *reqpage = 0;
1329                 marray[0] = m;
1330                 return 1;
1331         }
1332
1333         if (rahead > cahead) {
1334                 rahead = cahead;
1335         }
1336
1337         if (rbehind > cbehind) {
1338                 rbehind = cbehind;
1339         }
1340
1341         /*
1342          * scan backward for the read behind pages -- in memory
1343          */
1344         if (pindex > 0) {
1345                 if (rbehind > pindex) {
1346                         rbehind = pindex;
1347                         startpindex = 0;
1348                 } else {
1349                         startpindex = pindex - rbehind;
1350                 }
1351
1352                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1353                     rtm->pindex >= startpindex)
1354                         startpindex = rtm->pindex + 1;
1355
1356                 /* tpindex is unsigned; beware of numeric underflow. */
1357                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1358                     tpindex < pindex; i++, tpindex--) {
1359
1360                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1361                             VM_ALLOC_IFNOTCACHED);
1362                         if (rtm == NULL) {
1363                                 /*
1364                                  * Shift the allocated pages to the
1365                                  * beginning of the array.
1366                                  */
1367                                 for (j = 0; j < i; j++) {
1368                                         marray[j] = marray[j + tpindex + 1 -
1369                                             startpindex];
1370                                 }
1371                                 break;
1372                         }
1373
1374                         marray[tpindex - startpindex] = rtm;
1375                 }
1376         } else {
1377                 startpindex = 0;
1378                 i = 0;
1379         }
1380
1381         marray[i] = m;
1382         /* page offset of the required page */
1383         *reqpage = i;
1384
1385         tpindex = pindex + 1;
1386         i++;
1387
1388         /*
1389          * scan forward for the read ahead pages
1390          */
1391         endpindex = tpindex + rahead;
1392         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1393                 endpindex = rtm->pindex;
1394         if (endpindex > object->size)
1395                 endpindex = object->size;
1396
1397         for (; tpindex < endpindex; i++, tpindex++) {
1398
1399                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1400                     VM_ALLOC_IFNOTCACHED);
1401                 if (rtm == NULL) {
1402                         break;
1403                 }
1404
1405                 marray[i] = rtm;
1406         }
1407
1408         /* return number of pages */
1409         return i;
1410 }