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