sys/vm/vm_fault.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  *
   9  *
  10  * This code is derived from software contributed to Berkeley by
  11  * The Mach Operating System project at Carnegie-Mellon University.
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  * 1. Redistributions of source code must retain the above copyright
  17  *    notice, this list of conditions and the following disclaimer.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. All advertising materials mentioning features or use of this software
  22  *    must display the following acknowledgement:
  23  *      This product includes software developed by the University of
  24  *      California, Berkeley and its contributors.
  25  * 4. Neither the name of the University nor the names of its contributors
  26  *    may be used to endorse or promote products derived from this software
  27  *    without specific prior written permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  39  * SUCH DAMAGE.
  40  *
  41  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  42  *
  43  *
  44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  45  * All rights reserved.
  46  *
  47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  48  *
  49  * Permission to use, copy, modify and distribute this software and
  50  * its documentation is hereby granted, provided that both the copyright
  51  * notice and this permission notice appear in all copies of the
  52  * software, derivative works or modified versions, and any portions
  53  * thereof, and that both notices appear in supporting documentation.
  54  *
  55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  58  *
  59  * Carnegie Mellon requests users of this software to return to
  60  *
  61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  62  *  School of Computer Science
  63  *  Carnegie Mellon University
  64  *  Pittsburgh PA 15213-3890
  65  *
  66  * any improvements or extensions that they make and grant Carnegie the
  67  * rights to redistribute these changes.
  68  */
  69
  70 /*
  71  *      Page fault handling module.
  72  */
  73
  74 #include <sys/cdefs.h>
  75 __FBSDID("$FreeBSD$");
  76
  77 #include "opt_vm.h"
  78
  79 #include <sys/param.h>
  80 #include <sys/systm.h>
  81 #include <sys/kernel.h>
  82 #include <sys/lock.h>
  83 #include <sys/mutex.h>
  84 #include <sys/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/sysctl.h>
  87 #include <sys/vmmeter.h>
  88 #include <sys/vnode.h>
  89
  90 #include <vm/vm.h>
  91 #include <vm/vm_param.h>
  92 #include <vm/pmap.h>
  93 #include <vm/vm_map.h>
  94 #include <vm/vm_object.h>
  95 #include <vm/vm_page.h>
  96 #include <vm/vm_pageout.h>
  97 #include <vm/vm_kern.h>
  98 #include <vm/vm_pager.h>
  99 #include <vm/vnode_pager.h>
 100 #include <vm/vm_extern.h>
 101
 102 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 103
 104 #define PFBAK 4
 105 #define PFFOR 4
 106 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 107
 108 static int prefault_pageorder[] = {
 109         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 110         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 111         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 112         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 113 };
 114
 115 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 116 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 117
 118 #define VM_FAULT_READ_AHEAD 8
 119 #define VM_FAULT_READ_BEHIND 7
 120 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 121
 122 struct faultstate {
 123         vm_page_t m;
 124         vm_object_t object;
 125         vm_pindex_t pindex;
 126         vm_page_t first_m;
 127         vm_object_t     first_object;
 128         vm_pindex_t first_pindex;
 129         vm_map_t map;
 130         vm_map_entry_t entry;
 131         int lookup_still_valid;
 132         struct vnode *vp;
 133         int vfslocked;
 134 };
 135
 136 static inline void
 137 release_page(struct faultstate *fs)
 138 {
 139
 140         vm_page_wakeup(fs->m);
 141         vm_page_lock_queues();
 142         vm_page_deactivate(fs->m);
 143         vm_page_unlock_queues();
 144         fs->m = NULL;
 145 }
 146
 147 static inline void
 148 unlock_map(struct faultstate *fs)
 149 {
 150
 151         if (fs->lookup_still_valid) {
 152                 vm_map_lookup_done(fs->map, fs->entry);
 153                 fs->lookup_still_valid = FALSE;
 154         }
 155 }
 156
 157 static void
 158 unlock_and_deallocate(struct faultstate *fs)
 159 {
 160
 161         vm_object_pip_wakeup(fs->object);
 162         VM_OBJECT_UNLOCK(fs->object);
 163         if (fs->object != fs->first_object) {
 164                 VM_OBJECT_LOCK(fs->first_object);
 165                 vm_page_lock_queues();
 166                 vm_page_free(fs->first_m);
 167                 vm_page_unlock_queues();
 168                 vm_object_pip_wakeup(fs->first_object);
 169                 VM_OBJECT_UNLOCK(fs->first_object);
 170                 fs->first_m = NULL;
 171         }
 172         vm_object_deallocate(fs->first_object);
 173         unlock_map(fs);
 174         if (fs->vp != NULL) {
 175                 vput(fs->vp);
 176                 fs->vp = NULL;
 177         }
 178         VFS_UNLOCK_GIANT(fs->vfslocked);
 179         fs->vfslocked = 0;
 180 }
 181
 182 /*
 183  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 184  *            current object *might* contain the page.
 185  *
 186  *            default objects are zero-fill, there is no real pager.
 187  */
 188 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 189                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 190
 191 /*
 192  *      vm_fault:
 193  *
 194  *      Handle a page fault occurring at the given address,
 195  *      requiring the given permissions, in the map specified.
 196  *      If successful, the page is inserted into the
 197  *      associated physical map.
 198  *
 199  *      NOTE: the given address should be truncated to the
 200  *      proper page address.
 201  *
 202  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 203  *      a standard error specifying why the fault is fatal is returned.
 204  *
 205  *
 206  *      The map in question must be referenced, and remains so.
 207  *      Caller may hold no locks.
 208  */
 209 int
 210 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 211          int fault_flags)
 212 {
 213         vm_prot_t prot;
 214         int is_first_object_locked, result;
 215         boolean_t are_queues_locked, growstack, wired;
 216         int map_generation;
 217         vm_object_t next_object;
 218         vm_page_t marray[VM_FAULT_READ];
 219         int hardfault;
 220         int faultcount, ahead, behind;
 221         struct faultstate fs;
 222         struct vnode *vp;
 223         int locked, error;
 224
 225         hardfault = 0;
 226         growstack = TRUE;
 227         PCPU_INC(cnt.v_vm_faults);
 228         fs.vp = NULL;
 229         fs.vfslocked = 0;
 230         faultcount = behind = 0;
 231
 232 RetryFault:;
 233
 234         /*
 235          * Find the backing store object and offset into it to begin the
 236          * search.
 237          */
 238         fs.map = map;
 239         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 240             &fs.first_object, &fs.first_pindex, &prot, &wired);
 241         if (result != KERN_SUCCESS) {
 242                 if (result != KERN_PROTECTION_FAILURE ||
 243                     (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
 244                         if (growstack && result == KERN_INVALID_ADDRESS &&
 245                             map != kernel_map && curproc != NULL) {
 246                                 result = vm_map_growstack(curproc, vaddr);
 247                                 if (result != KERN_SUCCESS)
 248                                         return (KERN_FAILURE);
 249                                 growstack = FALSE;
 250                                 goto RetryFault;
 251                         }
 252                         return (result);
 253                 }
 254
 255                 /*
 256                  * If we are user-wiring a r/w segment, and it is COW, then
 257                  * we need to do the COW operation.  Note that we don't COW
 258                  * currently RO sections now, because it is NOT desirable
 259                  * to COW .text.  We simply keep .text from ever being COW'ed
 260                  * and take the heat that one cannot debug wired .text sections.
 261                  */
 262                 result = vm_map_lookup(&fs.map, vaddr,
 263                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 264                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 265                 if (result != KERN_SUCCESS)
 266                         return (result);
 267
 268                 /*
 269                  * If we don't COW now, on a user wire, the user will never
 270                  * be able to write to the mapping.  If we don't make this
 271                  * restriction, the bookkeeping would be nearly impossible.
 272                  *
 273                  * XXX The following assignment modifies the map without
 274                  * holding a write lock on it.
 275                  */
 276                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 277                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 278         }
 279
 280         map_generation = fs.map->timestamp;
 281
 282         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 283                 panic("vm_fault: fault on nofault entry, addr: %lx",
 284                     (u_long)vaddr);
 285         }
 286
 287         /*
 288          * Make a reference to this object to prevent its disposal while we
 289          * are messing with it.  Once we have the reference, the map is free
 290          * to be diddled.  Since objects reference their shadows (and copies),
 291          * they will stay around as well.
 292          *
 293          * Bump the paging-in-progress count to prevent size changes (e.g.
 294          * truncation operations) during I/O.  This must be done after
 295          * obtaining the vnode lock in order to avoid possible deadlocks.
 296          */
 297         VM_OBJECT_LOCK(fs.first_object);
 298         vm_object_reference_locked(fs.first_object);
 299         vm_object_pip_add(fs.first_object, 1);
 300
 301         fs.lookup_still_valid = TRUE;
 302
 303         if (wired)
 304                 fault_type = prot;
 305
 306         fs.first_m = NULL;
 307
 308         /*
 309          * Search for the page at object/offset.
 310          */
 311         fs.object = fs.first_object;
 312         fs.pindex = fs.first_pindex;
 313         while (TRUE) {
 314                 /*
 315                  * If the object is dead, we stop here
 316                  */
 317                 if (fs.object->flags & OBJ_DEAD) {
 318                         unlock_and_deallocate(&fs);
 319                         return (KERN_PROTECTION_FAILURE);
 320                 }
 321
 322                 /*
 323                  * See if page is resident
 324                  */
 325                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 326                 if (fs.m != NULL) {
 327                         /*
 328                          * check for page-based copy on write.
 329                          * We check fs.object == fs.first_object so
 330                          * as to ensure the legacy COW mechanism is
 331                          * used when the page in question is part of
 332                          * a shadow object.  Otherwise, vm_page_cowfault()
 333                          * removes the page from the backing object,
 334                          * which is not what we want.
 335                          */
 336                         vm_page_lock_queues();
 337                         if ((fs.m->cow) &&
 338                             (fault_type & VM_PROT_WRITE) &&
 339                             (fs.object == fs.first_object)) {
 340                                 vm_page_cowfault(fs.m);
 341                                 vm_page_unlock_queues();
 342                                 unlock_and_deallocate(&fs);
 343                                 goto RetryFault;
 344                         }
 345
 346                         /*
 347                          * Wait/Retry if the page is busy.  We have to do this
 348                          * if the page is busy via either VPO_BUSY or
 349                          * vm_page_t->busy because the vm_pager may be using
 350                          * vm_page_t->busy for pageouts ( and even pageins if
 351                          * it is the vnode pager ), and we could end up trying
 352                          * to pagein and pageout the same page simultaneously.
 353                          *
 354                          * We can theoretically allow the busy case on a read
 355                          * fault if the page is marked valid, but since such
 356                          * pages are typically already pmap'd, putting that
 357                          * special case in might be more effort then it is
 358                          * worth.  We cannot under any circumstances mess
 359                          * around with a vm_page_t->busy page except, perhaps,
 360                          * to pmap it.
 361                          */
 362                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 363                                 vm_page_unlock_queues();
 364                                 VM_OBJECT_UNLOCK(fs.object);
 365                                 if (fs.object != fs.first_object) {
 366                                         VM_OBJECT_LOCK(fs.first_object);
 367                                         vm_page_lock_queues();
 368                                         vm_page_free(fs.first_m);
 369                                         vm_page_unlock_queues();
 370                                         vm_object_pip_wakeup(fs.first_object);
 371                                         VM_OBJECT_UNLOCK(fs.first_object);
 372                                         fs.first_m = NULL;
 373                                 }
 374                                 unlock_map(&fs);
 375                                 VM_OBJECT_LOCK(fs.object);
 376                                 if (fs.m == vm_page_lookup(fs.object,
 377                                     fs.pindex)) {
 378                                         vm_page_sleep_if_busy(fs.m, TRUE,
 379                                             "vmpfw");
 380                                 }
 381                                 vm_object_pip_wakeup(fs.object);
 382                                 VM_OBJECT_UNLOCK(fs.object);
 383                                 PCPU_INC(cnt.v_intrans);
 384                                 vm_object_deallocate(fs.first_object);
 385                                 goto RetryFault;
 386                         }
 387                         vm_pageq_remove(fs.m);
 388                         vm_page_unlock_queues();
 389
 390                         /*
 391                          * Mark page busy for other processes, and the
 392                          * pagedaemon.  If it still isn't completely valid
 393                          * (readable), jump to readrest, else break-out ( we
 394                          * found the page ).
 395                          */
 396                         vm_page_busy(fs.m);
 397                         if (fs.m->valid != VM_PAGE_BITS_ALL &&
 398                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 399                                 goto readrest;
 400                         }
 401
 402                         break;
 403                 }
 404
 405                 /*
 406                  * Page is not resident, If this is the search termination
 407                  * or the pager might contain the page, allocate a new page.
 408                  */
 409                 if (TRYPAGER || fs.object == fs.first_object) {
 410                         if (fs.pindex >= fs.object->size) {
 411                                 unlock_and_deallocate(&fs);
 412                                 return (KERN_PROTECTION_FAILURE);
 413                         }
 414
 415                         /*
 416                          * Allocate a new page for this object/offset pair.
 417                          */
 418                         fs.m = NULL;
 419                         if (!vm_page_count_severe()) {
 420 #if VM_NRESERVLEVEL > 0
 421                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 422                                         fs.object->flags |= OBJ_COLORED;
 423                                         fs.object->pg_color = atop(vaddr) -
 424                                             fs.pindex;
 425                                 }
 426 #endif
 427                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 428                                     (fs.object->type == OBJT_VNODE ||
 429                                      fs.object->backing_object != NULL) ?
 430                                     VM_ALLOC_NORMAL : VM_ALLOC_ZERO);
 431                         }
 432                         if (fs.m == NULL) {
 433                                 unlock_and_deallocate(&fs);
 434                                 VM_WAITPFAULT;
 435                                 goto RetryFault;
 436                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 437                                 break;
 438                 }
 439
 440 readrest:
 441                 /*
 442                  * We have found a valid page or we have allocated a new page.
 443                  * The page thus may not be valid or may not be entirely
 444                  * valid.
 445                  *
 446                  * Attempt to fault-in the page if there is a chance that the
 447                  * pager has it, and potentially fault in additional pages
 448                  * at the same time.
 449                  */
 450                 if (TRYPAGER) {
 451                         int rv;
 452                         int reqpage = 0;
 453                         u_char behavior = vm_map_entry_behavior(fs.entry);
 454
 455                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 456                                 ahead = 0;
 457                                 behind = 0;
 458                         } else {
 459                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 460                                 if (behind > VM_FAULT_READ_BEHIND)
 461                                         behind = VM_FAULT_READ_BEHIND;
 462
 463                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 464                                 if (ahead > VM_FAULT_READ_AHEAD)
 465                                         ahead = VM_FAULT_READ_AHEAD;
 466                         }
 467                         is_first_object_locked = FALSE;
 468                         if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 469                              (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 470                               fs.pindex >= fs.entry->lastr &&
 471                               fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
 472                             (fs.first_object == fs.object ||
 473                              (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
 474                             fs.first_object->type != OBJT_DEVICE &&
 475                             fs.first_object->type != OBJT_PHYS) {
 476                                 vm_pindex_t firstpindex, tmppindex;
 477
 478                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 479                                         firstpindex = 0;
 480                                 else
 481                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 482
 483                                 are_queues_locked = FALSE;
 484                                 /*
 485                                  * note: partially valid pages cannot be
 486                                  * included in the lookahead - NFS piecemeal
 487                                  * writes will barf on it badly.
 488                                  */
 489                                 for (tmppindex = fs.first_pindex - 1;
 490                                         tmppindex >= firstpindex;
 491                                         --tmppindex) {
 492                                         vm_page_t mt;
 493
 494                                         mt = vm_page_lookup(fs.first_object, tmppindex);
 495                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 496                                                 break;
 497                                         if (mt->busy ||
 498                                             (mt->oflags & VPO_BUSY))
 499                                                 continue;
 500                                         if (!are_queues_locked) {
 501                                                 are_queues_locked = TRUE;
 502                                                 vm_page_lock_queues();
 503                                         }
 504                                         if (mt->hold_count ||
 505                                                 mt->wire_count)
 506                                                 continue;
 507                                         pmap_remove_all(mt);
 508                                         if (mt->dirty) {
 509                                                 vm_page_deactivate(mt);
 510                                         } else {
 511                                                 vm_page_cache(mt);
 512                                         }
 513                                 }
 514                                 if (are_queues_locked)
 515                                         vm_page_unlock_queues();
 516                                 ahead += behind;
 517                                 behind = 0;
 518                         }
 519                         if (is_first_object_locked)
 520                                 VM_OBJECT_UNLOCK(fs.first_object);
 521
 522                         /*
 523                          * Call the pager to retrieve the data, if any, after
 524                          * releasing the lock on the map.  We hold a ref on
 525                          * fs.object and the pages are VPO_BUSY'd.
 526                          */
 527                         unlock_map(&fs);
 528
 529 vnode_lock:
 530                         if (fs.object->type == OBJT_VNODE) {
 531                                 vp = fs.object->handle;
 532                                 if (vp == fs.vp)
 533                                         goto vnode_locked;
 534                                 else if (fs.vp != NULL) {
 535                                         vput(fs.vp);
 536                                         fs.vp = NULL;
 537                                 }
 538                                 locked = VOP_ISLOCKED(vp);
 539
 540                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 541                                         fs.vfslocked = 1;
 542                                         if (!mtx_trylock(&Giant)) {
 543                                                 VM_OBJECT_UNLOCK(fs.object);
 544                                                 mtx_lock(&Giant);
 545                                                 VM_OBJECT_LOCK(fs.object);
 546                                                 goto vnode_lock;
 547                                         }
 548                                 }
 549                                 if (locked != LK_EXCLUSIVE)
 550                                         locked = LK_SHARED;
 551                                 /* Do not sleep for vnode lock while fs.m is busy */
 552                                 error = vget(vp, locked | LK_CANRECURSE |
 553                                     LK_NOWAIT, curthread);
 554                                 if (error != 0) {
 555                                         int vfslocked;
 556
 557                                         vfslocked = fs.vfslocked;
 558                                         fs.vfslocked = 0; /* Keep Giant */
 559                                         vhold(vp);
 560                                         release_page(&fs);
 561                                         unlock_and_deallocate(&fs);
 562                                         error = vget(vp, locked | LK_RETRY |
 563                                             LK_CANRECURSE, curthread);
 564                                         vdrop(vp);
 565                                         fs.vp = vp;
 566                                         fs.vfslocked = vfslocked;
 567                                         KASSERT(error == 0,
 568                                             ("vm_fault: vget failed"));
 569                                         goto RetryFault;
 570                                 }
 571                                 fs.vp = vp;
 572                         }
 573 vnode_locked:
 574                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 575                             ("vm_fault: vnode-backed object mapped by system map"));
 576
 577                         /*
 578                          * now we find out if any other pages should be paged
 579                          * in at this time this routine checks to see if the
 580                          * pages surrounding this fault reside in the same
 581                          * object as the page for this fault.  If they do,
 582                          * then they are faulted in also into the object.  The
 583                          * array "marray" returned contains an array of
 584                          * vm_page_t structs where one of them is the
 585                          * vm_page_t passed to the routine.  The reqpage
 586                          * return value is the index into the marray for the
 587                          * vm_page_t passed to the routine.
 588                          *
 589                          * fs.m plus the additional pages are VPO_BUSY'd.
 590                          */
 591                         faultcount = vm_fault_additional_pages(
 592                             fs.m, behind, ahead, marray, &reqpage);
 593
 594                         rv = faultcount ?
 595                             vm_pager_get_pages(fs.object, marray, faultcount,
 596                                 reqpage) : VM_PAGER_FAIL;
 597
 598                         if (rv == VM_PAGER_OK) {
 599                                 /*
 600                                  * Found the page. Leave it busy while we play
 601                                  * with it.
 602                                  */
 603
 604                                 /*
 605                                  * Relookup in case pager changed page. Pager
 606                                  * is responsible for disposition of old page
 607                                  * if moved.
 608                                  */
 609                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 610                                 if (!fs.m) {
 611                                         unlock_and_deallocate(&fs);
 612                                         goto RetryFault;
 613                                 }
 614
 615                                 hardfault++;
 616                                 break; /* break to PAGE HAS BEEN FOUND */
 617                         }
 618                         /*
 619                          * Remove the bogus page (which does not exist at this
 620                          * object/offset); before doing so, we must get back
 621                          * our object lock to preserve our invariant.
 622                          *
 623                          * Also wake up any other process that may want to bring
 624                          * in this page.
 625                          *
 626                          * If this is the top-level object, we must leave the
 627                          * busy page to prevent another process from rushing
 628                          * past us, and inserting the page in that object at
 629                          * the same time that we are.
 630                          */
 631                         if (rv == VM_PAGER_ERROR)
 632                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 633                                     curproc->p_pid, curproc->p_comm);
 634                         /*
 635                          * Data outside the range of the pager or an I/O error
 636                          */
 637                         /*
 638                          * XXX - the check for kernel_map is a kludge to work
 639                          * around having the machine panic on a kernel space
 640                          * fault w/ I/O error.
 641                          */
 642                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 643                                 (rv == VM_PAGER_BAD)) {
 644                                 vm_page_lock_queues();
 645                                 vm_page_free(fs.m);
 646                                 vm_page_unlock_queues();
 647                                 fs.m = NULL;
 648                                 unlock_and_deallocate(&fs);
 649                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 650                         }
 651                         if (fs.object != fs.first_object) {
 652                                 vm_page_lock_queues();
 653                                 vm_page_free(fs.m);
 654                                 vm_page_unlock_queues();
 655                                 fs.m = NULL;
 656                                 /*
 657                                  * XXX - we cannot just fall out at this
 658                                  * point, m has been freed and is invalid!
 659                                  */
 660                         }
 661                 }
 662
 663                 /*
 664                  * We get here if the object has default pager (or unwiring)
 665                  * or the pager doesn't have the page.
 666                  */
 667                 if (fs.object == fs.first_object)
 668                         fs.first_m = fs.m;
 669
 670                 /*
 671                  * Move on to the next object.  Lock the next object before
 672                  * unlocking the current one.
 673                  */
 674                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 675                 next_object = fs.object->backing_object;
 676                 if (next_object == NULL) {
 677                         /*
 678                          * If there's no object left, fill the page in the top
 679                          * object with zeros.
 680                          */
 681                         if (fs.object != fs.first_object) {
 682                                 vm_object_pip_wakeup(fs.object);
 683                                 VM_OBJECT_UNLOCK(fs.object);
 684
 685                                 fs.object = fs.first_object;
 686                                 fs.pindex = fs.first_pindex;
 687                                 fs.m = fs.first_m;
 688                                 VM_OBJECT_LOCK(fs.object);
 689                         }
 690                         fs.first_m = NULL;
 691
 692                         /*
 693                          * Zero the page if necessary and mark it valid.
 694                          */
 695                         if ((fs.m->flags & PG_ZERO) == 0) {
 696                                 pmap_zero_page(fs.m);
 697                         } else {
 698                                 PCPU_INC(cnt.v_ozfod);
 699                         }
 700                         PCPU_INC(cnt.v_zfod);
 701                         fs.m->valid = VM_PAGE_BITS_ALL;
 702                         break;  /* break to PAGE HAS BEEN FOUND */
 703                 } else {
 704                         KASSERT(fs.object != next_object,
 705                             ("object loop %p", next_object));
 706                         VM_OBJECT_LOCK(next_object);
 707                         vm_object_pip_add(next_object, 1);
 708                         if (fs.object != fs.first_object)
 709                                 vm_object_pip_wakeup(fs.object);
 710                         VM_OBJECT_UNLOCK(fs.object);
 711                         fs.object = next_object;
 712                 }
 713         }
 714
 715         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 716             ("vm_fault: not busy after main loop"));
 717
 718         /*
 719          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 720          * is held.]
 721          */
 722
 723         /*
 724          * If the page is being written, but isn't already owned by the
 725          * top-level object, we have to copy it into a new page owned by the
 726          * top-level object.
 727          */
 728         if (fs.object != fs.first_object) {
 729                 /*
 730                  * We only really need to copy if we want to write it.
 731                  */
 732                 if (fault_type & VM_PROT_WRITE) {
 733                         /*
 734                          * This allows pages to be virtually copied from a
 735                          * backing_object into the first_object, where the
 736                          * backing object has no other refs to it, and cannot
 737                          * gain any more refs.  Instead of a bcopy, we just
 738                          * move the page from the backing object to the
 739                          * first object.  Note that we must mark the page
 740                          * dirty in the first object so that it will go out
 741                          * to swap when needed.
 742                          */
 743                         is_first_object_locked = FALSE;
 744                         if (
 745                                 /*
 746                                  * Only one shadow object
 747                                  */
 748                                 (fs.object->shadow_count == 1) &&
 749                                 /*
 750                                  * No COW refs, except us
 751                                  */
 752                                 (fs.object->ref_count == 1) &&
 753                                 /*
 754                                  * No one else can look this object up
 755                                  */
 756                                 (fs.object->handle == NULL) &&
 757                                 /*
 758                                  * No other ways to look the object up
 759                                  */
 760                                 ((fs.object->type == OBJT_DEFAULT) ||
 761                                  (fs.object->type == OBJT_SWAP)) &&
 762                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 763                                 /*
 764                                  * We don't chase down the shadow chain
 765                                  */
 766                             fs.object == fs.first_object->backing_object) {
 767                                 vm_page_lock_queues();
 768                                 /*
 769                                  * get rid of the unnecessary page
 770                                  */
 771                                 vm_page_free(fs.first_m);
 772                                 /*
 773                                  * grab the page and put it into the
 774                                  * process'es object.  The page is
 775                                  * automatically made dirty.
 776                                  */
 777                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 778                                 vm_page_unlock_queues();
 779                                 vm_page_busy(fs.m);
 780                                 fs.first_m = fs.m;
 781                                 fs.m = NULL;
 782                                 PCPU_INC(cnt.v_cow_optim);
 783                         } else {
 784                                 /*
 785                                  * Oh, well, lets copy it.
 786                                  */
 787                                 pmap_copy_page(fs.m, fs.first_m);
 788                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 789                         }
 790                         if (fs.m) {
 791                                 /*
 792                                  * We no longer need the old page or object.
 793                                  */
 794                                 release_page(&fs);
 795                         }
 796                         /*
 797                          * fs.object != fs.first_object due to above
 798                          * conditional
 799                          */
 800                         vm_object_pip_wakeup(fs.object);
 801                         VM_OBJECT_UNLOCK(fs.object);
 802                         /*
 803                          * Only use the new page below...
 804                          */
 805                         fs.object = fs.first_object;
 806                         fs.pindex = fs.first_pindex;
 807                         fs.m = fs.first_m;
 808                         if (!is_first_object_locked)
 809                                 VM_OBJECT_LOCK(fs.object);
 810                         PCPU_INC(cnt.v_cow_faults);
 811                 } else {
 812                         prot &= ~VM_PROT_WRITE;
 813                 }
 814         }
 815
 816         /*
 817          * We must verify that the maps have not changed since our last
 818          * lookup.
 819          */
 820         if (!fs.lookup_still_valid) {
 821                 vm_object_t retry_object;
 822                 vm_pindex_t retry_pindex;
 823                 vm_prot_t retry_prot;
 824
 825                 if (!vm_map_trylock_read(fs.map)) {
 826                         release_page(&fs);
 827                         unlock_and_deallocate(&fs);
 828                         goto RetryFault;
 829                 }
 830                 fs.lookup_still_valid = TRUE;
 831                 if (fs.map->timestamp != map_generation) {
 832                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 833                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 834
 835                         /*
 836                          * If we don't need the page any longer, put it on the inactive
 837                          * list (the easiest thing to do here).  If no one needs it,
 838                          * pageout will grab it eventually.
 839                          */
 840                         if (result != KERN_SUCCESS) {
 841                                 release_page(&fs);
 842                                 unlock_and_deallocate(&fs);
 843
 844                                 /*
 845                                  * If retry of map lookup would have blocked then
 846                                  * retry fault from start.
 847                                  */
 848                                 if (result == KERN_FAILURE)
 849                                         goto RetryFault;
 850                                 return (result);
 851                         }
 852                         if ((retry_object != fs.first_object) ||
 853                             (retry_pindex != fs.first_pindex)) {
 854                                 release_page(&fs);
 855                                 unlock_and_deallocate(&fs);
 856                                 goto RetryFault;
 857                         }
 858
 859                         /*
 860                          * Check whether the protection has changed or the object has
 861                          * been copied while we left the map unlocked. Changing from
 862                          * read to write permission is OK - we leave the page
 863                          * write-protected, and catch the write fault. Changing from
 864                          * write to read permission means that we can't mark the page
 865                          * write-enabled after all.
 866                          */
 867                         prot &= retry_prot;
 868                 }
 869         }
 870         /*
 871          * If the page was filled by a pager, update the map entry's
 872          * last read offset.  Since the pager does not return the
 873          * actual set of pages that it read, this update is based on
 874          * the requested set.  Typically, the requested and actual
 875          * sets are the same.
 876          *
 877          * XXX The following assignment modifies the map
 878          * without holding a write lock on it.
 879          */
 880         if (hardfault)
 881                 fs.entry->lastr = fs.pindex + faultcount - behind;
 882
 883         if (prot & VM_PROT_WRITE) {
 884                 vm_object_set_writeable_dirty(fs.object);
 885
 886                 /*
 887                  * If the fault is a write, we know that this page is being
 888                  * written NOW so dirty it explicitly to save on
 889                  * pmap_is_modified() calls later.
 890                  *
 891                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 892                  * if the page is already dirty to prevent data written with
 893                  * the expectation of being synced from not being synced.
 894                  * Likewise if this entry does not request NOSYNC then make
 895                  * sure the page isn't marked NOSYNC.  Applications sharing
 896                  * data should use the same flags to avoid ping ponging.
 897                  *
 898                  * Also tell the backing pager, if any, that it should remove
 899                  * any swap backing since the page is now dirty.
 900                  */
 901                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 902                         if (fs.m->dirty == 0)
 903                                 fs.m->oflags |= VPO_NOSYNC;
 904                 } else {
 905                         fs.m->oflags &= ~VPO_NOSYNC;
 906                 }
 907                 if (fault_flags & VM_FAULT_DIRTY) {
 908                         vm_page_dirty(fs.m);
 909                         vm_pager_page_unswapped(fs.m);
 910                 }
 911         }
 912
 913         /*
 914          * Page had better still be busy
 915          */
 916         KASSERT(fs.m->oflags & VPO_BUSY,
 917                 ("vm_fault: page %p not busy!", fs.m));
 918         /*
 919          * Page must be completely valid or it is not fit to
 920          * map into user space.  vm_pager_get_pages() ensures this.
 921          */
 922         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 923             ("vm_fault: page %p partially invalid", fs.m));
 924         VM_OBJECT_UNLOCK(fs.object);
 925
 926         /*
 927          * Put this page into the physical map.  We had to do the unlock above
 928          * because pmap_enter() may sleep.  We don't put the page
 929          * back on the active queue until later so that the pageout daemon
 930          * won't find it (yet).
 931          */
 932         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 933         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 934                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 935         }
 936         VM_OBJECT_LOCK(fs.object);
 937         vm_page_lock_queues();
 938         vm_page_flag_set(fs.m, PG_REFERENCED);
 939
 940         /*
 941          * If the page is not wired down, then put it where the pageout daemon
 942          * can find it.
 943          */
 944         if (fault_flags & VM_FAULT_WIRE_MASK) {
 945                 if (wired)
 946                         vm_page_wire(fs.m);
 947                 else
 948                         vm_page_unwire(fs.m, 1);
 949         } else {
 950                 vm_page_activate(fs.m);
 951         }
 952         vm_page_unlock_queues();
 953         vm_page_wakeup(fs.m);
 954
 955         /*
 956          * Unlock everything, and return
 957          */
 958         unlock_and_deallocate(&fs);
 959         if (hardfault)
 960                 curthread->td_ru.ru_majflt++;
 961         else
 962                 curthread->td_ru.ru_minflt++;
 963
 964         return (KERN_SUCCESS);
 965 }
 966
 967 /*
 968  * vm_fault_prefault provides a quick way of clustering
 969  * pagefaults into a processes address space.  It is a "cousin"
 970  * of vm_map_pmap_enter, except it runs at page fault time instead
 971  * of mmap time.
 972  */
 973 static void
 974 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
 975 {
 976         int i;
 977         vm_offset_t addr, starta;
 978         vm_pindex_t pindex;
 979         vm_page_t m;
 980         vm_object_t object;
 981
 982         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
 983                 return;
 984
 985         object = entry->object.vm_object;
 986
 987         starta = addra - PFBAK * PAGE_SIZE;
 988         if (starta < entry->start) {
 989                 starta = entry->start;
 990         } else if (starta > addra) {
 991                 starta = 0;
 992         }
 993
 994         for (i = 0; i < PAGEORDER_SIZE; i++) {
 995                 vm_object_t backing_object, lobject;
 996
 997                 addr = addra + prefault_pageorder[i];
 998                 if (addr > addra + (PFFOR * PAGE_SIZE))
 999                         addr = 0;
1000
1001                 if (addr < starta || addr >= entry->end)
1002                         continue;
1003
1004                 if (!pmap_is_prefaultable(pmap, addr))
1005                         continue;
1006
1007                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1008                 lobject = object;
1009                 VM_OBJECT_LOCK(lobject);
1010                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1011                     lobject->type == OBJT_DEFAULT &&
1012                     (backing_object = lobject->backing_object) != NULL) {
1013                         if (lobject->backing_object_offset & PAGE_MASK)
1014                                 break;
1015                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1016                         VM_OBJECT_LOCK(backing_object);
1017                         VM_OBJECT_UNLOCK(lobject);
1018                         lobject = backing_object;
1019                 }
1020                 /*
1021                  * give-up when a page is not in memory
1022                  */
1023                 if (m == NULL) {
1024                         VM_OBJECT_UNLOCK(lobject);
1025                         break;
1026                 }
1027                 if (m->valid == VM_PAGE_BITS_ALL &&
1028                     (m->flags & PG_FICTITIOUS) == 0) {
1029                         vm_page_lock_queues();
1030                         pmap_enter_quick(pmap, addr, m, entry->protection);
1031                         vm_page_unlock_queues();
1032                 }
1033                 VM_OBJECT_UNLOCK(lobject);
1034         }
1035 }
1036
1037 /*
1038  *      vm_fault_quick:
1039  *
1040  *      Ensure that the requested virtual address, which may be in userland,
1041  *      is valid.  Fault-in the page if necessary.  Return -1 on failure.
1042  */
1043 int
1044 vm_fault_quick(caddr_t v, int prot)
1045 {
1046         int r;
1047
1048         if (prot & VM_PROT_WRITE)
1049                 r = subyte(v, fubyte(v));
1050         else
1051                 r = fubyte(v);
1052         return(r);
1053 }
1054
1055 /*
1056  *      vm_fault_wire:
1057  *
1058  *      Wire down a range of virtual addresses in a map.
1059  */
1060 int
1061 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1062     boolean_t user_wire, boolean_t fictitious)
1063 {
1064         vm_offset_t va;
1065         int rv;
1066
1067         /*
1068          * We simulate a fault to get the page and enter it in the physical
1069          * map.  For user wiring, we only ask for read access on currently
1070          * read-only sections.
1071          */
1072         for (va = start; va < end; va += PAGE_SIZE) {
1073                 rv = vm_fault(map, va,
1074                     user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1075                     user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1076                 if (rv) {
1077                         if (va != start)
1078                                 vm_fault_unwire(map, start, va, fictitious);
1079                         return (rv);
1080                 }
1081         }
1082         return (KERN_SUCCESS);
1083 }
1084
1085 /*
1086  *      vm_fault_unwire:
1087  *
1088  *      Unwire a range of virtual addresses in a map.
1089  */
1090 void
1091 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1092     boolean_t fictitious)
1093 {
1094         vm_paddr_t pa;
1095         vm_offset_t va;
1096         pmap_t pmap;
1097
1098         pmap = vm_map_pmap(map);
1099
1100         /*
1101          * Since the pages are wired down, we must be able to get their
1102          * mappings from the physical map system.
1103          */
1104         for (va = start; va < end; va += PAGE_SIZE) {
1105                 pa = pmap_extract(pmap, va);
1106                 if (pa != 0) {
1107                         pmap_change_wiring(pmap, va, FALSE);
1108                         if (!fictitious) {
1109                                 vm_page_lock_queues();
1110                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1111                                 vm_page_unlock_queues();
1112                         }
1113                 }
1114         }
1115 }
1116
1117 /*
1118  *      Routine:
1119  *              vm_fault_copy_entry
1120  *      Function:
1121  *              Copy all of the pages from a wired-down map entry to another.
1122  *
1123  *      In/out conditions:
1124  *              The source and destination maps must be locked for write.
1125  *              The source map entry must be wired down (or be a sharing map
1126  *              entry corresponding to a main map entry that is wired down).
1127  */
1128 void
1129 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1130     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1131     vm_ooffset_t *fork_charge)
1132 {
1133         vm_object_t backing_object, dst_object, object;
1134         vm_object_t src_object;
1135         vm_ooffset_t dst_offset;
1136         vm_ooffset_t src_offset;
1137         vm_pindex_t pindex;
1138         vm_prot_t prot;
1139         vm_offset_t vaddr;
1140         vm_page_t dst_m;
1141         vm_page_t src_m;
1142
1143 #ifdef  lint
1144         src_map++;
1145 #endif  /* lint */
1146
1147         src_object = src_entry->object.vm_object;
1148         src_offset = src_entry->offset;
1149
1150         /*
1151          * Create the top-level object for the destination entry. (Doesn't
1152          * actually shadow anything - we copy the pages directly.)
1153          */
1154         dst_object = vm_object_allocate(OBJT_DEFAULT,
1155             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1156 #if VM_NRESERVLEVEL > 0
1157         dst_object->flags |= OBJ_COLORED;
1158         dst_object->pg_color = atop(dst_entry->start);
1159 #endif
1160
1161         VM_OBJECT_LOCK(dst_object);
1162         KASSERT(dst_entry->object.vm_object == NULL,
1163             ("vm_fault_copy_entry: vm_object not NULL"));
1164         dst_entry->object.vm_object = dst_object;
1165         dst_entry->offset = 0;
1166         dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1167         uihold(dst_object->uip);
1168         dst_object->charge = dst_entry->end - dst_entry->start;
1169         KASSERT(dst_entry->uip == NULL,
1170             ("vm_fault_copy_entry: leaked swp charge"));
1171         *fork_charge += dst_object->charge;
1172         prot = dst_entry->max_protection;
1173
1174         /*
1175          * Loop through all of the pages in the entry's range, copying each
1176          * one from the source object (it should be there) to the destination
1177          * object.
1178          */
1179         for (vaddr = dst_entry->start, dst_offset = 0;
1180             vaddr < dst_entry->end;
1181             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1182
1183                 /*
1184                  * Allocate a page in the destination object
1185                  */
1186                 do {
1187                         dst_m = vm_page_alloc(dst_object,
1188                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1189                         if (dst_m == NULL) {
1190                                 VM_OBJECT_UNLOCK(dst_object);
1191                                 VM_WAIT;
1192                                 VM_OBJECT_LOCK(dst_object);
1193                         }
1194                 } while (dst_m == NULL);
1195
1196                 /*
1197                  * Find the page in the source object, and copy it in.
1198                  * (Because the source is wired down, the page will be in
1199                  * memory.)
1200                  */
1201                 VM_OBJECT_LOCK(src_object);
1202                 object = src_object;
1203                 pindex = 0;
1204                 while ((src_m = vm_page_lookup(object, pindex +
1205                     OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1206                     (src_entry->protection & VM_PROT_WRITE) == 0 &&
1207                     (backing_object = object->backing_object) != NULL) {
1208                         /*
1209                          * Allow fallback to backing objects if we are reading.
1210                          */
1211                         VM_OBJECT_LOCK(backing_object);
1212                         pindex += OFF_TO_IDX(object->backing_object_offset);
1213                         VM_OBJECT_UNLOCK(object);
1214                         object = backing_object;
1215                 }
1216                 if (src_m == NULL)
1217                         panic("vm_fault_copy_wired: page missing");
1218                 pmap_copy_page(src_m, dst_m);
1219                 VM_OBJECT_UNLOCK(object);
1220                 dst_m->valid = VM_PAGE_BITS_ALL;
1221                 VM_OBJECT_UNLOCK(dst_object);
1222
1223                 /*
1224                  * Enter it in the pmap as a read and/or execute access.
1225                  */
1226                 pmap_enter(dst_map->pmap, vaddr, prot & ~VM_PROT_WRITE, dst_m,
1227                     prot, FALSE);
1228
1229                 /*
1230                  * Mark it no longer busy, and put it on the active list.
1231                  */
1232                 VM_OBJECT_LOCK(dst_object);
1233                 vm_page_lock_queues();
1234                 vm_page_activate(dst_m);
1235                 vm_page_unlock_queues();
1236                 vm_page_wakeup(dst_m);
1237         }
1238         VM_OBJECT_UNLOCK(dst_object);
1239 }
1240
1241
1242 /*
1243  * This routine checks around the requested page for other pages that
1244  * might be able to be faulted in.  This routine brackets the viable
1245  * pages for the pages to be paged in.
1246  *
1247  * Inputs:
1248  *      m, rbehind, rahead
1249  *
1250  * Outputs:
1251  *  marray (array of vm_page_t), reqpage (index of requested page)
1252  *
1253  * Return value:
1254  *  number of pages in marray
1255  */
1256 static int
1257 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1258         vm_page_t m;
1259         int rbehind;
1260         int rahead;
1261         vm_page_t *marray;
1262         int *reqpage;
1263 {
1264         int i,j;
1265         vm_object_t object;
1266         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1267         vm_page_t rtm;
1268         int cbehind, cahead;
1269
1270         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1271
1272         object = m->object;
1273         pindex = m->pindex;
1274         cbehind = cahead = 0;
1275
1276         /*
1277          * if the requested page is not available, then give up now
1278          */
1279         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1280                 return 0;
1281         }
1282
1283         if ((cbehind == 0) && (cahead == 0)) {
1284                 *reqpage = 0;
1285                 marray[0] = m;
1286                 return 1;
1287         }
1288
1289         if (rahead > cahead) {
1290                 rahead = cahead;
1291         }
1292
1293         if (rbehind > cbehind) {
1294                 rbehind = cbehind;
1295         }
1296
1297         /*
1298          * scan backward for the read behind pages -- in memory
1299          */
1300         if (pindex > 0) {
1301                 if (rbehind > pindex) {
1302                         rbehind = pindex;
1303                         startpindex = 0;
1304                 } else {
1305                         startpindex = pindex - rbehind;
1306                 }
1307
1308                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1309                     rtm->pindex >= startpindex)
1310                         startpindex = rtm->pindex + 1;
1311
1312                 /* tpindex is unsigned; beware of numeric underflow. */
1313                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1314                     tpindex < pindex; i++, tpindex--) {
1315
1316                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1317                             VM_ALLOC_IFNOTCACHED);
1318                         if (rtm == NULL) {
1319                                 /*
1320                                  * Shift the allocated pages to the
1321                                  * beginning of the array.
1322                                  */
1323                                 for (j = 0; j < i; j++) {
1324                                         marray[j] = marray[j + tpindex + 1 -
1325                                             startpindex];
1326                                 }
1327                                 break;
1328                         }
1329
1330                         marray[tpindex - startpindex] = rtm;
1331                 }
1332         } else {
1333                 startpindex = 0;
1334                 i = 0;
1335         }
1336
1337         marray[i] = m;
1338         /* page offset of the required page */
1339         *reqpage = i;
1340
1341         tpindex = pindex + 1;
1342         i++;
1343
1344         /*
1345          * scan forward for the read ahead pages
1346          */
1347         endpindex = tpindex + rahead;
1348         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1349                 endpindex = rtm->pindex;
1350         if (endpindex > object->size)
1351                 endpindex = object->size;
1352
1353         for (; tpindex < endpindex; i++, tpindex++) {
1354
1355                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1356                     VM_ALLOC_IFNOTCACHED);
1357                 if (rtm == NULL) {
1358                         break;
1359                 }
1360
1361                 marray[i] = rtm;
1362         }
1363
1364         /* return number of pages */
1365         return i;
1366 }