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/vm_extern.h>
 100
 101 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 102
 103 #define PFBAK 4
 104 #define PFFOR 4
 105 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 106
 107 static int prefault_pageorder[] = {
 108         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 109         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 110         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 111         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 112 };
 113
 114 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 115 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 116
 117 #define VM_FAULT_READ_AHEAD 8
 118 #define VM_FAULT_READ_BEHIND 7
 119 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 120
 121 struct faultstate {
 122         vm_page_t m;
 123         vm_object_t object;
 124         vm_pindex_t pindex;
 125         vm_page_t first_m;
 126         vm_object_t     first_object;
 127         vm_pindex_t first_pindex;
 128         vm_map_t map;
 129         vm_map_entry_t entry;
 130         int lookup_still_valid;
 131         struct vnode *vp;
 132         int vfslocked;
 133 };
 134
 135 static inline void
 136 release_page(struct faultstate *fs)
 137 {
 138
 139         vm_page_wakeup(fs->m);
 140         vm_page_lock(fs->m);
 141         vm_page_deactivate(fs->m);
 142         vm_page_unlock(fs->m);
 143         fs->m = NULL;
 144 }
 145
 146 static inline void
 147 unlock_map(struct faultstate *fs)
 148 {
 149
 150         if (fs->lookup_still_valid) {
 151                 vm_map_lookup_done(fs->map, fs->entry);
 152                 fs->lookup_still_valid = FALSE;
 153         }
 154 }
 155
 156 static void
 157 unlock_and_deallocate(struct faultstate *fs)
 158 {
 159
 160         vm_object_pip_wakeup(fs->object);
 161         VM_OBJECT_UNLOCK(fs->object);
 162         if (fs->object != fs->first_object) {
 163                 VM_OBJECT_LOCK(fs->first_object);
 164                 vm_page_lock(fs->first_m);
 165                 vm_page_free(fs->first_m);
 166                 vm_page_unlock(fs->first_m);
 167                 vm_object_pip_wakeup(fs->first_object);
 168                 VM_OBJECT_UNLOCK(fs->first_object);
 169                 fs->first_m = NULL;
 170         }
 171         vm_object_deallocate(fs->first_object);
 172         unlock_map(fs);
 173         if (fs->vp != NULL) {
 174                 vput(fs->vp);
 175                 fs->vp = NULL;
 176         }
 177         VFS_UNLOCK_GIANT(fs->vfslocked);
 178         fs->vfslocked = 0;
 179 }
 180
 181 /*
 182  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 183  *            current object *might* contain the page.
 184  *
 185  *            default objects are zero-fill, there is no real pager.
 186  */
 187 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 188                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 189
 190 /*
 191  *      vm_fault:
 192  *
 193  *      Handle a page fault occurring at the given address,
 194  *      requiring the given permissions, in the map specified.
 195  *      If successful, the page is inserted into the
 196  *      associated physical map.
 197  *
 198  *      NOTE: the given address should be truncated to the
 199  *      proper page address.
 200  *
 201  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 202  *      a standard error specifying why the fault is fatal is returned.
 203  *
 204  *      The map in question must be referenced, and remains so.
 205  *      Caller may hold no locks.
 206  */
 207 int
 208 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 209     int fault_flags)
 210 {
 211
 212         return (vm_fault_hold(map, vaddr, fault_type, fault_flags, NULL));
 213 }
 214
 215 int
 216 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 217     int fault_flags, vm_page_t *m_hold)
 218 {
 219         vm_prot_t prot;
 220         int is_first_object_locked, result;
 221         boolean_t growstack, wired;
 222         int map_generation;
 223         vm_object_t next_object;
 224         vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
 225         int hardfault;
 226         int faultcount, ahead, behind, alloc_req;
 227         struct faultstate fs;
 228         struct vnode *vp;
 229         int locked, error;
 230
 231         hardfault = 0;
 232         growstack = TRUE;
 233         PCPU_INC(cnt.v_vm_faults);
 234         fs.vp = NULL;
 235         fs.vfslocked = 0;
 236         faultcount = behind = 0;
 237
 238 RetryFault:;
 239
 240         /*
 241          * Find the backing store object and offset into it to begin the
 242          * search.
 243          */
 244         fs.map = map;
 245         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 246             &fs.first_object, &fs.first_pindex, &prot, &wired);
 247         if (result != KERN_SUCCESS) {
 248                 if (growstack && result == KERN_INVALID_ADDRESS &&
 249                     map != kernel_map) {
 250                         result = vm_map_growstack(curproc, vaddr);
 251                         if (result != KERN_SUCCESS)
 252                                 return (KERN_FAILURE);
 253                         growstack = FALSE;
 254                         goto RetryFault;
 255                 }
 256                 return (result);
 257         }
 258
 259         map_generation = fs.map->timestamp;
 260
 261         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 262                 panic("vm_fault: fault on nofault entry, addr: %lx",
 263                     (u_long)vaddr);
 264         }
 265
 266         /*
 267          * Make a reference to this object to prevent its disposal while we
 268          * are messing with it.  Once we have the reference, the map is free
 269          * to be diddled.  Since objects reference their shadows (and copies),
 270          * they will stay around as well.
 271          *
 272          * Bump the paging-in-progress count to prevent size changes (e.g.
 273          * truncation operations) during I/O.  This must be done after
 274          * obtaining the vnode lock in order to avoid possible deadlocks.
 275          */
 276         VM_OBJECT_LOCK(fs.first_object);
 277         vm_object_reference_locked(fs.first_object);
 278         vm_object_pip_add(fs.first_object, 1);
 279
 280         fs.lookup_still_valid = TRUE;
 281
 282         if (wired)
 283                 fault_type = prot | (fault_type & VM_PROT_COPY);
 284
 285         fs.first_m = NULL;
 286
 287         /*
 288          * Search for the page at object/offset.
 289          */
 290         fs.object = fs.first_object;
 291         fs.pindex = fs.first_pindex;
 292         while (TRUE) {
 293                 /*
 294                  * If the object is dead, we stop here
 295                  */
 296                 if (fs.object->flags & OBJ_DEAD) {
 297                         unlock_and_deallocate(&fs);
 298                         return (KERN_PROTECTION_FAILURE);
 299                 }
 300
 301                 /*
 302                  * See if page is resident
 303                  */
 304                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 305                 if (fs.m != NULL) {
 306                         /*
 307                          * check for page-based copy on write.
 308                          * We check fs.object == fs.first_object so
 309                          * as to ensure the legacy COW mechanism is
 310                          * used when the page in question is part of
 311                          * a shadow object.  Otherwise, vm_page_cowfault()
 312                          * removes the page from the backing object,
 313                          * which is not what we want.
 314                          */
 315                         vm_page_lock(fs.m);
 316                         if ((fs.m->cow) &&
 317                             (fault_type & VM_PROT_WRITE) &&
 318                             (fs.object == fs.first_object)) {
 319                                 vm_page_cowfault(fs.m);
 320                                 unlock_and_deallocate(&fs);
 321                                 goto RetryFault;
 322                         }
 323
 324                         /*
 325                          * Wait/Retry if the page is busy.  We have to do this
 326                          * if the page is busy via either VPO_BUSY or
 327                          * vm_page_t->busy because the vm_pager may be using
 328                          * vm_page_t->busy for pageouts ( and even pageins if
 329                          * it is the vnode pager ), and we could end up trying
 330                          * to pagein and pageout the same page simultaneously.
 331                          *
 332                          * We can theoretically allow the busy case on a read
 333                          * fault if the page is marked valid, but since such
 334                          * pages are typically already pmap'd, putting that
 335                          * special case in might be more effort then it is
 336                          * worth.  We cannot under any circumstances mess
 337                          * around with a vm_page_t->busy page except, perhaps,
 338                          * to pmap it.
 339                          */
 340                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 341                                 /*
 342                                  * Reference the page before unlocking and
 343                                  * sleeping so that the page daemon is less
 344                                  * likely to reclaim it.
 345                                  */
 346                                 vm_page_lock_queues();
 347                                 vm_page_flag_set(fs.m, PG_REFERENCED);
 348                                 vm_page_unlock_queues();
 349                                 vm_page_unlock(fs.m);
 350                                 if (fs.object != fs.first_object) {
 351                                         if (!VM_OBJECT_TRYLOCK(
 352                                             fs.first_object)) {
 353                                                 VM_OBJECT_UNLOCK(fs.object);
 354                                                 VM_OBJECT_LOCK(fs.first_object);
 355                                                 VM_OBJECT_LOCK(fs.object);
 356                                         }
 357                                         vm_page_lock(fs.first_m);
 358                                         vm_page_free(fs.first_m);
 359                                         vm_page_unlock(fs.first_m);
 360                                         vm_object_pip_wakeup(fs.first_object);
 361                                         VM_OBJECT_UNLOCK(fs.first_object);
 362                                         fs.first_m = NULL;
 363                                 }
 364                                 unlock_map(&fs);
 365                                 if (fs.m == vm_page_lookup(fs.object,
 366                                     fs.pindex)) {
 367                                         vm_page_sleep_if_busy(fs.m, TRUE,
 368                                             "vmpfw");
 369                                 }
 370                                 vm_object_pip_wakeup(fs.object);
 371                                 VM_OBJECT_UNLOCK(fs.object);
 372                                 PCPU_INC(cnt.v_intrans);
 373                                 vm_object_deallocate(fs.first_object);
 374                                 goto RetryFault;
 375                         }
 376                         vm_pageq_remove(fs.m);
 377                         vm_page_unlock(fs.m);
 378
 379                         /*
 380                          * Mark page busy for other processes, and the
 381                          * pagedaemon.  If it still isn't completely valid
 382                          * (readable), jump to readrest, else break-out ( we
 383                          * found the page ).
 384                          */
 385                         vm_page_busy(fs.m);
 386                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 387                                 goto readrest;
 388                         break;
 389                 }
 390
 391                 /*
 392                  * Page is not resident, If this is the search termination
 393                  * or the pager might contain the page, allocate a new page.
 394                  */
 395                 if (TRYPAGER || fs.object == fs.first_object) {
 396                         if (fs.pindex >= fs.object->size) {
 397                                 unlock_and_deallocate(&fs);
 398                                 return (KERN_PROTECTION_FAILURE);
 399                         }
 400
 401                         /*
 402                          * Allocate a new page for this object/offset pair.
 403                          *
 404                          * Unlocked read of the p_flag is harmless. At
 405                          * worst, the P_KILLED might be not observed
 406                          * there, and allocation can fail, causing
 407                          * restart and new reading of the p_flag.
 408                          */
 409                         fs.m = NULL;
 410                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 411 #if VM_NRESERVLEVEL > 0
 412                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 413                                         fs.object->flags |= OBJ_COLORED;
 414                                         fs.object->pg_color = atop(vaddr) -
 415                                             fs.pindex;
 416                                 }
 417 #endif
 418                                 alloc_req = P_KILLED(curproc) ?
 419                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 420                                 if (fs.object->type != OBJT_VNODE &&
 421                                     fs.object->backing_object == NULL)
 422                                         alloc_req |= VM_ALLOC_ZERO;
 423                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 424                                     alloc_req);
 425                         }
 426                         if (fs.m == NULL) {
 427                                 unlock_and_deallocate(&fs);
 428                                 VM_WAITPFAULT;
 429                                 goto RetryFault;
 430                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 431                                 break;
 432                 }
 433
 434 readrest:
 435                 /*
 436                  * We have found a valid page or we have allocated a new page.
 437                  * The page thus may not be valid or may not be entirely
 438                  * valid.
 439                  *
 440                  * Attempt to fault-in the page if there is a chance that the
 441                  * pager has it, and potentially fault in additional pages
 442                  * at the same time.
 443                  */
 444                 if (TRYPAGER) {
 445                         int rv;
 446                         int reqpage = 0;
 447                         u_char behavior = vm_map_entry_behavior(fs.entry);
 448
 449                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 450                             P_KILLED(curproc)) {
 451                                 ahead = 0;
 452                                 behind = 0;
 453                         } else {
 454                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 455                                 if (behind > VM_FAULT_READ_BEHIND)
 456                                         behind = VM_FAULT_READ_BEHIND;
 457
 458                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 459                                 if (ahead > VM_FAULT_READ_AHEAD)
 460                                         ahead = VM_FAULT_READ_AHEAD;
 461                         }
 462                         is_first_object_locked = FALSE;
 463                         if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 464                              (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 465                               fs.pindex >= fs.entry->lastr &&
 466                               fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
 467                             (fs.first_object == fs.object ||
 468                              (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
 469                             fs.first_object->type != OBJT_DEVICE &&
 470                             fs.first_object->type != OBJT_PHYS &&
 471                             fs.first_object->type != OBJT_SG) {
 472                                 vm_pindex_t firstpindex;
 473
 474                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 475                                         firstpindex = 0;
 476                                 else
 477                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 478                                 mt = fs.first_object != fs.object ?
 479                                     fs.first_m : fs.m;
 480                                 KASSERT(mt != NULL, ("vm_fault: missing mt"));
 481                                 KASSERT((mt->oflags & VPO_BUSY) != 0,
 482                                     ("vm_fault: mt %p not busy", mt));
 483                                 mt_prev = vm_page_prev(mt);
 484
 485                                 /*
 486                                  * note: partially valid pages cannot be
 487                                  * included in the lookahead - NFS piecemeal
 488                                  * writes will barf on it badly.
 489                                  */
 490                                 while ((mt = mt_prev) != NULL &&
 491                                     mt->pindex >= firstpindex &&
 492                                     mt->valid == VM_PAGE_BITS_ALL) {
 493                                         mt_prev = vm_page_prev(mt);
 494                                         if (mt->busy ||
 495                                             (mt->oflags & VPO_BUSY))
 496                                                 continue;
 497                                         vm_page_lock(mt);
 498                                         if (mt->hold_count ||
 499                                             mt->wire_count) {
 500                                                 vm_page_unlock(mt);
 501                                                 continue;
 502                                         }
 503                                         pmap_remove_all(mt);
 504                                         if (mt->dirty != 0)
 505                                                 vm_page_deactivate(mt);
 506                                         else
 507                                                 vm_page_cache(mt);
 508                                         vm_page_unlock(mt);
 509                                 }
 510                                 ahead += behind;
 511                                 behind = 0;
 512                         }
 513                         if (is_first_object_locked)
 514                                 VM_OBJECT_UNLOCK(fs.first_object);
 515
 516                         /*
 517                          * Call the pager to retrieve the data, if any, after
 518                          * releasing the lock on the map.  We hold a ref on
 519                          * fs.object and the pages are VPO_BUSY'd.
 520                          */
 521                         unlock_map(&fs);
 522
 523 vnode_lock:
 524                         if (fs.object->type == OBJT_VNODE) {
 525                                 vp = fs.object->handle;
 526                                 if (vp == fs.vp)
 527                                         goto vnode_locked;
 528                                 else if (fs.vp != NULL) {
 529                                         vput(fs.vp);
 530                                         fs.vp = NULL;
 531                                 }
 532                                 locked = VOP_ISLOCKED(vp);
 533
 534                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 535                                         fs.vfslocked = 1;
 536                                         if (!mtx_trylock(&Giant)) {
 537                                                 VM_OBJECT_UNLOCK(fs.object);
 538                                                 mtx_lock(&Giant);
 539                                                 VM_OBJECT_LOCK(fs.object);
 540                                                 goto vnode_lock;
 541                                         }
 542                                 }
 543                                 if (locked != LK_EXCLUSIVE)
 544                                         locked = LK_SHARED;
 545                                 /* Do not sleep for vnode lock while fs.m is busy */
 546                                 error = vget(vp, locked | LK_CANRECURSE |
 547                                     LK_NOWAIT, curthread);
 548                                 if (error != 0) {
 549                                         int vfslocked;
 550
 551                                         vfslocked = fs.vfslocked;
 552                                         fs.vfslocked = 0; /* Keep Giant */
 553                                         vhold(vp);
 554                                         release_page(&fs);
 555                                         unlock_and_deallocate(&fs);
 556                                         error = vget(vp, locked | LK_RETRY |
 557                                             LK_CANRECURSE, curthread);
 558                                         vdrop(vp);
 559                                         fs.vp = vp;
 560                                         fs.vfslocked = vfslocked;
 561                                         KASSERT(error == 0,
 562                                             ("vm_fault: vget failed"));
 563                                         goto RetryFault;
 564                                 }
 565                                 fs.vp = vp;
 566                         }
 567 vnode_locked:
 568                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 569                             ("vm_fault: vnode-backed object mapped by system map"));
 570
 571                         /*
 572                          * now we find out if any other pages should be paged
 573                          * in at this time this routine checks to see if the
 574                          * pages surrounding this fault reside in the same
 575                          * object as the page for this fault.  If they do,
 576                          * then they are faulted in also into the object.  The
 577                          * array "marray" returned contains an array of
 578                          * vm_page_t structs where one of them is the
 579                          * vm_page_t passed to the routine.  The reqpage
 580                          * return value is the index into the marray for the
 581                          * vm_page_t passed to the routine.
 582                          *
 583                          * fs.m plus the additional pages are VPO_BUSY'd.
 584                          */
 585                         faultcount = vm_fault_additional_pages(
 586                             fs.m, behind, ahead, marray, &reqpage);
 587
 588                         rv = faultcount ?
 589                             vm_pager_get_pages(fs.object, marray, faultcount,
 590                                 reqpage) : VM_PAGER_FAIL;
 591
 592                         if (rv == VM_PAGER_OK) {
 593                                 /*
 594                                  * Found the page. Leave it busy while we play
 595                                  * with it.
 596                                  */
 597
 598                                 /*
 599                                  * Relookup in case pager changed page. Pager
 600                                  * is responsible for disposition of old page
 601                                  * if moved.
 602                                  */
 603                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 604                                 if (!fs.m) {
 605                                         unlock_and_deallocate(&fs);
 606                                         goto RetryFault;
 607                                 }
 608
 609                                 hardfault++;
 610                                 break; /* break to PAGE HAS BEEN FOUND */
 611                         }
 612                         /*
 613                          * Remove the bogus page (which does not exist at this
 614                          * object/offset); before doing so, we must get back
 615                          * our object lock to preserve our invariant.
 616                          *
 617                          * Also wake up any other process that may want to bring
 618                          * in this page.
 619                          *
 620                          * If this is the top-level object, we must leave the
 621                          * busy page to prevent another process from rushing
 622                          * past us, and inserting the page in that object at
 623                          * the same time that we are.
 624                          */
 625                         if (rv == VM_PAGER_ERROR)
 626                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 627                                     curproc->p_pid, curproc->p_comm);
 628                         /*
 629                          * Data outside the range of the pager or an I/O error
 630                          */
 631                         /*
 632                          * XXX - the check for kernel_map is a kludge to work
 633                          * around having the machine panic on a kernel space
 634                          * fault w/ I/O error.
 635                          */
 636                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 637                                 (rv == VM_PAGER_BAD)) {
 638                                 vm_page_lock(fs.m);
 639                                 vm_page_free(fs.m);
 640                                 vm_page_unlock(fs.m);
 641                                 fs.m = NULL;
 642                                 unlock_and_deallocate(&fs);
 643                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 644                         }
 645                         if (fs.object != fs.first_object) {
 646                                 vm_page_lock(fs.m);
 647                                 vm_page_free(fs.m);
 648                                 vm_page_unlock(fs.m);
 649                                 fs.m = NULL;
 650                                 /*
 651                                  * XXX - we cannot just fall out at this
 652                                  * point, m has been freed and is invalid!
 653                                  */
 654                         }
 655                 }
 656
 657                 /*
 658                  * We get here if the object has default pager (or unwiring)
 659                  * or the pager doesn't have the page.
 660                  */
 661                 if (fs.object == fs.first_object)
 662                         fs.first_m = fs.m;
 663
 664                 /*
 665                  * Move on to the next object.  Lock the next object before
 666                  * unlocking the current one.
 667                  */
 668                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 669                 next_object = fs.object->backing_object;
 670                 if (next_object == NULL) {
 671                         /*
 672                          * If there's no object left, fill the page in the top
 673                          * object with zeros.
 674                          */
 675                         if (fs.object != fs.first_object) {
 676                                 vm_object_pip_wakeup(fs.object);
 677                                 VM_OBJECT_UNLOCK(fs.object);
 678
 679                                 fs.object = fs.first_object;
 680                                 fs.pindex = fs.first_pindex;
 681                                 fs.m = fs.first_m;
 682                                 VM_OBJECT_LOCK(fs.object);
 683                         }
 684                         fs.first_m = NULL;
 685
 686                         /*
 687                          * Zero the page if necessary and mark it valid.
 688                          */
 689                         if ((fs.m->flags & PG_ZERO) == 0) {
 690                                 pmap_zero_page(fs.m);
 691                         } else {
 692                                 PCPU_INC(cnt.v_ozfod);
 693                         }
 694                         PCPU_INC(cnt.v_zfod);
 695                         fs.m->valid = VM_PAGE_BITS_ALL;
 696                         break;  /* break to PAGE HAS BEEN FOUND */
 697                 } else {
 698                         KASSERT(fs.object != next_object,
 699                             ("object loop %p", next_object));
 700                         VM_OBJECT_LOCK(next_object);
 701                         vm_object_pip_add(next_object, 1);
 702                         if (fs.object != fs.first_object)
 703                                 vm_object_pip_wakeup(fs.object);
 704                         VM_OBJECT_UNLOCK(fs.object);
 705                         fs.object = next_object;
 706                 }
 707         }
 708
 709         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 710             ("vm_fault: not busy after main loop"));
 711
 712         /*
 713          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 714          * is held.]
 715          */
 716
 717         /*
 718          * If the page is being written, but isn't already owned by the
 719          * top-level object, we have to copy it into a new page owned by the
 720          * top-level object.
 721          */
 722         if (fs.object != fs.first_object) {
 723                 /*
 724                  * We only really need to copy if we want to write it.
 725                  */
 726                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 727                         /*
 728                          * This allows pages to be virtually copied from a
 729                          * backing_object into the first_object, where the
 730                          * backing object has no other refs to it, and cannot
 731                          * gain any more refs.  Instead of a bcopy, we just
 732                          * move the page from the backing object to the
 733                          * first object.  Note that we must mark the page
 734                          * dirty in the first object so that it will go out
 735                          * to swap when needed.
 736                          */
 737                         is_first_object_locked = FALSE;
 738                         if (
 739                                 /*
 740                                  * Only one shadow object
 741                                  */
 742                                 (fs.object->shadow_count == 1) &&
 743                                 /*
 744                                  * No COW refs, except us
 745                                  */
 746                                 (fs.object->ref_count == 1) &&
 747                                 /*
 748                                  * No one else can look this object up
 749                                  */
 750                                 (fs.object->handle == NULL) &&
 751                                 /*
 752                                  * No other ways to look the object up
 753                                  */
 754                                 ((fs.object->type == OBJT_DEFAULT) ||
 755                                  (fs.object->type == OBJT_SWAP)) &&
 756                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 757                                 /*
 758                                  * We don't chase down the shadow chain
 759                                  */
 760                             fs.object == fs.first_object->backing_object) {
 761                                 /*
 762                                  * get rid of the unnecessary page
 763                                  */
 764                                 vm_page_lock(fs.first_m);
 765                                 vm_page_free(fs.first_m);
 766                                 vm_page_unlock(fs.first_m);
 767                                 /*
 768                                  * grab the page and put it into the
 769                                  * process'es object.  The page is
 770                                  * automatically made dirty.
 771                                  */
 772                                 vm_page_lock(fs.m);
 773                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 774                                 vm_page_unlock(fs.m);
 775                                 vm_page_busy(fs.m);
 776                                 fs.first_m = fs.m;
 777                                 fs.m = NULL;
 778                                 PCPU_INC(cnt.v_cow_optim);
 779                         } else {
 780                                 /*
 781                                  * Oh, well, lets copy it.
 782                                  */
 783                                 pmap_copy_page(fs.m, fs.first_m);
 784                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 785                                 if (wired && (fault_flags &
 786                                     VM_FAULT_CHANGE_WIRING) == 0) {
 787                                         vm_page_lock(fs.first_m);
 788                                         vm_page_wire(fs.first_m);
 789                                         vm_page_unlock(fs.first_m);
 790
 791                                         vm_page_lock(fs.m);
 792                                         vm_page_unwire(fs.m, FALSE);
 793                                         vm_page_unlock(fs.m);
 794                                 }
 795                                 /*
 796                                  * We no longer need the old page or object.
 797                                  */
 798                                 release_page(&fs);
 799                         }
 800                         /*
 801                          * fs.object != fs.first_object due to above
 802                          * conditional
 803                          */
 804                         vm_object_pip_wakeup(fs.object);
 805                         VM_OBJECT_UNLOCK(fs.object);
 806                         /*
 807                          * Only use the new page below...
 808                          */
 809                         fs.object = fs.first_object;
 810                         fs.pindex = fs.first_pindex;
 811                         fs.m = fs.first_m;
 812                         if (!is_first_object_locked)
 813                                 VM_OBJECT_LOCK(fs.object);
 814                         PCPU_INC(cnt.v_cow_faults);
 815                 } else {
 816                         prot &= ~VM_PROT_WRITE;
 817                 }
 818         }
 819
 820         /*
 821          * We must verify that the maps have not changed since our last
 822          * lookup.
 823          */
 824         if (!fs.lookup_still_valid) {
 825                 vm_object_t retry_object;
 826                 vm_pindex_t retry_pindex;
 827                 vm_prot_t retry_prot;
 828
 829                 if (!vm_map_trylock_read(fs.map)) {
 830                         release_page(&fs);
 831                         unlock_and_deallocate(&fs);
 832                         goto RetryFault;
 833                 }
 834                 fs.lookup_still_valid = TRUE;
 835                 if (fs.map->timestamp != map_generation) {
 836                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 837                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 838
 839                         /*
 840                          * If we don't need the page any longer, put it on the inactive
 841                          * list (the easiest thing to do here).  If no one needs it,
 842                          * pageout will grab it eventually.
 843                          */
 844                         if (result != KERN_SUCCESS) {
 845                                 release_page(&fs);
 846                                 unlock_and_deallocate(&fs);
 847
 848                                 /*
 849                                  * If retry of map lookup would have blocked then
 850                                  * retry fault from start.
 851                                  */
 852                                 if (result == KERN_FAILURE)
 853                                         goto RetryFault;
 854                                 return (result);
 855                         }
 856                         if ((retry_object != fs.first_object) ||
 857                             (retry_pindex != fs.first_pindex)) {
 858                                 release_page(&fs);
 859                                 unlock_and_deallocate(&fs);
 860                                 goto RetryFault;
 861                         }
 862
 863                         /*
 864                          * Check whether the protection has changed or the object has
 865                          * been copied while we left the map unlocked. Changing from
 866                          * read to write permission is OK - we leave the page
 867                          * write-protected, and catch the write fault. Changing from
 868                          * write to read permission means that we can't mark the page
 869                          * write-enabled after all.
 870                          */
 871                         prot &= retry_prot;
 872                 }
 873         }
 874         /*
 875          * If the page was filled by a pager, update the map entry's
 876          * last read offset.  Since the pager does not return the
 877          * actual set of pages that it read, this update is based on
 878          * the requested set.  Typically, the requested and actual
 879          * sets are the same.
 880          *
 881          * XXX The following assignment modifies the map
 882          * without holding a write lock on it.
 883          */
 884         if (hardfault)
 885                 fs.entry->lastr = fs.pindex + faultcount - behind;
 886
 887         if ((prot & VM_PROT_WRITE) != 0 ||
 888             (fault_flags & VM_FAULT_DIRTY) != 0) {
 889                 vm_object_set_writeable_dirty(fs.object);
 890
 891                 /*
 892                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 893                  * if the page is already dirty to prevent data written with
 894                  * the expectation of being synced from not being synced.
 895                  * Likewise if this entry does not request NOSYNC then make
 896                  * sure the page isn't marked NOSYNC.  Applications sharing
 897                  * data should use the same flags to avoid ping ponging.
 898                  */
 899                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 900                         if (fs.m->dirty == 0)
 901                                 fs.m->oflags |= VPO_NOSYNC;
 902                 } else {
 903                         fs.m->oflags &= ~VPO_NOSYNC;
 904                 }
 905
 906                 /*
 907                  * If the fault is a write, we know that this page is being
 908                  * written NOW so dirty it explicitly to save on
 909                  * pmap_is_modified() calls later.
 910                  *
 911                  * Also tell the backing pager, if any, that it should remove
 912                  * any swap backing since the page is now dirty.
 913                  */
 914                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 915                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 916                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 917                         vm_page_dirty(fs.m);
 918                         vm_pager_page_unswapped(fs.m);
 919                 }
 920         }
 921
 922         /*
 923          * Page had better still be busy
 924          */
 925         KASSERT(fs.m->oflags & VPO_BUSY,
 926                 ("vm_fault: page %p not busy!", fs.m));
 927         /*
 928          * Page must be completely valid or it is not fit to
 929          * map into user space.  vm_pager_get_pages() ensures this.
 930          */
 931         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 932             ("vm_fault: page %p partially invalid", fs.m));
 933         VM_OBJECT_UNLOCK(fs.object);
 934
 935         /*
 936          * Put this page into the physical map.  We had to do the unlock above
 937          * because pmap_enter() may sleep.  We don't put the page
 938          * back on the active queue until later so that the pageout daemon
 939          * won't find it (yet).
 940          */
 941         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 942         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 943                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 944         VM_OBJECT_LOCK(fs.object);
 945         vm_page_lock(fs.m);
 946
 947         /*
 948          * If the page is not wired down, then put it where the pageout daemon
 949          * can find it.
 950          */
 951         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 952                 if (wired)
 953                         vm_page_wire(fs.m);
 954                 else
 955                         vm_page_unwire(fs.m, 1);
 956         } else
 957                 vm_page_activate(fs.m);
 958         if (m_hold != NULL) {
 959                 *m_hold = fs.m;
 960                 vm_page_hold(fs.m);
 961         }
 962         vm_page_unlock(fs.m);
 963         vm_page_wakeup(fs.m);
 964
 965         /*
 966          * Unlock everything, and return
 967          */
 968         unlock_and_deallocate(&fs);
 969         if (hardfault)
 970                 curthread->td_ru.ru_majflt++;
 971         else
 972                 curthread->td_ru.ru_minflt++;
 973
 974         return (KERN_SUCCESS);
 975 }
 976
 977 /*
 978  * vm_fault_prefault provides a quick way of clustering
 979  * pagefaults into a processes address space.  It is a "cousin"
 980  * of vm_map_pmap_enter, except it runs at page fault time instead
 981  * of mmap time.
 982  */
 983 static void
 984 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
 985 {
 986         int i;
 987         vm_offset_t addr, starta;
 988         vm_pindex_t pindex;
 989         vm_page_t m;
 990         vm_object_t object;
 991
 992         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
 993                 return;
 994
 995         object = entry->object.vm_object;
 996
 997         starta = addra - PFBAK * PAGE_SIZE;
 998         if (starta < entry->start) {
 999                 starta = entry->start;
1000         } else if (starta > addra) {
1001                 starta = 0;
1002         }
1003
1004         for (i = 0; i < PAGEORDER_SIZE; i++) {
1005                 vm_object_t backing_object, lobject;
1006
1007                 addr = addra + prefault_pageorder[i];
1008                 if (addr > addra + (PFFOR * PAGE_SIZE))
1009                         addr = 0;
1010
1011                 if (addr < starta || addr >= entry->end)
1012                         continue;
1013
1014                 if (!pmap_is_prefaultable(pmap, addr))
1015                         continue;
1016
1017                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1018                 lobject = object;
1019                 VM_OBJECT_LOCK(lobject);
1020                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1021                     lobject->type == OBJT_DEFAULT &&
1022                     (backing_object = lobject->backing_object) != NULL) {
1023                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1024                             0, ("vm_fault_prefault: unaligned object offset"));
1025                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1026                         VM_OBJECT_LOCK(backing_object);
1027                         VM_OBJECT_UNLOCK(lobject);
1028                         lobject = backing_object;
1029                 }
1030                 /*
1031                  * give-up when a page is not in memory
1032                  */
1033                 if (m == NULL) {
1034                         VM_OBJECT_UNLOCK(lobject);
1035                         break;
1036                 }
1037                 if (m->valid == VM_PAGE_BITS_ALL &&
1038                     (m->flags & PG_FICTITIOUS) == 0)
1039                         pmap_enter_quick(pmap, addr, m, entry->protection);
1040                 VM_OBJECT_UNLOCK(lobject);
1041         }
1042 }
1043
1044 /*
1045  * Hold each of the physical pages that are mapped by the specified range of
1046  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1047  * and allow the specified types of access, "prot".  If all of the implied
1048  * pages are successfully held, then the number of held pages is returned
1049  * together with pointers to those pages in the array "ma".  However, if any
1050  * of the pages cannot be held, -1 is returned.
1051  */
1052 int
1053 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1054     vm_prot_t prot, vm_page_t *ma, int max_count)
1055 {
1056         vm_offset_t end, va;
1057         vm_page_t *mp;
1058         int count;
1059         boolean_t pmap_failed;
1060
1061         if (len == 0)
1062                 return (0);
1063         end = round_page(addr + len);
1064         addr = trunc_page(addr);
1065
1066         /*
1067          * Check for illegal addresses.
1068          */
1069         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1070                 return (-1);
1071
1072         count = howmany(end - addr, PAGE_SIZE);
1073         if (count > max_count)
1074                 panic("vm_fault_quick_hold_pages: count > max_count");
1075
1076         /*
1077          * Most likely, the physical pages are resident in the pmap, so it is
1078          * faster to try pmap_extract_and_hold() first.
1079          */
1080         pmap_failed = FALSE;
1081         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1082                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1083                 if (*mp == NULL)
1084                         pmap_failed = TRUE;
1085                 else if ((prot & VM_PROT_WRITE) != 0 &&
1086                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1087                         /*
1088                          * Explicitly dirty the physical page.  Otherwise, the
1089                          * caller's changes may go unnoticed because they are
1090                          * performed through an unmanaged mapping or by a DMA
1091                          * operation.
1092                          */
1093                         vm_page_lock_queues();
1094                         vm_page_dirty(*mp);
1095                         vm_page_unlock_queues();
1096                 }
1097         }
1098         if (pmap_failed) {
1099                 /*
1100                  * One or more pages could not be held by the pmap.  Either no
1101                  * page was mapped at the specified virtual address or that
1102                  * mapping had insufficient permissions.  Attempt to fault in
1103                  * and hold these pages.
1104                  */
1105                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1106                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1107                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1108                                 goto error;
1109         }
1110         return (count);
1111 error:
1112         for (mp = ma; mp < ma + count; mp++)
1113                 if (*mp != NULL) {
1114                         vm_page_lock(*mp);
1115                         vm_page_unhold(*mp);
1116                         vm_page_unlock(*mp);
1117                 }
1118         return (-1);
1119 }
1120
1121 /*
1122  *      vm_fault_wire:
1123  *
1124  *      Wire down a range of virtual addresses in a map.
1125  */
1126 int
1127 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1128     boolean_t fictitious)
1129 {
1130         vm_offset_t va;
1131         int rv;
1132
1133         /*
1134          * We simulate a fault to get the page and enter it in the physical
1135          * map.  For user wiring, we only ask for read access on currently
1136          * read-only sections.
1137          */
1138         for (va = start; va < end; va += PAGE_SIZE) {
1139                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1140                 if (rv) {
1141                         if (va != start)
1142                                 vm_fault_unwire(map, start, va, fictitious);
1143                         return (rv);
1144                 }
1145         }
1146         return (KERN_SUCCESS);
1147 }
1148
1149 /*
1150  *      vm_fault_unwire:
1151  *
1152  *      Unwire a range of virtual addresses in a map.
1153  */
1154 void
1155 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1156     boolean_t fictitious)
1157 {
1158         vm_paddr_t pa;
1159         vm_offset_t va;
1160         vm_page_t m;
1161         pmap_t pmap;
1162
1163         pmap = vm_map_pmap(map);
1164
1165         /*
1166          * Since the pages are wired down, we must be able to get their
1167          * mappings from the physical map system.
1168          */
1169         for (va = start; va < end; va += PAGE_SIZE) {
1170                 pa = pmap_extract(pmap, va);
1171                 if (pa != 0) {
1172                         pmap_change_wiring(pmap, va, FALSE);
1173                         if (!fictitious) {
1174                                 m = PHYS_TO_VM_PAGE(pa);
1175                                 vm_page_lock(m);
1176                                 vm_page_unwire(m, TRUE);
1177                                 vm_page_unlock(m);
1178                         }
1179                 }
1180         }
1181 }
1182
1183 /*
1184  *      Routine:
1185  *              vm_fault_copy_entry
1186  *      Function:
1187  *              Create new shadow object backing dst_entry with private copy of
1188  *              all underlying pages. When src_entry is equal to dst_entry,
1189  *              function implements COW for wired-down map entry. Otherwise,
1190  *              it forks wired entry into dst_map.
1191  *
1192  *      In/out conditions:
1193  *              The source and destination maps must be locked for write.
1194  *              The source map entry must be wired down (or be a sharing map
1195  *              entry corresponding to a main map entry that is wired down).
1196  */
1197 void
1198 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1199     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1200     vm_ooffset_t *fork_charge)
1201 {
1202         vm_object_t backing_object, dst_object, object, src_object;
1203         vm_pindex_t dst_pindex, pindex, src_pindex;
1204         vm_prot_t access, prot;
1205         vm_offset_t vaddr;
1206         vm_page_t dst_m;
1207         vm_page_t src_m;
1208         boolean_t src_readonly, upgrade;
1209
1210 #ifdef  lint
1211         src_map++;
1212 #endif  /* lint */
1213
1214         upgrade = src_entry == dst_entry;
1215
1216         src_object = src_entry->object.vm_object;
1217         src_pindex = OFF_TO_IDX(src_entry->offset);
1218         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1219
1220         /*
1221          * Create the top-level object for the destination entry. (Doesn't
1222          * actually shadow anything - we copy the pages directly.)
1223          */
1224         dst_object = vm_object_allocate(OBJT_DEFAULT,
1225             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1226 #if VM_NRESERVLEVEL > 0
1227         dst_object->flags |= OBJ_COLORED;
1228         dst_object->pg_color = atop(dst_entry->start);
1229 #endif
1230
1231         VM_OBJECT_LOCK(dst_object);
1232         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1233             ("vm_fault_copy_entry: vm_object not NULL"));
1234         dst_entry->object.vm_object = dst_object;
1235         dst_entry->offset = 0;
1236         dst_object->charge = dst_entry->end - dst_entry->start;
1237         if (fork_charge != NULL) {
1238                 KASSERT(dst_entry->cred == NULL,
1239                     ("vm_fault_copy_entry: leaked swp charge"));
1240                 dst_object->cred = curthread->td_ucred;
1241                 crhold(dst_object->cred);
1242                 *fork_charge += dst_object->charge;
1243         } else {
1244                 dst_object->cred = dst_entry->cred;
1245                 dst_entry->cred = NULL;
1246         }
1247         access = prot = dst_entry->protection;
1248         /*
1249          * If not an upgrade, then enter the mappings in the pmap as
1250          * read and/or execute accesses.  Otherwise, enter them as
1251          * write accesses.
1252          *
1253          * A writeable large page mapping is only created if all of
1254          * the constituent small page mappings are modified. Marking
1255          * PTEs as modified on inception allows promotion to happen
1256          * without taking potentially large number of soft faults.
1257          */
1258         if (!upgrade)
1259                 access &= ~VM_PROT_WRITE;
1260
1261         /*
1262          * Loop through all of the pages in the entry's range, copying each
1263          * one from the source object (it should be there) to the destination
1264          * object.
1265          */
1266         for (vaddr = dst_entry->start, dst_pindex = 0;
1267             vaddr < dst_entry->end;
1268             vaddr += PAGE_SIZE, dst_pindex++) {
1269
1270                 /*
1271                  * Allocate a page in the destination object.
1272                  */
1273                 do {
1274                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1275                             VM_ALLOC_NORMAL);
1276                         if (dst_m == NULL) {
1277                                 VM_OBJECT_UNLOCK(dst_object);
1278                                 VM_WAIT;
1279                                 VM_OBJECT_LOCK(dst_object);
1280                         }
1281                 } while (dst_m == NULL);
1282
1283                 /*
1284                  * Find the page in the source object, and copy it in.
1285                  * (Because the source is wired down, the page will be in
1286                  * memory.)
1287                  */
1288                 VM_OBJECT_LOCK(src_object);
1289                 object = src_object;
1290                 pindex = src_pindex + dst_pindex;
1291                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1292                     src_readonly &&
1293                     (backing_object = object->backing_object) != NULL) {
1294                         /*
1295                          * Allow fallback to backing objects if we are reading.
1296                          */
1297                         VM_OBJECT_LOCK(backing_object);
1298                         pindex += OFF_TO_IDX(object->backing_object_offset);
1299                         VM_OBJECT_UNLOCK(object);
1300                         object = backing_object;
1301                 }
1302                 if (src_m == NULL)
1303                         panic("vm_fault_copy_wired: page missing");
1304                 pmap_copy_page(src_m, dst_m);
1305                 VM_OBJECT_UNLOCK(object);
1306                 dst_m->valid = VM_PAGE_BITS_ALL;
1307                 VM_OBJECT_UNLOCK(dst_object);
1308
1309                 /*
1310                  * Enter it in the pmap. If a wired, copy-on-write
1311                  * mapping is being replaced by a write-enabled
1312                  * mapping, then wire that new mapping.
1313                  */
1314                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1315
1316                 /*
1317                  * Mark it no longer busy, and put it on the active list.
1318                  */
1319                 VM_OBJECT_LOCK(dst_object);
1320
1321                 if (upgrade) {
1322                         vm_page_lock(src_m);
1323                         vm_page_unwire(src_m, 0);
1324                         vm_page_unlock(src_m);
1325
1326                         vm_page_lock(dst_m);
1327                         vm_page_wire(dst_m);
1328                         vm_page_unlock(dst_m);
1329                 } else {
1330                         vm_page_lock(dst_m);
1331                         vm_page_activate(dst_m);
1332                         vm_page_unlock(dst_m);
1333                 }
1334                 vm_page_wakeup(dst_m);
1335         }
1336         VM_OBJECT_UNLOCK(dst_object);
1337         if (upgrade) {
1338                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1339                 vm_object_deallocate(src_object);
1340         }
1341 }
1342
1343
1344 /*
1345  * This routine checks around the requested page for other pages that
1346  * might be able to be faulted in.  This routine brackets the viable
1347  * pages for the pages to be paged in.
1348  *
1349  * Inputs:
1350  *      m, rbehind, rahead
1351  *
1352  * Outputs:
1353  *  marray (array of vm_page_t), reqpage (index of requested page)
1354  *
1355  * Return value:
1356  *  number of pages in marray
1357  */
1358 static int
1359 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1360         vm_page_t m;
1361         int rbehind;
1362         int rahead;
1363         vm_page_t *marray;
1364         int *reqpage;
1365 {
1366         int i,j;
1367         vm_object_t object;
1368         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1369         vm_page_t rtm;
1370         int cbehind, cahead;
1371
1372         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1373
1374         object = m->object;
1375         pindex = m->pindex;
1376         cbehind = cahead = 0;
1377
1378         /*
1379          * if the requested page is not available, then give up now
1380          */
1381         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1382                 return 0;
1383         }
1384
1385         if ((cbehind == 0) && (cahead == 0)) {
1386                 *reqpage = 0;
1387                 marray[0] = m;
1388                 return 1;
1389         }
1390
1391         if (rahead > cahead) {
1392                 rahead = cahead;
1393         }
1394
1395         if (rbehind > cbehind) {
1396                 rbehind = cbehind;
1397         }
1398
1399         /*
1400          * scan backward for the read behind pages -- in memory
1401          */
1402         if (pindex > 0) {
1403                 if (rbehind > pindex) {
1404                         rbehind = pindex;
1405                         startpindex = 0;
1406                 } else {
1407                         startpindex = pindex - rbehind;
1408                 }
1409
1410                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1411                     rtm->pindex >= startpindex)
1412                         startpindex = rtm->pindex + 1;
1413
1414                 /* tpindex is unsigned; beware of numeric underflow. */
1415                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1416                     tpindex < pindex; i++, tpindex--) {
1417
1418                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1419                             VM_ALLOC_IFNOTCACHED);
1420                         if (rtm == NULL) {
1421                                 /*
1422                                  * Shift the allocated pages to the
1423                                  * beginning of the array.
1424                                  */
1425                                 for (j = 0; j < i; j++) {
1426                                         marray[j] = marray[j + tpindex + 1 -
1427                                             startpindex];
1428                                 }
1429                                 break;
1430                         }
1431
1432                         marray[tpindex - startpindex] = rtm;
1433                 }
1434         } else {
1435                 startpindex = 0;
1436                 i = 0;
1437         }
1438
1439         marray[i] = m;
1440         /* page offset of the required page */
1441         *reqpage = i;
1442
1443         tpindex = pindex + 1;
1444         i++;
1445
1446         /*
1447          * scan forward for the read ahead pages
1448          */
1449         endpindex = tpindex + rahead;
1450         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1451                 endpindex = rtm->pindex;
1452         if (endpindex > object->size)
1453                 endpindex = object->size;
1454
1455         for (; tpindex < endpindex; i++, tpindex++) {
1456
1457                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1458                     VM_ALLOC_IFNOTCACHED);
1459                 if (rtm == NULL) {
1460                         break;
1461                 }
1462
1463                 marray[i] = rtm;
1464         }
1465
1466         /* return number of pages */
1467         return i;
1468 }