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_ktrace.h"
  78 #include "opt_vm.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/kernel.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/resourcevar.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vm_extern.h>
 104
 105 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 106
 107 #define PFBAK 4
 108 #define PFFOR 4
 109 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 110
 111 static int prefault_pageorder[] = {
 112         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 113         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 114         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 115         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 116 };
 117
 118 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 119 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 120
 121 #define VM_FAULT_READ_BEHIND    8
 122 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 123 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 124 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 125 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 126
 127 struct faultstate {
 128         vm_page_t m;
 129         vm_object_t object;
 130         vm_pindex_t pindex;
 131         vm_page_t first_m;
 132         vm_object_t     first_object;
 133         vm_pindex_t first_pindex;
 134         vm_map_t map;
 135         vm_map_entry_t entry;
 136         int lookup_still_valid;
 137         struct vnode *vp;
 138         int vfslocked;
 139 };
 140
 141 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 142
 143 static inline void
 144 release_page(struct faultstate *fs)
 145 {
 146
 147         vm_page_wakeup(fs->m);
 148         vm_page_lock(fs->m);
 149         vm_page_deactivate(fs->m);
 150         vm_page_unlock(fs->m);
 151         fs->m = NULL;
 152 }
 153
 154 static inline void
 155 unlock_map(struct faultstate *fs)
 156 {
 157
 158         if (fs->lookup_still_valid) {
 159                 vm_map_lookup_done(fs->map, fs->entry);
 160                 fs->lookup_still_valid = FALSE;
 161         }
 162 }
 163
 164 static void
 165 unlock_and_deallocate(struct faultstate *fs)
 166 {
 167
 168         vm_object_pip_wakeup(fs->object);
 169         VM_OBJECT_UNLOCK(fs->object);
 170         if (fs->object != fs->first_object) {
 171                 VM_OBJECT_LOCK(fs->first_object);
 172                 vm_page_lock(fs->first_m);
 173                 vm_page_free(fs->first_m);
 174                 vm_page_unlock(fs->first_m);
 175                 vm_object_pip_wakeup(fs->first_object);
 176                 VM_OBJECT_UNLOCK(fs->first_object);
 177                 fs->first_m = NULL;
 178         }
 179         vm_object_deallocate(fs->first_object);
 180         unlock_map(fs);
 181         if (fs->vp != NULL) {
 182                 vput(fs->vp);
 183                 fs->vp = NULL;
 184         }
 185         VFS_UNLOCK_GIANT(fs->vfslocked);
 186         fs->vfslocked = 0;
 187 }
 188
 189 /*
 190  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 191  *            current object *might* contain the page.
 192  *
 193  *            default objects are zero-fill, there is no real pager.
 194  */
 195 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 196                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 197
 198 /*
 199  *      vm_fault:
 200  *
 201  *      Handle a page fault occurring at the given address,
 202  *      requiring the given permissions, in the map specified.
 203  *      If successful, the page is inserted into the
 204  *      associated physical map.
 205  *
 206  *      NOTE: the given address should be truncated to the
 207  *      proper page address.
 208  *
 209  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 210  *      a standard error specifying why the fault is fatal is returned.
 211  *
 212  *      The map in question must be referenced, and remains so.
 213  *      Caller may hold no locks.
 214  */
 215 int
 216 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 217     int fault_flags)
 218 {
 219         struct thread *td;
 220         int result;
 221
 222         td = curthread;
 223         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 224                 return (KERN_PROTECTION_FAILURE);
 225 #ifdef KTRACE
 226         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 227                 ktrfault(vaddr, fault_type);
 228 #endif
 229         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 230             NULL);
 231 #ifdef KTRACE
 232         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 233                 ktrfaultend(result);
 234 #endif
 235         return (result);
 236 }
 237
 238 int
 239 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 240     int fault_flags, vm_page_t *m_hold)
 241 {
 242         vm_prot_t prot;
 243         long ahead, behind;
 244         int alloc_req, era, faultcount, nera, reqpage, result;
 245         boolean_t growstack, is_first_object_locked, wired;
 246         int map_generation;
 247         vm_object_t next_object;
 248         vm_page_t marray[VM_FAULT_READ_MAX];
 249         int hardfault;
 250         struct faultstate fs;
 251         struct vnode *vp;
 252         int locked, error;
 253
 254         hardfault = 0;
 255         growstack = TRUE;
 256         PCPU_INC(cnt.v_vm_faults);
 257         fs.vp = NULL;
 258         fs.vfslocked = 0;
 259         faultcount = reqpage = 0;
 260
 261 RetryFault:;
 262
 263         /*
 264          * Find the backing store object and offset into it to begin the
 265          * search.
 266          */
 267         fs.map = map;
 268         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 269             &fs.first_object, &fs.first_pindex, &prot, &wired);
 270         if (result != KERN_SUCCESS) {
 271                 if (growstack && result == KERN_INVALID_ADDRESS &&
 272                     map != kernel_map) {
 273                         result = vm_map_growstack(curproc, vaddr);
 274                         if (result != KERN_SUCCESS)
 275                                 return (KERN_FAILURE);
 276                         growstack = FALSE;
 277                         goto RetryFault;
 278                 }
 279                 return (result);
 280         }
 281
 282         map_generation = fs.map->timestamp;
 283
 284         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 285                 panic("vm_fault: fault on nofault entry, addr: %lx",
 286                     (u_long)vaddr);
 287         }
 288
 289         /*
 290          * Make a reference to this object to prevent its disposal while we
 291          * are messing with it.  Once we have the reference, the map is free
 292          * to be diddled.  Since objects reference their shadows (and copies),
 293          * they will stay around as well.
 294          *
 295          * Bump the paging-in-progress count to prevent size changes (e.g.
 296          * truncation operations) during I/O.  This must be done after
 297          * obtaining the vnode lock in order to avoid possible deadlocks.
 298          */
 299         VM_OBJECT_LOCK(fs.first_object);
 300         vm_object_reference_locked(fs.first_object);
 301         vm_object_pip_add(fs.first_object, 1);
 302
 303         fs.lookup_still_valid = TRUE;
 304
 305         if (wired)
 306                 fault_type = prot | (fault_type & VM_PROT_COPY);
 307
 308         fs.first_m = NULL;
 309
 310         /*
 311          * Search for the page at object/offset.
 312          */
 313         fs.object = fs.first_object;
 314         fs.pindex = fs.first_pindex;
 315         while (TRUE) {
 316                 /*
 317                  * If the object is dead, we stop here
 318                  */
 319                 if (fs.object->flags & OBJ_DEAD) {
 320                         unlock_and_deallocate(&fs);
 321                         return (KERN_PROTECTION_FAILURE);
 322                 }
 323
 324                 /*
 325                  * See if page is resident
 326                  */
 327                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 328                 if (fs.m != NULL) {
 329                         /*
 330                          * check for page-based copy on write.
 331                          * We check fs.object == fs.first_object so
 332                          * as to ensure the legacy COW mechanism is
 333                          * used when the page in question is part of
 334                          * a shadow object.  Otherwise, vm_page_cowfault()
 335                          * removes the page from the backing object,
 336                          * which is not what we want.
 337                          */
 338                         vm_page_lock(fs.m);
 339                         if ((fs.m->cow) &&
 340                             (fault_type & VM_PROT_WRITE) &&
 341                             (fs.object == fs.first_object)) {
 342                                 vm_page_cowfault(fs.m);
 343                                 unlock_and_deallocate(&fs);
 344                                 goto RetryFault;
 345                         }
 346
 347                         /*
 348                          * Wait/Retry if the page is busy.  We have to do this
 349                          * if the page is busy via either VPO_BUSY or
 350                          * vm_page_t->busy because the vm_pager may be using
 351                          * vm_page_t->busy for pageouts ( and even pageins if
 352                          * it is the vnode pager ), and we could end up trying
 353                          * to pagein and pageout the same page simultaneously.
 354                          *
 355                          * We can theoretically allow the busy case on a read
 356                          * fault if the page is marked valid, but since such
 357                          * pages are typically already pmap'd, putting that
 358                          * special case in might be more effort then it is
 359                          * worth.  We cannot under any circumstances mess
 360                          * around with a vm_page_t->busy page except, perhaps,
 361                          * to pmap it.
 362                          */
 363                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 364                                 /*
 365                                  * Reference the page before unlocking and
 366                                  * sleeping so that the page daemon is less
 367                                  * likely to reclaim it.
 368                                  */
 369                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 370                                 vm_page_unlock(fs.m);
 371                                 if (fs.object != fs.first_object) {
 372                                         if (!VM_OBJECT_TRYLOCK(
 373                                             fs.first_object)) {
 374                                                 VM_OBJECT_UNLOCK(fs.object);
 375                                                 VM_OBJECT_LOCK(fs.first_object);
 376                                                 VM_OBJECT_LOCK(fs.object);
 377                                         }
 378                                         vm_page_lock(fs.first_m);
 379                                         vm_page_free(fs.first_m);
 380                                         vm_page_unlock(fs.first_m);
 381                                         vm_object_pip_wakeup(fs.first_object);
 382                                         VM_OBJECT_UNLOCK(fs.first_object);
 383                                         fs.first_m = NULL;
 384                                 }
 385                                 unlock_map(&fs);
 386                                 if (fs.m == vm_page_lookup(fs.object,
 387                                     fs.pindex)) {
 388                                         vm_page_sleep_if_busy(fs.m, TRUE,
 389                                             "vmpfw");
 390                                 }
 391                                 vm_object_pip_wakeup(fs.object);
 392                                 VM_OBJECT_UNLOCK(fs.object);
 393                                 PCPU_INC(cnt.v_intrans);
 394                                 vm_object_deallocate(fs.first_object);
 395                                 goto RetryFault;
 396                         }
 397                         vm_pageq_remove(fs.m);
 398                         vm_page_unlock(fs.m);
 399
 400                         /*
 401                          * Mark page busy for other processes, and the
 402                          * pagedaemon.  If it still isn't completely valid
 403                          * (readable), jump to readrest, else break-out ( we
 404                          * found the page ).
 405                          */
 406                         vm_page_busy(fs.m);
 407                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 408                                 goto readrest;
 409                         break;
 410                 }
 411
 412                 /*
 413                  * Page is not resident, If this is the search termination
 414                  * or the pager might contain the page, allocate a new page.
 415                  */
 416                 if (TRYPAGER || fs.object == fs.first_object) {
 417                         if (fs.pindex >= fs.object->size) {
 418                                 unlock_and_deallocate(&fs);
 419                                 return (KERN_PROTECTION_FAILURE);
 420                         }
 421
 422                         /*
 423                          * Allocate a new page for this object/offset pair.
 424                          *
 425                          * Unlocked read of the p_flag is harmless. At
 426                          * worst, the P_KILLED might be not observed
 427                          * there, and allocation can fail, causing
 428                          * restart and new reading of the p_flag.
 429                          */
 430                         fs.m = NULL;
 431                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 432 #if VM_NRESERVLEVEL > 0
 433                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 434                                         fs.object->flags |= OBJ_COLORED;
 435                                         fs.object->pg_color = atop(vaddr) -
 436                                             fs.pindex;
 437                                 }
 438 #endif
 439                                 alloc_req = P_KILLED(curproc) ?
 440                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 441                                 if (fs.object->type != OBJT_VNODE &&
 442                                     fs.object->backing_object == NULL)
 443                                         alloc_req |= VM_ALLOC_ZERO;
 444                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 445                                     alloc_req);
 446                         }
 447                         if (fs.m == NULL) {
 448                                 unlock_and_deallocate(&fs);
 449                                 VM_WAITPFAULT;
 450                                 goto RetryFault;
 451                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 452                                 break;
 453                 }
 454
 455 readrest:
 456                 /*
 457                  * We have found a valid page or we have allocated a new page.
 458                  * The page thus may not be valid or may not be entirely
 459                  * valid.
 460                  *
 461                  * Attempt to fault-in the page if there is a chance that the
 462                  * pager has it, and potentially fault in additional pages
 463                  * at the same time.
 464                  */
 465                 if (TRYPAGER) {
 466                         int rv;
 467                         u_char behavior = vm_map_entry_behavior(fs.entry);
 468
 469                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 470                             P_KILLED(curproc)) {
 471                                 behind = 0;
 472                                 ahead = 0;
 473                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 474                                 behind = 0;
 475                                 ahead = atop(fs.entry->end - vaddr) - 1;
 476                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 477                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 478                                 if (fs.pindex == fs.entry->next_read)
 479                                         vm_fault_cache_behind(&fs,
 480                                             VM_FAULT_READ_MAX);
 481                         } else {
 482                                 /*
 483                                  * If this is a sequential page fault, then
 484                                  * arithmetically increase the number of pages
 485                                  * in the read-ahead window.  Otherwise, reset
 486                                  * the read-ahead window to its smallest size.
 487                                  */
 488                                 behind = atop(vaddr - fs.entry->start);
 489                                 if (behind > VM_FAULT_READ_BEHIND)
 490                                         behind = VM_FAULT_READ_BEHIND;
 491                                 ahead = atop(fs.entry->end - vaddr) - 1;
 492                                 era = fs.entry->read_ahead;
 493                                 if (fs.pindex == fs.entry->next_read) {
 494                                         nera = era + behind;
 495                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 496                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 497                                         behind = 0;
 498                                         if (ahead > nera)
 499                                                 ahead = nera;
 500                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 501                                                 vm_fault_cache_behind(&fs,
 502                                                     VM_FAULT_CACHE_BEHIND);
 503                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 504                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 505                                 if (era != ahead)
 506                                         fs.entry->read_ahead = ahead;
 507                         }
 508
 509                         /*
 510                          * Call the pager to retrieve the data, if any, after
 511                          * releasing the lock on the map.  We hold a ref on
 512                          * fs.object and the pages are VPO_BUSY'd.
 513                          */
 514                         unlock_map(&fs);
 515
 516 vnode_lock:
 517                         if (fs.object->type == OBJT_VNODE) {
 518                                 vp = fs.object->handle;
 519                                 if (vp == fs.vp)
 520                                         goto vnode_locked;
 521                                 else if (fs.vp != NULL) {
 522                                         vput(fs.vp);
 523                                         fs.vp = NULL;
 524                                 }
 525                                 locked = VOP_ISLOCKED(vp);
 526
 527                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 528                                         fs.vfslocked = 1;
 529                                         if (!mtx_trylock(&Giant)) {
 530                                                 VM_OBJECT_UNLOCK(fs.object);
 531                                                 mtx_lock(&Giant);
 532                                                 VM_OBJECT_LOCK(fs.object);
 533                                                 goto vnode_lock;
 534                                         }
 535                                 }
 536                                 if (locked != LK_EXCLUSIVE)
 537                                         locked = LK_SHARED;
 538                                 /* Do not sleep for vnode lock while fs.m is busy */
 539                                 error = vget(vp, locked | LK_CANRECURSE |
 540                                     LK_NOWAIT, curthread);
 541                                 if (error != 0) {
 542                                         int vfslocked;
 543
 544                                         vfslocked = fs.vfslocked;
 545                                         fs.vfslocked = 0; /* Keep Giant */
 546                                         vhold(vp);
 547                                         release_page(&fs);
 548                                         unlock_and_deallocate(&fs);
 549                                         error = vget(vp, locked | LK_RETRY |
 550                                             LK_CANRECURSE, curthread);
 551                                         vdrop(vp);
 552                                         fs.vp = vp;
 553                                         fs.vfslocked = vfslocked;
 554                                         KASSERT(error == 0,
 555                                             ("vm_fault: vget failed"));
 556                                         goto RetryFault;
 557                                 }
 558                                 fs.vp = vp;
 559                         }
 560 vnode_locked:
 561                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 562                             ("vm_fault: vnode-backed object mapped by system map"));
 563
 564                         /*
 565                          * now we find out if any other pages should be paged
 566                          * in at this time this routine checks to see if the
 567                          * pages surrounding this fault reside in the same
 568                          * object as the page for this fault.  If they do,
 569                          * then they are faulted in also into the object.  The
 570                          * array "marray" returned contains an array of
 571                          * vm_page_t structs where one of them is the
 572                          * vm_page_t passed to the routine.  The reqpage
 573                          * return value is the index into the marray for the
 574                          * vm_page_t passed to the routine.
 575                          *
 576                          * fs.m plus the additional pages are VPO_BUSY'd.
 577                          */
 578                         faultcount = vm_fault_additional_pages(
 579                             fs.m, behind, ahead, marray, &reqpage);
 580
 581                         rv = faultcount ?
 582                             vm_pager_get_pages(fs.object, marray, faultcount,
 583                                 reqpage) : VM_PAGER_FAIL;
 584
 585                         if (rv == VM_PAGER_OK) {
 586                                 /*
 587                                  * Found the page. Leave it busy while we play
 588                                  * with it.
 589                                  */
 590
 591                                 /*
 592                                  * Relookup in case pager changed page. Pager
 593                                  * is responsible for disposition of old page
 594                                  * if moved.
 595                                  */
 596                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 597                                 if (!fs.m) {
 598                                         unlock_and_deallocate(&fs);
 599                                         goto RetryFault;
 600                                 }
 601
 602                                 hardfault++;
 603                                 break; /* break to PAGE HAS BEEN FOUND */
 604                         }
 605                         /*
 606                          * Remove the bogus page (which does not exist at this
 607                          * object/offset); before doing so, we must get back
 608                          * our object lock to preserve our invariant.
 609                          *
 610                          * Also wake up any other process that may want to bring
 611                          * in this page.
 612                          *
 613                          * If this is the top-level object, we must leave the
 614                          * busy page to prevent another process from rushing
 615                          * past us, and inserting the page in that object at
 616                          * the same time that we are.
 617                          */
 618                         if (rv == VM_PAGER_ERROR)
 619                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 620                                     curproc->p_pid, curproc->p_comm);
 621                         /*
 622                          * Data outside the range of the pager or an I/O error
 623                          */
 624                         /*
 625                          * XXX - the check for kernel_map is a kludge to work
 626                          * around having the machine panic on a kernel space
 627                          * fault w/ I/O error.
 628                          */
 629                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 630                                 (rv == VM_PAGER_BAD)) {
 631                                 vm_page_lock(fs.m);
 632                                 vm_page_free(fs.m);
 633                                 vm_page_unlock(fs.m);
 634                                 fs.m = NULL;
 635                                 unlock_and_deallocate(&fs);
 636                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 637                         }
 638                         if (fs.object != fs.first_object) {
 639                                 vm_page_lock(fs.m);
 640                                 vm_page_free(fs.m);
 641                                 vm_page_unlock(fs.m);
 642                                 fs.m = NULL;
 643                                 /*
 644                                  * XXX - we cannot just fall out at this
 645                                  * point, m has been freed and is invalid!
 646                                  */
 647                         }
 648                 }
 649
 650                 /*
 651                  * We get here if the object has default pager (or unwiring)
 652                  * or the pager doesn't have the page.
 653                  */
 654                 if (fs.object == fs.first_object)
 655                         fs.first_m = fs.m;
 656
 657                 /*
 658                  * Move on to the next object.  Lock the next object before
 659                  * unlocking the current one.
 660                  */
 661                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 662                 next_object = fs.object->backing_object;
 663                 if (next_object == NULL) {
 664                         /*
 665                          * If there's no object left, fill the page in the top
 666                          * object with zeros.
 667                          */
 668                         if (fs.object != fs.first_object) {
 669                                 vm_object_pip_wakeup(fs.object);
 670                                 VM_OBJECT_UNLOCK(fs.object);
 671
 672                                 fs.object = fs.first_object;
 673                                 fs.pindex = fs.first_pindex;
 674                                 fs.m = fs.first_m;
 675                                 VM_OBJECT_LOCK(fs.object);
 676                         }
 677                         fs.first_m = NULL;
 678
 679                         /*
 680                          * Zero the page if necessary and mark it valid.
 681                          */
 682                         if ((fs.m->flags & PG_ZERO) == 0) {
 683                                 pmap_zero_page(fs.m);
 684                         } else {
 685                                 PCPU_INC(cnt.v_ozfod);
 686                         }
 687                         PCPU_INC(cnt.v_zfod);
 688                         fs.m->valid = VM_PAGE_BITS_ALL;
 689                         break;  /* break to PAGE HAS BEEN FOUND */
 690                 } else {
 691                         KASSERT(fs.object != next_object,
 692                             ("object loop %p", next_object));
 693                         VM_OBJECT_LOCK(next_object);
 694                         vm_object_pip_add(next_object, 1);
 695                         if (fs.object != fs.first_object)
 696                                 vm_object_pip_wakeup(fs.object);
 697                         VM_OBJECT_UNLOCK(fs.object);
 698                         fs.object = next_object;
 699                 }
 700         }
 701
 702         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 703             ("vm_fault: not busy after main loop"));
 704
 705         /*
 706          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 707          * is held.]
 708          */
 709
 710         /*
 711          * If the page is being written, but isn't already owned by the
 712          * top-level object, we have to copy it into a new page owned by the
 713          * top-level object.
 714          */
 715         if (fs.object != fs.first_object) {
 716                 /*
 717                  * We only really need to copy if we want to write it.
 718                  */
 719                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 720                         /*
 721                          * This allows pages to be virtually copied from a
 722                          * backing_object into the first_object, where the
 723                          * backing object has no other refs to it, and cannot
 724                          * gain any more refs.  Instead of a bcopy, we just
 725                          * move the page from the backing object to the
 726                          * first object.  Note that we must mark the page
 727                          * dirty in the first object so that it will go out
 728                          * to swap when needed.
 729                          */
 730                         is_first_object_locked = FALSE;
 731                         if (
 732                                 /*
 733                                  * Only one shadow object
 734                                  */
 735                                 (fs.object->shadow_count == 1) &&
 736                                 /*
 737                                  * No COW refs, except us
 738                                  */
 739                                 (fs.object->ref_count == 1) &&
 740                                 /*
 741                                  * No one else can look this object up
 742                                  */
 743                                 (fs.object->handle == NULL) &&
 744                                 /*
 745                                  * No other ways to look the object up
 746                                  */
 747                                 ((fs.object->type == OBJT_DEFAULT) ||
 748                                  (fs.object->type == OBJT_SWAP)) &&
 749                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 750                                 /*
 751                                  * We don't chase down the shadow chain
 752                                  */
 753                             fs.object == fs.first_object->backing_object) {
 754                                 /*
 755                                  * get rid of the unnecessary page
 756                                  */
 757                                 vm_page_lock(fs.first_m);
 758                                 vm_page_free(fs.first_m);
 759                                 vm_page_unlock(fs.first_m);
 760                                 /*
 761                                  * grab the page and put it into the
 762                                  * process'es object.  The page is
 763                                  * automatically made dirty.
 764                                  */
 765                                 vm_page_lock(fs.m);
 766                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 767                                 vm_page_unlock(fs.m);
 768                                 vm_page_busy(fs.m);
 769                                 fs.first_m = fs.m;
 770                                 fs.m = NULL;
 771                                 PCPU_INC(cnt.v_cow_optim);
 772                         } else {
 773                                 /*
 774                                  * Oh, well, lets copy it.
 775                                  */
 776                                 pmap_copy_page(fs.m, fs.first_m);
 777                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 778                                 if (wired && (fault_flags &
 779                                     VM_FAULT_CHANGE_WIRING) == 0) {
 780                                         vm_page_lock(fs.first_m);
 781                                         vm_page_wire(fs.first_m);
 782                                         vm_page_unlock(fs.first_m);
 783
 784                                         vm_page_lock(fs.m);
 785                                         vm_page_unwire(fs.m, FALSE);
 786                                         vm_page_unlock(fs.m);
 787                                 }
 788                                 /*
 789                                  * We no longer need the old page or object.
 790                                  */
 791                                 release_page(&fs);
 792                         }
 793                         /*
 794                          * fs.object != fs.first_object due to above
 795                          * conditional
 796                          */
 797                         vm_object_pip_wakeup(fs.object);
 798                         VM_OBJECT_UNLOCK(fs.object);
 799                         /*
 800                          * Only use the new page below...
 801                          */
 802                         fs.object = fs.first_object;
 803                         fs.pindex = fs.first_pindex;
 804                         fs.m = fs.first_m;
 805                         if (!is_first_object_locked)
 806                                 VM_OBJECT_LOCK(fs.object);
 807                         PCPU_INC(cnt.v_cow_faults);
 808                         curthread->td_cow++;
 809                 } else {
 810                         prot &= ~VM_PROT_WRITE;
 811                 }
 812         }
 813
 814         /*
 815          * We must verify that the maps have not changed since our last
 816          * lookup.
 817          */
 818         if (!fs.lookup_still_valid) {
 819                 vm_object_t retry_object;
 820                 vm_pindex_t retry_pindex;
 821                 vm_prot_t retry_prot;
 822
 823                 if (!vm_map_trylock_read(fs.map)) {
 824                         release_page(&fs);
 825                         unlock_and_deallocate(&fs);
 826                         goto RetryFault;
 827                 }
 828                 fs.lookup_still_valid = TRUE;
 829                 if (fs.map->timestamp != map_generation) {
 830                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 831                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 832
 833                         /*
 834                          * If we don't need the page any longer, put it on the inactive
 835                          * list (the easiest thing to do here).  If no one needs it,
 836                          * pageout will grab it eventually.
 837                          */
 838                         if (result != KERN_SUCCESS) {
 839                                 release_page(&fs);
 840                                 unlock_and_deallocate(&fs);
 841
 842                                 /*
 843                                  * If retry of map lookup would have blocked then
 844                                  * retry fault from start.
 845                                  */
 846                                 if (result == KERN_FAILURE)
 847                                         goto RetryFault;
 848                                 return (result);
 849                         }
 850                         if ((retry_object != fs.first_object) ||
 851                             (retry_pindex != fs.first_pindex)) {
 852                                 release_page(&fs);
 853                                 unlock_and_deallocate(&fs);
 854                                 goto RetryFault;
 855                         }
 856
 857                         /*
 858                          * Check whether the protection has changed or the object has
 859                          * been copied while we left the map unlocked. Changing from
 860                          * read to write permission is OK - we leave the page
 861                          * write-protected, and catch the write fault. Changing from
 862                          * write to read permission means that we can't mark the page
 863                          * write-enabled after all.
 864                          */
 865                         prot &= retry_prot;
 866                 }
 867         }
 868         /*
 869          * If the page was filled by a pager, update the map entry's
 870          * last read offset.  Since the pager does not return the
 871          * actual set of pages that it read, this update is based on
 872          * the requested set.  Typically, the requested and actual
 873          * sets are the same.
 874          *
 875          * XXX The following assignment modifies the map
 876          * without holding a write lock on it.
 877          */
 878         if (hardfault)
 879                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 880
 881         if ((prot & VM_PROT_WRITE) != 0 ||
 882             (fault_flags & VM_FAULT_DIRTY) != 0) {
 883                 vm_object_set_writeable_dirty(fs.object);
 884
 885                 /*
 886                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 887                  * if the page is already dirty to prevent data written with
 888                  * the expectation of being synced from not being synced.
 889                  * Likewise if this entry does not request NOSYNC then make
 890                  * sure the page isn't marked NOSYNC.  Applications sharing
 891                  * data should use the same flags to avoid ping ponging.
 892                  */
 893                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 894                         if (fs.m->dirty == 0)
 895                                 fs.m->oflags |= VPO_NOSYNC;
 896                 } else {
 897                         fs.m->oflags &= ~VPO_NOSYNC;
 898                 }
 899
 900                 /*
 901                  * If the fault is a write, we know that this page is being
 902                  * written NOW so dirty it explicitly to save on
 903                  * pmap_is_modified() calls later.
 904                  *
 905                  * Also tell the backing pager, if any, that it should remove
 906                  * any swap backing since the page is now dirty.
 907                  */
 908                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 909                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 910                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 911                         vm_page_dirty(fs.m);
 912                         vm_pager_page_unswapped(fs.m);
 913                 }
 914         }
 915
 916         /*
 917          * Page had better still be busy
 918          */
 919         KASSERT(fs.m->oflags & VPO_BUSY,
 920                 ("vm_fault: page %p not busy!", fs.m));
 921         /*
 922          * Page must be completely valid or it is not fit to
 923          * map into user space.  vm_pager_get_pages() ensures this.
 924          */
 925         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 926             ("vm_fault: page %p partially invalid", fs.m));
 927         VM_OBJECT_UNLOCK(fs.object);
 928
 929         /*
 930          * Put this page into the physical map.  We had to do the unlock above
 931          * because pmap_enter() may sleep.  We don't put the page
 932          * back on the active queue until later so that the pageout daemon
 933          * won't find it (yet).
 934          */
 935         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 936         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 937                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 938         VM_OBJECT_LOCK(fs.object);
 939         vm_page_lock(fs.m);
 940
 941         /*
 942          * If the page is not wired down, then put it where the pageout daemon
 943          * can find it.
 944          */
 945         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 946                 if (wired)
 947                         vm_page_wire(fs.m);
 948                 else
 949                         vm_page_unwire(fs.m, 1);
 950         } else
 951                 vm_page_activate(fs.m);
 952         if (m_hold != NULL) {
 953                 *m_hold = fs.m;
 954                 vm_page_hold(fs.m);
 955         }
 956         vm_page_unlock(fs.m);
 957         vm_page_wakeup(fs.m);
 958
 959         /*
 960          * Unlock everything, and return
 961          */
 962         unlock_and_deallocate(&fs);
 963         if (hardfault)
 964                 curthread->td_ru.ru_majflt++;
 965         else
 966                 curthread->td_ru.ru_minflt++;
 967
 968         return (KERN_SUCCESS);
 969 }
 970
 971 /*
 972  * Speed up the reclamation of up to "distance" pages that precede the
 973  * faulting pindex within the first object of the shadow chain.
 974  */
 975 static void
 976 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 977 {
 978         vm_object_t first_object, object;
 979         vm_page_t m, m_prev;
 980         vm_pindex_t pindex;
 981
 982         object = fs->object;
 983         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 984         first_object = fs->first_object;
 985         if (first_object != object) {
 986                 if (!VM_OBJECT_TRYLOCK(first_object)) {
 987                         VM_OBJECT_UNLOCK(object);
 988                         VM_OBJECT_LOCK(first_object);
 989                         VM_OBJECT_LOCK(object);
 990                 }
 991         }
 992         if (first_object->type != OBJT_DEVICE &&
 993             first_object->type != OBJT_PHYS && first_object->type != OBJT_SG) {
 994                 if (fs->first_pindex < distance)
 995                         pindex = 0;
 996                 else
 997                         pindex = fs->first_pindex - distance;
 998                 if (pindex < OFF_TO_IDX(fs->entry->offset))
 999                         pindex = OFF_TO_IDX(fs->entry->offset);
1000                 m = first_object != object ? fs->first_m : fs->m;
1001                 KASSERT((m->oflags & VPO_BUSY) != 0,
1002                     ("vm_fault_cache_behind: page %p is not busy", m));
1003                 m_prev = vm_page_prev(m);
1004                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1005                     m->valid == VM_PAGE_BITS_ALL) {
1006                         m_prev = vm_page_prev(m);
1007                         if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0)
1008                                 continue;
1009                         vm_page_lock(m);
1010                         if (m->hold_count == 0 && m->wire_count == 0) {
1011                                 pmap_remove_all(m);
1012                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1013                                 if (m->dirty != 0)
1014                                         vm_page_deactivate(m);
1015                                 else
1016                                         vm_page_cache(m);
1017                         }
1018                         vm_page_unlock(m);
1019                 }
1020         }
1021         if (first_object != object)
1022                 VM_OBJECT_UNLOCK(first_object);
1023 }
1024
1025 /*
1026  * vm_fault_prefault provides a quick way of clustering
1027  * pagefaults into a processes address space.  It is a "cousin"
1028  * of vm_map_pmap_enter, except it runs at page fault time instead
1029  * of mmap time.
1030  */
1031 static void
1032 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1033 {
1034         int i;
1035         vm_offset_t addr, starta;
1036         vm_pindex_t pindex;
1037         vm_page_t m;
1038         vm_object_t object;
1039
1040         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1041                 return;
1042
1043         object = entry->object.vm_object;
1044
1045         starta = addra - PFBAK * PAGE_SIZE;
1046         if (starta < entry->start) {
1047                 starta = entry->start;
1048         } else if (starta > addra) {
1049                 starta = 0;
1050         }
1051
1052         for (i = 0; i < PAGEORDER_SIZE; i++) {
1053                 vm_object_t backing_object, lobject;
1054
1055                 addr = addra + prefault_pageorder[i];
1056                 if (addr > addra + (PFFOR * PAGE_SIZE))
1057                         addr = 0;
1058
1059                 if (addr < starta || addr >= entry->end)
1060                         continue;
1061
1062                 if (!pmap_is_prefaultable(pmap, addr))
1063                         continue;
1064
1065                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1066                 lobject = object;
1067                 VM_OBJECT_LOCK(lobject);
1068                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1069                     lobject->type == OBJT_DEFAULT &&
1070                     (backing_object = lobject->backing_object) != NULL) {
1071                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1072                             0, ("vm_fault_prefault: unaligned object offset"));
1073                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1074                         VM_OBJECT_LOCK(backing_object);
1075                         VM_OBJECT_UNLOCK(lobject);
1076                         lobject = backing_object;
1077                 }
1078                 /*
1079                  * give-up when a page is not in memory
1080                  */
1081                 if (m == NULL) {
1082                         VM_OBJECT_UNLOCK(lobject);
1083                         break;
1084                 }
1085                 if (m->valid == VM_PAGE_BITS_ALL &&
1086                     (m->flags & PG_FICTITIOUS) == 0)
1087                         pmap_enter_quick(pmap, addr, m, entry->protection);
1088                 VM_OBJECT_UNLOCK(lobject);
1089         }
1090 }
1091
1092 /*
1093  * Hold each of the physical pages that are mapped by the specified range of
1094  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1095  * and allow the specified types of access, "prot".  If all of the implied
1096  * pages are successfully held, then the number of held pages is returned
1097  * together with pointers to those pages in the array "ma".  However, if any
1098  * of the pages cannot be held, -1 is returned.
1099  */
1100 int
1101 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1102     vm_prot_t prot, vm_page_t *ma, int max_count)
1103 {
1104         vm_offset_t end, va;
1105         vm_page_t *mp;
1106         int count;
1107         boolean_t pmap_failed;
1108
1109         if (len == 0)
1110                 return (0);
1111         end = round_page(addr + len);
1112         addr = trunc_page(addr);
1113
1114         /*
1115          * Check for illegal addresses.
1116          */
1117         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1118                 return (-1);
1119
1120         count = howmany(end - addr, PAGE_SIZE);
1121         if (count > max_count)
1122                 panic("vm_fault_quick_hold_pages: count > max_count");
1123
1124         /*
1125          * Most likely, the physical pages are resident in the pmap, so it is
1126          * faster to try pmap_extract_and_hold() first.
1127          */
1128         pmap_failed = FALSE;
1129         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1130                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1131                 if (*mp == NULL)
1132                         pmap_failed = TRUE;
1133                 else if ((prot & VM_PROT_WRITE) != 0 &&
1134                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1135                         /*
1136                          * Explicitly dirty the physical page.  Otherwise, the
1137                          * caller's changes may go unnoticed because they are
1138                          * performed through an unmanaged mapping or by a DMA
1139                          * operation.
1140                          *
1141                          * The object lock is not held here.
1142                          * See vm_page_clear_dirty_mask().
1143                          */
1144                         vm_page_dirty(*mp);
1145                 }
1146         }
1147         if (pmap_failed) {
1148                 /*
1149                  * One or more pages could not be held by the pmap.  Either no
1150                  * page was mapped at the specified virtual address or that
1151                  * mapping had insufficient permissions.  Attempt to fault in
1152                  * and hold these pages.
1153                  */
1154                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1155                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1156                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1157                                 goto error;
1158         }
1159         return (count);
1160 error:
1161         for (mp = ma; mp < ma + count; mp++)
1162                 if (*mp != NULL) {
1163                         vm_page_lock(*mp);
1164                         vm_page_unhold(*mp);
1165                         vm_page_unlock(*mp);
1166                 }
1167         return (-1);
1168 }
1169
1170 /*
1171  *      vm_fault_wire:
1172  *
1173  *      Wire down a range of virtual addresses in a map.
1174  */
1175 int
1176 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1177     boolean_t fictitious)
1178 {
1179         vm_offset_t va;
1180         int rv;
1181
1182         /*
1183          * We simulate a fault to get the page and enter it in the physical
1184          * map.  For user wiring, we only ask for read access on currently
1185          * read-only sections.
1186          */
1187         for (va = start; va < end; va += PAGE_SIZE) {
1188                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1189                 if (rv) {
1190                         if (va != start)
1191                                 vm_fault_unwire(map, start, va, fictitious);
1192                         return (rv);
1193                 }
1194         }
1195         return (KERN_SUCCESS);
1196 }
1197
1198 /*
1199  *      vm_fault_unwire:
1200  *
1201  *      Unwire a range of virtual addresses in a map.
1202  */
1203 void
1204 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1205     boolean_t fictitious)
1206 {
1207         vm_paddr_t pa;
1208         vm_offset_t va;
1209         vm_page_t m;
1210         pmap_t pmap;
1211
1212         pmap = vm_map_pmap(map);
1213
1214         /*
1215          * Since the pages are wired down, we must be able to get their
1216          * mappings from the physical map system.
1217          */
1218         for (va = start; va < end; va += PAGE_SIZE) {
1219                 pa = pmap_extract(pmap, va);
1220                 if (pa != 0) {
1221                         pmap_change_wiring(pmap, va, FALSE);
1222                         if (!fictitious) {
1223                                 m = PHYS_TO_VM_PAGE(pa);
1224                                 vm_page_lock(m);
1225                                 vm_page_unwire(m, TRUE);
1226                                 vm_page_unlock(m);
1227                         }
1228                 }
1229         }
1230 }
1231
1232 /*
1233  *      Routine:
1234  *              vm_fault_copy_entry
1235  *      Function:
1236  *              Create new shadow object backing dst_entry with private copy of
1237  *              all underlying pages. When src_entry is equal to dst_entry,
1238  *              function implements COW for wired-down map entry. Otherwise,
1239  *              it forks wired entry into dst_map.
1240  *
1241  *      In/out conditions:
1242  *              The source and destination maps must be locked for write.
1243  *              The source map entry must be wired down (or be a sharing map
1244  *              entry corresponding to a main map entry that is wired down).
1245  */
1246 void
1247 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1248     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1249     vm_ooffset_t *fork_charge)
1250 {
1251         vm_object_t backing_object, dst_object, object, src_object;
1252         vm_pindex_t dst_pindex, pindex, src_pindex;
1253         vm_prot_t access, prot;
1254         vm_offset_t vaddr;
1255         vm_page_t dst_m;
1256         vm_page_t src_m;
1257         boolean_t src_readonly, upgrade;
1258
1259 #ifdef  lint
1260         src_map++;
1261 #endif  /* lint */
1262
1263         upgrade = src_entry == dst_entry;
1264
1265         src_object = src_entry->object.vm_object;
1266         src_pindex = OFF_TO_IDX(src_entry->offset);
1267         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1268
1269         /*
1270          * Create the top-level object for the destination entry. (Doesn't
1271          * actually shadow anything - we copy the pages directly.)
1272          */
1273         dst_object = vm_object_allocate(OBJT_DEFAULT,
1274             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1275 #if VM_NRESERVLEVEL > 0
1276         dst_object->flags |= OBJ_COLORED;
1277         dst_object->pg_color = atop(dst_entry->start);
1278 #endif
1279
1280         VM_OBJECT_LOCK(dst_object);
1281         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1282             ("vm_fault_copy_entry: vm_object not NULL"));
1283         dst_entry->object.vm_object = dst_object;
1284         dst_entry->offset = 0;
1285         dst_object->charge = dst_entry->end - dst_entry->start;
1286         if (fork_charge != NULL) {
1287                 KASSERT(dst_entry->cred == NULL,
1288                     ("vm_fault_copy_entry: leaked swp charge"));
1289                 dst_object->cred = curthread->td_ucred;
1290                 crhold(dst_object->cred);
1291                 *fork_charge += dst_object->charge;
1292         } else {
1293                 dst_object->cred = dst_entry->cred;
1294                 dst_entry->cred = NULL;
1295         }
1296         access = prot = dst_entry->protection;
1297         /*
1298          * If not an upgrade, then enter the mappings in the pmap as
1299          * read and/or execute accesses.  Otherwise, enter them as
1300          * write accesses.
1301          *
1302          * A writeable large page mapping is only created if all of
1303          * the constituent small page mappings are modified. Marking
1304          * PTEs as modified on inception allows promotion to happen
1305          * without taking potentially large number of soft faults.
1306          */
1307         if (!upgrade)
1308                 access &= ~VM_PROT_WRITE;
1309
1310         /*
1311          * Loop through all of the virtual pages within the entry's
1312          * range, copying each page from the source object to the
1313          * destination object.  Since the source is wired, those pages
1314          * must exist.  In contrast, the destination is pageable.
1315          * Since the destination object does share any backing storage
1316          * with the source object, all of its pages must be dirtied,
1317          * regardless of whether they can be written.
1318          */
1319         for (vaddr = dst_entry->start, dst_pindex = 0;
1320             vaddr < dst_entry->end;
1321             vaddr += PAGE_SIZE, dst_pindex++) {
1322
1323                 /*
1324                  * Allocate a page in the destination object.
1325                  */
1326                 do {
1327                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1328                             VM_ALLOC_NORMAL);
1329                         if (dst_m == NULL) {
1330                                 VM_OBJECT_UNLOCK(dst_object);
1331                                 VM_WAIT;
1332                                 VM_OBJECT_LOCK(dst_object);
1333                         }
1334                 } while (dst_m == NULL);
1335
1336                 /*
1337                  * Find the page in the source object, and copy it in.
1338                  * (Because the source is wired down, the page will be in
1339                  * memory.)
1340                  */
1341                 VM_OBJECT_LOCK(src_object);
1342                 object = src_object;
1343                 pindex = src_pindex + dst_pindex;
1344                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1345                     src_readonly &&
1346                     (backing_object = object->backing_object) != NULL) {
1347                         /*
1348                          * Allow fallback to backing objects if we are reading.
1349                          */
1350                         VM_OBJECT_LOCK(backing_object);
1351                         pindex += OFF_TO_IDX(object->backing_object_offset);
1352                         VM_OBJECT_UNLOCK(object);
1353                         object = backing_object;
1354                 }
1355                 if (src_m == NULL)
1356                         panic("vm_fault_copy_wired: page missing");
1357                 pmap_copy_page(src_m, dst_m);
1358                 VM_OBJECT_UNLOCK(object);
1359                 dst_m->valid = VM_PAGE_BITS_ALL;
1360                 dst_m->dirty = VM_PAGE_BITS_ALL;
1361                 VM_OBJECT_UNLOCK(dst_object);
1362
1363                 /*
1364                  * Enter it in the pmap. If a wired, copy-on-write
1365                  * mapping is being replaced by a write-enabled
1366                  * mapping, then wire that new mapping.
1367                  */
1368                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1369
1370                 /*
1371                  * Mark it no longer busy, and put it on the active list.
1372                  */
1373                 VM_OBJECT_LOCK(dst_object);
1374
1375                 if (upgrade) {
1376                         vm_page_lock(src_m);
1377                         vm_page_unwire(src_m, 0);
1378                         vm_page_unlock(src_m);
1379
1380                         vm_page_lock(dst_m);
1381                         vm_page_wire(dst_m);
1382                         vm_page_unlock(dst_m);
1383                 } else {
1384                         vm_page_lock(dst_m);
1385                         vm_page_activate(dst_m);
1386                         vm_page_unlock(dst_m);
1387                 }
1388                 vm_page_wakeup(dst_m);
1389         }
1390         VM_OBJECT_UNLOCK(dst_object);
1391         if (upgrade) {
1392                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1393                 vm_object_deallocate(src_object);
1394         }
1395 }
1396
1397
1398 /*
1399  * This routine checks around the requested page for other pages that
1400  * might be able to be faulted in.  This routine brackets the viable
1401  * pages for the pages to be paged in.
1402  *
1403  * Inputs:
1404  *      m, rbehind, rahead
1405  *
1406  * Outputs:
1407  *  marray (array of vm_page_t), reqpage (index of requested page)
1408  *
1409  * Return value:
1410  *  number of pages in marray
1411  */
1412 static int
1413 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1414         vm_page_t m;
1415         int rbehind;
1416         int rahead;
1417         vm_page_t *marray;
1418         int *reqpage;
1419 {
1420         int i,j;
1421         vm_object_t object;
1422         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1423         vm_page_t rtm;
1424         int cbehind, cahead;
1425
1426         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1427
1428         object = m->object;
1429         pindex = m->pindex;
1430         cbehind = cahead = 0;
1431
1432         /*
1433          * if the requested page is not available, then give up now
1434          */
1435         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1436                 return 0;
1437         }
1438
1439         if ((cbehind == 0) && (cahead == 0)) {
1440                 *reqpage = 0;
1441                 marray[0] = m;
1442                 return 1;
1443         }
1444
1445         if (rahead > cahead) {
1446                 rahead = cahead;
1447         }
1448
1449         if (rbehind > cbehind) {
1450                 rbehind = cbehind;
1451         }
1452
1453         /*
1454          * scan backward for the read behind pages -- in memory
1455          */
1456         if (pindex > 0) {
1457                 if (rbehind > pindex) {
1458                         rbehind = pindex;
1459                         startpindex = 0;
1460                 } else {
1461                         startpindex = pindex - rbehind;
1462                 }
1463
1464                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1465                     rtm->pindex >= startpindex)
1466                         startpindex = rtm->pindex + 1;
1467
1468                 /* tpindex is unsigned; beware of numeric underflow. */
1469                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1470                     tpindex < pindex; i++, tpindex--) {
1471
1472                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1473                             VM_ALLOC_IFNOTCACHED);
1474                         if (rtm == NULL) {
1475                                 /*
1476                                  * Shift the allocated pages to the
1477                                  * beginning of the array.
1478                                  */
1479                                 for (j = 0; j < i; j++) {
1480                                         marray[j] = marray[j + tpindex + 1 -
1481                                             startpindex];
1482                                 }
1483                                 break;
1484                         }
1485
1486                         marray[tpindex - startpindex] = rtm;
1487                 }
1488         } else {
1489                 startpindex = 0;
1490                 i = 0;
1491         }
1492
1493         marray[i] = m;
1494         /* page offset of the required page */
1495         *reqpage = i;
1496
1497         tpindex = pindex + 1;
1498         i++;
1499
1500         /*
1501          * scan forward for the read ahead pages
1502          */
1503         endpindex = tpindex + rahead;
1504         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1505                 endpindex = rtm->pindex;
1506         if (endpindex > object->size)
1507                 endpindex = object->size;
1508
1509         for (; tpindex < endpindex; i++, tpindex++) {
1510
1511                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1512                     VM_ALLOC_IFNOTCACHED);
1513                 if (rtm == NULL) {
1514                         break;
1515                 }
1516
1517                 marray[i] = rtm;
1518         }
1519
1520         /* return number of pages */
1521         return i;
1522 }
1523
1524 /*
1525  * Block entry into the machine-independent layer's page fault handler by
1526  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1527  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1528  * spurious page faults.
1529  */
1530 int
1531 vm_fault_disable_pagefaults(void)
1532 {
1533
1534         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1535 }
1536
1537 void
1538 vm_fault_enable_pagefaults(int save)
1539 {
1540
1541         curthread_pflags_restore(save);
1542 }