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                 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
 286                         vm_map_unlock_read(fs.map);
 287                         return (KERN_FAILURE);
 288                 }
 289                 panic("vm_fault: fault on nofault entry, addr: %lx",
 290                     (u_long)vaddr);
 291         }
 292
 293         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 294             fs.entry->wiring_thread != curthread) {
 295                 vm_map_unlock_read(fs.map);
 296                 vm_map_lock(fs.map);
 297                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 298                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 299                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 300                         vm_map_unlock_and_wait(fs.map, 0);
 301                 } else
 302                         vm_map_unlock(fs.map);
 303                 goto RetryFault;
 304         }
 305
 306         /*
 307          * Make a reference to this object to prevent its disposal while we
 308          * are messing with it.  Once we have the reference, the map is free
 309          * to be diddled.  Since objects reference their shadows (and copies),
 310          * they will stay around as well.
 311          *
 312          * Bump the paging-in-progress count to prevent size changes (e.g.
 313          * truncation operations) during I/O.  This must be done after
 314          * obtaining the vnode lock in order to avoid possible deadlocks.
 315          */
 316         VM_OBJECT_LOCK(fs.first_object);
 317         vm_object_reference_locked(fs.first_object);
 318         vm_object_pip_add(fs.first_object, 1);
 319
 320         fs.lookup_still_valid = TRUE;
 321
 322         if (wired)
 323                 fault_type = prot | (fault_type & VM_PROT_COPY);
 324
 325         fs.first_m = NULL;
 326
 327         /*
 328          * Search for the page at object/offset.
 329          */
 330         fs.object = fs.first_object;
 331         fs.pindex = fs.first_pindex;
 332         while (TRUE) {
 333                 /*
 334                  * If the object is dead, we stop here
 335                  */
 336                 if (fs.object->flags & OBJ_DEAD) {
 337                         unlock_and_deallocate(&fs);
 338                         return (KERN_PROTECTION_FAILURE);
 339                 }
 340
 341                 /*
 342                  * See if page is resident
 343                  */
 344                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 345                 if (fs.m != NULL) {
 346                         /*
 347                          * check for page-based copy on write.
 348                          * We check fs.object == fs.first_object so
 349                          * as to ensure the legacy COW mechanism is
 350                          * used when the page in question is part of
 351                          * a shadow object.  Otherwise, vm_page_cowfault()
 352                          * removes the page from the backing object,
 353                          * which is not what we want.
 354                          */
 355                         vm_page_lock(fs.m);
 356                         if ((fs.m->cow) &&
 357                             (fault_type & VM_PROT_WRITE) &&
 358                             (fs.object == fs.first_object)) {
 359                                 vm_page_cowfault(fs.m);
 360                                 unlock_and_deallocate(&fs);
 361                                 goto RetryFault;
 362                         }
 363
 364                         /*
 365                          * Wait/Retry if the page is busy.  We have to do this
 366                          * if the page is busy via either VPO_BUSY or
 367                          * vm_page_t->busy because the vm_pager may be using
 368                          * vm_page_t->busy for pageouts ( and even pageins if
 369                          * it is the vnode pager ), and we could end up trying
 370                          * to pagein and pageout the same page simultaneously.
 371                          *
 372                          * We can theoretically allow the busy case on a read
 373                          * fault if the page is marked valid, but since such
 374                          * pages are typically already pmap'd, putting that
 375                          * special case in might be more effort then it is
 376                          * worth.  We cannot under any circumstances mess
 377                          * around with a vm_page_t->busy page except, perhaps,
 378                          * to pmap it.
 379                          */
 380                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 381                                 /*
 382                                  * Reference the page before unlocking and
 383                                  * sleeping so that the page daemon is less
 384                                  * likely to reclaim it.
 385                                  */
 386                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 387                                 vm_page_unlock(fs.m);
 388                                 if (fs.object != fs.first_object) {
 389                                         if (!VM_OBJECT_TRYLOCK(
 390                                             fs.first_object)) {
 391                                                 VM_OBJECT_UNLOCK(fs.object);
 392                                                 VM_OBJECT_LOCK(fs.first_object);
 393                                                 VM_OBJECT_LOCK(fs.object);
 394                                         }
 395                                         vm_page_lock(fs.first_m);
 396                                         vm_page_free(fs.first_m);
 397                                         vm_page_unlock(fs.first_m);
 398                                         vm_object_pip_wakeup(fs.first_object);
 399                                         VM_OBJECT_UNLOCK(fs.first_object);
 400                                         fs.first_m = NULL;
 401                                 }
 402                                 unlock_map(&fs);
 403                                 if (fs.m == vm_page_lookup(fs.object,
 404                                     fs.pindex)) {
 405                                         vm_page_sleep_if_busy(fs.m, TRUE,
 406                                             "vmpfw");
 407                                 }
 408                                 vm_object_pip_wakeup(fs.object);
 409                                 VM_OBJECT_UNLOCK(fs.object);
 410                                 PCPU_INC(cnt.v_intrans);
 411                                 vm_object_deallocate(fs.first_object);
 412                                 goto RetryFault;
 413                         }
 414                         vm_pageq_remove(fs.m);
 415                         vm_page_unlock(fs.m);
 416
 417                         /*
 418                          * Mark page busy for other processes, and the
 419                          * pagedaemon.  If it still isn't completely valid
 420                          * (readable), jump to readrest, else break-out ( we
 421                          * found the page ).
 422                          */
 423                         vm_page_busy(fs.m);
 424                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 425                                 goto readrest;
 426                         break;
 427                 }
 428
 429                 /*
 430                  * Page is not resident, If this is the search termination
 431                  * or the pager might contain the page, allocate a new page.
 432                  */
 433                 if (TRYPAGER || fs.object == fs.first_object) {
 434                         if (fs.pindex >= fs.object->size) {
 435                                 unlock_and_deallocate(&fs);
 436                                 return (KERN_PROTECTION_FAILURE);
 437                         }
 438
 439                         /*
 440                          * Allocate a new page for this object/offset pair.
 441                          *
 442                          * Unlocked read of the p_flag is harmless. At
 443                          * worst, the P_KILLED might be not observed
 444                          * there, and allocation can fail, causing
 445                          * restart and new reading of the p_flag.
 446                          */
 447                         fs.m = NULL;
 448                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 449 #if VM_NRESERVLEVEL > 0
 450                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 451                                         fs.object->flags |= OBJ_COLORED;
 452                                         fs.object->pg_color = atop(vaddr) -
 453                                             fs.pindex;
 454                                 }
 455 #endif
 456                                 alloc_req = P_KILLED(curproc) ?
 457                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 458                                 if (fs.object->type != OBJT_VNODE &&
 459                                     fs.object->backing_object == NULL)
 460                                         alloc_req |= VM_ALLOC_ZERO;
 461                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 462                                     alloc_req);
 463                         }
 464                         if (fs.m == NULL) {
 465                                 unlock_and_deallocate(&fs);
 466                                 VM_WAITPFAULT;
 467                                 goto RetryFault;
 468                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 469                                 break;
 470                 }
 471
 472 readrest:
 473                 /*
 474                  * We have found a valid page or we have allocated a new page.
 475                  * The page thus may not be valid or may not be entirely
 476                  * valid.
 477                  *
 478                  * Attempt to fault-in the page if there is a chance that the
 479                  * pager has it, and potentially fault in additional pages
 480                  * at the same time.
 481                  */
 482                 if (TRYPAGER) {
 483                         int rv;
 484                         u_char behavior = vm_map_entry_behavior(fs.entry);
 485
 486                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 487                             P_KILLED(curproc)) {
 488                                 behind = 0;
 489                                 ahead = 0;
 490                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 491                                 behind = 0;
 492                                 ahead = atop(fs.entry->end - vaddr) - 1;
 493                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 494                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 495                                 if (fs.pindex == fs.entry->next_read)
 496                                         vm_fault_cache_behind(&fs,
 497                                             VM_FAULT_READ_MAX);
 498                         } else {
 499                                 /*
 500                                  * If this is a sequential page fault, then
 501                                  * arithmetically increase the number of pages
 502                                  * in the read-ahead window.  Otherwise, reset
 503                                  * the read-ahead window to its smallest size.
 504                                  */
 505                                 behind = atop(vaddr - fs.entry->start);
 506                                 if (behind > VM_FAULT_READ_BEHIND)
 507                                         behind = VM_FAULT_READ_BEHIND;
 508                                 ahead = atop(fs.entry->end - vaddr) - 1;
 509                                 era = fs.entry->read_ahead;
 510                                 if (fs.pindex == fs.entry->next_read) {
 511                                         nera = era + behind;
 512                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 513                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 514                                         behind = 0;
 515                                         if (ahead > nera)
 516                                                 ahead = nera;
 517                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 518                                                 vm_fault_cache_behind(&fs,
 519                                                     VM_FAULT_CACHE_BEHIND);
 520                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 521                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 522                                 if (era != ahead)
 523                                         fs.entry->read_ahead = ahead;
 524                         }
 525
 526                         /*
 527                          * Call the pager to retrieve the data, if any, after
 528                          * releasing the lock on the map.  We hold a ref on
 529                          * fs.object and the pages are VPO_BUSY'd.
 530                          */
 531                         unlock_map(&fs);
 532
 533 vnode_lock:
 534                         if (fs.object->type == OBJT_VNODE) {
 535                                 vp = fs.object->handle;
 536                                 if (vp == fs.vp)
 537                                         goto vnode_locked;
 538                                 else if (fs.vp != NULL) {
 539                                         vput(fs.vp);
 540                                         fs.vp = NULL;
 541                                 }
 542                                 locked = VOP_ISLOCKED(vp);
 543
 544                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 545                                         fs.vfslocked = 1;
 546                                         if (!mtx_trylock(&Giant)) {
 547                                                 VM_OBJECT_UNLOCK(fs.object);
 548                                                 mtx_lock(&Giant);
 549                                                 VM_OBJECT_LOCK(fs.object);
 550                                                 goto vnode_lock;
 551                                         }
 552                                 }
 553                                 if (locked != LK_EXCLUSIVE)
 554                                         locked = LK_SHARED;
 555                                 /* Do not sleep for vnode lock while fs.m is busy */
 556                                 error = vget(vp, locked | LK_CANRECURSE |
 557                                     LK_NOWAIT, curthread);
 558                                 if (error != 0) {
 559                                         int vfslocked;
 560
 561                                         vfslocked = fs.vfslocked;
 562                                         fs.vfslocked = 0; /* Keep Giant */
 563                                         vhold(vp);
 564                                         release_page(&fs);
 565                                         unlock_and_deallocate(&fs);
 566                                         error = vget(vp, locked | LK_RETRY |
 567                                             LK_CANRECURSE, curthread);
 568                                         vdrop(vp);
 569                                         fs.vp = vp;
 570                                         fs.vfslocked = vfslocked;
 571                                         KASSERT(error == 0,
 572                                             ("vm_fault: vget failed"));
 573                                         goto RetryFault;
 574                                 }
 575                                 fs.vp = vp;
 576                         }
 577 vnode_locked:
 578                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 579                             ("vm_fault: vnode-backed object mapped by system map"));
 580
 581                         /*
 582                          * now we find out if any other pages should be paged
 583                          * in at this time this routine checks to see if the
 584                          * pages surrounding this fault reside in the same
 585                          * object as the page for this fault.  If they do,
 586                          * then they are faulted in also into the object.  The
 587                          * array "marray" returned contains an array of
 588                          * vm_page_t structs where one of them is the
 589                          * vm_page_t passed to the routine.  The reqpage
 590                          * return value is the index into the marray for the
 591                          * vm_page_t passed to the routine.
 592                          *
 593                          * fs.m plus the additional pages are VPO_BUSY'd.
 594                          */
 595                         faultcount = vm_fault_additional_pages(
 596                             fs.m, behind, ahead, marray, &reqpage);
 597
 598                         rv = faultcount ?
 599                             vm_pager_get_pages(fs.object, marray, faultcount,
 600                                 reqpage) : VM_PAGER_FAIL;
 601
 602                         if (rv == VM_PAGER_OK) {
 603                                 /*
 604                                  * Found the page. Leave it busy while we play
 605                                  * with it.
 606                                  */
 607
 608                                 /*
 609                                  * Relookup in case pager changed page. Pager
 610                                  * is responsible for disposition of old page
 611                                  * if moved.
 612                                  */
 613                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 614                                 if (!fs.m) {
 615                                         unlock_and_deallocate(&fs);
 616                                         goto RetryFault;
 617                                 }
 618
 619                                 hardfault++;
 620                                 break; /* break to PAGE HAS BEEN FOUND */
 621                         }
 622                         /*
 623                          * Remove the bogus page (which does not exist at this
 624                          * object/offset); before doing so, we must get back
 625                          * our object lock to preserve our invariant.
 626                          *
 627                          * Also wake up any other process that may want to bring
 628                          * in this page.
 629                          *
 630                          * If this is the top-level object, we must leave the
 631                          * busy page to prevent another process from rushing
 632                          * past us, and inserting the page in that object at
 633                          * the same time that we are.
 634                          */
 635                         if (rv == VM_PAGER_ERROR)
 636                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 637                                     curproc->p_pid, curproc->p_comm);
 638                         /*
 639                          * Data outside the range of the pager or an I/O error
 640                          */
 641                         /*
 642                          * XXX - the check for kernel_map is a kludge to work
 643                          * around having the machine panic on a kernel space
 644                          * fault w/ I/O error.
 645                          */
 646                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 647                                 (rv == VM_PAGER_BAD)) {
 648                                 vm_page_lock(fs.m);
 649                                 vm_page_free(fs.m);
 650                                 vm_page_unlock(fs.m);
 651                                 fs.m = NULL;
 652                                 unlock_and_deallocate(&fs);
 653                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 654                         }
 655                         if (fs.object != fs.first_object) {
 656                                 vm_page_lock(fs.m);
 657                                 vm_page_free(fs.m);
 658                                 vm_page_unlock(fs.m);
 659                                 fs.m = NULL;
 660                                 /*
 661                                  * XXX - we cannot just fall out at this
 662                                  * point, m has been freed and is invalid!
 663                                  */
 664                         }
 665                 }
 666
 667                 /*
 668                  * We get here if the object has default pager (or unwiring)
 669                  * or the pager doesn't have the page.
 670                  */
 671                 if (fs.object == fs.first_object)
 672                         fs.first_m = fs.m;
 673
 674                 /*
 675                  * Move on to the next object.  Lock the next object before
 676                  * unlocking the current one.
 677                  */
 678                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 679                 next_object = fs.object->backing_object;
 680                 if (next_object == NULL) {
 681                         /*
 682                          * If there's no object left, fill the page in the top
 683                          * object with zeros.
 684                          */
 685                         if (fs.object != fs.first_object) {
 686                                 vm_object_pip_wakeup(fs.object);
 687                                 VM_OBJECT_UNLOCK(fs.object);
 688
 689                                 fs.object = fs.first_object;
 690                                 fs.pindex = fs.first_pindex;
 691                                 fs.m = fs.first_m;
 692                                 VM_OBJECT_LOCK(fs.object);
 693                         }
 694                         fs.first_m = NULL;
 695
 696                         /*
 697                          * Zero the page if necessary and mark it valid.
 698                          */
 699                         if ((fs.m->flags & PG_ZERO) == 0) {
 700                                 pmap_zero_page(fs.m);
 701                         } else {
 702                                 PCPU_INC(cnt.v_ozfod);
 703                         }
 704                         PCPU_INC(cnt.v_zfod);
 705                         fs.m->valid = VM_PAGE_BITS_ALL;
 706                         break;  /* break to PAGE HAS BEEN FOUND */
 707                 } else {
 708                         KASSERT(fs.object != next_object,
 709                             ("object loop %p", next_object));
 710                         VM_OBJECT_LOCK(next_object);
 711                         vm_object_pip_add(next_object, 1);
 712                         if (fs.object != fs.first_object)
 713                                 vm_object_pip_wakeup(fs.object);
 714                         VM_OBJECT_UNLOCK(fs.object);
 715                         fs.object = next_object;
 716                 }
 717         }
 718
 719         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 720             ("vm_fault: not busy after main loop"));
 721
 722         /*
 723          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 724          * is held.]
 725          */
 726
 727         /*
 728          * If the page is being written, but isn't already owned by the
 729          * top-level object, we have to copy it into a new page owned by the
 730          * top-level object.
 731          */
 732         if (fs.object != fs.first_object) {
 733                 /*
 734                  * We only really need to copy if we want to write it.
 735                  */
 736                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 737                         /*
 738                          * This allows pages to be virtually copied from a
 739                          * backing_object into the first_object, where the
 740                          * backing object has no other refs to it, and cannot
 741                          * gain any more refs.  Instead of a bcopy, we just
 742                          * move the page from the backing object to the
 743                          * first object.  Note that we must mark the page
 744                          * dirty in the first object so that it will go out
 745                          * to swap when needed.
 746                          */
 747                         is_first_object_locked = FALSE;
 748                         if (
 749                                 /*
 750                                  * Only one shadow object
 751                                  */
 752                                 (fs.object->shadow_count == 1) &&
 753                                 /*
 754                                  * No COW refs, except us
 755                                  */
 756                                 (fs.object->ref_count == 1) &&
 757                                 /*
 758                                  * No one else can look this object up
 759                                  */
 760                                 (fs.object->handle == NULL) &&
 761                                 /*
 762                                  * No other ways to look the object up
 763                                  */
 764                                 ((fs.object->type == OBJT_DEFAULT) ||
 765                                  (fs.object->type == OBJT_SWAP)) &&
 766                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 767                                 /*
 768                                  * We don't chase down the shadow chain
 769                                  */
 770                             fs.object == fs.first_object->backing_object) {
 771                                 /*
 772                                  * get rid of the unnecessary page
 773                                  */
 774                                 vm_page_lock(fs.first_m);
 775                                 vm_page_free(fs.first_m);
 776                                 vm_page_unlock(fs.first_m);
 777                                 /*
 778                                  * grab the page and put it into the
 779                                  * process'es object.  The page is
 780                                  * automatically made dirty.
 781                                  */
 782                                 vm_page_lock(fs.m);
 783                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 784                                 vm_page_unlock(fs.m);
 785                                 vm_page_busy(fs.m);
 786                                 fs.first_m = fs.m;
 787                                 fs.m = NULL;
 788                                 PCPU_INC(cnt.v_cow_optim);
 789                         } else {
 790                                 /*
 791                                  * Oh, well, lets copy it.
 792                                  */
 793                                 pmap_copy_page(fs.m, fs.first_m);
 794                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 795                                 if (wired && (fault_flags &
 796                                     VM_FAULT_CHANGE_WIRING) == 0) {
 797                                         vm_page_lock(fs.first_m);
 798                                         vm_page_wire(fs.first_m);
 799                                         vm_page_unlock(fs.first_m);
 800
 801                                         vm_page_lock(fs.m);
 802                                         vm_page_unwire(fs.m, FALSE);
 803                                         vm_page_unlock(fs.m);
 804                                 }
 805                                 /*
 806                                  * We no longer need the old page or object.
 807                                  */
 808                                 release_page(&fs);
 809                         }
 810                         /*
 811                          * fs.object != fs.first_object due to above
 812                          * conditional
 813                          */
 814                         vm_object_pip_wakeup(fs.object);
 815                         VM_OBJECT_UNLOCK(fs.object);
 816                         /*
 817                          * Only use the new page below...
 818                          */
 819                         fs.object = fs.first_object;
 820                         fs.pindex = fs.first_pindex;
 821                         fs.m = fs.first_m;
 822                         if (!is_first_object_locked)
 823                                 VM_OBJECT_LOCK(fs.object);
 824                         PCPU_INC(cnt.v_cow_faults);
 825                         curthread->td_cow++;
 826                 } else {
 827                         prot &= ~VM_PROT_WRITE;
 828                 }
 829         }
 830
 831         /*
 832          * We must verify that the maps have not changed since our last
 833          * lookup.
 834          */
 835         if (!fs.lookup_still_valid) {
 836                 vm_object_t retry_object;
 837                 vm_pindex_t retry_pindex;
 838                 vm_prot_t retry_prot;
 839
 840                 if (!vm_map_trylock_read(fs.map)) {
 841                         release_page(&fs);
 842                         unlock_and_deallocate(&fs);
 843                         goto RetryFault;
 844                 }
 845                 fs.lookup_still_valid = TRUE;
 846                 if (fs.map->timestamp != map_generation) {
 847                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 848                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 849
 850                         /*
 851                          * If we don't need the page any longer, put it on the inactive
 852                          * list (the easiest thing to do here).  If no one needs it,
 853                          * pageout will grab it eventually.
 854                          */
 855                         if (result != KERN_SUCCESS) {
 856                                 release_page(&fs);
 857                                 unlock_and_deallocate(&fs);
 858
 859                                 /*
 860                                  * If retry of map lookup would have blocked then
 861                                  * retry fault from start.
 862                                  */
 863                                 if (result == KERN_FAILURE)
 864                                         goto RetryFault;
 865                                 return (result);
 866                         }
 867                         if ((retry_object != fs.first_object) ||
 868                             (retry_pindex != fs.first_pindex)) {
 869                                 release_page(&fs);
 870                                 unlock_and_deallocate(&fs);
 871                                 goto RetryFault;
 872                         }
 873
 874                         /*
 875                          * Check whether the protection has changed or the object has
 876                          * been copied while we left the map unlocked. Changing from
 877                          * read to write permission is OK - we leave the page
 878                          * write-protected, and catch the write fault. Changing from
 879                          * write to read permission means that we can't mark the page
 880                          * write-enabled after all.
 881                          */
 882                         prot &= retry_prot;
 883                 }
 884         }
 885         /*
 886          * If the page was filled by a pager, update the map entry's
 887          * last read offset.  Since the pager does not return the
 888          * actual set of pages that it read, this update is based on
 889          * the requested set.  Typically, the requested and actual
 890          * sets are the same.
 891          *
 892          * XXX The following assignment modifies the map
 893          * without holding a write lock on it.
 894          */
 895         if (hardfault)
 896                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 897
 898         if ((prot & VM_PROT_WRITE) != 0 ||
 899             (fault_flags & VM_FAULT_DIRTY) != 0) {
 900                 vm_object_set_writeable_dirty(fs.object);
 901
 902                 /*
 903                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 904                  * if the page is already dirty to prevent data written with
 905                  * the expectation of being synced from not being synced.
 906                  * Likewise if this entry does not request NOSYNC then make
 907                  * sure the page isn't marked NOSYNC.  Applications sharing
 908                  * data should use the same flags to avoid ping ponging.
 909                  */
 910                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 911                         if (fs.m->dirty == 0)
 912                                 fs.m->oflags |= VPO_NOSYNC;
 913                 } else {
 914                         fs.m->oflags &= ~VPO_NOSYNC;
 915                 }
 916
 917                 /*
 918                  * If the fault is a write, we know that this page is being
 919                  * written NOW so dirty it explicitly to save on
 920                  * pmap_is_modified() calls later.
 921                  *
 922                  * Also tell the backing pager, if any, that it should remove
 923                  * any swap backing since the page is now dirty.
 924                  */
 925                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 926                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 927                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 928                         vm_page_dirty(fs.m);
 929                         vm_pager_page_unswapped(fs.m);
 930                 }
 931         }
 932
 933         /*
 934          * Page had better still be busy
 935          */
 936         KASSERT(fs.m->oflags & VPO_BUSY,
 937                 ("vm_fault: page %p not busy!", fs.m));
 938         /*
 939          * Page must be completely valid or it is not fit to
 940          * map into user space.  vm_pager_get_pages() ensures this.
 941          */
 942         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 943             ("vm_fault: page %p partially invalid", fs.m));
 944         VM_OBJECT_UNLOCK(fs.object);
 945
 946         /*
 947          * Put this page into the physical map.  We had to do the unlock above
 948          * because pmap_enter() may sleep.  We don't put the page
 949          * back on the active queue until later so that the pageout daemon
 950          * won't find it (yet).
 951          */
 952         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 953         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 954                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 955         VM_OBJECT_LOCK(fs.object);
 956         vm_page_lock(fs.m);
 957
 958         /*
 959          * If the page is not wired down, then put it where the pageout daemon
 960          * can find it.
 961          */
 962         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 963                 if (wired)
 964                         vm_page_wire(fs.m);
 965                 else
 966                         vm_page_unwire(fs.m, 1);
 967         } else
 968                 vm_page_activate(fs.m);
 969         if (m_hold != NULL) {
 970                 *m_hold = fs.m;
 971                 vm_page_hold(fs.m);
 972         }
 973         vm_page_unlock(fs.m);
 974         vm_page_wakeup(fs.m);
 975
 976         /*
 977          * Unlock everything, and return
 978          */
 979         unlock_and_deallocate(&fs);
 980         if (hardfault)
 981                 curthread->td_ru.ru_majflt++;
 982         else
 983                 curthread->td_ru.ru_minflt++;
 984
 985         return (KERN_SUCCESS);
 986 }
 987
 988 /*
 989  * Speed up the reclamation of up to "distance" pages that precede the
 990  * faulting pindex within the first object of the shadow chain.
 991  */
 992 static void
 993 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 994 {
 995         vm_object_t first_object, object;
 996         vm_page_t m, m_prev;
 997         vm_pindex_t pindex;
 998
 999         object = fs->object;
1000         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1001         first_object = fs->first_object;
1002         if (first_object != object) {
1003                 if (!VM_OBJECT_TRYLOCK(first_object)) {
1004                         VM_OBJECT_UNLOCK(object);
1005                         VM_OBJECT_LOCK(first_object);
1006                         VM_OBJECT_LOCK(object);
1007                 }
1008         }
1009         if (first_object->type != OBJT_DEVICE &&
1010             first_object->type != OBJT_PHYS && first_object->type != OBJT_SG) {
1011                 if (fs->first_pindex < distance)
1012                         pindex = 0;
1013                 else
1014                         pindex = fs->first_pindex - distance;
1015                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1016                         pindex = OFF_TO_IDX(fs->entry->offset);
1017                 m = first_object != object ? fs->first_m : fs->m;
1018                 KASSERT((m->oflags & VPO_BUSY) != 0,
1019                     ("vm_fault_cache_behind: page %p is not busy", m));
1020                 m_prev = vm_page_prev(m);
1021                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1022                     m->valid == VM_PAGE_BITS_ALL) {
1023                         m_prev = vm_page_prev(m);
1024                         if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0)
1025                                 continue;
1026                         vm_page_lock(m);
1027                         if (m->hold_count == 0 && m->wire_count == 0) {
1028                                 pmap_remove_all(m);
1029                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1030                                 if (m->dirty != 0)
1031                                         vm_page_deactivate(m);
1032                                 else
1033                                         vm_page_cache(m);
1034                         }
1035                         vm_page_unlock(m);
1036                 }
1037         }
1038         if (first_object != object)
1039                 VM_OBJECT_UNLOCK(first_object);
1040 }
1041
1042 /*
1043  * vm_fault_prefault provides a quick way of clustering
1044  * pagefaults into a processes address space.  It is a "cousin"
1045  * of vm_map_pmap_enter, except it runs at page fault time instead
1046  * of mmap time.
1047  */
1048 static void
1049 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1050 {
1051         int i;
1052         vm_offset_t addr, starta;
1053         vm_pindex_t pindex;
1054         vm_page_t m;
1055         vm_object_t object;
1056
1057         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1058                 return;
1059
1060         object = entry->object.vm_object;
1061
1062         starta = addra - PFBAK * PAGE_SIZE;
1063         if (starta < entry->start) {
1064                 starta = entry->start;
1065         } else if (starta > addra) {
1066                 starta = 0;
1067         }
1068
1069         for (i = 0; i < PAGEORDER_SIZE; i++) {
1070                 vm_object_t backing_object, lobject;
1071
1072                 addr = addra + prefault_pageorder[i];
1073                 if (addr > addra + (PFFOR * PAGE_SIZE))
1074                         addr = 0;
1075
1076                 if (addr < starta || addr >= entry->end)
1077                         continue;
1078
1079                 if (!pmap_is_prefaultable(pmap, addr))
1080                         continue;
1081
1082                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1083                 lobject = object;
1084                 VM_OBJECT_LOCK(lobject);
1085                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1086                     lobject->type == OBJT_DEFAULT &&
1087                     (backing_object = lobject->backing_object) != NULL) {
1088                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1089                             0, ("vm_fault_prefault: unaligned object offset"));
1090                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1091                         VM_OBJECT_LOCK(backing_object);
1092                         VM_OBJECT_UNLOCK(lobject);
1093                         lobject = backing_object;
1094                 }
1095                 /*
1096                  * give-up when a page is not in memory
1097                  */
1098                 if (m == NULL) {
1099                         VM_OBJECT_UNLOCK(lobject);
1100                         break;
1101                 }
1102                 if (m->valid == VM_PAGE_BITS_ALL &&
1103                     (m->flags & PG_FICTITIOUS) == 0)
1104                         pmap_enter_quick(pmap, addr, m, entry->protection);
1105                 VM_OBJECT_UNLOCK(lobject);
1106         }
1107 }
1108
1109 /*
1110  * Hold each of the physical pages that are mapped by the specified range of
1111  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1112  * and allow the specified types of access, "prot".  If all of the implied
1113  * pages are successfully held, then the number of held pages is returned
1114  * together with pointers to those pages in the array "ma".  However, if any
1115  * of the pages cannot be held, -1 is returned.
1116  */
1117 int
1118 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1119     vm_prot_t prot, vm_page_t *ma, int max_count)
1120 {
1121         vm_offset_t end, va;
1122         vm_page_t *mp;
1123         int count;
1124         boolean_t pmap_failed;
1125
1126         if (len == 0)
1127                 return (0);
1128         end = round_page(addr + len);
1129         addr = trunc_page(addr);
1130
1131         /*
1132          * Check for illegal addresses.
1133          */
1134         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1135                 return (-1);
1136
1137         if (atop(end - addr) > max_count)
1138                 panic("vm_fault_quick_hold_pages: count > max_count");
1139         count = atop(end - addr);
1140
1141         /*
1142          * Most likely, the physical pages are resident in the pmap, so it is
1143          * faster to try pmap_extract_and_hold() first.
1144          */
1145         pmap_failed = FALSE;
1146         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1147                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1148                 if (*mp == NULL)
1149                         pmap_failed = TRUE;
1150                 else if ((prot & VM_PROT_WRITE) != 0 &&
1151                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1152                         /*
1153                          * Explicitly dirty the physical page.  Otherwise, the
1154                          * caller's changes may go unnoticed because they are
1155                          * performed through an unmanaged mapping or by a DMA
1156                          * operation.
1157                          *
1158                          * The object lock is not held here.
1159                          * See vm_page_clear_dirty_mask().
1160                          */
1161                         vm_page_dirty(*mp);
1162                 }
1163         }
1164         if (pmap_failed) {
1165                 /*
1166                  * One or more pages could not be held by the pmap.  Either no
1167                  * page was mapped at the specified virtual address or that
1168                  * mapping had insufficient permissions.  Attempt to fault in
1169                  * and hold these pages.
1170                  */
1171                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1172                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1173                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1174                                 goto error;
1175         }
1176         return (count);
1177 error:
1178         for (mp = ma; mp < ma + count; mp++)
1179                 if (*mp != NULL) {
1180                         vm_page_lock(*mp);
1181                         vm_page_unhold(*mp);
1182                         vm_page_unlock(*mp);
1183                 }
1184         return (-1);
1185 }
1186
1187 /*
1188  *      vm_fault_wire:
1189  *
1190  *      Wire down a range of virtual addresses in a map.
1191  */
1192 int
1193 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1194     boolean_t fictitious)
1195 {
1196         vm_offset_t va;
1197         int rv;
1198
1199         /*
1200          * We simulate a fault to get the page and enter it in the physical
1201          * map.  For user wiring, we only ask for read access on currently
1202          * read-only sections.
1203          */
1204         for (va = start; va < end; va += PAGE_SIZE) {
1205                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1206                 if (rv) {
1207                         if (va != start)
1208                                 vm_fault_unwire(map, start, va, fictitious);
1209                         return (rv);
1210                 }
1211         }
1212         return (KERN_SUCCESS);
1213 }
1214
1215 /*
1216  *      vm_fault_unwire:
1217  *
1218  *      Unwire a range of virtual addresses in a map.
1219  */
1220 void
1221 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1222     boolean_t fictitious)
1223 {
1224         vm_paddr_t pa;
1225         vm_offset_t va;
1226         vm_page_t m;
1227         pmap_t pmap;
1228
1229         pmap = vm_map_pmap(map);
1230
1231         /*
1232          * Since the pages are wired down, we must be able to get their
1233          * mappings from the physical map system.
1234          */
1235         for (va = start; va < end; va += PAGE_SIZE) {
1236                 pa = pmap_extract(pmap, va);
1237                 if (pa != 0) {
1238                         pmap_change_wiring(pmap, va, FALSE);
1239                         if (!fictitious) {
1240                                 m = PHYS_TO_VM_PAGE(pa);
1241                                 vm_page_lock(m);
1242                                 vm_page_unwire(m, TRUE);
1243                                 vm_page_unlock(m);
1244                         }
1245                 }
1246         }
1247 }
1248
1249 /*
1250  *      Routine:
1251  *              vm_fault_copy_entry
1252  *      Function:
1253  *              Create new shadow object backing dst_entry with private copy of
1254  *              all underlying pages. When src_entry is equal to dst_entry,
1255  *              function implements COW for wired-down map entry. Otherwise,
1256  *              it forks wired entry into dst_map.
1257  *
1258  *      In/out conditions:
1259  *              The source and destination maps must be locked for write.
1260  *              The source map entry must be wired down (or be a sharing map
1261  *              entry corresponding to a main map entry that is wired down).
1262  */
1263 void
1264 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1265     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1266     vm_ooffset_t *fork_charge)
1267 {
1268         vm_object_t backing_object, dst_object, object, src_object;
1269         vm_pindex_t dst_pindex, pindex, src_pindex;
1270         vm_prot_t access, prot;
1271         vm_offset_t vaddr;
1272         vm_page_t dst_m;
1273         vm_page_t src_m;
1274         boolean_t upgrade;
1275
1276 #ifdef  lint
1277         src_map++;
1278 #endif  /* lint */
1279
1280         upgrade = src_entry == dst_entry;
1281
1282         src_object = src_entry->object.vm_object;
1283         src_pindex = OFF_TO_IDX(src_entry->offset);
1284
1285         /*
1286          * Create the top-level object for the destination entry. (Doesn't
1287          * actually shadow anything - we copy the pages directly.)
1288          */
1289         dst_object = vm_object_allocate(OBJT_DEFAULT,
1290             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1291 #if VM_NRESERVLEVEL > 0
1292         dst_object->flags |= OBJ_COLORED;
1293         dst_object->pg_color = atop(dst_entry->start);
1294 #endif
1295
1296         VM_OBJECT_LOCK(dst_object);
1297         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1298             ("vm_fault_copy_entry: vm_object not NULL"));
1299         dst_entry->object.vm_object = dst_object;
1300         dst_entry->offset = 0;
1301         dst_object->charge = dst_entry->end - dst_entry->start;
1302         if (fork_charge != NULL) {
1303                 KASSERT(dst_entry->cred == NULL,
1304                     ("vm_fault_copy_entry: leaked swp charge"));
1305                 dst_object->cred = curthread->td_ucred;
1306                 crhold(dst_object->cred);
1307                 *fork_charge += dst_object->charge;
1308         } else {
1309                 dst_object->cred = dst_entry->cred;
1310                 dst_entry->cred = NULL;
1311         }
1312         access = prot = dst_entry->protection;
1313         /*
1314          * If not an upgrade, then enter the mappings in the pmap as
1315          * read and/or execute accesses.  Otherwise, enter them as
1316          * write accesses.
1317          *
1318          * A writeable large page mapping is only created if all of
1319          * the constituent small page mappings are modified. Marking
1320          * PTEs as modified on inception allows promotion to happen
1321          * without taking potentially large number of soft faults.
1322          */
1323         if (!upgrade)
1324                 access &= ~VM_PROT_WRITE;
1325
1326         /*
1327          * Loop through all of the virtual pages within the entry's
1328          * range, copying each page from the source object to the
1329          * destination object.  Since the source is wired, those pages
1330          * must exist.  In contrast, the destination is pageable.
1331          * Since the destination object does share any backing storage
1332          * with the source object, all of its pages must be dirtied,
1333          * regardless of whether they can be written.
1334          */
1335         for (vaddr = dst_entry->start, dst_pindex = 0;
1336             vaddr < dst_entry->end;
1337             vaddr += PAGE_SIZE, dst_pindex++) {
1338
1339                 /*
1340                  * Allocate a page in the destination object.
1341                  */
1342                 do {
1343                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1344                             VM_ALLOC_NORMAL);
1345                         if (dst_m == NULL) {
1346                                 VM_OBJECT_UNLOCK(dst_object);
1347                                 VM_WAIT;
1348                                 VM_OBJECT_LOCK(dst_object);
1349                         }
1350                 } while (dst_m == NULL);
1351
1352                 /*
1353                  * Find the page in the source object, and copy it in.
1354                  * Because the source is wired down, the page will be
1355                  * in memory.
1356                  */
1357                 VM_OBJECT_LOCK(src_object);
1358                 object = src_object;
1359                 pindex = src_pindex + dst_pindex;
1360                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1361                     (backing_object = object->backing_object) != NULL) {
1362                         /*
1363                          * Unless the source mapping is read-only or
1364                          * it is presently being upgraded from
1365                          * read-only, the first object in the shadow
1366                          * chain should provide all of the pages.  In
1367                          * other words, this loop body should never be
1368                          * executed when the source mapping is already
1369                          * read/write.
1370                          */
1371                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1372                             upgrade,
1373                             ("vm_fault_copy_entry: main object missing page"));
1374
1375                         VM_OBJECT_LOCK(backing_object);
1376                         pindex += OFF_TO_IDX(object->backing_object_offset);
1377                         VM_OBJECT_UNLOCK(object);
1378                         object = backing_object;
1379                 }
1380                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1381                 pmap_copy_page(src_m, dst_m);
1382                 VM_OBJECT_UNLOCK(object);
1383                 dst_m->valid = VM_PAGE_BITS_ALL;
1384                 dst_m->dirty = VM_PAGE_BITS_ALL;
1385                 VM_OBJECT_UNLOCK(dst_object);
1386
1387                 /*
1388                  * Enter it in the pmap. If a wired, copy-on-write
1389                  * mapping is being replaced by a write-enabled
1390                  * mapping, then wire that new mapping.
1391                  */
1392                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1393
1394                 /*
1395                  * Mark it no longer busy, and put it on the active list.
1396                  */
1397                 VM_OBJECT_LOCK(dst_object);
1398
1399                 if (upgrade) {
1400                         vm_page_lock(src_m);
1401                         vm_page_unwire(src_m, 0);
1402                         vm_page_unlock(src_m);
1403
1404                         vm_page_lock(dst_m);
1405                         vm_page_wire(dst_m);
1406                         vm_page_unlock(dst_m);
1407                 } else {
1408                         vm_page_lock(dst_m);
1409                         vm_page_activate(dst_m);
1410                         vm_page_unlock(dst_m);
1411                 }
1412                 vm_page_wakeup(dst_m);
1413         }
1414         VM_OBJECT_UNLOCK(dst_object);
1415         if (upgrade) {
1416                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1417                 vm_object_deallocate(src_object);
1418         }
1419 }
1420
1421
1422 /*
1423  * This routine checks around the requested page for other pages that
1424  * might be able to be faulted in.  This routine brackets the viable
1425  * pages for the pages to be paged in.
1426  *
1427  * Inputs:
1428  *      m, rbehind, rahead
1429  *
1430  * Outputs:
1431  *  marray (array of vm_page_t), reqpage (index of requested page)
1432  *
1433  * Return value:
1434  *  number of pages in marray
1435  */
1436 static int
1437 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1438         vm_page_t m;
1439         int rbehind;
1440         int rahead;
1441         vm_page_t *marray;
1442         int *reqpage;
1443 {
1444         int i,j;
1445         vm_object_t object;
1446         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1447         vm_page_t rtm;
1448         int cbehind, cahead;
1449
1450         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1451
1452         object = m->object;
1453         pindex = m->pindex;
1454         cbehind = cahead = 0;
1455
1456         /*
1457          * if the requested page is not available, then give up now
1458          */
1459         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1460                 return 0;
1461         }
1462
1463         if ((cbehind == 0) && (cahead == 0)) {
1464                 *reqpage = 0;
1465                 marray[0] = m;
1466                 return 1;
1467         }
1468
1469         if (rahead > cahead) {
1470                 rahead = cahead;
1471         }
1472
1473         if (rbehind > cbehind) {
1474                 rbehind = cbehind;
1475         }
1476
1477         /*
1478          * scan backward for the read behind pages -- in memory
1479          */
1480         if (pindex > 0) {
1481                 if (rbehind > pindex) {
1482                         rbehind = pindex;
1483                         startpindex = 0;
1484                 } else {
1485                         startpindex = pindex - rbehind;
1486                 }
1487
1488                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1489                     rtm->pindex >= startpindex)
1490                         startpindex = rtm->pindex + 1;
1491
1492                 /* tpindex is unsigned; beware of numeric underflow. */
1493                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1494                     tpindex < pindex; i++, tpindex--) {
1495
1496                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1497                             VM_ALLOC_IFNOTCACHED);
1498                         if (rtm == NULL) {
1499                                 /*
1500                                  * Shift the allocated pages to the
1501                                  * beginning of the array.
1502                                  */
1503                                 for (j = 0; j < i; j++) {
1504                                         marray[j] = marray[j + tpindex + 1 -
1505                                             startpindex];
1506                                 }
1507                                 break;
1508                         }
1509
1510                         marray[tpindex - startpindex] = rtm;
1511                 }
1512         } else {
1513                 startpindex = 0;
1514                 i = 0;
1515         }
1516
1517         marray[i] = m;
1518         /* page offset of the required page */
1519         *reqpage = i;
1520
1521         tpindex = pindex + 1;
1522         i++;
1523
1524         /*
1525          * scan forward for the read ahead pages
1526          */
1527         endpindex = tpindex + rahead;
1528         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1529                 endpindex = rtm->pindex;
1530         if (endpindex > object->size)
1531                 endpindex = object->size;
1532
1533         for (; tpindex < endpindex; i++, tpindex++) {
1534
1535                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1536                     VM_ALLOC_IFNOTCACHED);
1537                 if (rtm == NULL) {
1538                         break;
1539                 }
1540
1541                 marray[i] = rtm;
1542         }
1543
1544         /* return number of pages */
1545         return i;
1546 }
1547
1548 /*
1549  * Block entry into the machine-independent layer's page fault handler by
1550  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1551  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1552  * spurious page faults.
1553  */
1554 int
1555 vm_fault_disable_pagefaults(void)
1556 {
1557
1558         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1559 }
1560
1561 void
1562 vm_fault_enable_pagefaults(int save)
1563 {
1564
1565         curthread_pflags_restore(save);
1566 }