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