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