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
 108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 109
 110 #define VM_FAULT_READ_BEHIND    8
 111 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 112 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 113 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 114 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 115
 116 struct faultstate {
 117         vm_page_t m;
 118         vm_object_t object;
 119         vm_pindex_t pindex;
 120         vm_page_t first_m;
 121         vm_object_t     first_object;
 122         vm_pindex_t first_pindex;
 123         vm_map_t map;
 124         vm_map_entry_t entry;
 125         int lookup_still_valid;
 126         struct vnode *vp;
 127 };
 128
 129 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 130 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
 131             int faultcount, int reqpage);
 132
 133 static inline void
 134 release_page(struct faultstate *fs)
 135 {
 136
 137         vm_page_xunbusy(fs->m);
 138         vm_page_lock(fs->m);
 139         vm_page_deactivate(fs->m);
 140         vm_page_unlock(fs->m);
 141         fs->m = NULL;
 142 }
 143
 144 static inline void
 145 unlock_map(struct faultstate *fs)
 146 {
 147
 148         if (fs->lookup_still_valid) {
 149                 vm_map_lookup_done(fs->map, fs->entry);
 150                 fs->lookup_still_valid = FALSE;
 151         }
 152 }
 153
 154 static void
 155 unlock_and_deallocate(struct faultstate *fs)
 156 {
 157
 158         vm_object_pip_wakeup(fs->object);
 159         VM_OBJECT_WUNLOCK(fs->object);
 160         if (fs->object != fs->first_object) {
 161                 VM_OBJECT_WLOCK(fs->first_object);
 162                 vm_page_lock(fs->first_m);
 163                 vm_page_free(fs->first_m);
 164                 vm_page_unlock(fs->first_m);
 165                 vm_object_pip_wakeup(fs->first_object);
 166                 VM_OBJECT_WUNLOCK(fs->first_object);
 167                 fs->first_m = NULL;
 168         }
 169         vm_object_deallocate(fs->first_object);
 170         unlock_map(fs);
 171         if (fs->vp != NULL) {
 172                 vput(fs->vp);
 173                 fs->vp = NULL;
 174         }
 175 }
 176
 177 /*
 178  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 179  *            current object *might* contain the page.
 180  *
 181  *            default objects are zero-fill, there is no real pager.
 182  */
 183 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 184                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 185
 186 /*
 187  *      vm_fault:
 188  *
 189  *      Handle a page fault occurring at the given address,
 190  *      requiring the given permissions, in the map specified.
 191  *      If successful, the page is inserted into the
 192  *      associated physical map.
 193  *
 194  *      NOTE: the given address should be truncated to the
 195  *      proper page address.
 196  *
 197  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 198  *      a standard error specifying why the fault is fatal is returned.
 199  *
 200  *      The map in question must be referenced, and remains so.
 201  *      Caller may hold no locks.
 202  */
 203 int
 204 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 205     int fault_flags)
 206 {
 207         struct thread *td;
 208         int result;
 209
 210         td = curthread;
 211         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 212                 return (KERN_PROTECTION_FAILURE);
 213 #ifdef KTRACE
 214         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 215                 ktrfault(vaddr, fault_type);
 216 #endif
 217         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 218             NULL);
 219 #ifdef KTRACE
 220         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 221                 ktrfaultend(result);
 222 #endif
 223         return (result);
 224 }
 225
 226 int
 227 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 228     int fault_flags, vm_page_t *m_hold)
 229 {
 230         vm_prot_t prot;
 231         long ahead, behind;
 232         int alloc_req, era, faultcount, nera, reqpage, result;
 233         boolean_t growstack, is_first_object_locked, wired;
 234         int map_generation;
 235         vm_object_t next_object;
 236         vm_page_t marray[VM_FAULT_READ_MAX];
 237         int hardfault;
 238         struct faultstate fs;
 239         struct vnode *vp;
 240         vm_page_t m;
 241         int locked, error;
 242
 243         hardfault = 0;
 244         growstack = TRUE;
 245         PCPU_INC(cnt.v_vm_faults);
 246         fs.vp = NULL;
 247         faultcount = reqpage = 0;
 248
 249 RetryFault:;
 250
 251         /*
 252          * Find the backing store object and offset into it to begin the
 253          * search.
 254          */
 255         fs.map = map;
 256         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 257             &fs.first_object, &fs.first_pindex, &prot, &wired);
 258         if (result != KERN_SUCCESS) {
 259                 if (growstack && result == KERN_INVALID_ADDRESS &&
 260                     map != kernel_map) {
 261                         result = vm_map_growstack(curproc, vaddr);
 262                         if (result != KERN_SUCCESS)
 263                                 return (KERN_FAILURE);
 264                         growstack = FALSE;
 265                         goto RetryFault;
 266                 }
 267                 return (result);
 268         }
 269
 270         map_generation = fs.map->timestamp;
 271
 272         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 273                 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
 274                         vm_map_unlock_read(fs.map);
 275                         return (KERN_FAILURE);
 276                 }
 277                 panic("vm_fault: fault on nofault entry, addr: %lx",
 278                     (u_long)vaddr);
 279         }
 280
 281         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 282             fs.entry->wiring_thread != curthread) {
 283                 vm_map_unlock_read(fs.map);
 284                 vm_map_lock(fs.map);
 285                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 286                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 287                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 288                         vm_map_unlock_and_wait(fs.map, 0);
 289                 } else
 290                         vm_map_unlock(fs.map);
 291                 goto RetryFault;
 292         }
 293
 294         if (wired)
 295                 fault_type = prot | (fault_type & VM_PROT_COPY);
 296
 297         if (fs.vp == NULL /* avoid locked vnode leak */ &&
 298             (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
 299             /* avoid calling vm_object_set_writeable_dirty() */
 300             ((prot & VM_PROT_WRITE) == 0 ||
 301             fs.first_object->type != OBJT_VNODE ||
 302             (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
 303                 VM_OBJECT_RLOCK(fs.first_object);
 304                 if ((prot & VM_PROT_WRITE) != 0 &&
 305                     fs.first_object->type == OBJT_VNODE &&
 306                     (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
 307                         goto fast_failed;
 308                 m = vm_page_lookup(fs.first_object, fs.first_pindex);
 309                 if (m == NULL || vm_page_busied(m) ||
 310                     m->valid != VM_PAGE_BITS_ALL)
 311                         goto fast_failed;
 312                 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
 313                    fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
 314                    0), 0);
 315                 if (result != KERN_SUCCESS)
 316                         goto fast_failed;
 317                 if (m_hold != NULL) {
 318                         *m_hold = m;
 319                         vm_page_lock(m);
 320                         vm_page_hold(m);
 321                         vm_page_unlock(m);
 322                 }
 323                 if ((fault_type & VM_PROT_WRITE) != 0 &&
 324                     (m->oflags & VPO_UNMANAGED) == 0) {
 325                         vm_page_dirty(m);
 326                         vm_pager_page_unswapped(m);
 327                 }
 328                 VM_OBJECT_RUNLOCK(fs.first_object);
 329                 if (!wired)
 330                         vm_fault_prefault(&fs, vaddr, 0, 0);
 331                 vm_map_lookup_done(fs.map, fs.entry);
 332                 curthread->td_ru.ru_minflt++;
 333                 return (KERN_SUCCESS);
 334 fast_failed:
 335                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 336                         VM_OBJECT_RUNLOCK(fs.first_object);
 337                         VM_OBJECT_WLOCK(fs.first_object);
 338                 }
 339         } else {
 340                 VM_OBJECT_WLOCK(fs.first_object);
 341         }
 342
 343         /*
 344          * Make a reference to this object to prevent its disposal while we
 345          * are messing with it.  Once we have the reference, the map is free
 346          * to be diddled.  Since objects reference their shadows (and copies),
 347          * they will stay around as well.
 348          *
 349          * Bump the paging-in-progress count to prevent size changes (e.g.
 350          * truncation operations) during I/O.  This must be done after
 351          * obtaining the vnode lock in order to avoid possible deadlocks.
 352          */
 353         vm_object_reference_locked(fs.first_object);
 354         vm_object_pip_add(fs.first_object, 1);
 355
 356         fs.lookup_still_valid = TRUE;
 357
 358         fs.first_m = NULL;
 359
 360         /*
 361          * Search for the page at object/offset.
 362          */
 363         fs.object = fs.first_object;
 364         fs.pindex = fs.first_pindex;
 365         while (TRUE) {
 366                 /*
 367                  * If the object is dead, we stop here
 368                  */
 369                 if (fs.object->flags & OBJ_DEAD) {
 370                         unlock_and_deallocate(&fs);
 371                         return (KERN_PROTECTION_FAILURE);
 372                 }
 373
 374                 /*
 375                  * See if page is resident
 376                  */
 377                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 378                 if (fs.m != NULL) {
 379                         /*
 380                          * Wait/Retry if the page is busy.  We have to do this
 381                          * if the page is either exclusive or shared busy
 382                          * because the vm_pager may be using read busy for
 383                          * pageouts (and even pageins if it is the vnode
 384                          * pager), and we could end up trying to pagein and
 385                          * pageout the same page simultaneously.
 386                          *
 387                          * We can theoretically allow the busy case on a read
 388                          * fault if the page is marked valid, but since such
 389                          * pages are typically already pmap'd, putting that
 390                          * special case in might be more effort then it is
 391                          * worth.  We cannot under any circumstances mess
 392                          * around with a shared busied page except, perhaps,
 393                          * to pmap it.
 394                          */
 395                         if (vm_page_busied(fs.m)) {
 396                                 /*
 397                                  * Reference the page before unlocking and
 398                                  * sleeping so that the page daemon is less
 399                                  * likely to reclaim it.
 400                                  */
 401                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 402                                 if (fs.object != fs.first_object) {
 403                                         if (!VM_OBJECT_TRYWLOCK(
 404                                             fs.first_object)) {
 405                                                 VM_OBJECT_WUNLOCK(fs.object);
 406                                                 VM_OBJECT_WLOCK(fs.first_object);
 407                                                 VM_OBJECT_WLOCK(fs.object);
 408                                         }
 409                                         vm_page_lock(fs.first_m);
 410                                         vm_page_free(fs.first_m);
 411                                         vm_page_unlock(fs.first_m);
 412                                         vm_object_pip_wakeup(fs.first_object);
 413                                         VM_OBJECT_WUNLOCK(fs.first_object);
 414                                         fs.first_m = NULL;
 415                                 }
 416                                 unlock_map(&fs);
 417                                 if (fs.m == vm_page_lookup(fs.object,
 418                                     fs.pindex)) {
 419                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 420                                 }
 421                                 vm_object_pip_wakeup(fs.object);
 422                                 VM_OBJECT_WUNLOCK(fs.object);
 423                                 PCPU_INC(cnt.v_intrans);
 424                                 vm_object_deallocate(fs.first_object);
 425                                 goto RetryFault;
 426                         }
 427                         vm_page_lock(fs.m);
 428                         vm_page_remque(fs.m);
 429                         vm_page_unlock(fs.m);
 430
 431                         /*
 432                          * Mark page busy for other processes, and the
 433                          * pagedaemon.  If it still isn't completely valid
 434                          * (readable), jump to readrest, else break-out ( we
 435                          * found the page ).
 436                          */
 437                         vm_page_xbusy(fs.m);
 438                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 439                                 goto readrest;
 440                         break;
 441                 }
 442
 443                 /*
 444                  * Page is not resident, If this is the search termination
 445                  * or the pager might contain the page, allocate a new page.
 446                  */
 447                 if (TRYPAGER || fs.object == fs.first_object) {
 448                         if (fs.pindex >= fs.object->size) {
 449                                 unlock_and_deallocate(&fs);
 450                                 return (KERN_PROTECTION_FAILURE);
 451                         }
 452
 453                         /*
 454                          * Allocate a new page for this object/offset pair.
 455                          *
 456                          * Unlocked read of the p_flag is harmless. At
 457                          * worst, the P_KILLED might be not observed
 458                          * there, and allocation can fail, causing
 459                          * restart and new reading of the p_flag.
 460                          */
 461                         fs.m = NULL;
 462                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 463 #if VM_NRESERVLEVEL > 0
 464                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 465                                         fs.object->flags |= OBJ_COLORED;
 466                                         fs.object->pg_color = atop(vaddr) -
 467                                             fs.pindex;
 468                                 }
 469 #endif
 470                                 alloc_req = P_KILLED(curproc) ?
 471                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 472                                 if (fs.object->type != OBJT_VNODE &&
 473                                     fs.object->backing_object == NULL)
 474                                         alloc_req |= VM_ALLOC_ZERO;
 475                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 476                                     alloc_req);
 477                         }
 478                         if (fs.m == NULL) {
 479                                 unlock_and_deallocate(&fs);
 480                                 VM_WAITPFAULT;
 481                                 goto RetryFault;
 482                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 483                                 break;
 484                 }
 485
 486 readrest:
 487                 /*
 488                  * We have found a valid page or we have allocated a new page.
 489                  * The page thus may not be valid or may not be entirely
 490                  * valid.
 491                  *
 492                  * Attempt to fault-in the page if there is a chance that the
 493                  * pager has it, and potentially fault in additional pages
 494                  * at the same time.
 495                  */
 496                 if (TRYPAGER) {
 497                         int rv;
 498                         u_char behavior = vm_map_entry_behavior(fs.entry);
 499
 500                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 501                             P_KILLED(curproc)) {
 502                                 behind = 0;
 503                                 ahead = 0;
 504                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 505                                 behind = 0;
 506                                 ahead = atop(fs.entry->end - vaddr) - 1;
 507                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 508                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 509                                 if (fs.pindex == fs.entry->next_read)
 510                                         vm_fault_cache_behind(&fs,
 511                                             VM_FAULT_READ_MAX);
 512                         } else {
 513                                 /*
 514                                  * If this is a sequential page fault, then
 515                                  * arithmetically increase the number of pages
 516                                  * in the read-ahead window.  Otherwise, reset
 517                                  * the read-ahead window to its smallest size.
 518                                  */
 519                                 behind = atop(vaddr - fs.entry->start);
 520                                 if (behind > VM_FAULT_READ_BEHIND)
 521                                         behind = VM_FAULT_READ_BEHIND;
 522                                 ahead = atop(fs.entry->end - vaddr) - 1;
 523                                 era = fs.entry->read_ahead;
 524                                 if (fs.pindex == fs.entry->next_read) {
 525                                         nera = era + behind;
 526                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 527                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 528                                         behind = 0;
 529                                         if (ahead > nera)
 530                                                 ahead = nera;
 531                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 532                                                 vm_fault_cache_behind(&fs,
 533                                                     VM_FAULT_CACHE_BEHIND);
 534                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 535                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 536                                 if (era != ahead)
 537                                         fs.entry->read_ahead = ahead;
 538                         }
 539
 540                         /*
 541                          * Call the pager to retrieve the data, if any, after
 542                          * releasing the lock on the map.  We hold a ref on
 543                          * fs.object and the pages are exclusive busied.
 544                          */
 545                         unlock_map(&fs);
 546
 547                         if (fs.object->type == OBJT_VNODE) {
 548                                 vp = fs.object->handle;
 549                                 if (vp == fs.vp)
 550                                         goto vnode_locked;
 551                                 else if (fs.vp != NULL) {
 552                                         vput(fs.vp);
 553                                         fs.vp = NULL;
 554                                 }
 555                                 locked = VOP_ISLOCKED(vp);
 556
 557                                 if (locked != LK_EXCLUSIVE)
 558                                         locked = LK_SHARED;
 559                                 /* Do not sleep for vnode lock while fs.m is busy */
 560                                 error = vget(vp, locked | LK_CANRECURSE |
 561                                     LK_NOWAIT, curthread);
 562                                 if (error != 0) {
 563                                         vhold(vp);
 564                                         release_page(&fs);
 565                                         unlock_and_deallocate(&fs);
 566                                         error = vget(vp, locked | LK_RETRY |
 567                                             LK_CANRECURSE, curthread);
 568                                         vdrop(vp);
 569                                         fs.vp = vp;
 570                                         KASSERT(error == 0,
 571                                             ("vm_fault: vget failed"));
 572                                         goto RetryFault;
 573                                 }
 574                                 fs.vp = vp;
 575                         }
 576 vnode_locked:
 577                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 578                             ("vm_fault: vnode-backed object mapped by system map"));
 579
 580                         /*
 581                          * now we find out if any other pages should be paged
 582                          * in at this time this routine checks to see if the
 583                          * pages surrounding this fault reside in the same
 584                          * object as the page for this fault.  If they do,
 585                          * then they are faulted in also into the object.  The
 586                          * array "marray" returned contains an array of
 587                          * vm_page_t structs where one of them is the
 588                          * vm_page_t passed to the routine.  The reqpage
 589                          * return value is the index into the marray for the
 590                          * vm_page_t passed to the routine.
 591                          *
 592                          * fs.m plus the additional pages are exclusive busied.
 593                          */
 594                         faultcount = vm_fault_additional_pages(
 595                             fs.m, behind, ahead, marray, &reqpage);
 596
 597                         rv = faultcount ?
 598                             vm_pager_get_pages(fs.object, marray, faultcount,
 599                                 reqpage) : VM_PAGER_FAIL;
 600
 601                         if (rv == VM_PAGER_OK) {
 602                                 /*
 603                                  * Found the page. Leave it busy while we play
 604                                  * with it.
 605                                  */
 606
 607                                 /*
 608                                  * Relookup in case pager changed page. Pager
 609                                  * is responsible for disposition of old page
 610                                  * if moved.
 611                                  */
 612                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 613                                 if (!fs.m) {
 614                                         unlock_and_deallocate(&fs);
 615                                         goto RetryFault;
 616                                 }
 617
 618                                 hardfault++;
 619                                 break; /* break to PAGE HAS BEEN FOUND */
 620                         }
 621                         /*
 622                          * Remove the bogus page (which does not exist at this
 623                          * object/offset); before doing so, we must get back
 624                          * our object lock to preserve our invariant.
 625                          *
 626                          * Also wake up any other process that may want to bring
 627                          * in this page.
 628                          *
 629                          * If this is the top-level object, we must leave the
 630                          * busy page to prevent another process from rushing
 631                          * past us, and inserting the page in that object at
 632                          * the same time that we are.
 633                          */
 634                         if (rv == VM_PAGER_ERROR)
 635                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 636                                     curproc->p_pid, curproc->p_comm);
 637                         /*
 638                          * Data outside the range of the pager or an I/O error
 639                          */
 640                         /*
 641                          * XXX - the check for kernel_map is a kludge to work
 642                          * around having the machine panic on a kernel space
 643                          * fault w/ I/O error.
 644                          */
 645                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 646                                 (rv == VM_PAGER_BAD)) {
 647                                 vm_page_lock(fs.m);
 648                                 vm_page_free(fs.m);
 649                                 vm_page_unlock(fs.m);
 650                                 fs.m = NULL;
 651                                 unlock_and_deallocate(&fs);
 652                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 653                         }
 654                         if (fs.object != fs.first_object) {
 655                                 vm_page_lock(fs.m);
 656                                 vm_page_free(fs.m);
 657                                 vm_page_unlock(fs.m);
 658                                 fs.m = NULL;
 659                                 /*
 660                                  * XXX - we cannot just fall out at this
 661                                  * point, m has been freed and is invalid!
 662                                  */
 663                         }
 664                 }
 665
 666                 /*
 667                  * We get here if the object has default pager (or unwiring)
 668                  * or the pager doesn't have the page.
 669                  */
 670                 if (fs.object == fs.first_object)
 671                         fs.first_m = fs.m;
 672
 673                 /*
 674                  * Move on to the next object.  Lock the next object before
 675                  * unlocking the current one.
 676                  */
 677                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 678                 next_object = fs.object->backing_object;
 679                 if (next_object == NULL) {
 680                         /*
 681                          * If there's no object left, fill the page in the top
 682                          * object with zeros.
 683                          */
 684                         if (fs.object != fs.first_object) {
 685                                 vm_object_pip_wakeup(fs.object);
 686                                 VM_OBJECT_WUNLOCK(fs.object);
 687
 688                                 fs.object = fs.first_object;
 689                                 fs.pindex = fs.first_pindex;
 690                                 fs.m = fs.first_m;
 691                                 VM_OBJECT_WLOCK(fs.object);
 692                         }
 693                         fs.first_m = NULL;
 694
 695                         /*
 696                          * Zero the page if necessary and mark it valid.
 697                          */
 698                         if ((fs.m->flags & PG_ZERO) == 0) {
 699                                 pmap_zero_page(fs.m);
 700                         } else {
 701                                 PCPU_INC(cnt.v_ozfod);
 702                         }
 703                         PCPU_INC(cnt.v_zfod);
 704                         fs.m->valid = VM_PAGE_BITS_ALL;
 705                         /* Don't try to prefault neighboring pages. */
 706                         faultcount = 1;
 707                         break;  /* break to PAGE HAS BEEN FOUND */
 708                 } else {
 709                         KASSERT(fs.object != next_object,
 710                             ("object loop %p", next_object));
 711                         VM_OBJECT_WLOCK(next_object);
 712                         vm_object_pip_add(next_object, 1);
 713                         if (fs.object != fs.first_object)
 714                                 vm_object_pip_wakeup(fs.object);
 715                         VM_OBJECT_WUNLOCK(fs.object);
 716                         fs.object = next_object;
 717                 }
 718         }
 719
 720         vm_page_assert_xbusied(fs.m);
 721
 722         /*
 723          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 724          * is held.]
 725          */
 726
 727         /*
 728          * If the page is being written, but isn't already owned by the
 729          * top-level object, we have to copy it into a new page owned by the
 730          * top-level object.
 731          */
 732         if (fs.object != fs.first_object) {
 733                 /*
 734                  * We only really need to copy if we want to write it.
 735                  */
 736                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 737                         /*
 738                          * This allows pages to be virtually copied from a
 739                          * backing_object into the first_object, where the
 740                          * backing object has no other refs to it, and cannot
 741                          * gain any more refs.  Instead of a bcopy, we just
 742                          * move the page from the backing object to the
 743                          * first object.  Note that we must mark the page
 744                          * dirty in the first object so that it will go out
 745                          * to swap when needed.
 746                          */
 747                         is_first_object_locked = FALSE;
 748                         if (
 749                                 /*
 750                                  * Only one shadow object
 751                                  */
 752                                 (fs.object->shadow_count == 1) &&
 753                                 /*
 754                                  * No COW refs, except us
 755                                  */
 756                                 (fs.object->ref_count == 1) &&
 757                                 /*
 758                                  * No one else can look this object up
 759                                  */
 760                                 (fs.object->handle == NULL) &&
 761                                 /*
 762                                  * No other ways to look the object up
 763                                  */
 764                                 ((fs.object->type == OBJT_DEFAULT) ||
 765                                  (fs.object->type == OBJT_SWAP)) &&
 766                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 767                                 /*
 768                                  * We don't chase down the shadow chain
 769                                  */
 770                             fs.object == fs.first_object->backing_object) {
 771                                 /*
 772                                  * get rid of the unnecessary page
 773                                  */
 774                                 vm_page_lock(fs.first_m);
 775                                 vm_page_free(fs.first_m);
 776                                 vm_page_unlock(fs.first_m);
 777                                 /*
 778                                  * grab the page and put it into the
 779                                  * process'es object.  The page is
 780                                  * automatically made dirty.
 781                                  */
 782                                 if (vm_page_rename(fs.m, fs.first_object,
 783                                     fs.first_pindex)) {
 784                                         unlock_and_deallocate(&fs);
 785                                         goto RetryFault;
 786                                 }
 787                                 vm_page_xbusy(fs.m);
 788                                 fs.first_m = fs.m;
 789                                 fs.m = NULL;
 790                                 PCPU_INC(cnt.v_cow_optim);
 791                         } else {
 792                                 /*
 793                                  * Oh, well, lets copy it.
 794                                  */
 795                                 pmap_copy_page(fs.m, fs.first_m);
 796                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 797                                 if (wired && (fault_flags &
 798                                     VM_FAULT_CHANGE_WIRING) == 0) {
 799                                         vm_page_lock(fs.first_m);
 800                                         vm_page_wire(fs.first_m);
 801                                         vm_page_unlock(fs.first_m);
 802
 803                                         vm_page_lock(fs.m);
 804                                         vm_page_unwire(fs.m, PQ_INACTIVE);
 805                                         vm_page_unlock(fs.m);
 806                                 }
 807                                 /*
 808                                  * We no longer need the old page or object.
 809                                  */
 810                                 release_page(&fs);
 811                         }
 812                         /*
 813                          * fs.object != fs.first_object due to above
 814                          * conditional
 815                          */
 816                         vm_object_pip_wakeup(fs.object);
 817                         VM_OBJECT_WUNLOCK(fs.object);
 818                         /*
 819                          * Only use the new page below...
 820                          */
 821                         fs.object = fs.first_object;
 822                         fs.pindex = fs.first_pindex;
 823                         fs.m = fs.first_m;
 824                         if (!is_first_object_locked)
 825                                 VM_OBJECT_WLOCK(fs.object);
 826                         PCPU_INC(cnt.v_cow_faults);
 827                         curthread->td_cow++;
 828                 } else {
 829                         prot &= ~VM_PROT_WRITE;
 830                 }
 831         }
 832
 833         /*
 834          * We must verify that the maps have not changed since our last
 835          * lookup.
 836          */
 837         if (!fs.lookup_still_valid) {
 838                 vm_object_t retry_object;
 839                 vm_pindex_t retry_pindex;
 840                 vm_prot_t retry_prot;
 841
 842                 if (!vm_map_trylock_read(fs.map)) {
 843                         release_page(&fs);
 844                         unlock_and_deallocate(&fs);
 845                         goto RetryFault;
 846                 }
 847                 fs.lookup_still_valid = TRUE;
 848                 if (fs.map->timestamp != map_generation) {
 849                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 850                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 851
 852                         /*
 853                          * If we don't need the page any longer, put it on the inactive
 854                          * list (the easiest thing to do here).  If no one needs it,
 855                          * pageout will grab it eventually.
 856                          */
 857                         if (result != KERN_SUCCESS) {
 858                                 release_page(&fs);
 859                                 unlock_and_deallocate(&fs);
 860
 861                                 /*
 862                                  * If retry of map lookup would have blocked then
 863                                  * retry fault from start.
 864                                  */
 865                                 if (result == KERN_FAILURE)
 866                                         goto RetryFault;
 867                                 return (result);
 868                         }
 869                         if ((retry_object != fs.first_object) ||
 870                             (retry_pindex != fs.first_pindex)) {
 871                                 release_page(&fs);
 872                                 unlock_and_deallocate(&fs);
 873                                 goto RetryFault;
 874                         }
 875
 876                         /*
 877                          * Check whether the protection has changed or the object has
 878                          * been copied while we left the map unlocked. Changing from
 879                          * read to write permission is OK - we leave the page
 880                          * write-protected, and catch the write fault. Changing from
 881                          * write to read permission means that we can't mark the page
 882                          * write-enabled after all.
 883                          */
 884                         prot &= retry_prot;
 885                 }
 886         }
 887         /*
 888          * If the page was filled by a pager, update the map entry's
 889          * last read offset.  Since the pager does not return the
 890          * actual set of pages that it read, this update is based on
 891          * the requested set.  Typically, the requested and actual
 892          * sets are the same.
 893          *
 894          * XXX The following assignment modifies the map
 895          * without holding a write lock on it.
 896          */
 897         if (hardfault)
 898                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 899
 900         if (((prot & VM_PROT_WRITE) != 0 ||
 901             (fault_flags & VM_FAULT_DIRTY) != 0) &&
 902             (fs.m->oflags & VPO_UNMANAGED) == 0) {
 903                 vm_object_set_writeable_dirty(fs.object);
 904
 905                 /*
 906                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 907                  * if the page is already dirty to prevent data written with
 908                  * the expectation of being synced from not being synced.
 909                  * Likewise if this entry does not request NOSYNC then make
 910                  * sure the page isn't marked NOSYNC.  Applications sharing
 911                  * data should use the same flags to avoid ping ponging.
 912                  */
 913                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 914                         if (fs.m->dirty == 0)
 915                                 fs.m->oflags |= VPO_NOSYNC;
 916                 } else {
 917                         fs.m->oflags &= ~VPO_NOSYNC;
 918                 }
 919
 920                 /*
 921                  * If the fault is a write, we know that this page is being
 922                  * written NOW so dirty it explicitly to save on
 923                  * pmap_is_modified() calls later.
 924                  *
 925                  * Also tell the backing pager, if any, that it should remove
 926                  * any swap backing since the page is now dirty.
 927                  */
 928                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 929                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 930                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 931                         vm_page_dirty(fs.m);
 932                         vm_pager_page_unswapped(fs.m);
 933                 }
 934         }
 935
 936         vm_page_assert_xbusied(fs.m);
 937
 938         /*
 939          * Page must be completely valid or it is not fit to
 940          * map into user space.  vm_pager_get_pages() ensures this.
 941          */
 942         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 943             ("vm_fault: page %p partially invalid", fs.m));
 944         VM_OBJECT_WUNLOCK(fs.object);
 945
 946         /*
 947          * Put this page into the physical map.  We had to do the unlock above
 948          * because pmap_enter() may sleep.  We don't put the page
 949          * back on the active queue until later so that the pageout daemon
 950          * won't find it (yet).
 951          */
 952         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
 953             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
 954         if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
 955             wired == 0)
 956                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
 957         VM_OBJECT_WLOCK(fs.object);
 958         vm_page_lock(fs.m);
 959
 960         /*
 961          * If the page is not wired down, then put it where the pageout daemon
 962          * can find it.
 963          */
 964         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 965                 if (wired)
 966                         vm_page_wire(fs.m);
 967                 else
 968                         vm_page_unwire(fs.m, PQ_ACTIVE);
 969         } else
 970                 vm_page_activate(fs.m);
 971         if (m_hold != NULL) {
 972                 *m_hold = fs.m;
 973                 vm_page_hold(fs.m);
 974         }
 975         vm_page_unlock(fs.m);
 976         vm_page_xunbusy(fs.m);
 977
 978         /*
 979          * Unlock everything, and return
 980          */
 981         unlock_and_deallocate(&fs);
 982         if (hardfault) {
 983                 PCPU_INC(cnt.v_io_faults);
 984                 curthread->td_ru.ru_majflt++;
 985         } else
 986                 curthread->td_ru.ru_minflt++;
 987
 988         return (KERN_SUCCESS);
 989 }
 990
 991 /*
 992  * Speed up the reclamation of up to "distance" pages that precede the
 993  * faulting pindex within the first object of the shadow chain.
 994  */
 995 static void
 996 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 997 {
 998         vm_object_t first_object, object;
 999         vm_page_t m, m_prev;
1000         vm_pindex_t pindex;
1001
1002         object = fs->object;
1003         VM_OBJECT_ASSERT_WLOCKED(object);
1004         first_object = fs->first_object;
1005         if (first_object != object) {
1006                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1007                         VM_OBJECT_WUNLOCK(object);
1008                         VM_OBJECT_WLOCK(first_object);
1009                         VM_OBJECT_WLOCK(object);
1010                 }
1011         }
1012         /* Neither fictitious nor unmanaged pages can be cached. */
1013         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1014                 if (fs->first_pindex < distance)
1015                         pindex = 0;
1016                 else
1017                         pindex = fs->first_pindex - distance;
1018                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1019                         pindex = OFF_TO_IDX(fs->entry->offset);
1020                 m = first_object != object ? fs->first_m : fs->m;
1021                 vm_page_assert_xbusied(m);
1022                 m_prev = vm_page_prev(m);
1023                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1024                     m->valid == VM_PAGE_BITS_ALL) {
1025                         m_prev = vm_page_prev(m);
1026                         if (vm_page_busied(m))
1027                                 continue;
1028                         vm_page_lock(m);
1029                         if (m->hold_count == 0 && m->wire_count == 0) {
1030                                 pmap_remove_all(m);
1031                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1032                                 if (m->dirty != 0)
1033                                         vm_page_deactivate(m);
1034                                 else
1035                                         vm_page_cache(m);
1036                         }
1037                         vm_page_unlock(m);
1038                 }
1039         }
1040         if (first_object != object)
1041                 VM_OBJECT_WUNLOCK(first_object);
1042 }
1043
1044 /*
1045  * vm_fault_prefault provides a quick way of clustering
1046  * pagefaults into a processes address space.  It is a "cousin"
1047  * of vm_map_pmap_enter, except it runs at page fault time instead
1048  * of mmap time.
1049  */
1050 static void
1051 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1052     int faultcount, int reqpage)
1053 {
1054         pmap_t pmap;
1055         vm_map_entry_t entry;
1056         vm_object_t backing_object, lobject;
1057         vm_offset_t addr, starta;
1058         vm_pindex_t pindex;
1059         vm_page_t m;
1060         int backward, forward, i;
1061
1062         pmap = fs->map->pmap;
1063         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1064                 return;
1065
1066         if (faultcount > 0) {
1067                 backward = reqpage;
1068                 forward = faultcount - reqpage - 1;
1069         } else {
1070                 backward = PFBAK;
1071                 forward = PFFOR;
1072         }
1073         entry = fs->entry;
1074
1075         starta = addra - backward * PAGE_SIZE;
1076         if (starta < entry->start) {
1077                 starta = entry->start;
1078         } else if (starta > addra) {
1079                 starta = 0;
1080         }
1081
1082         /*
1083          * Generate the sequence of virtual addresses that are candidates for
1084          * prefaulting in an outward spiral from the faulting virtual address,
1085          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1086          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1087          * If the candidate address doesn't have a backing physical page, then
1088          * the loop immediately terminates.
1089          */
1090         for (i = 0; i < 2 * imax(backward, forward); i++) {
1091                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1092                     PAGE_SIZE);
1093                 if (addr > addra + forward * PAGE_SIZE)
1094                         addr = 0;
1095
1096                 if (addr < starta || addr >= entry->end)
1097                         continue;
1098
1099                 if (!pmap_is_prefaultable(pmap, addr))
1100                         continue;
1101
1102                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1103                 lobject = entry->object.vm_object;
1104                 VM_OBJECT_RLOCK(lobject);
1105                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1106                     lobject->type == OBJT_DEFAULT &&
1107                     (backing_object = lobject->backing_object) != NULL) {
1108                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1109                             0, ("vm_fault_prefault: unaligned object offset"));
1110                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1111                         VM_OBJECT_RLOCK(backing_object);
1112                         VM_OBJECT_RUNLOCK(lobject);
1113                         lobject = backing_object;
1114                 }
1115                 if (m == NULL) {
1116                         VM_OBJECT_RUNLOCK(lobject);
1117                         break;
1118                 }
1119                 if (m->valid == VM_PAGE_BITS_ALL &&
1120                     (m->flags & PG_FICTITIOUS) == 0)
1121                         pmap_enter_quick(pmap, addr, m, entry->protection);
1122                 VM_OBJECT_RUNLOCK(lobject);
1123         }
1124 }
1125
1126 /*
1127  * Hold each of the physical pages that are mapped by the specified range of
1128  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1129  * and allow the specified types of access, "prot".  If all of the implied
1130  * pages are successfully held, then the number of held pages is returned
1131  * together with pointers to those pages in the array "ma".  However, if any
1132  * of the pages cannot be held, -1 is returned.
1133  */
1134 int
1135 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1136     vm_prot_t prot, vm_page_t *ma, int max_count)
1137 {
1138         vm_offset_t end, va;
1139         vm_page_t *mp;
1140         int count;
1141         boolean_t pmap_failed;
1142
1143         if (len == 0)
1144                 return (0);
1145         end = round_page(addr + len);
1146         addr = trunc_page(addr);
1147
1148         /*
1149          * Check for illegal addresses.
1150          */
1151         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1152                 return (-1);
1153
1154         if (atop(end - addr) > max_count)
1155                 panic("vm_fault_quick_hold_pages: count > max_count");
1156         count = atop(end - addr);
1157
1158         /*
1159          * Most likely, the physical pages are resident in the pmap, so it is
1160          * faster to try pmap_extract_and_hold() first.
1161          */
1162         pmap_failed = FALSE;
1163         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1164                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1165                 if (*mp == NULL)
1166                         pmap_failed = TRUE;
1167                 else if ((prot & VM_PROT_WRITE) != 0 &&
1168                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1169                         /*
1170                          * Explicitly dirty the physical page.  Otherwise, the
1171                          * caller's changes may go unnoticed because they are
1172                          * performed through an unmanaged mapping or by a DMA
1173                          * operation.
1174                          *
1175                          * The object lock is not held here.
1176                          * See vm_page_clear_dirty_mask().
1177                          */
1178                         vm_page_dirty(*mp);
1179                 }
1180         }
1181         if (pmap_failed) {
1182                 /*
1183                  * One or more pages could not be held by the pmap.  Either no
1184                  * page was mapped at the specified virtual address or that
1185                  * mapping had insufficient permissions.  Attempt to fault in
1186                  * and hold these pages.
1187                  */
1188                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1189                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1190                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1191                                 goto error;
1192         }
1193         return (count);
1194 error:
1195         for (mp = ma; mp < ma + count; mp++)
1196                 if (*mp != NULL) {
1197                         vm_page_lock(*mp);
1198                         vm_page_unhold(*mp);
1199                         vm_page_unlock(*mp);
1200                 }
1201         return (-1);
1202 }
1203
1204 /*
1205  *      Routine:
1206  *              vm_fault_copy_entry
1207  *      Function:
1208  *              Create new shadow object backing dst_entry with private copy of
1209  *              all underlying pages. When src_entry is equal to dst_entry,
1210  *              function implements COW for wired-down map entry. Otherwise,
1211  *              it forks wired entry into dst_map.
1212  *
1213  *      In/out conditions:
1214  *              The source and destination maps must be locked for write.
1215  *              The source map entry must be wired down (or be a sharing map
1216  *              entry corresponding to a main map entry that is wired down).
1217  */
1218 void
1219 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1220     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1221     vm_ooffset_t *fork_charge)
1222 {
1223         vm_object_t backing_object, dst_object, object, src_object;
1224         vm_pindex_t dst_pindex, pindex, src_pindex;
1225         vm_prot_t access, prot;
1226         vm_offset_t vaddr;
1227         vm_page_t dst_m;
1228         vm_page_t src_m;
1229         boolean_t upgrade;
1230
1231 #ifdef  lint
1232         src_map++;
1233 #endif  /* lint */
1234
1235         upgrade = src_entry == dst_entry;
1236         access = prot = dst_entry->protection;
1237
1238         src_object = src_entry->object.vm_object;
1239         src_pindex = OFF_TO_IDX(src_entry->offset);
1240
1241         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1242                 dst_object = src_object;
1243                 vm_object_reference(dst_object);
1244         } else {
1245                 /*
1246                  * Create the top-level object for the destination entry. (Doesn't
1247                  * actually shadow anything - we copy the pages directly.)
1248                  */
1249                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1250                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1251 #if VM_NRESERVLEVEL > 0
1252                 dst_object->flags |= OBJ_COLORED;
1253                 dst_object->pg_color = atop(dst_entry->start);
1254 #endif
1255         }
1256
1257         VM_OBJECT_WLOCK(dst_object);
1258         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1259             ("vm_fault_copy_entry: vm_object not NULL"));
1260         if (src_object != dst_object) {
1261                 dst_entry->object.vm_object = dst_object;
1262                 dst_entry->offset = 0;
1263                 dst_object->charge = dst_entry->end - dst_entry->start;
1264         }
1265         if (fork_charge != NULL) {
1266                 KASSERT(dst_entry->cred == NULL,
1267                     ("vm_fault_copy_entry: leaked swp charge"));
1268                 dst_object->cred = curthread->td_ucred;
1269                 crhold(dst_object->cred);
1270                 *fork_charge += dst_object->charge;
1271         } else if (dst_object->cred == NULL) {
1272                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1273                     dst_entry));
1274                 dst_object->cred = dst_entry->cred;
1275                 dst_entry->cred = NULL;
1276         }
1277
1278         /*
1279          * If not an upgrade, then enter the mappings in the pmap as
1280          * read and/or execute accesses.  Otherwise, enter them as
1281          * write accesses.
1282          *
1283          * A writeable large page mapping is only created if all of
1284          * the constituent small page mappings are modified. Marking
1285          * PTEs as modified on inception allows promotion to happen
1286          * without taking potentially large number of soft faults.
1287          */
1288         if (!upgrade)
1289                 access &= ~VM_PROT_WRITE;
1290
1291         /*
1292          * Loop through all of the virtual pages within the entry's
1293          * range, copying each page from the source object to the
1294          * destination object.  Since the source is wired, those pages
1295          * must exist.  In contrast, the destination is pageable.
1296          * Since the destination object does share any backing storage
1297          * with the source object, all of its pages must be dirtied,
1298          * regardless of whether they can be written.
1299          */
1300         for (vaddr = dst_entry->start, dst_pindex = 0;
1301             vaddr < dst_entry->end;
1302             vaddr += PAGE_SIZE, dst_pindex++) {
1303 again:
1304                 /*
1305                  * Find the page in the source object, and copy it in.
1306                  * Because the source is wired down, the page will be
1307                  * in memory.
1308                  */
1309                 if (src_object != dst_object)
1310                         VM_OBJECT_RLOCK(src_object);
1311                 object = src_object;
1312                 pindex = src_pindex + dst_pindex;
1313                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1314                     (backing_object = object->backing_object) != NULL) {
1315                         /*
1316                          * Unless the source mapping is read-only or
1317                          * it is presently being upgraded from
1318                          * read-only, the first object in the shadow
1319                          * chain should provide all of the pages.  In
1320                          * other words, this loop body should never be
1321                          * executed when the source mapping is already
1322                          * read/write.
1323                          */
1324                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1325                             upgrade,
1326                             ("vm_fault_copy_entry: main object missing page"));
1327
1328                         VM_OBJECT_RLOCK(backing_object);
1329                         pindex += OFF_TO_IDX(object->backing_object_offset);
1330                         if (object != dst_object)
1331                                 VM_OBJECT_RUNLOCK(object);
1332                         object = backing_object;
1333                 }
1334                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1335
1336                 if (object != dst_object) {
1337                         /*
1338                          * Allocate a page in the destination object.
1339                          */
1340                         dst_m = vm_page_alloc(dst_object, (src_object ==
1341                             dst_object ? src_pindex : 0) + dst_pindex,
1342                             VM_ALLOC_NORMAL);
1343                         if (dst_m == NULL) {
1344                                 VM_OBJECT_WUNLOCK(dst_object);
1345                                 VM_OBJECT_RUNLOCK(object);
1346                                 VM_WAIT;
1347                                 VM_OBJECT_WLOCK(dst_object);
1348                                 goto again;
1349                         }
1350                         pmap_copy_page(src_m, dst_m);
1351                         VM_OBJECT_RUNLOCK(object);
1352                         dst_m->valid = VM_PAGE_BITS_ALL;
1353                         dst_m->dirty = VM_PAGE_BITS_ALL;
1354                 } else {
1355                         dst_m = src_m;
1356                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1357                                 goto again;
1358                         vm_page_xbusy(dst_m);
1359                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1360                             ("invalid dst page %p", dst_m));
1361                 }
1362                 VM_OBJECT_WUNLOCK(dst_object);
1363
1364                 /*
1365                  * Enter it in the pmap. If a wired, copy-on-write
1366                  * mapping is being replaced by a write-enabled
1367                  * mapping, then wire that new mapping.
1368                  */
1369                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1370                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1371
1372                 /*
1373                  * Mark it no longer busy, and put it on the active list.
1374                  */
1375                 VM_OBJECT_WLOCK(dst_object);
1376
1377                 if (upgrade) {
1378                         if (src_m != dst_m) {
1379                                 vm_page_lock(src_m);
1380                                 vm_page_unwire(src_m, PQ_INACTIVE);
1381                                 vm_page_unlock(src_m);
1382                                 vm_page_lock(dst_m);
1383                                 vm_page_wire(dst_m);
1384                                 vm_page_unlock(dst_m);
1385                         } else {
1386                                 KASSERT(dst_m->wire_count > 0,
1387                                     ("dst_m %p is not wired", dst_m));
1388                         }
1389                 } else {
1390                         vm_page_lock(dst_m);
1391                         vm_page_activate(dst_m);
1392                         vm_page_unlock(dst_m);
1393                 }
1394                 vm_page_xunbusy(dst_m);
1395         }
1396         VM_OBJECT_WUNLOCK(dst_object);
1397         if (upgrade) {
1398                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1399                 vm_object_deallocate(src_object);
1400         }
1401 }
1402
1403
1404 /*
1405  * This routine checks around the requested page for other pages that
1406  * might be able to be faulted in.  This routine brackets the viable
1407  * pages for the pages to be paged in.
1408  *
1409  * Inputs:
1410  *      m, rbehind, rahead
1411  *
1412  * Outputs:
1413  *  marray (array of vm_page_t), reqpage (index of requested page)
1414  *
1415  * Return value:
1416  *  number of pages in marray
1417  */
1418 static int
1419 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1420         vm_page_t m;
1421         int rbehind;
1422         int rahead;
1423         vm_page_t *marray;
1424         int *reqpage;
1425 {
1426         int i,j;
1427         vm_object_t object;
1428         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1429         vm_page_t rtm;
1430         int cbehind, cahead;
1431
1432         VM_OBJECT_ASSERT_WLOCKED(m->object);
1433
1434         object = m->object;
1435         pindex = m->pindex;
1436         cbehind = cahead = 0;
1437
1438         /*
1439          * if the requested page is not available, then give up now
1440          */
1441         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1442                 return 0;
1443         }
1444
1445         if ((cbehind == 0) && (cahead == 0)) {
1446                 *reqpage = 0;
1447                 marray[0] = m;
1448                 return 1;
1449         }
1450
1451         if (rahead > cahead) {
1452                 rahead = cahead;
1453         }
1454
1455         if (rbehind > cbehind) {
1456                 rbehind = cbehind;
1457         }
1458
1459         /*
1460          * scan backward for the read behind pages -- in memory
1461          */
1462         if (pindex > 0) {
1463                 if (rbehind > pindex) {
1464                         rbehind = pindex;
1465                         startpindex = 0;
1466                 } else {
1467                         startpindex = pindex - rbehind;
1468                 }
1469
1470                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1471                     rtm->pindex >= startpindex)
1472                         startpindex = rtm->pindex + 1;
1473
1474                 /* tpindex is unsigned; beware of numeric underflow. */
1475                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1476                     tpindex < pindex; i++, tpindex--) {
1477
1478                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1479                             VM_ALLOC_IFNOTCACHED);
1480                         if (rtm == NULL) {
1481                                 /*
1482                                  * Shift the allocated pages to the
1483                                  * beginning of the array.
1484                                  */
1485                                 for (j = 0; j < i; j++) {
1486                                         marray[j] = marray[j + tpindex + 1 -
1487                                             startpindex];
1488                                 }
1489                                 break;
1490                         }
1491
1492                         marray[tpindex - startpindex] = rtm;
1493                 }
1494         } else {
1495                 startpindex = 0;
1496                 i = 0;
1497         }
1498
1499         marray[i] = m;
1500         /* page offset of the required page */
1501         *reqpage = i;
1502
1503         tpindex = pindex + 1;
1504         i++;
1505
1506         /*
1507          * scan forward for the read ahead pages
1508          */
1509         endpindex = tpindex + rahead;
1510         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1511                 endpindex = rtm->pindex;
1512         if (endpindex > object->size)
1513                 endpindex = object->size;
1514
1515         for (; tpindex < endpindex; i++, tpindex++) {
1516
1517                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1518                     VM_ALLOC_IFNOTCACHED);
1519                 if (rtm == NULL) {
1520                         break;
1521                 }
1522
1523                 marray[i] = rtm;
1524         }
1525
1526         /* return number of pages */
1527         return i;
1528 }
1529
1530 /*
1531  * Block entry into the machine-independent layer's page fault handler by
1532  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1533  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1534  * spurious page faults.
1535  */
1536 int
1537 vm_fault_disable_pagefaults(void)
1538 {
1539
1540         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1541 }
1542
1543 void
1544 vm_fault_enable_pagefaults(int save)
1545 {
1546
1547         curthread_pflags_restore(save);
1548 }