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                 /* A busy page can be mapped for read|execute access. */
 310                 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
 311                     vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
 312                         goto fast_failed;
 313                 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
 314                    fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
 315                    0), 0);
 316                 if (result != KERN_SUCCESS)
 317                         goto fast_failed;
 318                 if (m_hold != NULL) {
 319                         *m_hold = m;
 320                         vm_page_lock(m);
 321                         vm_page_hold(m);
 322                         vm_page_unlock(m);
 323                 }
 324                 if ((fault_type & VM_PROT_WRITE) != 0 &&
 325                     (m->oflags & VPO_UNMANAGED) == 0) {
 326                         vm_page_dirty(m);
 327                         vm_pager_page_unswapped(m);
 328                 }
 329                 VM_OBJECT_RUNLOCK(fs.first_object);
 330                 if (!wired)
 331                         vm_fault_prefault(&fs, vaddr, 0, 0);
 332                 vm_map_lookup_done(fs.map, fs.entry);
 333                 curthread->td_ru.ru_minflt++;
 334                 return (KERN_SUCCESS);
 335 fast_failed:
 336                 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
 337                         VM_OBJECT_RUNLOCK(fs.first_object);
 338                         VM_OBJECT_WLOCK(fs.first_object);
 339                 }
 340         } else {
 341                 VM_OBJECT_WLOCK(fs.first_object);
 342         }
 343
 344         /*
 345          * Make a reference to this object to prevent its disposal while we
 346          * are messing with it.  Once we have the reference, the map is free
 347          * to be diddled.  Since objects reference their shadows (and copies),
 348          * they will stay around as well.
 349          *
 350          * Bump the paging-in-progress count to prevent size changes (e.g.
 351          * truncation operations) during I/O.  This must be done after
 352          * obtaining the vnode lock in order to avoid possible deadlocks.
 353          */
 354         vm_object_reference_locked(fs.first_object);
 355         vm_object_pip_add(fs.first_object, 1);
 356
 357         fs.lookup_still_valid = TRUE;
 358
 359         fs.first_m = NULL;
 360
 361         /*
 362          * Search for the page at object/offset.
 363          */
 364         fs.object = fs.first_object;
 365         fs.pindex = fs.first_pindex;
 366         while (TRUE) {
 367                 /*
 368                  * If the object is dead, we stop here
 369                  */
 370                 if (fs.object->flags & OBJ_DEAD) {
 371                         unlock_and_deallocate(&fs);
 372                         return (KERN_PROTECTION_FAILURE);
 373                 }
 374
 375                 /*
 376                  * See if page is resident
 377                  */
 378                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 379                 if (fs.m != NULL) {
 380                         /*
 381                          * Wait/Retry if the page is busy.  We have to do this
 382                          * if the page is either exclusive or shared busy
 383                          * because the vm_pager may be using read busy for
 384                          * pageouts (and even pageins if it is the vnode
 385                          * pager), and we could end up trying to pagein and
 386                          * pageout the same page simultaneously.
 387                          *
 388                          * We can theoretically allow the busy case on a read
 389                          * fault if the page is marked valid, but since such
 390                          * pages are typically already pmap'd, putting that
 391                          * special case in might be more effort then it is
 392                          * worth.  We cannot under any circumstances mess
 393                          * around with a shared busied page except, perhaps,
 394                          * to pmap it.
 395                          */
 396                         if (vm_page_busied(fs.m)) {
 397                                 /*
 398                                  * Reference the page before unlocking and
 399                                  * sleeping so that the page daemon is less
 400                                  * likely to reclaim it.
 401                                  */
 402                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 403                                 if (fs.object != fs.first_object) {
 404                                         if (!VM_OBJECT_TRYWLOCK(
 405                                             fs.first_object)) {
 406                                                 VM_OBJECT_WUNLOCK(fs.object);
 407                                                 VM_OBJECT_WLOCK(fs.first_object);
 408                                                 VM_OBJECT_WLOCK(fs.object);
 409                                         }
 410                                         vm_page_lock(fs.first_m);
 411                                         vm_page_free(fs.first_m);
 412                                         vm_page_unlock(fs.first_m);
 413                                         vm_object_pip_wakeup(fs.first_object);
 414                                         VM_OBJECT_WUNLOCK(fs.first_object);
 415                                         fs.first_m = NULL;
 416                                 }
 417                                 unlock_map(&fs);
 418                                 if (fs.m == vm_page_lookup(fs.object,
 419                                     fs.pindex)) {
 420                                         vm_page_sleep_if_busy(fs.m, "vmpfw");
 421                                 }
 422                                 vm_object_pip_wakeup(fs.object);
 423                                 VM_OBJECT_WUNLOCK(fs.object);
 424                                 PCPU_INC(cnt.v_intrans);
 425                                 vm_object_deallocate(fs.first_object);
 426                                 goto RetryFault;
 427                         }
 428                         vm_page_lock(fs.m);
 429                         vm_page_remque(fs.m);
 430                         vm_page_unlock(fs.m);
 431
 432                         /*
 433                          * Mark page busy for other processes, and the
 434                          * pagedaemon.  If it still isn't completely valid
 435                          * (readable), jump to readrest, else break-out ( we
 436                          * found the page ).
 437                          */
 438                         vm_page_xbusy(fs.m);
 439                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 440                                 goto readrest;
 441                         break;
 442                 }
 443
 444                 /*
 445                  * Page is not resident, If this is the search termination
 446                  * or the pager might contain the page, allocate a new page.
 447                  */
 448                 if (TRYPAGER || fs.object == fs.first_object) {
 449                         if (fs.pindex >= fs.object->size) {
 450                                 unlock_and_deallocate(&fs);
 451                                 return (KERN_PROTECTION_FAILURE);
 452                         }
 453
 454                         /*
 455                          * Allocate a new page for this object/offset pair.
 456                          *
 457                          * Unlocked read of the p_flag is harmless. At
 458                          * worst, the P_KILLED might be not observed
 459                          * there, and allocation can fail, causing
 460                          * restart and new reading of the p_flag.
 461                          */
 462                         fs.m = NULL;
 463                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 464 #if VM_NRESERVLEVEL > 0
 465                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 466                                         fs.object->flags |= OBJ_COLORED;
 467                                         fs.object->pg_color = atop(vaddr) -
 468                                             fs.pindex;
 469                                 }
 470 #endif
 471                                 alloc_req = P_KILLED(curproc) ?
 472                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 473                                 if (fs.object->type != OBJT_VNODE &&
 474                                     fs.object->backing_object == NULL)
 475                                         alloc_req |= VM_ALLOC_ZERO;
 476                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 477                                     alloc_req);
 478                         }
 479                         if (fs.m == NULL) {
 480                                 unlock_and_deallocate(&fs);
 481                                 VM_WAITPFAULT;
 482                                 goto RetryFault;
 483                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 484                                 break;
 485                 }
 486
 487 readrest:
 488                 /*
 489                  * We have found a valid page or we have allocated a new page.
 490                  * The page thus may not be valid or may not be entirely
 491                  * valid.
 492                  *
 493                  * Attempt to fault-in the page if there is a chance that the
 494                  * pager has it, and potentially fault in additional pages
 495                  * at the same time.
 496                  */
 497                 if (TRYPAGER) {
 498                         int rv;
 499                         u_char behavior = vm_map_entry_behavior(fs.entry);
 500
 501                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 502                             P_KILLED(curproc)) {
 503                                 behind = 0;
 504                                 ahead = 0;
 505                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 506                                 behind = 0;
 507                                 ahead = atop(fs.entry->end - vaddr) - 1;
 508                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 509                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 510                                 if (fs.pindex == fs.entry->next_read)
 511                                         vm_fault_cache_behind(&fs,
 512                                             VM_FAULT_READ_MAX);
 513                         } else {
 514                                 /*
 515                                  * If this is a sequential page fault, then
 516                                  * arithmetically increase the number of pages
 517                                  * in the read-ahead window.  Otherwise, reset
 518                                  * the read-ahead window to its smallest size.
 519                                  */
 520                                 behind = atop(vaddr - fs.entry->start);
 521                                 if (behind > VM_FAULT_READ_BEHIND)
 522                                         behind = VM_FAULT_READ_BEHIND;
 523                                 ahead = atop(fs.entry->end - vaddr) - 1;
 524                                 era = fs.entry->read_ahead;
 525                                 if (fs.pindex == fs.entry->next_read) {
 526                                         nera = era + behind;
 527                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 528                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 529                                         behind = 0;
 530                                         if (ahead > nera)
 531                                                 ahead = nera;
 532                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 533                                                 vm_fault_cache_behind(&fs,
 534                                                     VM_FAULT_CACHE_BEHIND);
 535                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 536                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 537                                 if (era != ahead)
 538                                         fs.entry->read_ahead = ahead;
 539                         }
 540
 541                         /*
 542                          * Call the pager to retrieve the data, if any, after
 543                          * releasing the lock on the map.  We hold a ref on
 544                          * fs.object and the pages are exclusive busied.
 545                          */
 546                         unlock_map(&fs);
 547
 548                         if (fs.object->type == OBJT_VNODE) {
 549                                 vp = fs.object->handle;
 550                                 if (vp == fs.vp)
 551                                         goto vnode_locked;
 552                                 else if (fs.vp != NULL) {
 553                                         vput(fs.vp);
 554                                         fs.vp = NULL;
 555                                 }
 556                                 locked = VOP_ISLOCKED(vp);
 557
 558                                 if (locked != LK_EXCLUSIVE)
 559                                         locked = LK_SHARED;
 560                                 /* Do not sleep for vnode lock while fs.m is busy */
 561                                 error = vget(vp, locked | LK_CANRECURSE |
 562                                     LK_NOWAIT, curthread);
 563                                 if (error != 0) {
 564                                         vhold(vp);
 565                                         release_page(&fs);
 566                                         unlock_and_deallocate(&fs);
 567                                         error = vget(vp, locked | LK_RETRY |
 568                                             LK_CANRECURSE, curthread);
 569                                         vdrop(vp);
 570                                         fs.vp = vp;
 571                                         KASSERT(error == 0,
 572                                             ("vm_fault: vget failed"));
 573                                         goto RetryFault;
 574                                 }
 575                                 fs.vp = vp;
 576                         }
 577 vnode_locked:
 578                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 579                             ("vm_fault: vnode-backed object mapped by system map"));
 580
 581                         /*
 582                          * now we find out if any other pages should be paged
 583                          * in at this time this routine checks to see if the
 584                          * pages surrounding this fault reside in the same
 585                          * object as the page for this fault.  If they do,
 586                          * then they are faulted in also into the object.  The
 587                          * array "marray" returned contains an array of
 588                          * vm_page_t structs where one of them is the
 589                          * vm_page_t passed to the routine.  The reqpage
 590                          * return value is the index into the marray for the
 591                          * vm_page_t passed to the routine.
 592                          *
 593                          * fs.m plus the additional pages are exclusive busied.
 594                          */
 595                         faultcount = vm_fault_additional_pages(
 596                             fs.m, behind, ahead, marray, &reqpage);
 597
 598                         rv = faultcount ?
 599                             vm_pager_get_pages(fs.object, marray, faultcount,
 600                                 reqpage) : VM_PAGER_FAIL;
 601
 602                         if (rv == VM_PAGER_OK) {
 603                                 /*
 604                                  * Found the page. Leave it busy while we play
 605                                  * with it.
 606                                  */
 607
 608                                 /*
 609                                  * Relookup in case pager changed page. Pager
 610                                  * is responsible for disposition of old page
 611                                  * if moved.
 612                                  */
 613                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 614                                 if (!fs.m) {
 615                                         unlock_and_deallocate(&fs);
 616                                         goto RetryFault;
 617                                 }
 618
 619                                 hardfault++;
 620                                 break; /* break to PAGE HAS BEEN FOUND */
 621                         }
 622                         /*
 623                          * Remove the bogus page (which does not exist at this
 624                          * object/offset); before doing so, we must get back
 625                          * our object lock to preserve our invariant.
 626                          *
 627                          * Also wake up any other process that may want to bring
 628                          * in this page.
 629                          *
 630                          * If this is the top-level object, we must leave the
 631                          * busy page to prevent another process from rushing
 632                          * past us, and inserting the page in that object at
 633                          * the same time that we are.
 634                          */
 635                         if (rv == VM_PAGER_ERROR)
 636                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 637                                     curproc->p_pid, curproc->p_comm);
 638                         /*
 639                          * Data outside the range of the pager or an I/O error
 640                          */
 641                         /*
 642                          * XXX - the check for kernel_map is a kludge to work
 643                          * around having the machine panic on a kernel space
 644                          * fault w/ I/O error.
 645                          */
 646                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 647                                 (rv == VM_PAGER_BAD)) {
 648                                 vm_page_lock(fs.m);
 649                                 vm_page_free(fs.m);
 650                                 vm_page_unlock(fs.m);
 651                                 fs.m = NULL;
 652                                 unlock_and_deallocate(&fs);
 653                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 654                         }
 655                         if (fs.object != fs.first_object) {
 656                                 vm_page_lock(fs.m);
 657                                 vm_page_free(fs.m);
 658                                 vm_page_unlock(fs.m);
 659                                 fs.m = NULL;
 660                                 /*
 661                                  * XXX - we cannot just fall out at this
 662                                  * point, m has been freed and is invalid!
 663                                  */
 664                         }
 665                 }
 666
 667                 /*
 668                  * We get here if the object has default pager (or unwiring)
 669                  * or the pager doesn't have the page.
 670                  */
 671                 if (fs.object == fs.first_object)
 672                         fs.first_m = fs.m;
 673
 674                 /*
 675                  * Move on to the next object.  Lock the next object before
 676                  * unlocking the current one.
 677                  */
 678                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 679                 next_object = fs.object->backing_object;
 680                 if (next_object == NULL) {
 681                         /*
 682                          * If there's no object left, fill the page in the top
 683                          * object with zeros.
 684                          */
 685                         if (fs.object != fs.first_object) {
 686                                 vm_object_pip_wakeup(fs.object);
 687                                 VM_OBJECT_WUNLOCK(fs.object);
 688
 689                                 fs.object = fs.first_object;
 690                                 fs.pindex = fs.first_pindex;
 691                                 fs.m = fs.first_m;
 692                                 VM_OBJECT_WLOCK(fs.object);
 693                         }
 694                         fs.first_m = NULL;
 695
 696                         /*
 697                          * Zero the page if necessary and mark it valid.
 698                          */
 699                         if ((fs.m->flags & PG_ZERO) == 0) {
 700                                 pmap_zero_page(fs.m);
 701                         } else {
 702                                 PCPU_INC(cnt.v_ozfod);
 703                         }
 704                         PCPU_INC(cnt.v_zfod);
 705                         fs.m->valid = VM_PAGE_BITS_ALL;
 706                         /* Don't try to prefault neighboring pages. */
 707                         faultcount = 1;
 708                         break;  /* break to PAGE HAS BEEN FOUND */
 709                 } else {
 710                         KASSERT(fs.object != next_object,
 711                             ("object loop %p", next_object));
 712                         VM_OBJECT_WLOCK(next_object);
 713                         vm_object_pip_add(next_object, 1);
 714                         if (fs.object != fs.first_object)
 715                                 vm_object_pip_wakeup(fs.object);
 716                         VM_OBJECT_WUNLOCK(fs.object);
 717                         fs.object = next_object;
 718                 }
 719         }
 720
 721         vm_page_assert_xbusied(fs.m);
 722
 723         /*
 724          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 725          * is held.]
 726          */
 727
 728         /*
 729          * If the page is being written, but isn't already owned by the
 730          * top-level object, we have to copy it into a new page owned by the
 731          * top-level object.
 732          */
 733         if (fs.object != fs.first_object) {
 734                 /*
 735                  * We only really need to copy if we want to write it.
 736                  */
 737                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 738                         /*
 739                          * This allows pages to be virtually copied from a
 740                          * backing_object into the first_object, where the
 741                          * backing object has no other refs to it, and cannot
 742                          * gain any more refs.  Instead of a bcopy, we just
 743                          * move the page from the backing object to the
 744                          * first object.  Note that we must mark the page
 745                          * dirty in the first object so that it will go out
 746                          * to swap when needed.
 747                          */
 748                         is_first_object_locked = FALSE;
 749                         if (
 750                                 /*
 751                                  * Only one shadow object
 752                                  */
 753                                 (fs.object->shadow_count == 1) &&
 754                                 /*
 755                                  * No COW refs, except us
 756                                  */
 757                                 (fs.object->ref_count == 1) &&
 758                                 /*
 759                                  * No one else can look this object up
 760                                  */
 761                                 (fs.object->handle == NULL) &&
 762                                 /*
 763                                  * No other ways to look the object up
 764                                  */
 765                                 ((fs.object->type == OBJT_DEFAULT) ||
 766                                  (fs.object->type == OBJT_SWAP)) &&
 767                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 768                                 /*
 769                                  * We don't chase down the shadow chain
 770                                  */
 771                             fs.object == fs.first_object->backing_object) {
 772                                 /*
 773                                  * get rid of the unnecessary page
 774                                  */
 775                                 vm_page_lock(fs.first_m);
 776                                 vm_page_free(fs.first_m);
 777                                 vm_page_unlock(fs.first_m);
 778                                 /*
 779                                  * grab the page and put it into the
 780                                  * process'es object.  The page is
 781                                  * automatically made dirty.
 782                                  */
 783                                 if (vm_page_rename(fs.m, fs.first_object,
 784                                     fs.first_pindex)) {
 785                                         unlock_and_deallocate(&fs);
 786                                         goto RetryFault;
 787                                 }
 788                                 vm_page_xbusy(fs.m);
 789                                 fs.first_m = fs.m;
 790                                 fs.m = NULL;
 791                                 PCPU_INC(cnt.v_cow_optim);
 792                         } else {
 793                                 /*
 794                                  * Oh, well, lets copy it.
 795                                  */
 796                                 pmap_copy_page(fs.m, fs.first_m);
 797                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 798                                 if (wired && (fault_flags &
 799                                     VM_FAULT_CHANGE_WIRING) == 0) {
 800                                         vm_page_lock(fs.first_m);
 801                                         vm_page_wire(fs.first_m);
 802                                         vm_page_unlock(fs.first_m);
 803
 804                                         vm_page_lock(fs.m);
 805                                         vm_page_unwire(fs.m, FALSE);
 806                                         vm_page_unlock(fs.m);
 807                                 }
 808                                 /*
 809                                  * We no longer need the old page or object.
 810                                  */
 811                                 release_page(&fs);
 812                         }
 813                         /*
 814                          * fs.object != fs.first_object due to above
 815                          * conditional
 816                          */
 817                         vm_object_pip_wakeup(fs.object);
 818                         VM_OBJECT_WUNLOCK(fs.object);
 819                         /*
 820                          * Only use the new page below...
 821                          */
 822                         fs.object = fs.first_object;
 823                         fs.pindex = fs.first_pindex;
 824                         fs.m = fs.first_m;
 825                         if (!is_first_object_locked)
 826                                 VM_OBJECT_WLOCK(fs.object);
 827                         PCPU_INC(cnt.v_cow_faults);
 828                         curthread->td_cow++;
 829                 } else {
 830                         prot &= ~VM_PROT_WRITE;
 831                 }
 832         }
 833
 834         /*
 835          * We must verify that the maps have not changed since our last
 836          * lookup.
 837          */
 838         if (!fs.lookup_still_valid) {
 839                 vm_object_t retry_object;
 840                 vm_pindex_t retry_pindex;
 841                 vm_prot_t retry_prot;
 842
 843                 if (!vm_map_trylock_read(fs.map)) {
 844                         release_page(&fs);
 845                         unlock_and_deallocate(&fs);
 846                         goto RetryFault;
 847                 }
 848                 fs.lookup_still_valid = TRUE;
 849                 if (fs.map->timestamp != map_generation) {
 850                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 851                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 852
 853                         /*
 854                          * If we don't need the page any longer, put it on the inactive
 855                          * list (the easiest thing to do here).  If no one needs it,
 856                          * pageout will grab it eventually.
 857                          */
 858                         if (result != KERN_SUCCESS) {
 859                                 release_page(&fs);
 860                                 unlock_and_deallocate(&fs);
 861
 862                                 /*
 863                                  * If retry of map lookup would have blocked then
 864                                  * retry fault from start.
 865                                  */
 866                                 if (result == KERN_FAILURE)
 867                                         goto RetryFault;
 868                                 return (result);
 869                         }
 870                         if ((retry_object != fs.first_object) ||
 871                             (retry_pindex != fs.first_pindex)) {
 872                                 release_page(&fs);
 873                                 unlock_and_deallocate(&fs);
 874                                 goto RetryFault;
 875                         }
 876
 877                         /*
 878                          * Check whether the protection has changed or the object has
 879                          * been copied while we left the map unlocked. Changing from
 880                          * read to write permission is OK - we leave the page
 881                          * write-protected, and catch the write fault. Changing from
 882                          * write to read permission means that we can't mark the page
 883                          * write-enabled after all.
 884                          */
 885                         prot &= retry_prot;
 886                 }
 887         }
 888         /*
 889          * If the page was filled by a pager, update the map entry's
 890          * last read offset.  Since the pager does not return the
 891          * actual set of pages that it read, this update is based on
 892          * the requested set.  Typically, the requested and actual
 893          * sets are the same.
 894          *
 895          * XXX The following assignment modifies the map
 896          * without holding a write lock on it.
 897          */
 898         if (hardfault)
 899                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 900
 901         if (((prot & VM_PROT_WRITE) != 0 ||
 902             (fault_flags & VM_FAULT_DIRTY) != 0) &&
 903             (fs.m->oflags & VPO_UNMANAGED) == 0) {
 904                 vm_object_set_writeable_dirty(fs.object);
 905
 906                 /*
 907                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 908                  * if the page is already dirty to prevent data written with
 909                  * the expectation of being synced from not being synced.
 910                  * Likewise if this entry does not request NOSYNC then make
 911                  * sure the page isn't marked NOSYNC.  Applications sharing
 912                  * data should use the same flags to avoid ping ponging.
 913                  */
 914                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 915                         if (fs.m->dirty == 0)
 916                                 fs.m->oflags |= VPO_NOSYNC;
 917                 } else {
 918                         fs.m->oflags &= ~VPO_NOSYNC;
 919                 }
 920
 921                 /*
 922                  * If the fault is a write, we know that this page is being
 923                  * written NOW so dirty it explicitly to save on
 924                  * pmap_is_modified() calls later.
 925                  *
 926                  * Also tell the backing pager, if any, that it should remove
 927                  * any swap backing since the page is now dirty.
 928                  */
 929                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 930                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 931                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 932                         vm_page_dirty(fs.m);
 933                         vm_pager_page_unswapped(fs.m);
 934                 }
 935         }
 936
 937         vm_page_assert_xbusied(fs.m);
 938
 939         /*
 940          * Page must be completely valid or it is not fit to
 941          * map into user space.  vm_pager_get_pages() ensures this.
 942          */
 943         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 944             ("vm_fault: page %p partially invalid", fs.m));
 945         VM_OBJECT_WUNLOCK(fs.object);
 946
 947         /*
 948          * Put this page into the physical map.  We had to do the unlock above
 949          * because pmap_enter() may sleep.  We don't put the page
 950          * back on the active queue until later so that the pageout daemon
 951          * won't find it (yet).
 952          */
 953         pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
 954             fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
 955         if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
 956             wired == 0)
 957                 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
 958         VM_OBJECT_WLOCK(fs.object);
 959         vm_page_lock(fs.m);
 960
 961         /*
 962          * If the page is not wired down, then put it where the pageout daemon
 963          * can find it.
 964          */
 965         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 966                 if (wired)
 967                         vm_page_wire(fs.m);
 968                 else
 969                         vm_page_unwire(fs.m, 1);
 970         } else
 971                 vm_page_activate(fs.m);
 972         if (m_hold != NULL) {
 973                 *m_hold = fs.m;
 974                 vm_page_hold(fs.m);
 975         }
 976         vm_page_unlock(fs.m);
 977         vm_page_xunbusy(fs.m);
 978
 979         /*
 980          * Unlock everything, and return
 981          */
 982         unlock_and_deallocate(&fs);
 983         if (hardfault) {
 984                 PCPU_INC(cnt.v_io_faults);
 985                 curthread->td_ru.ru_majflt++;
 986         } else
 987                 curthread->td_ru.ru_minflt++;
 988
 989         return (KERN_SUCCESS);
 990 }
 991
 992 /*
 993  * Speed up the reclamation of up to "distance" pages that precede the
 994  * faulting pindex within the first object of the shadow chain.
 995  */
 996 static void
 997 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 998 {
 999         vm_object_t first_object, object;
1000         vm_page_t m, m_prev;
1001         vm_pindex_t pindex;
1002
1003         object = fs->object;
1004         VM_OBJECT_ASSERT_WLOCKED(object);
1005         first_object = fs->first_object;
1006         if (first_object != object) {
1007                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1008                         VM_OBJECT_WUNLOCK(object);
1009                         VM_OBJECT_WLOCK(first_object);
1010                         VM_OBJECT_WLOCK(object);
1011                 }
1012         }
1013         /* Neither fictitious nor unmanaged pages can be cached. */
1014         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1015                 if (fs->first_pindex < distance)
1016                         pindex = 0;
1017                 else
1018                         pindex = fs->first_pindex - distance;
1019                 if (pindex < OFF_TO_IDX(fs->entry->offset))
1020                         pindex = OFF_TO_IDX(fs->entry->offset);
1021                 m = first_object != object ? fs->first_m : fs->m;
1022                 vm_page_assert_xbusied(m);
1023                 m_prev = vm_page_prev(m);
1024                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1025                     m->valid == VM_PAGE_BITS_ALL) {
1026                         m_prev = vm_page_prev(m);
1027                         if (vm_page_busied(m))
1028                                 continue;
1029                         vm_page_lock(m);
1030                         if (m->hold_count == 0 && m->wire_count == 0) {
1031                                 pmap_remove_all(m);
1032                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1033                                 if (m->dirty != 0)
1034                                         vm_page_deactivate(m);
1035                                 else
1036                                         vm_page_cache(m);
1037                         }
1038                         vm_page_unlock(m);
1039                 }
1040         }
1041         if (first_object != object)
1042                 VM_OBJECT_WUNLOCK(first_object);
1043 }
1044
1045 /*
1046  * vm_fault_prefault provides a quick way of clustering
1047  * pagefaults into a processes address space.  It is a "cousin"
1048  * of vm_map_pmap_enter, except it runs at page fault time instead
1049  * of mmap time.
1050  */
1051 static void
1052 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1053     int faultcount, int reqpage)
1054 {
1055         pmap_t pmap;
1056         vm_map_entry_t entry;
1057         vm_object_t backing_object, lobject;
1058         vm_offset_t addr, starta;
1059         vm_pindex_t pindex;
1060         vm_page_t m;
1061         int backward, forward, i;
1062
1063         pmap = fs->map->pmap;
1064         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1065                 return;
1066
1067         if (faultcount > 0) {
1068                 backward = reqpage;
1069                 forward = faultcount - reqpage - 1;
1070         } else {
1071                 backward = PFBAK;
1072                 forward = PFFOR;
1073         }
1074         entry = fs->entry;
1075
1076         starta = addra - backward * PAGE_SIZE;
1077         if (starta < entry->start) {
1078                 starta = entry->start;
1079         } else if (starta > addra) {
1080                 starta = 0;
1081         }
1082
1083         /*
1084          * Generate the sequence of virtual addresses that are candidates for
1085          * prefaulting in an outward spiral from the faulting virtual address,
1086          * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1087          * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1088          * If the candidate address doesn't have a backing physical page, then
1089          * the loop immediately terminates.
1090          */
1091         for (i = 0; i < 2 * imax(backward, forward); i++) {
1092                 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1093                     PAGE_SIZE);
1094                 if (addr > addra + forward * PAGE_SIZE)
1095                         addr = 0;
1096
1097                 if (addr < starta || addr >= entry->end)
1098                         continue;
1099
1100                 if (!pmap_is_prefaultable(pmap, addr))
1101                         continue;
1102
1103                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1104                 lobject = entry->object.vm_object;
1105                 VM_OBJECT_RLOCK(lobject);
1106                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1107                     lobject->type == OBJT_DEFAULT &&
1108                     (backing_object = lobject->backing_object) != NULL) {
1109                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1110                             0, ("vm_fault_prefault: unaligned object offset"));
1111                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1112                         VM_OBJECT_RLOCK(backing_object);
1113                         VM_OBJECT_RUNLOCK(lobject);
1114                         lobject = backing_object;
1115                 }
1116                 if (m == NULL) {
1117                         VM_OBJECT_RUNLOCK(lobject);
1118                         break;
1119                 }
1120                 if (m->valid == VM_PAGE_BITS_ALL &&
1121                     (m->flags & PG_FICTITIOUS) == 0)
1122                         pmap_enter_quick(pmap, addr, m, entry->protection);
1123                 VM_OBJECT_RUNLOCK(lobject);
1124         }
1125 }
1126
1127 /*
1128  * Hold each of the physical pages that are mapped by the specified range of
1129  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1130  * and allow the specified types of access, "prot".  If all of the implied
1131  * pages are successfully held, then the number of held pages is returned
1132  * together with pointers to those pages in the array "ma".  However, if any
1133  * of the pages cannot be held, -1 is returned.
1134  */
1135 int
1136 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1137     vm_prot_t prot, vm_page_t *ma, int max_count)
1138 {
1139         vm_offset_t end, va;
1140         vm_page_t *mp;
1141         int count;
1142         boolean_t pmap_failed;
1143
1144         if (len == 0)
1145                 return (0);
1146         end = round_page(addr + len);
1147         addr = trunc_page(addr);
1148
1149         /*
1150          * Check for illegal addresses.
1151          */
1152         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1153                 return (-1);
1154
1155         if (atop(end - addr) > max_count)
1156                 panic("vm_fault_quick_hold_pages: count > max_count");
1157         count = atop(end - addr);
1158
1159         /*
1160          * Most likely, the physical pages are resident in the pmap, so it is
1161          * faster to try pmap_extract_and_hold() first.
1162          */
1163         pmap_failed = FALSE;
1164         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1165                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1166                 if (*mp == NULL)
1167                         pmap_failed = TRUE;
1168                 else if ((prot & VM_PROT_WRITE) != 0 &&
1169                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1170                         /*
1171                          * Explicitly dirty the physical page.  Otherwise, the
1172                          * caller's changes may go unnoticed because they are
1173                          * performed through an unmanaged mapping or by a DMA
1174                          * operation.
1175                          *
1176                          * The object lock is not held here.
1177                          * See vm_page_clear_dirty_mask().
1178                          */
1179                         vm_page_dirty(*mp);
1180                 }
1181         }
1182         if (pmap_failed) {
1183                 /*
1184                  * One or more pages could not be held by the pmap.  Either no
1185                  * page was mapped at the specified virtual address or that
1186                  * mapping had insufficient permissions.  Attempt to fault in
1187                  * and hold these pages.
1188                  */
1189                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1190                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1191                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1192                                 goto error;
1193         }
1194         return (count);
1195 error:
1196         for (mp = ma; mp < ma + count; mp++)
1197                 if (*mp != NULL) {
1198                         vm_page_lock(*mp);
1199                         vm_page_unhold(*mp);
1200                         vm_page_unlock(*mp);
1201                 }
1202         return (-1);
1203 }
1204
1205 /*
1206  *      Routine:
1207  *              vm_fault_copy_entry
1208  *      Function:
1209  *              Create new shadow object backing dst_entry with private copy of
1210  *              all underlying pages. When src_entry is equal to dst_entry,
1211  *              function implements COW for wired-down map entry. Otherwise,
1212  *              it forks wired entry into dst_map.
1213  *
1214  *      In/out conditions:
1215  *              The source and destination maps must be locked for write.
1216  *              The source map entry must be wired down (or be a sharing map
1217  *              entry corresponding to a main map entry that is wired down).
1218  */
1219 void
1220 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1221     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1222     vm_ooffset_t *fork_charge)
1223 {
1224         vm_object_t backing_object, dst_object, object, src_object;
1225         vm_pindex_t dst_pindex, pindex, src_pindex;
1226         vm_prot_t access, prot;
1227         vm_offset_t vaddr;
1228         vm_page_t dst_m;
1229         vm_page_t src_m;
1230         boolean_t upgrade;
1231
1232 #ifdef  lint
1233         src_map++;
1234 #endif  /* lint */
1235
1236         upgrade = src_entry == dst_entry;
1237         access = prot = dst_entry->protection;
1238
1239         src_object = src_entry->object.vm_object;
1240         src_pindex = OFF_TO_IDX(src_entry->offset);
1241
1242         if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1243                 dst_object = src_object;
1244                 vm_object_reference(dst_object);
1245         } else {
1246                 /*
1247                  * Create the top-level object for the destination entry. (Doesn't
1248                  * actually shadow anything - we copy the pages directly.)
1249                  */
1250                 dst_object = vm_object_allocate(OBJT_DEFAULT,
1251                     OFF_TO_IDX(dst_entry->end - dst_entry->start));
1252 #if VM_NRESERVLEVEL > 0
1253                 dst_object->flags |= OBJ_COLORED;
1254                 dst_object->pg_color = atop(dst_entry->start);
1255 #endif
1256         }
1257
1258         VM_OBJECT_WLOCK(dst_object);
1259         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1260             ("vm_fault_copy_entry: vm_object not NULL"));
1261         if (src_object != dst_object) {
1262                 dst_entry->object.vm_object = dst_object;
1263                 dst_entry->offset = 0;
1264                 dst_object->charge = dst_entry->end - dst_entry->start;
1265         }
1266         if (fork_charge != NULL) {
1267                 KASSERT(dst_entry->cred == NULL,
1268                     ("vm_fault_copy_entry: leaked swp charge"));
1269                 dst_object->cred = curthread->td_ucred;
1270                 crhold(dst_object->cred);
1271                 *fork_charge += dst_object->charge;
1272         } else if (dst_object->cred == NULL) {
1273                 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1274                     dst_entry));
1275                 dst_object->cred = dst_entry->cred;
1276                 dst_entry->cred = NULL;
1277         }
1278
1279         /*
1280          * If not an upgrade, then enter the mappings in the pmap as
1281          * read and/or execute accesses.  Otherwise, enter them as
1282          * write accesses.
1283          *
1284          * A writeable large page mapping is only created if all of
1285          * the constituent small page mappings are modified. Marking
1286          * PTEs as modified on inception allows promotion to happen
1287          * without taking potentially large number of soft faults.
1288          */
1289         if (!upgrade)
1290                 access &= ~VM_PROT_WRITE;
1291
1292         /*
1293          * Loop through all of the virtual pages within the entry's
1294          * range, copying each page from the source object to the
1295          * destination object.  Since the source is wired, those pages
1296          * must exist.  In contrast, the destination is pageable.
1297          * Since the destination object does share any backing storage
1298          * with the source object, all of its pages must be dirtied,
1299          * regardless of whether they can be written.
1300          */
1301         for (vaddr = dst_entry->start, dst_pindex = 0;
1302             vaddr < dst_entry->end;
1303             vaddr += PAGE_SIZE, dst_pindex++) {
1304 again:
1305                 /*
1306                  * Find the page in the source object, and copy it in.
1307                  * Because the source is wired down, the page will be
1308                  * in memory.
1309                  */
1310                 if (src_object != dst_object)
1311                         VM_OBJECT_RLOCK(src_object);
1312                 object = src_object;
1313                 pindex = src_pindex + dst_pindex;
1314                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1315                     (backing_object = object->backing_object) != NULL) {
1316                         /*
1317                          * Unless the source mapping is read-only or
1318                          * it is presently being upgraded from
1319                          * read-only, the first object in the shadow
1320                          * chain should provide all of the pages.  In
1321                          * other words, this loop body should never be
1322                          * executed when the source mapping is already
1323                          * read/write.
1324                          */
1325                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1326                             upgrade,
1327                             ("vm_fault_copy_entry: main object missing page"));
1328
1329                         VM_OBJECT_RLOCK(backing_object);
1330                         pindex += OFF_TO_IDX(object->backing_object_offset);
1331                         if (object != dst_object)
1332                                 VM_OBJECT_RUNLOCK(object);
1333                         object = backing_object;
1334                 }
1335                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1336
1337                 if (object != dst_object) {
1338                         /*
1339                          * Allocate a page in the destination object.
1340                          */
1341                         dst_m = vm_page_alloc(dst_object, (src_object ==
1342                             dst_object ? src_pindex : 0) + dst_pindex,
1343                             VM_ALLOC_NORMAL);
1344                         if (dst_m == NULL) {
1345                                 VM_OBJECT_WUNLOCK(dst_object);
1346                                 VM_OBJECT_RUNLOCK(object);
1347                                 VM_WAIT;
1348                                 VM_OBJECT_WLOCK(dst_object);
1349                                 goto again;
1350                         }
1351                         pmap_copy_page(src_m, dst_m);
1352                         VM_OBJECT_RUNLOCK(object);
1353                         dst_m->valid = VM_PAGE_BITS_ALL;
1354                         dst_m->dirty = VM_PAGE_BITS_ALL;
1355                 } else {
1356                         dst_m = src_m;
1357                         if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1358                                 goto again;
1359                         vm_page_xbusy(dst_m);
1360                         KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1361                             ("invalid dst page %p", dst_m));
1362                 }
1363                 VM_OBJECT_WUNLOCK(dst_object);
1364
1365                 /*
1366                  * Enter it in the pmap. If a wired, copy-on-write
1367                  * mapping is being replaced by a write-enabled
1368                  * mapping, then wire that new mapping.
1369                  */
1370                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1371                     access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1372
1373                 /*
1374                  * Mark it no longer busy, and put it on the active list.
1375                  */
1376                 VM_OBJECT_WLOCK(dst_object);
1377
1378                 if (upgrade) {
1379                         if (src_m != dst_m) {
1380                                 vm_page_lock(src_m);
1381                                 vm_page_unwire(src_m, 0);
1382                                 vm_page_unlock(src_m);
1383                                 vm_page_lock(dst_m);
1384                                 vm_page_wire(dst_m);
1385                                 vm_page_unlock(dst_m);
1386                         } else {
1387                                 KASSERT(dst_m->wire_count > 0,
1388                                     ("dst_m %p is not wired", dst_m));
1389                         }
1390                 } else {
1391                         vm_page_lock(dst_m);
1392                         vm_page_activate(dst_m);
1393                         vm_page_unlock(dst_m);
1394                 }
1395                 vm_page_xunbusy(dst_m);
1396         }
1397         VM_OBJECT_WUNLOCK(dst_object);
1398         if (upgrade) {
1399                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1400                 vm_object_deallocate(src_object);
1401         }
1402 }
1403
1404
1405 /*
1406  * This routine checks around the requested page for other pages that
1407  * might be able to be faulted in.  This routine brackets the viable
1408  * pages for the pages to be paged in.
1409  *
1410  * Inputs:
1411  *      m, rbehind, rahead
1412  *
1413  * Outputs:
1414  *  marray (array of vm_page_t), reqpage (index of requested page)
1415  *
1416  * Return value:
1417  *  number of pages in marray
1418  */
1419 static int
1420 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1421         vm_page_t m;
1422         int rbehind;
1423         int rahead;
1424         vm_page_t *marray;
1425         int *reqpage;
1426 {
1427         int i,j;
1428         vm_object_t object;
1429         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1430         vm_page_t rtm;
1431         int cbehind, cahead;
1432
1433         VM_OBJECT_ASSERT_WLOCKED(m->object);
1434
1435         object = m->object;
1436         pindex = m->pindex;
1437         cbehind = cahead = 0;
1438
1439         /*
1440          * if the requested page is not available, then give up now
1441          */
1442         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1443                 return 0;
1444         }
1445
1446         if ((cbehind == 0) && (cahead == 0)) {
1447                 *reqpage = 0;
1448                 marray[0] = m;
1449                 return 1;
1450         }
1451
1452         if (rahead > cahead) {
1453                 rahead = cahead;
1454         }
1455
1456         if (rbehind > cbehind) {
1457                 rbehind = cbehind;
1458         }
1459
1460         /*
1461          * scan backward for the read behind pages -- in memory
1462          */
1463         if (pindex > 0) {
1464                 if (rbehind > pindex) {
1465                         rbehind = pindex;
1466                         startpindex = 0;
1467                 } else {
1468                         startpindex = pindex - rbehind;
1469                 }
1470
1471                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1472                     rtm->pindex >= startpindex)
1473                         startpindex = rtm->pindex + 1;
1474
1475                 /* tpindex is unsigned; beware of numeric underflow. */
1476                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1477                     tpindex < pindex; i++, tpindex--) {
1478
1479                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1480                             VM_ALLOC_IFNOTCACHED);
1481                         if (rtm == NULL) {
1482                                 /*
1483                                  * Shift the allocated pages to the
1484                                  * beginning of the array.
1485                                  */
1486                                 for (j = 0; j < i; j++) {
1487                                         marray[j] = marray[j + tpindex + 1 -
1488                                             startpindex];
1489                                 }
1490                                 break;
1491                         }
1492
1493                         marray[tpindex - startpindex] = rtm;
1494                 }
1495         } else {
1496                 startpindex = 0;
1497                 i = 0;
1498         }
1499
1500         marray[i] = m;
1501         /* page offset of the required page */
1502         *reqpage = i;
1503
1504         tpindex = pindex + 1;
1505         i++;
1506
1507         /*
1508          * scan forward for the read ahead pages
1509          */
1510         endpindex = tpindex + rahead;
1511         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1512                 endpindex = rtm->pindex;
1513         if (endpindex > object->size)
1514                 endpindex = object->size;
1515
1516         for (; tpindex < endpindex; i++, tpindex++) {
1517
1518                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1519                     VM_ALLOC_IFNOTCACHED);
1520                 if (rtm == NULL) {
1521                         break;
1522                 }
1523
1524                 marray[i] = rtm;
1525         }
1526
1527         /* return number of pages */
1528         return i;
1529 }
1530
1531 /*
1532  * Block entry into the machine-independent layer's page fault handler by
1533  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1534  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1535  * spurious page faults.
1536  */
1537 int
1538 vm_fault_disable_pagefaults(void)
1539 {
1540
1541         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1542 }
1543
1544 void
1545 vm_fault_enable_pagefaults(int save)
1546 {
1547
1548         curthread_pflags_restore(save);
1549 }