sys/vm/vm_fault.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  *
   9  *
  10  * This code is derived from software contributed to Berkeley by
  11  * The Mach Operating System project at Carnegie-Mellon University.
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  * 1. Redistributions of source code must retain the above copyright
  17  *    notice, this list of conditions and the following disclaimer.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. All advertising materials mentioning features or use of this software
  22  *    must display the following acknowledgement:
  23  *      This product includes software developed by the University of
  24  *      California, Berkeley and its contributors.
  25  * 4. Neither the name of the University nor the names of its contributors
  26  *    may be used to endorse or promote products derived from this software
  27  *    without specific prior written permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  39  * SUCH DAMAGE.
  40  *
  41  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  42  *
  43  *
  44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  45  * All rights reserved.
  46  *
  47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  48  *
  49  * Permission to use, copy, modify and distribute this software and
  50  * its documentation is hereby granted, provided that both the copyright
  51  * notice and this permission notice appear in all copies of the
  52  * software, derivative works or modified versions, and any portions
  53  * thereof, and that both notices appear in supporting documentation.
  54  *
  55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  58  *
  59  * Carnegie Mellon requests users of this software to return to
  60  *
  61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  62  *  School of Computer Science
  63  *  Carnegie Mellon University
  64  *  Pittsburgh PA 15213-3890
  65  *
  66  * any improvements or extensions that they make and grant Carnegie the
  67  * rights to redistribute these changes.
  68  */
  69
  70 /*
  71  *      Page fault handling module.
  72  */
  73
  74 #include <sys/cdefs.h>
  75 __FBSDID("$FreeBSD$");
  76
  77 #include "opt_ktrace.h"
  78 #include "opt_vm.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/kernel.h>
  83 #include <sys/lock.h>
  84 #include <sys/mutex.h>
  85 #include <sys/proc.h>
  86 #include <sys/resourcevar.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vm_extern.h>
 104
 105 #include <sys/mount.h>  /* XXX Temporary for VFS_LOCK_GIANT() */
 106
 107 #define PFBAK 4
 108 #define PFFOR 4
 109 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 110
 111 static int prefault_pageorder[] = {
 112         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 113         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 114         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 115         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 116 };
 117
 118 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 119 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 120
 121 #define VM_FAULT_READ_BEHIND    8
 122 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 123 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 124 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 125 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 126
 127 struct faultstate {
 128         vm_page_t m;
 129         vm_object_t object;
 130         vm_pindex_t pindex;
 131         vm_page_t first_m;
 132         vm_object_t     first_object;
 133         vm_pindex_t first_pindex;
 134         vm_map_t map;
 135         vm_map_entry_t entry;
 136         int lookup_still_valid;
 137         struct vnode *vp;
 138         int vfslocked;
 139 };
 140
 141 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 142
 143 static inline void
 144 release_page(struct faultstate *fs)
 145 {
 146
 147         vm_page_wakeup(fs->m);
 148         vm_page_lock(fs->m);
 149         vm_page_deactivate(fs->m);
 150         vm_page_unlock(fs->m);
 151         fs->m = NULL;
 152 }
 153
 154 static inline void
 155 unlock_map(struct faultstate *fs)
 156 {
 157
 158         if (fs->lookup_still_valid) {
 159                 vm_map_lookup_done(fs->map, fs->entry);
 160                 fs->lookup_still_valid = FALSE;
 161         }
 162 }
 163
 164 static void
 165 unlock_and_deallocate(struct faultstate *fs)
 166 {
 167
 168         vm_object_pip_wakeup(fs->object);
 169         VM_OBJECT_UNLOCK(fs->object);
 170         if (fs->object != fs->first_object) {
 171                 VM_OBJECT_LOCK(fs->first_object);
 172                 vm_page_lock(fs->first_m);
 173                 vm_page_free(fs->first_m);
 174                 vm_page_unlock(fs->first_m);
 175                 vm_object_pip_wakeup(fs->first_object);
 176                 VM_OBJECT_UNLOCK(fs->first_object);
 177                 fs->first_m = NULL;
 178         }
 179         vm_object_deallocate(fs->first_object);
 180         unlock_map(fs);
 181         if (fs->vp != NULL) {
 182                 vput(fs->vp);
 183                 fs->vp = NULL;
 184         }
 185         VFS_UNLOCK_GIANT(fs->vfslocked);
 186         fs->vfslocked = 0;
 187 }
 188
 189 /*
 190  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 191  *            current object *might* contain the page.
 192  *
 193  *            default objects are zero-fill, there is no real pager.
 194  */
 195 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 196                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 197
 198 /*
 199  *      vm_fault:
 200  *
 201  *      Handle a page fault occurring at the given address,
 202  *      requiring the given permissions, in the map specified.
 203  *      If successful, the page is inserted into the
 204  *      associated physical map.
 205  *
 206  *      NOTE: the given address should be truncated to the
 207  *      proper page address.
 208  *
 209  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 210  *      a standard error specifying why the fault is fatal is returned.
 211  *
 212  *      The map in question must be referenced, and remains so.
 213  *      Caller may hold no locks.
 214  */
 215 int
 216 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 217     int fault_flags)
 218 {
 219         struct thread *td;
 220         int result;
 221
 222         td = curthread;
 223         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 224                 return (KERN_PROTECTION_FAILURE);
 225 #ifdef KTRACE
 226         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 227                 ktrfault(vaddr, fault_type);
 228 #endif
 229         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 230             NULL);
 231 #ifdef KTRACE
 232         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 233                 ktrfaultend(result);
 234 #endif
 235         return (result);
 236 }
 237
 238 int
 239 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 240     int fault_flags, vm_page_t *m_hold)
 241 {
 242         vm_prot_t prot;
 243         long ahead, behind;
 244         int alloc_req, era, faultcount, nera, reqpage, result;
 245         boolean_t growstack, is_first_object_locked, wired;
 246         int map_generation;
 247         vm_object_t next_object;
 248         vm_page_t marray[VM_FAULT_READ_MAX];
 249         int hardfault;
 250         struct faultstate fs;
 251         struct vnode *vp;
 252         int locked, error;
 253
 254         hardfault = 0;
 255         growstack = TRUE;
 256         PCPU_INC(cnt.v_vm_faults);
 257         fs.vp = NULL;
 258         fs.vfslocked = 0;
 259         faultcount = reqpage = 0;
 260
 261 RetryFault:;
 262
 263         /*
 264          * Find the backing store object and offset into it to begin the
 265          * search.
 266          */
 267         fs.map = map;
 268         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 269             &fs.first_object, &fs.first_pindex, &prot, &wired);
 270         if (result != KERN_SUCCESS) {
 271                 if (growstack && result == KERN_INVALID_ADDRESS &&
 272                     map != kernel_map) {
 273                         result = vm_map_growstack(curproc, vaddr);
 274                         if (result != KERN_SUCCESS)
 275                                 return (KERN_FAILURE);
 276                         growstack = FALSE;
 277                         goto RetryFault;
 278                 }
 279                 return (result);
 280         }
 281
 282         map_generation = fs.map->timestamp;
 283
 284         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 285                 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
 286                         vm_map_unlock_read(fs.map);
 287                         return (KERN_FAILURE);
 288                 }
 289                 panic("vm_fault: fault on nofault entry, addr: %lx",
 290                     (u_long)vaddr);
 291         }
 292
 293         if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
 294             fs.entry->wiring_thread != curthread) {
 295                 vm_map_unlock_read(fs.map);
 296                 vm_map_lock(fs.map);
 297                 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
 298                     (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
 299                         if (fs.vp != NULL) {
 300                                 vput(fs.vp);
 301                                 fs.vp = NULL;
 302                         }
 303                         fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
 304                         vm_map_unlock_and_wait(fs.map, 0);
 305                 } else
 306                         vm_map_unlock(fs.map);
 307                 goto RetryFault;
 308         }
 309
 310         /*
 311          * Make a reference to this object to prevent its disposal while we
 312          * are messing with it.  Once we have the reference, the map is free
 313          * to be diddled.  Since objects reference their shadows (and copies),
 314          * they will stay around as well.
 315          *
 316          * Bump the paging-in-progress count to prevent size changes (e.g.
 317          * truncation operations) during I/O.  This must be done after
 318          * obtaining the vnode lock in order to avoid possible deadlocks.
 319          */
 320         VM_OBJECT_LOCK(fs.first_object);
 321         vm_object_reference_locked(fs.first_object);
 322         vm_object_pip_add(fs.first_object, 1);
 323
 324         fs.lookup_still_valid = TRUE;
 325
 326         if (wired)
 327                 fault_type = prot | (fault_type & VM_PROT_COPY);
 328
 329         fs.first_m = NULL;
 330
 331         /*
 332          * Search for the page at object/offset.
 333          */
 334         fs.object = fs.first_object;
 335         fs.pindex = fs.first_pindex;
 336         while (TRUE) {
 337                 /*
 338                  * If the object is dead, we stop here
 339                  */
 340                 if (fs.object->flags & OBJ_DEAD) {
 341                         unlock_and_deallocate(&fs);
 342                         return (KERN_PROTECTION_FAILURE);
 343                 }
 344
 345                 /*
 346                  * See if page is resident
 347                  */
 348                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 349                 if (fs.m != NULL) {
 350                         /*
 351                          * check for page-based copy on write.
 352                          * We check fs.object == fs.first_object so
 353                          * as to ensure the legacy COW mechanism is
 354                          * used when the page in question is part of
 355                          * a shadow object.  Otherwise, vm_page_cowfault()
 356                          * removes the page from the backing object,
 357                          * which is not what we want.
 358                          */
 359                         vm_page_lock(fs.m);
 360                         if ((fs.m->cow) &&
 361                             (fault_type & VM_PROT_WRITE) &&
 362                             (fs.object == fs.first_object)) {
 363                                 vm_page_cowfault(fs.m);
 364                                 unlock_and_deallocate(&fs);
 365                                 goto RetryFault;
 366                         }
 367
 368                         /*
 369                          * Wait/Retry if the page is busy.  We have to do this
 370                          * if the page is busy via either VPO_BUSY or
 371                          * vm_page_t->busy because the vm_pager may be using
 372                          * vm_page_t->busy for pageouts ( and even pageins if
 373                          * it is the vnode pager ), and we could end up trying
 374                          * to pagein and pageout the same page simultaneously.
 375                          *
 376                          * We can theoretically allow the busy case on a read
 377                          * fault if the page is marked valid, but since such
 378                          * pages are typically already pmap'd, putting that
 379                          * special case in might be more effort then it is
 380                          * worth.  We cannot under any circumstances mess
 381                          * around with a vm_page_t->busy page except, perhaps,
 382                          * to pmap it.
 383                          */
 384                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 385                                 /*
 386                                  * Reference the page before unlocking and
 387                                  * sleeping so that the page daemon is less
 388                                  * likely to reclaim it.
 389                                  */
 390                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 391                                 vm_page_unlock(fs.m);
 392                                 if (fs.object != fs.first_object) {
 393                                         if (!VM_OBJECT_TRYLOCK(
 394                                             fs.first_object)) {
 395                                                 VM_OBJECT_UNLOCK(fs.object);
 396                                                 VM_OBJECT_LOCK(fs.first_object);
 397                                                 VM_OBJECT_LOCK(fs.object);
 398                                         }
 399                                         vm_page_lock(fs.first_m);
 400                                         vm_page_free(fs.first_m);
 401                                         vm_page_unlock(fs.first_m);
 402                                         vm_object_pip_wakeup(fs.first_object);
 403                                         VM_OBJECT_UNLOCK(fs.first_object);
 404                                         fs.first_m = NULL;
 405                                 }
 406                                 unlock_map(&fs);
 407                                 if (fs.m == vm_page_lookup(fs.object,
 408                                     fs.pindex)) {
 409                                         vm_page_sleep_if_busy(fs.m, TRUE,
 410                                             "vmpfw");
 411                                 }
 412                                 vm_object_pip_wakeup(fs.object);
 413                                 VM_OBJECT_UNLOCK(fs.object);
 414                                 PCPU_INC(cnt.v_intrans);
 415                                 vm_object_deallocate(fs.first_object);
 416                                 goto RetryFault;
 417                         }
 418                         vm_pageq_remove(fs.m);
 419                         vm_page_unlock(fs.m);
 420
 421                         /*
 422                          * Mark page busy for other processes, and the
 423                          * pagedaemon.  If it still isn't completely valid
 424                          * (readable), jump to readrest, else break-out ( we
 425                          * found the page ).
 426                          */
 427                         vm_page_busy(fs.m);
 428                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 429                                 goto readrest;
 430                         break;
 431                 }
 432
 433                 /*
 434                  * Page is not resident, If this is the search termination
 435                  * or the pager might contain the page, allocate a new page.
 436                  */
 437                 if (TRYPAGER || fs.object == fs.first_object) {
 438                         if (fs.pindex >= fs.object->size) {
 439                                 unlock_and_deallocate(&fs);
 440                                 return (KERN_PROTECTION_FAILURE);
 441                         }
 442
 443                         /*
 444                          * Allocate a new page for this object/offset pair.
 445                          *
 446                          * Unlocked read of the p_flag is harmless. At
 447                          * worst, the P_KILLED might be not observed
 448                          * there, and allocation can fail, causing
 449                          * restart and new reading of the p_flag.
 450                          */
 451                         fs.m = NULL;
 452                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 453 #if VM_NRESERVLEVEL > 0
 454                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 455                                         fs.object->flags |= OBJ_COLORED;
 456                                         fs.object->pg_color = atop(vaddr) -
 457                                             fs.pindex;
 458                                 }
 459 #endif
 460                                 alloc_req = P_KILLED(curproc) ?
 461                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 462                                 if (fs.object->type != OBJT_VNODE &&
 463                                     fs.object->backing_object == NULL)
 464                                         alloc_req |= VM_ALLOC_ZERO;
 465                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 466                                     alloc_req);
 467                         }
 468                         if (fs.m == NULL) {
 469                                 unlock_and_deallocate(&fs);
 470                                 VM_WAITPFAULT;
 471                                 goto RetryFault;
 472                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 473                                 break;
 474                 }
 475
 476 readrest:
 477                 /*
 478                  * We have found a valid page or we have allocated a new page.
 479                  * The page thus may not be valid or may not be entirely
 480                  * valid.
 481                  *
 482                  * Attempt to fault-in the page if there is a chance that the
 483                  * pager has it, and potentially fault in additional pages
 484                  * at the same time.
 485                  */
 486                 if (TRYPAGER) {
 487                         int rv;
 488                         u_char behavior = vm_map_entry_behavior(fs.entry);
 489
 490                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 491                             P_KILLED(curproc)) {
 492                                 behind = 0;
 493                                 ahead = 0;
 494                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 495                                 behind = 0;
 496                                 ahead = atop(fs.entry->end - vaddr) - 1;
 497                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 498                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 499                                 if (fs.pindex == fs.entry->next_read)
 500                                         vm_fault_cache_behind(&fs,
 501                                             VM_FAULT_READ_MAX);
 502                         } else {
 503                                 /*
 504                                  * If this is a sequential page fault, then
 505                                  * arithmetically increase the number of pages
 506                                  * in the read-ahead window.  Otherwise, reset
 507                                  * the read-ahead window to its smallest size.
 508                                  */
 509                                 behind = atop(vaddr - fs.entry->start);
 510                                 if (behind > VM_FAULT_READ_BEHIND)
 511                                         behind = VM_FAULT_READ_BEHIND;
 512                                 ahead = atop(fs.entry->end - vaddr) - 1;
 513                                 era = fs.entry->read_ahead;
 514                                 if (fs.pindex == fs.entry->next_read) {
 515                                         nera = era + behind;
 516                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 517                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 518                                         behind = 0;
 519                                         if (ahead > nera)
 520                                                 ahead = nera;
 521                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 522                                                 vm_fault_cache_behind(&fs,
 523                                                     VM_FAULT_CACHE_BEHIND);
 524                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 525                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 526                                 if (era != ahead)
 527                                         fs.entry->read_ahead = ahead;
 528                         }
 529
 530                         /*
 531                          * Call the pager to retrieve the data, if any, after
 532                          * releasing the lock on the map.  We hold a ref on
 533                          * fs.object and the pages are VPO_BUSY'd.
 534                          */
 535                         unlock_map(&fs);
 536
 537 vnode_lock:
 538                         if (fs.object->type == OBJT_VNODE) {
 539                                 vp = fs.object->handle;
 540                                 if (vp == fs.vp)
 541                                         goto vnode_locked;
 542                                 else if (fs.vp != NULL) {
 543                                         vput(fs.vp);
 544                                         fs.vp = NULL;
 545                                 }
 546                                 locked = VOP_ISLOCKED(vp);
 547
 548                                 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
 549                                         fs.vfslocked = 1;
 550                                         if (!mtx_trylock(&Giant)) {
 551                                                 VM_OBJECT_UNLOCK(fs.object);
 552                                                 mtx_lock(&Giant);
 553                                                 VM_OBJECT_LOCK(fs.object);
 554                                                 goto vnode_lock;
 555                                         }
 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                                         int vfslocked;
 564
 565                                         vfslocked = fs.vfslocked;
 566                                         fs.vfslocked = 0; /* Keep Giant */
 567                                         vhold(vp);
 568                                         release_page(&fs);
 569                                         unlock_and_deallocate(&fs);
 570                                         error = vget(vp, locked | LK_RETRY |
 571                                             LK_CANRECURSE, curthread);
 572                                         vdrop(vp);
 573                                         fs.vp = vp;
 574                                         fs.vfslocked = vfslocked;
 575                                         KASSERT(error == 0,
 576                                             ("vm_fault: vget failed"));
 577                                         goto RetryFault;
 578                                 }
 579                                 fs.vp = vp;
 580                         }
 581 vnode_locked:
 582                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 583                             ("vm_fault: vnode-backed object mapped by system map"));
 584
 585                         /*
 586                          * now we find out if any other pages should be paged
 587                          * in at this time this routine checks to see if the
 588                          * pages surrounding this fault reside in the same
 589                          * object as the page for this fault.  If they do,
 590                          * then they are faulted in also into the object.  The
 591                          * array "marray" returned contains an array of
 592                          * vm_page_t structs where one of them is the
 593                          * vm_page_t passed to the routine.  The reqpage
 594                          * return value is the index into the marray for the
 595                          * vm_page_t passed to the routine.
 596                          *
 597                          * fs.m plus the additional pages are VPO_BUSY'd.
 598                          */
 599                         faultcount = vm_fault_additional_pages(
 600                             fs.m, behind, ahead, marray, &reqpage);
 601
 602                         rv = faultcount ?
 603                             vm_pager_get_pages(fs.object, marray, faultcount,
 604                                 reqpage) : VM_PAGER_FAIL;
 605
 606                         if (rv == VM_PAGER_OK) {
 607                                 /*
 608                                  * Found the page. Leave it busy while we play
 609                                  * with it.
 610                                  */
 611
 612                                 /*
 613                                  * Relookup in case pager changed page. Pager
 614                                  * is responsible for disposition of old page
 615                                  * if moved.
 616                                  */
 617                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 618                                 if (!fs.m) {
 619                                         unlock_and_deallocate(&fs);
 620                                         goto RetryFault;
 621                                 }
 622
 623                                 hardfault++;
 624                                 break; /* break to PAGE HAS BEEN FOUND */
 625                         }
 626                         /*
 627                          * Remove the bogus page (which does not exist at this
 628                          * object/offset); before doing so, we must get back
 629                          * our object lock to preserve our invariant.
 630                          *
 631                          * Also wake up any other process that may want to bring
 632                          * in this page.
 633                          *
 634                          * If this is the top-level object, we must leave the
 635                          * busy page to prevent another process from rushing
 636                          * past us, and inserting the page in that object at
 637                          * the same time that we are.
 638                          */
 639                         if (rv == VM_PAGER_ERROR)
 640                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 641                                     curproc->p_pid, curproc->p_comm);
 642                         /*
 643                          * Data outside the range of the pager or an I/O error
 644                          */
 645                         /*
 646                          * XXX - the check for kernel_map is a kludge to work
 647                          * around having the machine panic on a kernel space
 648                          * fault w/ I/O error.
 649                          */
 650                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 651                                 (rv == VM_PAGER_BAD)) {
 652                                 vm_page_lock(fs.m);
 653                                 vm_page_free(fs.m);
 654                                 vm_page_unlock(fs.m);
 655                                 fs.m = NULL;
 656                                 unlock_and_deallocate(&fs);
 657                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 658                         }
 659                         if (fs.object != fs.first_object) {
 660                                 vm_page_lock(fs.m);
 661                                 vm_page_free(fs.m);
 662                                 vm_page_unlock(fs.m);
 663                                 fs.m = NULL;
 664                                 /*
 665                                  * XXX - we cannot just fall out at this
 666                                  * point, m has been freed and is invalid!
 667                                  */
 668                         }
 669                 }
 670
 671                 /*
 672                  * We get here if the object has default pager (or unwiring)
 673                  * or the pager doesn't have the page.
 674                  */
 675                 if (fs.object == fs.first_object)
 676                         fs.first_m = fs.m;
 677
 678                 /*
 679                  * Move on to the next object.  Lock the next object before
 680                  * unlocking the current one.
 681                  */
 682                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 683                 next_object = fs.object->backing_object;
 684                 if (next_object == NULL) {
 685                         /*
 686                          * If there's no object left, fill the page in the top
 687                          * object with zeros.
 688                          */
 689                         if (fs.object != fs.first_object) {
 690                                 vm_object_pip_wakeup(fs.object);
 691                                 VM_OBJECT_UNLOCK(fs.object);
 692
 693                                 fs.object = fs.first_object;
 694                                 fs.pindex = fs.first_pindex;
 695                                 fs.m = fs.first_m;
 696                                 VM_OBJECT_LOCK(fs.object);
 697                         }
 698                         fs.first_m = NULL;
 699
 700                         /*
 701                          * Zero the page if necessary and mark it valid.
 702                          */
 703                         if ((fs.m->flags & PG_ZERO) == 0) {
 704                                 pmap_zero_page(fs.m);
 705                         } else {
 706                                 PCPU_INC(cnt.v_ozfod);
 707                         }
 708                         PCPU_INC(cnt.v_zfod);
 709                         fs.m->valid = VM_PAGE_BITS_ALL;
 710                         break;  /* break to PAGE HAS BEEN FOUND */
 711                 } else {
 712                         KASSERT(fs.object != next_object,
 713                             ("object loop %p", next_object));
 714                         VM_OBJECT_LOCK(next_object);
 715                         vm_object_pip_add(next_object, 1);
 716                         if (fs.object != fs.first_object)
 717                                 vm_object_pip_wakeup(fs.object);
 718                         VM_OBJECT_UNLOCK(fs.object);
 719                         fs.object = next_object;
 720                 }
 721         }
 722
 723         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 724             ("vm_fault: not busy after main loop"));
 725
 726         /*
 727          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 728          * is held.]
 729          */
 730
 731         /*
 732          * If the page is being written, but isn't already owned by the
 733          * top-level object, we have to copy it into a new page owned by the
 734          * top-level object.
 735          */
 736         if (fs.object != fs.first_object) {
 737                 /*
 738                  * We only really need to copy if we want to write it.
 739                  */
 740                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 741                         /*
 742                          * This allows pages to be virtually copied from a
 743                          * backing_object into the first_object, where the
 744                          * backing object has no other refs to it, and cannot
 745                          * gain any more refs.  Instead of a bcopy, we just
 746                          * move the page from the backing object to the
 747                          * first object.  Note that we must mark the page
 748                          * dirty in the first object so that it will go out
 749                          * to swap when needed.
 750                          */
 751                         is_first_object_locked = FALSE;
 752                         if (
 753                                 /*
 754                                  * Only one shadow object
 755                                  */
 756                                 (fs.object->shadow_count == 1) &&
 757                                 /*
 758                                  * No COW refs, except us
 759                                  */
 760                                 (fs.object->ref_count == 1) &&
 761                                 /*
 762                                  * No one else can look this object up
 763                                  */
 764                                 (fs.object->handle == NULL) &&
 765                                 /*
 766                                  * No other ways to look the object up
 767                                  */
 768                                 ((fs.object->type == OBJT_DEFAULT) ||
 769                                  (fs.object->type == OBJT_SWAP)) &&
 770                             (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
 771                                 /*
 772                                  * We don't chase down the shadow chain
 773                                  */
 774                             fs.object == fs.first_object->backing_object) {
 775                                 /*
 776                                  * get rid of the unnecessary page
 777                                  */
 778                                 vm_page_lock(fs.first_m);
 779                                 vm_page_free(fs.first_m);
 780                                 vm_page_unlock(fs.first_m);
 781                                 /*
 782                                  * grab the page and put it into the
 783                                  * process'es object.  The page is
 784                                  * automatically made dirty.
 785                                  */
 786                                 vm_page_lock(fs.m);
 787                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 788                                 vm_page_unlock(fs.m);
 789                                 vm_page_busy(fs.m);
 790                                 fs.first_m = fs.m;
 791                                 fs.m = NULL;
 792                                 PCPU_INC(cnt.v_cow_optim);
 793                         } else {
 794                                 /*
 795                                  * Oh, well, lets copy it.
 796                                  */
 797                                 pmap_copy_page(fs.m, fs.first_m);
 798                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 799                                 if (wired && (fault_flags &
 800                                     VM_FAULT_CHANGE_WIRING) == 0) {
 801                                         vm_page_lock(fs.first_m);
 802                                         vm_page_wire(fs.first_m);
 803                                         vm_page_unlock(fs.first_m);
 804
 805                                         vm_page_lock(fs.m);
 806                                         vm_page_unwire(fs.m, FALSE);
 807                                         vm_page_unlock(fs.m);
 808                                 }
 809                                 /*
 810                                  * We no longer need the old page or object.
 811                                  */
 812                                 release_page(&fs);
 813                         }
 814                         /*
 815                          * fs.object != fs.first_object due to above
 816                          * conditional
 817                          */
 818                         vm_object_pip_wakeup(fs.object);
 819                         VM_OBJECT_UNLOCK(fs.object);
 820                         /*
 821                          * Only use the new page below...
 822                          */
 823                         fs.object = fs.first_object;
 824                         fs.pindex = fs.first_pindex;
 825                         fs.m = fs.first_m;
 826                         if (!is_first_object_locked)
 827                                 VM_OBJECT_LOCK(fs.object);
 828                         PCPU_INC(cnt.v_cow_faults);
 829                         curthread->td_cow++;
 830                 } else {
 831                         prot &= ~VM_PROT_WRITE;
 832                 }
 833         }
 834
 835         /*
 836          * We must verify that the maps have not changed since our last
 837          * lookup.
 838          */
 839         if (!fs.lookup_still_valid) {
 840                 vm_object_t retry_object;
 841                 vm_pindex_t retry_pindex;
 842                 vm_prot_t retry_prot;
 843
 844                 if (!vm_map_trylock_read(fs.map)) {
 845                         release_page(&fs);
 846                         unlock_and_deallocate(&fs);
 847                         goto RetryFault;
 848                 }
 849                 fs.lookup_still_valid = TRUE;
 850                 if (fs.map->timestamp != map_generation) {
 851                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 852                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 853
 854                         /*
 855                          * If we don't need the page any longer, put it on the inactive
 856                          * list (the easiest thing to do here).  If no one needs it,
 857                          * pageout will grab it eventually.
 858                          */
 859                         if (result != KERN_SUCCESS) {
 860                                 release_page(&fs);
 861                                 unlock_and_deallocate(&fs);
 862
 863                                 /*
 864                                  * If retry of map lookup would have blocked then
 865                                  * retry fault from start.
 866                                  */
 867                                 if (result == KERN_FAILURE)
 868                                         goto RetryFault;
 869                                 return (result);
 870                         }
 871                         if ((retry_object != fs.first_object) ||
 872                             (retry_pindex != fs.first_pindex)) {
 873                                 release_page(&fs);
 874                                 unlock_and_deallocate(&fs);
 875                                 goto RetryFault;
 876                         }
 877
 878                         /*
 879                          * Check whether the protection has changed or the object has
 880                          * been copied while we left the map unlocked. Changing from
 881                          * read to write permission is OK - we leave the page
 882                          * write-protected, and catch the write fault. Changing from
 883                          * write to read permission means that we can't mark the page
 884                          * write-enabled after all.
 885                          */
 886                         prot &= retry_prot;
 887                 }
 888         }
 889         /*
 890          * If the page was filled by a pager, update the map entry's
 891          * last read offset.  Since the pager does not return the
 892          * actual set of pages that it read, this update is based on
 893          * the requested set.  Typically, the requested and actual
 894          * sets are the same.
 895          *
 896          * XXX The following assignment modifies the map
 897          * without holding a write lock on it.
 898          */
 899         if (hardfault)
 900                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 901
 902         if ((prot & VM_PROT_WRITE) != 0 ||
 903             (fault_flags & VM_FAULT_DIRTY) != 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         /*
 938          * Page had better still be busy
 939          */
 940         KASSERT(fs.m->oflags & VPO_BUSY,
 941                 ("vm_fault: page %p not busy!", fs.m));
 942         /*
 943          * Page must be completely valid or it is not fit to
 944          * map into user space.  vm_pager_get_pages() ensures this.
 945          */
 946         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 947             ("vm_fault: page %p partially invalid", fs.m));
 948         VM_OBJECT_UNLOCK(fs.object);
 949
 950         /*
 951          * Put this page into the physical map.  We had to do the unlock above
 952          * because pmap_enter() may sleep.  We don't put the page
 953          * back on the active queue until later so that the pageout daemon
 954          * won't find it (yet).
 955          */
 956         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 957         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 958                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 959         VM_OBJECT_LOCK(fs.object);
 960         vm_page_lock(fs.m);
 961
 962         /*
 963          * If the page is not wired down, then put it where the pageout daemon
 964          * can find it.
 965          */
 966         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 967                 if (wired)
 968                         vm_page_wire(fs.m);
 969                 else
 970                         vm_page_unwire(fs.m, 1);
 971         } else
 972                 vm_page_activate(fs.m);
 973         if (m_hold != NULL) {
 974                 *m_hold = fs.m;
 975                 vm_page_hold(fs.m);
 976         }
 977         vm_page_unlock(fs.m);
 978         vm_page_wakeup(fs.m);
 979
 980         /*
 981          * Unlock everything, and return
 982          */
 983         unlock_and_deallocate(&fs);
 984         if (hardfault)
 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_LOCK_ASSERT(object, MA_OWNED);
1005         first_object = fs->first_object;
1006         if (first_object != object) {
1007                 if (!VM_OBJECT_TRYLOCK(first_object)) {
1008                         VM_OBJECT_UNLOCK(object);
1009                         VM_OBJECT_LOCK(first_object);
1010                         VM_OBJECT_LOCK(object);
1011                 }
1012         }
1013         if (first_object->type != OBJT_DEVICE &&
1014             first_object->type != OBJT_PHYS && first_object->type != OBJT_SG) {
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                 KASSERT((m->oflags & VPO_BUSY) != 0,
1023                     ("vm_fault_cache_behind: page %p is not busy", m));
1024                 m_prev = vm_page_prev(m);
1025                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1026                     m->valid == VM_PAGE_BITS_ALL) {
1027                         m_prev = vm_page_prev(m);
1028                         if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0)
1029                                 continue;
1030                         vm_page_lock(m);
1031                         if (m->hold_count == 0 && m->wire_count == 0) {
1032                                 pmap_remove_all(m);
1033                                 vm_page_aflag_clear(m, PGA_REFERENCED);
1034                                 if (m->dirty != 0)
1035                                         vm_page_deactivate(m);
1036                                 else
1037                                         vm_page_cache(m);
1038                         }
1039                         vm_page_unlock(m);
1040                 }
1041         }
1042         if (first_object != object)
1043                 VM_OBJECT_UNLOCK(first_object);
1044 }
1045
1046 /*
1047  * vm_fault_prefault provides a quick way of clustering
1048  * pagefaults into a processes address space.  It is a "cousin"
1049  * of vm_map_pmap_enter, except it runs at page fault time instead
1050  * of mmap time.
1051  */
1052 static void
1053 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1054 {
1055         int i;
1056         vm_offset_t addr, starta;
1057         vm_pindex_t pindex;
1058         vm_page_t m;
1059         vm_object_t object;
1060
1061         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1062                 return;
1063
1064         object = entry->object.vm_object;
1065
1066         starta = addra - PFBAK * PAGE_SIZE;
1067         if (starta < entry->start) {
1068                 starta = entry->start;
1069         } else if (starta > addra) {
1070                 starta = 0;
1071         }
1072
1073         for (i = 0; i < PAGEORDER_SIZE; i++) {
1074                 vm_object_t backing_object, lobject;
1075
1076                 addr = addra + prefault_pageorder[i];
1077                 if (addr > addra + (PFFOR * PAGE_SIZE))
1078                         addr = 0;
1079
1080                 if (addr < starta || addr >= entry->end)
1081                         continue;
1082
1083                 if (!pmap_is_prefaultable(pmap, addr))
1084                         continue;
1085
1086                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1087                 lobject = object;
1088                 VM_OBJECT_LOCK(lobject);
1089                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1090                     lobject->type == OBJT_DEFAULT &&
1091                     (backing_object = lobject->backing_object) != NULL) {
1092                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1093                             0, ("vm_fault_prefault: unaligned object offset"));
1094                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1095                         VM_OBJECT_LOCK(backing_object);
1096                         VM_OBJECT_UNLOCK(lobject);
1097                         lobject = backing_object;
1098                 }
1099                 /*
1100                  * give-up when a page is not in memory
1101                  */
1102                 if (m == NULL) {
1103                         VM_OBJECT_UNLOCK(lobject);
1104                         break;
1105                 }
1106                 if (m->valid == VM_PAGE_BITS_ALL &&
1107                     (m->flags & PG_FICTITIOUS) == 0)
1108                         pmap_enter_quick(pmap, addr, m, entry->protection);
1109                 VM_OBJECT_UNLOCK(lobject);
1110         }
1111 }
1112
1113 /*
1114  * Hold each of the physical pages that are mapped by the specified range of
1115  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1116  * and allow the specified types of access, "prot".  If all of the implied
1117  * pages are successfully held, then the number of held pages is returned
1118  * together with pointers to those pages in the array "ma".  However, if any
1119  * of the pages cannot be held, -1 is returned.
1120  */
1121 int
1122 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1123     vm_prot_t prot, vm_page_t *ma, int max_count)
1124 {
1125         vm_offset_t end, va;
1126         vm_page_t *mp;
1127         int count;
1128         boolean_t pmap_failed;
1129
1130         if (len == 0)
1131                 return (0);
1132         end = round_page(addr + len);
1133         addr = trunc_page(addr);
1134
1135         /*
1136          * Check for illegal addresses.
1137          */
1138         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1139                 return (-1);
1140
1141         if (atop(end - addr) > max_count)
1142                 panic("vm_fault_quick_hold_pages: count > max_count");
1143         count = atop(end - addr);
1144
1145         /*
1146          * Most likely, the physical pages are resident in the pmap, so it is
1147          * faster to try pmap_extract_and_hold() first.
1148          */
1149         pmap_failed = FALSE;
1150         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1151                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1152                 if (*mp == NULL)
1153                         pmap_failed = TRUE;
1154                 else if ((prot & VM_PROT_WRITE) != 0 &&
1155                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1156                         /*
1157                          * Explicitly dirty the physical page.  Otherwise, the
1158                          * caller's changes may go unnoticed because they are
1159                          * performed through an unmanaged mapping or by a DMA
1160                          * operation.
1161                          *
1162                          * The object lock is not held here.
1163                          * See vm_page_clear_dirty_mask().
1164                          */
1165                         vm_page_dirty(*mp);
1166                 }
1167         }
1168         if (pmap_failed) {
1169                 /*
1170                  * One or more pages could not be held by the pmap.  Either no
1171                  * page was mapped at the specified virtual address or that
1172                  * mapping had insufficient permissions.  Attempt to fault in
1173                  * and hold these pages.
1174                  */
1175                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1176                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1177                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1178                                 goto error;
1179         }
1180         return (count);
1181 error:
1182         for (mp = ma; mp < ma + count; mp++)
1183                 if (*mp != NULL) {
1184                         vm_page_lock(*mp);
1185                         vm_page_unhold(*mp);
1186                         vm_page_unlock(*mp);
1187                 }
1188         return (-1);
1189 }
1190
1191 /*
1192  *      vm_fault_wire:
1193  *
1194  *      Wire down a range of virtual addresses in a map.
1195  */
1196 int
1197 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1198     boolean_t fictitious)
1199 {
1200         vm_offset_t va;
1201         int rv;
1202
1203         /*
1204          * We simulate a fault to get the page and enter it in the physical
1205          * map.  For user wiring, we only ask for read access on currently
1206          * read-only sections.
1207          */
1208         for (va = start; va < end; va += PAGE_SIZE) {
1209                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1210                 if (rv) {
1211                         if (va != start)
1212                                 vm_fault_unwire(map, start, va, fictitious);
1213                         return (rv);
1214                 }
1215         }
1216         return (KERN_SUCCESS);
1217 }
1218
1219 /*
1220  *      vm_fault_unwire:
1221  *
1222  *      Unwire a range of virtual addresses in a map.
1223  */
1224 void
1225 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1226     boolean_t fictitious)
1227 {
1228         vm_paddr_t pa;
1229         vm_offset_t va;
1230         vm_page_t m;
1231         pmap_t pmap;
1232
1233         pmap = vm_map_pmap(map);
1234
1235         /*
1236          * Since the pages are wired down, we must be able to get their
1237          * mappings from the physical map system.
1238          */
1239         for (va = start; va < end; va += PAGE_SIZE) {
1240                 pa = pmap_extract(pmap, va);
1241                 if (pa != 0) {
1242                         pmap_change_wiring(pmap, va, FALSE);
1243                         if (!fictitious) {
1244                                 m = PHYS_TO_VM_PAGE(pa);
1245                                 vm_page_lock(m);
1246                                 vm_page_unwire(m, TRUE);
1247                                 vm_page_unlock(m);
1248                         }
1249                 }
1250         }
1251 }
1252
1253 /*
1254  *      Routine:
1255  *              vm_fault_copy_entry
1256  *      Function:
1257  *              Create new shadow object backing dst_entry with private copy of
1258  *              all underlying pages. When src_entry is equal to dst_entry,
1259  *              function implements COW for wired-down map entry. Otherwise,
1260  *              it forks wired entry into dst_map.
1261  *
1262  *      In/out conditions:
1263  *              The source and destination maps must be locked for write.
1264  *              The source map entry must be wired down (or be a sharing map
1265  *              entry corresponding to a main map entry that is wired down).
1266  */
1267 void
1268 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1269     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1270     vm_ooffset_t *fork_charge)
1271 {
1272         vm_object_t backing_object, dst_object, object, src_object;
1273         vm_pindex_t dst_pindex, pindex, src_pindex;
1274         vm_prot_t access, prot;
1275         vm_offset_t vaddr;
1276         vm_page_t dst_m;
1277         vm_page_t src_m;
1278         boolean_t upgrade;
1279
1280 #ifdef  lint
1281         src_map++;
1282 #endif  /* lint */
1283
1284         upgrade = src_entry == dst_entry;
1285
1286         src_object = src_entry->object.vm_object;
1287         src_pindex = OFF_TO_IDX(src_entry->offset);
1288
1289         /*
1290          * Create the top-level object for the destination entry. (Doesn't
1291          * actually shadow anything - we copy the pages directly.)
1292          */
1293         dst_object = vm_object_allocate(OBJT_DEFAULT,
1294             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1295 #if VM_NRESERVLEVEL > 0
1296         dst_object->flags |= OBJ_COLORED;
1297         dst_object->pg_color = atop(dst_entry->start);
1298 #endif
1299
1300         VM_OBJECT_LOCK(dst_object);
1301         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1302             ("vm_fault_copy_entry: vm_object not NULL"));
1303         dst_entry->object.vm_object = dst_object;
1304         dst_entry->offset = 0;
1305         dst_object->charge = dst_entry->end - dst_entry->start;
1306         if (fork_charge != NULL) {
1307                 KASSERT(dst_entry->cred == NULL,
1308                     ("vm_fault_copy_entry: leaked swp charge"));
1309                 dst_object->cred = curthread->td_ucred;
1310                 crhold(dst_object->cred);
1311                 *fork_charge += dst_object->charge;
1312         } else {
1313                 dst_object->cred = dst_entry->cred;
1314                 dst_entry->cred = NULL;
1315         }
1316         access = prot = dst_entry->protection;
1317         /*
1318          * If not an upgrade, then enter the mappings in the pmap as
1319          * read and/or execute accesses.  Otherwise, enter them as
1320          * write accesses.
1321          *
1322          * A writeable large page mapping is only created if all of
1323          * the constituent small page mappings are modified. Marking
1324          * PTEs as modified on inception allows promotion to happen
1325          * without taking potentially large number of soft faults.
1326          */
1327         if (!upgrade)
1328                 access &= ~VM_PROT_WRITE;
1329
1330         /*
1331          * Loop through all of the virtual pages within the entry's
1332          * range, copying each page from the source object to the
1333          * destination object.  Since the source is wired, those pages
1334          * must exist.  In contrast, the destination is pageable.
1335          * Since the destination object does share any backing storage
1336          * with the source object, all of its pages must be dirtied,
1337          * regardless of whether they can be written.
1338          */
1339         for (vaddr = dst_entry->start, dst_pindex = 0;
1340             vaddr < dst_entry->end;
1341             vaddr += PAGE_SIZE, dst_pindex++) {
1342
1343                 /*
1344                  * Allocate a page in the destination object.
1345                  */
1346                 do {
1347                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1348                             VM_ALLOC_NORMAL);
1349                         if (dst_m == NULL) {
1350                                 VM_OBJECT_UNLOCK(dst_object);
1351                                 VM_WAIT;
1352                                 VM_OBJECT_LOCK(dst_object);
1353                         }
1354                 } while (dst_m == NULL);
1355
1356                 /*
1357                  * Find the page in the source object, and copy it in.
1358                  * Because the source is wired down, the page will be
1359                  * in memory.
1360                  */
1361                 VM_OBJECT_LOCK(src_object);
1362                 object = src_object;
1363                 pindex = src_pindex + dst_pindex;
1364                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1365                     (backing_object = object->backing_object) != NULL) {
1366                         /*
1367                          * Unless the source mapping is read-only or
1368                          * it is presently being upgraded from
1369                          * read-only, the first object in the shadow
1370                          * chain should provide all of the pages.  In
1371                          * other words, this loop body should never be
1372                          * executed when the source mapping is already
1373                          * read/write.
1374                          */
1375                         KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1376                             upgrade,
1377                             ("vm_fault_copy_entry: main object missing page"));
1378
1379                         VM_OBJECT_LOCK(backing_object);
1380                         pindex += OFF_TO_IDX(object->backing_object_offset);
1381                         VM_OBJECT_UNLOCK(object);
1382                         object = backing_object;
1383                 }
1384                 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1385                 pmap_copy_page(src_m, dst_m);
1386                 VM_OBJECT_UNLOCK(object);
1387                 dst_m->valid = VM_PAGE_BITS_ALL;
1388                 dst_m->dirty = VM_PAGE_BITS_ALL;
1389                 VM_OBJECT_UNLOCK(dst_object);
1390
1391                 /*
1392                  * Enter it in the pmap. If a wired, copy-on-write
1393                  * mapping is being replaced by a write-enabled
1394                  * mapping, then wire that new mapping.
1395                  */
1396                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1397
1398                 /*
1399                  * Mark it no longer busy, and put it on the active list.
1400                  */
1401                 VM_OBJECT_LOCK(dst_object);
1402
1403                 if (upgrade) {
1404                         vm_page_lock(src_m);
1405                         vm_page_unwire(src_m, 0);
1406                         vm_page_unlock(src_m);
1407
1408                         vm_page_lock(dst_m);
1409                         vm_page_wire(dst_m);
1410                         vm_page_unlock(dst_m);
1411                 } else {
1412                         vm_page_lock(dst_m);
1413                         vm_page_activate(dst_m);
1414                         vm_page_unlock(dst_m);
1415                 }
1416                 vm_page_wakeup(dst_m);
1417         }
1418         VM_OBJECT_UNLOCK(dst_object);
1419         if (upgrade) {
1420                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1421                 vm_object_deallocate(src_object);
1422         }
1423 }
1424
1425
1426 /*
1427  * This routine checks around the requested page for other pages that
1428  * might be able to be faulted in.  This routine brackets the viable
1429  * pages for the pages to be paged in.
1430  *
1431  * Inputs:
1432  *      m, rbehind, rahead
1433  *
1434  * Outputs:
1435  *  marray (array of vm_page_t), reqpage (index of requested page)
1436  *
1437  * Return value:
1438  *  number of pages in marray
1439  */
1440 static int
1441 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1442         vm_page_t m;
1443         int rbehind;
1444         int rahead;
1445         vm_page_t *marray;
1446         int *reqpage;
1447 {
1448         int i,j;
1449         vm_object_t object;
1450         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1451         vm_page_t rtm;
1452         int cbehind, cahead;
1453
1454         VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1455
1456         object = m->object;
1457         pindex = m->pindex;
1458         cbehind = cahead = 0;
1459
1460         /*
1461          * if the requested page is not available, then give up now
1462          */
1463         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1464                 return 0;
1465         }
1466
1467         if ((cbehind == 0) && (cahead == 0)) {
1468                 *reqpage = 0;
1469                 marray[0] = m;
1470                 return 1;
1471         }
1472
1473         if (rahead > cahead) {
1474                 rahead = cahead;
1475         }
1476
1477         if (rbehind > cbehind) {
1478                 rbehind = cbehind;
1479         }
1480
1481         /*
1482          * scan backward for the read behind pages -- in memory
1483          */
1484         if (pindex > 0) {
1485                 if (rbehind > pindex) {
1486                         rbehind = pindex;
1487                         startpindex = 0;
1488                 } else {
1489                         startpindex = pindex - rbehind;
1490                 }
1491
1492                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1493                     rtm->pindex >= startpindex)
1494                         startpindex = rtm->pindex + 1;
1495
1496                 /* tpindex is unsigned; beware of numeric underflow. */
1497                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1498                     tpindex < pindex; i++, tpindex--) {
1499
1500                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1501                             VM_ALLOC_IFNOTCACHED);
1502                         if (rtm == NULL) {
1503                                 /*
1504                                  * Shift the allocated pages to the
1505                                  * beginning of the array.
1506                                  */
1507                                 for (j = 0; j < i; j++) {
1508                                         marray[j] = marray[j + tpindex + 1 -
1509                                             startpindex];
1510                                 }
1511                                 break;
1512                         }
1513
1514                         marray[tpindex - startpindex] = rtm;
1515                 }
1516         } else {
1517                 startpindex = 0;
1518                 i = 0;
1519         }
1520
1521         marray[i] = m;
1522         /* page offset of the required page */
1523         *reqpage = i;
1524
1525         tpindex = pindex + 1;
1526         i++;
1527
1528         /*
1529          * scan forward for the read ahead pages
1530          */
1531         endpindex = tpindex + rahead;
1532         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1533                 endpindex = rtm->pindex;
1534         if (endpindex > object->size)
1535                 endpindex = object->size;
1536
1537         for (; tpindex < endpindex; i++, tpindex++) {
1538
1539                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1540                     VM_ALLOC_IFNOTCACHED);
1541                 if (rtm == NULL) {
1542                         break;
1543                 }
1544
1545                 marray[i] = rtm;
1546         }
1547
1548         /* return number of pages */
1549         return i;
1550 }
1551
1552 /*
1553  * Block entry into the machine-independent layer's page fault handler by
1554  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1555  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1556  * spurious page faults.
1557  */
1558 int
1559 vm_fault_disable_pagefaults(void)
1560 {
1561
1562         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1563 }
1564
1565 void
1566 vm_fault_enable_pagefaults(int save)
1567 {
1568
1569         curthread_pflags_restore(save);
1570 }