sys/vm/vm_fault.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  *
   9  *
  10  * This code is derived from software contributed to Berkeley by
  11  * The Mach Operating System project at Carnegie-Mellon University.
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  * 1. Redistributions of source code must retain the above copyright
  17  *    notice, this list of conditions and the following disclaimer.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. All advertising materials mentioning features or use of this software
  22  *    must display the following acknowledgement:
  23  *      This product includes software developed by the University of
  24  *      California, Berkeley and its contributors.
  25  * 4. Neither the name of the University nor the names of its contributors
  26  *    may be used to endorse or promote products derived from this software
  27  *    without specific prior written permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  39  * SUCH DAMAGE.
  40  *
  41  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  42  *
  43  *
  44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  45  * All rights reserved.
  46  *
  47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  48  *
  49  * Permission to use, copy, modify and distribute this software and
  50  * its documentation is hereby granted, provided that both the copyright
  51  * notice and this permission notice appear in all copies of the
  52  * software, derivative works or modified versions, and any portions
  53  * thereof, and that both notices appear in supporting documentation.
  54  *
  55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  58  *
  59  * Carnegie Mellon requests users of this software to return to
  60  *
  61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  62  *  School of Computer Science
  63  *  Carnegie Mellon University
  64  *  Pittsburgh PA 15213-3890
  65  *
  66  * any improvements or extensions that they make and grant Carnegie the
  67  * rights to redistribute these changes.
  68  */
  69
  70 /*
  71  *      Page fault handling module.
  72  */
  73
  74 #include <sys/cdefs.h>
  75 __FBSDID("$FreeBSD$");
  76
  77 #include "opt_ktrace.h"
  78 #include "opt_vm.h"
  79
  80 #include <sys/param.h>
  81 #include <sys/systm.h>
  82 #include <sys/kernel.h>
  83 #include <sys/lock.h>
  84 #include <sys/proc.h>
  85 #include <sys/resourcevar.h>
  86 #include <sys/rwlock.h>
  87 #include <sys/sysctl.h>
  88 #include <sys/vmmeter.h>
  89 #include <sys/vnode.h>
  90 #ifdef KTRACE
  91 #include <sys/ktrace.h>
  92 #endif
  93
  94 #include <vm/vm.h>
  95 #include <vm/vm_param.h>
  96 #include <vm/pmap.h>
  97 #include <vm/vm_map.h>
  98 #include <vm/vm_object.h>
  99 #include <vm/vm_page.h>
 100 #include <vm/vm_pageout.h>
 101 #include <vm/vm_kern.h>
 102 #include <vm/vm_pager.h>
 103 #include <vm/vm_extern.h>
 104
 105 #define PFBAK 4
 106 #define PFFOR 4
 107 #define PAGEORDER_SIZE (PFBAK+PFFOR)
 108
 109 static int prefault_pageorder[] = {
 110         -1 * PAGE_SIZE, 1 * PAGE_SIZE,
 111         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
 112         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
 113         -4 * PAGE_SIZE, 4 * PAGE_SIZE
 114 };
 115
 116 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
 117 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
 118
 119 #define VM_FAULT_READ_BEHIND    8
 120 #define VM_FAULT_READ_MAX       (1 + VM_FAULT_READ_AHEAD_MAX)
 121 #define VM_FAULT_NINCR          (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
 122 #define VM_FAULT_SUM            (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
 123 #define VM_FAULT_CACHE_BEHIND   (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
 124
 125 struct faultstate {
 126         vm_page_t m;
 127         vm_object_t object;
 128         vm_pindex_t pindex;
 129         vm_page_t first_m;
 130         vm_object_t     first_object;
 131         vm_pindex_t first_pindex;
 132         vm_map_t map;
 133         vm_map_entry_t entry;
 134         int lookup_still_valid;
 135         struct vnode *vp;
 136 };
 137
 138 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
 139
 140 static inline void
 141 release_page(struct faultstate *fs)
 142 {
 143
 144         vm_page_wakeup(fs->m);
 145         vm_page_lock(fs->m);
 146         vm_page_deactivate(fs->m);
 147         vm_page_unlock(fs->m);
 148         fs->m = NULL;
 149 }
 150
 151 static inline void
 152 unlock_map(struct faultstate *fs)
 153 {
 154
 155         if (fs->lookup_still_valid) {
 156                 vm_map_lookup_done(fs->map, fs->entry);
 157                 fs->lookup_still_valid = FALSE;
 158         }
 159 }
 160
 161 static void
 162 unlock_and_deallocate(struct faultstate *fs)
 163 {
 164
 165         vm_object_pip_wakeup(fs->object);
 166         VM_OBJECT_WUNLOCK(fs->object);
 167         if (fs->object != fs->first_object) {
 168                 VM_OBJECT_WLOCK(fs->first_object);
 169                 vm_page_lock(fs->first_m);
 170                 vm_page_free(fs->first_m);
 171                 vm_page_unlock(fs->first_m);
 172                 vm_object_pip_wakeup(fs->first_object);
 173                 VM_OBJECT_WUNLOCK(fs->first_object);
 174                 fs->first_m = NULL;
 175         }
 176         vm_object_deallocate(fs->first_object);
 177         unlock_map(fs);
 178         if (fs->vp != NULL) {
 179                 vput(fs->vp);
 180                 fs->vp = NULL;
 181         }
 182 }
 183
 184 /*
 185  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 186  *            current object *might* contain the page.
 187  *
 188  *            default objects are zero-fill, there is no real pager.
 189  */
 190 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 191                         ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
 192
 193 /*
 194  *      vm_fault:
 195  *
 196  *      Handle a page fault occurring at the given address,
 197  *      requiring the given permissions, in the map specified.
 198  *      If successful, the page is inserted into the
 199  *      associated physical map.
 200  *
 201  *      NOTE: the given address should be truncated to the
 202  *      proper page address.
 203  *
 204  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 205  *      a standard error specifying why the fault is fatal is returned.
 206  *
 207  *      The map in question must be referenced, and remains so.
 208  *      Caller may hold no locks.
 209  */
 210 int
 211 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 212     int fault_flags)
 213 {
 214         struct thread *td;
 215         int result;
 216
 217         td = curthread;
 218         if ((td->td_pflags & TDP_NOFAULTING) != 0)
 219                 return (KERN_PROTECTION_FAILURE);
 220 #ifdef KTRACE
 221         if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
 222                 ktrfault(vaddr, fault_type);
 223 #endif
 224         result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
 225             NULL);
 226 #ifdef KTRACE
 227         if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
 228                 ktrfaultend(result);
 229 #endif
 230         return (result);
 231 }
 232
 233 int
 234 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
 235     int fault_flags, vm_page_t *m_hold)
 236 {
 237         vm_prot_t prot;
 238         long ahead, behind;
 239         int alloc_req, era, faultcount, nera, reqpage, result;
 240         boolean_t growstack, is_first_object_locked, wired;
 241         int map_generation;
 242         vm_object_t next_object;
 243         vm_page_t marray[VM_FAULT_READ_MAX];
 244         int hardfault;
 245         struct faultstate fs;
 246         struct vnode *vp;
 247         int locked, error;
 248
 249         hardfault = 0;
 250         growstack = TRUE;
 251         PCPU_INC(cnt.v_vm_faults);
 252         fs.vp = NULL;
 253         faultcount = reqpage = 0;
 254
 255 RetryFault:;
 256
 257         /*
 258          * Find the backing store object and offset into it to begin the
 259          * search.
 260          */
 261         fs.map = map;
 262         result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
 263             &fs.first_object, &fs.first_pindex, &prot, &wired);
 264         if (result != KERN_SUCCESS) {
 265                 if (growstack && result == KERN_INVALID_ADDRESS &&
 266                     map != kernel_map) {
 267                         result = vm_map_growstack(curproc, vaddr);
 268                         if (result != KERN_SUCCESS)
 269                                 return (KERN_FAILURE);
 270                         growstack = FALSE;
 271                         goto RetryFault;
 272                 }
 273                 return (result);
 274         }
 275
 276         map_generation = fs.map->timestamp;
 277
 278         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 279                 panic("vm_fault: fault on nofault entry, addr: %lx",
 280                     (u_long)vaddr);
 281         }
 282
 283         /*
 284          * Make a reference to this object to prevent its disposal while we
 285          * are messing with it.  Once we have the reference, the map is free
 286          * to be diddled.  Since objects reference their shadows (and copies),
 287          * they will stay around as well.
 288          *
 289          * Bump the paging-in-progress count to prevent size changes (e.g.
 290          * truncation operations) during I/O.  This must be done after
 291          * obtaining the vnode lock in order to avoid possible deadlocks.
 292          */
 293         VM_OBJECT_WLOCK(fs.first_object);
 294         vm_object_reference_locked(fs.first_object);
 295         vm_object_pip_add(fs.first_object, 1);
 296
 297         fs.lookup_still_valid = TRUE;
 298
 299         if (wired)
 300                 fault_type = prot | (fault_type & VM_PROT_COPY);
 301
 302         fs.first_m = NULL;
 303
 304         /*
 305          * Search for the page at object/offset.
 306          */
 307         fs.object = fs.first_object;
 308         fs.pindex = fs.first_pindex;
 309         while (TRUE) {
 310                 /*
 311                  * If the object is dead, we stop here
 312                  */
 313                 if (fs.object->flags & OBJ_DEAD) {
 314                         unlock_and_deallocate(&fs);
 315                         return (KERN_PROTECTION_FAILURE);
 316                 }
 317
 318                 /*
 319                  * See if page is resident
 320                  */
 321                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 322                 if (fs.m != NULL) {
 323                         /*
 324                          * check for page-based copy on write.
 325                          * We check fs.object == fs.first_object so
 326                          * as to ensure the legacy COW mechanism is
 327                          * used when the page in question is part of
 328                          * a shadow object.  Otherwise, vm_page_cowfault()
 329                          * removes the page from the backing object,
 330                          * which is not what we want.
 331                          */
 332                         vm_page_lock(fs.m);
 333                         if ((fs.m->cow) &&
 334                             (fault_type & VM_PROT_WRITE) &&
 335                             (fs.object == fs.first_object)) {
 336                                 vm_page_cowfault(fs.m);
 337                                 unlock_and_deallocate(&fs);
 338                                 goto RetryFault;
 339                         }
 340
 341                         /*
 342                          * Wait/Retry if the page is busy.  We have to do this
 343                          * if the page is busy via either VPO_BUSY or
 344                          * vm_page_t->busy because the vm_pager may be using
 345                          * vm_page_t->busy for pageouts ( and even pageins if
 346                          * it is the vnode pager ), and we could end up trying
 347                          * to pagein and pageout the same page simultaneously.
 348                          *
 349                          * We can theoretically allow the busy case on a read
 350                          * fault if the page is marked valid, but since such
 351                          * pages are typically already pmap'd, putting that
 352                          * special case in might be more effort then it is
 353                          * worth.  We cannot under any circumstances mess
 354                          * around with a vm_page_t->busy page except, perhaps,
 355                          * to pmap it.
 356                          */
 357                         if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
 358                                 /*
 359                                  * Reference the page before unlocking and
 360                                  * sleeping so that the page daemon is less
 361                                  * likely to reclaim it.
 362                                  */
 363                                 vm_page_aflag_set(fs.m, PGA_REFERENCED);
 364                                 vm_page_unlock(fs.m);
 365                                 if (fs.object != fs.first_object) {
 366                                         if (!VM_OBJECT_TRYWLOCK(
 367                                             fs.first_object)) {
 368                                                 VM_OBJECT_WUNLOCK(fs.object);
 369                                                 VM_OBJECT_WLOCK(fs.first_object);
 370                                                 VM_OBJECT_WLOCK(fs.object);
 371                                         }
 372                                         vm_page_lock(fs.first_m);
 373                                         vm_page_free(fs.first_m);
 374                                         vm_page_unlock(fs.first_m);
 375                                         vm_object_pip_wakeup(fs.first_object);
 376                                         VM_OBJECT_WUNLOCK(fs.first_object);
 377                                         fs.first_m = NULL;
 378                                 }
 379                                 unlock_map(&fs);
 380                                 if (fs.m == vm_page_lookup(fs.object,
 381                                     fs.pindex)) {
 382                                         vm_page_sleep_if_busy(fs.m, TRUE,
 383                                             "vmpfw");
 384                                 }
 385                                 vm_object_pip_wakeup(fs.object);
 386                                 VM_OBJECT_WUNLOCK(fs.object);
 387                                 PCPU_INC(cnt.v_intrans);
 388                                 vm_object_deallocate(fs.first_object);
 389                                 goto RetryFault;
 390                         }
 391                         vm_page_remque(fs.m);
 392                         vm_page_unlock(fs.m);
 393
 394                         /*
 395                          * Mark page busy for other processes, and the
 396                          * pagedaemon.  If it still isn't completely valid
 397                          * (readable), jump to readrest, else break-out ( we
 398                          * found the page ).
 399                          */
 400                         vm_page_busy(fs.m);
 401                         if (fs.m->valid != VM_PAGE_BITS_ALL)
 402                                 goto readrest;
 403                         break;
 404                 }
 405
 406                 /*
 407                  * Page is not resident, If this is the search termination
 408                  * or the pager might contain the page, allocate a new page.
 409                  */
 410                 if (TRYPAGER || fs.object == fs.first_object) {
 411                         if (fs.pindex >= fs.object->size) {
 412                                 unlock_and_deallocate(&fs);
 413                                 return (KERN_PROTECTION_FAILURE);
 414                         }
 415
 416                         /*
 417                          * Allocate a new page for this object/offset pair.
 418                          *
 419                          * Unlocked read of the p_flag is harmless. At
 420                          * worst, the P_KILLED might be not observed
 421                          * there, and allocation can fail, causing
 422                          * restart and new reading of the p_flag.
 423                          */
 424                         fs.m = NULL;
 425                         if (!vm_page_count_severe() || P_KILLED(curproc)) {
 426 #if VM_NRESERVLEVEL > 0
 427                                 if ((fs.object->flags & OBJ_COLORED) == 0) {
 428                                         fs.object->flags |= OBJ_COLORED;
 429                                         fs.object->pg_color = atop(vaddr) -
 430                                             fs.pindex;
 431                                 }
 432 #endif
 433                                 alloc_req = P_KILLED(curproc) ?
 434                                     VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
 435                                 if (fs.object->type != OBJT_VNODE &&
 436                                     fs.object->backing_object == NULL)
 437                                         alloc_req |= VM_ALLOC_ZERO;
 438                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 439                                     alloc_req);
 440                         }
 441                         if (fs.m == NULL) {
 442                                 unlock_and_deallocate(&fs);
 443                                 VM_WAITPFAULT;
 444                                 goto RetryFault;
 445                         } else if (fs.m->valid == VM_PAGE_BITS_ALL)
 446                                 break;
 447                 }
 448
 449 readrest:
 450                 /*
 451                  * We have found a valid page or we have allocated a new page.
 452                  * The page thus may not be valid or may not be entirely
 453                  * valid.
 454                  *
 455                  * Attempt to fault-in the page if there is a chance that the
 456                  * pager has it, and potentially fault in additional pages
 457                  * at the same time.
 458                  */
 459                 if (TRYPAGER) {
 460                         int rv;
 461                         u_char behavior = vm_map_entry_behavior(fs.entry);
 462
 463                         if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
 464                             P_KILLED(curproc)) {
 465                                 behind = 0;
 466                                 ahead = 0;
 467                         } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
 468                                 behind = 0;
 469                                 ahead = atop(fs.entry->end - vaddr) - 1;
 470                                 if (ahead > VM_FAULT_READ_AHEAD_MAX)
 471                                         ahead = VM_FAULT_READ_AHEAD_MAX;
 472                                 if (fs.pindex == fs.entry->next_read)
 473                                         vm_fault_cache_behind(&fs,
 474                                             VM_FAULT_READ_MAX);
 475                         } else {
 476                                 /*
 477                                  * If this is a sequential page fault, then
 478                                  * arithmetically increase the number of pages
 479                                  * in the read-ahead window.  Otherwise, reset
 480                                  * the read-ahead window to its smallest size.
 481                                  */
 482                                 behind = atop(vaddr - fs.entry->start);
 483                                 if (behind > VM_FAULT_READ_BEHIND)
 484                                         behind = VM_FAULT_READ_BEHIND;
 485                                 ahead = atop(fs.entry->end - vaddr) - 1;
 486                                 era = fs.entry->read_ahead;
 487                                 if (fs.pindex == fs.entry->next_read) {
 488                                         nera = era + behind;
 489                                         if (nera > VM_FAULT_READ_AHEAD_MAX)
 490                                                 nera = VM_FAULT_READ_AHEAD_MAX;
 491                                         behind = 0;
 492                                         if (ahead > nera)
 493                                                 ahead = nera;
 494                                         if (era == VM_FAULT_READ_AHEAD_MAX)
 495                                                 vm_fault_cache_behind(&fs,
 496                                                     VM_FAULT_CACHE_BEHIND);
 497                                 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
 498                                         ahead = VM_FAULT_READ_AHEAD_MIN;
 499                                 if (era != ahead)
 500                                         fs.entry->read_ahead = ahead;
 501                         }
 502
 503                         /*
 504                          * Call the pager to retrieve the data, if any, after
 505                          * releasing the lock on the map.  We hold a ref on
 506                          * fs.object and the pages are VPO_BUSY'd.
 507                          */
 508                         unlock_map(&fs);
 509
 510                         if (fs.object->type == OBJT_VNODE) {
 511                                 vp = fs.object->handle;
 512                                 if (vp == fs.vp)
 513                                         goto vnode_locked;
 514                                 else if (fs.vp != NULL) {
 515                                         vput(fs.vp);
 516                                         fs.vp = NULL;
 517                                 }
 518                                 locked = VOP_ISLOCKED(vp);
 519
 520                                 if (locked != LK_EXCLUSIVE)
 521                                         locked = LK_SHARED;
 522                                 /* Do not sleep for vnode lock while fs.m is busy */
 523                                 error = vget(vp, locked | LK_CANRECURSE |
 524                                     LK_NOWAIT, curthread);
 525                                 if (error != 0) {
 526                                         vhold(vp);
 527                                         release_page(&fs);
 528                                         unlock_and_deallocate(&fs);
 529                                         error = vget(vp, locked | LK_RETRY |
 530                                             LK_CANRECURSE, curthread);
 531                                         vdrop(vp);
 532                                         fs.vp = vp;
 533                                         KASSERT(error == 0,
 534                                             ("vm_fault: vget failed"));
 535                                         goto RetryFault;
 536                                 }
 537                                 fs.vp = vp;
 538                         }
 539 vnode_locked:
 540                         KASSERT(fs.vp == NULL || !fs.map->system_map,
 541                             ("vm_fault: vnode-backed object mapped by system map"));
 542
 543                         /*
 544                          * now we find out if any other pages should be paged
 545                          * in at this time this routine checks to see if the
 546                          * pages surrounding this fault reside in the same
 547                          * object as the page for this fault.  If they do,
 548                          * then they are faulted in also into the object.  The
 549                          * array "marray" returned contains an array of
 550                          * vm_page_t structs where one of them is the
 551                          * vm_page_t passed to the routine.  The reqpage
 552                          * return value is the index into the marray for the
 553                          * vm_page_t passed to the routine.
 554                          *
 555                          * fs.m plus the additional pages are VPO_BUSY'd.
 556                          */
 557                         faultcount = vm_fault_additional_pages(
 558                             fs.m, behind, ahead, marray, &reqpage);
 559
 560                         rv = faultcount ?
 561                             vm_pager_get_pages(fs.object, marray, faultcount,
 562                                 reqpage) : VM_PAGER_FAIL;
 563
 564                         if (rv == VM_PAGER_OK) {
 565                                 /*
 566                                  * Found the page. Leave it busy while we play
 567                                  * with it.
 568                                  */
 569
 570                                 /*
 571                                  * Relookup in case pager changed page. Pager
 572                                  * is responsible for disposition of old page
 573                                  * if moved.
 574                                  */
 575                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 576                                 if (!fs.m) {
 577                                         unlock_and_deallocate(&fs);
 578                                         goto RetryFault;
 579                                 }
 580
 581                                 hardfault++;
 582                                 break; /* break to PAGE HAS BEEN FOUND */
 583                         }
 584                         /*
 585                          * Remove the bogus page (which does not exist at this
 586                          * object/offset); before doing so, we must get back
 587                          * our object lock to preserve our invariant.
 588                          *
 589                          * Also wake up any other process that may want to bring
 590                          * in this page.
 591                          *
 592                          * If this is the top-level object, we must leave the
 593                          * busy page to prevent another process from rushing
 594                          * past us, and inserting the page in that object at
 595                          * the same time that we are.
 596                          */
 597                         if (rv == VM_PAGER_ERROR)
 598                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 599                                     curproc->p_pid, curproc->p_comm);
 600                         /*
 601                          * Data outside the range of the pager or an I/O error
 602                          */
 603                         /*
 604                          * XXX - the check for kernel_map is a kludge to work
 605                          * around having the machine panic on a kernel space
 606                          * fault w/ I/O error.
 607                          */
 608                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 609                                 (rv == VM_PAGER_BAD)) {
 610                                 vm_page_lock(fs.m);
 611                                 vm_page_free(fs.m);
 612                                 vm_page_unlock(fs.m);
 613                                 fs.m = NULL;
 614                                 unlock_and_deallocate(&fs);
 615                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 616                         }
 617                         if (fs.object != fs.first_object) {
 618                                 vm_page_lock(fs.m);
 619                                 vm_page_free(fs.m);
 620                                 vm_page_unlock(fs.m);
 621                                 fs.m = NULL;
 622                                 /*
 623                                  * XXX - we cannot just fall out at this
 624                                  * point, m has been freed and is invalid!
 625                                  */
 626                         }
 627                 }
 628
 629                 /*
 630                  * We get here if the object has default pager (or unwiring)
 631                  * or the pager doesn't have the page.
 632                  */
 633                 if (fs.object == fs.first_object)
 634                         fs.first_m = fs.m;
 635
 636                 /*
 637                  * Move on to the next object.  Lock the next object before
 638                  * unlocking the current one.
 639                  */
 640                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 641                 next_object = fs.object->backing_object;
 642                 if (next_object == NULL) {
 643                         /*
 644                          * If there's no object left, fill the page in the top
 645                          * object with zeros.
 646                          */
 647                         if (fs.object != fs.first_object) {
 648                                 vm_object_pip_wakeup(fs.object);
 649                                 VM_OBJECT_WUNLOCK(fs.object);
 650
 651                                 fs.object = fs.first_object;
 652                                 fs.pindex = fs.first_pindex;
 653                                 fs.m = fs.first_m;
 654                                 VM_OBJECT_WLOCK(fs.object);
 655                         }
 656                         fs.first_m = NULL;
 657
 658                         /*
 659                          * Zero the page if necessary and mark it valid.
 660                          */
 661                         if ((fs.m->flags & PG_ZERO) == 0) {
 662                                 pmap_zero_page(fs.m);
 663                         } else {
 664                                 PCPU_INC(cnt.v_ozfod);
 665                         }
 666                         PCPU_INC(cnt.v_zfod);
 667                         fs.m->valid = VM_PAGE_BITS_ALL;
 668                         break;  /* break to PAGE HAS BEEN FOUND */
 669                 } else {
 670                         KASSERT(fs.object != next_object,
 671                             ("object loop %p", next_object));
 672                         VM_OBJECT_WLOCK(next_object);
 673                         vm_object_pip_add(next_object, 1);
 674                         if (fs.object != fs.first_object)
 675                                 vm_object_pip_wakeup(fs.object);
 676                         VM_OBJECT_WUNLOCK(fs.object);
 677                         fs.object = next_object;
 678                 }
 679         }
 680
 681         KASSERT((fs.m->oflags & VPO_BUSY) != 0,
 682             ("vm_fault: not busy after main loop"));
 683
 684         /*
 685          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 686          * is held.]
 687          */
 688
 689         /*
 690          * If the page is being written, but isn't already owned by the
 691          * top-level object, we have to copy it into a new page owned by the
 692          * top-level object.
 693          */
 694         if (fs.object != fs.first_object) {
 695                 /*
 696                  * We only really need to copy if we want to write it.
 697                  */
 698                 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
 699                         /*
 700                          * This allows pages to be virtually copied from a
 701                          * backing_object into the first_object, where the
 702                          * backing object has no other refs to it, and cannot
 703                          * gain any more refs.  Instead of a bcopy, we just
 704                          * move the page from the backing object to the
 705                          * first object.  Note that we must mark the page
 706                          * dirty in the first object so that it will go out
 707                          * to swap when needed.
 708                          */
 709                         is_first_object_locked = FALSE;
 710                         if (
 711                                 /*
 712                                  * Only one shadow object
 713                                  */
 714                                 (fs.object->shadow_count == 1) &&
 715                                 /*
 716                                  * No COW refs, except us
 717                                  */
 718                                 (fs.object->ref_count == 1) &&
 719                                 /*
 720                                  * No one else can look this object up
 721                                  */
 722                                 (fs.object->handle == NULL) &&
 723                                 /*
 724                                  * No other ways to look the object up
 725                                  */
 726                                 ((fs.object->type == OBJT_DEFAULT) ||
 727                                  (fs.object->type == OBJT_SWAP)) &&
 728                             (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
 729                                 /*
 730                                  * We don't chase down the shadow chain
 731                                  */
 732                             fs.object == fs.first_object->backing_object) {
 733                                 /*
 734                                  * get rid of the unnecessary page
 735                                  */
 736                                 vm_page_lock(fs.first_m);
 737                                 vm_page_free(fs.first_m);
 738                                 vm_page_unlock(fs.first_m);
 739                                 /*
 740                                  * grab the page and put it into the
 741                                  * process'es object.  The page is
 742                                  * automatically made dirty.
 743                                  */
 744                                 vm_page_lock(fs.m);
 745                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 746                                 vm_page_unlock(fs.m);
 747                                 vm_page_busy(fs.m);
 748                                 fs.first_m = fs.m;
 749                                 fs.m = NULL;
 750                                 PCPU_INC(cnt.v_cow_optim);
 751                         } else {
 752                                 /*
 753                                  * Oh, well, lets copy it.
 754                                  */
 755                                 pmap_copy_page(fs.m, fs.first_m);
 756                                 fs.first_m->valid = VM_PAGE_BITS_ALL;
 757                                 if (wired && (fault_flags &
 758                                     VM_FAULT_CHANGE_WIRING) == 0) {
 759                                         vm_page_lock(fs.first_m);
 760                                         vm_page_wire(fs.first_m);
 761                                         vm_page_unlock(fs.first_m);
 762
 763                                         vm_page_lock(fs.m);
 764                                         vm_page_unwire(fs.m, FALSE);
 765                                         vm_page_unlock(fs.m);
 766                                 }
 767                                 /*
 768                                  * We no longer need the old page or object.
 769                                  */
 770                                 release_page(&fs);
 771                         }
 772                         /*
 773                          * fs.object != fs.first_object due to above
 774                          * conditional
 775                          */
 776                         vm_object_pip_wakeup(fs.object);
 777                         VM_OBJECT_WUNLOCK(fs.object);
 778                         /*
 779                          * Only use the new page below...
 780                          */
 781                         fs.object = fs.first_object;
 782                         fs.pindex = fs.first_pindex;
 783                         fs.m = fs.first_m;
 784                         if (!is_first_object_locked)
 785                                 VM_OBJECT_WLOCK(fs.object);
 786                         PCPU_INC(cnt.v_cow_faults);
 787                         curthread->td_cow++;
 788                 } else {
 789                         prot &= ~VM_PROT_WRITE;
 790                 }
 791         }
 792
 793         /*
 794          * We must verify that the maps have not changed since our last
 795          * lookup.
 796          */
 797         if (!fs.lookup_still_valid) {
 798                 vm_object_t retry_object;
 799                 vm_pindex_t retry_pindex;
 800                 vm_prot_t retry_prot;
 801
 802                 if (!vm_map_trylock_read(fs.map)) {
 803                         release_page(&fs);
 804                         unlock_and_deallocate(&fs);
 805                         goto RetryFault;
 806                 }
 807                 fs.lookup_still_valid = TRUE;
 808                 if (fs.map->timestamp != map_generation) {
 809                         result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
 810                             &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 811
 812                         /*
 813                          * If we don't need the page any longer, put it on the inactive
 814                          * list (the easiest thing to do here).  If no one needs it,
 815                          * pageout will grab it eventually.
 816                          */
 817                         if (result != KERN_SUCCESS) {
 818                                 release_page(&fs);
 819                                 unlock_and_deallocate(&fs);
 820
 821                                 /*
 822                                  * If retry of map lookup would have blocked then
 823                                  * retry fault from start.
 824                                  */
 825                                 if (result == KERN_FAILURE)
 826                                         goto RetryFault;
 827                                 return (result);
 828                         }
 829                         if ((retry_object != fs.first_object) ||
 830                             (retry_pindex != fs.first_pindex)) {
 831                                 release_page(&fs);
 832                                 unlock_and_deallocate(&fs);
 833                                 goto RetryFault;
 834                         }
 835
 836                         /*
 837                          * Check whether the protection has changed or the object has
 838                          * been copied while we left the map unlocked. Changing from
 839                          * read to write permission is OK - we leave the page
 840                          * write-protected, and catch the write fault. Changing from
 841                          * write to read permission means that we can't mark the page
 842                          * write-enabled after all.
 843                          */
 844                         prot &= retry_prot;
 845                 }
 846         }
 847         /*
 848          * If the page was filled by a pager, update the map entry's
 849          * last read offset.  Since the pager does not return the
 850          * actual set of pages that it read, this update is based on
 851          * the requested set.  Typically, the requested and actual
 852          * sets are the same.
 853          *
 854          * XXX The following assignment modifies the map
 855          * without holding a write lock on it.
 856          */
 857         if (hardfault)
 858                 fs.entry->next_read = fs.pindex + faultcount - reqpage;
 859
 860         if ((prot & VM_PROT_WRITE) != 0 ||
 861             (fault_flags & VM_FAULT_DIRTY) != 0) {
 862                 vm_object_set_writeable_dirty(fs.object);
 863
 864                 /*
 865                  * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
 866                  * if the page is already dirty to prevent data written with
 867                  * the expectation of being synced from not being synced.
 868                  * Likewise if this entry does not request NOSYNC then make
 869                  * sure the page isn't marked NOSYNC.  Applications sharing
 870                  * data should use the same flags to avoid ping ponging.
 871                  */
 872                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 873                         if (fs.m->dirty == 0)
 874                                 fs.m->oflags |= VPO_NOSYNC;
 875                 } else {
 876                         fs.m->oflags &= ~VPO_NOSYNC;
 877                 }
 878
 879                 /*
 880                  * If the fault is a write, we know that this page is being
 881                  * written NOW so dirty it explicitly to save on
 882                  * pmap_is_modified() calls later.
 883                  *
 884                  * Also tell the backing pager, if any, that it should remove
 885                  * any swap backing since the page is now dirty.
 886                  */
 887                 if (((fault_type & VM_PROT_WRITE) != 0 &&
 888                     (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
 889                     (fault_flags & VM_FAULT_DIRTY) != 0) {
 890                         vm_page_dirty(fs.m);
 891                         vm_pager_page_unswapped(fs.m);
 892                 }
 893         }
 894
 895         /*
 896          * Page had better still be busy
 897          */
 898         KASSERT(fs.m->oflags & VPO_BUSY,
 899                 ("vm_fault: page %p not busy!", fs.m));
 900         /*
 901          * Page must be completely valid or it is not fit to
 902          * map into user space.  vm_pager_get_pages() ensures this.
 903          */
 904         KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
 905             ("vm_fault: page %p partially invalid", fs.m));
 906         VM_OBJECT_WUNLOCK(fs.object);
 907
 908         /*
 909          * Put this page into the physical map.  We had to do the unlock above
 910          * because pmap_enter() may sleep.  We don't put the page
 911          * back on the active queue until later so that the pageout daemon
 912          * won't find it (yet).
 913          */
 914         pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
 915         if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
 916                 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
 917         VM_OBJECT_WLOCK(fs.object);
 918         vm_page_lock(fs.m);
 919
 920         /*
 921          * If the page is not wired down, then put it where the pageout daemon
 922          * can find it.
 923          */
 924         if (fault_flags & VM_FAULT_CHANGE_WIRING) {
 925                 if (wired)
 926                         vm_page_wire(fs.m);
 927                 else
 928                         vm_page_unwire(fs.m, 1);
 929         } else
 930                 vm_page_activate(fs.m);
 931         if (m_hold != NULL) {
 932                 *m_hold = fs.m;
 933                 vm_page_hold(fs.m);
 934         }
 935         vm_page_unlock(fs.m);
 936         vm_page_wakeup(fs.m);
 937
 938         /*
 939          * Unlock everything, and return
 940          */
 941         unlock_and_deallocate(&fs);
 942         if (hardfault) {
 943                 PCPU_INC(cnt.v_io_faults);
 944                 curthread->td_ru.ru_majflt++;
 945         } else
 946                 curthread->td_ru.ru_minflt++;
 947
 948         return (KERN_SUCCESS);
 949 }
 950
 951 /*
 952  * Speed up the reclamation of up to "distance" pages that precede the
 953  * faulting pindex within the first object of the shadow chain.
 954  */
 955 static void
 956 vm_fault_cache_behind(const struct faultstate *fs, int distance)
 957 {
 958         vm_object_t first_object, object;
 959         vm_page_t m, m_prev;
 960         vm_pindex_t pindex;
 961
 962         object = fs->object;
 963         VM_OBJECT_ASSERT_WLOCKED(object);
 964         first_object = fs->first_object;
 965         if (first_object != object) {
 966                 if (!VM_OBJECT_TRYWLOCK(first_object)) {
 967                         VM_OBJECT_WUNLOCK(object);
 968                         VM_OBJECT_WLOCK(first_object);
 969                         VM_OBJECT_WLOCK(object);
 970                 }
 971         }
 972         /* Neither fictitious nor unmanaged pages can be cached. */
 973         if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
 974                 if (fs->first_pindex < distance)
 975                         pindex = 0;
 976                 else
 977                         pindex = fs->first_pindex - distance;
 978                 if (pindex < OFF_TO_IDX(fs->entry->offset))
 979                         pindex = OFF_TO_IDX(fs->entry->offset);
 980                 m = first_object != object ? fs->first_m : fs->m;
 981                 KASSERT((m->oflags & VPO_BUSY) != 0,
 982                     ("vm_fault_cache_behind: page %p is not busy", m));
 983                 m_prev = vm_page_prev(m);
 984                 while ((m = m_prev) != NULL && m->pindex >= pindex &&
 985                     m->valid == VM_PAGE_BITS_ALL) {
 986                         m_prev = vm_page_prev(m);
 987                         if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0)
 988                                 continue;
 989                         vm_page_lock(m);
 990                         if (m->hold_count == 0 && m->wire_count == 0) {
 991                                 pmap_remove_all(m);
 992                                 vm_page_aflag_clear(m, PGA_REFERENCED);
 993                                 if (m->dirty != 0)
 994                                         vm_page_deactivate(m);
 995                                 else
 996                                         vm_page_cache(m);
 997                         }
 998                         vm_page_unlock(m);
 999                 }
1000         }
1001         if (first_object != object)
1002                 VM_OBJECT_WUNLOCK(first_object);
1003 }
1004
1005 /*
1006  * vm_fault_prefault provides a quick way of clustering
1007  * pagefaults into a processes address space.  It is a "cousin"
1008  * of vm_map_pmap_enter, except it runs at page fault time instead
1009  * of mmap time.
1010  */
1011 static void
1012 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1013 {
1014         int i;
1015         vm_offset_t addr, starta;
1016         vm_pindex_t pindex;
1017         vm_page_t m;
1018         vm_object_t object;
1019
1020         if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1021                 return;
1022
1023         object = entry->object.vm_object;
1024
1025         starta = addra - PFBAK * PAGE_SIZE;
1026         if (starta < entry->start) {
1027                 starta = entry->start;
1028         } else if (starta > addra) {
1029                 starta = 0;
1030         }
1031
1032         for (i = 0; i < PAGEORDER_SIZE; i++) {
1033                 vm_object_t backing_object, lobject;
1034
1035                 addr = addra + prefault_pageorder[i];
1036                 if (addr > addra + (PFFOR * PAGE_SIZE))
1037                         addr = 0;
1038
1039                 if (addr < starta || addr >= entry->end)
1040                         continue;
1041
1042                 if (!pmap_is_prefaultable(pmap, addr))
1043                         continue;
1044
1045                 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1046                 lobject = object;
1047                 VM_OBJECT_RLOCK(lobject);
1048                 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1049                     lobject->type == OBJT_DEFAULT &&
1050                     (backing_object = lobject->backing_object) != NULL) {
1051                         KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1052                             0, ("vm_fault_prefault: unaligned object offset"));
1053                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1054                         VM_OBJECT_RLOCK(backing_object);
1055                         VM_OBJECT_RUNLOCK(lobject);
1056                         lobject = backing_object;
1057                 }
1058                 /*
1059                  * give-up when a page is not in memory
1060                  */
1061                 if (m == NULL) {
1062                         VM_OBJECT_RUNLOCK(lobject);
1063                         break;
1064                 }
1065                 if (m->valid == VM_PAGE_BITS_ALL &&
1066                     (m->flags & PG_FICTITIOUS) == 0)
1067                         pmap_enter_quick(pmap, addr, m, entry->protection);
1068                 VM_OBJECT_RUNLOCK(lobject);
1069         }
1070 }
1071
1072 /*
1073  * Hold each of the physical pages that are mapped by the specified range of
1074  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1075  * and allow the specified types of access, "prot".  If all of the implied
1076  * pages are successfully held, then the number of held pages is returned
1077  * together with pointers to those pages in the array "ma".  However, if any
1078  * of the pages cannot be held, -1 is returned.
1079  */
1080 int
1081 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1082     vm_prot_t prot, vm_page_t *ma, int max_count)
1083 {
1084         vm_offset_t end, va;
1085         vm_page_t *mp;
1086         int count;
1087         boolean_t pmap_failed;
1088
1089         if (len == 0)
1090                 return (0);
1091         end = round_page(addr + len);
1092         addr = trunc_page(addr);
1093
1094         /*
1095          * Check for illegal addresses.
1096          */
1097         if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1098                 return (-1);
1099
1100         count = howmany(end - addr, PAGE_SIZE);
1101         if (count > max_count)
1102                 panic("vm_fault_quick_hold_pages: count > max_count");
1103
1104         /*
1105          * Most likely, the physical pages are resident in the pmap, so it is
1106          * faster to try pmap_extract_and_hold() first.
1107          */
1108         pmap_failed = FALSE;
1109         for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1110                 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1111                 if (*mp == NULL)
1112                         pmap_failed = TRUE;
1113                 else if ((prot & VM_PROT_WRITE) != 0 &&
1114                     (*mp)->dirty != VM_PAGE_BITS_ALL) {
1115                         /*
1116                          * Explicitly dirty the physical page.  Otherwise, the
1117                          * caller's changes may go unnoticed because they are
1118                          * performed through an unmanaged mapping or by a DMA
1119                          * operation.
1120                          *
1121                          * The object lock is not held here.
1122                          * See vm_page_clear_dirty_mask().
1123                          */
1124                         vm_page_dirty(*mp);
1125                 }
1126         }
1127         if (pmap_failed) {
1128                 /*
1129                  * One or more pages could not be held by the pmap.  Either no
1130                  * page was mapped at the specified virtual address or that
1131                  * mapping had insufficient permissions.  Attempt to fault in
1132                  * and hold these pages.
1133                  */
1134                 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1135                         if (*mp == NULL && vm_fault_hold(map, va, prot,
1136                             VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1137                                 goto error;
1138         }
1139         return (count);
1140 error:
1141         for (mp = ma; mp < ma + count; mp++)
1142                 if (*mp != NULL) {
1143                         vm_page_lock(*mp);
1144                         vm_page_unhold(*mp);
1145                         vm_page_unlock(*mp);
1146                 }
1147         return (-1);
1148 }
1149
1150 /*
1151  *      vm_fault_wire:
1152  *
1153  *      Wire down a range of virtual addresses in a map.
1154  */
1155 int
1156 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1157     boolean_t fictitious)
1158 {
1159         vm_offset_t va;
1160         int rv;
1161
1162         /*
1163          * We simulate a fault to get the page and enter it in the physical
1164          * map.  For user wiring, we only ask for read access on currently
1165          * read-only sections.
1166          */
1167         for (va = start; va < end; va += PAGE_SIZE) {
1168                 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1169                 if (rv) {
1170                         if (va != start)
1171                                 vm_fault_unwire(map, start, va, fictitious);
1172                         return (rv);
1173                 }
1174         }
1175         return (KERN_SUCCESS);
1176 }
1177
1178 /*
1179  *      vm_fault_unwire:
1180  *
1181  *      Unwire a range of virtual addresses in a map.
1182  */
1183 void
1184 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1185     boolean_t fictitious)
1186 {
1187         vm_paddr_t pa;
1188         vm_offset_t va;
1189         vm_page_t m;
1190         pmap_t pmap;
1191
1192         pmap = vm_map_pmap(map);
1193
1194         /*
1195          * Since the pages are wired down, we must be able to get their
1196          * mappings from the physical map system.
1197          */
1198         for (va = start; va < end; va += PAGE_SIZE) {
1199                 pa = pmap_extract(pmap, va);
1200                 if (pa != 0) {
1201                         pmap_change_wiring(pmap, va, FALSE);
1202                         if (!fictitious) {
1203                                 m = PHYS_TO_VM_PAGE(pa);
1204                                 vm_page_lock(m);
1205                                 vm_page_unwire(m, TRUE);
1206                                 vm_page_unlock(m);
1207                         }
1208                 }
1209         }
1210 }
1211
1212 /*
1213  *      Routine:
1214  *              vm_fault_copy_entry
1215  *      Function:
1216  *              Create new shadow object backing dst_entry with private copy of
1217  *              all underlying pages. When src_entry is equal to dst_entry,
1218  *              function implements COW for wired-down map entry. Otherwise,
1219  *              it forks wired entry into dst_map.
1220  *
1221  *      In/out conditions:
1222  *              The source and destination maps must be locked for write.
1223  *              The source map entry must be wired down (or be a sharing map
1224  *              entry corresponding to a main map entry that is wired down).
1225  */
1226 void
1227 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1228     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1229     vm_ooffset_t *fork_charge)
1230 {
1231         vm_object_t backing_object, dst_object, object, src_object;
1232         vm_pindex_t dst_pindex, pindex, src_pindex;
1233         vm_prot_t access, prot;
1234         vm_offset_t vaddr;
1235         vm_page_t dst_m;
1236         vm_page_t src_m;
1237         boolean_t src_readonly, upgrade;
1238
1239 #ifdef  lint
1240         src_map++;
1241 #endif  /* lint */
1242
1243         upgrade = src_entry == dst_entry;
1244
1245         src_object = src_entry->object.vm_object;
1246         src_pindex = OFF_TO_IDX(src_entry->offset);
1247         src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1248
1249         /*
1250          * Create the top-level object for the destination entry. (Doesn't
1251          * actually shadow anything - we copy the pages directly.)
1252          */
1253         dst_object = vm_object_allocate(OBJT_DEFAULT,
1254             OFF_TO_IDX(dst_entry->end - dst_entry->start));
1255 #if VM_NRESERVLEVEL > 0
1256         dst_object->flags |= OBJ_COLORED;
1257         dst_object->pg_color = atop(dst_entry->start);
1258 #endif
1259
1260         VM_OBJECT_WLOCK(dst_object);
1261         KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1262             ("vm_fault_copy_entry: vm_object not NULL"));
1263         dst_entry->object.vm_object = dst_object;
1264         dst_entry->offset = 0;
1265         dst_object->charge = dst_entry->end - dst_entry->start;
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 {
1273                 dst_object->cred = dst_entry->cred;
1274                 dst_entry->cred = NULL;
1275         }
1276         access = prot = dst_entry->protection;
1277         /*
1278          * If not an upgrade, then enter the mappings in the pmap as
1279          * read and/or execute accesses.  Otherwise, enter them as
1280          * write accesses.
1281          *
1282          * A writeable large page mapping is only created if all of
1283          * the constituent small page mappings are modified. Marking
1284          * PTEs as modified on inception allows promotion to happen
1285          * without taking potentially large number of soft faults.
1286          */
1287         if (!upgrade)
1288                 access &= ~VM_PROT_WRITE;
1289
1290         /*
1291          * Loop through all of the virtual pages within the entry's
1292          * range, copying each page from the source object to the
1293          * destination object.  Since the source is wired, those pages
1294          * must exist.  In contrast, the destination is pageable.
1295          * Since the destination object does share any backing storage
1296          * with the source object, all of its pages must be dirtied,
1297          * regardless of whether they can be written.
1298          */
1299         for (vaddr = dst_entry->start, dst_pindex = 0;
1300             vaddr < dst_entry->end;
1301             vaddr += PAGE_SIZE, dst_pindex++) {
1302
1303                 /*
1304                  * Allocate a page in the destination object.
1305                  */
1306                 do {
1307                         dst_m = vm_page_alloc(dst_object, dst_pindex,
1308                             VM_ALLOC_NORMAL);
1309                         if (dst_m == NULL) {
1310                                 VM_OBJECT_WUNLOCK(dst_object);
1311                                 VM_WAIT;
1312                                 VM_OBJECT_WLOCK(dst_object);
1313                         }
1314                 } while (dst_m == NULL);
1315
1316                 /*
1317                  * Find the page in the source object, and copy it in.
1318                  * (Because the source is wired down, the page will be in
1319                  * memory.)
1320                  */
1321                 VM_OBJECT_RLOCK(src_object);
1322                 object = src_object;
1323                 pindex = src_pindex + dst_pindex;
1324                 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1325                     src_readonly &&
1326                     (backing_object = object->backing_object) != NULL) {
1327                         /*
1328                          * Allow fallback to backing objects if we are reading.
1329                          */
1330                         VM_OBJECT_RLOCK(backing_object);
1331                         pindex += OFF_TO_IDX(object->backing_object_offset);
1332                         VM_OBJECT_RUNLOCK(object);
1333                         object = backing_object;
1334                 }
1335                 if (src_m == NULL)
1336                         panic("vm_fault_copy_wired: page missing");
1337                 pmap_copy_page(src_m, dst_m);
1338                 VM_OBJECT_RUNLOCK(object);
1339                 dst_m->valid = VM_PAGE_BITS_ALL;
1340                 dst_m->dirty = VM_PAGE_BITS_ALL;
1341                 VM_OBJECT_WUNLOCK(dst_object);
1342
1343                 /*
1344                  * Enter it in the pmap. If a wired, copy-on-write
1345                  * mapping is being replaced by a write-enabled
1346                  * mapping, then wire that new mapping.
1347                  */
1348                 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1349
1350                 /*
1351                  * Mark it no longer busy, and put it on the active list.
1352                  */
1353                 VM_OBJECT_WLOCK(dst_object);
1354
1355                 if (upgrade) {
1356                         vm_page_lock(src_m);
1357                         vm_page_unwire(src_m, 0);
1358                         vm_page_unlock(src_m);
1359
1360                         vm_page_lock(dst_m);
1361                         vm_page_wire(dst_m);
1362                         vm_page_unlock(dst_m);
1363                 } else {
1364                         vm_page_lock(dst_m);
1365                         vm_page_activate(dst_m);
1366                         vm_page_unlock(dst_m);
1367                 }
1368                 vm_page_wakeup(dst_m);
1369         }
1370         VM_OBJECT_WUNLOCK(dst_object);
1371         if (upgrade) {
1372                 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1373                 vm_object_deallocate(src_object);
1374         }
1375 }
1376
1377
1378 /*
1379  * This routine checks around the requested page for other pages that
1380  * might be able to be faulted in.  This routine brackets the viable
1381  * pages for the pages to be paged in.
1382  *
1383  * Inputs:
1384  *      m, rbehind, rahead
1385  *
1386  * Outputs:
1387  *  marray (array of vm_page_t), reqpage (index of requested page)
1388  *
1389  * Return value:
1390  *  number of pages in marray
1391  */
1392 static int
1393 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1394         vm_page_t m;
1395         int rbehind;
1396         int rahead;
1397         vm_page_t *marray;
1398         int *reqpage;
1399 {
1400         int i,j;
1401         vm_object_t object;
1402         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1403         vm_page_t rtm;
1404         int cbehind, cahead;
1405
1406         VM_OBJECT_ASSERT_WLOCKED(m->object);
1407
1408         object = m->object;
1409         pindex = m->pindex;
1410         cbehind = cahead = 0;
1411
1412         /*
1413          * if the requested page is not available, then give up now
1414          */
1415         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1416                 return 0;
1417         }
1418
1419         if ((cbehind == 0) && (cahead == 0)) {
1420                 *reqpage = 0;
1421                 marray[0] = m;
1422                 return 1;
1423         }
1424
1425         if (rahead > cahead) {
1426                 rahead = cahead;
1427         }
1428
1429         if (rbehind > cbehind) {
1430                 rbehind = cbehind;
1431         }
1432
1433         /*
1434          * scan backward for the read behind pages -- in memory
1435          */
1436         if (pindex > 0) {
1437                 if (rbehind > pindex) {
1438                         rbehind = pindex;
1439                         startpindex = 0;
1440                 } else {
1441                         startpindex = pindex - rbehind;
1442                 }
1443
1444                 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1445                     rtm->pindex >= startpindex)
1446                         startpindex = rtm->pindex + 1;
1447
1448                 /* tpindex is unsigned; beware of numeric underflow. */
1449                 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1450                     tpindex < pindex; i++, tpindex--) {
1451
1452                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1453                             VM_ALLOC_IFNOTCACHED);
1454                         if (rtm == NULL) {
1455                                 /*
1456                                  * Shift the allocated pages to the
1457                                  * beginning of the array.
1458                                  */
1459                                 for (j = 0; j < i; j++) {
1460                                         marray[j] = marray[j + tpindex + 1 -
1461                                             startpindex];
1462                                 }
1463                                 break;
1464                         }
1465
1466                         marray[tpindex - startpindex] = rtm;
1467                 }
1468         } else {
1469                 startpindex = 0;
1470                 i = 0;
1471         }
1472
1473         marray[i] = m;
1474         /* page offset of the required page */
1475         *reqpage = i;
1476
1477         tpindex = pindex + 1;
1478         i++;
1479
1480         /*
1481          * scan forward for the read ahead pages
1482          */
1483         endpindex = tpindex + rahead;
1484         if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1485                 endpindex = rtm->pindex;
1486         if (endpindex > object->size)
1487                 endpindex = object->size;
1488
1489         for (; tpindex < endpindex; i++, tpindex++) {
1490
1491                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1492                     VM_ALLOC_IFNOTCACHED);
1493                 if (rtm == NULL) {
1494                         break;
1495                 }
1496
1497                 marray[i] = rtm;
1498         }
1499
1500         /* return number of pages */
1501         return i;
1502 }
1503
1504 /*
1505  * Block entry into the machine-independent layer's page fault handler by
1506  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1507  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1508  * spurious page faults.
1509  */
1510 int
1511 vm_fault_disable_pagefaults(void)
1512 {
1513
1514         return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1515 }
1516
1517 void
1518 vm_fault_enable_pagefaults(int save)
1519 {
1520
1521         curthread_pflags_restore(save);
1522 }