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  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
  70  * $DragonFly: src/sys/vm/vm_fault.c,v 1.3 2003/06/25 03:56:12 dillon Exp $
  71  */
  72
  73 /*
  74  *      Page fault handling module.
  75  */
  76
  77 #include <sys/param.h>
  78 #include <sys/systm.h>
  79 #include <sys/proc.h>
  80 #include <sys/vnode.h>
  81 #include <sys/resourcevar.h>
  82 #include <sys/vmmeter.h>
  83
  84 #include <vm/vm.h>
  85 #include <vm/vm_param.h>
  86 #include <sys/lock.h>
  87 #include <vm/pmap.h>
  88 #include <vm/vm_map.h>
  89 #include <vm/vm_object.h>
  90 #include <vm/vm_page.h>
  91 #include <vm/vm_pageout.h>
  92 #include <vm/vm_kern.h>
  93 #include <vm/vm_pager.h>
  94 #include <vm/vnode_pager.h>
  95 #include <vm/vm_extern.h>
  96
  97 static int vm_fault_additional_pages __P((vm_page_t, int,
  98                                           int, vm_page_t *, int *));
  99
 100 #define VM_FAULT_READ_AHEAD 8
 101 #define VM_FAULT_READ_BEHIND 7
 102 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 103
 104 struct faultstate {
 105         vm_page_t m;
 106         vm_object_t object;
 107         vm_pindex_t pindex;
 108         vm_page_t first_m;
 109         vm_object_t     first_object;
 110         vm_pindex_t first_pindex;
 111         vm_map_t map;
 112         vm_map_entry_t entry;
 113         int lookup_still_valid;
 114         struct vnode *vp;
 115 };
 116
 117 static __inline void
 118 release_page(struct faultstate *fs)
 119 {
 120         vm_page_wakeup(fs->m);
 121         vm_page_deactivate(fs->m);
 122         fs->m = NULL;
 123 }
 124
 125 static __inline void
 126 unlock_map(struct faultstate *fs)
 127 {
 128         if (fs->lookup_still_valid) {
 129                 vm_map_lookup_done(fs->map, fs->entry);
 130                 fs->lookup_still_valid = FALSE;
 131         }
 132 }
 133
 134 static void
 135 _unlock_things(struct faultstate *fs, int dealloc)
 136 {
 137         vm_object_pip_wakeup(fs->object);
 138         if (fs->object != fs->first_object) {
 139                 vm_page_free(fs->first_m);
 140                 vm_object_pip_wakeup(fs->first_object);
 141                 fs->first_m = NULL;
 142         }
 143         if (dealloc) {
 144                 vm_object_deallocate(fs->first_object);
 145         }
 146         unlock_map(fs);
 147         if (fs->vp != NULL) {
 148                 vput(fs->vp);
 149                 fs->vp = NULL;
 150         }
 151 }
 152
 153 #define unlock_things(fs) _unlock_things(fs, 0)
 154 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 155
 156 /*
 157  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 158  *            current object *might* contain the page.
 159  *
 160  *            default objects are zero-fill, there is no real pager.
 161  */
 162
 163 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 164                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 165
 166 /*
 167  *      vm_fault:
 168  *
 169  *      Handle a page fault occurring at the given address,
 170  *      requiring the given permissions, in the map specified.
 171  *      If successful, the page is inserted into the
 172  *      associated physical map.
 173  *
 174  *      NOTE: the given address should be truncated to the
 175  *      proper page address.
 176  *
 177  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 178  *      a standard error specifying why the fault is fatal is returned.
 179  *
 180  *
 181  *      The map in question must be referenced, and remains so.
 182  *      Caller may hold no locks.
 183  */
 184 int
 185 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 186 {
 187         vm_prot_t prot;
 188         int result;
 189         boolean_t wired;
 190         int map_generation;
 191         vm_object_t next_object;
 192         vm_page_t marray[VM_FAULT_READ];
 193         int hardfault;
 194         int faultcount;
 195         struct faultstate fs;
 196
 197         cnt.v_vm_faults++;      /* needs lock XXX */
 198         hardfault = 0;
 199
 200 RetryFault:;
 201
 202         /*
 203          * Find the backing store object and offset into it to begin the
 204          * search.
 205          */
 206         fs.map = map;
 207         if ((result = vm_map_lookup(&fs.map, vaddr,
 208                 fault_type, &fs.entry, &fs.first_object,
 209                 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
 210                 if ((result != KERN_PROTECTION_FAILURE) ||
 211                         ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
 212                         return result;
 213                 }
 214
 215                 /*
 216                  * If we are user-wiring a r/w segment, and it is COW, then
 217                  * we need to do the COW operation.  Note that we don't COW
 218                  * currently RO sections now, because it is NOT desirable
 219                  * to COW .text.  We simply keep .text from ever being COW'ed
 220                  * and take the heat that one cannot debug wired .text sections.
 221                  */
 222                 result = vm_map_lookup(&fs.map, vaddr,
 223                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 224                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 225                 if (result != KERN_SUCCESS) {
 226                         return result;
 227                 }
 228
 229                 /*
 230                  * If we don't COW now, on a user wire, the user will never
 231                  * be able to write to the mapping.  If we don't make this
 232                  * restriction, the bookkeeping would be nearly impossible.
 233                  */
 234                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 235                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 236         }
 237
 238         map_generation = fs.map->timestamp;
 239
 240         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 241                 panic("vm_fault: fault on nofault entry, addr: %lx",
 242                     (u_long)vaddr);
 243         }
 244
 245         /*
 246          * Make a reference to this object to prevent its disposal while we
 247          * are messing with it.  Once we have the reference, the map is free
 248          * to be diddled.  Since objects reference their shadows (and copies),
 249          * they will stay around as well.
 250          *
 251          * Bump the paging-in-progress count to prevent size changes (e.g.
 252          * truncation operations) during I/O.  This must be done after
 253          * obtaining the vnode lock in order to avoid possible deadlocks.
 254          */
 255         vm_object_reference(fs.first_object);
 256         fs.vp = vnode_pager_lock(fs.first_object);
 257         vm_object_pip_add(fs.first_object, 1);
 258
 259         if ((fault_type & VM_PROT_WRITE) &&
 260                 (fs.first_object->type == OBJT_VNODE)) {
 261                 vm_freeze_copyopts(fs.first_object,
 262                         fs.first_pindex, fs.first_pindex + 1);
 263         }
 264
 265         fs.lookup_still_valid = TRUE;
 266
 267         if (wired)
 268                 fault_type = prot;
 269
 270         fs.first_m = NULL;
 271
 272         /*
 273          * Search for the page at object/offset.
 274          */
 275
 276         fs.object = fs.first_object;
 277         fs.pindex = fs.first_pindex;
 278
 279         while (TRUE) {
 280                 /*
 281                  * If the object is dead, we stop here
 282                  */
 283
 284                 if (fs.object->flags & OBJ_DEAD) {
 285                         unlock_and_deallocate(&fs);
 286                         return (KERN_PROTECTION_FAILURE);
 287                 }
 288
 289                 /*
 290                  * See if page is resident
 291                  */
 292
 293                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 294                 if (fs.m != NULL) {
 295                         int queue, s;
 296                         /*
 297                          * Wait/Retry if the page is busy.  We have to do this
 298                          * if the page is busy via either PG_BUSY or
 299                          * vm_page_t->busy because the vm_pager may be using
 300                          * vm_page_t->busy for pageouts ( and even pageins if
 301                          * it is the vnode pager ), and we could end up trying
 302                          * to pagein and pageout the same page simultaneously.
 303                          *
 304                          * We can theoretically allow the busy case on a read
 305                          * fault if the page is marked valid, but since such
 306                          * pages are typically already pmap'd, putting that
 307                          * special case in might be more effort then it is
 308                          * worth.  We cannot under any circumstances mess
 309                          * around with a vm_page_t->busy page except, perhaps,
 310                          * to pmap it.
 311                          */
 312                         if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
 313                                 unlock_things(&fs);
 314                                 (void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
 315                                 cnt.v_intrans++;
 316                                 vm_object_deallocate(fs.first_object);
 317                                 goto RetryFault;
 318                         }
 319
 320                         queue = fs.m->queue;
 321                         s = splvm();
 322                         vm_page_unqueue_nowakeup(fs.m);
 323                         splx(s);
 324
 325                         if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
 326                                 vm_page_activate(fs.m);
 327                                 unlock_and_deallocate(&fs);
 328                                 VM_WAITPFAULT;
 329                                 goto RetryFault;
 330                         }
 331
 332                         /*
 333                          * Mark page busy for other processes, and the
 334                          * pagedaemon.  If it still isn't completely valid
 335                          * (readable), jump to readrest, else break-out ( we
 336                          * found the page ).
 337                          */
 338
 339                         vm_page_busy(fs.m);
 340                         if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 341                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 342                                 goto readrest;
 343                         }
 344
 345                         break;
 346                 }
 347
 348                 /*
 349                  * Page is not resident, If this is the search termination
 350                  * or the pager might contain the page, allocate a new page.
 351                  */
 352
 353                 if (TRYPAGER || fs.object == fs.first_object) {
 354                         if (fs.pindex >= fs.object->size) {
 355                                 unlock_and_deallocate(&fs);
 356                                 return (KERN_PROTECTION_FAILURE);
 357                         }
 358
 359                         /*
 360                          * Allocate a new page for this object/offset pair.
 361                          */
 362                         fs.m = NULL;
 363                         if (!vm_page_count_severe()) {
 364                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 365                                     (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
 366                         }
 367                         if (fs.m == NULL) {
 368                                 unlock_and_deallocate(&fs);
 369                                 VM_WAITPFAULT;
 370                                 goto RetryFault;
 371                         }
 372                 }
 373
 374 readrest:
 375                 /*
 376                  * We have found a valid page or we have allocated a new page.
 377                  * The page thus may not be valid or may not be entirely
 378                  * valid.
 379                  *
 380                  * Attempt to fault-in the page if there is a chance that the
 381                  * pager has it, and potentially fault in additional pages
 382                  * at the same time.
 383                  */
 384
 385                 if (TRYPAGER) {
 386                         int rv;
 387                         int reqpage;
 388                         int ahead, behind;
 389                         u_char behavior = vm_map_entry_behavior(fs.entry);
 390
 391                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 392                                 ahead = 0;
 393                                 behind = 0;
 394                         } else {
 395                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 396                                 if (behind > VM_FAULT_READ_BEHIND)
 397                                         behind = VM_FAULT_READ_BEHIND;
 398
 399                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 400                                 if (ahead > VM_FAULT_READ_AHEAD)
 401                                         ahead = VM_FAULT_READ_AHEAD;
 402                         }
 403
 404                         if ((fs.first_object->type != OBJT_DEVICE) &&
 405                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 406                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 407                                 fs.pindex >= fs.entry->lastr &&
 408                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
 409                         ) {
 410                                 vm_pindex_t firstpindex, tmppindex;
 411
 412                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 413                                         firstpindex = 0;
 414                                 else
 415                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 416
 417                                 /*
 418                                  * note: partially valid pages cannot be
 419                                  * included in the lookahead - NFS piecemeal
 420                                  * writes will barf on it badly.
 421                                  */
 422
 423                                 for(tmppindex = fs.first_pindex - 1;
 424                                         tmppindex >= firstpindex;
 425                                         --tmppindex) {
 426                                         vm_page_t mt;
 427                                         mt = vm_page_lookup( fs.first_object, tmppindex);
 428                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 429                                                 break;
 430                                         if (mt->busy ||
 431                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
 432                                                 mt->hold_count ||
 433                                                 mt->wire_count)
 434                                                 continue;
 435                                         if (mt->dirty == 0)
 436                                                 vm_page_test_dirty(mt);
 437                                         if (mt->dirty) {
 438                                                 vm_page_protect(mt, VM_PROT_NONE);
 439                                                 vm_page_deactivate(mt);
 440                                         } else {
 441                                                 vm_page_cache(mt);
 442                                         }
 443                                 }
 444
 445                                 ahead += behind;
 446                                 behind = 0;
 447                         }
 448
 449                         /*
 450                          * now we find out if any other pages should be paged
 451                          * in at this time this routine checks to see if the
 452                          * pages surrounding this fault reside in the same
 453                          * object as the page for this fault.  If they do,
 454                          * then they are faulted in also into the object.  The
 455                          * array "marray" returned contains an array of
 456                          * vm_page_t structs where one of them is the
 457                          * vm_page_t passed to the routine.  The reqpage
 458                          * return value is the index into the marray for the
 459                          * vm_page_t passed to the routine.
 460                          *
 461                          * fs.m plus the additional pages are PG_BUSY'd.
 462                          */
 463                         faultcount = vm_fault_additional_pages(
 464                             fs.m, behind, ahead, marray, &reqpage);
 465
 466                         /*
 467                          * update lastr imperfectly (we do not know how much
 468                          * getpages will actually read), but good enough.
 469                          */
 470                         fs.entry->lastr = fs.pindex + faultcount - behind;
 471
 472                         /*
 473                          * Call the pager to retrieve the data, if any, after
 474                          * releasing the lock on the map.  We hold a ref on
 475                          * fs.object and the pages are PG_BUSY'd.
 476                          */
 477                         unlock_map(&fs);
 478
 479                         rv = faultcount ?
 480                             vm_pager_get_pages(fs.object, marray, faultcount,
 481                                 reqpage) : VM_PAGER_FAIL;
 482
 483                         if (rv == VM_PAGER_OK) {
 484                                 /*
 485                                  * Found the page. Leave it busy while we play
 486                                  * with it.
 487                                  */
 488
 489                                 /*
 490                                  * Relookup in case pager changed page. Pager
 491                                  * is responsible for disposition of old page
 492                                  * if moved.
 493                                  */
 494                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 495                                 if(!fs.m) {
 496                                         unlock_and_deallocate(&fs);
 497                                         goto RetryFault;
 498                                 }
 499
 500                                 hardfault++;
 501                                 break; /* break to PAGE HAS BEEN FOUND */
 502                         }
 503                         /*
 504                          * Remove the bogus page (which does not exist at this
 505                          * object/offset); before doing so, we must get back
 506                          * our object lock to preserve our invariant.
 507                          *
 508                          * Also wake up any other process that may want to bring
 509                          * in this page.
 510                          *
 511                          * If this is the top-level object, we must leave the
 512                          * busy page to prevent another process from rushing
 513                          * past us, and inserting the page in that object at
 514                          * the same time that we are.
 515                          */
 516
 517                         if (rv == VM_PAGER_ERROR)
 518                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 519                                     curproc->p_pid, curproc->p_comm);
 520                         /*
 521                          * Data outside the range of the pager or an I/O error
 522                          */
 523                         /*
 524                          * XXX - the check for kernel_map is a kludge to work
 525                          * around having the machine panic on a kernel space
 526                          * fault w/ I/O error.
 527                          */
 528                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 529                                 (rv == VM_PAGER_BAD)) {
 530                                 vm_page_free(fs.m);
 531                                 fs.m = NULL;
 532                                 unlock_and_deallocate(&fs);
 533                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 534                         }
 535                         if (fs.object != fs.first_object) {
 536                                 vm_page_free(fs.m);
 537                                 fs.m = NULL;
 538                                 /*
 539                                  * XXX - we cannot just fall out at this
 540                                  * point, m has been freed and is invalid!
 541                                  */
 542                         }
 543                 }
 544
 545                 /*
 546                  * We get here if the object has default pager (or unwiring)
 547                  * or the pager doesn't have the page.
 548                  */
 549                 if (fs.object == fs.first_object)
 550                         fs.first_m = fs.m;
 551
 552                 /*
 553                  * Move on to the next object.  Lock the next object before
 554                  * unlocking the current one.
 555                  */
 556
 557                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 558                 next_object = fs.object->backing_object;
 559                 if (next_object == NULL) {
 560                         /*
 561                          * If there's no object left, fill the page in the top
 562                          * object with zeros.
 563                          */
 564                         if (fs.object != fs.first_object) {
 565                                 vm_object_pip_wakeup(fs.object);
 566
 567                                 fs.object = fs.first_object;
 568                                 fs.pindex = fs.first_pindex;
 569                                 fs.m = fs.first_m;
 570                         }
 571                         fs.first_m = NULL;
 572
 573                         /*
 574                          * Zero the page if necessary and mark it valid.
 575                          */
 576                         if ((fs.m->flags & PG_ZERO) == 0) {
 577                                 vm_page_zero_fill(fs.m);
 578                         } else {
 579                                 cnt.v_ozfod++;
 580                         }
 581                         cnt.v_zfod++;
 582                         fs.m->valid = VM_PAGE_BITS_ALL;
 583                         break;  /* break to PAGE HAS BEEN FOUND */
 584                 } else {
 585                         if (fs.object != fs.first_object) {
 586                                 vm_object_pip_wakeup(fs.object);
 587                         }
 588                         KASSERT(fs.object != next_object, ("object loop %p", next_object));
 589                         fs.object = next_object;
 590                         vm_object_pip_add(fs.object, 1);
 591                 }
 592         }
 593
 594         KASSERT((fs.m->flags & PG_BUSY) != 0,
 595             ("vm_fault: not busy after main loop"));
 596
 597         /*
 598          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 599          * is held.]
 600          */
 601
 602         /*
 603          * If the page is being written, but isn't already owned by the
 604          * top-level object, we have to copy it into a new page owned by the
 605          * top-level object.
 606          */
 607
 608         if (fs.object != fs.first_object) {
 609                 /*
 610                  * We only really need to copy if we want to write it.
 611                  */
 612
 613                 if (fault_type & VM_PROT_WRITE) {
 614                         /*
 615                          * This allows pages to be virtually copied from a
 616                          * backing_object into the first_object, where the
 617                          * backing object has no other refs to it, and cannot
 618                          * gain any more refs.  Instead of a bcopy, we just
 619                          * move the page from the backing object to the
 620                          * first object.  Note that we must mark the page
 621                          * dirty in the first object so that it will go out
 622                          * to swap when needed.
 623                          */
 624                         if (map_generation == fs.map->timestamp &&
 625                                 /*
 626                                  * Only one shadow object
 627                                  */
 628                                 (fs.object->shadow_count == 1) &&
 629                                 /*
 630                                  * No COW refs, except us
 631                                  */
 632                                 (fs.object->ref_count == 1) &&
 633                                 /*
 634                                  * No one else can look this object up
 635                                  */
 636                                 (fs.object->handle == NULL) &&
 637                                 /*
 638                                  * No other ways to look the object up
 639                                  */
 640                                 ((fs.object->type == OBJT_DEFAULT) ||
 641                                  (fs.object->type == OBJT_SWAP)) &&
 642                                 /*
 643                                  * We don't chase down the shadow chain
 644                                  */
 645                                 (fs.object == fs.first_object->backing_object) &&
 646
 647                                 /*
 648                                  * grab the lock if we need to
 649                                  */
 650                                 (fs.lookup_still_valid ||
 651                                  lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curthread) == 0)
 652                             ) {
 653
 654                                 fs.lookup_still_valid = 1;
 655                                 /*
 656                                  * get rid of the unnecessary page
 657                                  */
 658                                 vm_page_protect(fs.first_m, VM_PROT_NONE);
 659                                 vm_page_free(fs.first_m);
 660                                 fs.first_m = NULL;
 661
 662                                 /*
 663                                  * grab the page and put it into the
 664                                  * process'es object.  The page is
 665                                  * automatically made dirty.
 666                                  */
 667                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 668                                 fs.first_m = fs.m;
 669                                 vm_page_busy(fs.first_m);
 670                                 fs.m = NULL;
 671                                 cnt.v_cow_optim++;
 672                         } else {
 673                                 /*
 674                                  * Oh, well, lets copy it.
 675                                  */
 676                                 vm_page_copy(fs.m, fs.first_m);
 677                         }
 678
 679                         if (fs.m) {
 680                                 /*
 681                                  * We no longer need the old page or object.
 682                                  */
 683                                 release_page(&fs);
 684                         }
 685
 686                         /*
 687                          * fs.object != fs.first_object due to above
 688                          * conditional
 689                          */
 690
 691                         vm_object_pip_wakeup(fs.object);
 692
 693                         /*
 694                          * Only use the new page below...
 695                          */
 696
 697                         cnt.v_cow_faults++;
 698                         fs.m = fs.first_m;
 699                         fs.object = fs.first_object;
 700                         fs.pindex = fs.first_pindex;
 701
 702                 } else {
 703                         prot &= ~VM_PROT_WRITE;
 704                 }
 705         }
 706
 707         /*
 708          * We must verify that the maps have not changed since our last
 709          * lookup.
 710          */
 711
 712         if (!fs.lookup_still_valid &&
 713                 (fs.map->timestamp != map_generation)) {
 714                 vm_object_t retry_object;
 715                 vm_pindex_t retry_pindex;
 716                 vm_prot_t retry_prot;
 717
 718                 /*
 719                  * Since map entries may be pageable, make sure we can take a
 720                  * page fault on them.
 721                  */
 722
 723                 /*
 724                  * Unlock vnode before the lookup to avoid deadlock.   E.G.
 725                  * avoid a deadlock between the inode and exec_map that can
 726                  * occur due to locks being obtained in different orders.
 727                  */
 728
 729                 if (fs.vp != NULL) {
 730                         vput(fs.vp);
 731                         fs.vp = NULL;
 732                 }
 733
 734                 if (fs.map->infork) {
 735                         release_page(&fs);
 736                         unlock_and_deallocate(&fs);
 737                         goto RetryFault;
 738                 }
 739
 740                 /*
 741                  * To avoid trying to write_lock the map while another process
 742                  * has it read_locked (in vm_map_pageable), we do not try for
 743                  * write permission.  If the page is still writable, we will
 744                  * get write permission.  If it is not, or has been marked
 745                  * needs_copy, we enter the mapping without write permission,
 746                  * and will merely take another fault.
 747                  */
 748                 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
 749                     &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 750                 map_generation = fs.map->timestamp;
 751
 752                 /*
 753                  * If we don't need the page any longer, put it on the active
 754                  * list (the easiest thing to do here).  If no one needs it,
 755                  * pageout will grab it eventually.
 756                  */
 757
 758                 if (result != KERN_SUCCESS) {
 759                         release_page(&fs);
 760                         unlock_and_deallocate(&fs);
 761                         return (result);
 762                 }
 763                 fs.lookup_still_valid = TRUE;
 764
 765                 if ((retry_object != fs.first_object) ||
 766                     (retry_pindex != fs.first_pindex)) {
 767                         release_page(&fs);
 768                         unlock_and_deallocate(&fs);
 769                         goto RetryFault;
 770                 }
 771                 /*
 772                  * Check whether the protection has changed or the object has
 773                  * been copied while we left the map unlocked. Changing from
 774                  * read to write permission is OK - we leave the page
 775                  * write-protected, and catch the write fault. Changing from
 776                  * write to read permission means that we can't mark the page
 777                  * write-enabled after all.
 778                  */
 779                 prot &= retry_prot;
 780         }
 781
 782         /*
 783          * Put this page into the physical map. We had to do the unlock above
 784          * because pmap_enter may cause other faults.   We don't put the page
 785          * back on the active queue until later so that the page-out daemon
 786          * won't find us (yet).
 787          */
 788
 789         if (prot & VM_PROT_WRITE) {
 790                 vm_page_flag_set(fs.m, PG_WRITEABLE);
 791                 vm_object_set_writeable_dirty(fs.m->object);
 792
 793                 /*
 794                  * If the fault is a write, we know that this page is being
 795                  * written NOW so dirty it explicitly to save on
 796                  * pmap_is_modified() calls later.
 797                  *
 798                  * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
 799                  * if the page is already dirty to prevent data written with
 800                  * the expectation of being synced from not being synced.
 801                  * Likewise if this entry does not request NOSYNC then make
 802                  * sure the page isn't marked NOSYNC.  Applications sharing
 803                  * data should use the same flags to avoid ping ponging.
 804                  *
 805                  * Also tell the backing pager, if any, that it should remove
 806                  * any swap backing since the page is now dirty.
 807                  */
 808                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 809                         if (fs.m->dirty == 0)
 810                                 vm_page_flag_set(fs.m, PG_NOSYNC);
 811                 } else {
 812                         vm_page_flag_clear(fs.m, PG_NOSYNC);
 813                 }
 814                 if (fault_flags & VM_FAULT_DIRTY) {
 815                         int s;
 816                         vm_page_dirty(fs.m);
 817                         s = splvm();
 818                         vm_pager_page_unswapped(fs.m);
 819                         splx(s);
 820                 }
 821         }
 822
 823         /*
 824          * Page had better still be busy
 825          */
 826
 827         KASSERT(fs.m->flags & PG_BUSY,
 828                 ("vm_fault: page %p not busy!", fs.m));
 829
 830         unlock_things(&fs);
 831
 832         /*
 833          * Sanity check: page must be completely valid or it is not fit to
 834          * map into user space.  vm_pager_get_pages() ensures this.
 835          */
 836
 837         if (fs.m->valid != VM_PAGE_BITS_ALL) {
 838                 vm_page_zero_invalid(fs.m, TRUE);
 839                 printf("Warning: page %p partially invalid on fault\n", fs.m);
 840         }
 841
 842         pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
 843
 844         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 845                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 846         }
 847
 848         vm_page_flag_clear(fs.m, PG_ZERO);
 849         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 850         if (fault_flags & VM_FAULT_HOLD)
 851                 vm_page_hold(fs.m);
 852
 853         /*
 854          * If the page is not wired down, then put it where the pageout daemon
 855          * can find it.
 856          */
 857
 858         if (fault_flags & VM_FAULT_WIRE_MASK) {
 859                 if (wired)
 860                         vm_page_wire(fs.m);
 861                 else
 862                         vm_page_unwire(fs.m, 1);
 863         } else {
 864                 vm_page_activate(fs.m);
 865         }
 866
 867         if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) {
 868                 if (hardfault) {
 869                         curproc->p_stats->p_ru.ru_majflt++;
 870                 } else {
 871                         curproc->p_stats->p_ru.ru_minflt++;
 872                 }
 873         }
 874
 875         /*
 876          * Unlock everything, and return
 877          */
 878
 879         vm_page_wakeup(fs.m);
 880         vm_object_deallocate(fs.first_object);
 881
 882         return (KERN_SUCCESS);
 883
 884 }
 885
 886 /*
 887  *      vm_fault_wire:
 888  *
 889  *      Wire down a range of virtual addresses in a map.
 890  */
 891 int
 892 vm_fault_wire(map, start, end)
 893         vm_map_t map;
 894         vm_offset_t start, end;
 895 {
 896
 897         register vm_offset_t va;
 898         register pmap_t pmap;
 899         int rv;
 900
 901         pmap = vm_map_pmap(map);
 902
 903         /*
 904          * Inform the physical mapping system that the range of addresses may
 905          * not fault, so that page tables and such can be locked down as well.
 906          */
 907
 908         pmap_pageable(pmap, start, end, FALSE);
 909
 910         /*
 911          * We simulate a fault to get the page and enter it in the physical
 912          * map.
 913          */
 914
 915         for (va = start; va < end; va += PAGE_SIZE) {
 916                 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
 917                         VM_FAULT_CHANGE_WIRING);
 918                 if (rv) {
 919                         if (va != start)
 920                                 vm_fault_unwire(map, start, va);
 921                         return (rv);
 922                 }
 923         }
 924         return (KERN_SUCCESS);
 925 }
 926
 927 /*
 928  *      vm_fault_user_wire:
 929  *
 930  *      Wire down a range of virtual addresses in a map.  This
 931  *      is for user mode though, so we only ask for read access
 932  *      on currently read only sections.
 933  */
 934 int
 935 vm_fault_user_wire(map, start, end)
 936         vm_map_t map;
 937         vm_offset_t start, end;
 938 {
 939
 940         register vm_offset_t va;
 941         register pmap_t pmap;
 942         int rv;
 943
 944         pmap = vm_map_pmap(map);
 945
 946         /*
 947          * Inform the physical mapping system that the range of addresses may
 948          * not fault, so that page tables and such can be locked down as well.
 949          */
 950
 951         pmap_pageable(pmap, start, end, FALSE);
 952
 953         /*
 954          * We simulate a fault to get the page and enter it in the physical
 955          * map.
 956          */
 957         for (va = start; va < end; va += PAGE_SIZE) {
 958                 rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
 959                 if (rv) {
 960                         if (va != start)
 961                                 vm_fault_unwire(map, start, va);
 962                         return (rv);
 963                 }
 964         }
 965         return (KERN_SUCCESS);
 966 }
 967
 968
 969 /*
 970  *      vm_fault_unwire:
 971  *
 972  *      Unwire a range of virtual addresses in a map.
 973  */
 974 void
 975 vm_fault_unwire(map, start, end)
 976         vm_map_t map;
 977         vm_offset_t start, end;
 978 {
 979
 980         register vm_offset_t va, pa;
 981         register pmap_t pmap;
 982
 983         pmap = vm_map_pmap(map);
 984
 985         /*
 986          * Since the pages are wired down, we must be able to get their
 987          * mappings from the physical map system.
 988          */
 989
 990         for (va = start; va < end; va += PAGE_SIZE) {
 991                 pa = pmap_extract(pmap, va);
 992                 if (pa != (vm_offset_t) 0) {
 993                         pmap_change_wiring(pmap, va, FALSE);
 994                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
 995                 }
 996         }
 997
 998         /*
 999          * Inform the physical mapping system that the range of addresses may
1000          * fault, so that page tables and such may be unwired themselves.
1001          */
1002
1003         pmap_pageable(pmap, start, end, TRUE);
1004
1005 }
1006
1007 /*
1008  *      Routine:
1009  *              vm_fault_copy_entry
1010  *      Function:
1011  *              Copy all of the pages from a wired-down map entry to another.
1012  *
1013  *      In/out conditions:
1014  *              The source and destination maps must be locked for write.
1015  *              The source map entry must be wired down (or be a sharing map
1016  *              entry corresponding to a main map entry that is wired down).
1017  */
1018
1019 void
1020 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1021         vm_map_t dst_map;
1022         vm_map_t src_map;
1023         vm_map_entry_t dst_entry;
1024         vm_map_entry_t src_entry;
1025 {
1026         vm_object_t dst_object;
1027         vm_object_t src_object;
1028         vm_ooffset_t dst_offset;
1029         vm_ooffset_t src_offset;
1030         vm_prot_t prot;
1031         vm_offset_t vaddr;
1032         vm_page_t dst_m;
1033         vm_page_t src_m;
1034
1035 #ifdef  lint
1036         src_map++;
1037 #endif  /* lint */
1038
1039         src_object = src_entry->object.vm_object;
1040         src_offset = src_entry->offset;
1041
1042         /*
1043          * Create the top-level object for the destination entry. (Doesn't
1044          * actually shadow anything - we copy the pages directly.)
1045          */
1046         dst_object = vm_object_allocate(OBJT_DEFAULT,
1047             (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1048
1049         dst_entry->object.vm_object = dst_object;
1050         dst_entry->offset = 0;
1051
1052         prot = dst_entry->max_protection;
1053
1054         /*
1055          * Loop through all of the pages in the entry's range, copying each
1056          * one from the source object (it should be there) to the destination
1057          * object.
1058          */
1059         for (vaddr = dst_entry->start, dst_offset = 0;
1060             vaddr < dst_entry->end;
1061             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1062
1063                 /*
1064                  * Allocate a page in the destination object
1065                  */
1066                 do {
1067                         dst_m = vm_page_alloc(dst_object,
1068                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1069                         if (dst_m == NULL) {
1070                                 VM_WAIT;
1071                         }
1072                 } while (dst_m == NULL);
1073
1074                 /*
1075                  * Find the page in the source object, and copy it in.
1076                  * (Because the source is wired down, the page will be in
1077                  * memory.)
1078                  */
1079                 src_m = vm_page_lookup(src_object,
1080                         OFF_TO_IDX(dst_offset + src_offset));
1081                 if (src_m == NULL)
1082                         panic("vm_fault_copy_wired: page missing");
1083
1084                 vm_page_copy(src_m, dst_m);
1085
1086                 /*
1087                  * Enter it in the pmap...
1088                  */
1089
1090                 vm_page_flag_clear(dst_m, PG_ZERO);
1091                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1092                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1093
1094                 /*
1095                  * Mark it no longer busy, and put it on the active list.
1096                  */
1097                 vm_page_activate(dst_m);
1098                 vm_page_wakeup(dst_m);
1099         }
1100 }
1101
1102
1103 /*
1104  * This routine checks around the requested page for other pages that
1105  * might be able to be faulted in.  This routine brackets the viable
1106  * pages for the pages to be paged in.
1107  *
1108  * Inputs:
1109  *      m, rbehind, rahead
1110  *
1111  * Outputs:
1112  *  marray (array of vm_page_t), reqpage (index of requested page)
1113  *
1114  * Return value:
1115  *  number of pages in marray
1116  */
1117 static int
1118 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1119         vm_page_t m;
1120         int rbehind;
1121         int rahead;
1122         vm_page_t *marray;
1123         int *reqpage;
1124 {
1125         int i,j;
1126         vm_object_t object;
1127         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1128         vm_page_t rtm;
1129         int cbehind, cahead;
1130
1131         object = m->object;
1132         pindex = m->pindex;
1133
1134         /*
1135          * we don't fault-ahead for device pager
1136          */
1137         if (object->type == OBJT_DEVICE) {
1138                 *reqpage = 0;
1139                 marray[0] = m;
1140                 return 1;
1141         }
1142
1143         /*
1144          * if the requested page is not available, then give up now
1145          */
1146
1147         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1148                 return 0;
1149         }
1150
1151         if ((cbehind == 0) && (cahead == 0)) {
1152                 *reqpage = 0;
1153                 marray[0] = m;
1154                 return 1;
1155         }
1156
1157         if (rahead > cahead) {
1158                 rahead = cahead;
1159         }
1160
1161         if (rbehind > cbehind) {
1162                 rbehind = cbehind;
1163         }
1164
1165         /*
1166          * try to do any readahead that we might have free pages for.
1167          */
1168         if ((rahead + rbehind) >
1169                 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1170                 pagedaemon_wakeup();
1171                 marray[0] = m;
1172                 *reqpage = 0;
1173                 return 1;
1174         }
1175
1176         /*
1177          * scan backward for the read behind pages -- in memory
1178          */
1179         if (pindex > 0) {
1180                 if (rbehind > pindex) {
1181                         rbehind = pindex;
1182                         startpindex = 0;
1183                 } else {
1184                         startpindex = pindex - rbehind;
1185                 }
1186
1187                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1188                         if (vm_page_lookup( object, tpindex)) {
1189                                 startpindex = tpindex + 1;
1190                                 break;
1191                         }
1192                         if (tpindex == 0)
1193                                 break;
1194                 }
1195
1196                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1197
1198                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1199                         if (rtm == NULL) {
1200                                 for (j = 0; j < i; j++) {
1201                                         vm_page_free(marray[j]);
1202                                 }
1203                                 marray[0] = m;
1204                                 *reqpage = 0;
1205                                 return 1;
1206                         }
1207
1208                         marray[i] = rtm;
1209                 }
1210         } else {
1211                 startpindex = 0;
1212                 i = 0;
1213         }
1214
1215         marray[i] = m;
1216         /* page offset of the required page */
1217         *reqpage = i;
1218
1219         tpindex = pindex + 1;
1220         i++;
1221
1222         /*
1223          * scan forward for the read ahead pages
1224          */
1225         endpindex = tpindex + rahead;
1226         if (endpindex > object->size)
1227                 endpindex = object->size;
1228
1229         for( ; tpindex < endpindex; i++, tpindex++) {
1230
1231                 if (vm_page_lookup(object, tpindex)) {
1232                         break;
1233                 }
1234
1235                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1236                 if (rtm == NULL) {
1237                         break;
1238                 }
1239
1240                 marray[i] = rtm;
1241         }
1242
1243         /* return number of bytes of pages */
1244         return i;
1245 }