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  * $Id: vm_fault.c,v 1.87 1998/08/24 08:39:37 dfr Exp $
  70  */
  71
  72 /*
  73  *      Page fault handling module.
  74  */
  75
  76 #include <sys/param.h>
  77 #include <sys/systm.h>
  78 #include <sys/proc.h>
  79 #include <sys/vnode.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/vmmeter.h>
  82
  83 #include <vm/vm.h>
  84 #include <vm/vm_param.h>
  85 #include <vm/vm_prot.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 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 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  *      vm_fault:
 158  *
 159  *      Handle a page fault occuring at the given address,
 160  *      requiring the given permissions, in the map specified.
 161  *      If successful, the page is inserted into the
 162  *      associated physical map.
 163  *
 164  *      NOTE: the given address should be truncated to the
 165  *      proper page address.
 166  *
 167  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 168  *      a standard error specifying why the fault is fatal is returned.
 169  *
 170  *
 171  *      The map in question must be referenced, and remains so.
 172  *      Caller may hold no locks.
 173  */
 174 int
 175 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 176 {
 177         vm_prot_t prot;
 178         int result;
 179         boolean_t wired;
 180         int map_generation;
 181         vm_page_t old_m;
 182         vm_object_t next_object;
 183         vm_page_t marray[VM_FAULT_READ];
 184         int hardfault;
 185         int faultcount;
 186         struct proc *p = curproc;       /* XXX */
 187         struct faultstate fs;
 188
 189         cnt.v_vm_faults++;      /* needs lock XXX */
 190         hardfault = 0;
 191
 192 RetryFault:;
 193         fs.map = map;
 194
 195         /*
 196          * Find the backing store object and offset into it to begin the
 197          * search.
 198          */
 199         if ((result = vm_map_lookup(&fs.map, vaddr,
 200                 fault_type, &fs.entry, &fs.first_object,
 201                 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
 202                 if ((result != KERN_PROTECTION_FAILURE) ||
 203                         ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
 204                         return result;
 205                 }
 206
 207                 /*
 208                  * If we are user-wiring a r/w segment, and it is COW, then
 209                  * we need to do the COW operation.  Note that we don't COW
 210                  * currently RO sections now, because it is NOT desirable
 211                  * to COW .text.  We simply keep .text from ever being COW'ed
 212                  * and take the heat that one cannot debug wired .text sections.
 213                  */
 214                 result = vm_map_lookup(&fs.map, vaddr,
 215                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 216                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 217                 if (result != KERN_SUCCESS) {
 218                         return result;
 219                 }
 220
 221                 /*
 222                  * If we don't COW now, on a user wire, the user will never
 223                  * be able to write to the mapping.  If we don't make this
 224                  * restriction, the bookkeeping would be nearly impossible.
 225                  */
 226                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 227                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 228         }
 229
 230         map_generation = fs.map->timestamp;
 231
 232         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 233                 panic("vm_fault: fault on nofault entry, addr: %lx",
 234                     (u_long)vaddr);
 235         }
 236
 237         /*
 238          * Make a reference to this object to prevent its disposal while we
 239          * are messing with it.  Once we have the reference, the map is free
 240          * to be diddled.  Since objects reference their shadows (and copies),
 241          * they will stay around as well.
 242          */
 243         vm_object_reference(fs.first_object);
 244         vm_object_pip_add(fs.first_object, 1);
 245
 246         fs.vp = vnode_pager_lock(fs.first_object);
 247         if ((fault_type & VM_PROT_WRITE) &&
 248                 (fs.first_object->type == OBJT_VNODE)) {
 249                 vm_freeze_copyopts(fs.first_object,
 250                         fs.first_pindex, fs.first_pindex + 1);
 251         }
 252
 253         fs.lookup_still_valid = TRUE;
 254
 255         if (wired)
 256                 fault_type = prot;
 257
 258         fs.first_m = NULL;
 259
 260         /*
 261          * Search for the page at object/offset.
 262          */
 263
 264         fs.object = fs.first_object;
 265         fs.pindex = fs.first_pindex;
 266
 267         /*
 268          * See whether this page is resident
 269          */
 270         while (TRUE) {
 271
 272                 if (fs.object->flags & OBJ_DEAD) {
 273                         unlock_and_deallocate(&fs);
 274                         return (KERN_PROTECTION_FAILURE);
 275                 }
 276
 277                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 278                 if (fs.m != NULL) {
 279                         int queue;
 280                         /*
 281                          * If the page is being brought in, wait for it and
 282                          * then retry.
 283                          */
 284                         if ((fs.m->flags & PG_BUSY) ||
 285                                 (fs.m->busy &&
 286                                  (fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL)) {
 287                                 int s;
 288
 289                                 unlock_things(&fs);
 290                                 s = splvm();
 291                                 if ((fs.m->flags & PG_BUSY) ||
 292                                         (fs.m->busy &&
 293                                          (fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL)) {
 294                                         vm_page_flag_set(fs.m, PG_WANTED | PG_REFERENCED);
 295                                         cnt.v_intrans++;
 296                                         tsleep(fs.m, PSWP, "vmpfw", 0);
 297                                 }
 298                                 splx(s);
 299                                 vm_object_deallocate(fs.first_object);
 300                                 goto RetryFault;
 301                         }
 302
 303                         queue = fs.m->queue;
 304                         vm_page_unqueue_nowakeup(fs.m);
 305
 306                         /*
 307                          * Mark page busy for other processes, and the pagedaemon.
 308                          */
 309                         if (((queue - fs.m->pc) == PQ_CACHE) &&
 310                             (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) {
 311                                 vm_page_activate(fs.m);
 312                                 unlock_and_deallocate(&fs);
 313                                 VM_WAIT;
 314                                 goto RetryFault;
 315                         }
 316
 317                         vm_page_busy(fs.m);
 318                         if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 319                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 320                                 goto readrest;
 321                         }
 322
 323                         break;
 324                 }
 325                 if (((fs.object->type != OBJT_DEFAULT) &&
 326                                 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 327                     || (fs.object == fs.first_object)) {
 328
 329                         if (fs.pindex >= fs.object->size) {
 330                                 unlock_and_deallocate(&fs);
 331                                 return (KERN_PROTECTION_FAILURE);
 332                         }
 333
 334                         /*
 335                          * Allocate a new page for this object/offset pair.
 336                          */
 337                         fs.m = vm_page_alloc(fs.object, fs.pindex,
 338                                 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
 339
 340                         if (fs.m == NULL) {
 341                                 unlock_and_deallocate(&fs);
 342                                 VM_WAIT;
 343                                 goto RetryFault;
 344                         }
 345                 }
 346
 347 readrest:
 348                 if (fs.object->type != OBJT_DEFAULT &&
 349                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) {
 350                         int rv;
 351                         int reqpage;
 352                         int ahead, behind;
 353
 354                         if (fs.first_object->behavior == OBJ_RANDOM) {
 355                                 ahead = 0;
 356                                 behind = 0;
 357                         } else {
 358                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 359                                 if (behind > VM_FAULT_READ_BEHIND)
 360                                         behind = VM_FAULT_READ_BEHIND;
 361
 362                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 363                                 if (ahead > VM_FAULT_READ_AHEAD)
 364                                         ahead = VM_FAULT_READ_AHEAD;
 365                         }
 366
 367                         if ((fs.first_object->type != OBJT_DEVICE) &&
 368                                 (fs.first_object->behavior == OBJ_SEQUENTIAL)) {
 369                                 vm_pindex_t firstpindex, tmppindex;
 370                                 if (fs.first_pindex <
 371                                         2*(VM_FAULT_READ_BEHIND + VM_FAULT_READ_AHEAD + 1))
 372                                         firstpindex = 0;
 373                                 else
 374                                         firstpindex = fs.first_pindex -
 375                                                 2*(VM_FAULT_READ_BEHIND + VM_FAULT_READ_AHEAD + 1);
 376
 377                                 for(tmppindex = fs.first_pindex - 1;
 378                                         tmppindex >= firstpindex;
 379                                         --tmppindex) {
 380                                         vm_page_t mt;
 381                                         mt = vm_page_lookup( fs.first_object, tmppindex);
 382                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 383                                                 break;
 384                                         if (mt->busy ||
 385                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS)) ||
 386                                                 mt->hold_count ||
 387                                                 mt->wire_count)
 388                                                 continue;
 389                                         if (mt->dirty == 0)
 390                                                 vm_page_test_dirty(mt);
 391                                         if (mt->dirty) {
 392                                                 vm_page_protect(mt, VM_PROT_NONE);
 393                                                 vm_page_deactivate(mt);
 394                                         } else {
 395                                                 vm_page_cache(mt);
 396                                         }
 397                                 }
 398
 399                                 ahead += behind;
 400                                 behind = 0;
 401                         }
 402
 403                         /*
 404                          * now we find out if any other pages should be paged
 405                          * in at this time this routine checks to see if the
 406                          * pages surrounding this fault reside in the same
 407                          * object as the page for this fault.  If they do,
 408                          * then they are faulted in also into the object.  The
 409                          * array "marray" returned contains an array of
 410                          * vm_page_t structs where one of them is the
 411                          * vm_page_t passed to the routine.  The reqpage
 412                          * return value is the index into the marray for the
 413                          * vm_page_t passed to the routine.
 414                          */
 415                         faultcount = vm_fault_additional_pages(
 416                             fs.m, behind, ahead, marray, &reqpage);
 417
 418                         /*
 419                          * Call the pager to retrieve the data, if any, after
 420                          * releasing the lock on the map.
 421                          */
 422                         unlock_map(&fs);
 423
 424                         rv = faultcount ?
 425                             vm_pager_get_pages(fs.object, marray, faultcount,
 426                                 reqpage) : VM_PAGER_FAIL;
 427
 428                         if (rv == VM_PAGER_OK) {
 429                                 /*
 430                                  * Found the page. Leave it busy while we play
 431                                  * with it.
 432                                  */
 433
 434                                 /*
 435                                  * Relookup in case pager changed page. Pager
 436                                  * is responsible for disposition of old page
 437                                  * if moved.
 438                                  */
 439                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 440                                 if(!fs.m) {
 441                                         unlock_and_deallocate(&fs);
 442                                         goto RetryFault;
 443                                 }
 444
 445                                 hardfault++;
 446                                 break;
 447                         }
 448                         /*
 449                          * Remove the bogus page (which does not exist at this
 450                          * object/offset); before doing so, we must get back
 451                          * our object lock to preserve our invariant.
 452                          *
 453                          * Also wake up any other process that may want to bring
 454                          * in this page.
 455                          *
 456                          * If this is the top-level object, we must leave the
 457                          * busy page to prevent another process from rushing
 458                          * past us, and inserting the page in that object at
 459                          * the same time that we are.
 460                          */
 461
 462                         if (rv == VM_PAGER_ERROR)
 463                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 464                                     curproc->p_pid, curproc->p_comm);
 465                         /*
 466                          * Data outside the range of the pager or an I/O error
 467                          */
 468                         /*
 469                          * XXX - the check for kernel_map is a kludge to work
 470                          * around having the machine panic on a kernel space
 471                          * fault w/ I/O error.
 472                          */
 473                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 474                                 (rv == VM_PAGER_BAD)) {
 475                                 vm_page_free(fs.m);
 476                                 fs.m = NULL;
 477                                 unlock_and_deallocate(&fs);
 478                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 479                         }
 480                         if (fs.object != fs.first_object) {
 481                                 vm_page_free(fs.m);
 482                                 fs.m = NULL;
 483                                 /*
 484                                  * XXX - we cannot just fall out at this
 485                                  * point, m has been freed and is invalid!
 486                                  */
 487                         }
 488                 }
 489                 /*
 490                  * We get here if the object has default pager (or unwiring) or the
 491                  * pager doesn't have the page.
 492                  */
 493                 if (fs.object == fs.first_object)
 494                         fs.first_m = fs.m;
 495
 496                 /*
 497                  * Move on to the next object.  Lock the next object before
 498                  * unlocking the current one.
 499                  */
 500
 501                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 502                 next_object = fs.object->backing_object;
 503                 if (next_object == NULL) {
 504                         /*
 505                          * If there's no object left, fill the page in the top
 506                          * object with zeros.
 507                          */
 508                         if (fs.object != fs.first_object) {
 509                                 vm_object_pip_wakeup(fs.object);
 510
 511                                 fs.object = fs.first_object;
 512                                 fs.pindex = fs.first_pindex;
 513                                 fs.m = fs.first_m;
 514                         }
 515                         fs.first_m = NULL;
 516
 517                         if ((fs.m->flags & PG_ZERO) == 0) {
 518                                 vm_page_zero_fill(fs.m);
 519                                 cnt.v_ozfod++;
 520                         }
 521                         cnt.v_zfod++;
 522                         break;
 523                 } else {
 524                         if (fs.object != fs.first_object) {
 525                                 vm_object_pip_wakeup(fs.object);
 526                         }
 527                         fs.object = next_object;
 528                         vm_object_pip_add(fs.object, 1);
 529                 }
 530         }
 531
 532 #if defined(DIAGNOSTIC)
 533         if ((fs.m->flags & PG_BUSY) == 0)
 534                 panic("vm_fault: not busy after main loop");
 535 #endif
 536
 537         /*
 538          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 539          * is held.]
 540          */
 541
 542         old_m = fs.m;   /* save page that would be copied */
 543
 544         /*
 545          * If the page is being written, but isn't already owned by the
 546          * top-level object, we have to copy it into a new page owned by the
 547          * top-level object.
 548          */
 549
 550         if (fs.object != fs.first_object) {
 551                 /*
 552                  * We only really need to copy if we want to write it.
 553                  */
 554
 555                 if (fault_type & VM_PROT_WRITE) {
 556
 557                         /*
 558                          * This allows pages to be virtually copied from a backing_object
 559                          * into the first_object, where the backing object has no other
 560                          * refs to it, and cannot gain any more refs.  Instead of a
 561                          * bcopy, we just move the page from the backing object to the
 562                          * first object.  Note that we must mark the page dirty in the
 563                          * first object so that it will go out to swap when needed.
 564                          */
 565                         if (map_generation == fs.map->timestamp &&
 566                                 /*
 567                                  * Only one shadow object
 568                                  */
 569                                 (fs.object->shadow_count == 1) &&
 570                                 /*
 571                                  * No COW refs, except us
 572                                  */
 573                                 (fs.object->ref_count == 1) &&
 574                                 /*
 575                                  * Noone else can look this object up
 576                                  */
 577                                 (fs.object->handle == NULL) &&
 578                                 /*
 579                                  * No other ways to look the object up
 580                                  */
 581                                 ((fs.object->type == OBJT_DEFAULT) ||
 582                                  (fs.object->type == OBJT_SWAP)) &&
 583                                 /*
 584                                  * We don't chase down the shadow chain
 585                                  */
 586                                 (fs.object == fs.first_object->backing_object) &&
 587
 588                                 /*
 589                                  * grab the lock if we need to
 590                                  */
 591                                 (fs.lookup_still_valid ||
 592                                                 (((fs.entry->eflags & MAP_ENTRY_IS_A_MAP) == 0) &&
 593                                                  lockmgr(&fs.map->lock,
 594                                                         LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curproc) == 0))) {
 595
 596                                 fs.lookup_still_valid = 1;
 597                                 /*
 598                                  * get rid of the unnecessary page
 599                                  */
 600                                 vm_page_protect(fs.first_m, VM_PROT_NONE);
 601                                 vm_page_free(fs.first_m);
 602                                 fs.first_m = NULL;
 603
 604                                 /*
 605                                  * grab the page and put it into the process'es object
 606                                  */
 607                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 608                                 fs.first_m = fs.m;
 609                                 fs.first_m->dirty = VM_PAGE_BITS_ALL;
 610                                 vm_page_busy(fs.first_m);
 611                                 fs.m = NULL;
 612                                 cnt.v_cow_optim++;
 613                         } else {
 614                                 /*
 615                                  * Oh, well, lets copy it.
 616                                  */
 617                                 vm_page_copy(fs.m, fs.first_m);
 618                         }
 619
 620                         if (fs.m) {
 621                                 /*
 622                                  * We no longer need the old page or object.
 623                                  */
 624                                 release_page(&fs);
 625                         }
 626
 627                         vm_object_pip_wakeup(fs.object);
 628                         /*
 629                          * Only use the new page below...
 630                          */
 631
 632                         cnt.v_cow_faults++;
 633                         fs.m = fs.first_m;
 634                         fs.object = fs.first_object;
 635                         fs.pindex = fs.first_pindex;
 636
 637                 } else {
 638                         prot &= ~VM_PROT_WRITE;
 639                 }
 640         }
 641
 642         /*
 643          * We must verify that the maps have not changed since our last
 644          * lookup.
 645          */
 646
 647         if (!fs.lookup_still_valid &&
 648                 (fs.map->timestamp != map_generation)) {
 649                 vm_object_t retry_object;
 650                 vm_pindex_t retry_pindex;
 651                 vm_prot_t retry_prot;
 652
 653                 /*
 654                  * Since map entries may be pageable, make sure we can take a
 655                  * page fault on them.
 656                  */
 657
 658                 /*
 659                  * To avoid trying to write_lock the map while another process
 660                  * has it read_locked (in vm_map_pageable), we do not try for
 661                  * write permission.  If the page is still writable, we will
 662                  * get write permission.  If it is not, or has been marked
 663                  * needs_copy, we enter the mapping without write permission,
 664                  * and will merely take another fault.
 665                  */
 666                 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
 667                     &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 668                 map_generation = fs.map->timestamp;
 669
 670                 /*
 671                  * If we don't need the page any longer, put it on the active
 672                  * list (the easiest thing to do here).  If no one needs it,
 673                  * pageout will grab it eventually.
 674                  */
 675
 676                 if (result != KERN_SUCCESS) {
 677                         release_page(&fs);
 678                         unlock_and_deallocate(&fs);
 679                         return (result);
 680                 }
 681                 fs.lookup_still_valid = TRUE;
 682
 683                 if ((retry_object != fs.first_object) ||
 684                     (retry_pindex != fs.first_pindex)) {
 685                         release_page(&fs);
 686                         unlock_and_deallocate(&fs);
 687                         goto RetryFault;
 688                 }
 689                 /*
 690                  * Check whether the protection has changed or the object has
 691                  * been copied while we left the map unlocked. Changing from
 692                  * read to write permission is OK - we leave the page
 693                  * write-protected, and catch the write fault. Changing from
 694                  * write to read permission means that we can't mark the page
 695                  * write-enabled after all.
 696                  */
 697                 prot &= retry_prot;
 698         }
 699
 700         /*
 701          * Put this page into the physical map. We had to do the unlock above
 702          * because pmap_enter may cause other faults.   We don't put the page
 703          * back on the active queue until later so that the page-out daemon
 704          * won't find us (yet).
 705          */
 706
 707         if (prot & VM_PROT_WRITE) {
 708                 vm_page_flag_set(fs.m, PG_WRITEABLE);
 709                 vm_object_set_flag(fs.m->object,
 710                                    OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
 711                 /*
 712                  * If the fault is a write, we know that this page is being
 713                  * written NOW. This will save on the pmap_is_modified() calls
 714                  * later.
 715                  */
 716                 if (fault_flags & VM_FAULT_DIRTY) {
 717                         fs.m->dirty = VM_PAGE_BITS_ALL;
 718                 }
 719         }
 720
 721         unlock_things(&fs);
 722         fs.m->valid = VM_PAGE_BITS_ALL;
 723         vm_page_flag_clear(fs.m, PG_ZERO);
 724
 725         pmap_enter(fs.map->pmap, vaddr, VM_PAGE_TO_PHYS(fs.m), prot, wired);
 726         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 727                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 728         }
 729
 730         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 731         if (fault_flags & VM_FAULT_HOLD)
 732                 vm_page_hold(fs.m);
 733
 734         /*
 735          * If the page is not wired down, then put it where the pageout daemon
 736          * can find it.
 737          */
 738         if (fault_flags & VM_FAULT_WIRE_MASK) {
 739                 if (wired)
 740                         vm_page_wire(fs.m);
 741                 else
 742                         vm_page_unwire(fs.m);
 743         } else {
 744                 vm_page_activate(fs.m);
 745         }
 746
 747         if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) {
 748                 if (hardfault) {
 749                         curproc->p_stats->p_ru.ru_majflt++;
 750                 } else {
 751                         curproc->p_stats->p_ru.ru_minflt++;
 752                 }
 753         }
 754
 755         /*
 756          * Unlock everything, and return
 757          */
 758
 759         vm_page_wakeup(fs.m);
 760         vm_object_deallocate(fs.first_object);
 761
 762         return (KERN_SUCCESS);
 763
 764 }
 765
 766 /*
 767  *      vm_fault_wire:
 768  *
 769  *      Wire down a range of virtual addresses in a map.
 770  */
 771 int
 772 vm_fault_wire(map, start, end)
 773         vm_map_t map;
 774         vm_offset_t start, end;
 775 {
 776
 777         register vm_offset_t va;
 778         register pmap_t pmap;
 779         int rv;
 780
 781         pmap = vm_map_pmap(map);
 782
 783         /*
 784          * Inform the physical mapping system that the range of addresses may
 785          * not fault, so that page tables and such can be locked down as well.
 786          */
 787
 788         pmap_pageable(pmap, start, end, FALSE);
 789
 790         /*
 791          * We simulate a fault to get the page and enter it in the physical
 792          * map.
 793          */
 794
 795         for (va = start; va < end; va += PAGE_SIZE) {
 796                 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
 797                         VM_FAULT_CHANGE_WIRING);
 798                 if (rv) {
 799                         if (va != start)
 800                                 vm_fault_unwire(map, start, va);
 801                         return (rv);
 802                 }
 803         }
 804         return (KERN_SUCCESS);
 805 }
 806
 807 /*
 808  *      vm_fault_user_wire:
 809  *
 810  *      Wire down a range of virtual addresses in a map.  This
 811  *      is for user mode though, so we only ask for read access
 812  *      on currently read only sections.
 813  */
 814 int
 815 vm_fault_user_wire(map, start, end)
 816         vm_map_t map;
 817         vm_offset_t start, end;
 818 {
 819
 820         register vm_offset_t va;
 821         register pmap_t pmap;
 822         int rv;
 823
 824         pmap = vm_map_pmap(map);
 825
 826         /*
 827          * Inform the physical mapping system that the range of addresses may
 828          * not fault, so that page tables and such can be locked down as well.
 829          */
 830
 831         pmap_pageable(pmap, start, end, FALSE);
 832
 833         /*
 834          * We simulate a fault to get the page and enter it in the physical
 835          * map.
 836          */
 837         for (va = start; va < end; va += PAGE_SIZE) {
 838                 rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
 839                 if (rv) {
 840                         if (va != start)
 841                                 vm_fault_unwire(map, start, va);
 842                         return (rv);
 843                 }
 844         }
 845         return (KERN_SUCCESS);
 846 }
 847
 848
 849 /*
 850  *      vm_fault_unwire:
 851  *
 852  *      Unwire a range of virtual addresses in a map.
 853  */
 854 void
 855 vm_fault_unwire(map, start, end)
 856         vm_map_t map;
 857         vm_offset_t start, end;
 858 {
 859
 860         register vm_offset_t va, pa;
 861         register pmap_t pmap;
 862
 863         pmap = vm_map_pmap(map);
 864
 865         /*
 866          * Since the pages are wired down, we must be able to get their
 867          * mappings from the physical map system.
 868          */
 869
 870         for (va = start; va < end; va += PAGE_SIZE) {
 871                 pa = pmap_extract(pmap, va);
 872                 if (pa != (vm_offset_t) 0) {
 873                         pmap_change_wiring(pmap, va, FALSE);
 874                         vm_page_unwire(PHYS_TO_VM_PAGE(pa));
 875                 }
 876         }
 877
 878         /*
 879          * Inform the physical mapping system that the range of addresses may
 880          * fault, so that page tables and such may be unwired themselves.
 881          */
 882
 883         pmap_pageable(pmap, start, end, TRUE);
 884
 885 }
 886
 887 /*
 888  *      Routine:
 889  *              vm_fault_copy_entry
 890  *      Function:
 891  *              Copy all of the pages from a wired-down map entry to another.
 892  *
 893  *      In/out conditions:
 894  *              The source and destination maps must be locked for write.
 895  *              The source map entry must be wired down (or be a sharing map
 896  *              entry corresponding to a main map entry that is wired down).
 897  */
 898
 899 void
 900 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
 901         vm_map_t dst_map;
 902         vm_map_t src_map;
 903         vm_map_entry_t dst_entry;
 904         vm_map_entry_t src_entry;
 905 {
 906         vm_object_t dst_object;
 907         vm_object_t src_object;
 908         vm_ooffset_t dst_offset;
 909         vm_ooffset_t src_offset;
 910         vm_prot_t prot;
 911         vm_offset_t vaddr;
 912         vm_page_t dst_m;
 913         vm_page_t src_m;
 914
 915 #ifdef  lint
 916         src_map++;
 917 #endif  /* lint */
 918
 919         src_object = src_entry->object.vm_object;
 920         src_offset = src_entry->offset;
 921
 922         /*
 923          * Create the top-level object for the destination entry. (Doesn't
 924          * actually shadow anything - we copy the pages directly.)
 925          */
 926         dst_object = vm_object_allocate(OBJT_DEFAULT,
 927             (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
 928
 929         dst_entry->object.vm_object = dst_object;
 930         dst_entry->offset = 0;
 931
 932         prot = dst_entry->max_protection;
 933
 934         /*
 935          * Loop through all of the pages in the entry's range, copying each
 936          * one from the source object (it should be there) to the destination
 937          * object.
 938          */
 939         for (vaddr = dst_entry->start, dst_offset = 0;
 940             vaddr < dst_entry->end;
 941             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
 942
 943                 /*
 944                  * Allocate a page in the destination object
 945                  */
 946                 do {
 947                         dst_m = vm_page_alloc(dst_object,
 948                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
 949                         if (dst_m == NULL) {
 950                                 VM_WAIT;
 951                         }
 952                 } while (dst_m == NULL);
 953
 954                 /*
 955                  * Find the page in the source object, and copy it in.
 956                  * (Because the source is wired down, the page will be in
 957                  * memory.)
 958                  */
 959                 src_m = vm_page_lookup(src_object,
 960                         OFF_TO_IDX(dst_offset + src_offset));
 961                 if (src_m == NULL)
 962                         panic("vm_fault_copy_wired: page missing");
 963
 964                 vm_page_copy(src_m, dst_m);
 965
 966                 /*
 967                  * Enter it in the pmap...
 968                  */
 969
 970                 vm_page_flag_clear(dst_m, PG_ZERO);
 971                 pmap_enter(dst_map->pmap, vaddr, VM_PAGE_TO_PHYS(dst_m),
 972                     prot, FALSE);
 973                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
 974
 975                 /*
 976                  * Mark it no longer busy, and put it on the active list.
 977                  */
 978                 vm_page_activate(dst_m);
 979                 vm_page_wakeup(dst_m);
 980         }
 981 }
 982
 983
 984 /*
 985  * This routine checks around the requested page for other pages that
 986  * might be able to be faulted in.  This routine brackets the viable
 987  * pages for the pages to be paged in.
 988  *
 989  * Inputs:
 990  *      m, rbehind, rahead
 991  *
 992  * Outputs:
 993  *  marray (array of vm_page_t), reqpage (index of requested page)
 994  *
 995  * Return value:
 996  *  number of pages in marray
 997  */
 998 static int
 999 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1000         vm_page_t m;
1001         int rbehind;
1002         int rahead;
1003         vm_page_t *marray;
1004         int *reqpage;
1005 {
1006         int i,j;
1007         vm_object_t object;
1008         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1009         vm_page_t rtm;
1010         int cbehind, cahead;
1011
1012         object = m->object;
1013         pindex = m->pindex;
1014
1015         /*
1016          * we don't fault-ahead for device pager
1017          */
1018         if (object->type == OBJT_DEVICE) {
1019                 *reqpage = 0;
1020                 marray[0] = m;
1021                 return 1;
1022         }
1023
1024         /*
1025          * if the requested page is not available, then give up now
1026          */
1027
1028         if (!vm_pager_has_page(object,
1029                 OFF_TO_IDX(object->paging_offset) + pindex, &cbehind, &cahead)) {
1030                 return 0;
1031         }
1032
1033         if ((cbehind == 0) && (cahead == 0)) {
1034                 *reqpage = 0;
1035                 marray[0] = m;
1036                 return 1;
1037         }
1038
1039         if (rahead > cahead) {
1040                 rahead = cahead;
1041         }
1042
1043         if (rbehind > cbehind) {
1044                 rbehind = cbehind;
1045         }
1046
1047         /*
1048          * try to do any readahead that we might have free pages for.
1049          */
1050         if ((rahead + rbehind) >
1051                 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1052                 pagedaemon_wakeup();
1053                 marray[0] = m;
1054                 *reqpage = 0;
1055                 return 1;
1056         }
1057
1058         /*
1059          * scan backward for the read behind pages -- in memory
1060          */
1061         if (pindex > 0) {
1062                 if (rbehind > pindex) {
1063                         rbehind = pindex;
1064                         startpindex = 0;
1065                 } else {
1066                         startpindex = pindex - rbehind;
1067                 }
1068
1069                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1070                         if (vm_page_lookup( object, tpindex)) {
1071                                 startpindex = tpindex + 1;
1072                                 break;
1073                         }
1074                         if (tpindex == 0)
1075                                 break;
1076                 }
1077
1078                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1079
1080                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1081                         if (rtm == NULL) {
1082                                 for (j = 0; j < i; j++) {
1083                                         vm_page_free(marray[j]);
1084                                 }
1085                                 marray[0] = m;
1086                                 *reqpage = 0;
1087                                 return 1;
1088                         }
1089
1090                         marray[i] = rtm;
1091                 }
1092         } else {
1093                 startpindex = 0;
1094                 i = 0;
1095         }
1096
1097         marray[i] = m;
1098         /* page offset of the required page */
1099         *reqpage = i;
1100
1101         tpindex = pindex + 1;
1102         i++;
1103
1104         /*
1105          * scan forward for the read ahead pages
1106          */
1107         endpindex = tpindex + rahead;
1108         if (endpindex > object->size)
1109                 endpindex = object->size;
1110
1111         for( ; tpindex < endpindex; i++, tpindex++) {
1112
1113                 if (vm_page_lookup(object, tpindex)) {
1114                         break;
1115                 }
1116
1117                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1118                 if (rtm == NULL) {
1119                         break;
1120                 }
1121
1122                 marray[i] = rtm;
1123         }
1124
1125         /* return number of bytes of pages */
1126         return i;
1127 }