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