sys/vm/vm_fault.c

   1 /*
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * Copyright (c) 1994 John S. Dyson
   5  * All rights reserved.
   6  * Copyright (c) 1994 David Greenman
   7  * All rights reserved.
   8  *
   9  *
  10  * This code is derived from software contributed to Berkeley by
  11  * The Mach Operating System project at Carnegie-Mellon University.
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  * 1. Redistributions of source code must retain the above copyright
  17  *    notice, this list of conditions and the following disclaimer.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. All advertising materials mentioning features or use of this software
  22  *    must display the following acknowledgement:
  23  *      This product includes software developed by the University of
  24  *      California, Berkeley and its contributors.
  25  * 4. Neither the name of the University nor the names of its contributors
  26  *    may be used to endorse or promote products derived from this software
  27  *    without specific prior written permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  39  * SUCH DAMAGE.
  40  *
  41  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
  42  *
  43  *
  44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  45  * All rights reserved.
  46  *
  47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  48  *
  49  * Permission to use, copy, modify and distribute this software and
  50  * its documentation is hereby granted, provided that both the copyright
  51  * notice and this permission notice appear in all copies of the
  52  * software, derivative works or modified versions, and any portions
  53  * thereof, and that both notices appear in supporting documentation.
  54  *
  55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  58  *
  59  * Carnegie Mellon requests users of this software to return to
  60  *
  61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  62  *  School of Computer Science
  63  *  Carnegie Mellon University
  64  *  Pittsburgh PA 15213-3890
  65  *
  66  * any improvements or extensions that they make and grant Carnegie the
  67  * rights to redistribute these changes.
  68  *
  69  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
  70  * $DragonFly: src/sys/vm/vm_fault.c,v 1.19 2005/10/24 20:02:09 dillon Exp $
  71  */
  72
  73 /*
  74  *      Page fault handling module.
  75  */
  76
  77 #include <sys/param.h>
  78 #include <sys/systm.h>
  79 #include <sys/proc.h>
  80 #include <sys/vnode.h>
  81 #include <sys/resourcevar.h>
  82 #include <sys/vmmeter.h>
  83
  84 #include <vm/vm.h>
  85 #include <vm/vm_param.h>
  86 #include <sys/lock.h>
  87 #include <vm/pmap.h>
  88 #include <vm/vm_map.h>
  89 #include <vm/vm_object.h>
  90 #include <vm/vm_page.h>
  91 #include <vm/vm_pageout.h>
  92 #include <vm/vm_kern.h>
  93 #include <vm/vm_pager.h>
  94 #include <vm/vnode_pager.h>
  95 #include <vm/vm_extern.h>
  96
  97 #include <sys/thread2.h>
  98 #include <vm/vm_page2.h>
  99
 100 static int vm_fault_additional_pages (vm_page_t, int,
 101                                           int, vm_page_t *, int *);
 102
 103 #define VM_FAULT_READ_AHEAD 8
 104 #define VM_FAULT_READ_BEHIND 7
 105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 106
 107 struct faultstate {
 108         vm_page_t m;
 109         vm_object_t object;
 110         vm_pindex_t pindex;
 111         vm_page_t first_m;
 112         vm_object_t     first_object;
 113         vm_pindex_t first_pindex;
 114         vm_map_t map;
 115         vm_map_entry_t entry;
 116         int lookup_still_valid;
 117         struct vnode *vp;
 118 };
 119
 120 static __inline void
 121 release_page(struct faultstate *fs)
 122 {
 123         vm_page_wakeup(fs->m);
 124         vm_page_deactivate(fs->m);
 125         fs->m = NULL;
 126 }
 127
 128 static __inline void
 129 unlock_map(struct faultstate *fs)
 130 {
 131         if (fs->lookup_still_valid) {
 132                 vm_map_lookup_done(fs->map, fs->entry, 0);
 133                 fs->lookup_still_valid = FALSE;
 134         }
 135 }
 136
 137 static void
 138 _unlock_things(struct faultstate *fs, int dealloc)
 139 {
 140         vm_object_pip_wakeup(fs->object);
 141         if (fs->object != fs->first_object) {
 142                 vm_page_free(fs->first_m);
 143                 vm_object_pip_wakeup(fs->first_object);
 144                 fs->first_m = NULL;
 145         }
 146         if (dealloc) {
 147                 vm_object_deallocate(fs->first_object);
 148         }
 149         unlock_map(fs);
 150         if (fs->vp != NULL) {
 151                 vput(fs->vp);
 152                 fs->vp = NULL;
 153         }
 154 }
 155
 156 #define unlock_things(fs) _unlock_things(fs, 0)
 157 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 158
 159 /*
 160  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 161  *            current object *might* contain the page.
 162  *
 163  *            default objects are zero-fill, there is no real pager.
 164  */
 165
 166 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 167                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 168
 169 /*
 170  *      vm_fault:
 171  *
 172  *      Handle a page fault occurring at the given address,
 173  *      requiring the given permissions, in the map specified.
 174  *      If successful, the page is inserted into the
 175  *      associated physical map.
 176  *
 177  *      NOTE: the given address should be truncated to the
 178  *      proper page address.
 179  *
 180  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 181  *      a standard error specifying why the fault is fatal is returned.
 182  *
 183  *
 184  *      The map in question must be referenced, and remains so.
 185  *      Caller may hold no locks.
 186  */
 187 int
 188 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 189 {
 190         vm_prot_t prot;
 191         int result;
 192         boolean_t wired;
 193         int map_generation;
 194         vm_object_t next_object;
 195         vm_page_t marray[VM_FAULT_READ];
 196         int hardfault;
 197         int faultcount;
 198         struct faultstate fs;
 199
 200         mycpu->gd_cnt.v_vm_faults++;
 201         hardfault = 0;
 202
 203 RetryFault:
 204         /*
 205          * Find the backing store object and offset into it to begin the
 206          * search.
 207          */
 208         fs.map = map;
 209         if ((result = vm_map_lookup(&fs.map, vaddr,
 210                 fault_type, &fs.entry, &fs.first_object,
 211                 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
 212                 if ((result != KERN_PROTECTION_FAILURE) ||
 213                         ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
 214                         return result;
 215                 }
 216
 217                 /*
 218                  * If we are user-wiring a r/w segment, and it is COW, then
 219                  * we need to do the COW operation.  Note that we don't COW
 220                  * currently RO sections now, because it is NOT desirable
 221                  * to COW .text.  We simply keep .text from ever being COW'ed
 222                  * and take the heat that one cannot debug wired .text sections.
 223                  */
 224                 result = vm_map_lookup(&fs.map, vaddr,
 225                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 226                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 227                 if (result != KERN_SUCCESS) {
 228                         return result;
 229                 }
 230
 231                 /*
 232                  * If we don't COW now, on a user wire, the user will never
 233                  * be able to write to the mapping.  If we don't make this
 234                  * restriction, the bookkeeping would be nearly impossible.
 235                  */
 236                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 237                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 238         }
 239
 240         map_generation = fs.map->timestamp;
 241
 242         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 243                 panic("vm_fault: fault on nofault entry, addr: %lx",
 244                     (u_long)vaddr);
 245         }
 246
 247         /*
 248          * A system map entry may return a NULL object.  No object means
 249          * no pager means an unrecoverable kernel fault.
 250          */
 251         if (fs.first_object == NULL) {
 252                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 253                         (void *)vaddr, fs.entry);
 254         }
 255
 256         /*
 257          * Make a reference to this object to prevent its disposal while we
 258          * are messing with it.  Once we have the reference, the map is free
 259          * to be diddled.  Since objects reference their shadows (and copies),
 260          * they will stay around as well.
 261          *
 262          * Bump the paging-in-progress count to prevent size changes (e.g.
 263          * truncation operations) during I/O.  This must be done after
 264          * obtaining the vnode lock in order to avoid possible deadlocks.
 265          */
 266         vm_object_reference(fs.first_object);
 267         fs.vp = vnode_pager_lock(fs.first_object);
 268         vm_object_pip_add(fs.first_object, 1);
 269
 270         if ((fault_type & VM_PROT_WRITE) &&
 271                 (fs.first_object->type == OBJT_VNODE)) {
 272                 vm_freeze_copyopts(fs.first_object,
 273                         fs.first_pindex, fs.first_pindex + 1);
 274         }
 275
 276         fs.lookup_still_valid = TRUE;
 277
 278         if (wired)
 279                 fault_type = prot;
 280
 281         fs.first_m = NULL;
 282
 283         /*
 284          * Search for the page at object/offset.
 285          */
 286
 287         fs.object = fs.first_object;
 288         fs.pindex = fs.first_pindex;
 289
 290         while (TRUE) {
 291                 /*
 292                  * If the object is dead, we stop here
 293                  */
 294
 295                 if (fs.object->flags & OBJ_DEAD) {
 296                         unlock_and_deallocate(&fs);
 297                         return (KERN_PROTECTION_FAILURE);
 298                 }
 299
 300                 /*
 301                  * See if page is resident.  spl protection is required
 302                  * to avoid an interrupt unbusy/free race against our
 303                  * lookup.  We must hold the protection through a page
 304                  * allocation or busy.
 305                  */
 306                 crit_enter();
 307                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 308                 if (fs.m != NULL) {
 309                         int queue;
 310                         /*
 311                          * Wait/Retry if the page is busy.  We have to do this
 312                          * if the page is busy via either PG_BUSY or
 313                          * vm_page_t->busy because the vm_pager may be using
 314                          * vm_page_t->busy for pageouts ( and even pageins if
 315                          * it is the vnode pager ), and we could end up trying
 316                          * to pagein and pageout the same page simultaneously.
 317                          *
 318                          * We can theoretically allow the busy case on a read
 319                          * fault if the page is marked valid, but since such
 320                          * pages are typically already pmap'd, putting that
 321                          * special case in might be more effort then it is
 322                          * worth.  We cannot under any circumstances mess
 323                          * around with a vm_page_t->busy page except, perhaps,
 324                          * to pmap it.
 325                          */
 326                         if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
 327                                 unlock_things(&fs);
 328                                 vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
 329                                 mycpu->gd_cnt.v_intrans++;
 330                                 vm_object_deallocate(fs.first_object);
 331                                 crit_exit();
 332                                 goto RetryFault;
 333                         }
 334
 335                         queue = fs.m->queue;
 336                         vm_page_unqueue_nowakeup(fs.m);
 337
 338                         if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
 339                                 vm_page_activate(fs.m);
 340                                 unlock_and_deallocate(&fs);
 341                                 vm_waitpfault();
 342                                 crit_exit();
 343                                 goto RetryFault;
 344                         }
 345
 346                         /*
 347                          * Mark page busy for other processes, and the
 348                          * pagedaemon.  If it still isn't completely valid
 349                          * (readable), jump to readrest, else break-out ( we
 350                          * found the page ).
 351                          *
 352                          * We can release the spl once we have marked the
 353                          * page busy.
 354                          */
 355
 356                         vm_page_busy(fs.m);
 357                         crit_exit();
 358
 359                         if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 360                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 361                                 goto readrest;
 362                         }
 363
 364                         break;
 365                 }
 366
 367                 /*
 368                  * Page is not resident, If this is the search termination
 369                  * or the pager might contain the page, allocate a new page.
 370                  *
 371                  * note: we are still in splvm().
 372                  */
 373
 374                 if (TRYPAGER || fs.object == fs.first_object) {
 375                         if (fs.pindex >= fs.object->size) {
 376                                 crit_exit();
 377                                 unlock_and_deallocate(&fs);
 378                                 return (KERN_PROTECTION_FAILURE);
 379                         }
 380
 381                         /*
 382                          * Allocate a new page for this object/offset pair.
 383                          */
 384                         fs.m = NULL;
 385                         if (!vm_page_count_severe()) {
 386                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 387                                     (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
 388                         }
 389                         if (fs.m == NULL) {
 390                                 crit_exit();
 391                                 unlock_and_deallocate(&fs);
 392                                 vm_waitpfault();
 393                                 goto RetryFault;
 394                         }
 395                 }
 396                 crit_exit();
 397
 398 readrest:
 399                 /*
 400                  * We have found a valid page or we have allocated a new page.
 401                  * The page thus may not be valid or may not be entirely
 402                  * valid.
 403                  *
 404                  * Attempt to fault-in the page if there is a chance that the
 405                  * pager has it, and potentially fault in additional pages
 406                  * at the same time.
 407                  *
 408                  * We are NOT in splvm here and if TRYPAGER is true then
 409                  * fs.m will be non-NULL and will be PG_BUSY for us.
 410                  */
 411
 412                 if (TRYPAGER) {
 413                         int rv;
 414                         int reqpage;
 415                         int ahead, behind;
 416                         u_char behavior = vm_map_entry_behavior(fs.entry);
 417
 418                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 419                                 ahead = 0;
 420                                 behind = 0;
 421                         } else {
 422                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 423                                 if (behind > VM_FAULT_READ_BEHIND)
 424                                         behind = VM_FAULT_READ_BEHIND;
 425
 426                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 427                                 if (ahead > VM_FAULT_READ_AHEAD)
 428                                         ahead = VM_FAULT_READ_AHEAD;
 429                         }
 430
 431                         if ((fs.first_object->type != OBJT_DEVICE) &&
 432                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 433                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 434                                 fs.pindex >= fs.entry->lastr &&
 435                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
 436                         ) {
 437                                 vm_pindex_t firstpindex, tmppindex;
 438
 439                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 440                                         firstpindex = 0;
 441                                 else
 442                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 443
 444                                 /*
 445                                  * note: partially valid pages cannot be
 446                                  * included in the lookahead - NFS piecemeal
 447                                  * writes will barf on it badly.
 448                                  *
 449                                  * spl protection is required to avoid races
 450                                  * between the lookup and an interrupt
 451                                  * unbusy/free sequence occuring prior to
 452                                  * our busy check.
 453                                  */
 454                                 crit_enter();
 455                                 for (tmppindex = fs.first_pindex - 1;
 456                                     tmppindex >= firstpindex;
 457                                     --tmppindex
 458                                 ) {
 459                                         vm_page_t mt;
 460                                         mt = vm_page_lookup( fs.first_object, tmppindex);
 461                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 462                                                 break;
 463                                         if (mt->busy ||
 464                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
 465                                                 mt->hold_count ||
 466                                                 mt->wire_count)
 467                                                 continue;
 468                                         if (mt->dirty == 0)
 469                                                 vm_page_test_dirty(mt);
 470                                         if (mt->dirty) {
 471                                                 vm_page_protect(mt, VM_PROT_NONE);
 472                                                 vm_page_deactivate(mt);
 473                                         } else {
 474                                                 vm_page_cache(mt);
 475                                         }
 476                                 }
 477                                 crit_exit();
 478
 479                                 ahead += behind;
 480                                 behind = 0;
 481                         }
 482
 483                         /*
 484                          * now we find out if any other pages should be paged
 485                          * in at this time this routine checks to see if the
 486                          * pages surrounding this fault reside in the same
 487                          * object as the page for this fault.  If they do,
 488                          * then they are faulted in also into the object.  The
 489                          * array "marray" returned contains an array of
 490                          * vm_page_t structs where one of them is the
 491                          * vm_page_t passed to the routine.  The reqpage
 492                          * return value is the index into the marray for the
 493                          * vm_page_t passed to the routine.
 494                          *
 495                          * fs.m plus the additional pages are PG_BUSY'd.
 496                          */
 497                         faultcount = vm_fault_additional_pages(
 498                             fs.m, behind, ahead, marray, &reqpage);
 499
 500                         /*
 501                          * update lastr imperfectly (we do not know how much
 502                          * getpages will actually read), but good enough.
 503                          */
 504                         fs.entry->lastr = fs.pindex + faultcount - behind;
 505
 506                         /*
 507                          * Call the pager to retrieve the data, if any, after
 508                          * releasing the lock on the map.  We hold a ref on
 509                          * fs.object and the pages are PG_BUSY'd.
 510                          */
 511                         unlock_map(&fs);
 512
 513                         rv = faultcount ?
 514                             vm_pager_get_pages(fs.object, marray, faultcount,
 515                                 reqpage) : VM_PAGER_FAIL;
 516
 517                         if (rv == VM_PAGER_OK) {
 518                                 /*
 519                                  * Found the page. Leave it busy while we play
 520                                  * with it.
 521                                  */
 522
 523                                 /*
 524                                  * Relookup in case pager changed page. Pager
 525                                  * is responsible for disposition of old page
 526                                  * if moved.
 527                                  *
 528                                  * XXX other code segments do relookups too.
 529                                  * It's a bad abstraction that needs to be
 530                                  * fixed/removed.
 531                                  */
 532                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 533                                 if (fs.m == NULL) {
 534                                         unlock_and_deallocate(&fs);
 535                                         goto RetryFault;
 536                                 }
 537
 538                                 hardfault++;
 539                                 break; /* break to PAGE HAS BEEN FOUND */
 540                         }
 541                         /*
 542                          * Remove the bogus page (which does not exist at this
 543                          * object/offset); before doing so, we must get back
 544                          * our object lock to preserve our invariant.
 545                          *
 546                          * Also wake up any other process that may want to bring
 547                          * in this page.
 548                          *
 549                          * If this is the top-level object, we must leave the
 550                          * busy page to prevent another process from rushing
 551                          * past us, and inserting the page in that object at
 552                          * the same time that we are.
 553                          */
 554
 555                         if (rv == VM_PAGER_ERROR)
 556                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 557                                     curproc->p_pid, curproc->p_comm);
 558                         /*
 559                          * Data outside the range of the pager or an I/O error
 560                          */
 561                         /*
 562                          * XXX - the check for kernel_map is a kludge to work
 563                          * around having the machine panic on a kernel space
 564                          * fault w/ I/O error.
 565                          */
 566                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 567                                 (rv == VM_PAGER_BAD)) {
 568                                 vm_page_free(fs.m);
 569                                 fs.m = NULL;
 570                                 unlock_and_deallocate(&fs);
 571                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 572                         }
 573                         if (fs.object != fs.first_object) {
 574                                 vm_page_free(fs.m);
 575                                 fs.m = NULL;
 576                                 /*
 577                                  * XXX - we cannot just fall out at this
 578                                  * point, m has been freed and is invalid!
 579                                  */
 580                         }
 581                 }
 582
 583                 /*
 584                  * We get here if the object has default pager (or unwiring)
 585                  * or the pager doesn't have the page.
 586                  */
 587                 if (fs.object == fs.first_object)
 588                         fs.first_m = fs.m;
 589
 590                 /*
 591                  * Move on to the next object.  Lock the next object before
 592                  * unlocking the current one.
 593                  */
 594
 595                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 596                 next_object = fs.object->backing_object;
 597                 if (next_object == NULL) {
 598                         /*
 599                          * If there's no object left, fill the page in the top
 600                          * object with zeros.
 601                          */
 602                         if (fs.object != fs.first_object) {
 603                                 vm_object_pip_wakeup(fs.object);
 604
 605                                 fs.object = fs.first_object;
 606                                 fs.pindex = fs.first_pindex;
 607                                 fs.m = fs.first_m;
 608                         }
 609                         fs.first_m = NULL;
 610
 611                         /*
 612                          * Zero the page if necessary and mark it valid.
 613                          */
 614                         if ((fs.m->flags & PG_ZERO) == 0) {
 615                                 vm_page_zero_fill(fs.m);
 616                         } else {
 617                                 mycpu->gd_cnt.v_ozfod++;
 618                         }
 619                         mycpu->gd_cnt.v_zfod++;
 620                         fs.m->valid = VM_PAGE_BITS_ALL;
 621                         break;  /* break to PAGE HAS BEEN FOUND */
 622                 } else {
 623                         if (fs.object != fs.first_object) {
 624                                 vm_object_pip_wakeup(fs.object);
 625                         }
 626                         KASSERT(fs.object != next_object, ("object loop %p", next_object));
 627                         fs.object = next_object;
 628                         vm_object_pip_add(fs.object, 1);
 629                 }
 630         }
 631
 632         KASSERT((fs.m->flags & PG_BUSY) != 0,
 633             ("vm_fault: not busy after main loop"));
 634
 635         /*
 636          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 637          * is held.]
 638          */
 639
 640         /*
 641          * If the page is being written, but isn't already owned by the
 642          * top-level object, we have to copy it into a new page owned by the
 643          * top-level object.
 644          */
 645
 646         if (fs.object != fs.first_object) {
 647                 /*
 648                  * We only really need to copy if we want to write it.
 649                  */
 650
 651                 if (fault_type & VM_PROT_WRITE) {
 652                         /*
 653                          * This allows pages to be virtually copied from a
 654                          * backing_object into the first_object, where the
 655                          * backing object has no other refs to it, and cannot
 656                          * gain any more refs.  Instead of a bcopy, we just
 657                          * move the page from the backing object to the
 658                          * first object.  Note that we must mark the page
 659                          * dirty in the first object so that it will go out
 660                          * to swap when needed.
 661                          */
 662                         if (map_generation == fs.map->timestamp &&
 663                                 /*
 664                                  * Only one shadow object
 665                                  */
 666                                 (fs.object->shadow_count == 1) &&
 667                                 /*
 668                                  * No COW refs, except us
 669                                  */
 670                                 (fs.object->ref_count == 1) &&
 671                                 /*
 672                                  * No one else can look this object up
 673                                  */
 674                                 (fs.object->handle == NULL) &&
 675                                 /*
 676                                  * No other ways to look the object up
 677                                  */
 678                                 ((fs.object->type == OBJT_DEFAULT) ||
 679                                  (fs.object->type == OBJT_SWAP)) &&
 680                                 /*
 681                                  * We don't chase down the shadow chain
 682                                  */
 683                                 (fs.object == fs.first_object->backing_object) &&
 684
 685                                 /*
 686                                  * grab the lock if we need to
 687                                  */
 688                                 (fs.lookup_still_valid ||
 689                                  lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, NULL, curthread) == 0)
 690                             ) {
 691
 692                                 fs.lookup_still_valid = 1;
 693                                 /*
 694                                  * get rid of the unnecessary page
 695                                  */
 696                                 vm_page_protect(fs.first_m, VM_PROT_NONE);
 697                                 vm_page_free(fs.first_m);
 698                                 fs.first_m = NULL;
 699
 700                                 /*
 701                                  * grab the page and put it into the
 702                                  * process'es object.  The page is
 703                                  * automatically made dirty.
 704                                  */
 705                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 706                                 fs.first_m = fs.m;
 707                                 vm_page_busy(fs.first_m);
 708                                 fs.m = NULL;
 709                                 mycpu->gd_cnt.v_cow_optim++;
 710                         } else {
 711                                 /*
 712                                  * Oh, well, lets copy it.
 713                                  */
 714                                 vm_page_copy(fs.m, fs.first_m);
 715                         }
 716
 717                         if (fs.m) {
 718                                 /*
 719                                  * We no longer need the old page or object.
 720                                  */
 721                                 release_page(&fs);
 722                         }
 723
 724                         /*
 725                          * fs.object != fs.first_object due to above
 726                          * conditional
 727                          */
 728
 729                         vm_object_pip_wakeup(fs.object);
 730
 731                         /*
 732                          * Only use the new page below...
 733                          */
 734
 735                         mycpu->gd_cnt.v_cow_faults++;
 736                         fs.m = fs.first_m;
 737                         fs.object = fs.first_object;
 738                         fs.pindex = fs.first_pindex;
 739
 740                 } else {
 741                         prot &= ~VM_PROT_WRITE;
 742                 }
 743         }
 744
 745         /*
 746          * We must verify that the maps have not changed since our last
 747          * lookup.
 748          */
 749
 750         if (!fs.lookup_still_valid &&
 751                 (fs.map->timestamp != map_generation)) {
 752                 vm_object_t retry_object;
 753                 vm_pindex_t retry_pindex;
 754                 vm_prot_t retry_prot;
 755
 756                 /*
 757                  * Since map entries may be pageable, make sure we can take a
 758                  * page fault on them.
 759                  */
 760
 761                 /*
 762                  * Unlock vnode before the lookup to avoid deadlock.   E.G.
 763                  * avoid a deadlock between the inode and exec_map that can
 764                  * occur due to locks being obtained in different orders.
 765                  */
 766
 767                 if (fs.vp != NULL) {
 768                         vput(fs.vp);
 769                         fs.vp = NULL;
 770                 }
 771
 772                 if (fs.map->infork) {
 773                         release_page(&fs);
 774                         unlock_and_deallocate(&fs);
 775                         goto RetryFault;
 776                 }
 777
 778                 /*
 779                  * To avoid trying to write_lock the map while another process
 780                  * has it read_locked (in vm_map_wire), we do not try for
 781                  * write permission.  If the page is still writable, we will
 782                  * get write permission.  If it is not, or has been marked
 783                  * needs_copy, we enter the mapping without write permission,
 784                  * and will merely take another fault.
 785                  */
 786                 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
 787                     &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 788                 map_generation = fs.map->timestamp;
 789
 790                 /*
 791                  * If we don't need the page any longer, put it on the active
 792                  * list (the easiest thing to do here).  If no one needs it,
 793                  * pageout will grab it eventually.
 794                  */
 795
 796                 if (result != KERN_SUCCESS) {
 797                         release_page(&fs);
 798                         unlock_and_deallocate(&fs);
 799                         return (result);
 800                 }
 801                 fs.lookup_still_valid = TRUE;
 802
 803                 if ((retry_object != fs.first_object) ||
 804                     (retry_pindex != fs.first_pindex)) {
 805                         release_page(&fs);
 806                         unlock_and_deallocate(&fs);
 807                         goto RetryFault;
 808                 }
 809                 /*
 810                  * Check whether the protection has changed or the object has
 811                  * been copied while we left the map unlocked. Changing from
 812                  * read to write permission is OK - we leave the page
 813                  * write-protected, and catch the write fault. Changing from
 814                  * write to read permission means that we can't mark the page
 815                  * write-enabled after all.
 816                  */
 817                 prot &= retry_prot;
 818         }
 819
 820         /*
 821          * Put this page into the physical map. We had to do the unlock above
 822          * because pmap_enter may cause other faults.   We don't put the page
 823          * back on the active queue until later so that the page-out daemon
 824          * won't find us (yet).
 825          */
 826
 827         if (prot & VM_PROT_WRITE) {
 828                 vm_page_flag_set(fs.m, PG_WRITEABLE);
 829                 vm_object_set_writeable_dirty(fs.m->object);
 830
 831                 /*
 832                  * If the fault is a write, we know that this page is being
 833                  * written NOW so dirty it explicitly to save on
 834                  * pmap_is_modified() calls later.
 835                  *
 836                  * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
 837                  * if the page is already dirty to prevent data written with
 838                  * the expectation of being synced from not being synced.
 839                  * Likewise if this entry does not request NOSYNC then make
 840                  * sure the page isn't marked NOSYNC.  Applications sharing
 841                  * data should use the same flags to avoid ping ponging.
 842                  *
 843                  * Also tell the backing pager, if any, that it should remove
 844                  * any swap backing since the page is now dirty.
 845                  */
 846                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 847                         if (fs.m->dirty == 0)
 848                                 vm_page_flag_set(fs.m, PG_NOSYNC);
 849                 } else {
 850                         vm_page_flag_clear(fs.m, PG_NOSYNC);
 851                 }
 852                 if (fault_flags & VM_FAULT_DIRTY) {
 853                         crit_enter();
 854                         vm_page_dirty(fs.m);
 855                         vm_pager_page_unswapped(fs.m);
 856                         crit_exit();
 857                 }
 858         }
 859
 860         /*
 861          * Page had better still be busy
 862          */
 863
 864         KASSERT(fs.m->flags & PG_BUSY,
 865                 ("vm_fault: page %p not busy!", fs.m));
 866
 867         unlock_things(&fs);
 868
 869         /*
 870          * Sanity check: page must be completely valid or it is not fit to
 871          * map into user space.  vm_pager_get_pages() ensures this.
 872          */
 873
 874         if (fs.m->valid != VM_PAGE_BITS_ALL) {
 875                 vm_page_zero_invalid(fs.m, TRUE);
 876                 printf("Warning: page %p partially invalid on fault\n", fs.m);
 877         }
 878
 879         pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
 880
 881         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 882                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 883         }
 884
 885         vm_page_flag_clear(fs.m, PG_ZERO);
 886         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 887         if (fault_flags & VM_FAULT_HOLD)
 888                 vm_page_hold(fs.m);
 889
 890         /*
 891          * If the page is not wired down, then put it where the pageout daemon
 892          * can find it.
 893          */
 894
 895         if (fault_flags & VM_FAULT_WIRE_MASK) {
 896                 if (wired)
 897                         vm_page_wire(fs.m);
 898                 else
 899                         vm_page_unwire(fs.m, 1);
 900         } else {
 901                 vm_page_activate(fs.m);
 902         }
 903
 904         if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) {
 905                 if (hardfault) {
 906                         curproc->p_stats->p_ru.ru_majflt++;
 907                 } else {
 908                         curproc->p_stats->p_ru.ru_minflt++;
 909                 }
 910         }
 911
 912         /*
 913          * Unlock everything, and return
 914          */
 915
 916         vm_page_wakeup(fs.m);
 917         vm_object_deallocate(fs.first_object);
 918
 919         return (KERN_SUCCESS);
 920
 921 }
 922
 923 /*
 924  * quick version of vm_fault
 925  */
 926 int
 927 vm_fault_quick(caddr_t v, int prot)
 928 {
 929         int r;
 930
 931         if (prot & VM_PROT_WRITE)
 932                 r = subyte(v, fubyte(v));
 933         else
 934                 r = fubyte(v);
 935         return(r);
 936 }
 937
 938 /*
 939  * Wire down a range of virtual addresses in a map.  The entry in question
 940  * should be marked in-transition and the map must be locked.  We must
 941  * release the map temporarily while faulting-in the page to avoid a
 942  * deadlock.  Note that the entry may be clipped while we are blocked but
 943  * will never be freed.
 944  */
 945 int
 946 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
 947 {
 948         boolean_t fictitious;
 949         vm_offset_t start;
 950         vm_offset_t end;
 951         vm_offset_t va;
 952         vm_paddr_t pa;
 953         pmap_t pmap;
 954         int rv;
 955
 956         pmap = vm_map_pmap(map);
 957         start = entry->start;
 958         end = entry->end;
 959         fictitious = entry->object.vm_object &&
 960                         (entry->object.vm_object->type == OBJT_DEVICE);
 961
 962         vm_map_unlock(map);
 963         map->timestamp++;
 964
 965         /*
 966          * We simulate a fault to get the page and enter it in the physical
 967          * map.
 968          */
 969         for (va = start; va < end; va += PAGE_SIZE) {
 970                 if (user_wire) {
 971                         rv = vm_fault(map, va, VM_PROT_READ,
 972                                         VM_FAULT_USER_WIRE);
 973                 } else {
 974                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
 975                                         VM_FAULT_CHANGE_WIRING);
 976                 }
 977                 if (rv) {
 978                         while (va > start) {
 979                                 va -= PAGE_SIZE;
 980                                 if ((pa = pmap_extract(pmap, va)) == 0)
 981                                         continue;
 982                                 pmap_change_wiring(pmap, va, FALSE);
 983                                 if (!fictitious)
 984                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
 985                         }
 986                         vm_map_lock(map);
 987                         return (rv);
 988                 }
 989         }
 990         vm_map_lock(map);
 991         return (KERN_SUCCESS);
 992 }
 993
 994 /*
 995  * Unwire a range of virtual addresses in a map.  The map should be
 996  * locked.
 997  */
 998 void
 999 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1000 {
1001         boolean_t fictitious;
1002         vm_offset_t start;
1003         vm_offset_t end;
1004         vm_offset_t va;
1005         vm_paddr_t pa;
1006         pmap_t pmap;
1007
1008         pmap = vm_map_pmap(map);
1009         start = entry->start;
1010         end = entry->end;
1011         fictitious = entry->object.vm_object &&
1012                         (entry->object.vm_object->type == OBJT_DEVICE);
1013
1014         /*
1015          * Since the pages are wired down, we must be able to get their
1016          * mappings from the physical map system.
1017          */
1018         for (va = start; va < end; va += PAGE_SIZE) {
1019                 pa = pmap_extract(pmap, va);
1020                 if (pa != 0) {
1021                         pmap_change_wiring(pmap, va, FALSE);
1022                         if (!fictitious)
1023                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1024                 }
1025         }
1026 }
1027
1028 /*
1029  *      Routine:
1030  *              vm_fault_copy_entry
1031  *      Function:
1032  *              Copy all of the pages from a wired-down map entry to another.
1033  *
1034  *      In/out conditions:
1035  *              The source and destination maps must be locked for write.
1036  *              The source map entry must be wired down (or be a sharing map
1037  *              entry corresponding to a main map entry that is wired down).
1038  */
1039
1040 void
1041 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1042     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1043 {
1044         vm_object_t dst_object;
1045         vm_object_t src_object;
1046         vm_ooffset_t dst_offset;
1047         vm_ooffset_t src_offset;
1048         vm_prot_t prot;
1049         vm_offset_t vaddr;
1050         vm_page_t dst_m;
1051         vm_page_t src_m;
1052
1053 #ifdef  lint
1054         src_map++;
1055 #endif  /* lint */
1056
1057         src_object = src_entry->object.vm_object;
1058         src_offset = src_entry->offset;
1059
1060         /*
1061          * Create the top-level object for the destination entry. (Doesn't
1062          * actually shadow anything - we copy the pages directly.)
1063          */
1064         dst_object = vm_object_allocate(OBJT_DEFAULT,
1065             (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1066
1067         dst_entry->object.vm_object = dst_object;
1068         dst_entry->offset = 0;
1069
1070         prot = dst_entry->max_protection;
1071
1072         /*
1073          * Loop through all of the pages in the entry's range, copying each
1074          * one from the source object (it should be there) to the destination
1075          * object.
1076          */
1077         for (vaddr = dst_entry->start, dst_offset = 0;
1078             vaddr < dst_entry->end;
1079             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1080
1081                 /*
1082                  * Allocate a page in the destination object
1083                  */
1084                 do {
1085                         dst_m = vm_page_alloc(dst_object,
1086                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1087                         if (dst_m == NULL) {
1088                                 vm_wait();
1089                         }
1090                 } while (dst_m == NULL);
1091
1092                 /*
1093                  * Find the page in the source object, and copy it in.
1094                  * (Because the source is wired down, the page will be in
1095                  * memory.)
1096                  */
1097                 src_m = vm_page_lookup(src_object,
1098                         OFF_TO_IDX(dst_offset + src_offset));
1099                 if (src_m == NULL)
1100                         panic("vm_fault_copy_wired: page missing");
1101
1102                 vm_page_copy(src_m, dst_m);
1103
1104                 /*
1105                  * Enter it in the pmap...
1106                  */
1107
1108                 vm_page_flag_clear(dst_m, PG_ZERO);
1109                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1110                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1111
1112                 /*
1113                  * Mark it no longer busy, and put it on the active list.
1114                  */
1115                 vm_page_activate(dst_m);
1116                 vm_page_wakeup(dst_m);
1117         }
1118 }
1119
1120
1121 /*
1122  * This routine checks around the requested page for other pages that
1123  * might be able to be faulted in.  This routine brackets the viable
1124  * pages for the pages to be paged in.
1125  *
1126  * Inputs:
1127  *      m, rbehind, rahead
1128  *
1129  * Outputs:
1130  *  marray (array of vm_page_t), reqpage (index of requested page)
1131  *
1132  * Return value:
1133  *  number of pages in marray
1134  */
1135 static int
1136 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1137     vm_page_t *marray, int *reqpage)
1138 {
1139         int i,j;
1140         vm_object_t object;
1141         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1142         vm_page_t rtm;
1143         int cbehind, cahead;
1144
1145         object = m->object;
1146         pindex = m->pindex;
1147
1148         /*
1149          * we don't fault-ahead for device pager
1150          */
1151         if (object->type == OBJT_DEVICE) {
1152                 *reqpage = 0;
1153                 marray[0] = m;
1154                 return 1;
1155         }
1156
1157         /*
1158          * if the requested page is not available, then give up now
1159          */
1160
1161         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1162                 return 0;
1163         }
1164
1165         if ((cbehind == 0) && (cahead == 0)) {
1166                 *reqpage = 0;
1167                 marray[0] = m;
1168                 return 1;
1169         }
1170
1171         if (rahead > cahead) {
1172                 rahead = cahead;
1173         }
1174
1175         if (rbehind > cbehind) {
1176                 rbehind = cbehind;
1177         }
1178
1179         /*
1180          * try to do any readahead that we might have free pages for.
1181          */
1182         if ((rahead + rbehind) >
1183                 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) {
1184                 pagedaemon_wakeup();
1185                 marray[0] = m;
1186                 *reqpage = 0;
1187                 return 1;
1188         }
1189
1190         /*
1191          * scan backward for the read behind pages -- in memory
1192          *
1193          * Assume that if the page is not found an interrupt will not
1194          * create it.  Theoretically interrupts can only remove (busy)
1195          * pages, not create new associations.
1196          */
1197         if (pindex > 0) {
1198                 if (rbehind > pindex) {
1199                         rbehind = pindex;
1200                         startpindex = 0;
1201                 } else {
1202                         startpindex = pindex - rbehind;
1203                 }
1204
1205                 crit_enter();
1206                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1207                         if (vm_page_lookup( object, tpindex)) {
1208                                 startpindex = tpindex + 1;
1209                                 break;
1210                         }
1211                         if (tpindex == 0)
1212                                 break;
1213                 }
1214
1215                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1216
1217                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1218                         if (rtm == NULL) {
1219                                 crit_exit();
1220                                 for (j = 0; j < i; j++) {
1221                                         vm_page_free(marray[j]);
1222                                 }
1223                                 marray[0] = m;
1224                                 *reqpage = 0;
1225                                 return 1;
1226                         }
1227
1228                         marray[i] = rtm;
1229                 }
1230                 crit_exit();
1231         } else {
1232                 startpindex = 0;
1233                 i = 0;
1234         }
1235
1236         marray[i] = m;
1237         /* page offset of the required page */
1238         *reqpage = i;
1239
1240         tpindex = pindex + 1;
1241         i++;
1242
1243         /*
1244          * scan forward for the read ahead pages
1245          */
1246         endpindex = tpindex + rahead;
1247         if (endpindex > object->size)
1248                 endpindex = object->size;
1249
1250         crit_enter();
1251         for( ; tpindex < endpindex; i++, tpindex++) {
1252
1253                 if (vm_page_lookup(object, tpindex)) {
1254                         break;
1255                 }
1256
1257                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1258                 if (rtm == NULL) {
1259                         break;
1260                 }
1261
1262                 marray[i] = rtm;
1263         }
1264         crit_exit();
1265
1266         /* return number of bytes of pages */
1267         return i;
1268 }