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 1.22 2006/04/23 03:08:04 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/kernel.h>
  80 #include <sys/proc.h>
  81 #include <sys/vnode.h>
  82 #include <sys/resourcevar.h>
  83 #include <sys/vmmeter.h>
  84
  85 #include <vm/vm.h>
  86 #include <vm/vm_param.h>
  87 #include <sys/lock.h>
  88 #include <vm/pmap.h>
  89 #include <vm/vm_map.h>
  90 #include <vm/vm_object.h>
  91 #include <vm/vm_page.h>
  92 #include <vm/vm_pageout.h>
  93 #include <vm/vm_kern.h>
  94 #include <vm/vm_pager.h>
  95 #include <vm/vnode_pager.h>
  96 #include <vm/vm_extern.h>
  97
  98 #include <sys/thread2.h>
  99 #include <vm/vm_page2.h>
 100
 101 static int vm_fault_additional_pages (vm_page_t, int,
 102                                           int, vm_page_t *, int *);
 103 static int vm_fault_ratelimit(struct vmspace *vmspace);
 104
 105 #define VM_FAULT_READ_AHEAD 8
 106 #define VM_FAULT_READ_BEHIND 7
 107 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
 108
 109 struct faultstate {
 110         vm_page_t m;
 111         vm_object_t object;
 112         vm_pindex_t pindex;
 113         vm_page_t first_m;
 114         vm_object_t     first_object;
 115         vm_pindex_t first_pindex;
 116         vm_map_t map;
 117         vm_map_entry_t entry;
 118         int lookup_still_valid;
 119         struct vnode *vp;
 120 };
 121
 122 static __inline void
 123 release_page(struct faultstate *fs)
 124 {
 125         vm_page_wakeup(fs->m);
 126         vm_page_deactivate(fs->m);
 127         fs->m = NULL;
 128 }
 129
 130 static __inline void
 131 unlock_map(struct faultstate *fs)
 132 {
 133         if (fs->lookup_still_valid) {
 134                 vm_map_lookup_done(fs->map, fs->entry, 0);
 135                 fs->lookup_still_valid = FALSE;
 136         }
 137 }
 138
 139 static void
 140 _unlock_things(struct faultstate *fs, int dealloc)
 141 {
 142         vm_object_pip_wakeup(fs->object);
 143         if (fs->object != fs->first_object) {
 144                 vm_page_free(fs->first_m);
 145                 vm_object_pip_wakeup(fs->first_object);
 146                 fs->first_m = NULL;
 147         }
 148         if (dealloc) {
 149                 vm_object_deallocate(fs->first_object);
 150         }
 151         unlock_map(fs);
 152         if (fs->vp != NULL) {
 153                 vput(fs->vp);
 154                 fs->vp = NULL;
 155         }
 156 }
 157
 158 #define unlock_things(fs) _unlock_things(fs, 0)
 159 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
 160
 161 /*
 162  * TRYPAGER - used by vm_fault to calculate whether the pager for the
 163  *            current object *might* contain the page.
 164  *
 165  *            default objects are zero-fill, there is no real pager.
 166  */
 167
 168 #define TRYPAGER        (fs.object->type != OBJT_DEFAULT && \
 169                         (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
 170
 171 /*
 172  *      vm_fault:
 173  *
 174  *      Handle a page fault occurring at the given address,
 175  *      requiring the given permissions, in the map specified.
 176  *      If successful, the page is inserted into the
 177  *      associated physical map.
 178  *
 179  *      NOTE: the given address should be truncated to the
 180  *      proper page address.
 181  *
 182  *      KERN_SUCCESS is returned if the page fault is handled; otherwise,
 183  *      a standard error specifying why the fault is fatal is returned.
 184  *
 185  *
 186  *      The map in question must be referenced, and remains so.
 187  *      Caller may hold no locks.
 188  */
 189 int
 190 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
 191 {
 192         vm_prot_t prot;
 193         int result;
 194         boolean_t wired;
 195         int map_generation;
 196         vm_object_t next_object;
 197         vm_page_t marray[VM_FAULT_READ];
 198         int hardfault;
 199         int faultcount;
 200         int limticks;
 201         int didlimit = 0;
 202         struct faultstate fs;
 203
 204         mycpu->gd_cnt.v_vm_faults++;
 205         hardfault = 0;
 206
 207 RetryFault:
 208         /*
 209          * Find the backing store object and offset into it to begin the
 210          * search.
 211          */
 212         fs.map = map;
 213         if ((result = vm_map_lookup(&fs.map, vaddr,
 214                 fault_type, &fs.entry, &fs.first_object,
 215                 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
 216                 if ((result != KERN_PROTECTION_FAILURE) ||
 217                         ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
 218                         return result;
 219                 }
 220
 221                 /*
 222                  * If we are user-wiring a r/w segment, and it is COW, then
 223                  * we need to do the COW operation.  Note that we don't COW
 224                  * currently RO sections now, because it is NOT desirable
 225                  * to COW .text.  We simply keep .text from ever being COW'ed
 226                  * and take the heat that one cannot debug wired .text sections.
 227                  */
 228                 result = vm_map_lookup(&fs.map, vaddr,
 229                         VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
 230                         &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
 231                 if (result != KERN_SUCCESS) {
 232                         return result;
 233                 }
 234
 235                 /*
 236                  * If we don't COW now, on a user wire, the user will never
 237                  * be able to write to the mapping.  If we don't make this
 238                  * restriction, the bookkeeping would be nearly impossible.
 239                  */
 240                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
 241                         fs.entry->max_protection &= ~VM_PROT_WRITE;
 242         }
 243
 244         map_generation = fs.map->timestamp;
 245
 246         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
 247                 panic("vm_fault: fault on nofault entry, addr: %lx",
 248                     (u_long)vaddr);
 249         }
 250
 251         /*
 252          * A system map entry may return a NULL object.  No object means
 253          * no pager means an unrecoverable kernel fault.
 254          */
 255         if (fs.first_object == NULL) {
 256                 panic("vm_fault: unrecoverable fault at %p in entry %p",
 257                         (void *)vaddr, fs.entry);
 258         }
 259
 260         /*
 261          * Make a reference to this object to prevent its disposal while we
 262          * are messing with it.  Once we have the reference, the map is free
 263          * to be diddled.  Since objects reference their shadows (and copies),
 264          * they will stay around as well.
 265          *
 266          * Bump the paging-in-progress count to prevent size changes (e.g.
 267          * truncation operations) during I/O.  This must be done after
 268          * obtaining the vnode lock in order to avoid possible deadlocks.
 269          */
 270         vm_object_reference(fs.first_object);
 271         fs.vp = vnode_pager_lock(fs.first_object);
 272         vm_object_pip_add(fs.first_object, 1);
 273
 274         if ((fault_type & VM_PROT_WRITE) &&
 275                 (fs.first_object->type == OBJT_VNODE)) {
 276                 vm_freeze_copyopts(fs.first_object,
 277                         fs.first_pindex, fs.first_pindex + 1);
 278         }
 279
 280         fs.lookup_still_valid = TRUE;
 281
 282         if (wired)
 283                 fault_type = prot;
 284
 285         fs.first_m = NULL;
 286
 287         /*
 288          * Search for the page at object/offset.
 289          */
 290
 291         fs.object = fs.first_object;
 292         fs.pindex = fs.first_pindex;
 293
 294         while (TRUE) {
 295                 /*
 296                  * If the object is dead, we stop here
 297                  */
 298
 299                 if (fs.object->flags & OBJ_DEAD) {
 300                         unlock_and_deallocate(&fs);
 301                         return (KERN_PROTECTION_FAILURE);
 302                 }
 303
 304                 /*
 305                  * See if page is resident.  spl protection is required
 306                  * to avoid an interrupt unbusy/free race against our
 307                  * lookup.  We must hold the protection through a page
 308                  * allocation or busy.
 309                  */
 310                 crit_enter();
 311                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 312                 if (fs.m != NULL) {
 313                         int queue;
 314                         /*
 315                          * Wait/Retry if the page is busy.  We have to do this
 316                          * if the page is busy via either PG_BUSY or
 317                          * vm_page_t->busy because the vm_pager may be using
 318                          * vm_page_t->busy for pageouts ( and even pageins if
 319                          * it is the vnode pager ), and we could end up trying
 320                          * to pagein and pageout the same page simultaneously.
 321                          *
 322                          * We can theoretically allow the busy case on a read
 323                          * fault if the page is marked valid, but since such
 324                          * pages are typically already pmap'd, putting that
 325                          * special case in might be more effort then it is
 326                          * worth.  We cannot under any circumstances mess
 327                          * around with a vm_page_t->busy page except, perhaps,
 328                          * to pmap it.
 329                          */
 330                         if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
 331                                 unlock_things(&fs);
 332                                 vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
 333                                 mycpu->gd_cnt.v_intrans++;
 334                                 vm_object_deallocate(fs.first_object);
 335                                 crit_exit();
 336                                 goto RetryFault;
 337                         }
 338
 339                         queue = fs.m->queue;
 340                         vm_page_unqueue_nowakeup(fs.m);
 341
 342                         if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
 343                                 vm_page_activate(fs.m);
 344                                 unlock_and_deallocate(&fs);
 345                                 vm_waitpfault();
 346                                 crit_exit();
 347                                 goto RetryFault;
 348                         }
 349
 350                         /*
 351                          * Mark page busy for other processes, and the
 352                          * pagedaemon.  If it still isn't completely valid
 353                          * (readable), jump to readrest, else break-out ( we
 354                          * found the page ).
 355                          *
 356                          * We can release the spl once we have marked the
 357                          * page busy.
 358                          */
 359
 360                         vm_page_busy(fs.m);
 361                         crit_exit();
 362
 363                         if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
 364                                 fs.m->object != kernel_object && fs.m->object != kmem_object) {
 365                                 goto readrest;
 366                         }
 367
 368                         break;
 369                 }
 370
 371                 /*
 372                  * Page is not resident, If this is the search termination
 373                  * or the pager might contain the page, allocate a new page.
 374                  *
 375                  * note: we are still in splvm().
 376                  */
 377
 378                 if (TRYPAGER || fs.object == fs.first_object) {
 379                         if (fs.pindex >= fs.object->size) {
 380                                 crit_exit();
 381                                 unlock_and_deallocate(&fs);
 382                                 return (KERN_PROTECTION_FAILURE);
 383                         }
 384
 385                         /*
 386                          * Ratelimit.
 387                          */
 388                         if (didlimit == 0) {
 389                                 limticks =
 390                                         vm_fault_ratelimit(curproc->p_vmspace);
 391                                 if (limticks) {
 392                                         crit_exit();
 393                                         unlock_and_deallocate(&fs);
 394                                         tsleep(curproc, 0, "vmrate", limticks);
 395                                         didlimit = 1;
 396                                         goto RetryFault;
 397                                 }
 398                         }
 399
 400                         /*
 401                          * Allocate a new page for this object/offset pair.
 402                          */
 403                         fs.m = NULL;
 404                         if (!vm_page_count_severe()) {
 405                                 fs.m = vm_page_alloc(fs.object, fs.pindex,
 406                                     (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
 407                         }
 408                         if (fs.m == NULL) {
 409                                 crit_exit();
 410                                 unlock_and_deallocate(&fs);
 411                                 vm_waitpfault();
 412                                 goto RetryFault;
 413                         }
 414                 }
 415                 crit_exit();
 416
 417 readrest:
 418                 /*
 419                  * We have found a valid page or we have allocated a new page.
 420                  * The page thus may not be valid or may not be entirely
 421                  * valid.
 422                  *
 423                  * Attempt to fault-in the page if there is a chance that the
 424                  * pager has it, and potentially fault in additional pages
 425                  * at the same time.
 426                  *
 427                  * We are NOT in splvm here and if TRYPAGER is true then
 428                  * fs.m will be non-NULL and will be PG_BUSY for us.
 429                  */
 430
 431                 if (TRYPAGER) {
 432                         int rv;
 433                         int reqpage;
 434                         int ahead, behind;
 435                         u_char behavior = vm_map_entry_behavior(fs.entry);
 436
 437                         if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
 438                                 ahead = 0;
 439                                 behind = 0;
 440                         } else {
 441                                 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
 442                                 if (behind > VM_FAULT_READ_BEHIND)
 443                                         behind = VM_FAULT_READ_BEHIND;
 444
 445                                 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
 446                                 if (ahead > VM_FAULT_READ_AHEAD)
 447                                         ahead = VM_FAULT_READ_AHEAD;
 448                         }
 449
 450                         if ((fs.first_object->type != OBJT_DEVICE) &&
 451                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
 452                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
 453                                 fs.pindex >= fs.entry->lastr &&
 454                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
 455                         ) {
 456                                 vm_pindex_t firstpindex, tmppindex;
 457
 458                                 if (fs.first_pindex < 2 * VM_FAULT_READ)
 459                                         firstpindex = 0;
 460                                 else
 461                                         firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
 462
 463                                 /*
 464                                  * note: partially valid pages cannot be
 465                                  * included in the lookahead - NFS piecemeal
 466                                  * writes will barf on it badly.
 467                                  *
 468                                  * spl protection is required to avoid races
 469                                  * between the lookup and an interrupt
 470                                  * unbusy/free sequence occuring prior to
 471                                  * our busy check.
 472                                  */
 473                                 crit_enter();
 474                                 for (tmppindex = fs.first_pindex - 1;
 475                                     tmppindex >= firstpindex;
 476                                     --tmppindex
 477                                 ) {
 478                                         vm_page_t mt;
 479                                         mt = vm_page_lookup( fs.first_object, tmppindex);
 480                                         if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
 481                                                 break;
 482                                         if (mt->busy ||
 483                                                 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
 484                                                 mt->hold_count ||
 485                                                 mt->wire_count)
 486                                                 continue;
 487                                         if (mt->dirty == 0)
 488                                                 vm_page_test_dirty(mt);
 489                                         if (mt->dirty) {
 490                                                 vm_page_protect(mt, VM_PROT_NONE);
 491                                                 vm_page_deactivate(mt);
 492                                         } else {
 493                                                 vm_page_cache(mt);
 494                                         }
 495                                 }
 496                                 crit_exit();
 497
 498                                 ahead += behind;
 499                                 behind = 0;
 500                         }
 501
 502                         /*
 503                          * now we find out if any other pages should be paged
 504                          * in at this time this routine checks to see if the
 505                          * pages surrounding this fault reside in the same
 506                          * object as the page for this fault.  If they do,
 507                          * then they are faulted in also into the object.  The
 508                          * array "marray" returned contains an array of
 509                          * vm_page_t structs where one of them is the
 510                          * vm_page_t passed to the routine.  The reqpage
 511                          * return value is the index into the marray for the
 512                          * vm_page_t passed to the routine.
 513                          *
 514                          * fs.m plus the additional pages are PG_BUSY'd.
 515                          */
 516                         faultcount = vm_fault_additional_pages(
 517                             fs.m, behind, ahead, marray, &reqpage);
 518
 519                         /*
 520                          * update lastr imperfectly (we do not know how much
 521                          * getpages will actually read), but good enough.
 522                          */
 523                         fs.entry->lastr = fs.pindex + faultcount - behind;
 524
 525                         /*
 526                          * Call the pager to retrieve the data, if any, after
 527                          * releasing the lock on the map.  We hold a ref on
 528                          * fs.object and the pages are PG_BUSY'd.
 529                          */
 530                         unlock_map(&fs);
 531
 532                         rv = faultcount ?
 533                             vm_pager_get_pages(fs.object, marray, faultcount,
 534                                 reqpage) : VM_PAGER_FAIL;
 535
 536                         if (rv == VM_PAGER_OK) {
 537                                 /*
 538                                  * Found the page. Leave it busy while we play
 539                                  * with it.
 540                                  */
 541
 542                                 /*
 543                                  * Relookup in case pager changed page. Pager
 544                                  * is responsible for disposition of old page
 545                                  * if moved.
 546                                  *
 547                                  * XXX other code segments do relookups too.
 548                                  * It's a bad abstraction that needs to be
 549                                  * fixed/removed.
 550                                  */
 551                                 fs.m = vm_page_lookup(fs.object, fs.pindex);
 552                                 if (fs.m == NULL) {
 553                                         unlock_and_deallocate(&fs);
 554                                         goto RetryFault;
 555                                 }
 556
 557                                 hardfault++;
 558                                 break; /* break to PAGE HAS BEEN FOUND */
 559                         }
 560                         /*
 561                          * Remove the bogus page (which does not exist at this
 562                          * object/offset); before doing so, we must get back
 563                          * our object lock to preserve our invariant.
 564                          *
 565                          * Also wake up any other process that may want to bring
 566                          * in this page.
 567                          *
 568                          * If this is the top-level object, we must leave the
 569                          * busy page to prevent another process from rushing
 570                          * past us, and inserting the page in that object at
 571                          * the same time that we are.
 572                          */
 573
 574                         if (rv == VM_PAGER_ERROR)
 575                                 printf("vm_fault: pager read error, pid %d (%s)\n",
 576                                     curproc->p_pid, curproc->p_comm);
 577                         /*
 578                          * Data outside the range of the pager or an I/O error
 579                          */
 580                         /*
 581                          * XXX - the check for kernel_map is a kludge to work
 582                          * around having the machine panic on a kernel space
 583                          * fault w/ I/O error.
 584                          */
 585                         if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
 586                                 (rv == VM_PAGER_BAD)) {
 587                                 vm_page_free(fs.m);
 588                                 fs.m = NULL;
 589                                 unlock_and_deallocate(&fs);
 590                                 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
 591                         }
 592                         if (fs.object != fs.first_object) {
 593                                 vm_page_free(fs.m);
 594                                 fs.m = NULL;
 595                                 /*
 596                                  * XXX - we cannot just fall out at this
 597                                  * point, m has been freed and is invalid!
 598                                  */
 599                         }
 600                 }
 601
 602                 /*
 603                  * We get here if the object has default pager (or unwiring)
 604                  * or the pager doesn't have the page.
 605                  */
 606                 if (fs.object == fs.first_object)
 607                         fs.first_m = fs.m;
 608
 609                 /*
 610                  * Move on to the next object.  Lock the next object before
 611                  * unlocking the current one.
 612                  */
 613
 614                 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
 615                 next_object = fs.object->backing_object;
 616                 if (next_object == NULL) {
 617                         /*
 618                          * If there's no object left, fill the page in the top
 619                          * object with zeros.
 620                          */
 621                         if (fs.object != fs.first_object) {
 622                                 vm_object_pip_wakeup(fs.object);
 623
 624                                 fs.object = fs.first_object;
 625                                 fs.pindex = fs.first_pindex;
 626                                 fs.m = fs.first_m;
 627                         }
 628                         fs.first_m = NULL;
 629
 630                         /*
 631                          * Zero the page if necessary and mark it valid.
 632                          */
 633                         if ((fs.m->flags & PG_ZERO) == 0) {
 634                                 vm_page_zero_fill(fs.m);
 635                         } else {
 636                                 mycpu->gd_cnt.v_ozfod++;
 637                         }
 638                         mycpu->gd_cnt.v_zfod++;
 639                         fs.m->valid = VM_PAGE_BITS_ALL;
 640                         break;  /* break to PAGE HAS BEEN FOUND */
 641                 } else {
 642                         if (fs.object != fs.first_object) {
 643                                 vm_object_pip_wakeup(fs.object);
 644                         }
 645                         KASSERT(fs.object != next_object, ("object loop %p", next_object));
 646                         fs.object = next_object;
 647                         vm_object_pip_add(fs.object, 1);
 648                 }
 649         }
 650
 651         KASSERT((fs.m->flags & PG_BUSY) != 0,
 652             ("vm_fault: not busy after main loop"));
 653
 654         /*
 655          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
 656          * is held.]
 657          */
 658
 659         /*
 660          * If the page is being written, but isn't already owned by the
 661          * top-level object, we have to copy it into a new page owned by the
 662          * top-level object.
 663          */
 664
 665         if (fs.object != fs.first_object) {
 666                 /*
 667                  * We only really need to copy if we want to write it.
 668                  */
 669
 670                 if (fault_type & VM_PROT_WRITE) {
 671                         /*
 672                          * This allows pages to be virtually copied from a
 673                          * backing_object into the first_object, where the
 674                          * backing object has no other refs to it, and cannot
 675                          * gain any more refs.  Instead of a bcopy, we just
 676                          * move the page from the backing object to the
 677                          * first object.  Note that we must mark the page
 678                          * dirty in the first object so that it will go out
 679                          * to swap when needed.
 680                          */
 681                         if (map_generation == fs.map->timestamp &&
 682                                 /*
 683                                  * Only one shadow object
 684                                  */
 685                                 (fs.object->shadow_count == 1) &&
 686                                 /*
 687                                  * No COW refs, except us
 688                                  */
 689                                 (fs.object->ref_count == 1) &&
 690                                 /*
 691                                  * No one else can look this object up
 692                                  */
 693                                 (fs.object->handle == NULL) &&
 694                                 /*
 695                                  * No other ways to look the object up
 696                                  */
 697                                 ((fs.object->type == OBJT_DEFAULT) ||
 698                                  (fs.object->type == OBJT_SWAP)) &&
 699                                 /*
 700                                  * We don't chase down the shadow chain
 701                                  */
 702                                 (fs.object == fs.first_object->backing_object) &&
 703
 704                                 /*
 705                                  * grab the lock if we need to
 706                                  */
 707                                 (fs.lookup_still_valid ||
 708                                  lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, curthread) == 0)
 709                             ) {
 710
 711                                 fs.lookup_still_valid = 1;
 712                                 /*
 713                                  * get rid of the unnecessary page
 714                                  */
 715                                 vm_page_protect(fs.first_m, VM_PROT_NONE);
 716                                 vm_page_free(fs.first_m);
 717                                 fs.first_m = NULL;
 718
 719                                 /*
 720                                  * grab the page and put it into the
 721                                  * process'es object.  The page is
 722                                  * automatically made dirty.
 723                                  */
 724                                 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
 725                                 fs.first_m = fs.m;
 726                                 vm_page_busy(fs.first_m);
 727                                 fs.m = NULL;
 728                                 mycpu->gd_cnt.v_cow_optim++;
 729                         } else {
 730                                 /*
 731                                  * Oh, well, lets copy it.
 732                                  */
 733                                 vm_page_copy(fs.m, fs.first_m);
 734                         }
 735
 736                         if (fs.m) {
 737                                 /*
 738                                  * We no longer need the old page or object.
 739                                  */
 740                                 release_page(&fs);
 741                         }
 742
 743                         /*
 744                          * fs.object != fs.first_object due to above
 745                          * conditional
 746                          */
 747
 748                         vm_object_pip_wakeup(fs.object);
 749
 750                         /*
 751                          * Only use the new page below...
 752                          */
 753
 754                         mycpu->gd_cnt.v_cow_faults++;
 755                         fs.m = fs.first_m;
 756                         fs.object = fs.first_object;
 757                         fs.pindex = fs.first_pindex;
 758
 759                 } else {
 760                         prot &= ~VM_PROT_WRITE;
 761                 }
 762         }
 763
 764         /*
 765          * We must verify that the maps have not changed since our last
 766          * lookup.
 767          */
 768
 769         if (!fs.lookup_still_valid &&
 770                 (fs.map->timestamp != map_generation)) {
 771                 vm_object_t retry_object;
 772                 vm_pindex_t retry_pindex;
 773                 vm_prot_t retry_prot;
 774
 775                 /*
 776                  * Since map entries may be pageable, make sure we can take a
 777                  * page fault on them.
 778                  */
 779
 780                 /*
 781                  * Unlock vnode before the lookup to avoid deadlock.   E.G.
 782                  * avoid a deadlock between the inode and exec_map that can
 783                  * occur due to locks being obtained in different orders.
 784                  */
 785
 786                 if (fs.vp != NULL) {
 787                         vput(fs.vp);
 788                         fs.vp = NULL;
 789                 }
 790
 791                 if (fs.map->infork) {
 792                         release_page(&fs);
 793                         unlock_and_deallocate(&fs);
 794                         goto RetryFault;
 795                 }
 796
 797                 /*
 798                  * To avoid trying to write_lock the map while another process
 799                  * has it read_locked (in vm_map_wire), we do not try for
 800                  * write permission.  If the page is still writable, we will
 801                  * get write permission.  If it is not, or has been marked
 802                  * needs_copy, we enter the mapping without write permission,
 803                  * and will merely take another fault.
 804                  */
 805                 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
 806                     &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
 807                 map_generation = fs.map->timestamp;
 808
 809                 /*
 810                  * If we don't need the page any longer, put it on the active
 811                  * list (the easiest thing to do here).  If no one needs it,
 812                  * pageout will grab it eventually.
 813                  */
 814
 815                 if (result != KERN_SUCCESS) {
 816                         release_page(&fs);
 817                         unlock_and_deallocate(&fs);
 818                         return (result);
 819                 }
 820                 fs.lookup_still_valid = TRUE;
 821
 822                 if ((retry_object != fs.first_object) ||
 823                     (retry_pindex != fs.first_pindex)) {
 824                         release_page(&fs);
 825                         unlock_and_deallocate(&fs);
 826                         goto RetryFault;
 827                 }
 828                 /*
 829                  * Check whether the protection has changed or the object has
 830                  * been copied while we left the map unlocked. Changing from
 831                  * read to write permission is OK - we leave the page
 832                  * write-protected, and catch the write fault. Changing from
 833                  * write to read permission means that we can't mark the page
 834                  * write-enabled after all.
 835                  */
 836                 prot &= retry_prot;
 837         }
 838
 839         /*
 840          * Put this page into the physical map. We had to do the unlock above
 841          * because pmap_enter may cause other faults.   We don't put the page
 842          * back on the active queue until later so that the page-out daemon
 843          * won't find us (yet).
 844          */
 845
 846         if (prot & VM_PROT_WRITE) {
 847                 vm_page_flag_set(fs.m, PG_WRITEABLE);
 848                 vm_object_set_writeable_dirty(fs.m->object);
 849
 850                 /*
 851                  * If the fault is a write, we know that this page is being
 852                  * written NOW so dirty it explicitly to save on
 853                  * pmap_is_modified() calls later.
 854                  *
 855                  * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
 856                  * if the page is already dirty to prevent data written with
 857                  * the expectation of being synced from not being synced.
 858                  * Likewise if this entry does not request NOSYNC then make
 859                  * sure the page isn't marked NOSYNC.  Applications sharing
 860                  * data should use the same flags to avoid ping ponging.
 861                  *
 862                  * Also tell the backing pager, if any, that it should remove
 863                  * any swap backing since the page is now dirty.
 864                  */
 865                 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
 866                         if (fs.m->dirty == 0)
 867                                 vm_page_flag_set(fs.m, PG_NOSYNC);
 868                 } else {
 869                         vm_page_flag_clear(fs.m, PG_NOSYNC);
 870                 }
 871                 if (fault_flags & VM_FAULT_DIRTY) {
 872                         crit_enter();
 873                         vm_page_dirty(fs.m);
 874                         vm_pager_page_unswapped(fs.m);
 875                         crit_exit();
 876                 }
 877         }
 878
 879         /*
 880          * Page had better still be busy
 881          */
 882
 883         KASSERT(fs.m->flags & PG_BUSY,
 884                 ("vm_fault: page %p not busy!", fs.m));
 885
 886         unlock_things(&fs);
 887
 888         /*
 889          * Sanity check: page must be completely valid or it is not fit to
 890          * map into user space.  vm_pager_get_pages() ensures this.
 891          */
 892
 893         if (fs.m->valid != VM_PAGE_BITS_ALL) {
 894                 vm_page_zero_invalid(fs.m, TRUE);
 895                 printf("Warning: page %p partially invalid on fault\n", fs.m);
 896         }
 897
 898         pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
 899
 900         if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
 901                 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
 902         }
 903
 904         vm_page_flag_clear(fs.m, PG_ZERO);
 905         vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
 906         if (fault_flags & VM_FAULT_HOLD)
 907                 vm_page_hold(fs.m);
 908
 909         /*
 910          * If the page is not wired down, then put it where the pageout daemon
 911          * can find it.
 912          */
 913
 914         if (fault_flags & VM_FAULT_WIRE_MASK) {
 915                 if (wired)
 916                         vm_page_wire(fs.m);
 917                 else
 918                         vm_page_unwire(fs.m, 1);
 919         } else {
 920                 vm_page_activate(fs.m);
 921         }
 922
 923         if (curproc && (curproc->p_flag & P_SWAPPEDOUT) == 0 &&
 924             curproc->p_stats) {
 925                 if (hardfault) {
 926                         curproc->p_stats->p_ru.ru_majflt++;
 927                 } else {
 928                         curproc->p_stats->p_ru.ru_minflt++;
 929                 }
 930         }
 931
 932         /*
 933          * Unlock everything, and return
 934          */
 935
 936         vm_page_wakeup(fs.m);
 937         vm_object_deallocate(fs.first_object);
 938
 939         return (KERN_SUCCESS);
 940
 941 }
 942
 943 /*
 944  * quick version of vm_fault
 945  */
 946 int
 947 vm_fault_quick(caddr_t v, int prot)
 948 {
 949         int r;
 950
 951         if (prot & VM_PROT_WRITE)
 952                 r = subyte(v, fubyte(v));
 953         else
 954                 r = fubyte(v);
 955         return(r);
 956 }
 957
 958 /*
 959  * Wire down a range of virtual addresses in a map.  The entry in question
 960  * should be marked in-transition and the map must be locked.  We must
 961  * release the map temporarily while faulting-in the page to avoid a
 962  * deadlock.  Note that the entry may be clipped while we are blocked but
 963  * will never be freed.
 964  */
 965 int
 966 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
 967 {
 968         boolean_t fictitious;
 969         vm_offset_t start;
 970         vm_offset_t end;
 971         vm_offset_t va;
 972         vm_paddr_t pa;
 973         pmap_t pmap;
 974         int rv;
 975
 976         pmap = vm_map_pmap(map);
 977         start = entry->start;
 978         end = entry->end;
 979         fictitious = entry->object.vm_object &&
 980                         (entry->object.vm_object->type == OBJT_DEVICE);
 981
 982         vm_map_unlock(map);
 983         map->timestamp++;
 984
 985         /*
 986          * We simulate a fault to get the page and enter it in the physical
 987          * map.
 988          */
 989         for (va = start; va < end; va += PAGE_SIZE) {
 990                 if (user_wire) {
 991                         rv = vm_fault(map, va, VM_PROT_READ,
 992                                         VM_FAULT_USER_WIRE);
 993                 } else {
 994                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
 995                                         VM_FAULT_CHANGE_WIRING);
 996                 }
 997                 if (rv) {
 998                         while (va > start) {
 999                                 va -= PAGE_SIZE;
1000                                 if ((pa = pmap_extract(pmap, va)) == 0)
1001                                         continue;
1002                                 pmap_change_wiring(pmap, va, FALSE);
1003                                 if (!fictitious)
1004                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1005                         }
1006                         vm_map_lock(map);
1007                         return (rv);
1008                 }
1009         }
1010         vm_map_lock(map);
1011         return (KERN_SUCCESS);
1012 }
1013
1014 /*
1015  * Unwire a range of virtual addresses in a map.  The map should be
1016  * locked.
1017  */
1018 void
1019 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1020 {
1021         boolean_t fictitious;
1022         vm_offset_t start;
1023         vm_offset_t end;
1024         vm_offset_t va;
1025         vm_paddr_t pa;
1026         pmap_t pmap;
1027
1028         pmap = vm_map_pmap(map);
1029         start = entry->start;
1030         end = entry->end;
1031         fictitious = entry->object.vm_object &&
1032                         (entry->object.vm_object->type == OBJT_DEVICE);
1033
1034         /*
1035          * Since the pages are wired down, we must be able to get their
1036          * mappings from the physical map system.
1037          */
1038         for (va = start; va < end; va += PAGE_SIZE) {
1039                 pa = pmap_extract(pmap, va);
1040                 if (pa != 0) {
1041                         pmap_change_wiring(pmap, va, FALSE);
1042                         if (!fictitious)
1043                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1044                 }
1045         }
1046 }
1047
1048 /*
1049  * Reduce the rate at which memory is allocated to a process based
1050  * on the perceived load on the VM system. As the load increases
1051  * the allocation burst rate goes down and the delay increases.
1052  *
1053  * Rate limiting does not apply when faulting active or inactive
1054  * pages.  When faulting 'cache' pages, rate limiting only applies
1055  * if the system currently has a severe page deficit.
1056  *
1057  * XXX vm_pagesupply should be increased when a page is freed.
1058  *
1059  * We sleep up to 1/10 of a second.
1060  */
1061 static int
1062 vm_fault_ratelimit(struct vmspace *vmspace)
1063 {
1064         if (vm_load_enable == 0)
1065                 return(0);
1066         if (vmspace->vm_pagesupply > 0) {
1067                 --vmspace->vm_pagesupply;
1068                 return(0);
1069         }
1070 #ifdef INVARIANTS
1071         if (vm_load_debug) {
1072                 printf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1073                         vm_load,
1074                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1075                         curproc->p_pid, curproc->p_comm);
1076         }
1077 #endif
1078         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1079         return(vm_load * hz / 10000);
1080 }
1081
1082 /*
1083  *      Routine:
1084  *              vm_fault_copy_entry
1085  *      Function:
1086  *              Copy all of the pages from a wired-down map entry to another.
1087  *
1088  *      In/out conditions:
1089  *              The source and destination maps must be locked for write.
1090  *              The source map entry must be wired down (or be a sharing map
1091  *              entry corresponding to a main map entry that is wired down).
1092  */
1093
1094 void
1095 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1096     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1097 {
1098         vm_object_t dst_object;
1099         vm_object_t src_object;
1100         vm_ooffset_t dst_offset;
1101         vm_ooffset_t src_offset;
1102         vm_prot_t prot;
1103         vm_offset_t vaddr;
1104         vm_page_t dst_m;
1105         vm_page_t src_m;
1106
1107 #ifdef  lint
1108         src_map++;
1109 #endif  /* lint */
1110
1111         src_object = src_entry->object.vm_object;
1112         src_offset = src_entry->offset;
1113
1114         /*
1115          * Create the top-level object for the destination entry. (Doesn't
1116          * actually shadow anything - we copy the pages directly.)
1117          */
1118         dst_object = vm_object_allocate(OBJT_DEFAULT,
1119             (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1120
1121         dst_entry->object.vm_object = dst_object;
1122         dst_entry->offset = 0;
1123
1124         prot = dst_entry->max_protection;
1125
1126         /*
1127          * Loop through all of the pages in the entry's range, copying each
1128          * one from the source object (it should be there) to the destination
1129          * object.
1130          */
1131         for (vaddr = dst_entry->start, dst_offset = 0;
1132             vaddr < dst_entry->end;
1133             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1134
1135                 /*
1136                  * Allocate a page in the destination object
1137                  */
1138                 do {
1139                         dst_m = vm_page_alloc(dst_object,
1140                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1141                         if (dst_m == NULL) {
1142                                 vm_wait();
1143                         }
1144                 } while (dst_m == NULL);
1145
1146                 /*
1147                  * Find the page in the source object, and copy it in.
1148                  * (Because the source is wired down, the page will be in
1149                  * memory.)
1150                  */
1151                 src_m = vm_page_lookup(src_object,
1152                         OFF_TO_IDX(dst_offset + src_offset));
1153                 if (src_m == NULL)
1154                         panic("vm_fault_copy_wired: page missing");
1155
1156                 vm_page_copy(src_m, dst_m);
1157
1158                 /*
1159                  * Enter it in the pmap...
1160                  */
1161
1162                 vm_page_flag_clear(dst_m, PG_ZERO);
1163                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1164                 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1165
1166                 /*
1167                  * Mark it no longer busy, and put it on the active list.
1168                  */
1169                 vm_page_activate(dst_m);
1170                 vm_page_wakeup(dst_m);
1171         }
1172 }
1173
1174
1175 /*
1176  * This routine checks around the requested page for other pages that
1177  * might be able to be faulted in.  This routine brackets the viable
1178  * pages for the pages to be paged in.
1179  *
1180  * Inputs:
1181  *      m, rbehind, rahead
1182  *
1183  * Outputs:
1184  *  marray (array of vm_page_t), reqpage (index of requested page)
1185  *
1186  * Return value:
1187  *  number of pages in marray
1188  */
1189 static int
1190 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1191     vm_page_t *marray, int *reqpage)
1192 {
1193         int i,j;
1194         vm_object_t object;
1195         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1196         vm_page_t rtm;
1197         int cbehind, cahead;
1198
1199         object = m->object;
1200         pindex = m->pindex;
1201
1202         /*
1203          * we don't fault-ahead for device pager
1204          */
1205         if (object->type == OBJT_DEVICE) {
1206                 *reqpage = 0;
1207                 marray[0] = m;
1208                 return 1;
1209         }
1210
1211         /*
1212          * if the requested page is not available, then give up now
1213          */
1214
1215         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1216                 return 0;
1217         }
1218
1219         if ((cbehind == 0) && (cahead == 0)) {
1220                 *reqpage = 0;
1221                 marray[0] = m;
1222                 return 1;
1223         }
1224
1225         if (rahead > cahead) {
1226                 rahead = cahead;
1227         }
1228
1229         if (rbehind > cbehind) {
1230                 rbehind = cbehind;
1231         }
1232
1233         /*
1234          * try to do any readahead that we might have free pages for.
1235          */
1236         if ((rahead + rbehind) >
1237                 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) {
1238                 pagedaemon_wakeup();
1239                 marray[0] = m;
1240                 *reqpage = 0;
1241                 return 1;
1242         }
1243
1244         /*
1245          * scan backward for the read behind pages -- in memory
1246          *
1247          * Assume that if the page is not found an interrupt will not
1248          * create it.  Theoretically interrupts can only remove (busy)
1249          * pages, not create new associations.
1250          */
1251         if (pindex > 0) {
1252                 if (rbehind > pindex) {
1253                         rbehind = pindex;
1254                         startpindex = 0;
1255                 } else {
1256                         startpindex = pindex - rbehind;
1257                 }
1258
1259                 crit_enter();
1260                 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1261                         if (vm_page_lookup( object, tpindex)) {
1262                                 startpindex = tpindex + 1;
1263                                 break;
1264                         }
1265                         if (tpindex == 0)
1266                                 break;
1267                 }
1268
1269                 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1270
1271                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1272                         if (rtm == NULL) {
1273                                 crit_exit();
1274                                 for (j = 0; j < i; j++) {
1275                                         vm_page_free(marray[j]);
1276                                 }
1277                                 marray[0] = m;
1278                                 *reqpage = 0;
1279                                 return 1;
1280                         }
1281
1282                         marray[i] = rtm;
1283                 }
1284                 crit_exit();
1285         } else {
1286                 startpindex = 0;
1287                 i = 0;
1288         }
1289
1290         marray[i] = m;
1291         /* page offset of the required page */
1292         *reqpage = i;
1293
1294         tpindex = pindex + 1;
1295         i++;
1296
1297         /*
1298          * scan forward for the read ahead pages
1299          */
1300         endpindex = tpindex + rahead;
1301         if (endpindex > object->size)
1302                 endpindex = object->size;
1303
1304         crit_enter();
1305         for( ; tpindex < endpindex; i++, tpindex++) {
1306
1307                 if (vm_page_lookup(object, tpindex)) {
1308                         break;
1309                 }
1310
1311                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1312                 if (rtm == NULL) {
1313                         break;
1314                 }
1315
1316                 marray[i] = rtm;
1317         }
1318         crit_exit();
1319
1320         /* return number of bytes of pages */
1321         return i;
1322 }