sys/vm/vm_object.c

   1 /*
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * The Mach Operating System project at Carnegie-Mellon University.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. All advertising materials mentioning features or use of this software
  17  *    must display the following acknowledgement:
  18  *      This product includes software developed by the University of
  19  *      California, Berkeley and its contributors.
  20  * 4. Neither the name of the University nor the names of its contributors
  21  *    may be used to endorse or promote products derived from this software
  22  *    without specific prior written permission.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  *
  36  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  37  *
  38  *
  39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  40  * All rights reserved.
  41  *
  42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  43  *
  44  * Permission to use, copy, modify and distribute this software and
  45  * its documentation is hereby granted, provided that both the copyright
  46  * notice and this permission notice appear in all copies of the
  47  * software, derivative works or modified versions, and any portions
  48  * thereof, and that both notices appear in supporting documentation.
  49  *
  50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  53  *
  54  * Carnegie Mellon requests users of this software to return to
  55  *
  56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  57  *  School of Computer Science
  58  *  Carnegie Mellon University
  59  *  Pittsburgh PA 15213-3890
  60  *
  61  * any improvements or extensions that they make and grant Carnegie the
  62  * rights to redistribute these changes.
  63  *
  64  * $Id: vm_object.c,v 1.153 1999/03/14 06:36:00 alc Exp $
  65  */
  66
  67 /*
  68  *      Virtual memory object module.
  69  */
  70
  71 #include <sys/param.h>
  72 #include <sys/systm.h>
  73 #include <sys/proc.h>           /* for curproc, pageproc */
  74 #include <sys/vnode.h>
  75 #include <sys/vmmeter.h>
  76 #include <sys/mman.h>
  77 #include <sys/mount.h>
  78
  79 #include <vm/vm.h>
  80 #include <vm/vm_param.h>
  81 #include <vm/vm_prot.h>
  82 #include <vm/pmap.h>
  83 #include <vm/vm_map.h>
  84 #include <vm/vm_object.h>
  85 #include <vm/vm_page.h>
  86 #include <vm/vm_pageout.h>
  87 #include <vm/vm_pager.h>
  88 #include <vm/swap_pager.h>
  89 #include <vm/vm_kern.h>
  90 #include <vm/vm_extern.h>
  91 #include <vm/vm_zone.h>
  92
  93 static void     vm_object_qcollapse __P((vm_object_t object));
  94
  95 /*
  96  *      Virtual memory objects maintain the actual data
  97  *      associated with allocated virtual memory.  A given
  98  *      page of memory exists within exactly one object.
  99  *
 100  *      An object is only deallocated when all "references"
 101  *      are given up.  Only one "reference" to a given
 102  *      region of an object should be writeable.
 103  *
 104  *      Associated with each object is a list of all resident
 105  *      memory pages belonging to that object; this list is
 106  *      maintained by the "vm_page" module, and locked by the object's
 107  *      lock.
 108  *
 109  *      Each object also records a "pager" routine which is
 110  *      used to retrieve (and store) pages to the proper backing
 111  *      storage.  In addition, objects may be backed by other
 112  *      objects from which they were virtual-copied.
 113  *
 114  *      The only items within the object structure which are
 115  *      modified after time of creation are:
 116  *              reference count         locked by object's lock
 117  *              pager routine           locked by object's lock
 118  *
 119  */
 120
 121 struct object_q vm_object_list;
 122 #ifndef NULL_SIMPLELOCKS
 123 static struct simplelock vm_object_list_lock;
 124 #endif
 125 static long vm_object_count;            /* count of all objects */
 126 vm_object_t kernel_object;
 127 vm_object_t kmem_object;
 128 static struct vm_object kernel_object_store;
 129 static struct vm_object kmem_object_store;
 130 extern int vm_pageout_page_count;
 131
 132 static long object_collapses;
 133 static long object_bypasses;
 134 static int next_index;
 135 static vm_zone_t obj_zone;
 136 static struct vm_zone obj_zone_store;
 137 static int object_hash_rand;
 138 #define VM_OBJECTS_INIT 256
 139 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
 140
 141 void
 142 _vm_object_allocate(type, size, object)
 143         objtype_t type;
 144         vm_size_t size;
 145         vm_object_t object;
 146 {
 147         int incr;
 148         TAILQ_INIT(&object->memq);
 149         TAILQ_INIT(&object->shadow_head);
 150
 151         object->type = type;
 152         object->size = size;
 153         object->ref_count = 1;
 154         object->flags = 0;
 155         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
 156                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 157         object->behavior = OBJ_NORMAL;
 158         object->paging_in_progress = 0;
 159         object->resident_page_count = 0;
 160         object->cache_count = 0;
 161         object->wire_count = 0;
 162         object->shadow_count = 0;
 163         object->pg_color = next_index;
 164         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
 165                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
 166         else
 167                 incr = size;
 168         next_index = (next_index + incr) & PQ_L2_MASK;
 169         object->handle = NULL;
 170         object->backing_object = NULL;
 171         object->backing_object_offset = (vm_ooffset_t) 0;
 172         /*
 173          * Try to generate a number that will spread objects out in the
 174          * hash table.  We 'wipe' new objects across the hash in 128 page
 175          * increments plus 1 more to offset it a little more by the time
 176          * it wraps around.
 177          */
 178         object->hash_rand = object_hash_rand - 129;
 179
 180         object->last_read = 0;
 181         object->generation++;
 182
 183         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 184         vm_object_count++;
 185         object_hash_rand = object->hash_rand;
 186 }
 187
 188 /*
 189  *      vm_object_init:
 190  *
 191  *      Initialize the VM objects module.
 192  */
 193 void
 194 vm_object_init()
 195 {
 196         TAILQ_INIT(&vm_object_list);
 197         simple_lock_init(&vm_object_list_lock);
 198         vm_object_count = 0;
 199
 200         kernel_object = &kernel_object_store;
 201         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 202             kernel_object);
 203
 204         kmem_object = &kmem_object_store;
 205         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 206             kmem_object);
 207
 208         obj_zone = &obj_zone_store;
 209         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
 210                 vm_objects_init, VM_OBJECTS_INIT);
 211 }
 212
 213 void
 214 vm_object_init2() {
 215         zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
 216 }
 217
 218 /*
 219  *      vm_object_allocate:
 220  *
 221  *      Returns a new object with the given size.
 222  */
 223
 224 vm_object_t
 225 vm_object_allocate(type, size)
 226         objtype_t type;
 227         vm_size_t size;
 228 {
 229         vm_object_t result;
 230
 231         result = (vm_object_t) zalloc(obj_zone);
 232
 233         _vm_object_allocate(type, size, result);
 234
 235         return (result);
 236 }
 237
 238
 239 /*
 240  *      vm_object_reference:
 241  *
 242  *      Gets another reference to the given object.
 243  */
 244 void
 245 vm_object_reference(object)
 246         vm_object_t object;
 247 {
 248         if (object == NULL)
 249                 return;
 250
 251         KASSERT(!(object->flags & OBJ_DEAD),
 252             ("vm_object_reference: attempting to reference dead obj"));
 253
 254         object->ref_count++;
 255         if (object->type == OBJT_VNODE) {
 256                 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curproc)) {
 257 #if !defined(MAX_PERF)
 258                         printf("vm_object_reference: delay in getting object\n");
 259 #endif
 260                 }
 261         }
 262 }
 263
 264 void
 265 vm_object_vndeallocate(object)
 266         vm_object_t object;
 267 {
 268         struct vnode *vp = (struct vnode *) object->handle;
 269
 270         KASSERT(object->type == OBJT_VNODE,
 271             ("vm_object_vndeallocate: not a vnode object"));
 272         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 273 #ifdef INVARIANTS
 274         if (object->ref_count == 0) {
 275                 vprint("vm_object_vndeallocate", vp);
 276                 panic("vm_object_vndeallocate: bad object reference count");
 277         }
 278 #endif
 279
 280         object->ref_count--;
 281         if (object->ref_count == 0) {
 282                 vp->v_flag &= ~VTEXT;
 283                 vm_object_clear_flag(object, OBJ_OPT);
 284         }
 285         vrele(vp);
 286 }
 287
 288 /*
 289  *      vm_object_deallocate:
 290  *
 291  *      Release a reference to the specified object,
 292  *      gained either through a vm_object_allocate
 293  *      or a vm_object_reference call.  When all references
 294  *      are gone, storage associated with this object
 295  *      may be relinquished.
 296  *
 297  *      No object may be locked.
 298  */
 299 void
 300 vm_object_deallocate(object)
 301         vm_object_t object;
 302 {
 303         vm_object_t temp;
 304
 305         while (object != NULL) {
 306
 307                 if (object->type == OBJT_VNODE) {
 308                         vm_object_vndeallocate(object);
 309                         return;
 310                 }
 311
 312                 if (object->ref_count == 0) {
 313                         panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
 314                 } else if (object->ref_count > 2) {
 315                         object->ref_count--;
 316                         return;
 317                 }
 318
 319                 /*
 320                  * Here on ref_count of one or two, which are special cases for
 321                  * objects.
 322                  */
 323                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
 324                         vm_object_set_flag(object, OBJ_ONEMAPPING);
 325                         object->ref_count--;
 326                         return;
 327                 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
 328                         object->ref_count--;
 329                         if ((object->handle == NULL) &&
 330                             (object->type == OBJT_DEFAULT ||
 331                              object->type == OBJT_SWAP)) {
 332                                 vm_object_t robject;
 333
 334                                 robject = TAILQ_FIRST(&object->shadow_head);
 335                                 KASSERT(robject != NULL,
 336                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 337                                          object->ref_count,
 338                                          object->shadow_count));
 339                                 if ((robject->handle == NULL) &&
 340                                     (robject->type == OBJT_DEFAULT ||
 341                                      robject->type == OBJT_SWAP)) {
 342
 343                                         robject->ref_count++;
 344
 345                                         while (
 346                                                 robject->paging_in_progress ||
 347                                                 object->paging_in_progress
 348                                         ) {
 349                                                 vm_object_pip_sleep(robject, "objde1");
 350                                                 vm_object_pip_sleep(object, "objde2");
 351                                         }
 352
 353                                         if (robject->ref_count == 1) {
 354                                                 robject->ref_count--;
 355                                                 object = robject;
 356                                                 goto doterm;
 357                                         }
 358
 359                                         object = robject;
 360                                         vm_object_collapse(object);
 361                                         continue;
 362                                 }
 363                         }
 364
 365                         return;
 366
 367                 } else {
 368                         object->ref_count--;
 369                         if (object->ref_count != 0)
 370                                 return;
 371                 }
 372
 373 doterm:
 374
 375                 temp = object->backing_object;
 376                 if (temp) {
 377                         TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
 378                         temp->shadow_count--;
 379                         if (temp->ref_count == 0)
 380                                 vm_object_clear_flag(temp, OBJ_OPT);
 381                         temp->generation++;
 382                         object->backing_object = NULL;
 383                 }
 384                 vm_object_terminate(object);
 385                 /* unlocks and deallocates object */
 386                 object = temp;
 387         }
 388 }
 389
 390 /*
 391  *      vm_object_terminate actually destroys the specified object, freeing
 392  *      up all previously used resources.
 393  *
 394  *      The object must be locked.
 395  *      This routine may block.
 396  */
 397 void
 398 vm_object_terminate(object)
 399         vm_object_t object;
 400 {
 401         vm_page_t p;
 402         int s;
 403
 404         /*
 405          * Make sure no one uses us.
 406          */
 407         vm_object_set_flag(object, OBJ_DEAD);
 408
 409         /*
 410          * wait for the pageout daemon to be done with the object
 411          */
 412         vm_object_pip_wait(object, "objtrm");
 413
 414         KASSERT(!object->paging_in_progress,
 415                 ("vm_object_terminate: pageout in progress"));
 416
 417         /*
 418          * Clean and free the pages, as appropriate. All references to the
 419          * object are gone, so we don't need to lock it.
 420          */
 421         if (object->type == OBJT_VNODE) {
 422                 struct vnode *vp;
 423
 424                 /*
 425                  * Freeze optimized copies.
 426                  */
 427                 vm_freeze_copyopts(object, 0, object->size);
 428
 429                 /*
 430                  * Clean pages and flush buffers.
 431                  */
 432                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 433
 434                 vp = (struct vnode *) object->handle;
 435                 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
 436         }
 437
 438         if (object->ref_count != 0)
 439                 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
 440
 441         /*
 442          * Now free any remaining pages. For internal objects, this also
 443          * removes them from paging queues. Don't free wired pages, just
 444          * remove them from the object.
 445          */
 446         s = splvm();
 447         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
 448 #if !defined(MAX_PERF)
 449                 if (p->busy || (p->flags & PG_BUSY))
 450                         panic("vm_object_terminate: freeing busy page %p\n", p);
 451 #endif
 452                 if (p->wire_count == 0) {
 453                         vm_page_busy(p);
 454                         vm_page_free(p);
 455                         cnt.v_pfree++;
 456                 } else {
 457                         vm_page_busy(p);
 458                         vm_page_remove(p);
 459                 }
 460         }
 461         splx(s);
 462
 463         /*
 464          * Let the pager know object is dead.
 465          */
 466         vm_pager_deallocate(object);
 467
 468         /*
 469          * Remove the object from the global object list.
 470          */
 471         simple_lock(&vm_object_list_lock);
 472         TAILQ_REMOVE(&vm_object_list, object, object_list);
 473         simple_unlock(&vm_object_list_lock);
 474
 475         wakeup(object);
 476
 477         /*
 478          * Free the space for the object.
 479          */
 480         zfree(obj_zone, object);
 481 }
 482
 483 /*
 484  *      vm_object_page_clean
 485  *
 486  *      Clean all dirty pages in the specified range of object.
 487  *      Leaves page on whatever queue it is currently on.
 488  *
 489  *      Odd semantics: if start == end, we clean everything.
 490  *
 491  *      The object must be locked.
 492  */
 493
 494 void
 495 vm_object_page_clean(object, start, end, flags)
 496         vm_object_t object;
 497         vm_pindex_t start;
 498         vm_pindex_t end;
 499         int flags;
 500 {
 501         vm_page_t p, np, tp;
 502         vm_offset_t tstart, tend;
 503         vm_pindex_t pi;
 504         int s;
 505         struct vnode *vp;
 506         int runlen;
 507         int maxf;
 508         int chkb;
 509         int maxb;
 510         int i;
 511         int pagerflags;
 512         vm_page_t maf[vm_pageout_page_count];
 513         vm_page_t mab[vm_pageout_page_count];
 514         vm_page_t ma[vm_pageout_page_count];
 515         int curgeneration;
 516
 517         if (object->type != OBJT_VNODE ||
 518                 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
 519                 return;
 520
 521         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0;
 522         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
 523
 524         vp = object->handle;
 525
 526         vm_object_set_flag(object, OBJ_CLEANING);
 527
 528         tstart = start;
 529         if (end == 0) {
 530                 tend = object->size;
 531         } else {
 532                 tend = end;
 533         }
 534
 535         for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
 536                 vm_page_flag_set(p, PG_CLEANCHK);
 537                 vm_page_protect(p, VM_PROT_READ);
 538         }
 539
 540         if ((tstart == 0) && (tend == object->size)) {
 541                 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
 542         }
 543
 544 rescan:
 545         curgeneration = object->generation;
 546
 547         for(p = TAILQ_FIRST(&object->memq); p; p = np) {
 548                 np = TAILQ_NEXT(p, listq);
 549
 550                 pi = p->pindex;
 551                 if (((p->flags & PG_CLEANCHK) == 0) ||
 552                         (pi < tstart) || (pi >= tend) ||
 553                         (p->valid == 0) ||
 554                         ((p->queue - p->pc) == PQ_CACHE)) {
 555                         vm_page_flag_clear(p, PG_CLEANCHK);
 556                         continue;
 557                 }
 558
 559                 vm_page_test_dirty(p);
 560                 if ((p->dirty & p->valid) == 0) {
 561                         vm_page_flag_clear(p, PG_CLEANCHK);
 562                         continue;
 563                 }
 564
 565                 s = splvm();
 566                 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
 567                         if (object->generation != curgeneration) {
 568                                 splx(s);
 569                                 goto rescan;
 570                         }
 571                 }
 572
 573                 maxf = 0;
 574                 for(i=1;i<vm_pageout_page_count;i++) {
 575                         if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
 576                                 if ((tp->flags & PG_BUSY) ||
 577                                         (tp->flags & PG_CLEANCHK) == 0 ||
 578                                         (tp->busy != 0))
 579                                         break;
 580                                 if((tp->queue - tp->pc) == PQ_CACHE) {
 581                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 582                                         break;
 583                                 }
 584                                 vm_page_test_dirty(tp);
 585                                 if ((tp->dirty & tp->valid) == 0) {
 586                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 587                                         break;
 588                                 }
 589                                 maf[ i - 1 ] = tp;
 590                                 maxf++;
 591                                 continue;
 592                         }
 593                         break;
 594                 }
 595
 596                 maxb = 0;
 597                 chkb = vm_pageout_page_count -  maxf;
 598                 if (chkb) {
 599                         for(i = 1; i < chkb;i++) {
 600                                 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
 601                                         if ((tp->flags & PG_BUSY) ||
 602                                                 (tp->flags & PG_CLEANCHK) == 0 ||
 603                                                 (tp->busy != 0))
 604                                                 break;
 605                                         if((tp->queue - tp->pc) == PQ_CACHE) {
 606                                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 607                                                 break;
 608                                         }
 609                                         vm_page_test_dirty(tp);
 610                                         if ((tp->dirty & tp->valid) == 0) {
 611                                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 612                                                 break;
 613                                         }
 614                                         mab[ i - 1 ] = tp;
 615                                         maxb++;
 616                                         continue;
 617                                 }
 618                                 break;
 619                         }
 620                 }
 621
 622                 for(i=0;i<maxb;i++) {
 623                         int index = (maxb - i) - 1;
 624                         ma[index] = mab[i];
 625                         vm_page_flag_clear(ma[index], PG_CLEANCHK);
 626                 }
 627                 vm_page_flag_clear(p, PG_CLEANCHK);
 628                 ma[maxb] = p;
 629                 for(i=0;i<maxf;i++) {
 630                         int index = (maxb + i) + 1;
 631                         ma[index] = maf[i];
 632                         vm_page_flag_clear(ma[index], PG_CLEANCHK);
 633                 }
 634                 runlen = maxb + maxf + 1;
 635
 636                 splx(s);
 637                 vm_pageout_flush(ma, runlen, pagerflags);
 638                 for (i = 0; i<runlen; i++) {
 639                         if (ma[i]->valid & ma[i]->dirty) {
 640                                 vm_page_protect(ma[i], VM_PROT_READ);
 641                                 vm_page_flag_set(ma[i], PG_CLEANCHK);
 642                         }
 643                 }
 644                 if (object->generation != curgeneration)
 645                         goto rescan;
 646         }
 647
 648         VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
 649
 650         vm_object_clear_flag(object, OBJ_CLEANING);
 651         return;
 652 }
 653
 654 #ifdef not_used
 655 /* XXX I cannot tell if this should be an exported symbol */
 656 /*
 657  *      vm_object_deactivate_pages
 658  *
 659  *      Deactivate all pages in the specified object.  (Keep its pages
 660  *      in memory even though it is no longer referenced.)
 661  *
 662  *      The object must be locked.
 663  */
 664 static void
 665 vm_object_deactivate_pages(object)
 666         vm_object_t object;
 667 {
 668         vm_page_t p, next;
 669
 670         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
 671                 next = TAILQ_NEXT(p, listq);
 672                 vm_page_deactivate(p);
 673         }
 674 }
 675 #endif
 676
 677 /*
 678  * Same as vm_object_pmap_copy, except range checking really
 679  * works, and is meant for small sections of an object.
 680  *
 681  * This code protects resident pages by making them read-only
 682  * and is typically called on a fork or split when a page
 683  * is converted to copy-on-write.
 684  *
 685  * NOTE: If the page is already at VM_PROT_NONE, calling
 686  * vm_page_protect will have no effect.
 687  */
 688
 689 void
 690 vm_object_pmap_copy_1(object, start, end)
 691         vm_object_t object;
 692         vm_pindex_t start;
 693         vm_pindex_t end;
 694 {
 695         vm_pindex_t idx;
 696         vm_page_t p;
 697
 698         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
 699                 return;
 700
 701         for (idx = start; idx < end; idx++) {
 702                 p = vm_page_lookup(object, idx);
 703                 if (p == NULL)
 704                         continue;
 705                 vm_page_protect(p, VM_PROT_READ);
 706         }
 707 }
 708
 709 /*
 710  *      vm_object_pmap_remove:
 711  *
 712  *      Removes all physical pages in the specified
 713  *      object range from all physical maps.
 714  *
 715  *      The object must *not* be locked.
 716  */
 717 void
 718 vm_object_pmap_remove(object, start, end)
 719         vm_object_t object;
 720         vm_pindex_t start;
 721         vm_pindex_t end;
 722 {
 723         vm_page_t p;
 724
 725         if (object == NULL)
 726                 return;
 727         for (p = TAILQ_FIRST(&object->memq);
 728                 p != NULL;
 729                 p = TAILQ_NEXT(p, listq)) {
 730                 if (p->pindex >= start && p->pindex < end)
 731                         vm_page_protect(p, VM_PROT_NONE);
 732         }
 733         if ((start == 0) && (object->size == end))
 734                 vm_object_clear_flag(object, OBJ_WRITEABLE);
 735 }
 736
 737 /*
 738  *      vm_object_madvise:
 739  *
 740  *      Implements the madvise function at the object/page level.
 741  *
 742  *      Currently, madvise() functions are limited to the default and
 743  *      swap object types only, and also limited to only the unshared portions
 744  *      of a process's address space.  MADV_FREE, certainly, could never be
 745  *      run on anything else.  The others are more flexible and the code could
 746  *      be adjusted in the future to handle expanded cases for them.
 747  */
 748 void
 749 vm_object_madvise(object, pindex, count, advise)
 750         vm_object_t object;
 751         vm_pindex_t pindex;
 752         int count;
 753         int advise;
 754 {
 755         vm_pindex_t end, tpindex;
 756         vm_object_t tobject;
 757         vm_page_t m;
 758
 759         if (object == NULL)
 760                 return;
 761
 762         end = pindex + count;
 763
 764         /*
 765          * Locate and adjust resident pages
 766          */
 767
 768         for (; pindex < end; pindex += 1) {
 769 relookup:
 770                 tobject = object;
 771                 tpindex = pindex;
 772 shadowlookup:
 773
 774                 if (tobject->type != OBJT_DEFAULT &&
 775                     tobject->type != OBJT_SWAP
 776                 ) {
 777                         continue;
 778                 }
 779
 780                 if ((tobject->flags & OBJ_ONEMAPPING) == 0)
 781                         continue;
 782
 783                 m = vm_page_lookup(tobject, tpindex);
 784
 785                 if (m == NULL) {
 786                         /*
 787                          * There may be swap even if there is no backing page
 788                          */
 789                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
 790                                 swap_pager_freespace(tobject, tpindex, 1);
 791
 792                         /*
 793                          * next object
 794                          */
 795                         tobject = tobject->backing_object;
 796                         if (tobject == NULL)
 797                                 continue;
 798                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
 799                         goto shadowlookup;
 800                 }
 801
 802                 /*
 803                  * If the page is busy or not in a normal active state,
 804                  * we skip it.  Things can break if we mess with pages
 805                  * in any of the below states.
 806                  */
 807                 if (
 808                     m->hold_count ||
 809                     m->wire_count ||
 810                     m->valid != VM_PAGE_BITS_ALL
 811                 ) {
 812                         continue;
 813                 }
 814
 815                 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
 816                         goto relookup;
 817
 818                 if (advise == MADV_WILLNEED) {
 819                         vm_page_activate(m);
 820                 } else if (advise == MADV_DONTNEED) {
 821                         vm_page_deactivate(m);
 822                 } else if (advise == MADV_FREE) {
 823                         /*
 824                          * Mark the page clean.  This will allow the page
 825                          * to be freed up by the system.  However, such pages
 826                          * are often reused quickly by malloc()/free()
 827                          * so we do not do anything that would cause
 828                          * a page fault if we can help it.
 829                          *
 830                          * Specifically, we do not try to actually free
 831                          * the page now nor do we try to put it in the
 832                          * cache (which would cause a page fault on reuse).
 833                          *
 834                          * But we do make the page is freeable as we
 835                          * can without actually taking the step of unmapping
 836                          * it.
 837                          */
 838                         pmap_clear_modify(VM_PAGE_TO_PHYS(m));
 839                         m->dirty = 0;
 840                         m->act_count = 0;
 841                         vm_page_deactivate(m);
 842                         if (tobject->type == OBJT_SWAP)
 843                                 swap_pager_freespace(tobject, tpindex, 1);
 844                 }
 845         }
 846 }
 847
 848 /*
 849  *      vm_object_shadow:
 850  *
 851  *      Create a new object which is backed by the
 852  *      specified existing object range.  The source
 853  *      object reference is deallocated.
 854  *
 855  *      The new object and offset into that object
 856  *      are returned in the source parameters.
 857  */
 858
 859 void
 860 vm_object_shadow(object, offset, length)
 861         vm_object_t *object;    /* IN/OUT */
 862         vm_ooffset_t *offset;   /* IN/OUT */
 863         vm_size_t length;
 864 {
 865         vm_object_t source;
 866         vm_object_t result;
 867
 868         source = *object;
 869
 870         /*
 871          * Allocate a new object with the given length
 872          */
 873
 874         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
 875                 panic("vm_object_shadow: no object for shadowing");
 876
 877         /*
 878          * The new object shadows the source object, adding a reference to it.
 879          * Our caller changes his reference to point to the new object,
 880          * removing a reference to the source object.  Net result: no change
 881          * of reference count.
 882          *
 883          * Try to optimize the result object's page color when shadowing
 884          * in order to maintain page coloring consistancy in the combined
 885          * shadowed object.
 886          */
 887         result->backing_object = source;
 888         if (source) {
 889                 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list);
 890                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
 891                 source->shadow_count++;
 892                 source->generation++;
 893                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
 894         }
 895
 896         /*
 897          * Store the offset into the source object, and fix up the offset into
 898          * the new object.
 899          */
 900
 901         result->backing_object_offset = *offset;
 902
 903         /*
 904          * Return the new things
 905          */
 906
 907         *offset = 0;
 908         *object = result;
 909 }
 910
 911 #define OBSC_TEST_ALL_SHADOWED  0x0001
 912 #define OBSC_COLLAPSE_NOWAIT    0x0002
 913 #define OBSC_COLLAPSE_WAIT      0x0004
 914
 915 static __inline int
 916 vm_object_backing_scan(vm_object_t object, int op)
 917 {
 918         int s;
 919         int r = 1;
 920         vm_page_t p;
 921         vm_object_t backing_object;
 922         vm_pindex_t backing_offset_index;
 923
 924         s = splvm();
 925
 926         backing_object = object->backing_object;
 927         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
 928
 929         /*
 930          * Initial conditions
 931          */
 932
 933         if (op & OBSC_TEST_ALL_SHADOWED) {
 934                 /*
 935                  * We do not want to have to test for the existance of
 936                  * swap pages in the backing object.  XXX but with the
 937                  * new swapper this would be pretty easy to do.
 938                  *
 939                  * XXX what about anonymous MAP_SHARED memory that hasn't
 940                  * been ZFOD faulted yet?  If we do not test for this, the
 941                  * shadow test may succeed! XXX
 942                  */
 943                 if (backing_object->type != OBJT_DEFAULT) {
 944                         splx(s);
 945                         return(0);
 946                 }
 947         }
 948         if (op & OBSC_COLLAPSE_WAIT) {
 949                 vm_object_set_flag(backing_object, OBJ_DEAD);
 950         }
 951
 952         /*
 953          * Our scan
 954          */
 955
 956         p = TAILQ_FIRST(&backing_object->memq);
 957         while (p) {
 958                 vm_page_t next = TAILQ_NEXT(p, listq);
 959                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
 960
 961                 if (op & OBSC_TEST_ALL_SHADOWED) {
 962                         vm_page_t pp;
 963
 964                         /*
 965                          * Ignore pages outside the parent object's range
 966                          * and outside the parent object's mapping of the
 967                          * backing object.
 968                          *
 969                          * note that we do not busy the backing object's
 970                          * page.
 971                          */
 972
 973                         if (
 974                             p->pindex < backing_offset_index ||
 975                             new_pindex >= object->size
 976                         ) {
 977                                 p = next;
 978                                 continue;
 979                         }
 980
 981                         /*
 982                          * See if the parent has the page or if the parent's
 983                          * object pager has the page.  If the parent has the
 984                          * page but the page is not valid, the parent's
 985                          * object pager must have the page.
 986                          *
 987                          * If this fails, the parent does not completely shadow
 988                          * the object and we might as well give up now.
 989                          */
 990
 991                         pp = vm_page_lookup(object, new_pindex);
 992                         if (
 993                             (pp == NULL || pp->valid == 0) &&
 994                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
 995                         ) {
 996                                 r = 0;
 997                                 break;
 998                         }
 999                 }
1000
1001                 /*
1002                  * Check for busy page
1003                  */
1004
1005                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1006                         vm_page_t pp;
1007
1008                         if (op & OBSC_COLLAPSE_NOWAIT) {
1009                                 if (
1010                                     (p->flags & PG_BUSY) ||
1011                                     !p->valid ||
1012                                     p->hold_count ||
1013                                     p->wire_count ||
1014                                     p->busy
1015                                 ) {
1016                                         p = next;
1017                                         continue;
1018                                 }
1019                         } else if (op & OBSC_COLLAPSE_WAIT) {
1020                                 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1021                                         /*
1022                                          * If we slept, anything could have
1023                                          * happened.  Since the object is
1024                                          * marked dead, the backing offset
1025                                          * should not have changed so we
1026                                          * just restart our scan.
1027                                          */
1028                                         p = TAILQ_FIRST(&backing_object->memq);
1029                                         continue;
1030                                 }
1031                         }
1032
1033                         /*
1034                          * Busy the page
1035                          */
1036                         vm_page_busy(p);
1037
1038                         KASSERT(
1039                             p->object == backing_object,
1040                             ("vm_object_qcollapse(): object mismatch")
1041                         );
1042
1043                         /*
1044                          * Destroy any associated swap
1045                          */
1046                         if (backing_object->type == OBJT_SWAP) {
1047                                 swap_pager_freespace(
1048                                     backing_object,
1049                                     p->pindex,
1050                                     1
1051                                 );
1052                         }
1053
1054                         if (
1055                             p->pindex < backing_offset_index ||
1056                             new_pindex >= object->size
1057                         ) {
1058                                 /*
1059                                  * Page is out of the parent object's range, we
1060                                  * can simply destroy it.
1061                                  */
1062                                 vm_page_protect(p, VM_PROT_NONE);
1063                                 vm_page_free(p);
1064                                 p = next;
1065                                 continue;
1066                         }
1067
1068                         pp = vm_page_lookup(object, new_pindex);
1069                         if (
1070                             pp != NULL ||
1071                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1072                         ) {
1073                                 /*
1074                                  * page already exists in parent OR swap exists
1075                                  * for this location in the parent.  Destroy
1076                                  * the original page from the backing object.
1077                                  *
1078                                  * Leave the parent's page alone
1079                                  */
1080                                 vm_page_protect(p, VM_PROT_NONE);
1081                                 vm_page_free(p);
1082                                 p = next;
1083                                 continue;
1084                         }
1085
1086                         /*
1087                          * Page does not exist in parent, rename the
1088                          * page from the backing object to the main object.
1089                          *
1090                          * If the page was mapped to a process, it can remain
1091                          * mapped through the rename.
1092                          */
1093                         if ((p->queue - p->pc) == PQ_CACHE)
1094                                 vm_page_deactivate(p);
1095
1096                         vm_page_rename(p, object, new_pindex);
1097                         /* page automatically made dirty by rename */
1098                 }
1099                 p = next;
1100         }
1101         splx(s);
1102         return(r);
1103 }
1104
1105
1106 /*
1107  * this version of collapse allows the operation to occur earlier and
1108  * when paging_in_progress is true for an object...  This is not a complete
1109  * operation, but should plug 99.9% of the rest of the leaks.
1110  */
1111 static void
1112 vm_object_qcollapse(object)
1113         vm_object_t object;
1114 {
1115         vm_object_t backing_object = object->backing_object;
1116
1117         if (backing_object->ref_count != 1)
1118                 return;
1119
1120         backing_object->ref_count += 2;
1121
1122         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1123
1124         backing_object->ref_count -= 2;
1125 }
1126
1127 /*
1128  *      vm_object_collapse:
1129  *
1130  *      Collapse an object with the object backing it.
1131  *      Pages in the backing object are moved into the
1132  *      parent, and the backing object is deallocated.
1133  */
1134 void
1135 vm_object_collapse(object)
1136         vm_object_t object;
1137 {
1138         while (TRUE) {
1139                 vm_object_t backing_object;
1140
1141                 /*
1142                  * Verify that the conditions are right for collapse:
1143                  *
1144                  * The object exists and the backing object exists.
1145                  */
1146                 if (object == NULL)
1147                         break;
1148
1149                 if ((backing_object = object->backing_object) == NULL)
1150                         break;
1151
1152                 /*
1153                  * we check the backing object first, because it is most likely
1154                  * not collapsable.
1155                  */
1156                 if (backing_object->handle != NULL ||
1157                     (backing_object->type != OBJT_DEFAULT &&
1158                      backing_object->type != OBJT_SWAP) ||
1159                     (backing_object->flags & OBJ_DEAD) ||
1160                     object->handle != NULL ||
1161                     (object->type != OBJT_DEFAULT &&
1162                      object->type != OBJT_SWAP) ||
1163                     (object->flags & OBJ_DEAD)) {
1164                         break;
1165                 }
1166
1167                 if (
1168                     object->paging_in_progress != 0 ||
1169                     backing_object->paging_in_progress != 0
1170                 ) {
1171                         vm_object_qcollapse(object);
1172                         break;
1173                 }
1174
1175                 /*
1176                  * We know that we can either collapse the backing object (if
1177                  * the parent is the only reference to it) or (perhaps) have
1178                  * the parent bypass the object if the parent happens to shadow
1179                  * all the resident pages in the entire backing object.
1180                  *
1181                  * This is ignoring pager-backed pages such as swap pages.
1182                  * vm_object_backing_scan fails the shadowing test in this
1183                  * case.
1184                  */
1185
1186                 if (backing_object->ref_count == 1) {
1187                         /*
1188                          * If there is exactly one reference to the backing
1189                          * object, we can collapse it into the parent.
1190                          */
1191
1192                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1193
1194                         /*
1195                          * Move the pager from backing_object to object.
1196                          */
1197
1198                         if (backing_object->type == OBJT_SWAP) {
1199                                 vm_object_pip_add(backing_object, 1);
1200
1201                                 /*
1202                                  * scrap the paging_offset junk and do a
1203                                  * discrete copy.  This also removes major
1204                                  * assumptions about how the swap-pager
1205                                  * works from where it doesn't belong.  The
1206                                  * new swapper is able to optimize the
1207                                  * destroy-source case.
1208                                  */
1209
1210                                 vm_object_pip_add(object, 1);
1211                                 swap_pager_copy(
1212                                     backing_object,
1213                                     object,
1214                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1215                                 vm_object_pip_wakeup(object);
1216
1217                                 vm_object_pip_wakeup(backing_object);
1218                         }
1219                         /*
1220                          * Object now shadows whatever backing_object did.
1221                          * Note that the reference to
1222                          * backing_object->backing_object moves from within
1223                          * backing_object to within object.
1224                          */
1225
1226                         TAILQ_REMOVE(
1227                             &object->backing_object->shadow_head,
1228                             object,
1229                             shadow_list
1230                         );
1231                         object->backing_object->shadow_count--;
1232                         object->backing_object->generation++;
1233                         if (backing_object->backing_object) {
1234                                 TAILQ_REMOVE(
1235                                     &backing_object->backing_object->shadow_head,
1236                                     backing_object,
1237                                     shadow_list
1238                                 );
1239                                 backing_object->backing_object->shadow_count--;
1240                                 backing_object->backing_object->generation++;
1241                         }
1242                         object->backing_object = backing_object->backing_object;
1243                         if (object->backing_object) {
1244                                 TAILQ_INSERT_TAIL(
1245                                     &object->backing_object->shadow_head,
1246                                     object,
1247                                     shadow_list
1248                                 );
1249                                 object->backing_object->shadow_count++;
1250                                 object->backing_object->generation++;
1251                         }
1252
1253                         object->backing_object_offset +=
1254                             backing_object->backing_object_offset;
1255
1256                         /*
1257                          * Discard backing_object.
1258                          *
1259                          * Since the backing object has no pages, no pager left,
1260                          * and no object references within it, all that is
1261                          * necessary is to dispose of it.
1262                          */
1263
1264                         TAILQ_REMOVE(
1265                             &vm_object_list,
1266                             backing_object,
1267                             object_list
1268                         );
1269                         vm_object_count--;
1270
1271                         zfree(obj_zone, backing_object);
1272
1273                         object_collapses++;
1274                 } else {
1275                         vm_object_t new_backing_object;
1276
1277                         /*
1278                          * If we do not entirely shadow the backing object,
1279                          * there is nothing we can do so we give up.
1280                          */
1281
1282                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1283                                 break;
1284                         }
1285
1286                         /*
1287                          * Make the parent shadow the next object in the
1288                          * chain.  Deallocating backing_object will not remove
1289                          * it, since its reference count is at least 2.
1290                          */
1291
1292                         TAILQ_REMOVE(
1293                             &backing_object->shadow_head,
1294                             object,
1295                             shadow_list
1296                         );
1297                         backing_object->shadow_count--;
1298                         backing_object->generation++;
1299
1300                         new_backing_object = backing_object->backing_object;
1301                         if ((object->backing_object = new_backing_object) != NULL) {
1302                                 vm_object_reference(new_backing_object);
1303                                 TAILQ_INSERT_TAIL(
1304                                     &new_backing_object->shadow_head,
1305                                     object,
1306                                     shadow_list
1307                                 );
1308                                 new_backing_object->shadow_count++;
1309                                 new_backing_object->generation++;
1310                                 object->backing_object_offset +=
1311                                         backing_object->backing_object_offset;
1312                         }
1313
1314                         /*
1315                          * Drop the reference count on backing_object. Since
1316                          * its ref_count was at least 2, it will not vanish;
1317                          * so we don't need to call vm_object_deallocate, but
1318                          * we do anyway.
1319                          */
1320                         vm_object_deallocate(backing_object);
1321                         object_bypasses++;
1322                 }
1323
1324                 /*
1325                  * Try again with this object's new backing object.
1326                  */
1327         }
1328 }
1329
1330 /*
1331  *      vm_object_page_remove: [internal]
1332  *
1333  *      Removes all physical pages in the specified
1334  *      object range from the object's list of pages.
1335  *
1336  *      The object must be locked.
1337  */
1338 void
1339 vm_object_page_remove(object, start, end, clean_only)
1340         vm_object_t object;
1341         vm_pindex_t start;
1342         vm_pindex_t end;
1343         boolean_t clean_only;
1344 {
1345         vm_page_t p, next;
1346         unsigned int size;
1347         int all;
1348
1349         if (object == NULL ||
1350             object->resident_page_count == 0)
1351                 return;
1352
1353         all = ((end == 0) && (start == 0));
1354
1355         vm_object_pip_add(object, 1);
1356 again:
1357         size = end - start;
1358         if (all || size > object->resident_page_count / 4) {
1359                 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1360                         next = TAILQ_NEXT(p, listq);
1361                         if (all || ((start <= p->pindex) && (p->pindex < end))) {
1362                                 if (p->wire_count != 0) {
1363                                         vm_page_protect(p, VM_PROT_NONE);
1364                                         if (!clean_only)
1365                                                 p->valid = 0;
1366                                         continue;
1367                                 }
1368
1369                                 /*
1370                                  * The busy flags are only cleared at
1371                                  * interrupt -- minimize the spl transitions
1372                                  */
1373
1374                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1375                                         goto again;
1376
1377                                 if (clean_only && p->valid) {
1378                                         vm_page_test_dirty(p);
1379                                         if (p->valid & p->dirty)
1380                                                 continue;
1381                                 }
1382
1383                                 vm_page_busy(p);
1384                                 vm_page_protect(p, VM_PROT_NONE);
1385                                 vm_page_free(p);
1386                         }
1387                 }
1388         } else {
1389                 while (size > 0) {
1390                         if ((p = vm_page_lookup(object, start)) != 0) {
1391
1392                                 if (p->wire_count != 0) {
1393                                         vm_page_protect(p, VM_PROT_NONE);
1394                                         if (!clean_only)
1395                                                 p->valid = 0;
1396                                         start += 1;
1397                                         size -= 1;
1398                                         continue;
1399                                 }
1400
1401                                 /*
1402                                  * The busy flags are only cleared at
1403                                  * interrupt -- minimize the spl transitions
1404                                  */
1405                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1406                                         goto again;
1407
1408                                 if (clean_only && p->valid) {
1409                                         vm_page_test_dirty(p);
1410                                         if (p->valid & p->dirty) {
1411                                                 start += 1;
1412                                                 size -= 1;
1413                                                 continue;
1414                                         }
1415                                 }
1416
1417                                 vm_page_busy(p);
1418                                 vm_page_protect(p, VM_PROT_NONE);
1419                                 vm_page_free(p);
1420                         }
1421                         start += 1;
1422                         size -= 1;
1423                 }
1424         }
1425         vm_object_pip_wakeup(object);
1426 }
1427
1428 /*
1429  *      Routine:        vm_object_coalesce
1430  *      Function:       Coalesces two objects backing up adjoining
1431  *                      regions of memory into a single object.
1432  *
1433  *      returns TRUE if objects were combined.
1434  *
1435  *      NOTE:   Only works at the moment if the second object is NULL -
1436  *              if it's not, which object do we lock first?
1437  *
1438  *      Parameters:
1439  *              prev_object     First object to coalesce
1440  *              prev_offset     Offset into prev_object
1441  *              next_object     Second object into coalesce
1442  *              next_offset     Offset into next_object
1443  *
1444  *              prev_size       Size of reference to prev_object
1445  *              next_size       Size of reference to next_object
1446  *
1447  *      Conditions:
1448  *      The object must *not* be locked.
1449  */
1450 boolean_t
1451 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1452         vm_object_t prev_object;
1453         vm_pindex_t prev_pindex;
1454         vm_size_t prev_size, next_size;
1455 {
1456         vm_pindex_t next_pindex;
1457
1458         if (prev_object == NULL) {
1459                 return (TRUE);
1460         }
1461
1462         if (prev_object->type != OBJT_DEFAULT &&
1463             prev_object->type != OBJT_SWAP) {
1464                 return (FALSE);
1465         }
1466
1467         /*
1468          * Try to collapse the object first
1469          */
1470         vm_object_collapse(prev_object);
1471
1472         /*
1473          * Can't coalesce if: . more than one reference . paged out . shadows
1474          * another object . has a copy elsewhere (any of which mean that the
1475          * pages not mapped to prev_entry may be in use anyway)
1476          */
1477
1478         if (prev_object->backing_object != NULL) {
1479                 return (FALSE);
1480         }
1481
1482         prev_size >>= PAGE_SHIFT;
1483         next_size >>= PAGE_SHIFT;
1484         next_pindex = prev_pindex + prev_size;
1485
1486         if ((prev_object->ref_count > 1) &&
1487             (prev_object->size != next_pindex)) {
1488                 return (FALSE);
1489         }
1490
1491         /*
1492          * Remove any pages that may still be in the object from a previous
1493          * deallocation.
1494          */
1495         if (next_pindex < prev_object->size) {
1496                 vm_object_page_remove(prev_object,
1497                                       next_pindex,
1498                                       next_pindex + next_size, FALSE);
1499                 if (prev_object->type == OBJT_SWAP)
1500                         swap_pager_freespace(prev_object,
1501                                              next_pindex, next_size);
1502         }
1503
1504         /*
1505          * Extend the object if necessary.
1506          */
1507         if (next_pindex + next_size > prev_object->size)
1508                 prev_object->size = next_pindex + next_size;
1509
1510         return (TRUE);
1511 }
1512
1513 #include "opt_ddb.h"
1514 #ifdef DDB
1515 #include <sys/kernel.h>
1516
1517 #include <machine/cons.h>
1518
1519 #include <ddb/ddb.h>
1520
1521 static int      _vm_object_in_map __P((vm_map_t map, vm_object_t object,
1522                                        vm_map_entry_t entry));
1523 static int      vm_object_in_map __P((vm_object_t object));
1524
1525 static int
1526 _vm_object_in_map(map, object, entry)
1527         vm_map_t map;
1528         vm_object_t object;
1529         vm_map_entry_t entry;
1530 {
1531         vm_map_t tmpm;
1532         vm_map_entry_t tmpe;
1533         vm_object_t obj;
1534         int entcount;
1535
1536         if (map == 0)
1537                 return 0;
1538
1539         if (entry == 0) {
1540                 tmpe = map->header.next;
1541                 entcount = map->nentries;
1542                 while (entcount-- && (tmpe != &map->header)) {
1543                         if( _vm_object_in_map(map, object, tmpe)) {
1544                                 return 1;
1545                         }
1546                         tmpe = tmpe->next;
1547                 }
1548         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1549                 tmpm = entry->object.sub_map;
1550                 tmpe = tmpm->header.next;
1551                 entcount = tmpm->nentries;
1552                 while (entcount-- && tmpe != &tmpm->header) {
1553                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1554                                 return 1;
1555                         }
1556                         tmpe = tmpe->next;
1557                 }
1558         } else if ((obj = entry->object.vm_object) != NULL) {
1559                 for(; obj; obj=obj->backing_object)
1560                         if( obj == object) {
1561                                 return 1;
1562                         }
1563         }
1564         return 0;
1565 }
1566
1567 static int
1568 vm_object_in_map( object)
1569         vm_object_t object;
1570 {
1571         struct proc *p;
1572         for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1573                 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1574                         continue;
1575                 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1576                         return 1;
1577         }
1578         if( _vm_object_in_map( kernel_map, object, 0))
1579                 return 1;
1580         if( _vm_object_in_map( kmem_map, object, 0))
1581                 return 1;
1582         if( _vm_object_in_map( pager_map, object, 0))
1583                 return 1;
1584         if( _vm_object_in_map( buffer_map, object, 0))
1585                 return 1;
1586         if( _vm_object_in_map( io_map, object, 0))
1587                 return 1;
1588         if( _vm_object_in_map( phys_map, object, 0))
1589                 return 1;
1590         if( _vm_object_in_map( mb_map, object, 0))
1591                 return 1;
1592         if( _vm_object_in_map( u_map, object, 0))
1593                 return 1;
1594         return 0;
1595 }
1596
1597 DB_SHOW_COMMAND(vmochk, vm_object_check)
1598 {
1599         vm_object_t object;
1600
1601         /*
1602          * make sure that internal objs are in a map somewhere
1603          * and none have zero ref counts.
1604          */
1605         for (object = TAILQ_FIRST(&vm_object_list);
1606                         object != NULL;
1607                         object = TAILQ_NEXT(object, object_list)) {
1608                 if (object->handle == NULL &&
1609                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1610                         if (object->ref_count == 0) {
1611                                 db_printf("vmochk: internal obj has zero ref count: %d\n",
1612                                         object->size);
1613                         }
1614                         if (!vm_object_in_map(object)) {
1615                                 db_printf(
1616                         "vmochk: internal obj is not in a map: "
1617                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1618                                     object->ref_count, (u_long)object->size,
1619                                     (u_long)object->size,
1620                                     (void *)object->backing_object);
1621                         }
1622                 }
1623         }
1624 }
1625
1626 /*
1627  *      vm_object_print:        [ debug ]
1628  */
1629 DB_SHOW_COMMAND(object, vm_object_print_static)
1630 {
1631         /* XXX convert args. */
1632         vm_object_t object = (vm_object_t)addr;
1633         boolean_t full = have_addr;
1634
1635         vm_page_t p;
1636
1637         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1638 #define count   was_count
1639
1640         int count;
1641
1642         if (object == NULL)
1643                 return;
1644
1645         db_iprintf(
1646             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1647             object, (int)object->type, (u_long)object->size,
1648             object->resident_page_count, object->ref_count, object->flags);
1649         /*
1650          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1651          */
1652         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1653             object->shadow_count,
1654             object->backing_object ? object->backing_object->ref_count : 0,
1655             object->backing_object, (long)object->backing_object_offset);
1656
1657         if (!full)
1658                 return;
1659
1660         db_indent += 2;
1661         count = 0;
1662         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1663                 if (count == 0)
1664                         db_iprintf("memory:=");
1665                 else if (count == 6) {
1666                         db_printf("\n");
1667                         db_iprintf(" ...");
1668                         count = 0;
1669                 } else
1670                         db_printf(",");
1671                 count++;
1672
1673                 db_printf("(off=0x%lx,page=0x%lx)",
1674                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1675         }
1676         if (count != 0)
1677                 db_printf("\n");
1678         db_indent -= 2;
1679 }
1680
1681 /* XXX. */
1682 #undef count
1683
1684 /* XXX need this non-static entry for calling from vm_map_print. */
1685 void
1686 vm_object_print(addr, have_addr, count, modif)
1687         /* db_expr_t */ long addr;
1688         boolean_t have_addr;
1689         /* db_expr_t */ long count;
1690         char *modif;
1691 {
1692         vm_object_print_static(addr, have_addr, count, modif);
1693 }
1694
1695 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1696 {
1697         vm_object_t object;
1698         int nl = 0;
1699         int c;
1700         for (object = TAILQ_FIRST(&vm_object_list);
1701                         object != NULL;
1702                         object = TAILQ_NEXT(object, object_list)) {
1703                 vm_pindex_t idx, fidx;
1704                 vm_pindex_t osize;
1705                 vm_offset_t pa = -1, padiff;
1706                 int rcount;
1707                 vm_page_t m;
1708
1709                 db_printf("new object: %p\n", (void *)object);
1710                 if ( nl > 18) {
1711                         c = cngetc();
1712                         if (c != ' ')
1713                                 return;
1714                         nl = 0;
1715                 }
1716                 nl++;
1717                 rcount = 0;
1718                 fidx = 0;
1719                 osize = object->size;
1720                 if (osize > 128)
1721                         osize = 128;
1722                 for(idx=0;idx<osize;idx++) {
1723                         m = vm_page_lookup(object, idx);
1724                         if (m == NULL) {
1725                                 if (rcount) {
1726                                         db_printf(" index(%d)run(%d)pa(0x%x)\n",
1727                                                 fidx, rcount, pa);
1728                                         if ( nl > 18) {
1729                                                 c = cngetc();
1730                                                 if (c != ' ')
1731                                                         return;
1732                                                 nl = 0;
1733                                         }
1734                                         nl++;
1735                                         rcount = 0;
1736                                 }
1737                                 continue;
1738                         }
1739
1740
1741                         if (rcount &&
1742                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1743                                 ++rcount;
1744                                 continue;
1745                         }
1746                         if (rcount) {
1747                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1748                                 padiff >>= PAGE_SHIFT;
1749                                 padiff &= PQ_L2_MASK;
1750                                 if (padiff == 0) {
1751                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1752                                         ++rcount;
1753                                         continue;
1754                                 }
1755                                 db_printf(" index(%d)run(%d)pa(0x%x)", fidx, rcount, pa);
1756                                 db_printf("pd(%d)\n", padiff);
1757                                 if ( nl > 18) {
1758                                         c = cngetc();
1759                                         if (c != ' ')
1760                                                 return;
1761                                         nl = 0;
1762                                 }
1763                                 nl++;
1764                         }
1765                         fidx = idx;
1766                         pa = VM_PAGE_TO_PHYS(m);
1767                         rcount = 1;
1768                 }
1769                 if (rcount) {
1770                         db_printf(" index(%d)run(%d)pa(0x%x)\n", fidx, rcount, pa);
1771                         if ( nl > 18) {
1772                                 c = cngetc();
1773                                 if (c != ' ')
1774                                         return;
1775                                 nl = 0;
1776                         }
1777                         nl++;
1778                 }
1779         }
1780 }
1781 #endif /* DDB */