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  * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
  65  * $DragonFly: src/sys/vm/vm_object.c,v 1.13 2004/03/01 06:33:24 dillon Exp $
  66  */
  67
  68 /*
  69  *      Virtual memory object module.
  70  */
  71
  72 #include <sys/param.h>
  73 #include <sys/systm.h>
  74 #include <sys/proc.h>           /* for curproc, pageproc */
  75 #include <sys/vnode.h>
  76 #include <sys/vmmeter.h>
  77 #include <sys/mman.h>
  78 #include <sys/mount.h>
  79 #include <sys/kernel.h>
  80 #include <sys/sysctl.h>
  81
  82 #include <vm/vm.h>
  83 #include <vm/vm_param.h>
  84 #include <vm/pmap.h>
  85 #include <vm/vm_map.h>
  86 #include <vm/vm_object.h>
  87 #include <vm/vm_page.h>
  88 #include <vm/vm_pageout.h>
  89 #include <vm/vm_pager.h>
  90 #include <vm/swap_pager.h>
  91 #include <vm/vm_kern.h>
  92 #include <vm/vm_extern.h>
  93 #include <vm/vm_zone.h>
  94
  95 #define EASY_SCAN_FACTOR        8
  96
  97 #define MSYNC_FLUSH_HARDSEQ     0x01
  98 #define MSYNC_FLUSH_SOFTSEQ     0x02
  99
 100 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
 101 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
 102         CTLFLAG_RW, &msync_flush_flags, 0, "");
 103
 104 static void     vm_object_qcollapse (vm_object_t object);
 105 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
 106
 107 /*
 108  *      Virtual memory objects maintain the actual data
 109  *      associated with allocated virtual memory.  A given
 110  *      page of memory exists within exactly one object.
 111  *
 112  *      An object is only deallocated when all "references"
 113  *      are given up.  Only one "reference" to a given
 114  *      region of an object should be writeable.
 115  *
 116  *      Associated with each object is a list of all resident
 117  *      memory pages belonging to that object; this list is
 118  *      maintained by the "vm_page" module, and locked by the object's
 119  *      lock.
 120  *
 121  *      Each object also records a "pager" routine which is
 122  *      used to retrieve (and store) pages to the proper backing
 123  *      storage.  In addition, objects may be backed by other
 124  *      objects from which they were virtual-copied.
 125  *
 126  *      The only items within the object structure which are
 127  *      modified after time of creation are:
 128  *              reference count         locked by object's lock
 129  *              pager routine           locked by object's lock
 130  *
 131  */
 132
 133 struct object_q vm_object_list;
 134 static struct lwkt_token vm_object_list_token;
 135 static long vm_object_count;            /* count of all objects */
 136 vm_object_t kernel_object;
 137 vm_object_t kmem_object;
 138 static struct vm_object kernel_object_store;
 139 static struct vm_object kmem_object_store;
 140 extern int vm_pageout_page_count;
 141
 142 static long object_collapses;
 143 static long object_bypasses;
 144 static int next_index;
 145 static vm_zone_t obj_zone;
 146 static struct vm_zone obj_zone_store;
 147 static int object_hash_rand;
 148 #define VM_OBJECTS_INIT 256
 149 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
 150
 151 void
 152 _vm_object_allocate(type, size, object)
 153         objtype_t type;
 154         vm_size_t size;
 155         vm_object_t object;
 156 {
 157         int incr;
 158         TAILQ_INIT(&object->memq);
 159         LIST_INIT(&object->shadow_head);
 160
 161         object->type = type;
 162         object->size = size;
 163         object->ref_count = 1;
 164         object->flags = 0;
 165         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
 166                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 167         object->paging_in_progress = 0;
 168         object->resident_page_count = 0;
 169         object->shadow_count = 0;
 170         object->pg_color = next_index;
 171         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
 172                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
 173         else
 174                 incr = size;
 175         next_index = (next_index + incr) & PQ_L2_MASK;
 176         object->handle = NULL;
 177         object->backing_object = NULL;
 178         object->backing_object_offset = (vm_ooffset_t) 0;
 179         /*
 180          * Try to generate a number that will spread objects out in the
 181          * hash table.  We 'wipe' new objects across the hash in 128 page
 182          * increments plus 1 more to offset it a little more by the time
 183          * it wraps around.
 184          */
 185         object->hash_rand = object_hash_rand - 129;
 186
 187         object->generation++;
 188
 189         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 190         vm_object_count++;
 191         object_hash_rand = object->hash_rand;
 192 }
 193
 194 /*
 195  *      vm_object_init:
 196  *
 197  *      Initialize the VM objects module.
 198  */
 199 void
 200 vm_object_init()
 201 {
 202         TAILQ_INIT(&vm_object_list);
 203         lwkt_token_init(&vm_object_list_token);
 204         vm_object_count = 0;
 205
 206         kernel_object = &kernel_object_store;
 207         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 208             kernel_object);
 209
 210         kmem_object = &kmem_object_store;
 211         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 212             kmem_object);
 213
 214         obj_zone = &obj_zone_store;
 215         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
 216                 vm_objects_init, VM_OBJECTS_INIT);
 217 }
 218
 219 void
 220 vm_object_init2()
 221 {
 222         zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
 223 }
 224
 225 /*
 226  *      vm_object_allocate:
 227  *
 228  *      Returns a new object with the given size.
 229  */
 230
 231 vm_object_t
 232 vm_object_allocate(type, size)
 233         objtype_t type;
 234         vm_size_t size;
 235 {
 236         vm_object_t result;
 237
 238         result = (vm_object_t) zalloc(obj_zone);
 239
 240         _vm_object_allocate(type, size, result);
 241
 242         return (result);
 243 }
 244
 245
 246 /*
 247  *      vm_object_reference:
 248  *
 249  *      Gets another reference to the given object.
 250  */
 251 void
 252 vm_object_reference(object)
 253         vm_object_t object;
 254 {
 255         if (object == NULL)
 256                 return;
 257
 258 #if 0
 259         /* object can be re-referenced during final cleaning */
 260         KASSERT(!(object->flags & OBJ_DEAD),
 261             ("vm_object_reference: attempting to reference dead obj"));
 262 #endif
 263
 264         object->ref_count++;
 265         if (object->type == OBJT_VNODE) {
 266                 while (vget((struct vnode *) object->handle, NULL,
 267                     LK_RETRY|LK_NOOBJ, curthread)) {
 268                         printf("vm_object_reference: delay in getting object\n");
 269                 }
 270         }
 271 }
 272
 273 void
 274 vm_object_vndeallocate(object)
 275         vm_object_t object;
 276 {
 277         struct vnode *vp = (struct vnode *) object->handle;
 278
 279         KASSERT(object->type == OBJT_VNODE,
 280             ("vm_object_vndeallocate: not a vnode object"));
 281         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 282 #ifdef INVARIANTS
 283         if (object->ref_count == 0) {
 284                 vprint("vm_object_vndeallocate", vp);
 285                 panic("vm_object_vndeallocate: bad object reference count");
 286         }
 287 #endif
 288
 289         object->ref_count--;
 290         if (object->ref_count == 0) {
 291                 vp->v_flag &= ~VTEXT;
 292                 vm_object_clear_flag(object, OBJ_OPT);
 293         }
 294         vrele(vp);
 295 }
 296
 297 /*
 298  *      vm_object_deallocate:
 299  *
 300  *      Release a reference to the specified object,
 301  *      gained either through a vm_object_allocate
 302  *      or a vm_object_reference call.  When all references
 303  *      are gone, storage associated with this object
 304  *      may be relinquished.
 305  *
 306  *      No object may be locked.
 307  */
 308 void
 309 vm_object_deallocate(object)
 310         vm_object_t object;
 311 {
 312         vm_object_t temp;
 313
 314         while (object != NULL) {
 315
 316                 if (object->type == OBJT_VNODE) {
 317                         vm_object_vndeallocate(object);
 318                         return;
 319                 }
 320
 321                 if (object->ref_count == 0) {
 322                         panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
 323                 } else if (object->ref_count > 2) {
 324                         object->ref_count--;
 325                         return;
 326                 }
 327
 328                 /*
 329                  * Here on ref_count of one or two, which are special cases for
 330                  * objects.
 331                  */
 332                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
 333                         vm_object_set_flag(object, OBJ_ONEMAPPING);
 334                         object->ref_count--;
 335                         return;
 336                 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
 337                         object->ref_count--;
 338                         if ((object->handle == NULL) &&
 339                             (object->type == OBJT_DEFAULT ||
 340                              object->type == OBJT_SWAP)) {
 341                                 vm_object_t robject;
 342
 343                                 robject = LIST_FIRST(&object->shadow_head);
 344                                 KASSERT(robject != NULL,
 345                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 346                                          object->ref_count,
 347                                          object->shadow_count));
 348                                 if ((robject->handle == NULL) &&
 349                                     (robject->type == OBJT_DEFAULT ||
 350                                      robject->type == OBJT_SWAP)) {
 351
 352                                         robject->ref_count++;
 353
 354                                         while (
 355                                                 robject->paging_in_progress ||
 356                                                 object->paging_in_progress
 357                                         ) {
 358                                                 vm_object_pip_sleep(robject, "objde1");
 359                                                 vm_object_pip_sleep(object, "objde2");
 360                                         }
 361
 362                                         if (robject->ref_count == 1) {
 363                                                 robject->ref_count--;
 364                                                 object = robject;
 365                                                 goto doterm;
 366                                         }
 367
 368                                         object = robject;
 369                                         vm_object_collapse(object);
 370                                         continue;
 371                                 }
 372                         }
 373
 374                         return;
 375
 376                 } else {
 377                         object->ref_count--;
 378                         if (object->ref_count != 0)
 379                                 return;
 380                 }
 381
 382 doterm:
 383
 384                 temp = object->backing_object;
 385                 if (temp) {
 386                         LIST_REMOVE(object, shadow_list);
 387                         temp->shadow_count--;
 388                         if (temp->ref_count == 0)
 389                                 vm_object_clear_flag(temp, OBJ_OPT);
 390                         temp->generation++;
 391                         object->backing_object = NULL;
 392                 }
 393
 394                 /*
 395                  * Don't double-terminate, we could be in a termination
 396                  * recursion due to the terminate having to sync data
 397                  * to disk.
 398                  */
 399                 if ((object->flags & OBJ_DEAD) == 0)
 400                         vm_object_terminate(object);
 401                 object = temp;
 402         }
 403 }
 404
 405 /*
 406  *      vm_object_terminate actually destroys the specified object, freeing
 407  *      up all previously used resources.
 408  *
 409  *      The object must be locked.
 410  *      This routine may block.
 411  */
 412 void
 413 vm_object_terminate(object)
 414         vm_object_t object;
 415 {
 416         lwkt_tokref ilock;
 417         vm_page_t p;
 418         int s;
 419
 420         /*
 421          * Make sure no one uses us.
 422          */
 423         vm_object_set_flag(object, OBJ_DEAD);
 424
 425         /*
 426          * wait for the pageout daemon to be done with the object
 427          */
 428         vm_object_pip_wait(object, "objtrm");
 429
 430         KASSERT(!object->paging_in_progress,
 431                 ("vm_object_terminate: pageout in progress"));
 432
 433         /*
 434          * Clean and free the pages, as appropriate. All references to the
 435          * object are gone, so we don't need to lock it.
 436          */
 437         if (object->type == OBJT_VNODE) {
 438                 struct vnode *vp;
 439
 440                 /*
 441                  * Freeze optimized copies.
 442                  */
 443                 vm_freeze_copyopts(object, 0, object->size);
 444
 445                 /*
 446                  * Clean pages and flush buffers.
 447                  */
 448                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 449
 450                 vp = (struct vnode *) object->handle;
 451                 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
 452         }
 453
 454         /*
 455          * Wait for any I/O to complete, after which there had better not
 456          * be any references left on the object.
 457          */
 458         vm_object_pip_wait(object, "objtrm");
 459
 460         if (object->ref_count != 0)
 461                 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
 462
 463         /*
 464          * Now free any remaining pages. For internal objects, this also
 465          * removes them from paging queues. Don't free wired pages, just
 466          * remove them from the object.
 467          */
 468         s = splvm();
 469         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
 470                 if (p->busy || (p->flags & PG_BUSY))
 471                         panic("vm_object_terminate: freeing busy page %p\n", p);
 472                 if (p->wire_count == 0) {
 473                         vm_page_busy(p);
 474                         vm_page_free(p);
 475                         mycpu->gd_cnt.v_pfree++;
 476                 } else {
 477                         vm_page_busy(p);
 478                         vm_page_remove(p);
 479                 }
 480         }
 481         splx(s);
 482
 483         /*
 484          * Let the pager know object is dead.
 485          */
 486         vm_pager_deallocate(object);
 487
 488         /*
 489          * Remove the object from the global object list.
 490          */
 491         lwkt_gettoken(&ilock, &vm_object_list_token);
 492         TAILQ_REMOVE(&vm_object_list, object, object_list);
 493         lwkt_reltoken(&ilock);
 494
 495         wakeup(object);
 496
 497         /*
 498          * Free the space for the object.
 499          */
 500         zfree(obj_zone, object);
 501 }
 502
 503 /*
 504  *      vm_object_page_clean
 505  *
 506  *      Clean all dirty pages in the specified range of object.  Leaves page
 507  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 508  *      write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
 509  *      leaving the object dirty.
 510  *
 511  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 512  *      synchronous clustering mode implementation.
 513  *
 514  *      Odd semantics: if start == end, we clean everything.
 515  *
 516  *      The object must be locked.
 517  */
 518
 519 void
 520 vm_object_page_clean(object, start, end, flags)
 521         vm_object_t object;
 522         vm_pindex_t start;
 523         vm_pindex_t end;
 524         int flags;
 525 {
 526         vm_page_t p, np;
 527         vm_offset_t tstart, tend;
 528         vm_pindex_t pi;
 529         struct vnode *vp;
 530         int clearobjflags;
 531         int pagerflags;
 532         int curgeneration;
 533         lwkt_tokref vlock;
 534
 535         if (object->type != OBJT_VNODE ||
 536                 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
 537                 return;
 538
 539         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 540         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
 541
 542         vp = object->handle;
 543
 544         vm_object_set_flag(object, OBJ_CLEANING);
 545
 546         /*
 547          * Handle 'entire object' case
 548          */
 549         tstart = start;
 550         if (end == 0) {
 551                 tend = object->size;
 552         } else {
 553                 tend = end;
 554         }
 555
 556         /*
 557          * If the caller is smart and only msync()s a range he knows is
 558          * dirty, we may be able to avoid an object scan.  This results in
 559          * a phenominal improvement in performance.  We cannot do this
 560          * as a matter of course because the object may be huge - e.g.
 561          * the size might be in the gigabytes or terrabytes.
 562          */
 563         if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
 564                 vm_offset_t tscan;
 565                 int scanlimit;
 566                 int scanreset;
 567
 568                 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
 569                 if (scanreset < 16)
 570                         scanreset = 16;
 571                 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
 572
 573                 scanlimit = scanreset;
 574                 tscan = tstart;
 575                 while (tscan < tend) {
 576                         curgeneration = object->generation;
 577                         p = vm_page_lookup(object, tscan);
 578                         if (p == NULL || p->valid == 0 ||
 579                             (p->queue - p->pc) == PQ_CACHE) {
 580                                 if (--scanlimit == 0)
 581                                         break;
 582                                 ++tscan;
 583                                 continue;
 584                         }
 585                         vm_page_test_dirty(p);
 586                         if ((p->dirty & p->valid) == 0) {
 587                                 if (--scanlimit == 0)
 588                                         break;
 589                                 ++tscan;
 590                                 continue;
 591                         }
 592                         /*
 593                          * If we have been asked to skip nosync pages and
 594                          * this is a nosync page, we can't continue.
 595                          */
 596                         if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
 597                                 if (--scanlimit == 0)
 598                                         break;
 599                                 ++tscan;
 600                                 continue;
 601                         }
 602                         scanlimit = scanreset;
 603
 604                         /*
 605                          * This returns 0 if it was unable to busy the first
 606                          * page (i.e. had to sleep).
 607                          */
 608                         tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
 609                 }
 610
 611                 /*
 612                  * If everything was dirty and we flushed it successfully,
 613                  * and the requested range is not the entire object, we
 614                  * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
 615                  * return immediately.
 616                  */
 617                 if (tscan >= tend && (tstart || tend < object->size)) {
 618                         vm_object_clear_flag(object, OBJ_CLEANING);
 619                         return;
 620                 }
 621                 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
 622         }
 623
 624         /*
 625          * Generally set CLEANCHK interlock and make the page read-only so
 626          * we can then clear the object flags.
 627          *
 628          * However, if this is a nosync mmap then the object is likely to
 629          * stay dirty so do not mess with the page and do not clear the
 630          * object flags.
 631          */
 632
 633         clearobjflags = 1;
 634
 635         for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
 636                 vm_page_flag_set(p, PG_CLEANCHK);
 637                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
 638                         clearobjflags = 0;
 639                 else
 640                         vm_page_protect(p, VM_PROT_READ);
 641         }
 642
 643         if (clearobjflags && (tstart == 0) && (tend == object->size)) {
 644                 struct vnode *vp;
 645
 646                 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
 647                 if (object->type == OBJT_VNODE &&
 648                     (vp = (struct vnode *)object->handle) != NULL) {
 649                         if (vp->v_flag & VOBJDIRTY) {
 650                                 lwkt_gettoken(&vlock, vp->v_interlock);
 651                                 vp->v_flag &= ~VOBJDIRTY;
 652                                 lwkt_reltoken(&vlock);
 653                         }
 654                 }
 655         }
 656
 657 rescan:
 658         curgeneration = object->generation;
 659
 660         for(p = TAILQ_FIRST(&object->memq); p; p = np) {
 661                 int n;
 662
 663                 np = TAILQ_NEXT(p, listq);
 664
 665 again:
 666                 pi = p->pindex;
 667                 if (((p->flags & PG_CLEANCHK) == 0) ||
 668                         (pi < tstart) || (pi >= tend) ||
 669                         (p->valid == 0) ||
 670                         ((p->queue - p->pc) == PQ_CACHE)) {
 671                         vm_page_flag_clear(p, PG_CLEANCHK);
 672                         continue;
 673                 }
 674
 675                 vm_page_test_dirty(p);
 676                 if ((p->dirty & p->valid) == 0) {
 677                         vm_page_flag_clear(p, PG_CLEANCHK);
 678                         continue;
 679                 }
 680
 681                 /*
 682                  * If we have been asked to skip nosync pages and this is a
 683                  * nosync page, skip it.  Note that the object flags were
 684                  * not cleared in this case so we do not have to set them.
 685                  */
 686                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
 687                         vm_page_flag_clear(p, PG_CLEANCHK);
 688                         continue;
 689                 }
 690
 691                 n = vm_object_page_collect_flush(object, p,
 692                         curgeneration, pagerflags);
 693                 if (n == 0)
 694                         goto rescan;
 695                 if (object->generation != curgeneration)
 696                         goto rescan;
 697
 698                 /*
 699                  * Try to optimize the next page.  If we can't we pick up
 700                  * our (random) scan where we left off.
 701                  */
 702                 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
 703                         if ((p = vm_page_lookup(object, pi + n)) != NULL)
 704                                 goto again;
 705                 }
 706         }
 707
 708 #if 0
 709         VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
 710 #endif
 711
 712         vm_object_clear_flag(object, OBJ_CLEANING);
 713         return;
 714 }
 715
 716 static int
 717 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
 718 {
 719         int runlen;
 720         int s;
 721         int maxf;
 722         int chkb;
 723         int maxb;
 724         int i;
 725         vm_pindex_t pi;
 726         vm_page_t maf[vm_pageout_page_count];
 727         vm_page_t mab[vm_pageout_page_count];
 728         vm_page_t ma[vm_pageout_page_count];
 729
 730         s = splvm();
 731         pi = p->pindex;
 732         while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
 733                 if (object->generation != curgeneration) {
 734                         splx(s);
 735                         return(0);
 736                 }
 737         }
 738
 739         maxf = 0;
 740         for(i = 1; i < vm_pageout_page_count; i++) {
 741                 vm_page_t tp;
 742
 743                 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
 744                         if ((tp->flags & PG_BUSY) ||
 745                                 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
 746                                  (tp->flags & PG_CLEANCHK) == 0) ||
 747                                 (tp->busy != 0))
 748                                 break;
 749                         if((tp->queue - tp->pc) == PQ_CACHE) {
 750                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 751                                 break;
 752                         }
 753                         vm_page_test_dirty(tp);
 754                         if ((tp->dirty & tp->valid) == 0) {
 755                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 756                                 break;
 757                         }
 758                         maf[ i - 1 ] = tp;
 759                         maxf++;
 760                         continue;
 761                 }
 762                 break;
 763         }
 764
 765         maxb = 0;
 766         chkb = vm_pageout_page_count -  maxf;
 767         if (chkb) {
 768                 for(i = 1; i < chkb;i++) {
 769                         vm_page_t tp;
 770
 771                         if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
 772                                 if ((tp->flags & PG_BUSY) ||
 773                                         ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
 774                                          (tp->flags & PG_CLEANCHK) == 0) ||
 775                                         (tp->busy != 0))
 776                                         break;
 777                                 if((tp->queue - tp->pc) == PQ_CACHE) {
 778                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 779                                         break;
 780                                 }
 781                                 vm_page_test_dirty(tp);
 782                                 if ((tp->dirty & tp->valid) == 0) {
 783                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 784                                         break;
 785                                 }
 786                                 mab[ i - 1 ] = tp;
 787                                 maxb++;
 788                                 continue;
 789                         }
 790                         break;
 791                 }
 792         }
 793
 794         for(i = 0; i < maxb; i++) {
 795                 int index = (maxb - i) - 1;
 796                 ma[index] = mab[i];
 797                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
 798         }
 799         vm_page_flag_clear(p, PG_CLEANCHK);
 800         ma[maxb] = p;
 801         for(i = 0; i < maxf; i++) {
 802                 int index = (maxb + i) + 1;
 803                 ma[index] = maf[i];
 804                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
 805         }
 806         runlen = maxb + maxf + 1;
 807
 808         splx(s);
 809         vm_pageout_flush(ma, runlen, pagerflags);
 810         for (i = 0; i < runlen; i++) {
 811                 if (ma[i]->valid & ma[i]->dirty) {
 812                         vm_page_protect(ma[i], VM_PROT_READ);
 813                         vm_page_flag_set(ma[i], PG_CLEANCHK);
 814
 815                         /*
 816                          * maxf will end up being the actual number of pages
 817                          * we wrote out contiguously, non-inclusive of the
 818                          * first page.  We do not count look-behind pages.
 819                          */
 820                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
 821                                 maxf = i - maxb - 1;
 822                 }
 823         }
 824         return(maxf + 1);
 825 }
 826
 827 #ifdef not_used
 828 /* XXX I cannot tell if this should be an exported symbol */
 829 /*
 830  *      vm_object_deactivate_pages
 831  *
 832  *      Deactivate all pages in the specified object.  (Keep its pages
 833  *      in memory even though it is no longer referenced.)
 834  *
 835  *      The object must be locked.
 836  */
 837 static void
 838 vm_object_deactivate_pages(object)
 839         vm_object_t object;
 840 {
 841         vm_page_t p, next;
 842
 843         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
 844                 next = TAILQ_NEXT(p, listq);
 845                 vm_page_deactivate(p);
 846         }
 847 }
 848 #endif
 849
 850 /*
 851  * Same as vm_object_pmap_copy, except range checking really
 852  * works, and is meant for small sections of an object.
 853  *
 854  * This code protects resident pages by making them read-only
 855  * and is typically called on a fork or split when a page
 856  * is converted to copy-on-write.
 857  *
 858  * NOTE: If the page is already at VM_PROT_NONE, calling
 859  * vm_page_protect will have no effect.
 860  */
 861
 862 void
 863 vm_object_pmap_copy_1(object, start, end)
 864         vm_object_t object;
 865         vm_pindex_t start;
 866         vm_pindex_t end;
 867 {
 868         vm_pindex_t idx;
 869         vm_page_t p;
 870
 871         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
 872                 return;
 873
 874         for (idx = start; idx < end; idx++) {
 875                 p = vm_page_lookup(object, idx);
 876                 if (p == NULL)
 877                         continue;
 878                 vm_page_protect(p, VM_PROT_READ);
 879         }
 880 }
 881
 882 /*
 883  *      vm_object_pmap_remove:
 884  *
 885  *      Removes all physical pages in the specified
 886  *      object range from all physical maps.
 887  *
 888  *      The object must *not* be locked.
 889  */
 890 void
 891 vm_object_pmap_remove(object, start, end)
 892         vm_object_t object;
 893         vm_pindex_t start;
 894         vm_pindex_t end;
 895 {
 896         vm_page_t p;
 897
 898         if (object == NULL)
 899                 return;
 900         for (p = TAILQ_FIRST(&object->memq);
 901                 p != NULL;
 902                 p = TAILQ_NEXT(p, listq)) {
 903                 if (p->pindex >= start && p->pindex < end)
 904                         vm_page_protect(p, VM_PROT_NONE);
 905         }
 906         if ((start == 0) && (object->size == end))
 907                 vm_object_clear_flag(object, OBJ_WRITEABLE);
 908 }
 909
 910 /*
 911  *      vm_object_madvise:
 912  *
 913  *      Implements the madvise function at the object/page level.
 914  *
 915  *      MADV_WILLNEED   (any object)
 916  *
 917  *          Activate the specified pages if they are resident.
 918  *
 919  *      MADV_DONTNEED   (any object)
 920  *
 921  *          Deactivate the specified pages if they are resident.
 922  *
 923  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
 924  *                       OBJ_ONEMAPPING only)
 925  *
 926  *          Deactivate and clean the specified pages if they are
 927  *          resident.  This permits the process to reuse the pages
 928  *          without faulting or the kernel to reclaim the pages
 929  *          without I/O.
 930  */
 931 void
 932 vm_object_madvise(object, pindex, count, advise)
 933         vm_object_t object;
 934         vm_pindex_t pindex;
 935         int count;
 936         int advise;
 937 {
 938         vm_pindex_t end, tpindex;
 939         vm_object_t tobject;
 940         vm_page_t m;
 941
 942         if (object == NULL)
 943                 return;
 944
 945         end = pindex + count;
 946
 947         /*
 948          * Locate and adjust resident pages
 949          */
 950
 951         for (; pindex < end; pindex += 1) {
 952 relookup:
 953                 tobject = object;
 954                 tpindex = pindex;
 955 shadowlookup:
 956                 /*
 957                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
 958                  * and those pages must be OBJ_ONEMAPPING.
 959                  */
 960                 if (advise == MADV_FREE) {
 961                         if ((tobject->type != OBJT_DEFAULT &&
 962                              tobject->type != OBJT_SWAP) ||
 963                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
 964                                 continue;
 965                         }
 966                 }
 967
 968                 m = vm_page_lookup(tobject, tpindex);
 969
 970                 if (m == NULL) {
 971                         /*
 972                          * There may be swap even if there is no backing page
 973                          */
 974                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
 975                                 swap_pager_freespace(tobject, tpindex, 1);
 976
 977                         /*
 978                          * next object
 979                          */
 980                         tobject = tobject->backing_object;
 981                         if (tobject == NULL)
 982                                 continue;
 983                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
 984                         goto shadowlookup;
 985                 }
 986
 987                 /*
 988                  * If the page is busy or not in a normal active state,
 989                  * we skip it.  If the page is not managed there are no
 990                  * page queues to mess with.  Things can break if we mess
 991                  * with pages in any of the below states.
 992                  */
 993                 if (
 994                     m->hold_count ||
 995                     m->wire_count ||
 996                     (m->flags & PG_UNMANAGED) ||
 997                     m->valid != VM_PAGE_BITS_ALL
 998                 ) {
 999                         continue;
1000                 }
1001
1002                 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1003                         goto relookup;
1004
1005                 if (advise == MADV_WILLNEED) {
1006                         vm_page_activate(m);
1007                 } else if (advise == MADV_DONTNEED) {
1008                         vm_page_dontneed(m);
1009                 } else if (advise == MADV_FREE) {
1010                         /*
1011                          * Mark the page clean.  This will allow the page
1012                          * to be freed up by the system.  However, such pages
1013                          * are often reused quickly by malloc()/free()
1014                          * so we do not do anything that would cause
1015                          * a page fault if we can help it.
1016                          *
1017                          * Specifically, we do not try to actually free
1018                          * the page now nor do we try to put it in the
1019                          * cache (which would cause a page fault on reuse).
1020                          *
1021                          * But we do make the page is freeable as we
1022                          * can without actually taking the step of unmapping
1023                          * it.
1024                          */
1025                         pmap_clear_modify(m);
1026                         m->dirty = 0;
1027                         m->act_count = 0;
1028                         vm_page_dontneed(m);
1029                         if (tobject->type == OBJT_SWAP)
1030                                 swap_pager_freespace(tobject, tpindex, 1);
1031                 }
1032         }
1033 }
1034
1035 /*
1036  *      vm_object_shadow:
1037  *
1038  *      Create a new object which is backed by the
1039  *      specified existing object range.  The source
1040  *      object reference is deallocated.
1041  *
1042  *      The new object and offset into that object
1043  *      are returned in the source parameters.
1044  */
1045
1046 void
1047 vm_object_shadow(object, offset, length)
1048         vm_object_t *object;    /* IN/OUT */
1049         vm_ooffset_t *offset;   /* IN/OUT */
1050         vm_size_t length;
1051 {
1052         vm_object_t source;
1053         vm_object_t result;
1054
1055         source = *object;
1056
1057         /*
1058          * Don't create the new object if the old object isn't shared.
1059          */
1060
1061         if (source != NULL &&
1062             source->ref_count == 1 &&
1063             source->handle == NULL &&
1064             (source->type == OBJT_DEFAULT ||
1065              source->type == OBJT_SWAP))
1066                 return;
1067
1068         /*
1069          * Allocate a new object with the given length
1070          */
1071
1072         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1073                 panic("vm_object_shadow: no object for shadowing");
1074
1075         /*
1076          * The new object shadows the source object, adding a reference to it.
1077          * Our caller changes his reference to point to the new object,
1078          * removing a reference to the source object.  Net result: no change
1079          * of reference count.
1080          *
1081          * Try to optimize the result object's page color when shadowing
1082          * in order to maintain page coloring consistency in the combined
1083          * shadowed object.
1084          */
1085         result->backing_object = source;
1086         if (source) {
1087                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1088                 source->shadow_count++;
1089                 source->generation++;
1090                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1091         }
1092
1093         /*
1094          * Store the offset into the source object, and fix up the offset into
1095          * the new object.
1096          */
1097
1098         result->backing_object_offset = *offset;
1099
1100         /*
1101          * Return the new things
1102          */
1103
1104         *offset = 0;
1105         *object = result;
1106 }
1107
1108 #define OBSC_TEST_ALL_SHADOWED  0x0001
1109 #define OBSC_COLLAPSE_NOWAIT    0x0002
1110 #define OBSC_COLLAPSE_WAIT      0x0004
1111
1112 static __inline int
1113 vm_object_backing_scan(vm_object_t object, int op)
1114 {
1115         int s;
1116         int r = 1;
1117         vm_page_t p;
1118         vm_object_t backing_object;
1119         vm_pindex_t backing_offset_index;
1120
1121         s = splvm();
1122
1123         backing_object = object->backing_object;
1124         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1125
1126         /*
1127          * Initial conditions
1128          */
1129
1130         if (op & OBSC_TEST_ALL_SHADOWED) {
1131                 /*
1132                  * We do not want to have to test for the existence of
1133                  * swap pages in the backing object.  XXX but with the
1134                  * new swapper this would be pretty easy to do.
1135                  *
1136                  * XXX what about anonymous MAP_SHARED memory that hasn't
1137                  * been ZFOD faulted yet?  If we do not test for this, the
1138                  * shadow test may succeed! XXX
1139                  */
1140                 if (backing_object->type != OBJT_DEFAULT) {
1141                         splx(s);
1142                         return(0);
1143                 }
1144         }
1145         if (op & OBSC_COLLAPSE_WAIT) {
1146                 vm_object_set_flag(backing_object, OBJ_DEAD);
1147         }
1148
1149         /*
1150          * Our scan
1151          */
1152
1153         p = TAILQ_FIRST(&backing_object->memq);
1154         while (p) {
1155                 vm_page_t next = TAILQ_NEXT(p, listq);
1156                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1157
1158                 if (op & OBSC_TEST_ALL_SHADOWED) {
1159                         vm_page_t pp;
1160
1161                         /*
1162                          * Ignore pages outside the parent object's range
1163                          * and outside the parent object's mapping of the
1164                          * backing object.
1165                          *
1166                          * note that we do not busy the backing object's
1167                          * page.
1168                          */
1169
1170                         if (
1171                             p->pindex < backing_offset_index ||
1172                             new_pindex >= object->size
1173                         ) {
1174                                 p = next;
1175                                 continue;
1176                         }
1177
1178                         /*
1179                          * See if the parent has the page or if the parent's
1180                          * object pager has the page.  If the parent has the
1181                          * page but the page is not valid, the parent's
1182                          * object pager must have the page.
1183                          *
1184                          * If this fails, the parent does not completely shadow
1185                          * the object and we might as well give up now.
1186                          */
1187
1188                         pp = vm_page_lookup(object, new_pindex);
1189                         if (
1190                             (pp == NULL || pp->valid == 0) &&
1191                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1192                         ) {
1193                                 r = 0;
1194                                 break;
1195                         }
1196                 }
1197
1198                 /*
1199                  * Check for busy page
1200                  */
1201
1202                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1203                         vm_page_t pp;
1204
1205                         if (op & OBSC_COLLAPSE_NOWAIT) {
1206                                 if (
1207                                     (p->flags & PG_BUSY) ||
1208                                     !p->valid ||
1209                                     p->hold_count ||
1210                                     p->wire_count ||
1211                                     p->busy
1212                                 ) {
1213                                         p = next;
1214                                         continue;
1215                                 }
1216                         } else if (op & OBSC_COLLAPSE_WAIT) {
1217                                 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1218                                         /*
1219                                          * If we slept, anything could have
1220                                          * happened.  Since the object is
1221                                          * marked dead, the backing offset
1222                                          * should not have changed so we
1223                                          * just restart our scan.
1224                                          */
1225                                         p = TAILQ_FIRST(&backing_object->memq);
1226                                         continue;
1227                                 }
1228                         }
1229
1230                         /*
1231                          * Busy the page
1232                          */
1233                         vm_page_busy(p);
1234
1235                         KASSERT(
1236                             p->object == backing_object,
1237                             ("vm_object_qcollapse(): object mismatch")
1238                         );
1239
1240                         /*
1241                          * Destroy any associated swap
1242                          */
1243                         if (backing_object->type == OBJT_SWAP) {
1244                                 swap_pager_freespace(
1245                                     backing_object,
1246                                     p->pindex,
1247                                     1
1248                                 );
1249                         }
1250
1251                         if (
1252                             p->pindex < backing_offset_index ||
1253                             new_pindex >= object->size
1254                         ) {
1255                                 /*
1256                                  * Page is out of the parent object's range, we
1257                                  * can simply destroy it.
1258                                  */
1259                                 vm_page_protect(p, VM_PROT_NONE);
1260                                 vm_page_free(p);
1261                                 p = next;
1262                                 continue;
1263                         }
1264
1265                         pp = vm_page_lookup(object, new_pindex);
1266                         if (
1267                             pp != NULL ||
1268                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1269                         ) {
1270                                 /*
1271                                  * page already exists in parent OR swap exists
1272                                  * for this location in the parent.  Destroy
1273                                  * the original page from the backing object.
1274                                  *
1275                                  * Leave the parent's page alone
1276                                  */
1277                                 vm_page_protect(p, VM_PROT_NONE);
1278                                 vm_page_free(p);
1279                                 p = next;
1280                                 continue;
1281                         }
1282
1283                         /*
1284                          * Page does not exist in parent, rename the
1285                          * page from the backing object to the main object.
1286                          *
1287                          * If the page was mapped to a process, it can remain
1288                          * mapped through the rename.
1289                          */
1290                         if ((p->queue - p->pc) == PQ_CACHE)
1291                                 vm_page_deactivate(p);
1292
1293                         vm_page_rename(p, object, new_pindex);
1294                         /* page automatically made dirty by rename */
1295                 }
1296                 p = next;
1297         }
1298         splx(s);
1299         return(r);
1300 }
1301
1302
1303 /*
1304  * this version of collapse allows the operation to occur earlier and
1305  * when paging_in_progress is true for an object...  This is not a complete
1306  * operation, but should plug 99.9% of the rest of the leaks.
1307  */
1308 static void
1309 vm_object_qcollapse(object)
1310         vm_object_t object;
1311 {
1312         vm_object_t backing_object = object->backing_object;
1313
1314         if (backing_object->ref_count != 1)
1315                 return;
1316
1317         backing_object->ref_count += 2;
1318
1319         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1320
1321         backing_object->ref_count -= 2;
1322 }
1323
1324 /*
1325  *      vm_object_collapse:
1326  *
1327  *      Collapse an object with the object backing it.
1328  *      Pages in the backing object are moved into the
1329  *      parent, and the backing object is deallocated.
1330  */
1331 void
1332 vm_object_collapse(object)
1333         vm_object_t object;
1334 {
1335         while (TRUE) {
1336                 vm_object_t backing_object;
1337
1338                 /*
1339                  * Verify that the conditions are right for collapse:
1340                  *
1341                  * The object exists and the backing object exists.
1342                  */
1343                 if (object == NULL)
1344                         break;
1345
1346                 if ((backing_object = object->backing_object) == NULL)
1347                         break;
1348
1349                 /*
1350                  * we check the backing object first, because it is most likely
1351                  * not collapsable.
1352                  */
1353                 if (backing_object->handle != NULL ||
1354                     (backing_object->type != OBJT_DEFAULT &&
1355                      backing_object->type != OBJT_SWAP) ||
1356                     (backing_object->flags & OBJ_DEAD) ||
1357                     object->handle != NULL ||
1358                     (object->type != OBJT_DEFAULT &&
1359                      object->type != OBJT_SWAP) ||
1360                     (object->flags & OBJ_DEAD)) {
1361                         break;
1362                 }
1363
1364                 if (
1365                     object->paging_in_progress != 0 ||
1366                     backing_object->paging_in_progress != 0
1367                 ) {
1368                         vm_object_qcollapse(object);
1369                         break;
1370                 }
1371
1372                 /*
1373                  * We know that we can either collapse the backing object (if
1374                  * the parent is the only reference to it) or (perhaps) have
1375                  * the parent bypass the object if the parent happens to shadow
1376                  * all the resident pages in the entire backing object.
1377                  *
1378                  * This is ignoring pager-backed pages such as swap pages.
1379                  * vm_object_backing_scan fails the shadowing test in this
1380                  * case.
1381                  */
1382
1383                 if (backing_object->ref_count == 1) {
1384                         /*
1385                          * If there is exactly one reference to the backing
1386                          * object, we can collapse it into the parent.
1387                          */
1388
1389                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1390
1391                         /*
1392                          * Move the pager from backing_object to object.
1393                          */
1394
1395                         if (backing_object->type == OBJT_SWAP) {
1396                                 vm_object_pip_add(backing_object, 1);
1397
1398                                 /*
1399                                  * scrap the paging_offset junk and do a
1400                                  * discrete copy.  This also removes major
1401                                  * assumptions about how the swap-pager
1402                                  * works from where it doesn't belong.  The
1403                                  * new swapper is able to optimize the
1404                                  * destroy-source case.
1405                                  */
1406
1407                                 vm_object_pip_add(object, 1);
1408                                 swap_pager_copy(
1409                                     backing_object,
1410                                     object,
1411                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1412                                 vm_object_pip_wakeup(object);
1413
1414                                 vm_object_pip_wakeup(backing_object);
1415                         }
1416                         /*
1417                          * Object now shadows whatever backing_object did.
1418                          * Note that the reference to
1419                          * backing_object->backing_object moves from within
1420                          * backing_object to within object.
1421                          */
1422
1423                         LIST_REMOVE(object, shadow_list);
1424                         object->backing_object->shadow_count--;
1425                         object->backing_object->generation++;
1426                         if (backing_object->backing_object) {
1427                                 LIST_REMOVE(backing_object, shadow_list);
1428                                 backing_object->backing_object->shadow_count--;
1429                                 backing_object->backing_object->generation++;
1430                         }
1431                         object->backing_object = backing_object->backing_object;
1432                         if (object->backing_object) {
1433                                 LIST_INSERT_HEAD(
1434                                     &object->backing_object->shadow_head,
1435                                     object,
1436                                     shadow_list
1437                                 );
1438                                 object->backing_object->shadow_count++;
1439                                 object->backing_object->generation++;
1440                         }
1441
1442                         object->backing_object_offset +=
1443                             backing_object->backing_object_offset;
1444
1445                         /*
1446                          * Discard backing_object.
1447                          *
1448                          * Since the backing object has no pages, no pager left,
1449                          * and no object references within it, all that is
1450                          * necessary is to dispose of it.
1451                          */
1452
1453                         KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1454                         KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1455                         TAILQ_REMOVE(
1456                             &vm_object_list,
1457                             backing_object,
1458                             object_list
1459                         );
1460                         vm_object_count--;
1461
1462                         zfree(obj_zone, backing_object);
1463
1464                         object_collapses++;
1465                 } else {
1466                         vm_object_t new_backing_object;
1467
1468                         /*
1469                          * If we do not entirely shadow the backing object,
1470                          * there is nothing we can do so we give up.
1471                          */
1472
1473                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1474                                 break;
1475                         }
1476
1477                         /*
1478                          * Make the parent shadow the next object in the
1479                          * chain.  Deallocating backing_object will not remove
1480                          * it, since its reference count is at least 2.
1481                          */
1482
1483                         LIST_REMOVE(object, shadow_list);
1484                         backing_object->shadow_count--;
1485                         backing_object->generation++;
1486
1487                         new_backing_object = backing_object->backing_object;
1488                         if ((object->backing_object = new_backing_object) != NULL) {
1489                                 vm_object_reference(new_backing_object);
1490                                 LIST_INSERT_HEAD(
1491                                     &new_backing_object->shadow_head,
1492                                     object,
1493                                     shadow_list
1494                                 );
1495                                 new_backing_object->shadow_count++;
1496                                 new_backing_object->generation++;
1497                                 object->backing_object_offset +=
1498                                         backing_object->backing_object_offset;
1499                         }
1500
1501                         /*
1502                          * Drop the reference count on backing_object. Since
1503                          * its ref_count was at least 2, it will not vanish;
1504                          * so we don't need to call vm_object_deallocate, but
1505                          * we do anyway.
1506                          */
1507                         vm_object_deallocate(backing_object);
1508                         object_bypasses++;
1509                 }
1510
1511                 /*
1512                  * Try again with this object's new backing object.
1513                  */
1514         }
1515 }
1516
1517 /*
1518  *      vm_object_page_remove: [internal]
1519  *
1520  *      Removes all physical pages in the specified
1521  *      object range from the object's list of pages.
1522  *
1523  *      The object must be locked.
1524  */
1525 void
1526 vm_object_page_remove(object, start, end, clean_only)
1527         vm_object_t object;
1528         vm_pindex_t start;
1529         vm_pindex_t end;
1530         boolean_t clean_only;
1531 {
1532         vm_page_t p, next;
1533         unsigned int size;
1534         int all;
1535
1536         if (object == NULL ||
1537             object->resident_page_count == 0)
1538                 return;
1539
1540         all = ((end == 0) && (start == 0));
1541
1542         /*
1543          * Since physically-backed objects do not use managed pages, we can't
1544          * remove pages from the object (we must instead remove the page
1545          * references, and then destroy the object).
1546          */
1547         KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1548
1549         vm_object_pip_add(object, 1);
1550 again:
1551         size = end - start;
1552         if (all || size > object->resident_page_count / 4) {
1553                 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1554                         next = TAILQ_NEXT(p, listq);
1555                         if (all || ((start <= p->pindex) && (p->pindex < end))) {
1556                                 if (p->wire_count != 0) {
1557                                         vm_page_protect(p, VM_PROT_NONE);
1558                                         if (!clean_only)
1559                                                 p->valid = 0;
1560                                         continue;
1561                                 }
1562
1563                                 /*
1564                                  * The busy flags are only cleared at
1565                                  * interrupt -- minimize the spl transitions
1566                                  */
1567
1568                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1569                                         goto again;
1570
1571                                 if (clean_only && p->valid) {
1572                                         vm_page_test_dirty(p);
1573                                         if (p->valid & p->dirty)
1574                                                 continue;
1575                                 }
1576
1577                                 vm_page_busy(p);
1578                                 vm_page_protect(p, VM_PROT_NONE);
1579                                 vm_page_free(p);
1580                         }
1581                 }
1582         } else {
1583                 while (size > 0) {
1584                         if ((p = vm_page_lookup(object, start)) != 0) {
1585
1586                                 if (p->wire_count != 0) {
1587                                         vm_page_protect(p, VM_PROT_NONE);
1588                                         if (!clean_only)
1589                                                 p->valid = 0;
1590                                         start += 1;
1591                                         size -= 1;
1592                                         continue;
1593                                 }
1594
1595                                 /*
1596                                  * The busy flags are only cleared at
1597                                  * interrupt -- minimize the spl transitions
1598                                  */
1599                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1600                                         goto again;
1601
1602                                 if (clean_only && p->valid) {
1603                                         vm_page_test_dirty(p);
1604                                         if (p->valid & p->dirty) {
1605                                                 start += 1;
1606                                                 size -= 1;
1607                                                 continue;
1608                                         }
1609                                 }
1610
1611                                 vm_page_busy(p);
1612                                 vm_page_protect(p, VM_PROT_NONE);
1613                                 vm_page_free(p);
1614                         }
1615                         start += 1;
1616                         size -= 1;
1617                 }
1618         }
1619         vm_object_pip_wakeup(object);
1620 }
1621
1622 /*
1623  *      Routine:        vm_object_coalesce
1624  *      Function:       Coalesces two objects backing up adjoining
1625  *                      regions of memory into a single object.
1626  *
1627  *      returns TRUE if objects were combined.
1628  *
1629  *      NOTE:   Only works at the moment if the second object is NULL -
1630  *              if it's not, which object do we lock first?
1631  *
1632  *      Parameters:
1633  *              prev_object     First object to coalesce
1634  *              prev_offset     Offset into prev_object
1635  *              next_object     Second object into coalesce
1636  *              next_offset     Offset into next_object
1637  *
1638  *              prev_size       Size of reference to prev_object
1639  *              next_size       Size of reference to next_object
1640  *
1641  *      Conditions:
1642  *      The object must *not* be locked.
1643  */
1644 boolean_t
1645 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1646         vm_object_t prev_object;
1647         vm_pindex_t prev_pindex;
1648         vm_size_t prev_size, next_size;
1649 {
1650         vm_pindex_t next_pindex;
1651
1652         if (prev_object == NULL) {
1653                 return (TRUE);
1654         }
1655
1656         if (prev_object->type != OBJT_DEFAULT &&
1657             prev_object->type != OBJT_SWAP) {
1658                 return (FALSE);
1659         }
1660
1661         /*
1662          * Try to collapse the object first
1663          */
1664         vm_object_collapse(prev_object);
1665
1666         /*
1667          * Can't coalesce if: . more than one reference . paged out . shadows
1668          * another object . has a copy elsewhere (any of which mean that the
1669          * pages not mapped to prev_entry may be in use anyway)
1670          */
1671
1672         if (prev_object->backing_object != NULL) {
1673                 return (FALSE);
1674         }
1675
1676         prev_size >>= PAGE_SHIFT;
1677         next_size >>= PAGE_SHIFT;
1678         next_pindex = prev_pindex + prev_size;
1679
1680         if ((prev_object->ref_count > 1) &&
1681             (prev_object->size != next_pindex)) {
1682                 return (FALSE);
1683         }
1684
1685         /*
1686          * Remove any pages that may still be in the object from a previous
1687          * deallocation.
1688          */
1689         if (next_pindex < prev_object->size) {
1690                 vm_object_page_remove(prev_object,
1691                                       next_pindex,
1692                                       next_pindex + next_size, FALSE);
1693                 if (prev_object->type == OBJT_SWAP)
1694                         swap_pager_freespace(prev_object,
1695                                              next_pindex, next_size);
1696         }
1697
1698         /*
1699          * Extend the object if necessary.
1700          */
1701         if (next_pindex + next_size > prev_object->size)
1702                 prev_object->size = next_pindex + next_size;
1703
1704         return (TRUE);
1705 }
1706
1707 void
1708 vm_object_set_writeable_dirty(vm_object_t object)
1709 {
1710         struct vnode *vp;
1711         lwkt_tokref vlock;
1712
1713         vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1714         if (object->type == OBJT_VNODE &&
1715             (vp = (struct vnode *)object->handle) != NULL) {
1716                 if ((vp->v_flag & VOBJDIRTY) == 0) {
1717                         lwkt_gettoken(&vlock, vp->v_interlock);
1718                         vp->v_flag |= VOBJDIRTY;
1719                         lwkt_reltoken(&vlock);
1720                 }
1721         }
1722 }
1723
1724
1725
1726 #include "opt_ddb.h"
1727 #ifdef DDB
1728 #include <sys/kernel.h>
1729
1730 #include <sys/cons.h>
1731
1732 #include <ddb/ddb.h>
1733
1734 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
1735                                        vm_map_entry_t entry);
1736 static int      vm_object_in_map (vm_object_t object);
1737
1738 static int
1739 _vm_object_in_map(map, object, entry)
1740         vm_map_t map;
1741         vm_object_t object;
1742         vm_map_entry_t entry;
1743 {
1744         vm_map_t tmpm;
1745         vm_map_entry_t tmpe;
1746         vm_object_t obj;
1747         int entcount;
1748
1749         if (map == 0)
1750                 return 0;
1751
1752         if (entry == 0) {
1753                 tmpe = map->header.next;
1754                 entcount = map->nentries;
1755                 while (entcount-- && (tmpe != &map->header)) {
1756                         if( _vm_object_in_map(map, object, tmpe)) {
1757                                 return 1;
1758                         }
1759                         tmpe = tmpe->next;
1760                 }
1761         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1762                 tmpm = entry->object.sub_map;
1763                 tmpe = tmpm->header.next;
1764                 entcount = tmpm->nentries;
1765                 while (entcount-- && tmpe != &tmpm->header) {
1766                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1767                                 return 1;
1768                         }
1769                         tmpe = tmpe->next;
1770                 }
1771         } else if ((obj = entry->object.vm_object) != NULL) {
1772                 for(; obj; obj=obj->backing_object)
1773                         if( obj == object) {
1774                                 return 1;
1775                         }
1776         }
1777         return 0;
1778 }
1779
1780 static int
1781 vm_object_in_map( object)
1782         vm_object_t object;
1783 {
1784         struct proc *p;
1785         for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1786                 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1787                         continue;
1788                 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1789                         return 1;
1790         }
1791         if( _vm_object_in_map( kernel_map, object, 0))
1792                 return 1;
1793         if( _vm_object_in_map( pager_map, object, 0))
1794                 return 1;
1795         if( _vm_object_in_map( buffer_map, object, 0))
1796                 return 1;
1797         if( _vm_object_in_map( mb_map, object, 0))
1798                 return 1;
1799         return 0;
1800 }
1801
1802 DB_SHOW_COMMAND(vmochk, vm_object_check)
1803 {
1804         vm_object_t object;
1805
1806         /*
1807          * make sure that internal objs are in a map somewhere
1808          * and none have zero ref counts.
1809          */
1810         for (object = TAILQ_FIRST(&vm_object_list);
1811                         object != NULL;
1812                         object = TAILQ_NEXT(object, object_list)) {
1813                 if (object->handle == NULL &&
1814                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1815                         if (object->ref_count == 0) {
1816                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1817                                         (long)object->size);
1818                         }
1819                         if (!vm_object_in_map(object)) {
1820                                 db_printf(
1821                         "vmochk: internal obj is not in a map: "
1822                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1823                                     object->ref_count, (u_long)object->size,
1824                                     (u_long)object->size,
1825                                     (void *)object->backing_object);
1826                         }
1827                 }
1828         }
1829 }
1830
1831 /*
1832  *      vm_object_print:        [ debug ]
1833  */
1834 DB_SHOW_COMMAND(object, vm_object_print_static)
1835 {
1836         /* XXX convert args. */
1837         vm_object_t object = (vm_object_t)addr;
1838         boolean_t full = have_addr;
1839
1840         vm_page_t p;
1841
1842         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1843 #define count   was_count
1844
1845         int count;
1846
1847         if (object == NULL)
1848                 return;
1849
1850         db_iprintf(
1851             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1852             object, (int)object->type, (u_long)object->size,
1853             object->resident_page_count, object->ref_count, object->flags);
1854         /*
1855          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1856          */
1857         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1858             object->shadow_count,
1859             object->backing_object ? object->backing_object->ref_count : 0,
1860             object->backing_object, (long)object->backing_object_offset);
1861
1862         if (!full)
1863                 return;
1864
1865         db_indent += 2;
1866         count = 0;
1867         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1868                 if (count == 0)
1869                         db_iprintf("memory:=");
1870                 else if (count == 6) {
1871                         db_printf("\n");
1872                         db_iprintf(" ...");
1873                         count = 0;
1874                 } else
1875                         db_printf(",");
1876                 count++;
1877
1878                 db_printf("(off=0x%lx,page=0x%lx)",
1879                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1880         }
1881         if (count != 0)
1882                 db_printf("\n");
1883         db_indent -= 2;
1884 }
1885
1886 /* XXX. */
1887 #undef count
1888
1889 /* XXX need this non-static entry for calling from vm_map_print. */
1890 void
1891 vm_object_print(addr, have_addr, count, modif)
1892         /* db_expr_t */ long addr;
1893         boolean_t have_addr;
1894         /* db_expr_t */ long count;
1895         char *modif;
1896 {
1897         vm_object_print_static(addr, have_addr, count, modif);
1898 }
1899
1900 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1901 {
1902         vm_object_t object;
1903         int nl = 0;
1904         int c;
1905         for (object = TAILQ_FIRST(&vm_object_list);
1906                         object != NULL;
1907                         object = TAILQ_NEXT(object, object_list)) {
1908                 vm_pindex_t idx, fidx;
1909                 vm_pindex_t osize;
1910                 vm_paddr_t pa = -1, padiff;
1911                 int rcount;
1912                 vm_page_t m;
1913
1914                 db_printf("new object: %p\n", (void *)object);
1915                 if ( nl > 18) {
1916                         c = cngetc();
1917                         if (c != ' ')
1918                                 return;
1919                         nl = 0;
1920                 }
1921                 nl++;
1922                 rcount = 0;
1923                 fidx = 0;
1924                 osize = object->size;
1925                 if (osize > 128)
1926                         osize = 128;
1927                 for(idx=0;idx<osize;idx++) {
1928                         m = vm_page_lookup(object, idx);
1929                         if (m == NULL) {
1930                                 if (rcount) {
1931                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1932                                                 (long)fidx, rcount, (long)pa);
1933                                         if ( nl > 18) {
1934                                                 c = cngetc();
1935                                                 if (c != ' ')
1936                                                         return;
1937                                                 nl = 0;
1938                                         }
1939                                         nl++;
1940                                         rcount = 0;
1941                                 }
1942                                 continue;
1943                         }
1944
1945
1946                         if (rcount &&
1947                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1948                                 ++rcount;
1949                                 continue;
1950                         }
1951                         if (rcount) {
1952                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1953                                 padiff >>= PAGE_SHIFT;
1954                                 padiff &= PQ_L2_MASK;
1955                                 if (padiff == 0) {
1956                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1957                                         ++rcount;
1958                                         continue;
1959                                 }
1960                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1961                                         (long)fidx, rcount, (long)pa);
1962                                 db_printf("pd(%ld)\n", (long)padiff);
1963                                 if ( nl > 18) {
1964                                         c = cngetc();
1965                                         if (c != ' ')
1966                                                 return;
1967                                         nl = 0;
1968                                 }
1969                                 nl++;
1970                         }
1971                         fidx = idx;
1972                         pa = VM_PAGE_TO_PHYS(m);
1973                         rcount = 1;
1974                 }
1975                 if (rcount) {
1976                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1977                                 (long)fidx, rcount, (long)pa);
1978                         if ( nl > 18) {
1979                                 c = cngetc();
1980                                 if (c != ' ')
1981                                         return;
1982                                 nl = 0;
1983                         }
1984                         nl++;
1985                 }
1986         }
1987 }
1988 #endif /* DDB */