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 1.11 2003/11/03 17:11:23 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_inittoken(&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         zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
 222 }
 223
 224 /*
 225  *      vm_object_allocate:
 226  *
 227  *      Returns a new object with the given size.
 228  */
 229
 230 vm_object_t
 231 vm_object_allocate(type, size)
 232         objtype_t type;
 233         vm_size_t size;
 234 {
 235         vm_object_t result;
 236
 237         result = (vm_object_t) zalloc(obj_zone);
 238
 239         _vm_object_allocate(type, size, result);
 240
 241         return (result);
 242 }
 243
 244
 245 /*
 246  *      vm_object_reference:
 247  *
 248  *      Gets another reference to the given object.
 249  */
 250 void
 251 vm_object_reference(object)
 252         vm_object_t object;
 253 {
 254         if (object == NULL)
 255                 return;
 256
 257 #if 0
 258         /* object can be re-referenced during final cleaning */
 259         KASSERT(!(object->flags & OBJ_DEAD),
 260             ("vm_object_reference: attempting to reference dead obj"));
 261 #endif
 262
 263         object->ref_count++;
 264         if (object->type == OBJT_VNODE) {
 265                 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
 266                         printf("vm_object_reference: delay in getting object\n");
 267                 }
 268         }
 269 }
 270
 271 void
 272 vm_object_vndeallocate(object)
 273         vm_object_t object;
 274 {
 275         struct vnode *vp = (struct vnode *) object->handle;
 276
 277         KASSERT(object->type == OBJT_VNODE,
 278             ("vm_object_vndeallocate: not a vnode object"));
 279         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 280 #ifdef INVARIANTS
 281         if (object->ref_count == 0) {
 282                 vprint("vm_object_vndeallocate", vp);
 283                 panic("vm_object_vndeallocate: bad object reference count");
 284         }
 285 #endif
 286
 287         object->ref_count--;
 288         if (object->ref_count == 0) {
 289                 vp->v_flag &= ~VTEXT;
 290                 vm_object_clear_flag(object, OBJ_OPT);
 291         }
 292         vrele(vp);
 293 }
 294
 295 /*
 296  *      vm_object_deallocate:
 297  *
 298  *      Release a reference to the specified object,
 299  *      gained either through a vm_object_allocate
 300  *      or a vm_object_reference call.  When all references
 301  *      are gone, storage associated with this object
 302  *      may be relinquished.
 303  *
 304  *      No object may be locked.
 305  */
 306 void
 307 vm_object_deallocate(object)
 308         vm_object_t object;
 309 {
 310         vm_object_t temp;
 311
 312         while (object != NULL) {
 313
 314                 if (object->type == OBJT_VNODE) {
 315                         vm_object_vndeallocate(object);
 316                         return;
 317                 }
 318
 319                 if (object->ref_count == 0) {
 320                         panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
 321                 } else if (object->ref_count > 2) {
 322                         object->ref_count--;
 323                         return;
 324                 }
 325
 326                 /*
 327                  * Here on ref_count of one or two, which are special cases for
 328                  * objects.
 329                  */
 330                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
 331                         vm_object_set_flag(object, OBJ_ONEMAPPING);
 332                         object->ref_count--;
 333                         return;
 334                 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
 335                         object->ref_count--;
 336                         if ((object->handle == NULL) &&
 337                             (object->type == OBJT_DEFAULT ||
 338                              object->type == OBJT_SWAP)) {
 339                                 vm_object_t robject;
 340
 341                                 robject = LIST_FIRST(&object->shadow_head);
 342                                 KASSERT(robject != NULL,
 343                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 344                                          object->ref_count,
 345                                          object->shadow_count));
 346                                 if ((robject->handle == NULL) &&
 347                                     (robject->type == OBJT_DEFAULT ||
 348                                      robject->type == OBJT_SWAP)) {
 349
 350                                         robject->ref_count++;
 351
 352                                         while (
 353                                                 robject->paging_in_progress ||
 354                                                 object->paging_in_progress
 355                                         ) {
 356                                                 vm_object_pip_sleep(robject, "objde1");
 357                                                 vm_object_pip_sleep(object, "objde2");
 358                                         }
 359
 360                                         if (robject->ref_count == 1) {
 361                                                 robject->ref_count--;
 362                                                 object = robject;
 363                                                 goto doterm;
 364                                         }
 365
 366                                         object = robject;
 367                                         vm_object_collapse(object);
 368                                         continue;
 369                                 }
 370                         }
 371
 372                         return;
 373
 374                 } else {
 375                         object->ref_count--;
 376                         if (object->ref_count != 0)
 377                                 return;
 378                 }
 379
 380 doterm:
 381
 382                 temp = object->backing_object;
 383                 if (temp) {
 384                         LIST_REMOVE(object, shadow_list);
 385                         temp->shadow_count--;
 386                         if (temp->ref_count == 0)
 387                                 vm_object_clear_flag(temp, OBJ_OPT);
 388                         temp->generation++;
 389                         object->backing_object = NULL;
 390                 }
 391
 392                 /*
 393                  * Don't double-terminate, we could be in a termination
 394                  * recursion due to the terminate having to sync data
 395                  * to disk.
 396                  */
 397                 if ((object->flags & OBJ_DEAD) == 0)
 398                         vm_object_terminate(object);
 399                 object = temp;
 400         }
 401 }
 402
 403 /*
 404  *      vm_object_terminate actually destroys the specified object, freeing
 405  *      up all previously used resources.
 406  *
 407  *      The object must be locked.
 408  *      This routine may block.
 409  */
 410 void
 411 vm_object_terminate(object)
 412         vm_object_t object;
 413 {
 414         vm_page_t p;
 415         int s;
 416
 417         /*
 418          * Make sure no one uses us.
 419          */
 420         vm_object_set_flag(object, OBJ_DEAD);
 421
 422         /*
 423          * wait for the pageout daemon to be done with the object
 424          */
 425         vm_object_pip_wait(object, "objtrm");
 426
 427         KASSERT(!object->paging_in_progress,
 428                 ("vm_object_terminate: pageout in progress"));
 429
 430         /*
 431          * Clean and free the pages, as appropriate. All references to the
 432          * object are gone, so we don't need to lock it.
 433          */
 434         if (object->type == OBJT_VNODE) {
 435                 struct vnode *vp;
 436
 437                 /*
 438                  * Freeze optimized copies.
 439                  */
 440                 vm_freeze_copyopts(object, 0, object->size);
 441
 442                 /*
 443                  * Clean pages and flush buffers.
 444                  */
 445                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 446
 447                 vp = (struct vnode *) object->handle;
 448                 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
 449         }
 450
 451         /*
 452          * Wait for any I/O to complete, after which there had better not
 453          * be any references left on the object.
 454          */
 455         vm_object_pip_wait(object, "objtrm");
 456
 457         if (object->ref_count != 0)
 458                 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
 459
 460         /*
 461          * Now free any remaining pages. For internal objects, this also
 462          * removes them from paging queues. Don't free wired pages, just
 463          * remove them from the object.
 464          */
 465         s = splvm();
 466         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
 467                 if (p->busy || (p->flags & PG_BUSY))
 468                         panic("vm_object_terminate: freeing busy page %p\n", p);
 469                 if (p->wire_count == 0) {
 470                         vm_page_busy(p);
 471                         vm_page_free(p);
 472                         mycpu->gd_cnt.v_pfree++;
 473                 } else {
 474                         vm_page_busy(p);
 475                         vm_page_remove(p);
 476                 }
 477         }
 478         splx(s);
 479
 480         /*
 481          * Let the pager know object is dead.
 482          */
 483         vm_pager_deallocate(object);
 484
 485         /*
 486          * Remove the object from the global object list.
 487          */
 488         lwkt_gettoken(&vm_object_list_token);
 489         TAILQ_REMOVE(&vm_object_list, object, object_list);
 490         lwkt_reltoken(&vm_object_list_token);
 491
 492         wakeup(object);
 493
 494         /*
 495          * Free the space for the object.
 496          */
 497         zfree(obj_zone, object);
 498 }
 499
 500 /*
 501  *      vm_object_page_clean
 502  *
 503  *      Clean all dirty pages in the specified range of object.  Leaves page
 504  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 505  *      write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
 506  *      leaving the object dirty.
 507  *
 508  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 509  *      synchronous clustering mode implementation.
 510  *
 511  *      Odd semantics: if start == end, we clean everything.
 512  *
 513  *      The object must be locked.
 514  */
 515
 516 void
 517 vm_object_page_clean(object, start, end, flags)
 518         vm_object_t object;
 519         vm_pindex_t start;
 520         vm_pindex_t end;
 521         int flags;
 522 {
 523         vm_page_t p, np;
 524         vm_offset_t tstart, tend;
 525         vm_pindex_t pi;
 526         struct vnode *vp;
 527         int clearobjflags;
 528         int pagerflags;
 529         int curgeneration;
 530
 531         if (object->type != OBJT_VNODE ||
 532                 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
 533                 return;
 534
 535         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 536         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
 537
 538         vp = object->handle;
 539
 540         vm_object_set_flag(object, OBJ_CLEANING);
 541
 542         /*
 543          * Handle 'entire object' case
 544          */
 545         tstart = start;
 546         if (end == 0) {
 547                 tend = object->size;
 548         } else {
 549                 tend = end;
 550         }
 551
 552         /*
 553          * If the caller is smart and only msync()s a range he knows is
 554          * dirty, we may be able to avoid an object scan.  This results in
 555          * a phenominal improvement in performance.  We cannot do this
 556          * as a matter of course because the object may be huge - e.g.
 557          * the size might be in the gigabytes or terrabytes.
 558          */
 559         if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
 560                 vm_offset_t tscan;
 561                 int scanlimit;
 562                 int scanreset;
 563
 564                 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
 565                 if (scanreset < 16)
 566                         scanreset = 16;
 567                 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
 568
 569                 scanlimit = scanreset;
 570                 tscan = tstart;
 571                 while (tscan < tend) {
 572                         curgeneration = object->generation;
 573                         p = vm_page_lookup(object, tscan);
 574                         if (p == NULL || p->valid == 0 ||
 575                             (p->queue - p->pc) == PQ_CACHE) {
 576                                 if (--scanlimit == 0)
 577                                         break;
 578                                 ++tscan;
 579                                 continue;
 580                         }
 581                         vm_page_test_dirty(p);
 582                         if ((p->dirty & p->valid) == 0) {
 583                                 if (--scanlimit == 0)
 584                                         break;
 585                                 ++tscan;
 586                                 continue;
 587                         }
 588                         /*
 589                          * If we have been asked to skip nosync pages and
 590                          * this is a nosync page, we can't continue.
 591                          */
 592                         if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
 593                                 if (--scanlimit == 0)
 594                                         break;
 595                                 ++tscan;
 596                                 continue;
 597                         }
 598                         scanlimit = scanreset;
 599
 600                         /*
 601                          * This returns 0 if it was unable to busy the first
 602                          * page (i.e. had to sleep).
 603                          */
 604                         tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
 605                 }
 606
 607                 /*
 608                  * If everything was dirty and we flushed it successfully,
 609                  * and the requested range is not the entire object, we
 610                  * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
 611                  * return immediately.
 612                  */
 613                 if (tscan >= tend && (tstart || tend < object->size)) {
 614                         vm_object_clear_flag(object, OBJ_CLEANING);
 615                         return;
 616                 }
 617                 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
 618         }
 619
 620         /*
 621          * Generally set CLEANCHK interlock and make the page read-only so
 622          * we can then clear the object flags.
 623          *
 624          * However, if this is a nosync mmap then the object is likely to
 625          * stay dirty so do not mess with the page and do not clear the
 626          * object flags.
 627          */
 628
 629         clearobjflags = 1;
 630
 631         for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
 632                 vm_page_flag_set(p, PG_CLEANCHK);
 633                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
 634                         clearobjflags = 0;
 635                 else
 636                         vm_page_protect(p, VM_PROT_READ);
 637         }
 638
 639         if (clearobjflags && (tstart == 0) && (tend == object->size)) {
 640                 struct vnode *vp;
 641
 642                 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
 643                 if (object->type == OBJT_VNODE &&
 644                     (vp = (struct vnode *)object->handle) != NULL) {
 645                         if (vp->v_flag & VOBJDIRTY) {
 646                                 lwkt_gettoken(&vp->v_interlock);
 647                                 vp->v_flag &= ~VOBJDIRTY;
 648                                 lwkt_reltoken(&vp->v_interlock);
 649                         }
 650                 }
 651         }
 652
 653 rescan:
 654         curgeneration = object->generation;
 655
 656         for(p = TAILQ_FIRST(&object->memq); p; p = np) {
 657                 int n;
 658
 659                 np = TAILQ_NEXT(p, listq);
 660
 661 again:
 662                 pi = p->pindex;
 663                 if (((p->flags & PG_CLEANCHK) == 0) ||
 664                         (pi < tstart) || (pi >= tend) ||
 665                         (p->valid == 0) ||
 666                         ((p->queue - p->pc) == PQ_CACHE)) {
 667                         vm_page_flag_clear(p, PG_CLEANCHK);
 668                         continue;
 669                 }
 670
 671                 vm_page_test_dirty(p);
 672                 if ((p->dirty & p->valid) == 0) {
 673                         vm_page_flag_clear(p, PG_CLEANCHK);
 674                         continue;
 675                 }
 676
 677                 /*
 678                  * If we have been asked to skip nosync pages and this is a
 679                  * nosync page, skip it.  Note that the object flags were
 680                  * not cleared in this case so we do not have to set them.
 681                  */
 682                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
 683                         vm_page_flag_clear(p, PG_CLEANCHK);
 684                         continue;
 685                 }
 686
 687                 n = vm_object_page_collect_flush(object, p,
 688                         curgeneration, pagerflags);
 689                 if (n == 0)
 690                         goto rescan;
 691                 if (object->generation != curgeneration)
 692                         goto rescan;
 693
 694                 /*
 695                  * Try to optimize the next page.  If we can't we pick up
 696                  * our (random) scan where we left off.
 697                  */
 698                 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
 699                         if ((p = vm_page_lookup(object, pi + n)) != NULL)
 700                                 goto again;
 701                 }
 702         }
 703
 704 #if 0
 705         VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
 706 #endif
 707
 708         vm_object_clear_flag(object, OBJ_CLEANING);
 709         return;
 710 }
 711
 712 static int
 713 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
 714 {
 715         int runlen;
 716         int s;
 717         int maxf;
 718         int chkb;
 719         int maxb;
 720         int i;
 721         vm_pindex_t pi;
 722         vm_page_t maf[vm_pageout_page_count];
 723         vm_page_t mab[vm_pageout_page_count];
 724         vm_page_t ma[vm_pageout_page_count];
 725
 726         s = splvm();
 727         pi = p->pindex;
 728         while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
 729                 if (object->generation != curgeneration) {
 730                         splx(s);
 731                         return(0);
 732                 }
 733         }
 734
 735         maxf = 0;
 736         for(i = 1; i < vm_pageout_page_count; i++) {
 737                 vm_page_t tp;
 738
 739                 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
 740                         if ((tp->flags & PG_BUSY) ||
 741                                 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
 742                                  (tp->flags & PG_CLEANCHK) == 0) ||
 743                                 (tp->busy != 0))
 744                                 break;
 745                         if((tp->queue - tp->pc) == PQ_CACHE) {
 746                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 747                                 break;
 748                         }
 749                         vm_page_test_dirty(tp);
 750                         if ((tp->dirty & tp->valid) == 0) {
 751                                 vm_page_flag_clear(tp, PG_CLEANCHK);
 752                                 break;
 753                         }
 754                         maf[ i - 1 ] = tp;
 755                         maxf++;
 756                         continue;
 757                 }
 758                 break;
 759         }
 760
 761         maxb = 0;
 762         chkb = vm_pageout_page_count -  maxf;
 763         if (chkb) {
 764                 for(i = 1; i < chkb;i++) {
 765                         vm_page_t tp;
 766
 767                         if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
 768                                 if ((tp->flags & PG_BUSY) ||
 769                                         ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
 770                                          (tp->flags & PG_CLEANCHK) == 0) ||
 771                                         (tp->busy != 0))
 772                                         break;
 773                                 if((tp->queue - tp->pc) == PQ_CACHE) {
 774                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 775                                         break;
 776                                 }
 777                                 vm_page_test_dirty(tp);
 778                                 if ((tp->dirty & tp->valid) == 0) {
 779                                         vm_page_flag_clear(tp, PG_CLEANCHK);
 780                                         break;
 781                                 }
 782                                 mab[ i - 1 ] = tp;
 783                                 maxb++;
 784                                 continue;
 785                         }
 786                         break;
 787                 }
 788         }
 789
 790         for(i = 0; i < maxb; i++) {
 791                 int index = (maxb - i) - 1;
 792                 ma[index] = mab[i];
 793                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
 794         }
 795         vm_page_flag_clear(p, PG_CLEANCHK);
 796         ma[maxb] = p;
 797         for(i = 0; i < maxf; i++) {
 798                 int index = (maxb + i) + 1;
 799                 ma[index] = maf[i];
 800                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
 801         }
 802         runlen = maxb + maxf + 1;
 803
 804         splx(s);
 805         vm_pageout_flush(ma, runlen, pagerflags);
 806         for (i = 0; i < runlen; i++) {
 807                 if (ma[i]->valid & ma[i]->dirty) {
 808                         vm_page_protect(ma[i], VM_PROT_READ);
 809                         vm_page_flag_set(ma[i], PG_CLEANCHK);
 810
 811                         /*
 812                          * maxf will end up being the actual number of pages
 813                          * we wrote out contiguously, non-inclusive of the
 814                          * first page.  We do not count look-behind pages.
 815                          */
 816                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
 817                                 maxf = i - maxb - 1;
 818                 }
 819         }
 820         return(maxf + 1);
 821 }
 822
 823 #ifdef not_used
 824 /* XXX I cannot tell if this should be an exported symbol */
 825 /*
 826  *      vm_object_deactivate_pages
 827  *
 828  *      Deactivate all pages in the specified object.  (Keep its pages
 829  *      in memory even though it is no longer referenced.)
 830  *
 831  *      The object must be locked.
 832  */
 833 static void
 834 vm_object_deactivate_pages(object)
 835         vm_object_t object;
 836 {
 837         vm_page_t p, next;
 838
 839         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
 840                 next = TAILQ_NEXT(p, listq);
 841                 vm_page_deactivate(p);
 842         }
 843 }
 844 #endif
 845
 846 /*
 847  * Same as vm_object_pmap_copy, except range checking really
 848  * works, and is meant for small sections of an object.
 849  *
 850  * This code protects resident pages by making them read-only
 851  * and is typically called on a fork or split when a page
 852  * is converted to copy-on-write.
 853  *
 854  * NOTE: If the page is already at VM_PROT_NONE, calling
 855  * vm_page_protect will have no effect.
 856  */
 857
 858 void
 859 vm_object_pmap_copy_1(object, start, end)
 860         vm_object_t object;
 861         vm_pindex_t start;
 862         vm_pindex_t end;
 863 {
 864         vm_pindex_t idx;
 865         vm_page_t p;
 866
 867         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
 868                 return;
 869
 870         for (idx = start; idx < end; idx++) {
 871                 p = vm_page_lookup(object, idx);
 872                 if (p == NULL)
 873                         continue;
 874                 vm_page_protect(p, VM_PROT_READ);
 875         }
 876 }
 877
 878 /*
 879  *      vm_object_pmap_remove:
 880  *
 881  *      Removes all physical pages in the specified
 882  *      object range from all physical maps.
 883  *
 884  *      The object must *not* be locked.
 885  */
 886 void
 887 vm_object_pmap_remove(object, start, end)
 888         vm_object_t object;
 889         vm_pindex_t start;
 890         vm_pindex_t end;
 891 {
 892         vm_page_t p;
 893
 894         if (object == NULL)
 895                 return;
 896         for (p = TAILQ_FIRST(&object->memq);
 897                 p != NULL;
 898                 p = TAILQ_NEXT(p, listq)) {
 899                 if (p->pindex >= start && p->pindex < end)
 900                         vm_page_protect(p, VM_PROT_NONE);
 901         }
 902         if ((start == 0) && (object->size == end))
 903                 vm_object_clear_flag(object, OBJ_WRITEABLE);
 904 }
 905
 906 /*
 907  *      vm_object_madvise:
 908  *
 909  *      Implements the madvise function at the object/page level.
 910  *
 911  *      MADV_WILLNEED   (any object)
 912  *
 913  *          Activate the specified pages if they are resident.
 914  *
 915  *      MADV_DONTNEED   (any object)
 916  *
 917  *          Deactivate the specified pages if they are resident.
 918  *
 919  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
 920  *                       OBJ_ONEMAPPING only)
 921  *
 922  *          Deactivate and clean the specified pages if they are
 923  *          resident.  This permits the process to reuse the pages
 924  *          without faulting or the kernel to reclaim the pages
 925  *          without I/O.
 926  */
 927 void
 928 vm_object_madvise(object, pindex, count, advise)
 929         vm_object_t object;
 930         vm_pindex_t pindex;
 931         int count;
 932         int advise;
 933 {
 934         vm_pindex_t end, tpindex;
 935         vm_object_t tobject;
 936         vm_page_t m;
 937
 938         if (object == NULL)
 939                 return;
 940
 941         end = pindex + count;
 942
 943         /*
 944          * Locate and adjust resident pages
 945          */
 946
 947         for (; pindex < end; pindex += 1) {
 948 relookup:
 949                 tobject = object;
 950                 tpindex = pindex;
 951 shadowlookup:
 952                 /*
 953                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
 954                  * and those pages must be OBJ_ONEMAPPING.
 955                  */
 956                 if (advise == MADV_FREE) {
 957                         if ((tobject->type != OBJT_DEFAULT &&
 958                              tobject->type != OBJT_SWAP) ||
 959                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
 960                                 continue;
 961                         }
 962                 }
 963
 964                 m = vm_page_lookup(tobject, tpindex);
 965
 966                 if (m == NULL) {
 967                         /*
 968                          * There may be swap even if there is no backing page
 969                          */
 970                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
 971                                 swap_pager_freespace(tobject, tpindex, 1);
 972
 973                         /*
 974                          * next object
 975                          */
 976                         tobject = tobject->backing_object;
 977                         if (tobject == NULL)
 978                                 continue;
 979                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
 980                         goto shadowlookup;
 981                 }
 982
 983                 /*
 984                  * If the page is busy or not in a normal active state,
 985                  * we skip it.  If the page is not managed there are no
 986                  * page queues to mess with.  Things can break if we mess
 987                  * with pages in any of the below states.
 988                  */
 989                 if (
 990                     m->hold_count ||
 991                     m->wire_count ||
 992                     (m->flags & PG_UNMANAGED) ||
 993                     m->valid != VM_PAGE_BITS_ALL
 994                 ) {
 995                         continue;
 996                 }
 997
 998                 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
 999                         goto relookup;
1000
1001                 if (advise == MADV_WILLNEED) {
1002                         vm_page_activate(m);
1003                 } else if (advise == MADV_DONTNEED) {
1004                         vm_page_dontneed(m);
1005                 } else if (advise == MADV_FREE) {
1006                         /*
1007                          * Mark the page clean.  This will allow the page
1008                          * to be freed up by the system.  However, such pages
1009                          * are often reused quickly by malloc()/free()
1010                          * so we do not do anything that would cause
1011                          * a page fault if we can help it.
1012                          *
1013                          * Specifically, we do not try to actually free
1014                          * the page now nor do we try to put it in the
1015                          * cache (which would cause a page fault on reuse).
1016                          *
1017                          * But we do make the page is freeable as we
1018                          * can without actually taking the step of unmapping
1019                          * it.
1020                          */
1021                         pmap_clear_modify(m);
1022                         m->dirty = 0;
1023                         m->act_count = 0;
1024                         vm_page_dontneed(m);
1025                         if (tobject->type == OBJT_SWAP)
1026                                 swap_pager_freespace(tobject, tpindex, 1);
1027                 }
1028         }
1029 }
1030
1031 /*
1032  *      vm_object_shadow:
1033  *
1034  *      Create a new object which is backed by the
1035  *      specified existing object range.  The source
1036  *      object reference is deallocated.
1037  *
1038  *      The new object and offset into that object
1039  *      are returned in the source parameters.
1040  */
1041
1042 void
1043 vm_object_shadow(object, offset, length)
1044         vm_object_t *object;    /* IN/OUT */
1045         vm_ooffset_t *offset;   /* IN/OUT */
1046         vm_size_t length;
1047 {
1048         vm_object_t source;
1049         vm_object_t result;
1050
1051         source = *object;
1052
1053         /*
1054          * Don't create the new object if the old object isn't shared.
1055          */
1056
1057         if (source != NULL &&
1058             source->ref_count == 1 &&
1059             source->handle == NULL &&
1060             (source->type == OBJT_DEFAULT ||
1061              source->type == OBJT_SWAP))
1062                 return;
1063
1064         /*
1065          * Allocate a new object with the given length
1066          */
1067
1068         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1069                 panic("vm_object_shadow: no object for shadowing");
1070
1071         /*
1072          * The new object shadows the source object, adding a reference to it.
1073          * Our caller changes his reference to point to the new object,
1074          * removing a reference to the source object.  Net result: no change
1075          * of reference count.
1076          *
1077          * Try to optimize the result object's page color when shadowing
1078          * in order to maintain page coloring consistency in the combined
1079          * shadowed object.
1080          */
1081         result->backing_object = source;
1082         if (source) {
1083                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1084                 source->shadow_count++;
1085                 source->generation++;
1086                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1087         }
1088
1089         /*
1090          * Store the offset into the source object, and fix up the offset into
1091          * the new object.
1092          */
1093
1094         result->backing_object_offset = *offset;
1095
1096         /*
1097          * Return the new things
1098          */
1099
1100         *offset = 0;
1101         *object = result;
1102 }
1103
1104 #define OBSC_TEST_ALL_SHADOWED  0x0001
1105 #define OBSC_COLLAPSE_NOWAIT    0x0002
1106 #define OBSC_COLLAPSE_WAIT      0x0004
1107
1108 static __inline int
1109 vm_object_backing_scan(vm_object_t object, int op)
1110 {
1111         int s;
1112         int r = 1;
1113         vm_page_t p;
1114         vm_object_t backing_object;
1115         vm_pindex_t backing_offset_index;
1116
1117         s = splvm();
1118
1119         backing_object = object->backing_object;
1120         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1121
1122         /*
1123          * Initial conditions
1124          */
1125
1126         if (op & OBSC_TEST_ALL_SHADOWED) {
1127                 /*
1128                  * We do not want to have to test for the existence of
1129                  * swap pages in the backing object.  XXX but with the
1130                  * new swapper this would be pretty easy to do.
1131                  *
1132                  * XXX what about anonymous MAP_SHARED memory that hasn't
1133                  * been ZFOD faulted yet?  If we do not test for this, the
1134                  * shadow test may succeed! XXX
1135                  */
1136                 if (backing_object->type != OBJT_DEFAULT) {
1137                         splx(s);
1138                         return(0);
1139                 }
1140         }
1141         if (op & OBSC_COLLAPSE_WAIT) {
1142                 vm_object_set_flag(backing_object, OBJ_DEAD);
1143         }
1144
1145         /*
1146          * Our scan
1147          */
1148
1149         p = TAILQ_FIRST(&backing_object->memq);
1150         while (p) {
1151                 vm_page_t next = TAILQ_NEXT(p, listq);
1152                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1153
1154                 if (op & OBSC_TEST_ALL_SHADOWED) {
1155                         vm_page_t pp;
1156
1157                         /*
1158                          * Ignore pages outside the parent object's range
1159                          * and outside the parent object's mapping of the
1160                          * backing object.
1161                          *
1162                          * note that we do not busy the backing object's
1163                          * page.
1164                          */
1165
1166                         if (
1167                             p->pindex < backing_offset_index ||
1168                             new_pindex >= object->size
1169                         ) {
1170                                 p = next;
1171                                 continue;
1172                         }
1173
1174                         /*
1175                          * See if the parent has the page or if the parent's
1176                          * object pager has the page.  If the parent has the
1177                          * page but the page is not valid, the parent's
1178                          * object pager must have the page.
1179                          *
1180                          * If this fails, the parent does not completely shadow
1181                          * the object and we might as well give up now.
1182                          */
1183
1184                         pp = vm_page_lookup(object, new_pindex);
1185                         if (
1186                             (pp == NULL || pp->valid == 0) &&
1187                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1188                         ) {
1189                                 r = 0;
1190                                 break;
1191                         }
1192                 }
1193
1194                 /*
1195                  * Check for busy page
1196                  */
1197
1198                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1199                         vm_page_t pp;
1200
1201                         if (op & OBSC_COLLAPSE_NOWAIT) {
1202                                 if (
1203                                     (p->flags & PG_BUSY) ||
1204                                     !p->valid ||
1205                                     p->hold_count ||
1206                                     p->wire_count ||
1207                                     p->busy
1208                                 ) {
1209                                         p = next;
1210                                         continue;
1211                                 }
1212                         } else if (op & OBSC_COLLAPSE_WAIT) {
1213                                 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1214                                         /*
1215                                          * If we slept, anything could have
1216                                          * happened.  Since the object is
1217                                          * marked dead, the backing offset
1218                                          * should not have changed so we
1219                                          * just restart our scan.
1220                                          */
1221                                         p = TAILQ_FIRST(&backing_object->memq);
1222                                         continue;
1223                                 }
1224                         }
1225
1226                         /*
1227                          * Busy the page
1228                          */
1229                         vm_page_busy(p);
1230
1231                         KASSERT(
1232                             p->object == backing_object,
1233                             ("vm_object_qcollapse(): object mismatch")
1234                         );
1235
1236                         /*
1237                          * Destroy any associated swap
1238                          */
1239                         if (backing_object->type == OBJT_SWAP) {
1240                                 swap_pager_freespace(
1241                                     backing_object,
1242                                     p->pindex,
1243                                     1
1244                                 );
1245                         }
1246
1247                         if (
1248                             p->pindex < backing_offset_index ||
1249                             new_pindex >= object->size
1250                         ) {
1251                                 /*
1252                                  * Page is out of the parent object's range, we
1253                                  * can simply destroy it.
1254                                  */
1255                                 vm_page_protect(p, VM_PROT_NONE);
1256                                 vm_page_free(p);
1257                                 p = next;
1258                                 continue;
1259                         }
1260
1261                         pp = vm_page_lookup(object, new_pindex);
1262                         if (
1263                             pp != NULL ||
1264                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1265                         ) {
1266                                 /*
1267                                  * page already exists in parent OR swap exists
1268                                  * for this location in the parent.  Destroy
1269                                  * the original page from the backing object.
1270                                  *
1271                                  * Leave the parent's page alone
1272                                  */
1273                                 vm_page_protect(p, VM_PROT_NONE);
1274                                 vm_page_free(p);
1275                                 p = next;
1276                                 continue;
1277                         }
1278
1279                         /*
1280                          * Page does not exist in parent, rename the
1281                          * page from the backing object to the main object.
1282                          *
1283                          * If the page was mapped to a process, it can remain
1284                          * mapped through the rename.
1285                          */
1286                         if ((p->queue - p->pc) == PQ_CACHE)
1287                                 vm_page_deactivate(p);
1288
1289                         vm_page_rename(p, object, new_pindex);
1290                         /* page automatically made dirty by rename */
1291                 }
1292                 p = next;
1293         }
1294         splx(s);
1295         return(r);
1296 }
1297
1298
1299 /*
1300  * this version of collapse allows the operation to occur earlier and
1301  * when paging_in_progress is true for an object...  This is not a complete
1302  * operation, but should plug 99.9% of the rest of the leaks.
1303  */
1304 static void
1305 vm_object_qcollapse(object)
1306         vm_object_t object;
1307 {
1308         vm_object_t backing_object = object->backing_object;
1309
1310         if (backing_object->ref_count != 1)
1311                 return;
1312
1313         backing_object->ref_count += 2;
1314
1315         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1316
1317         backing_object->ref_count -= 2;
1318 }
1319
1320 /*
1321  *      vm_object_collapse:
1322  *
1323  *      Collapse an object with the object backing it.
1324  *      Pages in the backing object are moved into the
1325  *      parent, and the backing object is deallocated.
1326  */
1327 void
1328 vm_object_collapse(object)
1329         vm_object_t object;
1330 {
1331         while (TRUE) {
1332                 vm_object_t backing_object;
1333
1334                 /*
1335                  * Verify that the conditions are right for collapse:
1336                  *
1337                  * The object exists and the backing object exists.
1338                  */
1339                 if (object == NULL)
1340                         break;
1341
1342                 if ((backing_object = object->backing_object) == NULL)
1343                         break;
1344
1345                 /*
1346                  * we check the backing object first, because it is most likely
1347                  * not collapsable.
1348                  */
1349                 if (backing_object->handle != NULL ||
1350                     (backing_object->type != OBJT_DEFAULT &&
1351                      backing_object->type != OBJT_SWAP) ||
1352                     (backing_object->flags & OBJ_DEAD) ||
1353                     object->handle != NULL ||
1354                     (object->type != OBJT_DEFAULT &&
1355                      object->type != OBJT_SWAP) ||
1356                     (object->flags & OBJ_DEAD)) {
1357                         break;
1358                 }
1359
1360                 if (
1361                     object->paging_in_progress != 0 ||
1362                     backing_object->paging_in_progress != 0
1363                 ) {
1364                         vm_object_qcollapse(object);
1365                         break;
1366                 }
1367
1368                 /*
1369                  * We know that we can either collapse the backing object (if
1370                  * the parent is the only reference to it) or (perhaps) have
1371                  * the parent bypass the object if the parent happens to shadow
1372                  * all the resident pages in the entire backing object.
1373                  *
1374                  * This is ignoring pager-backed pages such as swap pages.
1375                  * vm_object_backing_scan fails the shadowing test in this
1376                  * case.
1377                  */
1378
1379                 if (backing_object->ref_count == 1) {
1380                         /*
1381                          * If there is exactly one reference to the backing
1382                          * object, we can collapse it into the parent.
1383                          */
1384
1385                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1386
1387                         /*
1388                          * Move the pager from backing_object to object.
1389                          */
1390
1391                         if (backing_object->type == OBJT_SWAP) {
1392                                 vm_object_pip_add(backing_object, 1);
1393
1394                                 /*
1395                                  * scrap the paging_offset junk and do a
1396                                  * discrete copy.  This also removes major
1397                                  * assumptions about how the swap-pager
1398                                  * works from where it doesn't belong.  The
1399                                  * new swapper is able to optimize the
1400                                  * destroy-source case.
1401                                  */
1402
1403                                 vm_object_pip_add(object, 1);
1404                                 swap_pager_copy(
1405                                     backing_object,
1406                                     object,
1407                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1408                                 vm_object_pip_wakeup(object);
1409
1410                                 vm_object_pip_wakeup(backing_object);
1411                         }
1412                         /*
1413                          * Object now shadows whatever backing_object did.
1414                          * Note that the reference to
1415                          * backing_object->backing_object moves from within
1416                          * backing_object to within object.
1417                          */
1418
1419                         LIST_REMOVE(object, shadow_list);
1420                         object->backing_object->shadow_count--;
1421                         object->backing_object->generation++;
1422                         if (backing_object->backing_object) {
1423                                 LIST_REMOVE(backing_object, shadow_list);
1424                                 backing_object->backing_object->shadow_count--;
1425                                 backing_object->backing_object->generation++;
1426                         }
1427                         object->backing_object = backing_object->backing_object;
1428                         if (object->backing_object) {
1429                                 LIST_INSERT_HEAD(
1430                                     &object->backing_object->shadow_head,
1431                                     object,
1432                                     shadow_list
1433                                 );
1434                                 object->backing_object->shadow_count++;
1435                                 object->backing_object->generation++;
1436                         }
1437
1438                         object->backing_object_offset +=
1439                             backing_object->backing_object_offset;
1440
1441                         /*
1442                          * Discard backing_object.
1443                          *
1444                          * Since the backing object has no pages, no pager left,
1445                          * and no object references within it, all that is
1446                          * necessary is to dispose of it.
1447                          */
1448
1449                         KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1450                         KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1451                         TAILQ_REMOVE(
1452                             &vm_object_list,
1453                             backing_object,
1454                             object_list
1455                         );
1456                         vm_object_count--;
1457
1458                         zfree(obj_zone, backing_object);
1459
1460                         object_collapses++;
1461                 } else {
1462                         vm_object_t new_backing_object;
1463
1464                         /*
1465                          * If we do not entirely shadow the backing object,
1466                          * there is nothing we can do so we give up.
1467                          */
1468
1469                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1470                                 break;
1471                         }
1472
1473                         /*
1474                          * Make the parent shadow the next object in the
1475                          * chain.  Deallocating backing_object will not remove
1476                          * it, since its reference count is at least 2.
1477                          */
1478
1479                         LIST_REMOVE(object, shadow_list);
1480                         backing_object->shadow_count--;
1481                         backing_object->generation++;
1482
1483                         new_backing_object = backing_object->backing_object;
1484                         if ((object->backing_object = new_backing_object) != NULL) {
1485                                 vm_object_reference(new_backing_object);
1486                                 LIST_INSERT_HEAD(
1487                                     &new_backing_object->shadow_head,
1488                                     object,
1489                                     shadow_list
1490                                 );
1491                                 new_backing_object->shadow_count++;
1492                                 new_backing_object->generation++;
1493                                 object->backing_object_offset +=
1494                                         backing_object->backing_object_offset;
1495                         }
1496
1497                         /*
1498                          * Drop the reference count on backing_object. Since
1499                          * its ref_count was at least 2, it will not vanish;
1500                          * so we don't need to call vm_object_deallocate, but
1501                          * we do anyway.
1502                          */
1503                         vm_object_deallocate(backing_object);
1504                         object_bypasses++;
1505                 }
1506
1507                 /*
1508                  * Try again with this object's new backing object.
1509                  */
1510         }
1511 }
1512
1513 /*
1514  *      vm_object_page_remove: [internal]
1515  *
1516  *      Removes all physical pages in the specified
1517  *      object range from the object's list of pages.
1518  *
1519  *      The object must be locked.
1520  */
1521 void
1522 vm_object_page_remove(object, start, end, clean_only)
1523         vm_object_t object;
1524         vm_pindex_t start;
1525         vm_pindex_t end;
1526         boolean_t clean_only;
1527 {
1528         vm_page_t p, next;
1529         unsigned int size;
1530         int all;
1531
1532         if (object == NULL ||
1533             object->resident_page_count == 0)
1534                 return;
1535
1536         all = ((end == 0) && (start == 0));
1537
1538         /*
1539          * Since physically-backed objects do not use managed pages, we can't
1540          * remove pages from the object (we must instead remove the page
1541          * references, and then destroy the object).
1542          */
1543         KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1544
1545         vm_object_pip_add(object, 1);
1546 again:
1547         size = end - start;
1548         if (all || size > object->resident_page_count / 4) {
1549                 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1550                         next = TAILQ_NEXT(p, listq);
1551                         if (all || ((start <= p->pindex) && (p->pindex < end))) {
1552                                 if (p->wire_count != 0) {
1553                                         vm_page_protect(p, VM_PROT_NONE);
1554                                         if (!clean_only)
1555                                                 p->valid = 0;
1556                                         continue;
1557                                 }
1558
1559                                 /*
1560                                  * The busy flags are only cleared at
1561                                  * interrupt -- minimize the spl transitions
1562                                  */
1563
1564                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1565                                         goto again;
1566
1567                                 if (clean_only && p->valid) {
1568                                         vm_page_test_dirty(p);
1569                                         if (p->valid & p->dirty)
1570                                                 continue;
1571                                 }
1572
1573                                 vm_page_busy(p);
1574                                 vm_page_protect(p, VM_PROT_NONE);
1575                                 vm_page_free(p);
1576                         }
1577                 }
1578         } else {
1579                 while (size > 0) {
1580                         if ((p = vm_page_lookup(object, start)) != 0) {
1581
1582                                 if (p->wire_count != 0) {
1583                                         vm_page_protect(p, VM_PROT_NONE);
1584                                         if (!clean_only)
1585                                                 p->valid = 0;
1586                                         start += 1;
1587                                         size -= 1;
1588                                         continue;
1589                                 }
1590
1591                                 /*
1592                                  * The busy flags are only cleared at
1593                                  * interrupt -- minimize the spl transitions
1594                                  */
1595                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1596                                         goto again;
1597
1598                                 if (clean_only && p->valid) {
1599                                         vm_page_test_dirty(p);
1600                                         if (p->valid & p->dirty) {
1601                                                 start += 1;
1602                                                 size -= 1;
1603                                                 continue;
1604                                         }
1605                                 }
1606
1607                                 vm_page_busy(p);
1608                                 vm_page_protect(p, VM_PROT_NONE);
1609                                 vm_page_free(p);
1610                         }
1611                         start += 1;
1612                         size -= 1;
1613                 }
1614         }
1615         vm_object_pip_wakeup(object);
1616 }
1617
1618 /*
1619  *      Routine:        vm_object_coalesce
1620  *      Function:       Coalesces two objects backing up adjoining
1621  *                      regions of memory into a single object.
1622  *
1623  *      returns TRUE if objects were combined.
1624  *
1625  *      NOTE:   Only works at the moment if the second object is NULL -
1626  *              if it's not, which object do we lock first?
1627  *
1628  *      Parameters:
1629  *              prev_object     First object to coalesce
1630  *              prev_offset     Offset into prev_object
1631  *              next_object     Second object into coalesce
1632  *              next_offset     Offset into next_object
1633  *
1634  *              prev_size       Size of reference to prev_object
1635  *              next_size       Size of reference to next_object
1636  *
1637  *      Conditions:
1638  *      The object must *not* be locked.
1639  */
1640 boolean_t
1641 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1642         vm_object_t prev_object;
1643         vm_pindex_t prev_pindex;
1644         vm_size_t prev_size, next_size;
1645 {
1646         vm_pindex_t next_pindex;
1647
1648         if (prev_object == NULL) {
1649                 return (TRUE);
1650         }
1651
1652         if (prev_object->type != OBJT_DEFAULT &&
1653             prev_object->type != OBJT_SWAP) {
1654                 return (FALSE);
1655         }
1656
1657         /*
1658          * Try to collapse the object first
1659          */
1660         vm_object_collapse(prev_object);
1661
1662         /*
1663          * Can't coalesce if: . more than one reference . paged out . shadows
1664          * another object . has a copy elsewhere (any of which mean that the
1665          * pages not mapped to prev_entry may be in use anyway)
1666          */
1667
1668         if (prev_object->backing_object != NULL) {
1669                 return (FALSE);
1670         }
1671
1672         prev_size >>= PAGE_SHIFT;
1673         next_size >>= PAGE_SHIFT;
1674         next_pindex = prev_pindex + prev_size;
1675
1676         if ((prev_object->ref_count > 1) &&
1677             (prev_object->size != next_pindex)) {
1678                 return (FALSE);
1679         }
1680
1681         /*
1682          * Remove any pages that may still be in the object from a previous
1683          * deallocation.
1684          */
1685         if (next_pindex < prev_object->size) {
1686                 vm_object_page_remove(prev_object,
1687                                       next_pindex,
1688                                       next_pindex + next_size, FALSE);
1689                 if (prev_object->type == OBJT_SWAP)
1690                         swap_pager_freespace(prev_object,
1691                                              next_pindex, next_size);
1692         }
1693
1694         /*
1695          * Extend the object if necessary.
1696          */
1697         if (next_pindex + next_size > prev_object->size)
1698                 prev_object->size = next_pindex + next_size;
1699
1700         return (TRUE);
1701 }
1702
1703 void
1704 vm_object_set_writeable_dirty(vm_object_t object)
1705 {
1706         struct vnode *vp;
1707
1708         vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1709         if (object->type == OBJT_VNODE &&
1710             (vp = (struct vnode *)object->handle) != NULL) {
1711                 if ((vp->v_flag & VOBJDIRTY) == 0) {
1712                         lwkt_gettoken(&vp->v_interlock);
1713                         vp->v_flag |= VOBJDIRTY;
1714                         lwkt_reltoken(&vp->v_interlock);
1715                 }
1716         }
1717 }
1718
1719
1720
1721 #include "opt_ddb.h"
1722 #ifdef DDB
1723 #include <sys/kernel.h>
1724
1725 #include <sys/cons.h>
1726
1727 #include <ddb/ddb.h>
1728
1729 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
1730                                        vm_map_entry_t entry);
1731 static int      vm_object_in_map (vm_object_t object);
1732
1733 static int
1734 _vm_object_in_map(map, object, entry)
1735         vm_map_t map;
1736         vm_object_t object;
1737         vm_map_entry_t entry;
1738 {
1739         vm_map_t tmpm;
1740         vm_map_entry_t tmpe;
1741         vm_object_t obj;
1742         int entcount;
1743
1744         if (map == 0)
1745                 return 0;
1746
1747         if (entry == 0) {
1748                 tmpe = map->header.next;
1749                 entcount = map->nentries;
1750                 while (entcount-- && (tmpe != &map->header)) {
1751                         if( _vm_object_in_map(map, object, tmpe)) {
1752                                 return 1;
1753                         }
1754                         tmpe = tmpe->next;
1755                 }
1756         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1757                 tmpm = entry->object.sub_map;
1758                 tmpe = tmpm->header.next;
1759                 entcount = tmpm->nentries;
1760                 while (entcount-- && tmpe != &tmpm->header) {
1761                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1762                                 return 1;
1763                         }
1764                         tmpe = tmpe->next;
1765                 }
1766         } else if ((obj = entry->object.vm_object) != NULL) {
1767                 for(; obj; obj=obj->backing_object)
1768                         if( obj == object) {
1769                                 return 1;
1770                         }
1771         }
1772         return 0;
1773 }
1774
1775 static int
1776 vm_object_in_map( object)
1777         vm_object_t object;
1778 {
1779         struct proc *p;
1780         for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1781                 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1782                         continue;
1783                 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1784                         return 1;
1785         }
1786         if( _vm_object_in_map( kernel_map, object, 0))
1787                 return 1;
1788         if( _vm_object_in_map( pager_map, object, 0))
1789                 return 1;
1790         if( _vm_object_in_map( buffer_map, object, 0))
1791                 return 1;
1792         if( _vm_object_in_map( mb_map, object, 0))
1793                 return 1;
1794         return 0;
1795 }
1796
1797 DB_SHOW_COMMAND(vmochk, vm_object_check)
1798 {
1799         vm_object_t object;
1800
1801         /*
1802          * make sure that internal objs are in a map somewhere
1803          * and none have zero ref counts.
1804          */
1805         for (object = TAILQ_FIRST(&vm_object_list);
1806                         object != NULL;
1807                         object = TAILQ_NEXT(object, object_list)) {
1808                 if (object->handle == NULL &&
1809                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1810                         if (object->ref_count == 0) {
1811                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1812                                         (long)object->size);
1813                         }
1814                         if (!vm_object_in_map(object)) {
1815                                 db_printf(
1816                         "vmochk: internal obj is not in a map: "
1817                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1818                                     object->ref_count, (u_long)object->size,
1819                                     (u_long)object->size,
1820                                     (void *)object->backing_object);
1821                         }
1822                 }
1823         }
1824 }
1825
1826 /*
1827  *      vm_object_print:        [ debug ]
1828  */
1829 DB_SHOW_COMMAND(object, vm_object_print_static)
1830 {
1831         /* XXX convert args. */
1832         vm_object_t object = (vm_object_t)addr;
1833         boolean_t full = have_addr;
1834
1835         vm_page_t p;
1836
1837         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1838 #define count   was_count
1839
1840         int count;
1841
1842         if (object == NULL)
1843                 return;
1844
1845         db_iprintf(
1846             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1847             object, (int)object->type, (u_long)object->size,
1848             object->resident_page_count, object->ref_count, object->flags);
1849         /*
1850          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1851          */
1852         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1853             object->shadow_count,
1854             object->backing_object ? object->backing_object->ref_count : 0,
1855             object->backing_object, (long)object->backing_object_offset);
1856
1857         if (!full)
1858                 return;
1859
1860         db_indent += 2;
1861         count = 0;
1862         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1863                 if (count == 0)
1864                         db_iprintf("memory:=");
1865                 else if (count == 6) {
1866                         db_printf("\n");
1867                         db_iprintf(" ...");
1868                         count = 0;
1869                 } else
1870                         db_printf(",");
1871                 count++;
1872
1873                 db_printf("(off=0x%lx,page=0x%lx)",
1874                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1875         }
1876         if (count != 0)
1877                 db_printf("\n");
1878         db_indent -= 2;
1879 }
1880
1881 /* XXX. */
1882 #undef count
1883
1884 /* XXX need this non-static entry for calling from vm_map_print. */
1885 void
1886 vm_object_print(addr, have_addr, count, modif)
1887         /* db_expr_t */ long addr;
1888         boolean_t have_addr;
1889         /* db_expr_t */ long count;
1890         char *modif;
1891 {
1892         vm_object_print_static(addr, have_addr, count, modif);
1893 }
1894
1895 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1896 {
1897         vm_object_t object;
1898         int nl = 0;
1899         int c;
1900         for (object = TAILQ_FIRST(&vm_object_list);
1901                         object != NULL;
1902                         object = TAILQ_NEXT(object, object_list)) {
1903                 vm_pindex_t idx, fidx;
1904                 vm_pindex_t osize;
1905                 vm_paddr_t pa = -1, padiff;
1906                 int rcount;
1907                 vm_page_t m;
1908
1909                 db_printf("new object: %p\n", (void *)object);
1910                 if ( nl > 18) {
1911                         c = cngetc();
1912                         if (c != ' ')
1913                                 return;
1914                         nl = 0;
1915                 }
1916                 nl++;
1917                 rcount = 0;
1918                 fidx = 0;
1919                 osize = object->size;
1920                 if (osize > 128)
1921                         osize = 128;
1922                 for(idx=0;idx<osize;idx++) {
1923                         m = vm_page_lookup(object, idx);
1924                         if (m == NULL) {
1925                                 if (rcount) {
1926                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1927                                                 (long)fidx, rcount, (long)pa);
1928                                         if ( nl > 18) {
1929                                                 c = cngetc();
1930                                                 if (c != ' ')
1931                                                         return;
1932                                                 nl = 0;
1933                                         }
1934                                         nl++;
1935                                         rcount = 0;
1936                                 }
1937                                 continue;
1938                         }
1939
1940
1941                         if (rcount &&
1942                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1943                                 ++rcount;
1944                                 continue;
1945                         }
1946                         if (rcount) {
1947                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1948                                 padiff >>= PAGE_SHIFT;
1949                                 padiff &= PQ_L2_MASK;
1950                                 if (padiff == 0) {
1951                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1952                                         ++rcount;
1953                                         continue;
1954                                 }
1955                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1956                                         (long)fidx, rcount, (long)pa);
1957                                 db_printf("pd(%ld)\n", (long)padiff);
1958                                 if ( nl > 18) {
1959                                         c = cngetc();
1960                                         if (c != ' ')
1961                                                 return;
1962                                         nl = 0;
1963                                 }
1964                                 nl++;
1965                         }
1966                         fidx = idx;
1967                         pa = VM_PAGE_TO_PHYS(m);
1968                         rcount = 1;
1969                 }
1970                 if (rcount) {
1971                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1972                                 (long)fidx, rcount, (long)pa);
1973                         if ( nl > 18) {
1974                                 c = cngetc();
1975                                 if (c != ' ')
1976                                         return;
1977                                 nl = 0;
1978                         }
1979                         nl++;
1980                 }
1981         }
1982 }
1983 #endif /* DDB */