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  * 4. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  33  *
  34  *
  35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  36  * All rights reserved.
  37  *
  38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  39  *
  40  * Permission to use, copy, modify and distribute this software and
  41  * its documentation is hereby granted, provided that both the copyright
  42  * notice and this permission notice appear in all copies of the
  43  * software, derivative works or modified versions, and any portions
  44  * thereof, and that both notices appear in supporting documentation.
  45  *
  46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  49  *
  50  * Carnegie Mellon requests users of this software to return to
  51  *
  52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  53  *  School of Computer Science
  54  *  Carnegie Mellon University
  55  *  Pittsburgh PA 15213-3890
  56  *
  57  * any improvements or extensions that they make and grant Carnegie the
  58  * rights to redistribute these changes.
  59  */
  60
  61 /*
  62  *      Virtual memory object module.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_vm.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/lock.h>
  73 #include <sys/mman.h>
  74 #include <sys/mount.h>
  75 #include <sys/kernel.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/mutex.h>
  78 #include <sys/proc.h>           /* for curproc, pageproc */
  79 #include <sys/socket.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/vnode.h>
  82 #include <sys/vmmeter.h>
  83 #include <sys/sx.h>
  84
  85 #include <vm/vm.h>
  86 #include <vm/vm_param.h>
  87 #include <vm/pmap.h>
  88 #include <vm/vm_map.h>
  89 #include <vm/vm_object.h>
  90 #include <vm/vm_page.h>
  91 #include <vm/vm_pageout.h>
  92 #include <vm/vm_pager.h>
  93 #include <vm/swap_pager.h>
  94 #include <vm/vm_kern.h>
  95 #include <vm/vm_extern.h>
  96 #include <vm/vm_reserv.h>
  97 #include <vm/uma.h>
  98
  99 static int msync_flush_flags = 0;
 100 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0,
 101     "Does nothing; kept for backward compatibility");
 102
 103 static int old_msync;
 104 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 105     "Use old (insecure) msync behavior");
 106
 107 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 108                     int pagerflags, int flags, int *clearobjflags);
 109 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 110                     int *clearobjflags);
 111 static void     vm_object_qcollapse(vm_object_t object);
 112 static void     vm_object_vndeallocate(vm_object_t object);
 113
 114 /*
 115  *      Virtual memory objects maintain the actual data
 116  *      associated with allocated virtual memory.  A given
 117  *      page of memory exists within exactly one object.
 118  *
 119  *      An object is only deallocated when all "references"
 120  *      are given up.  Only one "reference" to a given
 121  *      region of an object should be writeable.
 122  *
 123  *      Associated with each object is a list of all resident
 124  *      memory pages belonging to that object; this list is
 125  *      maintained by the "vm_page" module, and locked by the object's
 126  *      lock.
 127  *
 128  *      Each object also records a "pager" routine which is
 129  *      used to retrieve (and store) pages to the proper backing
 130  *      storage.  In addition, objects may be backed by other
 131  *      objects from which they were virtual-copied.
 132  *
 133  *      The only items within the object structure which are
 134  *      modified after time of creation are:
 135  *              reference count         locked by object's lock
 136  *              pager routine           locked by object's lock
 137  *
 138  */
 139
 140 struct object_q vm_object_list;
 141 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 142
 143 struct vm_object kernel_object_store;
 144 struct vm_object kmem_object_store;
 145
 146 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
 147
 148 static long object_collapses;
 149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 150     &object_collapses, 0, "VM object collapses");
 151
 152 static long object_bypasses;
 153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 154     &object_bypasses, 0, "VM object bypasses");
 155
 156 static uma_zone_t obj_zone;
 157
 158 static int vm_object_zinit(void *mem, int size, int flags);
 159
 160 #ifdef INVARIANTS
 161 static void vm_object_zdtor(void *mem, int size, void *arg);
 162
 163 static void
 164 vm_object_zdtor(void *mem, int size, void *arg)
 165 {
 166         vm_object_t object;
 167
 168         object = (vm_object_t)mem;
 169         KASSERT(TAILQ_EMPTY(&object->memq),
 170             ("object %p has resident pages",
 171             object));
 172 #if VM_NRESERVLEVEL > 0
 173         KASSERT(LIST_EMPTY(&object->rvq),
 174             ("object %p has reservations",
 175             object));
 176 #endif
 177         KASSERT(object->cache == NULL,
 178             ("object %p has cached pages",
 179             object));
 180         KASSERT(object->paging_in_progress == 0,
 181             ("object %p paging_in_progress = %d",
 182             object, object->paging_in_progress));
 183         KASSERT(object->resident_page_count == 0,
 184             ("object %p resident_page_count = %d",
 185             object, object->resident_page_count));
 186         KASSERT(object->shadow_count == 0,
 187             ("object %p shadow_count = %d",
 188             object, object->shadow_count));
 189 }
 190 #endif
 191
 192 static int
 193 vm_object_zinit(void *mem, int size, int flags)
 194 {
 195         vm_object_t object;
 196
 197         object = (vm_object_t)mem;
 198         bzero(&object->mtx, sizeof(object->mtx));
 199         VM_OBJECT_LOCK_INIT(object, "standard object");
 200
 201         /* These are true for any object that has been freed */
 202         object->paging_in_progress = 0;
 203         object->resident_page_count = 0;
 204         object->shadow_count = 0;
 205         return (0);
 206 }
 207
 208 void
 209 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 210 {
 211
 212         TAILQ_INIT(&object->memq);
 213         LIST_INIT(&object->shadow_head);
 214
 215         object->root = NULL;
 216         object->type = type;
 217         object->size = size;
 218         object->generation = 1;
 219         object->ref_count = 1;
 220         object->memattr = VM_MEMATTR_DEFAULT;
 221         object->flags = 0;
 222         object->uip = NULL;
 223         object->charge = 0;
 224         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
 225                 object->flags = OBJ_ONEMAPPING;
 226         object->pg_color = 0;
 227         object->handle = NULL;
 228         object->backing_object = NULL;
 229         object->backing_object_offset = (vm_ooffset_t) 0;
 230 #if VM_NRESERVLEVEL > 0
 231         LIST_INIT(&object->rvq);
 232 #endif
 233         object->cache = NULL;
 234
 235         mtx_lock(&vm_object_list_mtx);
 236         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 237         mtx_unlock(&vm_object_list_mtx);
 238 }
 239
 240 /*
 241  *      vm_object_init:
 242  *
 243  *      Initialize the VM objects module.
 244  */
 245 void
 246 vm_object_init(void)
 247 {
 248         TAILQ_INIT(&vm_object_list);
 249         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 250
 251         VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
 252         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 253             kernel_object);
 254 #if VM_NRESERVLEVEL > 0
 255         kernel_object->flags |= OBJ_COLORED;
 256         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 257 #endif
 258
 259         VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
 260         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 261             kmem_object);
 262 #if VM_NRESERVLEVEL > 0
 263         kmem_object->flags |= OBJ_COLORED;
 264         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 265 #endif
 266
 267         /*
 268          * The lock portion of struct vm_object must be type stable due
 269          * to vm_pageout_fallback_object_lock locking a vm object
 270          * without holding any references to it.
 271          */
 272         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 273 #ifdef INVARIANTS
 274             vm_object_zdtor,
 275 #else
 276             NULL,
 277 #endif
 278             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
 279 }
 280
 281 void
 282 vm_object_clear_flag(vm_object_t object, u_short bits)
 283 {
 284
 285         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 286         object->flags &= ~bits;
 287 }
 288
 289 /*
 290  *      Sets the default memory attribute for the specified object.  Pages
 291  *      that are allocated to this object are by default assigned this memory
 292  *      attribute.
 293  *
 294  *      Presently, this function must be called before any pages are allocated
 295  *      to the object.  In the future, this requirement may be relaxed for
 296  *      "default" and "swap" objects.
 297  */
 298 int
 299 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 300 {
 301
 302         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 303         switch (object->type) {
 304         case OBJT_DEFAULT:
 305         case OBJT_DEVICE:
 306         case OBJT_PHYS:
 307         case OBJT_SG:
 308         case OBJT_SWAP:
 309         case OBJT_VNODE:
 310                 if (!TAILQ_EMPTY(&object->memq))
 311                         return (KERN_FAILURE);
 312                 break;
 313         case OBJT_DEAD:
 314                 return (KERN_INVALID_ARGUMENT);
 315         }
 316         object->memattr = memattr;
 317         return (KERN_SUCCESS);
 318 }
 319
 320 void
 321 vm_object_pip_add(vm_object_t object, short i)
 322 {
 323
 324         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 325         object->paging_in_progress += i;
 326 }
 327
 328 void
 329 vm_object_pip_subtract(vm_object_t object, short i)
 330 {
 331
 332         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 333         object->paging_in_progress -= i;
 334 }
 335
 336 void
 337 vm_object_pip_wakeup(vm_object_t object)
 338 {
 339
 340         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 341         object->paging_in_progress--;
 342         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 343                 vm_object_clear_flag(object, OBJ_PIPWNT);
 344                 wakeup(object);
 345         }
 346 }
 347
 348 void
 349 vm_object_pip_wakeupn(vm_object_t object, short i)
 350 {
 351
 352         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 353         if (i)
 354                 object->paging_in_progress -= i;
 355         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 356                 vm_object_clear_flag(object, OBJ_PIPWNT);
 357                 wakeup(object);
 358         }
 359 }
 360
 361 void
 362 vm_object_pip_wait(vm_object_t object, char *waitid)
 363 {
 364
 365         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 366         while (object->paging_in_progress) {
 367                 object->flags |= OBJ_PIPWNT;
 368                 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
 369         }
 370 }
 371
 372 /*
 373  *      vm_object_allocate:
 374  *
 375  *      Returns a new object with the given size.
 376  */
 377 vm_object_t
 378 vm_object_allocate(objtype_t type, vm_pindex_t size)
 379 {
 380         vm_object_t object;
 381
 382         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 383         _vm_object_allocate(type, size, object);
 384         return (object);
 385 }
 386
 387
 388 /*
 389  *      vm_object_reference:
 390  *
 391  *      Gets another reference to the given object.  Note: OBJ_DEAD
 392  *      objects can be referenced during final cleaning.
 393  */
 394 void
 395 vm_object_reference(vm_object_t object)
 396 {
 397         if (object == NULL)
 398                 return;
 399         VM_OBJECT_LOCK(object);
 400         vm_object_reference_locked(object);
 401         VM_OBJECT_UNLOCK(object);
 402 }
 403
 404 /*
 405  *      vm_object_reference_locked:
 406  *
 407  *      Gets another reference to the given object.
 408  *
 409  *      The object must be locked.
 410  */
 411 void
 412 vm_object_reference_locked(vm_object_t object)
 413 {
 414         struct vnode *vp;
 415
 416         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 417         object->ref_count++;
 418         if (object->type == OBJT_VNODE) {
 419                 vp = object->handle;
 420                 vref(vp);
 421         }
 422 }
 423
 424 /*
 425  * Handle deallocating an object of type OBJT_VNODE.
 426  */
 427 static void
 428 vm_object_vndeallocate(vm_object_t object)
 429 {
 430         struct vnode *vp = (struct vnode *) object->handle;
 431
 432         VFS_ASSERT_GIANT(vp->v_mount);
 433         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 434         KASSERT(object->type == OBJT_VNODE,
 435             ("vm_object_vndeallocate: not a vnode object"));
 436         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 437 #ifdef INVARIANTS
 438         if (object->ref_count == 0) {
 439                 vprint("vm_object_vndeallocate", vp);
 440                 panic("vm_object_vndeallocate: bad object reference count");
 441         }
 442 #endif
 443
 444         if (object->ref_count > 1) {
 445                 object->ref_count--;
 446                 VM_OBJECT_UNLOCK(object);
 447                 /* vrele may need the vnode lock. */
 448                 vrele(vp);
 449         } else {
 450                 vhold(vp);
 451                 VM_OBJECT_UNLOCK(object);
 452                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 453                 vdrop(vp);
 454                 VM_OBJECT_LOCK(object);
 455                 object->ref_count--;
 456                 if (object->type == OBJT_DEAD) {
 457                         VM_OBJECT_UNLOCK(object);
 458                         VOP_UNLOCK(vp, 0);
 459                 } else {
 460                         if (object->ref_count == 0)
 461                                 vp->v_vflag &= ~VV_TEXT;
 462                         VM_OBJECT_UNLOCK(object);
 463                         vput(vp);
 464                 }
 465         }
 466 }
 467
 468 /*
 469  *      vm_object_deallocate:
 470  *
 471  *      Release a reference to the specified object,
 472  *      gained either through a vm_object_allocate
 473  *      or a vm_object_reference call.  When all references
 474  *      are gone, storage associated with this object
 475  *      may be relinquished.
 476  *
 477  *      No object may be locked.
 478  */
 479 void
 480 vm_object_deallocate(vm_object_t object)
 481 {
 482         vm_object_t temp;
 483
 484         while (object != NULL) {
 485                 int vfslocked;
 486
 487                 vfslocked = 0;
 488         restart:
 489                 VM_OBJECT_LOCK(object);
 490                 if (object->type == OBJT_VNODE) {
 491                         struct vnode *vp = (struct vnode *) object->handle;
 492
 493                         /*
 494                          * Conditionally acquire Giant for a vnode-backed
 495                          * object.  We have to be careful since the type of
 496                          * a vnode object can change while the object is
 497                          * unlocked.
 498                          */
 499                         if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
 500                                 vfslocked = 1;
 501                                 if (!mtx_trylock(&Giant)) {
 502                                         VM_OBJECT_UNLOCK(object);
 503                                         mtx_lock(&Giant);
 504                                         goto restart;
 505                                 }
 506                         }
 507                         vm_object_vndeallocate(object);
 508                         VFS_UNLOCK_GIANT(vfslocked);
 509                         return;
 510                 } else
 511                         /*
 512                          * This is to handle the case that the object
 513                          * changed type while we dropped its lock to
 514                          * obtain Giant.
 515                          */
 516                         VFS_UNLOCK_GIANT(vfslocked);
 517
 518                 KASSERT(object->ref_count != 0,
 519                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 520
 521                 /*
 522                  * If the reference count goes to 0 we start calling
 523                  * vm_object_terminate() on the object chain.
 524                  * A ref count of 1 may be a special case depending on the
 525                  * shadow count being 0 or 1.
 526                  */
 527                 object->ref_count--;
 528                 if (object->ref_count > 1) {
 529                         VM_OBJECT_UNLOCK(object);
 530                         return;
 531                 } else if (object->ref_count == 1) {
 532                         if (object->shadow_count == 0 &&
 533                             object->handle == NULL &&
 534                             (object->type == OBJT_DEFAULT ||
 535                              object->type == OBJT_SWAP)) {
 536                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 537                         } else if ((object->shadow_count == 1) &&
 538                             (object->handle == NULL) &&
 539                             (object->type == OBJT_DEFAULT ||
 540                              object->type == OBJT_SWAP)) {
 541                                 vm_object_t robject;
 542
 543                                 robject = LIST_FIRST(&object->shadow_head);
 544                                 KASSERT(robject != NULL,
 545                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 546                                          object->ref_count,
 547                                          object->shadow_count));
 548                                 if (!VM_OBJECT_TRYLOCK(robject)) {
 549                                         /*
 550                                          * Avoid a potential deadlock.
 551                                          */
 552                                         object->ref_count++;
 553                                         VM_OBJECT_UNLOCK(object);
 554                                         /*
 555                                          * More likely than not the thread
 556                                          * holding robject's lock has lower
 557                                          * priority than the current thread.
 558                                          * Let the lower priority thread run.
 559                                          */
 560                                         pause("vmo_de", 1);
 561                                         continue;
 562                                 }
 563                                 /*
 564                                  * Collapse object into its shadow unless its
 565                                  * shadow is dead.  In that case, object will
 566                                  * be deallocated by the thread that is
 567                                  * deallocating its shadow.
 568                                  */
 569                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 570                                     (robject->handle == NULL) &&
 571                                     (robject->type == OBJT_DEFAULT ||
 572                                      robject->type == OBJT_SWAP)) {
 573
 574                                         robject->ref_count++;
 575 retry:
 576                                         if (robject->paging_in_progress) {
 577                                                 VM_OBJECT_UNLOCK(object);
 578                                                 vm_object_pip_wait(robject,
 579                                                     "objde1");
 580                                                 temp = robject->backing_object;
 581                                                 if (object == temp) {
 582                                                         VM_OBJECT_LOCK(object);
 583                                                         goto retry;
 584                                                 }
 585                                         } else if (object->paging_in_progress) {
 586                                                 VM_OBJECT_UNLOCK(robject);
 587                                                 object->flags |= OBJ_PIPWNT;
 588                                                 msleep(object,
 589                                                     VM_OBJECT_MTX(object),
 590                                                     PDROP | PVM, "objde2", 0);
 591                                                 VM_OBJECT_LOCK(robject);
 592                                                 temp = robject->backing_object;
 593                                                 if (object == temp) {
 594                                                         VM_OBJECT_LOCK(object);
 595                                                         goto retry;
 596                                                 }
 597                                         } else
 598                                                 VM_OBJECT_UNLOCK(object);
 599
 600                                         if (robject->ref_count == 1) {
 601                                                 robject->ref_count--;
 602                                                 object = robject;
 603                                                 goto doterm;
 604                                         }
 605                                         object = robject;
 606                                         vm_object_collapse(object);
 607                                         VM_OBJECT_UNLOCK(object);
 608                                         continue;
 609                                 }
 610                                 VM_OBJECT_UNLOCK(robject);
 611                         }
 612                         VM_OBJECT_UNLOCK(object);
 613                         return;
 614                 }
 615 doterm:
 616                 temp = object->backing_object;
 617                 if (temp != NULL) {
 618                         VM_OBJECT_LOCK(temp);
 619                         LIST_REMOVE(object, shadow_list);
 620                         temp->shadow_count--;
 621                         VM_OBJECT_UNLOCK(temp);
 622                         object->backing_object = NULL;
 623                 }
 624                 /*
 625                  * Don't double-terminate, we could be in a termination
 626                  * recursion due to the terminate having to sync data
 627                  * to disk.
 628                  */
 629                 if ((object->flags & OBJ_DEAD) == 0)
 630                         vm_object_terminate(object);
 631                 else
 632                         VM_OBJECT_UNLOCK(object);
 633                 object = temp;
 634         }
 635 }
 636
 637 /*
 638  *      vm_object_destroy removes the object from the global object list
 639  *      and frees the space for the object.
 640  */
 641 void
 642 vm_object_destroy(vm_object_t object)
 643 {
 644
 645         /*
 646          * Remove the object from the global object list.
 647          */
 648         mtx_lock(&vm_object_list_mtx);
 649         TAILQ_REMOVE(&vm_object_list, object, object_list);
 650         mtx_unlock(&vm_object_list_mtx);
 651
 652         /*
 653          * Release the allocation charge.
 654          */
 655         if (object->uip != NULL) {
 656                 KASSERT(object->type == OBJT_DEFAULT ||
 657                     object->type == OBJT_SWAP,
 658                     ("vm_object_terminate: non-swap obj %p has uip",
 659                      object));
 660                 swap_release_by_uid(object->charge, object->uip);
 661                 object->charge = 0;
 662                 uifree(object->uip);
 663                 object->uip = NULL;
 664         }
 665
 666         /*
 667          * Free the space for the object.
 668          */
 669         uma_zfree(obj_zone, object);
 670 }
 671
 672 /*
 673  *      vm_object_terminate actually destroys the specified object, freeing
 674  *      up all previously used resources.
 675  *
 676  *      The object must be locked.
 677  *      This routine may block.
 678  */
 679 void
 680 vm_object_terminate(vm_object_t object)
 681 {
 682         vm_page_t p;
 683
 684         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 685
 686         /*
 687          * Make sure no one uses us.
 688          */
 689         vm_object_set_flag(object, OBJ_DEAD);
 690
 691         /*
 692          * wait for the pageout daemon to be done with the object
 693          */
 694         vm_object_pip_wait(object, "objtrm");
 695
 696         KASSERT(!object->paging_in_progress,
 697                 ("vm_object_terminate: pageout in progress"));
 698
 699         /*
 700          * Clean and free the pages, as appropriate. All references to the
 701          * object are gone, so we don't need to lock it.
 702          */
 703         if (object->type == OBJT_VNODE) {
 704                 struct vnode *vp = (struct vnode *)object->handle;
 705
 706                 /*
 707                  * Clean pages and flush buffers.
 708                  */
 709                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 710                 VM_OBJECT_UNLOCK(object);
 711
 712                 vinvalbuf(vp, V_SAVE, 0, 0);
 713
 714                 VM_OBJECT_LOCK(object);
 715         }
 716
 717         KASSERT(object->ref_count == 0,
 718                 ("vm_object_terminate: object with references, ref_count=%d",
 719                 object->ref_count));
 720
 721         /*
 722          * Now free any remaining pages. For internal objects, this also
 723          * removes them from paging queues. Don't free wired pages, just
 724          * remove them from the object.
 725          */
 726         vm_page_lock_queues();
 727         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
 728                 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
 729                         ("vm_object_terminate: freeing busy page %p "
 730                         "p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
 731                 if (p->wire_count == 0) {
 732                         vm_page_free(p);
 733                         cnt.v_pfree++;
 734                 } else {
 735                         vm_page_remove(p);
 736                 }
 737         }
 738         vm_page_unlock_queues();
 739
 740 #if VM_NRESERVLEVEL > 0
 741         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 742                 vm_reserv_break_all(object);
 743 #endif
 744         if (__predict_false(object->cache != NULL))
 745                 vm_page_cache_free(object, 0, 0);
 746
 747         /*
 748          * Let the pager know object is dead.
 749          */
 750         vm_pager_deallocate(object);
 751         VM_OBJECT_UNLOCK(object);
 752
 753         vm_object_destroy(object);
 754 }
 755
 756 static boolean_t
 757 vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags)
 758 {
 759
 760         /*
 761          * If we have been asked to skip nosync pages and this is a
 762          * nosync page, skip it.  Note that the object flags were not
 763          * cleared in this case so we do not have to set them.
 764          */
 765         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 766                 *clearobjflags = 0;
 767                 return (FALSE);
 768         } else {
 769                 pmap_remove_write(p);
 770                 return (p->dirty != 0);
 771         }
 772 }
 773
 774 /*
 775  *      vm_object_page_clean
 776  *
 777  *      Clean all dirty pages in the specified range of object.  Leaves page
 778  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 779  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 780  *      leaving the object dirty.
 781  *
 782  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 783  *      synchronous clustering mode implementation.
 784  *
 785  *      Odd semantics: if start == end, we clean everything.
 786  *
 787  *      The object must be locked.
 788  */
 789 void
 790 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
 791     int flags)
 792 {
 793         vm_page_t np, p;
 794         vm_pindex_t pi, tend;
 795         int clearobjflags, curgeneration, n, pagerflags;
 796
 797         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 798         KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
 799         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 800             object->resident_page_count == 0)
 801                 return;
 802
 803         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 804             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 805         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 806
 807         tend = (end == 0) ? object->size : end;
 808
 809         vm_object_set_flag(object, OBJ_CLEANING);
 810
 811         vm_page_lock_queues();
 812
 813         /*
 814          * Make the page read-only so we can then clear the object flags.
 815          *
 816          * However, if this is a nosync mmap then the object is likely to
 817          * stay dirty so do not mess with the page and do not clear the
 818          * object flags.
 819          */
 820         clearobjflags = 1;
 821
 822 rescan:
 823         curgeneration = object->generation;
 824
 825         for (p = vm_page_find_least(object, start); p != NULL; p = np) {
 826                 pi = p->pindex;
 827                 if (pi >= tend)
 828                         break;
 829                 np = TAILQ_NEXT(p, listq);
 830                 if (p->valid == 0)
 831                         continue;
 832                 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
 833                         vm_page_lock_queues();
 834                         if (object->generation != curgeneration)
 835                                 goto rescan;
 836                         np = vm_page_find_least(object, pi);
 837                         continue;
 838                 }
 839                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 840                         continue;
 841
 842                 n = vm_object_page_collect_flush(object, p, pagerflags,
 843                     flags, &clearobjflags);
 844                 if (object->generation != curgeneration)
 845                         goto rescan;
 846                 np = vm_page_find_least(object, pi + n);
 847         }
 848         vm_page_unlock_queues();
 849 #if 0
 850         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 851 #endif
 852
 853         vm_object_clear_flag(object, OBJ_CLEANING);
 854         if (clearobjflags && start == 0 && tend == object->size)
 855                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 856 }
 857
 858 static int
 859 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 860     int flags, int *clearobjflags)
 861 {
 862         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 863         int count, i, mreq, runlen;
 864
 865         mtx_assert(&vm_page_queue_mtx, MA_OWNED);
 866
 867         count = 1;
 868         mreq = 0;
 869
 870         for (tp = p; count < vm_pageout_page_count; count++) {
 871                 tp = vm_page_next(tp);
 872                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 873                         break;
 874                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 875                         break;
 876         }
 877
 878         for (p_first = p; count < vm_pageout_page_count; count++) {
 879                 tp = vm_page_prev(p_first);
 880                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 881                         break;
 882                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 883                         break;
 884                 p_first = tp;
 885                 mreq++;
 886         }
 887
 888         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 889                 ma[i] = tp;
 890
 891         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
 892         return (runlen);
 893 }
 894
 895 /*
 896  * Note that there is absolutely no sense in writing out
 897  * anonymous objects, so we track down the vnode object
 898  * to write out.
 899  * We invalidate (remove) all pages from the address space
 900  * for semantic correctness.
 901  *
 902  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 903  * may start out with a NULL object.
 904  */
 905 void
 906 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 907     boolean_t syncio, boolean_t invalidate)
 908 {
 909         vm_object_t backing_object;
 910         struct vnode *vp;
 911         struct mount *mp;
 912         int flags;
 913
 914         if (object == NULL)
 915                 return;
 916         VM_OBJECT_LOCK(object);
 917         while ((backing_object = object->backing_object) != NULL) {
 918                 VM_OBJECT_LOCK(backing_object);
 919                 offset += object->backing_object_offset;
 920                 VM_OBJECT_UNLOCK(object);
 921                 object = backing_object;
 922                 if (object->size < OFF_TO_IDX(offset + size))
 923                         size = IDX_TO_OFF(object->size) - offset;
 924         }
 925         /*
 926          * Flush pages if writing is allowed, invalidate them
 927          * if invalidation requested.  Pages undergoing I/O
 928          * will be ignored by vm_object_page_remove().
 929          *
 930          * We cannot lock the vnode and then wait for paging
 931          * to complete without deadlocking against vm_fault.
 932          * Instead we simply call vm_object_page_remove() and
 933          * allow it to block internally on a page-by-page
 934          * basis when it encounters pages undergoing async
 935          * I/O.
 936          */
 937         if (object->type == OBJT_VNODE &&
 938             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 939                 int vfslocked;
 940                 vp = object->handle;
 941                 VM_OBJECT_UNLOCK(object);
 942                 (void) vn_start_write(vp, &mp, V_WAIT);
 943                 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 944                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 945                 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
 946                 flags |= invalidate ? OBJPC_INVAL : 0;
 947                 VM_OBJECT_LOCK(object);
 948                 vm_object_page_clean(object,
 949                     OFF_TO_IDX(offset),
 950                     OFF_TO_IDX(offset + size + PAGE_MASK),
 951                     flags);
 952                 VM_OBJECT_UNLOCK(object);
 953                 VOP_UNLOCK(vp, 0);
 954                 VFS_UNLOCK_GIANT(vfslocked);
 955                 vn_finished_write(mp);
 956                 VM_OBJECT_LOCK(object);
 957         }
 958         if ((object->type == OBJT_VNODE ||
 959              object->type == OBJT_DEVICE) && invalidate) {
 960                 boolean_t purge;
 961                 purge = old_msync || (object->type == OBJT_DEVICE);
 962                 vm_object_page_remove(object,
 963                     OFF_TO_IDX(offset),
 964                     OFF_TO_IDX(offset + size + PAGE_MASK),
 965                     purge ? FALSE : TRUE);
 966         }
 967         VM_OBJECT_UNLOCK(object);
 968 }
 969
 970 /*
 971  *      vm_object_madvise:
 972  *
 973  *      Implements the madvise function at the object/page level.
 974  *
 975  *      MADV_WILLNEED   (any object)
 976  *
 977  *          Activate the specified pages if they are resident.
 978  *
 979  *      MADV_DONTNEED   (any object)
 980  *
 981  *          Deactivate the specified pages if they are resident.
 982  *
 983  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
 984  *                       OBJ_ONEMAPPING only)
 985  *
 986  *          Deactivate and clean the specified pages if they are
 987  *          resident.  This permits the process to reuse the pages
 988  *          without faulting or the kernel to reclaim the pages
 989  *          without I/O.
 990  */
 991 void
 992 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
 993 {
 994         vm_pindex_t end, tpindex;
 995         vm_object_t backing_object, tobject;
 996         vm_page_t m;
 997
 998         if (object == NULL)
 999                 return;
1000         VM_OBJECT_LOCK(object);
1001         end = pindex + count;
1002         /*
1003          * Locate and adjust resident pages
1004          */
1005         for (; pindex < end; pindex += 1) {
1006 relookup:
1007                 tobject = object;
1008                 tpindex = pindex;
1009 shadowlookup:
1010                 /*
1011                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1012                  * and those pages must be OBJ_ONEMAPPING.
1013                  */
1014                 if (advise == MADV_FREE) {
1015                         if ((tobject->type != OBJT_DEFAULT &&
1016                              tobject->type != OBJT_SWAP) ||
1017                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1018                                 goto unlock_tobject;
1019                         }
1020                 } else if (tobject->type == OBJT_PHYS)
1021                         goto unlock_tobject;
1022                 m = vm_page_lookup(tobject, tpindex);
1023                 if (m == NULL && advise == MADV_WILLNEED) {
1024                         /*
1025                          * If the page is cached, reactivate it.
1026                          */
1027                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1028                             VM_ALLOC_NOBUSY);
1029                 }
1030                 if (m == NULL) {
1031                         /*
1032                          * There may be swap even if there is no backing page
1033                          */
1034                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1035                                 swap_pager_freespace(tobject, tpindex, 1);
1036                         /*
1037                          * next object
1038                          */
1039                         backing_object = tobject->backing_object;
1040                         if (backing_object == NULL)
1041                                 goto unlock_tobject;
1042                         VM_OBJECT_LOCK(backing_object);
1043                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1044                         if (tobject != object)
1045                                 VM_OBJECT_UNLOCK(tobject);
1046                         tobject = backing_object;
1047                         goto shadowlookup;
1048                 } else if (m->valid != VM_PAGE_BITS_ALL)
1049                         goto unlock_tobject;
1050                 /*
1051                  * If the page is not in a normal state, skip it.
1052                  */
1053                 vm_page_lock_queues();
1054                 if (m->hold_count != 0 || m->wire_count != 0) {
1055                         vm_page_unlock_queues();
1056                         goto unlock_tobject;
1057                 }
1058                 if ((m->oflags & VPO_BUSY) || m->busy) {
1059                         if (advise == MADV_WILLNEED)
1060                                 /*
1061                                  * Reference the page before unlocking and
1062                                  * sleeping so that the page daemon is less
1063                                  * likely to reclaim it.
1064                                  */
1065                                 vm_page_flag_set(m, PG_REFERENCED);
1066                         vm_page_unlock_queues();
1067                         if (object != tobject)
1068                                 VM_OBJECT_UNLOCK(object);
1069                         m->oflags |= VPO_WANTED;
1070                         msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1071                             0);
1072                         VM_OBJECT_LOCK(object);
1073                         goto relookup;
1074                 }
1075                 if (advise == MADV_WILLNEED) {
1076                         vm_page_activate(m);
1077                 } else if (advise == MADV_DONTNEED) {
1078                         vm_page_dontneed(m);
1079                 } else if (advise == MADV_FREE) {
1080                         /*
1081                          * Mark the page clean.  This will allow the page
1082                          * to be freed up by the system.  However, such pages
1083                          * are often reused quickly by malloc()/free()
1084                          * so we do not do anything that would cause
1085                          * a page fault if we can help it.
1086                          *
1087                          * Specifically, we do not try to actually free
1088                          * the page now nor do we try to put it in the
1089                          * cache (which would cause a page fault on reuse).
1090                          *
1091                          * But we do make the page is freeable as we
1092                          * can without actually taking the step of unmapping
1093                          * it.
1094                          */
1095                         pmap_clear_modify(m);
1096                         m->dirty = 0;
1097                         m->act_count = 0;
1098                         vm_page_dontneed(m);
1099                 }
1100                 vm_page_unlock_queues();
1101                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1102                         swap_pager_freespace(tobject, tpindex, 1);
1103 unlock_tobject:
1104                 if (tobject != object)
1105                         VM_OBJECT_UNLOCK(tobject);
1106         }
1107         VM_OBJECT_UNLOCK(object);
1108 }
1109
1110 /*
1111  *      vm_object_shadow:
1112  *
1113  *      Create a new object which is backed by the
1114  *      specified existing object range.  The source
1115  *      object reference is deallocated.
1116  *
1117  *      The new object and offset into that object
1118  *      are returned in the source parameters.
1119  */
1120 void
1121 vm_object_shadow(
1122         vm_object_t *object,    /* IN/OUT */
1123         vm_ooffset_t *offset,   /* IN/OUT */
1124         vm_size_t length)
1125 {
1126         vm_object_t source;
1127         vm_object_t result;
1128
1129         source = *object;
1130
1131         /*
1132          * Don't create the new object if the old object isn't shared.
1133          */
1134         if (source != NULL) {
1135                 VM_OBJECT_LOCK(source);
1136                 if (source->ref_count == 1 &&
1137                     source->handle == NULL &&
1138                     (source->type == OBJT_DEFAULT ||
1139                      source->type == OBJT_SWAP)) {
1140                         VM_OBJECT_UNLOCK(source);
1141                         return;
1142                 }
1143                 VM_OBJECT_UNLOCK(source);
1144         }
1145
1146         /*
1147          * Allocate a new object with the given length.
1148          */
1149         result = vm_object_allocate(OBJT_DEFAULT, length);
1150
1151         /*
1152          * The new object shadows the source object, adding a reference to it.
1153          * Our caller changes his reference to point to the new object,
1154          * removing a reference to the source object.  Net result: no change
1155          * of reference count.
1156          *
1157          * Try to optimize the result object's page color when shadowing
1158          * in order to maintain page coloring consistency in the combined
1159          * shadowed object.
1160          */
1161         result->backing_object = source;
1162         /*
1163          * Store the offset into the source object, and fix up the offset into
1164          * the new object.
1165          */
1166         result->backing_object_offset = *offset;
1167         if (source != NULL) {
1168                 VM_OBJECT_LOCK(source);
1169                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1170                 source->shadow_count++;
1171 #if VM_NRESERVLEVEL > 0
1172                 result->flags |= source->flags & OBJ_COLORED;
1173                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1174                     ((1 << (VM_NFREEORDER - 1)) - 1);
1175 #endif
1176                 VM_OBJECT_UNLOCK(source);
1177         }
1178
1179
1180         /*
1181          * Return the new things
1182          */
1183         *offset = 0;
1184         *object = result;
1185 }
1186
1187 /*
1188  *      vm_object_split:
1189  *
1190  * Split the pages in a map entry into a new object.  This affords
1191  * easier removal of unused pages, and keeps object inheritance from
1192  * being a negative impact on memory usage.
1193  */
1194 void
1195 vm_object_split(vm_map_entry_t entry)
1196 {
1197         vm_page_t m, m_next;
1198         vm_object_t orig_object, new_object, source;
1199         vm_pindex_t idx, offidxstart;
1200         vm_size_t size;
1201
1202         orig_object = entry->object.vm_object;
1203         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1204                 return;
1205         if (orig_object->ref_count <= 1)
1206                 return;
1207         VM_OBJECT_UNLOCK(orig_object);
1208
1209         offidxstart = OFF_TO_IDX(entry->offset);
1210         size = atop(entry->end - entry->start);
1211
1212         /*
1213          * If swap_pager_copy() is later called, it will convert new_object
1214          * into a swap object.
1215          */
1216         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1217
1218         /*
1219          * At this point, the new object is still private, so the order in
1220          * which the original and new objects are locked does not matter.
1221          */
1222         VM_OBJECT_LOCK(new_object);
1223         VM_OBJECT_LOCK(orig_object);
1224         source = orig_object->backing_object;
1225         if (source != NULL) {
1226                 VM_OBJECT_LOCK(source);
1227                 if ((source->flags & OBJ_DEAD) != 0) {
1228                         VM_OBJECT_UNLOCK(source);
1229                         VM_OBJECT_UNLOCK(orig_object);
1230                         VM_OBJECT_UNLOCK(new_object);
1231                         vm_object_deallocate(new_object);
1232                         VM_OBJECT_LOCK(orig_object);
1233                         return;
1234                 }
1235                 LIST_INSERT_HEAD(&source->shadow_head,
1236                                   new_object, shadow_list);
1237                 source->shadow_count++;
1238                 vm_object_reference_locked(source);     /* for new_object */
1239                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1240                 VM_OBJECT_UNLOCK(source);
1241                 new_object->backing_object_offset =
1242                         orig_object->backing_object_offset + entry->offset;
1243                 new_object->backing_object = source;
1244         }
1245         if (orig_object->uip != NULL) {
1246                 new_object->uip = orig_object->uip;
1247                 uihold(orig_object->uip);
1248                 new_object->charge = ptoa(size);
1249                 KASSERT(orig_object->charge >= ptoa(size),
1250                     ("orig_object->charge < 0"));
1251                 orig_object->charge -= ptoa(size);
1252         }
1253 retry:
1254         m = vm_page_find_least(orig_object, offidxstart);
1255         vm_page_lock_queues();
1256         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1257             m = m_next) {
1258                 m_next = TAILQ_NEXT(m, listq);
1259
1260                 /*
1261                  * We must wait for pending I/O to complete before we can
1262                  * rename the page.
1263                  *
1264                  * We do not have to VM_PROT_NONE the page as mappings should
1265                  * not be changed by this operation.
1266                  */
1267                 if ((m->oflags & VPO_BUSY) || m->busy) {
1268                         vm_page_unlock_queues();
1269                         VM_OBJECT_UNLOCK(new_object);
1270                         m->oflags |= VPO_WANTED;
1271                         msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1272                         VM_OBJECT_LOCK(new_object);
1273                         goto retry;
1274                 }
1275                 vm_page_rename(m, new_object, idx);
1276                 /* page automatically made dirty by rename and cache handled */
1277                 vm_page_busy(m);
1278         }
1279         vm_page_unlock_queues();
1280         if (orig_object->type == OBJT_SWAP) {
1281                 /*
1282                  * swap_pager_copy() can sleep, in which case the orig_object's
1283                  * and new_object's locks are released and reacquired.
1284                  */
1285                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1286
1287                 /*
1288                  * Transfer any cached pages from orig_object to new_object.
1289                  */
1290                 if (__predict_false(orig_object->cache != NULL))
1291                         vm_page_cache_transfer(orig_object, offidxstart,
1292                             new_object);
1293         }
1294         VM_OBJECT_UNLOCK(orig_object);
1295         TAILQ_FOREACH(m, &new_object->memq, listq)
1296                 vm_page_wakeup(m);
1297         VM_OBJECT_UNLOCK(new_object);
1298         entry->object.vm_object = new_object;
1299         entry->offset = 0LL;
1300         vm_object_deallocate(orig_object);
1301         VM_OBJECT_LOCK(new_object);
1302 }
1303
1304 #define OBSC_TEST_ALL_SHADOWED  0x0001
1305 #define OBSC_COLLAPSE_NOWAIT    0x0002
1306 #define OBSC_COLLAPSE_WAIT      0x0004
1307
1308 static int
1309 vm_object_backing_scan(vm_object_t object, int op)
1310 {
1311         int r = 1;
1312         vm_page_t p;
1313         vm_object_t backing_object;
1314         vm_pindex_t backing_offset_index;
1315
1316         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1317         VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1318
1319         backing_object = object->backing_object;
1320         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1321
1322         /*
1323          * Initial conditions
1324          */
1325         if (op & OBSC_TEST_ALL_SHADOWED) {
1326                 /*
1327                  * We do not want to have to test for the existence of cache
1328                  * or swap pages in the backing object.  XXX but with the
1329                  * new swapper this would be pretty easy to do.
1330                  *
1331                  * XXX what about anonymous MAP_SHARED memory that hasn't
1332                  * been ZFOD faulted yet?  If we do not test for this, the
1333                  * shadow test may succeed! XXX
1334                  */
1335                 if (backing_object->type != OBJT_DEFAULT) {
1336                         return (0);
1337                 }
1338         }
1339         if (op & OBSC_COLLAPSE_WAIT) {
1340                 vm_object_set_flag(backing_object, OBJ_DEAD);
1341         }
1342
1343         /*
1344          * Our scan
1345          */
1346         p = TAILQ_FIRST(&backing_object->memq);
1347         while (p) {
1348                 vm_page_t next = TAILQ_NEXT(p, listq);
1349                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1350
1351                 if (op & OBSC_TEST_ALL_SHADOWED) {
1352                         vm_page_t pp;
1353
1354                         /*
1355                          * Ignore pages outside the parent object's range
1356                          * and outside the parent object's mapping of the
1357                          * backing object.
1358                          *
1359                          * note that we do not busy the backing object's
1360                          * page.
1361                          */
1362                         if (
1363                             p->pindex < backing_offset_index ||
1364                             new_pindex >= object->size
1365                         ) {
1366                                 p = next;
1367                                 continue;
1368                         }
1369
1370                         /*
1371                          * See if the parent has the page or if the parent's
1372                          * object pager has the page.  If the parent has the
1373                          * page but the page is not valid, the parent's
1374                          * object pager must have the page.
1375                          *
1376                          * If this fails, the parent does not completely shadow
1377                          * the object and we might as well give up now.
1378                          */
1379
1380                         pp = vm_page_lookup(object, new_pindex);
1381                         if (
1382                             (pp == NULL || pp->valid == 0) &&
1383                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1384                         ) {
1385                                 r = 0;
1386                                 break;
1387                         }
1388                 }
1389
1390                 /*
1391                  * Check for busy page
1392                  */
1393                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1394                         vm_page_t pp;
1395
1396                         if (op & OBSC_COLLAPSE_NOWAIT) {
1397                                 if ((p->oflags & VPO_BUSY) ||
1398                                     !p->valid ||
1399                                     p->busy) {
1400                                         p = next;
1401                                         continue;
1402                                 }
1403                         } else if (op & OBSC_COLLAPSE_WAIT) {
1404                                 if ((p->oflags & VPO_BUSY) || p->busy) {
1405                                         VM_OBJECT_UNLOCK(object);
1406                                         p->oflags |= VPO_WANTED;
1407                                         msleep(p, VM_OBJECT_MTX(backing_object),
1408                                             PDROP | PVM, "vmocol", 0);
1409                                         VM_OBJECT_LOCK(object);
1410                                         VM_OBJECT_LOCK(backing_object);
1411                                         /*
1412                                          * If we slept, anything could have
1413                                          * happened.  Since the object is
1414                                          * marked dead, the backing offset
1415                                          * should not have changed so we
1416                                          * just restart our scan.
1417                                          */
1418                                         p = TAILQ_FIRST(&backing_object->memq);
1419                                         continue;
1420                                 }
1421                         }
1422
1423                         KASSERT(
1424                             p->object == backing_object,
1425                             ("vm_object_backing_scan: object mismatch")
1426                         );
1427
1428                         /*
1429                          * Destroy any associated swap
1430                          */
1431                         if (backing_object->type == OBJT_SWAP) {
1432                                 swap_pager_freespace(
1433                                     backing_object,
1434                                     p->pindex,
1435                                     1
1436                                 );
1437                         }
1438
1439                         if (
1440                             p->pindex < backing_offset_index ||
1441                             new_pindex >= object->size
1442                         ) {
1443                                 /*
1444                                  * Page is out of the parent object's range, we
1445                                  * can simply destroy it.
1446                                  */
1447                                 vm_page_lock_queues();
1448                                 KASSERT(!pmap_page_is_mapped(p),
1449                                     ("freeing mapped page %p", p));
1450                                 if (p->wire_count == 0)
1451                                         vm_page_free(p);
1452                                 else
1453                                         vm_page_remove(p);
1454                                 vm_page_unlock_queues();
1455                                 p = next;
1456                                 continue;
1457                         }
1458
1459                         pp = vm_page_lookup(object, new_pindex);
1460                         if (
1461                             pp != NULL ||
1462                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1463                         ) {
1464                                 /*
1465                                  * page already exists in parent OR swap exists
1466                                  * for this location in the parent.  Destroy
1467                                  * the original page from the backing object.
1468                                  *
1469                                  * Leave the parent's page alone
1470                                  */
1471                                 vm_page_lock_queues();
1472                                 KASSERT(!pmap_page_is_mapped(p),
1473                                     ("freeing mapped page %p", p));
1474                                 if (p->wire_count == 0)
1475                                         vm_page_free(p);
1476                                 else
1477                                         vm_page_remove(p);
1478                                 vm_page_unlock_queues();
1479                                 p = next;
1480                                 continue;
1481                         }
1482
1483 #if VM_NRESERVLEVEL > 0
1484                         /*
1485                          * Rename the reservation.
1486                          */
1487                         vm_reserv_rename(p, object, backing_object,
1488                             backing_offset_index);
1489 #endif
1490
1491                         /*
1492                          * Page does not exist in parent, rename the
1493                          * page from the backing object to the main object.
1494                          *
1495                          * If the page was mapped to a process, it can remain
1496                          * mapped through the rename.
1497                          */
1498                         vm_page_lock_queues();
1499                         vm_page_rename(p, object, new_pindex);
1500                         vm_page_unlock_queues();
1501                         /* page automatically made dirty by rename */
1502                 }
1503                 p = next;
1504         }
1505         return (r);
1506 }
1507
1508
1509 /*
1510  * this version of collapse allows the operation to occur earlier and
1511  * when paging_in_progress is true for an object...  This is not a complete
1512  * operation, but should plug 99.9% of the rest of the leaks.
1513  */
1514 static void
1515 vm_object_qcollapse(vm_object_t object)
1516 {
1517         vm_object_t backing_object = object->backing_object;
1518
1519         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1520         VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1521
1522         if (backing_object->ref_count != 1)
1523                 return;
1524
1525         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1526 }
1527
1528 /*
1529  *      vm_object_collapse:
1530  *
1531  *      Collapse an object with the object backing it.
1532  *      Pages in the backing object are moved into the
1533  *      parent, and the backing object is deallocated.
1534  */
1535 void
1536 vm_object_collapse(vm_object_t object)
1537 {
1538         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1539
1540         while (TRUE) {
1541                 vm_object_t backing_object;
1542
1543                 /*
1544                  * Verify that the conditions are right for collapse:
1545                  *
1546                  * The object exists and the backing object exists.
1547                  */
1548                 if ((backing_object = object->backing_object) == NULL)
1549                         break;
1550
1551                 /*
1552                  * we check the backing object first, because it is most likely
1553                  * not collapsable.
1554                  */
1555                 VM_OBJECT_LOCK(backing_object);
1556                 if (backing_object->handle != NULL ||
1557                     (backing_object->type != OBJT_DEFAULT &&
1558                      backing_object->type != OBJT_SWAP) ||
1559                     (backing_object->flags & OBJ_DEAD) ||
1560                     object->handle != NULL ||
1561                     (object->type != OBJT_DEFAULT &&
1562                      object->type != OBJT_SWAP) ||
1563                     (object->flags & OBJ_DEAD)) {
1564                         VM_OBJECT_UNLOCK(backing_object);
1565                         break;
1566                 }
1567
1568                 if (
1569                     object->paging_in_progress != 0 ||
1570                     backing_object->paging_in_progress != 0
1571                 ) {
1572                         vm_object_qcollapse(object);
1573                         VM_OBJECT_UNLOCK(backing_object);
1574                         break;
1575                 }
1576                 /*
1577                  * We know that we can either collapse the backing object (if
1578                  * the parent is the only reference to it) or (perhaps) have
1579                  * the parent bypass the object if the parent happens to shadow
1580                  * all the resident pages in the entire backing object.
1581                  *
1582                  * This is ignoring pager-backed pages such as swap pages.
1583                  * vm_object_backing_scan fails the shadowing test in this
1584                  * case.
1585                  */
1586                 if (backing_object->ref_count == 1) {
1587                         /*
1588                          * If there is exactly one reference to the backing
1589                          * object, we can collapse it into the parent.
1590                          */
1591                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1592
1593 #if VM_NRESERVLEVEL > 0
1594                         /*
1595                          * Break any reservations from backing_object.
1596                          */
1597                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1598                                 vm_reserv_break_all(backing_object);
1599 #endif
1600
1601                         /*
1602                          * Move the pager from backing_object to object.
1603                          */
1604                         if (backing_object->type == OBJT_SWAP) {
1605                                 /*
1606                                  * swap_pager_copy() can sleep, in which case
1607                                  * the backing_object's and object's locks are
1608                                  * released and reacquired.
1609                                  */
1610                                 swap_pager_copy(
1611                                     backing_object,
1612                                     object,
1613                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1614
1615                                 /*
1616                                  * Free any cached pages from backing_object.
1617                                  */
1618                                 if (__predict_false(backing_object->cache != NULL))
1619                                         vm_page_cache_free(backing_object, 0, 0);
1620                         }
1621                         /*
1622                          * Object now shadows whatever backing_object did.
1623                          * Note that the reference to
1624                          * backing_object->backing_object moves from within
1625                          * backing_object to within object.
1626                          */
1627                         LIST_REMOVE(object, shadow_list);
1628                         backing_object->shadow_count--;
1629                         if (backing_object->backing_object) {
1630                                 VM_OBJECT_LOCK(backing_object->backing_object);
1631                                 LIST_REMOVE(backing_object, shadow_list);
1632                                 LIST_INSERT_HEAD(
1633                                     &backing_object->backing_object->shadow_head,
1634                                     object, shadow_list);
1635                                 /*
1636                                  * The shadow_count has not changed.
1637                                  */
1638                                 VM_OBJECT_UNLOCK(backing_object->backing_object);
1639                         }
1640                         object->backing_object = backing_object->backing_object;
1641                         object->backing_object_offset +=
1642                             backing_object->backing_object_offset;
1643
1644                         /*
1645                          * Discard backing_object.
1646                          *
1647                          * Since the backing object has no pages, no pager left,
1648                          * and no object references within it, all that is
1649                          * necessary is to dispose of it.
1650                          */
1651                         KASSERT(backing_object->ref_count == 1, (
1652 "backing_object %p was somehow re-referenced during collapse!",
1653                             backing_object));
1654                         VM_OBJECT_UNLOCK(backing_object);
1655                         vm_object_destroy(backing_object);
1656
1657                         object_collapses++;
1658                 } else {
1659                         vm_object_t new_backing_object;
1660
1661                         /*
1662                          * If we do not entirely shadow the backing object,
1663                          * there is nothing we can do so we give up.
1664                          */
1665                         if (object->resident_page_count != object->size &&
1666                             vm_object_backing_scan(object,
1667                             OBSC_TEST_ALL_SHADOWED) == 0) {
1668                                 VM_OBJECT_UNLOCK(backing_object);
1669                                 break;
1670                         }
1671
1672                         /*
1673                          * Make the parent shadow the next object in the
1674                          * chain.  Deallocating backing_object will not remove
1675                          * it, since its reference count is at least 2.
1676                          */
1677                         LIST_REMOVE(object, shadow_list);
1678                         backing_object->shadow_count--;
1679
1680                         new_backing_object = backing_object->backing_object;
1681                         if ((object->backing_object = new_backing_object) != NULL) {
1682                                 VM_OBJECT_LOCK(new_backing_object);
1683                                 LIST_INSERT_HEAD(
1684                                     &new_backing_object->shadow_head,
1685                                     object,
1686                                     shadow_list
1687                                 );
1688                                 new_backing_object->shadow_count++;
1689                                 vm_object_reference_locked(new_backing_object);
1690                                 VM_OBJECT_UNLOCK(new_backing_object);
1691                                 object->backing_object_offset +=
1692                                         backing_object->backing_object_offset;
1693                         }
1694
1695                         /*
1696                          * Drop the reference count on backing_object. Since
1697                          * its ref_count was at least 2, it will not vanish.
1698                          */
1699                         backing_object->ref_count--;
1700                         VM_OBJECT_UNLOCK(backing_object);
1701                         object_bypasses++;
1702                 }
1703
1704                 /*
1705                  * Try again with this object's new backing object.
1706                  */
1707         }
1708 }
1709
1710 /*
1711  *      vm_object_page_remove:
1712  *
1713  *      For the given object, either frees or invalidates each of the
1714  *      specified pages.  In general, a page is freed.  However, if a
1715  *      page is wired for any reason other than the existence of a
1716  *      managed, wired mapping, then it may be invalidated but not
1717  *      removed from the object.  Pages are specified by the given
1718  *      range ["start", "end") and Boolean "clean_only".  As a
1719  *      special case, if "end" is zero, then the range extends from
1720  *      "start" to the end of the object.  If "clean_only" is TRUE,
1721  *      then only the non-dirty pages within the specified range are
1722  *      affected.
1723  *
1724  *      In general, this operation should only be performed on objects
1725  *      that contain managed pages.  There are two exceptions.  First,
1726  *      it may be performed on the kernel and kmem objects.  Second,
1727  *      it may be used by msync(..., MS_INVALIDATE) to invalidate
1728  *      device-backed pages.  In both of these cases, "clean_only"
1729  *      must be FALSE.
1730  *
1731  *      The object must be locked.
1732  */
1733 void
1734 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1735     boolean_t clean_only)
1736 {
1737         vm_page_t p, next;
1738         int wirings;
1739
1740         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1741         if (object->resident_page_count == 0)
1742                 goto skipmemq;
1743
1744         /*
1745          * Since physically-backed objects do not use managed pages, we can't
1746          * remove pages from the object (we must instead remove the page
1747          * references, and then destroy the object).
1748          */
1749         KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1750             object == kmem_object,
1751             ("attempt to remove pages from a physical object"));
1752
1753         vm_object_pip_add(object, 1);
1754 again:
1755         p = vm_page_find_least(object, start);
1756         vm_page_lock_queues();
1757         /*
1758          * Assert: the variable p is either (1) the page with the
1759          * least pindex greater than or equal to the parameter pindex
1760          * or (2) NULL.
1761          */
1762         for (;
1763              p != NULL && (p->pindex < end || end == 0);
1764              p = next) {
1765                 next = TAILQ_NEXT(p, listq);
1766
1767                 /*
1768                  * If the page is wired for any reason besides the
1769                  * existence of managed, wired mappings, then it cannot
1770                  * be freed.  For example, fictitious pages, which
1771                  * represent device memory, are inherently wired and
1772                  * cannot be freed.  They can, however, be invalidated
1773                  * if "clean_only" is FALSE.
1774                  */
1775                 if ((wirings = p->wire_count) != 0 &&
1776                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1777                         /* Fictitious pages do not have managed mappings. */
1778                         if ((p->flags & PG_FICTITIOUS) == 0)
1779                                 pmap_remove_all(p);
1780                         /* Account for removal of managed, wired mappings. */
1781                         p->wire_count -= wirings;
1782                         if (!clean_only) {
1783                                 p->valid = 0;
1784                                 vm_page_undirty(p);
1785                         }
1786                         continue;
1787                 }
1788                 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1789                         goto again;
1790                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1791                     ("vm_object_page_remove: page %p is fictitious", p));
1792                 if (clean_only && p->valid) {
1793                         pmap_remove_write(p);
1794                         if (p->dirty)
1795                                 continue;
1796                 }
1797                 pmap_remove_all(p);
1798                 /* Account for removal of managed, wired mappings. */
1799                 if (wirings != 0)
1800                         p->wire_count -= wirings;
1801                 vm_page_free(p);
1802         }
1803         vm_page_unlock_queues();
1804         vm_object_pip_wakeup(object);
1805 skipmemq:
1806         if (__predict_false(object->cache != NULL))
1807                 vm_page_cache_free(object, start, end);
1808 }
1809
1810 /*
1811  *      Populate the specified range of the object with valid pages.  Returns
1812  *      TRUE if the range is successfully populated and FALSE otherwise.
1813  *
1814  *      Note: This function should be optimized to pass a larger array of
1815  *      pages to vm_pager_get_pages() before it is applied to a non-
1816  *      OBJT_DEVICE object.
1817  *
1818  *      The object must be locked.
1819  */
1820 boolean_t
1821 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1822 {
1823         vm_page_t m, ma[1];
1824         vm_pindex_t pindex;
1825         int rv;
1826
1827         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1828         for (pindex = start; pindex < end; pindex++) {
1829                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1830                     VM_ALLOC_RETRY);
1831                 if (m->valid != VM_PAGE_BITS_ALL) {
1832                         ma[0] = m;
1833                         rv = vm_pager_get_pages(object, ma, 1, 0);
1834                         m = vm_page_lookup(object, pindex);
1835                         if (m == NULL)
1836                                 break;
1837                         if (rv != VM_PAGER_OK) {
1838                                 vm_page_lock_queues();
1839                                 vm_page_free(m);
1840                                 vm_page_unlock_queues();
1841                                 break;
1842                         }
1843                 }
1844                 /*
1845                  * Keep "m" busy because a subsequent iteration may unlock
1846                  * the object.
1847                  */
1848         }
1849         if (pindex > start) {
1850                 m = vm_page_lookup(object, start);
1851                 while (m != NULL && m->pindex < pindex) {
1852                         vm_page_wakeup(m);
1853                         m = TAILQ_NEXT(m, listq);
1854                 }
1855         }
1856         return (pindex == end);
1857 }
1858
1859 /*
1860  *      Routine:        vm_object_coalesce
1861  *      Function:       Coalesces two objects backing up adjoining
1862  *                      regions of memory into a single object.
1863  *
1864  *      returns TRUE if objects were combined.
1865  *
1866  *      NOTE:   Only works at the moment if the second object is NULL -
1867  *              if it's not, which object do we lock first?
1868  *
1869  *      Parameters:
1870  *              prev_object     First object to coalesce
1871  *              prev_offset     Offset into prev_object
1872  *              prev_size       Size of reference to prev_object
1873  *              next_size       Size of reference to the second object
1874  *              reserved        Indicator that extension region has
1875  *                              swap accounted for
1876  *
1877  *      Conditions:
1878  *      The object must *not* be locked.
1879  */
1880 boolean_t
1881 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1882     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1883 {
1884         vm_pindex_t next_pindex;
1885
1886         if (prev_object == NULL)
1887                 return (TRUE);
1888         VM_OBJECT_LOCK(prev_object);
1889         if (prev_object->type != OBJT_DEFAULT &&
1890             prev_object->type != OBJT_SWAP) {
1891                 VM_OBJECT_UNLOCK(prev_object);
1892                 return (FALSE);
1893         }
1894
1895         /*
1896          * Try to collapse the object first
1897          */
1898         vm_object_collapse(prev_object);
1899
1900         /*
1901          * Can't coalesce if: . more than one reference . paged out . shadows
1902          * another object . has a copy elsewhere (any of which mean that the
1903          * pages not mapped to prev_entry may be in use anyway)
1904          */
1905         if (prev_object->backing_object != NULL) {
1906                 VM_OBJECT_UNLOCK(prev_object);
1907                 return (FALSE);
1908         }
1909
1910         prev_size >>= PAGE_SHIFT;
1911         next_size >>= PAGE_SHIFT;
1912         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1913
1914         if ((prev_object->ref_count > 1) &&
1915             (prev_object->size != next_pindex)) {
1916                 VM_OBJECT_UNLOCK(prev_object);
1917                 return (FALSE);
1918         }
1919
1920         /*
1921          * Account for the charge.
1922          */
1923         if (prev_object->uip != NULL) {
1924
1925                 /*
1926                  * If prev_object was charged, then this mapping,
1927                  * althought not charged now, may become writable
1928                  * later. Non-NULL uip in the object would prevent
1929                  * swap reservation during enabling of the write
1930                  * access, so reserve swap now. Failed reservation
1931                  * cause allocation of the separate object for the map
1932                  * entry, and swap reservation for this entry is
1933                  * managed in appropriate time.
1934                  */
1935                 if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
1936                     prev_object->uip)) {
1937                         return (FALSE);
1938                 }
1939                 prev_object->charge += ptoa(next_size);
1940         }
1941
1942         /*
1943          * Remove any pages that may still be in the object from a previous
1944          * deallocation.
1945          */
1946         if (next_pindex < prev_object->size) {
1947                 vm_object_page_remove(prev_object,
1948                                       next_pindex,
1949                                       next_pindex + next_size, FALSE);
1950                 if (prev_object->type == OBJT_SWAP)
1951                         swap_pager_freespace(prev_object,
1952                                              next_pindex, next_size);
1953 #if 0
1954                 if (prev_object->uip != NULL) {
1955                         KASSERT(prev_object->charge >=
1956                             ptoa(prev_object->size - next_pindex),
1957                             ("object %p overcharged 1 %jx %jx", prev_object,
1958                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
1959                         prev_object->charge -= ptoa(prev_object->size -
1960                             next_pindex);
1961                 }
1962 #endif
1963         }
1964
1965         /*
1966          * Extend the object if necessary.
1967          */
1968         if (next_pindex + next_size > prev_object->size)
1969                 prev_object->size = next_pindex + next_size;
1970
1971         VM_OBJECT_UNLOCK(prev_object);
1972         return (TRUE);
1973 }
1974
1975 void
1976 vm_object_set_writeable_dirty(vm_object_t object)
1977 {
1978
1979         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1980         if (object->type != OBJT_VNODE)
1981                 return;
1982         object->generation++;
1983         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
1984                 return;
1985         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
1986 }
1987
1988 #include "opt_ddb.h"
1989 #ifdef DDB
1990 #include <sys/kernel.h>
1991
1992 #include <sys/cons.h>
1993
1994 #include <ddb/ddb.h>
1995
1996 static int
1997 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1998 {
1999         vm_map_t tmpm;
2000         vm_map_entry_t tmpe;
2001         vm_object_t obj;
2002         int entcount;
2003
2004         if (map == 0)
2005                 return 0;
2006
2007         if (entry == 0) {
2008                 tmpe = map->header.next;
2009                 entcount = map->nentries;
2010                 while (entcount-- && (tmpe != &map->header)) {
2011                         if (_vm_object_in_map(map, object, tmpe)) {
2012                                 return 1;
2013                         }
2014                         tmpe = tmpe->next;
2015                 }
2016         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2017                 tmpm = entry->object.sub_map;
2018                 tmpe = tmpm->header.next;
2019                 entcount = tmpm->nentries;
2020                 while (entcount-- && tmpe != &tmpm->header) {
2021                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2022                                 return 1;
2023                         }
2024                         tmpe = tmpe->next;
2025                 }
2026         } else if ((obj = entry->object.vm_object) != NULL) {
2027                 for (; obj; obj = obj->backing_object)
2028                         if (obj == object) {
2029                                 return 1;
2030                         }
2031         }
2032         return 0;
2033 }
2034
2035 static int
2036 vm_object_in_map(vm_object_t object)
2037 {
2038         struct proc *p;
2039
2040         /* sx_slock(&allproc_lock); */
2041         FOREACH_PROC_IN_SYSTEM(p) {
2042                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2043                         continue;
2044                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2045                         /* sx_sunlock(&allproc_lock); */
2046                         return 1;
2047                 }
2048         }
2049         /* sx_sunlock(&allproc_lock); */
2050         if (_vm_object_in_map(kernel_map, object, 0))
2051                 return 1;
2052         if (_vm_object_in_map(kmem_map, object, 0))
2053                 return 1;
2054         if (_vm_object_in_map(pager_map, object, 0))
2055                 return 1;
2056         if (_vm_object_in_map(buffer_map, object, 0))
2057                 return 1;
2058         return 0;
2059 }
2060
2061 DB_SHOW_COMMAND(vmochk, vm_object_check)
2062 {
2063         vm_object_t object;
2064
2065         /*
2066          * make sure that internal objs are in a map somewhere
2067          * and none have zero ref counts.
2068          */
2069         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2070                 if (object->handle == NULL &&
2071                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2072                         if (object->ref_count == 0) {
2073                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2074                                         (long)object->size);
2075                         }
2076                         if (!vm_object_in_map(object)) {
2077                                 db_printf(
2078                         "vmochk: internal obj is not in a map: "
2079                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2080                                     object->ref_count, (u_long)object->size,
2081                                     (u_long)object->size,
2082                                     (void *)object->backing_object);
2083                         }
2084                 }
2085         }
2086 }
2087
2088 /*
2089  *      vm_object_print:        [ debug ]
2090  */
2091 DB_SHOW_COMMAND(object, vm_object_print_static)
2092 {
2093         /* XXX convert args. */
2094         vm_object_t object = (vm_object_t)addr;
2095         boolean_t full = have_addr;
2096
2097         vm_page_t p;
2098
2099         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2100 #define count   was_count
2101
2102         int count;
2103
2104         if (object == NULL)
2105                 return;
2106
2107         db_iprintf(
2108             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2109             object, (int)object->type, (uintmax_t)object->size,
2110             object->resident_page_count, object->ref_count, object->flags,
2111             object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2112         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2113             object->shadow_count,
2114             object->backing_object ? object->backing_object->ref_count : 0,
2115             object->backing_object, (uintmax_t)object->backing_object_offset);
2116
2117         if (!full)
2118                 return;
2119
2120         db_indent += 2;
2121         count = 0;
2122         TAILQ_FOREACH(p, &object->memq, listq) {
2123                 if (count == 0)
2124                         db_iprintf("memory:=");
2125                 else if (count == 6) {
2126                         db_printf("\n");
2127                         db_iprintf(" ...");
2128                         count = 0;
2129                 } else
2130                         db_printf(",");
2131                 count++;
2132
2133                 db_printf("(off=0x%jx,page=0x%jx)",
2134                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2135         }
2136         if (count != 0)
2137                 db_printf("\n");
2138         db_indent -= 2;
2139 }
2140
2141 /* XXX. */
2142 #undef count
2143
2144 /* XXX need this non-static entry for calling from vm_map_print. */
2145 void
2146 vm_object_print(
2147         /* db_expr_t */ long addr,
2148         boolean_t have_addr,
2149         /* db_expr_t */ long count,
2150         char *modif)
2151 {
2152         vm_object_print_static(addr, have_addr, count, modif);
2153 }
2154
2155 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2156 {
2157         vm_object_t object;
2158         vm_pindex_t fidx;
2159         vm_paddr_t pa;
2160         vm_page_t m, prev_m;
2161         int rcount, nl, c;
2162
2163         nl = 0;
2164         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2165                 db_printf("new object: %p\n", (void *)object);
2166                 if (nl > 18) {
2167                         c = cngetc();
2168                         if (c != ' ')
2169                                 return;
2170                         nl = 0;
2171                 }
2172                 nl++;
2173                 rcount = 0;
2174                 fidx = 0;
2175                 pa = -1;
2176                 TAILQ_FOREACH(m, &object->memq, listq) {
2177                         if (m->pindex > 128)
2178                                 break;
2179                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2180                             prev_m->pindex + 1 != m->pindex) {
2181                                 if (rcount) {
2182                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2183                                                 (long)fidx, rcount, (long)pa);
2184                                         if (nl > 18) {
2185                                                 c = cngetc();
2186                                                 if (c != ' ')
2187                                                         return;
2188                                                 nl = 0;
2189                                         }
2190                                         nl++;
2191                                         rcount = 0;
2192                                 }
2193                         }
2194                         if (rcount &&
2195                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2196                                 ++rcount;
2197                                 continue;
2198                         }
2199                         if (rcount) {
2200                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2201                                         (long)fidx, rcount, (long)pa);
2202                                 if (nl > 18) {
2203                                         c = cngetc();
2204                                         if (c != ' ')
2205                                                 return;
2206                                         nl = 0;
2207                                 }
2208                                 nl++;
2209                         }
2210                         fidx = m->pindex;
2211                         pa = VM_PAGE_TO_PHYS(m);
2212                         rcount = 1;
2213                 }
2214                 if (rcount) {
2215                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2216                                 (long)fidx, rcount, (long)pa);
2217                         if (nl > 18) {
2218                                 c = cngetc();
2219                                 if (c != ' ')
2220                                         return;
2221                                 nl = 0;
2222                         }
2223                         nl++;
2224                 }
2225         }
2226 }
2227 #endif /* DDB */