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 old_msync;
 100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 101     "Use old (insecure) msync behavior");
 102
 103 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 104                     int pagerflags, int flags, int *clearobjflags);
 105 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 106                     int *clearobjflags);
 107 static void     vm_object_qcollapse(vm_object_t object);
 108 static void     vm_object_vndeallocate(vm_object_t object);
 109
 110 /*
 111  *      Virtual memory objects maintain the actual data
 112  *      associated with allocated virtual memory.  A given
 113  *      page of memory exists within exactly one object.
 114  *
 115  *      An object is only deallocated when all "references"
 116  *      are given up.  Only one "reference" to a given
 117  *      region of an object should be writeable.
 118  *
 119  *      Associated with each object is a list of all resident
 120  *      memory pages belonging to that object; this list is
 121  *      maintained by the "vm_page" module, and locked by the object's
 122  *      lock.
 123  *
 124  *      Each object also records a "pager" routine which is
 125  *      used to retrieve (and store) pages to the proper backing
 126  *      storage.  In addition, objects may be backed by other
 127  *      objects from which they were virtual-copied.
 128  *
 129  *      The only items within the object structure which are
 130  *      modified after time of creation are:
 131  *              reference count         locked by object's lock
 132  *              pager routine           locked by object's lock
 133  *
 134  */
 135
 136 struct object_q vm_object_list;
 137 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 138
 139 struct vm_object kernel_object_store;
 140 struct vm_object kmem_object_store;
 141
 142 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
 143
 144 static long object_collapses;
 145 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 146     &object_collapses, 0, "VM object collapses");
 147
 148 static long object_bypasses;
 149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 150     &object_bypasses, 0, "VM object bypasses");
 151
 152 static uma_zone_t obj_zone;
 153
 154 static int vm_object_zinit(void *mem, int size, int flags);
 155
 156 #ifdef INVARIANTS
 157 static void vm_object_zdtor(void *mem, int size, void *arg);
 158
 159 static void
 160 vm_object_zdtor(void *mem, int size, void *arg)
 161 {
 162         vm_object_t object;
 163
 164         object = (vm_object_t)mem;
 165         KASSERT(TAILQ_EMPTY(&object->memq),
 166             ("object %p has resident pages",
 167             object));
 168 #if VM_NRESERVLEVEL > 0
 169         KASSERT(LIST_EMPTY(&object->rvq),
 170             ("object %p has reservations",
 171             object));
 172 #endif
 173         KASSERT(object->cache == NULL,
 174             ("object %p has cached pages",
 175             object));
 176         KASSERT(object->paging_in_progress == 0,
 177             ("object %p paging_in_progress = %d",
 178             object, object->paging_in_progress));
 179         KASSERT(object->resident_page_count == 0,
 180             ("object %p resident_page_count = %d",
 181             object, object->resident_page_count));
 182         KASSERT(object->shadow_count == 0,
 183             ("object %p shadow_count = %d",
 184             object, object->shadow_count));
 185 }
 186 #endif
 187
 188 static int
 189 vm_object_zinit(void *mem, int size, int flags)
 190 {
 191         vm_object_t object;
 192
 193         object = (vm_object_t)mem;
 194         bzero(&object->mtx, sizeof(object->mtx));
 195         VM_OBJECT_LOCK_INIT(object, "standard object");
 196
 197         /* These are true for any object that has been freed */
 198         object->paging_in_progress = 0;
 199         object->resident_page_count = 0;
 200         object->shadow_count = 0;
 201         return (0);
 202 }
 203
 204 void
 205 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 206 {
 207
 208         TAILQ_INIT(&object->memq);
 209         LIST_INIT(&object->shadow_head);
 210
 211         object->root = NULL;
 212         object->type = type;
 213         object->size = size;
 214         object->generation = 1;
 215         object->ref_count = 1;
 216         object->memattr = VM_MEMATTR_DEFAULT;
 217         object->flags = 0;
 218         object->cred = NULL;
 219         object->charge = 0;
 220         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
 221                 object->flags = OBJ_ONEMAPPING;
 222         object->pg_color = 0;
 223         object->handle = NULL;
 224         object->backing_object = NULL;
 225         object->backing_object_offset = (vm_ooffset_t) 0;
 226 #if VM_NRESERVLEVEL > 0
 227         LIST_INIT(&object->rvq);
 228 #endif
 229         object->cache = NULL;
 230
 231         mtx_lock(&vm_object_list_mtx);
 232         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 233         mtx_unlock(&vm_object_list_mtx);
 234 }
 235
 236 /*
 237  *      vm_object_init:
 238  *
 239  *      Initialize the VM objects module.
 240  */
 241 void
 242 vm_object_init(void)
 243 {
 244         TAILQ_INIT(&vm_object_list);
 245         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 246
 247         VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
 248         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 249             kernel_object);
 250 #if VM_NRESERVLEVEL > 0
 251         kernel_object->flags |= OBJ_COLORED;
 252         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 253 #endif
 254
 255         VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
 256         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 257             kmem_object);
 258 #if VM_NRESERVLEVEL > 0
 259         kmem_object->flags |= OBJ_COLORED;
 260         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 261 #endif
 262
 263         /*
 264          * The lock portion of struct vm_object must be type stable due
 265          * to vm_pageout_fallback_object_lock locking a vm object
 266          * without holding any references to it.
 267          */
 268         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 269 #ifdef INVARIANTS
 270             vm_object_zdtor,
 271 #else
 272             NULL,
 273 #endif
 274             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
 275 }
 276
 277 void
 278 vm_object_clear_flag(vm_object_t object, u_short bits)
 279 {
 280
 281         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 282         object->flags &= ~bits;
 283 }
 284
 285 /*
 286  *      Sets the default memory attribute for the specified object.  Pages
 287  *      that are allocated to this object are by default assigned this memory
 288  *      attribute.
 289  *
 290  *      Presently, this function must be called before any pages are allocated
 291  *      to the object.  In the future, this requirement may be relaxed for
 292  *      "default" and "swap" objects.
 293  */
 294 int
 295 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 296 {
 297
 298         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 299         switch (object->type) {
 300         case OBJT_DEFAULT:
 301         case OBJT_DEVICE:
 302         case OBJT_PHYS:
 303         case OBJT_SG:
 304         case OBJT_SWAP:
 305         case OBJT_VNODE:
 306                 if (!TAILQ_EMPTY(&object->memq))
 307                         return (KERN_FAILURE);
 308                 break;
 309         case OBJT_DEAD:
 310                 return (KERN_INVALID_ARGUMENT);
 311         }
 312         object->memattr = memattr;
 313         return (KERN_SUCCESS);
 314 }
 315
 316 void
 317 vm_object_pip_add(vm_object_t object, short i)
 318 {
 319
 320         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 321         object->paging_in_progress += i;
 322 }
 323
 324 void
 325 vm_object_pip_subtract(vm_object_t object, short i)
 326 {
 327
 328         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 329         object->paging_in_progress -= i;
 330 }
 331
 332 void
 333 vm_object_pip_wakeup(vm_object_t object)
 334 {
 335
 336         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 337         object->paging_in_progress--;
 338         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 339                 vm_object_clear_flag(object, OBJ_PIPWNT);
 340                 wakeup(object);
 341         }
 342 }
 343
 344 void
 345 vm_object_pip_wakeupn(vm_object_t object, short i)
 346 {
 347
 348         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 349         if (i)
 350                 object->paging_in_progress -= i;
 351         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 352                 vm_object_clear_flag(object, OBJ_PIPWNT);
 353                 wakeup(object);
 354         }
 355 }
 356
 357 void
 358 vm_object_pip_wait(vm_object_t object, char *waitid)
 359 {
 360
 361         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 362         while (object->paging_in_progress) {
 363                 object->flags |= OBJ_PIPWNT;
 364                 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
 365         }
 366 }
 367
 368 /*
 369  *      vm_object_allocate:
 370  *
 371  *      Returns a new object with the given size.
 372  */
 373 vm_object_t
 374 vm_object_allocate(objtype_t type, vm_pindex_t size)
 375 {
 376         vm_object_t object;
 377
 378         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 379         _vm_object_allocate(type, size, object);
 380         return (object);
 381 }
 382
 383
 384 /*
 385  *      vm_object_reference:
 386  *
 387  *      Gets another reference to the given object.  Note: OBJ_DEAD
 388  *      objects can be referenced during final cleaning.
 389  */
 390 void
 391 vm_object_reference(vm_object_t object)
 392 {
 393         if (object == NULL)
 394                 return;
 395         VM_OBJECT_LOCK(object);
 396         vm_object_reference_locked(object);
 397         VM_OBJECT_UNLOCK(object);
 398 }
 399
 400 /*
 401  *      vm_object_reference_locked:
 402  *
 403  *      Gets another reference to the given object.
 404  *
 405  *      The object must be locked.
 406  */
 407 void
 408 vm_object_reference_locked(vm_object_t object)
 409 {
 410         struct vnode *vp;
 411
 412         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 413         object->ref_count++;
 414         if (object->type == OBJT_VNODE) {
 415                 vp = object->handle;
 416                 vref(vp);
 417         }
 418 }
 419
 420 /*
 421  * Handle deallocating an object of type OBJT_VNODE.
 422  */
 423 static void
 424 vm_object_vndeallocate(vm_object_t object)
 425 {
 426         struct vnode *vp = (struct vnode *) object->handle;
 427
 428         VFS_ASSERT_GIANT(vp->v_mount);
 429         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 430         KASSERT(object->type == OBJT_VNODE,
 431             ("vm_object_vndeallocate: not a vnode object"));
 432         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 433 #ifdef INVARIANTS
 434         if (object->ref_count == 0) {
 435                 vprint("vm_object_vndeallocate", vp);
 436                 panic("vm_object_vndeallocate: bad object reference count");
 437         }
 438 #endif
 439
 440         if (object->ref_count > 1) {
 441                 object->ref_count--;
 442                 VM_OBJECT_UNLOCK(object);
 443                 /* vrele may need the vnode lock. */
 444                 vrele(vp);
 445         } else {
 446                 VM_OBJECT_UNLOCK(object);
 447                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 448                 VM_OBJECT_LOCK(object);
 449                 object->ref_count--;
 450                 if (object->ref_count == 0)
 451                         vp->v_vflag &= ~VV_TEXT;
 452                 VM_OBJECT_UNLOCK(object);
 453                 vput(vp);
 454         }
 455 }
 456
 457 /*
 458  *      vm_object_deallocate:
 459  *
 460  *      Release a reference to the specified object,
 461  *      gained either through a vm_object_allocate
 462  *      or a vm_object_reference call.  When all references
 463  *      are gone, storage associated with this object
 464  *      may be relinquished.
 465  *
 466  *      No object may be locked.
 467  */
 468 void
 469 vm_object_deallocate(vm_object_t object)
 470 {
 471         vm_object_t temp;
 472
 473         while (object != NULL) {
 474                 int vfslocked;
 475
 476                 vfslocked = 0;
 477         restart:
 478                 VM_OBJECT_LOCK(object);
 479                 if (object->type == OBJT_VNODE) {
 480                         struct vnode *vp = (struct vnode *) object->handle;
 481
 482                         /*
 483                          * Conditionally acquire Giant for a vnode-backed
 484                          * object.  We have to be careful since the type of
 485                          * a vnode object can change while the object is
 486                          * unlocked.
 487                          */
 488                         if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
 489                                 vfslocked = 1;
 490                                 if (!mtx_trylock(&Giant)) {
 491                                         VM_OBJECT_UNLOCK(object);
 492                                         mtx_lock(&Giant);
 493                                         goto restart;
 494                                 }
 495                         }
 496                         vm_object_vndeallocate(object);
 497                         VFS_UNLOCK_GIANT(vfslocked);
 498                         return;
 499                 } else
 500                         /*
 501                          * This is to handle the case that the object
 502                          * changed type while we dropped its lock to
 503                          * obtain Giant.
 504                          */
 505                         VFS_UNLOCK_GIANT(vfslocked);
 506
 507                 KASSERT(object->ref_count != 0,
 508                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 509
 510                 /*
 511                  * If the reference count goes to 0 we start calling
 512                  * vm_object_terminate() on the object chain.
 513                  * A ref count of 1 may be a special case depending on the
 514                  * shadow count being 0 or 1.
 515                  */
 516                 object->ref_count--;
 517                 if (object->ref_count > 1) {
 518                         VM_OBJECT_UNLOCK(object);
 519                         return;
 520                 } else if (object->ref_count == 1) {
 521                         if (object->shadow_count == 0 &&
 522                             object->handle == NULL &&
 523                             (object->type == OBJT_DEFAULT ||
 524                              object->type == OBJT_SWAP)) {
 525                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 526                         } else if ((object->shadow_count == 1) &&
 527                             (object->handle == NULL) &&
 528                             (object->type == OBJT_DEFAULT ||
 529                              object->type == OBJT_SWAP)) {
 530                                 vm_object_t robject;
 531
 532                                 robject = LIST_FIRST(&object->shadow_head);
 533                                 KASSERT(robject != NULL,
 534                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 535                                          object->ref_count,
 536                                          object->shadow_count));
 537                                 if (!VM_OBJECT_TRYLOCK(robject)) {
 538                                         /*
 539                                          * Avoid a potential deadlock.
 540                                          */
 541                                         object->ref_count++;
 542                                         VM_OBJECT_UNLOCK(object);
 543                                         /*
 544                                          * More likely than not the thread
 545                                          * holding robject's lock has lower
 546                                          * priority than the current thread.
 547                                          * Let the lower priority thread run.
 548                                          */
 549                                         pause("vmo_de", 1);
 550                                         continue;
 551                                 }
 552                                 /*
 553                                  * Collapse object into its shadow unless its
 554                                  * shadow is dead.  In that case, object will
 555                                  * be deallocated by the thread that is
 556                                  * deallocating its shadow.
 557                                  */
 558                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 559                                     (robject->handle == NULL) &&
 560                                     (robject->type == OBJT_DEFAULT ||
 561                                      robject->type == OBJT_SWAP)) {
 562
 563                                         robject->ref_count++;
 564 retry:
 565                                         if (robject->paging_in_progress) {
 566                                                 VM_OBJECT_UNLOCK(object);
 567                                                 vm_object_pip_wait(robject,
 568                                                     "objde1");
 569                                                 temp = robject->backing_object;
 570                                                 if (object == temp) {
 571                                                         VM_OBJECT_LOCK(object);
 572                                                         goto retry;
 573                                                 }
 574                                         } else if (object->paging_in_progress) {
 575                                                 VM_OBJECT_UNLOCK(robject);
 576                                                 object->flags |= OBJ_PIPWNT;
 577                                                 msleep(object,
 578                                                     VM_OBJECT_MTX(object),
 579                                                     PDROP | PVM, "objde2", 0);
 580                                                 VM_OBJECT_LOCK(robject);
 581                                                 temp = robject->backing_object;
 582                                                 if (object == temp) {
 583                                                         VM_OBJECT_LOCK(object);
 584                                                         goto retry;
 585                                                 }
 586                                         } else
 587                                                 VM_OBJECT_UNLOCK(object);
 588
 589                                         if (robject->ref_count == 1) {
 590                                                 robject->ref_count--;
 591                                                 object = robject;
 592                                                 goto doterm;
 593                                         }
 594                                         object = robject;
 595                                         vm_object_collapse(object);
 596                                         VM_OBJECT_UNLOCK(object);
 597                                         continue;
 598                                 }
 599                                 VM_OBJECT_UNLOCK(robject);
 600                         }
 601                         VM_OBJECT_UNLOCK(object);
 602                         return;
 603                 }
 604 doterm:
 605                 temp = object->backing_object;
 606                 if (temp != NULL) {
 607                         VM_OBJECT_LOCK(temp);
 608                         LIST_REMOVE(object, shadow_list);
 609                         temp->shadow_count--;
 610                         VM_OBJECT_UNLOCK(temp);
 611                         object->backing_object = NULL;
 612                 }
 613                 /*
 614                  * Don't double-terminate, we could be in a termination
 615                  * recursion due to the terminate having to sync data
 616                  * to disk.
 617                  */
 618                 if ((object->flags & OBJ_DEAD) == 0)
 619                         vm_object_terminate(object);
 620                 else
 621                         VM_OBJECT_UNLOCK(object);
 622                 object = temp;
 623         }
 624 }
 625
 626 /*
 627  *      vm_object_destroy removes the object from the global object list
 628  *      and frees the space for the object.
 629  */
 630 void
 631 vm_object_destroy(vm_object_t object)
 632 {
 633
 634         /*
 635          * Remove the object from the global object list.
 636          */
 637         mtx_lock(&vm_object_list_mtx);
 638         TAILQ_REMOVE(&vm_object_list, object, object_list);
 639         mtx_unlock(&vm_object_list_mtx);
 640
 641         /*
 642          * Release the allocation charge.
 643          */
 644         if (object->cred != NULL) {
 645                 KASSERT(object->type == OBJT_DEFAULT ||
 646                     object->type == OBJT_SWAP,
 647                     ("vm_object_terminate: non-swap obj %p has cred",
 648                      object));
 649                 swap_release_by_cred(object->charge, object->cred);
 650                 object->charge = 0;
 651                 crfree(object->cred);
 652                 object->cred = NULL;
 653         }
 654
 655         /*
 656          * Free the space for the object.
 657          */
 658         uma_zfree(obj_zone, object);
 659 }
 660
 661 /*
 662  *      vm_object_terminate actually destroys the specified object, freeing
 663  *      up all previously used resources.
 664  *
 665  *      The object must be locked.
 666  *      This routine may block.
 667  */
 668 void
 669 vm_object_terminate(vm_object_t object)
 670 {
 671         vm_page_t p, p_next;
 672
 673         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 674
 675         /*
 676          * Make sure no one uses us.
 677          */
 678         vm_object_set_flag(object, OBJ_DEAD);
 679
 680         /*
 681          * wait for the pageout daemon to be done with the object
 682          */
 683         vm_object_pip_wait(object, "objtrm");
 684
 685         KASSERT(!object->paging_in_progress,
 686                 ("vm_object_terminate: pageout in progress"));
 687
 688         /*
 689          * Clean and free the pages, as appropriate. All references to the
 690          * object are gone, so we don't need to lock it.
 691          */
 692         if (object->type == OBJT_VNODE) {
 693                 struct vnode *vp = (struct vnode *)object->handle;
 694
 695                 /*
 696                  * Clean pages and flush buffers.
 697                  */
 698                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 699                 VM_OBJECT_UNLOCK(object);
 700
 701                 vinvalbuf(vp, V_SAVE, 0, 0);
 702
 703                 VM_OBJECT_LOCK(object);
 704         }
 705
 706         KASSERT(object->ref_count == 0,
 707                 ("vm_object_terminate: object with references, ref_count=%d",
 708                 object->ref_count));
 709
 710         /*
 711          * Free any remaining pageable pages.  This also removes them from the
 712          * paging queues.  However, don't free wired pages, just remove them
 713          * from the object.  Rather than incrementally removing each page from
 714          * the object, the page and object are reset to any empty state.
 715          */
 716         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
 717                 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
 718                     ("vm_object_terminate: freeing busy page %p", p));
 719                 vm_page_lock(p);
 720                 /*
 721                  * Optimize the page's removal from the object by resetting
 722                  * its "object" field.  Specifically, if the page is not
 723                  * wired, then the effect of this assignment is that
 724                  * vm_page_free()'s call to vm_page_remove() will return
 725                  * immediately without modifying the page or the object.
 726                  */
 727                 p->object = NULL;
 728                 if (p->wire_count == 0) {
 729                         vm_page_free(p);
 730                         PCPU_INC(cnt.v_pfree);
 731                 }
 732                 vm_page_unlock(p);
 733         }
 734         /*
 735          * If the object contained any pages, then reset it to an empty state.
 736          * None of the object's fields, including "resident_page_count", were
 737          * modified by the preceding loop.
 738          */
 739         if (object->resident_page_count != 0) {
 740                 object->root = NULL;
 741                 TAILQ_INIT(&object->memq);
 742                 object->resident_page_count = 0;
 743                 if (object->type == OBJT_VNODE)
 744                         vdrop(object->handle);
 745         }
 746
 747 #if VM_NRESERVLEVEL > 0
 748         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 749                 vm_reserv_break_all(object);
 750 #endif
 751         if (__predict_false(object->cache != NULL))
 752                 vm_page_cache_free(object, 0, 0);
 753
 754         /*
 755          * Let the pager know object is dead.
 756          */
 757         vm_pager_deallocate(object);
 758         VM_OBJECT_UNLOCK(object);
 759
 760         vm_object_destroy(object);
 761 }
 762
 763 /*
 764  * Make the page read-only so that we can clear the object flags.  However, if
 765  * this is a nosync mmap then the object is likely to stay dirty so do not
 766  * mess with the page and do not clear the object flags.  Returns TRUE if the
 767  * page should be flushed, and FALSE otherwise.
 768  */
 769 static boolean_t
 770 vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags)
 771 {
 772
 773         /*
 774          * If we have been asked to skip nosync pages and this is a
 775          * nosync page, skip it.  Note that the object flags were not
 776          * cleared in this case so we do not have to set them.
 777          */
 778         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 779                 *clearobjflags = 0;
 780                 return (FALSE);
 781         } else {
 782                 pmap_remove_write(p);
 783                 return (p->dirty != 0);
 784         }
 785 }
 786
 787 /*
 788  *      vm_object_page_clean
 789  *
 790  *      Clean all dirty pages in the specified range of object.  Leaves page
 791  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 792  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 793  *      leaving the object dirty.
 794  *
 795  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 796  *      synchronous clustering mode implementation.
 797  *
 798  *      Odd semantics: if start == end, we clean everything.
 799  *
 800  *      The object must be locked.
 801  */
 802 void
 803 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 804     int flags)
 805 {
 806         vm_page_t np, p;
 807         vm_pindex_t pi, tend, tstart;
 808         int clearobjflags, curgeneration, n, pagerflags;
 809
 810         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 811         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 812         KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
 813         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 814             object->resident_page_count == 0)
 815                 return;
 816
 817         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 818             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 819         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 820
 821         tstart = OFF_TO_IDX(start);
 822         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 823         clearobjflags = tstart == 0 && tend >= object->size;
 824
 825 rescan:
 826         curgeneration = object->generation;
 827
 828         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 829                 pi = p->pindex;
 830                 if (pi >= tend)
 831                         break;
 832                 np = TAILQ_NEXT(p, listq);
 833                 if (p->valid == 0)
 834                         continue;
 835                 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
 836                         if (object->generation != curgeneration)
 837                                 goto rescan;
 838                         np = vm_page_find_least(object, pi);
 839                         continue;
 840                 }
 841                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 842                         continue;
 843
 844                 n = vm_object_page_collect_flush(object, p, pagerflags,
 845                     flags, &clearobjflags);
 846                 if (object->generation != curgeneration)
 847                         goto rescan;
 848                 np = vm_page_find_least(object, pi + n);
 849         }
 850 #if 0
 851         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 852 #endif
 853
 854         if (clearobjflags)
 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_NOTOWNED);
 866         vm_page_lock_assert(p, MA_NOTOWNED);
 867         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 868
 869         count = 1;
 870         mreq = 0;
 871
 872         for (tp = p; count < vm_pageout_page_count; count++) {
 873                 tp = vm_page_next(tp);
 874                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 875                         break;
 876                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 877                         break;
 878         }
 879
 880         for (p_first = p; count < vm_pageout_page_count; count++) {
 881                 tp = vm_page_prev(p_first);
 882                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 883                         break;
 884                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 885                         break;
 886                 p_first = tp;
 887                 mreq++;
 888         }
 889
 890         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 891                 ma[i] = tp;
 892
 893         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
 894         return (runlen);
 895 }
 896
 897 /*
 898  * Note that there is absolutely no sense in writing out
 899  * anonymous objects, so we track down the vnode object
 900  * to write out.
 901  * We invalidate (remove) all pages from the address space
 902  * for semantic correctness.
 903  *
 904  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 905  * may start out with a NULL object.
 906  */
 907 void
 908 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 909     boolean_t syncio, boolean_t invalidate)
 910 {
 911         vm_object_t backing_object;
 912         struct vnode *vp;
 913         struct mount *mp;
 914         int flags;
 915
 916         if (object == NULL)
 917                 return;
 918         VM_OBJECT_LOCK(object);
 919         while ((backing_object = object->backing_object) != NULL) {
 920                 VM_OBJECT_LOCK(backing_object);
 921                 offset += object->backing_object_offset;
 922                 VM_OBJECT_UNLOCK(object);
 923                 object = backing_object;
 924                 if (object->size < OFF_TO_IDX(offset + size))
 925                         size = IDX_TO_OFF(object->size) - offset;
 926         }
 927         /*
 928          * Flush pages if writing is allowed, invalidate them
 929          * if invalidation requested.  Pages undergoing I/O
 930          * will be ignored by vm_object_page_remove().
 931          *
 932          * We cannot lock the vnode and then wait for paging
 933          * to complete without deadlocking against vm_fault.
 934          * Instead we simply call vm_object_page_remove() and
 935          * allow it to block internally on a page-by-page
 936          * basis when it encounters pages undergoing async
 937          * I/O.
 938          */
 939         if (object->type == OBJT_VNODE &&
 940             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 941                 int vfslocked;
 942                 vp = object->handle;
 943                 VM_OBJECT_UNLOCK(object);
 944                 (void) vn_start_write(vp, &mp, V_WAIT);
 945                 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 946                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 947                 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
 948                 flags |= invalidate ? OBJPC_INVAL : 0;
 949                 VM_OBJECT_LOCK(object);
 950                 vm_object_page_clean(object, offset, offset + size, flags);
 951                 VM_OBJECT_UNLOCK(object);
 952                 VOP_UNLOCK(vp, 0);
 953                 VFS_UNLOCK_GIANT(vfslocked);
 954                 vn_finished_write(mp);
 955                 VM_OBJECT_LOCK(object);
 956         }
 957         if ((object->type == OBJT_VNODE ||
 958              object->type == OBJT_DEVICE) && invalidate) {
 959                 boolean_t purge;
 960                 purge = old_msync || (object->type == OBJT_DEVICE);
 961                 vm_object_page_remove(object,
 962                     OFF_TO_IDX(offset),
 963                     OFF_TO_IDX(offset + size + PAGE_MASK),
 964                     purge ? FALSE : TRUE);
 965         }
 966         VM_OBJECT_UNLOCK(object);
 967 }
 968
 969 /*
 970  *      vm_object_madvise:
 971  *
 972  *      Implements the madvise function at the object/page level.
 973  *
 974  *      MADV_WILLNEED   (any object)
 975  *
 976  *          Activate the specified pages if they are resident.
 977  *
 978  *      MADV_DONTNEED   (any object)
 979  *
 980  *          Deactivate the specified pages if they are resident.
 981  *
 982  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
 983  *                       OBJ_ONEMAPPING only)
 984  *
 985  *          Deactivate and clean the specified pages if they are
 986  *          resident.  This permits the process to reuse the pages
 987  *          without faulting or the kernel to reclaim the pages
 988  *          without I/O.
 989  */
 990 void
 991 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
 992 {
 993         vm_pindex_t end, tpindex;
 994         vm_object_t backing_object, tobject;
 995         vm_page_t m;
 996
 997         if (object == NULL)
 998                 return;
 999         VM_OBJECT_LOCK(object);
1000         end = pindex + count;
1001         /*
1002          * Locate and adjust resident pages
1003          */
1004         for (; pindex < end; pindex += 1) {
1005 relookup:
1006                 tobject = object;
1007                 tpindex = pindex;
1008 shadowlookup:
1009                 /*
1010                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1011                  * and those pages must be OBJ_ONEMAPPING.
1012                  */
1013                 if (advise == MADV_FREE) {
1014                         if ((tobject->type != OBJT_DEFAULT &&
1015                              tobject->type != OBJT_SWAP) ||
1016                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1017                                 goto unlock_tobject;
1018                         }
1019                 } else if (tobject->type == OBJT_PHYS)
1020                         goto unlock_tobject;
1021                 m = vm_page_lookup(tobject, tpindex);
1022                 if (m == NULL && advise == MADV_WILLNEED) {
1023                         /*
1024                          * If the page is cached, reactivate it.
1025                          */
1026                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1027                             VM_ALLOC_NOBUSY);
1028                 }
1029                 if (m == NULL) {
1030                         /*
1031                          * There may be swap even if there is no backing page
1032                          */
1033                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1034                                 swap_pager_freespace(tobject, tpindex, 1);
1035                         /*
1036                          * next object
1037                          */
1038                         backing_object = tobject->backing_object;
1039                         if (backing_object == NULL)
1040                                 goto unlock_tobject;
1041                         VM_OBJECT_LOCK(backing_object);
1042                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1043                         if (tobject != object)
1044                                 VM_OBJECT_UNLOCK(tobject);
1045                         tobject = backing_object;
1046                         goto shadowlookup;
1047                 } else if (m->valid != VM_PAGE_BITS_ALL)
1048                         goto unlock_tobject;
1049                 /*
1050                  * If the page is not in a normal state, skip it.
1051                  */
1052                 vm_page_lock(m);
1053                 if (m->hold_count != 0 || m->wire_count != 0) {
1054                         vm_page_unlock(m);
1055                         goto unlock_tobject;
1056                 }
1057                 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1058                     ("vm_object_madvise: page %p is not managed", m));
1059                 if ((m->oflags & VPO_BUSY) || m->busy) {
1060                         if (advise == MADV_WILLNEED) {
1061                                 /*
1062                                  * Reference the page before unlocking and
1063                                  * sleeping so that the page daemon is less
1064                                  * likely to reclaim it.
1065                                  */
1066                                 vm_page_lock_queues();
1067                                 vm_page_flag_set(m, PG_REFERENCED);
1068                                 vm_page_unlock_queues();
1069                         }
1070                         vm_page_unlock(m);
1071                         if (object != tobject)
1072                                 VM_OBJECT_UNLOCK(object);
1073                         m->oflags |= VPO_WANTED;
1074                         msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1075                             0);
1076                         VM_OBJECT_LOCK(object);
1077                         goto relookup;
1078                 }
1079                 if (advise == MADV_WILLNEED) {
1080                         vm_page_activate(m);
1081                 } else if (advise == MADV_DONTNEED) {
1082                         vm_page_dontneed(m);
1083                 } else if (advise == MADV_FREE) {
1084                         /*
1085                          * Mark the page clean.  This will allow the page
1086                          * to be freed up by the system.  However, such pages
1087                          * are often reused quickly by malloc()/free()
1088                          * so we do not do anything that would cause
1089                          * a page fault if we can help it.
1090                          *
1091                          * Specifically, we do not try to actually free
1092                          * the page now nor do we try to put it in the
1093                          * cache (which would cause a page fault on reuse).
1094                          *
1095                          * But we do make the page is freeable as we
1096                          * can without actually taking the step of unmapping
1097                          * it.
1098                          */
1099                         pmap_clear_modify(m);
1100                         m->dirty = 0;
1101                         m->act_count = 0;
1102                         vm_page_dontneed(m);
1103                 }
1104                 vm_page_unlock(m);
1105                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1106                         swap_pager_freespace(tobject, tpindex, 1);
1107 unlock_tobject:
1108                 if (tobject != object)
1109                         VM_OBJECT_UNLOCK(tobject);
1110         }
1111         VM_OBJECT_UNLOCK(object);
1112 }
1113
1114 /*
1115  *      vm_object_shadow:
1116  *
1117  *      Create a new object which is backed by the
1118  *      specified existing object range.  The source
1119  *      object reference is deallocated.
1120  *
1121  *      The new object and offset into that object
1122  *      are returned in the source parameters.
1123  */
1124 void
1125 vm_object_shadow(
1126         vm_object_t *object,    /* IN/OUT */
1127         vm_ooffset_t *offset,   /* IN/OUT */
1128         vm_size_t length)
1129 {
1130         vm_object_t source;
1131         vm_object_t result;
1132
1133         source = *object;
1134
1135         /*
1136          * Don't create the new object if the old object isn't shared.
1137          */
1138         if (source != NULL) {
1139                 VM_OBJECT_LOCK(source);
1140                 if (source->ref_count == 1 &&
1141                     source->handle == NULL &&
1142                     (source->type == OBJT_DEFAULT ||
1143                      source->type == OBJT_SWAP)) {
1144                         VM_OBJECT_UNLOCK(source);
1145                         return;
1146                 }
1147                 VM_OBJECT_UNLOCK(source);
1148         }
1149
1150         /*
1151          * Allocate a new object with the given length.
1152          */
1153         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1154
1155         /*
1156          * The new object shadows the source object, adding a reference to it.
1157          * Our caller changes his reference to point to the new object,
1158          * removing a reference to the source object.  Net result: no change
1159          * of reference count.
1160          *
1161          * Try to optimize the result object's page color when shadowing
1162          * in order to maintain page coloring consistency in the combined
1163          * shadowed object.
1164          */
1165         result->backing_object = source;
1166         /*
1167          * Store the offset into the source object, and fix up the offset into
1168          * the new object.
1169          */
1170         result->backing_object_offset = *offset;
1171         if (source != NULL) {
1172                 VM_OBJECT_LOCK(source);
1173                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1174                 source->shadow_count++;
1175 #if VM_NRESERVLEVEL > 0
1176                 result->flags |= source->flags & OBJ_COLORED;
1177                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1178                     ((1 << (VM_NFREEORDER - 1)) - 1);
1179 #endif
1180                 VM_OBJECT_UNLOCK(source);
1181         }
1182
1183
1184         /*
1185          * Return the new things
1186          */
1187         *offset = 0;
1188         *object = result;
1189 }
1190
1191 /*
1192  *      vm_object_split:
1193  *
1194  * Split the pages in a map entry into a new object.  This affords
1195  * easier removal of unused pages, and keeps object inheritance from
1196  * being a negative impact on memory usage.
1197  */
1198 void
1199 vm_object_split(vm_map_entry_t entry)
1200 {
1201         vm_page_t m, m_next;
1202         vm_object_t orig_object, new_object, source;
1203         vm_pindex_t idx, offidxstart;
1204         vm_size_t size;
1205
1206         orig_object = entry->object.vm_object;
1207         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1208                 return;
1209         if (orig_object->ref_count <= 1)
1210                 return;
1211         VM_OBJECT_UNLOCK(orig_object);
1212
1213         offidxstart = OFF_TO_IDX(entry->offset);
1214         size = atop(entry->end - entry->start);
1215
1216         /*
1217          * If swap_pager_copy() is later called, it will convert new_object
1218          * into a swap object.
1219          */
1220         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1221
1222         /*
1223          * At this point, the new object is still private, so the order in
1224          * which the original and new objects are locked does not matter.
1225          */
1226         VM_OBJECT_LOCK(new_object);
1227         VM_OBJECT_LOCK(orig_object);
1228         source = orig_object->backing_object;
1229         if (source != NULL) {
1230                 VM_OBJECT_LOCK(source);
1231                 if ((source->flags & OBJ_DEAD) != 0) {
1232                         VM_OBJECT_UNLOCK(source);
1233                         VM_OBJECT_UNLOCK(orig_object);
1234                         VM_OBJECT_UNLOCK(new_object);
1235                         vm_object_deallocate(new_object);
1236                         VM_OBJECT_LOCK(orig_object);
1237                         return;
1238                 }
1239                 LIST_INSERT_HEAD(&source->shadow_head,
1240                                   new_object, shadow_list);
1241                 source->shadow_count++;
1242                 vm_object_reference_locked(source);     /* for new_object */
1243                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1244                 VM_OBJECT_UNLOCK(source);
1245                 new_object->backing_object_offset =
1246                         orig_object->backing_object_offset + entry->offset;
1247                 new_object->backing_object = source;
1248         }
1249         if (orig_object->cred != NULL) {
1250                 new_object->cred = orig_object->cred;
1251                 crhold(orig_object->cred);
1252                 new_object->charge = ptoa(size);
1253                 KASSERT(orig_object->charge >= ptoa(size),
1254                     ("orig_object->charge < 0"));
1255                 orig_object->charge -= ptoa(size);
1256         }
1257 retry:
1258         m = vm_page_find_least(orig_object, offidxstart);
1259         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1260             m = m_next) {
1261                 m_next = TAILQ_NEXT(m, listq);
1262
1263                 /*
1264                  * We must wait for pending I/O to complete before we can
1265                  * rename the page.
1266                  *
1267                  * We do not have to VM_PROT_NONE the page as mappings should
1268                  * not be changed by this operation.
1269                  */
1270                 if ((m->oflags & VPO_BUSY) || m->busy) {
1271                         VM_OBJECT_UNLOCK(new_object);
1272                         m->oflags |= VPO_WANTED;
1273                         msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1274                         VM_OBJECT_LOCK(new_object);
1275                         goto retry;
1276                 }
1277                 vm_page_lock(m);
1278                 vm_page_rename(m, new_object, idx);
1279                 vm_page_unlock(m);
1280                 /* page automatically made dirty by rename and cache handled */
1281                 vm_page_busy(m);
1282         }
1283         if (orig_object->type == OBJT_SWAP) {
1284                 /*
1285                  * swap_pager_copy() can sleep, in which case the orig_object's
1286                  * and new_object's locks are released and reacquired.
1287                  */
1288                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1289
1290                 /*
1291                  * Transfer any cached pages from orig_object to new_object.
1292                  */
1293                 if (__predict_false(orig_object->cache != NULL))
1294                         vm_page_cache_transfer(orig_object, offidxstart,
1295                             new_object);
1296         }
1297         VM_OBJECT_UNLOCK(orig_object);
1298         TAILQ_FOREACH(m, &new_object->memq, listq)
1299                 vm_page_wakeup(m);
1300         VM_OBJECT_UNLOCK(new_object);
1301         entry->object.vm_object = new_object;
1302         entry->offset = 0LL;
1303         vm_object_deallocate(orig_object);
1304         VM_OBJECT_LOCK(new_object);
1305 }
1306
1307 #define OBSC_TEST_ALL_SHADOWED  0x0001
1308 #define OBSC_COLLAPSE_NOWAIT    0x0002
1309 #define OBSC_COLLAPSE_WAIT      0x0004
1310
1311 static int
1312 vm_object_backing_scan(vm_object_t object, int op)
1313 {
1314         int r = 1;
1315         vm_page_t p;
1316         vm_object_t backing_object;
1317         vm_pindex_t backing_offset_index;
1318
1319         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1320         VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1321
1322         backing_object = object->backing_object;
1323         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1324
1325         /*
1326          * Initial conditions
1327          */
1328         if (op & OBSC_TEST_ALL_SHADOWED) {
1329                 /*
1330                  * We do not want to have to test for the existence of cache
1331                  * or swap pages in the backing object.  XXX but with the
1332                  * new swapper this would be pretty easy to do.
1333                  *
1334                  * XXX what about anonymous MAP_SHARED memory that hasn't
1335                  * been ZFOD faulted yet?  If we do not test for this, the
1336                  * shadow test may succeed! XXX
1337                  */
1338                 if (backing_object->type != OBJT_DEFAULT) {
1339                         return (0);
1340                 }
1341         }
1342         if (op & OBSC_COLLAPSE_WAIT) {
1343                 vm_object_set_flag(backing_object, OBJ_DEAD);
1344         }
1345
1346         /*
1347          * Our scan
1348          */
1349         p = TAILQ_FIRST(&backing_object->memq);
1350         while (p) {
1351                 vm_page_t next = TAILQ_NEXT(p, listq);
1352                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1353
1354                 if (op & OBSC_TEST_ALL_SHADOWED) {
1355                         vm_page_t pp;
1356
1357                         /*
1358                          * Ignore pages outside the parent object's range
1359                          * and outside the parent object's mapping of the
1360                          * backing object.
1361                          *
1362                          * note that we do not busy the backing object's
1363                          * page.
1364                          */
1365                         if (
1366                             p->pindex < backing_offset_index ||
1367                             new_pindex >= object->size
1368                         ) {
1369                                 p = next;
1370                                 continue;
1371                         }
1372
1373                         /*
1374                          * See if the parent has the page or if the parent's
1375                          * object pager has the page.  If the parent has the
1376                          * page but the page is not valid, the parent's
1377                          * object pager must have the page.
1378                          *
1379                          * If this fails, the parent does not completely shadow
1380                          * the object and we might as well give up now.
1381                          */
1382
1383                         pp = vm_page_lookup(object, new_pindex);
1384                         if (
1385                             (pp == NULL || pp->valid == 0) &&
1386                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1387                         ) {
1388                                 r = 0;
1389                                 break;
1390                         }
1391                 }
1392
1393                 /*
1394                  * Check for busy page
1395                  */
1396                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1397                         vm_page_t pp;
1398
1399                         if (op & OBSC_COLLAPSE_NOWAIT) {
1400                                 if ((p->oflags & VPO_BUSY) ||
1401                                     !p->valid ||
1402                                     p->busy) {
1403                                         p = next;
1404                                         continue;
1405                                 }
1406                         } else if (op & OBSC_COLLAPSE_WAIT) {
1407                                 if ((p->oflags & VPO_BUSY) || p->busy) {
1408                                         VM_OBJECT_UNLOCK(object);
1409                                         p->oflags |= VPO_WANTED;
1410                                         msleep(p, VM_OBJECT_MTX(backing_object),
1411                                             PDROP | PVM, "vmocol", 0);
1412                                         VM_OBJECT_LOCK(object);
1413                                         VM_OBJECT_LOCK(backing_object);
1414                                         /*
1415                                          * If we slept, anything could have
1416                                          * happened.  Since the object is
1417                                          * marked dead, the backing offset
1418                                          * should not have changed so we
1419                                          * just restart our scan.
1420                                          */
1421                                         p = TAILQ_FIRST(&backing_object->memq);
1422                                         continue;
1423                                 }
1424                         }
1425
1426                         KASSERT(
1427                             p->object == backing_object,
1428                             ("vm_object_backing_scan: object mismatch")
1429                         );
1430
1431                         /*
1432                          * Destroy any associated swap
1433                          */
1434                         if (backing_object->type == OBJT_SWAP) {
1435                                 swap_pager_freespace(
1436                                     backing_object,
1437                                     p->pindex,
1438                                     1
1439                                 );
1440                         }
1441
1442                         if (
1443                             p->pindex < backing_offset_index ||
1444                             new_pindex >= object->size
1445                         ) {
1446                                 /*
1447                                  * Page is out of the parent object's range, we
1448                                  * can simply destroy it.
1449                                  */
1450                                 vm_page_lock(p);
1451                                 KASSERT(!pmap_page_is_mapped(p),
1452                                     ("freeing mapped page %p", p));
1453                                 if (p->wire_count == 0)
1454                                         vm_page_free(p);
1455                                 else
1456                                         vm_page_remove(p);
1457                                 vm_page_unlock(p);
1458                                 p = next;
1459                                 continue;
1460                         }
1461
1462                         pp = vm_page_lookup(object, new_pindex);
1463                         if (
1464                             pp != NULL ||
1465                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1466                         ) {
1467                                 /*
1468                                  * page already exists in parent OR swap exists
1469                                  * for this location in the parent.  Destroy
1470                                  * the original page from the backing object.
1471                                  *
1472                                  * Leave the parent's page alone
1473                                  */
1474                                 vm_page_lock(p);
1475                                 KASSERT(!pmap_page_is_mapped(p),
1476                                     ("freeing mapped page %p", p));
1477                                 if (p->wire_count == 0)
1478                                         vm_page_free(p);
1479                                 else
1480                                         vm_page_remove(p);
1481                                 vm_page_unlock(p);
1482                                 p = next;
1483                                 continue;
1484                         }
1485
1486 #if VM_NRESERVLEVEL > 0
1487                         /*
1488                          * Rename the reservation.
1489                          */
1490                         vm_reserv_rename(p, object, backing_object,
1491                             backing_offset_index);
1492 #endif
1493
1494                         /*
1495                          * Page does not exist in parent, rename the
1496                          * page from the backing object to the main object.
1497                          *
1498                          * If the page was mapped to a process, it can remain
1499                          * mapped through the rename.
1500                          */
1501                         vm_page_lock(p);
1502                         vm_page_rename(p, object, new_pindex);
1503                         vm_page_unlock(p);
1504                         /* page automatically made dirty by rename */
1505                 }
1506                 p = next;
1507         }
1508         return (r);
1509 }
1510
1511
1512 /*
1513  * this version of collapse allows the operation to occur earlier and
1514  * when paging_in_progress is true for an object...  This is not a complete
1515  * operation, but should plug 99.9% of the rest of the leaks.
1516  */
1517 static void
1518 vm_object_qcollapse(vm_object_t object)
1519 {
1520         vm_object_t backing_object = object->backing_object;
1521
1522         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1523         VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1524
1525         if (backing_object->ref_count != 1)
1526                 return;
1527
1528         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1529 }
1530
1531 /*
1532  *      vm_object_collapse:
1533  *
1534  *      Collapse an object with the object backing it.
1535  *      Pages in the backing object are moved into the
1536  *      parent, and the backing object is deallocated.
1537  */
1538 void
1539 vm_object_collapse(vm_object_t object)
1540 {
1541         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1542
1543         while (TRUE) {
1544                 vm_object_t backing_object;
1545
1546                 /*
1547                  * Verify that the conditions are right for collapse:
1548                  *
1549                  * The object exists and the backing object exists.
1550                  */
1551                 if ((backing_object = object->backing_object) == NULL)
1552                         break;
1553
1554                 /*
1555                  * we check the backing object first, because it is most likely
1556                  * not collapsable.
1557                  */
1558                 VM_OBJECT_LOCK(backing_object);
1559                 if (backing_object->handle != NULL ||
1560                     (backing_object->type != OBJT_DEFAULT &&
1561                      backing_object->type != OBJT_SWAP) ||
1562                     (backing_object->flags & OBJ_DEAD) ||
1563                     object->handle != NULL ||
1564                     (object->type != OBJT_DEFAULT &&
1565                      object->type != OBJT_SWAP) ||
1566                     (object->flags & OBJ_DEAD)) {
1567                         VM_OBJECT_UNLOCK(backing_object);
1568                         break;
1569                 }
1570
1571                 if (
1572                     object->paging_in_progress != 0 ||
1573                     backing_object->paging_in_progress != 0
1574                 ) {
1575                         vm_object_qcollapse(object);
1576                         VM_OBJECT_UNLOCK(backing_object);
1577                         break;
1578                 }
1579                 /*
1580                  * We know that we can either collapse the backing object (if
1581                  * the parent is the only reference to it) or (perhaps) have
1582                  * the parent bypass the object if the parent happens to shadow
1583                  * all the resident pages in the entire backing object.
1584                  *
1585                  * This is ignoring pager-backed pages such as swap pages.
1586                  * vm_object_backing_scan fails the shadowing test in this
1587                  * case.
1588                  */
1589                 if (backing_object->ref_count == 1) {
1590                         /*
1591                          * If there is exactly one reference to the backing
1592                          * object, we can collapse it into the parent.
1593                          */
1594                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1595
1596 #if VM_NRESERVLEVEL > 0
1597                         /*
1598                          * Break any reservations from backing_object.
1599                          */
1600                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1601                                 vm_reserv_break_all(backing_object);
1602 #endif
1603
1604                         /*
1605                          * Move the pager from backing_object to object.
1606                          */
1607                         if (backing_object->type == OBJT_SWAP) {
1608                                 /*
1609                                  * swap_pager_copy() can sleep, in which case
1610                                  * the backing_object's and object's locks are
1611                                  * released and reacquired.
1612                                  */
1613                                 swap_pager_copy(
1614                                     backing_object,
1615                                     object,
1616                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1617
1618                                 /*
1619                                  * Free any cached pages from backing_object.
1620                                  */
1621                                 if (__predict_false(backing_object->cache != NULL))
1622                                         vm_page_cache_free(backing_object, 0, 0);
1623                         }
1624                         /*
1625                          * Object now shadows whatever backing_object did.
1626                          * Note that the reference to
1627                          * backing_object->backing_object moves from within
1628                          * backing_object to within object.
1629                          */
1630                         LIST_REMOVE(object, shadow_list);
1631                         backing_object->shadow_count--;
1632                         if (backing_object->backing_object) {
1633                                 VM_OBJECT_LOCK(backing_object->backing_object);
1634                                 LIST_REMOVE(backing_object, shadow_list);
1635                                 LIST_INSERT_HEAD(
1636                                     &backing_object->backing_object->shadow_head,
1637                                     object, shadow_list);
1638                                 /*
1639                                  * The shadow_count has not changed.
1640                                  */
1641                                 VM_OBJECT_UNLOCK(backing_object->backing_object);
1642                         }
1643                         object->backing_object = backing_object->backing_object;
1644                         object->backing_object_offset +=
1645                             backing_object->backing_object_offset;
1646
1647                         /*
1648                          * Discard backing_object.
1649                          *
1650                          * Since the backing object has no pages, no pager left,
1651                          * and no object references within it, all that is
1652                          * necessary is to dispose of it.
1653                          */
1654                         KASSERT(backing_object->ref_count == 1, (
1655 "backing_object %p was somehow re-referenced during collapse!",
1656                             backing_object));
1657                         VM_OBJECT_UNLOCK(backing_object);
1658                         vm_object_destroy(backing_object);
1659
1660                         object_collapses++;
1661                 } else {
1662                         vm_object_t new_backing_object;
1663
1664                         /*
1665                          * If we do not entirely shadow the backing object,
1666                          * there is nothing we can do so we give up.
1667                          */
1668                         if (object->resident_page_count != object->size &&
1669                             vm_object_backing_scan(object,
1670                             OBSC_TEST_ALL_SHADOWED) == 0) {
1671                                 VM_OBJECT_UNLOCK(backing_object);
1672                                 break;
1673                         }
1674
1675                         /*
1676                          * Make the parent shadow the next object in the
1677                          * chain.  Deallocating backing_object will not remove
1678                          * it, since its reference count is at least 2.
1679                          */
1680                         LIST_REMOVE(object, shadow_list);
1681                         backing_object->shadow_count--;
1682
1683                         new_backing_object = backing_object->backing_object;
1684                         if ((object->backing_object = new_backing_object) != NULL) {
1685                                 VM_OBJECT_LOCK(new_backing_object);
1686                                 LIST_INSERT_HEAD(
1687                                     &new_backing_object->shadow_head,
1688                                     object,
1689                                     shadow_list
1690                                 );
1691                                 new_backing_object->shadow_count++;
1692                                 vm_object_reference_locked(new_backing_object);
1693                                 VM_OBJECT_UNLOCK(new_backing_object);
1694                                 object->backing_object_offset +=
1695                                         backing_object->backing_object_offset;
1696                         }
1697
1698                         /*
1699                          * Drop the reference count on backing_object. Since
1700                          * its ref_count was at least 2, it will not vanish.
1701                          */
1702                         backing_object->ref_count--;
1703                         VM_OBJECT_UNLOCK(backing_object);
1704                         object_bypasses++;
1705                 }
1706
1707                 /*
1708                  * Try again with this object's new backing object.
1709                  */
1710         }
1711 }
1712
1713 /*
1714  *      vm_object_page_remove:
1715  *
1716  *      For the given object, either frees or invalidates each of the
1717  *      specified pages.  In general, a page is freed.  However, if a
1718  *      page is wired for any reason other than the existence of a
1719  *      managed, wired mapping, then it may be invalidated but not
1720  *      removed from the object.  Pages are specified by the given
1721  *      range ["start", "end") and Boolean "clean_only".  As a
1722  *      special case, if "end" is zero, then the range extends from
1723  *      "start" to the end of the object.  If "clean_only" is TRUE,
1724  *      then only the non-dirty pages within the specified range are
1725  *      affected.
1726  *
1727  *      In general, this operation should only be performed on objects
1728  *      that contain managed pages.  There are two exceptions.  First,
1729  *      it may be performed on the kernel and kmem objects.  Second,
1730  *      it may be used by msync(..., MS_INVALIDATE) to invalidate
1731  *      device-backed pages.  In both of these cases, "clean_only"
1732  *      must be FALSE.
1733  *
1734  *      The object must be locked.
1735  */
1736 void
1737 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1738     boolean_t clean_only)
1739 {
1740         vm_page_t p, next;
1741         int wirings;
1742
1743         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1744         if (object->resident_page_count == 0)
1745                 goto skipmemq;
1746
1747         /*
1748          * Since physically-backed objects do not use managed pages, we can't
1749          * remove pages from the object (we must instead remove the page
1750          * references, and then destroy the object).
1751          */
1752         KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1753             object == kmem_object,
1754             ("attempt to remove pages from a physical object"));
1755
1756         vm_object_pip_add(object, 1);
1757 again:
1758         p = vm_page_find_least(object, start);
1759
1760         /*
1761          * Assert: the variable p is either (1) the page with the
1762          * least pindex greater than or equal to the parameter pindex
1763          * or (2) NULL.
1764          */
1765         for (;
1766              p != NULL && (p->pindex < end || end == 0);
1767              p = next) {
1768                 next = TAILQ_NEXT(p, listq);
1769
1770                 /*
1771                  * If the page is wired for any reason besides the
1772                  * existence of managed, wired mappings, then it cannot
1773                  * be freed.  For example, fictitious pages, which
1774                  * represent device memory, are inherently wired and
1775                  * cannot be freed.  They can, however, be invalidated
1776                  * if "clean_only" is FALSE.
1777                  */
1778                 vm_page_lock(p);
1779                 if ((wirings = p->wire_count) != 0 &&
1780                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1781                         /* Fictitious pages do not have managed mappings. */
1782                         if ((p->flags & PG_FICTITIOUS) == 0)
1783                                 pmap_remove_all(p);
1784                         /* Account for removal of managed, wired mappings. */
1785                         p->wire_count -= wirings;
1786                         if (!clean_only) {
1787                                 p->valid = 0;
1788                                 vm_page_undirty(p);
1789                         }
1790                         vm_page_unlock(p);
1791                         continue;
1792                 }
1793                 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1794                         goto again;
1795                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1796                     ("vm_object_page_remove: page %p is fictitious", p));
1797                 if (clean_only && p->valid) {
1798                         pmap_remove_write(p);
1799                         if (p->dirty) {
1800                                 vm_page_unlock(p);
1801                                 continue;
1802                         }
1803                 }
1804                 pmap_remove_all(p);
1805                 /* Account for removal of managed, wired mappings. */
1806                 if (wirings != 0)
1807                         p->wire_count -= wirings;
1808                 vm_page_free(p);
1809                 vm_page_unlock(p);
1810         }
1811         vm_object_pip_wakeup(object);
1812 skipmemq:
1813         if (__predict_false(object->cache != NULL))
1814                 vm_page_cache_free(object, start, end);
1815 }
1816
1817 /*
1818  *      Populate the specified range of the object with valid pages.  Returns
1819  *      TRUE if the range is successfully populated and FALSE otherwise.
1820  *
1821  *      Note: This function should be optimized to pass a larger array of
1822  *      pages to vm_pager_get_pages() before it is applied to a non-
1823  *      OBJT_DEVICE object.
1824  *
1825  *      The object must be locked.
1826  */
1827 boolean_t
1828 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1829 {
1830         vm_page_t m, ma[1];
1831         vm_pindex_t pindex;
1832         int rv;
1833
1834         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1835         for (pindex = start; pindex < end; pindex++) {
1836                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1837                     VM_ALLOC_RETRY);
1838                 if (m->valid != VM_PAGE_BITS_ALL) {
1839                         ma[0] = m;
1840                         rv = vm_pager_get_pages(object, ma, 1, 0);
1841                         m = vm_page_lookup(object, pindex);
1842                         if (m == NULL)
1843                                 break;
1844                         if (rv != VM_PAGER_OK) {
1845                                 vm_page_lock(m);
1846                                 vm_page_free(m);
1847                                 vm_page_unlock(m);
1848                                 break;
1849                         }
1850                 }
1851                 /*
1852                  * Keep "m" busy because a subsequent iteration may unlock
1853                  * the object.
1854                  */
1855         }
1856         if (pindex > start) {
1857                 m = vm_page_lookup(object, start);
1858                 while (m != NULL && m->pindex < pindex) {
1859                         vm_page_wakeup(m);
1860                         m = TAILQ_NEXT(m, listq);
1861                 }
1862         }
1863         return (pindex == end);
1864 }
1865
1866 /*
1867  *      Routine:        vm_object_coalesce
1868  *      Function:       Coalesces two objects backing up adjoining
1869  *                      regions of memory into a single object.
1870  *
1871  *      returns TRUE if objects were combined.
1872  *
1873  *      NOTE:   Only works at the moment if the second object is NULL -
1874  *              if it's not, which object do we lock first?
1875  *
1876  *      Parameters:
1877  *              prev_object     First object to coalesce
1878  *              prev_offset     Offset into prev_object
1879  *              prev_size       Size of reference to prev_object
1880  *              next_size       Size of reference to the second object
1881  *              reserved        Indicator that extension region has
1882  *                              swap accounted for
1883  *
1884  *      Conditions:
1885  *      The object must *not* be locked.
1886  */
1887 boolean_t
1888 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1889     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1890 {
1891         vm_pindex_t next_pindex;
1892
1893         if (prev_object == NULL)
1894                 return (TRUE);
1895         VM_OBJECT_LOCK(prev_object);
1896         if (prev_object->type != OBJT_DEFAULT &&
1897             prev_object->type != OBJT_SWAP) {
1898                 VM_OBJECT_UNLOCK(prev_object);
1899                 return (FALSE);
1900         }
1901
1902         /*
1903          * Try to collapse the object first
1904          */
1905         vm_object_collapse(prev_object);
1906
1907         /*
1908          * Can't coalesce if: . more than one reference . paged out . shadows
1909          * another object . has a copy elsewhere (any of which mean that the
1910          * pages not mapped to prev_entry may be in use anyway)
1911          */
1912         if (prev_object->backing_object != NULL) {
1913                 VM_OBJECT_UNLOCK(prev_object);
1914                 return (FALSE);
1915         }
1916
1917         prev_size >>= PAGE_SHIFT;
1918         next_size >>= PAGE_SHIFT;
1919         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1920
1921         if ((prev_object->ref_count > 1) &&
1922             (prev_object->size != next_pindex)) {
1923                 VM_OBJECT_UNLOCK(prev_object);
1924                 return (FALSE);
1925         }
1926
1927         /*
1928          * Account for the charge.
1929          */
1930         if (prev_object->cred != NULL) {
1931
1932                 /*
1933                  * If prev_object was charged, then this mapping,
1934                  * althought not charged now, may become writable
1935                  * later. Non-NULL cred in the object would prevent
1936                  * swap reservation during enabling of the write
1937                  * access, so reserve swap now. Failed reservation
1938                  * cause allocation of the separate object for the map
1939                  * entry, and swap reservation for this entry is
1940                  * managed in appropriate time.
1941                  */
1942                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
1943                     prev_object->cred)) {
1944                         return (FALSE);
1945                 }
1946                 prev_object->charge += ptoa(next_size);
1947         }
1948
1949         /*
1950          * Remove any pages that may still be in the object from a previous
1951          * deallocation.
1952          */
1953         if (next_pindex < prev_object->size) {
1954                 vm_object_page_remove(prev_object,
1955                                       next_pindex,
1956                                       next_pindex + next_size, FALSE);
1957                 if (prev_object->type == OBJT_SWAP)
1958                         swap_pager_freespace(prev_object,
1959                                              next_pindex, next_size);
1960 #if 0
1961                 if (prev_object->cred != NULL) {
1962                         KASSERT(prev_object->charge >=
1963                             ptoa(prev_object->size - next_pindex),
1964                             ("object %p overcharged 1 %jx %jx", prev_object,
1965                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
1966                         prev_object->charge -= ptoa(prev_object->size -
1967                             next_pindex);
1968                 }
1969 #endif
1970         }
1971
1972         /*
1973          * Extend the object if necessary.
1974          */
1975         if (next_pindex + next_size > prev_object->size)
1976                 prev_object->size = next_pindex + next_size;
1977
1978         VM_OBJECT_UNLOCK(prev_object);
1979         return (TRUE);
1980 }
1981
1982 void
1983 vm_object_set_writeable_dirty(vm_object_t object)
1984 {
1985
1986         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1987         if (object->type != OBJT_VNODE)
1988                 return;
1989         object->generation++;
1990         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
1991                 return;
1992         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
1993 }
1994
1995 #include "opt_ddb.h"
1996 #ifdef DDB
1997 #include <sys/kernel.h>
1998
1999 #include <sys/cons.h>
2000
2001 #include <ddb/ddb.h>
2002
2003 static int
2004 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2005 {
2006         vm_map_t tmpm;
2007         vm_map_entry_t tmpe;
2008         vm_object_t obj;
2009         int entcount;
2010
2011         if (map == 0)
2012                 return 0;
2013
2014         if (entry == 0) {
2015                 tmpe = map->header.next;
2016                 entcount = map->nentries;
2017                 while (entcount-- && (tmpe != &map->header)) {
2018                         if (_vm_object_in_map(map, object, tmpe)) {
2019                                 return 1;
2020                         }
2021                         tmpe = tmpe->next;
2022                 }
2023         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2024                 tmpm = entry->object.sub_map;
2025                 tmpe = tmpm->header.next;
2026                 entcount = tmpm->nentries;
2027                 while (entcount-- && tmpe != &tmpm->header) {
2028                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2029                                 return 1;
2030                         }
2031                         tmpe = tmpe->next;
2032                 }
2033         } else if ((obj = entry->object.vm_object) != NULL) {
2034                 for (; obj; obj = obj->backing_object)
2035                         if (obj == object) {
2036                                 return 1;
2037                         }
2038         }
2039         return 0;
2040 }
2041
2042 static int
2043 vm_object_in_map(vm_object_t object)
2044 {
2045         struct proc *p;
2046
2047         /* sx_slock(&allproc_lock); */
2048         FOREACH_PROC_IN_SYSTEM(p) {
2049                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2050                         continue;
2051                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2052                         /* sx_sunlock(&allproc_lock); */
2053                         return 1;
2054                 }
2055         }
2056         /* sx_sunlock(&allproc_lock); */
2057         if (_vm_object_in_map(kernel_map, object, 0))
2058                 return 1;
2059         if (_vm_object_in_map(kmem_map, object, 0))
2060                 return 1;
2061         if (_vm_object_in_map(pager_map, object, 0))
2062                 return 1;
2063         if (_vm_object_in_map(buffer_map, object, 0))
2064                 return 1;
2065         return 0;
2066 }
2067
2068 DB_SHOW_COMMAND(vmochk, vm_object_check)
2069 {
2070         vm_object_t object;
2071
2072         /*
2073          * make sure that internal objs are in a map somewhere
2074          * and none have zero ref counts.
2075          */
2076         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2077                 if (object->handle == NULL &&
2078                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2079                         if (object->ref_count == 0) {
2080                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2081                                         (long)object->size);
2082                         }
2083                         if (!vm_object_in_map(object)) {
2084                                 db_printf(
2085                         "vmochk: internal obj is not in a map: "
2086                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2087                                     object->ref_count, (u_long)object->size,
2088                                     (u_long)object->size,
2089                                     (void *)object->backing_object);
2090                         }
2091                 }
2092         }
2093 }
2094
2095 /*
2096  *      vm_object_print:        [ debug ]
2097  */
2098 DB_SHOW_COMMAND(object, vm_object_print_static)
2099 {
2100         /* XXX convert args. */
2101         vm_object_t object = (vm_object_t)addr;
2102         boolean_t full = have_addr;
2103
2104         vm_page_t p;
2105
2106         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2107 #define count   was_count
2108
2109         int count;
2110
2111         if (object == NULL)
2112                 return;
2113
2114         db_iprintf(
2115             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2116             object, (int)object->type, (uintmax_t)object->size,
2117             object->resident_page_count, object->ref_count, object->flags,
2118             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2119         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2120             object->shadow_count,
2121             object->backing_object ? object->backing_object->ref_count : 0,
2122             object->backing_object, (uintmax_t)object->backing_object_offset);
2123
2124         if (!full)
2125                 return;
2126
2127         db_indent += 2;
2128         count = 0;
2129         TAILQ_FOREACH(p, &object->memq, listq) {
2130                 if (count == 0)
2131                         db_iprintf("memory:=");
2132                 else if (count == 6) {
2133                         db_printf("\n");
2134                         db_iprintf(" ...");
2135                         count = 0;
2136                 } else
2137                         db_printf(",");
2138                 count++;
2139
2140                 db_printf("(off=0x%jx,page=0x%jx)",
2141                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2142         }
2143         if (count != 0)
2144                 db_printf("\n");
2145         db_indent -= 2;
2146 }
2147
2148 /* XXX. */
2149 #undef count
2150
2151 /* XXX need this non-static entry for calling from vm_map_print. */
2152 void
2153 vm_object_print(
2154         /* db_expr_t */ long addr,
2155         boolean_t have_addr,
2156         /* db_expr_t */ long count,
2157         char *modif)
2158 {
2159         vm_object_print_static(addr, have_addr, count, modif);
2160 }
2161
2162 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2163 {
2164         vm_object_t object;
2165         vm_pindex_t fidx;
2166         vm_paddr_t pa;
2167         vm_page_t m, prev_m;
2168         int rcount, nl, c;
2169
2170         nl = 0;
2171         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2172                 db_printf("new object: %p\n", (void *)object);
2173                 if (nl > 18) {
2174                         c = cngetc();
2175                         if (c != ' ')
2176                                 return;
2177                         nl = 0;
2178                 }
2179                 nl++;
2180                 rcount = 0;
2181                 fidx = 0;
2182                 pa = -1;
2183                 TAILQ_FOREACH(m, &object->memq, listq) {
2184                         if (m->pindex > 128)
2185                                 break;
2186                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2187                             prev_m->pindex + 1 != m->pindex) {
2188                                 if (rcount) {
2189                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2190                                                 (long)fidx, rcount, (long)pa);
2191                                         if (nl > 18) {
2192                                                 c = cngetc();
2193                                                 if (c != ' ')
2194                                                         return;
2195                                                 nl = 0;
2196                                         }
2197                                         nl++;
2198                                         rcount = 0;
2199                                 }
2200                         }
2201                         if (rcount &&
2202                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2203                                 ++rcount;
2204                                 continue;
2205                         }
2206                         if (rcount) {
2207                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2208                                         (long)fidx, rcount, (long)pa);
2209                                 if (nl > 18) {
2210                                         c = cngetc();
2211                                         if (c != ' ')
2212                                                 return;
2213                                         nl = 0;
2214                                 }
2215                                 nl++;
2216                         }
2217                         fidx = m->pindex;
2218                         pa = VM_PAGE_TO_PHYS(m);
2219                         rcount = 1;
2220                 }
2221                 if (rcount) {
2222                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2223                                 (long)fidx, rcount, (long)pa);
2224                         if (nl > 18) {
2225                                 c = cngetc();
2226                                 if (c != ' ')
2227                                         return;
2228                                 nl = 0;
2229                         }
2230                         nl++;
2231                 }
2232         }
2233 }
2234 #endif /* DDB */