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/rwlock.h>
  82 #include <sys/user.h>
  83 #include <sys/vnode.h>
  84 #include <sys/vmmeter.h>
  85 #include <sys/sx.h>
  86
  87 #include <vm/vm.h>
  88 #include <vm/vm_param.h>
  89 #include <vm/pmap.h>
  90 #include <vm/vm_map.h>
  91 #include <vm/vm_object.h>
  92 #include <vm/vm_page.h>
  93 #include <vm/vm_pageout.h>
  94 #include <vm/vm_pager.h>
  95 #include <vm/swap_pager.h>
  96 #include <vm/vm_kern.h>
  97 #include <vm/vm_extern.h>
  98 #include <vm/vm_radix.h>
  99 #include <vm/vm_reserv.h>
 100 #include <vm/uma.h>
 101
 102 static int old_msync;
 103 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 104     "Use old (insecure) msync behavior");
 105
 106 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 107                     int pagerflags, int flags, boolean_t *clearobjflags,
 108                     boolean_t *eio);
 109 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 110                     boolean_t *clearobjflags);
 111 static void     vm_object_qcollapse(vm_object_t object);
 112 static void     vm_object_vndeallocate(vm_object_t object);
 113
 114 /*
 115  *      Virtual memory objects maintain the actual data
 116  *      associated with allocated virtual memory.  A given
 117  *      page of memory exists within exactly one object.
 118  *
 119  *      An object is only deallocated when all "references"
 120  *      are given up.  Only one "reference" to a given
 121  *      region of an object should be writeable.
 122  *
 123  *      Associated with each object is a list of all resident
 124  *      memory pages belonging to that object; this list is
 125  *      maintained by the "vm_page" module, and locked by the object's
 126  *      lock.
 127  *
 128  *      Each object also records a "pager" routine which is
 129  *      used to retrieve (and store) pages to the proper backing
 130  *      storage.  In addition, objects may be backed by other
 131  *      objects from which they were virtual-copied.
 132  *
 133  *      The only items within the object structure which are
 134  *      modified after time of creation are:
 135  *              reference count         locked by object's lock
 136  *              pager routine           locked by object's lock
 137  *
 138  */
 139
 140 struct object_q vm_object_list;
 141 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 142
 143 struct vm_object kernel_object_store;
 144 struct vm_object kmem_object_store;
 145
 146 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
 147     "VM object stats");
 148
 149 static long object_collapses;
 150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 151     &object_collapses, 0, "VM object collapses");
 152
 153 static long object_bypasses;
 154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 155     &object_bypasses, 0, "VM object bypasses");
 156
 157 static uma_zone_t obj_zone;
 158
 159 static int vm_object_zinit(void *mem, int size, int flags);
 160
 161 #ifdef INVARIANTS
 162 static void vm_object_zdtor(void *mem, int size, void *arg);
 163
 164 static void
 165 vm_object_zdtor(void *mem, int size, void *arg)
 166 {
 167         vm_object_t object;
 168
 169         object = (vm_object_t)mem;
 170         KASSERT(object->ref_count == 0,
 171             ("object %p ref_count = %d", object, object->ref_count));
 172         KASSERT(TAILQ_EMPTY(&object->memq),
 173             ("object %p has resident pages in its memq", object));
 174         KASSERT(vm_radix_is_empty(&object->rtree),
 175             ("object %p has resident pages in its trie", object));
 176 #if VM_NRESERVLEVEL > 0
 177         KASSERT(LIST_EMPTY(&object->rvq),
 178             ("object %p has reservations",
 179             object));
 180 #endif
 181         KASSERT(vm_object_cache_is_empty(object),
 182             ("object %p has cached pages",
 183             object));
 184         KASSERT(object->paging_in_progress == 0,
 185             ("object %p paging_in_progress = %d",
 186             object, object->paging_in_progress));
 187         KASSERT(object->resident_page_count == 0,
 188             ("object %p resident_page_count = %d",
 189             object, object->resident_page_count));
 190         KASSERT(object->shadow_count == 0,
 191             ("object %p shadow_count = %d",
 192             object, object->shadow_count));
 193         KASSERT(object->type == OBJT_DEAD,
 194             ("object %p has non-dead type %d",
 195             object, object->type));
 196 }
 197 #endif
 198
 199 static int
 200 vm_object_zinit(void *mem, int size, int flags)
 201 {
 202         vm_object_t object;
 203
 204         object = (vm_object_t)mem;
 205         bzero(&object->lock, sizeof(object->lock));
 206         rw_init_flags(&object->lock, "vm object", RW_DUPOK);
 207
 208         /* These are true for any object that has been freed */
 209         object->type = OBJT_DEAD;
 210         object->ref_count = 0;
 211         object->rtree.rt_root = 0;
 212         object->rtree.rt_flags = 0;
 213         object->paging_in_progress = 0;
 214         object->resident_page_count = 0;
 215         object->shadow_count = 0;
 216         object->cache.rt_root = 0;
 217         object->cache.rt_flags = 0;
 218
 219         mtx_lock(&vm_object_list_mtx);
 220         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 221         mtx_unlock(&vm_object_list_mtx);
 222         return (0);
 223 }
 224
 225 static void
 226 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 227 {
 228
 229         TAILQ_INIT(&object->memq);
 230         LIST_INIT(&object->shadow_head);
 231
 232         object->type = type;
 233         switch (type) {
 234         case OBJT_DEAD:
 235                 panic("_vm_object_allocate: can't create OBJT_DEAD");
 236         case OBJT_DEFAULT:
 237         case OBJT_SWAP:
 238                 object->flags = OBJ_ONEMAPPING;
 239                 break;
 240         case OBJT_DEVICE:
 241         case OBJT_SG:
 242                 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
 243                 break;
 244         case OBJT_MGTDEVICE:
 245                 object->flags = OBJ_FICTITIOUS;
 246                 break;
 247         case OBJT_PHYS:
 248                 object->flags = OBJ_UNMANAGED;
 249                 break;
 250         case OBJT_VNODE:
 251                 object->flags = 0;
 252                 break;
 253         default:
 254                 panic("_vm_object_allocate: type %d is undefined", type);
 255         }
 256         object->size = size;
 257         object->generation = 1;
 258         object->ref_count = 1;
 259         object->memattr = VM_MEMATTR_DEFAULT;
 260         object->cred = NULL;
 261         object->charge = 0;
 262         object->handle = NULL;
 263         object->backing_object = NULL;
 264         object->backing_object_offset = (vm_ooffset_t) 0;
 265 #if VM_NRESERVLEVEL > 0
 266         LIST_INIT(&object->rvq);
 267 #endif
 268 }
 269
 270 /*
 271  *      vm_object_init:
 272  *
 273  *      Initialize the VM objects module.
 274  */
 275 void
 276 vm_object_init(void)
 277 {
 278         TAILQ_INIT(&vm_object_list);
 279         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 280
 281         rw_init(&kernel_object->lock, "kernel vm object");
 282         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 283             kernel_object);
 284 #if VM_NRESERVLEVEL > 0
 285         kernel_object->flags |= OBJ_COLORED;
 286         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 287 #endif
 288
 289         rw_init(&kmem_object->lock, "kmem vm object");
 290         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 291             kmem_object);
 292 #if VM_NRESERVLEVEL > 0
 293         kmem_object->flags |= OBJ_COLORED;
 294         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 295 #endif
 296
 297         /*
 298          * The lock portion of struct vm_object must be type stable due
 299          * to vm_pageout_fallback_object_lock locking a vm object
 300          * without holding any references to it.
 301          */
 302         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 303 #ifdef INVARIANTS
 304             vm_object_zdtor,
 305 #else
 306             NULL,
 307 #endif
 308             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 309
 310         vm_radix_init();
 311 }
 312
 313 void
 314 vm_object_clear_flag(vm_object_t object, u_short bits)
 315 {
 316
 317         VM_OBJECT_ASSERT_WLOCKED(object);
 318         object->flags &= ~bits;
 319 }
 320
 321 /*
 322  *      Sets the default memory attribute for the specified object.  Pages
 323  *      that are allocated to this object are by default assigned this memory
 324  *      attribute.
 325  *
 326  *      Presently, this function must be called before any pages are allocated
 327  *      to the object.  In the future, this requirement may be relaxed for
 328  *      "default" and "swap" objects.
 329  */
 330 int
 331 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 332 {
 333
 334         VM_OBJECT_ASSERT_WLOCKED(object);
 335         switch (object->type) {
 336         case OBJT_DEFAULT:
 337         case OBJT_DEVICE:
 338         case OBJT_MGTDEVICE:
 339         case OBJT_PHYS:
 340         case OBJT_SG:
 341         case OBJT_SWAP:
 342         case OBJT_VNODE:
 343                 if (!TAILQ_EMPTY(&object->memq))
 344                         return (KERN_FAILURE);
 345                 break;
 346         case OBJT_DEAD:
 347                 return (KERN_INVALID_ARGUMENT);
 348         default:
 349                 panic("vm_object_set_memattr: object %p is of undefined type",
 350                     object);
 351         }
 352         object->memattr = memattr;
 353         return (KERN_SUCCESS);
 354 }
 355
 356 void
 357 vm_object_pip_add(vm_object_t object, short i)
 358 {
 359
 360         VM_OBJECT_ASSERT_WLOCKED(object);
 361         object->paging_in_progress += i;
 362 }
 363
 364 void
 365 vm_object_pip_subtract(vm_object_t object, short i)
 366 {
 367
 368         VM_OBJECT_ASSERT_WLOCKED(object);
 369         object->paging_in_progress -= i;
 370 }
 371
 372 void
 373 vm_object_pip_wakeup(vm_object_t object)
 374 {
 375
 376         VM_OBJECT_ASSERT_WLOCKED(object);
 377         object->paging_in_progress--;
 378         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 379                 vm_object_clear_flag(object, OBJ_PIPWNT);
 380                 wakeup(object);
 381         }
 382 }
 383
 384 void
 385 vm_object_pip_wakeupn(vm_object_t object, short i)
 386 {
 387
 388         VM_OBJECT_ASSERT_WLOCKED(object);
 389         if (i)
 390                 object->paging_in_progress -= i;
 391         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 392                 vm_object_clear_flag(object, OBJ_PIPWNT);
 393                 wakeup(object);
 394         }
 395 }
 396
 397 void
 398 vm_object_pip_wait(vm_object_t object, char *waitid)
 399 {
 400
 401         VM_OBJECT_ASSERT_WLOCKED(object);
 402         while (object->paging_in_progress) {
 403                 object->flags |= OBJ_PIPWNT;
 404                 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
 405         }
 406 }
 407
 408 /*
 409  *      vm_object_allocate:
 410  *
 411  *      Returns a new object with the given size.
 412  */
 413 vm_object_t
 414 vm_object_allocate(objtype_t type, vm_pindex_t size)
 415 {
 416         vm_object_t object;
 417
 418         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 419         _vm_object_allocate(type, size, object);
 420         return (object);
 421 }
 422
 423
 424 /*
 425  *      vm_object_reference:
 426  *
 427  *      Gets another reference to the given object.  Note: OBJ_DEAD
 428  *      objects can be referenced during final cleaning.
 429  */
 430 void
 431 vm_object_reference(vm_object_t object)
 432 {
 433         if (object == NULL)
 434                 return;
 435         VM_OBJECT_WLOCK(object);
 436         vm_object_reference_locked(object);
 437         VM_OBJECT_WUNLOCK(object);
 438 }
 439
 440 /*
 441  *      vm_object_reference_locked:
 442  *
 443  *      Gets another reference to the given object.
 444  *
 445  *      The object must be locked.
 446  */
 447 void
 448 vm_object_reference_locked(vm_object_t object)
 449 {
 450         struct vnode *vp;
 451
 452         VM_OBJECT_ASSERT_WLOCKED(object);
 453         object->ref_count++;
 454         if (object->type == OBJT_VNODE) {
 455                 vp = object->handle;
 456                 vref(vp);
 457         }
 458 }
 459
 460 /*
 461  * Handle deallocating an object of type OBJT_VNODE.
 462  */
 463 static void
 464 vm_object_vndeallocate(vm_object_t object)
 465 {
 466         struct vnode *vp = (struct vnode *) object->handle;
 467
 468         VM_OBJECT_ASSERT_WLOCKED(object);
 469         KASSERT(object->type == OBJT_VNODE,
 470             ("vm_object_vndeallocate: not a vnode object"));
 471         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 472 #ifdef INVARIANTS
 473         if (object->ref_count == 0) {
 474                 vprint("vm_object_vndeallocate", vp);
 475                 panic("vm_object_vndeallocate: bad object reference count");
 476         }
 477 #endif
 478
 479         /*
 480          * The test for text of vp vnode does not need a bypass to
 481          * reach right VV_TEXT there, since it is obtained from
 482          * object->handle.
 483          */
 484         if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
 485                 object->ref_count--;
 486                 VM_OBJECT_WUNLOCK(object);
 487                 /* vrele may need the vnode lock. */
 488                 vrele(vp);
 489         } else {
 490                 vhold(vp);
 491                 VM_OBJECT_WUNLOCK(object);
 492                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 493                 vdrop(vp);
 494                 VM_OBJECT_WLOCK(object);
 495                 object->ref_count--;
 496                 if (object->type == OBJT_DEAD) {
 497                         VM_OBJECT_WUNLOCK(object);
 498                         VOP_UNLOCK(vp, 0);
 499                 } else {
 500                         if (object->ref_count == 0)
 501                                 VOP_UNSET_TEXT(vp);
 502                         VM_OBJECT_WUNLOCK(object);
 503                         vput(vp);
 504                 }
 505         }
 506 }
 507
 508 /*
 509  *      vm_object_deallocate:
 510  *
 511  *      Release a reference to the specified object,
 512  *      gained either through a vm_object_allocate
 513  *      or a vm_object_reference call.  When all references
 514  *      are gone, storage associated with this object
 515  *      may be relinquished.
 516  *
 517  *      No object may be locked.
 518  */
 519 void
 520 vm_object_deallocate(vm_object_t object)
 521 {
 522         vm_object_t temp;
 523         struct vnode *vp;
 524
 525         while (object != NULL) {
 526                 VM_OBJECT_WLOCK(object);
 527                 if (object->type == OBJT_VNODE) {
 528                         vm_object_vndeallocate(object);
 529                         return;
 530                 }
 531
 532                 KASSERT(object->ref_count != 0,
 533                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 534
 535                 /*
 536                  * If the reference count goes to 0 we start calling
 537                  * vm_object_terminate() on the object chain.
 538                  * A ref count of 1 may be a special case depending on the
 539                  * shadow count being 0 or 1.
 540                  */
 541                 object->ref_count--;
 542                 if (object->ref_count > 1) {
 543                         VM_OBJECT_WUNLOCK(object);
 544                         return;
 545                 } else if (object->ref_count == 1) {
 546                         if (object->type == OBJT_SWAP &&
 547                             (object->flags & OBJ_TMPFS) != 0) {
 548                                 vp = object->un_pager.swp.swp_tmpfs;
 549                                 vhold(vp);
 550                                 VM_OBJECT_WUNLOCK(object);
 551                                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 552                                 VM_OBJECT_WLOCK(object);
 553                                 if (object->type == OBJT_DEAD ||
 554                                     object->ref_count != 1) {
 555                                         VM_OBJECT_WUNLOCK(object);
 556                                         VOP_UNLOCK(vp, 0);
 557                                         vdrop(vp);
 558                                         return;
 559                                 }
 560                                 if ((object->flags & OBJ_TMPFS) != 0)
 561                                         VOP_UNSET_TEXT(vp);
 562                                 VOP_UNLOCK(vp, 0);
 563                                 vdrop(vp);
 564                         }
 565                         if (object->shadow_count == 0 &&
 566                             object->handle == NULL &&
 567                             (object->type == OBJT_DEFAULT ||
 568                             (object->type == OBJT_SWAP &&
 569                             (object->flags & OBJ_TMPFS_NODE) == 0))) {
 570                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 571                         } else if ((object->shadow_count == 1) &&
 572                             (object->handle == NULL) &&
 573                             (object->type == OBJT_DEFAULT ||
 574                              object->type == OBJT_SWAP)) {
 575                                 vm_object_t robject;
 576
 577                                 robject = LIST_FIRST(&object->shadow_head);
 578                                 KASSERT(robject != NULL,
 579                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 580                                          object->ref_count,
 581                                          object->shadow_count));
 582                                 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
 583                                     ("shadowed tmpfs v_object %p", object));
 584                                 if (!VM_OBJECT_TRYWLOCK(robject)) {
 585                                         /*
 586                                          * Avoid a potential deadlock.
 587                                          */
 588                                         object->ref_count++;
 589                                         VM_OBJECT_WUNLOCK(object);
 590                                         /*
 591                                          * More likely than not the thread
 592                                          * holding robject's lock has lower
 593                                          * priority than the current thread.
 594                                          * Let the lower priority thread run.
 595                                          */
 596                                         pause("vmo_de", 1);
 597                                         continue;
 598                                 }
 599                                 /*
 600                                  * Collapse object into its shadow unless its
 601                                  * shadow is dead.  In that case, object will
 602                                  * be deallocated by the thread that is
 603                                  * deallocating its shadow.
 604                                  */
 605                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 606                                     (robject->handle == NULL) &&
 607                                     (robject->type == OBJT_DEFAULT ||
 608                                      robject->type == OBJT_SWAP)) {
 609
 610                                         robject->ref_count++;
 611 retry:
 612                                         if (robject->paging_in_progress) {
 613                                                 VM_OBJECT_WUNLOCK(object);
 614                                                 vm_object_pip_wait(robject,
 615                                                     "objde1");
 616                                                 temp = robject->backing_object;
 617                                                 if (object == temp) {
 618                                                         VM_OBJECT_WLOCK(object);
 619                                                         goto retry;
 620                                                 }
 621                                         } else if (object->paging_in_progress) {
 622                                                 VM_OBJECT_WUNLOCK(robject);
 623                                                 object->flags |= OBJ_PIPWNT;
 624                                                 VM_OBJECT_SLEEP(object, object,
 625                                                     PDROP | PVM, "objde2", 0);
 626                                                 VM_OBJECT_WLOCK(robject);
 627                                                 temp = robject->backing_object;
 628                                                 if (object == temp) {
 629                                                         VM_OBJECT_WLOCK(object);
 630                                                         goto retry;
 631                                                 }
 632                                         } else
 633                                                 VM_OBJECT_WUNLOCK(object);
 634
 635                                         if (robject->ref_count == 1) {
 636                                                 robject->ref_count--;
 637                                                 object = robject;
 638                                                 goto doterm;
 639                                         }
 640                                         object = robject;
 641                                         vm_object_collapse(object);
 642                                         VM_OBJECT_WUNLOCK(object);
 643                                         continue;
 644                                 }
 645                                 VM_OBJECT_WUNLOCK(robject);
 646                         }
 647                         VM_OBJECT_WUNLOCK(object);
 648                         return;
 649                 }
 650 doterm:
 651                 temp = object->backing_object;
 652                 if (temp != NULL) {
 653                         KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
 654                             ("shadowed tmpfs v_object 2 %p", object));
 655                         VM_OBJECT_WLOCK(temp);
 656                         LIST_REMOVE(object, shadow_list);
 657                         temp->shadow_count--;
 658                         VM_OBJECT_WUNLOCK(temp);
 659                         object->backing_object = NULL;
 660                 }
 661                 /*
 662                  * Don't double-terminate, we could be in a termination
 663                  * recursion due to the terminate having to sync data
 664                  * to disk.
 665                  */
 666                 if ((object->flags & OBJ_DEAD) == 0)
 667                         vm_object_terminate(object);
 668                 else
 669                         VM_OBJECT_WUNLOCK(object);
 670                 object = temp;
 671         }
 672 }
 673
 674 /*
 675  *      vm_object_destroy removes the object from the global object list
 676  *      and frees the space for the object.
 677  */
 678 void
 679 vm_object_destroy(vm_object_t object)
 680 {
 681
 682         /*
 683          * Release the allocation charge.
 684          */
 685         if (object->cred != NULL) {
 686                 swap_release_by_cred(object->charge, object->cred);
 687                 object->charge = 0;
 688                 crfree(object->cred);
 689                 object->cred = NULL;
 690         }
 691
 692         /*
 693          * Free the space for the object.
 694          */
 695         uma_zfree(obj_zone, object);
 696 }
 697
 698 /*
 699  *      vm_object_terminate actually destroys the specified object, freeing
 700  *      up all previously used resources.
 701  *
 702  *      The object must be locked.
 703  *      This routine may block.
 704  */
 705 void
 706 vm_object_terminate(vm_object_t object)
 707 {
 708         vm_page_t p, p_next;
 709
 710         VM_OBJECT_ASSERT_WLOCKED(object);
 711
 712         /*
 713          * Make sure no one uses us.
 714          */
 715         vm_object_set_flag(object, OBJ_DEAD);
 716
 717         /*
 718          * wait for the pageout daemon to be done with the object
 719          */
 720         vm_object_pip_wait(object, "objtrm");
 721
 722         KASSERT(!object->paging_in_progress,
 723                 ("vm_object_terminate: pageout in progress"));
 724
 725         /*
 726          * Clean and free the pages, as appropriate. All references to the
 727          * object are gone, so we don't need to lock it.
 728          */
 729         if (object->type == OBJT_VNODE) {
 730                 struct vnode *vp = (struct vnode *)object->handle;
 731
 732                 /*
 733                  * Clean pages and flush buffers.
 734                  */
 735                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 736                 VM_OBJECT_WUNLOCK(object);
 737
 738                 vinvalbuf(vp, V_SAVE, 0, 0);
 739
 740                 VM_OBJECT_WLOCK(object);
 741         }
 742
 743         KASSERT(object->ref_count == 0,
 744                 ("vm_object_terminate: object with references, ref_count=%d",
 745                 object->ref_count));
 746
 747         /*
 748          * Free any remaining pageable pages.  This also removes them from the
 749          * paging queues.  However, don't free wired pages, just remove them
 750          * from the object.  Rather than incrementally removing each page from
 751          * the object, the page and object are reset to any empty state.
 752          */
 753         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
 754                 vm_page_assert_unbusied(p);
 755                 vm_page_lock(p);
 756                 /*
 757                  * Optimize the page's removal from the object by resetting
 758                  * its "object" field.  Specifically, if the page is not
 759                  * wired, then the effect of this assignment is that
 760                  * vm_page_free()'s call to vm_page_remove() will return
 761                  * immediately without modifying the page or the object.
 762                  */
 763                 p->object = NULL;
 764                 if (p->wire_count == 0) {
 765                         vm_page_free(p);
 766                         PCPU_INC(cnt.v_pfree);
 767                 }
 768                 vm_page_unlock(p);
 769         }
 770         /*
 771          * If the object contained any pages, then reset it to an empty state.
 772          * None of the object's fields, including "resident_page_count", were
 773          * modified by the preceding loop.
 774          */
 775         if (object->resident_page_count != 0) {
 776                 vm_radix_reclaim_allnodes(&object->rtree);
 777                 TAILQ_INIT(&object->memq);
 778                 object->resident_page_count = 0;
 779                 if (object->type == OBJT_VNODE)
 780                         vdrop(object->handle);
 781         }
 782
 783 #if VM_NRESERVLEVEL > 0
 784         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 785                 vm_reserv_break_all(object);
 786 #endif
 787         if (__predict_false(!vm_object_cache_is_empty(object)))
 788                 vm_page_cache_free(object, 0, 0);
 789
 790         KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
 791             object->type == OBJT_SWAP,
 792             ("%s: non-swap obj %p has cred", __func__, object));
 793
 794         /*
 795          * Let the pager know object is dead.
 796          */
 797         vm_pager_deallocate(object);
 798         VM_OBJECT_WUNLOCK(object);
 799
 800         vm_object_destroy(object);
 801 }
 802
 803 /*
 804  * Make the page read-only so that we can clear the object flags.  However, if
 805  * this is a nosync mmap then the object is likely to stay dirty so do not
 806  * mess with the page and do not clear the object flags.  Returns TRUE if the
 807  * page should be flushed, and FALSE otherwise.
 808  */
 809 static boolean_t
 810 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
 811 {
 812
 813         /*
 814          * If we have been asked to skip nosync pages and this is a
 815          * nosync page, skip it.  Note that the object flags were not
 816          * cleared in this case so we do not have to set them.
 817          */
 818         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 819                 *clearobjflags = FALSE;
 820                 return (FALSE);
 821         } else {
 822                 pmap_remove_write(p);
 823                 return (p->dirty != 0);
 824         }
 825 }
 826
 827 /*
 828  *      vm_object_page_clean
 829  *
 830  *      Clean all dirty pages in the specified range of object.  Leaves page
 831  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 832  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 833  *      leaving the object dirty.
 834  *
 835  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 836  *      synchronous clustering mode implementation.
 837  *
 838  *      Odd semantics: if start == end, we clean everything.
 839  *
 840  *      The object must be locked.
 841  *
 842  *      Returns FALSE if some page from the range was not written, as
 843  *      reported by the pager, and TRUE otherwise.
 844  */
 845 boolean_t
 846 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 847     int flags)
 848 {
 849         vm_page_t np, p;
 850         vm_pindex_t pi, tend, tstart;
 851         int curgeneration, n, pagerflags;
 852         boolean_t clearobjflags, eio, res;
 853
 854         VM_OBJECT_ASSERT_WLOCKED(object);
 855
 856         /*
 857          * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
 858          * objects.  The check below prevents the function from
 859          * operating on non-vnode objects.
 860          */
 861         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 862             object->resident_page_count == 0)
 863                 return (TRUE);
 864
 865         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 866             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 867         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 868
 869         tstart = OFF_TO_IDX(start);
 870         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 871         clearobjflags = tstart == 0 && tend >= object->size;
 872         res = TRUE;
 873
 874 rescan:
 875         curgeneration = object->generation;
 876
 877         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 878                 pi = p->pindex;
 879                 if (pi >= tend)
 880                         break;
 881                 np = TAILQ_NEXT(p, listq);
 882                 if (p->valid == 0)
 883                         continue;
 884                 if (vm_page_sleep_if_busy(p, "vpcwai")) {
 885                         if (object->generation != curgeneration) {
 886                                 if ((flags & OBJPC_SYNC) != 0)
 887                                         goto rescan;
 888                                 else
 889                                         clearobjflags = FALSE;
 890                         }
 891                         np = vm_page_find_least(object, pi);
 892                         continue;
 893                 }
 894                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 895                         continue;
 896
 897                 n = vm_object_page_collect_flush(object, p, pagerflags,
 898                     flags, &clearobjflags, &eio);
 899                 if (eio) {
 900                         res = FALSE;
 901                         clearobjflags = FALSE;
 902                 }
 903                 if (object->generation != curgeneration) {
 904                         if ((flags & OBJPC_SYNC) != 0)
 905                                 goto rescan;
 906                         else
 907                                 clearobjflags = FALSE;
 908                 }
 909
 910                 /*
 911                  * If the VOP_PUTPAGES() did a truncated write, so
 912                  * that even the first page of the run is not fully
 913                  * written, vm_pageout_flush() returns 0 as the run
 914                  * length.  Since the condition that caused truncated
 915                  * write may be permanent, e.g. exhausted free space,
 916                  * accepting n == 0 would cause an infinite loop.
 917                  *
 918                  * Forwarding the iterator leaves the unwritten page
 919                  * behind, but there is not much we can do there if
 920                  * filesystem refuses to write it.
 921                  */
 922                 if (n == 0) {
 923                         n = 1;
 924                         clearobjflags = FALSE;
 925                 }
 926                 np = vm_page_find_least(object, pi + n);
 927         }
 928 #if 0
 929         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 930 #endif
 931
 932         if (clearobjflags)
 933                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 934         return (res);
 935 }
 936
 937 static int
 938 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 939     int flags, boolean_t *clearobjflags, boolean_t *eio)
 940 {
 941         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 942         int count, i, mreq, runlen;
 943
 944         vm_page_lock_assert(p, MA_NOTOWNED);
 945         VM_OBJECT_ASSERT_WLOCKED(object);
 946
 947         count = 1;
 948         mreq = 0;
 949
 950         for (tp = p; count < vm_pageout_page_count; count++) {
 951                 tp = vm_page_next(tp);
 952                 if (tp == NULL || vm_page_busied(tp))
 953                         break;
 954                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 955                         break;
 956         }
 957
 958         for (p_first = p; count < vm_pageout_page_count; count++) {
 959                 tp = vm_page_prev(p_first);
 960                 if (tp == NULL || vm_page_busied(tp))
 961                         break;
 962                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 963                         break;
 964                 p_first = tp;
 965                 mreq++;
 966         }
 967
 968         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 969                 ma[i] = tp;
 970
 971         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
 972         return (runlen);
 973 }
 974
 975 /*
 976  * Note that there is absolutely no sense in writing out
 977  * anonymous objects, so we track down the vnode object
 978  * to write out.
 979  * We invalidate (remove) all pages from the address space
 980  * for semantic correctness.
 981  *
 982  * If the backing object is a device object with unmanaged pages, then any
 983  * mappings to the specified range of pages must be removed before this
 984  * function is called.
 985  *
 986  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 987  * may start out with a NULL object.
 988  */
 989 boolean_t
 990 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 991     boolean_t syncio, boolean_t invalidate)
 992 {
 993         vm_object_t backing_object;
 994         struct vnode *vp;
 995         struct mount *mp;
 996         int error, flags, fsync_after;
 997         boolean_t res;
 998
 999         if (object == NULL)
1000                 return (TRUE);
1001         res = TRUE;
1002         error = 0;
1003         VM_OBJECT_WLOCK(object);
1004         while ((backing_object = object->backing_object) != NULL) {
1005                 VM_OBJECT_WLOCK(backing_object);
1006                 offset += object->backing_object_offset;
1007                 VM_OBJECT_WUNLOCK(object);
1008                 object = backing_object;
1009                 if (object->size < OFF_TO_IDX(offset + size))
1010                         size = IDX_TO_OFF(object->size) - offset;
1011         }
1012         /*
1013          * Flush pages if writing is allowed, invalidate them
1014          * if invalidation requested.  Pages undergoing I/O
1015          * will be ignored by vm_object_page_remove().
1016          *
1017          * We cannot lock the vnode and then wait for paging
1018          * to complete without deadlocking against vm_fault.
1019          * Instead we simply call vm_object_page_remove() and
1020          * allow it to block internally on a page-by-page
1021          * basis when it encounters pages undergoing async
1022          * I/O.
1023          */
1024         if (object->type == OBJT_VNODE &&
1025             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1026                 vp = object->handle;
1027                 VM_OBJECT_WUNLOCK(object);
1028                 (void) vn_start_write(vp, &mp, V_WAIT);
1029                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1030                 if (syncio && !invalidate && offset == 0 &&
1031                     OFF_TO_IDX(size) == object->size) {
1032                         /*
1033                          * If syncing the whole mapping of the file,
1034                          * it is faster to schedule all the writes in
1035                          * async mode, also allowing the clustering,
1036                          * and then wait for i/o to complete.
1037                          */
1038                         flags = 0;
1039                         fsync_after = TRUE;
1040                 } else {
1041                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1042                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1043                         fsync_after = FALSE;
1044                 }
1045                 VM_OBJECT_WLOCK(object);
1046                 res = vm_object_page_clean(object, offset, offset + size,
1047                     flags);
1048                 VM_OBJECT_WUNLOCK(object);
1049                 if (fsync_after)
1050                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1051                 VOP_UNLOCK(vp, 0);
1052                 vn_finished_write(mp);
1053                 if (error != 0)
1054                         res = FALSE;
1055                 VM_OBJECT_WLOCK(object);
1056         }
1057         if ((object->type == OBJT_VNODE ||
1058              object->type == OBJT_DEVICE) && invalidate) {
1059                 if (object->type == OBJT_DEVICE)
1060                         /*
1061                          * The option OBJPR_NOTMAPPED must be passed here
1062                          * because vm_object_page_remove() cannot remove
1063                          * unmanaged mappings.
1064                          */
1065                         flags = OBJPR_NOTMAPPED;
1066                 else if (old_msync)
1067                         flags = 0;
1068                 else
1069                         flags = OBJPR_CLEANONLY;
1070                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1071                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1072         }
1073         VM_OBJECT_WUNLOCK(object);
1074         return (res);
1075 }
1076
1077 /*
1078  *      vm_object_madvise:
1079  *
1080  *      Implements the madvise function at the object/page level.
1081  *
1082  *      MADV_WILLNEED   (any object)
1083  *
1084  *          Activate the specified pages if they are resident.
1085  *
1086  *      MADV_DONTNEED   (any object)
1087  *
1088  *          Deactivate the specified pages if they are resident.
1089  *
1090  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1091  *                       OBJ_ONEMAPPING only)
1092  *
1093  *          Deactivate and clean the specified pages if they are
1094  *          resident.  This permits the process to reuse the pages
1095  *          without faulting or the kernel to reclaim the pages
1096  *          without I/O.
1097  */
1098 void
1099 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1100     int advise)
1101 {
1102         vm_pindex_t tpindex;
1103         vm_object_t backing_object, tobject;
1104         vm_page_t m;
1105
1106         if (object == NULL)
1107                 return;
1108         VM_OBJECT_WLOCK(object);
1109         /*
1110          * Locate and adjust resident pages
1111          */
1112         for (; pindex < end; pindex += 1) {
1113 relookup:
1114                 tobject = object;
1115                 tpindex = pindex;
1116 shadowlookup:
1117                 /*
1118                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1119                  * and those pages must be OBJ_ONEMAPPING.
1120                  */
1121                 if (advise == MADV_FREE) {
1122                         if ((tobject->type != OBJT_DEFAULT &&
1123                              tobject->type != OBJT_SWAP) ||
1124                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1125                                 goto unlock_tobject;
1126                         }
1127                 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1128                         goto unlock_tobject;
1129                 m = vm_page_lookup(tobject, tpindex);
1130                 if (m == NULL && advise == MADV_WILLNEED) {
1131                         /*
1132                          * If the page is cached, reactivate it.
1133                          */
1134                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1135                             VM_ALLOC_NOBUSY);
1136                 }
1137                 if (m == NULL) {
1138                         /*
1139                          * There may be swap even if there is no backing page
1140                          */
1141                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1142                                 swap_pager_freespace(tobject, tpindex, 1);
1143                         /*
1144                          * next object
1145                          */
1146                         backing_object = tobject->backing_object;
1147                         if (backing_object == NULL)
1148                                 goto unlock_tobject;
1149                         VM_OBJECT_WLOCK(backing_object);
1150                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1151                         if (tobject != object)
1152                                 VM_OBJECT_WUNLOCK(tobject);
1153                         tobject = backing_object;
1154                         goto shadowlookup;
1155                 } else if (m->valid != VM_PAGE_BITS_ALL)
1156                         goto unlock_tobject;
1157                 /*
1158                  * If the page is not in a normal state, skip it.
1159                  */
1160                 vm_page_lock(m);
1161                 if (m->hold_count != 0 || m->wire_count != 0) {
1162                         vm_page_unlock(m);
1163                         goto unlock_tobject;
1164                 }
1165                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1166                     ("vm_object_madvise: page %p is fictitious", m));
1167                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1168                     ("vm_object_madvise: page %p is not managed", m));
1169                 if (vm_page_busied(m)) {
1170                         if (advise == MADV_WILLNEED) {
1171                                 /*
1172                                  * Reference the page before unlocking and
1173                                  * sleeping so that the page daemon is less
1174                                  * likely to reclaim it.
1175                                  */
1176                                 vm_page_aflag_set(m, PGA_REFERENCED);
1177                         }
1178                         if (object != tobject)
1179                                 VM_OBJECT_WUNLOCK(object);
1180                         VM_OBJECT_WUNLOCK(tobject);
1181                         vm_page_busy_sleep(m, "madvpo");
1182                         VM_OBJECT_WLOCK(object);
1183                         goto relookup;
1184                 }
1185                 if (advise == MADV_WILLNEED) {
1186                         vm_page_activate(m);
1187                 } else {
1188                         vm_page_advise(m, advise);
1189                 }
1190                 vm_page_unlock(m);
1191                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1192                         swap_pager_freespace(tobject, tpindex, 1);
1193 unlock_tobject:
1194                 if (tobject != object)
1195                         VM_OBJECT_WUNLOCK(tobject);
1196         }
1197         VM_OBJECT_WUNLOCK(object);
1198 }
1199
1200 /*
1201  *      vm_object_shadow:
1202  *
1203  *      Create a new object which is backed by the
1204  *      specified existing object range.  The source
1205  *      object reference is deallocated.
1206  *
1207  *      The new object and offset into that object
1208  *      are returned in the source parameters.
1209  */
1210 void
1211 vm_object_shadow(
1212         vm_object_t *object,    /* IN/OUT */
1213         vm_ooffset_t *offset,   /* IN/OUT */
1214         vm_size_t length)
1215 {
1216         vm_object_t source;
1217         vm_object_t result;
1218
1219         source = *object;
1220
1221         /*
1222          * Don't create the new object if the old object isn't shared.
1223          */
1224         if (source != NULL) {
1225                 VM_OBJECT_WLOCK(source);
1226                 if (source->ref_count == 1 &&
1227                     source->handle == NULL &&
1228                     (source->type == OBJT_DEFAULT ||
1229                      source->type == OBJT_SWAP)) {
1230                         VM_OBJECT_WUNLOCK(source);
1231                         return;
1232                 }
1233                 VM_OBJECT_WUNLOCK(source);
1234         }
1235
1236         /*
1237          * Allocate a new object with the given length.
1238          */
1239         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1240
1241         /*
1242          * The new object shadows the source object, adding a reference to it.
1243          * Our caller changes his reference to point to the new object,
1244          * removing a reference to the source object.  Net result: no change
1245          * of reference count.
1246          *
1247          * Try to optimize the result object's page color when shadowing
1248          * in order to maintain page coloring consistency in the combined
1249          * shadowed object.
1250          */
1251         result->backing_object = source;
1252         /*
1253          * Store the offset into the source object, and fix up the offset into
1254          * the new object.
1255          */
1256         result->backing_object_offset = *offset;
1257         if (source != NULL) {
1258                 VM_OBJECT_WLOCK(source);
1259                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1260                 source->shadow_count++;
1261 #if VM_NRESERVLEVEL > 0
1262                 result->flags |= source->flags & OBJ_COLORED;
1263                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1264                     ((1 << (VM_NFREEORDER - 1)) - 1);
1265 #endif
1266                 VM_OBJECT_WUNLOCK(source);
1267         }
1268
1269
1270         /*
1271          * Return the new things
1272          */
1273         *offset = 0;
1274         *object = result;
1275 }
1276
1277 /*
1278  *      vm_object_split:
1279  *
1280  * Split the pages in a map entry into a new object.  This affords
1281  * easier removal of unused pages, and keeps object inheritance from
1282  * being a negative impact on memory usage.
1283  */
1284 void
1285 vm_object_split(vm_map_entry_t entry)
1286 {
1287         vm_page_t m, m_next;
1288         vm_object_t orig_object, new_object, source;
1289         vm_pindex_t idx, offidxstart;
1290         vm_size_t size;
1291
1292         orig_object = entry->object.vm_object;
1293         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1294                 return;
1295         if (orig_object->ref_count <= 1)
1296                 return;
1297         VM_OBJECT_WUNLOCK(orig_object);
1298
1299         offidxstart = OFF_TO_IDX(entry->offset);
1300         size = atop(entry->end - entry->start);
1301
1302         /*
1303          * If swap_pager_copy() is later called, it will convert new_object
1304          * into a swap object.
1305          */
1306         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1307
1308         /*
1309          * At this point, the new object is still private, so the order in
1310          * which the original and new objects are locked does not matter.
1311          */
1312         VM_OBJECT_WLOCK(new_object);
1313         VM_OBJECT_WLOCK(orig_object);
1314         source = orig_object->backing_object;
1315         if (source != NULL) {
1316                 VM_OBJECT_WLOCK(source);
1317                 if ((source->flags & OBJ_DEAD) != 0) {
1318                         VM_OBJECT_WUNLOCK(source);
1319                         VM_OBJECT_WUNLOCK(orig_object);
1320                         VM_OBJECT_WUNLOCK(new_object);
1321                         vm_object_deallocate(new_object);
1322                         VM_OBJECT_WLOCK(orig_object);
1323                         return;
1324                 }
1325                 LIST_INSERT_HEAD(&source->shadow_head,
1326                                   new_object, shadow_list);
1327                 source->shadow_count++;
1328                 vm_object_reference_locked(source);     /* for new_object */
1329                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1330                 VM_OBJECT_WUNLOCK(source);
1331                 new_object->backing_object_offset =
1332                         orig_object->backing_object_offset + entry->offset;
1333                 new_object->backing_object = source;
1334         }
1335         if (orig_object->cred != NULL) {
1336                 new_object->cred = orig_object->cred;
1337                 crhold(orig_object->cred);
1338                 new_object->charge = ptoa(size);
1339                 KASSERT(orig_object->charge >= ptoa(size),
1340                     ("orig_object->charge < 0"));
1341                 orig_object->charge -= ptoa(size);
1342         }
1343 retry:
1344         m = vm_page_find_least(orig_object, offidxstart);
1345         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1346             m = m_next) {
1347                 m_next = TAILQ_NEXT(m, listq);
1348
1349                 /*
1350                  * We must wait for pending I/O to complete before we can
1351                  * rename the page.
1352                  *
1353                  * We do not have to VM_PROT_NONE the page as mappings should
1354                  * not be changed by this operation.
1355                  */
1356                 if (vm_page_busied(m)) {
1357                         VM_OBJECT_WUNLOCK(new_object);
1358                         vm_page_lock(m);
1359                         VM_OBJECT_WUNLOCK(orig_object);
1360                         vm_page_busy_sleep(m, "spltwt");
1361                         VM_OBJECT_WLOCK(orig_object);
1362                         VM_OBJECT_WLOCK(new_object);
1363                         goto retry;
1364                 }
1365
1366                 /* vm_page_rename() will handle dirty and cache. */
1367                 if (vm_page_rename(m, new_object, idx)) {
1368                         VM_OBJECT_WUNLOCK(new_object);
1369                         VM_OBJECT_WUNLOCK(orig_object);
1370                         VM_WAIT;
1371                         VM_OBJECT_WLOCK(orig_object);
1372                         VM_OBJECT_WLOCK(new_object);
1373                         goto retry;
1374                 }
1375 #if VM_NRESERVLEVEL > 0
1376                 /*
1377                  * If some of the reservation's allocated pages remain with
1378                  * the original object, then transferring the reservation to
1379                  * the new object is neither particularly beneficial nor
1380                  * particularly harmful as compared to leaving the reservation
1381                  * with the original object.  If, however, all of the
1382                  * reservation's allocated pages are transferred to the new
1383                  * object, then transferring the reservation is typically
1384                  * beneficial.  Determining which of these two cases applies
1385                  * would be more costly than unconditionally renaming the
1386                  * reservation.
1387                  */
1388                 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1389 #endif
1390                 if (orig_object->type == OBJT_SWAP)
1391                         vm_page_xbusy(m);
1392         }
1393         if (orig_object->type == OBJT_SWAP) {
1394                 /*
1395                  * swap_pager_copy() can sleep, in which case the orig_object's
1396                  * and new_object's locks are released and reacquired.
1397                  */
1398                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1399                 TAILQ_FOREACH(m, &new_object->memq, listq)
1400                         vm_page_xunbusy(m);
1401
1402                 /*
1403                  * Transfer any cached pages from orig_object to new_object.
1404                  * If swap_pager_copy() found swapped out pages within the
1405                  * specified range of orig_object, then it changed
1406                  * new_object's type to OBJT_SWAP when it transferred those
1407                  * pages to new_object.  Otherwise, new_object's type
1408                  * should still be OBJT_DEFAULT and orig_object should not
1409                  * contain any cached pages within the specified range.
1410                  */
1411                 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1412                         vm_page_cache_transfer(orig_object, offidxstart,
1413                             new_object);
1414         }
1415         VM_OBJECT_WUNLOCK(orig_object);
1416         VM_OBJECT_WUNLOCK(new_object);
1417         entry->object.vm_object = new_object;
1418         entry->offset = 0LL;
1419         vm_object_deallocate(orig_object);
1420         VM_OBJECT_WLOCK(new_object);
1421 }
1422
1423 #define OBSC_TEST_ALL_SHADOWED  0x0001
1424 #define OBSC_COLLAPSE_NOWAIT    0x0002
1425 #define OBSC_COLLAPSE_WAIT      0x0004
1426
1427 static vm_page_t
1428 vm_object_backing_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1429     int op)
1430 {
1431         vm_object_t backing_object;
1432
1433         VM_OBJECT_ASSERT_WLOCKED(object);
1434         backing_object = object->backing_object;
1435         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1436
1437         KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1438         KASSERT(p == NULL || p->object == object || p->object == backing_object,
1439             ("invalid ownership %p %p %p", p, object, backing_object));
1440         if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1441                 return (next);
1442         if (p != NULL)
1443                 vm_page_lock(p);
1444         VM_OBJECT_WUNLOCK(object);
1445         VM_OBJECT_WUNLOCK(backing_object);
1446         if (p == NULL)
1447                 VM_WAIT;
1448         else
1449                 vm_page_busy_sleep(p, "vmocol");
1450         VM_OBJECT_WLOCK(object);
1451         VM_OBJECT_WLOCK(backing_object);
1452         return (TAILQ_FIRST(&backing_object->memq));
1453 }
1454
1455 static bool
1456 vm_object_backing_scan(vm_object_t object, int op)
1457 {
1458         vm_object_t backing_object;
1459         vm_page_t next, p, pp;
1460         vm_pindex_t backing_offset_index, new_pindex;
1461
1462         VM_OBJECT_ASSERT_WLOCKED(object);
1463         VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1464
1465         backing_object = object->backing_object;
1466         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1467
1468         /*
1469          * Initial conditions
1470          */
1471         if (op & OBSC_TEST_ALL_SHADOWED) {
1472                 /*
1473                  * We do not want to have to test for the existence of cache
1474                  * or swap pages in the backing object.  XXX but with the
1475                  * new swapper this would be pretty easy to do.
1476                  *
1477                  * XXX what about anonymous MAP_SHARED memory that hasn't
1478                  * been ZFOD faulted yet?  If we do not test for this, the
1479                  * shadow test may succeed! XXX
1480                  */
1481                 if (backing_object->type != OBJT_DEFAULT) {
1482                         return (false);
1483                 }
1484         }
1485         if (op & OBSC_COLLAPSE_WAIT) {
1486                 vm_object_set_flag(backing_object, OBJ_DEAD);
1487         }
1488
1489         /*
1490          * Our scan
1491          */
1492         p = TAILQ_FIRST(&backing_object->memq);
1493         while (p) {
1494                 next = TAILQ_NEXT(p, listq);
1495                 new_pindex = p->pindex - backing_offset_index;
1496                 if (op & OBSC_TEST_ALL_SHADOWED) {
1497                         /*
1498                          * Ignore pages outside the parent object's range
1499                          * and outside the parent object's mapping of the
1500                          * backing object.
1501                          *
1502                          * Note that we do not busy the backing object's
1503                          * page.
1504                          */
1505                         if (p->pindex < backing_offset_index ||
1506                             new_pindex >= object->size) {
1507                                 p = next;
1508                                 continue;
1509                         }
1510
1511                         /*
1512                          * See if the parent has the page or if the parent's
1513                          * object pager has the page.  If the parent has the
1514                          * page but the page is not valid, the parent's
1515                          * object pager must have the page.
1516                          *
1517                          * If this fails, the parent does not completely shadow
1518                          * the object and we might as well give up now.
1519                          */
1520
1521                         pp = vm_page_lookup(object, new_pindex);
1522                         if ((pp == NULL || pp->valid == 0) &&
1523                             !vm_pager_has_page(object, new_pindex, NULL, NULL))
1524                                 return (false);
1525                 }
1526
1527                 /*
1528                  * Check for busy page
1529                  */
1530                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1531                         if (vm_page_busied(p)) {
1532                                 p = vm_object_backing_scan_wait(object, p,
1533                                     next, op);
1534                                 continue;
1535                         }
1536
1537                         KASSERT(p->object == backing_object,
1538                             ("vm_object_backing_scan: object mismatch"));
1539
1540                         if (p->pindex < backing_offset_index ||
1541                             new_pindex >= object->size) {
1542                                 if (backing_object->type == OBJT_SWAP)
1543                                         swap_pager_freespace(backing_object,
1544                                             p->pindex, 1);
1545
1546                                 /*
1547                                  * Page is out of the parent object's range, we
1548                                  * can simply destroy it.
1549                                  */
1550                                 vm_page_lock(p);
1551                                 KASSERT(!pmap_page_is_mapped(p),
1552                                     ("freeing mapped page %p", p));
1553                                 if (p->wire_count == 0)
1554                                         vm_page_free(p);
1555                                 else
1556                                         vm_page_remove(p);
1557                                 vm_page_unlock(p);
1558                                 p = next;
1559                                 continue;
1560                         }
1561
1562                         pp = vm_page_lookup(object, new_pindex);
1563                         if (pp != NULL && vm_page_busied(pp)) {
1564                                 /*
1565                                  * The page in the parent is busy and
1566                                  * possibly not (yet) valid.  Until
1567                                  * its state is finalized by the busy
1568                                  * bit owner, we can't tell whether it
1569                                  * shadows the original page.
1570                                  * Therefore, we must either skip it
1571                                  * and the original (backing_object)
1572                                  * page or wait for its state to be
1573                                  * finalized.
1574                                  *
1575                                  * This is due to a race with vm_fault()
1576                                  * where we must unbusy the original
1577                                  * (backing_obj) page before we can
1578                                  * (re)lock the parent.  Hence we can
1579                                  * get here.
1580                                  */
1581                                 p = vm_object_backing_scan_wait(object, pp,
1582                                     next, op);
1583                                 continue;
1584                         }
1585
1586                         KASSERT(pp == NULL || pp->valid != 0,
1587                             ("unbusy invalid page %p", pp));
1588
1589                         if (pp != NULL || vm_pager_has_page(object,
1590                             new_pindex, NULL, NULL)) {
1591                                 /*
1592                                  * The page already exists in the
1593                                  * parent OR swap exists for this
1594                                  * location in the parent.  Leave the
1595                                  * parent's page alone.  Destroy the
1596                                  * original page from the backing
1597                                  * object.
1598                                  */
1599                                 if (backing_object->type == OBJT_SWAP)
1600                                         swap_pager_freespace(backing_object,
1601                                             p->pindex, 1);
1602                                 vm_page_lock(p);
1603                                 KASSERT(!pmap_page_is_mapped(p),
1604                                     ("freeing mapped page %p", p));
1605                                 if (p->wire_count == 0)
1606                                         vm_page_free(p);
1607                                 else
1608                                         vm_page_remove(p);
1609                                 vm_page_unlock(p);
1610                                 p = next;
1611                                 continue;
1612                         }
1613
1614                         /*
1615                          * Page does not exist in parent, rename the
1616                          * page from the backing object to the main object.
1617                          *
1618                          * If the page was mapped to a process, it can remain
1619                          * mapped through the rename.
1620                          * vm_page_rename() will handle dirty and cache.
1621                          */
1622                         if (vm_page_rename(p, object, new_pindex)) {
1623                                 p = vm_object_backing_scan_wait(object, NULL,
1624                                     next, op);
1625                                 continue;
1626                         }
1627
1628                         /* Use the old pindex to free the right page. */
1629                         if (backing_object->type == OBJT_SWAP)
1630                                 swap_pager_freespace(backing_object,
1631                                     new_pindex + backing_offset_index, 1);
1632
1633 #if VM_NRESERVLEVEL > 0
1634                         /*
1635                          * Rename the reservation.
1636                          */
1637                         vm_reserv_rename(p, object, backing_object,
1638                             backing_offset_index);
1639 #endif
1640                 }
1641                 p = next;
1642         }
1643         return (true);
1644 }
1645
1646
1647 /*
1648  * this version of collapse allows the operation to occur earlier and
1649  * when paging_in_progress is true for an object...  This is not a complete
1650  * operation, but should plug 99.9% of the rest of the leaks.
1651  */
1652 static void
1653 vm_object_qcollapse(vm_object_t object)
1654 {
1655         vm_object_t backing_object = object->backing_object;
1656
1657         VM_OBJECT_ASSERT_WLOCKED(object);
1658         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1659
1660         if (backing_object->ref_count != 1)
1661                 return;
1662
1663         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1664 }
1665
1666 /*
1667  *      vm_object_collapse:
1668  *
1669  *      Collapse an object with the object backing it.
1670  *      Pages in the backing object are moved into the
1671  *      parent, and the backing object is deallocated.
1672  */
1673 void
1674 vm_object_collapse(vm_object_t object)
1675 {
1676         VM_OBJECT_ASSERT_WLOCKED(object);
1677
1678         while (TRUE) {
1679                 vm_object_t backing_object;
1680
1681                 /*
1682                  * Verify that the conditions are right for collapse:
1683                  *
1684                  * The object exists and the backing object exists.
1685                  */
1686                 if ((backing_object = object->backing_object) == NULL)
1687                         break;
1688
1689                 /*
1690                  * we check the backing object first, because it is most likely
1691                  * not collapsable.
1692                  */
1693                 VM_OBJECT_WLOCK(backing_object);
1694                 if (backing_object->handle != NULL ||
1695                     (backing_object->type != OBJT_DEFAULT &&
1696                      backing_object->type != OBJT_SWAP) ||
1697                     (backing_object->flags & OBJ_DEAD) ||
1698                     object->handle != NULL ||
1699                     (object->type != OBJT_DEFAULT &&
1700                      object->type != OBJT_SWAP) ||
1701                     (object->flags & OBJ_DEAD)) {
1702                         VM_OBJECT_WUNLOCK(backing_object);
1703                         break;
1704                 }
1705
1706                 if (
1707                     object->paging_in_progress != 0 ||
1708                     backing_object->paging_in_progress != 0
1709                 ) {
1710                         vm_object_qcollapse(object);
1711                         VM_OBJECT_WUNLOCK(backing_object);
1712                         break;
1713                 }
1714                 /*
1715                  * We know that we can either collapse the backing object (if
1716                  * the parent is the only reference to it) or (perhaps) have
1717                  * the parent bypass the object if the parent happens to shadow
1718                  * all the resident pages in the entire backing object.
1719                  *
1720                  * This is ignoring pager-backed pages such as swap pages.
1721                  * vm_object_backing_scan fails the shadowing test in this
1722                  * case.
1723                  */
1724                 if (backing_object->ref_count == 1) {
1725                         /*
1726                          * If there is exactly one reference to the backing
1727                          * object, we can collapse it into the parent.
1728                          */
1729                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1730
1731 #if VM_NRESERVLEVEL > 0
1732                         /*
1733                          * Break any reservations from backing_object.
1734                          */
1735                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1736                                 vm_reserv_break_all(backing_object);
1737 #endif
1738
1739                         /*
1740                          * Move the pager from backing_object to object.
1741                          */
1742                         if (backing_object->type == OBJT_SWAP) {
1743                                 /*
1744                                  * swap_pager_copy() can sleep, in which case
1745                                  * the backing_object's and object's locks are
1746                                  * released and reacquired.
1747                                  * Since swap_pager_copy() is being asked to
1748                                  * destroy the source, it will change the
1749                                  * backing_object's type to OBJT_DEFAULT.
1750                                  */
1751                                 swap_pager_copy(
1752                                     backing_object,
1753                                     object,
1754                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1755
1756                                 /*
1757                                  * Free any cached pages from backing_object.
1758                                  */
1759                                 if (__predict_false(
1760                                     !vm_object_cache_is_empty(backing_object)))
1761                                         vm_page_cache_free(backing_object, 0, 0);
1762                         }
1763                         /*
1764                          * Object now shadows whatever backing_object did.
1765                          * Note that the reference to
1766                          * backing_object->backing_object moves from within
1767                          * backing_object to within object.
1768                          */
1769                         LIST_REMOVE(object, shadow_list);
1770                         backing_object->shadow_count--;
1771                         if (backing_object->backing_object) {
1772                                 VM_OBJECT_WLOCK(backing_object->backing_object);
1773                                 LIST_REMOVE(backing_object, shadow_list);
1774                                 LIST_INSERT_HEAD(
1775                                     &backing_object->backing_object->shadow_head,
1776                                     object, shadow_list);
1777                                 /*
1778                                  * The shadow_count has not changed.
1779                                  */
1780                                 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1781                         }
1782                         object->backing_object = backing_object->backing_object;
1783                         object->backing_object_offset +=
1784                             backing_object->backing_object_offset;
1785
1786                         /*
1787                          * Discard backing_object.
1788                          *
1789                          * Since the backing object has no pages, no pager left,
1790                          * and no object references within it, all that is
1791                          * necessary is to dispose of it.
1792                          */
1793                         KASSERT(backing_object->ref_count == 1, (
1794 "backing_object %p was somehow re-referenced during collapse!",
1795                             backing_object));
1796                         backing_object->type = OBJT_DEAD;
1797                         backing_object->ref_count = 0;
1798                         VM_OBJECT_WUNLOCK(backing_object);
1799                         vm_object_destroy(backing_object);
1800
1801                         object_collapses++;
1802                 } else {
1803                         vm_object_t new_backing_object;
1804
1805                         /*
1806                          * If we do not entirely shadow the backing object,
1807                          * there is nothing we can do so we give up.
1808                          */
1809                         if (object->resident_page_count != object->size &&
1810                             !vm_object_backing_scan(object,
1811                             OBSC_TEST_ALL_SHADOWED)) {
1812                                 VM_OBJECT_WUNLOCK(backing_object);
1813                                 break;
1814                         }
1815
1816                         /*
1817                          * Make the parent shadow the next object in the
1818                          * chain.  Deallocating backing_object will not remove
1819                          * it, since its reference count is at least 2.
1820                          */
1821                         LIST_REMOVE(object, shadow_list);
1822                         backing_object->shadow_count--;
1823
1824                         new_backing_object = backing_object->backing_object;
1825                         if ((object->backing_object = new_backing_object) != NULL) {
1826                                 VM_OBJECT_WLOCK(new_backing_object);
1827                                 LIST_INSERT_HEAD(
1828                                     &new_backing_object->shadow_head,
1829                                     object,
1830                                     shadow_list
1831                                 );
1832                                 new_backing_object->shadow_count++;
1833                                 vm_object_reference_locked(new_backing_object);
1834                                 VM_OBJECT_WUNLOCK(new_backing_object);
1835                                 object->backing_object_offset +=
1836                                         backing_object->backing_object_offset;
1837                         }
1838
1839                         /*
1840                          * Drop the reference count on backing_object. Since
1841                          * its ref_count was at least 2, it will not vanish.
1842                          */
1843                         backing_object->ref_count--;
1844                         VM_OBJECT_WUNLOCK(backing_object);
1845                         object_bypasses++;
1846                 }
1847
1848                 /*
1849                  * Try again with this object's new backing object.
1850                  */
1851         }
1852 }
1853
1854 /*
1855  *      vm_object_page_remove:
1856  *
1857  *      For the given object, either frees or invalidates each of the
1858  *      specified pages.  In general, a page is freed.  However, if a page is
1859  *      wired for any reason other than the existence of a managed, wired
1860  *      mapping, then it may be invalidated but not removed from the object.
1861  *      Pages are specified by the given range ["start", "end") and the option
1862  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1863  *      extends from "start" to the end of the object.  If the option
1864  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1865  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1866  *      specified, then the pages within the specified range must have no
1867  *      mappings.  Otherwise, if this option is not specified, any mappings to
1868  *      the specified pages are removed before the pages are freed or
1869  *      invalidated.
1870  *
1871  *      In general, this operation should only be performed on objects that
1872  *      contain managed pages.  There are, however, two exceptions.  First, it
1873  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1874  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1875  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1876  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1877  *
1878  *      The object must be locked.
1879  */
1880 void
1881 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1882     int options)
1883 {
1884         vm_page_t p, next;
1885
1886         VM_OBJECT_ASSERT_WLOCKED(object);
1887         KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1888             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1889             ("vm_object_page_remove: illegal options for object %p", object));
1890         if (object->resident_page_count == 0)
1891                 goto skipmemq;
1892         vm_object_pip_add(object, 1);
1893 again:
1894         p = vm_page_find_least(object, start);
1895
1896         /*
1897          * Here, the variable "p" is either (1) the page with the least pindex
1898          * greater than or equal to the parameter "start" or (2) NULL.
1899          */
1900         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1901                 next = TAILQ_NEXT(p, listq);
1902
1903                 /*
1904                  * If the page is wired for any reason besides the existence
1905                  * of managed, wired mappings, then it cannot be freed.  For
1906                  * example, fictitious pages, which represent device memory,
1907                  * are inherently wired and cannot be freed.  They can,
1908                  * however, be invalidated if the option OBJPR_CLEANONLY is
1909                  * not specified.
1910                  */
1911                 vm_page_lock(p);
1912                 if (vm_page_xbusied(p)) {
1913                         VM_OBJECT_WUNLOCK(object);
1914                         vm_page_busy_sleep(p, "vmopax");
1915                         VM_OBJECT_WLOCK(object);
1916                         goto again;
1917                 }
1918                 if (p->wire_count != 0) {
1919                         if ((options & OBJPR_NOTMAPPED) == 0)
1920                                 pmap_remove_all(p);
1921                         if ((options & OBJPR_CLEANONLY) == 0) {
1922                                 p->valid = 0;
1923                                 vm_page_undirty(p);
1924                         }
1925                         goto next;
1926                 }
1927                 if (vm_page_busied(p)) {
1928                         VM_OBJECT_WUNLOCK(object);
1929                         vm_page_busy_sleep(p, "vmopar");
1930                         VM_OBJECT_WLOCK(object);
1931                         goto again;
1932                 }
1933                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1934                     ("vm_object_page_remove: page %p is fictitious", p));
1935                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1936                         if ((options & OBJPR_NOTMAPPED) == 0)
1937                                 pmap_remove_write(p);
1938                         if (p->dirty)
1939                                 goto next;
1940                 }
1941                 if ((options & OBJPR_NOTMAPPED) == 0)
1942                         pmap_remove_all(p);
1943                 vm_page_free(p);
1944 next:
1945                 vm_page_unlock(p);
1946         }
1947         vm_object_pip_wakeup(object);
1948 skipmemq:
1949         if (__predict_false(!vm_object_cache_is_empty(object)))
1950                 vm_page_cache_free(object, start, end);
1951 }
1952
1953 /*
1954  *      vm_object_page_cache:
1955  *
1956  *      For the given object, attempt to move the specified clean
1957  *      pages to the cache queue.  If a page is wired for any reason,
1958  *      then it will not be changed.  Pages are specified by the given
1959  *      range ["start", "end").  As a special case, if "end" is zero,
1960  *      then the range extends from "start" to the end of the object.
1961  *      Any mappings to the specified pages are removed before the
1962  *      pages are moved to the cache queue.
1963  *
1964  *      This operation should only be performed on objects that
1965  *      contain non-fictitious, managed pages.
1966  *
1967  *      The object must be locked.
1968  */
1969 void
1970 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1971 {
1972         struct mtx *mtx, *new_mtx;
1973         vm_page_t p, next;
1974
1975         VM_OBJECT_ASSERT_WLOCKED(object);
1976         KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1977             ("vm_object_page_cache: illegal object %p", object));
1978         if (object->resident_page_count == 0)
1979                 return;
1980         p = vm_page_find_least(object, start);
1981
1982         /*
1983          * Here, the variable "p" is either (1) the page with the least pindex
1984          * greater than or equal to the parameter "start" or (2) NULL.
1985          */
1986         mtx = NULL;
1987         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1988                 next = TAILQ_NEXT(p, listq);
1989
1990                 /*
1991                  * Avoid releasing and reacquiring the same page lock.
1992                  */
1993                 new_mtx = vm_page_lockptr(p);
1994                 if (mtx != new_mtx) {
1995                         if (mtx != NULL)
1996                                 mtx_unlock(mtx);
1997                         mtx = new_mtx;
1998                         mtx_lock(mtx);
1999                 }
2000                 vm_page_try_to_cache(p);
2001         }
2002         if (mtx != NULL)
2003                 mtx_unlock(mtx);
2004 }
2005
2006 /*
2007  *      Populate the specified range of the object with valid pages.  Returns
2008  *      TRUE if the range is successfully populated and FALSE otherwise.
2009  *
2010  *      Note: This function should be optimized to pass a larger array of
2011  *      pages to vm_pager_get_pages() before it is applied to a non-
2012  *      OBJT_DEVICE object.
2013  *
2014  *      The object must be locked.
2015  */
2016 boolean_t
2017 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2018 {
2019         vm_page_t m, ma[1];
2020         vm_pindex_t pindex;
2021         int rv;
2022
2023         VM_OBJECT_ASSERT_WLOCKED(object);
2024         for (pindex = start; pindex < end; pindex++) {
2025                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2026                 if (m->valid != VM_PAGE_BITS_ALL) {
2027                         ma[0] = m;
2028                         rv = vm_pager_get_pages(object, ma, 1, 0);
2029                         m = vm_page_lookup(object, pindex);
2030                         if (m == NULL)
2031                                 break;
2032                         if (rv != VM_PAGER_OK) {
2033                                 vm_page_lock(m);
2034                                 vm_page_free(m);
2035                                 vm_page_unlock(m);
2036                                 break;
2037                         }
2038                 }
2039                 /*
2040                  * Keep "m" busy because a subsequent iteration may unlock
2041                  * the object.
2042                  */
2043         }
2044         if (pindex > start) {
2045                 m = vm_page_lookup(object, start);
2046                 while (m != NULL && m->pindex < pindex) {
2047                         vm_page_xunbusy(m);
2048                         m = TAILQ_NEXT(m, listq);
2049                 }
2050         }
2051         return (pindex == end);
2052 }
2053
2054 /*
2055  *      Routine:        vm_object_coalesce
2056  *      Function:       Coalesces two objects backing up adjoining
2057  *                      regions of memory into a single object.
2058  *
2059  *      returns TRUE if objects were combined.
2060  *
2061  *      NOTE:   Only works at the moment if the second object is NULL -
2062  *              if it's not, which object do we lock first?
2063  *
2064  *      Parameters:
2065  *              prev_object     First object to coalesce
2066  *              prev_offset     Offset into prev_object
2067  *              prev_size       Size of reference to prev_object
2068  *              next_size       Size of reference to the second object
2069  *              reserved        Indicator that extension region has
2070  *                              swap accounted for
2071  *
2072  *      Conditions:
2073  *      The object must *not* be locked.
2074  */
2075 boolean_t
2076 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2077     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2078 {
2079         vm_pindex_t next_pindex;
2080
2081         if (prev_object == NULL)
2082                 return (TRUE);
2083         VM_OBJECT_WLOCK(prev_object);
2084         if ((prev_object->type != OBJT_DEFAULT &&
2085             prev_object->type != OBJT_SWAP) ||
2086             (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2087                 VM_OBJECT_WUNLOCK(prev_object);
2088                 return (FALSE);
2089         }
2090
2091         /*
2092          * Try to collapse the object first
2093          */
2094         vm_object_collapse(prev_object);
2095
2096         /*
2097          * Can't coalesce if: . more than one reference . paged out . shadows
2098          * another object . has a copy elsewhere (any of which mean that the
2099          * pages not mapped to prev_entry may be in use anyway)
2100          */
2101         if (prev_object->backing_object != NULL) {
2102                 VM_OBJECT_WUNLOCK(prev_object);
2103                 return (FALSE);
2104         }
2105
2106         prev_size >>= PAGE_SHIFT;
2107         next_size >>= PAGE_SHIFT;
2108         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2109
2110         if ((prev_object->ref_count > 1) &&
2111             (prev_object->size != next_pindex)) {
2112                 VM_OBJECT_WUNLOCK(prev_object);
2113                 return (FALSE);
2114         }
2115
2116         /*
2117          * Account for the charge.
2118          */
2119         if (prev_object->cred != NULL) {
2120
2121                 /*
2122                  * If prev_object was charged, then this mapping,
2123                  * althought not charged now, may become writable
2124                  * later. Non-NULL cred in the object would prevent
2125                  * swap reservation during enabling of the write
2126                  * access, so reserve swap now. Failed reservation
2127                  * cause allocation of the separate object for the map
2128                  * entry, and swap reservation for this entry is
2129                  * managed in appropriate time.
2130                  */
2131                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2132                     prev_object->cred)) {
2133                         return (FALSE);
2134                 }
2135                 prev_object->charge += ptoa(next_size);
2136         }
2137
2138         /*
2139          * Remove any pages that may still be in the object from a previous
2140          * deallocation.
2141          */
2142         if (next_pindex < prev_object->size) {
2143                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2144                     next_size, 0);
2145                 if (prev_object->type == OBJT_SWAP)
2146                         swap_pager_freespace(prev_object,
2147                                              next_pindex, next_size);
2148 #if 0
2149                 if (prev_object->cred != NULL) {
2150                         KASSERT(prev_object->charge >=
2151                             ptoa(prev_object->size - next_pindex),
2152                             ("object %p overcharged 1 %jx %jx", prev_object,
2153                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2154                         prev_object->charge -= ptoa(prev_object->size -
2155                             next_pindex);
2156                 }
2157 #endif
2158         }
2159
2160         /*
2161          * Extend the object if necessary.
2162          */
2163         if (next_pindex + next_size > prev_object->size)
2164                 prev_object->size = next_pindex + next_size;
2165
2166         VM_OBJECT_WUNLOCK(prev_object);
2167         return (TRUE);
2168 }
2169
2170 void
2171 vm_object_set_writeable_dirty(vm_object_t object)
2172 {
2173
2174         VM_OBJECT_ASSERT_WLOCKED(object);
2175         if (object->type != OBJT_VNODE) {
2176                 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2177                         KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2178                         vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2179                 }
2180                 return;
2181         }
2182         object->generation++;
2183         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2184                 return;
2185         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2186 }
2187
2188 /*
2189  *      vm_object_unwire:
2190  *
2191  *      For each page offset within the specified range of the given object,
2192  *      find the highest-level page in the shadow chain and unwire it.  A page
2193  *      must exist at every page offset, and the highest-level page must be
2194  *      wired.
2195  */
2196 void
2197 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2198     uint8_t queue)
2199 {
2200         vm_object_t tobject;
2201         vm_page_t m, tm;
2202         vm_pindex_t end_pindex, pindex, tpindex;
2203         int depth, locked_depth;
2204
2205         KASSERT((offset & PAGE_MASK) == 0,
2206             ("vm_object_unwire: offset is not page aligned"));
2207         KASSERT((length & PAGE_MASK) == 0,
2208             ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2209         /* The wired count of a fictitious page never changes. */
2210         if ((object->flags & OBJ_FICTITIOUS) != 0)
2211                 return;
2212         pindex = OFF_TO_IDX(offset);
2213         end_pindex = pindex + atop(length);
2214         locked_depth = 1;
2215         VM_OBJECT_RLOCK(object);
2216         m = vm_page_find_least(object, pindex);
2217         while (pindex < end_pindex) {
2218                 if (m == NULL || pindex < m->pindex) {
2219                         /*
2220                          * The first object in the shadow chain doesn't
2221                          * contain a page at the current index.  Therefore,
2222                          * the page must exist in a backing object.
2223                          */
2224                         tobject = object;
2225                         tpindex = pindex;
2226                         depth = 0;
2227                         do {
2228                                 tpindex +=
2229                                     OFF_TO_IDX(tobject->backing_object_offset);
2230                                 tobject = tobject->backing_object;
2231                                 KASSERT(tobject != NULL,
2232                                     ("vm_object_unwire: missing page"));
2233                                 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2234                                         goto next_page;
2235                                 depth++;
2236                                 if (depth == locked_depth) {
2237                                         locked_depth++;
2238                                         VM_OBJECT_RLOCK(tobject);
2239                                 }
2240                         } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2241                             NULL);
2242                 } else {
2243                         tm = m;
2244                         m = TAILQ_NEXT(m, listq);
2245                 }
2246                 vm_page_lock(tm);
2247                 vm_page_unwire(tm, queue);
2248                 vm_page_unlock(tm);
2249 next_page:
2250                 pindex++;
2251         }
2252         /* Release the accumulated object locks. */
2253         for (depth = 0; depth < locked_depth; depth++) {
2254                 tobject = object->backing_object;
2255                 VM_OBJECT_RUNLOCK(object);
2256                 object = tobject;
2257         }
2258 }
2259
2260 struct vnode *
2261 vm_object_vnode(vm_object_t object)
2262 {
2263
2264         VM_OBJECT_ASSERT_LOCKED(object);
2265         if (object->type == OBJT_VNODE)
2266                 return (object->handle);
2267         if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2268                 return (object->un_pager.swp.swp_tmpfs);
2269         return (NULL);
2270 }
2271
2272 static int
2273 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2274 {
2275         struct kinfo_vmobject kvo;
2276         char *fullpath, *freepath;
2277         struct vnode *vp;
2278         struct vattr va;
2279         vm_object_t obj;
2280         vm_page_t m;
2281         int count, error;
2282
2283         if (req->oldptr == NULL) {
2284                 /*
2285                  * If an old buffer has not been provided, generate an
2286                  * estimate of the space needed for a subsequent call.
2287                  */
2288                 mtx_lock(&vm_object_list_mtx);
2289                 count = 0;
2290                 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2291                         if (obj->type == OBJT_DEAD)
2292                                 continue;
2293                         count++;
2294                 }
2295                 mtx_unlock(&vm_object_list_mtx);
2296                 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2297                     count * 11 / 10));
2298         }
2299
2300         error = 0;
2301
2302         /*
2303          * VM objects are type stable and are never removed from the
2304          * list once added.  This allows us to safely read obj->object_list
2305          * after reacquiring the VM object lock.
2306          */
2307         mtx_lock(&vm_object_list_mtx);
2308         TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2309                 if (obj->type == OBJT_DEAD)
2310                         continue;
2311                 VM_OBJECT_RLOCK(obj);
2312                 if (obj->type == OBJT_DEAD) {
2313                         VM_OBJECT_RUNLOCK(obj);
2314                         continue;
2315                 }
2316                 mtx_unlock(&vm_object_list_mtx);
2317                 kvo.kvo_size = ptoa(obj->size);
2318                 kvo.kvo_resident = obj->resident_page_count;
2319                 kvo.kvo_ref_count = obj->ref_count;
2320                 kvo.kvo_shadow_count = obj->shadow_count;
2321                 kvo.kvo_memattr = obj->memattr;
2322                 kvo.kvo_active = 0;
2323                 kvo.kvo_inactive = 0;
2324                 TAILQ_FOREACH(m, &obj->memq, listq) {
2325                         /*
2326                          * A page may belong to the object but be
2327                          * dequeued and set to PQ_NONE while the
2328                          * object lock is not held.  This makes the
2329                          * reads of m->queue below racy, and we do not
2330                          * count pages set to PQ_NONE.  However, this
2331                          * sysctl is only meant to give an
2332                          * approximation of the system anyway.
2333                          */
2334                         if (m->queue == PQ_ACTIVE)
2335                                 kvo.kvo_active++;
2336                         else if (m->queue == PQ_INACTIVE)
2337                                 kvo.kvo_inactive++;
2338                 }
2339
2340                 kvo.kvo_vn_fileid = 0;
2341                 kvo.kvo_vn_fsid = 0;
2342                 freepath = NULL;
2343                 fullpath = "";
2344                 vp = NULL;
2345                 switch (obj->type) {
2346                 case OBJT_DEFAULT:
2347                         kvo.kvo_type = KVME_TYPE_DEFAULT;
2348                         break;
2349                 case OBJT_VNODE:
2350                         kvo.kvo_type = KVME_TYPE_VNODE;
2351                         vp = obj->handle;
2352                         vref(vp);
2353                         break;
2354                 case OBJT_SWAP:
2355                         kvo.kvo_type = KVME_TYPE_SWAP;
2356                         break;
2357                 case OBJT_DEVICE:
2358                         kvo.kvo_type = KVME_TYPE_DEVICE;
2359                         break;
2360                 case OBJT_PHYS:
2361                         kvo.kvo_type = KVME_TYPE_PHYS;
2362                         break;
2363                 case OBJT_DEAD:
2364                         kvo.kvo_type = KVME_TYPE_DEAD;
2365                         break;
2366                 case OBJT_SG:
2367                         kvo.kvo_type = KVME_TYPE_SG;
2368                         break;
2369                 case OBJT_MGTDEVICE:
2370                         kvo.kvo_type = KVME_TYPE_MGTDEVICE;
2371                         break;
2372                 default:
2373                         kvo.kvo_type = KVME_TYPE_UNKNOWN;
2374                         break;
2375                 }
2376                 VM_OBJECT_RUNLOCK(obj);
2377                 if (vp != NULL) {
2378                         vn_fullpath(curthread, vp, &fullpath, &freepath);
2379                         vn_lock(vp, LK_SHARED | LK_RETRY);
2380                         if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2381                                 kvo.kvo_vn_fileid = va.va_fileid;
2382                                 kvo.kvo_vn_fsid = va.va_fsid;
2383                         }
2384                         vput(vp);
2385                 }
2386
2387                 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path));
2388                 if (freepath != NULL)
2389                         free(freepath, M_TEMP);
2390
2391                 /* Pack record size down */
2392                 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) +
2393                     strlen(kvo.kvo_path) + 1;
2394                 kvo.kvo_structsize = roundup(kvo.kvo_structsize,
2395                     sizeof(uint64_t));
2396                 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize);
2397                 mtx_lock(&vm_object_list_mtx);
2398                 if (error)
2399                         break;
2400         }
2401         mtx_unlock(&vm_object_list_mtx);
2402         return (error);
2403 }
2404 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2405     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2406     "List of VM objects");
2407
2408 #include "opt_ddb.h"
2409 #ifdef DDB
2410 #include <sys/kernel.h>
2411
2412 #include <sys/cons.h>
2413
2414 #include <ddb/ddb.h>
2415
2416 static int
2417 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2418 {
2419         vm_map_t tmpm;
2420         vm_map_entry_t tmpe;
2421         vm_object_t obj;
2422         int entcount;
2423
2424         if (map == 0)
2425                 return 0;
2426
2427         if (entry == 0) {
2428                 tmpe = map->header.next;
2429                 entcount = map->nentries;
2430                 while (entcount-- && (tmpe != &map->header)) {
2431                         if (_vm_object_in_map(map, object, tmpe)) {
2432                                 return 1;
2433                         }
2434                         tmpe = tmpe->next;
2435                 }
2436         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2437                 tmpm = entry->object.sub_map;
2438                 tmpe = tmpm->header.next;
2439                 entcount = tmpm->nentries;
2440                 while (entcount-- && tmpe != &tmpm->header) {
2441                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2442                                 return 1;
2443                         }
2444                         tmpe = tmpe->next;
2445                 }
2446         } else if ((obj = entry->object.vm_object) != NULL) {
2447                 for (; obj; obj = obj->backing_object)
2448                         if (obj == object) {
2449                                 return 1;
2450                         }
2451         }
2452         return 0;
2453 }
2454
2455 static int
2456 vm_object_in_map(vm_object_t object)
2457 {
2458         struct proc *p;
2459
2460         /* sx_slock(&allproc_lock); */
2461         FOREACH_PROC_IN_SYSTEM(p) {
2462                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2463                         continue;
2464                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2465                         /* sx_sunlock(&allproc_lock); */
2466                         return 1;
2467                 }
2468         }
2469         /* sx_sunlock(&allproc_lock); */
2470         if (_vm_object_in_map(kernel_map, object, 0))
2471                 return 1;
2472         return 0;
2473 }
2474
2475 DB_SHOW_COMMAND(vmochk, vm_object_check)
2476 {
2477         vm_object_t object;
2478
2479         /*
2480          * make sure that internal objs are in a map somewhere
2481          * and none have zero ref counts.
2482          */
2483         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2484                 if (object->handle == NULL &&
2485                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2486                         if (object->ref_count == 0) {
2487                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2488                                         (long)object->size);
2489                         }
2490                         if (!vm_object_in_map(object)) {
2491                                 db_printf(
2492                         "vmochk: internal obj is not in a map: "
2493                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2494                                     object->ref_count, (u_long)object->size,
2495                                     (u_long)object->size,
2496                                     (void *)object->backing_object);
2497                         }
2498                 }
2499         }
2500 }
2501
2502 /*
2503  *      vm_object_print:        [ debug ]
2504  */
2505 DB_SHOW_COMMAND(object, vm_object_print_static)
2506 {
2507         /* XXX convert args. */
2508         vm_object_t object = (vm_object_t)addr;
2509         boolean_t full = have_addr;
2510
2511         vm_page_t p;
2512
2513         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2514 #define count   was_count
2515
2516         int count;
2517
2518         if (object == NULL)
2519                 return;
2520
2521         db_iprintf(
2522             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2523             object, (int)object->type, (uintmax_t)object->size,
2524             object->resident_page_count, object->ref_count, object->flags,
2525             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2526         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2527             object->shadow_count,
2528             object->backing_object ? object->backing_object->ref_count : 0,
2529             object->backing_object, (uintmax_t)object->backing_object_offset);
2530
2531         if (!full)
2532                 return;
2533
2534         db_indent += 2;
2535         count = 0;
2536         TAILQ_FOREACH(p, &object->memq, listq) {
2537                 if (count == 0)
2538                         db_iprintf("memory:=");
2539                 else if (count == 6) {
2540                         db_printf("\n");
2541                         db_iprintf(" ...");
2542                         count = 0;
2543                 } else
2544                         db_printf(",");
2545                 count++;
2546
2547                 db_printf("(off=0x%jx,page=0x%jx)",
2548                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2549         }
2550         if (count != 0)
2551                 db_printf("\n");
2552         db_indent -= 2;
2553 }
2554
2555 /* XXX. */
2556 #undef count
2557
2558 /* XXX need this non-static entry for calling from vm_map_print. */
2559 void
2560 vm_object_print(
2561         /* db_expr_t */ long addr,
2562         boolean_t have_addr,
2563         /* db_expr_t */ long count,
2564         char *modif)
2565 {
2566         vm_object_print_static(addr, have_addr, count, modif);
2567 }
2568
2569 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2570 {
2571         vm_object_t object;
2572         vm_pindex_t fidx;
2573         vm_paddr_t pa;
2574         vm_page_t m, prev_m;
2575         int rcount, nl, c;
2576
2577         nl = 0;
2578         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2579                 db_printf("new object: %p\n", (void *)object);
2580                 if (nl > 18) {
2581                         c = cngetc();
2582                         if (c != ' ')
2583                                 return;
2584                         nl = 0;
2585                 }
2586                 nl++;
2587                 rcount = 0;
2588                 fidx = 0;
2589                 pa = -1;
2590                 TAILQ_FOREACH(m, &object->memq, listq) {
2591                         if (m->pindex > 128)
2592                                 break;
2593                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2594                             prev_m->pindex + 1 != m->pindex) {
2595                                 if (rcount) {
2596                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2597                                                 (long)fidx, rcount, (long)pa);
2598                                         if (nl > 18) {
2599                                                 c = cngetc();
2600                                                 if (c != ' ')
2601                                                         return;
2602                                                 nl = 0;
2603                                         }
2604                                         nl++;
2605                                         rcount = 0;
2606                                 }
2607                         }
2608                         if (rcount &&
2609                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2610                                 ++rcount;
2611                                 continue;
2612                         }
2613                         if (rcount) {
2614                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2615                                         (long)fidx, rcount, (long)pa);
2616                                 if (nl > 18) {
2617                                         c = cngetc();
2618                                         if (c != ' ')
2619                                                 return;
2620                                         nl = 0;
2621                                 }
2622                                 nl++;
2623                         }
2624                         fidx = m->pindex;
2625                         pa = VM_PAGE_TO_PHYS(m);
2626                         rcount = 1;
2627                 }
2628                 if (rcount) {
2629                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2630                                 (long)fidx, rcount, (long)pa);
2631                         if (nl > 18) {
2632                                 c = cngetc();
2633                                 if (c != ' ')
2634                                         return;
2635                                 nl = 0;
2636                         }
2637                         nl++;
2638                 }
2639         }
2640 }
2641 #endif /* DDB */