sys/vm/vm_object.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * The Mach Operating System project at Carnegie-Mellon University.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 4. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  33  *
  34  *
  35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  36  * All rights reserved.
  37  *
  38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  39  *
  40  * Permission to use, copy, modify and distribute this software and
  41  * its documentation is hereby granted, provided that both the copyright
  42  * notice and this permission notice appear in all copies of the
  43  * software, derivative works or modified versions, and any portions
  44  * thereof, and that both notices appear in supporting documentation.
  45  *
  46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  49  *
  50  * Carnegie Mellon requests users of this software to return to
  51  *
  52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  53  *  School of Computer Science
  54  *  Carnegie Mellon University
  55  *  Pittsburgh PA 15213-3890
  56  *
  57  * any improvements or extensions that they make and grant Carnegie the
  58  * rights to redistribute these changes.
  59  */
  60
  61 /*
  62  *      Virtual memory object module.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_vm.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/lock.h>
  73 #include <sys/mman.h>
  74 #include <sys/mount.h>
  75 #include <sys/kernel.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/mutex.h>
  78 #include <sys/proc.h>           /* for curproc, pageproc */
  79 #include <sys/socket.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/vnode.h>
  82 #include <sys/vmmeter.h>
  83 #include <sys/sx.h>
  84
  85 #include <vm/vm.h>
  86 #include <vm/vm_param.h>
  87 #include <vm/pmap.h>
  88 #include <vm/vm_map.h>
  89 #include <vm/vm_object.h>
  90 #include <vm/vm_page.h>
  91 #include <vm/vm_pageout.h>
  92 #include <vm/vm_pager.h>
  93 #include <vm/swap_pager.h>
  94 #include <vm/vm_kern.h>
  95 #include <vm/vm_extern.h>
  96 #include <vm/vm_reserv.h>
  97 #include <vm/uma.h>
  98
  99 static int old_msync;
 100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 101     "Use old (insecure) msync behavior");
 102
 103 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 104                     int pagerflags, int flags, boolean_t *clearobjflags,
 105                     boolean_t *eio);
 106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 107                     boolean_t *clearobjflags);
 108 static void     vm_object_qcollapse(vm_object_t object);
 109 static void     vm_object_vndeallocate(vm_object_t object);
 110
 111 /*
 112  *      Virtual memory objects maintain the actual data
 113  *      associated with allocated virtual memory.  A given
 114  *      page of memory exists within exactly one object.
 115  *
 116  *      An object is only deallocated when all "references"
 117  *      are given up.  Only one "reference" to a given
 118  *      region of an object should be writeable.
 119  *
 120  *      Associated with each object is a list of all resident
 121  *      memory pages belonging to that object; this list is
 122  *      maintained by the "vm_page" module, and locked by the object's
 123  *      lock.
 124  *
 125  *      Each object also records a "pager" routine which is
 126  *      used to retrieve (and store) pages to the proper backing
 127  *      storage.  In addition, objects may be backed by other
 128  *      objects from which they were virtual-copied.
 129  *
 130  *      The only items within the object structure which are
 131  *      modified after time of creation are:
 132  *              reference count         locked by object's lock
 133  *              pager routine           locked by object's lock
 134  *
 135  */
 136
 137 struct object_q vm_object_list;
 138 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 139
 140 struct vm_object kernel_object_store;
 141 struct vm_object kmem_object_store;
 142
 143 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
 144
 145 static long object_collapses;
 146 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 147     &object_collapses, 0, "VM object collapses");
 148
 149 static long object_bypasses;
 150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 151     &object_bypasses, 0, "VM object bypasses");
 152
 153 static uma_zone_t obj_zone;
 154
 155 static int vm_object_zinit(void *mem, int size, int flags);
 156
 157 #ifdef INVARIANTS
 158 static void vm_object_zdtor(void *mem, int size, void *arg);
 159
 160 static void
 161 vm_object_zdtor(void *mem, int size, void *arg)
 162 {
 163         vm_object_t object;
 164
 165         object = (vm_object_t)mem;
 166         KASSERT(TAILQ_EMPTY(&object->memq),
 167             ("object %p has resident pages",
 168             object));
 169 #if VM_NRESERVLEVEL > 0
 170         KASSERT(LIST_EMPTY(&object->rvq),
 171             ("object %p has reservations",
 172             object));
 173 #endif
 174         KASSERT(object->cache == NULL,
 175             ("object %p has cached pages",
 176             object));
 177         KASSERT(object->paging_in_progress == 0,
 178             ("object %p paging_in_progress = %d",
 179             object, object->paging_in_progress));
 180         KASSERT(object->resident_page_count == 0,
 181             ("object %p resident_page_count = %d",
 182             object, object->resident_page_count));
 183         KASSERT(object->shadow_count == 0,
 184             ("object %p shadow_count = %d",
 185             object, object->shadow_count));
 186 }
 187 #endif
 188
 189 static int
 190 vm_object_zinit(void *mem, int size, int flags)
 191 {
 192         vm_object_t object;
 193
 194         object = (vm_object_t)mem;
 195         bzero(&object->mtx, sizeof(object->mtx));
 196         VM_OBJECT_LOCK_INIT(object, "standard object");
 197
 198         /* These are true for any object that has been freed */
 199         object->paging_in_progress = 0;
 200         object->resident_page_count = 0;
 201         object->shadow_count = 0;
 202         return (0);
 203 }
 204
 205 void
 206 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 207 {
 208
 209         TAILQ_INIT(&object->memq);
 210         LIST_INIT(&object->shadow_head);
 211
 212         object->root = NULL;
 213         object->type = type;
 214         object->size = size;
 215         object->generation = 1;
 216         object->ref_count = 1;
 217         object->memattr = VM_MEMATTR_DEFAULT;
 218         object->flags = 0;
 219         object->cred = NULL;
 220         object->charge = 0;
 221         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
 222                 object->flags = OBJ_ONEMAPPING;
 223         object->pg_color = 0;
 224         object->handle = NULL;
 225         object->backing_object = NULL;
 226         object->backing_object_offset = (vm_ooffset_t) 0;
 227 #if VM_NRESERVLEVEL > 0
 228         LIST_INIT(&object->rvq);
 229 #endif
 230         object->cache = NULL;
 231
 232         mtx_lock(&vm_object_list_mtx);
 233         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 234         mtx_unlock(&vm_object_list_mtx);
 235 }
 236
 237 /*
 238  *      vm_object_init:
 239  *
 240  *      Initialize the VM objects module.
 241  */
 242 void
 243 vm_object_init(void)
 244 {
 245         TAILQ_INIT(&vm_object_list);
 246         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 247
 248         VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
 249         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 250             kernel_object);
 251 #if VM_NRESERVLEVEL > 0
 252         kernel_object->flags |= OBJ_COLORED;
 253         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 254 #endif
 255
 256         VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
 257         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 258             kmem_object);
 259 #if VM_NRESERVLEVEL > 0
 260         kmem_object->flags |= OBJ_COLORED;
 261         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 262 #endif
 263
 264         /*
 265          * The lock portion of struct vm_object must be type stable due
 266          * to vm_pageout_fallback_object_lock locking a vm object
 267          * without holding any references to it.
 268          */
 269         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 270 #ifdef INVARIANTS
 271             vm_object_zdtor,
 272 #else
 273             NULL,
 274 #endif
 275             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
 276 }
 277
 278 void
 279 vm_object_clear_flag(vm_object_t object, u_short bits)
 280 {
 281
 282         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 283         object->flags &= ~bits;
 284 }
 285
 286 /*
 287  *      Sets the default memory attribute for the specified object.  Pages
 288  *      that are allocated to this object are by default assigned this memory
 289  *      attribute.
 290  *
 291  *      Presently, this function must be called before any pages are allocated
 292  *      to the object.  In the future, this requirement may be relaxed for
 293  *      "default" and "swap" objects.
 294  */
 295 int
 296 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 297 {
 298
 299         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 300         switch (object->type) {
 301         case OBJT_DEFAULT:
 302         case OBJT_DEVICE:
 303         case OBJT_PHYS:
 304         case OBJT_SG:
 305         case OBJT_SWAP:
 306         case OBJT_VNODE:
 307                 if (!TAILQ_EMPTY(&object->memq))
 308                         return (KERN_FAILURE);
 309                 break;
 310         case OBJT_DEAD:
 311                 return (KERN_INVALID_ARGUMENT);
 312         }
 313         object->memattr = memattr;
 314         return (KERN_SUCCESS);
 315 }
 316
 317 void
 318 vm_object_pip_add(vm_object_t object, short i)
 319 {
 320
 321         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 322         object->paging_in_progress += i;
 323 }
 324
 325 void
 326 vm_object_pip_subtract(vm_object_t object, short i)
 327 {
 328
 329         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 330         object->paging_in_progress -= i;
 331 }
 332
 333 void
 334 vm_object_pip_wakeup(vm_object_t object)
 335 {
 336
 337         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 338         object->paging_in_progress--;
 339         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 340                 vm_object_clear_flag(object, OBJ_PIPWNT);
 341                 wakeup(object);
 342         }
 343 }
 344
 345 void
 346 vm_object_pip_wakeupn(vm_object_t object, short i)
 347 {
 348
 349         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 350         if (i)
 351                 object->paging_in_progress -= i;
 352         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 353                 vm_object_clear_flag(object, OBJ_PIPWNT);
 354                 wakeup(object);
 355         }
 356 }
 357
 358 void
 359 vm_object_pip_wait(vm_object_t object, char *waitid)
 360 {
 361
 362         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 363         while (object->paging_in_progress) {
 364                 object->flags |= OBJ_PIPWNT;
 365                 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
 366         }
 367 }
 368
 369 /*
 370  *      vm_object_allocate:
 371  *
 372  *      Returns a new object with the given size.
 373  */
 374 vm_object_t
 375 vm_object_allocate(objtype_t type, vm_pindex_t size)
 376 {
 377         vm_object_t object;
 378
 379         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 380         _vm_object_allocate(type, size, object);
 381         return (object);
 382 }
 383
 384
 385 /*
 386  *      vm_object_reference:
 387  *
 388  *      Gets another reference to the given object.  Note: OBJ_DEAD
 389  *      objects can be referenced during final cleaning.
 390  */
 391 void
 392 vm_object_reference(vm_object_t object)
 393 {
 394         if (object == NULL)
 395                 return;
 396         VM_OBJECT_LOCK(object);
 397         vm_object_reference_locked(object);
 398         VM_OBJECT_UNLOCK(object);
 399 }
 400
 401 /*
 402  *      vm_object_reference_locked:
 403  *
 404  *      Gets another reference to the given object.
 405  *
 406  *      The object must be locked.
 407  */
 408 void
 409 vm_object_reference_locked(vm_object_t object)
 410 {
 411         struct vnode *vp;
 412
 413         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 414         object->ref_count++;
 415         if (object->type == OBJT_VNODE) {
 416                 vp = object->handle;
 417                 vref(vp);
 418         }
 419 }
 420
 421 /*
 422  * Handle deallocating an object of type OBJT_VNODE.
 423  */
 424 static void
 425 vm_object_vndeallocate(vm_object_t object)
 426 {
 427         struct vnode *vp = (struct vnode *) object->handle;
 428
 429         VFS_ASSERT_GIANT(vp->v_mount);
 430         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 431         KASSERT(object->type == OBJT_VNODE,
 432             ("vm_object_vndeallocate: not a vnode object"));
 433         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 434 #ifdef INVARIANTS
 435         if (object->ref_count == 0) {
 436                 vprint("vm_object_vndeallocate", vp);
 437                 panic("vm_object_vndeallocate: bad object reference count");
 438         }
 439 #endif
 440
 441         if (object->ref_count > 1) {
 442                 object->ref_count--;
 443                 VM_OBJECT_UNLOCK(object);
 444                 /* vrele may need the vnode lock. */
 445                 vrele(vp);
 446         } else {
 447                 vhold(vp);
 448                 VM_OBJECT_UNLOCK(object);
 449                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 450                 vdrop(vp);
 451                 VM_OBJECT_LOCK(object);
 452                 object->ref_count--;
 453                 if (object->type == OBJT_DEAD) {
 454                         VM_OBJECT_UNLOCK(object);
 455                         VOP_UNLOCK(vp, 0);
 456                 } else {
 457                         if (object->ref_count == 0)
 458                                 vp->v_vflag &= ~VV_TEXT;
 459                         VM_OBJECT_UNLOCK(object);
 460                         vput(vp);
 461                 }
 462         }
 463 }
 464
 465 /*
 466  *      vm_object_deallocate:
 467  *
 468  *      Release a reference to the specified object,
 469  *      gained either through a vm_object_allocate
 470  *      or a vm_object_reference call.  When all references
 471  *      are gone, storage associated with this object
 472  *      may be relinquished.
 473  *
 474  *      No object may be locked.
 475  */
 476 void
 477 vm_object_deallocate(vm_object_t object)
 478 {
 479         vm_object_t temp;
 480
 481         while (object != NULL) {
 482                 int vfslocked;
 483
 484                 vfslocked = 0;
 485         restart:
 486                 VM_OBJECT_LOCK(object);
 487                 if (object->type == OBJT_VNODE) {
 488                         struct vnode *vp = (struct vnode *) object->handle;
 489
 490                         /*
 491                          * Conditionally acquire Giant for a vnode-backed
 492                          * object.  We have to be careful since the type of
 493                          * a vnode object can change while the object is
 494                          * unlocked.
 495                          */
 496                         if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
 497                                 vfslocked = 1;
 498                                 if (!mtx_trylock(&Giant)) {
 499                                         VM_OBJECT_UNLOCK(object);
 500                                         mtx_lock(&Giant);
 501                                         goto restart;
 502                                 }
 503                         }
 504                         vm_object_vndeallocate(object);
 505                         VFS_UNLOCK_GIANT(vfslocked);
 506                         return;
 507                 } else
 508                         /*
 509                          * This is to handle the case that the object
 510                          * changed type while we dropped its lock to
 511                          * obtain Giant.
 512                          */
 513                         VFS_UNLOCK_GIANT(vfslocked);
 514
 515                 KASSERT(object->ref_count != 0,
 516                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 517
 518                 /*
 519                  * If the reference count goes to 0 we start calling
 520                  * vm_object_terminate() on the object chain.
 521                  * A ref count of 1 may be a special case depending on the
 522                  * shadow count being 0 or 1.
 523                  */
 524                 object->ref_count--;
 525                 if (object->ref_count > 1) {
 526                         VM_OBJECT_UNLOCK(object);
 527                         return;
 528                 } else if (object->ref_count == 1) {
 529                         if (object->shadow_count == 0 &&
 530                             object->handle == NULL &&
 531                             (object->type == OBJT_DEFAULT ||
 532                              object->type == OBJT_SWAP)) {
 533                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 534                         } else if ((object->shadow_count == 1) &&
 535                             (object->handle == NULL) &&
 536                             (object->type == OBJT_DEFAULT ||
 537                              object->type == OBJT_SWAP)) {
 538                                 vm_object_t robject;
 539
 540                                 robject = LIST_FIRST(&object->shadow_head);
 541                                 KASSERT(robject != NULL,
 542                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 543                                          object->ref_count,
 544                                          object->shadow_count));
 545                                 if (!VM_OBJECT_TRYLOCK(robject)) {
 546                                         /*
 547                                          * Avoid a potential deadlock.
 548                                          */
 549                                         object->ref_count++;
 550                                         VM_OBJECT_UNLOCK(object);
 551                                         /*
 552                                          * More likely than not the thread
 553                                          * holding robject's lock has lower
 554                                          * priority than the current thread.
 555                                          * Let the lower priority thread run.
 556                                          */
 557                                         pause("vmo_de", 1);
 558                                         continue;
 559                                 }
 560                                 /*
 561                                  * Collapse object into its shadow unless its
 562                                  * shadow is dead.  In that case, object will
 563                                  * be deallocated by the thread that is
 564                                  * deallocating its shadow.
 565                                  */
 566                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 567                                     (robject->handle == NULL) &&
 568                                     (robject->type == OBJT_DEFAULT ||
 569                                      robject->type == OBJT_SWAP)) {
 570
 571                                         robject->ref_count++;
 572 retry:
 573                                         if (robject->paging_in_progress) {
 574                                                 VM_OBJECT_UNLOCK(object);
 575                                                 vm_object_pip_wait(robject,
 576                                                     "objde1");
 577                                                 temp = robject->backing_object;
 578                                                 if (object == temp) {
 579                                                         VM_OBJECT_LOCK(object);
 580                                                         goto retry;
 581                                                 }
 582                                         } else if (object->paging_in_progress) {
 583                                                 VM_OBJECT_UNLOCK(robject);
 584                                                 object->flags |= OBJ_PIPWNT;
 585                                                 msleep(object,
 586                                                     VM_OBJECT_MTX(object),
 587                                                     PDROP | PVM, "objde2", 0);
 588                                                 VM_OBJECT_LOCK(robject);
 589                                                 temp = robject->backing_object;
 590                                                 if (object == temp) {
 591                                                         VM_OBJECT_LOCK(object);
 592                                                         goto retry;
 593                                                 }
 594                                         } else
 595                                                 VM_OBJECT_UNLOCK(object);
 596
 597                                         if (robject->ref_count == 1) {
 598                                                 robject->ref_count--;
 599                                                 object = robject;
 600                                                 goto doterm;
 601                                         }
 602                                         object = robject;
 603                                         vm_object_collapse(object);
 604                                         VM_OBJECT_UNLOCK(object);
 605                                         continue;
 606                                 }
 607                                 VM_OBJECT_UNLOCK(robject);
 608                         }
 609                         VM_OBJECT_UNLOCK(object);
 610                         return;
 611                 }
 612 doterm:
 613                 temp = object->backing_object;
 614                 if (temp != NULL) {
 615                         VM_OBJECT_LOCK(temp);
 616                         LIST_REMOVE(object, shadow_list);
 617                         temp->shadow_count--;
 618                         VM_OBJECT_UNLOCK(temp);
 619                         object->backing_object = NULL;
 620                 }
 621                 /*
 622                  * Don't double-terminate, we could be in a termination
 623                  * recursion due to the terminate having to sync data
 624                  * to disk.
 625                  */
 626                 if ((object->flags & OBJ_DEAD) == 0)
 627                         vm_object_terminate(object);
 628                 else
 629                         VM_OBJECT_UNLOCK(object);
 630                 object = temp;
 631         }
 632 }
 633
 634 /*
 635  *      vm_object_destroy removes the object from the global object list
 636  *      and frees the space for the object.
 637  */
 638 void
 639 vm_object_destroy(vm_object_t object)
 640 {
 641
 642         /*
 643          * Remove the object from the global object list.
 644          */
 645         mtx_lock(&vm_object_list_mtx);
 646         TAILQ_REMOVE(&vm_object_list, object, object_list);
 647         mtx_unlock(&vm_object_list_mtx);
 648
 649         /*
 650          * Release the allocation charge.
 651          */
 652         if (object->cred != NULL) {
 653                 KASSERT(object->type == OBJT_DEFAULT ||
 654                     object->type == OBJT_SWAP,
 655                     ("vm_object_terminate: non-swap obj %p has cred",
 656                      object));
 657                 swap_release_by_cred(object->charge, object->cred);
 658                 object->charge = 0;
 659                 crfree(object->cred);
 660                 object->cred = NULL;
 661         }
 662
 663         /*
 664          * Free the space for the object.
 665          */
 666         uma_zfree(obj_zone, object);
 667 }
 668
 669 /*
 670  *      vm_object_terminate actually destroys the specified object, freeing
 671  *      up all previously used resources.
 672  *
 673  *      The object must be locked.
 674  *      This routine may block.
 675  */
 676 void
 677 vm_object_terminate(vm_object_t object)
 678 {
 679         vm_page_t p, p_next;
 680
 681         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 682
 683         /*
 684          * Make sure no one uses us.
 685          */
 686         vm_object_set_flag(object, OBJ_DEAD);
 687
 688         /*
 689          * wait for the pageout daemon to be done with the object
 690          */
 691         vm_object_pip_wait(object, "objtrm");
 692
 693         KASSERT(!object->paging_in_progress,
 694                 ("vm_object_terminate: pageout in progress"));
 695
 696         /*
 697          * Clean and free the pages, as appropriate. All references to the
 698          * object are gone, so we don't need to lock it.
 699          */
 700         if (object->type == OBJT_VNODE) {
 701                 struct vnode *vp = (struct vnode *)object->handle;
 702
 703                 /*
 704                  * Clean pages and flush buffers.
 705                  */
 706                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 707                 VM_OBJECT_UNLOCK(object);
 708
 709                 vinvalbuf(vp, V_SAVE, 0, 0);
 710
 711                 VM_OBJECT_LOCK(object);
 712         }
 713
 714         KASSERT(object->ref_count == 0,
 715                 ("vm_object_terminate: object with references, ref_count=%d",
 716                 object->ref_count));
 717
 718         /*
 719          * Free any remaining pageable pages.  This also removes them from the
 720          * paging queues.  However, don't free wired pages, just remove them
 721          * from the object.  Rather than incrementally removing each page from
 722          * the object, the page and object are reset to any empty state.
 723          */
 724         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
 725                 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
 726                     ("vm_object_terminate: freeing busy page %p", p));
 727                 vm_page_lock(p);
 728                 /*
 729                  * Optimize the page's removal from the object by resetting
 730                  * its "object" field.  Specifically, if the page is not
 731                  * wired, then the effect of this assignment is that
 732                  * vm_page_free()'s call to vm_page_remove() will return
 733                  * immediately without modifying the page or the object.
 734                  */
 735                 p->object = NULL;
 736                 if (p->wire_count == 0) {
 737                         vm_page_free(p);
 738                         PCPU_INC(cnt.v_pfree);
 739                 }
 740                 vm_page_unlock(p);
 741         }
 742         /*
 743          * If the object contained any pages, then reset it to an empty state.
 744          * None of the object's fields, including "resident_page_count", were
 745          * modified by the preceding loop.
 746          */
 747         if (object->resident_page_count != 0) {
 748                 object->root = NULL;
 749                 TAILQ_INIT(&object->memq);
 750                 object->resident_page_count = 0;
 751                 if (object->type == OBJT_VNODE)
 752                         vdrop(object->handle);
 753         }
 754
 755 #if VM_NRESERVLEVEL > 0
 756         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 757                 vm_reserv_break_all(object);
 758 #endif
 759         if (__predict_false(object->cache != NULL))
 760                 vm_page_cache_free(object, 0, 0);
 761
 762         /*
 763          * Let the pager know object is dead.
 764          */
 765         vm_pager_deallocate(object);
 766         VM_OBJECT_UNLOCK(object);
 767
 768         vm_object_destroy(object);
 769 }
 770
 771 /*
 772  * Make the page read-only so that we can clear the object flags.  However, if
 773  * this is a nosync mmap then the object is likely to stay dirty so do not
 774  * mess with the page and do not clear the object flags.  Returns TRUE if the
 775  * page should be flushed, and FALSE otherwise.
 776  */
 777 static boolean_t
 778 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
 779 {
 780
 781         /*
 782          * If we have been asked to skip nosync pages and this is a
 783          * nosync page, skip it.  Note that the object flags were not
 784          * cleared in this case so we do not have to set them.
 785          */
 786         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 787                 *clearobjflags = FALSE;
 788                 return (FALSE);
 789         } else {
 790                 pmap_remove_write(p);
 791                 return (p->dirty != 0);
 792         }
 793 }
 794
 795 /*
 796  *      vm_object_page_clean
 797  *
 798  *      Clean all dirty pages in the specified range of object.  Leaves page
 799  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 800  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 801  *      leaving the object dirty.
 802  *
 803  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 804  *      synchronous clustering mode implementation.
 805  *
 806  *      Odd semantics: if start == end, we clean everything.
 807  *
 808  *      The object must be locked.
 809  *
 810  *      Returns FALSE if some page from the range was not written, as
 811  *      reported by the pager, and TRUE otherwise.
 812  */
 813 boolean_t
 814 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 815     int flags)
 816 {
 817         vm_page_t np, p;
 818         vm_pindex_t pi, tend, tstart;
 819         int curgeneration, n, pagerflags;
 820         boolean_t clearobjflags, eio, res;
 821
 822         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 823         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 824         KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
 825         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 826             object->resident_page_count == 0)
 827                 return (TRUE);
 828
 829         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 830             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 831         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 832
 833         tstart = OFF_TO_IDX(start);
 834         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 835         clearobjflags = tstart == 0 && tend >= object->size;
 836         res = TRUE;
 837
 838 rescan:
 839         curgeneration = object->generation;
 840
 841         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 842                 pi = p->pindex;
 843                 if (pi >= tend)
 844                         break;
 845                 np = TAILQ_NEXT(p, listq);
 846                 if (p->valid == 0)
 847                         continue;
 848                 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
 849                         if (object->generation != curgeneration) {
 850                                 if ((flags & OBJPC_SYNC) != 0)
 851                                         goto rescan;
 852                                 else
 853                                         clearobjflags = FALSE;
 854                         }
 855                         np = vm_page_find_least(object, pi);
 856                         continue;
 857                 }
 858                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 859                         continue;
 860
 861                 n = vm_object_page_collect_flush(object, p, pagerflags,
 862                     flags, &clearobjflags, &eio);
 863                 if (eio) {
 864                         res = FALSE;
 865                         clearobjflags = FALSE;
 866                 }
 867                 if (object->generation != curgeneration) {
 868                         if ((flags & OBJPC_SYNC) != 0)
 869                                 goto rescan;
 870                         else
 871                                 clearobjflags = FALSE;
 872                 }
 873
 874                 /*
 875                  * If the VOP_PUTPAGES() did a truncated write, so
 876                  * that even the first page of the run is not fully
 877                  * written, vm_pageout_flush() returns 0 as the run
 878                  * length.  Since the condition that caused truncated
 879                  * write may be permanent, e.g. exhausted free space,
 880                  * accepting n == 0 would cause an infinite loop.
 881                  *
 882                  * Forwarding the iterator leaves the unwritten page
 883                  * behind, but there is not much we can do there if
 884                  * filesystem refuses to write it.
 885                  */
 886                 if (n == 0) {
 887                         n = 1;
 888                         clearobjflags = FALSE;
 889                 }
 890                 np = vm_page_find_least(object, pi + n);
 891         }
 892 #if 0
 893         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 894 #endif
 895
 896         if (clearobjflags)
 897                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 898         return (res);
 899 }
 900
 901 static int
 902 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 903     int flags, boolean_t *clearobjflags, boolean_t *eio)
 904 {
 905         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 906         int count, i, mreq, runlen;
 907
 908         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 909         vm_page_lock_assert(p, MA_NOTOWNED);
 910         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 911
 912         count = 1;
 913         mreq = 0;
 914
 915         for (tp = p; count < vm_pageout_page_count; count++) {
 916                 tp = vm_page_next(tp);
 917                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 918                         break;
 919                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 920                         break;
 921         }
 922
 923         for (p_first = p; count < vm_pageout_page_count; count++) {
 924                 tp = vm_page_prev(p_first);
 925                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 926                         break;
 927                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 928                         break;
 929                 p_first = tp;
 930                 mreq++;
 931         }
 932
 933         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 934                 ma[i] = tp;
 935
 936         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
 937         return (runlen);
 938 }
 939
 940 /*
 941  * Note that there is absolutely no sense in writing out
 942  * anonymous objects, so we track down the vnode object
 943  * to write out.
 944  * We invalidate (remove) all pages from the address space
 945  * for semantic correctness.
 946  *
 947  * If the backing object is a device object with unmanaged pages, then any
 948  * mappings to the specified range of pages must be removed before this
 949  * function is called.
 950  *
 951  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 952  * may start out with a NULL object.
 953  */
 954 boolean_t
 955 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 956     boolean_t syncio, boolean_t invalidate)
 957 {
 958         vm_object_t backing_object;
 959         struct vnode *vp;
 960         struct mount *mp;
 961         int error, flags, fsync_after;
 962         boolean_t res;
 963
 964         if (object == NULL)
 965                 return (TRUE);
 966         res = TRUE;
 967         error = 0;
 968         VM_OBJECT_LOCK(object);
 969         while ((backing_object = object->backing_object) != NULL) {
 970                 VM_OBJECT_LOCK(backing_object);
 971                 offset += object->backing_object_offset;
 972                 VM_OBJECT_UNLOCK(object);
 973                 object = backing_object;
 974                 if (object->size < OFF_TO_IDX(offset + size))
 975                         size = IDX_TO_OFF(object->size) - offset;
 976         }
 977         /*
 978          * Flush pages if writing is allowed, invalidate them
 979          * if invalidation requested.  Pages undergoing I/O
 980          * will be ignored by vm_object_page_remove().
 981          *
 982          * We cannot lock the vnode and then wait for paging
 983          * to complete without deadlocking against vm_fault.
 984          * Instead we simply call vm_object_page_remove() and
 985          * allow it to block internally on a page-by-page
 986          * basis when it encounters pages undergoing async
 987          * I/O.
 988          */
 989         if (object->type == OBJT_VNODE &&
 990             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 991                 int vfslocked;
 992                 vp = object->handle;
 993                 VM_OBJECT_UNLOCK(object);
 994                 (void) vn_start_write(vp, &mp, V_WAIT);
 995                 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 996                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 997                 if (syncio && !invalidate && offset == 0 &&
 998                     OFF_TO_IDX(size) == object->size) {
 999                         /*
1000                          * If syncing the whole mapping of the file,
1001                          * it is faster to schedule all the writes in
1002                          * async mode, also allowing the clustering,
1003                          * and then wait for i/o to complete.
1004                          */
1005                         flags = 0;
1006                         fsync_after = TRUE;
1007                 } else {
1008                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1009                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1010                         fsync_after = FALSE;
1011                 }
1012                 VM_OBJECT_LOCK(object);
1013                 res = vm_object_page_clean(object, offset, offset + size,
1014                     flags);
1015                 VM_OBJECT_UNLOCK(object);
1016                 if (fsync_after)
1017                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1018                 VOP_UNLOCK(vp, 0);
1019                 VFS_UNLOCK_GIANT(vfslocked);
1020                 vn_finished_write(mp);
1021                 if (error != 0)
1022                         res = FALSE;
1023                 VM_OBJECT_LOCK(object);
1024         }
1025         if ((object->type == OBJT_VNODE ||
1026              object->type == OBJT_DEVICE) && invalidate) {
1027                 if (object->type == OBJT_DEVICE)
1028                         /*
1029                          * The option OBJPR_NOTMAPPED must be passed here
1030                          * because vm_object_page_remove() cannot remove
1031                          * unmanaged mappings.
1032                          */
1033                         flags = OBJPR_NOTMAPPED;
1034                 else if (old_msync)
1035                         flags = 0;
1036                 else
1037                         flags = OBJPR_CLEANONLY;
1038                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1039                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1040         }
1041         VM_OBJECT_UNLOCK(object);
1042         return (res);
1043 }
1044
1045 /*
1046  *      vm_object_madvise:
1047  *
1048  *      Implements the madvise function at the object/page level.
1049  *
1050  *      MADV_WILLNEED   (any object)
1051  *
1052  *          Activate the specified pages if they are resident.
1053  *
1054  *      MADV_DONTNEED   (any object)
1055  *
1056  *          Deactivate the specified pages if they are resident.
1057  *
1058  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1059  *                       OBJ_ONEMAPPING only)
1060  *
1061  *          Deactivate and clean the specified pages if they are
1062  *          resident.  This permits the process to reuse the pages
1063  *          without faulting or the kernel to reclaim the pages
1064  *          without I/O.
1065  */
1066 void
1067 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1068     int advise)
1069 {
1070         vm_pindex_t tpindex;
1071         vm_object_t backing_object, tobject;
1072         vm_page_t m;
1073
1074         if (object == NULL)
1075                 return;
1076         VM_OBJECT_LOCK(object);
1077         /*
1078          * Locate and adjust resident pages
1079          */
1080         for (; pindex < end; pindex += 1) {
1081 relookup:
1082                 tobject = object;
1083                 tpindex = pindex;
1084 shadowlookup:
1085                 /*
1086                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1087                  * and those pages must be OBJ_ONEMAPPING.
1088                  */
1089                 if (advise == MADV_FREE) {
1090                         if ((tobject->type != OBJT_DEFAULT &&
1091                              tobject->type != OBJT_SWAP) ||
1092                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1093                                 goto unlock_tobject;
1094                         }
1095                 } else if (tobject->type == OBJT_PHYS)
1096                         goto unlock_tobject;
1097                 m = vm_page_lookup(tobject, tpindex);
1098                 if (m == NULL && advise == MADV_WILLNEED) {
1099                         /*
1100                          * If the page is cached, reactivate it.
1101                          */
1102                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1103                             VM_ALLOC_NOBUSY);
1104                 }
1105                 if (m == NULL) {
1106                         /*
1107                          * There may be swap even if there is no backing page
1108                          */
1109                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1110                                 swap_pager_freespace(tobject, tpindex, 1);
1111                         /*
1112                          * next object
1113                          */
1114                         backing_object = tobject->backing_object;
1115                         if (backing_object == NULL)
1116                                 goto unlock_tobject;
1117                         VM_OBJECT_LOCK(backing_object);
1118                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1119                         if (tobject != object)
1120                                 VM_OBJECT_UNLOCK(tobject);
1121                         tobject = backing_object;
1122                         goto shadowlookup;
1123                 } else if (m->valid != VM_PAGE_BITS_ALL)
1124                         goto unlock_tobject;
1125                 /*
1126                  * If the page is not in a normal state, skip it.
1127                  */
1128                 vm_page_lock(m);
1129                 if (m->hold_count != 0 || m->wire_count != 0) {
1130                         vm_page_unlock(m);
1131                         goto unlock_tobject;
1132                 }
1133                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1134                     ("vm_object_madvise: page %p is fictitious", m));
1135                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1136                     ("vm_object_madvise: page %p is not managed", m));
1137                 if ((m->oflags & VPO_BUSY) || m->busy) {
1138                         if (advise == MADV_WILLNEED) {
1139                                 /*
1140                                  * Reference the page before unlocking and
1141                                  * sleeping so that the page daemon is less
1142                                  * likely to reclaim it.
1143                                  */
1144                                 vm_page_aflag_set(m, PGA_REFERENCED);
1145                         }
1146                         vm_page_unlock(m);
1147                         if (object != tobject)
1148                                 VM_OBJECT_UNLOCK(object);
1149                         m->oflags |= VPO_WANTED;
1150                         msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1151                             0);
1152                         VM_OBJECT_LOCK(object);
1153                         goto relookup;
1154                 }
1155                 if (advise == MADV_WILLNEED) {
1156                         vm_page_activate(m);
1157                 } else if (advise == MADV_DONTNEED) {
1158                         vm_page_dontneed(m);
1159                 } else if (advise == MADV_FREE) {
1160                         /*
1161                          * Mark the page clean.  This will allow the page
1162                          * to be freed up by the system.  However, such pages
1163                          * are often reused quickly by malloc()/free()
1164                          * so we do not do anything that would cause
1165                          * a page fault if we can help it.
1166                          *
1167                          * Specifically, we do not try to actually free
1168                          * the page now nor do we try to put it in the
1169                          * cache (which would cause a page fault on reuse).
1170                          *
1171                          * But we do make the page is freeable as we
1172                          * can without actually taking the step of unmapping
1173                          * it.
1174                          */
1175                         pmap_clear_modify(m);
1176                         m->dirty = 0;
1177                         m->act_count = 0;
1178                         vm_page_dontneed(m);
1179                 }
1180                 vm_page_unlock(m);
1181                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1182                         swap_pager_freespace(tobject, tpindex, 1);
1183 unlock_tobject:
1184                 if (tobject != object)
1185                         VM_OBJECT_UNLOCK(tobject);
1186         }
1187         VM_OBJECT_UNLOCK(object);
1188 }
1189
1190 /*
1191  *      vm_object_shadow:
1192  *
1193  *      Create a new object which is backed by the
1194  *      specified existing object range.  The source
1195  *      object reference is deallocated.
1196  *
1197  *      The new object and offset into that object
1198  *      are returned in the source parameters.
1199  */
1200 void
1201 vm_object_shadow(
1202         vm_object_t *object,    /* IN/OUT */
1203         vm_ooffset_t *offset,   /* IN/OUT */
1204         vm_size_t length)
1205 {
1206         vm_object_t source;
1207         vm_object_t result;
1208
1209         source = *object;
1210
1211         /*
1212          * Don't create the new object if the old object isn't shared.
1213          */
1214         if (source != NULL) {
1215                 VM_OBJECT_LOCK(source);
1216                 if (source->ref_count == 1 &&
1217                     source->handle == NULL &&
1218                     (source->type == OBJT_DEFAULT ||
1219                      source->type == OBJT_SWAP)) {
1220                         VM_OBJECT_UNLOCK(source);
1221                         return;
1222                 }
1223                 VM_OBJECT_UNLOCK(source);
1224         }
1225
1226         /*
1227          * Allocate a new object with the given length.
1228          */
1229         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1230
1231         /*
1232          * The new object shadows the source object, adding a reference to it.
1233          * Our caller changes his reference to point to the new object,
1234          * removing a reference to the source object.  Net result: no change
1235          * of reference count.
1236          *
1237          * Try to optimize the result object's page color when shadowing
1238          * in order to maintain page coloring consistency in the combined
1239          * shadowed object.
1240          */
1241         result->backing_object = source;
1242         /*
1243          * Store the offset into the source object, and fix up the offset into
1244          * the new object.
1245          */
1246         result->backing_object_offset = *offset;
1247         if (source != NULL) {
1248                 VM_OBJECT_LOCK(source);
1249                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1250                 source->shadow_count++;
1251 #if VM_NRESERVLEVEL > 0
1252                 result->flags |= source->flags & OBJ_COLORED;
1253                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1254                     ((1 << (VM_NFREEORDER - 1)) - 1);
1255 #endif
1256                 VM_OBJECT_UNLOCK(source);
1257         }
1258
1259
1260         /*
1261          * Return the new things
1262          */
1263         *offset = 0;
1264         *object = result;
1265 }
1266
1267 /*
1268  *      vm_object_split:
1269  *
1270  * Split the pages in a map entry into a new object.  This affords
1271  * easier removal of unused pages, and keeps object inheritance from
1272  * being a negative impact on memory usage.
1273  */
1274 void
1275 vm_object_split(vm_map_entry_t entry)
1276 {
1277         vm_page_t m, m_next;
1278         vm_object_t orig_object, new_object, source;
1279         vm_pindex_t idx, offidxstart;
1280         vm_size_t size;
1281
1282         orig_object = entry->object.vm_object;
1283         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1284                 return;
1285         if (orig_object->ref_count <= 1)
1286                 return;
1287         VM_OBJECT_UNLOCK(orig_object);
1288
1289         offidxstart = OFF_TO_IDX(entry->offset);
1290         size = atop(entry->end - entry->start);
1291
1292         /*
1293          * If swap_pager_copy() is later called, it will convert new_object
1294          * into a swap object.
1295          */
1296         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1297
1298         /*
1299          * At this point, the new object is still private, so the order in
1300          * which the original and new objects are locked does not matter.
1301          */
1302         VM_OBJECT_LOCK(new_object);
1303         VM_OBJECT_LOCK(orig_object);
1304         source = orig_object->backing_object;
1305         if (source != NULL) {
1306                 VM_OBJECT_LOCK(source);
1307                 if ((source->flags & OBJ_DEAD) != 0) {
1308                         VM_OBJECT_UNLOCK(source);
1309                         VM_OBJECT_UNLOCK(orig_object);
1310                         VM_OBJECT_UNLOCK(new_object);
1311                         vm_object_deallocate(new_object);
1312                         VM_OBJECT_LOCK(orig_object);
1313                         return;
1314                 }
1315                 LIST_INSERT_HEAD(&source->shadow_head,
1316                                   new_object, shadow_list);
1317                 source->shadow_count++;
1318                 vm_object_reference_locked(source);     /* for new_object */
1319                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1320                 VM_OBJECT_UNLOCK(source);
1321                 new_object->backing_object_offset =
1322                         orig_object->backing_object_offset + entry->offset;
1323                 new_object->backing_object = source;
1324         }
1325         if (orig_object->cred != NULL) {
1326                 new_object->cred = orig_object->cred;
1327                 crhold(orig_object->cred);
1328                 new_object->charge = ptoa(size);
1329                 KASSERT(orig_object->charge >= ptoa(size),
1330                     ("orig_object->charge < 0"));
1331                 orig_object->charge -= ptoa(size);
1332         }
1333 retry:
1334         m = vm_page_find_least(orig_object, offidxstart);
1335         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1336             m = m_next) {
1337                 m_next = TAILQ_NEXT(m, listq);
1338
1339                 /*
1340                  * We must wait for pending I/O to complete before we can
1341                  * rename the page.
1342                  *
1343                  * We do not have to VM_PROT_NONE the page as mappings should
1344                  * not be changed by this operation.
1345                  */
1346                 if ((m->oflags & VPO_BUSY) || m->busy) {
1347                         VM_OBJECT_UNLOCK(new_object);
1348                         m->oflags |= VPO_WANTED;
1349                         msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1350                         VM_OBJECT_LOCK(new_object);
1351                         goto retry;
1352                 }
1353                 vm_page_lock(m);
1354                 vm_page_rename(m, new_object, idx);
1355                 vm_page_unlock(m);
1356                 /* page automatically made dirty by rename and cache handled */
1357                 vm_page_busy(m);
1358         }
1359         if (orig_object->type == OBJT_SWAP) {
1360                 /*
1361                  * swap_pager_copy() can sleep, in which case the orig_object's
1362                  * and new_object's locks are released and reacquired.
1363                  */
1364                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1365
1366                 /*
1367                  * Transfer any cached pages from orig_object to new_object.
1368                  */
1369                 if (__predict_false(orig_object->cache != NULL))
1370                         vm_page_cache_transfer(orig_object, offidxstart,
1371                             new_object);
1372         }
1373         VM_OBJECT_UNLOCK(orig_object);
1374         TAILQ_FOREACH(m, &new_object->memq, listq)
1375                 vm_page_wakeup(m);
1376         VM_OBJECT_UNLOCK(new_object);
1377         entry->object.vm_object = new_object;
1378         entry->offset = 0LL;
1379         vm_object_deallocate(orig_object);
1380         VM_OBJECT_LOCK(new_object);
1381 }
1382
1383 #define OBSC_TEST_ALL_SHADOWED  0x0001
1384 #define OBSC_COLLAPSE_NOWAIT    0x0002
1385 #define OBSC_COLLAPSE_WAIT      0x0004
1386
1387 static int
1388 vm_object_backing_scan(vm_object_t object, int op)
1389 {
1390         int r = 1;
1391         vm_page_t p;
1392         vm_object_t backing_object;
1393         vm_pindex_t backing_offset_index;
1394
1395         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1396         VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1397
1398         backing_object = object->backing_object;
1399         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1400
1401         /*
1402          * Initial conditions
1403          */
1404         if (op & OBSC_TEST_ALL_SHADOWED) {
1405                 /*
1406                  * We do not want to have to test for the existence of cache
1407                  * or swap pages in the backing object.  XXX but with the
1408                  * new swapper this would be pretty easy to do.
1409                  *
1410                  * XXX what about anonymous MAP_SHARED memory that hasn't
1411                  * been ZFOD faulted yet?  If we do not test for this, the
1412                  * shadow test may succeed! XXX
1413                  */
1414                 if (backing_object->type != OBJT_DEFAULT) {
1415                         return (0);
1416                 }
1417         }
1418         if (op & OBSC_COLLAPSE_WAIT) {
1419                 vm_object_set_flag(backing_object, OBJ_DEAD);
1420         }
1421
1422         /*
1423          * Our scan
1424          */
1425         p = TAILQ_FIRST(&backing_object->memq);
1426         while (p) {
1427                 vm_page_t next = TAILQ_NEXT(p, listq);
1428                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1429
1430                 if (op & OBSC_TEST_ALL_SHADOWED) {
1431                         vm_page_t pp;
1432
1433                         /*
1434                          * Ignore pages outside the parent object's range
1435                          * and outside the parent object's mapping of the
1436                          * backing object.
1437                          *
1438                          * note that we do not busy the backing object's
1439                          * page.
1440                          */
1441                         if (
1442                             p->pindex < backing_offset_index ||
1443                             new_pindex >= object->size
1444                         ) {
1445                                 p = next;
1446                                 continue;
1447                         }
1448
1449                         /*
1450                          * See if the parent has the page or if the parent's
1451                          * object pager has the page.  If the parent has the
1452                          * page but the page is not valid, the parent's
1453                          * object pager must have the page.
1454                          *
1455                          * If this fails, the parent does not completely shadow
1456                          * the object and we might as well give up now.
1457                          */
1458
1459                         pp = vm_page_lookup(object, new_pindex);
1460                         if (
1461                             (pp == NULL || pp->valid == 0) &&
1462                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1463                         ) {
1464                                 r = 0;
1465                                 break;
1466                         }
1467                 }
1468
1469                 /*
1470                  * Check for busy page
1471                  */
1472                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1473                         vm_page_t pp;
1474
1475                         if (op & OBSC_COLLAPSE_NOWAIT) {
1476                                 if ((p->oflags & VPO_BUSY) ||
1477                                     !p->valid ||
1478                                     p->busy) {
1479                                         p = next;
1480                                         continue;
1481                                 }
1482                         } else if (op & OBSC_COLLAPSE_WAIT) {
1483                                 if ((p->oflags & VPO_BUSY) || p->busy) {
1484                                         VM_OBJECT_UNLOCK(object);
1485                                         p->oflags |= VPO_WANTED;
1486                                         msleep(p, VM_OBJECT_MTX(backing_object),
1487                                             PDROP | PVM, "vmocol", 0);
1488                                         VM_OBJECT_LOCK(object);
1489                                         VM_OBJECT_LOCK(backing_object);
1490                                         /*
1491                                          * If we slept, anything could have
1492                                          * happened.  Since the object is
1493                                          * marked dead, the backing offset
1494                                          * should not have changed so we
1495                                          * just restart our scan.
1496                                          */
1497                                         p = TAILQ_FIRST(&backing_object->memq);
1498                                         continue;
1499                                 }
1500                         }
1501
1502                         KASSERT(
1503                             p->object == backing_object,
1504                             ("vm_object_backing_scan: object mismatch")
1505                         );
1506
1507                         /*
1508                          * Destroy any associated swap
1509                          */
1510                         if (backing_object->type == OBJT_SWAP) {
1511                                 swap_pager_freespace(
1512                                     backing_object,
1513                                     p->pindex,
1514                                     1
1515                                 );
1516                         }
1517
1518                         if (
1519                             p->pindex < backing_offset_index ||
1520                             new_pindex >= object->size
1521                         ) {
1522                                 /*
1523                                  * Page is out of the parent object's range, we
1524                                  * can simply destroy it.
1525                                  */
1526                                 vm_page_lock(p);
1527                                 KASSERT(!pmap_page_is_mapped(p),
1528                                     ("freeing mapped page %p", p));
1529                                 if (p->wire_count == 0)
1530                                         vm_page_free(p);
1531                                 else
1532                                         vm_page_remove(p);
1533                                 vm_page_unlock(p);
1534                                 p = next;
1535                                 continue;
1536                         }
1537
1538                         pp = vm_page_lookup(object, new_pindex);
1539                         if (
1540                             (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1541                             (pp != NULL && pp->valid == 0)
1542                         ) {
1543                                 /*
1544                                  * The page in the parent is not (yet) valid.
1545                                  * We don't know anything about the state of
1546                                  * the original page.  It might be mapped,
1547                                  * so we must avoid the next if here.
1548                                  *
1549                                  * This is due to a race in vm_fault() where
1550                                  * we must unbusy the original (backing_obj)
1551                                  * page before we can (re)lock the parent.
1552                                  * Hence we can get here.
1553                                  */
1554                                 p = next;
1555                                 continue;
1556                         }
1557                         if (
1558                             pp != NULL ||
1559                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1560                         ) {
1561                                 /*
1562                                  * page already exists in parent OR swap exists
1563                                  * for this location in the parent.  Destroy
1564                                  * the original page from the backing object.
1565                                  *
1566                                  * Leave the parent's page alone
1567                                  */
1568                                 vm_page_lock(p);
1569                                 KASSERT(!pmap_page_is_mapped(p),
1570                                     ("freeing mapped page %p", p));
1571                                 if (p->wire_count == 0)
1572                                         vm_page_free(p);
1573                                 else
1574                                         vm_page_remove(p);
1575                                 vm_page_unlock(p);
1576                                 p = next;
1577                                 continue;
1578                         }
1579
1580 #if VM_NRESERVLEVEL > 0
1581                         /*
1582                          * Rename the reservation.
1583                          */
1584                         vm_reserv_rename(p, object, backing_object,
1585                             backing_offset_index);
1586 #endif
1587
1588                         /*
1589                          * Page does not exist in parent, rename the
1590                          * page from the backing object to the main object.
1591                          *
1592                          * If the page was mapped to a process, it can remain
1593                          * mapped through the rename.
1594                          */
1595                         vm_page_lock(p);
1596                         vm_page_rename(p, object, new_pindex);
1597                         vm_page_unlock(p);
1598                         /* page automatically made dirty by rename */
1599                 }
1600                 p = next;
1601         }
1602         return (r);
1603 }
1604
1605
1606 /*
1607  * this version of collapse allows the operation to occur earlier and
1608  * when paging_in_progress is true for an object...  This is not a complete
1609  * operation, but should plug 99.9% of the rest of the leaks.
1610  */
1611 static void
1612 vm_object_qcollapse(vm_object_t object)
1613 {
1614         vm_object_t backing_object = object->backing_object;
1615
1616         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1617         VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1618
1619         if (backing_object->ref_count != 1)
1620                 return;
1621
1622         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1623 }
1624
1625 /*
1626  *      vm_object_collapse:
1627  *
1628  *      Collapse an object with the object backing it.
1629  *      Pages in the backing object are moved into the
1630  *      parent, and the backing object is deallocated.
1631  */
1632 void
1633 vm_object_collapse(vm_object_t object)
1634 {
1635         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1636
1637         while (TRUE) {
1638                 vm_object_t backing_object;
1639
1640                 /*
1641                  * Verify that the conditions are right for collapse:
1642                  *
1643                  * The object exists and the backing object exists.
1644                  */
1645                 if ((backing_object = object->backing_object) == NULL)
1646                         break;
1647
1648                 /*
1649                  * we check the backing object first, because it is most likely
1650                  * not collapsable.
1651                  */
1652                 VM_OBJECT_LOCK(backing_object);
1653                 if (backing_object->handle != NULL ||
1654                     (backing_object->type != OBJT_DEFAULT &&
1655                      backing_object->type != OBJT_SWAP) ||
1656                     (backing_object->flags & OBJ_DEAD) ||
1657                     object->handle != NULL ||
1658                     (object->type != OBJT_DEFAULT &&
1659                      object->type != OBJT_SWAP) ||
1660                     (object->flags & OBJ_DEAD)) {
1661                         VM_OBJECT_UNLOCK(backing_object);
1662                         break;
1663                 }
1664
1665                 if (
1666                     object->paging_in_progress != 0 ||
1667                     backing_object->paging_in_progress != 0
1668                 ) {
1669                         vm_object_qcollapse(object);
1670                         VM_OBJECT_UNLOCK(backing_object);
1671                         break;
1672                 }
1673                 /*
1674                  * We know that we can either collapse the backing object (if
1675                  * the parent is the only reference to it) or (perhaps) have
1676                  * the parent bypass the object if the parent happens to shadow
1677                  * all the resident pages in the entire backing object.
1678                  *
1679                  * This is ignoring pager-backed pages such as swap pages.
1680                  * vm_object_backing_scan fails the shadowing test in this
1681                  * case.
1682                  */
1683                 if (backing_object->ref_count == 1) {
1684                         /*
1685                          * If there is exactly one reference to the backing
1686                          * object, we can collapse it into the parent.
1687                          */
1688                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1689
1690 #if VM_NRESERVLEVEL > 0
1691                         /*
1692                          * Break any reservations from backing_object.
1693                          */
1694                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1695                                 vm_reserv_break_all(backing_object);
1696 #endif
1697
1698                         /*
1699                          * Move the pager from backing_object to object.
1700                          */
1701                         if (backing_object->type == OBJT_SWAP) {
1702                                 /*
1703                                  * swap_pager_copy() can sleep, in which case
1704                                  * the backing_object's and object's locks are
1705                                  * released and reacquired.
1706                                  */
1707                                 swap_pager_copy(
1708                                     backing_object,
1709                                     object,
1710                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1711
1712                                 /*
1713                                  * Free any cached pages from backing_object.
1714                                  */
1715                                 if (__predict_false(backing_object->cache != NULL))
1716                                         vm_page_cache_free(backing_object, 0, 0);
1717                         }
1718                         /*
1719                          * Object now shadows whatever backing_object did.
1720                          * Note that the reference to
1721                          * backing_object->backing_object moves from within
1722                          * backing_object to within object.
1723                          */
1724                         LIST_REMOVE(object, shadow_list);
1725                         backing_object->shadow_count--;
1726                         if (backing_object->backing_object) {
1727                                 VM_OBJECT_LOCK(backing_object->backing_object);
1728                                 LIST_REMOVE(backing_object, shadow_list);
1729                                 LIST_INSERT_HEAD(
1730                                     &backing_object->backing_object->shadow_head,
1731                                     object, shadow_list);
1732                                 /*
1733                                  * The shadow_count has not changed.
1734                                  */
1735                                 VM_OBJECT_UNLOCK(backing_object->backing_object);
1736                         }
1737                         object->backing_object = backing_object->backing_object;
1738                         object->backing_object_offset +=
1739                             backing_object->backing_object_offset;
1740
1741                         /*
1742                          * Discard backing_object.
1743                          *
1744                          * Since the backing object has no pages, no pager left,
1745                          * and no object references within it, all that is
1746                          * necessary is to dispose of it.
1747                          */
1748                         KASSERT(backing_object->ref_count == 1, (
1749 "backing_object %p was somehow re-referenced during collapse!",
1750                             backing_object));
1751                         VM_OBJECT_UNLOCK(backing_object);
1752                         vm_object_destroy(backing_object);
1753
1754                         object_collapses++;
1755                 } else {
1756                         vm_object_t new_backing_object;
1757
1758                         /*
1759                          * If we do not entirely shadow the backing object,
1760                          * there is nothing we can do so we give up.
1761                          */
1762                         if (object->resident_page_count != object->size &&
1763                             vm_object_backing_scan(object,
1764                             OBSC_TEST_ALL_SHADOWED) == 0) {
1765                                 VM_OBJECT_UNLOCK(backing_object);
1766                                 break;
1767                         }
1768
1769                         /*
1770                          * Make the parent shadow the next object in the
1771                          * chain.  Deallocating backing_object will not remove
1772                          * it, since its reference count is at least 2.
1773                          */
1774                         LIST_REMOVE(object, shadow_list);
1775                         backing_object->shadow_count--;
1776
1777                         new_backing_object = backing_object->backing_object;
1778                         if ((object->backing_object = new_backing_object) != NULL) {
1779                                 VM_OBJECT_LOCK(new_backing_object);
1780                                 LIST_INSERT_HEAD(
1781                                     &new_backing_object->shadow_head,
1782                                     object,
1783                                     shadow_list
1784                                 );
1785                                 new_backing_object->shadow_count++;
1786                                 vm_object_reference_locked(new_backing_object);
1787                                 VM_OBJECT_UNLOCK(new_backing_object);
1788                                 object->backing_object_offset +=
1789                                         backing_object->backing_object_offset;
1790                         }
1791
1792                         /*
1793                          * Drop the reference count on backing_object. Since
1794                          * its ref_count was at least 2, it will not vanish.
1795                          */
1796                         backing_object->ref_count--;
1797                         VM_OBJECT_UNLOCK(backing_object);
1798                         object_bypasses++;
1799                 }
1800
1801                 /*
1802                  * Try again with this object's new backing object.
1803                  */
1804         }
1805 }
1806
1807 /*
1808  *      vm_object_page_remove:
1809  *
1810  *      For the given object, either frees or invalidates each of the
1811  *      specified pages.  In general, a page is freed.  However, if a page is
1812  *      wired for any reason other than the existence of a managed, wired
1813  *      mapping, then it may be invalidated but not removed from the object.
1814  *      Pages are specified by the given range ["start", "end") and the option
1815  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1816  *      extends from "start" to the end of the object.  If the option
1817  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1818  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1819  *      specified, then the pages within the specified range must have no
1820  *      mappings.  Otherwise, if this option is not specified, any mappings to
1821  *      the specified pages are removed before the pages are freed or
1822  *      invalidated.
1823  *
1824  *      In general, this operation should only be performed on objects that
1825  *      contain managed pages.  There are, however, two exceptions.  First, it
1826  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1827  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1828  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1829  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1830  *
1831  *      The object must be locked.
1832  */
1833 void
1834 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1835     int options)
1836 {
1837         vm_page_t p, next;
1838         int wirings;
1839
1840         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1841         KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1842             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1843             ("vm_object_page_remove: illegal options for object %p", object));
1844         if (object->resident_page_count == 0)
1845                 goto skipmemq;
1846         vm_object_pip_add(object, 1);
1847 again:
1848         p = vm_page_find_least(object, start);
1849
1850         /*
1851          * Here, the variable "p" is either (1) the page with the least pindex
1852          * greater than or equal to the parameter "start" or (2) NULL.
1853          */
1854         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1855                 next = TAILQ_NEXT(p, listq);
1856
1857                 /*
1858                  * If the page is wired for any reason besides the existence
1859                  * of managed, wired mappings, then it cannot be freed.  For
1860                  * example, fictitious pages, which represent device memory,
1861                  * are inherently wired and cannot be freed.  They can,
1862                  * however, be invalidated if the option OBJPR_CLEANONLY is
1863                  * not specified.
1864                  */
1865                 vm_page_lock(p);
1866                 if ((wirings = p->wire_count) != 0 &&
1867                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1868                         if ((options & OBJPR_NOTMAPPED) == 0) {
1869                                 pmap_remove_all(p);
1870                                 /* Account for removal of wired mappings. */
1871                                 if (wirings != 0)
1872                                         p->wire_count -= wirings;
1873                         }
1874                         if ((options & OBJPR_CLEANONLY) == 0) {
1875                                 p->valid = 0;
1876                                 vm_page_undirty(p);
1877                         }
1878                         vm_page_unlock(p);
1879                         continue;
1880                 }
1881                 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1882                         goto again;
1883                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1884                     ("vm_object_page_remove: page %p is fictitious", p));
1885                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1886                         if ((options & OBJPR_NOTMAPPED) == 0)
1887                                 pmap_remove_write(p);
1888                         if (p->dirty) {
1889                                 vm_page_unlock(p);
1890                                 continue;
1891                         }
1892                 }
1893                 if ((options & OBJPR_NOTMAPPED) == 0) {
1894                         pmap_remove_all(p);
1895                         /* Account for removal of wired mappings. */
1896                         if (wirings != 0)
1897                                 p->wire_count -= wirings;
1898                 }
1899                 vm_page_free(p);
1900                 vm_page_unlock(p);
1901         }
1902         vm_object_pip_wakeup(object);
1903 skipmemq:
1904         if (__predict_false(object->cache != NULL))
1905                 vm_page_cache_free(object, start, end);
1906 }
1907
1908 /*
1909  *      vm_object_page_cache:
1910  *
1911  *      For the given object, attempt to move the specified clean
1912  *      pages to the cache queue.  If a page is wired for any reason,
1913  *      then it will not be changed.  Pages are specified by the given
1914  *      range ["start", "end").  As a special case, if "end" is zero,
1915  *      then the range extends from "start" to the end of the object.
1916  *      Any mappings to the specified pages are removed before the
1917  *      pages are moved to the cache queue.
1918  *
1919  *      This operation should only be performed on objects that
1920  *      contain managed pages.
1921  *
1922  *      The object must be locked.
1923  */
1924 void
1925 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1926 {
1927         struct mtx *mtx, *new_mtx;
1928         vm_page_t p, next;
1929
1930         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1931         KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1932             object->type != OBJT_PHYS),
1933             ("vm_object_page_cache: illegal object %p", object));
1934         if (object->resident_page_count == 0)
1935                 return;
1936         p = vm_page_find_least(object, start);
1937
1938         /*
1939          * Here, the variable "p" is either (1) the page with the least pindex
1940          * greater than or equal to the parameter "start" or (2) NULL.
1941          */
1942         mtx = NULL;
1943         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1944                 next = TAILQ_NEXT(p, listq);
1945
1946                 /*
1947                  * Avoid releasing and reacquiring the same page lock.
1948                  */
1949                 new_mtx = vm_page_lockptr(p);
1950                 if (mtx != new_mtx) {
1951                         if (mtx != NULL)
1952                                 mtx_unlock(mtx);
1953                         mtx = new_mtx;
1954                         mtx_lock(mtx);
1955                 }
1956                 vm_page_try_to_cache(p);
1957         }
1958         if (mtx != NULL)
1959                 mtx_unlock(mtx);
1960 }
1961
1962 /*
1963  *      Populate the specified range of the object with valid pages.  Returns
1964  *      TRUE if the range is successfully populated and FALSE otherwise.
1965  *
1966  *      Note: This function should be optimized to pass a larger array of
1967  *      pages to vm_pager_get_pages() before it is applied to a non-
1968  *      OBJT_DEVICE object.
1969  *
1970  *      The object must be locked.
1971  */
1972 boolean_t
1973 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1974 {
1975         vm_page_t m, ma[1];
1976         vm_pindex_t pindex;
1977         int rv;
1978
1979         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1980         for (pindex = start; pindex < end; pindex++) {
1981                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1982                     VM_ALLOC_RETRY);
1983                 if (m->valid != VM_PAGE_BITS_ALL) {
1984                         ma[0] = m;
1985                         rv = vm_pager_get_pages(object, ma, 1, 0);
1986                         m = vm_page_lookup(object, pindex);
1987                         if (m == NULL)
1988                                 break;
1989                         if (rv != VM_PAGER_OK) {
1990                                 vm_page_lock(m);
1991                                 vm_page_free(m);
1992                                 vm_page_unlock(m);
1993                                 break;
1994                         }
1995                 }
1996                 /*
1997                  * Keep "m" busy because a subsequent iteration may unlock
1998                  * the object.
1999                  */
2000         }
2001         if (pindex > start) {
2002                 m = vm_page_lookup(object, start);
2003                 while (m != NULL && m->pindex < pindex) {
2004                         vm_page_wakeup(m);
2005                         m = TAILQ_NEXT(m, listq);
2006                 }
2007         }
2008         return (pindex == end);
2009 }
2010
2011 /*
2012  *      Routine:        vm_object_coalesce
2013  *      Function:       Coalesces two objects backing up adjoining
2014  *                      regions of memory into a single object.
2015  *
2016  *      returns TRUE if objects were combined.
2017  *
2018  *      NOTE:   Only works at the moment if the second object is NULL -
2019  *              if it's not, which object do we lock first?
2020  *
2021  *      Parameters:
2022  *              prev_object     First object to coalesce
2023  *              prev_offset     Offset into prev_object
2024  *              prev_size       Size of reference to prev_object
2025  *              next_size       Size of reference to the second object
2026  *              reserved        Indicator that extension region has
2027  *                              swap accounted for
2028  *
2029  *      Conditions:
2030  *      The object must *not* be locked.
2031  */
2032 boolean_t
2033 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2034     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2035 {
2036         vm_pindex_t next_pindex;
2037
2038         if (prev_object == NULL)
2039                 return (TRUE);
2040         VM_OBJECT_LOCK(prev_object);
2041         if (prev_object->type != OBJT_DEFAULT &&
2042             prev_object->type != OBJT_SWAP) {
2043                 VM_OBJECT_UNLOCK(prev_object);
2044                 return (FALSE);
2045         }
2046
2047         /*
2048          * Try to collapse the object first
2049          */
2050         vm_object_collapse(prev_object);
2051
2052         /*
2053          * Can't coalesce if: . more than one reference . paged out . shadows
2054          * another object . has a copy elsewhere (any of which mean that the
2055          * pages not mapped to prev_entry may be in use anyway)
2056          */
2057         if (prev_object->backing_object != NULL) {
2058                 VM_OBJECT_UNLOCK(prev_object);
2059                 return (FALSE);
2060         }
2061
2062         prev_size >>= PAGE_SHIFT;
2063         next_size >>= PAGE_SHIFT;
2064         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2065
2066         if ((prev_object->ref_count > 1) &&
2067             (prev_object->size != next_pindex)) {
2068                 VM_OBJECT_UNLOCK(prev_object);
2069                 return (FALSE);
2070         }
2071
2072         /*
2073          * Account for the charge.
2074          */
2075         if (prev_object->cred != NULL) {
2076
2077                 /*
2078                  * If prev_object was charged, then this mapping,
2079                  * althought not charged now, may become writable
2080                  * later. Non-NULL cred in the object would prevent
2081                  * swap reservation during enabling of the write
2082                  * access, so reserve swap now. Failed reservation
2083                  * cause allocation of the separate object for the map
2084                  * entry, and swap reservation for this entry is
2085                  * managed in appropriate time.
2086                  */
2087                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2088                     prev_object->cred)) {
2089                         return (FALSE);
2090                 }
2091                 prev_object->charge += ptoa(next_size);
2092         }
2093
2094         /*
2095          * Remove any pages that may still be in the object from a previous
2096          * deallocation.
2097          */
2098         if (next_pindex < prev_object->size) {
2099                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2100                     next_size, 0);
2101                 if (prev_object->type == OBJT_SWAP)
2102                         swap_pager_freespace(prev_object,
2103                                              next_pindex, next_size);
2104 #if 0
2105                 if (prev_object->cred != NULL) {
2106                         KASSERT(prev_object->charge >=
2107                             ptoa(prev_object->size - next_pindex),
2108                             ("object %p overcharged 1 %jx %jx", prev_object,
2109                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2110                         prev_object->charge -= ptoa(prev_object->size -
2111                             next_pindex);
2112                 }
2113 #endif
2114         }
2115
2116         /*
2117          * Extend the object if necessary.
2118          */
2119         if (next_pindex + next_size > prev_object->size)
2120                 prev_object->size = next_pindex + next_size;
2121
2122         VM_OBJECT_UNLOCK(prev_object);
2123         return (TRUE);
2124 }
2125
2126 void
2127 vm_object_set_writeable_dirty(vm_object_t object)
2128 {
2129
2130         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2131         if (object->type != OBJT_VNODE)
2132                 return;
2133         object->generation++;
2134         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2135                 return;
2136         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2137 }
2138
2139 #include "opt_ddb.h"
2140 #ifdef DDB
2141 #include <sys/kernel.h>
2142
2143 #include <sys/cons.h>
2144
2145 #include <ddb/ddb.h>
2146
2147 static int
2148 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2149 {
2150         vm_map_t tmpm;
2151         vm_map_entry_t tmpe;
2152         vm_object_t obj;
2153         int entcount;
2154
2155         if (map == 0)
2156                 return 0;
2157
2158         if (entry == 0) {
2159                 tmpe = map->header.next;
2160                 entcount = map->nentries;
2161                 while (entcount-- && (tmpe != &map->header)) {
2162                         if (_vm_object_in_map(map, object, tmpe)) {
2163                                 return 1;
2164                         }
2165                         tmpe = tmpe->next;
2166                 }
2167         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2168                 tmpm = entry->object.sub_map;
2169                 tmpe = tmpm->header.next;
2170                 entcount = tmpm->nentries;
2171                 while (entcount-- && tmpe != &tmpm->header) {
2172                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2173                                 return 1;
2174                         }
2175                         tmpe = tmpe->next;
2176                 }
2177         } else if ((obj = entry->object.vm_object) != NULL) {
2178                 for (; obj; obj = obj->backing_object)
2179                         if (obj == object) {
2180                                 return 1;
2181                         }
2182         }
2183         return 0;
2184 }
2185
2186 static int
2187 vm_object_in_map(vm_object_t object)
2188 {
2189         struct proc *p;
2190
2191         /* sx_slock(&allproc_lock); */
2192         FOREACH_PROC_IN_SYSTEM(p) {
2193                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2194                         continue;
2195                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2196                         /* sx_sunlock(&allproc_lock); */
2197                         return 1;
2198                 }
2199         }
2200         /* sx_sunlock(&allproc_lock); */
2201         if (_vm_object_in_map(kernel_map, object, 0))
2202                 return 1;
2203         if (_vm_object_in_map(kmem_map, object, 0))
2204                 return 1;
2205         if (_vm_object_in_map(pager_map, object, 0))
2206                 return 1;
2207         if (_vm_object_in_map(buffer_map, object, 0))
2208                 return 1;
2209         return 0;
2210 }
2211
2212 DB_SHOW_COMMAND(vmochk, vm_object_check)
2213 {
2214         vm_object_t object;
2215
2216         /*
2217          * make sure that internal objs are in a map somewhere
2218          * and none have zero ref counts.
2219          */
2220         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2221                 if (object->handle == NULL &&
2222                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2223                         if (object->ref_count == 0) {
2224                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2225                                         (long)object->size);
2226                         }
2227                         if (!vm_object_in_map(object)) {
2228                                 db_printf(
2229                         "vmochk: internal obj is not in a map: "
2230                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2231                                     object->ref_count, (u_long)object->size,
2232                                     (u_long)object->size,
2233                                     (void *)object->backing_object);
2234                         }
2235                 }
2236         }
2237 }
2238
2239 /*
2240  *      vm_object_print:        [ debug ]
2241  */
2242 DB_SHOW_COMMAND(object, vm_object_print_static)
2243 {
2244         /* XXX convert args. */
2245         vm_object_t object = (vm_object_t)addr;
2246         boolean_t full = have_addr;
2247
2248         vm_page_t p;
2249
2250         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2251 #define count   was_count
2252
2253         int count;
2254
2255         if (object == NULL)
2256                 return;
2257
2258         db_iprintf(
2259             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2260             object, (int)object->type, (uintmax_t)object->size,
2261             object->resident_page_count, object->ref_count, object->flags,
2262             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2263         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2264             object->shadow_count,
2265             object->backing_object ? object->backing_object->ref_count : 0,
2266             object->backing_object, (uintmax_t)object->backing_object_offset);
2267
2268         if (!full)
2269                 return;
2270
2271         db_indent += 2;
2272         count = 0;
2273         TAILQ_FOREACH(p, &object->memq, listq) {
2274                 if (count == 0)
2275                         db_iprintf("memory:=");
2276                 else if (count == 6) {
2277                         db_printf("\n");
2278                         db_iprintf(" ...");
2279                         count = 0;
2280                 } else
2281                         db_printf(",");
2282                 count++;
2283
2284                 db_printf("(off=0x%jx,page=0x%jx)",
2285                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2286         }
2287         if (count != 0)
2288                 db_printf("\n");
2289         db_indent -= 2;
2290 }
2291
2292 /* XXX. */
2293 #undef count
2294
2295 /* XXX need this non-static entry for calling from vm_map_print. */
2296 void
2297 vm_object_print(
2298         /* db_expr_t */ long addr,
2299         boolean_t have_addr,
2300         /* db_expr_t */ long count,
2301         char *modif)
2302 {
2303         vm_object_print_static(addr, have_addr, count, modif);
2304 }
2305
2306 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2307 {
2308         vm_object_t object;
2309         vm_pindex_t fidx;
2310         vm_paddr_t pa;
2311         vm_page_t m, prev_m;
2312         int rcount, nl, c;
2313
2314         nl = 0;
2315         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2316                 db_printf("new object: %p\n", (void *)object);
2317                 if (nl > 18) {
2318                         c = cngetc();
2319                         if (c != ' ')
2320                                 return;
2321                         nl = 0;
2322                 }
2323                 nl++;
2324                 rcount = 0;
2325                 fidx = 0;
2326                 pa = -1;
2327                 TAILQ_FOREACH(m, &object->memq, listq) {
2328                         if (m->pindex > 128)
2329                                 break;
2330                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2331                             prev_m->pindex + 1 != m->pindex) {
2332                                 if (rcount) {
2333                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2334                                                 (long)fidx, rcount, (long)pa);
2335                                         if (nl > 18) {
2336                                                 c = cngetc();
2337                                                 if (c != ' ')
2338                                                         return;
2339                                                 nl = 0;
2340                                         }
2341                                         nl++;
2342                                         rcount = 0;
2343                                 }
2344                         }
2345                         if (rcount &&
2346                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2347                                 ++rcount;
2348                                 continue;
2349                         }
2350                         if (rcount) {
2351                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2352                                         (long)fidx, rcount, (long)pa);
2353                                 if (nl > 18) {
2354                                         c = cngetc();
2355                                         if (c != ' ')
2356                                                 return;
2357                                         nl = 0;
2358                                 }
2359                                 nl++;
2360                         }
2361                         fidx = m->pindex;
2362                         pa = VM_PAGE_TO_PHYS(m);
2363                         rcount = 1;
2364                 }
2365                 if (rcount) {
2366                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2367                                 (long)fidx, rcount, (long)pa);
2368                         if (nl > 18) {
2369                                 c = cngetc();
2370                                 if (c != ' ')
2371                                         return;
2372                                 nl = 0;
2373                         }
2374                         nl++;
2375                 }
2376         }
2377 }
2378 #endif /* DDB */