sys/vm/vm_object.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * The Mach Operating System project at Carnegie-Mellon University.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 4. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  33  *
  34  *
  35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  36  * All rights reserved.
  37  *
  38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  39  *
  40  * Permission to use, copy, modify and distribute this software and
  41  * its documentation is hereby granted, provided that both the copyright
  42  * notice and this permission notice appear in all copies of the
  43  * software, derivative works or modified versions, and any portions
  44  * thereof, and that both notices appear in supporting documentation.
  45  *
  46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  49  *
  50  * Carnegie Mellon requests users of this software to return to
  51  *
  52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  53  *  School of Computer Science
  54  *  Carnegie Mellon University
  55  *  Pittsburgh PA 15213-3890
  56  *
  57  * any improvements or extensions that they make and grant Carnegie the
  58  * rights to redistribute these changes.
  59  */
  60
  61 /*
  62  *      Virtual memory object module.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_vm.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/lock.h>
  73 #include <sys/mman.h>
  74 #include <sys/mount.h>
  75 #include <sys/kernel.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/mutex.h>
  78 #include <sys/proc.h>           /* for curproc, pageproc */
  79 #include <sys/socket.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/vnode.h>
  82 #include <sys/vmmeter.h>
  83 #include <sys/sx.h>
  84
  85 #include <vm/vm.h>
  86 #include <vm/vm_param.h>
  87 #include <vm/pmap.h>
  88 #include <vm/vm_map.h>
  89 #include <vm/vm_object.h>
  90 #include <vm/vm_page.h>
  91 #include <vm/vm_pageout.h>
  92 #include <vm/vm_pager.h>
  93 #include <vm/swap_pager.h>
  94 #include <vm/vm_kern.h>
  95 #include <vm/vm_extern.h>
  96 #include <vm/vm_reserv.h>
  97 #include <vm/uma.h>
  98
  99 static int old_msync;
 100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 101     "Use old (insecure) msync behavior");
 102
 103 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 104                     int pagerflags, int flags, int *clearobjflags);
 105 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 106                     int *clearobjflags);
 107 static void     vm_object_qcollapse(vm_object_t object);
 108 static void     vm_object_vndeallocate(vm_object_t object);
 109
 110 /*
 111  *      Virtual memory objects maintain the actual data
 112  *      associated with allocated virtual memory.  A given
 113  *      page of memory exists within exactly one object.
 114  *
 115  *      An object is only deallocated when all "references"
 116  *      are given up.  Only one "reference" to a given
 117  *      region of an object should be writeable.
 118  *
 119  *      Associated with each object is a list of all resident
 120  *      memory pages belonging to that object; this list is
 121  *      maintained by the "vm_page" module, and locked by the object's
 122  *      lock.
 123  *
 124  *      Each object also records a "pager" routine which is
 125  *      used to retrieve (and store) pages to the proper backing
 126  *      storage.  In addition, objects may be backed by other
 127  *      objects from which they were virtual-copied.
 128  *
 129  *      The only items within the object structure which are
 130  *      modified after time of creation are:
 131  *              reference count         locked by object's lock
 132  *              pager routine           locked by object's lock
 133  *
 134  */
 135
 136 struct object_q vm_object_list;
 137 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 138
 139 struct vm_object kernel_object_store;
 140 struct vm_object kmem_object_store;
 141
 142 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
 143     "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, int *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 = 0;
 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 void
 811 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 812     int flags)
 813 {
 814         vm_page_t np, p;
 815         vm_pindex_t pi, tend, tstart;
 816         int clearobjflags, curgeneration, n, pagerflags;
 817
 818         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 819         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 820         KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
 821         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 822             object->resident_page_count == 0)
 823                 return;
 824
 825         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 826             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 827         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 828
 829         tstart = OFF_TO_IDX(start);
 830         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 831         clearobjflags = tstart == 0 && tend >= object->size;
 832
 833 rescan:
 834         curgeneration = object->generation;
 835
 836         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 837                 pi = p->pindex;
 838                 if (pi >= tend)
 839                         break;
 840                 np = TAILQ_NEXT(p, listq);
 841                 if (p->valid == 0)
 842                         continue;
 843                 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
 844                         if (object->generation != curgeneration)
 845                                 goto rescan;
 846                         np = vm_page_find_least(object, pi);
 847                         continue;
 848                 }
 849                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 850                         continue;
 851
 852                 n = vm_object_page_collect_flush(object, p, pagerflags,
 853                     flags, &clearobjflags);
 854                 if (object->generation != curgeneration)
 855                         goto rescan;
 856
 857                 /*
 858                  * If the VOP_PUTPAGES() did a truncated write, so
 859                  * that even the first page of the run is not fully
 860                  * written, vm_pageout_flush() returns 0 as the run
 861                  * length.  Since the condition that caused truncated
 862                  * write may be permanent, e.g. exhausted free space,
 863                  * accepting n == 0 would cause an infinite loop.
 864                  *
 865                  * Forwarding the iterator leaves the unwritten page
 866                  * behind, but there is not much we can do there if
 867                  * filesystem refuses to write it.
 868                  */
 869                 if (n == 0)
 870                         n = 1;
 871                 np = vm_page_find_least(object, pi + n);
 872         }
 873 #if 0
 874         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 875 #endif
 876
 877         if (clearobjflags)
 878                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 879 }
 880
 881 static int
 882 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 883     int flags, int *clearobjflags)
 884 {
 885         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 886         int count, i, mreq, runlen;
 887
 888         mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
 889         vm_page_lock_assert(p, MA_NOTOWNED);
 890         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
 891
 892         count = 1;
 893         mreq = 0;
 894
 895         for (tp = p; count < vm_pageout_page_count; count++) {
 896                 tp = vm_page_next(tp);
 897                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 898                         break;
 899                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 900                         break;
 901         }
 902
 903         for (p_first = p; count < vm_pageout_page_count; count++) {
 904                 tp = vm_page_prev(p_first);
 905                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
 906                         break;
 907                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 908                         break;
 909                 p_first = tp;
 910                 mreq++;
 911         }
 912
 913         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 914                 ma[i] = tp;
 915
 916         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
 917         return (runlen);
 918 }
 919
 920 /*
 921  * Note that there is absolutely no sense in writing out
 922  * anonymous objects, so we track down the vnode object
 923  * to write out.
 924  * We invalidate (remove) all pages from the address space
 925  * for semantic correctness.
 926  *
 927  * If the backing object is a device object with unmanaged pages, then any
 928  * mappings to the specified range of pages must be removed before this
 929  * function is called.
 930  *
 931  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 932  * may start out with a NULL object.
 933  */
 934 void
 935 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 936     boolean_t syncio, boolean_t invalidate)
 937 {
 938         vm_object_t backing_object;
 939         struct vnode *vp;
 940         struct mount *mp;
 941         int flags;
 942
 943         if (object == NULL)
 944                 return;
 945         VM_OBJECT_LOCK(object);
 946         while ((backing_object = object->backing_object) != NULL) {
 947                 VM_OBJECT_LOCK(backing_object);
 948                 offset += object->backing_object_offset;
 949                 VM_OBJECT_UNLOCK(object);
 950                 object = backing_object;
 951                 if (object->size < OFF_TO_IDX(offset + size))
 952                         size = IDX_TO_OFF(object->size) - offset;
 953         }
 954         /*
 955          * Flush pages if writing is allowed, invalidate them
 956          * if invalidation requested.  Pages undergoing I/O
 957          * will be ignored by vm_object_page_remove().
 958          *
 959          * We cannot lock the vnode and then wait for paging
 960          * to complete without deadlocking against vm_fault.
 961          * Instead we simply call vm_object_page_remove() and
 962          * allow it to block internally on a page-by-page
 963          * basis when it encounters pages undergoing async
 964          * I/O.
 965          */
 966         if (object->type == OBJT_VNODE &&
 967             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
 968                 int vfslocked;
 969                 vp = object->handle;
 970                 VM_OBJECT_UNLOCK(object);
 971                 (void) vn_start_write(vp, &mp, V_WAIT);
 972                 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 973                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 974                 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
 975                 flags |= invalidate ? OBJPC_INVAL : 0;
 976                 VM_OBJECT_LOCK(object);
 977                 vm_object_page_clean(object, offset, offset + size, flags);
 978                 VM_OBJECT_UNLOCK(object);
 979                 VOP_UNLOCK(vp, 0);
 980                 VFS_UNLOCK_GIANT(vfslocked);
 981                 vn_finished_write(mp);
 982                 VM_OBJECT_LOCK(object);
 983         }
 984         if ((object->type == OBJT_VNODE ||
 985              object->type == OBJT_DEVICE) && invalidate) {
 986                 if (object->type == OBJT_DEVICE)
 987                         /*
 988                          * The option OBJPR_NOTMAPPED must be passed here
 989                          * because vm_object_page_remove() cannot remove
 990                          * unmanaged mappings.
 991                          */
 992                         flags = OBJPR_NOTMAPPED;
 993                 else if (old_msync)
 994                         flags = 0;
 995                 else
 996                         flags = OBJPR_CLEANONLY;
 997                 vm_object_page_remove(object, OFF_TO_IDX(offset),
 998                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
 999         }
1000         VM_OBJECT_UNLOCK(object);
1001 }
1002
1003 /*
1004  *      vm_object_madvise:
1005  *
1006  *      Implements the madvise function at the object/page level.
1007  *
1008  *      MADV_WILLNEED   (any object)
1009  *
1010  *          Activate the specified pages if they are resident.
1011  *
1012  *      MADV_DONTNEED   (any object)
1013  *
1014  *          Deactivate the specified pages if they are resident.
1015  *
1016  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1017  *                       OBJ_ONEMAPPING only)
1018  *
1019  *          Deactivate and clean the specified pages if they are
1020  *          resident.  This permits the process to reuse the pages
1021  *          without faulting or the kernel to reclaim the pages
1022  *          without I/O.
1023  */
1024 void
1025 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1026 {
1027         vm_pindex_t end, tpindex;
1028         vm_object_t backing_object, tobject;
1029         vm_page_t m;
1030
1031         if (object == NULL)
1032                 return;
1033         VM_OBJECT_LOCK(object);
1034         end = pindex + count;
1035         /*
1036          * Locate and adjust resident pages
1037          */
1038         for (; pindex < end; pindex += 1) {
1039 relookup:
1040                 tobject = object;
1041                 tpindex = pindex;
1042 shadowlookup:
1043                 /*
1044                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1045                  * and those pages must be OBJ_ONEMAPPING.
1046                  */
1047                 if (advise == MADV_FREE) {
1048                         if ((tobject->type != OBJT_DEFAULT &&
1049                              tobject->type != OBJT_SWAP) ||
1050                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1051                                 goto unlock_tobject;
1052                         }
1053                 } else if (tobject->type == OBJT_PHYS)
1054                         goto unlock_tobject;
1055                 m = vm_page_lookup(tobject, tpindex);
1056                 if (m == NULL && advise == MADV_WILLNEED) {
1057                         /*
1058                          * If the page is cached, reactivate it.
1059                          */
1060                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1061                             VM_ALLOC_NOBUSY);
1062                 }
1063                 if (m == NULL) {
1064                         /*
1065                          * There may be swap even if there is no backing page
1066                          */
1067                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1068                                 swap_pager_freespace(tobject, tpindex, 1);
1069                         /*
1070                          * next object
1071                          */
1072                         backing_object = tobject->backing_object;
1073                         if (backing_object == NULL)
1074                                 goto unlock_tobject;
1075                         VM_OBJECT_LOCK(backing_object);
1076                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1077                         if (tobject != object)
1078                                 VM_OBJECT_UNLOCK(tobject);
1079                         tobject = backing_object;
1080                         goto shadowlookup;
1081                 } else if (m->valid != VM_PAGE_BITS_ALL)
1082                         goto unlock_tobject;
1083                 /*
1084                  * If the page is not in a normal state, skip it.
1085                  */
1086                 vm_page_lock(m);
1087                 if (m->hold_count != 0 || m->wire_count != 0) {
1088                         vm_page_unlock(m);
1089                         goto unlock_tobject;
1090                 }
1091                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1092                     ("vm_object_madvise: page %p is fictitious", m));
1093                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1094                     ("vm_object_madvise: page %p is not managed", m));
1095                 if ((m->oflags & VPO_BUSY) || m->busy) {
1096                         if (advise == MADV_WILLNEED) {
1097                                 /*
1098                                  * Reference the page before unlocking and
1099                                  * sleeping so that the page daemon is less
1100                                  * likely to reclaim it.
1101                                  */
1102                                 vm_page_aflag_set(m, PGA_REFERENCED);
1103                         }
1104                         vm_page_unlock(m);
1105                         if (object != tobject)
1106                                 VM_OBJECT_UNLOCK(object);
1107                         m->oflags |= VPO_WANTED;
1108                         msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1109                             0);
1110                         VM_OBJECT_LOCK(object);
1111                         goto relookup;
1112                 }
1113                 if (advise == MADV_WILLNEED) {
1114                         vm_page_activate(m);
1115                 } else if (advise == MADV_DONTNEED) {
1116                         vm_page_dontneed(m);
1117                 } else if (advise == MADV_FREE) {
1118                         /*
1119                          * Mark the page clean.  This will allow the page
1120                          * to be freed up by the system.  However, such pages
1121                          * are often reused quickly by malloc()/free()
1122                          * so we do not do anything that would cause
1123                          * a page fault if we can help it.
1124                          *
1125                          * Specifically, we do not try to actually free
1126                          * the page now nor do we try to put it in the
1127                          * cache (which would cause a page fault on reuse).
1128                          *
1129                          * But we do make the page is freeable as we
1130                          * can without actually taking the step of unmapping
1131                          * it.
1132                          */
1133                         pmap_clear_modify(m);
1134                         m->dirty = 0;
1135                         m->act_count = 0;
1136                         vm_page_dontneed(m);
1137                 }
1138                 vm_page_unlock(m);
1139                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1140                         swap_pager_freespace(tobject, tpindex, 1);
1141 unlock_tobject:
1142                 if (tobject != object)
1143                         VM_OBJECT_UNLOCK(tobject);
1144         }
1145         VM_OBJECT_UNLOCK(object);
1146 }
1147
1148 /*
1149  *      vm_object_shadow:
1150  *
1151  *      Create a new object which is backed by the
1152  *      specified existing object range.  The source
1153  *      object reference is deallocated.
1154  *
1155  *      The new object and offset into that object
1156  *      are returned in the source parameters.
1157  */
1158 void
1159 vm_object_shadow(
1160         vm_object_t *object,    /* IN/OUT */
1161         vm_ooffset_t *offset,   /* IN/OUT */
1162         vm_size_t length)
1163 {
1164         vm_object_t source;
1165         vm_object_t result;
1166
1167         source = *object;
1168
1169         /*
1170          * Don't create the new object if the old object isn't shared.
1171          */
1172         if (source != NULL) {
1173                 VM_OBJECT_LOCK(source);
1174                 if (source->ref_count == 1 &&
1175                     source->handle == NULL &&
1176                     (source->type == OBJT_DEFAULT ||
1177                      source->type == OBJT_SWAP)) {
1178                         VM_OBJECT_UNLOCK(source);
1179                         return;
1180                 }
1181                 VM_OBJECT_UNLOCK(source);
1182         }
1183
1184         /*
1185          * Allocate a new object with the given length.
1186          */
1187         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1188
1189         /*
1190          * The new object shadows the source object, adding a reference to it.
1191          * Our caller changes his reference to point to the new object,
1192          * removing a reference to the source object.  Net result: no change
1193          * of reference count.
1194          *
1195          * Try to optimize the result object's page color when shadowing
1196          * in order to maintain page coloring consistency in the combined
1197          * shadowed object.
1198          */
1199         result->backing_object = source;
1200         /*
1201          * Store the offset into the source object, and fix up the offset into
1202          * the new object.
1203          */
1204         result->backing_object_offset = *offset;
1205         if (source != NULL) {
1206                 VM_OBJECT_LOCK(source);
1207                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1208                 source->shadow_count++;
1209 #if VM_NRESERVLEVEL > 0
1210                 result->flags |= source->flags & OBJ_COLORED;
1211                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1212                     ((1 << (VM_NFREEORDER - 1)) - 1);
1213 #endif
1214                 VM_OBJECT_UNLOCK(source);
1215         }
1216
1217
1218         /*
1219          * Return the new things
1220          */
1221         *offset = 0;
1222         *object = result;
1223 }
1224
1225 /*
1226  *      vm_object_split:
1227  *
1228  * Split the pages in a map entry into a new object.  This affords
1229  * easier removal of unused pages, and keeps object inheritance from
1230  * being a negative impact on memory usage.
1231  */
1232 void
1233 vm_object_split(vm_map_entry_t entry)
1234 {
1235         vm_page_t m, m_next;
1236         vm_object_t orig_object, new_object, source;
1237         vm_pindex_t idx, offidxstart;
1238         vm_size_t size;
1239
1240         orig_object = entry->object.vm_object;
1241         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1242                 return;
1243         if (orig_object->ref_count <= 1)
1244                 return;
1245         VM_OBJECT_UNLOCK(orig_object);
1246
1247         offidxstart = OFF_TO_IDX(entry->offset);
1248         size = atop(entry->end - entry->start);
1249
1250         /*
1251          * If swap_pager_copy() is later called, it will convert new_object
1252          * into a swap object.
1253          */
1254         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1255
1256         /*
1257          * At this point, the new object is still private, so the order in
1258          * which the original and new objects are locked does not matter.
1259          */
1260         VM_OBJECT_LOCK(new_object);
1261         VM_OBJECT_LOCK(orig_object);
1262         source = orig_object->backing_object;
1263         if (source != NULL) {
1264                 VM_OBJECT_LOCK(source);
1265                 if ((source->flags & OBJ_DEAD) != 0) {
1266                         VM_OBJECT_UNLOCK(source);
1267                         VM_OBJECT_UNLOCK(orig_object);
1268                         VM_OBJECT_UNLOCK(new_object);
1269                         vm_object_deallocate(new_object);
1270                         VM_OBJECT_LOCK(orig_object);
1271                         return;
1272                 }
1273                 LIST_INSERT_HEAD(&source->shadow_head,
1274                                   new_object, shadow_list);
1275                 source->shadow_count++;
1276                 vm_object_reference_locked(source);     /* for new_object */
1277                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1278                 VM_OBJECT_UNLOCK(source);
1279                 new_object->backing_object_offset =
1280                         orig_object->backing_object_offset + entry->offset;
1281                 new_object->backing_object = source;
1282         }
1283         if (orig_object->cred != NULL) {
1284                 new_object->cred = orig_object->cred;
1285                 crhold(orig_object->cred);
1286                 new_object->charge = ptoa(size);
1287                 KASSERT(orig_object->charge >= ptoa(size),
1288                     ("orig_object->charge < 0"));
1289                 orig_object->charge -= ptoa(size);
1290         }
1291 retry:
1292         m = vm_page_find_least(orig_object, offidxstart);
1293         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1294             m = m_next) {
1295                 m_next = TAILQ_NEXT(m, listq);
1296
1297                 /*
1298                  * We must wait for pending I/O to complete before we can
1299                  * rename the page.
1300                  *
1301                  * We do not have to VM_PROT_NONE the page as mappings should
1302                  * not be changed by this operation.
1303                  */
1304                 if ((m->oflags & VPO_BUSY) || m->busy) {
1305                         VM_OBJECT_UNLOCK(new_object);
1306                         m->oflags |= VPO_WANTED;
1307                         msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1308                         VM_OBJECT_LOCK(new_object);
1309                         goto retry;
1310                 }
1311                 vm_page_lock(m);
1312                 vm_page_rename(m, new_object, idx);
1313                 vm_page_unlock(m);
1314                 /* page automatically made dirty by rename and cache handled */
1315                 vm_page_busy(m);
1316         }
1317         if (orig_object->type == OBJT_SWAP) {
1318                 /*
1319                  * swap_pager_copy() can sleep, in which case the orig_object's
1320                  * and new_object's locks are released and reacquired.
1321                  */
1322                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1323
1324                 /*
1325                  * Transfer any cached pages from orig_object to new_object.
1326                  */
1327                 if (__predict_false(orig_object->cache != NULL))
1328                         vm_page_cache_transfer(orig_object, offidxstart,
1329                             new_object);
1330         }
1331         VM_OBJECT_UNLOCK(orig_object);
1332         TAILQ_FOREACH(m, &new_object->memq, listq)
1333                 vm_page_wakeup(m);
1334         VM_OBJECT_UNLOCK(new_object);
1335         entry->object.vm_object = new_object;
1336         entry->offset = 0LL;
1337         vm_object_deallocate(orig_object);
1338         VM_OBJECT_LOCK(new_object);
1339 }
1340
1341 #define OBSC_TEST_ALL_SHADOWED  0x0001
1342 #define OBSC_COLLAPSE_NOWAIT    0x0002
1343 #define OBSC_COLLAPSE_WAIT      0x0004
1344
1345 static int
1346 vm_object_backing_scan(vm_object_t object, int op)
1347 {
1348         int r = 1;
1349         vm_page_t p;
1350         vm_object_t backing_object;
1351         vm_pindex_t backing_offset_index;
1352
1353         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1354         VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1355
1356         backing_object = object->backing_object;
1357         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1358
1359         /*
1360          * Initial conditions
1361          */
1362         if (op & OBSC_TEST_ALL_SHADOWED) {
1363                 /*
1364                  * We do not want to have to test for the existence of cache
1365                  * or swap pages in the backing object.  XXX but with the
1366                  * new swapper this would be pretty easy to do.
1367                  *
1368                  * XXX what about anonymous MAP_SHARED memory that hasn't
1369                  * been ZFOD faulted yet?  If we do not test for this, the
1370                  * shadow test may succeed! XXX
1371                  */
1372                 if (backing_object->type != OBJT_DEFAULT) {
1373                         return (0);
1374                 }
1375         }
1376         if (op & OBSC_COLLAPSE_WAIT) {
1377                 vm_object_set_flag(backing_object, OBJ_DEAD);
1378         }
1379
1380         /*
1381          * Our scan
1382          */
1383         p = TAILQ_FIRST(&backing_object->memq);
1384         while (p) {
1385                 vm_page_t next = TAILQ_NEXT(p, listq);
1386                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1387
1388                 if (op & OBSC_TEST_ALL_SHADOWED) {
1389                         vm_page_t pp;
1390
1391                         /*
1392                          * Ignore pages outside the parent object's range
1393                          * and outside the parent object's mapping of the
1394                          * backing object.
1395                          *
1396                          * note that we do not busy the backing object's
1397                          * page.
1398                          */
1399                         if (
1400                             p->pindex < backing_offset_index ||
1401                             new_pindex >= object->size
1402                         ) {
1403                                 p = next;
1404                                 continue;
1405                         }
1406
1407                         /*
1408                          * See if the parent has the page or if the parent's
1409                          * object pager has the page.  If the parent has the
1410                          * page but the page is not valid, the parent's
1411                          * object pager must have the page.
1412                          *
1413                          * If this fails, the parent does not completely shadow
1414                          * the object and we might as well give up now.
1415                          */
1416
1417                         pp = vm_page_lookup(object, new_pindex);
1418                         if (
1419                             (pp == NULL || pp->valid == 0) &&
1420                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1421                         ) {
1422                                 r = 0;
1423                                 break;
1424                         }
1425                 }
1426
1427                 /*
1428                  * Check for busy page
1429                  */
1430                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1431                         vm_page_t pp;
1432
1433                         if (op & OBSC_COLLAPSE_NOWAIT) {
1434                                 if ((p->oflags & VPO_BUSY) ||
1435                                     !p->valid ||
1436                                     p->busy) {
1437                                         p = next;
1438                                         continue;
1439                                 }
1440                         } else if (op & OBSC_COLLAPSE_WAIT) {
1441                                 if ((p->oflags & VPO_BUSY) || p->busy) {
1442                                         VM_OBJECT_UNLOCK(object);
1443                                         p->oflags |= VPO_WANTED;
1444                                         msleep(p, VM_OBJECT_MTX(backing_object),
1445                                             PDROP | PVM, "vmocol", 0);
1446                                         VM_OBJECT_LOCK(object);
1447                                         VM_OBJECT_LOCK(backing_object);
1448                                         /*
1449                                          * If we slept, anything could have
1450                                          * happened.  Since the object is
1451                                          * marked dead, the backing offset
1452                                          * should not have changed so we
1453                                          * just restart our scan.
1454                                          */
1455                                         p = TAILQ_FIRST(&backing_object->memq);
1456                                         continue;
1457                                 }
1458                         }
1459
1460                         KASSERT(
1461                             p->object == backing_object,
1462                             ("vm_object_backing_scan: object mismatch")
1463                         );
1464
1465                         /*
1466                          * Destroy any associated swap
1467                          */
1468                         if (backing_object->type == OBJT_SWAP) {
1469                                 swap_pager_freespace(
1470                                     backing_object,
1471                                     p->pindex,
1472                                     1
1473                                 );
1474                         }
1475
1476                         if (
1477                             p->pindex < backing_offset_index ||
1478                             new_pindex >= object->size
1479                         ) {
1480                                 /*
1481                                  * Page is out of the parent object's range, we
1482                                  * can simply destroy it.
1483                                  */
1484                                 vm_page_lock(p);
1485                                 KASSERT(!pmap_page_is_mapped(p),
1486                                     ("freeing mapped page %p", p));
1487                                 if (p->wire_count == 0)
1488                                         vm_page_free(p);
1489                                 else
1490                                         vm_page_remove(p);
1491                                 vm_page_unlock(p);
1492                                 p = next;
1493                                 continue;
1494                         }
1495
1496                         pp = vm_page_lookup(object, new_pindex);
1497                         if (
1498                             (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1499                             (pp != NULL && pp->valid == 0)
1500                         ) {
1501                                 /*
1502                                  * The page in the parent is not (yet) valid.
1503                                  * We don't know anything about the state of
1504                                  * the original page.  It might be mapped,
1505                                  * so we must avoid the next if here.
1506                                  *
1507                                  * This is due to a race in vm_fault() where
1508                                  * we must unbusy the original (backing_obj)
1509                                  * page before we can (re)lock the parent.
1510                                  * Hence we can get here.
1511                                  */
1512                                 p = next;
1513                                 continue;
1514                         }
1515                         if (
1516                             pp != NULL ||
1517                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1518                         ) {
1519                                 /*
1520                                  * page already exists in parent OR swap exists
1521                                  * for this location in the parent.  Destroy
1522                                  * the original page from the backing object.
1523                                  *
1524                                  * Leave the parent's page alone
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 #if VM_NRESERVLEVEL > 0
1539                         /*
1540                          * Rename the reservation.
1541                          */
1542                         vm_reserv_rename(p, object, backing_object,
1543                             backing_offset_index);
1544 #endif
1545
1546                         /*
1547                          * Page does not exist in parent, rename the
1548                          * page from the backing object to the main object.
1549                          *
1550                          * If the page was mapped to a process, it can remain
1551                          * mapped through the rename.
1552                          */
1553                         vm_page_lock(p);
1554                         vm_page_rename(p, object, new_pindex);
1555                         vm_page_unlock(p);
1556                         /* page automatically made dirty by rename */
1557                 }
1558                 p = next;
1559         }
1560         return (r);
1561 }
1562
1563
1564 /*
1565  * this version of collapse allows the operation to occur earlier and
1566  * when paging_in_progress is true for an object...  This is not a complete
1567  * operation, but should plug 99.9% of the rest of the leaks.
1568  */
1569 static void
1570 vm_object_qcollapse(vm_object_t object)
1571 {
1572         vm_object_t backing_object = object->backing_object;
1573
1574         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1575         VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1576
1577         if (backing_object->ref_count != 1)
1578                 return;
1579
1580         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1581 }
1582
1583 /*
1584  *      vm_object_collapse:
1585  *
1586  *      Collapse an object with the object backing it.
1587  *      Pages in the backing object are moved into the
1588  *      parent, and the backing object is deallocated.
1589  */
1590 void
1591 vm_object_collapse(vm_object_t object)
1592 {
1593         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1594
1595         while (TRUE) {
1596                 vm_object_t backing_object;
1597
1598                 /*
1599                  * Verify that the conditions are right for collapse:
1600                  *
1601                  * The object exists and the backing object exists.
1602                  */
1603                 if ((backing_object = object->backing_object) == NULL)
1604                         break;
1605
1606                 /*
1607                  * we check the backing object first, because it is most likely
1608                  * not collapsable.
1609                  */
1610                 VM_OBJECT_LOCK(backing_object);
1611                 if (backing_object->handle != NULL ||
1612                     (backing_object->type != OBJT_DEFAULT &&
1613                      backing_object->type != OBJT_SWAP) ||
1614                     (backing_object->flags & OBJ_DEAD) ||
1615                     object->handle != NULL ||
1616                     (object->type != OBJT_DEFAULT &&
1617                      object->type != OBJT_SWAP) ||
1618                     (object->flags & OBJ_DEAD)) {
1619                         VM_OBJECT_UNLOCK(backing_object);
1620                         break;
1621                 }
1622
1623                 if (
1624                     object->paging_in_progress != 0 ||
1625                     backing_object->paging_in_progress != 0
1626                 ) {
1627                         vm_object_qcollapse(object);
1628                         VM_OBJECT_UNLOCK(backing_object);
1629                         break;
1630                 }
1631                 /*
1632                  * We know that we can either collapse the backing object (if
1633                  * the parent is the only reference to it) or (perhaps) have
1634                  * the parent bypass the object if the parent happens to shadow
1635                  * all the resident pages in the entire backing object.
1636                  *
1637                  * This is ignoring pager-backed pages such as swap pages.
1638                  * vm_object_backing_scan fails the shadowing test in this
1639                  * case.
1640                  */
1641                 if (backing_object->ref_count == 1) {
1642                         /*
1643                          * If there is exactly one reference to the backing
1644                          * object, we can collapse it into the parent.
1645                          */
1646                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1647
1648 #if VM_NRESERVLEVEL > 0
1649                         /*
1650                          * Break any reservations from backing_object.
1651                          */
1652                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1653                                 vm_reserv_break_all(backing_object);
1654 #endif
1655
1656                         /*
1657                          * Move the pager from backing_object to object.
1658                          */
1659                         if (backing_object->type == OBJT_SWAP) {
1660                                 /*
1661                                  * swap_pager_copy() can sleep, in which case
1662                                  * the backing_object's and object's locks are
1663                                  * released and reacquired.
1664                                  */
1665                                 swap_pager_copy(
1666                                     backing_object,
1667                                     object,
1668                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1669
1670                                 /*
1671                                  * Free any cached pages from backing_object.
1672                                  */
1673                                 if (__predict_false(backing_object->cache != NULL))
1674                                         vm_page_cache_free(backing_object, 0, 0);
1675                         }
1676                         /*
1677                          * Object now shadows whatever backing_object did.
1678                          * Note that the reference to
1679                          * backing_object->backing_object moves from within
1680                          * backing_object to within object.
1681                          */
1682                         LIST_REMOVE(object, shadow_list);
1683                         backing_object->shadow_count--;
1684                         if (backing_object->backing_object) {
1685                                 VM_OBJECT_LOCK(backing_object->backing_object);
1686                                 LIST_REMOVE(backing_object, shadow_list);
1687                                 LIST_INSERT_HEAD(
1688                                     &backing_object->backing_object->shadow_head,
1689                                     object, shadow_list);
1690                                 /*
1691                                  * The shadow_count has not changed.
1692                                  */
1693                                 VM_OBJECT_UNLOCK(backing_object->backing_object);
1694                         }
1695                         object->backing_object = backing_object->backing_object;
1696                         object->backing_object_offset +=
1697                             backing_object->backing_object_offset;
1698
1699                         /*
1700                          * Discard backing_object.
1701                          *
1702                          * Since the backing object has no pages, no pager left,
1703                          * and no object references within it, all that is
1704                          * necessary is to dispose of it.
1705                          */
1706                         KASSERT(backing_object->ref_count == 1, (
1707 "backing_object %p was somehow re-referenced during collapse!",
1708                             backing_object));
1709                         VM_OBJECT_UNLOCK(backing_object);
1710                         vm_object_destroy(backing_object);
1711
1712                         object_collapses++;
1713                 } else {
1714                         vm_object_t new_backing_object;
1715
1716                         /*
1717                          * If we do not entirely shadow the backing object,
1718                          * there is nothing we can do so we give up.
1719                          */
1720                         if (object->resident_page_count != object->size &&
1721                             vm_object_backing_scan(object,
1722                             OBSC_TEST_ALL_SHADOWED) == 0) {
1723                                 VM_OBJECT_UNLOCK(backing_object);
1724                                 break;
1725                         }
1726
1727                         /*
1728                          * Make the parent shadow the next object in the
1729                          * chain.  Deallocating backing_object will not remove
1730                          * it, since its reference count is at least 2.
1731                          */
1732                         LIST_REMOVE(object, shadow_list);
1733                         backing_object->shadow_count--;
1734
1735                         new_backing_object = backing_object->backing_object;
1736                         if ((object->backing_object = new_backing_object) != NULL) {
1737                                 VM_OBJECT_LOCK(new_backing_object);
1738                                 LIST_INSERT_HEAD(
1739                                     &new_backing_object->shadow_head,
1740                                     object,
1741                                     shadow_list
1742                                 );
1743                                 new_backing_object->shadow_count++;
1744                                 vm_object_reference_locked(new_backing_object);
1745                                 VM_OBJECT_UNLOCK(new_backing_object);
1746                                 object->backing_object_offset +=
1747                                         backing_object->backing_object_offset;
1748                         }
1749
1750                         /*
1751                          * Drop the reference count on backing_object. Since
1752                          * its ref_count was at least 2, it will not vanish.
1753                          */
1754                         backing_object->ref_count--;
1755                         VM_OBJECT_UNLOCK(backing_object);
1756                         object_bypasses++;
1757                 }
1758
1759                 /*
1760                  * Try again with this object's new backing object.
1761                  */
1762         }
1763 }
1764
1765 /*
1766  *      vm_object_page_remove:
1767  *
1768  *      For the given object, either frees or invalidates each of the
1769  *      specified pages.  In general, a page is freed.  However, if a page is
1770  *      wired for any reason other than the existence of a managed, wired
1771  *      mapping, then it may be invalidated but not removed from the object.
1772  *      Pages are specified by the given range ["start", "end") and the option
1773  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1774  *      extends from "start" to the end of the object.  If the option
1775  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1776  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1777  *      specified, then the pages within the specified range must have no
1778  *      mappings.  Otherwise, if this option is not specified, any mappings to
1779  *      the specified pages are removed before the pages are freed or
1780  *      invalidated.
1781  *
1782  *      In general, this operation should only be performed on objects that
1783  *      contain managed pages.  There are, however, two exceptions.  First, it
1784  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1785  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1786  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1787  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1788  *
1789  *      The object must be locked.
1790  */
1791 void
1792 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1793     int options)
1794 {
1795         vm_page_t p, next;
1796         int wirings;
1797
1798         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1799         KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1800             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1801             ("vm_object_page_remove: illegal options for object %p", object));
1802         if (object->resident_page_count == 0)
1803                 goto skipmemq;
1804         vm_object_pip_add(object, 1);
1805 again:
1806         p = vm_page_find_least(object, start);
1807
1808         /*
1809          * Here, the variable "p" is either (1) the page with the least pindex
1810          * greater than or equal to the parameter "start" or (2) NULL.
1811          */
1812         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1813                 next = TAILQ_NEXT(p, listq);
1814
1815                 /*
1816                  * If the page is wired for any reason besides the existence
1817                  * of managed, wired mappings, then it cannot be freed.  For
1818                  * example, fictitious pages, which represent device memory,
1819                  * are inherently wired and cannot be freed.  They can,
1820                  * however, be invalidated if the option OBJPR_CLEANONLY is
1821                  * not specified.
1822                  */
1823                 vm_page_lock(p);
1824                 if ((wirings = p->wire_count) != 0 &&
1825                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1826                         if ((options & OBJPR_NOTMAPPED) == 0) {
1827                                 pmap_remove_all(p);
1828                                 /* Account for removal of wired mappings. */
1829                                 if (wirings != 0)
1830                                         p->wire_count -= wirings;
1831                         }
1832                         if ((options & OBJPR_CLEANONLY) == 0) {
1833                                 p->valid = 0;
1834                                 vm_page_undirty(p);
1835                         }
1836                         vm_page_unlock(p);
1837                         continue;
1838                 }
1839                 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1840                         goto again;
1841                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1842                     ("vm_object_page_remove: page %p is fictitious", p));
1843                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1844                         if ((options & OBJPR_NOTMAPPED) == 0)
1845                                 pmap_remove_write(p);
1846                         if (p->dirty) {
1847                                 vm_page_unlock(p);
1848                                 continue;
1849                         }
1850                 }
1851                 if ((options & OBJPR_NOTMAPPED) == 0) {
1852                         pmap_remove_all(p);
1853                         /* Account for removal of wired mappings. */
1854                         if (wirings != 0)
1855                                 p->wire_count -= wirings;
1856                 }
1857                 vm_page_free(p);
1858                 vm_page_unlock(p);
1859         }
1860         vm_object_pip_wakeup(object);
1861 skipmemq:
1862         if (__predict_false(object->cache != NULL))
1863                 vm_page_cache_free(object, start, end);
1864 }
1865
1866 /*
1867  *      vm_object_page_cache:
1868  *
1869  *      For the given object, attempt to move the specified clean
1870  *      pages to the cache queue.  If a page is wired for any reason,
1871  *      then it will not be changed.  Pages are specified by the given
1872  *      range ["start", "end").  As a special case, if "end" is zero,
1873  *      then the range extends from "start" to the end of the object.
1874  *      Any mappings to the specified pages are removed before the
1875  *      pages are moved to the cache queue.
1876  *
1877  *      This operation should only be performed on objects that
1878  *      contain managed pages.
1879  *
1880  *      The object must be locked.
1881  */
1882 void
1883 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1884 {
1885         struct mtx *mtx, *new_mtx;
1886         vm_page_t p, next;
1887
1888         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1889         KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1890             object->type != OBJT_PHYS),
1891             ("vm_object_page_cache: illegal object %p", object));
1892         if (object->resident_page_count == 0)
1893                 return;
1894         p = vm_page_find_least(object, start);
1895
1896         /*
1897          * Here, the variable "p" is either (1) the page with the least pindex
1898          * greater than or equal to the parameter "start" or (2) NULL.
1899          */
1900         mtx = NULL;
1901         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1902                 next = TAILQ_NEXT(p, listq);
1903
1904                 /*
1905                  * Avoid releasing and reacquiring the same page lock.
1906                  */
1907                 new_mtx = vm_page_lockptr(p);
1908                 if (mtx != new_mtx) {
1909                         if (mtx != NULL)
1910                                 mtx_unlock(mtx);
1911                         mtx = new_mtx;
1912                         mtx_lock(mtx);
1913                 }
1914                 vm_page_try_to_cache(p);
1915         }
1916         if (mtx != NULL)
1917                 mtx_unlock(mtx);
1918 }
1919
1920 /*
1921  *      Populate the specified range of the object with valid pages.  Returns
1922  *      TRUE if the range is successfully populated and FALSE otherwise.
1923  *
1924  *      Note: This function should be optimized to pass a larger array of
1925  *      pages to vm_pager_get_pages() before it is applied to a non-
1926  *      OBJT_DEVICE object.
1927  *
1928  *      The object must be locked.
1929  */
1930 boolean_t
1931 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1932 {
1933         vm_page_t m, ma[1];
1934         vm_pindex_t pindex;
1935         int rv;
1936
1937         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1938         for (pindex = start; pindex < end; pindex++) {
1939                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1940                     VM_ALLOC_RETRY);
1941                 if (m->valid != VM_PAGE_BITS_ALL) {
1942                         ma[0] = m;
1943                         rv = vm_pager_get_pages(object, ma, 1, 0);
1944                         m = vm_page_lookup(object, pindex);
1945                         if (m == NULL)
1946                                 break;
1947                         if (rv != VM_PAGER_OK) {
1948                                 vm_page_lock(m);
1949                                 vm_page_free(m);
1950                                 vm_page_unlock(m);
1951                                 break;
1952                         }
1953                 }
1954                 /*
1955                  * Keep "m" busy because a subsequent iteration may unlock
1956                  * the object.
1957                  */
1958         }
1959         if (pindex > start) {
1960                 m = vm_page_lookup(object, start);
1961                 while (m != NULL && m->pindex < pindex) {
1962                         vm_page_wakeup(m);
1963                         m = TAILQ_NEXT(m, listq);
1964                 }
1965         }
1966         return (pindex == end);
1967 }
1968
1969 /*
1970  *      Routine:        vm_object_coalesce
1971  *      Function:       Coalesces two objects backing up adjoining
1972  *                      regions of memory into a single object.
1973  *
1974  *      returns TRUE if objects were combined.
1975  *
1976  *      NOTE:   Only works at the moment if the second object is NULL -
1977  *              if it's not, which object do we lock first?
1978  *
1979  *      Parameters:
1980  *              prev_object     First object to coalesce
1981  *              prev_offset     Offset into prev_object
1982  *              prev_size       Size of reference to prev_object
1983  *              next_size       Size of reference to the second object
1984  *              reserved        Indicator that extension region has
1985  *                              swap accounted for
1986  *
1987  *      Conditions:
1988  *      The object must *not* be locked.
1989  */
1990 boolean_t
1991 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1992     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1993 {
1994         vm_pindex_t next_pindex;
1995
1996         if (prev_object == NULL)
1997                 return (TRUE);
1998         VM_OBJECT_LOCK(prev_object);
1999         if (prev_object->type != OBJT_DEFAULT &&
2000             prev_object->type != OBJT_SWAP) {
2001                 VM_OBJECT_UNLOCK(prev_object);
2002                 return (FALSE);
2003         }
2004
2005         /*
2006          * Try to collapse the object first
2007          */
2008         vm_object_collapse(prev_object);
2009
2010         /*
2011          * Can't coalesce if: . more than one reference . paged out . shadows
2012          * another object . has a copy elsewhere (any of which mean that the
2013          * pages not mapped to prev_entry may be in use anyway)
2014          */
2015         if (prev_object->backing_object != NULL) {
2016                 VM_OBJECT_UNLOCK(prev_object);
2017                 return (FALSE);
2018         }
2019
2020         prev_size >>= PAGE_SHIFT;
2021         next_size >>= PAGE_SHIFT;
2022         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2023
2024         if ((prev_object->ref_count > 1) &&
2025             (prev_object->size != next_pindex)) {
2026                 VM_OBJECT_UNLOCK(prev_object);
2027                 return (FALSE);
2028         }
2029
2030         /*
2031          * Account for the charge.
2032          */
2033         if (prev_object->cred != NULL) {
2034
2035                 /*
2036                  * If prev_object was charged, then this mapping,
2037                  * althought not charged now, may become writable
2038                  * later. Non-NULL cred in the object would prevent
2039                  * swap reservation during enabling of the write
2040                  * access, so reserve swap now. Failed reservation
2041                  * cause allocation of the separate object for the map
2042                  * entry, and swap reservation for this entry is
2043                  * managed in appropriate time.
2044                  */
2045                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2046                     prev_object->cred)) {
2047                         return (FALSE);
2048                 }
2049                 prev_object->charge += ptoa(next_size);
2050         }
2051
2052         /*
2053          * Remove any pages that may still be in the object from a previous
2054          * deallocation.
2055          */
2056         if (next_pindex < prev_object->size) {
2057                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2058                     next_size, 0);
2059                 if (prev_object->type == OBJT_SWAP)
2060                         swap_pager_freespace(prev_object,
2061                                              next_pindex, next_size);
2062 #if 0
2063                 if (prev_object->cred != NULL) {
2064                         KASSERT(prev_object->charge >=
2065                             ptoa(prev_object->size - next_pindex),
2066                             ("object %p overcharged 1 %jx %jx", prev_object,
2067                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2068                         prev_object->charge -= ptoa(prev_object->size -
2069                             next_pindex);
2070                 }
2071 #endif
2072         }
2073
2074         /*
2075          * Extend the object if necessary.
2076          */
2077         if (next_pindex + next_size > prev_object->size)
2078                 prev_object->size = next_pindex + next_size;
2079
2080         VM_OBJECT_UNLOCK(prev_object);
2081         return (TRUE);
2082 }
2083
2084 void
2085 vm_object_set_writeable_dirty(vm_object_t object)
2086 {
2087
2088         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2089         if (object->type != OBJT_VNODE)
2090                 return;
2091         object->generation++;
2092         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2093                 return;
2094         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2095 }
2096
2097 #include "opt_ddb.h"
2098 #ifdef DDB
2099 #include <sys/kernel.h>
2100
2101 #include <sys/cons.h>
2102
2103 #include <ddb/ddb.h>
2104
2105 static int
2106 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2107 {
2108         vm_map_t tmpm;
2109         vm_map_entry_t tmpe;
2110         vm_object_t obj;
2111         int entcount;
2112
2113         if (map == 0)
2114                 return 0;
2115
2116         if (entry == 0) {
2117                 tmpe = map->header.next;
2118                 entcount = map->nentries;
2119                 while (entcount-- && (tmpe != &map->header)) {
2120                         if (_vm_object_in_map(map, object, tmpe)) {
2121                                 return 1;
2122                         }
2123                         tmpe = tmpe->next;
2124                 }
2125         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2126                 tmpm = entry->object.sub_map;
2127                 tmpe = tmpm->header.next;
2128                 entcount = tmpm->nentries;
2129                 while (entcount-- && tmpe != &tmpm->header) {
2130                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2131                                 return 1;
2132                         }
2133                         tmpe = tmpe->next;
2134                 }
2135         } else if ((obj = entry->object.vm_object) != NULL) {
2136                 for (; obj; obj = obj->backing_object)
2137                         if (obj == object) {
2138                                 return 1;
2139                         }
2140         }
2141         return 0;
2142 }
2143
2144 static int
2145 vm_object_in_map(vm_object_t object)
2146 {
2147         struct proc *p;
2148
2149         /* sx_slock(&allproc_lock); */
2150         FOREACH_PROC_IN_SYSTEM(p) {
2151                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2152                         continue;
2153                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2154                         /* sx_sunlock(&allproc_lock); */
2155                         return 1;
2156                 }
2157         }
2158         /* sx_sunlock(&allproc_lock); */
2159         if (_vm_object_in_map(kernel_map, object, 0))
2160                 return 1;
2161         if (_vm_object_in_map(kmem_map, object, 0))
2162                 return 1;
2163         if (_vm_object_in_map(pager_map, object, 0))
2164                 return 1;
2165         if (_vm_object_in_map(buffer_map, object, 0))
2166                 return 1;
2167         return 0;
2168 }
2169
2170 DB_SHOW_COMMAND(vmochk, vm_object_check)
2171 {
2172         vm_object_t object;
2173
2174         /*
2175          * make sure that internal objs are in a map somewhere
2176          * and none have zero ref counts.
2177          */
2178         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2179                 if (object->handle == NULL &&
2180                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2181                         if (object->ref_count == 0) {
2182                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2183                                         (long)object->size);
2184                         }
2185                         if (!vm_object_in_map(object)) {
2186                                 db_printf(
2187                         "vmochk: internal obj is not in a map: "
2188                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2189                                     object->ref_count, (u_long)object->size,
2190                                     (u_long)object->size,
2191                                     (void *)object->backing_object);
2192                         }
2193                 }
2194         }
2195 }
2196
2197 /*
2198  *      vm_object_print:        [ debug ]
2199  */
2200 DB_SHOW_COMMAND(object, vm_object_print_static)
2201 {
2202         /* XXX convert args. */
2203         vm_object_t object = (vm_object_t)addr;
2204         boolean_t full = have_addr;
2205
2206         vm_page_t p;
2207
2208         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2209 #define count   was_count
2210
2211         int count;
2212
2213         if (object == NULL)
2214                 return;
2215
2216         db_iprintf(
2217             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2218             object, (int)object->type, (uintmax_t)object->size,
2219             object->resident_page_count, object->ref_count, object->flags,
2220             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2221         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2222             object->shadow_count,
2223             object->backing_object ? object->backing_object->ref_count : 0,
2224             object->backing_object, (uintmax_t)object->backing_object_offset);
2225
2226         if (!full)
2227                 return;
2228
2229         db_indent += 2;
2230         count = 0;
2231         TAILQ_FOREACH(p, &object->memq, listq) {
2232                 if (count == 0)
2233                         db_iprintf("memory:=");
2234                 else if (count == 6) {
2235                         db_printf("\n");
2236                         db_iprintf(" ...");
2237                         count = 0;
2238                 } else
2239                         db_printf(",");
2240                 count++;
2241
2242                 db_printf("(off=0x%jx,page=0x%jx)",
2243                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2244         }
2245         if (count != 0)
2246                 db_printf("\n");
2247         db_indent -= 2;
2248 }
2249
2250 /* XXX. */
2251 #undef count
2252
2253 /* XXX need this non-static entry for calling from vm_map_print. */
2254 void
2255 vm_object_print(
2256         /* db_expr_t */ long addr,
2257         boolean_t have_addr,
2258         /* db_expr_t */ long count,
2259         char *modif)
2260 {
2261         vm_object_print_static(addr, have_addr, count, modif);
2262 }
2263
2264 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2265 {
2266         vm_object_t object;
2267         vm_pindex_t fidx;
2268         vm_paddr_t pa;
2269         vm_page_t m, prev_m;
2270         int rcount, nl, c;
2271
2272         nl = 0;
2273         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2274                 db_printf("new object: %p\n", (void *)object);
2275                 if (nl > 18) {
2276                         c = cngetc();
2277                         if (c != ' ')
2278                                 return;
2279                         nl = 0;
2280                 }
2281                 nl++;
2282                 rcount = 0;
2283                 fidx = 0;
2284                 pa = -1;
2285                 TAILQ_FOREACH(m, &object->memq, listq) {
2286                         if (m->pindex > 128)
2287                                 break;
2288                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2289                             prev_m->pindex + 1 != m->pindex) {
2290                                 if (rcount) {
2291                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2292                                                 (long)fidx, rcount, (long)pa);
2293                                         if (nl > 18) {
2294                                                 c = cngetc();
2295                                                 if (c != ' ')
2296                                                         return;
2297                                                 nl = 0;
2298                                         }
2299                                         nl++;
2300                                         rcount = 0;
2301                                 }
2302                         }
2303                         if (rcount &&
2304                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2305                                 ++rcount;
2306                                 continue;
2307                         }
2308                         if (rcount) {
2309                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2310                                         (long)fidx, rcount, (long)pa);
2311                                 if (nl > 18) {
2312                                         c = cngetc();
2313                                         if (c != ' ')
2314                                                 return;
2315                                         nl = 0;
2316                                 }
2317                                 nl++;
2318                         }
2319                         fidx = m->pindex;
2320                         pa = VM_PAGE_TO_PHYS(m);
2321                         rcount = 1;
2322                 }
2323                 if (rcount) {
2324                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2325                                 (long)fidx, rcount, (long)pa);
2326                         if (nl > 18) {
2327                                 c = cngetc();
2328                                 if (c != ' ')
2329                                         return;
2330                                 nl = 0;
2331                         }
2332                         nl++;
2333                 }
2334         }
2335 }
2336 #endif /* DDB */