sys/vm/vm_object.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * The Mach Operating System project at Carnegie-Mellon University.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 4. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  33  *
  34  *
  35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  36  * All rights reserved.
  37  *
  38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  39  *
  40  * Permission to use, copy, modify and distribute this software and
  41  * its documentation is hereby granted, provided that both the copyright
  42  * notice and this permission notice appear in all copies of the
  43  * software, derivative works or modified versions, and any portions
  44  * thereof, and that both notices appear in supporting documentation.
  45  *
  46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  49  *
  50  * Carnegie Mellon requests users of this software to return to
  51  *
  52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  53  *  School of Computer Science
  54  *  Carnegie Mellon University
  55  *  Pittsburgh PA 15213-3890
  56  *
  57  * any improvements or extensions that they make and grant Carnegie the
  58  * rights to redistribute these changes.
  59  */
  60
  61 /*
  62  *      Virtual memory object module.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_vm.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/lock.h>
  73 #include <sys/mman.h>
  74 #include <sys/mount.h>
  75 #include <sys/kernel.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/mutex.h>
  78 #include <sys/proc.h>           /* for curproc, pageproc */
  79 #include <sys/socket.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/rwlock.h>
  82 #include <sys/vnode.h>
  83 #include <sys/vmmeter.h>
  84 #include <sys/sx.h>
  85
  86 #include <vm/vm.h>
  87 #include <vm/vm_param.h>
  88 #include <vm/pmap.h>
  89 #include <vm/vm_map.h>
  90 #include <vm/vm_object.h>
  91 #include <vm/vm_page.h>
  92 #include <vm/vm_pageout.h>
  93 #include <vm/vm_pager.h>
  94 #include <vm/swap_pager.h>
  95 #include <vm/vm_kern.h>
  96 #include <vm/vm_extern.h>
  97 #include <vm/vm_radix.h>
  98 #include <vm/vm_reserv.h>
  99 #include <vm/uma.h>
 100
 101 static int old_msync;
 102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 103     "Use old (insecure) msync behavior");
 104
 105 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 106                     int pagerflags, int flags, boolean_t *clearobjflags,
 107                     boolean_t *eio);
 108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 109                     boolean_t *clearobjflags);
 110 static void     vm_object_qcollapse(vm_object_t object);
 111 static void     vm_object_vndeallocate(vm_object_t object);
 112
 113 /*
 114  *      Virtual memory objects maintain the actual data
 115  *      associated with allocated virtual memory.  A given
 116  *      page of memory exists within exactly one object.
 117  *
 118  *      An object is only deallocated when all "references"
 119  *      are given up.  Only one "reference" to a given
 120  *      region of an object should be writeable.
 121  *
 122  *      Associated with each object is a list of all resident
 123  *      memory pages belonging to that object; this list is
 124  *      maintained by the "vm_page" module, and locked by the object's
 125  *      lock.
 126  *
 127  *      Each object also records a "pager" routine which is
 128  *      used to retrieve (and store) pages to the proper backing
 129  *      storage.  In addition, objects may be backed by other
 130  *      objects from which they were virtual-copied.
 131  *
 132  *      The only items within the object structure which are
 133  *      modified after time of creation are:
 134  *              reference count         locked by object's lock
 135  *              pager routine           locked by object's lock
 136  *
 137  */
 138
 139 struct object_q vm_object_list;
 140 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 141
 142 struct vm_object kernel_object_store;
 143 struct vm_object kmem_object_store;
 144
 145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
 146     "VM object stats");
 147
 148 static long object_collapses;
 149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 150     &object_collapses, 0, "VM object collapses");
 151
 152 static long object_bypasses;
 153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 154     &object_bypasses, 0, "VM object bypasses");
 155
 156 static uma_zone_t obj_zone;
 157
 158 static int vm_object_zinit(void *mem, int size, int flags);
 159
 160 #ifdef INVARIANTS
 161 static void vm_object_zdtor(void *mem, int size, void *arg);
 162
 163 static void
 164 vm_object_zdtor(void *mem, int size, void *arg)
 165 {
 166         vm_object_t object;
 167
 168         object = (vm_object_t)mem;
 169         KASSERT(TAILQ_EMPTY(&object->memq),
 170             ("object %p has resident pages in its memq", object));
 171         KASSERT(vm_radix_is_empty(&object->rtree),
 172             ("object %p has resident pages in its trie", object));
 173 #if VM_NRESERVLEVEL > 0
 174         KASSERT(LIST_EMPTY(&object->rvq),
 175             ("object %p has reservations",
 176             object));
 177 #endif
 178         KASSERT(vm_object_cache_is_empty(object),
 179             ("object %p has cached pages",
 180             object));
 181         KASSERT(object->paging_in_progress == 0,
 182             ("object %p paging_in_progress = %d",
 183             object, object->paging_in_progress));
 184         KASSERT(object->resident_page_count == 0,
 185             ("object %p resident_page_count = %d",
 186             object, object->resident_page_count));
 187         KASSERT(object->shadow_count == 0,
 188             ("object %p shadow_count = %d",
 189             object, object->shadow_count));
 190 }
 191 #endif
 192
 193 static int
 194 vm_object_zinit(void *mem, int size, int flags)
 195 {
 196         vm_object_t object;
 197
 198         object = (vm_object_t)mem;
 199         bzero(&object->lock, sizeof(object->lock));
 200         rw_init_flags(&object->lock, "vm object", RW_DUPOK);
 201
 202         /* These are true for any object that has been freed */
 203         object->rtree.rt_root = 0;
 204         object->rtree.rt_flags = 0;
 205         object->paging_in_progress = 0;
 206         object->resident_page_count = 0;
 207         object->shadow_count = 0;
 208         object->cache.rt_root = 0;
 209         object->cache.rt_flags = 0;
 210         return (0);
 211 }
 212
 213 static void
 214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 215 {
 216
 217         TAILQ_INIT(&object->memq);
 218         LIST_INIT(&object->shadow_head);
 219
 220         object->type = type;
 221         switch (type) {
 222         case OBJT_DEAD:
 223                 panic("_vm_object_allocate: can't create OBJT_DEAD");
 224         case OBJT_DEFAULT:
 225         case OBJT_SWAP:
 226                 object->flags = OBJ_ONEMAPPING;
 227                 break;
 228         case OBJT_DEVICE:
 229         case OBJT_SG:
 230                 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
 231                 break;
 232         case OBJT_MGTDEVICE:
 233                 object->flags = OBJ_FICTITIOUS;
 234                 break;
 235         case OBJT_PHYS:
 236                 object->flags = OBJ_UNMANAGED;
 237                 break;
 238         case OBJT_VNODE:
 239                 object->flags = 0;
 240                 break;
 241         default:
 242                 panic("_vm_object_allocate: type %d is undefined", type);
 243         }
 244         object->size = size;
 245         object->generation = 1;
 246         object->ref_count = 1;
 247         object->memattr = VM_MEMATTR_DEFAULT;
 248         object->cred = NULL;
 249         object->charge = 0;
 250         object->handle = NULL;
 251         object->backing_object = NULL;
 252         object->backing_object_offset = (vm_ooffset_t) 0;
 253 #if VM_NRESERVLEVEL > 0
 254         LIST_INIT(&object->rvq);
 255 #endif
 256
 257         mtx_lock(&vm_object_list_mtx);
 258         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 259         mtx_unlock(&vm_object_list_mtx);
 260 }
 261
 262 /*
 263  *      vm_object_init:
 264  *
 265  *      Initialize the VM objects module.
 266  */
 267 void
 268 vm_object_init(void)
 269 {
 270         TAILQ_INIT(&vm_object_list);
 271         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 272
 273         rw_init(&kernel_object->lock, "kernel vm object");
 274         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 275             kernel_object);
 276 #if VM_NRESERVLEVEL > 0
 277         kernel_object->flags |= OBJ_COLORED;
 278         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 279 #endif
 280
 281         rw_init(&kmem_object->lock, "kmem vm object");
 282         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 283             kmem_object);
 284 #if VM_NRESERVLEVEL > 0
 285         kmem_object->flags |= OBJ_COLORED;
 286         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 287 #endif
 288
 289         /*
 290          * The lock portion of struct vm_object must be type stable due
 291          * to vm_pageout_fallback_object_lock locking a vm object
 292          * without holding any references to it.
 293          */
 294         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 295 #ifdef INVARIANTS
 296             vm_object_zdtor,
 297 #else
 298             NULL,
 299 #endif
 300             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 301
 302         vm_radix_init();
 303 }
 304
 305 void
 306 vm_object_clear_flag(vm_object_t object, u_short bits)
 307 {
 308
 309         VM_OBJECT_ASSERT_WLOCKED(object);
 310         object->flags &= ~bits;
 311 }
 312
 313 /*
 314  *      Sets the default memory attribute for the specified object.  Pages
 315  *      that are allocated to this object are by default assigned this memory
 316  *      attribute.
 317  *
 318  *      Presently, this function must be called before any pages are allocated
 319  *      to the object.  In the future, this requirement may be relaxed for
 320  *      "default" and "swap" objects.
 321  */
 322 int
 323 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 324 {
 325
 326         VM_OBJECT_ASSERT_WLOCKED(object);
 327         switch (object->type) {
 328         case OBJT_DEFAULT:
 329         case OBJT_DEVICE:
 330         case OBJT_MGTDEVICE:
 331         case OBJT_PHYS:
 332         case OBJT_SG:
 333         case OBJT_SWAP:
 334         case OBJT_VNODE:
 335                 if (!TAILQ_EMPTY(&object->memq))
 336                         return (KERN_FAILURE);
 337                 break;
 338         case OBJT_DEAD:
 339                 return (KERN_INVALID_ARGUMENT);
 340         default:
 341                 panic("vm_object_set_memattr: object %p is of undefined type",
 342                     object);
 343         }
 344         object->memattr = memattr;
 345         return (KERN_SUCCESS);
 346 }
 347
 348 void
 349 vm_object_pip_add(vm_object_t object, short i)
 350 {
 351
 352         VM_OBJECT_ASSERT_WLOCKED(object);
 353         object->paging_in_progress += i;
 354 }
 355
 356 void
 357 vm_object_pip_subtract(vm_object_t object, short i)
 358 {
 359
 360         VM_OBJECT_ASSERT_WLOCKED(object);
 361         object->paging_in_progress -= i;
 362 }
 363
 364 void
 365 vm_object_pip_wakeup(vm_object_t object)
 366 {
 367
 368         VM_OBJECT_ASSERT_WLOCKED(object);
 369         object->paging_in_progress--;
 370         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 371                 vm_object_clear_flag(object, OBJ_PIPWNT);
 372                 wakeup(object);
 373         }
 374 }
 375
 376 void
 377 vm_object_pip_wakeupn(vm_object_t object, short i)
 378 {
 379
 380         VM_OBJECT_ASSERT_WLOCKED(object);
 381         if (i)
 382                 object->paging_in_progress -= i;
 383         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 384                 vm_object_clear_flag(object, OBJ_PIPWNT);
 385                 wakeup(object);
 386         }
 387 }
 388
 389 void
 390 vm_object_pip_wait(vm_object_t object, char *waitid)
 391 {
 392
 393         VM_OBJECT_ASSERT_WLOCKED(object);
 394         while (object->paging_in_progress) {
 395                 object->flags |= OBJ_PIPWNT;
 396                 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
 397         }
 398 }
 399
 400 /*
 401  *      vm_object_allocate:
 402  *
 403  *      Returns a new object with the given size.
 404  */
 405 vm_object_t
 406 vm_object_allocate(objtype_t type, vm_pindex_t size)
 407 {
 408         vm_object_t object;
 409
 410         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 411         _vm_object_allocate(type, size, object);
 412         return (object);
 413 }
 414
 415
 416 /*
 417  *      vm_object_reference:
 418  *
 419  *      Gets another reference to the given object.  Note: OBJ_DEAD
 420  *      objects can be referenced during final cleaning.
 421  */
 422 void
 423 vm_object_reference(vm_object_t object)
 424 {
 425         if (object == NULL)
 426                 return;
 427         VM_OBJECT_WLOCK(object);
 428         vm_object_reference_locked(object);
 429         VM_OBJECT_WUNLOCK(object);
 430 }
 431
 432 /*
 433  *      vm_object_reference_locked:
 434  *
 435  *      Gets another reference to the given object.
 436  *
 437  *      The object must be locked.
 438  */
 439 void
 440 vm_object_reference_locked(vm_object_t object)
 441 {
 442         struct vnode *vp;
 443
 444         VM_OBJECT_ASSERT_WLOCKED(object);
 445         object->ref_count++;
 446         if (object->type == OBJT_VNODE) {
 447                 vp = object->handle;
 448                 vref(vp);
 449         }
 450 }
 451
 452 /*
 453  * Handle deallocating an object of type OBJT_VNODE.
 454  */
 455 static void
 456 vm_object_vndeallocate(vm_object_t object)
 457 {
 458         struct vnode *vp = (struct vnode *) object->handle;
 459
 460         VM_OBJECT_ASSERT_WLOCKED(object);
 461         KASSERT(object->type == OBJT_VNODE,
 462             ("vm_object_vndeallocate: not a vnode object"));
 463         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 464 #ifdef INVARIANTS
 465         if (object->ref_count == 0) {
 466                 vprint("vm_object_vndeallocate", vp);
 467                 panic("vm_object_vndeallocate: bad object reference count");
 468         }
 469 #endif
 470
 471         if (object->ref_count > 1) {
 472                 object->ref_count--;
 473                 VM_OBJECT_WUNLOCK(object);
 474                 /* vrele may need the vnode lock. */
 475                 vrele(vp);
 476         } else {
 477                 vhold(vp);
 478                 VM_OBJECT_WUNLOCK(object);
 479                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 480                 vdrop(vp);
 481                 VM_OBJECT_WLOCK(object);
 482                 object->ref_count--;
 483                 if (object->type == OBJT_DEAD) {
 484                         VM_OBJECT_WUNLOCK(object);
 485                         VOP_UNLOCK(vp, 0);
 486                 } else {
 487                         if (object->ref_count == 0)
 488                                 VOP_UNSET_TEXT(vp);
 489                         VM_OBJECT_WUNLOCK(object);
 490                         vput(vp);
 491                 }
 492         }
 493 }
 494
 495 /*
 496  *      vm_object_deallocate:
 497  *
 498  *      Release a reference to the specified object,
 499  *      gained either through a vm_object_allocate
 500  *      or a vm_object_reference call.  When all references
 501  *      are gone, storage associated with this object
 502  *      may be relinquished.
 503  *
 504  *      No object may be locked.
 505  */
 506 void
 507 vm_object_deallocate(vm_object_t object)
 508 {
 509         vm_object_t temp;
 510         struct vnode *vp;
 511
 512         while (object != NULL) {
 513                 VM_OBJECT_WLOCK(object);
 514                 if (object->type == OBJT_VNODE) {
 515                         vm_object_vndeallocate(object);
 516                         return;
 517                 }
 518
 519                 KASSERT(object->ref_count != 0,
 520                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 521
 522                 /*
 523                  * If the reference count goes to 0 we start calling
 524                  * vm_object_terminate() on the object chain.
 525                  * A ref count of 1 may be a special case depending on the
 526                  * shadow count being 0 or 1.
 527                  */
 528                 object->ref_count--;
 529                 if (object->ref_count > 1) {
 530                         VM_OBJECT_WUNLOCK(object);
 531                         return;
 532                 } else if (object->ref_count == 1) {
 533                         if (object->type == OBJT_SWAP &&
 534                             (object->flags & OBJ_TMPFS) != 0) {
 535                                 vp = object->un_pager.swp.swp_tmpfs;
 536                                 vhold(vp);
 537                                 VM_OBJECT_WUNLOCK(object);
 538                                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 539                                 vdrop(vp);
 540                                 VM_OBJECT_WLOCK(object);
 541                                 if (object->type == OBJT_DEAD ||
 542                                     object->ref_count != 1) {
 543                                         VM_OBJECT_WUNLOCK(object);
 544                                         VOP_UNLOCK(vp, 0);
 545                                         return;
 546                                 }
 547                                 if ((object->flags & OBJ_TMPFS) != 0)
 548                                         VOP_UNSET_TEXT(vp);
 549                                 VOP_UNLOCK(vp, 0);
 550                         }
 551                         if (object->shadow_count == 0 &&
 552                             object->handle == NULL &&
 553                             (object->type == OBJT_DEFAULT ||
 554                             (object->type == OBJT_SWAP &&
 555                             (object->flags & OBJ_TMPFS) == 0))) {
 556                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 557                         } else if ((object->shadow_count == 1) &&
 558                             (object->handle == NULL) &&
 559                             (object->type == OBJT_DEFAULT ||
 560                              object->type == OBJT_SWAP)) {
 561                                 KASSERT((object->flags & OBJ_TMPFS) == 0,
 562                                     ("shadowed tmpfs v_object %p", object));
 563                                 vm_object_t robject;
 564
 565                                 robject = LIST_FIRST(&object->shadow_head);
 566                                 KASSERT(robject != NULL,
 567                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 568                                          object->ref_count,
 569                                          object->shadow_count));
 570                                 if (!VM_OBJECT_TRYWLOCK(robject)) {
 571                                         /*
 572                                          * Avoid a potential deadlock.
 573                                          */
 574                                         object->ref_count++;
 575                                         VM_OBJECT_WUNLOCK(object);
 576                                         /*
 577                                          * More likely than not the thread
 578                                          * holding robject's lock has lower
 579                                          * priority than the current thread.
 580                                          * Let the lower priority thread run.
 581                                          */
 582                                         pause("vmo_de", 1);
 583                                         continue;
 584                                 }
 585                                 /*
 586                                  * Collapse object into its shadow unless its
 587                                  * shadow is dead.  In that case, object will
 588                                  * be deallocated by the thread that is
 589                                  * deallocating its shadow.
 590                                  */
 591                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 592                                     (robject->handle == NULL) &&
 593                                     (robject->type == OBJT_DEFAULT ||
 594                                      robject->type == OBJT_SWAP)) {
 595
 596                                         robject->ref_count++;
 597 retry:
 598                                         if (robject->paging_in_progress) {
 599                                                 VM_OBJECT_WUNLOCK(object);
 600                                                 vm_object_pip_wait(robject,
 601                                                     "objde1");
 602                                                 temp = robject->backing_object;
 603                                                 if (object == temp) {
 604                                                         VM_OBJECT_WLOCK(object);
 605                                                         goto retry;
 606                                                 }
 607                                         } else if (object->paging_in_progress) {
 608                                                 VM_OBJECT_WUNLOCK(robject);
 609                                                 object->flags |= OBJ_PIPWNT;
 610                                                 VM_OBJECT_SLEEP(object, object,
 611                                                     PDROP | PVM, "objde2", 0);
 612                                                 VM_OBJECT_WLOCK(robject);
 613                                                 temp = robject->backing_object;
 614                                                 if (object == temp) {
 615                                                         VM_OBJECT_WLOCK(object);
 616                                                         goto retry;
 617                                                 }
 618                                         } else
 619                                                 VM_OBJECT_WUNLOCK(object);
 620
 621                                         if (robject->ref_count == 1) {
 622                                                 robject->ref_count--;
 623                                                 object = robject;
 624                                                 goto doterm;
 625                                         }
 626                                         object = robject;
 627                                         vm_object_collapse(object);
 628                                         VM_OBJECT_WUNLOCK(object);
 629                                         continue;
 630                                 }
 631                                 VM_OBJECT_WUNLOCK(robject);
 632                         }
 633                         VM_OBJECT_WUNLOCK(object);
 634                         return;
 635                 }
 636 doterm:
 637                 temp = object->backing_object;
 638                 if (temp != NULL) {
 639                         VM_OBJECT_WLOCK(temp);
 640                         LIST_REMOVE(object, shadow_list);
 641                         temp->shadow_count--;
 642                         VM_OBJECT_WUNLOCK(temp);
 643                         object->backing_object = NULL;
 644                 }
 645                 /*
 646                  * Don't double-terminate, we could be in a termination
 647                  * recursion due to the terminate having to sync data
 648                  * to disk.
 649                  */
 650                 if ((object->flags & OBJ_DEAD) == 0)
 651                         vm_object_terminate(object);
 652                 else
 653                         VM_OBJECT_WUNLOCK(object);
 654                 object = temp;
 655         }
 656 }
 657
 658 /*
 659  *      vm_object_destroy removes the object from the global object list
 660  *      and frees the space for the object.
 661  */
 662 void
 663 vm_object_destroy(vm_object_t object)
 664 {
 665
 666         /*
 667          * Remove the object from the global object list.
 668          */
 669         mtx_lock(&vm_object_list_mtx);
 670         TAILQ_REMOVE(&vm_object_list, object, object_list);
 671         mtx_unlock(&vm_object_list_mtx);
 672
 673         /*
 674          * Release the allocation charge.
 675          */
 676         if (object->cred != NULL) {
 677                 KASSERT(object->type == OBJT_DEFAULT ||
 678                     object->type == OBJT_SWAP,
 679                     ("vm_object_terminate: non-swap obj %p has cred",
 680                      object));
 681                 swap_release_by_cred(object->charge, object->cred);
 682                 object->charge = 0;
 683                 crfree(object->cred);
 684                 object->cred = NULL;
 685         }
 686
 687         /*
 688          * Free the space for the object.
 689          */
 690         uma_zfree(obj_zone, object);
 691 }
 692
 693 /*
 694  *      vm_object_terminate actually destroys the specified object, freeing
 695  *      up all previously used resources.
 696  *
 697  *      The object must be locked.
 698  *      This routine may block.
 699  */
 700 void
 701 vm_object_terminate(vm_object_t object)
 702 {
 703         vm_page_t p, p_next;
 704
 705         VM_OBJECT_ASSERT_WLOCKED(object);
 706
 707         /*
 708          * Make sure no one uses us.
 709          */
 710         vm_object_set_flag(object, OBJ_DEAD);
 711
 712         /*
 713          * wait for the pageout daemon to be done with the object
 714          */
 715         vm_object_pip_wait(object, "objtrm");
 716
 717         KASSERT(!object->paging_in_progress,
 718                 ("vm_object_terminate: pageout in progress"));
 719
 720         /*
 721          * Clean and free the pages, as appropriate. All references to the
 722          * object are gone, so we don't need to lock it.
 723          */
 724         if (object->type == OBJT_VNODE) {
 725                 struct vnode *vp = (struct vnode *)object->handle;
 726
 727                 /*
 728                  * Clean pages and flush buffers.
 729                  */
 730                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 731                 VM_OBJECT_WUNLOCK(object);
 732
 733                 vinvalbuf(vp, V_SAVE, 0, 0);
 734
 735                 VM_OBJECT_WLOCK(object);
 736         }
 737
 738         KASSERT(object->ref_count == 0,
 739                 ("vm_object_terminate: object with references, ref_count=%d",
 740                 object->ref_count));
 741
 742         /*
 743          * Free any remaining pageable pages.  This also removes them from the
 744          * paging queues.  However, don't free wired pages, just remove them
 745          * from the object.  Rather than incrementally removing each page from
 746          * the object, the page and object are reset to any empty state.
 747          */
 748         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
 749                 vm_page_assert_unbusied(p);
 750                 vm_page_lock(p);
 751                 /*
 752                  * Optimize the page's removal from the object by resetting
 753                  * its "object" field.  Specifically, if the page is not
 754                  * wired, then the effect of this assignment is that
 755                  * vm_page_free()'s call to vm_page_remove() will return
 756                  * immediately without modifying the page or the object.
 757                  */
 758                 p->object = NULL;
 759                 if (p->wire_count == 0) {
 760                         vm_page_free(p);
 761                         PCPU_INC(cnt.v_pfree);
 762                 }
 763                 vm_page_unlock(p);
 764         }
 765         /*
 766          * If the object contained any pages, then reset it to an empty state.
 767          * None of the object's fields, including "resident_page_count", were
 768          * modified by the preceding loop.
 769          */
 770         if (object->resident_page_count != 0) {
 771                 vm_radix_reclaim_allnodes(&object->rtree);
 772                 TAILQ_INIT(&object->memq);
 773                 object->resident_page_count = 0;
 774                 if (object->type == OBJT_VNODE)
 775                         vdrop(object->handle);
 776         }
 777
 778 #if VM_NRESERVLEVEL > 0
 779         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 780                 vm_reserv_break_all(object);
 781 #endif
 782         if (__predict_false(!vm_object_cache_is_empty(object)))
 783                 vm_page_cache_free(object, 0, 0);
 784
 785         /*
 786          * Let the pager know object is dead.
 787          */
 788         vm_pager_deallocate(object);
 789         VM_OBJECT_WUNLOCK(object);
 790
 791         vm_object_destroy(object);
 792 }
 793
 794 /*
 795  * Make the page read-only so that we can clear the object flags.  However, if
 796  * this is a nosync mmap then the object is likely to stay dirty so do not
 797  * mess with the page and do not clear the object flags.  Returns TRUE if the
 798  * page should be flushed, and FALSE otherwise.
 799  */
 800 static boolean_t
 801 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
 802 {
 803
 804         /*
 805          * If we have been asked to skip nosync pages and this is a
 806          * nosync page, skip it.  Note that the object flags were not
 807          * cleared in this case so we do not have to set them.
 808          */
 809         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 810                 *clearobjflags = FALSE;
 811                 return (FALSE);
 812         } else {
 813                 pmap_remove_write(p);
 814                 return (p->dirty != 0);
 815         }
 816 }
 817
 818 /*
 819  *      vm_object_page_clean
 820  *
 821  *      Clean all dirty pages in the specified range of object.  Leaves page
 822  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 823  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 824  *      leaving the object dirty.
 825  *
 826  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 827  *      synchronous clustering mode implementation.
 828  *
 829  *      Odd semantics: if start == end, we clean everything.
 830  *
 831  *      The object must be locked.
 832  *
 833  *      Returns FALSE if some page from the range was not written, as
 834  *      reported by the pager, and TRUE otherwise.
 835  */
 836 boolean_t
 837 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 838     int flags)
 839 {
 840         vm_page_t np, p;
 841         vm_pindex_t pi, tend, tstart;
 842         int curgeneration, n, pagerflags;
 843         boolean_t clearobjflags, eio, res;
 844
 845         VM_OBJECT_ASSERT_WLOCKED(object);
 846
 847         /*
 848          * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
 849          * objects.  The check below prevents the function from
 850          * operating on non-vnode objects.
 851          */
 852         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 853             object->resident_page_count == 0)
 854                 return (TRUE);
 855
 856         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 857             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 858         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 859
 860         tstart = OFF_TO_IDX(start);
 861         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 862         clearobjflags = tstart == 0 && tend >= object->size;
 863         res = TRUE;
 864
 865 rescan:
 866         curgeneration = object->generation;
 867
 868         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 869                 pi = p->pindex;
 870                 if (pi >= tend)
 871                         break;
 872                 np = TAILQ_NEXT(p, listq);
 873                 if (p->valid == 0)
 874                         continue;
 875                 if (vm_page_sleep_if_busy(p, "vpcwai")) {
 876                         if (object->generation != curgeneration) {
 877                                 if ((flags & OBJPC_SYNC) != 0)
 878                                         goto rescan;
 879                                 else
 880                                         clearobjflags = FALSE;
 881                         }
 882                         np = vm_page_find_least(object, pi);
 883                         continue;
 884                 }
 885                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 886                         continue;
 887
 888                 n = vm_object_page_collect_flush(object, p, pagerflags,
 889                     flags, &clearobjflags, &eio);
 890                 if (eio) {
 891                         res = FALSE;
 892                         clearobjflags = FALSE;
 893                 }
 894                 if (object->generation != curgeneration) {
 895                         if ((flags & OBJPC_SYNC) != 0)
 896                                 goto rescan;
 897                         else
 898                                 clearobjflags = FALSE;
 899                 }
 900
 901                 /*
 902                  * If the VOP_PUTPAGES() did a truncated write, so
 903                  * that even the first page of the run is not fully
 904                  * written, vm_pageout_flush() returns 0 as the run
 905                  * length.  Since the condition that caused truncated
 906                  * write may be permanent, e.g. exhausted free space,
 907                  * accepting n == 0 would cause an infinite loop.
 908                  *
 909                  * Forwarding the iterator leaves the unwritten page
 910                  * behind, but there is not much we can do there if
 911                  * filesystem refuses to write it.
 912                  */
 913                 if (n == 0) {
 914                         n = 1;
 915                         clearobjflags = FALSE;
 916                 }
 917                 np = vm_page_find_least(object, pi + n);
 918         }
 919 #if 0
 920         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 921 #endif
 922
 923         if (clearobjflags)
 924                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 925         return (res);
 926 }
 927
 928 static int
 929 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 930     int flags, boolean_t *clearobjflags, boolean_t *eio)
 931 {
 932         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 933         int count, i, mreq, runlen;
 934
 935         vm_page_lock_assert(p, MA_NOTOWNED);
 936         VM_OBJECT_ASSERT_WLOCKED(object);
 937
 938         count = 1;
 939         mreq = 0;
 940
 941         for (tp = p; count < vm_pageout_page_count; count++) {
 942                 tp = vm_page_next(tp);
 943                 if (tp == NULL || vm_page_busied(tp))
 944                         break;
 945                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 946                         break;
 947         }
 948
 949         for (p_first = p; count < vm_pageout_page_count; count++) {
 950                 tp = vm_page_prev(p_first);
 951                 if (tp == NULL || vm_page_busied(tp))
 952                         break;
 953                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 954                         break;
 955                 p_first = tp;
 956                 mreq++;
 957         }
 958
 959         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 960                 ma[i] = tp;
 961
 962         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
 963         return (runlen);
 964 }
 965
 966 /*
 967  * Note that there is absolutely no sense in writing out
 968  * anonymous objects, so we track down the vnode object
 969  * to write out.
 970  * We invalidate (remove) all pages from the address space
 971  * for semantic correctness.
 972  *
 973  * If the backing object is a device object with unmanaged pages, then any
 974  * mappings to the specified range of pages must be removed before this
 975  * function is called.
 976  *
 977  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 978  * may start out with a NULL object.
 979  */
 980 boolean_t
 981 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 982     boolean_t syncio, boolean_t invalidate)
 983 {
 984         vm_object_t backing_object;
 985         struct vnode *vp;
 986         struct mount *mp;
 987         int error, flags, fsync_after;
 988         boolean_t res;
 989
 990         if (object == NULL)
 991                 return (TRUE);
 992         res = TRUE;
 993         error = 0;
 994         VM_OBJECT_WLOCK(object);
 995         while ((backing_object = object->backing_object) != NULL) {
 996                 VM_OBJECT_WLOCK(backing_object);
 997                 offset += object->backing_object_offset;
 998                 VM_OBJECT_WUNLOCK(object);
 999                 object = backing_object;
1000                 if (object->size < OFF_TO_IDX(offset + size))
1001                         size = IDX_TO_OFF(object->size) - offset;
1002         }
1003         /*
1004          * Flush pages if writing is allowed, invalidate them
1005          * if invalidation requested.  Pages undergoing I/O
1006          * will be ignored by vm_object_page_remove().
1007          *
1008          * We cannot lock the vnode and then wait for paging
1009          * to complete without deadlocking against vm_fault.
1010          * Instead we simply call vm_object_page_remove() and
1011          * allow it to block internally on a page-by-page
1012          * basis when it encounters pages undergoing async
1013          * I/O.
1014          */
1015         if (object->type == OBJT_VNODE &&
1016             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1017                 vp = object->handle;
1018                 VM_OBJECT_WUNLOCK(object);
1019                 (void) vn_start_write(vp, &mp, V_WAIT);
1020                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1021                 if (syncio && !invalidate && offset == 0 &&
1022                     OFF_TO_IDX(size) == object->size) {
1023                         /*
1024                          * If syncing the whole mapping of the file,
1025                          * it is faster to schedule all the writes in
1026                          * async mode, also allowing the clustering,
1027                          * and then wait for i/o to complete.
1028                          */
1029                         flags = 0;
1030                         fsync_after = TRUE;
1031                 } else {
1032                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1033                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1034                         fsync_after = FALSE;
1035                 }
1036                 VM_OBJECT_WLOCK(object);
1037                 res = vm_object_page_clean(object, offset, offset + size,
1038                     flags);
1039                 VM_OBJECT_WUNLOCK(object);
1040                 if (fsync_after)
1041                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1042                 VOP_UNLOCK(vp, 0);
1043                 vn_finished_write(mp);
1044                 if (error != 0)
1045                         res = FALSE;
1046                 VM_OBJECT_WLOCK(object);
1047         }
1048         if ((object->type == OBJT_VNODE ||
1049              object->type == OBJT_DEVICE) && invalidate) {
1050                 if (object->type == OBJT_DEVICE)
1051                         /*
1052                          * The option OBJPR_NOTMAPPED must be passed here
1053                          * because vm_object_page_remove() cannot remove
1054                          * unmanaged mappings.
1055                          */
1056                         flags = OBJPR_NOTMAPPED;
1057                 else if (old_msync)
1058                         flags = OBJPR_NOTWIRED;
1059                 else
1060                         flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1061                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1062                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1063         }
1064         VM_OBJECT_WUNLOCK(object);
1065         return (res);
1066 }
1067
1068 /*
1069  *      vm_object_madvise:
1070  *
1071  *      Implements the madvise function at the object/page level.
1072  *
1073  *      MADV_WILLNEED   (any object)
1074  *
1075  *          Activate the specified pages if they are resident.
1076  *
1077  *      MADV_DONTNEED   (any object)
1078  *
1079  *          Deactivate the specified pages if they are resident.
1080  *
1081  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1082  *                       OBJ_ONEMAPPING only)
1083  *
1084  *          Deactivate and clean the specified pages if they are
1085  *          resident.  This permits the process to reuse the pages
1086  *          without faulting or the kernel to reclaim the pages
1087  *          without I/O.
1088  */
1089 void
1090 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1091     int advise)
1092 {
1093         vm_pindex_t tpindex;
1094         vm_object_t backing_object, tobject;
1095         vm_page_t m;
1096
1097         if (object == NULL)
1098                 return;
1099         VM_OBJECT_WLOCK(object);
1100         /*
1101          * Locate and adjust resident pages
1102          */
1103         for (; pindex < end; pindex += 1) {
1104 relookup:
1105                 tobject = object;
1106                 tpindex = pindex;
1107 shadowlookup:
1108                 /*
1109                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1110                  * and those pages must be OBJ_ONEMAPPING.
1111                  */
1112                 if (advise == MADV_FREE) {
1113                         if ((tobject->type != OBJT_DEFAULT &&
1114                              tobject->type != OBJT_SWAP) ||
1115                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1116                                 goto unlock_tobject;
1117                         }
1118                 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1119                         goto unlock_tobject;
1120                 m = vm_page_lookup(tobject, tpindex);
1121                 if (m == NULL && advise == MADV_WILLNEED) {
1122                         /*
1123                          * If the page is cached, reactivate it.
1124                          */
1125                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1126                             VM_ALLOC_NOBUSY);
1127                 }
1128                 if (m == NULL) {
1129                         /*
1130                          * There may be swap even if there is no backing page
1131                          */
1132                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1133                                 swap_pager_freespace(tobject, tpindex, 1);
1134                         /*
1135                          * next object
1136                          */
1137                         backing_object = tobject->backing_object;
1138                         if (backing_object == NULL)
1139                                 goto unlock_tobject;
1140                         VM_OBJECT_WLOCK(backing_object);
1141                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1142                         if (tobject != object)
1143                                 VM_OBJECT_WUNLOCK(tobject);
1144                         tobject = backing_object;
1145                         goto shadowlookup;
1146                 } else if (m->valid != VM_PAGE_BITS_ALL)
1147                         goto unlock_tobject;
1148                 /*
1149                  * If the page is not in a normal state, skip it.
1150                  */
1151                 vm_page_lock(m);
1152                 if (m->hold_count != 0 || m->wire_count != 0) {
1153                         vm_page_unlock(m);
1154                         goto unlock_tobject;
1155                 }
1156                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1157                     ("vm_object_madvise: page %p is fictitious", m));
1158                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1159                     ("vm_object_madvise: page %p is not managed", m));
1160                 if (vm_page_busied(m)) {
1161                         if (advise == MADV_WILLNEED) {
1162                                 /*
1163                                  * Reference the page before unlocking and
1164                                  * sleeping so that the page daemon is less
1165                                  * likely to reclaim it.
1166                                  */
1167                                 vm_page_aflag_set(m, PGA_REFERENCED);
1168                         }
1169                         if (object != tobject)
1170                                 VM_OBJECT_WUNLOCK(object);
1171                         VM_OBJECT_WUNLOCK(tobject);
1172                         vm_page_busy_sleep(m, "madvpo");
1173                         VM_OBJECT_WLOCK(object);
1174                         goto relookup;
1175                 }
1176                 if (advise == MADV_WILLNEED) {
1177                         vm_page_activate(m);
1178                 } else {
1179                         vm_page_advise(m, advise);
1180                 }
1181                 vm_page_unlock(m);
1182                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1183                         swap_pager_freespace(tobject, tpindex, 1);
1184 unlock_tobject:
1185                 if (tobject != object)
1186                         VM_OBJECT_WUNLOCK(tobject);
1187         }
1188         VM_OBJECT_WUNLOCK(object);
1189 }
1190
1191 /*
1192  *      vm_object_shadow:
1193  *
1194  *      Create a new object which is backed by the
1195  *      specified existing object range.  The source
1196  *      object reference is deallocated.
1197  *
1198  *      The new object and offset into that object
1199  *      are returned in the source parameters.
1200  */
1201 void
1202 vm_object_shadow(
1203         vm_object_t *object,    /* IN/OUT */
1204         vm_ooffset_t *offset,   /* IN/OUT */
1205         vm_size_t length)
1206 {
1207         vm_object_t source;
1208         vm_object_t result;
1209
1210         source = *object;
1211
1212         /*
1213          * Don't create the new object if the old object isn't shared.
1214          */
1215         if (source != NULL) {
1216                 VM_OBJECT_WLOCK(source);
1217                 if (source->ref_count == 1 &&
1218                     source->handle == NULL &&
1219                     (source->type == OBJT_DEFAULT ||
1220                      source->type == OBJT_SWAP)) {
1221                         VM_OBJECT_WUNLOCK(source);
1222                         return;
1223                 }
1224                 VM_OBJECT_WUNLOCK(source);
1225         }
1226
1227         /*
1228          * Allocate a new object with the given length.
1229          */
1230         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1231
1232         /*
1233          * The new object shadows the source object, adding a reference to it.
1234          * Our caller changes his reference to point to the new object,
1235          * removing a reference to the source object.  Net result: no change
1236          * of reference count.
1237          *
1238          * Try to optimize the result object's page color when shadowing
1239          * in order to maintain page coloring consistency in the combined
1240          * shadowed object.
1241          */
1242         result->backing_object = source;
1243         /*
1244          * Store the offset into the source object, and fix up the offset into
1245          * the new object.
1246          */
1247         result->backing_object_offset = *offset;
1248         if (source != NULL) {
1249                 VM_OBJECT_WLOCK(source);
1250                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1251                 source->shadow_count++;
1252 #if VM_NRESERVLEVEL > 0
1253                 result->flags |= source->flags & OBJ_COLORED;
1254                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1255                     ((1 << (VM_NFREEORDER - 1)) - 1);
1256 #endif
1257                 VM_OBJECT_WUNLOCK(source);
1258         }
1259
1260
1261         /*
1262          * Return the new things
1263          */
1264         *offset = 0;
1265         *object = result;
1266 }
1267
1268 /*
1269  *      vm_object_split:
1270  *
1271  * Split the pages in a map entry into a new object.  This affords
1272  * easier removal of unused pages, and keeps object inheritance from
1273  * being a negative impact on memory usage.
1274  */
1275 void
1276 vm_object_split(vm_map_entry_t entry)
1277 {
1278         vm_page_t m, m_next;
1279         vm_object_t orig_object, new_object, source;
1280         vm_pindex_t idx, offidxstart;
1281         vm_size_t size;
1282
1283         orig_object = entry->object.vm_object;
1284         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1285                 return;
1286         if (orig_object->ref_count <= 1)
1287                 return;
1288         VM_OBJECT_WUNLOCK(orig_object);
1289
1290         offidxstart = OFF_TO_IDX(entry->offset);
1291         size = atop(entry->end - entry->start);
1292
1293         /*
1294          * If swap_pager_copy() is later called, it will convert new_object
1295          * into a swap object.
1296          */
1297         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1298
1299         /*
1300          * At this point, the new object is still private, so the order in
1301          * which the original and new objects are locked does not matter.
1302          */
1303         VM_OBJECT_WLOCK(new_object);
1304         VM_OBJECT_WLOCK(orig_object);
1305         source = orig_object->backing_object;
1306         if (source != NULL) {
1307                 VM_OBJECT_WLOCK(source);
1308                 if ((source->flags & OBJ_DEAD) != 0) {
1309                         VM_OBJECT_WUNLOCK(source);
1310                         VM_OBJECT_WUNLOCK(orig_object);
1311                         VM_OBJECT_WUNLOCK(new_object);
1312                         vm_object_deallocate(new_object);
1313                         VM_OBJECT_WLOCK(orig_object);
1314                         return;
1315                 }
1316                 LIST_INSERT_HEAD(&source->shadow_head,
1317                                   new_object, shadow_list);
1318                 source->shadow_count++;
1319                 vm_object_reference_locked(source);     /* for new_object */
1320                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1321                 VM_OBJECT_WUNLOCK(source);
1322                 new_object->backing_object_offset =
1323                         orig_object->backing_object_offset + entry->offset;
1324                 new_object->backing_object = source;
1325         }
1326         if (orig_object->cred != NULL) {
1327                 new_object->cred = orig_object->cred;
1328                 crhold(orig_object->cred);
1329                 new_object->charge = ptoa(size);
1330                 KASSERT(orig_object->charge >= ptoa(size),
1331                     ("orig_object->charge < 0"));
1332                 orig_object->charge -= ptoa(size);
1333         }
1334 retry:
1335         m = vm_page_find_least(orig_object, offidxstart);
1336         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1337             m = m_next) {
1338                 m_next = TAILQ_NEXT(m, listq);
1339
1340                 /*
1341                  * We must wait for pending I/O to complete before we can
1342                  * rename the page.
1343                  *
1344                  * We do not have to VM_PROT_NONE the page as mappings should
1345                  * not be changed by this operation.
1346                  */
1347                 if (vm_page_busied(m)) {
1348                         VM_OBJECT_WUNLOCK(new_object);
1349                         vm_page_lock(m);
1350                         VM_OBJECT_WUNLOCK(orig_object);
1351                         vm_page_busy_sleep(m, "spltwt");
1352                         VM_OBJECT_WLOCK(orig_object);
1353                         VM_OBJECT_WLOCK(new_object);
1354                         goto retry;
1355                 }
1356
1357                 /* vm_page_rename() will handle dirty and cache. */
1358                 if (vm_page_rename(m, new_object, idx)) {
1359                         VM_OBJECT_WUNLOCK(new_object);
1360                         VM_OBJECT_WUNLOCK(orig_object);
1361                         VM_WAIT;
1362                         VM_OBJECT_WLOCK(orig_object);
1363                         VM_OBJECT_WLOCK(new_object);
1364                         goto retry;
1365                 }
1366 #if VM_NRESERVLEVEL > 0
1367                 /*
1368                  * If some of the reservation's allocated pages remain with
1369                  * the original object, then transferring the reservation to
1370                  * the new object is neither particularly beneficial nor
1371                  * particularly harmful as compared to leaving the reservation
1372                  * with the original object.  If, however, all of the
1373                  * reservation's allocated pages are transferred to the new
1374                  * object, then transferring the reservation is typically
1375                  * beneficial.  Determining which of these two cases applies
1376                  * would be more costly than unconditionally renaming the
1377                  * reservation.
1378                  */
1379                 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1380 #endif
1381                 if (orig_object->type == OBJT_SWAP)
1382                         vm_page_xbusy(m);
1383         }
1384         if (orig_object->type == OBJT_SWAP) {
1385                 /*
1386                  * swap_pager_copy() can sleep, in which case the orig_object's
1387                  * and new_object's locks are released and reacquired.
1388                  */
1389                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1390                 TAILQ_FOREACH(m, &new_object->memq, listq)
1391                         vm_page_xunbusy(m);
1392
1393                 /*
1394                  * Transfer any cached pages from orig_object to new_object.
1395                  * If swap_pager_copy() found swapped out pages within the
1396                  * specified range of orig_object, then it changed
1397                  * new_object's type to OBJT_SWAP when it transferred those
1398                  * pages to new_object.  Otherwise, new_object's type
1399                  * should still be OBJT_DEFAULT and orig_object should not
1400                  * contain any cached pages within the specified range.
1401                  */
1402                 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1403                         vm_page_cache_transfer(orig_object, offidxstart,
1404                             new_object);
1405         }
1406         VM_OBJECT_WUNLOCK(orig_object);
1407         VM_OBJECT_WUNLOCK(new_object);
1408         entry->object.vm_object = new_object;
1409         entry->offset = 0LL;
1410         vm_object_deallocate(orig_object);
1411         VM_OBJECT_WLOCK(new_object);
1412 }
1413
1414 #define OBSC_TEST_ALL_SHADOWED  0x0001
1415 #define OBSC_COLLAPSE_NOWAIT    0x0002
1416 #define OBSC_COLLAPSE_WAIT      0x0004
1417
1418 static int
1419 vm_object_backing_scan(vm_object_t object, int op)
1420 {
1421         int r = 1;
1422         vm_page_t p;
1423         vm_object_t backing_object;
1424         vm_pindex_t backing_offset_index;
1425
1426         VM_OBJECT_ASSERT_WLOCKED(object);
1427         VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1428
1429         backing_object = object->backing_object;
1430         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1431
1432         /*
1433          * Initial conditions
1434          */
1435         if (op & OBSC_TEST_ALL_SHADOWED) {
1436                 /*
1437                  * We do not want to have to test for the existence of cache
1438                  * or swap pages in the backing object.  XXX but with the
1439                  * new swapper this would be pretty easy to do.
1440                  *
1441                  * XXX what about anonymous MAP_SHARED memory that hasn't
1442                  * been ZFOD faulted yet?  If we do not test for this, the
1443                  * shadow test may succeed! XXX
1444                  */
1445                 if (backing_object->type != OBJT_DEFAULT) {
1446                         return (0);
1447                 }
1448         }
1449         if (op & OBSC_COLLAPSE_WAIT) {
1450                 vm_object_set_flag(backing_object, OBJ_DEAD);
1451         }
1452
1453         /*
1454          * Our scan
1455          */
1456         p = TAILQ_FIRST(&backing_object->memq);
1457         while (p) {
1458                 vm_page_t next = TAILQ_NEXT(p, listq);
1459                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1460
1461                 if (op & OBSC_TEST_ALL_SHADOWED) {
1462                         vm_page_t pp;
1463
1464                         /*
1465                          * Ignore pages outside the parent object's range
1466                          * and outside the parent object's mapping of the
1467                          * backing object.
1468                          *
1469                          * note that we do not busy the backing object's
1470                          * page.
1471                          */
1472                         if (
1473                             p->pindex < backing_offset_index ||
1474                             new_pindex >= object->size
1475                         ) {
1476                                 p = next;
1477                                 continue;
1478                         }
1479
1480                         /*
1481                          * See if the parent has the page or if the parent's
1482                          * object pager has the page.  If the parent has the
1483                          * page but the page is not valid, the parent's
1484                          * object pager must have the page.
1485                          *
1486                          * If this fails, the parent does not completely shadow
1487                          * the object and we might as well give up now.
1488                          */
1489
1490                         pp = vm_page_lookup(object, new_pindex);
1491                         if (
1492                             (pp == NULL || pp->valid == 0) &&
1493                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1494                         ) {
1495                                 r = 0;
1496                                 break;
1497                         }
1498                 }
1499
1500                 /*
1501                  * Check for busy page
1502                  */
1503                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1504                         vm_page_t pp;
1505
1506                         if (op & OBSC_COLLAPSE_NOWAIT) {
1507                                 if (!p->valid || vm_page_busied(p)) {
1508                                         p = next;
1509                                         continue;
1510                                 }
1511                         } else if (op & OBSC_COLLAPSE_WAIT) {
1512                                 if (vm_page_busied(p)) {
1513                                         VM_OBJECT_WUNLOCK(object);
1514                                         vm_page_lock(p);
1515                                         VM_OBJECT_WUNLOCK(backing_object);
1516                                         vm_page_busy_sleep(p, "vmocol");
1517                                         VM_OBJECT_WLOCK(object);
1518                                         VM_OBJECT_WLOCK(backing_object);
1519                                         /*
1520                                          * If we slept, anything could have
1521                                          * happened.  Since the object is
1522                                          * marked dead, the backing offset
1523                                          * should not have changed so we
1524                                          * just restart our scan.
1525                                          */
1526                                         p = TAILQ_FIRST(&backing_object->memq);
1527                                         continue;
1528                                 }
1529                         }
1530
1531                         KASSERT(
1532                             p->object == backing_object,
1533                             ("vm_object_backing_scan: object mismatch")
1534                         );
1535
1536                         if (
1537                             p->pindex < backing_offset_index ||
1538                             new_pindex >= object->size
1539                         ) {
1540                                 if (backing_object->type == OBJT_SWAP)
1541                                         swap_pager_freespace(backing_object,
1542                                             p->pindex, 1);
1543
1544                                 /*
1545                                  * Page is out of the parent object's range, we
1546                                  * can simply destroy it.
1547                                  */
1548                                 vm_page_lock(p);
1549                                 KASSERT(!pmap_page_is_mapped(p),
1550                                     ("freeing mapped page %p", p));
1551                                 if (p->wire_count == 0)
1552                                         vm_page_free(p);
1553                                 else
1554                                         vm_page_remove(p);
1555                                 vm_page_unlock(p);
1556                                 p = next;
1557                                 continue;
1558                         }
1559
1560                         pp = vm_page_lookup(object, new_pindex);
1561                         if (
1562                             (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1563                             (pp != NULL && pp->valid == 0)
1564                         ) {
1565                                 if (backing_object->type == OBJT_SWAP)
1566                                         swap_pager_freespace(backing_object,
1567                                             p->pindex, 1);
1568
1569                                 /*
1570                                  * The page in the parent is not (yet) valid.
1571                                  * We don't know anything about the state of
1572                                  * the original page.  It might be mapped,
1573                                  * so we must avoid the next if here.
1574                                  *
1575                                  * This is due to a race in vm_fault() where
1576                                  * we must unbusy the original (backing_obj)
1577                                  * page before we can (re)lock the parent.
1578                                  * Hence we can get here.
1579                                  */
1580                                 p = next;
1581                                 continue;
1582                         }
1583                         if (
1584                             pp != NULL ||
1585                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1586                         ) {
1587                                 if (backing_object->type == OBJT_SWAP)
1588                                         swap_pager_freespace(backing_object,
1589                                             p->pindex, 1);
1590
1591                                 /*
1592                                  * page already exists in parent OR swap exists
1593                                  * for this location in the parent.  Destroy
1594                                  * the original page from the backing object.
1595                                  *
1596                                  * Leave the parent's page alone
1597                                  */
1598                                 vm_page_lock(p);
1599                                 KASSERT(!pmap_page_is_mapped(p),
1600                                     ("freeing mapped page %p", p));
1601                                 if (p->wire_count == 0)
1602                                         vm_page_free(p);
1603                                 else
1604                                         vm_page_remove(p);
1605                                 vm_page_unlock(p);
1606                                 p = next;
1607                                 continue;
1608                         }
1609
1610                         /*
1611                          * Page does not exist in parent, rename the
1612                          * page from the backing object to the main object.
1613                          *
1614                          * If the page was mapped to a process, it can remain
1615                          * mapped through the rename.
1616                          * vm_page_rename() will handle dirty and cache.
1617                          */
1618                         if (vm_page_rename(p, object, new_pindex)) {
1619                                 if (op & OBSC_COLLAPSE_NOWAIT) {
1620                                         p = next;
1621                                         continue;
1622                                 }
1623                                 VM_OBJECT_WLOCK(backing_object);
1624                                 VM_OBJECT_WUNLOCK(object);
1625                                 VM_WAIT;
1626                                 VM_OBJECT_WLOCK(object);
1627                                 VM_OBJECT_WLOCK(backing_object);
1628                                 p = TAILQ_FIRST(&backing_object->memq);
1629                                 continue;
1630                         }
1631                         if (backing_object->type == OBJT_SWAP)
1632                                 swap_pager_freespace(backing_object, p->pindex,
1633                                     1);
1634
1635 #if VM_NRESERVLEVEL > 0
1636                         /*
1637                          * Rename the reservation.
1638                          */
1639                         vm_reserv_rename(p, object, backing_object,
1640                             backing_offset_index);
1641 #endif
1642                 }
1643                 p = next;
1644         }
1645         return (r);
1646 }
1647
1648
1649 /*
1650  * this version of collapse allows the operation to occur earlier and
1651  * when paging_in_progress is true for an object...  This is not a complete
1652  * operation, but should plug 99.9% of the rest of the leaks.
1653  */
1654 static void
1655 vm_object_qcollapse(vm_object_t object)
1656 {
1657         vm_object_t backing_object = object->backing_object;
1658
1659         VM_OBJECT_ASSERT_WLOCKED(object);
1660         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1661
1662         if (backing_object->ref_count != 1)
1663                 return;
1664
1665         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1666 }
1667
1668 /*
1669  *      vm_object_collapse:
1670  *
1671  *      Collapse an object with the object backing it.
1672  *      Pages in the backing object are moved into the
1673  *      parent, and the backing object is deallocated.
1674  */
1675 void
1676 vm_object_collapse(vm_object_t object)
1677 {
1678         VM_OBJECT_ASSERT_WLOCKED(object);
1679
1680         while (TRUE) {
1681                 vm_object_t backing_object;
1682
1683                 /*
1684                  * Verify that the conditions are right for collapse:
1685                  *
1686                  * The object exists and the backing object exists.
1687                  */
1688                 if ((backing_object = object->backing_object) == NULL)
1689                         break;
1690
1691                 /*
1692                  * we check the backing object first, because it is most likely
1693                  * not collapsable.
1694                  */
1695                 VM_OBJECT_WLOCK(backing_object);
1696                 if (backing_object->handle != NULL ||
1697                     (backing_object->type != OBJT_DEFAULT &&
1698                      backing_object->type != OBJT_SWAP) ||
1699                     (backing_object->flags & OBJ_DEAD) ||
1700                     object->handle != NULL ||
1701                     (object->type != OBJT_DEFAULT &&
1702                      object->type != OBJT_SWAP) ||
1703                     (object->flags & OBJ_DEAD)) {
1704                         VM_OBJECT_WUNLOCK(backing_object);
1705                         break;
1706                 }
1707
1708                 if (
1709                     object->paging_in_progress != 0 ||
1710                     backing_object->paging_in_progress != 0
1711                 ) {
1712                         vm_object_qcollapse(object);
1713                         VM_OBJECT_WUNLOCK(backing_object);
1714                         break;
1715                 }
1716                 /*
1717                  * We know that we can either collapse the backing object (if
1718                  * the parent is the only reference to it) or (perhaps) have
1719                  * the parent bypass the object if the parent happens to shadow
1720                  * all the resident pages in the entire backing object.
1721                  *
1722                  * This is ignoring pager-backed pages such as swap pages.
1723                  * vm_object_backing_scan fails the shadowing test in this
1724                  * case.
1725                  */
1726                 if (backing_object->ref_count == 1) {
1727                         /*
1728                          * If there is exactly one reference to the backing
1729                          * object, we can collapse it into the parent.
1730                          */
1731                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1732
1733 #if VM_NRESERVLEVEL > 0
1734                         /*
1735                          * Break any reservations from backing_object.
1736                          */
1737                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1738                                 vm_reserv_break_all(backing_object);
1739 #endif
1740
1741                         /*
1742                          * Move the pager from backing_object to object.
1743                          */
1744                         if (backing_object->type == OBJT_SWAP) {
1745                                 /*
1746                                  * swap_pager_copy() can sleep, in which case
1747                                  * the backing_object's and object's locks are
1748                                  * released and reacquired.
1749                                  * Since swap_pager_copy() is being asked to
1750                                  * destroy the source, it will change the
1751                                  * backing_object's type to OBJT_DEFAULT.
1752                                  */
1753                                 swap_pager_copy(
1754                                     backing_object,
1755                                     object,
1756                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1757
1758                                 /*
1759                                  * Free any cached pages from backing_object.
1760                                  */
1761                                 if (__predict_false(
1762                                     !vm_object_cache_is_empty(backing_object)))
1763                                         vm_page_cache_free(backing_object, 0, 0);
1764                         }
1765                         /*
1766                          * Object now shadows whatever backing_object did.
1767                          * Note that the reference to
1768                          * backing_object->backing_object moves from within
1769                          * backing_object to within object.
1770                          */
1771                         LIST_REMOVE(object, shadow_list);
1772                         backing_object->shadow_count--;
1773                         if (backing_object->backing_object) {
1774                                 VM_OBJECT_WLOCK(backing_object->backing_object);
1775                                 LIST_REMOVE(backing_object, shadow_list);
1776                                 LIST_INSERT_HEAD(
1777                                     &backing_object->backing_object->shadow_head,
1778                                     object, shadow_list);
1779                                 /*
1780                                  * The shadow_count has not changed.
1781                                  */
1782                                 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1783                         }
1784                         object->backing_object = backing_object->backing_object;
1785                         object->backing_object_offset +=
1786                             backing_object->backing_object_offset;
1787
1788                         /*
1789                          * Discard backing_object.
1790                          *
1791                          * Since the backing object has no pages, no pager left,
1792                          * and no object references within it, all that is
1793                          * necessary is to dispose of it.
1794                          */
1795                         KASSERT(backing_object->ref_count == 1, (
1796 "backing_object %p was somehow re-referenced during collapse!",
1797                             backing_object));
1798                         VM_OBJECT_WUNLOCK(backing_object);
1799                         vm_object_destroy(backing_object);
1800
1801                         object_collapses++;
1802                 } else {
1803                         vm_object_t new_backing_object;
1804
1805                         /*
1806                          * If we do not entirely shadow the backing object,
1807                          * there is nothing we can do so we give up.
1808                          */
1809                         if (object->resident_page_count != object->size &&
1810                             vm_object_backing_scan(object,
1811                             OBSC_TEST_ALL_SHADOWED) == 0) {
1812                                 VM_OBJECT_WUNLOCK(backing_object);
1813                                 break;
1814                         }
1815
1816                         /*
1817                          * Make the parent shadow the next object in the
1818                          * chain.  Deallocating backing_object will not remove
1819                          * it, since its reference count is at least 2.
1820                          */
1821                         LIST_REMOVE(object, shadow_list);
1822                         backing_object->shadow_count--;
1823
1824                         new_backing_object = backing_object->backing_object;
1825                         if ((object->backing_object = new_backing_object) != NULL) {
1826                                 VM_OBJECT_WLOCK(new_backing_object);
1827                                 LIST_INSERT_HEAD(
1828                                     &new_backing_object->shadow_head,
1829                                     object,
1830                                     shadow_list
1831                                 );
1832                                 new_backing_object->shadow_count++;
1833                                 vm_object_reference_locked(new_backing_object);
1834                                 VM_OBJECT_WUNLOCK(new_backing_object);
1835                                 object->backing_object_offset +=
1836                                         backing_object->backing_object_offset;
1837                         }
1838
1839                         /*
1840                          * Drop the reference count on backing_object. Since
1841                          * its ref_count was at least 2, it will not vanish.
1842                          */
1843                         backing_object->ref_count--;
1844                         VM_OBJECT_WUNLOCK(backing_object);
1845                         object_bypasses++;
1846                 }
1847
1848                 /*
1849                  * Try again with this object's new backing object.
1850                  */
1851         }
1852 }
1853
1854 /*
1855  *      vm_object_page_remove:
1856  *
1857  *      For the given object, either frees or invalidates each of the
1858  *      specified pages.  In general, a page is freed.  However, if a page is
1859  *      wired for any reason other than the existence of a managed, wired
1860  *      mapping, then it may be invalidated but not removed from the object.
1861  *      Pages are specified by the given range ["start", "end") and the option
1862  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1863  *      extends from "start" to the end of the object.  If the option
1864  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1865  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1866  *      specified, then the pages within the specified range must have no
1867  *      mappings.  Otherwise, if this option is not specified, any mappings to
1868  *      the specified pages are removed before the pages are freed or
1869  *      invalidated.
1870  *
1871  *      In general, this operation should only be performed on objects that
1872  *      contain managed pages.  There are, however, two exceptions.  First, it
1873  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1874  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1875  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1876  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1877  *
1878  *      The object must be locked.
1879  */
1880 void
1881 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1882     int options)
1883 {
1884         vm_page_t p, next;
1885         int wirings;
1886
1887         VM_OBJECT_ASSERT_WLOCKED(object);
1888         KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1889             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1890             ("vm_object_page_remove: illegal options for object %p", object));
1891         if (object->resident_page_count == 0)
1892                 goto skipmemq;
1893         vm_object_pip_add(object, 1);
1894 again:
1895         p = vm_page_find_least(object, start);
1896
1897         /*
1898          * Here, the variable "p" is either (1) the page with the least pindex
1899          * greater than or equal to the parameter "start" or (2) NULL.
1900          */
1901         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1902                 next = TAILQ_NEXT(p, listq);
1903
1904                 /*
1905                  * If the page is wired for any reason besides the existence
1906                  * of managed, wired mappings, then it cannot be freed.  For
1907                  * example, fictitious pages, which represent device memory,
1908                  * are inherently wired and cannot be freed.  They can,
1909                  * however, be invalidated if the option OBJPR_CLEANONLY is
1910                  * not specified.
1911                  */
1912                 vm_page_lock(p);
1913                 if (vm_page_xbusied(p)) {
1914                         VM_OBJECT_WUNLOCK(object);
1915                         vm_page_busy_sleep(p, "vmopax");
1916                         VM_OBJECT_WLOCK(object);
1917                         goto again;
1918                 }
1919                 if ((wirings = p->wire_count) != 0 &&
1920                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1921                         if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1922                             0) {
1923                                 pmap_remove_all(p);
1924                                 /* Account for removal of wired mappings. */
1925                                 if (wirings != 0)
1926                                         p->wire_count -= wirings;
1927                         }
1928                         if ((options & OBJPR_CLEANONLY) == 0) {
1929                                 p->valid = 0;
1930                                 vm_page_undirty(p);
1931                         }
1932                         goto next;
1933                 }
1934                 if (vm_page_busied(p)) {
1935                         VM_OBJECT_WUNLOCK(object);
1936                         vm_page_busy_sleep(p, "vmopar");
1937                         VM_OBJECT_WLOCK(object);
1938                         goto again;
1939                 }
1940                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1941                     ("vm_object_page_remove: page %p is fictitious", p));
1942                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1943                         if ((options & OBJPR_NOTMAPPED) == 0)
1944                                 pmap_remove_write(p);
1945                         if (p->dirty)
1946                                 goto next;
1947                 }
1948                 if ((options & OBJPR_NOTMAPPED) == 0) {
1949                         if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1950                                 goto next;
1951                         pmap_remove_all(p);
1952                         /* Account for removal of wired mappings. */
1953                         if (wirings != 0) {
1954                                 KASSERT(p->wire_count == wirings,
1955                                     ("inconsistent wire count %d %d %p",
1956                                     p->wire_count, wirings, p));
1957                                 p->wire_count = 0;
1958                                 atomic_subtract_int(&cnt.v_wire_count, 1);
1959                         }
1960                 }
1961                 vm_page_free(p);
1962 next:
1963                 vm_page_unlock(p);
1964         }
1965         vm_object_pip_wakeup(object);
1966 skipmemq:
1967         if (__predict_false(!vm_object_cache_is_empty(object)))
1968                 vm_page_cache_free(object, start, end);
1969 }
1970
1971 /*
1972  *      vm_object_page_cache:
1973  *
1974  *      For the given object, attempt to move the specified clean
1975  *      pages to the cache queue.  If a page is wired for any reason,
1976  *      then it will not be changed.  Pages are specified by the given
1977  *      range ["start", "end").  As a special case, if "end" is zero,
1978  *      then the range extends from "start" to the end of the object.
1979  *      Any mappings to the specified pages are removed before the
1980  *      pages are moved to the cache queue.
1981  *
1982  *      This operation should only be performed on objects that
1983  *      contain non-fictitious, managed pages.
1984  *
1985  *      The object must be locked.
1986  */
1987 void
1988 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1989 {
1990         struct mtx *mtx, *new_mtx;
1991         vm_page_t p, next;
1992
1993         VM_OBJECT_ASSERT_WLOCKED(object);
1994         KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1995             ("vm_object_page_cache: illegal object %p", object));
1996         if (object->resident_page_count == 0)
1997                 return;
1998         p = vm_page_find_least(object, start);
1999
2000         /*
2001          * Here, the variable "p" is either (1) the page with the least pindex
2002          * greater than or equal to the parameter "start" or (2) NULL.
2003          */
2004         mtx = NULL;
2005         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2006                 next = TAILQ_NEXT(p, listq);
2007
2008                 /*
2009                  * Avoid releasing and reacquiring the same page lock.
2010                  */
2011                 new_mtx = vm_page_lockptr(p);
2012                 if (mtx != new_mtx) {
2013                         if (mtx != NULL)
2014                                 mtx_unlock(mtx);
2015                         mtx = new_mtx;
2016                         mtx_lock(mtx);
2017                 }
2018                 vm_page_try_to_cache(p);
2019         }
2020         if (mtx != NULL)
2021                 mtx_unlock(mtx);
2022 }
2023
2024 /*
2025  *      Populate the specified range of the object with valid pages.  Returns
2026  *      TRUE if the range is successfully populated and FALSE otherwise.
2027  *
2028  *      Note: This function should be optimized to pass a larger array of
2029  *      pages to vm_pager_get_pages() before it is applied to a non-
2030  *      OBJT_DEVICE object.
2031  *
2032  *      The object must be locked.
2033  */
2034 boolean_t
2035 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2036 {
2037         vm_page_t m, ma[1];
2038         vm_pindex_t pindex;
2039         int rv;
2040
2041         VM_OBJECT_ASSERT_WLOCKED(object);
2042         for (pindex = start; pindex < end; pindex++) {
2043                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2044                 if (m->valid != VM_PAGE_BITS_ALL) {
2045                         ma[0] = m;
2046                         rv = vm_pager_get_pages(object, ma, 1, 0);
2047                         m = vm_page_lookup(object, pindex);
2048                         if (m == NULL)
2049                                 break;
2050                         if (rv != VM_PAGER_OK) {
2051                                 vm_page_lock(m);
2052                                 vm_page_free(m);
2053                                 vm_page_unlock(m);
2054                                 break;
2055                         }
2056                 }
2057                 /*
2058                  * Keep "m" busy because a subsequent iteration may unlock
2059                  * the object.
2060                  */
2061         }
2062         if (pindex > start) {
2063                 m = vm_page_lookup(object, start);
2064                 while (m != NULL && m->pindex < pindex) {
2065                         vm_page_xunbusy(m);
2066                         m = TAILQ_NEXT(m, listq);
2067                 }
2068         }
2069         return (pindex == end);
2070 }
2071
2072 /*
2073  *      Routine:        vm_object_coalesce
2074  *      Function:       Coalesces two objects backing up adjoining
2075  *                      regions of memory into a single object.
2076  *
2077  *      returns TRUE if objects were combined.
2078  *
2079  *      NOTE:   Only works at the moment if the second object is NULL -
2080  *              if it's not, which object do we lock first?
2081  *
2082  *      Parameters:
2083  *              prev_object     First object to coalesce
2084  *              prev_offset     Offset into prev_object
2085  *              prev_size       Size of reference to prev_object
2086  *              next_size       Size of reference to the second object
2087  *              reserved        Indicator that extension region has
2088  *                              swap accounted for
2089  *
2090  *      Conditions:
2091  *      The object must *not* be locked.
2092  */
2093 boolean_t
2094 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2095     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2096 {
2097         vm_pindex_t next_pindex;
2098
2099         if (prev_object == NULL)
2100                 return (TRUE);
2101         VM_OBJECT_WLOCK(prev_object);
2102         if (prev_object->type != OBJT_DEFAULT &&
2103             prev_object->type != OBJT_SWAP) {
2104                 VM_OBJECT_WUNLOCK(prev_object);
2105                 return (FALSE);
2106         }
2107
2108         /*
2109          * Try to collapse the object first
2110          */
2111         vm_object_collapse(prev_object);
2112
2113         /*
2114          * Can't coalesce if: . more than one reference . paged out . shadows
2115          * another object . has a copy elsewhere (any of which mean that the
2116          * pages not mapped to prev_entry may be in use anyway)
2117          */
2118         if (prev_object->backing_object != NULL) {
2119                 VM_OBJECT_WUNLOCK(prev_object);
2120                 return (FALSE);
2121         }
2122
2123         prev_size >>= PAGE_SHIFT;
2124         next_size >>= PAGE_SHIFT;
2125         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2126
2127         if ((prev_object->ref_count > 1) &&
2128             (prev_object->size != next_pindex)) {
2129                 VM_OBJECT_WUNLOCK(prev_object);
2130                 return (FALSE);
2131         }
2132
2133         /*
2134          * Account for the charge.
2135          */
2136         if (prev_object->cred != NULL) {
2137
2138                 /*
2139                  * If prev_object was charged, then this mapping,
2140                  * althought not charged now, may become writable
2141                  * later. Non-NULL cred in the object would prevent
2142                  * swap reservation during enabling of the write
2143                  * access, so reserve swap now. Failed reservation
2144                  * cause allocation of the separate object for the map
2145                  * entry, and swap reservation for this entry is
2146                  * managed in appropriate time.
2147                  */
2148                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2149                     prev_object->cred)) {
2150                         return (FALSE);
2151                 }
2152                 prev_object->charge += ptoa(next_size);
2153         }
2154
2155         /*
2156          * Remove any pages that may still be in the object from a previous
2157          * deallocation.
2158          */
2159         if (next_pindex < prev_object->size) {
2160                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2161                     next_size, 0);
2162                 if (prev_object->type == OBJT_SWAP)
2163                         swap_pager_freespace(prev_object,
2164                                              next_pindex, next_size);
2165 #if 0
2166                 if (prev_object->cred != NULL) {
2167                         KASSERT(prev_object->charge >=
2168                             ptoa(prev_object->size - next_pindex),
2169                             ("object %p overcharged 1 %jx %jx", prev_object,
2170                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2171                         prev_object->charge -= ptoa(prev_object->size -
2172                             next_pindex);
2173                 }
2174 #endif
2175         }
2176
2177         /*
2178          * Extend the object if necessary.
2179          */
2180         if (next_pindex + next_size > prev_object->size)
2181                 prev_object->size = next_pindex + next_size;
2182
2183         VM_OBJECT_WUNLOCK(prev_object);
2184         return (TRUE);
2185 }
2186
2187 void
2188 vm_object_set_writeable_dirty(vm_object_t object)
2189 {
2190
2191         VM_OBJECT_ASSERT_WLOCKED(object);
2192         if (object->type != OBJT_VNODE)
2193                 return;
2194         object->generation++;
2195         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2196                 return;
2197         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2198 }
2199
2200 #include "opt_ddb.h"
2201 #ifdef DDB
2202 #include <sys/kernel.h>
2203
2204 #include <sys/cons.h>
2205
2206 #include <ddb/ddb.h>
2207
2208 static int
2209 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2210 {
2211         vm_map_t tmpm;
2212         vm_map_entry_t tmpe;
2213         vm_object_t obj;
2214         int entcount;
2215
2216         if (map == 0)
2217                 return 0;
2218
2219         if (entry == 0) {
2220                 tmpe = map->header.next;
2221                 entcount = map->nentries;
2222                 while (entcount-- && (tmpe != &map->header)) {
2223                         if (_vm_object_in_map(map, object, tmpe)) {
2224                                 return 1;
2225                         }
2226                         tmpe = tmpe->next;
2227                 }
2228         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2229                 tmpm = entry->object.sub_map;
2230                 tmpe = tmpm->header.next;
2231                 entcount = tmpm->nentries;
2232                 while (entcount-- && tmpe != &tmpm->header) {
2233                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2234                                 return 1;
2235                         }
2236                         tmpe = tmpe->next;
2237                 }
2238         } else if ((obj = entry->object.vm_object) != NULL) {
2239                 for (; obj; obj = obj->backing_object)
2240                         if (obj == object) {
2241                                 return 1;
2242                         }
2243         }
2244         return 0;
2245 }
2246
2247 static int
2248 vm_object_in_map(vm_object_t object)
2249 {
2250         struct proc *p;
2251
2252         /* sx_slock(&allproc_lock); */
2253         FOREACH_PROC_IN_SYSTEM(p) {
2254                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2255                         continue;
2256                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2257                         /* sx_sunlock(&allproc_lock); */
2258                         return 1;
2259                 }
2260         }
2261         /* sx_sunlock(&allproc_lock); */
2262         if (_vm_object_in_map(kernel_map, object, 0))
2263                 return 1;
2264         return 0;
2265 }
2266
2267 DB_SHOW_COMMAND(vmochk, vm_object_check)
2268 {
2269         vm_object_t object;
2270
2271         /*
2272          * make sure that internal objs are in a map somewhere
2273          * and none have zero ref counts.
2274          */
2275         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2276                 if (object->handle == NULL &&
2277                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2278                         if (object->ref_count == 0) {
2279                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2280                                         (long)object->size);
2281                         }
2282                         if (!vm_object_in_map(object)) {
2283                                 db_printf(
2284                         "vmochk: internal obj is not in a map: "
2285                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2286                                     object->ref_count, (u_long)object->size,
2287                                     (u_long)object->size,
2288                                     (void *)object->backing_object);
2289                         }
2290                 }
2291         }
2292 }
2293
2294 /*
2295  *      vm_object_print:        [ debug ]
2296  */
2297 DB_SHOW_COMMAND(object, vm_object_print_static)
2298 {
2299         /* XXX convert args. */
2300         vm_object_t object = (vm_object_t)addr;
2301         boolean_t full = have_addr;
2302
2303         vm_page_t p;
2304
2305         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2306 #define count   was_count
2307
2308         int count;
2309
2310         if (object == NULL)
2311                 return;
2312
2313         db_iprintf(
2314             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2315             object, (int)object->type, (uintmax_t)object->size,
2316             object->resident_page_count, object->ref_count, object->flags,
2317             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2318         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2319             object->shadow_count,
2320             object->backing_object ? object->backing_object->ref_count : 0,
2321             object->backing_object, (uintmax_t)object->backing_object_offset);
2322
2323         if (!full)
2324                 return;
2325
2326         db_indent += 2;
2327         count = 0;
2328         TAILQ_FOREACH(p, &object->memq, listq) {
2329                 if (count == 0)
2330                         db_iprintf("memory:=");
2331                 else if (count == 6) {
2332                         db_printf("\n");
2333                         db_iprintf(" ...");
2334                         count = 0;
2335                 } else
2336                         db_printf(",");
2337                 count++;
2338
2339                 db_printf("(off=0x%jx,page=0x%jx)",
2340                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2341         }
2342         if (count != 0)
2343                 db_printf("\n");
2344         db_indent -= 2;
2345 }
2346
2347 /* XXX. */
2348 #undef count
2349
2350 /* XXX need this non-static entry for calling from vm_map_print. */
2351 void
2352 vm_object_print(
2353         /* db_expr_t */ long addr,
2354         boolean_t have_addr,
2355         /* db_expr_t */ long count,
2356         char *modif)
2357 {
2358         vm_object_print_static(addr, have_addr, count, modif);
2359 }
2360
2361 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2362 {
2363         vm_object_t object;
2364         vm_pindex_t fidx;
2365         vm_paddr_t pa;
2366         vm_page_t m, prev_m;
2367         int rcount, nl, c;
2368
2369         nl = 0;
2370         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2371                 db_printf("new object: %p\n", (void *)object);
2372                 if (nl > 18) {
2373                         c = cngetc();
2374                         if (c != ' ')
2375                                 return;
2376                         nl = 0;
2377                 }
2378                 nl++;
2379                 rcount = 0;
2380                 fidx = 0;
2381                 pa = -1;
2382                 TAILQ_FOREACH(m, &object->memq, listq) {
2383                         if (m->pindex > 128)
2384                                 break;
2385                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2386                             prev_m->pindex + 1 != m->pindex) {
2387                                 if (rcount) {
2388                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2389                                                 (long)fidx, rcount, (long)pa);
2390                                         if (nl > 18) {
2391                                                 c = cngetc();
2392                                                 if (c != ' ')
2393                                                         return;
2394                                                 nl = 0;
2395                                         }
2396                                         nl++;
2397                                         rcount = 0;
2398                                 }
2399                         }
2400                         if (rcount &&
2401                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2402                                 ++rcount;
2403                                 continue;
2404                         }
2405                         if (rcount) {
2406                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2407                                         (long)fidx, rcount, (long)pa);
2408                                 if (nl > 18) {
2409                                         c = cngetc();
2410                                         if (c != ' ')
2411                                                 return;
2412                                         nl = 0;
2413                                 }
2414                                 nl++;
2415                         }
2416                         fidx = m->pindex;
2417                         pa = VM_PAGE_TO_PHYS(m);
2418                         rcount = 1;
2419                 }
2420                 if (rcount) {
2421                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2422                                 (long)fidx, rcount, (long)pa);
2423                         if (nl > 18) {
2424                                 c = cngetc();
2425                                 if (c != ' ')
2426                                         return;
2427                                 nl = 0;
2428                         }
2429                         nl++;
2430                 }
2431         }
2432 }
2433 #endif /* DDB */