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