sys/vm/vm_object.c

   1 /*-
   2  * Copyright (c) 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * The Mach Operating System project at Carnegie-Mellon University.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 4. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
  33  *
  34  *
  35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  36  * All rights reserved.
  37  *
  38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  39  *
  40  * Permission to use, copy, modify and distribute this software and
  41  * its documentation is hereby granted, provided that both the copyright
  42  * notice and this permission notice appear in all copies of the
  43  * software, derivative works or modified versions, and any portions
  44  * thereof, and that both notices appear in supporting documentation.
  45  *
  46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  49  *
  50  * Carnegie Mellon requests users of this software to return to
  51  *
  52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  53  *  School of Computer Science
  54  *  Carnegie Mellon University
  55  *  Pittsburgh PA 15213-3890
  56  *
  57  * any improvements or extensions that they make and grant Carnegie the
  58  * rights to redistribute these changes.
  59  */
  60
  61 /*
  62  *      Virtual memory object module.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_vm.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/lock.h>
  73 #include <sys/mman.h>
  74 #include <sys/mount.h>
  75 #include <sys/kernel.h>
  76 #include <sys/sysctl.h>
  77 #include <sys/mutex.h>
  78 #include <sys/proc.h>           /* for curproc, pageproc */
  79 #include <sys/socket.h>
  80 #include <sys/resourcevar.h>
  81 #include <sys/rwlock.h>
  82 #include <sys/vnode.h>
  83 #include <sys/vmmeter.h>
  84 #include <sys/sx.h>
  85
  86 #include <vm/vm.h>
  87 #include <vm/vm_param.h>
  88 #include <vm/pmap.h>
  89 #include <vm/vm_map.h>
  90 #include <vm/vm_object.h>
  91 #include <vm/vm_page.h>
  92 #include <vm/vm_pageout.h>
  93 #include <vm/vm_pager.h>
  94 #include <vm/swap_pager.h>
  95 #include <vm/vm_kern.h>
  96 #include <vm/vm_extern.h>
  97 #include <vm/vm_radix.h>
  98 #include <vm/vm_reserv.h>
  99 #include <vm/uma.h>
 100
 101 static int old_msync;
 102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
 103     "Use old (insecure) msync behavior");
 104
 105 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
 106                     int pagerflags, int flags, boolean_t *clearobjflags,
 107                     boolean_t *eio);
 108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
 109                     boolean_t *clearobjflags);
 110 static void     vm_object_qcollapse(vm_object_t object);
 111 static void     vm_object_vndeallocate(vm_object_t object);
 112
 113 /*
 114  *      Virtual memory objects maintain the actual data
 115  *      associated with allocated virtual memory.  A given
 116  *      page of memory exists within exactly one object.
 117  *
 118  *      An object is only deallocated when all "references"
 119  *      are given up.  Only one "reference" to a given
 120  *      region of an object should be writeable.
 121  *
 122  *      Associated with each object is a list of all resident
 123  *      memory pages belonging to that object; this list is
 124  *      maintained by the "vm_page" module, and locked by the object's
 125  *      lock.
 126  *
 127  *      Each object also records a "pager" routine which is
 128  *      used to retrieve (and store) pages to the proper backing
 129  *      storage.  In addition, objects may be backed by other
 130  *      objects from which they were virtual-copied.
 131  *
 132  *      The only items within the object structure which are
 133  *      modified after time of creation are:
 134  *              reference count         locked by object's lock
 135  *              pager routine           locked by object's lock
 136  *
 137  */
 138
 139 struct object_q vm_object_list;
 140 struct mtx vm_object_list_mtx;  /* lock for object list and count */
 141
 142 struct vm_object kernel_object_store;
 143 struct vm_object kmem_object_store;
 144
 145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
 146     "VM object stats");
 147
 148 static long object_collapses;
 149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
 150     &object_collapses, 0, "VM object collapses");
 151
 152 static long object_bypasses;
 153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
 154     &object_bypasses, 0, "VM object bypasses");
 155
 156 static uma_zone_t obj_zone;
 157
 158 static int vm_object_zinit(void *mem, int size, int flags);
 159
 160 #ifdef INVARIANTS
 161 static void vm_object_zdtor(void *mem, int size, void *arg);
 162
 163 static void
 164 vm_object_zdtor(void *mem, int size, void *arg)
 165 {
 166         vm_object_t object;
 167
 168         object = (vm_object_t)mem;
 169         KASSERT(TAILQ_EMPTY(&object->memq),
 170             ("object %p has resident pages in its memq", object));
 171         KASSERT(vm_radix_is_empty(&object->rtree),
 172             ("object %p has resident pages in its trie", object));
 173 #if VM_NRESERVLEVEL > 0
 174         KASSERT(LIST_EMPTY(&object->rvq),
 175             ("object %p has reservations",
 176             object));
 177 #endif
 178         KASSERT(vm_object_cache_is_empty(object),
 179             ("object %p has cached pages",
 180             object));
 181         KASSERT(object->paging_in_progress == 0,
 182             ("object %p paging_in_progress = %d",
 183             object, object->paging_in_progress));
 184         KASSERT(object->resident_page_count == 0,
 185             ("object %p resident_page_count = %d",
 186             object, object->resident_page_count));
 187         KASSERT(object->shadow_count == 0,
 188             ("object %p shadow_count = %d",
 189             object, object->shadow_count));
 190 }
 191 #endif
 192
 193 static int
 194 vm_object_zinit(void *mem, int size, int flags)
 195 {
 196         vm_object_t object;
 197
 198         object = (vm_object_t)mem;
 199         bzero(&object->lock, sizeof(object->lock));
 200         rw_init_flags(&object->lock, "vm object", RW_DUPOK);
 201
 202         /* These are true for any object that has been freed */
 203         object->rtree.rt_root = 0;
 204         object->rtree.rt_flags = 0;
 205         object->paging_in_progress = 0;
 206         object->resident_page_count = 0;
 207         object->shadow_count = 0;
 208         object->cache.rt_root = 0;
 209         object->cache.rt_flags = 0;
 210         return (0);
 211 }
 212
 213 static void
 214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
 215 {
 216
 217         TAILQ_INIT(&object->memq);
 218         LIST_INIT(&object->shadow_head);
 219
 220         object->type = type;
 221         switch (type) {
 222         case OBJT_DEAD:
 223                 panic("_vm_object_allocate: can't create OBJT_DEAD");
 224         case OBJT_DEFAULT:
 225         case OBJT_SWAP:
 226                 object->flags = OBJ_ONEMAPPING;
 227                 break;
 228         case OBJT_DEVICE:
 229         case OBJT_SG:
 230                 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
 231                 break;
 232         case OBJT_MGTDEVICE:
 233                 object->flags = OBJ_FICTITIOUS;
 234                 break;
 235         case OBJT_PHYS:
 236                 object->flags = OBJ_UNMANAGED;
 237                 break;
 238         case OBJT_VNODE:
 239                 object->flags = 0;
 240                 break;
 241         default:
 242                 panic("_vm_object_allocate: type %d is undefined", type);
 243         }
 244         object->size = size;
 245         object->generation = 1;
 246         object->ref_count = 1;
 247         object->memattr = VM_MEMATTR_DEFAULT;
 248         object->cred = NULL;
 249         object->charge = 0;
 250         object->handle = NULL;
 251         object->backing_object = NULL;
 252         object->backing_object_offset = (vm_ooffset_t) 0;
 253 #if VM_NRESERVLEVEL > 0
 254         LIST_INIT(&object->rvq);
 255 #endif
 256
 257         mtx_lock(&vm_object_list_mtx);
 258         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
 259         mtx_unlock(&vm_object_list_mtx);
 260 }
 261
 262 /*
 263  *      vm_object_init:
 264  *
 265  *      Initialize the VM objects module.
 266  */
 267 void
 268 vm_object_init(void)
 269 {
 270         TAILQ_INIT(&vm_object_list);
 271         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
 272
 273         rw_init(&kernel_object->lock, "kernel vm object");
 274         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 275             kernel_object);
 276 #if VM_NRESERVLEVEL > 0
 277         kernel_object->flags |= OBJ_COLORED;
 278         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 279 #endif
 280
 281         rw_init(&kmem_object->lock, "kmem vm object");
 282         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
 283             kmem_object);
 284 #if VM_NRESERVLEVEL > 0
 285         kmem_object->flags |= OBJ_COLORED;
 286         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
 287 #endif
 288
 289         /*
 290          * The lock portion of struct vm_object must be type stable due
 291          * to vm_pageout_fallback_object_lock locking a vm object
 292          * without holding any references to it.
 293          */
 294         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
 295 #ifdef INVARIANTS
 296             vm_object_zdtor,
 297 #else
 298             NULL,
 299 #endif
 300             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 301
 302         vm_radix_init();
 303 }
 304
 305 void
 306 vm_object_clear_flag(vm_object_t object, u_short bits)
 307 {
 308
 309         VM_OBJECT_ASSERT_WLOCKED(object);
 310         object->flags &= ~bits;
 311 }
 312
 313 /*
 314  *      Sets the default memory attribute for the specified object.  Pages
 315  *      that are allocated to this object are by default assigned this memory
 316  *      attribute.
 317  *
 318  *      Presently, this function must be called before any pages are allocated
 319  *      to the object.  In the future, this requirement may be relaxed for
 320  *      "default" and "swap" objects.
 321  */
 322 int
 323 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
 324 {
 325
 326         VM_OBJECT_ASSERT_WLOCKED(object);
 327         switch (object->type) {
 328         case OBJT_DEFAULT:
 329         case OBJT_DEVICE:
 330         case OBJT_MGTDEVICE:
 331         case OBJT_PHYS:
 332         case OBJT_SG:
 333         case OBJT_SWAP:
 334         case OBJT_VNODE:
 335                 if (!TAILQ_EMPTY(&object->memq))
 336                         return (KERN_FAILURE);
 337                 break;
 338         case OBJT_DEAD:
 339                 return (KERN_INVALID_ARGUMENT);
 340         default:
 341                 panic("vm_object_set_memattr: object %p is of undefined type",
 342                     object);
 343         }
 344         object->memattr = memattr;
 345         return (KERN_SUCCESS);
 346 }
 347
 348 void
 349 vm_object_pip_add(vm_object_t object, short i)
 350 {
 351
 352         VM_OBJECT_ASSERT_WLOCKED(object);
 353         object->paging_in_progress += i;
 354 }
 355
 356 void
 357 vm_object_pip_subtract(vm_object_t object, short i)
 358 {
 359
 360         VM_OBJECT_ASSERT_WLOCKED(object);
 361         object->paging_in_progress -= i;
 362 }
 363
 364 void
 365 vm_object_pip_wakeup(vm_object_t object)
 366 {
 367
 368         VM_OBJECT_ASSERT_WLOCKED(object);
 369         object->paging_in_progress--;
 370         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 371                 vm_object_clear_flag(object, OBJ_PIPWNT);
 372                 wakeup(object);
 373         }
 374 }
 375
 376 void
 377 vm_object_pip_wakeupn(vm_object_t object, short i)
 378 {
 379
 380         VM_OBJECT_ASSERT_WLOCKED(object);
 381         if (i)
 382                 object->paging_in_progress -= i;
 383         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
 384                 vm_object_clear_flag(object, OBJ_PIPWNT);
 385                 wakeup(object);
 386         }
 387 }
 388
 389 void
 390 vm_object_pip_wait(vm_object_t object, char *waitid)
 391 {
 392
 393         VM_OBJECT_ASSERT_WLOCKED(object);
 394         while (object->paging_in_progress) {
 395                 object->flags |= OBJ_PIPWNT;
 396                 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
 397         }
 398 }
 399
 400 /*
 401  *      vm_object_allocate:
 402  *
 403  *      Returns a new object with the given size.
 404  */
 405 vm_object_t
 406 vm_object_allocate(objtype_t type, vm_pindex_t size)
 407 {
 408         vm_object_t object;
 409
 410         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
 411         _vm_object_allocate(type, size, object);
 412         return (object);
 413 }
 414
 415
 416 /*
 417  *      vm_object_reference:
 418  *
 419  *      Gets another reference to the given object.  Note: OBJ_DEAD
 420  *      objects can be referenced during final cleaning.
 421  */
 422 void
 423 vm_object_reference(vm_object_t object)
 424 {
 425         if (object == NULL)
 426                 return;
 427         VM_OBJECT_WLOCK(object);
 428         vm_object_reference_locked(object);
 429         VM_OBJECT_WUNLOCK(object);
 430 }
 431
 432 /*
 433  *      vm_object_reference_locked:
 434  *
 435  *      Gets another reference to the given object.
 436  *
 437  *      The object must be locked.
 438  */
 439 void
 440 vm_object_reference_locked(vm_object_t object)
 441 {
 442         struct vnode *vp;
 443
 444         VM_OBJECT_ASSERT_WLOCKED(object);
 445         object->ref_count++;
 446         if (object->type == OBJT_VNODE) {
 447                 vp = object->handle;
 448                 vref(vp);
 449         }
 450 }
 451
 452 /*
 453  * Handle deallocating an object of type OBJT_VNODE.
 454  */
 455 static void
 456 vm_object_vndeallocate(vm_object_t object)
 457 {
 458         struct vnode *vp = (struct vnode *) object->handle;
 459
 460         VM_OBJECT_ASSERT_WLOCKED(object);
 461         KASSERT(object->type == OBJT_VNODE,
 462             ("vm_object_vndeallocate: not a vnode object"));
 463         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
 464 #ifdef INVARIANTS
 465         if (object->ref_count == 0) {
 466                 vprint("vm_object_vndeallocate", vp);
 467                 panic("vm_object_vndeallocate: bad object reference count");
 468         }
 469 #endif
 470
 471         if (object->ref_count > 1) {
 472                 object->ref_count--;
 473                 VM_OBJECT_WUNLOCK(object);
 474                 /* vrele may need the vnode lock. */
 475                 vrele(vp);
 476         } else {
 477                 vhold(vp);
 478                 VM_OBJECT_WUNLOCK(object);
 479                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 480                 vdrop(vp);
 481                 VM_OBJECT_WLOCK(object);
 482                 object->ref_count--;
 483                 if (object->type == OBJT_DEAD) {
 484                         VM_OBJECT_WUNLOCK(object);
 485                         VOP_UNLOCK(vp, 0);
 486                 } else {
 487                         if (object->ref_count == 0)
 488                                 VOP_UNSET_TEXT(vp);
 489                         VM_OBJECT_WUNLOCK(object);
 490                         vput(vp);
 491                 }
 492         }
 493 }
 494
 495 /*
 496  *      vm_object_deallocate:
 497  *
 498  *      Release a reference to the specified object,
 499  *      gained either through a vm_object_allocate
 500  *      or a vm_object_reference call.  When all references
 501  *      are gone, storage associated with this object
 502  *      may be relinquished.
 503  *
 504  *      No object may be locked.
 505  */
 506 void
 507 vm_object_deallocate(vm_object_t object)
 508 {
 509         vm_object_t temp;
 510         struct vnode *vp;
 511
 512         while (object != NULL) {
 513                 VM_OBJECT_WLOCK(object);
 514                 if (object->type == OBJT_VNODE) {
 515                         vm_object_vndeallocate(object);
 516                         return;
 517                 }
 518
 519                 KASSERT(object->ref_count != 0,
 520                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
 521
 522                 /*
 523                  * If the reference count goes to 0 we start calling
 524                  * vm_object_terminate() on the object chain.
 525                  * A ref count of 1 may be a special case depending on the
 526                  * shadow count being 0 or 1.
 527                  */
 528                 object->ref_count--;
 529                 if (object->ref_count > 1) {
 530                         VM_OBJECT_WUNLOCK(object);
 531                         return;
 532                 } else if (object->ref_count == 1) {
 533                         if (object->type == OBJT_SWAP &&
 534                             (object->flags & OBJ_TMPFS) != 0) {
 535                                 vp = object->un_pager.swp.swp_tmpfs;
 536                                 vhold(vp);
 537                                 VM_OBJECT_WUNLOCK(object);
 538                                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 539                                 VM_OBJECT_WLOCK(object);
 540                                 if (object->type == OBJT_DEAD ||
 541                                     object->ref_count != 1) {
 542                                         VM_OBJECT_WUNLOCK(object);
 543                                         VOP_UNLOCK(vp, 0);
 544                                         vdrop(vp);
 545                                         return;
 546                                 }
 547                                 if ((object->flags & OBJ_TMPFS) != 0)
 548                                         VOP_UNSET_TEXT(vp);
 549                                 VOP_UNLOCK(vp, 0);
 550                                 vdrop(vp);
 551                         }
 552                         if (object->shadow_count == 0 &&
 553                             object->handle == NULL &&
 554                             (object->type == OBJT_DEFAULT ||
 555                             (object->type == OBJT_SWAP &&
 556                             (object->flags & OBJ_TMPFS_NODE) == 0))) {
 557                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
 558                         } else if ((object->shadow_count == 1) &&
 559                             (object->handle == NULL) &&
 560                             (object->type == OBJT_DEFAULT ||
 561                              object->type == OBJT_SWAP)) {
 562                                 vm_object_t robject;
 563
 564                                 robject = LIST_FIRST(&object->shadow_head);
 565                                 KASSERT(robject != NULL,
 566                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
 567                                          object->ref_count,
 568                                          object->shadow_count));
 569                                 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
 570                                     ("shadowed tmpfs v_object %p", object));
 571                                 if (!VM_OBJECT_TRYWLOCK(robject)) {
 572                                         /*
 573                                          * Avoid a potential deadlock.
 574                                          */
 575                                         object->ref_count++;
 576                                         VM_OBJECT_WUNLOCK(object);
 577                                         /*
 578                                          * More likely than not the thread
 579                                          * holding robject's lock has lower
 580                                          * priority than the current thread.
 581                                          * Let the lower priority thread run.
 582                                          */
 583                                         pause("vmo_de", 1);
 584                                         continue;
 585                                 }
 586                                 /*
 587                                  * Collapse object into its shadow unless its
 588                                  * shadow is dead.  In that case, object will
 589                                  * be deallocated by the thread that is
 590                                  * deallocating its shadow.
 591                                  */
 592                                 if ((robject->flags & OBJ_DEAD) == 0 &&
 593                                     (robject->handle == NULL) &&
 594                                     (robject->type == OBJT_DEFAULT ||
 595                                      robject->type == OBJT_SWAP)) {
 596
 597                                         robject->ref_count++;
 598 retry:
 599                                         if (robject->paging_in_progress) {
 600                                                 VM_OBJECT_WUNLOCK(object);
 601                                                 vm_object_pip_wait(robject,
 602                                                     "objde1");
 603                                                 temp = robject->backing_object;
 604                                                 if (object == temp) {
 605                                                         VM_OBJECT_WLOCK(object);
 606                                                         goto retry;
 607                                                 }
 608                                         } else if (object->paging_in_progress) {
 609                                                 VM_OBJECT_WUNLOCK(robject);
 610                                                 object->flags |= OBJ_PIPWNT;
 611                                                 VM_OBJECT_SLEEP(object, object,
 612                                                     PDROP | PVM, "objde2", 0);
 613                                                 VM_OBJECT_WLOCK(robject);
 614                                                 temp = robject->backing_object;
 615                                                 if (object == temp) {
 616                                                         VM_OBJECT_WLOCK(object);
 617                                                         goto retry;
 618                                                 }
 619                                         } else
 620                                                 VM_OBJECT_WUNLOCK(object);
 621
 622                                         if (robject->ref_count == 1) {
 623                                                 robject->ref_count--;
 624                                                 object = robject;
 625                                                 goto doterm;
 626                                         }
 627                                         object = robject;
 628                                         vm_object_collapse(object);
 629                                         VM_OBJECT_WUNLOCK(object);
 630                                         continue;
 631                                 }
 632                                 VM_OBJECT_WUNLOCK(robject);
 633                         }
 634                         VM_OBJECT_WUNLOCK(object);
 635                         return;
 636                 }
 637 doterm:
 638                 temp = object->backing_object;
 639                 if (temp != NULL) {
 640                         KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
 641                             ("shadowed tmpfs v_object 2 %p", object));
 642                         VM_OBJECT_WLOCK(temp);
 643                         LIST_REMOVE(object, shadow_list);
 644                         temp->shadow_count--;
 645                         VM_OBJECT_WUNLOCK(temp);
 646                         object->backing_object = NULL;
 647                 }
 648                 /*
 649                  * Don't double-terminate, we could be in a termination
 650                  * recursion due to the terminate having to sync data
 651                  * to disk.
 652                  */
 653                 if ((object->flags & OBJ_DEAD) == 0)
 654                         vm_object_terminate(object);
 655                 else
 656                         VM_OBJECT_WUNLOCK(object);
 657                 object = temp;
 658         }
 659 }
 660
 661 /*
 662  *      vm_object_destroy removes the object from the global object list
 663  *      and frees the space for the object.
 664  */
 665 void
 666 vm_object_destroy(vm_object_t object)
 667 {
 668
 669         /*
 670          * Remove the object from the global object list.
 671          */
 672         mtx_lock(&vm_object_list_mtx);
 673         TAILQ_REMOVE(&vm_object_list, object, object_list);
 674         mtx_unlock(&vm_object_list_mtx);
 675
 676         /*
 677          * Release the allocation charge.
 678          */
 679         if (object->cred != NULL) {
 680                 KASSERT(object->type == OBJT_DEFAULT ||
 681                     object->type == OBJT_SWAP,
 682                     ("vm_object_terminate: non-swap obj %p has cred",
 683                      object));
 684                 swap_release_by_cred(object->charge, object->cred);
 685                 object->charge = 0;
 686                 crfree(object->cred);
 687                 object->cred = NULL;
 688         }
 689
 690         /*
 691          * Free the space for the object.
 692          */
 693         uma_zfree(obj_zone, object);
 694 }
 695
 696 /*
 697  *      vm_object_terminate actually destroys the specified object, freeing
 698  *      up all previously used resources.
 699  *
 700  *      The object must be locked.
 701  *      This routine may block.
 702  */
 703 void
 704 vm_object_terminate(vm_object_t object)
 705 {
 706         vm_page_t p, p_next;
 707
 708         VM_OBJECT_ASSERT_WLOCKED(object);
 709
 710         /*
 711          * Make sure no one uses us.
 712          */
 713         vm_object_set_flag(object, OBJ_DEAD);
 714
 715         /*
 716          * wait for the pageout daemon to be done with the object
 717          */
 718         vm_object_pip_wait(object, "objtrm");
 719
 720         KASSERT(!object->paging_in_progress,
 721                 ("vm_object_terminate: pageout in progress"));
 722
 723         /*
 724          * Clean and free the pages, as appropriate. All references to the
 725          * object are gone, so we don't need to lock it.
 726          */
 727         if (object->type == OBJT_VNODE) {
 728                 struct vnode *vp = (struct vnode *)object->handle;
 729
 730                 /*
 731                  * Clean pages and flush buffers.
 732                  */
 733                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 734                 VM_OBJECT_WUNLOCK(object);
 735
 736                 vinvalbuf(vp, V_SAVE, 0, 0);
 737
 738                 VM_OBJECT_WLOCK(object);
 739         }
 740
 741         KASSERT(object->ref_count == 0,
 742                 ("vm_object_terminate: object with references, ref_count=%d",
 743                 object->ref_count));
 744
 745         /*
 746          * Free any remaining pageable pages.  This also removes them from the
 747          * paging queues.  However, don't free wired pages, just remove them
 748          * from the object.  Rather than incrementally removing each page from
 749          * the object, the page and object are reset to any empty state.
 750          */
 751         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
 752                 vm_page_assert_unbusied(p);
 753                 vm_page_lock(p);
 754                 /*
 755                  * Optimize the page's removal from the object by resetting
 756                  * its "object" field.  Specifically, if the page is not
 757                  * wired, then the effect of this assignment is that
 758                  * vm_page_free()'s call to vm_page_remove() will return
 759                  * immediately without modifying the page or the object.
 760                  */
 761                 p->object = NULL;
 762                 if (p->wire_count == 0) {
 763                         vm_page_free(p);
 764                         PCPU_INC(cnt.v_pfree);
 765                 }
 766                 vm_page_unlock(p);
 767         }
 768         /*
 769          * If the object contained any pages, then reset it to an empty state.
 770          * None of the object's fields, including "resident_page_count", were
 771          * modified by the preceding loop.
 772          */
 773         if (object->resident_page_count != 0) {
 774                 vm_radix_reclaim_allnodes(&object->rtree);
 775                 TAILQ_INIT(&object->memq);
 776                 object->resident_page_count = 0;
 777                 if (object->type == OBJT_VNODE)
 778                         vdrop(object->handle);
 779         }
 780
 781 #if VM_NRESERVLEVEL > 0
 782         if (__predict_false(!LIST_EMPTY(&object->rvq)))
 783                 vm_reserv_break_all(object);
 784 #endif
 785         if (__predict_false(!vm_object_cache_is_empty(object)))
 786                 vm_page_cache_free(object, 0, 0);
 787
 788         /*
 789          * Let the pager know object is dead.
 790          */
 791         vm_pager_deallocate(object);
 792         VM_OBJECT_WUNLOCK(object);
 793
 794         vm_object_destroy(object);
 795 }
 796
 797 /*
 798  * Make the page read-only so that we can clear the object flags.  However, if
 799  * this is a nosync mmap then the object is likely to stay dirty so do not
 800  * mess with the page and do not clear the object flags.  Returns TRUE if the
 801  * page should be flushed, and FALSE otherwise.
 802  */
 803 static boolean_t
 804 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
 805 {
 806
 807         /*
 808          * If we have been asked to skip nosync pages and this is a
 809          * nosync page, skip it.  Note that the object flags were not
 810          * cleared in this case so we do not have to set them.
 811          */
 812         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
 813                 *clearobjflags = FALSE;
 814                 return (FALSE);
 815         } else {
 816                 pmap_remove_write(p);
 817                 return (p->dirty != 0);
 818         }
 819 }
 820
 821 /*
 822  *      vm_object_page_clean
 823  *
 824  *      Clean all dirty pages in the specified range of object.  Leaves page
 825  *      on whatever queue it is currently on.   If NOSYNC is set then do not
 826  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
 827  *      leaving the object dirty.
 828  *
 829  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
 830  *      synchronous clustering mode implementation.
 831  *
 832  *      Odd semantics: if start == end, we clean everything.
 833  *
 834  *      The object must be locked.
 835  *
 836  *      Returns FALSE if some page from the range was not written, as
 837  *      reported by the pager, and TRUE otherwise.
 838  */
 839 boolean_t
 840 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
 841     int flags)
 842 {
 843         vm_page_t np, p;
 844         vm_pindex_t pi, tend, tstart;
 845         int curgeneration, n, pagerflags;
 846         boolean_t clearobjflags, eio, res;
 847
 848         VM_OBJECT_ASSERT_WLOCKED(object);
 849
 850         /*
 851          * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
 852          * objects.  The check below prevents the function from
 853          * operating on non-vnode objects.
 854          */
 855         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
 856             object->resident_page_count == 0)
 857                 return (TRUE);
 858
 859         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
 860             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
 861         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
 862
 863         tstart = OFF_TO_IDX(start);
 864         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
 865         clearobjflags = tstart == 0 && tend >= object->size;
 866         res = TRUE;
 867
 868 rescan:
 869         curgeneration = object->generation;
 870
 871         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
 872                 pi = p->pindex;
 873                 if (pi >= tend)
 874                         break;
 875                 np = TAILQ_NEXT(p, listq);
 876                 if (p->valid == 0)
 877                         continue;
 878                 if (vm_page_sleep_if_busy(p, "vpcwai")) {
 879                         if (object->generation != curgeneration) {
 880                                 if ((flags & OBJPC_SYNC) != 0)
 881                                         goto rescan;
 882                                 else
 883                                         clearobjflags = FALSE;
 884                         }
 885                         np = vm_page_find_least(object, pi);
 886                         continue;
 887                 }
 888                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
 889                         continue;
 890
 891                 n = vm_object_page_collect_flush(object, p, pagerflags,
 892                     flags, &clearobjflags, &eio);
 893                 if (eio) {
 894                         res = FALSE;
 895                         clearobjflags = FALSE;
 896                 }
 897                 if (object->generation != curgeneration) {
 898                         if ((flags & OBJPC_SYNC) != 0)
 899                                 goto rescan;
 900                         else
 901                                 clearobjflags = FALSE;
 902                 }
 903
 904                 /*
 905                  * If the VOP_PUTPAGES() did a truncated write, so
 906                  * that even the first page of the run is not fully
 907                  * written, vm_pageout_flush() returns 0 as the run
 908                  * length.  Since the condition that caused truncated
 909                  * write may be permanent, e.g. exhausted free space,
 910                  * accepting n == 0 would cause an infinite loop.
 911                  *
 912                  * Forwarding the iterator leaves the unwritten page
 913                  * behind, but there is not much we can do there if
 914                  * filesystem refuses to write it.
 915                  */
 916                 if (n == 0) {
 917                         n = 1;
 918                         clearobjflags = FALSE;
 919                 }
 920                 np = vm_page_find_least(object, pi + n);
 921         }
 922 #if 0
 923         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
 924 #endif
 925
 926         if (clearobjflags)
 927                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
 928         return (res);
 929 }
 930
 931 static int
 932 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
 933     int flags, boolean_t *clearobjflags, boolean_t *eio)
 934 {
 935         vm_page_t ma[vm_pageout_page_count], p_first, tp;
 936         int count, i, mreq, runlen;
 937
 938         vm_page_lock_assert(p, MA_NOTOWNED);
 939         VM_OBJECT_ASSERT_WLOCKED(object);
 940
 941         count = 1;
 942         mreq = 0;
 943
 944         for (tp = p; count < vm_pageout_page_count; count++) {
 945                 tp = vm_page_next(tp);
 946                 if (tp == NULL || vm_page_busied(tp))
 947                         break;
 948                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 949                         break;
 950         }
 951
 952         for (p_first = p; count < vm_pageout_page_count; count++) {
 953                 tp = vm_page_prev(p_first);
 954                 if (tp == NULL || vm_page_busied(tp))
 955                         break;
 956                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
 957                         break;
 958                 p_first = tp;
 959                 mreq++;
 960         }
 961
 962         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
 963                 ma[i] = tp;
 964
 965         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
 966         return (runlen);
 967 }
 968
 969 /*
 970  * Note that there is absolutely no sense in writing out
 971  * anonymous objects, so we track down the vnode object
 972  * to write out.
 973  * We invalidate (remove) all pages from the address space
 974  * for semantic correctness.
 975  *
 976  * If the backing object is a device object with unmanaged pages, then any
 977  * mappings to the specified range of pages must be removed before this
 978  * function is called.
 979  *
 980  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
 981  * may start out with a NULL object.
 982  */
 983 boolean_t
 984 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
 985     boolean_t syncio, boolean_t invalidate)
 986 {
 987         vm_object_t backing_object;
 988         struct vnode *vp;
 989         struct mount *mp;
 990         int error, flags, fsync_after;
 991         boolean_t res;
 992
 993         if (object == NULL)
 994                 return (TRUE);
 995         res = TRUE;
 996         error = 0;
 997         VM_OBJECT_WLOCK(object);
 998         while ((backing_object = object->backing_object) != NULL) {
 999                 VM_OBJECT_WLOCK(backing_object);
1000                 offset += object->backing_object_offset;
1001                 VM_OBJECT_WUNLOCK(object);
1002                 object = backing_object;
1003                 if (object->size < OFF_TO_IDX(offset + size))
1004                         size = IDX_TO_OFF(object->size) - offset;
1005         }
1006         /*
1007          * Flush pages if writing is allowed, invalidate them
1008          * if invalidation requested.  Pages undergoing I/O
1009          * will be ignored by vm_object_page_remove().
1010          *
1011          * We cannot lock the vnode and then wait for paging
1012          * to complete without deadlocking against vm_fault.
1013          * Instead we simply call vm_object_page_remove() and
1014          * allow it to block internally on a page-by-page
1015          * basis when it encounters pages undergoing async
1016          * I/O.
1017          */
1018         if (object->type == OBJT_VNODE &&
1019             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1020                 vp = object->handle;
1021                 VM_OBJECT_WUNLOCK(object);
1022                 (void) vn_start_write(vp, &mp, V_WAIT);
1023                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1024                 if (syncio && !invalidate && offset == 0 &&
1025                     OFF_TO_IDX(size) == object->size) {
1026                         /*
1027                          * If syncing the whole mapping of the file,
1028                          * it is faster to schedule all the writes in
1029                          * async mode, also allowing the clustering,
1030                          * and then wait for i/o to complete.
1031                          */
1032                         flags = 0;
1033                         fsync_after = TRUE;
1034                 } else {
1035                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1036                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1037                         fsync_after = FALSE;
1038                 }
1039                 VM_OBJECT_WLOCK(object);
1040                 res = vm_object_page_clean(object, offset, offset + size,
1041                     flags);
1042                 VM_OBJECT_WUNLOCK(object);
1043                 if (fsync_after)
1044                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1045                 VOP_UNLOCK(vp, 0);
1046                 vn_finished_write(mp);
1047                 if (error != 0)
1048                         res = FALSE;
1049                 VM_OBJECT_WLOCK(object);
1050         }
1051         if ((object->type == OBJT_VNODE ||
1052              object->type == OBJT_DEVICE) && invalidate) {
1053                 if (object->type == OBJT_DEVICE)
1054                         /*
1055                          * The option OBJPR_NOTMAPPED must be passed here
1056                          * because vm_object_page_remove() cannot remove
1057                          * unmanaged mappings.
1058                          */
1059                         flags = OBJPR_NOTMAPPED;
1060                 else if (old_msync)
1061                         flags = OBJPR_NOTWIRED;
1062                 else
1063                         flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1064                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1065                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1066         }
1067         VM_OBJECT_WUNLOCK(object);
1068         return (res);
1069 }
1070
1071 /*
1072  *      vm_object_madvise:
1073  *
1074  *      Implements the madvise function at the object/page level.
1075  *
1076  *      MADV_WILLNEED   (any object)
1077  *
1078  *          Activate the specified pages if they are resident.
1079  *
1080  *      MADV_DONTNEED   (any object)
1081  *
1082  *          Deactivate the specified pages if they are resident.
1083  *
1084  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1085  *                       OBJ_ONEMAPPING only)
1086  *
1087  *          Deactivate and clean the specified pages if they are
1088  *          resident.  This permits the process to reuse the pages
1089  *          without faulting or the kernel to reclaim the pages
1090  *          without I/O.
1091  */
1092 void
1093 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1094     int advise)
1095 {
1096         vm_pindex_t tpindex;
1097         vm_object_t backing_object, tobject;
1098         vm_page_t m;
1099
1100         if (object == NULL)
1101                 return;
1102         VM_OBJECT_WLOCK(object);
1103         /*
1104          * Locate and adjust resident pages
1105          */
1106         for (; pindex < end; pindex += 1) {
1107 relookup:
1108                 tobject = object;
1109                 tpindex = pindex;
1110 shadowlookup:
1111                 /*
1112                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1113                  * and those pages must be OBJ_ONEMAPPING.
1114                  */
1115                 if (advise == MADV_FREE) {
1116                         if ((tobject->type != OBJT_DEFAULT &&
1117                              tobject->type != OBJT_SWAP) ||
1118                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1119                                 goto unlock_tobject;
1120                         }
1121                 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1122                         goto unlock_tobject;
1123                 m = vm_page_lookup(tobject, tpindex);
1124                 if (m == NULL && advise == MADV_WILLNEED) {
1125                         /*
1126                          * If the page is cached, reactivate it.
1127                          */
1128                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1129                             VM_ALLOC_NOBUSY);
1130                 }
1131                 if (m == NULL) {
1132                         /*
1133                          * There may be swap even if there is no backing page
1134                          */
1135                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1136                                 swap_pager_freespace(tobject, tpindex, 1);
1137                         /*
1138                          * next object
1139                          */
1140                         backing_object = tobject->backing_object;
1141                         if (backing_object == NULL)
1142                                 goto unlock_tobject;
1143                         VM_OBJECT_WLOCK(backing_object);
1144                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1145                         if (tobject != object)
1146                                 VM_OBJECT_WUNLOCK(tobject);
1147                         tobject = backing_object;
1148                         goto shadowlookup;
1149                 } else if (m->valid != VM_PAGE_BITS_ALL)
1150                         goto unlock_tobject;
1151                 /*
1152                  * If the page is not in a normal state, skip it.
1153                  */
1154                 vm_page_lock(m);
1155                 if (m->hold_count != 0 || m->wire_count != 0) {
1156                         vm_page_unlock(m);
1157                         goto unlock_tobject;
1158                 }
1159                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1160                     ("vm_object_madvise: page %p is fictitious", m));
1161                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1162                     ("vm_object_madvise: page %p is not managed", m));
1163                 if (vm_page_busied(m)) {
1164                         if (advise == MADV_WILLNEED) {
1165                                 /*
1166                                  * Reference the page before unlocking and
1167                                  * sleeping so that the page daemon is less
1168                                  * likely to reclaim it.
1169                                  */
1170                                 vm_page_aflag_set(m, PGA_REFERENCED);
1171                         }
1172                         if (object != tobject)
1173                                 VM_OBJECT_WUNLOCK(object);
1174                         VM_OBJECT_WUNLOCK(tobject);
1175                         vm_page_busy_sleep(m, "madvpo");
1176                         VM_OBJECT_WLOCK(object);
1177                         goto relookup;
1178                 }
1179                 if (advise == MADV_WILLNEED) {
1180                         vm_page_activate(m);
1181                 } else {
1182                         vm_page_advise(m, advise);
1183                 }
1184                 vm_page_unlock(m);
1185                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1186                         swap_pager_freespace(tobject, tpindex, 1);
1187 unlock_tobject:
1188                 if (tobject != object)
1189                         VM_OBJECT_WUNLOCK(tobject);
1190         }
1191         VM_OBJECT_WUNLOCK(object);
1192 }
1193
1194 /*
1195  *      vm_object_shadow:
1196  *
1197  *      Create a new object which is backed by the
1198  *      specified existing object range.  The source
1199  *      object reference is deallocated.
1200  *
1201  *      The new object and offset into that object
1202  *      are returned in the source parameters.
1203  */
1204 void
1205 vm_object_shadow(
1206         vm_object_t *object,    /* IN/OUT */
1207         vm_ooffset_t *offset,   /* IN/OUT */
1208         vm_size_t length)
1209 {
1210         vm_object_t source;
1211         vm_object_t result;
1212
1213         source = *object;
1214
1215         /*
1216          * Don't create the new object if the old object isn't shared.
1217          */
1218         if (source != NULL) {
1219                 VM_OBJECT_WLOCK(source);
1220                 if (source->ref_count == 1 &&
1221                     source->handle == NULL &&
1222                     (source->type == OBJT_DEFAULT ||
1223                      source->type == OBJT_SWAP)) {
1224                         VM_OBJECT_WUNLOCK(source);
1225                         return;
1226                 }
1227                 VM_OBJECT_WUNLOCK(source);
1228         }
1229
1230         /*
1231          * Allocate a new object with the given length.
1232          */
1233         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1234
1235         /*
1236          * The new object shadows the source object, adding a reference to it.
1237          * Our caller changes his reference to point to the new object,
1238          * removing a reference to the source object.  Net result: no change
1239          * of reference count.
1240          *
1241          * Try to optimize the result object's page color when shadowing
1242          * in order to maintain page coloring consistency in the combined
1243          * shadowed object.
1244          */
1245         result->backing_object = source;
1246         /*
1247          * Store the offset into the source object, and fix up the offset into
1248          * the new object.
1249          */
1250         result->backing_object_offset = *offset;
1251         if (source != NULL) {
1252                 VM_OBJECT_WLOCK(source);
1253                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1254                 source->shadow_count++;
1255 #if VM_NRESERVLEVEL > 0
1256                 result->flags |= source->flags & OBJ_COLORED;
1257                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1258                     ((1 << (VM_NFREEORDER - 1)) - 1);
1259 #endif
1260                 VM_OBJECT_WUNLOCK(source);
1261         }
1262
1263
1264         /*
1265          * Return the new things
1266          */
1267         *offset = 0;
1268         *object = result;
1269 }
1270
1271 /*
1272  *      vm_object_split:
1273  *
1274  * Split the pages in a map entry into a new object.  This affords
1275  * easier removal of unused pages, and keeps object inheritance from
1276  * being a negative impact on memory usage.
1277  */
1278 void
1279 vm_object_split(vm_map_entry_t entry)
1280 {
1281         vm_page_t m, m_next;
1282         vm_object_t orig_object, new_object, source;
1283         vm_pindex_t idx, offidxstart;
1284         vm_size_t size;
1285
1286         orig_object = entry->object.vm_object;
1287         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1288                 return;
1289         if (orig_object->ref_count <= 1)
1290                 return;
1291         VM_OBJECT_WUNLOCK(orig_object);
1292
1293         offidxstart = OFF_TO_IDX(entry->offset);
1294         size = atop(entry->end - entry->start);
1295
1296         /*
1297          * If swap_pager_copy() is later called, it will convert new_object
1298          * into a swap object.
1299          */
1300         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1301
1302         /*
1303          * At this point, the new object is still private, so the order in
1304          * which the original and new objects are locked does not matter.
1305          */
1306         VM_OBJECT_WLOCK(new_object);
1307         VM_OBJECT_WLOCK(orig_object);
1308         source = orig_object->backing_object;
1309         if (source != NULL) {
1310                 VM_OBJECT_WLOCK(source);
1311                 if ((source->flags & OBJ_DEAD) != 0) {
1312                         VM_OBJECT_WUNLOCK(source);
1313                         VM_OBJECT_WUNLOCK(orig_object);
1314                         VM_OBJECT_WUNLOCK(new_object);
1315                         vm_object_deallocate(new_object);
1316                         VM_OBJECT_WLOCK(orig_object);
1317                         return;
1318                 }
1319                 LIST_INSERT_HEAD(&source->shadow_head,
1320                                   new_object, shadow_list);
1321                 source->shadow_count++;
1322                 vm_object_reference_locked(source);     /* for new_object */
1323                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1324                 VM_OBJECT_WUNLOCK(source);
1325                 new_object->backing_object_offset =
1326                         orig_object->backing_object_offset + entry->offset;
1327                 new_object->backing_object = source;
1328         }
1329         if (orig_object->cred != NULL) {
1330                 new_object->cred = orig_object->cred;
1331                 crhold(orig_object->cred);
1332                 new_object->charge = ptoa(size);
1333                 KASSERT(orig_object->charge >= ptoa(size),
1334                     ("orig_object->charge < 0"));
1335                 orig_object->charge -= ptoa(size);
1336         }
1337 retry:
1338         m = vm_page_find_least(orig_object, offidxstart);
1339         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1340             m = m_next) {
1341                 m_next = TAILQ_NEXT(m, listq);
1342
1343                 /*
1344                  * We must wait for pending I/O to complete before we can
1345                  * rename the page.
1346                  *
1347                  * We do not have to VM_PROT_NONE the page as mappings should
1348                  * not be changed by this operation.
1349                  */
1350                 if (vm_page_busied(m)) {
1351                         VM_OBJECT_WUNLOCK(new_object);
1352                         vm_page_lock(m);
1353                         VM_OBJECT_WUNLOCK(orig_object);
1354                         vm_page_busy_sleep(m, "spltwt");
1355                         VM_OBJECT_WLOCK(orig_object);
1356                         VM_OBJECT_WLOCK(new_object);
1357                         goto retry;
1358                 }
1359
1360                 /* vm_page_rename() will handle dirty and cache. */
1361                 if (vm_page_rename(m, new_object, idx)) {
1362                         VM_OBJECT_WUNLOCK(new_object);
1363                         VM_OBJECT_WUNLOCK(orig_object);
1364                         VM_WAIT;
1365                         VM_OBJECT_WLOCK(orig_object);
1366                         VM_OBJECT_WLOCK(new_object);
1367                         goto retry;
1368                 }
1369 #if VM_NRESERVLEVEL > 0
1370                 /*
1371                  * If some of the reservation's allocated pages remain with
1372                  * the original object, then transferring the reservation to
1373                  * the new object is neither particularly beneficial nor
1374                  * particularly harmful as compared to leaving the reservation
1375                  * with the original object.  If, however, all of the
1376                  * reservation's allocated pages are transferred to the new
1377                  * object, then transferring the reservation is typically
1378                  * beneficial.  Determining which of these two cases applies
1379                  * would be more costly than unconditionally renaming the
1380                  * reservation.
1381                  */
1382                 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1383 #endif
1384                 if (orig_object->type == OBJT_SWAP)
1385                         vm_page_xbusy(m);
1386         }
1387         if (orig_object->type == OBJT_SWAP) {
1388                 /*
1389                  * swap_pager_copy() can sleep, in which case the orig_object's
1390                  * and new_object's locks are released and reacquired.
1391                  */
1392                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1393                 TAILQ_FOREACH(m, &new_object->memq, listq)
1394                         vm_page_xunbusy(m);
1395
1396                 /*
1397                  * Transfer any cached pages from orig_object to new_object.
1398                  * If swap_pager_copy() found swapped out pages within the
1399                  * specified range of orig_object, then it changed
1400                  * new_object's type to OBJT_SWAP when it transferred those
1401                  * pages to new_object.  Otherwise, new_object's type
1402                  * should still be OBJT_DEFAULT and orig_object should not
1403                  * contain any cached pages within the specified range.
1404                  */
1405                 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1406                         vm_page_cache_transfer(orig_object, offidxstart,
1407                             new_object);
1408         }
1409         VM_OBJECT_WUNLOCK(orig_object);
1410         VM_OBJECT_WUNLOCK(new_object);
1411         entry->object.vm_object = new_object;
1412         entry->offset = 0LL;
1413         vm_object_deallocate(orig_object);
1414         VM_OBJECT_WLOCK(new_object);
1415 }
1416
1417 #define OBSC_TEST_ALL_SHADOWED  0x0001
1418 #define OBSC_COLLAPSE_NOWAIT    0x0002
1419 #define OBSC_COLLAPSE_WAIT      0x0004
1420
1421 static int
1422 vm_object_backing_scan(vm_object_t object, int op)
1423 {
1424         int r = 1;
1425         vm_page_t p;
1426         vm_object_t backing_object;
1427         vm_pindex_t backing_offset_index;
1428
1429         VM_OBJECT_ASSERT_WLOCKED(object);
1430         VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1431
1432         backing_object = object->backing_object;
1433         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1434
1435         /*
1436          * Initial conditions
1437          */
1438         if (op & OBSC_TEST_ALL_SHADOWED) {
1439                 /*
1440                  * We do not want to have to test for the existence of cache
1441                  * or swap pages in the backing object.  XXX but with the
1442                  * new swapper this would be pretty easy to do.
1443                  *
1444                  * XXX what about anonymous MAP_SHARED memory that hasn't
1445                  * been ZFOD faulted yet?  If we do not test for this, the
1446                  * shadow test may succeed! XXX
1447                  */
1448                 if (backing_object->type != OBJT_DEFAULT) {
1449                         return (0);
1450                 }
1451         }
1452         if (op & OBSC_COLLAPSE_WAIT) {
1453                 vm_object_set_flag(backing_object, OBJ_DEAD);
1454         }
1455
1456         /*
1457          * Our scan
1458          */
1459         p = TAILQ_FIRST(&backing_object->memq);
1460         while (p) {
1461                 vm_page_t next = TAILQ_NEXT(p, listq);
1462                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1463
1464                 if (op & OBSC_TEST_ALL_SHADOWED) {
1465                         vm_page_t pp;
1466
1467                         /*
1468                          * Ignore pages outside the parent object's range
1469                          * and outside the parent object's mapping of the
1470                          * backing object.
1471                          *
1472                          * note that we do not busy the backing object's
1473                          * page.
1474                          */
1475                         if (
1476                             p->pindex < backing_offset_index ||
1477                             new_pindex >= object->size
1478                         ) {
1479                                 p = next;
1480                                 continue;
1481                         }
1482
1483                         /*
1484                          * See if the parent has the page or if the parent's
1485                          * object pager has the page.  If the parent has the
1486                          * page but the page is not valid, the parent's
1487                          * object pager must have the page.
1488                          *
1489                          * If this fails, the parent does not completely shadow
1490                          * the object and we might as well give up now.
1491                          */
1492
1493                         pp = vm_page_lookup(object, new_pindex);
1494                         if (
1495                             (pp == NULL || pp->valid == 0) &&
1496                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1497                         ) {
1498                                 r = 0;
1499                                 break;
1500                         }
1501                 }
1502
1503                 /*
1504                  * Check for busy page
1505                  */
1506                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1507                         vm_page_t pp;
1508
1509                         if (op & OBSC_COLLAPSE_NOWAIT) {
1510                                 if (!p->valid || vm_page_busied(p)) {
1511                                         p = next;
1512                                         continue;
1513                                 }
1514                         } else if (op & OBSC_COLLAPSE_WAIT) {
1515                                 if (vm_page_busied(p)) {
1516                                         VM_OBJECT_WUNLOCK(object);
1517                                         vm_page_lock(p);
1518                                         VM_OBJECT_WUNLOCK(backing_object);
1519                                         vm_page_busy_sleep(p, "vmocol");
1520                                         VM_OBJECT_WLOCK(object);
1521                                         VM_OBJECT_WLOCK(backing_object);
1522                                         /*
1523                                          * If we slept, anything could have
1524                                          * happened.  Since the object is
1525                                          * marked dead, the backing offset
1526                                          * should not have changed so we
1527                                          * just restart our scan.
1528                                          */
1529                                         p = TAILQ_FIRST(&backing_object->memq);
1530                                         continue;
1531                                 }
1532                         }
1533
1534                         KASSERT(
1535                             p->object == backing_object,
1536                             ("vm_object_backing_scan: object mismatch")
1537                         );
1538
1539                         if (
1540                             p->pindex < backing_offset_index ||
1541                             new_pindex >= object->size
1542                         ) {
1543                                 if (backing_object->type == OBJT_SWAP)
1544                                         swap_pager_freespace(backing_object,
1545                                             p->pindex, 1);
1546
1547                                 /*
1548                                  * Page is out of the parent object's range, we
1549                                  * can simply destroy it.
1550                                  */
1551                                 vm_page_lock(p);
1552                                 KASSERT(!pmap_page_is_mapped(p),
1553                                     ("freeing mapped page %p", p));
1554                                 if (p->wire_count == 0)
1555                                         vm_page_free(p);
1556                                 else
1557                                         vm_page_remove(p);
1558                                 vm_page_unlock(p);
1559                                 p = next;
1560                                 continue;
1561                         }
1562
1563                         pp = vm_page_lookup(object, new_pindex);
1564                         if (
1565                             (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1566                             (pp != NULL && pp->valid == 0)
1567                         ) {
1568                                 if (backing_object->type == OBJT_SWAP)
1569                                         swap_pager_freespace(backing_object,
1570                                             p->pindex, 1);
1571
1572                                 /*
1573                                  * The page in the parent is not (yet) valid.
1574                                  * We don't know anything about the state of
1575                                  * the original page.  It might be mapped,
1576                                  * so we must avoid the next if here.
1577                                  *
1578                                  * This is due to a race in vm_fault() where
1579                                  * we must unbusy the original (backing_obj)
1580                                  * page before we can (re)lock the parent.
1581                                  * Hence we can get here.
1582                                  */
1583                                 p = next;
1584                                 continue;
1585                         }
1586                         if (
1587                             pp != NULL ||
1588                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1589                         ) {
1590                                 if (backing_object->type == OBJT_SWAP)
1591                                         swap_pager_freespace(backing_object,
1592                                             p->pindex, 1);
1593
1594                                 /*
1595                                  * page already exists in parent OR swap exists
1596                                  * for this location in the parent.  Destroy
1597                                  * the original page from the backing object.
1598                                  *
1599                                  * Leave the parent's page alone
1600                                  */
1601                                 vm_page_lock(p);
1602                                 KASSERT(!pmap_page_is_mapped(p),
1603                                     ("freeing mapped page %p", p));
1604                                 if (p->wire_count == 0)
1605                                         vm_page_free(p);
1606                                 else
1607                                         vm_page_remove(p);
1608                                 vm_page_unlock(p);
1609                                 p = next;
1610                                 continue;
1611                         }
1612
1613                         /*
1614                          * Page does not exist in parent, rename the
1615                          * page from the backing object to the main object.
1616                          *
1617                          * If the page was mapped to a process, it can remain
1618                          * mapped through the rename.
1619                          * vm_page_rename() will handle dirty and cache.
1620                          */
1621                         if (vm_page_rename(p, object, new_pindex)) {
1622                                 if (op & OBSC_COLLAPSE_NOWAIT) {
1623                                         p = next;
1624                                         continue;
1625                                 }
1626                                 VM_OBJECT_WLOCK(backing_object);
1627                                 VM_OBJECT_WUNLOCK(object);
1628                                 VM_WAIT;
1629                                 VM_OBJECT_WLOCK(object);
1630                                 VM_OBJECT_WLOCK(backing_object);
1631                                 p = TAILQ_FIRST(&backing_object->memq);
1632                                 continue;
1633                         }
1634
1635                         /* Use the old pindex to free the right page. */
1636                         if (backing_object->type == OBJT_SWAP)
1637                                 swap_pager_freespace(backing_object,
1638                                     new_pindex + backing_offset_index, 1);
1639
1640 #if VM_NRESERVLEVEL > 0
1641                         /*
1642                          * Rename the reservation.
1643                          */
1644                         vm_reserv_rename(p, object, backing_object,
1645                             backing_offset_index);
1646 #endif
1647                 }
1648                 p = next;
1649         }
1650         return (r);
1651 }
1652
1653
1654 /*
1655  * this version of collapse allows the operation to occur earlier and
1656  * when paging_in_progress is true for an object...  This is not a complete
1657  * operation, but should plug 99.9% of the rest of the leaks.
1658  */
1659 static void
1660 vm_object_qcollapse(vm_object_t object)
1661 {
1662         vm_object_t backing_object = object->backing_object;
1663
1664         VM_OBJECT_ASSERT_WLOCKED(object);
1665         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1666
1667         if (backing_object->ref_count != 1)
1668                 return;
1669
1670         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1671 }
1672
1673 /*
1674  *      vm_object_collapse:
1675  *
1676  *      Collapse an object with the object backing it.
1677  *      Pages in the backing object are moved into the
1678  *      parent, and the backing object is deallocated.
1679  */
1680 void
1681 vm_object_collapse(vm_object_t object)
1682 {
1683         VM_OBJECT_ASSERT_WLOCKED(object);
1684
1685         while (TRUE) {
1686                 vm_object_t backing_object;
1687
1688                 /*
1689                  * Verify that the conditions are right for collapse:
1690                  *
1691                  * The object exists and the backing object exists.
1692                  */
1693                 if ((backing_object = object->backing_object) == NULL)
1694                         break;
1695
1696                 /*
1697                  * we check the backing object first, because it is most likely
1698                  * not collapsable.
1699                  */
1700                 VM_OBJECT_WLOCK(backing_object);
1701                 if (backing_object->handle != NULL ||
1702                     (backing_object->type != OBJT_DEFAULT &&
1703                      backing_object->type != OBJT_SWAP) ||
1704                     (backing_object->flags & OBJ_DEAD) ||
1705                     object->handle != NULL ||
1706                     (object->type != OBJT_DEFAULT &&
1707                      object->type != OBJT_SWAP) ||
1708                     (object->flags & OBJ_DEAD)) {
1709                         VM_OBJECT_WUNLOCK(backing_object);
1710                         break;
1711                 }
1712
1713                 if (
1714                     object->paging_in_progress != 0 ||
1715                     backing_object->paging_in_progress != 0
1716                 ) {
1717                         vm_object_qcollapse(object);
1718                         VM_OBJECT_WUNLOCK(backing_object);
1719                         break;
1720                 }
1721                 /*
1722                  * We know that we can either collapse the backing object (if
1723                  * the parent is the only reference to it) or (perhaps) have
1724                  * the parent bypass the object if the parent happens to shadow
1725                  * all the resident pages in the entire backing object.
1726                  *
1727                  * This is ignoring pager-backed pages such as swap pages.
1728                  * vm_object_backing_scan fails the shadowing test in this
1729                  * case.
1730                  */
1731                 if (backing_object->ref_count == 1) {
1732                         /*
1733                          * If there is exactly one reference to the backing
1734                          * object, we can collapse it into the parent.
1735                          */
1736                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1737
1738 #if VM_NRESERVLEVEL > 0
1739                         /*
1740                          * Break any reservations from backing_object.
1741                          */
1742                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1743                                 vm_reserv_break_all(backing_object);
1744 #endif
1745
1746                         /*
1747                          * Move the pager from backing_object to object.
1748                          */
1749                         if (backing_object->type == OBJT_SWAP) {
1750                                 /*
1751                                  * swap_pager_copy() can sleep, in which case
1752                                  * the backing_object's and object's locks are
1753                                  * released and reacquired.
1754                                  * Since swap_pager_copy() is being asked to
1755                                  * destroy the source, it will change the
1756                                  * backing_object's type to OBJT_DEFAULT.
1757                                  */
1758                                 swap_pager_copy(
1759                                     backing_object,
1760                                     object,
1761                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1762
1763                                 /*
1764                                  * Free any cached pages from backing_object.
1765                                  */
1766                                 if (__predict_false(
1767                                     !vm_object_cache_is_empty(backing_object)))
1768                                         vm_page_cache_free(backing_object, 0, 0);
1769                         }
1770                         /*
1771                          * Object now shadows whatever backing_object did.
1772                          * Note that the reference to
1773                          * backing_object->backing_object moves from within
1774                          * backing_object to within object.
1775                          */
1776                         LIST_REMOVE(object, shadow_list);
1777                         backing_object->shadow_count--;
1778                         if (backing_object->backing_object) {
1779                                 VM_OBJECT_WLOCK(backing_object->backing_object);
1780                                 LIST_REMOVE(backing_object, shadow_list);
1781                                 LIST_INSERT_HEAD(
1782                                     &backing_object->backing_object->shadow_head,
1783                                     object, shadow_list);
1784                                 /*
1785                                  * The shadow_count has not changed.
1786                                  */
1787                                 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1788                         }
1789                         object->backing_object = backing_object->backing_object;
1790                         object->backing_object_offset +=
1791                             backing_object->backing_object_offset;
1792
1793                         /*
1794                          * Discard backing_object.
1795                          *
1796                          * Since the backing object has no pages, no pager left,
1797                          * and no object references within it, all that is
1798                          * necessary is to dispose of it.
1799                          */
1800                         KASSERT(backing_object->ref_count == 1, (
1801 "backing_object %p was somehow re-referenced during collapse!",
1802                             backing_object));
1803                         VM_OBJECT_WUNLOCK(backing_object);
1804                         vm_object_destroy(backing_object);
1805
1806                         object_collapses++;
1807                 } else {
1808                         vm_object_t new_backing_object;
1809
1810                         /*
1811                          * If we do not entirely shadow the backing object,
1812                          * there is nothing we can do so we give up.
1813                          */
1814                         if (object->resident_page_count != object->size &&
1815                             vm_object_backing_scan(object,
1816                             OBSC_TEST_ALL_SHADOWED) == 0) {
1817                                 VM_OBJECT_WUNLOCK(backing_object);
1818                                 break;
1819                         }
1820
1821                         /*
1822                          * Make the parent shadow the next object in the
1823                          * chain.  Deallocating backing_object will not remove
1824                          * it, since its reference count is at least 2.
1825                          */
1826                         LIST_REMOVE(object, shadow_list);
1827                         backing_object->shadow_count--;
1828
1829                         new_backing_object = backing_object->backing_object;
1830                         if ((object->backing_object = new_backing_object) != NULL) {
1831                                 VM_OBJECT_WLOCK(new_backing_object);
1832                                 LIST_INSERT_HEAD(
1833                                     &new_backing_object->shadow_head,
1834                                     object,
1835                                     shadow_list
1836                                 );
1837                                 new_backing_object->shadow_count++;
1838                                 vm_object_reference_locked(new_backing_object);
1839                                 VM_OBJECT_WUNLOCK(new_backing_object);
1840                                 object->backing_object_offset +=
1841                                         backing_object->backing_object_offset;
1842                         }
1843
1844                         /*
1845                          * Drop the reference count on backing_object. Since
1846                          * its ref_count was at least 2, it will not vanish.
1847                          */
1848                         backing_object->ref_count--;
1849                         VM_OBJECT_WUNLOCK(backing_object);
1850                         object_bypasses++;
1851                 }
1852
1853                 /*
1854                  * Try again with this object's new backing object.
1855                  */
1856         }
1857 }
1858
1859 /*
1860  *      vm_object_page_remove:
1861  *
1862  *      For the given object, either frees or invalidates each of the
1863  *      specified pages.  In general, a page is freed.  However, if a page is
1864  *      wired for any reason other than the existence of a managed, wired
1865  *      mapping, then it may be invalidated but not removed from the object.
1866  *      Pages are specified by the given range ["start", "end") and the option
1867  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1868  *      extends from "start" to the end of the object.  If the option
1869  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1870  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1871  *      specified, then the pages within the specified range must have no
1872  *      mappings.  Otherwise, if this option is not specified, any mappings to
1873  *      the specified pages are removed before the pages are freed or
1874  *      invalidated.
1875  *
1876  *      In general, this operation should only be performed on objects that
1877  *      contain managed pages.  There are, however, two exceptions.  First, it
1878  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1879  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1880  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1881  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1882  *
1883  *      The object must be locked.
1884  */
1885 void
1886 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1887     int options)
1888 {
1889         vm_page_t p, next;
1890         int wirings;
1891
1892         VM_OBJECT_ASSERT_WLOCKED(object);
1893         KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1894             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1895             ("vm_object_page_remove: illegal options for object %p", object));
1896         if (object->resident_page_count == 0)
1897                 goto skipmemq;
1898         vm_object_pip_add(object, 1);
1899 again:
1900         p = vm_page_find_least(object, start);
1901
1902         /*
1903          * Here, the variable "p" is either (1) the page with the least pindex
1904          * greater than or equal to the parameter "start" or (2) NULL.
1905          */
1906         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1907                 next = TAILQ_NEXT(p, listq);
1908
1909                 /*
1910                  * If the page is wired for any reason besides the existence
1911                  * of managed, wired mappings, then it cannot be freed.  For
1912                  * example, fictitious pages, which represent device memory,
1913                  * are inherently wired and cannot be freed.  They can,
1914                  * however, be invalidated if the option OBJPR_CLEANONLY is
1915                  * not specified.
1916                  */
1917                 vm_page_lock(p);
1918                 if (vm_page_xbusied(p)) {
1919                         VM_OBJECT_WUNLOCK(object);
1920                         vm_page_busy_sleep(p, "vmopax");
1921                         VM_OBJECT_WLOCK(object);
1922                         goto again;
1923                 }
1924                 if ((wirings = p->wire_count) != 0 &&
1925                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1926                         if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1927                             0) {
1928                                 pmap_remove_all(p);
1929                                 /* Account for removal of wired mappings. */
1930                                 if (wirings != 0)
1931                                         p->wire_count -= wirings;
1932                         }
1933                         if ((options & OBJPR_CLEANONLY) == 0) {
1934                                 p->valid = 0;
1935                                 vm_page_undirty(p);
1936                         }
1937                         goto next;
1938                 }
1939                 if (vm_page_busied(p)) {
1940                         VM_OBJECT_WUNLOCK(object);
1941                         vm_page_busy_sleep(p, "vmopar");
1942                         VM_OBJECT_WLOCK(object);
1943                         goto again;
1944                 }
1945                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1946                     ("vm_object_page_remove: page %p is fictitious", p));
1947                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1948                         if ((options & OBJPR_NOTMAPPED) == 0)
1949                                 pmap_remove_write(p);
1950                         if (p->dirty)
1951                                 goto next;
1952                 }
1953                 if ((options & OBJPR_NOTMAPPED) == 0) {
1954                         if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1955                                 goto next;
1956                         pmap_remove_all(p);
1957                         /* Account for removal of wired mappings. */
1958                         if (wirings != 0) {
1959                                 KASSERT(p->wire_count == wirings,
1960                                     ("inconsistent wire count %d %d %p",
1961                                     p->wire_count, wirings, p));
1962                                 p->wire_count = 0;
1963                                 atomic_subtract_int(&cnt.v_wire_count, 1);
1964                         }
1965                 }
1966                 vm_page_free(p);
1967 next:
1968                 vm_page_unlock(p);
1969         }
1970         vm_object_pip_wakeup(object);
1971 skipmemq:
1972         if (__predict_false(!vm_object_cache_is_empty(object)))
1973                 vm_page_cache_free(object, start, end);
1974 }
1975
1976 /*
1977  *      vm_object_page_cache:
1978  *
1979  *      For the given object, attempt to move the specified clean
1980  *      pages to the cache queue.  If a page is wired for any reason,
1981  *      then it will not be changed.  Pages are specified by the given
1982  *      range ["start", "end").  As a special case, if "end" is zero,
1983  *      then the range extends from "start" to the end of the object.
1984  *      Any mappings to the specified pages are removed before the
1985  *      pages are moved to the cache queue.
1986  *
1987  *      This operation should only be performed on objects that
1988  *      contain non-fictitious, managed pages.
1989  *
1990  *      The object must be locked.
1991  */
1992 void
1993 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1994 {
1995         struct mtx *mtx, *new_mtx;
1996         vm_page_t p, next;
1997
1998         VM_OBJECT_ASSERT_WLOCKED(object);
1999         KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2000             ("vm_object_page_cache: illegal object %p", object));
2001         if (object->resident_page_count == 0)
2002                 return;
2003         p = vm_page_find_least(object, start);
2004
2005         /*
2006          * Here, the variable "p" is either (1) the page with the least pindex
2007          * greater than or equal to the parameter "start" or (2) NULL.
2008          */
2009         mtx = NULL;
2010         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2011                 next = TAILQ_NEXT(p, listq);
2012
2013                 /*
2014                  * Avoid releasing and reacquiring the same page lock.
2015                  */
2016                 new_mtx = vm_page_lockptr(p);
2017                 if (mtx != new_mtx) {
2018                         if (mtx != NULL)
2019                                 mtx_unlock(mtx);
2020                         mtx = new_mtx;
2021                         mtx_lock(mtx);
2022                 }
2023                 vm_page_try_to_cache(p);
2024         }
2025         if (mtx != NULL)
2026                 mtx_unlock(mtx);
2027 }
2028
2029 /*
2030  *      Populate the specified range of the object with valid pages.  Returns
2031  *      TRUE if the range is successfully populated and FALSE otherwise.
2032  *
2033  *      Note: This function should be optimized to pass a larger array of
2034  *      pages to vm_pager_get_pages() before it is applied to a non-
2035  *      OBJT_DEVICE object.
2036  *
2037  *      The object must be locked.
2038  */
2039 boolean_t
2040 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2041 {
2042         vm_page_t m, ma[1];
2043         vm_pindex_t pindex;
2044         int rv;
2045
2046         VM_OBJECT_ASSERT_WLOCKED(object);
2047         for (pindex = start; pindex < end; pindex++) {
2048                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2049                 if (m->valid != VM_PAGE_BITS_ALL) {
2050                         ma[0] = m;
2051                         rv = vm_pager_get_pages(object, ma, 1, 0);
2052                         m = vm_page_lookup(object, pindex);
2053                         if (m == NULL)
2054                                 break;
2055                         if (rv != VM_PAGER_OK) {
2056                                 vm_page_lock(m);
2057                                 vm_page_free(m);
2058                                 vm_page_unlock(m);
2059                                 break;
2060                         }
2061                 }
2062                 /*
2063                  * Keep "m" busy because a subsequent iteration may unlock
2064                  * the object.
2065                  */
2066         }
2067         if (pindex > start) {
2068                 m = vm_page_lookup(object, start);
2069                 while (m != NULL && m->pindex < pindex) {
2070                         vm_page_xunbusy(m);
2071                         m = TAILQ_NEXT(m, listq);
2072                 }
2073         }
2074         return (pindex == end);
2075 }
2076
2077 /*
2078  *      Routine:        vm_object_coalesce
2079  *      Function:       Coalesces two objects backing up adjoining
2080  *                      regions of memory into a single object.
2081  *
2082  *      returns TRUE if objects were combined.
2083  *
2084  *      NOTE:   Only works at the moment if the second object is NULL -
2085  *              if it's not, which object do we lock first?
2086  *
2087  *      Parameters:
2088  *              prev_object     First object to coalesce
2089  *              prev_offset     Offset into prev_object
2090  *              prev_size       Size of reference to prev_object
2091  *              next_size       Size of reference to the second object
2092  *              reserved        Indicator that extension region has
2093  *                              swap accounted for
2094  *
2095  *      Conditions:
2096  *      The object must *not* be locked.
2097  */
2098 boolean_t
2099 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2100     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2101 {
2102         vm_pindex_t next_pindex;
2103
2104         if (prev_object == NULL)
2105                 return (TRUE);
2106         VM_OBJECT_WLOCK(prev_object);
2107         if ((prev_object->type != OBJT_DEFAULT &&
2108             prev_object->type != OBJT_SWAP) ||
2109             (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2110                 VM_OBJECT_WUNLOCK(prev_object);
2111                 return (FALSE);
2112         }
2113
2114         /*
2115          * Try to collapse the object first
2116          */
2117         vm_object_collapse(prev_object);
2118
2119         /*
2120          * Can't coalesce if: . more than one reference . paged out . shadows
2121          * another object . has a copy elsewhere (any of which mean that the
2122          * pages not mapped to prev_entry may be in use anyway)
2123          */
2124         if (prev_object->backing_object != NULL) {
2125                 VM_OBJECT_WUNLOCK(prev_object);
2126                 return (FALSE);
2127         }
2128
2129         prev_size >>= PAGE_SHIFT;
2130         next_size >>= PAGE_SHIFT;
2131         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2132
2133         if ((prev_object->ref_count > 1) &&
2134             (prev_object->size != next_pindex)) {
2135                 VM_OBJECT_WUNLOCK(prev_object);
2136                 return (FALSE);
2137         }
2138
2139         /*
2140          * Account for the charge.
2141          */
2142         if (prev_object->cred != NULL) {
2143
2144                 /*
2145                  * If prev_object was charged, then this mapping,
2146                  * althought not charged now, may become writable
2147                  * later. Non-NULL cred in the object would prevent
2148                  * swap reservation during enabling of the write
2149                  * access, so reserve swap now. Failed reservation
2150                  * cause allocation of the separate object for the map
2151                  * entry, and swap reservation for this entry is
2152                  * managed in appropriate time.
2153                  */
2154                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2155                     prev_object->cred)) {
2156                         return (FALSE);
2157                 }
2158                 prev_object->charge += ptoa(next_size);
2159         }
2160
2161         /*
2162          * Remove any pages that may still be in the object from a previous
2163          * deallocation.
2164          */
2165         if (next_pindex < prev_object->size) {
2166                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2167                     next_size, 0);
2168                 if (prev_object->type == OBJT_SWAP)
2169                         swap_pager_freespace(prev_object,
2170                                              next_pindex, next_size);
2171 #if 0
2172                 if (prev_object->cred != NULL) {
2173                         KASSERT(prev_object->charge >=
2174                             ptoa(prev_object->size - next_pindex),
2175                             ("object %p overcharged 1 %jx %jx", prev_object,
2176                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2177                         prev_object->charge -= ptoa(prev_object->size -
2178                             next_pindex);
2179                 }
2180 #endif
2181         }
2182
2183         /*
2184          * Extend the object if necessary.
2185          */
2186         if (next_pindex + next_size > prev_object->size)
2187                 prev_object->size = next_pindex + next_size;
2188
2189         VM_OBJECT_WUNLOCK(prev_object);
2190         return (TRUE);
2191 }
2192
2193 void
2194 vm_object_set_writeable_dirty(vm_object_t object)
2195 {
2196
2197         VM_OBJECT_ASSERT_WLOCKED(object);
2198         if (object->type != OBJT_VNODE)
2199                 return;
2200         object->generation++;
2201         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2202                 return;
2203         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2204 }
2205
2206 /*
2207  *      vm_object_unwire:
2208  *
2209  *      For each page offset within the specified range of the given object,
2210  *      find the highest-level page in the shadow chain and unwire it.  A page
2211  *      must exist at every page offset, and the highest-level page must be
2212  *      wired.
2213  */
2214 void
2215 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2216     uint8_t queue)
2217 {
2218         vm_object_t tobject;
2219         vm_page_t m, tm;
2220         vm_pindex_t end_pindex, pindex, tpindex;
2221         int depth, locked_depth;
2222
2223         KASSERT((offset & PAGE_MASK) == 0,
2224             ("vm_object_unwire: offset is not page aligned"));
2225         KASSERT((length & PAGE_MASK) == 0,
2226             ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2227         /* The wired count of a fictitious page never changes. */
2228         if ((object->flags & OBJ_FICTITIOUS) != 0)
2229                 return;
2230         pindex = OFF_TO_IDX(offset);
2231         end_pindex = pindex + atop(length);
2232         locked_depth = 1;
2233         VM_OBJECT_RLOCK(object);
2234         m = vm_page_find_least(object, pindex);
2235         while (pindex < end_pindex) {
2236                 if (m == NULL || pindex < m->pindex) {
2237                         /*
2238                          * The first object in the shadow chain doesn't
2239                          * contain a page at the current index.  Therefore,
2240                          * the page must exist in a backing object.
2241                          */
2242                         tobject = object;
2243                         tpindex = pindex;
2244                         depth = 0;
2245                         do {
2246                                 tpindex +=
2247                                     OFF_TO_IDX(tobject->backing_object_offset);
2248                                 tobject = tobject->backing_object;
2249                                 KASSERT(tobject != NULL,
2250                                     ("vm_object_unwire: missing page"));
2251                                 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2252                                         goto next_page;
2253                                 depth++;
2254                                 if (depth == locked_depth) {
2255                                         locked_depth++;
2256                                         VM_OBJECT_RLOCK(tobject);
2257                                 }
2258                         } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2259                             NULL);
2260                 } else {
2261                         tm = m;
2262                         m = TAILQ_NEXT(m, listq);
2263                 }
2264                 vm_page_lock(tm);
2265                 vm_page_unwire(tm, queue);
2266                 vm_page_unlock(tm);
2267 next_page:
2268                 pindex++;
2269         }
2270         /* Release the accumulated object locks. */
2271         for (depth = 0; depth < locked_depth; depth++) {
2272                 tobject = object->backing_object;
2273                 VM_OBJECT_RUNLOCK(object);
2274                 object = tobject;
2275         }
2276 }
2277
2278 #include "opt_ddb.h"
2279 #ifdef DDB
2280 #include <sys/kernel.h>
2281
2282 #include <sys/cons.h>
2283
2284 #include <ddb/ddb.h>
2285
2286 static int
2287 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2288 {
2289         vm_map_t tmpm;
2290         vm_map_entry_t tmpe;
2291         vm_object_t obj;
2292         int entcount;
2293
2294         if (map == 0)
2295                 return 0;
2296
2297         if (entry == 0) {
2298                 tmpe = map->header.next;
2299                 entcount = map->nentries;
2300                 while (entcount-- && (tmpe != &map->header)) {
2301                         if (_vm_object_in_map(map, object, tmpe)) {
2302                                 return 1;
2303                         }
2304                         tmpe = tmpe->next;
2305                 }
2306         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2307                 tmpm = entry->object.sub_map;
2308                 tmpe = tmpm->header.next;
2309                 entcount = tmpm->nentries;
2310                 while (entcount-- && tmpe != &tmpm->header) {
2311                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2312                                 return 1;
2313                         }
2314                         tmpe = tmpe->next;
2315                 }
2316         } else if ((obj = entry->object.vm_object) != NULL) {
2317                 for (; obj; obj = obj->backing_object)
2318                         if (obj == object) {
2319                                 return 1;
2320                         }
2321         }
2322         return 0;
2323 }
2324
2325 static int
2326 vm_object_in_map(vm_object_t object)
2327 {
2328         struct proc *p;
2329
2330         /* sx_slock(&allproc_lock); */
2331         FOREACH_PROC_IN_SYSTEM(p) {
2332                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2333                         continue;
2334                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2335                         /* sx_sunlock(&allproc_lock); */
2336                         return 1;
2337                 }
2338         }
2339         /* sx_sunlock(&allproc_lock); */
2340         if (_vm_object_in_map(kernel_map, object, 0))
2341                 return 1;
2342         return 0;
2343 }
2344
2345 DB_SHOW_COMMAND(vmochk, vm_object_check)
2346 {
2347         vm_object_t object;
2348
2349         /*
2350          * make sure that internal objs are in a map somewhere
2351          * and none have zero ref counts.
2352          */
2353         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2354                 if (object->handle == NULL &&
2355                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2356                         if (object->ref_count == 0) {
2357                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2358                                         (long)object->size);
2359                         }
2360                         if (!vm_object_in_map(object)) {
2361                                 db_printf(
2362                         "vmochk: internal obj is not in a map: "
2363                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2364                                     object->ref_count, (u_long)object->size,
2365                                     (u_long)object->size,
2366                                     (void *)object->backing_object);
2367                         }
2368                 }
2369         }
2370 }
2371
2372 /*
2373  *      vm_object_print:        [ debug ]
2374  */
2375 DB_SHOW_COMMAND(object, vm_object_print_static)
2376 {
2377         /* XXX convert args. */
2378         vm_object_t object = (vm_object_t)addr;
2379         boolean_t full = have_addr;
2380
2381         vm_page_t p;
2382
2383         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2384 #define count   was_count
2385
2386         int count;
2387
2388         if (object == NULL)
2389                 return;
2390
2391         db_iprintf(
2392             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2393             object, (int)object->type, (uintmax_t)object->size,
2394             object->resident_page_count, object->ref_count, object->flags,
2395             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2396         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2397             object->shadow_count,
2398             object->backing_object ? object->backing_object->ref_count : 0,
2399             object->backing_object, (uintmax_t)object->backing_object_offset);
2400
2401         if (!full)
2402                 return;
2403
2404         db_indent += 2;
2405         count = 0;
2406         TAILQ_FOREACH(p, &object->memq, listq) {
2407                 if (count == 0)
2408                         db_iprintf("memory:=");
2409                 else if (count == 6) {
2410                         db_printf("\n");
2411                         db_iprintf(" ...");
2412                         count = 0;
2413                 } else
2414                         db_printf(",");
2415                 count++;
2416
2417                 db_printf("(off=0x%jx,page=0x%jx)",
2418                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2419         }
2420         if (count != 0)
2421                 db_printf("\n");
2422         db_indent -= 2;
2423 }
2424
2425 /* XXX. */
2426 #undef count
2427
2428 /* XXX need this non-static entry for calling from vm_map_print. */
2429 void
2430 vm_object_print(
2431         /* db_expr_t */ long addr,
2432         boolean_t have_addr,
2433         /* db_expr_t */ long count,
2434         char *modif)
2435 {
2436         vm_object_print_static(addr, have_addr, count, modif);
2437 }
2438
2439 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2440 {
2441         vm_object_t object;
2442         vm_pindex_t fidx;
2443         vm_paddr_t pa;
2444         vm_page_t m, prev_m;
2445         int rcount, nl, c;
2446
2447         nl = 0;
2448         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2449                 db_printf("new object: %p\n", (void *)object);
2450                 if (nl > 18) {
2451                         c = cngetc();
2452                         if (c != ' ')
2453                                 return;
2454                         nl = 0;
2455                 }
2456                 nl++;
2457                 rcount = 0;
2458                 fidx = 0;
2459                 pa = -1;
2460                 TAILQ_FOREACH(m, &object->memq, listq) {
2461                         if (m->pindex > 128)
2462                                 break;
2463                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2464                             prev_m->pindex + 1 != m->pindex) {
2465                                 if (rcount) {
2466                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2467                                                 (long)fidx, rcount, (long)pa);
2468                                         if (nl > 18) {
2469                                                 c = cngetc();
2470                                                 if (c != ' ')
2471                                                         return;
2472                                                 nl = 0;
2473                                         }
2474                                         nl++;
2475                                         rcount = 0;
2476                                 }
2477                         }
2478                         if (rcount &&
2479                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2480                                 ++rcount;
2481                                 continue;
2482                         }
2483                         if (rcount) {
2484                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2485                                         (long)fidx, rcount, (long)pa);
2486                                 if (nl > 18) {
2487                                         c = cngetc();
2488                                         if (c != ' ')
2489                                                 return;
2490                                         nl = 0;
2491                                 }
2492                                 nl++;
2493                         }
2494                         fidx = m->pindex;
2495                         pa = VM_PAGE_TO_PHYS(m);
2496                         rcount = 1;
2497                 }
2498                 if (rcount) {
2499                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2500                                 (long)fidx, rcount, (long)pa);
2501                         if (nl > 18) {
2502                                 c = cngetc();
2503                                 if (c != ' ')
2504                                         return;
2505                                 nl = 0;
2506                         }
2507                         nl++;
2508                 }
2509         }
2510 }
2511 #endif /* DDB */