sys/vm/vm_object.c

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