vm_object.c revision 275726
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: stable/10/sys/vm/vm_object.c 275726 2014-12-12 09:22:52Z kib $"); 67 68#include "opt_vm.h" 69 70#include <sys/param.h> 71#include <sys/systm.h> 72#include <sys/lock.h> 73#include <sys/mman.h> 74#include <sys/mount.h> 75#include <sys/kernel.h> 76#include <sys/sysctl.h> 77#include <sys/mutex.h> 78#include <sys/proc.h> /* for curproc, pageproc */ 79#include <sys/socket.h> 80#include <sys/resourcevar.h> 81#include <sys/rwlock.h> 82#include <sys/vnode.h> 83#include <sys/vmmeter.h> 84#include <sys/sx.h> 85 86#include <vm/vm.h> 87#include <vm/vm_param.h> 88#include <vm/pmap.h> 89#include <vm/vm_map.h> 90#include <vm/vm_object.h> 91#include <vm/vm_page.h> 92#include <vm/vm_pageout.h> 93#include <vm/vm_pager.h> 94#include <vm/swap_pager.h> 95#include <vm/vm_kern.h> 96#include <vm/vm_extern.h> 97#include <vm/vm_radix.h> 98#include <vm/vm_reserv.h> 99#include <vm/uma.h> 100 101static int old_msync; 102SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 103 "Use old (insecure) msync behavior"); 104 105static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 106 int pagerflags, int flags, boolean_t *clearobjflags, 107 boolean_t *eio); 108static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 109 boolean_t *clearobjflags); 110static void vm_object_qcollapse(vm_object_t object); 111static void vm_object_vndeallocate(vm_object_t object); 112 113/* 114 * Virtual memory objects maintain the actual data 115 * associated with allocated virtual memory. A given 116 * page of memory exists within exactly one object. 117 * 118 * An object is only deallocated when all "references" 119 * are given up. Only one "reference" to a given 120 * region of an object should be writeable. 121 * 122 * Associated with each object is a list of all resident 123 * memory pages belonging to that object; this list is 124 * maintained by the "vm_page" module, and locked by the object's 125 * lock. 126 * 127 * Each object also records a "pager" routine which is 128 * used to retrieve (and store) pages to the proper backing 129 * storage. In addition, objects may be backed by other 130 * objects from which they were virtual-copied. 131 * 132 * The only items within the object structure which are 133 * modified after time of creation are: 134 * reference count locked by object's lock 135 * pager routine locked by object's lock 136 * 137 */ 138 139struct object_q vm_object_list; 140struct mtx vm_object_list_mtx; /* lock for object list and count */ 141 142struct vm_object kernel_object_store; 143struct vm_object kmem_object_store; 144 145static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 146 "VM object stats"); 147 148static long object_collapses; 149SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 150 &object_collapses, 0, "VM object collapses"); 151 152static long object_bypasses; 153SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 154 &object_bypasses, 0, "VM object bypasses"); 155 156static uma_zone_t obj_zone; 157 158static int vm_object_zinit(void *mem, int size, int flags); 159 160#ifdef INVARIANTS 161static void vm_object_zdtor(void *mem, int size, void *arg); 162 163static void 164vm_object_zdtor(void *mem, int size, void *arg) 165{ 166 vm_object_t object; 167 168 object = (vm_object_t)mem; 169 KASSERT(TAILQ_EMPTY(&object->memq), 170 ("object %p has resident pages in its memq", object)); 171 KASSERT(vm_radix_is_empty(&object->rtree), 172 ("object %p has resident pages in its trie", object)); 173#if VM_NRESERVLEVEL > 0 174 KASSERT(LIST_EMPTY(&object->rvq), 175 ("object %p has reservations", 176 object)); 177#endif 178 KASSERT(vm_object_cache_is_empty(object), 179 ("object %p has cached pages", 180 object)); 181 KASSERT(object->paging_in_progress == 0, 182 ("object %p paging_in_progress = %d", 183 object, object->paging_in_progress)); 184 KASSERT(object->resident_page_count == 0, 185 ("object %p resident_page_count = %d", 186 object, object->resident_page_count)); 187 KASSERT(object->shadow_count == 0, 188 ("object %p shadow_count = %d", 189 object, object->shadow_count)); 190} 191#endif 192 193static int 194vm_object_zinit(void *mem, int size, int flags) 195{ 196 vm_object_t object; 197 198 object = (vm_object_t)mem; 199 bzero(&object->lock, sizeof(object->lock)); 200 rw_init_flags(&object->lock, "vm object", RW_DUPOK); 201 202 /* These are true for any object that has been freed */ 203 object->rtree.rt_root = 0; 204 object->rtree.rt_flags = 0; 205 object->paging_in_progress = 0; 206 object->resident_page_count = 0; 207 object->shadow_count = 0; 208 object->cache.rt_root = 0; 209 object->cache.rt_flags = 0; 210 return (0); 211} 212 213static void 214_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 215{ 216 217 TAILQ_INIT(&object->memq); 218 LIST_INIT(&object->shadow_head); 219 220 object->type = type; 221 switch (type) { 222 case OBJT_DEAD: 223 panic("_vm_object_allocate: can't create OBJT_DEAD"); 224 case OBJT_DEFAULT: 225 case OBJT_SWAP: 226 object->flags = OBJ_ONEMAPPING; 227 break; 228 case OBJT_DEVICE: 229 case OBJT_SG: 230 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 231 break; 232 case OBJT_MGTDEVICE: 233 object->flags = OBJ_FICTITIOUS; 234 break; 235 case OBJT_PHYS: 236 object->flags = OBJ_UNMANAGED; 237 break; 238 case OBJT_VNODE: 239 object->flags = 0; 240 break; 241 default: 242 panic("_vm_object_allocate: type %d is undefined", type); 243 } 244 object->size = size; 245 object->generation = 1; 246 object->ref_count = 1; 247 object->memattr = VM_MEMATTR_DEFAULT; 248 object->cred = NULL; 249 object->charge = 0; 250 object->handle = NULL; 251 object->backing_object = NULL; 252 object->backing_object_offset = (vm_ooffset_t) 0; 253#if VM_NRESERVLEVEL > 0 254 LIST_INIT(&object->rvq); 255#endif 256 257 mtx_lock(&vm_object_list_mtx); 258 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 259 mtx_unlock(&vm_object_list_mtx); 260} 261 262/* 263 * vm_object_init: 264 * 265 * Initialize the VM objects module. 266 */ 267void 268vm_object_init(void) 269{ 270 TAILQ_INIT(&vm_object_list); 271 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 272 273 rw_init(&kernel_object->lock, "kernel vm object"); 274 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 275 kernel_object); 276#if VM_NRESERVLEVEL > 0 277 kernel_object->flags |= OBJ_COLORED; 278 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 279#endif 280 281 rw_init(&kmem_object->lock, "kmem vm object"); 282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 283 kmem_object); 284#if VM_NRESERVLEVEL > 0 285 kmem_object->flags |= OBJ_COLORED; 286 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 287#endif 288 289 /* 290 * The lock portion of struct vm_object must be type stable due 291 * to vm_pageout_fallback_object_lock locking a vm object 292 * without holding any references to it. 293 */ 294 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 295#ifdef INVARIANTS 296 vm_object_zdtor, 297#else 298 NULL, 299#endif 300 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 301 302 vm_radix_init(); 303} 304 305void 306vm_object_clear_flag(vm_object_t object, u_short bits) 307{ 308 309 VM_OBJECT_ASSERT_WLOCKED(object); 310 object->flags &= ~bits; 311} 312 313/* 314 * Sets the default memory attribute for the specified object. Pages 315 * that are allocated to this object are by default assigned this memory 316 * attribute. 317 * 318 * Presently, this function must be called before any pages are allocated 319 * to the object. In the future, this requirement may be relaxed for 320 * "default" and "swap" objects. 321 */ 322int 323vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 324{ 325 326 VM_OBJECT_ASSERT_WLOCKED(object); 327 switch (object->type) { 328 case OBJT_DEFAULT: 329 case OBJT_DEVICE: 330 case OBJT_MGTDEVICE: 331 case OBJT_PHYS: 332 case OBJT_SG: 333 case OBJT_SWAP: 334 case OBJT_VNODE: 335 if (!TAILQ_EMPTY(&object->memq)) 336 return (KERN_FAILURE); 337 break; 338 case OBJT_DEAD: 339 return (KERN_INVALID_ARGUMENT); 340 default: 341 panic("vm_object_set_memattr: object %p is of undefined type", 342 object); 343 } 344 object->memattr = memattr; 345 return (KERN_SUCCESS); 346} 347 348void 349vm_object_pip_add(vm_object_t object, short i) 350{ 351 352 VM_OBJECT_ASSERT_WLOCKED(object); 353 object->paging_in_progress += i; 354} 355 356void 357vm_object_pip_subtract(vm_object_t object, short i) 358{ 359 360 VM_OBJECT_ASSERT_WLOCKED(object); 361 object->paging_in_progress -= i; 362} 363 364void 365vm_object_pip_wakeup(vm_object_t object) 366{ 367 368 VM_OBJECT_ASSERT_WLOCKED(object); 369 object->paging_in_progress--; 370 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 371 vm_object_clear_flag(object, OBJ_PIPWNT); 372 wakeup(object); 373 } 374} 375 376void 377vm_object_pip_wakeupn(vm_object_t object, short i) 378{ 379 380 VM_OBJECT_ASSERT_WLOCKED(object); 381 if (i) 382 object->paging_in_progress -= i; 383 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 384 vm_object_clear_flag(object, OBJ_PIPWNT); 385 wakeup(object); 386 } 387} 388 389void 390vm_object_pip_wait(vm_object_t object, char *waitid) 391{ 392 393 VM_OBJECT_ASSERT_WLOCKED(object); 394 while (object->paging_in_progress) { 395 object->flags |= OBJ_PIPWNT; 396 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 397 } 398} 399 400/* 401 * vm_object_allocate: 402 * 403 * Returns a new object with the given size. 404 */ 405vm_object_t 406vm_object_allocate(objtype_t type, vm_pindex_t size) 407{ 408 vm_object_t object; 409 410 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 411 _vm_object_allocate(type, size, object); 412 return (object); 413} 414 415 416/* 417 * vm_object_reference: 418 * 419 * Gets another reference to the given object. Note: OBJ_DEAD 420 * objects can be referenced during final cleaning. 421 */ 422void 423vm_object_reference(vm_object_t object) 424{ 425 if (object == NULL) 426 return; 427 VM_OBJECT_WLOCK(object); 428 vm_object_reference_locked(object); 429 VM_OBJECT_WUNLOCK(object); 430} 431 432/* 433 * vm_object_reference_locked: 434 * 435 * Gets another reference to the given object. 436 * 437 * The object must be locked. 438 */ 439void 440vm_object_reference_locked(vm_object_t object) 441{ 442 struct vnode *vp; 443 444 VM_OBJECT_ASSERT_WLOCKED(object); 445 object->ref_count++; 446 if (object->type == OBJT_VNODE) { 447 vp = object->handle; 448 vref(vp); 449 } 450} 451 452/* 453 * Handle deallocating an object of type OBJT_VNODE. 454 */ 455static void 456vm_object_vndeallocate(vm_object_t object) 457{ 458 struct vnode *vp = (struct vnode *) object->handle; 459 460 VM_OBJECT_ASSERT_WLOCKED(object); 461 KASSERT(object->type == OBJT_VNODE, 462 ("vm_object_vndeallocate: not a vnode object")); 463 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 464#ifdef INVARIANTS 465 if (object->ref_count == 0) { 466 vprint("vm_object_vndeallocate", vp); 467 panic("vm_object_vndeallocate: bad object reference count"); 468 } 469#endif 470 471 /* 472 * The test for text of vp vnode does not need a bypass to 473 * reach right VV_TEXT there, since it is obtained from 474 * object->handle. 475 */ 476 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 477 object->ref_count--; 478 VM_OBJECT_WUNLOCK(object); 479 /* vrele may need the vnode lock. */ 480 vrele(vp); 481 } else { 482 vhold(vp); 483 VM_OBJECT_WUNLOCK(object); 484 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 485 vdrop(vp); 486 VM_OBJECT_WLOCK(object); 487 object->ref_count--; 488 if (object->type == OBJT_DEAD) { 489 VM_OBJECT_WUNLOCK(object); 490 VOP_UNLOCK(vp, 0); 491 } else { 492 if (object->ref_count == 0) 493 VOP_UNSET_TEXT(vp); 494 VM_OBJECT_WUNLOCK(object); 495 vput(vp); 496 } 497 } 498} 499 500/* 501 * vm_object_deallocate: 502 * 503 * Release a reference to the specified object, 504 * gained either through a vm_object_allocate 505 * or a vm_object_reference call. When all references 506 * are gone, storage associated with this object 507 * may be relinquished. 508 * 509 * No object may be locked. 510 */ 511void 512vm_object_deallocate(vm_object_t object) 513{ 514 vm_object_t temp; 515 struct vnode *vp; 516 517 while (object != NULL) { 518 VM_OBJECT_WLOCK(object); 519 if (object->type == OBJT_VNODE) { 520 vm_object_vndeallocate(object); 521 return; 522 } 523 524 KASSERT(object->ref_count != 0, 525 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 526 527 /* 528 * If the reference count goes to 0 we start calling 529 * vm_object_terminate() on the object chain. 530 * A ref count of 1 may be a special case depending on the 531 * shadow count being 0 or 1. 532 */ 533 object->ref_count--; 534 if (object->ref_count > 1) { 535 VM_OBJECT_WUNLOCK(object); 536 return; 537 } else if (object->ref_count == 1) { 538 if (object->type == OBJT_SWAP && 539 (object->flags & OBJ_TMPFS) != 0) { 540 vp = object->un_pager.swp.swp_tmpfs; 541 vhold(vp); 542 VM_OBJECT_WUNLOCK(object); 543 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 544 VM_OBJECT_WLOCK(object); 545 if (object->type == OBJT_DEAD || 546 object->ref_count != 1) { 547 VM_OBJECT_WUNLOCK(object); 548 VOP_UNLOCK(vp, 0); 549 vdrop(vp); 550 return; 551 } 552 if ((object->flags & OBJ_TMPFS) != 0) 553 VOP_UNSET_TEXT(vp); 554 VOP_UNLOCK(vp, 0); 555 vdrop(vp); 556 } 557 if (object->shadow_count == 0 && 558 object->handle == NULL && 559 (object->type == OBJT_DEFAULT || 560 (object->type == OBJT_SWAP && 561 (object->flags & OBJ_TMPFS_NODE) == 0))) { 562 vm_object_set_flag(object, OBJ_ONEMAPPING); 563 } else if ((object->shadow_count == 1) && 564 (object->handle == NULL) && 565 (object->type == OBJT_DEFAULT || 566 object->type == OBJT_SWAP)) { 567 vm_object_t robject; 568 569 robject = LIST_FIRST(&object->shadow_head); 570 KASSERT(robject != NULL, 571 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 572 object->ref_count, 573 object->shadow_count)); 574 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 575 ("shadowed tmpfs v_object %p", object)); 576 if (!VM_OBJECT_TRYWLOCK(robject)) { 577 /* 578 * Avoid a potential deadlock. 579 */ 580 object->ref_count++; 581 VM_OBJECT_WUNLOCK(object); 582 /* 583 * More likely than not the thread 584 * holding robject's lock has lower 585 * priority than the current thread. 586 * Let the lower priority thread run. 587 */ 588 pause("vmo_de", 1); 589 continue; 590 } 591 /* 592 * Collapse object into its shadow unless its 593 * shadow is dead. In that case, object will 594 * be deallocated by the thread that is 595 * deallocating its shadow. 596 */ 597 if ((robject->flags & OBJ_DEAD) == 0 && 598 (robject->handle == NULL) && 599 (robject->type == OBJT_DEFAULT || 600 robject->type == OBJT_SWAP)) { 601 602 robject->ref_count++; 603retry: 604 if (robject->paging_in_progress) { 605 VM_OBJECT_WUNLOCK(object); 606 vm_object_pip_wait(robject, 607 "objde1"); 608 temp = robject->backing_object; 609 if (object == temp) { 610 VM_OBJECT_WLOCK(object); 611 goto retry; 612 } 613 } else if (object->paging_in_progress) { 614 VM_OBJECT_WUNLOCK(robject); 615 object->flags |= OBJ_PIPWNT; 616 VM_OBJECT_SLEEP(object, object, 617 PDROP | PVM, "objde2", 0); 618 VM_OBJECT_WLOCK(robject); 619 temp = robject->backing_object; 620 if (object == temp) { 621 VM_OBJECT_WLOCK(object); 622 goto retry; 623 } 624 } else 625 VM_OBJECT_WUNLOCK(object); 626 627 if (robject->ref_count == 1) { 628 robject->ref_count--; 629 object = robject; 630 goto doterm; 631 } 632 object = robject; 633 vm_object_collapse(object); 634 VM_OBJECT_WUNLOCK(object); 635 continue; 636 } 637 VM_OBJECT_WUNLOCK(robject); 638 } 639 VM_OBJECT_WUNLOCK(object); 640 return; 641 } 642doterm: 643 temp = object->backing_object; 644 if (temp != NULL) { 645 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 646 ("shadowed tmpfs v_object 2 %p", object)); 647 VM_OBJECT_WLOCK(temp); 648 LIST_REMOVE(object, shadow_list); 649 temp->shadow_count--; 650 VM_OBJECT_WUNLOCK(temp); 651 object->backing_object = NULL; 652 } 653 /* 654 * Don't double-terminate, we could be in a termination 655 * recursion due to the terminate having to sync data 656 * to disk. 657 */ 658 if ((object->flags & OBJ_DEAD) == 0) 659 vm_object_terminate(object); 660 else 661 VM_OBJECT_WUNLOCK(object); 662 object = temp; 663 } 664} 665 666/* 667 * vm_object_destroy removes the object from the global object list 668 * and frees the space for the object. 669 */ 670void 671vm_object_destroy(vm_object_t object) 672{ 673 674 /* 675 * Remove the object from the global object list. 676 */ 677 mtx_lock(&vm_object_list_mtx); 678 TAILQ_REMOVE(&vm_object_list, object, object_list); 679 mtx_unlock(&vm_object_list_mtx); 680 681 /* 682 * Release the allocation charge. 683 */ 684 if (object->cred != NULL) { 685 KASSERT(object->type == OBJT_DEFAULT || 686 object->type == OBJT_SWAP, 687 ("vm_object_terminate: non-swap obj %p has cred", 688 object)); 689 swap_release_by_cred(object->charge, object->cred); 690 object->charge = 0; 691 crfree(object->cred); 692 object->cred = NULL; 693 } 694 695 /* 696 * Free the space for the object. 697 */ 698 uma_zfree(obj_zone, object); 699} 700 701/* 702 * vm_object_terminate actually destroys the specified object, freeing 703 * up all previously used resources. 704 * 705 * The object must be locked. 706 * This routine may block. 707 */ 708void 709vm_object_terminate(vm_object_t object) 710{ 711 vm_page_t p, p_next; 712 713 VM_OBJECT_ASSERT_WLOCKED(object); 714 715 /* 716 * Make sure no one uses us. 717 */ 718 vm_object_set_flag(object, OBJ_DEAD); 719 720 /* 721 * wait for the pageout daemon to be done with the object 722 */ 723 vm_object_pip_wait(object, "objtrm"); 724 725 KASSERT(!object->paging_in_progress, 726 ("vm_object_terminate: pageout in progress")); 727 728 /* 729 * Clean and free the pages, as appropriate. All references to the 730 * object are gone, so we don't need to lock it. 731 */ 732 if (object->type == OBJT_VNODE) { 733 struct vnode *vp = (struct vnode *)object->handle; 734 735 /* 736 * Clean pages and flush buffers. 737 */ 738 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 739 VM_OBJECT_WUNLOCK(object); 740 741 vinvalbuf(vp, V_SAVE, 0, 0); 742 743 VM_OBJECT_WLOCK(object); 744 } 745 746 KASSERT(object->ref_count == 0, 747 ("vm_object_terminate: object with references, ref_count=%d", 748 object->ref_count)); 749 750 /* 751 * Free any remaining pageable pages. This also removes them from the 752 * paging queues. However, don't free wired pages, just remove them 753 * from the object. Rather than incrementally removing each page from 754 * the object, the page and object are reset to any empty state. 755 */ 756 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 757 vm_page_assert_unbusied(p); 758 vm_page_lock(p); 759 /* 760 * Optimize the page's removal from the object by resetting 761 * its "object" field. Specifically, if the page is not 762 * wired, then the effect of this assignment is that 763 * vm_page_free()'s call to vm_page_remove() will return 764 * immediately without modifying the page or the object. 765 */ 766 p->object = NULL; 767 if (p->wire_count == 0) { 768 vm_page_free(p); 769 PCPU_INC(cnt.v_pfree); 770 } 771 vm_page_unlock(p); 772 } 773 /* 774 * If the object contained any pages, then reset it to an empty state. 775 * None of the object's fields, including "resident_page_count", were 776 * modified by the preceding loop. 777 */ 778 if (object->resident_page_count != 0) { 779 vm_radix_reclaim_allnodes(&object->rtree); 780 TAILQ_INIT(&object->memq); 781 object->resident_page_count = 0; 782 if (object->type == OBJT_VNODE) 783 vdrop(object->handle); 784 } 785 786#if VM_NRESERVLEVEL > 0 787 if (__predict_false(!LIST_EMPTY(&object->rvq))) 788 vm_reserv_break_all(object); 789#endif 790 if (__predict_false(!vm_object_cache_is_empty(object))) 791 vm_page_cache_free(object, 0, 0); 792 793 /* 794 * Let the pager know object is dead. 795 */ 796 vm_pager_deallocate(object); 797 VM_OBJECT_WUNLOCK(object); 798 799 vm_object_destroy(object); 800} 801 802/* 803 * Make the page read-only so that we can clear the object flags. However, if 804 * this is a nosync mmap then the object is likely to stay dirty so do not 805 * mess with the page and do not clear the object flags. Returns TRUE if the 806 * page should be flushed, and FALSE otherwise. 807 */ 808static boolean_t 809vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 810{ 811 812 /* 813 * If we have been asked to skip nosync pages and this is a 814 * nosync page, skip it. Note that the object flags were not 815 * cleared in this case so we do not have to set them. 816 */ 817 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 818 *clearobjflags = FALSE; 819 return (FALSE); 820 } else { 821 pmap_remove_write(p); 822 return (p->dirty != 0); 823 } 824} 825 826/* 827 * vm_object_page_clean 828 * 829 * Clean all dirty pages in the specified range of object. Leaves page 830 * on whatever queue it is currently on. If NOSYNC is set then do not 831 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 832 * leaving the object dirty. 833 * 834 * When stuffing pages asynchronously, allow clustering. XXX we need a 835 * synchronous clustering mode implementation. 836 * 837 * Odd semantics: if start == end, we clean everything. 838 * 839 * The object must be locked. 840 * 841 * Returns FALSE if some page from the range was not written, as 842 * reported by the pager, and TRUE otherwise. 843 */ 844boolean_t 845vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 846 int flags) 847{ 848 vm_page_t np, p; 849 vm_pindex_t pi, tend, tstart; 850 int curgeneration, n, pagerflags; 851 boolean_t clearobjflags, eio, res; 852 853 VM_OBJECT_ASSERT_WLOCKED(object); 854 855 /* 856 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 857 * objects. The check below prevents the function from 858 * operating on non-vnode objects. 859 */ 860 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 861 object->resident_page_count == 0) 862 return (TRUE); 863 864 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 865 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 866 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 867 868 tstart = OFF_TO_IDX(start); 869 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 870 clearobjflags = tstart == 0 && tend >= object->size; 871 res = TRUE; 872 873rescan: 874 curgeneration = object->generation; 875 876 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 877 pi = p->pindex; 878 if (pi >= tend) 879 break; 880 np = TAILQ_NEXT(p, listq); 881 if (p->valid == 0) 882 continue; 883 if (vm_page_sleep_if_busy(p, "vpcwai")) { 884 if (object->generation != curgeneration) { 885 if ((flags & OBJPC_SYNC) != 0) 886 goto rescan; 887 else 888 clearobjflags = FALSE; 889 } 890 np = vm_page_find_least(object, pi); 891 continue; 892 } 893 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 894 continue; 895 896 n = vm_object_page_collect_flush(object, p, pagerflags, 897 flags, &clearobjflags, &eio); 898 if (eio) { 899 res = FALSE; 900 clearobjflags = FALSE; 901 } 902 if (object->generation != curgeneration) { 903 if ((flags & OBJPC_SYNC) != 0) 904 goto rescan; 905 else 906 clearobjflags = FALSE; 907 } 908 909 /* 910 * If the VOP_PUTPAGES() did a truncated write, so 911 * that even the first page of the run is not fully 912 * written, vm_pageout_flush() returns 0 as the run 913 * length. Since the condition that caused truncated 914 * write may be permanent, e.g. exhausted free space, 915 * accepting n == 0 would cause an infinite loop. 916 * 917 * Forwarding the iterator leaves the unwritten page 918 * behind, but there is not much we can do there if 919 * filesystem refuses to write it. 920 */ 921 if (n == 0) { 922 n = 1; 923 clearobjflags = FALSE; 924 } 925 np = vm_page_find_least(object, pi + n); 926 } 927#if 0 928 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 929#endif 930 931 if (clearobjflags) 932 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 933 return (res); 934} 935 936static int 937vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 938 int flags, boolean_t *clearobjflags, boolean_t *eio) 939{ 940 vm_page_t ma[vm_pageout_page_count], p_first, tp; 941 int count, i, mreq, runlen; 942 943 vm_page_lock_assert(p, MA_NOTOWNED); 944 VM_OBJECT_ASSERT_WLOCKED(object); 945 946 count = 1; 947 mreq = 0; 948 949 for (tp = p; count < vm_pageout_page_count; count++) { 950 tp = vm_page_next(tp); 951 if (tp == NULL || vm_page_busied(tp)) 952 break; 953 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 954 break; 955 } 956 957 for (p_first = p; count < vm_pageout_page_count; count++) { 958 tp = vm_page_prev(p_first); 959 if (tp == NULL || vm_page_busied(tp)) 960 break; 961 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 962 break; 963 p_first = tp; 964 mreq++; 965 } 966 967 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 968 ma[i] = tp; 969 970 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 971 return (runlen); 972} 973 974/* 975 * Note that there is absolutely no sense in writing out 976 * anonymous objects, so we track down the vnode object 977 * to write out. 978 * We invalidate (remove) all pages from the address space 979 * for semantic correctness. 980 * 981 * If the backing object is a device object with unmanaged pages, then any 982 * mappings to the specified range of pages must be removed before this 983 * function is called. 984 * 985 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 986 * may start out with a NULL object. 987 */ 988boolean_t 989vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 990 boolean_t syncio, boolean_t invalidate) 991{ 992 vm_object_t backing_object; 993 struct vnode *vp; 994 struct mount *mp; 995 int error, flags, fsync_after; 996 boolean_t res; 997 998 if (object == NULL) 999 return (TRUE); 1000 res = TRUE; 1001 error = 0; 1002 VM_OBJECT_WLOCK(object); 1003 while ((backing_object = object->backing_object) != NULL) { 1004 VM_OBJECT_WLOCK(backing_object); 1005 offset += object->backing_object_offset; 1006 VM_OBJECT_WUNLOCK(object); 1007 object = backing_object; 1008 if (object->size < OFF_TO_IDX(offset + size)) 1009 size = IDX_TO_OFF(object->size) - offset; 1010 } 1011 /* 1012 * Flush pages if writing is allowed, invalidate them 1013 * if invalidation requested. Pages undergoing I/O 1014 * will be ignored by vm_object_page_remove(). 1015 * 1016 * We cannot lock the vnode and then wait for paging 1017 * to complete without deadlocking against vm_fault. 1018 * Instead we simply call vm_object_page_remove() and 1019 * allow it to block internally on a page-by-page 1020 * basis when it encounters pages undergoing async 1021 * I/O. 1022 */ 1023 if (object->type == OBJT_VNODE && 1024 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1025 vp = object->handle; 1026 VM_OBJECT_WUNLOCK(object); 1027 (void) vn_start_write(vp, &mp, V_WAIT); 1028 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1029 if (syncio && !invalidate && offset == 0 && 1030 OFF_TO_IDX(size) == object->size) { 1031 /* 1032 * If syncing the whole mapping of the file, 1033 * it is faster to schedule all the writes in 1034 * async mode, also allowing the clustering, 1035 * and then wait for i/o to complete. 1036 */ 1037 flags = 0; 1038 fsync_after = TRUE; 1039 } else { 1040 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1041 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1042 fsync_after = FALSE; 1043 } 1044 VM_OBJECT_WLOCK(object); 1045 res = vm_object_page_clean(object, offset, offset + size, 1046 flags); 1047 VM_OBJECT_WUNLOCK(object); 1048 if (fsync_after) 1049 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1050 VOP_UNLOCK(vp, 0); 1051 vn_finished_write(mp); 1052 if (error != 0) 1053 res = FALSE; 1054 VM_OBJECT_WLOCK(object); 1055 } 1056 if ((object->type == OBJT_VNODE || 1057 object->type == OBJT_DEVICE) && invalidate) { 1058 if (object->type == OBJT_DEVICE) 1059 /* 1060 * The option OBJPR_NOTMAPPED must be passed here 1061 * because vm_object_page_remove() cannot remove 1062 * unmanaged mappings. 1063 */ 1064 flags = OBJPR_NOTMAPPED; 1065 else if (old_msync) 1066 flags = OBJPR_NOTWIRED; 1067 else 1068 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED; 1069 vm_object_page_remove(object, OFF_TO_IDX(offset), 1070 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1071 } 1072 VM_OBJECT_WUNLOCK(object); 1073 return (res); 1074} 1075 1076/* 1077 * vm_object_madvise: 1078 * 1079 * Implements the madvise function at the object/page level. 1080 * 1081 * MADV_WILLNEED (any object) 1082 * 1083 * Activate the specified pages if they are resident. 1084 * 1085 * MADV_DONTNEED (any object) 1086 * 1087 * Deactivate the specified pages if they are resident. 1088 * 1089 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1090 * OBJ_ONEMAPPING only) 1091 * 1092 * Deactivate and clean the specified pages if they are 1093 * resident. This permits the process to reuse the pages 1094 * without faulting or the kernel to reclaim the pages 1095 * without I/O. 1096 */ 1097void 1098vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1099 int advise) 1100{ 1101 vm_pindex_t tpindex; 1102 vm_object_t backing_object, tobject; 1103 vm_page_t m; 1104 1105 if (object == NULL) 1106 return; 1107 VM_OBJECT_WLOCK(object); 1108 /* 1109 * Locate and adjust resident pages 1110 */ 1111 for (; pindex < end; pindex += 1) { 1112relookup: 1113 tobject = object; 1114 tpindex = pindex; 1115shadowlookup: 1116 /* 1117 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1118 * and those pages must be OBJ_ONEMAPPING. 1119 */ 1120 if (advise == MADV_FREE) { 1121 if ((tobject->type != OBJT_DEFAULT && 1122 tobject->type != OBJT_SWAP) || 1123 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1124 goto unlock_tobject; 1125 } 1126 } else if ((tobject->flags & OBJ_UNMANAGED) != 0) 1127 goto unlock_tobject; 1128 m = vm_page_lookup(tobject, tpindex); 1129 if (m == NULL && advise == MADV_WILLNEED) { 1130 /* 1131 * If the page is cached, reactivate it. 1132 */ 1133 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1134 VM_ALLOC_NOBUSY); 1135 } 1136 if (m == NULL) { 1137 /* 1138 * There may be swap even if there is no backing page 1139 */ 1140 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1141 swap_pager_freespace(tobject, tpindex, 1); 1142 /* 1143 * next object 1144 */ 1145 backing_object = tobject->backing_object; 1146 if (backing_object == NULL) 1147 goto unlock_tobject; 1148 VM_OBJECT_WLOCK(backing_object); 1149 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1150 if (tobject != object) 1151 VM_OBJECT_WUNLOCK(tobject); 1152 tobject = backing_object; 1153 goto shadowlookup; 1154 } else if (m->valid != VM_PAGE_BITS_ALL) 1155 goto unlock_tobject; 1156 /* 1157 * If the page is not in a normal state, skip it. 1158 */ 1159 vm_page_lock(m); 1160 if (m->hold_count != 0 || m->wire_count != 0) { 1161 vm_page_unlock(m); 1162 goto unlock_tobject; 1163 } 1164 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1165 ("vm_object_madvise: page %p is fictitious", m)); 1166 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1167 ("vm_object_madvise: page %p is not managed", m)); 1168 if (vm_page_busied(m)) { 1169 if (advise == MADV_WILLNEED) { 1170 /* 1171 * Reference the page before unlocking and 1172 * sleeping so that the page daemon is less 1173 * likely to reclaim it. 1174 */ 1175 vm_page_aflag_set(m, PGA_REFERENCED); 1176 } 1177 if (object != tobject) 1178 VM_OBJECT_WUNLOCK(object); 1179 VM_OBJECT_WUNLOCK(tobject); 1180 vm_page_busy_sleep(m, "madvpo"); 1181 VM_OBJECT_WLOCK(object); 1182 goto relookup; 1183 } 1184 if (advise == MADV_WILLNEED) { 1185 vm_page_activate(m); 1186 } else { 1187 vm_page_advise(m, advise); 1188 } 1189 vm_page_unlock(m); 1190 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1191 swap_pager_freespace(tobject, tpindex, 1); 1192unlock_tobject: 1193 if (tobject != object) 1194 VM_OBJECT_WUNLOCK(tobject); 1195 } 1196 VM_OBJECT_WUNLOCK(object); 1197} 1198 1199/* 1200 * vm_object_shadow: 1201 * 1202 * Create a new object which is backed by the 1203 * specified existing object range. The source 1204 * object reference is deallocated. 1205 * 1206 * The new object and offset into that object 1207 * are returned in the source parameters. 1208 */ 1209void 1210vm_object_shadow( 1211 vm_object_t *object, /* IN/OUT */ 1212 vm_ooffset_t *offset, /* IN/OUT */ 1213 vm_size_t length) 1214{ 1215 vm_object_t source; 1216 vm_object_t result; 1217 1218 source = *object; 1219 1220 /* 1221 * Don't create the new object if the old object isn't shared. 1222 */ 1223 if (source != NULL) { 1224 VM_OBJECT_WLOCK(source); 1225 if (source->ref_count == 1 && 1226 source->handle == NULL && 1227 (source->type == OBJT_DEFAULT || 1228 source->type == OBJT_SWAP)) { 1229 VM_OBJECT_WUNLOCK(source); 1230 return; 1231 } 1232 VM_OBJECT_WUNLOCK(source); 1233 } 1234 1235 /* 1236 * Allocate a new object with the given length. 1237 */ 1238 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1239 1240 /* 1241 * The new object shadows the source object, adding a reference to it. 1242 * Our caller changes his reference to point to the new object, 1243 * removing a reference to the source object. Net result: no change 1244 * of reference count. 1245 * 1246 * Try to optimize the result object's page color when shadowing 1247 * in order to maintain page coloring consistency in the combined 1248 * shadowed object. 1249 */ 1250 result->backing_object = source; 1251 /* 1252 * Store the offset into the source object, and fix up the offset into 1253 * the new object. 1254 */ 1255 result->backing_object_offset = *offset; 1256 if (source != NULL) { 1257 VM_OBJECT_WLOCK(source); 1258 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1259 source->shadow_count++; 1260#if VM_NRESERVLEVEL > 0 1261 result->flags |= source->flags & OBJ_COLORED; 1262 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1263 ((1 << (VM_NFREEORDER - 1)) - 1); 1264#endif 1265 VM_OBJECT_WUNLOCK(source); 1266 } 1267 1268 1269 /* 1270 * Return the new things 1271 */ 1272 *offset = 0; 1273 *object = result; 1274} 1275 1276/* 1277 * vm_object_split: 1278 * 1279 * Split the pages in a map entry into a new object. This affords 1280 * easier removal of unused pages, and keeps object inheritance from 1281 * being a negative impact on memory usage. 1282 */ 1283void 1284vm_object_split(vm_map_entry_t entry) 1285{ 1286 vm_page_t m, m_next; 1287 vm_object_t orig_object, new_object, source; 1288 vm_pindex_t idx, offidxstart; 1289 vm_size_t size; 1290 1291 orig_object = entry->object.vm_object; 1292 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1293 return; 1294 if (orig_object->ref_count <= 1) 1295 return; 1296 VM_OBJECT_WUNLOCK(orig_object); 1297 1298 offidxstart = OFF_TO_IDX(entry->offset); 1299 size = atop(entry->end - entry->start); 1300 1301 /* 1302 * If swap_pager_copy() is later called, it will convert new_object 1303 * into a swap object. 1304 */ 1305 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1306 1307 /* 1308 * At this point, the new object is still private, so the order in 1309 * which the original and new objects are locked does not matter. 1310 */ 1311 VM_OBJECT_WLOCK(new_object); 1312 VM_OBJECT_WLOCK(orig_object); 1313 source = orig_object->backing_object; 1314 if (source != NULL) { 1315 VM_OBJECT_WLOCK(source); 1316 if ((source->flags & OBJ_DEAD) != 0) { 1317 VM_OBJECT_WUNLOCK(source); 1318 VM_OBJECT_WUNLOCK(orig_object); 1319 VM_OBJECT_WUNLOCK(new_object); 1320 vm_object_deallocate(new_object); 1321 VM_OBJECT_WLOCK(orig_object); 1322 return; 1323 } 1324 LIST_INSERT_HEAD(&source->shadow_head, 1325 new_object, shadow_list); 1326 source->shadow_count++; 1327 vm_object_reference_locked(source); /* for new_object */ 1328 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1329 VM_OBJECT_WUNLOCK(source); 1330 new_object->backing_object_offset = 1331 orig_object->backing_object_offset + entry->offset; 1332 new_object->backing_object = source; 1333 } 1334 if (orig_object->cred != NULL) { 1335 new_object->cred = orig_object->cred; 1336 crhold(orig_object->cred); 1337 new_object->charge = ptoa(size); 1338 KASSERT(orig_object->charge >= ptoa(size), 1339 ("orig_object->charge < 0")); 1340 orig_object->charge -= ptoa(size); 1341 } 1342retry: 1343 m = vm_page_find_least(orig_object, offidxstart); 1344 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1345 m = m_next) { 1346 m_next = TAILQ_NEXT(m, listq); 1347 1348 /* 1349 * We must wait for pending I/O to complete before we can 1350 * rename the page. 1351 * 1352 * We do not have to VM_PROT_NONE the page as mappings should 1353 * not be changed by this operation. 1354 */ 1355 if (vm_page_busied(m)) { 1356 VM_OBJECT_WUNLOCK(new_object); 1357 vm_page_lock(m); 1358 VM_OBJECT_WUNLOCK(orig_object); 1359 vm_page_busy_sleep(m, "spltwt"); 1360 VM_OBJECT_WLOCK(orig_object); 1361 VM_OBJECT_WLOCK(new_object); 1362 goto retry; 1363 } 1364 1365 /* vm_page_rename() will handle dirty and cache. */ 1366 if (vm_page_rename(m, new_object, idx)) { 1367 VM_OBJECT_WUNLOCK(new_object); 1368 VM_OBJECT_WUNLOCK(orig_object); 1369 VM_WAIT; 1370 VM_OBJECT_WLOCK(orig_object); 1371 VM_OBJECT_WLOCK(new_object); 1372 goto retry; 1373 } 1374#if VM_NRESERVLEVEL > 0 1375 /* 1376 * If some of the reservation's allocated pages remain with 1377 * the original object, then transferring the reservation to 1378 * the new object is neither particularly beneficial nor 1379 * particularly harmful as compared to leaving the reservation 1380 * with the original object. If, however, all of the 1381 * reservation's allocated pages are transferred to the new 1382 * object, then transferring the reservation is typically 1383 * beneficial. Determining which of these two cases applies 1384 * would be more costly than unconditionally renaming the 1385 * reservation. 1386 */ 1387 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1388#endif 1389 if (orig_object->type == OBJT_SWAP) 1390 vm_page_xbusy(m); 1391 } 1392 if (orig_object->type == OBJT_SWAP) { 1393 /* 1394 * swap_pager_copy() can sleep, in which case the orig_object's 1395 * and new_object's locks are released and reacquired. 1396 */ 1397 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1398 TAILQ_FOREACH(m, &new_object->memq, listq) 1399 vm_page_xunbusy(m); 1400 1401 /* 1402 * Transfer any cached pages from orig_object to new_object. 1403 * If swap_pager_copy() found swapped out pages within the 1404 * specified range of orig_object, then it changed 1405 * new_object's type to OBJT_SWAP when it transferred those 1406 * pages to new_object. Otherwise, new_object's type 1407 * should still be OBJT_DEFAULT and orig_object should not 1408 * contain any cached pages within the specified range. 1409 */ 1410 if (__predict_false(!vm_object_cache_is_empty(orig_object))) 1411 vm_page_cache_transfer(orig_object, offidxstart, 1412 new_object); 1413 } 1414 VM_OBJECT_WUNLOCK(orig_object); 1415 VM_OBJECT_WUNLOCK(new_object); 1416 entry->object.vm_object = new_object; 1417 entry->offset = 0LL; 1418 vm_object_deallocate(orig_object); 1419 VM_OBJECT_WLOCK(new_object); 1420} 1421 1422#define OBSC_TEST_ALL_SHADOWED 0x0001 1423#define OBSC_COLLAPSE_NOWAIT 0x0002 1424#define OBSC_COLLAPSE_WAIT 0x0004 1425 1426static int 1427vm_object_backing_scan(vm_object_t object, int op) 1428{ 1429 int r = 1; 1430 vm_page_t p; 1431 vm_object_t backing_object; 1432 vm_pindex_t backing_offset_index; 1433 1434 VM_OBJECT_ASSERT_WLOCKED(object); 1435 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1436 1437 backing_object = object->backing_object; 1438 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1439 1440 /* 1441 * Initial conditions 1442 */ 1443 if (op & OBSC_TEST_ALL_SHADOWED) { 1444 /* 1445 * We do not want to have to test for the existence of cache 1446 * or swap pages in the backing object. XXX but with the 1447 * new swapper this would be pretty easy to do. 1448 * 1449 * XXX what about anonymous MAP_SHARED memory that hasn't 1450 * been ZFOD faulted yet? If we do not test for this, the 1451 * shadow test may succeed! XXX 1452 */ 1453 if (backing_object->type != OBJT_DEFAULT) { 1454 return (0); 1455 } 1456 } 1457 if (op & OBSC_COLLAPSE_WAIT) { 1458 vm_object_set_flag(backing_object, OBJ_DEAD); 1459 } 1460 1461 /* 1462 * Our scan 1463 */ 1464 p = TAILQ_FIRST(&backing_object->memq); 1465 while (p) { 1466 vm_page_t next = TAILQ_NEXT(p, listq); 1467 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1468 1469 if (op & OBSC_TEST_ALL_SHADOWED) { 1470 vm_page_t pp; 1471 1472 /* 1473 * Ignore pages outside the parent object's range 1474 * and outside the parent object's mapping of the 1475 * backing object. 1476 * 1477 * note that we do not busy the backing object's 1478 * page. 1479 */ 1480 if ( 1481 p->pindex < backing_offset_index || 1482 new_pindex >= object->size 1483 ) { 1484 p = next; 1485 continue; 1486 } 1487 1488 /* 1489 * See if the parent has the page or if the parent's 1490 * object pager has the page. If the parent has the 1491 * page but the page is not valid, the parent's 1492 * object pager must have the page. 1493 * 1494 * If this fails, the parent does not completely shadow 1495 * the object and we might as well give up now. 1496 */ 1497 1498 pp = vm_page_lookup(object, new_pindex); 1499 if ( 1500 (pp == NULL || pp->valid == 0) && 1501 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1502 ) { 1503 r = 0; 1504 break; 1505 } 1506 } 1507 1508 /* 1509 * Check for busy page 1510 */ 1511 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1512 vm_page_t pp; 1513 1514 if (op & OBSC_COLLAPSE_NOWAIT) { 1515 if (!p->valid || vm_page_busied(p)) { 1516 p = next; 1517 continue; 1518 } 1519 } else if (op & OBSC_COLLAPSE_WAIT) { 1520 if (vm_page_busied(p)) { 1521 VM_OBJECT_WUNLOCK(object); 1522 vm_page_lock(p); 1523 VM_OBJECT_WUNLOCK(backing_object); 1524 vm_page_busy_sleep(p, "vmocol"); 1525 VM_OBJECT_WLOCK(object); 1526 VM_OBJECT_WLOCK(backing_object); 1527 /* 1528 * If we slept, anything could have 1529 * happened. Since the object is 1530 * marked dead, the backing offset 1531 * should not have changed so we 1532 * just restart our scan. 1533 */ 1534 p = TAILQ_FIRST(&backing_object->memq); 1535 continue; 1536 } 1537 } 1538 1539 KASSERT( 1540 p->object == backing_object, 1541 ("vm_object_backing_scan: object mismatch") 1542 ); 1543 1544 if ( 1545 p->pindex < backing_offset_index || 1546 new_pindex >= object->size 1547 ) { 1548 if (backing_object->type == OBJT_SWAP) 1549 swap_pager_freespace(backing_object, 1550 p->pindex, 1); 1551 1552 /* 1553 * Page is out of the parent object's range, we 1554 * can simply destroy it. 1555 */ 1556 vm_page_lock(p); 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(p); 1564 p = next; 1565 continue; 1566 } 1567 1568 pp = vm_page_lookup(object, new_pindex); 1569 if ( 1570 (op & OBSC_COLLAPSE_NOWAIT) != 0 && 1571 (pp != NULL && pp->valid == 0) 1572 ) { 1573 if (backing_object->type == OBJT_SWAP) 1574 swap_pager_freespace(backing_object, 1575 p->pindex, 1); 1576 1577 /* 1578 * The page in the parent is not (yet) valid. 1579 * We don't know anything about the state of 1580 * the original page. It might be mapped, 1581 * so we must avoid the next if here. 1582 * 1583 * This is due to a race in vm_fault() where 1584 * we must unbusy the original (backing_obj) 1585 * page before we can (re)lock the parent. 1586 * Hence we can get here. 1587 */ 1588 p = next; 1589 continue; 1590 } 1591 if ( 1592 pp != NULL || 1593 vm_pager_has_page(object, new_pindex, NULL, NULL) 1594 ) { 1595 if (backing_object->type == OBJT_SWAP) 1596 swap_pager_freespace(backing_object, 1597 p->pindex, 1); 1598 1599 /* 1600 * page already exists in parent OR swap exists 1601 * for this location in the parent. Destroy 1602 * the original page from the backing object. 1603 * 1604 * Leave the parent's page alone 1605 */ 1606 vm_page_lock(p); 1607 KASSERT(!pmap_page_is_mapped(p), 1608 ("freeing mapped page %p", p)); 1609 if (p->wire_count == 0) 1610 vm_page_free(p); 1611 else 1612 vm_page_remove(p); 1613 vm_page_unlock(p); 1614 p = next; 1615 continue; 1616 } 1617 1618 /* 1619 * Page does not exist in parent, rename the 1620 * page from the backing object to the main object. 1621 * 1622 * If the page was mapped to a process, it can remain 1623 * mapped through the rename. 1624 * vm_page_rename() will handle dirty and cache. 1625 */ 1626 if (vm_page_rename(p, object, new_pindex)) { 1627 if (op & OBSC_COLLAPSE_NOWAIT) { 1628 p = next; 1629 continue; 1630 } 1631 VM_OBJECT_WLOCK(backing_object); 1632 VM_OBJECT_WUNLOCK(object); 1633 VM_WAIT; 1634 VM_OBJECT_WLOCK(object); 1635 VM_OBJECT_WLOCK(backing_object); 1636 p = TAILQ_FIRST(&backing_object->memq); 1637 continue; 1638 } 1639 1640 /* Use the old pindex to free the right page. */ 1641 if (backing_object->type == OBJT_SWAP) 1642 swap_pager_freespace(backing_object, 1643 new_pindex + backing_offset_index, 1); 1644 1645#if VM_NRESERVLEVEL > 0 1646 /* 1647 * Rename the reservation. 1648 */ 1649 vm_reserv_rename(p, object, backing_object, 1650 backing_offset_index); 1651#endif 1652 } 1653 p = next; 1654 } 1655 return (r); 1656} 1657 1658 1659/* 1660 * this version of collapse allows the operation to occur earlier and 1661 * when paging_in_progress is true for an object... This is not a complete 1662 * operation, but should plug 99.9% of the rest of the leaks. 1663 */ 1664static void 1665vm_object_qcollapse(vm_object_t object) 1666{ 1667 vm_object_t backing_object = object->backing_object; 1668 1669 VM_OBJECT_ASSERT_WLOCKED(object); 1670 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1671 1672 if (backing_object->ref_count != 1) 1673 return; 1674 1675 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1676} 1677 1678/* 1679 * vm_object_collapse: 1680 * 1681 * Collapse an object with the object backing it. 1682 * Pages in the backing object are moved into the 1683 * parent, and the backing object is deallocated. 1684 */ 1685void 1686vm_object_collapse(vm_object_t object) 1687{ 1688 VM_OBJECT_ASSERT_WLOCKED(object); 1689 1690 while (TRUE) { 1691 vm_object_t backing_object; 1692 1693 /* 1694 * Verify that the conditions are right for collapse: 1695 * 1696 * The object exists and the backing object exists. 1697 */ 1698 if ((backing_object = object->backing_object) == NULL) 1699 break; 1700 1701 /* 1702 * we check the backing object first, because it is most likely 1703 * not collapsable. 1704 */ 1705 VM_OBJECT_WLOCK(backing_object); 1706 if (backing_object->handle != NULL || 1707 (backing_object->type != OBJT_DEFAULT && 1708 backing_object->type != OBJT_SWAP) || 1709 (backing_object->flags & OBJ_DEAD) || 1710 object->handle != NULL || 1711 (object->type != OBJT_DEFAULT && 1712 object->type != OBJT_SWAP) || 1713 (object->flags & OBJ_DEAD)) { 1714 VM_OBJECT_WUNLOCK(backing_object); 1715 break; 1716 } 1717 1718 if ( 1719 object->paging_in_progress != 0 || 1720 backing_object->paging_in_progress != 0 1721 ) { 1722 vm_object_qcollapse(object); 1723 VM_OBJECT_WUNLOCK(backing_object); 1724 break; 1725 } 1726 /* 1727 * We know that we can either collapse the backing object (if 1728 * the parent is the only reference to it) or (perhaps) have 1729 * the parent bypass the object if the parent happens to shadow 1730 * all the resident pages in the entire backing object. 1731 * 1732 * This is ignoring pager-backed pages such as swap pages. 1733 * vm_object_backing_scan fails the shadowing test in this 1734 * case. 1735 */ 1736 if (backing_object->ref_count == 1) { 1737 /* 1738 * If there is exactly one reference to the backing 1739 * object, we can collapse it into the parent. 1740 */ 1741 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1742 1743#if VM_NRESERVLEVEL > 0 1744 /* 1745 * Break any reservations from backing_object. 1746 */ 1747 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1748 vm_reserv_break_all(backing_object); 1749#endif 1750 1751 /* 1752 * Move the pager from backing_object to object. 1753 */ 1754 if (backing_object->type == OBJT_SWAP) { 1755 /* 1756 * swap_pager_copy() can sleep, in which case 1757 * the backing_object's and object's locks are 1758 * released and reacquired. 1759 * Since swap_pager_copy() is being asked to 1760 * destroy the source, it will change the 1761 * backing_object's type to OBJT_DEFAULT. 1762 */ 1763 swap_pager_copy( 1764 backing_object, 1765 object, 1766 OFF_TO_IDX(object->backing_object_offset), TRUE); 1767 1768 /* 1769 * Free any cached pages from backing_object. 1770 */ 1771 if (__predict_false( 1772 !vm_object_cache_is_empty(backing_object))) 1773 vm_page_cache_free(backing_object, 0, 0); 1774 } 1775 /* 1776 * Object now shadows whatever backing_object did. 1777 * Note that the reference to 1778 * backing_object->backing_object moves from within 1779 * backing_object to within object. 1780 */ 1781 LIST_REMOVE(object, shadow_list); 1782 backing_object->shadow_count--; 1783 if (backing_object->backing_object) { 1784 VM_OBJECT_WLOCK(backing_object->backing_object); 1785 LIST_REMOVE(backing_object, shadow_list); 1786 LIST_INSERT_HEAD( 1787 &backing_object->backing_object->shadow_head, 1788 object, shadow_list); 1789 /* 1790 * The shadow_count has not changed. 1791 */ 1792 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1793 } 1794 object->backing_object = backing_object->backing_object; 1795 object->backing_object_offset += 1796 backing_object->backing_object_offset; 1797 1798 /* 1799 * Discard backing_object. 1800 * 1801 * Since the backing object has no pages, no pager left, 1802 * and no object references within it, all that is 1803 * necessary is to dispose of it. 1804 */ 1805 KASSERT(backing_object->ref_count == 1, ( 1806"backing_object %p was somehow re-referenced during collapse!", 1807 backing_object)); 1808 VM_OBJECT_WUNLOCK(backing_object); 1809 vm_object_destroy(backing_object); 1810 1811 object_collapses++; 1812 } else { 1813 vm_object_t new_backing_object; 1814 1815 /* 1816 * If we do not entirely shadow the backing object, 1817 * there is nothing we can do so we give up. 1818 */ 1819 if (object->resident_page_count != object->size && 1820 vm_object_backing_scan(object, 1821 OBSC_TEST_ALL_SHADOWED) == 0) { 1822 VM_OBJECT_WUNLOCK(backing_object); 1823 break; 1824 } 1825 1826 /* 1827 * Make the parent shadow the next object in the 1828 * chain. Deallocating backing_object will not remove 1829 * it, since its reference count is at least 2. 1830 */ 1831 LIST_REMOVE(object, shadow_list); 1832 backing_object->shadow_count--; 1833 1834 new_backing_object = backing_object->backing_object; 1835 if ((object->backing_object = new_backing_object) != NULL) { 1836 VM_OBJECT_WLOCK(new_backing_object); 1837 LIST_INSERT_HEAD( 1838 &new_backing_object->shadow_head, 1839 object, 1840 shadow_list 1841 ); 1842 new_backing_object->shadow_count++; 1843 vm_object_reference_locked(new_backing_object); 1844 VM_OBJECT_WUNLOCK(new_backing_object); 1845 object->backing_object_offset += 1846 backing_object->backing_object_offset; 1847 } 1848 1849 /* 1850 * Drop the reference count on backing_object. Since 1851 * its ref_count was at least 2, it will not vanish. 1852 */ 1853 backing_object->ref_count--; 1854 VM_OBJECT_WUNLOCK(backing_object); 1855 object_bypasses++; 1856 } 1857 1858 /* 1859 * Try again with this object's new backing object. 1860 */ 1861 } 1862} 1863 1864/* 1865 * vm_object_page_remove: 1866 * 1867 * For the given object, either frees or invalidates each of the 1868 * specified pages. In general, a page is freed. However, if a page is 1869 * wired for any reason other than the existence of a managed, wired 1870 * mapping, then it may be invalidated but not removed from the object. 1871 * Pages are specified by the given range ["start", "end") and the option 1872 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1873 * extends from "start" to the end of the object. If the option 1874 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1875 * specified range are affected. If the option OBJPR_NOTMAPPED is 1876 * specified, then the pages within the specified range must have no 1877 * mappings. Otherwise, if this option is not specified, any mappings to 1878 * the specified pages are removed before the pages are freed or 1879 * invalidated. 1880 * 1881 * In general, this operation should only be performed on objects that 1882 * contain managed pages. There are, however, two exceptions. First, it 1883 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1884 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1885 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1886 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1887 * 1888 * The object must be locked. 1889 */ 1890void 1891vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1892 int options) 1893{ 1894 vm_page_t p, next; 1895 int wirings; 1896 1897 VM_OBJECT_ASSERT_WLOCKED(object); 1898 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1899 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1900 ("vm_object_page_remove: illegal options for object %p", object)); 1901 if (object->resident_page_count == 0) 1902 goto skipmemq; 1903 vm_object_pip_add(object, 1); 1904again: 1905 p = vm_page_find_least(object, start); 1906 1907 /* 1908 * Here, the variable "p" is either (1) the page with the least pindex 1909 * greater than or equal to the parameter "start" or (2) NULL. 1910 */ 1911 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1912 next = TAILQ_NEXT(p, listq); 1913 1914 /* 1915 * If the page is wired for any reason besides the existence 1916 * of managed, wired mappings, then it cannot be freed. For 1917 * example, fictitious pages, which represent device memory, 1918 * are inherently wired and cannot be freed. They can, 1919 * however, be invalidated if the option OBJPR_CLEANONLY is 1920 * not specified. 1921 */ 1922 vm_page_lock(p); 1923 if (vm_page_xbusied(p)) { 1924 VM_OBJECT_WUNLOCK(object); 1925 vm_page_busy_sleep(p, "vmopax"); 1926 VM_OBJECT_WLOCK(object); 1927 goto again; 1928 } 1929 if ((wirings = p->wire_count) != 0 && 1930 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) { 1931 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) == 1932 0) { 1933 pmap_remove_all(p); 1934 /* Account for removal of wired mappings. */ 1935 if (wirings != 0) 1936 p->wire_count -= wirings; 1937 } 1938 if ((options & OBJPR_CLEANONLY) == 0) { 1939 p->valid = 0; 1940 vm_page_undirty(p); 1941 } 1942 goto next; 1943 } 1944 if (vm_page_busied(p)) { 1945 VM_OBJECT_WUNLOCK(object); 1946 vm_page_busy_sleep(p, "vmopar"); 1947 VM_OBJECT_WLOCK(object); 1948 goto again; 1949 } 1950 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1951 ("vm_object_page_remove: page %p is fictitious", p)); 1952 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1953 if ((options & OBJPR_NOTMAPPED) == 0) 1954 pmap_remove_write(p); 1955 if (p->dirty) 1956 goto next; 1957 } 1958 if ((options & OBJPR_NOTMAPPED) == 0) { 1959 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0) 1960 goto next; 1961 pmap_remove_all(p); 1962 /* Account for removal of wired mappings. */ 1963 if (wirings != 0) { 1964 KASSERT(p->wire_count == wirings, 1965 ("inconsistent wire count %d %d %p", 1966 p->wire_count, wirings, p)); 1967 p->wire_count = 0; 1968 atomic_subtract_int(&cnt.v_wire_count, 1); 1969 } 1970 } 1971 vm_page_free(p); 1972next: 1973 vm_page_unlock(p); 1974 } 1975 vm_object_pip_wakeup(object); 1976skipmemq: 1977 if (__predict_false(!vm_object_cache_is_empty(object))) 1978 vm_page_cache_free(object, start, end); 1979} 1980 1981/* 1982 * vm_object_page_cache: 1983 * 1984 * For the given object, attempt to move the specified clean 1985 * pages to the cache queue. If a page is wired for any reason, 1986 * then it will not be changed. Pages are specified by the given 1987 * range ["start", "end"). As a special case, if "end" is zero, 1988 * then the range extends from "start" to the end of the object. 1989 * Any mappings to the specified pages are removed before the 1990 * pages are moved to the cache queue. 1991 * 1992 * This operation should only be performed on objects that 1993 * contain non-fictitious, managed pages. 1994 * 1995 * The object must be locked. 1996 */ 1997void 1998vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1999{ 2000 struct mtx *mtx, *new_mtx; 2001 vm_page_t p, next; 2002 2003 VM_OBJECT_ASSERT_WLOCKED(object); 2004 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2005 ("vm_object_page_cache: illegal object %p", object)); 2006 if (object->resident_page_count == 0) 2007 return; 2008 p = vm_page_find_least(object, start); 2009 2010 /* 2011 * Here, the variable "p" is either (1) the page with the least pindex 2012 * greater than or equal to the parameter "start" or (2) NULL. 2013 */ 2014 mtx = NULL; 2015 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2016 next = TAILQ_NEXT(p, listq); 2017 2018 /* 2019 * Avoid releasing and reacquiring the same page lock. 2020 */ 2021 new_mtx = vm_page_lockptr(p); 2022 if (mtx != new_mtx) { 2023 if (mtx != NULL) 2024 mtx_unlock(mtx); 2025 mtx = new_mtx; 2026 mtx_lock(mtx); 2027 } 2028 vm_page_try_to_cache(p); 2029 } 2030 if (mtx != NULL) 2031 mtx_unlock(mtx); 2032} 2033 2034/* 2035 * Populate the specified range of the object with valid pages. Returns 2036 * TRUE if the range is successfully populated and FALSE otherwise. 2037 * 2038 * Note: This function should be optimized to pass a larger array of 2039 * pages to vm_pager_get_pages() before it is applied to a non- 2040 * OBJT_DEVICE object. 2041 * 2042 * The object must be locked. 2043 */ 2044boolean_t 2045vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2046{ 2047 vm_page_t m, ma[1]; 2048 vm_pindex_t pindex; 2049 int rv; 2050 2051 VM_OBJECT_ASSERT_WLOCKED(object); 2052 for (pindex = start; pindex < end; pindex++) { 2053 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2054 if (m->valid != VM_PAGE_BITS_ALL) { 2055 ma[0] = m; 2056 rv = vm_pager_get_pages(object, ma, 1, 0); 2057 m = vm_page_lookup(object, pindex); 2058 if (m == NULL) 2059 break; 2060 if (rv != VM_PAGER_OK) { 2061 vm_page_lock(m); 2062 vm_page_free(m); 2063 vm_page_unlock(m); 2064 break; 2065 } 2066 } 2067 /* 2068 * Keep "m" busy because a subsequent iteration may unlock 2069 * the object. 2070 */ 2071 } 2072 if (pindex > start) { 2073 m = vm_page_lookup(object, start); 2074 while (m != NULL && m->pindex < pindex) { 2075 vm_page_xunbusy(m); 2076 m = TAILQ_NEXT(m, listq); 2077 } 2078 } 2079 return (pindex == end); 2080} 2081 2082/* 2083 * Routine: vm_object_coalesce 2084 * Function: Coalesces two objects backing up adjoining 2085 * regions of memory into a single object. 2086 * 2087 * returns TRUE if objects were combined. 2088 * 2089 * NOTE: Only works at the moment if the second object is NULL - 2090 * if it's not, which object do we lock first? 2091 * 2092 * Parameters: 2093 * prev_object First object to coalesce 2094 * prev_offset Offset into prev_object 2095 * prev_size Size of reference to prev_object 2096 * next_size Size of reference to the second object 2097 * reserved Indicator that extension region has 2098 * swap accounted for 2099 * 2100 * Conditions: 2101 * The object must *not* be locked. 2102 */ 2103boolean_t 2104vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2105 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2106{ 2107 vm_pindex_t next_pindex; 2108 2109 if (prev_object == NULL) 2110 return (TRUE); 2111 VM_OBJECT_WLOCK(prev_object); 2112 if ((prev_object->type != OBJT_DEFAULT && 2113 prev_object->type != OBJT_SWAP) || 2114 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2115 VM_OBJECT_WUNLOCK(prev_object); 2116 return (FALSE); 2117 } 2118 2119 /* 2120 * Try to collapse the object first 2121 */ 2122 vm_object_collapse(prev_object); 2123 2124 /* 2125 * Can't coalesce if: . more than one reference . paged out . shadows 2126 * another object . has a copy elsewhere (any of which mean that the 2127 * pages not mapped to prev_entry may be in use anyway) 2128 */ 2129 if (prev_object->backing_object != NULL) { 2130 VM_OBJECT_WUNLOCK(prev_object); 2131 return (FALSE); 2132 } 2133 2134 prev_size >>= PAGE_SHIFT; 2135 next_size >>= PAGE_SHIFT; 2136 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2137 2138 if ((prev_object->ref_count > 1) && 2139 (prev_object->size != next_pindex)) { 2140 VM_OBJECT_WUNLOCK(prev_object); 2141 return (FALSE); 2142 } 2143 2144 /* 2145 * Account for the charge. 2146 */ 2147 if (prev_object->cred != NULL) { 2148 2149 /* 2150 * If prev_object was charged, then this mapping, 2151 * althought not charged now, may become writable 2152 * later. Non-NULL cred in the object would prevent 2153 * swap reservation during enabling of the write 2154 * access, so reserve swap now. Failed reservation 2155 * cause allocation of the separate object for the map 2156 * entry, and swap reservation for this entry is 2157 * managed in appropriate time. 2158 */ 2159 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2160 prev_object->cred)) { 2161 return (FALSE); 2162 } 2163 prev_object->charge += ptoa(next_size); 2164 } 2165 2166 /* 2167 * Remove any pages that may still be in the object from a previous 2168 * deallocation. 2169 */ 2170 if (next_pindex < prev_object->size) { 2171 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2172 next_size, 0); 2173 if (prev_object->type == OBJT_SWAP) 2174 swap_pager_freespace(prev_object, 2175 next_pindex, next_size); 2176#if 0 2177 if (prev_object->cred != NULL) { 2178 KASSERT(prev_object->charge >= 2179 ptoa(prev_object->size - next_pindex), 2180 ("object %p overcharged 1 %jx %jx", prev_object, 2181 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2182 prev_object->charge -= ptoa(prev_object->size - 2183 next_pindex); 2184 } 2185#endif 2186 } 2187 2188 /* 2189 * Extend the object if necessary. 2190 */ 2191 if (next_pindex + next_size > prev_object->size) 2192 prev_object->size = next_pindex + next_size; 2193 2194 VM_OBJECT_WUNLOCK(prev_object); 2195 return (TRUE); 2196} 2197 2198void 2199vm_object_set_writeable_dirty(vm_object_t object) 2200{ 2201 2202 VM_OBJECT_ASSERT_WLOCKED(object); 2203 if (object->type != OBJT_VNODE) 2204 return; 2205 object->generation++; 2206 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2207 return; 2208 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2209} 2210 2211/* 2212 * vm_object_unwire: 2213 * 2214 * For each page offset within the specified range of the given object, 2215 * find the highest-level page in the shadow chain and unwire it. A page 2216 * must exist at every page offset, and the highest-level page must be 2217 * wired. 2218 */ 2219void 2220vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2221 uint8_t queue) 2222{ 2223 vm_object_t tobject; 2224 vm_page_t m, tm; 2225 vm_pindex_t end_pindex, pindex, tpindex; 2226 int depth, locked_depth; 2227 2228 KASSERT((offset & PAGE_MASK) == 0, 2229 ("vm_object_unwire: offset is not page aligned")); 2230 KASSERT((length & PAGE_MASK) == 0, 2231 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2232 /* The wired count of a fictitious page never changes. */ 2233 if ((object->flags & OBJ_FICTITIOUS) != 0) 2234 return; 2235 pindex = OFF_TO_IDX(offset); 2236 end_pindex = pindex + atop(length); 2237 locked_depth = 1; 2238 VM_OBJECT_RLOCK(object); 2239 m = vm_page_find_least(object, pindex); 2240 while (pindex < end_pindex) { 2241 if (m == NULL || pindex < m->pindex) { 2242 /* 2243 * The first object in the shadow chain doesn't 2244 * contain a page at the current index. Therefore, 2245 * the page must exist in a backing object. 2246 */ 2247 tobject = object; 2248 tpindex = pindex; 2249 depth = 0; 2250 do { 2251 tpindex += 2252 OFF_TO_IDX(tobject->backing_object_offset); 2253 tobject = tobject->backing_object; 2254 KASSERT(tobject != NULL, 2255 ("vm_object_unwire: missing page")); 2256 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2257 goto next_page; 2258 depth++; 2259 if (depth == locked_depth) { 2260 locked_depth++; 2261 VM_OBJECT_RLOCK(tobject); 2262 } 2263 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2264 NULL); 2265 } else { 2266 tm = m; 2267 m = TAILQ_NEXT(m, listq); 2268 } 2269 vm_page_lock(tm); 2270 vm_page_unwire(tm, queue); 2271 vm_page_unlock(tm); 2272next_page: 2273 pindex++; 2274 } 2275 /* Release the accumulated object locks. */ 2276 for (depth = 0; depth < locked_depth; depth++) { 2277 tobject = object->backing_object; 2278 VM_OBJECT_RUNLOCK(object); 2279 object = tobject; 2280 } 2281} 2282 2283#include "opt_ddb.h" 2284#ifdef DDB 2285#include <sys/kernel.h> 2286 2287#include <sys/cons.h> 2288 2289#include <ddb/ddb.h> 2290 2291static int 2292_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2293{ 2294 vm_map_t tmpm; 2295 vm_map_entry_t tmpe; 2296 vm_object_t obj; 2297 int entcount; 2298 2299 if (map == 0) 2300 return 0; 2301 2302 if (entry == 0) { 2303 tmpe = map->header.next; 2304 entcount = map->nentries; 2305 while (entcount-- && (tmpe != &map->header)) { 2306 if (_vm_object_in_map(map, object, tmpe)) { 2307 return 1; 2308 } 2309 tmpe = tmpe->next; 2310 } 2311 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2312 tmpm = entry->object.sub_map; 2313 tmpe = tmpm->header.next; 2314 entcount = tmpm->nentries; 2315 while (entcount-- && tmpe != &tmpm->header) { 2316 if (_vm_object_in_map(tmpm, object, tmpe)) { 2317 return 1; 2318 } 2319 tmpe = tmpe->next; 2320 } 2321 } else if ((obj = entry->object.vm_object) != NULL) { 2322 for (; obj; obj = obj->backing_object) 2323 if (obj == object) { 2324 return 1; 2325 } 2326 } 2327 return 0; 2328} 2329 2330static int 2331vm_object_in_map(vm_object_t object) 2332{ 2333 struct proc *p; 2334 2335 /* sx_slock(&allproc_lock); */ 2336 FOREACH_PROC_IN_SYSTEM(p) { 2337 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2338 continue; 2339 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2340 /* sx_sunlock(&allproc_lock); */ 2341 return 1; 2342 } 2343 } 2344 /* sx_sunlock(&allproc_lock); */ 2345 if (_vm_object_in_map(kernel_map, object, 0)) 2346 return 1; 2347 return 0; 2348} 2349 2350DB_SHOW_COMMAND(vmochk, vm_object_check) 2351{ 2352 vm_object_t object; 2353 2354 /* 2355 * make sure that internal objs are in a map somewhere 2356 * and none have zero ref counts. 2357 */ 2358 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2359 if (object->handle == NULL && 2360 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2361 if (object->ref_count == 0) { 2362 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2363 (long)object->size); 2364 } 2365 if (!vm_object_in_map(object)) { 2366 db_printf( 2367 "vmochk: internal obj is not in a map: " 2368 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2369 object->ref_count, (u_long)object->size, 2370 (u_long)object->size, 2371 (void *)object->backing_object); 2372 } 2373 } 2374 } 2375} 2376 2377/* 2378 * vm_object_print: [ debug ] 2379 */ 2380DB_SHOW_COMMAND(object, vm_object_print_static) 2381{ 2382 /* XXX convert args. */ 2383 vm_object_t object = (vm_object_t)addr; 2384 boolean_t full = have_addr; 2385 2386 vm_page_t p; 2387 2388 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2389#define count was_count 2390 2391 int count; 2392 2393 if (object == NULL) 2394 return; 2395 2396 db_iprintf( 2397 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2398 object, (int)object->type, (uintmax_t)object->size, 2399 object->resident_page_count, object->ref_count, object->flags, 2400 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2401 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2402 object->shadow_count, 2403 object->backing_object ? object->backing_object->ref_count : 0, 2404 object->backing_object, (uintmax_t)object->backing_object_offset); 2405 2406 if (!full) 2407 return; 2408 2409 db_indent += 2; 2410 count = 0; 2411 TAILQ_FOREACH(p, &object->memq, listq) { 2412 if (count == 0) 2413 db_iprintf("memory:="); 2414 else if (count == 6) { 2415 db_printf("\n"); 2416 db_iprintf(" ..."); 2417 count = 0; 2418 } else 2419 db_printf(","); 2420 count++; 2421 2422 db_printf("(off=0x%jx,page=0x%jx)", 2423 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2424 } 2425 if (count != 0) 2426 db_printf("\n"); 2427 db_indent -= 2; 2428} 2429 2430/* XXX. */ 2431#undef count 2432 2433/* XXX need this non-static entry for calling from vm_map_print. */ 2434void 2435vm_object_print( 2436 /* db_expr_t */ long addr, 2437 boolean_t have_addr, 2438 /* db_expr_t */ long count, 2439 char *modif) 2440{ 2441 vm_object_print_static(addr, have_addr, count, modif); 2442} 2443 2444DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2445{ 2446 vm_object_t object; 2447 vm_pindex_t fidx; 2448 vm_paddr_t pa; 2449 vm_page_t m, prev_m; 2450 int rcount, nl, c; 2451 2452 nl = 0; 2453 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2454 db_printf("new object: %p\n", (void *)object); 2455 if (nl > 18) { 2456 c = cngetc(); 2457 if (c != ' ') 2458 return; 2459 nl = 0; 2460 } 2461 nl++; 2462 rcount = 0; 2463 fidx = 0; 2464 pa = -1; 2465 TAILQ_FOREACH(m, &object->memq, listq) { 2466 if (m->pindex > 128) 2467 break; 2468 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2469 prev_m->pindex + 1 != m->pindex) { 2470 if (rcount) { 2471 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2472 (long)fidx, rcount, (long)pa); 2473 if (nl > 18) { 2474 c = cngetc(); 2475 if (c != ' ') 2476 return; 2477 nl = 0; 2478 } 2479 nl++; 2480 rcount = 0; 2481 } 2482 } 2483 if (rcount && 2484 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2485 ++rcount; 2486 continue; 2487 } 2488 if (rcount) { 2489 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2490 (long)fidx, rcount, (long)pa); 2491 if (nl > 18) { 2492 c = cngetc(); 2493 if (c != ' ') 2494 return; 2495 nl = 0; 2496 } 2497 nl++; 2498 } 2499 fidx = m->pindex; 2500 pa = VM_PAGE_TO_PHYS(m); 2501 rcount = 1; 2502 } 2503 if (rcount) { 2504 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2505 (long)fidx, rcount, (long)pa); 2506 if (nl > 18) { 2507 c = cngetc(); 2508 if (c != ' ') 2509 return; 2510 nl = 0; 2511 } 2512 nl++; 2513 } 2514 } 2515} 2516#endif /* DDB */ 2517