vm_object.c revision 288453
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 288453 2015-10-01 17:09:20Z jhb $"); 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/user.h> 83#include <sys/vnode.h> 84#include <sys/vmmeter.h> 85#include <sys/sx.h> 86 87#include <vm/vm.h> 88#include <vm/vm_param.h> 89#include <vm/pmap.h> 90#include <vm/vm_map.h> 91#include <vm/vm_object.h> 92#include <vm/vm_page.h> 93#include <vm/vm_pageout.h> 94#include <vm/vm_pager.h> 95#include <vm/swap_pager.h> 96#include <vm/vm_kern.h> 97#include <vm/vm_extern.h> 98#include <vm/vm_radix.h> 99#include <vm/vm_reserv.h> 100#include <vm/uma.h> 101 102static int old_msync; 103SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 104 "Use old (insecure) msync behavior"); 105 106static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 107 int pagerflags, int flags, boolean_t *clearobjflags, 108 boolean_t *eio); 109static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 110 boolean_t *clearobjflags); 111static void vm_object_qcollapse(vm_object_t object); 112static void vm_object_vndeallocate(vm_object_t object); 113 114/* 115 * Virtual memory objects maintain the actual data 116 * associated with allocated virtual memory. A given 117 * page of memory exists within exactly one object. 118 * 119 * An object is only deallocated when all "references" 120 * are given up. Only one "reference" to a given 121 * region of an object should be writeable. 122 * 123 * Associated with each object is a list of all resident 124 * memory pages belonging to that object; this list is 125 * maintained by the "vm_page" module, and locked by the object's 126 * lock. 127 * 128 * Each object also records a "pager" routine which is 129 * used to retrieve (and store) pages to the proper backing 130 * storage. In addition, objects may be backed by other 131 * objects from which they were virtual-copied. 132 * 133 * The only items within the object structure which are 134 * modified after time of creation are: 135 * reference count locked by object's lock 136 * pager routine locked by object's lock 137 * 138 */ 139 140struct object_q vm_object_list; 141struct mtx vm_object_list_mtx; /* lock for object list and count */ 142 143struct vm_object kernel_object_store; 144struct vm_object kmem_object_store; 145 146static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 147 "VM object stats"); 148 149static long object_collapses; 150SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 151 &object_collapses, 0, "VM object collapses"); 152 153static long object_bypasses; 154SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 155 &object_bypasses, 0, "VM object bypasses"); 156 157static uma_zone_t obj_zone; 158 159static int vm_object_zinit(void *mem, int size, int flags); 160 161#ifdef INVARIANTS 162static void vm_object_zdtor(void *mem, int size, void *arg); 163 164static void 165vm_object_zdtor(void *mem, int size, void *arg) 166{ 167 vm_object_t object; 168 169 object = (vm_object_t)mem; 170 KASSERT(object->ref_count == 0, 171 ("object %p ref_count = %d", object, object->ref_count)); 172 KASSERT(TAILQ_EMPTY(&object->memq), 173 ("object %p has resident pages in its memq", object)); 174 KASSERT(vm_radix_is_empty(&object->rtree), 175 ("object %p has resident pages in its trie", object)); 176#if VM_NRESERVLEVEL > 0 177 KASSERT(LIST_EMPTY(&object->rvq), 178 ("object %p has reservations", 179 object)); 180#endif 181 KASSERT(vm_object_cache_is_empty(object), 182 ("object %p has cached pages", 183 object)); 184 KASSERT(object->paging_in_progress == 0, 185 ("object %p paging_in_progress = %d", 186 object, object->paging_in_progress)); 187 KASSERT(object->resident_page_count == 0, 188 ("object %p resident_page_count = %d", 189 object, object->resident_page_count)); 190 KASSERT(object->shadow_count == 0, 191 ("object %p shadow_count = %d", 192 object, object->shadow_count)); 193 KASSERT(object->type == OBJT_DEAD, 194 ("object %p has non-dead type %d", 195 object, object->type)); 196} 197#endif 198 199static int 200vm_object_zinit(void *mem, int size, int flags) 201{ 202 vm_object_t object; 203 204 object = (vm_object_t)mem; 205 bzero(&object->lock, sizeof(object->lock)); 206 rw_init_flags(&object->lock, "vm object", RW_DUPOK); 207 208 /* These are true for any object that has been freed */ 209 object->type = OBJT_DEAD; 210 object->ref_count = 0; 211 object->rtree.rt_root = 0; 212 object->rtree.rt_flags = 0; 213 object->paging_in_progress = 0; 214 object->resident_page_count = 0; 215 object->shadow_count = 0; 216 object->cache.rt_root = 0; 217 object->cache.rt_flags = 0; 218 219 mtx_lock(&vm_object_list_mtx); 220 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 221 mtx_unlock(&vm_object_list_mtx); 222 return (0); 223} 224 225static void 226_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 227{ 228 229 TAILQ_INIT(&object->memq); 230 LIST_INIT(&object->shadow_head); 231 232 object->type = type; 233 switch (type) { 234 case OBJT_DEAD: 235 panic("_vm_object_allocate: can't create OBJT_DEAD"); 236 case OBJT_DEFAULT: 237 case OBJT_SWAP: 238 object->flags = OBJ_ONEMAPPING; 239 break; 240 case OBJT_DEVICE: 241 case OBJT_SG: 242 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 243 break; 244 case OBJT_MGTDEVICE: 245 object->flags = OBJ_FICTITIOUS; 246 break; 247 case OBJT_PHYS: 248 object->flags = OBJ_UNMANAGED; 249 break; 250 case OBJT_VNODE: 251 object->flags = 0; 252 break; 253 default: 254 panic("_vm_object_allocate: type %d is undefined", type); 255 } 256 object->size = size; 257 object->generation = 1; 258 object->ref_count = 1; 259 object->memattr = VM_MEMATTR_DEFAULT; 260 object->cred = NULL; 261 object->charge = 0; 262 object->handle = NULL; 263 object->backing_object = NULL; 264 object->backing_object_offset = (vm_ooffset_t) 0; 265#if VM_NRESERVLEVEL > 0 266 LIST_INIT(&object->rvq); 267#endif 268} 269 270/* 271 * vm_object_init: 272 * 273 * Initialize the VM objects module. 274 */ 275void 276vm_object_init(void) 277{ 278 TAILQ_INIT(&vm_object_list); 279 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 280 281 rw_init(&kernel_object->lock, "kernel vm object"); 282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 283 kernel_object); 284#if VM_NRESERVLEVEL > 0 285 kernel_object->flags |= OBJ_COLORED; 286 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 287#endif 288 289 rw_init(&kmem_object->lock, "kmem vm object"); 290 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 291 kmem_object); 292#if VM_NRESERVLEVEL > 0 293 kmem_object->flags |= OBJ_COLORED; 294 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 295#endif 296 297 /* 298 * The lock portion of struct vm_object must be type stable due 299 * to vm_pageout_fallback_object_lock locking a vm object 300 * without holding any references to it. 301 */ 302 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 303#ifdef INVARIANTS 304 vm_object_zdtor, 305#else 306 NULL, 307#endif 308 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 309 310 vm_radix_init(); 311} 312 313void 314vm_object_clear_flag(vm_object_t object, u_short bits) 315{ 316 317 VM_OBJECT_ASSERT_WLOCKED(object); 318 object->flags &= ~bits; 319} 320 321/* 322 * Sets the default memory attribute for the specified object. Pages 323 * that are allocated to this object are by default assigned this memory 324 * attribute. 325 * 326 * Presently, this function must be called before any pages are allocated 327 * to the object. In the future, this requirement may be relaxed for 328 * "default" and "swap" objects. 329 */ 330int 331vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 332{ 333 334 VM_OBJECT_ASSERT_WLOCKED(object); 335 switch (object->type) { 336 case OBJT_DEFAULT: 337 case OBJT_DEVICE: 338 case OBJT_MGTDEVICE: 339 case OBJT_PHYS: 340 case OBJT_SG: 341 case OBJT_SWAP: 342 case OBJT_VNODE: 343 if (!TAILQ_EMPTY(&object->memq)) 344 return (KERN_FAILURE); 345 break; 346 case OBJT_DEAD: 347 return (KERN_INVALID_ARGUMENT); 348 default: 349 panic("vm_object_set_memattr: object %p is of undefined type", 350 object); 351 } 352 object->memattr = memattr; 353 return (KERN_SUCCESS); 354} 355 356void 357vm_object_pip_add(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_subtract(vm_object_t object, short i) 366{ 367 368 VM_OBJECT_ASSERT_WLOCKED(object); 369 object->paging_in_progress -= i; 370} 371 372void 373vm_object_pip_wakeup(vm_object_t object) 374{ 375 376 VM_OBJECT_ASSERT_WLOCKED(object); 377 object->paging_in_progress--; 378 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 379 vm_object_clear_flag(object, OBJ_PIPWNT); 380 wakeup(object); 381 } 382} 383 384void 385vm_object_pip_wakeupn(vm_object_t object, short i) 386{ 387 388 VM_OBJECT_ASSERT_WLOCKED(object); 389 if (i) 390 object->paging_in_progress -= i; 391 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 392 vm_object_clear_flag(object, OBJ_PIPWNT); 393 wakeup(object); 394 } 395} 396 397void 398vm_object_pip_wait(vm_object_t object, char *waitid) 399{ 400 401 VM_OBJECT_ASSERT_WLOCKED(object); 402 while (object->paging_in_progress) { 403 object->flags |= OBJ_PIPWNT; 404 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 405 } 406} 407 408/* 409 * vm_object_allocate: 410 * 411 * Returns a new object with the given size. 412 */ 413vm_object_t 414vm_object_allocate(objtype_t type, vm_pindex_t size) 415{ 416 vm_object_t object; 417 418 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 419 _vm_object_allocate(type, size, object); 420 return (object); 421} 422 423 424/* 425 * vm_object_reference: 426 * 427 * Gets another reference to the given object. Note: OBJ_DEAD 428 * objects can be referenced during final cleaning. 429 */ 430void 431vm_object_reference(vm_object_t object) 432{ 433 if (object == NULL) 434 return; 435 VM_OBJECT_WLOCK(object); 436 vm_object_reference_locked(object); 437 VM_OBJECT_WUNLOCK(object); 438} 439 440/* 441 * vm_object_reference_locked: 442 * 443 * Gets another reference to the given object. 444 * 445 * The object must be locked. 446 */ 447void 448vm_object_reference_locked(vm_object_t object) 449{ 450 struct vnode *vp; 451 452 VM_OBJECT_ASSERT_WLOCKED(object); 453 object->ref_count++; 454 if (object->type == OBJT_VNODE) { 455 vp = object->handle; 456 vref(vp); 457 } 458} 459 460/* 461 * Handle deallocating an object of type OBJT_VNODE. 462 */ 463static void 464vm_object_vndeallocate(vm_object_t object) 465{ 466 struct vnode *vp = (struct vnode *) object->handle; 467 468 VM_OBJECT_ASSERT_WLOCKED(object); 469 KASSERT(object->type == OBJT_VNODE, 470 ("vm_object_vndeallocate: not a vnode object")); 471 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 472#ifdef INVARIANTS 473 if (object->ref_count == 0) { 474 vprint("vm_object_vndeallocate", vp); 475 panic("vm_object_vndeallocate: bad object reference count"); 476 } 477#endif 478 479 /* 480 * The test for text of vp vnode does not need a bypass to 481 * reach right VV_TEXT there, since it is obtained from 482 * object->handle. 483 */ 484 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 485 object->ref_count--; 486 VM_OBJECT_WUNLOCK(object); 487 /* vrele may need the vnode lock. */ 488 vrele(vp); 489 } else { 490 vhold(vp); 491 VM_OBJECT_WUNLOCK(object); 492 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 493 vdrop(vp); 494 VM_OBJECT_WLOCK(object); 495 object->ref_count--; 496 if (object->type == OBJT_DEAD) { 497 VM_OBJECT_WUNLOCK(object); 498 VOP_UNLOCK(vp, 0); 499 } else { 500 if (object->ref_count == 0) 501 VOP_UNSET_TEXT(vp); 502 VM_OBJECT_WUNLOCK(object); 503 vput(vp); 504 } 505 } 506} 507 508/* 509 * vm_object_deallocate: 510 * 511 * Release a reference to the specified object, 512 * gained either through a vm_object_allocate 513 * or a vm_object_reference call. When all references 514 * are gone, storage associated with this object 515 * may be relinquished. 516 * 517 * No object may be locked. 518 */ 519void 520vm_object_deallocate(vm_object_t object) 521{ 522 vm_object_t temp; 523 struct vnode *vp; 524 525 while (object != NULL) { 526 VM_OBJECT_WLOCK(object); 527 if (object->type == OBJT_VNODE) { 528 vm_object_vndeallocate(object); 529 return; 530 } 531 532 KASSERT(object->ref_count != 0, 533 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 534 535 /* 536 * If the reference count goes to 0 we start calling 537 * vm_object_terminate() on the object chain. 538 * A ref count of 1 may be a special case depending on the 539 * shadow count being 0 or 1. 540 */ 541 object->ref_count--; 542 if (object->ref_count > 1) { 543 VM_OBJECT_WUNLOCK(object); 544 return; 545 } else if (object->ref_count == 1) { 546 if (object->type == OBJT_SWAP && 547 (object->flags & OBJ_TMPFS) != 0) { 548 vp = object->un_pager.swp.swp_tmpfs; 549 vhold(vp); 550 VM_OBJECT_WUNLOCK(object); 551 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 552 VM_OBJECT_WLOCK(object); 553 if (object->type == OBJT_DEAD || 554 object->ref_count != 1) { 555 VM_OBJECT_WUNLOCK(object); 556 VOP_UNLOCK(vp, 0); 557 vdrop(vp); 558 return; 559 } 560 if ((object->flags & OBJ_TMPFS) != 0) 561 VOP_UNSET_TEXT(vp); 562 VOP_UNLOCK(vp, 0); 563 vdrop(vp); 564 } 565 if (object->shadow_count == 0 && 566 object->handle == NULL && 567 (object->type == OBJT_DEFAULT || 568 (object->type == OBJT_SWAP && 569 (object->flags & OBJ_TMPFS_NODE) == 0))) { 570 vm_object_set_flag(object, OBJ_ONEMAPPING); 571 } else if ((object->shadow_count == 1) && 572 (object->handle == NULL) && 573 (object->type == OBJT_DEFAULT || 574 object->type == OBJT_SWAP)) { 575 vm_object_t robject; 576 577 robject = LIST_FIRST(&object->shadow_head); 578 KASSERT(robject != NULL, 579 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 580 object->ref_count, 581 object->shadow_count)); 582 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 583 ("shadowed tmpfs v_object %p", object)); 584 if (!VM_OBJECT_TRYWLOCK(robject)) { 585 /* 586 * Avoid a potential deadlock. 587 */ 588 object->ref_count++; 589 VM_OBJECT_WUNLOCK(object); 590 /* 591 * More likely than not the thread 592 * holding robject's lock has lower 593 * priority than the current thread. 594 * Let the lower priority thread run. 595 */ 596 pause("vmo_de", 1); 597 continue; 598 } 599 /* 600 * Collapse object into its shadow unless its 601 * shadow is dead. In that case, object will 602 * be deallocated by the thread that is 603 * deallocating its shadow. 604 */ 605 if ((robject->flags & OBJ_DEAD) == 0 && 606 (robject->handle == NULL) && 607 (robject->type == OBJT_DEFAULT || 608 robject->type == OBJT_SWAP)) { 609 610 robject->ref_count++; 611retry: 612 if (robject->paging_in_progress) { 613 VM_OBJECT_WUNLOCK(object); 614 vm_object_pip_wait(robject, 615 "objde1"); 616 temp = robject->backing_object; 617 if (object == temp) { 618 VM_OBJECT_WLOCK(object); 619 goto retry; 620 } 621 } else if (object->paging_in_progress) { 622 VM_OBJECT_WUNLOCK(robject); 623 object->flags |= OBJ_PIPWNT; 624 VM_OBJECT_SLEEP(object, object, 625 PDROP | PVM, "objde2", 0); 626 VM_OBJECT_WLOCK(robject); 627 temp = robject->backing_object; 628 if (object == temp) { 629 VM_OBJECT_WLOCK(object); 630 goto retry; 631 } 632 } else 633 VM_OBJECT_WUNLOCK(object); 634 635 if (robject->ref_count == 1) { 636 robject->ref_count--; 637 object = robject; 638 goto doterm; 639 } 640 object = robject; 641 vm_object_collapse(object); 642 VM_OBJECT_WUNLOCK(object); 643 continue; 644 } 645 VM_OBJECT_WUNLOCK(robject); 646 } 647 VM_OBJECT_WUNLOCK(object); 648 return; 649 } 650doterm: 651 temp = object->backing_object; 652 if (temp != NULL) { 653 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 654 ("shadowed tmpfs v_object 2 %p", object)); 655 VM_OBJECT_WLOCK(temp); 656 LIST_REMOVE(object, shadow_list); 657 temp->shadow_count--; 658 VM_OBJECT_WUNLOCK(temp); 659 object->backing_object = NULL; 660 } 661 /* 662 * Don't double-terminate, we could be in a termination 663 * recursion due to the terminate having to sync data 664 * to disk. 665 */ 666 if ((object->flags & OBJ_DEAD) == 0) 667 vm_object_terminate(object); 668 else 669 VM_OBJECT_WUNLOCK(object); 670 object = temp; 671 } 672} 673 674/* 675 * vm_object_destroy removes the object from the global object list 676 * and frees the space for the object. 677 */ 678void 679vm_object_destroy(vm_object_t object) 680{ 681 682 /* 683 * Release the allocation charge. 684 */ 685 if (object->cred != NULL) { 686 swap_release_by_cred(object->charge, object->cred); 687 object->charge = 0; 688 crfree(object->cred); 689 object->cred = NULL; 690 } 691 692 /* 693 * Free the space for the object. 694 */ 695 uma_zfree(obj_zone, object); 696} 697 698/* 699 * vm_object_terminate actually destroys the specified object, freeing 700 * up all previously used resources. 701 * 702 * The object must be locked. 703 * This routine may block. 704 */ 705void 706vm_object_terminate(vm_object_t object) 707{ 708 vm_page_t p, p_next; 709 710 VM_OBJECT_ASSERT_WLOCKED(object); 711 712 /* 713 * Make sure no one uses us. 714 */ 715 vm_object_set_flag(object, OBJ_DEAD); 716 717 /* 718 * wait for the pageout daemon to be done with the object 719 */ 720 vm_object_pip_wait(object, "objtrm"); 721 722 KASSERT(!object->paging_in_progress, 723 ("vm_object_terminate: pageout in progress")); 724 725 /* 726 * Clean and free the pages, as appropriate. All references to the 727 * object are gone, so we don't need to lock it. 728 */ 729 if (object->type == OBJT_VNODE) { 730 struct vnode *vp = (struct vnode *)object->handle; 731 732 /* 733 * Clean pages and flush buffers. 734 */ 735 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 736 VM_OBJECT_WUNLOCK(object); 737 738 vinvalbuf(vp, V_SAVE, 0, 0); 739 740 VM_OBJECT_WLOCK(object); 741 } 742 743 KASSERT(object->ref_count == 0, 744 ("vm_object_terminate: object with references, ref_count=%d", 745 object->ref_count)); 746 747 /* 748 * Free any remaining pageable pages. This also removes them from the 749 * paging queues. However, don't free wired pages, just remove them 750 * from the object. Rather than incrementally removing each page from 751 * the object, the page and object are reset to any empty state. 752 */ 753 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 754 vm_page_assert_unbusied(p); 755 vm_page_lock(p); 756 /* 757 * Optimize the page's removal from the object by resetting 758 * its "object" field. Specifically, if the page is not 759 * wired, then the effect of this assignment is that 760 * vm_page_free()'s call to vm_page_remove() will return 761 * immediately without modifying the page or the object. 762 */ 763 p->object = NULL; 764 if (p->wire_count == 0) { 765 vm_page_free(p); 766 PCPU_INC(cnt.v_pfree); 767 } 768 vm_page_unlock(p); 769 } 770 /* 771 * If the object contained any pages, then reset it to an empty state. 772 * None of the object's fields, including "resident_page_count", were 773 * modified by the preceding loop. 774 */ 775 if (object->resident_page_count != 0) { 776 vm_radix_reclaim_allnodes(&object->rtree); 777 TAILQ_INIT(&object->memq); 778 object->resident_page_count = 0; 779 if (object->type == OBJT_VNODE) 780 vdrop(object->handle); 781 } 782 783#if VM_NRESERVLEVEL > 0 784 if (__predict_false(!LIST_EMPTY(&object->rvq))) 785 vm_reserv_break_all(object); 786#endif 787 if (__predict_false(!vm_object_cache_is_empty(object))) 788 vm_page_cache_free(object, 0, 0); 789 790 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 791 object->type == OBJT_SWAP, 792 ("%s: non-swap obj %p has cred", __func__, object)); 793 794 /* 795 * Let the pager know object is dead. 796 */ 797 vm_pager_deallocate(object); 798 VM_OBJECT_WUNLOCK(object); 799 800 vm_object_destroy(object); 801} 802 803/* 804 * Make the page read-only so that we can clear the object flags. However, if 805 * this is a nosync mmap then the object is likely to stay dirty so do not 806 * mess with the page and do not clear the object flags. Returns TRUE if the 807 * page should be flushed, and FALSE otherwise. 808 */ 809static boolean_t 810vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 811{ 812 813 /* 814 * If we have been asked to skip nosync pages and this is a 815 * nosync page, skip it. Note that the object flags were not 816 * cleared in this case so we do not have to set them. 817 */ 818 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 819 *clearobjflags = FALSE; 820 return (FALSE); 821 } else { 822 pmap_remove_write(p); 823 return (p->dirty != 0); 824 } 825} 826 827/* 828 * vm_object_page_clean 829 * 830 * Clean all dirty pages in the specified range of object. Leaves page 831 * on whatever queue it is currently on. If NOSYNC is set then do not 832 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 833 * leaving the object dirty. 834 * 835 * When stuffing pages asynchronously, allow clustering. XXX we need a 836 * synchronous clustering mode implementation. 837 * 838 * Odd semantics: if start == end, we clean everything. 839 * 840 * The object must be locked. 841 * 842 * Returns FALSE if some page from the range was not written, as 843 * reported by the pager, and TRUE otherwise. 844 */ 845boolean_t 846vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 847 int flags) 848{ 849 vm_page_t np, p; 850 vm_pindex_t pi, tend, tstart; 851 int curgeneration, n, pagerflags; 852 boolean_t clearobjflags, eio, res; 853 854 VM_OBJECT_ASSERT_WLOCKED(object); 855 856 /* 857 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 858 * objects. The check below prevents the function from 859 * operating on non-vnode objects. 860 */ 861 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 862 object->resident_page_count == 0) 863 return (TRUE); 864 865 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 866 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 867 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 868 869 tstart = OFF_TO_IDX(start); 870 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 871 clearobjflags = tstart == 0 && tend >= object->size; 872 res = TRUE; 873 874rescan: 875 curgeneration = object->generation; 876 877 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 878 pi = p->pindex; 879 if (pi >= tend) 880 break; 881 np = TAILQ_NEXT(p, listq); 882 if (p->valid == 0) 883 continue; 884 if (vm_page_sleep_if_busy(p, "vpcwai")) { 885 if (object->generation != curgeneration) { 886 if ((flags & OBJPC_SYNC) != 0) 887 goto rescan; 888 else 889 clearobjflags = FALSE; 890 } 891 np = vm_page_find_least(object, pi); 892 continue; 893 } 894 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 895 continue; 896 897 n = vm_object_page_collect_flush(object, p, pagerflags, 898 flags, &clearobjflags, &eio); 899 if (eio) { 900 res = FALSE; 901 clearobjflags = FALSE; 902 } 903 if (object->generation != curgeneration) { 904 if ((flags & OBJPC_SYNC) != 0) 905 goto rescan; 906 else 907 clearobjflags = FALSE; 908 } 909 910 /* 911 * If the VOP_PUTPAGES() did a truncated write, so 912 * that even the first page of the run is not fully 913 * written, vm_pageout_flush() returns 0 as the run 914 * length. Since the condition that caused truncated 915 * write may be permanent, e.g. exhausted free space, 916 * accepting n == 0 would cause an infinite loop. 917 * 918 * Forwarding the iterator leaves the unwritten page 919 * behind, but there is not much we can do there if 920 * filesystem refuses to write it. 921 */ 922 if (n == 0) { 923 n = 1; 924 clearobjflags = FALSE; 925 } 926 np = vm_page_find_least(object, pi + n); 927 } 928#if 0 929 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 930#endif 931 932 if (clearobjflags) 933 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 934 return (res); 935} 936 937static int 938vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 939 int flags, boolean_t *clearobjflags, boolean_t *eio) 940{ 941 vm_page_t ma[vm_pageout_page_count], p_first, tp; 942 int count, i, mreq, runlen; 943 944 vm_page_lock_assert(p, MA_NOTOWNED); 945 VM_OBJECT_ASSERT_WLOCKED(object); 946 947 count = 1; 948 mreq = 0; 949 950 for (tp = p; count < vm_pageout_page_count; count++) { 951 tp = vm_page_next(tp); 952 if (tp == NULL || vm_page_busied(tp)) 953 break; 954 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 955 break; 956 } 957 958 for (p_first = p; count < vm_pageout_page_count; count++) { 959 tp = vm_page_prev(p_first); 960 if (tp == NULL || vm_page_busied(tp)) 961 break; 962 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 963 break; 964 p_first = tp; 965 mreq++; 966 } 967 968 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 969 ma[i] = tp; 970 971 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 972 return (runlen); 973} 974 975/* 976 * Note that there is absolutely no sense in writing out 977 * anonymous objects, so we track down the vnode object 978 * to write out. 979 * We invalidate (remove) all pages from the address space 980 * for semantic correctness. 981 * 982 * If the backing object is a device object with unmanaged pages, then any 983 * mappings to the specified range of pages must be removed before this 984 * function is called. 985 * 986 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 987 * may start out with a NULL object. 988 */ 989boolean_t 990vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 991 boolean_t syncio, boolean_t invalidate) 992{ 993 vm_object_t backing_object; 994 struct vnode *vp; 995 struct mount *mp; 996 int error, flags, fsync_after; 997 boolean_t res; 998 999 if (object == NULL) 1000 return (TRUE); 1001 res = TRUE; 1002 error = 0; 1003 VM_OBJECT_WLOCK(object); 1004 while ((backing_object = object->backing_object) != NULL) { 1005 VM_OBJECT_WLOCK(backing_object); 1006 offset += object->backing_object_offset; 1007 VM_OBJECT_WUNLOCK(object); 1008 object = backing_object; 1009 if (object->size < OFF_TO_IDX(offset + size)) 1010 size = IDX_TO_OFF(object->size) - offset; 1011 } 1012 /* 1013 * Flush pages if writing is allowed, invalidate them 1014 * if invalidation requested. Pages undergoing I/O 1015 * will be ignored by vm_object_page_remove(). 1016 * 1017 * We cannot lock the vnode and then wait for paging 1018 * to complete without deadlocking against vm_fault. 1019 * Instead we simply call vm_object_page_remove() and 1020 * allow it to block internally on a page-by-page 1021 * basis when it encounters pages undergoing async 1022 * I/O. 1023 */ 1024 if (object->type == OBJT_VNODE && 1025 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1026 vp = object->handle; 1027 VM_OBJECT_WUNLOCK(object); 1028 (void) vn_start_write(vp, &mp, V_WAIT); 1029 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1030 if (syncio && !invalidate && offset == 0 && 1031 OFF_TO_IDX(size) == object->size) { 1032 /* 1033 * If syncing the whole mapping of the file, 1034 * it is faster to schedule all the writes in 1035 * async mode, also allowing the clustering, 1036 * and then wait for i/o to complete. 1037 */ 1038 flags = 0; 1039 fsync_after = TRUE; 1040 } else { 1041 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1042 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1043 fsync_after = FALSE; 1044 } 1045 VM_OBJECT_WLOCK(object); 1046 res = vm_object_page_clean(object, offset, offset + size, 1047 flags); 1048 VM_OBJECT_WUNLOCK(object); 1049 if (fsync_after) 1050 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1051 VOP_UNLOCK(vp, 0); 1052 vn_finished_write(mp); 1053 if (error != 0) 1054 res = FALSE; 1055 VM_OBJECT_WLOCK(object); 1056 } 1057 if ((object->type == OBJT_VNODE || 1058 object->type == OBJT_DEVICE) && invalidate) { 1059 if (object->type == OBJT_DEVICE) 1060 /* 1061 * The option OBJPR_NOTMAPPED must be passed here 1062 * because vm_object_page_remove() cannot remove 1063 * unmanaged mappings. 1064 */ 1065 flags = OBJPR_NOTMAPPED; 1066 else if (old_msync) 1067 flags = 0; 1068 else 1069 flags = OBJPR_CLEANONLY; 1070 vm_object_page_remove(object, OFF_TO_IDX(offset), 1071 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1072 } 1073 VM_OBJECT_WUNLOCK(object); 1074 return (res); 1075} 1076 1077/* 1078 * vm_object_madvise: 1079 * 1080 * Implements the madvise function at the object/page level. 1081 * 1082 * MADV_WILLNEED (any object) 1083 * 1084 * Activate the specified pages if they are resident. 1085 * 1086 * MADV_DONTNEED (any object) 1087 * 1088 * Deactivate the specified pages if they are resident. 1089 * 1090 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1091 * OBJ_ONEMAPPING only) 1092 * 1093 * Deactivate and clean the specified pages if they are 1094 * resident. This permits the process to reuse the pages 1095 * without faulting or the kernel to reclaim the pages 1096 * without I/O. 1097 */ 1098void 1099vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1100 int advise) 1101{ 1102 vm_pindex_t tpindex; 1103 vm_object_t backing_object, tobject; 1104 vm_page_t m; 1105 1106 if (object == NULL) 1107 return; 1108 VM_OBJECT_WLOCK(object); 1109 /* 1110 * Locate and adjust resident pages 1111 */ 1112 for (; pindex < end; pindex += 1) { 1113relookup: 1114 tobject = object; 1115 tpindex = pindex; 1116shadowlookup: 1117 /* 1118 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1119 * and those pages must be OBJ_ONEMAPPING. 1120 */ 1121 if (advise == MADV_FREE) { 1122 if ((tobject->type != OBJT_DEFAULT && 1123 tobject->type != OBJT_SWAP) || 1124 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1125 goto unlock_tobject; 1126 } 1127 } else if ((tobject->flags & OBJ_UNMANAGED) != 0) 1128 goto unlock_tobject; 1129 m = vm_page_lookup(tobject, tpindex); 1130 if (m == NULL && advise == MADV_WILLNEED) { 1131 /* 1132 * If the page is cached, reactivate it. 1133 */ 1134 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1135 VM_ALLOC_NOBUSY); 1136 } 1137 if (m == NULL) { 1138 /* 1139 * There may be swap even if there is no backing page 1140 */ 1141 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1142 swap_pager_freespace(tobject, tpindex, 1); 1143 /* 1144 * next object 1145 */ 1146 backing_object = tobject->backing_object; 1147 if (backing_object == NULL) 1148 goto unlock_tobject; 1149 VM_OBJECT_WLOCK(backing_object); 1150 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1151 if (tobject != object) 1152 VM_OBJECT_WUNLOCK(tobject); 1153 tobject = backing_object; 1154 goto shadowlookup; 1155 } else if (m->valid != VM_PAGE_BITS_ALL) 1156 goto unlock_tobject; 1157 /* 1158 * If the page is not in a normal state, skip it. 1159 */ 1160 vm_page_lock(m); 1161 if (m->hold_count != 0 || m->wire_count != 0) { 1162 vm_page_unlock(m); 1163 goto unlock_tobject; 1164 } 1165 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1166 ("vm_object_madvise: page %p is fictitious", m)); 1167 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1168 ("vm_object_madvise: page %p is not managed", m)); 1169 if (vm_page_busied(m)) { 1170 if (advise == MADV_WILLNEED) { 1171 /* 1172 * Reference the page before unlocking and 1173 * sleeping so that the page daemon is less 1174 * likely to reclaim it. 1175 */ 1176 vm_page_aflag_set(m, PGA_REFERENCED); 1177 } 1178 if (object != tobject) 1179 VM_OBJECT_WUNLOCK(object); 1180 VM_OBJECT_WUNLOCK(tobject); 1181 vm_page_busy_sleep(m, "madvpo"); 1182 VM_OBJECT_WLOCK(object); 1183 goto relookup; 1184 } 1185 if (advise == MADV_WILLNEED) { 1186 vm_page_activate(m); 1187 } else { 1188 vm_page_advise(m, advise); 1189 } 1190 vm_page_unlock(m); 1191 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1192 swap_pager_freespace(tobject, tpindex, 1); 1193unlock_tobject: 1194 if (tobject != object) 1195 VM_OBJECT_WUNLOCK(tobject); 1196 } 1197 VM_OBJECT_WUNLOCK(object); 1198} 1199 1200/* 1201 * vm_object_shadow: 1202 * 1203 * Create a new object which is backed by the 1204 * specified existing object range. The source 1205 * object reference is deallocated. 1206 * 1207 * The new object and offset into that object 1208 * are returned in the source parameters. 1209 */ 1210void 1211vm_object_shadow( 1212 vm_object_t *object, /* IN/OUT */ 1213 vm_ooffset_t *offset, /* IN/OUT */ 1214 vm_size_t length) 1215{ 1216 vm_object_t source; 1217 vm_object_t result; 1218 1219 source = *object; 1220 1221 /* 1222 * Don't create the new object if the old object isn't shared. 1223 */ 1224 if (source != NULL) { 1225 VM_OBJECT_WLOCK(source); 1226 if (source->ref_count == 1 && 1227 source->handle == NULL && 1228 (source->type == OBJT_DEFAULT || 1229 source->type == OBJT_SWAP)) { 1230 VM_OBJECT_WUNLOCK(source); 1231 return; 1232 } 1233 VM_OBJECT_WUNLOCK(source); 1234 } 1235 1236 /* 1237 * Allocate a new object with the given length. 1238 */ 1239 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1240 1241 /* 1242 * The new object shadows the source object, adding a reference to it. 1243 * Our caller changes his reference to point to the new object, 1244 * removing a reference to the source object. Net result: no change 1245 * of reference count. 1246 * 1247 * Try to optimize the result object's page color when shadowing 1248 * in order to maintain page coloring consistency in the combined 1249 * shadowed object. 1250 */ 1251 result->backing_object = source; 1252 /* 1253 * Store the offset into the source object, and fix up the offset into 1254 * the new object. 1255 */ 1256 result->backing_object_offset = *offset; 1257 if (source != NULL) { 1258 VM_OBJECT_WLOCK(source); 1259 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1260 source->shadow_count++; 1261#if VM_NRESERVLEVEL > 0 1262 result->flags |= source->flags & OBJ_COLORED; 1263 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1264 ((1 << (VM_NFREEORDER - 1)) - 1); 1265#endif 1266 VM_OBJECT_WUNLOCK(source); 1267 } 1268 1269 1270 /* 1271 * Return the new things 1272 */ 1273 *offset = 0; 1274 *object = result; 1275} 1276 1277/* 1278 * vm_object_split: 1279 * 1280 * Split the pages in a map entry into a new object. This affords 1281 * easier removal of unused pages, and keeps object inheritance from 1282 * being a negative impact on memory usage. 1283 */ 1284void 1285vm_object_split(vm_map_entry_t entry) 1286{ 1287 vm_page_t m, m_next; 1288 vm_object_t orig_object, new_object, source; 1289 vm_pindex_t idx, offidxstart; 1290 vm_size_t size; 1291 1292 orig_object = entry->object.vm_object; 1293 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1294 return; 1295 if (orig_object->ref_count <= 1) 1296 return; 1297 VM_OBJECT_WUNLOCK(orig_object); 1298 1299 offidxstart = OFF_TO_IDX(entry->offset); 1300 size = atop(entry->end - entry->start); 1301 1302 /* 1303 * If swap_pager_copy() is later called, it will convert new_object 1304 * into a swap object. 1305 */ 1306 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1307 1308 /* 1309 * At this point, the new object is still private, so the order in 1310 * which the original and new objects are locked does not matter. 1311 */ 1312 VM_OBJECT_WLOCK(new_object); 1313 VM_OBJECT_WLOCK(orig_object); 1314 source = orig_object->backing_object; 1315 if (source != NULL) { 1316 VM_OBJECT_WLOCK(source); 1317 if ((source->flags & OBJ_DEAD) != 0) { 1318 VM_OBJECT_WUNLOCK(source); 1319 VM_OBJECT_WUNLOCK(orig_object); 1320 VM_OBJECT_WUNLOCK(new_object); 1321 vm_object_deallocate(new_object); 1322 VM_OBJECT_WLOCK(orig_object); 1323 return; 1324 } 1325 LIST_INSERT_HEAD(&source->shadow_head, 1326 new_object, shadow_list); 1327 source->shadow_count++; 1328 vm_object_reference_locked(source); /* for new_object */ 1329 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1330 VM_OBJECT_WUNLOCK(source); 1331 new_object->backing_object_offset = 1332 orig_object->backing_object_offset + entry->offset; 1333 new_object->backing_object = source; 1334 } 1335 if (orig_object->cred != NULL) { 1336 new_object->cred = orig_object->cred; 1337 crhold(orig_object->cred); 1338 new_object->charge = ptoa(size); 1339 KASSERT(orig_object->charge >= ptoa(size), 1340 ("orig_object->charge < 0")); 1341 orig_object->charge -= ptoa(size); 1342 } 1343retry: 1344 m = vm_page_find_least(orig_object, offidxstart); 1345 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1346 m = m_next) { 1347 m_next = TAILQ_NEXT(m, listq); 1348 1349 /* 1350 * We must wait for pending I/O to complete before we can 1351 * rename the page. 1352 * 1353 * We do not have to VM_PROT_NONE the page as mappings should 1354 * not be changed by this operation. 1355 */ 1356 if (vm_page_busied(m)) { 1357 VM_OBJECT_WUNLOCK(new_object); 1358 vm_page_lock(m); 1359 VM_OBJECT_WUNLOCK(orig_object); 1360 vm_page_busy_sleep(m, "spltwt"); 1361 VM_OBJECT_WLOCK(orig_object); 1362 VM_OBJECT_WLOCK(new_object); 1363 goto retry; 1364 } 1365 1366 /* vm_page_rename() will handle dirty and cache. */ 1367 if (vm_page_rename(m, new_object, idx)) { 1368 VM_OBJECT_WUNLOCK(new_object); 1369 VM_OBJECT_WUNLOCK(orig_object); 1370 VM_WAIT; 1371 VM_OBJECT_WLOCK(orig_object); 1372 VM_OBJECT_WLOCK(new_object); 1373 goto retry; 1374 } 1375#if VM_NRESERVLEVEL > 0 1376 /* 1377 * If some of the reservation's allocated pages remain with 1378 * the original object, then transferring the reservation to 1379 * the new object is neither particularly beneficial nor 1380 * particularly harmful as compared to leaving the reservation 1381 * with the original object. If, however, all of the 1382 * reservation's allocated pages are transferred to the new 1383 * object, then transferring the reservation is typically 1384 * beneficial. Determining which of these two cases applies 1385 * would be more costly than unconditionally renaming the 1386 * reservation. 1387 */ 1388 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1389#endif 1390 if (orig_object->type == OBJT_SWAP) 1391 vm_page_xbusy(m); 1392 } 1393 if (orig_object->type == OBJT_SWAP) { 1394 /* 1395 * swap_pager_copy() can sleep, in which case the orig_object's 1396 * and new_object's locks are released and reacquired. 1397 */ 1398 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1399 TAILQ_FOREACH(m, &new_object->memq, listq) 1400 vm_page_xunbusy(m); 1401 1402 /* 1403 * Transfer any cached pages from orig_object to new_object. 1404 * If swap_pager_copy() found swapped out pages within the 1405 * specified range of orig_object, then it changed 1406 * new_object's type to OBJT_SWAP when it transferred those 1407 * pages to new_object. Otherwise, new_object's type 1408 * should still be OBJT_DEFAULT and orig_object should not 1409 * contain any cached pages within the specified range. 1410 */ 1411 if (__predict_false(!vm_object_cache_is_empty(orig_object))) 1412 vm_page_cache_transfer(orig_object, offidxstart, 1413 new_object); 1414 } 1415 VM_OBJECT_WUNLOCK(orig_object); 1416 VM_OBJECT_WUNLOCK(new_object); 1417 entry->object.vm_object = new_object; 1418 entry->offset = 0LL; 1419 vm_object_deallocate(orig_object); 1420 VM_OBJECT_WLOCK(new_object); 1421} 1422 1423#define OBSC_TEST_ALL_SHADOWED 0x0001 1424#define OBSC_COLLAPSE_NOWAIT 0x0002 1425#define OBSC_COLLAPSE_WAIT 0x0004 1426 1427static int 1428vm_object_backing_scan(vm_object_t object, int op) 1429{ 1430 int r = 1; 1431 vm_page_t p; 1432 vm_object_t backing_object; 1433 vm_pindex_t backing_offset_index; 1434 1435 VM_OBJECT_ASSERT_WLOCKED(object); 1436 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1437 1438 backing_object = object->backing_object; 1439 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1440 1441 /* 1442 * Initial conditions 1443 */ 1444 if (op & OBSC_TEST_ALL_SHADOWED) { 1445 /* 1446 * We do not want to have to test for the existence of cache 1447 * or swap pages in the backing object. XXX but with the 1448 * new swapper this would be pretty easy to do. 1449 * 1450 * XXX what about anonymous MAP_SHARED memory that hasn't 1451 * been ZFOD faulted yet? If we do not test for this, the 1452 * shadow test may succeed! XXX 1453 */ 1454 if (backing_object->type != OBJT_DEFAULT) { 1455 return (0); 1456 } 1457 } 1458 if (op & OBSC_COLLAPSE_WAIT) { 1459 vm_object_set_flag(backing_object, OBJ_DEAD); 1460 } 1461 1462 /* 1463 * Our scan 1464 */ 1465 p = TAILQ_FIRST(&backing_object->memq); 1466 while (p) { 1467 vm_page_t next = TAILQ_NEXT(p, listq); 1468 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1469 1470 if (op & OBSC_TEST_ALL_SHADOWED) { 1471 vm_page_t pp; 1472 1473 /* 1474 * Ignore pages outside the parent object's range 1475 * and outside the parent object's mapping of the 1476 * backing object. 1477 * 1478 * note that we do not busy the backing object's 1479 * page. 1480 */ 1481 if ( 1482 p->pindex < backing_offset_index || 1483 new_pindex >= object->size 1484 ) { 1485 p = next; 1486 continue; 1487 } 1488 1489 /* 1490 * See if the parent has the page or if the parent's 1491 * object pager has the page. If the parent has the 1492 * page but the page is not valid, the parent's 1493 * object pager must have the page. 1494 * 1495 * If this fails, the parent does not completely shadow 1496 * the object and we might as well give up now. 1497 */ 1498 1499 pp = vm_page_lookup(object, new_pindex); 1500 if ( 1501 (pp == NULL || pp->valid == 0) && 1502 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1503 ) { 1504 r = 0; 1505 break; 1506 } 1507 } 1508 1509 /* 1510 * Check for busy page 1511 */ 1512 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1513 vm_page_t pp; 1514 1515 if (op & OBSC_COLLAPSE_NOWAIT) { 1516 if (!p->valid || vm_page_busied(p)) { 1517 p = next; 1518 continue; 1519 } 1520 } else if (op & OBSC_COLLAPSE_WAIT) { 1521 if (vm_page_busied(p)) { 1522 VM_OBJECT_WUNLOCK(object); 1523 vm_page_lock(p); 1524 VM_OBJECT_WUNLOCK(backing_object); 1525 vm_page_busy_sleep(p, "vmocol"); 1526 VM_OBJECT_WLOCK(object); 1527 VM_OBJECT_WLOCK(backing_object); 1528 /* 1529 * If we slept, anything could have 1530 * happened. Since the object is 1531 * marked dead, the backing offset 1532 * should not have changed so we 1533 * just restart our scan. 1534 */ 1535 p = TAILQ_FIRST(&backing_object->memq); 1536 continue; 1537 } 1538 } 1539 1540 KASSERT( 1541 p->object == backing_object, 1542 ("vm_object_backing_scan: object mismatch") 1543 ); 1544 1545 if ( 1546 p->pindex < backing_offset_index || 1547 new_pindex >= object->size 1548 ) { 1549 if (backing_object->type == OBJT_SWAP) 1550 swap_pager_freespace(backing_object, 1551 p->pindex, 1); 1552 1553 /* 1554 * Page is out of the parent object's range, we 1555 * can simply destroy it. 1556 */ 1557 vm_page_lock(p); 1558 KASSERT(!pmap_page_is_mapped(p), 1559 ("freeing mapped page %p", p)); 1560 if (p->wire_count == 0) 1561 vm_page_free(p); 1562 else 1563 vm_page_remove(p); 1564 vm_page_unlock(p); 1565 p = next; 1566 continue; 1567 } 1568 1569 pp = vm_page_lookup(object, new_pindex); 1570 if ( 1571 (op & OBSC_COLLAPSE_NOWAIT) != 0 && 1572 (pp != NULL && pp->valid == 0) 1573 ) { 1574 if (backing_object->type == OBJT_SWAP) 1575 swap_pager_freespace(backing_object, 1576 p->pindex, 1); 1577 1578 /* 1579 * The page in the parent is not (yet) valid. 1580 * We don't know anything about the state of 1581 * the original page. It might be mapped, 1582 * so we must avoid the next if here. 1583 * 1584 * This is due to a race in vm_fault() where 1585 * we must unbusy the original (backing_obj) 1586 * page before we can (re)lock the parent. 1587 * Hence we can get here. 1588 */ 1589 p = next; 1590 continue; 1591 } 1592 if ( 1593 pp != NULL || 1594 vm_pager_has_page(object, new_pindex, NULL, NULL) 1595 ) { 1596 if (backing_object->type == OBJT_SWAP) 1597 swap_pager_freespace(backing_object, 1598 p->pindex, 1); 1599 1600 /* 1601 * page already exists in parent OR swap exists 1602 * for this location in the parent. Destroy 1603 * the original page from the backing object. 1604 * 1605 * Leave the parent's page alone 1606 */ 1607 vm_page_lock(p); 1608 KASSERT(!pmap_page_is_mapped(p), 1609 ("freeing mapped page %p", p)); 1610 if (p->wire_count == 0) 1611 vm_page_free(p); 1612 else 1613 vm_page_remove(p); 1614 vm_page_unlock(p); 1615 p = next; 1616 continue; 1617 } 1618 1619 /* 1620 * Page does not exist in parent, rename the 1621 * page from the backing object to the main object. 1622 * 1623 * If the page was mapped to a process, it can remain 1624 * mapped through the rename. 1625 * vm_page_rename() will handle dirty and cache. 1626 */ 1627 if (vm_page_rename(p, object, new_pindex)) { 1628 if (op & OBSC_COLLAPSE_NOWAIT) { 1629 p = next; 1630 continue; 1631 } 1632 VM_OBJECT_WUNLOCK(backing_object); 1633 VM_OBJECT_WUNLOCK(object); 1634 VM_WAIT; 1635 VM_OBJECT_WLOCK(object); 1636 VM_OBJECT_WLOCK(backing_object); 1637 p = TAILQ_FIRST(&backing_object->memq); 1638 continue; 1639 } 1640 1641 /* Use the old pindex to free the right page. */ 1642 if (backing_object->type == OBJT_SWAP) 1643 swap_pager_freespace(backing_object, 1644 new_pindex + backing_offset_index, 1); 1645 1646#if VM_NRESERVLEVEL > 0 1647 /* 1648 * Rename the reservation. 1649 */ 1650 vm_reserv_rename(p, object, backing_object, 1651 backing_offset_index); 1652#endif 1653 } 1654 p = next; 1655 } 1656 return (r); 1657} 1658 1659 1660/* 1661 * this version of collapse allows the operation to occur earlier and 1662 * when paging_in_progress is true for an object... This is not a complete 1663 * operation, but should plug 99.9% of the rest of the leaks. 1664 */ 1665static void 1666vm_object_qcollapse(vm_object_t object) 1667{ 1668 vm_object_t backing_object = object->backing_object; 1669 1670 VM_OBJECT_ASSERT_WLOCKED(object); 1671 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1672 1673 if (backing_object->ref_count != 1) 1674 return; 1675 1676 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1677} 1678 1679/* 1680 * vm_object_collapse: 1681 * 1682 * Collapse an object with the object backing it. 1683 * Pages in the backing object are moved into the 1684 * parent, and the backing object is deallocated. 1685 */ 1686void 1687vm_object_collapse(vm_object_t object) 1688{ 1689 VM_OBJECT_ASSERT_WLOCKED(object); 1690 1691 while (TRUE) { 1692 vm_object_t backing_object; 1693 1694 /* 1695 * Verify that the conditions are right for collapse: 1696 * 1697 * The object exists and the backing object exists. 1698 */ 1699 if ((backing_object = object->backing_object) == NULL) 1700 break; 1701 1702 /* 1703 * we check the backing object first, because it is most likely 1704 * not collapsable. 1705 */ 1706 VM_OBJECT_WLOCK(backing_object); 1707 if (backing_object->handle != NULL || 1708 (backing_object->type != OBJT_DEFAULT && 1709 backing_object->type != OBJT_SWAP) || 1710 (backing_object->flags & OBJ_DEAD) || 1711 object->handle != NULL || 1712 (object->type != OBJT_DEFAULT && 1713 object->type != OBJT_SWAP) || 1714 (object->flags & OBJ_DEAD)) { 1715 VM_OBJECT_WUNLOCK(backing_object); 1716 break; 1717 } 1718 1719 if ( 1720 object->paging_in_progress != 0 || 1721 backing_object->paging_in_progress != 0 1722 ) { 1723 vm_object_qcollapse(object); 1724 VM_OBJECT_WUNLOCK(backing_object); 1725 break; 1726 } 1727 /* 1728 * We know that we can either collapse the backing object (if 1729 * the parent is the only reference to it) or (perhaps) have 1730 * the parent bypass the object if the parent happens to shadow 1731 * all the resident pages in the entire backing object. 1732 * 1733 * This is ignoring pager-backed pages such as swap pages. 1734 * vm_object_backing_scan fails the shadowing test in this 1735 * case. 1736 */ 1737 if (backing_object->ref_count == 1) { 1738 /* 1739 * If there is exactly one reference to the backing 1740 * object, we can collapse it into the parent. 1741 */ 1742 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1743 1744#if VM_NRESERVLEVEL > 0 1745 /* 1746 * Break any reservations from backing_object. 1747 */ 1748 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1749 vm_reserv_break_all(backing_object); 1750#endif 1751 1752 /* 1753 * Move the pager from backing_object to object. 1754 */ 1755 if (backing_object->type == OBJT_SWAP) { 1756 /* 1757 * swap_pager_copy() can sleep, in which case 1758 * the backing_object's and object's locks are 1759 * released and reacquired. 1760 * Since swap_pager_copy() is being asked to 1761 * destroy the source, it will change the 1762 * backing_object's type to OBJT_DEFAULT. 1763 */ 1764 swap_pager_copy( 1765 backing_object, 1766 object, 1767 OFF_TO_IDX(object->backing_object_offset), TRUE); 1768 1769 /* 1770 * Free any cached pages from backing_object. 1771 */ 1772 if (__predict_false( 1773 !vm_object_cache_is_empty(backing_object))) 1774 vm_page_cache_free(backing_object, 0, 0); 1775 } 1776 /* 1777 * Object now shadows whatever backing_object did. 1778 * Note that the reference to 1779 * backing_object->backing_object moves from within 1780 * backing_object to within object. 1781 */ 1782 LIST_REMOVE(object, shadow_list); 1783 backing_object->shadow_count--; 1784 if (backing_object->backing_object) { 1785 VM_OBJECT_WLOCK(backing_object->backing_object); 1786 LIST_REMOVE(backing_object, shadow_list); 1787 LIST_INSERT_HEAD( 1788 &backing_object->backing_object->shadow_head, 1789 object, shadow_list); 1790 /* 1791 * The shadow_count has not changed. 1792 */ 1793 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1794 } 1795 object->backing_object = backing_object->backing_object; 1796 object->backing_object_offset += 1797 backing_object->backing_object_offset; 1798 1799 /* 1800 * Discard backing_object. 1801 * 1802 * Since the backing object has no pages, no pager left, 1803 * and no object references within it, all that is 1804 * necessary is to dispose of it. 1805 */ 1806 KASSERT(backing_object->ref_count == 1, ( 1807"backing_object %p was somehow re-referenced during collapse!", 1808 backing_object)); 1809 backing_object->type = OBJT_DEAD; 1810 backing_object->ref_count = 0; 1811 VM_OBJECT_WUNLOCK(backing_object); 1812 vm_object_destroy(backing_object); 1813 1814 object_collapses++; 1815 } else { 1816 vm_object_t new_backing_object; 1817 1818 /* 1819 * If we do not entirely shadow the backing object, 1820 * there is nothing we can do so we give up. 1821 */ 1822 if (object->resident_page_count != object->size && 1823 vm_object_backing_scan(object, 1824 OBSC_TEST_ALL_SHADOWED) == 0) { 1825 VM_OBJECT_WUNLOCK(backing_object); 1826 break; 1827 } 1828 1829 /* 1830 * Make the parent shadow the next object in the 1831 * chain. Deallocating backing_object will not remove 1832 * it, since its reference count is at least 2. 1833 */ 1834 LIST_REMOVE(object, shadow_list); 1835 backing_object->shadow_count--; 1836 1837 new_backing_object = backing_object->backing_object; 1838 if ((object->backing_object = new_backing_object) != NULL) { 1839 VM_OBJECT_WLOCK(new_backing_object); 1840 LIST_INSERT_HEAD( 1841 &new_backing_object->shadow_head, 1842 object, 1843 shadow_list 1844 ); 1845 new_backing_object->shadow_count++; 1846 vm_object_reference_locked(new_backing_object); 1847 VM_OBJECT_WUNLOCK(new_backing_object); 1848 object->backing_object_offset += 1849 backing_object->backing_object_offset; 1850 } 1851 1852 /* 1853 * Drop the reference count on backing_object. Since 1854 * its ref_count was at least 2, it will not vanish. 1855 */ 1856 backing_object->ref_count--; 1857 VM_OBJECT_WUNLOCK(backing_object); 1858 object_bypasses++; 1859 } 1860 1861 /* 1862 * Try again with this object's new backing object. 1863 */ 1864 } 1865} 1866 1867/* 1868 * vm_object_page_remove: 1869 * 1870 * For the given object, either frees or invalidates each of the 1871 * specified pages. In general, a page is freed. However, if a page is 1872 * wired for any reason other than the existence of a managed, wired 1873 * mapping, then it may be invalidated but not removed from the object. 1874 * Pages are specified by the given range ["start", "end") and the option 1875 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1876 * extends from "start" to the end of the object. If the option 1877 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1878 * specified range are affected. If the option OBJPR_NOTMAPPED is 1879 * specified, then the pages within the specified range must have no 1880 * mappings. Otherwise, if this option is not specified, any mappings to 1881 * the specified pages are removed before the pages are freed or 1882 * invalidated. 1883 * 1884 * In general, this operation should only be performed on objects that 1885 * contain managed pages. There are, however, two exceptions. First, it 1886 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1887 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1888 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1889 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1890 * 1891 * The object must be locked. 1892 */ 1893void 1894vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1895 int options) 1896{ 1897 vm_page_t p, next; 1898 1899 VM_OBJECT_ASSERT_WLOCKED(object); 1900 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1901 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1902 ("vm_object_page_remove: illegal options for object %p", object)); 1903 if (object->resident_page_count == 0) 1904 goto skipmemq; 1905 vm_object_pip_add(object, 1); 1906again: 1907 p = vm_page_find_least(object, start); 1908 1909 /* 1910 * Here, the variable "p" is either (1) the page with the least pindex 1911 * greater than or equal to the parameter "start" or (2) NULL. 1912 */ 1913 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1914 next = TAILQ_NEXT(p, listq); 1915 1916 /* 1917 * If the page is wired for any reason besides the existence 1918 * of managed, wired mappings, then it cannot be freed. For 1919 * example, fictitious pages, which represent device memory, 1920 * are inherently wired and cannot be freed. They can, 1921 * however, be invalidated if the option OBJPR_CLEANONLY is 1922 * not specified. 1923 */ 1924 vm_page_lock(p); 1925 if (vm_page_xbusied(p)) { 1926 VM_OBJECT_WUNLOCK(object); 1927 vm_page_busy_sleep(p, "vmopax"); 1928 VM_OBJECT_WLOCK(object); 1929 goto again; 1930 } 1931 if (p->wire_count != 0) { 1932 if ((options & OBJPR_NOTMAPPED) == 0) 1933 pmap_remove_all(p); 1934 if ((options & OBJPR_CLEANONLY) == 0) { 1935 p->valid = 0; 1936 vm_page_undirty(p); 1937 } 1938 goto next; 1939 } 1940 if (vm_page_busied(p)) { 1941 VM_OBJECT_WUNLOCK(object); 1942 vm_page_busy_sleep(p, "vmopar"); 1943 VM_OBJECT_WLOCK(object); 1944 goto again; 1945 } 1946 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1947 ("vm_object_page_remove: page %p is fictitious", p)); 1948 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1949 if ((options & OBJPR_NOTMAPPED) == 0) 1950 pmap_remove_write(p); 1951 if (p->dirty) 1952 goto next; 1953 } 1954 if ((options & OBJPR_NOTMAPPED) == 0) 1955 pmap_remove_all(p); 1956 vm_page_free(p); 1957next: 1958 vm_page_unlock(p); 1959 } 1960 vm_object_pip_wakeup(object); 1961skipmemq: 1962 if (__predict_false(!vm_object_cache_is_empty(object))) 1963 vm_page_cache_free(object, start, end); 1964} 1965 1966/* 1967 * vm_object_page_cache: 1968 * 1969 * For the given object, attempt to move the specified clean 1970 * pages to the cache queue. If a page is wired for any reason, 1971 * then it will not be changed. Pages are specified by the given 1972 * range ["start", "end"). As a special case, if "end" is zero, 1973 * then the range extends from "start" to the end of the object. 1974 * Any mappings to the specified pages are removed before the 1975 * pages are moved to the cache queue. 1976 * 1977 * This operation should only be performed on objects that 1978 * contain non-fictitious, managed pages. 1979 * 1980 * The object must be locked. 1981 */ 1982void 1983vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1984{ 1985 struct mtx *mtx, *new_mtx; 1986 vm_page_t p, next; 1987 1988 VM_OBJECT_ASSERT_WLOCKED(object); 1989 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 1990 ("vm_object_page_cache: illegal object %p", object)); 1991 if (object->resident_page_count == 0) 1992 return; 1993 p = vm_page_find_least(object, start); 1994 1995 /* 1996 * Here, the variable "p" is either (1) the page with the least pindex 1997 * greater than or equal to the parameter "start" or (2) NULL. 1998 */ 1999 mtx = NULL; 2000 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2001 next = TAILQ_NEXT(p, listq); 2002 2003 /* 2004 * Avoid releasing and reacquiring the same page lock. 2005 */ 2006 new_mtx = vm_page_lockptr(p); 2007 if (mtx != new_mtx) { 2008 if (mtx != NULL) 2009 mtx_unlock(mtx); 2010 mtx = new_mtx; 2011 mtx_lock(mtx); 2012 } 2013 vm_page_try_to_cache(p); 2014 } 2015 if (mtx != NULL) 2016 mtx_unlock(mtx); 2017} 2018 2019/* 2020 * Populate the specified range of the object with valid pages. Returns 2021 * TRUE if the range is successfully populated and FALSE otherwise. 2022 * 2023 * Note: This function should be optimized to pass a larger array of 2024 * pages to vm_pager_get_pages() before it is applied to a non- 2025 * OBJT_DEVICE object. 2026 * 2027 * The object must be locked. 2028 */ 2029boolean_t 2030vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2031{ 2032 vm_page_t m, ma[1]; 2033 vm_pindex_t pindex; 2034 int rv; 2035 2036 VM_OBJECT_ASSERT_WLOCKED(object); 2037 for (pindex = start; pindex < end; pindex++) { 2038 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2039 if (m->valid != VM_PAGE_BITS_ALL) { 2040 ma[0] = m; 2041 rv = vm_pager_get_pages(object, ma, 1, 0); 2042 m = vm_page_lookup(object, pindex); 2043 if (m == NULL) 2044 break; 2045 if (rv != VM_PAGER_OK) { 2046 vm_page_lock(m); 2047 vm_page_free(m); 2048 vm_page_unlock(m); 2049 break; 2050 } 2051 } 2052 /* 2053 * Keep "m" busy because a subsequent iteration may unlock 2054 * the object. 2055 */ 2056 } 2057 if (pindex > start) { 2058 m = vm_page_lookup(object, start); 2059 while (m != NULL && m->pindex < pindex) { 2060 vm_page_xunbusy(m); 2061 m = TAILQ_NEXT(m, listq); 2062 } 2063 } 2064 return (pindex == end); 2065} 2066 2067/* 2068 * Routine: vm_object_coalesce 2069 * Function: Coalesces two objects backing up adjoining 2070 * regions of memory into a single object. 2071 * 2072 * returns TRUE if objects were combined. 2073 * 2074 * NOTE: Only works at the moment if the second object is NULL - 2075 * if it's not, which object do we lock first? 2076 * 2077 * Parameters: 2078 * prev_object First object to coalesce 2079 * prev_offset Offset into prev_object 2080 * prev_size Size of reference to prev_object 2081 * next_size Size of reference to the second object 2082 * reserved Indicator that extension region has 2083 * swap accounted for 2084 * 2085 * Conditions: 2086 * The object must *not* be locked. 2087 */ 2088boolean_t 2089vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2090 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2091{ 2092 vm_pindex_t next_pindex; 2093 2094 if (prev_object == NULL) 2095 return (TRUE); 2096 VM_OBJECT_WLOCK(prev_object); 2097 if ((prev_object->type != OBJT_DEFAULT && 2098 prev_object->type != OBJT_SWAP) || 2099 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2100 VM_OBJECT_WUNLOCK(prev_object); 2101 return (FALSE); 2102 } 2103 2104 /* 2105 * Try to collapse the object first 2106 */ 2107 vm_object_collapse(prev_object); 2108 2109 /* 2110 * Can't coalesce if: . more than one reference . paged out . shadows 2111 * another object . has a copy elsewhere (any of which mean that the 2112 * pages not mapped to prev_entry may be in use anyway) 2113 */ 2114 if (prev_object->backing_object != NULL) { 2115 VM_OBJECT_WUNLOCK(prev_object); 2116 return (FALSE); 2117 } 2118 2119 prev_size >>= PAGE_SHIFT; 2120 next_size >>= PAGE_SHIFT; 2121 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2122 2123 if ((prev_object->ref_count > 1) && 2124 (prev_object->size != next_pindex)) { 2125 VM_OBJECT_WUNLOCK(prev_object); 2126 return (FALSE); 2127 } 2128 2129 /* 2130 * Account for the charge. 2131 */ 2132 if (prev_object->cred != NULL) { 2133 2134 /* 2135 * If prev_object was charged, then this mapping, 2136 * althought not charged now, may become writable 2137 * later. Non-NULL cred in the object would prevent 2138 * swap reservation during enabling of the write 2139 * access, so reserve swap now. Failed reservation 2140 * cause allocation of the separate object for the map 2141 * entry, and swap reservation for this entry is 2142 * managed in appropriate time. 2143 */ 2144 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2145 prev_object->cred)) { 2146 return (FALSE); 2147 } 2148 prev_object->charge += ptoa(next_size); 2149 } 2150 2151 /* 2152 * Remove any pages that may still be in the object from a previous 2153 * deallocation. 2154 */ 2155 if (next_pindex < prev_object->size) { 2156 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2157 next_size, 0); 2158 if (prev_object->type == OBJT_SWAP) 2159 swap_pager_freespace(prev_object, 2160 next_pindex, next_size); 2161#if 0 2162 if (prev_object->cred != NULL) { 2163 KASSERT(prev_object->charge >= 2164 ptoa(prev_object->size - next_pindex), 2165 ("object %p overcharged 1 %jx %jx", prev_object, 2166 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2167 prev_object->charge -= ptoa(prev_object->size - 2168 next_pindex); 2169 } 2170#endif 2171 } 2172 2173 /* 2174 * Extend the object if necessary. 2175 */ 2176 if (next_pindex + next_size > prev_object->size) 2177 prev_object->size = next_pindex + next_size; 2178 2179 VM_OBJECT_WUNLOCK(prev_object); 2180 return (TRUE); 2181} 2182 2183void 2184vm_object_set_writeable_dirty(vm_object_t object) 2185{ 2186 2187 VM_OBJECT_ASSERT_WLOCKED(object); 2188 if (object->type != OBJT_VNODE) { 2189 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2190 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2191 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2192 } 2193 return; 2194 } 2195 object->generation++; 2196 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2197 return; 2198 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2199} 2200 2201/* 2202 * vm_object_unwire: 2203 * 2204 * For each page offset within the specified range of the given object, 2205 * find the highest-level page in the shadow chain and unwire it. A page 2206 * must exist at every page offset, and the highest-level page must be 2207 * wired. 2208 */ 2209void 2210vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2211 uint8_t queue) 2212{ 2213 vm_object_t tobject; 2214 vm_page_t m, tm; 2215 vm_pindex_t end_pindex, pindex, tpindex; 2216 int depth, locked_depth; 2217 2218 KASSERT((offset & PAGE_MASK) == 0, 2219 ("vm_object_unwire: offset is not page aligned")); 2220 KASSERT((length & PAGE_MASK) == 0, 2221 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2222 /* The wired count of a fictitious page never changes. */ 2223 if ((object->flags & OBJ_FICTITIOUS) != 0) 2224 return; 2225 pindex = OFF_TO_IDX(offset); 2226 end_pindex = pindex + atop(length); 2227 locked_depth = 1; 2228 VM_OBJECT_RLOCK(object); 2229 m = vm_page_find_least(object, pindex); 2230 while (pindex < end_pindex) { 2231 if (m == NULL || pindex < m->pindex) { 2232 /* 2233 * The first object in the shadow chain doesn't 2234 * contain a page at the current index. Therefore, 2235 * the page must exist in a backing object. 2236 */ 2237 tobject = object; 2238 tpindex = pindex; 2239 depth = 0; 2240 do { 2241 tpindex += 2242 OFF_TO_IDX(tobject->backing_object_offset); 2243 tobject = tobject->backing_object; 2244 KASSERT(tobject != NULL, 2245 ("vm_object_unwire: missing page")); 2246 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2247 goto next_page; 2248 depth++; 2249 if (depth == locked_depth) { 2250 locked_depth++; 2251 VM_OBJECT_RLOCK(tobject); 2252 } 2253 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2254 NULL); 2255 } else { 2256 tm = m; 2257 m = TAILQ_NEXT(m, listq); 2258 } 2259 vm_page_lock(tm); 2260 vm_page_unwire(tm, queue); 2261 vm_page_unlock(tm); 2262next_page: 2263 pindex++; 2264 } 2265 /* Release the accumulated object locks. */ 2266 for (depth = 0; depth < locked_depth; depth++) { 2267 tobject = object->backing_object; 2268 VM_OBJECT_RUNLOCK(object); 2269 object = tobject; 2270 } 2271} 2272 2273static int 2274sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2275{ 2276 struct kinfo_vmobject kvo; 2277 char *fullpath, *freepath; 2278 struct vnode *vp; 2279 struct vattr va; 2280 vm_object_t obj; 2281 vm_page_t m; 2282 int count, error; 2283 2284 if (req->oldptr == NULL) { 2285 /* 2286 * If an old buffer has not been provided, generate an 2287 * estimate of the space needed for a subsequent call. 2288 */ 2289 mtx_lock(&vm_object_list_mtx); 2290 count = 0; 2291 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2292 if (obj->type == OBJT_DEAD) 2293 continue; 2294 count++; 2295 } 2296 mtx_unlock(&vm_object_list_mtx); 2297 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2298 count * 11 / 10)); 2299 } 2300 2301 error = 0; 2302 2303 /* 2304 * VM objects are type stable and are never removed from the 2305 * list once added. This allows us to safely read obj->object_list 2306 * after reacquiring the VM object lock. 2307 */ 2308 mtx_lock(&vm_object_list_mtx); 2309 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2310 if (obj->type == OBJT_DEAD) 2311 continue; 2312 VM_OBJECT_RLOCK(obj); 2313 if (obj->type == OBJT_DEAD) { 2314 VM_OBJECT_RUNLOCK(obj); 2315 continue; 2316 } 2317 mtx_unlock(&vm_object_list_mtx); 2318 kvo.kvo_size = ptoa(obj->size); 2319 kvo.kvo_resident = obj->resident_page_count; 2320 kvo.kvo_ref_count = obj->ref_count; 2321 kvo.kvo_shadow_count = obj->shadow_count; 2322 kvo.kvo_memattr = obj->memattr; 2323 kvo.kvo_active = 0; 2324 kvo.kvo_inactive = 0; 2325 TAILQ_FOREACH(m, &obj->memq, listq) { 2326 /* 2327 * A page may belong to the object but be 2328 * dequeued and set to PQ_NONE while the 2329 * object lock is not held. This makes the 2330 * reads of m->queue below racy, and we do not 2331 * count pages set to PQ_NONE. However, this 2332 * sysctl is only meant to give an 2333 * approximation of the system anyway. 2334 */ 2335 if (m->queue == PQ_ACTIVE) 2336 kvo.kvo_active++; 2337 else if (m->queue == PQ_INACTIVE) 2338 kvo.kvo_inactive++; 2339 } 2340 2341 kvo.kvo_vn_fileid = 0; 2342 kvo.kvo_vn_fsid = 0; 2343 freepath = NULL; 2344 fullpath = ""; 2345 vp = NULL; 2346 switch (obj->type) { 2347 case OBJT_DEFAULT: 2348 kvo.kvo_type = KVME_TYPE_DEFAULT; 2349 break; 2350 case OBJT_VNODE: 2351 kvo.kvo_type = KVME_TYPE_VNODE; 2352 vp = obj->handle; 2353 vref(vp); 2354 break; 2355 case OBJT_SWAP: 2356 kvo.kvo_type = KVME_TYPE_SWAP; 2357 break; 2358 case OBJT_DEVICE: 2359 kvo.kvo_type = KVME_TYPE_DEVICE; 2360 break; 2361 case OBJT_PHYS: 2362 kvo.kvo_type = KVME_TYPE_PHYS; 2363 break; 2364 case OBJT_DEAD: 2365 kvo.kvo_type = KVME_TYPE_DEAD; 2366 break; 2367 case OBJT_SG: 2368 kvo.kvo_type = KVME_TYPE_SG; 2369 break; 2370 case OBJT_MGTDEVICE: 2371 kvo.kvo_type = KVME_TYPE_MGTDEVICE; 2372 break; 2373 default: 2374 kvo.kvo_type = KVME_TYPE_UNKNOWN; 2375 break; 2376 } 2377 VM_OBJECT_RUNLOCK(obj); 2378 if (vp != NULL) { 2379 vn_fullpath(curthread, vp, &fullpath, &freepath); 2380 vn_lock(vp, LK_SHARED | LK_RETRY); 2381 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2382 kvo.kvo_vn_fileid = va.va_fileid; 2383 kvo.kvo_vn_fsid = va.va_fsid; 2384 } 2385 vput(vp); 2386 } 2387 2388 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path)); 2389 if (freepath != NULL) 2390 free(freepath, M_TEMP); 2391 2392 /* Pack record size down */ 2393 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) + 2394 strlen(kvo.kvo_path) + 1; 2395 kvo.kvo_structsize = roundup(kvo.kvo_structsize, 2396 sizeof(uint64_t)); 2397 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize); 2398 mtx_lock(&vm_object_list_mtx); 2399 if (error) 2400 break; 2401 } 2402 mtx_unlock(&vm_object_list_mtx); 2403 return (error); 2404} 2405SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2406 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2407 "List of VM objects"); 2408 2409#include "opt_ddb.h" 2410#ifdef DDB 2411#include <sys/kernel.h> 2412 2413#include <sys/cons.h> 2414 2415#include <ddb/ddb.h> 2416 2417static int 2418_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2419{ 2420 vm_map_t tmpm; 2421 vm_map_entry_t tmpe; 2422 vm_object_t obj; 2423 int entcount; 2424 2425 if (map == 0) 2426 return 0; 2427 2428 if (entry == 0) { 2429 tmpe = map->header.next; 2430 entcount = map->nentries; 2431 while (entcount-- && (tmpe != &map->header)) { 2432 if (_vm_object_in_map(map, object, tmpe)) { 2433 return 1; 2434 } 2435 tmpe = tmpe->next; 2436 } 2437 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2438 tmpm = entry->object.sub_map; 2439 tmpe = tmpm->header.next; 2440 entcount = tmpm->nentries; 2441 while (entcount-- && tmpe != &tmpm->header) { 2442 if (_vm_object_in_map(tmpm, object, tmpe)) { 2443 return 1; 2444 } 2445 tmpe = tmpe->next; 2446 } 2447 } else if ((obj = entry->object.vm_object) != NULL) { 2448 for (; obj; obj = obj->backing_object) 2449 if (obj == object) { 2450 return 1; 2451 } 2452 } 2453 return 0; 2454} 2455 2456static int 2457vm_object_in_map(vm_object_t object) 2458{ 2459 struct proc *p; 2460 2461 /* sx_slock(&allproc_lock); */ 2462 FOREACH_PROC_IN_SYSTEM(p) { 2463 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2464 continue; 2465 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2466 /* sx_sunlock(&allproc_lock); */ 2467 return 1; 2468 } 2469 } 2470 /* sx_sunlock(&allproc_lock); */ 2471 if (_vm_object_in_map(kernel_map, object, 0)) 2472 return 1; 2473 return 0; 2474} 2475 2476DB_SHOW_COMMAND(vmochk, vm_object_check) 2477{ 2478 vm_object_t object; 2479 2480 /* 2481 * make sure that internal objs are in a map somewhere 2482 * and none have zero ref counts. 2483 */ 2484 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2485 if (object->handle == NULL && 2486 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2487 if (object->ref_count == 0) { 2488 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2489 (long)object->size); 2490 } 2491 if (!vm_object_in_map(object)) { 2492 db_printf( 2493 "vmochk: internal obj is not in a map: " 2494 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2495 object->ref_count, (u_long)object->size, 2496 (u_long)object->size, 2497 (void *)object->backing_object); 2498 } 2499 } 2500 } 2501} 2502 2503/* 2504 * vm_object_print: [ debug ] 2505 */ 2506DB_SHOW_COMMAND(object, vm_object_print_static) 2507{ 2508 /* XXX convert args. */ 2509 vm_object_t object = (vm_object_t)addr; 2510 boolean_t full = have_addr; 2511 2512 vm_page_t p; 2513 2514 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2515#define count was_count 2516 2517 int count; 2518 2519 if (object == NULL) 2520 return; 2521 2522 db_iprintf( 2523 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2524 object, (int)object->type, (uintmax_t)object->size, 2525 object->resident_page_count, object->ref_count, object->flags, 2526 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2527 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2528 object->shadow_count, 2529 object->backing_object ? object->backing_object->ref_count : 0, 2530 object->backing_object, (uintmax_t)object->backing_object_offset); 2531 2532 if (!full) 2533 return; 2534 2535 db_indent += 2; 2536 count = 0; 2537 TAILQ_FOREACH(p, &object->memq, listq) { 2538 if (count == 0) 2539 db_iprintf("memory:="); 2540 else if (count == 6) { 2541 db_printf("\n"); 2542 db_iprintf(" ..."); 2543 count = 0; 2544 } else 2545 db_printf(","); 2546 count++; 2547 2548 db_printf("(off=0x%jx,page=0x%jx)", 2549 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2550 } 2551 if (count != 0) 2552 db_printf("\n"); 2553 db_indent -= 2; 2554} 2555 2556/* XXX. */ 2557#undef count 2558 2559/* XXX need this non-static entry for calling from vm_map_print. */ 2560void 2561vm_object_print( 2562 /* db_expr_t */ long addr, 2563 boolean_t have_addr, 2564 /* db_expr_t */ long count, 2565 char *modif) 2566{ 2567 vm_object_print_static(addr, have_addr, count, modif); 2568} 2569 2570DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2571{ 2572 vm_object_t object; 2573 vm_pindex_t fidx; 2574 vm_paddr_t pa; 2575 vm_page_t m, prev_m; 2576 int rcount, nl, c; 2577 2578 nl = 0; 2579 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2580 db_printf("new object: %p\n", (void *)object); 2581 if (nl > 18) { 2582 c = cngetc(); 2583 if (c != ' ') 2584 return; 2585 nl = 0; 2586 } 2587 nl++; 2588 rcount = 0; 2589 fidx = 0; 2590 pa = -1; 2591 TAILQ_FOREACH(m, &object->memq, listq) { 2592 if (m->pindex > 128) 2593 break; 2594 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2595 prev_m->pindex + 1 != m->pindex) { 2596 if (rcount) { 2597 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2598 (long)fidx, rcount, (long)pa); 2599 if (nl > 18) { 2600 c = cngetc(); 2601 if (c != ' ') 2602 return; 2603 nl = 0; 2604 } 2605 nl++; 2606 rcount = 0; 2607 } 2608 } 2609 if (rcount && 2610 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2611 ++rcount; 2612 continue; 2613 } 2614 if (rcount) { 2615 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2616 (long)fidx, rcount, (long)pa); 2617 if (nl > 18) { 2618 c = cngetc(); 2619 if (c != ' ') 2620 return; 2621 nl = 0; 2622 } 2623 nl++; 2624 } 2625 fidx = m->pindex; 2626 pa = VM_PAGE_TO_PHYS(m); 2627 rcount = 1; 2628 } 2629 if (rcount) { 2630 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2631 (long)fidx, rcount, (long)pa); 2632 if (nl > 18) { 2633 c = cngetc(); 2634 if (c != ' ') 2635 return; 2636 nl = 0; 2637 } 2638 nl++; 2639 } 2640 } 2641} 2642#endif /* DDB */ 2643