vm_object.c revision 330897
1/*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63/* 64 * Virtual memory object module. 65 */ 66 67#include <sys/cdefs.h> 68__FBSDID("$FreeBSD: stable/11/sys/vm/vm_object.c 330897 2018-03-14 03:19:51Z eadler $"); 69 70#include "opt_vm.h" 71 72#include <sys/param.h> 73#include <sys/systm.h> 74#include <sys/lock.h> 75#include <sys/mman.h> 76#include <sys/mount.h> 77#include <sys/kernel.h> 78#include <sys/pctrie.h> 79#include <sys/sysctl.h> 80#include <sys/mutex.h> 81#include <sys/proc.h> /* for curproc, pageproc */ 82#include <sys/socket.h> 83#include <sys/resourcevar.h> 84#include <sys/rwlock.h> 85#include <sys/user.h> 86#include <sys/vnode.h> 87#include <sys/vmmeter.h> 88#include <sys/sx.h> 89 90#include <vm/vm.h> 91#include <vm/vm_param.h> 92#include <vm/pmap.h> 93#include <vm/vm_map.h> 94#include <vm/vm_object.h> 95#include <vm/vm_page.h> 96#include <vm/vm_pageout.h> 97#include <vm/vm_pager.h> 98#include <vm/swap_pager.h> 99#include <vm/vm_kern.h> 100#include <vm/vm_extern.h> 101#include <vm/vm_radix.h> 102#include <vm/vm_reserv.h> 103#include <vm/uma.h> 104 105static int old_msync; 106SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 107 "Use old (insecure) msync behavior"); 108 109static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 110 int pagerflags, int flags, boolean_t *clearobjflags, 111 boolean_t *eio); 112static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 113 boolean_t *clearobjflags); 114static void vm_object_qcollapse(vm_object_t object); 115static void vm_object_vndeallocate(vm_object_t object); 116 117/* 118 * Virtual memory objects maintain the actual data 119 * associated with allocated virtual memory. A given 120 * page of memory exists within exactly one object. 121 * 122 * An object is only deallocated when all "references" 123 * are given up. Only one "reference" to a given 124 * region of an object should be writeable. 125 * 126 * Associated with each object is a list of all resident 127 * memory pages belonging to that object; this list is 128 * maintained by the "vm_page" module, and locked by the object's 129 * lock. 130 * 131 * Each object also records a "pager" routine which is 132 * used to retrieve (and store) pages to the proper backing 133 * storage. In addition, objects may be backed by other 134 * objects from which they were virtual-copied. 135 * 136 * The only items within the object structure which are 137 * modified after time of creation are: 138 * reference count locked by object's lock 139 * pager routine locked by object's lock 140 * 141 */ 142 143struct object_q vm_object_list; 144struct mtx vm_object_list_mtx; /* lock for object list and count */ 145 146struct vm_object kernel_object_store; 147struct vm_object kmem_object_store; 148 149static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 150 "VM object stats"); 151 152static long object_collapses; 153SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 154 &object_collapses, 0, "VM object collapses"); 155 156static long object_bypasses; 157SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 158 &object_bypasses, 0, "VM object bypasses"); 159 160static uma_zone_t obj_zone; 161 162static int vm_object_zinit(void *mem, int size, int flags); 163 164#ifdef INVARIANTS 165static void vm_object_zdtor(void *mem, int size, void *arg); 166 167static void 168vm_object_zdtor(void *mem, int size, void *arg) 169{ 170 vm_object_t object; 171 172 object = (vm_object_t)mem; 173 KASSERT(object->ref_count == 0, 174 ("object %p ref_count = %d", object, object->ref_count)); 175 KASSERT(TAILQ_EMPTY(&object->memq), 176 ("object %p has resident pages in its memq", object)); 177 KASSERT(vm_radix_is_empty(&object->rtree), 178 ("object %p has resident pages in its trie", object)); 179#if VM_NRESERVLEVEL > 0 180 KASSERT(LIST_EMPTY(&object->rvq), 181 ("object %p has reservations", 182 object)); 183#endif 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 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 206 207 /* These are true for any object that has been freed */ 208 object->type = OBJT_DEAD; 209 object->ref_count = 0; 210 vm_radix_init(&object->rtree); 211 object->paging_in_progress = 0; 212 object->resident_page_count = 0; 213 object->shadow_count = 0; 214 object->flags = OBJ_DEAD; 215 216 mtx_lock(&vm_object_list_mtx); 217 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 218 mtx_unlock(&vm_object_list_mtx); 219 return (0); 220} 221 222static void 223_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 224{ 225 226 TAILQ_INIT(&object->memq); 227 LIST_INIT(&object->shadow_head); 228 229 object->type = type; 230 if (type == OBJT_SWAP) 231 pctrie_init(&object->un_pager.swp.swp_blks); 232 233 /* 234 * Ensure that swap_pager_swapoff() iteration over object_list 235 * sees up to date type and pctrie head if it observed 236 * non-dead object. 237 */ 238 atomic_thread_fence_rel(); 239 240 switch (type) { 241 case OBJT_DEAD: 242 panic("_vm_object_allocate: can't create OBJT_DEAD"); 243 case OBJT_DEFAULT: 244 case OBJT_SWAP: 245 object->flags = OBJ_ONEMAPPING; 246 break; 247 case OBJT_DEVICE: 248 case OBJT_SG: 249 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 250 break; 251 case OBJT_MGTDEVICE: 252 object->flags = OBJ_FICTITIOUS; 253 break; 254 case OBJT_PHYS: 255 object->flags = OBJ_UNMANAGED; 256 break; 257 case OBJT_VNODE: 258 object->flags = 0; 259 break; 260 default: 261 panic("_vm_object_allocate: type %d is undefined", type); 262 } 263 object->size = size; 264 object->generation = 1; 265 object->ref_count = 1; 266 object->memattr = VM_MEMATTR_DEFAULT; 267 object->cred = NULL; 268 object->charge = 0; 269 object->handle = NULL; 270 object->backing_object = NULL; 271 object->backing_object_offset = (vm_ooffset_t) 0; 272#if VM_NRESERVLEVEL > 0 273 LIST_INIT(&object->rvq); 274#endif 275 umtx_shm_object_init(object); 276} 277 278/* 279 * vm_object_init: 280 * 281 * Initialize the VM objects module. 282 */ 283void 284vm_object_init(void) 285{ 286 TAILQ_INIT(&vm_object_list); 287 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 288 289 rw_init(&kernel_object->lock, "kernel vm object"); 290 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 291 VM_MIN_KERNEL_ADDRESS), kernel_object); 292#if VM_NRESERVLEVEL > 0 293 kernel_object->flags |= OBJ_COLORED; 294 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 295#endif 296 297 rw_init(&kmem_object->lock, "kmem vm object"); 298 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 299 VM_MIN_KERNEL_ADDRESS), kmem_object); 300#if VM_NRESERVLEVEL > 0 301 kmem_object->flags |= OBJ_COLORED; 302 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 303#endif 304 305 /* 306 * The lock portion of struct vm_object must be type stable due 307 * to vm_pageout_fallback_object_lock locking a vm object 308 * without holding any references to it. 309 */ 310 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 311#ifdef INVARIANTS 312 vm_object_zdtor, 313#else 314 NULL, 315#endif 316 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 317 318 vm_radix_zinit(); 319} 320 321void 322vm_object_clear_flag(vm_object_t object, u_short bits) 323{ 324 325 VM_OBJECT_ASSERT_WLOCKED(object); 326 object->flags &= ~bits; 327} 328 329/* 330 * Sets the default memory attribute for the specified object. Pages 331 * that are allocated to this object are by default assigned this memory 332 * attribute. 333 * 334 * Presently, this function must be called before any pages are allocated 335 * to the object. In the future, this requirement may be relaxed for 336 * "default" and "swap" objects. 337 */ 338int 339vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 340{ 341 342 VM_OBJECT_ASSERT_WLOCKED(object); 343 switch (object->type) { 344 case OBJT_DEFAULT: 345 case OBJT_DEVICE: 346 case OBJT_MGTDEVICE: 347 case OBJT_PHYS: 348 case OBJT_SG: 349 case OBJT_SWAP: 350 case OBJT_VNODE: 351 if (!TAILQ_EMPTY(&object->memq)) 352 return (KERN_FAILURE); 353 break; 354 case OBJT_DEAD: 355 return (KERN_INVALID_ARGUMENT); 356 default: 357 panic("vm_object_set_memattr: object %p is of undefined type", 358 object); 359 } 360 object->memattr = memattr; 361 return (KERN_SUCCESS); 362} 363 364void 365vm_object_pip_add(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_subtract(vm_object_t object, short i) 374{ 375 376 VM_OBJECT_ASSERT_WLOCKED(object); 377 object->paging_in_progress -= i; 378} 379 380void 381vm_object_pip_wakeup(vm_object_t object) 382{ 383 384 VM_OBJECT_ASSERT_WLOCKED(object); 385 object->paging_in_progress--; 386 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 387 vm_object_clear_flag(object, OBJ_PIPWNT); 388 wakeup(object); 389 } 390} 391 392void 393vm_object_pip_wakeupn(vm_object_t object, short i) 394{ 395 396 VM_OBJECT_ASSERT_WLOCKED(object); 397 if (i) 398 object->paging_in_progress -= i; 399 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 400 vm_object_clear_flag(object, OBJ_PIPWNT); 401 wakeup(object); 402 } 403} 404 405void 406vm_object_pip_wait(vm_object_t object, char *waitid) 407{ 408 409 VM_OBJECT_ASSERT_WLOCKED(object); 410 while (object->paging_in_progress) { 411 object->flags |= OBJ_PIPWNT; 412 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 413 } 414} 415 416/* 417 * vm_object_allocate: 418 * 419 * Returns a new object with the given size. 420 */ 421vm_object_t 422vm_object_allocate(objtype_t type, vm_pindex_t size) 423{ 424 vm_object_t object; 425 426 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 427 _vm_object_allocate(type, size, object); 428 return (object); 429} 430 431 432/* 433 * vm_object_reference: 434 * 435 * Gets another reference to the given object. Note: OBJ_DEAD 436 * objects can be referenced during final cleaning. 437 */ 438void 439vm_object_reference(vm_object_t object) 440{ 441 if (object == NULL) 442 return; 443 VM_OBJECT_WLOCK(object); 444 vm_object_reference_locked(object); 445 VM_OBJECT_WUNLOCK(object); 446} 447 448/* 449 * vm_object_reference_locked: 450 * 451 * Gets another reference to the given object. 452 * 453 * The object must be locked. 454 */ 455void 456vm_object_reference_locked(vm_object_t object) 457{ 458 struct vnode *vp; 459 460 VM_OBJECT_ASSERT_WLOCKED(object); 461 object->ref_count++; 462 if (object->type == OBJT_VNODE) { 463 vp = object->handle; 464 vref(vp); 465 } 466} 467 468/* 469 * Handle deallocating an object of type OBJT_VNODE. 470 */ 471static void 472vm_object_vndeallocate(vm_object_t object) 473{ 474 struct vnode *vp = (struct vnode *) object->handle; 475 476 VM_OBJECT_ASSERT_WLOCKED(object); 477 KASSERT(object->type == OBJT_VNODE, 478 ("vm_object_vndeallocate: not a vnode object")); 479 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 480#ifdef INVARIANTS 481 if (object->ref_count == 0) { 482 vn_printf(vp, "vm_object_vndeallocate "); 483 panic("vm_object_vndeallocate: bad object reference count"); 484 } 485#endif 486 487 if (!umtx_shm_vnobj_persistent && object->ref_count == 1) 488 umtx_shm_object_terminated(object); 489 490 /* 491 * The test for text of vp vnode does not need a bypass to 492 * reach right VV_TEXT there, since it is obtained from 493 * object->handle. 494 */ 495 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 496 object->ref_count--; 497 VM_OBJECT_WUNLOCK(object); 498 /* vrele may need the vnode lock. */ 499 vrele(vp); 500 } else { 501 vhold(vp); 502 VM_OBJECT_WUNLOCK(object); 503 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 504 vdrop(vp); 505 VM_OBJECT_WLOCK(object); 506 object->ref_count--; 507 if (object->type == OBJT_DEAD) { 508 VM_OBJECT_WUNLOCK(object); 509 VOP_UNLOCK(vp, 0); 510 } else { 511 if (object->ref_count == 0) 512 VOP_UNSET_TEXT(vp); 513 VM_OBJECT_WUNLOCK(object); 514 vput(vp); 515 } 516 } 517} 518 519/* 520 * vm_object_deallocate: 521 * 522 * Release a reference to the specified object, 523 * gained either through a vm_object_allocate 524 * or a vm_object_reference call. When all references 525 * are gone, storage associated with this object 526 * may be relinquished. 527 * 528 * No object may be locked. 529 */ 530void 531vm_object_deallocate(vm_object_t object) 532{ 533 vm_object_t temp; 534 struct vnode *vp; 535 536 while (object != NULL) { 537 VM_OBJECT_WLOCK(object); 538 if (object->type == OBJT_VNODE) { 539 vm_object_vndeallocate(object); 540 return; 541 } 542 543 KASSERT(object->ref_count != 0, 544 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 545 546 /* 547 * If the reference count goes to 0 we start calling 548 * vm_object_terminate() on the object chain. 549 * A ref count of 1 may be a special case depending on the 550 * shadow count being 0 or 1. 551 */ 552 object->ref_count--; 553 if (object->ref_count > 1) { 554 VM_OBJECT_WUNLOCK(object); 555 return; 556 } else if (object->ref_count == 1) { 557 if (object->type == OBJT_SWAP && 558 (object->flags & OBJ_TMPFS) != 0) { 559 vp = object->un_pager.swp.swp_tmpfs; 560 vhold(vp); 561 VM_OBJECT_WUNLOCK(object); 562 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 563 VM_OBJECT_WLOCK(object); 564 if (object->type == OBJT_DEAD || 565 object->ref_count != 1) { 566 VM_OBJECT_WUNLOCK(object); 567 VOP_UNLOCK(vp, 0); 568 vdrop(vp); 569 return; 570 } 571 if ((object->flags & OBJ_TMPFS) != 0) 572 VOP_UNSET_TEXT(vp); 573 VOP_UNLOCK(vp, 0); 574 vdrop(vp); 575 } 576 if (object->shadow_count == 0 && 577 object->handle == NULL && 578 (object->type == OBJT_DEFAULT || 579 (object->type == OBJT_SWAP && 580 (object->flags & OBJ_TMPFS_NODE) == 0))) { 581 vm_object_set_flag(object, OBJ_ONEMAPPING); 582 } else if ((object->shadow_count == 1) && 583 (object->handle == NULL) && 584 (object->type == OBJT_DEFAULT || 585 object->type == OBJT_SWAP)) { 586 vm_object_t robject; 587 588 robject = LIST_FIRST(&object->shadow_head); 589 KASSERT(robject != NULL, 590 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 591 object->ref_count, 592 object->shadow_count)); 593 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 594 ("shadowed tmpfs v_object %p", object)); 595 if (!VM_OBJECT_TRYWLOCK(robject)) { 596 /* 597 * Avoid a potential deadlock. 598 */ 599 object->ref_count++; 600 VM_OBJECT_WUNLOCK(object); 601 /* 602 * More likely than not the thread 603 * holding robject's lock has lower 604 * priority than the current thread. 605 * Let the lower priority thread run. 606 */ 607 pause("vmo_de", 1); 608 continue; 609 } 610 /* 611 * Collapse object into its shadow unless its 612 * shadow is dead. In that case, object will 613 * be deallocated by the thread that is 614 * deallocating its shadow. 615 */ 616 if ((robject->flags & OBJ_DEAD) == 0 && 617 (robject->handle == NULL) && 618 (robject->type == OBJT_DEFAULT || 619 robject->type == OBJT_SWAP)) { 620 621 robject->ref_count++; 622retry: 623 if (robject->paging_in_progress) { 624 VM_OBJECT_WUNLOCK(object); 625 vm_object_pip_wait(robject, 626 "objde1"); 627 temp = robject->backing_object; 628 if (object == temp) { 629 VM_OBJECT_WLOCK(object); 630 goto retry; 631 } 632 } else if (object->paging_in_progress) { 633 VM_OBJECT_WUNLOCK(robject); 634 object->flags |= OBJ_PIPWNT; 635 VM_OBJECT_SLEEP(object, object, 636 PDROP | PVM, "objde2", 0); 637 VM_OBJECT_WLOCK(robject); 638 temp = robject->backing_object; 639 if (object == temp) { 640 VM_OBJECT_WLOCK(object); 641 goto retry; 642 } 643 } else 644 VM_OBJECT_WUNLOCK(object); 645 646 if (robject->ref_count == 1) { 647 robject->ref_count--; 648 object = robject; 649 goto doterm; 650 } 651 object = robject; 652 vm_object_collapse(object); 653 VM_OBJECT_WUNLOCK(object); 654 continue; 655 } 656 VM_OBJECT_WUNLOCK(robject); 657 } 658 VM_OBJECT_WUNLOCK(object); 659 return; 660 } 661doterm: 662 umtx_shm_object_terminated(object); 663 temp = object->backing_object; 664 if (temp != NULL) { 665 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 666 ("shadowed tmpfs v_object 2 %p", object)); 667 VM_OBJECT_WLOCK(temp); 668 LIST_REMOVE(object, shadow_list); 669 temp->shadow_count--; 670 VM_OBJECT_WUNLOCK(temp); 671 object->backing_object = NULL; 672 } 673 /* 674 * Don't double-terminate, we could be in a termination 675 * recursion due to the terminate having to sync data 676 * to disk. 677 */ 678 if ((object->flags & OBJ_DEAD) == 0) 679 vm_object_terminate(object); 680 else 681 VM_OBJECT_WUNLOCK(object); 682 object = temp; 683 } 684} 685 686/* 687 * vm_object_destroy removes the object from the global object list 688 * and frees the space for the object. 689 */ 690void 691vm_object_destroy(vm_object_t object) 692{ 693 694 /* 695 * Release the allocation charge. 696 */ 697 if (object->cred != NULL) { 698 swap_release_by_cred(object->charge, object->cred); 699 object->charge = 0; 700 crfree(object->cred); 701 object->cred = NULL; 702 } 703 704 /* 705 * Free the space for the object. 706 */ 707 uma_zfree(obj_zone, object); 708} 709 710/* 711 * vm_object_terminate_pages removes any remaining pageable pages 712 * from the object and resets the object to an empty state. 713 */ 714static void 715vm_object_terminate_pages(vm_object_t object) 716{ 717 vm_page_t p, p_next; 718 struct mtx *mtx, *mtx1; 719 struct vm_pagequeue *pq, *pq1; 720 721 VM_OBJECT_ASSERT_WLOCKED(object); 722 723 mtx = NULL; 724 pq = NULL; 725 726 /* 727 * Free any remaining pageable pages. This also removes them from the 728 * paging queues. However, don't free wired pages, just remove them 729 * from the object. Rather than incrementally removing each page from 730 * the object, the page and object are reset to any empty state. 731 */ 732 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 733 vm_page_assert_unbusied(p); 734 if ((object->flags & OBJ_UNMANAGED) == 0) { 735 /* 736 * vm_page_free_prep() only needs the page 737 * lock for managed pages. 738 */ 739 mtx1 = vm_page_lockptr(p); 740 if (mtx1 != mtx) { 741 if (mtx != NULL) 742 mtx_unlock(mtx); 743 if (pq != NULL) { 744 vm_pagequeue_unlock(pq); 745 pq = NULL; 746 } 747 mtx = mtx1; 748 mtx_lock(mtx); 749 } 750 } 751 p->object = NULL; 752 if (p->wire_count != 0) 753 goto unlist; 754 PCPU_INC(cnt.v_pfree); 755 p->flags &= ~PG_ZERO; 756 if (p->queue != PQ_NONE) { 757 KASSERT(p->queue < PQ_COUNT, ("vm_object_terminate: " 758 "page %p is not queued", p)); 759 pq1 = vm_page_pagequeue(p); 760 if (pq != pq1) { 761 if (pq != NULL) 762 vm_pagequeue_unlock(pq); 763 pq = pq1; 764 vm_pagequeue_lock(pq); 765 } 766 } 767 if (vm_page_free_prep(p, true)) 768 continue; 769unlist: 770 TAILQ_REMOVE(&object->memq, p, listq); 771 } 772 if (pq != NULL) 773 vm_pagequeue_unlock(pq); 774 if (mtx != NULL) 775 mtx_unlock(mtx); 776 777 vm_page_free_phys_pglist(&object->memq); 778 779 /* 780 * If the object contained any pages, then reset it to an empty state. 781 * None of the object's fields, including "resident_page_count", were 782 * modified by the preceding loop. 783 */ 784 if (object->resident_page_count != 0) { 785 vm_radix_reclaim_allnodes(&object->rtree); 786 TAILQ_INIT(&object->memq); 787 object->resident_page_count = 0; 788 if (object->type == OBJT_VNODE) 789 vdrop(object->handle); 790 } 791} 792 793/* 794 * vm_object_terminate actually destroys the specified object, freeing 795 * up all previously used resources. 796 * 797 * The object must be locked. 798 * This routine may block. 799 */ 800void 801vm_object_terminate(vm_object_t object) 802{ 803 804 VM_OBJECT_ASSERT_WLOCKED(object); 805 806 /* 807 * Make sure no one uses us. 808 */ 809 vm_object_set_flag(object, OBJ_DEAD); 810 811 /* 812 * wait for the pageout daemon to be done with the object 813 */ 814 vm_object_pip_wait(object, "objtrm"); 815 816 KASSERT(!object->paging_in_progress, 817 ("vm_object_terminate: pageout in progress")); 818 819 /* 820 * Clean and free the pages, as appropriate. All references to the 821 * object are gone, so we don't need to lock it. 822 */ 823 if (object->type == OBJT_VNODE) { 824 struct vnode *vp = (struct vnode *)object->handle; 825 826 /* 827 * Clean pages and flush buffers. 828 */ 829 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 830 VM_OBJECT_WUNLOCK(object); 831 832 vinvalbuf(vp, V_SAVE, 0, 0); 833 834 BO_LOCK(&vp->v_bufobj); 835 vp->v_bufobj.bo_flag |= BO_DEAD; 836 BO_UNLOCK(&vp->v_bufobj); 837 838 VM_OBJECT_WLOCK(object); 839 } 840 841 KASSERT(object->ref_count == 0, 842 ("vm_object_terminate: object with references, ref_count=%d", 843 object->ref_count)); 844 845 if ((object->flags & OBJ_PG_DTOR) == 0) 846 vm_object_terminate_pages(object); 847 848#if VM_NRESERVLEVEL > 0 849 if (__predict_false(!LIST_EMPTY(&object->rvq))) 850 vm_reserv_break_all(object); 851#endif 852 853 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 854 object->type == OBJT_SWAP, 855 ("%s: non-swap obj %p has cred", __func__, object)); 856 857 /* 858 * Let the pager know object is dead. 859 */ 860 vm_pager_deallocate(object); 861 VM_OBJECT_WUNLOCK(object); 862 863 vm_object_destroy(object); 864} 865 866/* 867 * Make the page read-only so that we can clear the object flags. However, if 868 * this is a nosync mmap then the object is likely to stay dirty so do not 869 * mess with the page and do not clear the object flags. Returns TRUE if the 870 * page should be flushed, and FALSE otherwise. 871 */ 872static boolean_t 873vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 874{ 875 876 /* 877 * If we have been asked to skip nosync pages and this is a 878 * nosync page, skip it. Note that the object flags were not 879 * cleared in this case so we do not have to set them. 880 */ 881 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 882 *clearobjflags = FALSE; 883 return (FALSE); 884 } else { 885 pmap_remove_write(p); 886 return (p->dirty != 0); 887 } 888} 889 890/* 891 * vm_object_page_clean 892 * 893 * Clean all dirty pages in the specified range of object. Leaves page 894 * on whatever queue it is currently on. If NOSYNC is set then do not 895 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 896 * leaving the object dirty. 897 * 898 * When stuffing pages asynchronously, allow clustering. XXX we need a 899 * synchronous clustering mode implementation. 900 * 901 * Odd semantics: if start == end, we clean everything. 902 * 903 * The object must be locked. 904 * 905 * Returns FALSE if some page from the range was not written, as 906 * reported by the pager, and TRUE otherwise. 907 */ 908boolean_t 909vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 910 int flags) 911{ 912 vm_page_t np, p; 913 vm_pindex_t pi, tend, tstart; 914 int curgeneration, n, pagerflags; 915 boolean_t clearobjflags, eio, res; 916 917 VM_OBJECT_ASSERT_WLOCKED(object); 918 919 /* 920 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 921 * objects. The check below prevents the function from 922 * operating on non-vnode objects. 923 */ 924 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 925 object->resident_page_count == 0) 926 return (TRUE); 927 928 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 929 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 930 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 931 932 tstart = OFF_TO_IDX(start); 933 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 934 clearobjflags = tstart == 0 && tend >= object->size; 935 res = TRUE; 936 937rescan: 938 curgeneration = object->generation; 939 940 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 941 pi = p->pindex; 942 if (pi >= tend) 943 break; 944 np = TAILQ_NEXT(p, listq); 945 if (p->valid == 0) 946 continue; 947 if (vm_page_sleep_if_busy(p, "vpcwai")) { 948 if (object->generation != curgeneration) { 949 if ((flags & OBJPC_SYNC) != 0) 950 goto rescan; 951 else 952 clearobjflags = FALSE; 953 } 954 np = vm_page_find_least(object, pi); 955 continue; 956 } 957 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 958 continue; 959 960 n = vm_object_page_collect_flush(object, p, pagerflags, 961 flags, &clearobjflags, &eio); 962 if (eio) { 963 res = FALSE; 964 clearobjflags = FALSE; 965 } 966 if (object->generation != curgeneration) { 967 if ((flags & OBJPC_SYNC) != 0) 968 goto rescan; 969 else 970 clearobjflags = FALSE; 971 } 972 973 /* 974 * If the VOP_PUTPAGES() did a truncated write, so 975 * that even the first page of the run is not fully 976 * written, vm_pageout_flush() returns 0 as the run 977 * length. Since the condition that caused truncated 978 * write may be permanent, e.g. exhausted free space, 979 * accepting n == 0 would cause an infinite loop. 980 * 981 * Forwarding the iterator leaves the unwritten page 982 * behind, but there is not much we can do there if 983 * filesystem refuses to write it. 984 */ 985 if (n == 0) { 986 n = 1; 987 clearobjflags = FALSE; 988 } 989 np = vm_page_find_least(object, pi + n); 990 } 991#if 0 992 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 993#endif 994 995 if (clearobjflags) 996 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 997 return (res); 998} 999 1000static int 1001vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1002 int flags, boolean_t *clearobjflags, boolean_t *eio) 1003{ 1004 vm_page_t ma[vm_pageout_page_count], p_first, tp; 1005 int count, i, mreq, runlen; 1006 1007 vm_page_lock_assert(p, MA_NOTOWNED); 1008 VM_OBJECT_ASSERT_WLOCKED(object); 1009 1010 count = 1; 1011 mreq = 0; 1012 1013 for (tp = p; count < vm_pageout_page_count; count++) { 1014 tp = vm_page_next(tp); 1015 if (tp == NULL || vm_page_busied(tp)) 1016 break; 1017 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1018 break; 1019 } 1020 1021 for (p_first = p; count < vm_pageout_page_count; count++) { 1022 tp = vm_page_prev(p_first); 1023 if (tp == NULL || vm_page_busied(tp)) 1024 break; 1025 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1026 break; 1027 p_first = tp; 1028 mreq++; 1029 } 1030 1031 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 1032 ma[i] = tp; 1033 1034 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 1035 return (runlen); 1036} 1037 1038/* 1039 * Note that there is absolutely no sense in writing out 1040 * anonymous objects, so we track down the vnode object 1041 * to write out. 1042 * We invalidate (remove) all pages from the address space 1043 * for semantic correctness. 1044 * 1045 * If the backing object is a device object with unmanaged pages, then any 1046 * mappings to the specified range of pages must be removed before this 1047 * function is called. 1048 * 1049 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1050 * may start out with a NULL object. 1051 */ 1052boolean_t 1053vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1054 boolean_t syncio, boolean_t invalidate) 1055{ 1056 vm_object_t backing_object; 1057 struct vnode *vp; 1058 struct mount *mp; 1059 int error, flags, fsync_after; 1060 boolean_t res; 1061 1062 if (object == NULL) 1063 return (TRUE); 1064 res = TRUE; 1065 error = 0; 1066 VM_OBJECT_WLOCK(object); 1067 while ((backing_object = object->backing_object) != NULL) { 1068 VM_OBJECT_WLOCK(backing_object); 1069 offset += object->backing_object_offset; 1070 VM_OBJECT_WUNLOCK(object); 1071 object = backing_object; 1072 if (object->size < OFF_TO_IDX(offset + size)) 1073 size = IDX_TO_OFF(object->size) - offset; 1074 } 1075 /* 1076 * Flush pages if writing is allowed, invalidate them 1077 * if invalidation requested. Pages undergoing I/O 1078 * will be ignored by vm_object_page_remove(). 1079 * 1080 * We cannot lock the vnode and then wait for paging 1081 * to complete without deadlocking against vm_fault. 1082 * Instead we simply call vm_object_page_remove() and 1083 * allow it to block internally on a page-by-page 1084 * basis when it encounters pages undergoing async 1085 * I/O. 1086 */ 1087 if (object->type == OBJT_VNODE && 1088 (object->flags & OBJ_MIGHTBEDIRTY) != 0 && 1089 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1090 VM_OBJECT_WUNLOCK(object); 1091 (void) vn_start_write(vp, &mp, V_WAIT); 1092 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1093 if (syncio && !invalidate && offset == 0 && 1094 atop(size) == object->size) { 1095 /* 1096 * If syncing the whole mapping of the file, 1097 * it is faster to schedule all the writes in 1098 * async mode, also allowing the clustering, 1099 * and then wait for i/o to complete. 1100 */ 1101 flags = 0; 1102 fsync_after = TRUE; 1103 } else { 1104 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1105 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1106 fsync_after = FALSE; 1107 } 1108 VM_OBJECT_WLOCK(object); 1109 res = vm_object_page_clean(object, offset, offset + size, 1110 flags); 1111 VM_OBJECT_WUNLOCK(object); 1112 if (fsync_after) 1113 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1114 VOP_UNLOCK(vp, 0); 1115 vn_finished_write(mp); 1116 if (error != 0) 1117 res = FALSE; 1118 VM_OBJECT_WLOCK(object); 1119 } 1120 if ((object->type == OBJT_VNODE || 1121 object->type == OBJT_DEVICE) && invalidate) { 1122 if (object->type == OBJT_DEVICE) 1123 /* 1124 * The option OBJPR_NOTMAPPED must be passed here 1125 * because vm_object_page_remove() cannot remove 1126 * unmanaged mappings. 1127 */ 1128 flags = OBJPR_NOTMAPPED; 1129 else if (old_msync) 1130 flags = 0; 1131 else 1132 flags = OBJPR_CLEANONLY; 1133 vm_object_page_remove(object, OFF_TO_IDX(offset), 1134 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1135 } 1136 VM_OBJECT_WUNLOCK(object); 1137 return (res); 1138} 1139 1140/* 1141 * Determine whether the given advice can be applied to the object. Advice is 1142 * not applied to unmanaged pages since they never belong to page queues, and 1143 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1144 * have been mapped at most once. 1145 */ 1146static bool 1147vm_object_advice_applies(vm_object_t object, int advice) 1148{ 1149 1150 if ((object->flags & OBJ_UNMANAGED) != 0) 1151 return (false); 1152 if (advice != MADV_FREE) 1153 return (true); 1154 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && 1155 (object->flags & OBJ_ONEMAPPING) != 0); 1156} 1157 1158static void 1159vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1160 vm_size_t size) 1161{ 1162 1163 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1164 swap_pager_freespace(object, pindex, size); 1165} 1166 1167/* 1168 * vm_object_madvise: 1169 * 1170 * Implements the madvise function at the object/page level. 1171 * 1172 * MADV_WILLNEED (any object) 1173 * 1174 * Activate the specified pages if they are resident. 1175 * 1176 * MADV_DONTNEED (any object) 1177 * 1178 * Deactivate the specified pages if they are resident. 1179 * 1180 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1181 * OBJ_ONEMAPPING only) 1182 * 1183 * Deactivate and clean the specified pages if they are 1184 * resident. This permits the process to reuse the pages 1185 * without faulting or the kernel to reclaim the pages 1186 * without I/O. 1187 */ 1188void 1189vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1190 int advice) 1191{ 1192 vm_pindex_t tpindex; 1193 vm_object_t backing_object, tobject; 1194 vm_page_t m, tm; 1195 1196 if (object == NULL) 1197 return; 1198 1199relookup: 1200 VM_OBJECT_WLOCK(object); 1201 if (!vm_object_advice_applies(object, advice)) { 1202 VM_OBJECT_WUNLOCK(object); 1203 return; 1204 } 1205 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1206 tobject = object; 1207 1208 /* 1209 * If the next page isn't resident in the top-level object, we 1210 * need to search the shadow chain. When applying MADV_FREE, we 1211 * take care to release any swap space used to store 1212 * non-resident pages. 1213 */ 1214 if (m == NULL || pindex < m->pindex) { 1215 /* 1216 * Optimize a common case: if the top-level object has 1217 * no backing object, we can skip over the non-resident 1218 * range in constant time. 1219 */ 1220 if (object->backing_object == NULL) { 1221 tpindex = (m != NULL && m->pindex < end) ? 1222 m->pindex : end; 1223 vm_object_madvise_freespace(object, advice, 1224 pindex, tpindex - pindex); 1225 if ((pindex = tpindex) == end) 1226 break; 1227 goto next_page; 1228 } 1229 1230 tpindex = pindex; 1231 do { 1232 vm_object_madvise_freespace(tobject, advice, 1233 tpindex, 1); 1234 /* 1235 * Prepare to search the next object in the 1236 * chain. 1237 */ 1238 backing_object = tobject->backing_object; 1239 if (backing_object == NULL) 1240 goto next_pindex; 1241 VM_OBJECT_WLOCK(backing_object); 1242 tpindex += 1243 OFF_TO_IDX(tobject->backing_object_offset); 1244 if (tobject != object) 1245 VM_OBJECT_WUNLOCK(tobject); 1246 tobject = backing_object; 1247 if (!vm_object_advice_applies(tobject, advice)) 1248 goto next_pindex; 1249 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1250 NULL); 1251 } else { 1252next_page: 1253 tm = m; 1254 m = TAILQ_NEXT(m, listq); 1255 } 1256 1257 /* 1258 * If the page is not in a normal state, skip it. 1259 */ 1260 if (tm->valid != VM_PAGE_BITS_ALL) 1261 goto next_pindex; 1262 vm_page_lock(tm); 1263 if (tm->hold_count != 0 || tm->wire_count != 0) { 1264 vm_page_unlock(tm); 1265 goto next_pindex; 1266 } 1267 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1268 ("vm_object_madvise: page %p is fictitious", tm)); 1269 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1270 ("vm_object_madvise: page %p is not managed", tm)); 1271 if (vm_page_busied(tm)) { 1272 if (object != tobject) 1273 VM_OBJECT_WUNLOCK(tobject); 1274 VM_OBJECT_WUNLOCK(object); 1275 if (advice == MADV_WILLNEED) { 1276 /* 1277 * Reference the page before unlocking and 1278 * sleeping so that the page daemon is less 1279 * likely to reclaim it. 1280 */ 1281 vm_page_aflag_set(tm, PGA_REFERENCED); 1282 } 1283 vm_page_busy_sleep(tm, "madvpo", false); 1284 goto relookup; 1285 } 1286 vm_page_advise(tm, advice); 1287 vm_page_unlock(tm); 1288 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1289next_pindex: 1290 if (tobject != object) 1291 VM_OBJECT_WUNLOCK(tobject); 1292 } 1293 VM_OBJECT_WUNLOCK(object); 1294} 1295 1296/* 1297 * vm_object_shadow: 1298 * 1299 * Create a new object which is backed by the 1300 * specified existing object range. The source 1301 * object reference is deallocated. 1302 * 1303 * The new object and offset into that object 1304 * are returned in the source parameters. 1305 */ 1306void 1307vm_object_shadow( 1308 vm_object_t *object, /* IN/OUT */ 1309 vm_ooffset_t *offset, /* IN/OUT */ 1310 vm_size_t length) 1311{ 1312 vm_object_t source; 1313 vm_object_t result; 1314 1315 source = *object; 1316 1317 /* 1318 * Don't create the new object if the old object isn't shared. 1319 */ 1320 if (source != NULL) { 1321 VM_OBJECT_WLOCK(source); 1322 if (source->ref_count == 1 && 1323 source->handle == NULL && 1324 (source->type == OBJT_DEFAULT || 1325 source->type == OBJT_SWAP)) { 1326 VM_OBJECT_WUNLOCK(source); 1327 return; 1328 } 1329 VM_OBJECT_WUNLOCK(source); 1330 } 1331 1332 /* 1333 * Allocate a new object with the given length. 1334 */ 1335 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1336 1337 /* 1338 * The new object shadows the source object, adding a reference to it. 1339 * Our caller changes his reference to point to the new object, 1340 * removing a reference to the source object. Net result: no change 1341 * of reference count. 1342 * 1343 * Try to optimize the result object's page color when shadowing 1344 * in order to maintain page coloring consistency in the combined 1345 * shadowed object. 1346 */ 1347 result->backing_object = source; 1348 /* 1349 * Store the offset into the source object, and fix up the offset into 1350 * the new object. 1351 */ 1352 result->backing_object_offset = *offset; 1353 if (source != NULL) { 1354 VM_OBJECT_WLOCK(source); 1355 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1356 source->shadow_count++; 1357#if VM_NRESERVLEVEL > 0 1358 result->flags |= source->flags & OBJ_COLORED; 1359 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1360 ((1 << (VM_NFREEORDER - 1)) - 1); 1361#endif 1362 VM_OBJECT_WUNLOCK(source); 1363 } 1364 1365 1366 /* 1367 * Return the new things 1368 */ 1369 *offset = 0; 1370 *object = result; 1371} 1372 1373/* 1374 * vm_object_split: 1375 * 1376 * Split the pages in a map entry into a new object. This affords 1377 * easier removal of unused pages, and keeps object inheritance from 1378 * being a negative impact on memory usage. 1379 */ 1380void 1381vm_object_split(vm_map_entry_t entry) 1382{ 1383 vm_page_t m, m_next; 1384 vm_object_t orig_object, new_object, source; 1385 vm_pindex_t idx, offidxstart; 1386 vm_size_t size; 1387 1388 orig_object = entry->object.vm_object; 1389 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1390 return; 1391 if (orig_object->ref_count <= 1) 1392 return; 1393 VM_OBJECT_WUNLOCK(orig_object); 1394 1395 offidxstart = OFF_TO_IDX(entry->offset); 1396 size = atop(entry->end - entry->start); 1397 1398 /* 1399 * If swap_pager_copy() is later called, it will convert new_object 1400 * into a swap object. 1401 */ 1402 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1403 1404 /* 1405 * At this point, the new object is still private, so the order in 1406 * which the original and new objects are locked does not matter. 1407 */ 1408 VM_OBJECT_WLOCK(new_object); 1409 VM_OBJECT_WLOCK(orig_object); 1410 source = orig_object->backing_object; 1411 if (source != NULL) { 1412 VM_OBJECT_WLOCK(source); 1413 if ((source->flags & OBJ_DEAD) != 0) { 1414 VM_OBJECT_WUNLOCK(source); 1415 VM_OBJECT_WUNLOCK(orig_object); 1416 VM_OBJECT_WUNLOCK(new_object); 1417 vm_object_deallocate(new_object); 1418 VM_OBJECT_WLOCK(orig_object); 1419 return; 1420 } 1421 LIST_INSERT_HEAD(&source->shadow_head, 1422 new_object, shadow_list); 1423 source->shadow_count++; 1424 vm_object_reference_locked(source); /* for new_object */ 1425 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1426 VM_OBJECT_WUNLOCK(source); 1427 new_object->backing_object_offset = 1428 orig_object->backing_object_offset + entry->offset; 1429 new_object->backing_object = source; 1430 } 1431 if (orig_object->cred != NULL) { 1432 new_object->cred = orig_object->cred; 1433 crhold(orig_object->cred); 1434 new_object->charge = ptoa(size); 1435 KASSERT(orig_object->charge >= ptoa(size), 1436 ("orig_object->charge < 0")); 1437 orig_object->charge -= ptoa(size); 1438 } 1439retry: 1440 m = vm_page_find_least(orig_object, offidxstart); 1441 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1442 m = m_next) { 1443 m_next = TAILQ_NEXT(m, listq); 1444 1445 /* 1446 * We must wait for pending I/O to complete before we can 1447 * rename the page. 1448 * 1449 * We do not have to VM_PROT_NONE the page as mappings should 1450 * not be changed by this operation. 1451 */ 1452 if (vm_page_busied(m)) { 1453 VM_OBJECT_WUNLOCK(new_object); 1454 vm_page_lock(m); 1455 VM_OBJECT_WUNLOCK(orig_object); 1456 vm_page_busy_sleep(m, "spltwt", false); 1457 VM_OBJECT_WLOCK(orig_object); 1458 VM_OBJECT_WLOCK(new_object); 1459 goto retry; 1460 } 1461 1462 /* vm_page_rename() will dirty the page. */ 1463 if (vm_page_rename(m, new_object, idx)) { 1464 VM_OBJECT_WUNLOCK(new_object); 1465 VM_OBJECT_WUNLOCK(orig_object); 1466 vm_radix_wait(); 1467 VM_OBJECT_WLOCK(orig_object); 1468 VM_OBJECT_WLOCK(new_object); 1469 goto retry; 1470 } 1471#if VM_NRESERVLEVEL > 0 1472 /* 1473 * If some of the reservation's allocated pages remain with 1474 * the original object, then transferring the reservation to 1475 * the new object is neither particularly beneficial nor 1476 * particularly harmful as compared to leaving the reservation 1477 * with the original object. If, however, all of the 1478 * reservation's allocated pages are transferred to the new 1479 * object, then transferring the reservation is typically 1480 * beneficial. Determining which of these two cases applies 1481 * would be more costly than unconditionally renaming the 1482 * reservation. 1483 */ 1484 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1485#endif 1486 if (orig_object->type == OBJT_SWAP) 1487 vm_page_xbusy(m); 1488 } 1489 if (orig_object->type == OBJT_SWAP) { 1490 /* 1491 * swap_pager_copy() can sleep, in which case the orig_object's 1492 * and new_object's locks are released and reacquired. 1493 */ 1494 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1495 TAILQ_FOREACH(m, &new_object->memq, listq) 1496 vm_page_xunbusy(m); 1497 } 1498 VM_OBJECT_WUNLOCK(orig_object); 1499 VM_OBJECT_WUNLOCK(new_object); 1500 entry->object.vm_object = new_object; 1501 entry->offset = 0LL; 1502 vm_object_deallocate(orig_object); 1503 VM_OBJECT_WLOCK(new_object); 1504} 1505 1506#define OBSC_COLLAPSE_NOWAIT 0x0002 1507#define OBSC_COLLAPSE_WAIT 0x0004 1508 1509static vm_page_t 1510vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1511 int op) 1512{ 1513 vm_object_t backing_object; 1514 1515 VM_OBJECT_ASSERT_WLOCKED(object); 1516 backing_object = object->backing_object; 1517 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1518 1519 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1520 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1521 ("invalid ownership %p %p %p", p, object, backing_object)); 1522 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1523 return (next); 1524 if (p != NULL) 1525 vm_page_lock(p); 1526 VM_OBJECT_WUNLOCK(object); 1527 VM_OBJECT_WUNLOCK(backing_object); 1528 /* The page is only NULL when rename fails. */ 1529 if (p == NULL) 1530 vm_radix_wait(); 1531 else 1532 vm_page_busy_sleep(p, "vmocol", false); 1533 VM_OBJECT_WLOCK(object); 1534 VM_OBJECT_WLOCK(backing_object); 1535 return (TAILQ_FIRST(&backing_object->memq)); 1536} 1537 1538static bool 1539vm_object_scan_all_shadowed(vm_object_t object) 1540{ 1541 vm_object_t backing_object; 1542 vm_page_t p, pp; 1543 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1544 1545 VM_OBJECT_ASSERT_WLOCKED(object); 1546 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1547 1548 backing_object = object->backing_object; 1549 1550 /* 1551 * Initial conditions: 1552 * 1553 * We do not want to have to test for the existence of swap 1554 * pages in the backing object. XXX but with the new swapper this 1555 * would be pretty easy to do. 1556 */ 1557 if (backing_object->type != OBJT_DEFAULT && 1558 backing_object->type != OBJT_SWAP) 1559 return (false); 1560 1561 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1562 p = vm_page_find_least(backing_object, pi); 1563 ps = swap_pager_find_least(backing_object, pi); 1564 1565 /* 1566 * Only check pages inside the parent object's range and 1567 * inside the parent object's mapping of the backing object. 1568 */ 1569 for (;; pi++) { 1570 if (p != NULL && p->pindex < pi) 1571 p = TAILQ_NEXT(p, listq); 1572 if (ps < pi) 1573 ps = swap_pager_find_least(backing_object, pi); 1574 if (p == NULL && ps >= backing_object->size) 1575 break; 1576 else if (p == NULL) 1577 pi = ps; 1578 else 1579 pi = MIN(p->pindex, ps); 1580 1581 new_pindex = pi - backing_offset_index; 1582 if (new_pindex >= object->size) 1583 break; 1584 1585 /* 1586 * See if the parent has the page or if the parent's object 1587 * pager has the page. If the parent has the page but the page 1588 * is not valid, the parent's object pager must have the page. 1589 * 1590 * If this fails, the parent does not completely shadow the 1591 * object and we might as well give up now. 1592 */ 1593 pp = vm_page_lookup(object, new_pindex); 1594 if ((pp == NULL || pp->valid == 0) && 1595 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1596 return (false); 1597 } 1598 return (true); 1599} 1600 1601static bool 1602vm_object_collapse_scan(vm_object_t object, int op) 1603{ 1604 vm_object_t backing_object; 1605 vm_page_t next, p, pp; 1606 vm_pindex_t backing_offset_index, new_pindex; 1607 1608 VM_OBJECT_ASSERT_WLOCKED(object); 1609 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1610 1611 backing_object = object->backing_object; 1612 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1613 1614 /* 1615 * Initial conditions 1616 */ 1617 if ((op & OBSC_COLLAPSE_WAIT) != 0) 1618 vm_object_set_flag(backing_object, OBJ_DEAD); 1619 1620 /* 1621 * Our scan 1622 */ 1623 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1624 next = TAILQ_NEXT(p, listq); 1625 new_pindex = p->pindex - backing_offset_index; 1626 1627 /* 1628 * Check for busy page 1629 */ 1630 if (vm_page_busied(p)) { 1631 next = vm_object_collapse_scan_wait(object, p, next, op); 1632 continue; 1633 } 1634 1635 KASSERT(p->object == backing_object, 1636 ("vm_object_collapse_scan: object mismatch")); 1637 1638 if (p->pindex < backing_offset_index || 1639 new_pindex >= object->size) { 1640 if (backing_object->type == OBJT_SWAP) 1641 swap_pager_freespace(backing_object, p->pindex, 1642 1); 1643 1644 /* 1645 * Page is out of the parent object's range, we can 1646 * simply destroy it. 1647 */ 1648 vm_page_lock(p); 1649 KASSERT(!pmap_page_is_mapped(p), 1650 ("freeing mapped page %p", p)); 1651 if (p->wire_count == 0) 1652 vm_page_free(p); 1653 else 1654 vm_page_remove(p); 1655 vm_page_unlock(p); 1656 continue; 1657 } 1658 1659 pp = vm_page_lookup(object, new_pindex); 1660 if (pp != NULL && vm_page_busied(pp)) { 1661 /* 1662 * The page in the parent is busy and possibly not 1663 * (yet) valid. Until its state is finalized by the 1664 * busy bit owner, we can't tell whether it shadows the 1665 * original page. Therefore, we must either skip it 1666 * and the original (backing_object) page or wait for 1667 * its state to be finalized. 1668 * 1669 * This is due to a race with vm_fault() where we must 1670 * unbusy the original (backing_obj) page before we can 1671 * (re)lock the parent. Hence we can get here. 1672 */ 1673 next = vm_object_collapse_scan_wait(object, pp, next, 1674 op); 1675 continue; 1676 } 1677 1678 KASSERT(pp == NULL || pp->valid != 0, 1679 ("unbusy invalid page %p", pp)); 1680 1681 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1682 NULL)) { 1683 /* 1684 * The page already exists in the parent OR swap exists 1685 * for this location in the parent. Leave the parent's 1686 * page alone. Destroy the original page from the 1687 * backing object. 1688 */ 1689 if (backing_object->type == OBJT_SWAP) 1690 swap_pager_freespace(backing_object, p->pindex, 1691 1); 1692 vm_page_lock(p); 1693 KASSERT(!pmap_page_is_mapped(p), 1694 ("freeing mapped page %p", p)); 1695 if (p->wire_count == 0) 1696 vm_page_free(p); 1697 else 1698 vm_page_remove(p); 1699 vm_page_unlock(p); 1700 continue; 1701 } 1702 1703 /* 1704 * Page does not exist in parent, rename the page from the 1705 * backing object to the main object. 1706 * 1707 * If the page was mapped to a process, it can remain mapped 1708 * through the rename. vm_page_rename() will dirty the page. 1709 */ 1710 if (vm_page_rename(p, object, new_pindex)) { 1711 next = vm_object_collapse_scan_wait(object, NULL, next, 1712 op); 1713 continue; 1714 } 1715 1716 /* Use the old pindex to free the right page. */ 1717 if (backing_object->type == OBJT_SWAP) 1718 swap_pager_freespace(backing_object, 1719 new_pindex + backing_offset_index, 1); 1720 1721#if VM_NRESERVLEVEL > 0 1722 /* 1723 * Rename the reservation. 1724 */ 1725 vm_reserv_rename(p, object, backing_object, 1726 backing_offset_index); 1727#endif 1728 } 1729 return (true); 1730} 1731 1732 1733/* 1734 * this version of collapse allows the operation to occur earlier and 1735 * when paging_in_progress is true for an object... This is not a complete 1736 * operation, but should plug 99.9% of the rest of the leaks. 1737 */ 1738static void 1739vm_object_qcollapse(vm_object_t object) 1740{ 1741 vm_object_t backing_object = object->backing_object; 1742 1743 VM_OBJECT_ASSERT_WLOCKED(object); 1744 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1745 1746 if (backing_object->ref_count != 1) 1747 return; 1748 1749 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); 1750} 1751 1752/* 1753 * vm_object_collapse: 1754 * 1755 * Collapse an object with the object backing it. 1756 * Pages in the backing object are moved into the 1757 * parent, and the backing object is deallocated. 1758 */ 1759void 1760vm_object_collapse(vm_object_t object) 1761{ 1762 vm_object_t backing_object, new_backing_object; 1763 1764 VM_OBJECT_ASSERT_WLOCKED(object); 1765 1766 while (TRUE) { 1767 /* 1768 * Verify that the conditions are right for collapse: 1769 * 1770 * The object exists and the backing object exists. 1771 */ 1772 if ((backing_object = object->backing_object) == NULL) 1773 break; 1774 1775 /* 1776 * we check the backing object first, because it is most likely 1777 * not collapsable. 1778 */ 1779 VM_OBJECT_WLOCK(backing_object); 1780 if (backing_object->handle != NULL || 1781 (backing_object->type != OBJT_DEFAULT && 1782 backing_object->type != OBJT_SWAP) || 1783 (backing_object->flags & OBJ_DEAD) || 1784 object->handle != NULL || 1785 (object->type != OBJT_DEFAULT && 1786 object->type != OBJT_SWAP) || 1787 (object->flags & OBJ_DEAD)) { 1788 VM_OBJECT_WUNLOCK(backing_object); 1789 break; 1790 } 1791 1792 if (object->paging_in_progress != 0 || 1793 backing_object->paging_in_progress != 0) { 1794 vm_object_qcollapse(object); 1795 VM_OBJECT_WUNLOCK(backing_object); 1796 break; 1797 } 1798 1799 /* 1800 * We know that we can either collapse the backing object (if 1801 * the parent is the only reference to it) or (perhaps) have 1802 * the parent bypass the object if the parent happens to shadow 1803 * all the resident pages in the entire backing object. 1804 * 1805 * This is ignoring pager-backed pages such as swap pages. 1806 * vm_object_collapse_scan fails the shadowing test in this 1807 * case. 1808 */ 1809 if (backing_object->ref_count == 1) { 1810 vm_object_pip_add(object, 1); 1811 vm_object_pip_add(backing_object, 1); 1812 1813 /* 1814 * If there is exactly one reference to the backing 1815 * object, we can collapse it into the parent. 1816 */ 1817 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); 1818 1819#if VM_NRESERVLEVEL > 0 1820 /* 1821 * Break any reservations from backing_object. 1822 */ 1823 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1824 vm_reserv_break_all(backing_object); 1825#endif 1826 1827 /* 1828 * Move the pager from backing_object to object. 1829 */ 1830 if (backing_object->type == OBJT_SWAP) { 1831 /* 1832 * swap_pager_copy() can sleep, in which case 1833 * the backing_object's and object's locks are 1834 * released and reacquired. 1835 * Since swap_pager_copy() is being asked to 1836 * destroy the source, it will change the 1837 * backing_object's type to OBJT_DEFAULT. 1838 */ 1839 swap_pager_copy( 1840 backing_object, 1841 object, 1842 OFF_TO_IDX(object->backing_object_offset), TRUE); 1843 } 1844 /* 1845 * Object now shadows whatever backing_object did. 1846 * Note that the reference to 1847 * backing_object->backing_object moves from within 1848 * backing_object to within object. 1849 */ 1850 LIST_REMOVE(object, shadow_list); 1851 backing_object->shadow_count--; 1852 if (backing_object->backing_object) { 1853 VM_OBJECT_WLOCK(backing_object->backing_object); 1854 LIST_REMOVE(backing_object, shadow_list); 1855 LIST_INSERT_HEAD( 1856 &backing_object->backing_object->shadow_head, 1857 object, shadow_list); 1858 /* 1859 * The shadow_count has not changed. 1860 */ 1861 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1862 } 1863 object->backing_object = backing_object->backing_object; 1864 object->backing_object_offset += 1865 backing_object->backing_object_offset; 1866 1867 /* 1868 * Discard backing_object. 1869 * 1870 * Since the backing object has no pages, no pager left, 1871 * and no object references within it, all that is 1872 * necessary is to dispose of it. 1873 */ 1874 KASSERT(backing_object->ref_count == 1, ( 1875"backing_object %p was somehow re-referenced during collapse!", 1876 backing_object)); 1877 vm_object_pip_wakeup(backing_object); 1878 backing_object->type = OBJT_DEAD; 1879 backing_object->ref_count = 0; 1880 VM_OBJECT_WUNLOCK(backing_object); 1881 vm_object_destroy(backing_object); 1882 1883 vm_object_pip_wakeup(object); 1884 object_collapses++; 1885 } else { 1886 /* 1887 * If we do not entirely shadow the backing object, 1888 * there is nothing we can do so we give up. 1889 */ 1890 if (object->resident_page_count != object->size && 1891 !vm_object_scan_all_shadowed(object)) { 1892 VM_OBJECT_WUNLOCK(backing_object); 1893 break; 1894 } 1895 1896 /* 1897 * Make the parent shadow the next object in the 1898 * chain. Deallocating backing_object will not remove 1899 * it, since its reference count is at least 2. 1900 */ 1901 LIST_REMOVE(object, shadow_list); 1902 backing_object->shadow_count--; 1903 1904 new_backing_object = backing_object->backing_object; 1905 if ((object->backing_object = new_backing_object) != NULL) { 1906 VM_OBJECT_WLOCK(new_backing_object); 1907 LIST_INSERT_HEAD( 1908 &new_backing_object->shadow_head, 1909 object, 1910 shadow_list 1911 ); 1912 new_backing_object->shadow_count++; 1913 vm_object_reference_locked(new_backing_object); 1914 VM_OBJECT_WUNLOCK(new_backing_object); 1915 object->backing_object_offset += 1916 backing_object->backing_object_offset; 1917 } 1918 1919 /* 1920 * Drop the reference count on backing_object. Since 1921 * its ref_count was at least 2, it will not vanish. 1922 */ 1923 backing_object->ref_count--; 1924 VM_OBJECT_WUNLOCK(backing_object); 1925 object_bypasses++; 1926 } 1927 1928 /* 1929 * Try again with this object's new backing object. 1930 */ 1931 } 1932} 1933 1934/* 1935 * vm_object_page_remove: 1936 * 1937 * For the given object, either frees or invalidates each of the 1938 * specified pages. In general, a page is freed. However, if a page is 1939 * wired for any reason other than the existence of a managed, wired 1940 * mapping, then it may be invalidated but not removed from the object. 1941 * Pages are specified by the given range ["start", "end") and the option 1942 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1943 * extends from "start" to the end of the object. If the option 1944 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1945 * specified range are affected. If the option OBJPR_NOTMAPPED is 1946 * specified, then the pages within the specified range must have no 1947 * mappings. Otherwise, if this option is not specified, any mappings to 1948 * the specified pages are removed before the pages are freed or 1949 * invalidated. 1950 * 1951 * In general, this operation should only be performed on objects that 1952 * contain managed pages. There are, however, two exceptions. First, it 1953 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1954 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1955 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1956 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1957 * 1958 * The object must be locked. 1959 */ 1960void 1961vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1962 int options) 1963{ 1964 vm_page_t p, next; 1965 struct mtx *mtx; 1966 struct pglist pgl; 1967 1968 VM_OBJECT_ASSERT_WLOCKED(object); 1969 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1970 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1971 ("vm_object_page_remove: illegal options for object %p", object)); 1972 if (object->resident_page_count == 0) 1973 return; 1974 vm_object_pip_add(object, 1); 1975 TAILQ_INIT(&pgl); 1976again: 1977 p = vm_page_find_least(object, start); 1978 mtx = NULL; 1979 1980 /* 1981 * Here, the variable "p" is either (1) the page with the least pindex 1982 * greater than or equal to the parameter "start" or (2) NULL. 1983 */ 1984 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1985 next = TAILQ_NEXT(p, listq); 1986 1987 /* 1988 * If the page is wired for any reason besides the existence 1989 * of managed, wired mappings, then it cannot be freed. For 1990 * example, fictitious pages, which represent device memory, 1991 * are inherently wired and cannot be freed. They can, 1992 * however, be invalidated if the option OBJPR_CLEANONLY is 1993 * not specified. 1994 */ 1995 vm_page_change_lock(p, &mtx); 1996 if (vm_page_xbusied(p)) { 1997 VM_OBJECT_WUNLOCK(object); 1998 vm_page_busy_sleep(p, "vmopax", true); 1999 VM_OBJECT_WLOCK(object); 2000 goto again; 2001 } 2002 if (p->wire_count != 0) { 2003 if ((options & OBJPR_NOTMAPPED) == 0 && 2004 object->ref_count != 0) 2005 pmap_remove_all(p); 2006 if ((options & OBJPR_CLEANONLY) == 0) { 2007 p->valid = 0; 2008 vm_page_undirty(p); 2009 } 2010 continue; 2011 } 2012 if (vm_page_busied(p)) { 2013 VM_OBJECT_WUNLOCK(object); 2014 vm_page_busy_sleep(p, "vmopar", false); 2015 VM_OBJECT_WLOCK(object); 2016 goto again; 2017 } 2018 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2019 ("vm_object_page_remove: page %p is fictitious", p)); 2020 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 2021 if ((options & OBJPR_NOTMAPPED) == 0 && 2022 object->ref_count != 0) 2023 pmap_remove_write(p); 2024 if (p->dirty != 0) 2025 continue; 2026 } 2027 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) 2028 pmap_remove_all(p); 2029 p->flags &= ~PG_ZERO; 2030 if (vm_page_free_prep(p, false)) 2031 TAILQ_INSERT_TAIL(&pgl, p, listq); 2032 } 2033 if (mtx != NULL) 2034 mtx_unlock(mtx); 2035 vm_page_free_phys_pglist(&pgl); 2036 vm_object_pip_wakeup(object); 2037} 2038 2039/* 2040 * vm_object_page_noreuse: 2041 * 2042 * For the given object, attempt to move the specified pages to 2043 * the head of the inactive queue. This bypasses regular LRU 2044 * operation and allows the pages to be reused quickly under memory 2045 * pressure. If a page is wired for any reason, then it will not 2046 * be queued. Pages are specified by the range ["start", "end"). 2047 * As a special case, if "end" is zero, then the range extends from 2048 * "start" to the end of the object. 2049 * 2050 * This operation should only be performed on objects that 2051 * contain non-fictitious, managed pages. 2052 * 2053 * The object must be locked. 2054 */ 2055void 2056vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2057{ 2058 struct mtx *mtx; 2059 vm_page_t p, next; 2060 2061 VM_OBJECT_ASSERT_LOCKED(object); 2062 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2063 ("vm_object_page_noreuse: illegal object %p", object)); 2064 if (object->resident_page_count == 0) 2065 return; 2066 p = vm_page_find_least(object, start); 2067 2068 /* 2069 * Here, the variable "p" is either (1) the page with the least pindex 2070 * greater than or equal to the parameter "start" or (2) NULL. 2071 */ 2072 mtx = NULL; 2073 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2074 next = TAILQ_NEXT(p, listq); 2075 vm_page_change_lock(p, &mtx); 2076 vm_page_deactivate_noreuse(p); 2077 } 2078 if (mtx != NULL) 2079 mtx_unlock(mtx); 2080} 2081 2082/* 2083 * Populate the specified range of the object with valid pages. Returns 2084 * TRUE if the range is successfully populated and FALSE otherwise. 2085 * 2086 * Note: This function should be optimized to pass a larger array of 2087 * pages to vm_pager_get_pages() before it is applied to a non- 2088 * OBJT_DEVICE object. 2089 * 2090 * The object must be locked. 2091 */ 2092boolean_t 2093vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2094{ 2095 vm_page_t m; 2096 vm_pindex_t pindex; 2097 int rv; 2098 2099 VM_OBJECT_ASSERT_WLOCKED(object); 2100 for (pindex = start; pindex < end; pindex++) { 2101 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2102 if (m->valid != VM_PAGE_BITS_ALL) { 2103 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 2104 if (rv != VM_PAGER_OK) { 2105 vm_page_lock(m); 2106 vm_page_free(m); 2107 vm_page_unlock(m); 2108 break; 2109 } 2110 } 2111 /* 2112 * Keep "m" busy because a subsequent iteration may unlock 2113 * the object. 2114 */ 2115 } 2116 if (pindex > start) { 2117 m = vm_page_lookup(object, start); 2118 while (m != NULL && m->pindex < pindex) { 2119 vm_page_xunbusy(m); 2120 m = TAILQ_NEXT(m, listq); 2121 } 2122 } 2123 return (pindex == end); 2124} 2125 2126/* 2127 * Routine: vm_object_coalesce 2128 * Function: Coalesces two objects backing up adjoining 2129 * regions of memory into a single object. 2130 * 2131 * returns TRUE if objects were combined. 2132 * 2133 * NOTE: Only works at the moment if the second object is NULL - 2134 * if it's not, which object do we lock first? 2135 * 2136 * Parameters: 2137 * prev_object First object to coalesce 2138 * prev_offset Offset into prev_object 2139 * prev_size Size of reference to prev_object 2140 * next_size Size of reference to the second object 2141 * reserved Indicator that extension region has 2142 * swap accounted for 2143 * 2144 * Conditions: 2145 * The object must *not* be locked. 2146 */ 2147boolean_t 2148vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2149 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2150{ 2151 vm_pindex_t next_pindex; 2152 2153 if (prev_object == NULL) 2154 return (TRUE); 2155 VM_OBJECT_WLOCK(prev_object); 2156 if ((prev_object->type != OBJT_DEFAULT && 2157 prev_object->type != OBJT_SWAP) || 2158 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2159 VM_OBJECT_WUNLOCK(prev_object); 2160 return (FALSE); 2161 } 2162 2163 /* 2164 * Try to collapse the object first 2165 */ 2166 vm_object_collapse(prev_object); 2167 2168 /* 2169 * Can't coalesce if: . more than one reference . paged out . shadows 2170 * another object . has a copy elsewhere (any of which mean that the 2171 * pages not mapped to prev_entry may be in use anyway) 2172 */ 2173 if (prev_object->backing_object != NULL) { 2174 VM_OBJECT_WUNLOCK(prev_object); 2175 return (FALSE); 2176 } 2177 2178 prev_size >>= PAGE_SHIFT; 2179 next_size >>= PAGE_SHIFT; 2180 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2181 2182 if ((prev_object->ref_count > 1) && 2183 (prev_object->size != next_pindex)) { 2184 VM_OBJECT_WUNLOCK(prev_object); 2185 return (FALSE); 2186 } 2187 2188 /* 2189 * Account for the charge. 2190 */ 2191 if (prev_object->cred != NULL) { 2192 2193 /* 2194 * If prev_object was charged, then this mapping, 2195 * although not charged now, may become writable 2196 * later. Non-NULL cred in the object would prevent 2197 * swap reservation during enabling of the write 2198 * access, so reserve swap now. Failed reservation 2199 * cause allocation of the separate object for the map 2200 * entry, and swap reservation for this entry is 2201 * managed in appropriate time. 2202 */ 2203 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2204 prev_object->cred)) { 2205 VM_OBJECT_WUNLOCK(prev_object); 2206 return (FALSE); 2207 } 2208 prev_object->charge += ptoa(next_size); 2209 } 2210 2211 /* 2212 * Remove any pages that may still be in the object from a previous 2213 * deallocation. 2214 */ 2215 if (next_pindex < prev_object->size) { 2216 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2217 next_size, 0); 2218 if (prev_object->type == OBJT_SWAP) 2219 swap_pager_freespace(prev_object, 2220 next_pindex, next_size); 2221#if 0 2222 if (prev_object->cred != NULL) { 2223 KASSERT(prev_object->charge >= 2224 ptoa(prev_object->size - next_pindex), 2225 ("object %p overcharged 1 %jx %jx", prev_object, 2226 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2227 prev_object->charge -= ptoa(prev_object->size - 2228 next_pindex); 2229 } 2230#endif 2231 } 2232 2233 /* 2234 * Extend the object if necessary. 2235 */ 2236 if (next_pindex + next_size > prev_object->size) 2237 prev_object->size = next_pindex + next_size; 2238 2239 VM_OBJECT_WUNLOCK(prev_object); 2240 return (TRUE); 2241} 2242 2243void 2244vm_object_set_writeable_dirty(vm_object_t object) 2245{ 2246 2247 VM_OBJECT_ASSERT_WLOCKED(object); 2248 if (object->type != OBJT_VNODE) { 2249 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2250 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2251 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2252 } 2253 return; 2254 } 2255 object->generation++; 2256 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2257 return; 2258 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2259} 2260 2261/* 2262 * vm_object_unwire: 2263 * 2264 * For each page offset within the specified range of the given object, 2265 * find the highest-level page in the shadow chain and unwire it. A page 2266 * must exist at every page offset, and the highest-level page must be 2267 * wired. 2268 */ 2269void 2270vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2271 uint8_t queue) 2272{ 2273 vm_object_t tobject, t1object; 2274 vm_page_t m, tm; 2275 vm_pindex_t end_pindex, pindex, tpindex; 2276 int depth, locked_depth; 2277 2278 KASSERT((offset & PAGE_MASK) == 0, 2279 ("vm_object_unwire: offset is not page aligned")); 2280 KASSERT((length & PAGE_MASK) == 0, 2281 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2282 /* The wired count of a fictitious page never changes. */ 2283 if ((object->flags & OBJ_FICTITIOUS) != 0) 2284 return; 2285 pindex = OFF_TO_IDX(offset); 2286 end_pindex = pindex + atop(length); 2287again: 2288 locked_depth = 1; 2289 VM_OBJECT_RLOCK(object); 2290 m = vm_page_find_least(object, pindex); 2291 while (pindex < end_pindex) { 2292 if (m == NULL || pindex < m->pindex) { 2293 /* 2294 * The first object in the shadow chain doesn't 2295 * contain a page at the current index. Therefore, 2296 * the page must exist in a backing object. 2297 */ 2298 tobject = object; 2299 tpindex = pindex; 2300 depth = 0; 2301 do { 2302 tpindex += 2303 OFF_TO_IDX(tobject->backing_object_offset); 2304 tobject = tobject->backing_object; 2305 KASSERT(tobject != NULL, 2306 ("vm_object_unwire: missing page")); 2307 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2308 goto next_page; 2309 depth++; 2310 if (depth == locked_depth) { 2311 locked_depth++; 2312 VM_OBJECT_RLOCK(tobject); 2313 } 2314 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2315 NULL); 2316 } else { 2317 tm = m; 2318 m = TAILQ_NEXT(m, listq); 2319 } 2320 vm_page_lock(tm); 2321 if (vm_page_xbusied(tm)) { 2322 for (tobject = object; locked_depth >= 1; 2323 locked_depth--) { 2324 t1object = tobject->backing_object; 2325 VM_OBJECT_RUNLOCK(tobject); 2326 tobject = t1object; 2327 } 2328 vm_page_busy_sleep(tm, "unwbo", true); 2329 goto again; 2330 } 2331 vm_page_unwire(tm, queue); 2332 vm_page_unlock(tm); 2333next_page: 2334 pindex++; 2335 } 2336 /* Release the accumulated object locks. */ 2337 for (tobject = object; locked_depth >= 1; locked_depth--) { 2338 t1object = tobject->backing_object; 2339 VM_OBJECT_RUNLOCK(tobject); 2340 tobject = t1object; 2341 } 2342} 2343 2344struct vnode * 2345vm_object_vnode(vm_object_t object) 2346{ 2347 2348 VM_OBJECT_ASSERT_LOCKED(object); 2349 if (object->type == OBJT_VNODE) 2350 return (object->handle); 2351 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) 2352 return (object->un_pager.swp.swp_tmpfs); 2353 return (NULL); 2354} 2355 2356static int 2357sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2358{ 2359 struct kinfo_vmobject *kvo; 2360 char *fullpath, *freepath; 2361 struct vnode *vp; 2362 struct vattr va; 2363 vm_object_t obj; 2364 vm_page_t m; 2365 int count, error; 2366 2367 if (req->oldptr == NULL) { 2368 /* 2369 * If an old buffer has not been provided, generate an 2370 * estimate of the space needed for a subsequent call. 2371 */ 2372 mtx_lock(&vm_object_list_mtx); 2373 count = 0; 2374 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2375 if (obj->type == OBJT_DEAD) 2376 continue; 2377 count++; 2378 } 2379 mtx_unlock(&vm_object_list_mtx); 2380 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2381 count * 11 / 10)); 2382 } 2383 2384 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2385 error = 0; 2386 2387 /* 2388 * VM objects are type stable and are never removed from the 2389 * list once added. This allows us to safely read obj->object_list 2390 * after reacquiring the VM object lock. 2391 */ 2392 mtx_lock(&vm_object_list_mtx); 2393 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2394 if (obj->type == OBJT_DEAD) 2395 continue; 2396 VM_OBJECT_RLOCK(obj); 2397 if (obj->type == OBJT_DEAD) { 2398 VM_OBJECT_RUNLOCK(obj); 2399 continue; 2400 } 2401 mtx_unlock(&vm_object_list_mtx); 2402 kvo->kvo_size = ptoa(obj->size); 2403 kvo->kvo_resident = obj->resident_page_count; 2404 kvo->kvo_ref_count = obj->ref_count; 2405 kvo->kvo_shadow_count = obj->shadow_count; 2406 kvo->kvo_memattr = obj->memattr; 2407 kvo->kvo_active = 0; 2408 kvo->kvo_inactive = 0; 2409 TAILQ_FOREACH(m, &obj->memq, listq) { 2410 /* 2411 * A page may belong to the object but be 2412 * dequeued and set to PQ_NONE while the 2413 * object lock is not held. This makes the 2414 * reads of m->queue below racy, and we do not 2415 * count pages set to PQ_NONE. However, this 2416 * sysctl is only meant to give an 2417 * approximation of the system anyway. 2418 */ 2419 if (vm_page_active(m)) 2420 kvo->kvo_active++; 2421 else if (vm_page_inactive(m)) 2422 kvo->kvo_inactive++; 2423 } 2424 2425 kvo->kvo_vn_fileid = 0; 2426 kvo->kvo_vn_fsid = 0; 2427 freepath = NULL; 2428 fullpath = ""; 2429 vp = NULL; 2430 switch (obj->type) { 2431 case OBJT_DEFAULT: 2432 kvo->kvo_type = KVME_TYPE_DEFAULT; 2433 break; 2434 case OBJT_VNODE: 2435 kvo->kvo_type = KVME_TYPE_VNODE; 2436 vp = obj->handle; 2437 vref(vp); 2438 break; 2439 case OBJT_SWAP: 2440 kvo->kvo_type = KVME_TYPE_SWAP; 2441 break; 2442 case OBJT_DEVICE: 2443 kvo->kvo_type = KVME_TYPE_DEVICE; 2444 break; 2445 case OBJT_PHYS: 2446 kvo->kvo_type = KVME_TYPE_PHYS; 2447 break; 2448 case OBJT_DEAD: 2449 kvo->kvo_type = KVME_TYPE_DEAD; 2450 break; 2451 case OBJT_SG: 2452 kvo->kvo_type = KVME_TYPE_SG; 2453 break; 2454 case OBJT_MGTDEVICE: 2455 kvo->kvo_type = KVME_TYPE_MGTDEVICE; 2456 break; 2457 default: 2458 kvo->kvo_type = KVME_TYPE_UNKNOWN; 2459 break; 2460 } 2461 VM_OBJECT_RUNLOCK(obj); 2462 if (vp != NULL) { 2463 vn_fullpath(curthread, vp, &fullpath, &freepath); 2464 vn_lock(vp, LK_SHARED | LK_RETRY); 2465 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2466 kvo->kvo_vn_fileid = va.va_fileid; 2467 kvo->kvo_vn_fsid = va.va_fsid; 2468 } 2469 vput(vp); 2470 } 2471 2472 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2473 if (freepath != NULL) 2474 free(freepath, M_TEMP); 2475 2476 /* Pack record size down */ 2477 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2478 + strlen(kvo->kvo_path) + 1; 2479 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2480 sizeof(uint64_t)); 2481 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2482 mtx_lock(&vm_object_list_mtx); 2483 if (error) 2484 break; 2485 } 2486 mtx_unlock(&vm_object_list_mtx); 2487 free(kvo, M_TEMP); 2488 return (error); 2489} 2490SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2491 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2492 "List of VM objects"); 2493 2494#include "opt_ddb.h" 2495#ifdef DDB 2496#include <sys/kernel.h> 2497 2498#include <sys/cons.h> 2499 2500#include <ddb/ddb.h> 2501 2502static int 2503_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2504{ 2505 vm_map_t tmpm; 2506 vm_map_entry_t tmpe; 2507 vm_object_t obj; 2508 int entcount; 2509 2510 if (map == 0) 2511 return 0; 2512 2513 if (entry == 0) { 2514 tmpe = map->header.next; 2515 entcount = map->nentries; 2516 while (entcount-- && (tmpe != &map->header)) { 2517 if (_vm_object_in_map(map, object, tmpe)) { 2518 return 1; 2519 } 2520 tmpe = tmpe->next; 2521 } 2522 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2523 tmpm = entry->object.sub_map; 2524 tmpe = tmpm->header.next; 2525 entcount = tmpm->nentries; 2526 while (entcount-- && tmpe != &tmpm->header) { 2527 if (_vm_object_in_map(tmpm, object, tmpe)) { 2528 return 1; 2529 } 2530 tmpe = tmpe->next; 2531 } 2532 } else if ((obj = entry->object.vm_object) != NULL) { 2533 for (; obj; obj = obj->backing_object) 2534 if (obj == object) { 2535 return 1; 2536 } 2537 } 2538 return 0; 2539} 2540 2541static int 2542vm_object_in_map(vm_object_t object) 2543{ 2544 struct proc *p; 2545 2546 /* sx_slock(&allproc_lock); */ 2547 FOREACH_PROC_IN_SYSTEM(p) { 2548 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2549 continue; 2550 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2551 /* sx_sunlock(&allproc_lock); */ 2552 return 1; 2553 } 2554 } 2555 /* sx_sunlock(&allproc_lock); */ 2556 if (_vm_object_in_map(kernel_map, object, 0)) 2557 return 1; 2558 return 0; 2559} 2560 2561DB_SHOW_COMMAND(vmochk, vm_object_check) 2562{ 2563 vm_object_t object; 2564 2565 /* 2566 * make sure that internal objs are in a map somewhere 2567 * and none have zero ref counts. 2568 */ 2569 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2570 if (object->handle == NULL && 2571 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2572 if (object->ref_count == 0) { 2573 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2574 (long)object->size); 2575 } 2576 if (!vm_object_in_map(object)) { 2577 db_printf( 2578 "vmochk: internal obj is not in a map: " 2579 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2580 object->ref_count, (u_long)object->size, 2581 (u_long)object->size, 2582 (void *)object->backing_object); 2583 } 2584 } 2585 } 2586} 2587 2588/* 2589 * vm_object_print: [ debug ] 2590 */ 2591DB_SHOW_COMMAND(object, vm_object_print_static) 2592{ 2593 /* XXX convert args. */ 2594 vm_object_t object = (vm_object_t)addr; 2595 boolean_t full = have_addr; 2596 2597 vm_page_t p; 2598 2599 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2600#define count was_count 2601 2602 int count; 2603 2604 if (object == NULL) 2605 return; 2606 2607 db_iprintf( 2608 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2609 object, (int)object->type, (uintmax_t)object->size, 2610 object->resident_page_count, object->ref_count, object->flags, 2611 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2612 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2613 object->shadow_count, 2614 object->backing_object ? object->backing_object->ref_count : 0, 2615 object->backing_object, (uintmax_t)object->backing_object_offset); 2616 2617 if (!full) 2618 return; 2619 2620 db_indent += 2; 2621 count = 0; 2622 TAILQ_FOREACH(p, &object->memq, listq) { 2623 if (count == 0) 2624 db_iprintf("memory:="); 2625 else if (count == 6) { 2626 db_printf("\n"); 2627 db_iprintf(" ..."); 2628 count = 0; 2629 } else 2630 db_printf(","); 2631 count++; 2632 2633 db_printf("(off=0x%jx,page=0x%jx)", 2634 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2635 } 2636 if (count != 0) 2637 db_printf("\n"); 2638 db_indent -= 2; 2639} 2640 2641/* XXX. */ 2642#undef count 2643 2644/* XXX need this non-static entry for calling from vm_map_print. */ 2645void 2646vm_object_print( 2647 /* db_expr_t */ long addr, 2648 boolean_t have_addr, 2649 /* db_expr_t */ long count, 2650 char *modif) 2651{ 2652 vm_object_print_static(addr, have_addr, count, modif); 2653} 2654 2655DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2656{ 2657 vm_object_t object; 2658 vm_pindex_t fidx; 2659 vm_paddr_t pa; 2660 vm_page_t m, prev_m; 2661 int rcount, nl, c; 2662 2663 nl = 0; 2664 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2665 db_printf("new object: %p\n", (void *)object); 2666 if (nl > 18) { 2667 c = cngetc(); 2668 if (c != ' ') 2669 return; 2670 nl = 0; 2671 } 2672 nl++; 2673 rcount = 0; 2674 fidx = 0; 2675 pa = -1; 2676 TAILQ_FOREACH(m, &object->memq, listq) { 2677 if (m->pindex > 128) 2678 break; 2679 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2680 prev_m->pindex + 1 != m->pindex) { 2681 if (rcount) { 2682 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2683 (long)fidx, rcount, (long)pa); 2684 if (nl > 18) { 2685 c = cngetc(); 2686 if (c != ' ') 2687 return; 2688 nl = 0; 2689 } 2690 nl++; 2691 rcount = 0; 2692 } 2693 } 2694 if (rcount && 2695 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2696 ++rcount; 2697 continue; 2698 } 2699 if (rcount) { 2700 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2701 (long)fidx, rcount, (long)pa); 2702 if (nl > 18) { 2703 c = cngetc(); 2704 if (c != ' ') 2705 return; 2706 nl = 0; 2707 } 2708 nl++; 2709 } 2710 fidx = m->pindex; 2711 pa = VM_PAGE_TO_PHYS(m); 2712 rcount = 1; 2713 } 2714 if (rcount) { 2715 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2716 (long)fidx, rcount, (long)pa); 2717 if (nl > 18) { 2718 c = cngetc(); 2719 if (c != ' ') 2720 return; 2721 nl = 0; 2722 } 2723 nl++; 2724 } 2725 } 2726} 2727#endif /* DDB */ 2728