vm_object.c revision 327785
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/11/sys/vm/vm_object.c 327785 2018-01-10 20:39:26Z markj $"); 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/pctrie.h> 77#include <sys/sysctl.h> 78#include <sys/mutex.h> 79#include <sys/proc.h> /* for curproc, pageproc */ 80#include <sys/socket.h> 81#include <sys/resourcevar.h> 82#include <sys/rwlock.h> 83#include <sys/user.h> 84#include <sys/vnode.h> 85#include <sys/vmmeter.h> 86#include <sys/sx.h> 87 88#include <vm/vm.h> 89#include <vm/vm_param.h> 90#include <vm/pmap.h> 91#include <vm/vm_map.h> 92#include <vm/vm_object.h> 93#include <vm/vm_page.h> 94#include <vm/vm_pageout.h> 95#include <vm/vm_pager.h> 96#include <vm/swap_pager.h> 97#include <vm/vm_kern.h> 98#include <vm/vm_extern.h> 99#include <vm/vm_radix.h> 100#include <vm/vm_reserv.h> 101#include <vm/uma.h> 102 103static int old_msync; 104SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 105 "Use old (insecure) msync behavior"); 106 107static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 108 int pagerflags, int flags, boolean_t *clearobjflags, 109 boolean_t *eio); 110static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 111 boolean_t *clearobjflags); 112static void vm_object_qcollapse(vm_object_t object); 113static void vm_object_vndeallocate(vm_object_t object); 114 115/* 116 * Virtual memory objects maintain the actual data 117 * associated with allocated virtual memory. A given 118 * page of memory exists within exactly one object. 119 * 120 * An object is only deallocated when all "references" 121 * are given up. Only one "reference" to a given 122 * region of an object should be writeable. 123 * 124 * Associated with each object is a list of all resident 125 * memory pages belonging to that object; this list is 126 * maintained by the "vm_page" module, and locked by the object's 127 * lock. 128 * 129 * Each object also records a "pager" routine which is 130 * used to retrieve (and store) pages to the proper backing 131 * storage. In addition, objects may be backed by other 132 * objects from which they were virtual-copied. 133 * 134 * The only items within the object structure which are 135 * modified after time of creation are: 136 * reference count locked by object's lock 137 * pager routine locked by object's lock 138 * 139 */ 140 141struct object_q vm_object_list; 142struct mtx vm_object_list_mtx; /* lock for object list and count */ 143 144struct vm_object kernel_object_store; 145struct vm_object kmem_object_store; 146 147static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 148 "VM object stats"); 149 150static long object_collapses; 151SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 152 &object_collapses, 0, "VM object collapses"); 153 154static long object_bypasses; 155SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 156 &object_bypasses, 0, "VM object bypasses"); 157 158static uma_zone_t obj_zone; 159 160static int vm_object_zinit(void *mem, int size, int flags); 161 162#ifdef INVARIANTS 163static void vm_object_zdtor(void *mem, int size, void *arg); 164 165static void 166vm_object_zdtor(void *mem, int size, void *arg) 167{ 168 vm_object_t object; 169 170 object = (vm_object_t)mem; 171 KASSERT(object->ref_count == 0, 172 ("object %p ref_count = %d", object, object->ref_count)); 173 KASSERT(TAILQ_EMPTY(&object->memq), 174 ("object %p has resident pages in its memq", object)); 175 KASSERT(vm_radix_is_empty(&object->rtree), 176 ("object %p has resident pages in its trie", object)); 177#if VM_NRESERVLEVEL > 0 178 KASSERT(LIST_EMPTY(&object->rvq), 179 ("object %p has reservations", 180 object)); 181#endif 182 KASSERT(object->paging_in_progress == 0, 183 ("object %p paging_in_progress = %d", 184 object, object->paging_in_progress)); 185 KASSERT(object->resident_page_count == 0, 186 ("object %p resident_page_count = %d", 187 object, object->resident_page_count)); 188 KASSERT(object->shadow_count == 0, 189 ("object %p shadow_count = %d", 190 object, object->shadow_count)); 191 KASSERT(object->type == OBJT_DEAD, 192 ("object %p has non-dead type %d", 193 object, object->type)); 194} 195#endif 196 197static int 198vm_object_zinit(void *mem, int size, int flags) 199{ 200 vm_object_t object; 201 202 object = (vm_object_t)mem; 203 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 204 205 /* These are true for any object that has been freed */ 206 object->type = OBJT_DEAD; 207 object->ref_count = 0; 208 vm_radix_init(&object->rtree); 209 object->paging_in_progress = 0; 210 object->resident_page_count = 0; 211 object->shadow_count = 0; 212 object->flags = OBJ_DEAD; 213 214 mtx_lock(&vm_object_list_mtx); 215 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 216 mtx_unlock(&vm_object_list_mtx); 217 return (0); 218} 219 220static void 221_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 222{ 223 224 TAILQ_INIT(&object->memq); 225 LIST_INIT(&object->shadow_head); 226 227 object->type = type; 228 if (type == OBJT_SWAP) 229 pctrie_init(&object->un_pager.swp.swp_blks); 230 231 /* 232 * Ensure that swap_pager_swapoff() iteration over object_list 233 * sees up to date type and pctrie head if it observed 234 * non-dead object. 235 */ 236 atomic_thread_fence_rel(); 237 238 switch (type) { 239 case OBJT_DEAD: 240 panic("_vm_object_allocate: can't create OBJT_DEAD"); 241 case OBJT_DEFAULT: 242 case OBJT_SWAP: 243 object->flags = OBJ_ONEMAPPING; 244 break; 245 case OBJT_DEVICE: 246 case OBJT_SG: 247 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 248 break; 249 case OBJT_MGTDEVICE: 250 object->flags = OBJ_FICTITIOUS; 251 break; 252 case OBJT_PHYS: 253 object->flags = OBJ_UNMANAGED; 254 break; 255 case OBJT_VNODE: 256 object->flags = 0; 257 break; 258 default: 259 panic("_vm_object_allocate: type %d is undefined", type); 260 } 261 object->size = size; 262 object->generation = 1; 263 object->ref_count = 1; 264 object->memattr = VM_MEMATTR_DEFAULT; 265 object->cred = NULL; 266 object->charge = 0; 267 object->handle = NULL; 268 object->backing_object = NULL; 269 object->backing_object_offset = (vm_ooffset_t) 0; 270#if VM_NRESERVLEVEL > 0 271 LIST_INIT(&object->rvq); 272#endif 273 umtx_shm_object_init(object); 274} 275 276/* 277 * vm_object_init: 278 * 279 * Initialize the VM objects module. 280 */ 281void 282vm_object_init(void) 283{ 284 TAILQ_INIT(&vm_object_list); 285 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 286 287 rw_init(&kernel_object->lock, "kernel vm object"); 288 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 289 VM_MIN_KERNEL_ADDRESS), kernel_object); 290#if VM_NRESERVLEVEL > 0 291 kernel_object->flags |= OBJ_COLORED; 292 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 293#endif 294 295 rw_init(&kmem_object->lock, "kmem vm object"); 296 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 297 VM_MIN_KERNEL_ADDRESS), kmem_object); 298#if VM_NRESERVLEVEL > 0 299 kmem_object->flags |= OBJ_COLORED; 300 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 301#endif 302 303 /* 304 * The lock portion of struct vm_object must be type stable due 305 * to vm_pageout_fallback_object_lock locking a vm object 306 * without holding any references to it. 307 */ 308 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 309#ifdef INVARIANTS 310 vm_object_zdtor, 311#else 312 NULL, 313#endif 314 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 315 316 vm_radix_zinit(); 317} 318 319void 320vm_object_clear_flag(vm_object_t object, u_short bits) 321{ 322 323 VM_OBJECT_ASSERT_WLOCKED(object); 324 object->flags &= ~bits; 325} 326 327/* 328 * Sets the default memory attribute for the specified object. Pages 329 * that are allocated to this object are by default assigned this memory 330 * attribute. 331 * 332 * Presently, this function must be called before any pages are allocated 333 * to the object. In the future, this requirement may be relaxed for 334 * "default" and "swap" objects. 335 */ 336int 337vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 338{ 339 340 VM_OBJECT_ASSERT_WLOCKED(object); 341 switch (object->type) { 342 case OBJT_DEFAULT: 343 case OBJT_DEVICE: 344 case OBJT_MGTDEVICE: 345 case OBJT_PHYS: 346 case OBJT_SG: 347 case OBJT_SWAP: 348 case OBJT_VNODE: 349 if (!TAILQ_EMPTY(&object->memq)) 350 return (KERN_FAILURE); 351 break; 352 case OBJT_DEAD: 353 return (KERN_INVALID_ARGUMENT); 354 default: 355 panic("vm_object_set_memattr: object %p is of undefined type", 356 object); 357 } 358 object->memattr = memattr; 359 return (KERN_SUCCESS); 360} 361 362void 363vm_object_pip_add(vm_object_t object, short i) 364{ 365 366 VM_OBJECT_ASSERT_WLOCKED(object); 367 object->paging_in_progress += i; 368} 369 370void 371vm_object_pip_subtract(vm_object_t object, short i) 372{ 373 374 VM_OBJECT_ASSERT_WLOCKED(object); 375 object->paging_in_progress -= i; 376} 377 378void 379vm_object_pip_wakeup(vm_object_t object) 380{ 381 382 VM_OBJECT_ASSERT_WLOCKED(object); 383 object->paging_in_progress--; 384 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 385 vm_object_clear_flag(object, OBJ_PIPWNT); 386 wakeup(object); 387 } 388} 389 390void 391vm_object_pip_wakeupn(vm_object_t object, short i) 392{ 393 394 VM_OBJECT_ASSERT_WLOCKED(object); 395 if (i) 396 object->paging_in_progress -= i; 397 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 398 vm_object_clear_flag(object, OBJ_PIPWNT); 399 wakeup(object); 400 } 401} 402 403void 404vm_object_pip_wait(vm_object_t object, char *waitid) 405{ 406 407 VM_OBJECT_ASSERT_WLOCKED(object); 408 while (object->paging_in_progress) { 409 object->flags |= OBJ_PIPWNT; 410 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 411 } 412} 413 414/* 415 * vm_object_allocate: 416 * 417 * Returns a new object with the given size. 418 */ 419vm_object_t 420vm_object_allocate(objtype_t type, vm_pindex_t size) 421{ 422 vm_object_t object; 423 424 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 425 _vm_object_allocate(type, size, object); 426 return (object); 427} 428 429 430/* 431 * vm_object_reference: 432 * 433 * Gets another reference to the given object. Note: OBJ_DEAD 434 * objects can be referenced during final cleaning. 435 */ 436void 437vm_object_reference(vm_object_t object) 438{ 439 if (object == NULL) 440 return; 441 VM_OBJECT_WLOCK(object); 442 vm_object_reference_locked(object); 443 VM_OBJECT_WUNLOCK(object); 444} 445 446/* 447 * vm_object_reference_locked: 448 * 449 * Gets another reference to the given object. 450 * 451 * The object must be locked. 452 */ 453void 454vm_object_reference_locked(vm_object_t object) 455{ 456 struct vnode *vp; 457 458 VM_OBJECT_ASSERT_WLOCKED(object); 459 object->ref_count++; 460 if (object->type == OBJT_VNODE) { 461 vp = object->handle; 462 vref(vp); 463 } 464} 465 466/* 467 * Handle deallocating an object of type OBJT_VNODE. 468 */ 469static void 470vm_object_vndeallocate(vm_object_t object) 471{ 472 struct vnode *vp = (struct vnode *) object->handle; 473 474 VM_OBJECT_ASSERT_WLOCKED(object); 475 KASSERT(object->type == OBJT_VNODE, 476 ("vm_object_vndeallocate: not a vnode object")); 477 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 478#ifdef INVARIANTS 479 if (object->ref_count == 0) { 480 vn_printf(vp, "vm_object_vndeallocate "); 481 panic("vm_object_vndeallocate: bad object reference count"); 482 } 483#endif 484 485 if (!umtx_shm_vnobj_persistent && object->ref_count == 1) 486 umtx_shm_object_terminated(object); 487 488 /* 489 * The test for text of vp vnode does not need a bypass to 490 * reach right VV_TEXT there, since it is obtained from 491 * object->handle. 492 */ 493 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 494 object->ref_count--; 495 VM_OBJECT_WUNLOCK(object); 496 /* vrele may need the vnode lock. */ 497 vrele(vp); 498 } else { 499 vhold(vp); 500 VM_OBJECT_WUNLOCK(object); 501 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 502 vdrop(vp); 503 VM_OBJECT_WLOCK(object); 504 object->ref_count--; 505 if (object->type == OBJT_DEAD) { 506 VM_OBJECT_WUNLOCK(object); 507 VOP_UNLOCK(vp, 0); 508 } else { 509 if (object->ref_count == 0) 510 VOP_UNSET_TEXT(vp); 511 VM_OBJECT_WUNLOCK(object); 512 vput(vp); 513 } 514 } 515} 516 517/* 518 * vm_object_deallocate: 519 * 520 * Release a reference to the specified object, 521 * gained either through a vm_object_allocate 522 * or a vm_object_reference call. When all references 523 * are gone, storage associated with this object 524 * may be relinquished. 525 * 526 * No object may be locked. 527 */ 528void 529vm_object_deallocate(vm_object_t object) 530{ 531 vm_object_t temp; 532 struct vnode *vp; 533 534 while (object != NULL) { 535 VM_OBJECT_WLOCK(object); 536 if (object->type == OBJT_VNODE) { 537 vm_object_vndeallocate(object); 538 return; 539 } 540 541 KASSERT(object->ref_count != 0, 542 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 543 544 /* 545 * If the reference count goes to 0 we start calling 546 * vm_object_terminate() on the object chain. 547 * A ref count of 1 may be a special case depending on the 548 * shadow count being 0 or 1. 549 */ 550 object->ref_count--; 551 if (object->ref_count > 1) { 552 VM_OBJECT_WUNLOCK(object); 553 return; 554 } else if (object->ref_count == 1) { 555 if (object->type == OBJT_SWAP && 556 (object->flags & OBJ_TMPFS) != 0) { 557 vp = object->un_pager.swp.swp_tmpfs; 558 vhold(vp); 559 VM_OBJECT_WUNLOCK(object); 560 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 561 VM_OBJECT_WLOCK(object); 562 if (object->type == OBJT_DEAD || 563 object->ref_count != 1) { 564 VM_OBJECT_WUNLOCK(object); 565 VOP_UNLOCK(vp, 0); 566 vdrop(vp); 567 return; 568 } 569 if ((object->flags & OBJ_TMPFS) != 0) 570 VOP_UNSET_TEXT(vp); 571 VOP_UNLOCK(vp, 0); 572 vdrop(vp); 573 } 574 if (object->shadow_count == 0 && 575 object->handle == NULL && 576 (object->type == OBJT_DEFAULT || 577 (object->type == OBJT_SWAP && 578 (object->flags & OBJ_TMPFS_NODE) == 0))) { 579 vm_object_set_flag(object, OBJ_ONEMAPPING); 580 } else if ((object->shadow_count == 1) && 581 (object->handle == NULL) && 582 (object->type == OBJT_DEFAULT || 583 object->type == OBJT_SWAP)) { 584 vm_object_t robject; 585 586 robject = LIST_FIRST(&object->shadow_head); 587 KASSERT(robject != NULL, 588 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 589 object->ref_count, 590 object->shadow_count)); 591 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 592 ("shadowed tmpfs v_object %p", object)); 593 if (!VM_OBJECT_TRYWLOCK(robject)) { 594 /* 595 * Avoid a potential deadlock. 596 */ 597 object->ref_count++; 598 VM_OBJECT_WUNLOCK(object); 599 /* 600 * More likely than not the thread 601 * holding robject's lock has lower 602 * priority than the current thread. 603 * Let the lower priority thread run. 604 */ 605 pause("vmo_de", 1); 606 continue; 607 } 608 /* 609 * Collapse object into its shadow unless its 610 * shadow is dead. In that case, object will 611 * be deallocated by the thread that is 612 * deallocating its shadow. 613 */ 614 if ((robject->flags & OBJ_DEAD) == 0 && 615 (robject->handle == NULL) && 616 (robject->type == OBJT_DEFAULT || 617 robject->type == OBJT_SWAP)) { 618 619 robject->ref_count++; 620retry: 621 if (robject->paging_in_progress) { 622 VM_OBJECT_WUNLOCK(object); 623 vm_object_pip_wait(robject, 624 "objde1"); 625 temp = robject->backing_object; 626 if (object == temp) { 627 VM_OBJECT_WLOCK(object); 628 goto retry; 629 } 630 } else if (object->paging_in_progress) { 631 VM_OBJECT_WUNLOCK(robject); 632 object->flags |= OBJ_PIPWNT; 633 VM_OBJECT_SLEEP(object, object, 634 PDROP | PVM, "objde2", 0); 635 VM_OBJECT_WLOCK(robject); 636 temp = robject->backing_object; 637 if (object == temp) { 638 VM_OBJECT_WLOCK(object); 639 goto retry; 640 } 641 } else 642 VM_OBJECT_WUNLOCK(object); 643 644 if (robject->ref_count == 1) { 645 robject->ref_count--; 646 object = robject; 647 goto doterm; 648 } 649 object = robject; 650 vm_object_collapse(object); 651 VM_OBJECT_WUNLOCK(object); 652 continue; 653 } 654 VM_OBJECT_WUNLOCK(robject); 655 } 656 VM_OBJECT_WUNLOCK(object); 657 return; 658 } 659doterm: 660 umtx_shm_object_terminated(object); 661 temp = object->backing_object; 662 if (temp != NULL) { 663 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 664 ("shadowed tmpfs v_object 2 %p", object)); 665 VM_OBJECT_WLOCK(temp); 666 LIST_REMOVE(object, shadow_list); 667 temp->shadow_count--; 668 VM_OBJECT_WUNLOCK(temp); 669 object->backing_object = NULL; 670 } 671 /* 672 * Don't double-terminate, we could be in a termination 673 * recursion due to the terminate having to sync data 674 * to disk. 675 */ 676 if ((object->flags & OBJ_DEAD) == 0) 677 vm_object_terminate(object); 678 else 679 VM_OBJECT_WUNLOCK(object); 680 object = temp; 681 } 682} 683 684/* 685 * vm_object_destroy removes the object from the global object list 686 * and frees the space for the object. 687 */ 688void 689vm_object_destroy(vm_object_t object) 690{ 691 692 /* 693 * Release the allocation charge. 694 */ 695 if (object->cred != NULL) { 696 swap_release_by_cred(object->charge, object->cred); 697 object->charge = 0; 698 crfree(object->cred); 699 object->cred = NULL; 700 } 701 702 /* 703 * Free the space for the object. 704 */ 705 uma_zfree(obj_zone, object); 706} 707 708/* 709 * vm_object_terminate_pages removes any remaining pageable pages 710 * from the object and resets the object to an empty state. 711 */ 712static void 713vm_object_terminate_pages(vm_object_t object) 714{ 715 vm_page_t p, p_next; 716 struct mtx *mtx, *mtx1; 717 struct vm_pagequeue *pq, *pq1; 718 719 VM_OBJECT_ASSERT_WLOCKED(object); 720 721 mtx = NULL; 722 pq = NULL; 723 724 /* 725 * Free any remaining pageable pages. This also removes them from the 726 * paging queues. However, don't free wired pages, just remove them 727 * from the object. Rather than incrementally removing each page from 728 * the object, the page and object are reset to any empty state. 729 */ 730 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 731 vm_page_assert_unbusied(p); 732 if ((object->flags & OBJ_UNMANAGED) == 0) { 733 /* 734 * vm_page_free_prep() only needs the page 735 * lock for managed pages. 736 */ 737 mtx1 = vm_page_lockptr(p); 738 if (mtx1 != mtx) { 739 if (mtx != NULL) 740 mtx_unlock(mtx); 741 if (pq != NULL) { 742 vm_pagequeue_unlock(pq); 743 pq = NULL; 744 } 745 mtx = mtx1; 746 mtx_lock(mtx); 747 } 748 } 749 p->object = NULL; 750 if (p->wire_count != 0) 751 goto unlist; 752 PCPU_INC(cnt.v_pfree); 753 p->flags &= ~PG_ZERO; 754 if (p->queue != PQ_NONE) { 755 KASSERT(p->queue < PQ_COUNT, ("vm_object_terminate: " 756 "page %p is not queued", p)); 757 pq1 = vm_page_pagequeue(p); 758 if (pq != pq1) { 759 if (pq != NULL) 760 vm_pagequeue_unlock(pq); 761 pq = pq1; 762 vm_pagequeue_lock(pq); 763 } 764 } 765 if (vm_page_free_prep(p, true)) 766 continue; 767unlist: 768 TAILQ_REMOVE(&object->memq, p, listq); 769 } 770 if (pq != NULL) 771 vm_pagequeue_unlock(pq); 772 if (mtx != NULL) 773 mtx_unlock(mtx); 774 775 vm_page_free_phys_pglist(&object->memq); 776 777 /* 778 * If the object contained any pages, then reset it to an empty state. 779 * None of the object's fields, including "resident_page_count", were 780 * modified by the preceding loop. 781 */ 782 if (object->resident_page_count != 0) { 783 vm_radix_reclaim_allnodes(&object->rtree); 784 TAILQ_INIT(&object->memq); 785 object->resident_page_count = 0; 786 if (object->type == OBJT_VNODE) 787 vdrop(object->handle); 788 } 789} 790 791/* 792 * vm_object_terminate actually destroys the specified object, freeing 793 * up all previously used resources. 794 * 795 * The object must be locked. 796 * This routine may block. 797 */ 798void 799vm_object_terminate(vm_object_t object) 800{ 801 802 VM_OBJECT_ASSERT_WLOCKED(object); 803 804 /* 805 * Make sure no one uses us. 806 */ 807 vm_object_set_flag(object, OBJ_DEAD); 808 809 /* 810 * wait for the pageout daemon to be done with the object 811 */ 812 vm_object_pip_wait(object, "objtrm"); 813 814 KASSERT(!object->paging_in_progress, 815 ("vm_object_terminate: pageout in progress")); 816 817 /* 818 * Clean and free the pages, as appropriate. All references to the 819 * object are gone, so we don't need to lock it. 820 */ 821 if (object->type == OBJT_VNODE) { 822 struct vnode *vp = (struct vnode *)object->handle; 823 824 /* 825 * Clean pages and flush buffers. 826 */ 827 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 828 VM_OBJECT_WUNLOCK(object); 829 830 vinvalbuf(vp, V_SAVE, 0, 0); 831 832 BO_LOCK(&vp->v_bufobj); 833 vp->v_bufobj.bo_flag |= BO_DEAD; 834 BO_UNLOCK(&vp->v_bufobj); 835 836 VM_OBJECT_WLOCK(object); 837 } 838 839 KASSERT(object->ref_count == 0, 840 ("vm_object_terminate: object with references, ref_count=%d", 841 object->ref_count)); 842 843 if ((object->flags & OBJ_PG_DTOR) == 0) 844 vm_object_terminate_pages(object); 845 846#if VM_NRESERVLEVEL > 0 847 if (__predict_false(!LIST_EMPTY(&object->rvq))) 848 vm_reserv_break_all(object); 849#endif 850 851 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 852 object->type == OBJT_SWAP, 853 ("%s: non-swap obj %p has cred", __func__, object)); 854 855 /* 856 * Let the pager know object is dead. 857 */ 858 vm_pager_deallocate(object); 859 VM_OBJECT_WUNLOCK(object); 860 861 vm_object_destroy(object); 862} 863 864/* 865 * Make the page read-only so that we can clear the object flags. However, if 866 * this is a nosync mmap then the object is likely to stay dirty so do not 867 * mess with the page and do not clear the object flags. Returns TRUE if the 868 * page should be flushed, and FALSE otherwise. 869 */ 870static boolean_t 871vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 872{ 873 874 /* 875 * If we have been asked to skip nosync pages and this is a 876 * nosync page, skip it. Note that the object flags were not 877 * cleared in this case so we do not have to set them. 878 */ 879 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 880 *clearobjflags = FALSE; 881 return (FALSE); 882 } else { 883 pmap_remove_write(p); 884 return (p->dirty != 0); 885 } 886} 887 888/* 889 * vm_object_page_clean 890 * 891 * Clean all dirty pages in the specified range of object. Leaves page 892 * on whatever queue it is currently on. If NOSYNC is set then do not 893 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 894 * leaving the object dirty. 895 * 896 * When stuffing pages asynchronously, allow clustering. XXX we need a 897 * synchronous clustering mode implementation. 898 * 899 * Odd semantics: if start == end, we clean everything. 900 * 901 * The object must be locked. 902 * 903 * Returns FALSE if some page from the range was not written, as 904 * reported by the pager, and TRUE otherwise. 905 */ 906boolean_t 907vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 908 int flags) 909{ 910 vm_page_t np, p; 911 vm_pindex_t pi, tend, tstart; 912 int curgeneration, n, pagerflags; 913 boolean_t clearobjflags, eio, res; 914 915 VM_OBJECT_ASSERT_WLOCKED(object); 916 917 /* 918 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 919 * objects. The check below prevents the function from 920 * operating on non-vnode objects. 921 */ 922 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 923 object->resident_page_count == 0) 924 return (TRUE); 925 926 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 927 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 928 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 929 930 tstart = OFF_TO_IDX(start); 931 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 932 clearobjflags = tstart == 0 && tend >= object->size; 933 res = TRUE; 934 935rescan: 936 curgeneration = object->generation; 937 938 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 939 pi = p->pindex; 940 if (pi >= tend) 941 break; 942 np = TAILQ_NEXT(p, listq); 943 if (p->valid == 0) 944 continue; 945 if (vm_page_sleep_if_busy(p, "vpcwai")) { 946 if (object->generation != curgeneration) { 947 if ((flags & OBJPC_SYNC) != 0) 948 goto rescan; 949 else 950 clearobjflags = FALSE; 951 } 952 np = vm_page_find_least(object, pi); 953 continue; 954 } 955 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 956 continue; 957 958 n = vm_object_page_collect_flush(object, p, pagerflags, 959 flags, &clearobjflags, &eio); 960 if (eio) { 961 res = FALSE; 962 clearobjflags = FALSE; 963 } 964 if (object->generation != curgeneration) { 965 if ((flags & OBJPC_SYNC) != 0) 966 goto rescan; 967 else 968 clearobjflags = FALSE; 969 } 970 971 /* 972 * If the VOP_PUTPAGES() did a truncated write, so 973 * that even the first page of the run is not fully 974 * written, vm_pageout_flush() returns 0 as the run 975 * length. Since the condition that caused truncated 976 * write may be permanent, e.g. exhausted free space, 977 * accepting n == 0 would cause an infinite loop. 978 * 979 * Forwarding the iterator leaves the unwritten page 980 * behind, but there is not much we can do there if 981 * filesystem refuses to write it. 982 */ 983 if (n == 0) { 984 n = 1; 985 clearobjflags = FALSE; 986 } 987 np = vm_page_find_least(object, pi + n); 988 } 989#if 0 990 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 991#endif 992 993 if (clearobjflags) 994 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 995 return (res); 996} 997 998static int 999vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1000 int flags, boolean_t *clearobjflags, boolean_t *eio) 1001{ 1002 vm_page_t ma[vm_pageout_page_count], p_first, tp; 1003 int count, i, mreq, runlen; 1004 1005 vm_page_lock_assert(p, MA_NOTOWNED); 1006 VM_OBJECT_ASSERT_WLOCKED(object); 1007 1008 count = 1; 1009 mreq = 0; 1010 1011 for (tp = p; count < vm_pageout_page_count; count++) { 1012 tp = vm_page_next(tp); 1013 if (tp == NULL || vm_page_busied(tp)) 1014 break; 1015 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1016 break; 1017 } 1018 1019 for (p_first = p; count < vm_pageout_page_count; count++) { 1020 tp = vm_page_prev(p_first); 1021 if (tp == NULL || vm_page_busied(tp)) 1022 break; 1023 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1024 break; 1025 p_first = tp; 1026 mreq++; 1027 } 1028 1029 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 1030 ma[i] = tp; 1031 1032 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 1033 return (runlen); 1034} 1035 1036/* 1037 * Note that there is absolutely no sense in writing out 1038 * anonymous objects, so we track down the vnode object 1039 * to write out. 1040 * We invalidate (remove) all pages from the address space 1041 * for semantic correctness. 1042 * 1043 * If the backing object is a device object with unmanaged pages, then any 1044 * mappings to the specified range of pages must be removed before this 1045 * function is called. 1046 * 1047 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1048 * may start out with a NULL object. 1049 */ 1050boolean_t 1051vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1052 boolean_t syncio, boolean_t invalidate) 1053{ 1054 vm_object_t backing_object; 1055 struct vnode *vp; 1056 struct mount *mp; 1057 int error, flags, fsync_after; 1058 boolean_t res; 1059 1060 if (object == NULL) 1061 return (TRUE); 1062 res = TRUE; 1063 error = 0; 1064 VM_OBJECT_WLOCK(object); 1065 while ((backing_object = object->backing_object) != NULL) { 1066 VM_OBJECT_WLOCK(backing_object); 1067 offset += object->backing_object_offset; 1068 VM_OBJECT_WUNLOCK(object); 1069 object = backing_object; 1070 if (object->size < OFF_TO_IDX(offset + size)) 1071 size = IDX_TO_OFF(object->size) - offset; 1072 } 1073 /* 1074 * Flush pages if writing is allowed, invalidate them 1075 * if invalidation requested. Pages undergoing I/O 1076 * will be ignored by vm_object_page_remove(). 1077 * 1078 * We cannot lock the vnode and then wait for paging 1079 * to complete without deadlocking against vm_fault. 1080 * Instead we simply call vm_object_page_remove() and 1081 * allow it to block internally on a page-by-page 1082 * basis when it encounters pages undergoing async 1083 * I/O. 1084 */ 1085 if (object->type == OBJT_VNODE && 1086 (object->flags & OBJ_MIGHTBEDIRTY) != 0 && 1087 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1088 VM_OBJECT_WUNLOCK(object); 1089 (void) vn_start_write(vp, &mp, V_WAIT); 1090 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1091 if (syncio && !invalidate && offset == 0 && 1092 atop(size) == object->size) { 1093 /* 1094 * If syncing the whole mapping of the file, 1095 * it is faster to schedule all the writes in 1096 * async mode, also allowing the clustering, 1097 * and then wait for i/o to complete. 1098 */ 1099 flags = 0; 1100 fsync_after = TRUE; 1101 } else { 1102 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1103 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1104 fsync_after = FALSE; 1105 } 1106 VM_OBJECT_WLOCK(object); 1107 res = vm_object_page_clean(object, offset, offset + size, 1108 flags); 1109 VM_OBJECT_WUNLOCK(object); 1110 if (fsync_after) 1111 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1112 VOP_UNLOCK(vp, 0); 1113 vn_finished_write(mp); 1114 if (error != 0) 1115 res = FALSE; 1116 VM_OBJECT_WLOCK(object); 1117 } 1118 if ((object->type == OBJT_VNODE || 1119 object->type == OBJT_DEVICE) && invalidate) { 1120 if (object->type == OBJT_DEVICE) 1121 /* 1122 * The option OBJPR_NOTMAPPED must be passed here 1123 * because vm_object_page_remove() cannot remove 1124 * unmanaged mappings. 1125 */ 1126 flags = OBJPR_NOTMAPPED; 1127 else if (old_msync) 1128 flags = 0; 1129 else 1130 flags = OBJPR_CLEANONLY; 1131 vm_object_page_remove(object, OFF_TO_IDX(offset), 1132 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1133 } 1134 VM_OBJECT_WUNLOCK(object); 1135 return (res); 1136} 1137 1138/* 1139 * Determine whether the given advice can be applied to the object. Advice is 1140 * not applied to unmanaged pages since they never belong to page queues, and 1141 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1142 * have been mapped at most once. 1143 */ 1144static bool 1145vm_object_advice_applies(vm_object_t object, int advice) 1146{ 1147 1148 if ((object->flags & OBJ_UNMANAGED) != 0) 1149 return (false); 1150 if (advice != MADV_FREE) 1151 return (true); 1152 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && 1153 (object->flags & OBJ_ONEMAPPING) != 0); 1154} 1155 1156static void 1157vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1158 vm_size_t size) 1159{ 1160 1161 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1162 swap_pager_freespace(object, pindex, size); 1163} 1164 1165/* 1166 * vm_object_madvise: 1167 * 1168 * Implements the madvise function at the object/page level. 1169 * 1170 * MADV_WILLNEED (any object) 1171 * 1172 * Activate the specified pages if they are resident. 1173 * 1174 * MADV_DONTNEED (any object) 1175 * 1176 * Deactivate the specified pages if they are resident. 1177 * 1178 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1179 * OBJ_ONEMAPPING only) 1180 * 1181 * Deactivate and clean the specified pages if they are 1182 * resident. This permits the process to reuse the pages 1183 * without faulting or the kernel to reclaim the pages 1184 * without I/O. 1185 */ 1186void 1187vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1188 int advice) 1189{ 1190 vm_pindex_t tpindex; 1191 vm_object_t backing_object, tobject; 1192 vm_page_t m, tm; 1193 1194 if (object == NULL) 1195 return; 1196 1197relookup: 1198 VM_OBJECT_WLOCK(object); 1199 if (!vm_object_advice_applies(object, advice)) { 1200 VM_OBJECT_WUNLOCK(object); 1201 return; 1202 } 1203 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1204 tobject = object; 1205 1206 /* 1207 * If the next page isn't resident in the top-level object, we 1208 * need to search the shadow chain. When applying MADV_FREE, we 1209 * take care to release any swap space used to store 1210 * non-resident pages. 1211 */ 1212 if (m == NULL || pindex < m->pindex) { 1213 /* 1214 * Optimize a common case: if the top-level object has 1215 * no backing object, we can skip over the non-resident 1216 * range in constant time. 1217 */ 1218 if (object->backing_object == NULL) { 1219 tpindex = (m != NULL && m->pindex < end) ? 1220 m->pindex : end; 1221 vm_object_madvise_freespace(object, advice, 1222 pindex, tpindex - pindex); 1223 if ((pindex = tpindex) == end) 1224 break; 1225 goto next_page; 1226 } 1227 1228 tpindex = pindex; 1229 do { 1230 vm_object_madvise_freespace(tobject, advice, 1231 tpindex, 1); 1232 /* 1233 * Prepare to search the next object in the 1234 * chain. 1235 */ 1236 backing_object = tobject->backing_object; 1237 if (backing_object == NULL) 1238 goto next_pindex; 1239 VM_OBJECT_WLOCK(backing_object); 1240 tpindex += 1241 OFF_TO_IDX(tobject->backing_object_offset); 1242 if (tobject != object) 1243 VM_OBJECT_WUNLOCK(tobject); 1244 tobject = backing_object; 1245 if (!vm_object_advice_applies(tobject, advice)) 1246 goto next_pindex; 1247 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1248 NULL); 1249 } else { 1250next_page: 1251 tm = m; 1252 m = TAILQ_NEXT(m, listq); 1253 } 1254 1255 /* 1256 * If the page is not in a normal state, skip it. 1257 */ 1258 if (tm->valid != VM_PAGE_BITS_ALL) 1259 goto next_pindex; 1260 vm_page_lock(tm); 1261 if (tm->hold_count != 0 || tm->wire_count != 0) { 1262 vm_page_unlock(tm); 1263 goto next_pindex; 1264 } 1265 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1266 ("vm_object_madvise: page %p is fictitious", tm)); 1267 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1268 ("vm_object_madvise: page %p is not managed", tm)); 1269 if (vm_page_busied(tm)) { 1270 if (object != tobject) 1271 VM_OBJECT_WUNLOCK(tobject); 1272 VM_OBJECT_WUNLOCK(object); 1273 if (advice == MADV_WILLNEED) { 1274 /* 1275 * Reference the page before unlocking and 1276 * sleeping so that the page daemon is less 1277 * likely to reclaim it. 1278 */ 1279 vm_page_aflag_set(tm, PGA_REFERENCED); 1280 } 1281 vm_page_busy_sleep(tm, "madvpo", false); 1282 goto relookup; 1283 } 1284 vm_page_advise(tm, advice); 1285 vm_page_unlock(tm); 1286 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1287next_pindex: 1288 if (tobject != object) 1289 VM_OBJECT_WUNLOCK(tobject); 1290 } 1291 VM_OBJECT_WUNLOCK(object); 1292} 1293 1294/* 1295 * vm_object_shadow: 1296 * 1297 * Create a new object which is backed by the 1298 * specified existing object range. The source 1299 * object reference is deallocated. 1300 * 1301 * The new object and offset into that object 1302 * are returned in the source parameters. 1303 */ 1304void 1305vm_object_shadow( 1306 vm_object_t *object, /* IN/OUT */ 1307 vm_ooffset_t *offset, /* IN/OUT */ 1308 vm_size_t length) 1309{ 1310 vm_object_t source; 1311 vm_object_t result; 1312 1313 source = *object; 1314 1315 /* 1316 * Don't create the new object if the old object isn't shared. 1317 */ 1318 if (source != NULL) { 1319 VM_OBJECT_WLOCK(source); 1320 if (source->ref_count == 1 && 1321 source->handle == NULL && 1322 (source->type == OBJT_DEFAULT || 1323 source->type == OBJT_SWAP)) { 1324 VM_OBJECT_WUNLOCK(source); 1325 return; 1326 } 1327 VM_OBJECT_WUNLOCK(source); 1328 } 1329 1330 /* 1331 * Allocate a new object with the given length. 1332 */ 1333 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1334 1335 /* 1336 * The new object shadows the source object, adding a reference to it. 1337 * Our caller changes his reference to point to the new object, 1338 * removing a reference to the source object. Net result: no change 1339 * of reference count. 1340 * 1341 * Try to optimize the result object's page color when shadowing 1342 * in order to maintain page coloring consistency in the combined 1343 * shadowed object. 1344 */ 1345 result->backing_object = source; 1346 /* 1347 * Store the offset into the source object, and fix up the offset into 1348 * the new object. 1349 */ 1350 result->backing_object_offset = *offset; 1351 if (source != NULL) { 1352 VM_OBJECT_WLOCK(source); 1353 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1354 source->shadow_count++; 1355#if VM_NRESERVLEVEL > 0 1356 result->flags |= source->flags & OBJ_COLORED; 1357 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1358 ((1 << (VM_NFREEORDER - 1)) - 1); 1359#endif 1360 VM_OBJECT_WUNLOCK(source); 1361 } 1362 1363 1364 /* 1365 * Return the new things 1366 */ 1367 *offset = 0; 1368 *object = result; 1369} 1370 1371/* 1372 * vm_object_split: 1373 * 1374 * Split the pages in a map entry into a new object. This affords 1375 * easier removal of unused pages, and keeps object inheritance from 1376 * being a negative impact on memory usage. 1377 */ 1378void 1379vm_object_split(vm_map_entry_t entry) 1380{ 1381 vm_page_t m, m_next; 1382 vm_object_t orig_object, new_object, source; 1383 vm_pindex_t idx, offidxstart; 1384 vm_size_t size; 1385 1386 orig_object = entry->object.vm_object; 1387 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1388 return; 1389 if (orig_object->ref_count <= 1) 1390 return; 1391 VM_OBJECT_WUNLOCK(orig_object); 1392 1393 offidxstart = OFF_TO_IDX(entry->offset); 1394 size = atop(entry->end - entry->start); 1395 1396 /* 1397 * If swap_pager_copy() is later called, it will convert new_object 1398 * into a swap object. 1399 */ 1400 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1401 1402 /* 1403 * At this point, the new object is still private, so the order in 1404 * which the original and new objects are locked does not matter. 1405 */ 1406 VM_OBJECT_WLOCK(new_object); 1407 VM_OBJECT_WLOCK(orig_object); 1408 source = orig_object->backing_object; 1409 if (source != NULL) { 1410 VM_OBJECT_WLOCK(source); 1411 if ((source->flags & OBJ_DEAD) != 0) { 1412 VM_OBJECT_WUNLOCK(source); 1413 VM_OBJECT_WUNLOCK(orig_object); 1414 VM_OBJECT_WUNLOCK(new_object); 1415 vm_object_deallocate(new_object); 1416 VM_OBJECT_WLOCK(orig_object); 1417 return; 1418 } 1419 LIST_INSERT_HEAD(&source->shadow_head, 1420 new_object, shadow_list); 1421 source->shadow_count++; 1422 vm_object_reference_locked(source); /* for new_object */ 1423 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1424 VM_OBJECT_WUNLOCK(source); 1425 new_object->backing_object_offset = 1426 orig_object->backing_object_offset + entry->offset; 1427 new_object->backing_object = source; 1428 } 1429 if (orig_object->cred != NULL) { 1430 new_object->cred = orig_object->cred; 1431 crhold(orig_object->cred); 1432 new_object->charge = ptoa(size); 1433 KASSERT(orig_object->charge >= ptoa(size), 1434 ("orig_object->charge < 0")); 1435 orig_object->charge -= ptoa(size); 1436 } 1437retry: 1438 m = vm_page_find_least(orig_object, offidxstart); 1439 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1440 m = m_next) { 1441 m_next = TAILQ_NEXT(m, listq); 1442 1443 /* 1444 * We must wait for pending I/O to complete before we can 1445 * rename the page. 1446 * 1447 * We do not have to VM_PROT_NONE the page as mappings should 1448 * not be changed by this operation. 1449 */ 1450 if (vm_page_busied(m)) { 1451 VM_OBJECT_WUNLOCK(new_object); 1452 vm_page_lock(m); 1453 VM_OBJECT_WUNLOCK(orig_object); 1454 vm_page_busy_sleep(m, "spltwt", false); 1455 VM_OBJECT_WLOCK(orig_object); 1456 VM_OBJECT_WLOCK(new_object); 1457 goto retry; 1458 } 1459 1460 /* vm_page_rename() will dirty the page. */ 1461 if (vm_page_rename(m, new_object, idx)) { 1462 VM_OBJECT_WUNLOCK(new_object); 1463 VM_OBJECT_WUNLOCK(orig_object); 1464 vm_radix_wait(); 1465 VM_OBJECT_WLOCK(orig_object); 1466 VM_OBJECT_WLOCK(new_object); 1467 goto retry; 1468 } 1469#if VM_NRESERVLEVEL > 0 1470 /* 1471 * If some of the reservation's allocated pages remain with 1472 * the original object, then transferring the reservation to 1473 * the new object is neither particularly beneficial nor 1474 * particularly harmful as compared to leaving the reservation 1475 * with the original object. If, however, all of the 1476 * reservation's allocated pages are transferred to the new 1477 * object, then transferring the reservation is typically 1478 * beneficial. Determining which of these two cases applies 1479 * would be more costly than unconditionally renaming the 1480 * reservation. 1481 */ 1482 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1483#endif 1484 if (orig_object->type == OBJT_SWAP) 1485 vm_page_xbusy(m); 1486 } 1487 if (orig_object->type == OBJT_SWAP) { 1488 /* 1489 * swap_pager_copy() can sleep, in which case the orig_object's 1490 * and new_object's locks are released and reacquired. 1491 */ 1492 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1493 TAILQ_FOREACH(m, &new_object->memq, listq) 1494 vm_page_xunbusy(m); 1495 } 1496 VM_OBJECT_WUNLOCK(orig_object); 1497 VM_OBJECT_WUNLOCK(new_object); 1498 entry->object.vm_object = new_object; 1499 entry->offset = 0LL; 1500 vm_object_deallocate(orig_object); 1501 VM_OBJECT_WLOCK(new_object); 1502} 1503 1504#define OBSC_COLLAPSE_NOWAIT 0x0002 1505#define OBSC_COLLAPSE_WAIT 0x0004 1506 1507static vm_page_t 1508vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1509 int op) 1510{ 1511 vm_object_t backing_object; 1512 1513 VM_OBJECT_ASSERT_WLOCKED(object); 1514 backing_object = object->backing_object; 1515 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1516 1517 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1518 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1519 ("invalid ownership %p %p %p", p, object, backing_object)); 1520 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1521 return (next); 1522 if (p != NULL) 1523 vm_page_lock(p); 1524 VM_OBJECT_WUNLOCK(object); 1525 VM_OBJECT_WUNLOCK(backing_object); 1526 /* The page is only NULL when rename fails. */ 1527 if (p == NULL) 1528 vm_radix_wait(); 1529 else 1530 vm_page_busy_sleep(p, "vmocol", false); 1531 VM_OBJECT_WLOCK(object); 1532 VM_OBJECT_WLOCK(backing_object); 1533 return (TAILQ_FIRST(&backing_object->memq)); 1534} 1535 1536static bool 1537vm_object_scan_all_shadowed(vm_object_t object) 1538{ 1539 vm_object_t backing_object; 1540 vm_page_t p, pp; 1541 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1542 1543 VM_OBJECT_ASSERT_WLOCKED(object); 1544 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1545 1546 backing_object = object->backing_object; 1547 1548 /* 1549 * Initial conditions: 1550 * 1551 * We do not want to have to test for the existence of swap 1552 * pages in the backing object. XXX but with the new swapper this 1553 * would be pretty easy to do. 1554 */ 1555 if (backing_object->type != OBJT_DEFAULT && 1556 backing_object->type != OBJT_SWAP) 1557 return (false); 1558 1559 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1560 p = vm_page_find_least(backing_object, pi); 1561 ps = swap_pager_find_least(backing_object, pi); 1562 1563 /* 1564 * Only check pages inside the parent object's range and 1565 * inside the parent object's mapping of the backing object. 1566 */ 1567 for (;; pi++) { 1568 if (p != NULL && p->pindex < pi) 1569 p = TAILQ_NEXT(p, listq); 1570 if (ps < pi) 1571 ps = swap_pager_find_least(backing_object, pi); 1572 if (p == NULL && ps >= backing_object->size) 1573 break; 1574 else if (p == NULL) 1575 pi = ps; 1576 else 1577 pi = MIN(p->pindex, ps); 1578 1579 new_pindex = pi - backing_offset_index; 1580 if (new_pindex >= object->size) 1581 break; 1582 1583 /* 1584 * See if the parent has the page or if the parent's object 1585 * pager has the page. If the parent has the page but the page 1586 * is not valid, the parent's object pager must have the page. 1587 * 1588 * If this fails, the parent does not completely shadow the 1589 * object and we might as well give up now. 1590 */ 1591 pp = vm_page_lookup(object, new_pindex); 1592 if ((pp == NULL || pp->valid == 0) && 1593 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1594 return (false); 1595 } 1596 return (true); 1597} 1598 1599static bool 1600vm_object_collapse_scan(vm_object_t object, int op) 1601{ 1602 vm_object_t backing_object; 1603 vm_page_t next, p, pp; 1604 vm_pindex_t backing_offset_index, new_pindex; 1605 1606 VM_OBJECT_ASSERT_WLOCKED(object); 1607 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1608 1609 backing_object = object->backing_object; 1610 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1611 1612 /* 1613 * Initial conditions 1614 */ 1615 if ((op & OBSC_COLLAPSE_WAIT) != 0) 1616 vm_object_set_flag(backing_object, OBJ_DEAD); 1617 1618 /* 1619 * Our scan 1620 */ 1621 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1622 next = TAILQ_NEXT(p, listq); 1623 new_pindex = p->pindex - backing_offset_index; 1624 1625 /* 1626 * Check for busy page 1627 */ 1628 if (vm_page_busied(p)) { 1629 next = vm_object_collapse_scan_wait(object, p, next, op); 1630 continue; 1631 } 1632 1633 KASSERT(p->object == backing_object, 1634 ("vm_object_collapse_scan: object mismatch")); 1635 1636 if (p->pindex < backing_offset_index || 1637 new_pindex >= object->size) { 1638 if (backing_object->type == OBJT_SWAP) 1639 swap_pager_freespace(backing_object, p->pindex, 1640 1); 1641 1642 /* 1643 * Page is out of the parent object's range, we can 1644 * simply destroy it. 1645 */ 1646 vm_page_lock(p); 1647 KASSERT(!pmap_page_is_mapped(p), 1648 ("freeing mapped page %p", p)); 1649 if (p->wire_count == 0) 1650 vm_page_free(p); 1651 else 1652 vm_page_remove(p); 1653 vm_page_unlock(p); 1654 continue; 1655 } 1656 1657 pp = vm_page_lookup(object, new_pindex); 1658 if (pp != NULL && vm_page_busied(pp)) { 1659 /* 1660 * The page in the parent is busy and possibly not 1661 * (yet) valid. Until its state is finalized by the 1662 * busy bit owner, we can't tell whether it shadows the 1663 * original page. Therefore, we must either skip it 1664 * and the original (backing_object) page or wait for 1665 * its state to be finalized. 1666 * 1667 * This is due to a race with vm_fault() where we must 1668 * unbusy the original (backing_obj) page before we can 1669 * (re)lock the parent. Hence we can get here. 1670 */ 1671 next = vm_object_collapse_scan_wait(object, pp, next, 1672 op); 1673 continue; 1674 } 1675 1676 KASSERT(pp == NULL || pp->valid != 0, 1677 ("unbusy invalid page %p", pp)); 1678 1679 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1680 NULL)) { 1681 /* 1682 * The page already exists in the parent OR swap exists 1683 * for this location in the parent. Leave the parent's 1684 * page alone. Destroy the original page from the 1685 * backing object. 1686 */ 1687 if (backing_object->type == OBJT_SWAP) 1688 swap_pager_freespace(backing_object, p->pindex, 1689 1); 1690 vm_page_lock(p); 1691 KASSERT(!pmap_page_is_mapped(p), 1692 ("freeing mapped page %p", p)); 1693 if (p->wire_count == 0) 1694 vm_page_free(p); 1695 else 1696 vm_page_remove(p); 1697 vm_page_unlock(p); 1698 continue; 1699 } 1700 1701 /* 1702 * Page does not exist in parent, rename the page from the 1703 * backing object to the main object. 1704 * 1705 * If the page was mapped to a process, it can remain mapped 1706 * through the rename. vm_page_rename() will dirty the page. 1707 */ 1708 if (vm_page_rename(p, object, new_pindex)) { 1709 next = vm_object_collapse_scan_wait(object, NULL, next, 1710 op); 1711 continue; 1712 } 1713 1714 /* Use the old pindex to free the right page. */ 1715 if (backing_object->type == OBJT_SWAP) 1716 swap_pager_freespace(backing_object, 1717 new_pindex + backing_offset_index, 1); 1718 1719#if VM_NRESERVLEVEL > 0 1720 /* 1721 * Rename the reservation. 1722 */ 1723 vm_reserv_rename(p, object, backing_object, 1724 backing_offset_index); 1725#endif 1726 } 1727 return (true); 1728} 1729 1730 1731/* 1732 * this version of collapse allows the operation to occur earlier and 1733 * when paging_in_progress is true for an object... This is not a complete 1734 * operation, but should plug 99.9% of the rest of the leaks. 1735 */ 1736static void 1737vm_object_qcollapse(vm_object_t object) 1738{ 1739 vm_object_t backing_object = object->backing_object; 1740 1741 VM_OBJECT_ASSERT_WLOCKED(object); 1742 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1743 1744 if (backing_object->ref_count != 1) 1745 return; 1746 1747 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); 1748} 1749 1750/* 1751 * vm_object_collapse: 1752 * 1753 * Collapse an object with the object backing it. 1754 * Pages in the backing object are moved into the 1755 * parent, and the backing object is deallocated. 1756 */ 1757void 1758vm_object_collapse(vm_object_t object) 1759{ 1760 vm_object_t backing_object, new_backing_object; 1761 1762 VM_OBJECT_ASSERT_WLOCKED(object); 1763 1764 while (TRUE) { 1765 /* 1766 * Verify that the conditions are right for collapse: 1767 * 1768 * The object exists and the backing object exists. 1769 */ 1770 if ((backing_object = object->backing_object) == NULL) 1771 break; 1772 1773 /* 1774 * we check the backing object first, because it is most likely 1775 * not collapsable. 1776 */ 1777 VM_OBJECT_WLOCK(backing_object); 1778 if (backing_object->handle != NULL || 1779 (backing_object->type != OBJT_DEFAULT && 1780 backing_object->type != OBJT_SWAP) || 1781 (backing_object->flags & OBJ_DEAD) || 1782 object->handle != NULL || 1783 (object->type != OBJT_DEFAULT && 1784 object->type != OBJT_SWAP) || 1785 (object->flags & OBJ_DEAD)) { 1786 VM_OBJECT_WUNLOCK(backing_object); 1787 break; 1788 } 1789 1790 if (object->paging_in_progress != 0 || 1791 backing_object->paging_in_progress != 0) { 1792 vm_object_qcollapse(object); 1793 VM_OBJECT_WUNLOCK(backing_object); 1794 break; 1795 } 1796 1797 /* 1798 * We know that we can either collapse the backing object (if 1799 * the parent is the only reference to it) or (perhaps) have 1800 * the parent bypass the object if the parent happens to shadow 1801 * all the resident pages in the entire backing object. 1802 * 1803 * This is ignoring pager-backed pages such as swap pages. 1804 * vm_object_collapse_scan fails the shadowing test in this 1805 * case. 1806 */ 1807 if (backing_object->ref_count == 1) { 1808 vm_object_pip_add(object, 1); 1809 vm_object_pip_add(backing_object, 1); 1810 1811 /* 1812 * If there is exactly one reference to the backing 1813 * object, we can collapse it into the parent. 1814 */ 1815 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); 1816 1817#if VM_NRESERVLEVEL > 0 1818 /* 1819 * Break any reservations from backing_object. 1820 */ 1821 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1822 vm_reserv_break_all(backing_object); 1823#endif 1824 1825 /* 1826 * Move the pager from backing_object to object. 1827 */ 1828 if (backing_object->type == OBJT_SWAP) { 1829 /* 1830 * swap_pager_copy() can sleep, in which case 1831 * the backing_object's and object's locks are 1832 * released and reacquired. 1833 * Since swap_pager_copy() is being asked to 1834 * destroy the source, it will change the 1835 * backing_object's type to OBJT_DEFAULT. 1836 */ 1837 swap_pager_copy( 1838 backing_object, 1839 object, 1840 OFF_TO_IDX(object->backing_object_offset), TRUE); 1841 } 1842 /* 1843 * Object now shadows whatever backing_object did. 1844 * Note that the reference to 1845 * backing_object->backing_object moves from within 1846 * backing_object to within object. 1847 */ 1848 LIST_REMOVE(object, shadow_list); 1849 backing_object->shadow_count--; 1850 if (backing_object->backing_object) { 1851 VM_OBJECT_WLOCK(backing_object->backing_object); 1852 LIST_REMOVE(backing_object, shadow_list); 1853 LIST_INSERT_HEAD( 1854 &backing_object->backing_object->shadow_head, 1855 object, shadow_list); 1856 /* 1857 * The shadow_count has not changed. 1858 */ 1859 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1860 } 1861 object->backing_object = backing_object->backing_object; 1862 object->backing_object_offset += 1863 backing_object->backing_object_offset; 1864 1865 /* 1866 * Discard backing_object. 1867 * 1868 * Since the backing object has no pages, no pager left, 1869 * and no object references within it, all that is 1870 * necessary is to dispose of it. 1871 */ 1872 KASSERT(backing_object->ref_count == 1, ( 1873"backing_object %p was somehow re-referenced during collapse!", 1874 backing_object)); 1875 vm_object_pip_wakeup(backing_object); 1876 backing_object->type = OBJT_DEAD; 1877 backing_object->ref_count = 0; 1878 VM_OBJECT_WUNLOCK(backing_object); 1879 vm_object_destroy(backing_object); 1880 1881 vm_object_pip_wakeup(object); 1882 object_collapses++; 1883 } else { 1884 /* 1885 * If we do not entirely shadow the backing object, 1886 * there is nothing we can do so we give up. 1887 */ 1888 if (object->resident_page_count != object->size && 1889 !vm_object_scan_all_shadowed(object)) { 1890 VM_OBJECT_WUNLOCK(backing_object); 1891 break; 1892 } 1893 1894 /* 1895 * Make the parent shadow the next object in the 1896 * chain. Deallocating backing_object will not remove 1897 * it, since its reference count is at least 2. 1898 */ 1899 LIST_REMOVE(object, shadow_list); 1900 backing_object->shadow_count--; 1901 1902 new_backing_object = backing_object->backing_object; 1903 if ((object->backing_object = new_backing_object) != NULL) { 1904 VM_OBJECT_WLOCK(new_backing_object); 1905 LIST_INSERT_HEAD( 1906 &new_backing_object->shadow_head, 1907 object, 1908 shadow_list 1909 ); 1910 new_backing_object->shadow_count++; 1911 vm_object_reference_locked(new_backing_object); 1912 VM_OBJECT_WUNLOCK(new_backing_object); 1913 object->backing_object_offset += 1914 backing_object->backing_object_offset; 1915 } 1916 1917 /* 1918 * Drop the reference count on backing_object. Since 1919 * its ref_count was at least 2, it will not vanish. 1920 */ 1921 backing_object->ref_count--; 1922 VM_OBJECT_WUNLOCK(backing_object); 1923 object_bypasses++; 1924 } 1925 1926 /* 1927 * Try again with this object's new backing object. 1928 */ 1929 } 1930} 1931 1932/* 1933 * vm_object_page_remove: 1934 * 1935 * For the given object, either frees or invalidates each of the 1936 * specified pages. In general, a page is freed. However, if a page is 1937 * wired for any reason other than the existence of a managed, wired 1938 * mapping, then it may be invalidated but not removed from the object. 1939 * Pages are specified by the given range ["start", "end") and the option 1940 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1941 * extends from "start" to the end of the object. If the option 1942 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1943 * specified range are affected. If the option OBJPR_NOTMAPPED is 1944 * specified, then the pages within the specified range must have no 1945 * mappings. Otherwise, if this option is not specified, any mappings to 1946 * the specified pages are removed before the pages are freed or 1947 * invalidated. 1948 * 1949 * In general, this operation should only be performed on objects that 1950 * contain managed pages. There are, however, two exceptions. First, it 1951 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1952 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1953 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1954 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1955 * 1956 * The object must be locked. 1957 */ 1958void 1959vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1960 int options) 1961{ 1962 vm_page_t p, next; 1963 struct mtx *mtx; 1964 struct pglist pgl; 1965 1966 VM_OBJECT_ASSERT_WLOCKED(object); 1967 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1968 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1969 ("vm_object_page_remove: illegal options for object %p", object)); 1970 if (object->resident_page_count == 0) 1971 return; 1972 vm_object_pip_add(object, 1); 1973 TAILQ_INIT(&pgl); 1974again: 1975 p = vm_page_find_least(object, start); 1976 mtx = NULL; 1977 1978 /* 1979 * Here, the variable "p" is either (1) the page with the least pindex 1980 * greater than or equal to the parameter "start" or (2) NULL. 1981 */ 1982 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1983 next = TAILQ_NEXT(p, listq); 1984 1985 /* 1986 * If the page is wired for any reason besides the existence 1987 * of managed, wired mappings, then it cannot be freed. For 1988 * example, fictitious pages, which represent device memory, 1989 * are inherently wired and cannot be freed. They can, 1990 * however, be invalidated if the option OBJPR_CLEANONLY is 1991 * not specified. 1992 */ 1993 vm_page_change_lock(p, &mtx); 1994 if (vm_page_xbusied(p)) { 1995 VM_OBJECT_WUNLOCK(object); 1996 vm_page_busy_sleep(p, "vmopax", true); 1997 VM_OBJECT_WLOCK(object); 1998 goto again; 1999 } 2000 if (p->wire_count != 0) { 2001 if ((options & OBJPR_NOTMAPPED) == 0 && 2002 object->ref_count != 0) 2003 pmap_remove_all(p); 2004 if ((options & OBJPR_CLEANONLY) == 0) { 2005 p->valid = 0; 2006 vm_page_undirty(p); 2007 } 2008 continue; 2009 } 2010 if (vm_page_busied(p)) { 2011 VM_OBJECT_WUNLOCK(object); 2012 vm_page_busy_sleep(p, "vmopar", false); 2013 VM_OBJECT_WLOCK(object); 2014 goto again; 2015 } 2016 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2017 ("vm_object_page_remove: page %p is fictitious", p)); 2018 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 2019 if ((options & OBJPR_NOTMAPPED) == 0 && 2020 object->ref_count != 0) 2021 pmap_remove_write(p); 2022 if (p->dirty != 0) 2023 continue; 2024 } 2025 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) 2026 pmap_remove_all(p); 2027 p->flags &= ~PG_ZERO; 2028 if (vm_page_free_prep(p, false)) 2029 TAILQ_INSERT_TAIL(&pgl, p, listq); 2030 } 2031 if (mtx != NULL) 2032 mtx_unlock(mtx); 2033 vm_page_free_phys_pglist(&pgl); 2034 vm_object_pip_wakeup(object); 2035} 2036 2037/* 2038 * vm_object_page_noreuse: 2039 * 2040 * For the given object, attempt to move the specified pages to 2041 * the head of the inactive queue. This bypasses regular LRU 2042 * operation and allows the pages to be reused quickly under memory 2043 * pressure. If a page is wired for any reason, then it will not 2044 * be queued. Pages are specified by the range ["start", "end"). 2045 * As a special case, if "end" is zero, then the range extends from 2046 * "start" to the end of the object. 2047 * 2048 * This operation should only be performed on objects that 2049 * contain non-fictitious, managed pages. 2050 * 2051 * The object must be locked. 2052 */ 2053void 2054vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2055{ 2056 struct mtx *mtx; 2057 vm_page_t p, next; 2058 2059 VM_OBJECT_ASSERT_LOCKED(object); 2060 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2061 ("vm_object_page_noreuse: illegal object %p", object)); 2062 if (object->resident_page_count == 0) 2063 return; 2064 p = vm_page_find_least(object, start); 2065 2066 /* 2067 * Here, the variable "p" is either (1) the page with the least pindex 2068 * greater than or equal to the parameter "start" or (2) NULL. 2069 */ 2070 mtx = NULL; 2071 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2072 next = TAILQ_NEXT(p, listq); 2073 vm_page_change_lock(p, &mtx); 2074 vm_page_deactivate_noreuse(p); 2075 } 2076 if (mtx != NULL) 2077 mtx_unlock(mtx); 2078} 2079 2080/* 2081 * Populate the specified range of the object with valid pages. Returns 2082 * TRUE if the range is successfully populated and FALSE otherwise. 2083 * 2084 * Note: This function should be optimized to pass a larger array of 2085 * pages to vm_pager_get_pages() before it is applied to a non- 2086 * OBJT_DEVICE object. 2087 * 2088 * The object must be locked. 2089 */ 2090boolean_t 2091vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2092{ 2093 vm_page_t m; 2094 vm_pindex_t pindex; 2095 int rv; 2096 2097 VM_OBJECT_ASSERT_WLOCKED(object); 2098 for (pindex = start; pindex < end; pindex++) { 2099 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2100 if (m->valid != VM_PAGE_BITS_ALL) { 2101 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 2102 if (rv != VM_PAGER_OK) { 2103 vm_page_lock(m); 2104 vm_page_free(m); 2105 vm_page_unlock(m); 2106 break; 2107 } 2108 } 2109 /* 2110 * Keep "m" busy because a subsequent iteration may unlock 2111 * the object. 2112 */ 2113 } 2114 if (pindex > start) { 2115 m = vm_page_lookup(object, start); 2116 while (m != NULL && m->pindex < pindex) { 2117 vm_page_xunbusy(m); 2118 m = TAILQ_NEXT(m, listq); 2119 } 2120 } 2121 return (pindex == end); 2122} 2123 2124/* 2125 * Routine: vm_object_coalesce 2126 * Function: Coalesces two objects backing up adjoining 2127 * regions of memory into a single object. 2128 * 2129 * returns TRUE if objects were combined. 2130 * 2131 * NOTE: Only works at the moment if the second object is NULL - 2132 * if it's not, which object do we lock first? 2133 * 2134 * Parameters: 2135 * prev_object First object to coalesce 2136 * prev_offset Offset into prev_object 2137 * prev_size Size of reference to prev_object 2138 * next_size Size of reference to the second object 2139 * reserved Indicator that extension region has 2140 * swap accounted for 2141 * 2142 * Conditions: 2143 * The object must *not* be locked. 2144 */ 2145boolean_t 2146vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2147 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2148{ 2149 vm_pindex_t next_pindex; 2150 2151 if (prev_object == NULL) 2152 return (TRUE); 2153 VM_OBJECT_WLOCK(prev_object); 2154 if ((prev_object->type != OBJT_DEFAULT && 2155 prev_object->type != OBJT_SWAP) || 2156 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2157 VM_OBJECT_WUNLOCK(prev_object); 2158 return (FALSE); 2159 } 2160 2161 /* 2162 * Try to collapse the object first 2163 */ 2164 vm_object_collapse(prev_object); 2165 2166 /* 2167 * Can't coalesce if: . more than one reference . paged out . shadows 2168 * another object . has a copy elsewhere (any of which mean that the 2169 * pages not mapped to prev_entry may be in use anyway) 2170 */ 2171 if (prev_object->backing_object != NULL) { 2172 VM_OBJECT_WUNLOCK(prev_object); 2173 return (FALSE); 2174 } 2175 2176 prev_size >>= PAGE_SHIFT; 2177 next_size >>= PAGE_SHIFT; 2178 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2179 2180 if ((prev_object->ref_count > 1) && 2181 (prev_object->size != next_pindex)) { 2182 VM_OBJECT_WUNLOCK(prev_object); 2183 return (FALSE); 2184 } 2185 2186 /* 2187 * Account for the charge. 2188 */ 2189 if (prev_object->cred != NULL) { 2190 2191 /* 2192 * If prev_object was charged, then this mapping, 2193 * although not charged now, may become writable 2194 * later. Non-NULL cred in the object would prevent 2195 * swap reservation during enabling of the write 2196 * access, so reserve swap now. Failed reservation 2197 * cause allocation of the separate object for the map 2198 * entry, and swap reservation for this entry is 2199 * managed in appropriate time. 2200 */ 2201 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2202 prev_object->cred)) { 2203 VM_OBJECT_WUNLOCK(prev_object); 2204 return (FALSE); 2205 } 2206 prev_object->charge += ptoa(next_size); 2207 } 2208 2209 /* 2210 * Remove any pages that may still be in the object from a previous 2211 * deallocation. 2212 */ 2213 if (next_pindex < prev_object->size) { 2214 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2215 next_size, 0); 2216 if (prev_object->type == OBJT_SWAP) 2217 swap_pager_freespace(prev_object, 2218 next_pindex, next_size); 2219#if 0 2220 if (prev_object->cred != NULL) { 2221 KASSERT(prev_object->charge >= 2222 ptoa(prev_object->size - next_pindex), 2223 ("object %p overcharged 1 %jx %jx", prev_object, 2224 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2225 prev_object->charge -= ptoa(prev_object->size - 2226 next_pindex); 2227 } 2228#endif 2229 } 2230 2231 /* 2232 * Extend the object if necessary. 2233 */ 2234 if (next_pindex + next_size > prev_object->size) 2235 prev_object->size = next_pindex + next_size; 2236 2237 VM_OBJECT_WUNLOCK(prev_object); 2238 return (TRUE); 2239} 2240 2241void 2242vm_object_set_writeable_dirty(vm_object_t object) 2243{ 2244 2245 VM_OBJECT_ASSERT_WLOCKED(object); 2246 if (object->type != OBJT_VNODE) { 2247 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2248 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2249 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2250 } 2251 return; 2252 } 2253 object->generation++; 2254 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2255 return; 2256 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2257} 2258 2259/* 2260 * vm_object_unwire: 2261 * 2262 * For each page offset within the specified range of the given object, 2263 * find the highest-level page in the shadow chain and unwire it. A page 2264 * must exist at every page offset, and the highest-level page must be 2265 * wired. 2266 */ 2267void 2268vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2269 uint8_t queue) 2270{ 2271 vm_object_t tobject; 2272 vm_page_t m, tm; 2273 vm_pindex_t end_pindex, pindex, tpindex; 2274 int depth, locked_depth; 2275 2276 KASSERT((offset & PAGE_MASK) == 0, 2277 ("vm_object_unwire: offset is not page aligned")); 2278 KASSERT((length & PAGE_MASK) == 0, 2279 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2280 /* The wired count of a fictitious page never changes. */ 2281 if ((object->flags & OBJ_FICTITIOUS) != 0) 2282 return; 2283 pindex = OFF_TO_IDX(offset); 2284 end_pindex = pindex + atop(length); 2285 locked_depth = 1; 2286 VM_OBJECT_RLOCK(object); 2287 m = vm_page_find_least(object, pindex); 2288 while (pindex < end_pindex) { 2289 if (m == NULL || pindex < m->pindex) { 2290 /* 2291 * The first object in the shadow chain doesn't 2292 * contain a page at the current index. Therefore, 2293 * the page must exist in a backing object. 2294 */ 2295 tobject = object; 2296 tpindex = pindex; 2297 depth = 0; 2298 do { 2299 tpindex += 2300 OFF_TO_IDX(tobject->backing_object_offset); 2301 tobject = tobject->backing_object; 2302 KASSERT(tobject != NULL, 2303 ("vm_object_unwire: missing page")); 2304 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2305 goto next_page; 2306 depth++; 2307 if (depth == locked_depth) { 2308 locked_depth++; 2309 VM_OBJECT_RLOCK(tobject); 2310 } 2311 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2312 NULL); 2313 } else { 2314 tm = m; 2315 m = TAILQ_NEXT(m, listq); 2316 } 2317 vm_page_lock(tm); 2318 vm_page_unwire(tm, queue); 2319 vm_page_unlock(tm); 2320next_page: 2321 pindex++; 2322 } 2323 /* Release the accumulated object locks. */ 2324 for (depth = 0; depth < locked_depth; depth++) { 2325 tobject = object->backing_object; 2326 VM_OBJECT_RUNLOCK(object); 2327 object = tobject; 2328 } 2329} 2330 2331struct vnode * 2332vm_object_vnode(vm_object_t object) 2333{ 2334 2335 VM_OBJECT_ASSERT_LOCKED(object); 2336 if (object->type == OBJT_VNODE) 2337 return (object->handle); 2338 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) 2339 return (object->un_pager.swp.swp_tmpfs); 2340 return (NULL); 2341} 2342 2343static int 2344sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2345{ 2346 struct kinfo_vmobject *kvo; 2347 char *fullpath, *freepath; 2348 struct vnode *vp; 2349 struct vattr va; 2350 vm_object_t obj; 2351 vm_page_t m; 2352 int count, error; 2353 2354 if (req->oldptr == NULL) { 2355 /* 2356 * If an old buffer has not been provided, generate an 2357 * estimate of the space needed for a subsequent call. 2358 */ 2359 mtx_lock(&vm_object_list_mtx); 2360 count = 0; 2361 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2362 if (obj->type == OBJT_DEAD) 2363 continue; 2364 count++; 2365 } 2366 mtx_unlock(&vm_object_list_mtx); 2367 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2368 count * 11 / 10)); 2369 } 2370 2371 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2372 error = 0; 2373 2374 /* 2375 * VM objects are type stable and are never removed from the 2376 * list once added. This allows us to safely read obj->object_list 2377 * after reacquiring the VM object lock. 2378 */ 2379 mtx_lock(&vm_object_list_mtx); 2380 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2381 if (obj->type == OBJT_DEAD) 2382 continue; 2383 VM_OBJECT_RLOCK(obj); 2384 if (obj->type == OBJT_DEAD) { 2385 VM_OBJECT_RUNLOCK(obj); 2386 continue; 2387 } 2388 mtx_unlock(&vm_object_list_mtx); 2389 kvo->kvo_size = ptoa(obj->size); 2390 kvo->kvo_resident = obj->resident_page_count; 2391 kvo->kvo_ref_count = obj->ref_count; 2392 kvo->kvo_shadow_count = obj->shadow_count; 2393 kvo->kvo_memattr = obj->memattr; 2394 kvo->kvo_active = 0; 2395 kvo->kvo_inactive = 0; 2396 TAILQ_FOREACH(m, &obj->memq, listq) { 2397 /* 2398 * A page may belong to the object but be 2399 * dequeued and set to PQ_NONE while the 2400 * object lock is not held. This makes the 2401 * reads of m->queue below racy, and we do not 2402 * count pages set to PQ_NONE. However, this 2403 * sysctl is only meant to give an 2404 * approximation of the system anyway. 2405 */ 2406 if (vm_page_active(m)) 2407 kvo->kvo_active++; 2408 else if (vm_page_inactive(m)) 2409 kvo->kvo_inactive++; 2410 } 2411 2412 kvo->kvo_vn_fileid = 0; 2413 kvo->kvo_vn_fsid = 0; 2414 freepath = NULL; 2415 fullpath = ""; 2416 vp = NULL; 2417 switch (obj->type) { 2418 case OBJT_DEFAULT: 2419 kvo->kvo_type = KVME_TYPE_DEFAULT; 2420 break; 2421 case OBJT_VNODE: 2422 kvo->kvo_type = KVME_TYPE_VNODE; 2423 vp = obj->handle; 2424 vref(vp); 2425 break; 2426 case OBJT_SWAP: 2427 kvo->kvo_type = KVME_TYPE_SWAP; 2428 break; 2429 case OBJT_DEVICE: 2430 kvo->kvo_type = KVME_TYPE_DEVICE; 2431 break; 2432 case OBJT_PHYS: 2433 kvo->kvo_type = KVME_TYPE_PHYS; 2434 break; 2435 case OBJT_DEAD: 2436 kvo->kvo_type = KVME_TYPE_DEAD; 2437 break; 2438 case OBJT_SG: 2439 kvo->kvo_type = KVME_TYPE_SG; 2440 break; 2441 case OBJT_MGTDEVICE: 2442 kvo->kvo_type = KVME_TYPE_MGTDEVICE; 2443 break; 2444 default: 2445 kvo->kvo_type = KVME_TYPE_UNKNOWN; 2446 break; 2447 } 2448 VM_OBJECT_RUNLOCK(obj); 2449 if (vp != NULL) { 2450 vn_fullpath(curthread, vp, &fullpath, &freepath); 2451 vn_lock(vp, LK_SHARED | LK_RETRY); 2452 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2453 kvo->kvo_vn_fileid = va.va_fileid; 2454 kvo->kvo_vn_fsid = va.va_fsid; 2455 } 2456 vput(vp); 2457 } 2458 2459 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2460 if (freepath != NULL) 2461 free(freepath, M_TEMP); 2462 2463 /* Pack record size down */ 2464 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2465 + strlen(kvo->kvo_path) + 1; 2466 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2467 sizeof(uint64_t)); 2468 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2469 mtx_lock(&vm_object_list_mtx); 2470 if (error) 2471 break; 2472 } 2473 mtx_unlock(&vm_object_list_mtx); 2474 free(kvo, M_TEMP); 2475 return (error); 2476} 2477SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2478 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2479 "List of VM objects"); 2480 2481#include "opt_ddb.h" 2482#ifdef DDB 2483#include <sys/kernel.h> 2484 2485#include <sys/cons.h> 2486 2487#include <ddb/ddb.h> 2488 2489static int 2490_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2491{ 2492 vm_map_t tmpm; 2493 vm_map_entry_t tmpe; 2494 vm_object_t obj; 2495 int entcount; 2496 2497 if (map == 0) 2498 return 0; 2499 2500 if (entry == 0) { 2501 tmpe = map->header.next; 2502 entcount = map->nentries; 2503 while (entcount-- && (tmpe != &map->header)) { 2504 if (_vm_object_in_map(map, object, tmpe)) { 2505 return 1; 2506 } 2507 tmpe = tmpe->next; 2508 } 2509 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2510 tmpm = entry->object.sub_map; 2511 tmpe = tmpm->header.next; 2512 entcount = tmpm->nentries; 2513 while (entcount-- && tmpe != &tmpm->header) { 2514 if (_vm_object_in_map(tmpm, object, tmpe)) { 2515 return 1; 2516 } 2517 tmpe = tmpe->next; 2518 } 2519 } else if ((obj = entry->object.vm_object) != NULL) { 2520 for (; obj; obj = obj->backing_object) 2521 if (obj == object) { 2522 return 1; 2523 } 2524 } 2525 return 0; 2526} 2527 2528static int 2529vm_object_in_map(vm_object_t object) 2530{ 2531 struct proc *p; 2532 2533 /* sx_slock(&allproc_lock); */ 2534 FOREACH_PROC_IN_SYSTEM(p) { 2535 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2536 continue; 2537 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2538 /* sx_sunlock(&allproc_lock); */ 2539 return 1; 2540 } 2541 } 2542 /* sx_sunlock(&allproc_lock); */ 2543 if (_vm_object_in_map(kernel_map, object, 0)) 2544 return 1; 2545 return 0; 2546} 2547 2548DB_SHOW_COMMAND(vmochk, vm_object_check) 2549{ 2550 vm_object_t object; 2551 2552 /* 2553 * make sure that internal objs are in a map somewhere 2554 * and none have zero ref counts. 2555 */ 2556 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2557 if (object->handle == NULL && 2558 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2559 if (object->ref_count == 0) { 2560 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2561 (long)object->size); 2562 } 2563 if (!vm_object_in_map(object)) { 2564 db_printf( 2565 "vmochk: internal obj is not in a map: " 2566 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2567 object->ref_count, (u_long)object->size, 2568 (u_long)object->size, 2569 (void *)object->backing_object); 2570 } 2571 } 2572 } 2573} 2574 2575/* 2576 * vm_object_print: [ debug ] 2577 */ 2578DB_SHOW_COMMAND(object, vm_object_print_static) 2579{ 2580 /* XXX convert args. */ 2581 vm_object_t object = (vm_object_t)addr; 2582 boolean_t full = have_addr; 2583 2584 vm_page_t p; 2585 2586 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2587#define count was_count 2588 2589 int count; 2590 2591 if (object == NULL) 2592 return; 2593 2594 db_iprintf( 2595 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2596 object, (int)object->type, (uintmax_t)object->size, 2597 object->resident_page_count, object->ref_count, object->flags, 2598 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2599 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2600 object->shadow_count, 2601 object->backing_object ? object->backing_object->ref_count : 0, 2602 object->backing_object, (uintmax_t)object->backing_object_offset); 2603 2604 if (!full) 2605 return; 2606 2607 db_indent += 2; 2608 count = 0; 2609 TAILQ_FOREACH(p, &object->memq, listq) { 2610 if (count == 0) 2611 db_iprintf("memory:="); 2612 else if (count == 6) { 2613 db_printf("\n"); 2614 db_iprintf(" ..."); 2615 count = 0; 2616 } else 2617 db_printf(","); 2618 count++; 2619 2620 db_printf("(off=0x%jx,page=0x%jx)", 2621 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2622 } 2623 if (count != 0) 2624 db_printf("\n"); 2625 db_indent -= 2; 2626} 2627 2628/* XXX. */ 2629#undef count 2630 2631/* XXX need this non-static entry for calling from vm_map_print. */ 2632void 2633vm_object_print( 2634 /* db_expr_t */ long addr, 2635 boolean_t have_addr, 2636 /* db_expr_t */ long count, 2637 char *modif) 2638{ 2639 vm_object_print_static(addr, have_addr, count, modif); 2640} 2641 2642DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2643{ 2644 vm_object_t object; 2645 vm_pindex_t fidx; 2646 vm_paddr_t pa; 2647 vm_page_t m, prev_m; 2648 int rcount, nl, c; 2649 2650 nl = 0; 2651 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2652 db_printf("new object: %p\n", (void *)object); 2653 if (nl > 18) { 2654 c = cngetc(); 2655 if (c != ' ') 2656 return; 2657 nl = 0; 2658 } 2659 nl++; 2660 rcount = 0; 2661 fidx = 0; 2662 pa = -1; 2663 TAILQ_FOREACH(m, &object->memq, listq) { 2664 if (m->pindex > 128) 2665 break; 2666 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2667 prev_m->pindex + 1 != m->pindex) { 2668 if (rcount) { 2669 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2670 (long)fidx, rcount, (long)pa); 2671 if (nl > 18) { 2672 c = cngetc(); 2673 if (c != ' ') 2674 return; 2675 nl = 0; 2676 } 2677 nl++; 2678 rcount = 0; 2679 } 2680 } 2681 if (rcount && 2682 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2683 ++rcount; 2684 continue; 2685 } 2686 if (rcount) { 2687 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2688 (long)fidx, rcount, (long)pa); 2689 if (nl > 18) { 2690 c = cngetc(); 2691 if (c != ' ') 2692 return; 2693 nl = 0; 2694 } 2695 nl++; 2696 } 2697 fidx = m->pindex; 2698 pa = VM_PAGE_TO_PHYS(m); 2699 rcount = 1; 2700 } 2701 if (rcount) { 2702 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2703 (long)fidx, rcount, (long)pa); 2704 if (nl > 18) { 2705 c = cngetc(); 2706 if (c != ' ') 2707 return; 2708 nl = 0; 2709 } 2710 nl++; 2711 } 2712 } 2713} 2714#endif /* DDB */ 2715