1/*- 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61/* 62 * Virtual memory object module. 63 */ 64 65#include <sys/cdefs.h> 66__FBSDID("$FreeBSD: stable/10/sys/vm/vm_object.c 321677 2017-07-29 08:24:51Z kib $"); 67 68#include "opt_vm.h" 69 70#include <sys/param.h> 71#include <sys/systm.h> 72#include <sys/lock.h> 73#include <sys/mman.h> 74#include <sys/mount.h> 75#include <sys/kernel.h> 76#include <sys/sysctl.h> 77#include <sys/mutex.h> 78#include <sys/proc.h> /* for curproc, pageproc */ 79#include <sys/socket.h> 80#include <sys/resourcevar.h> 81#include <sys/rwlock.h> 82#include <sys/user.h> 83#include <sys/vnode.h> 84#include <sys/vmmeter.h> 85#include <sys/sx.h> 86 87#include <vm/vm.h> 88#include <vm/vm_param.h> 89#include <vm/pmap.h> 90#include <vm/vm_map.h> 91#include <vm/vm_object.h> 92#include <vm/vm_page.h> 93#include <vm/vm_pageout.h> 94#include <vm/vm_pager.h> 95#include <vm/swap_pager.h> 96#include <vm/vm_kern.h> 97#include <vm/vm_extern.h> 98#include <vm/vm_radix.h> 99#include <vm/vm_reserv.h> 100#include <vm/uma.h> 101 102static int old_msync; 103SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 104 "Use old (insecure) msync behavior"); 105 106static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 107 int pagerflags, int flags, boolean_t *clearobjflags, 108 boolean_t *eio); 109static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 110 boolean_t *clearobjflags); 111static void vm_object_qcollapse(vm_object_t object); 112static void vm_object_vndeallocate(vm_object_t object); 113 114/* 115 * Virtual memory objects maintain the actual data 116 * associated with allocated virtual memory. A given 117 * page of memory exists within exactly one object. 118 * 119 * An object is only deallocated when all "references" 120 * are given up. Only one "reference" to a given 121 * region of an object should be writeable. 122 * 123 * Associated with each object is a list of all resident 124 * memory pages belonging to that object; this list is 125 * maintained by the "vm_page" module, and locked by the object's 126 * lock. 127 * 128 * Each object also records a "pager" routine which is 129 * used to retrieve (and store) pages to the proper backing 130 * storage. In addition, objects may be backed by other 131 * objects from which they were virtual-copied. 132 * 133 * The only items within the object structure which are 134 * modified after time of creation are: 135 * reference count locked by object's lock 136 * pager routine locked by object's lock 137 * 138 */ 139 140struct object_q vm_object_list; 141struct mtx vm_object_list_mtx; /* lock for object list and count */ 142 143struct vm_object kernel_object_store; 144struct vm_object kmem_object_store; 145 146static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 147 "VM object stats"); 148 149static long object_collapses; 150SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 151 &object_collapses, 0, "VM object collapses"); 152 153static long object_bypasses; 154SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 155 &object_bypasses, 0, "VM object bypasses"); 156 157static uma_zone_t obj_zone; 158 159static int vm_object_zinit(void *mem, int size, int flags); 160 161#ifdef INVARIANTS 162static void vm_object_zdtor(void *mem, int size, void *arg); 163 164static void 165vm_object_zdtor(void *mem, int size, void *arg) 166{ 167 vm_object_t object; 168 169 object = (vm_object_t)mem; 170 KASSERT(object->ref_count == 0, 171 ("object %p ref_count = %d", object, object->ref_count)); 172 KASSERT(TAILQ_EMPTY(&object->memq), 173 ("object %p has resident pages in its memq", object)); 174 KASSERT(vm_radix_is_empty(&object->rtree), 175 ("object %p has resident pages in its trie", object)); 176#if VM_NRESERVLEVEL > 0 177 KASSERT(LIST_EMPTY(&object->rvq), 178 ("object %p has reservations", 179 object)); 180#endif 181 KASSERT(vm_object_cache_is_empty(object), 182 ("object %p has cached pages", 183 object)); 184 KASSERT(object->paging_in_progress == 0, 185 ("object %p paging_in_progress = %d", 186 object, object->paging_in_progress)); 187 KASSERT(object->resident_page_count == 0, 188 ("object %p resident_page_count = %d", 189 object, object->resident_page_count)); 190 KASSERT(object->shadow_count == 0, 191 ("object %p shadow_count = %d", 192 object, object->shadow_count)); 193 KASSERT(object->type == OBJT_DEAD, 194 ("object %p has non-dead type %d", 195 object, object->type)); 196} 197#endif 198 199static int 200vm_object_zinit(void *mem, int size, int flags) 201{ 202 vm_object_t object; 203 204 object = (vm_object_t)mem; 205 bzero(&object->lock, sizeof(object->lock)); 206 rw_init_flags(&object->lock, "vm object", RW_DUPOK); 207 208 /* These are true for any object that has been freed */ 209 object->type = OBJT_DEAD; 210 object->ref_count = 0; 211 object->rtree.rt_root = 0; 212 object->rtree.rt_flags = 0; 213 object->paging_in_progress = 0; 214 object->resident_page_count = 0; 215 object->shadow_count = 0; 216 object->cache.rt_root = 0; 217 object->cache.rt_flags = 0; 218 219 mtx_lock(&vm_object_list_mtx); 220 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 221 mtx_unlock(&vm_object_list_mtx); 222 return (0); 223} 224 225static void 226_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 227{ 228 229 TAILQ_INIT(&object->memq); 230 LIST_INIT(&object->shadow_head); 231 232 object->type = type; 233 switch (type) { 234 case OBJT_DEAD: 235 panic("_vm_object_allocate: can't create OBJT_DEAD"); 236 case OBJT_DEFAULT: 237 case OBJT_SWAP: 238 object->flags = OBJ_ONEMAPPING; 239 break; 240 case OBJT_DEVICE: 241 case OBJT_SG: 242 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 243 break; 244 case OBJT_MGTDEVICE: 245 object->flags = OBJ_FICTITIOUS; 246 break; 247 case OBJT_PHYS: 248 object->flags = OBJ_UNMANAGED; 249 break; 250 case OBJT_VNODE: 251 object->flags = 0; 252 break; 253 default: 254 panic("_vm_object_allocate: type %d is undefined", type); 255 } 256 object->size = size; 257 object->generation = 1; 258 object->ref_count = 1; 259 object->memattr = VM_MEMATTR_DEFAULT; 260 object->cred = NULL; 261 object->charge = 0; 262 object->handle = NULL; 263 object->backing_object = NULL; 264 object->backing_object_offset = (vm_ooffset_t) 0; 265#if VM_NRESERVLEVEL > 0 266 LIST_INIT(&object->rvq); 267#endif 268} 269 270/* 271 * vm_object_init: 272 * 273 * Initialize the VM objects module. 274 */ 275void 276vm_object_init(void) 277{ 278 TAILQ_INIT(&vm_object_list); 279 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 280 281 rw_init(&kernel_object->lock, "kernel vm object"); 282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 283 kernel_object); 284#if VM_NRESERVLEVEL > 0 285 kernel_object->flags |= OBJ_COLORED; 286 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 287#endif 288 289 rw_init(&kmem_object->lock, "kmem vm object"); 290 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 291 kmem_object); 292#if VM_NRESERVLEVEL > 0 293 kmem_object->flags |= OBJ_COLORED; 294 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 295#endif 296 297 /* 298 * The lock portion of struct vm_object must be type stable due 299 * to vm_pageout_fallback_object_lock locking a vm object 300 * without holding any references to it. 301 */ 302 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 303#ifdef INVARIANTS 304 vm_object_zdtor, 305#else 306 NULL, 307#endif 308 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 309 310 vm_radix_init(); 311} 312 313void 314vm_object_clear_flag(vm_object_t object, u_short bits) 315{ 316 317 VM_OBJECT_ASSERT_WLOCKED(object); 318 object->flags &= ~bits; 319} 320 321/* 322 * Sets the default memory attribute for the specified object. Pages 323 * that are allocated to this object are by default assigned this memory 324 * attribute. 325 * 326 * Presently, this function must be called before any pages are allocated 327 * to the object. In the future, this requirement may be relaxed for 328 * "default" and "swap" objects. 329 */ 330int 331vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 332{ 333 334 VM_OBJECT_ASSERT_WLOCKED(object); 335 switch (object->type) { 336 case OBJT_DEFAULT: 337 case OBJT_DEVICE: 338 case OBJT_MGTDEVICE: 339 case OBJT_PHYS: 340 case OBJT_SG: 341 case OBJT_SWAP: 342 case OBJT_VNODE: 343 if (!TAILQ_EMPTY(&object->memq)) 344 return (KERN_FAILURE); 345 break; 346 case OBJT_DEAD: 347 return (KERN_INVALID_ARGUMENT); 348 default: 349 panic("vm_object_set_memattr: object %p is of undefined type", 350 object); 351 } 352 object->memattr = memattr; 353 return (KERN_SUCCESS); 354} 355 356void 357vm_object_pip_add(vm_object_t object, short i) 358{ 359 360 VM_OBJECT_ASSERT_WLOCKED(object); 361 object->paging_in_progress += i; 362} 363 364void 365vm_object_pip_subtract(vm_object_t object, short i) 366{ 367 368 VM_OBJECT_ASSERT_WLOCKED(object); 369 object->paging_in_progress -= i; 370} 371 372void 373vm_object_pip_wakeup(vm_object_t object) 374{ 375 376 VM_OBJECT_ASSERT_WLOCKED(object); 377 object->paging_in_progress--; 378 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 379 vm_object_clear_flag(object, OBJ_PIPWNT); 380 wakeup(object); 381 } 382} 383 384void 385vm_object_pip_wakeupn(vm_object_t object, short i) 386{ 387 388 VM_OBJECT_ASSERT_WLOCKED(object); 389 if (i) 390 object->paging_in_progress -= i; 391 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 392 vm_object_clear_flag(object, OBJ_PIPWNT); 393 wakeup(object); 394 } 395} 396 397void 398vm_object_pip_wait(vm_object_t object, char *waitid) 399{ 400 401 VM_OBJECT_ASSERT_WLOCKED(object); 402 while (object->paging_in_progress) { 403 object->flags |= OBJ_PIPWNT; 404 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 405 } 406} 407 408/* 409 * vm_object_allocate: 410 * 411 * Returns a new object with the given size. 412 */ 413vm_object_t 414vm_object_allocate(objtype_t type, vm_pindex_t size) 415{ 416 vm_object_t object; 417 418 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 419 _vm_object_allocate(type, size, object); 420 return (object); 421} 422 423 424/* 425 * vm_object_reference: 426 * 427 * Gets another reference to the given object. Note: OBJ_DEAD 428 * objects can be referenced during final cleaning. 429 */ 430void 431vm_object_reference(vm_object_t object) 432{ 433 if (object == NULL) 434 return; 435 VM_OBJECT_WLOCK(object); 436 vm_object_reference_locked(object); 437 VM_OBJECT_WUNLOCK(object); 438} 439 440/* 441 * vm_object_reference_locked: 442 * 443 * Gets another reference to the given object. 444 * 445 * The object must be locked. 446 */ 447void 448vm_object_reference_locked(vm_object_t object) 449{ 450 struct vnode *vp; 451 452 VM_OBJECT_ASSERT_WLOCKED(object); 453 object->ref_count++; 454 if (object->type == OBJT_VNODE) { 455 vp = object->handle; 456 vref(vp); 457 } 458} 459 460/* 461 * Handle deallocating an object of type OBJT_VNODE. 462 */ 463static void 464vm_object_vndeallocate(vm_object_t object) 465{ 466 struct vnode *vp = (struct vnode *) object->handle; 467 468 VM_OBJECT_ASSERT_WLOCKED(object); 469 KASSERT(object->type == OBJT_VNODE, 470 ("vm_object_vndeallocate: not a vnode object")); 471 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 472#ifdef INVARIANTS 473 if (object->ref_count == 0) { 474 vprint("vm_object_vndeallocate", vp); 475 panic("vm_object_vndeallocate: bad object reference count"); 476 } 477#endif 478 479 /* 480 * The test for text of vp vnode does not need a bypass to 481 * reach right VV_TEXT there, since it is obtained from 482 * object->handle. 483 */ 484 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 485 object->ref_count--; 486 VM_OBJECT_WUNLOCK(object); 487 /* vrele may need the vnode lock. */ 488 vrele(vp); 489 } else { 490 vhold(vp); 491 VM_OBJECT_WUNLOCK(object); 492 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 493 vdrop(vp); 494 VM_OBJECT_WLOCK(object); 495 object->ref_count--; 496 if (object->type == OBJT_DEAD) { 497 VM_OBJECT_WUNLOCK(object); 498 VOP_UNLOCK(vp, 0); 499 } else { 500 if (object->ref_count == 0) 501 VOP_UNSET_TEXT(vp); 502 VM_OBJECT_WUNLOCK(object); 503 vput(vp); 504 } 505 } 506} 507 508/* 509 * vm_object_deallocate: 510 * 511 * Release a reference to the specified object, 512 * gained either through a vm_object_allocate 513 * or a vm_object_reference call. When all references 514 * are gone, storage associated with this object 515 * may be relinquished. 516 * 517 * No object may be locked. 518 */ 519void 520vm_object_deallocate(vm_object_t object) 521{ 522 vm_object_t temp; 523 struct vnode *vp; 524 525 while (object != NULL) { 526 VM_OBJECT_WLOCK(object); 527 if (object->type == OBJT_VNODE) { 528 vm_object_vndeallocate(object); 529 return; 530 } 531 532 KASSERT(object->ref_count != 0, 533 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 534 535 /* 536 * If the reference count goes to 0 we start calling 537 * vm_object_terminate() on the object chain. 538 * A ref count of 1 may be a special case depending on the 539 * shadow count being 0 or 1. 540 */ 541 object->ref_count--; 542 if (object->ref_count > 1) { 543 VM_OBJECT_WUNLOCK(object); 544 return; 545 } else if (object->ref_count == 1) { 546 if (object->type == OBJT_SWAP && 547 (object->flags & OBJ_TMPFS) != 0) { 548 vp = object->un_pager.swp.swp_tmpfs; 549 vhold(vp); 550 VM_OBJECT_WUNLOCK(object); 551 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 552 VM_OBJECT_WLOCK(object); 553 if (object->type == OBJT_DEAD || 554 object->ref_count != 1) { 555 VM_OBJECT_WUNLOCK(object); 556 VOP_UNLOCK(vp, 0); 557 vdrop(vp); 558 return; 559 } 560 if ((object->flags & OBJ_TMPFS) != 0) 561 VOP_UNSET_TEXT(vp); 562 VOP_UNLOCK(vp, 0); 563 vdrop(vp); 564 } 565 if (object->shadow_count == 0 && 566 object->handle == NULL && 567 (object->type == OBJT_DEFAULT || 568 (object->type == OBJT_SWAP && 569 (object->flags & OBJ_TMPFS_NODE) == 0))) { 570 vm_object_set_flag(object, OBJ_ONEMAPPING); 571 } else if ((object->shadow_count == 1) && 572 (object->handle == NULL) && 573 (object->type == OBJT_DEFAULT || 574 object->type == OBJT_SWAP)) { 575 vm_object_t robject; 576 577 robject = LIST_FIRST(&object->shadow_head); 578 KASSERT(robject != NULL, 579 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 580 object->ref_count, 581 object->shadow_count)); 582 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 583 ("shadowed tmpfs v_object %p", object)); 584 if (!VM_OBJECT_TRYWLOCK(robject)) { 585 /* 586 * Avoid a potential deadlock. 587 */ 588 object->ref_count++; 589 VM_OBJECT_WUNLOCK(object); 590 /* 591 * More likely than not the thread 592 * holding robject's lock has lower 593 * priority than the current thread. 594 * Let the lower priority thread run. 595 */ 596 pause("vmo_de", 1); 597 continue; 598 } 599 /* 600 * Collapse object into its shadow unless its 601 * shadow is dead. In that case, object will 602 * be deallocated by the thread that is 603 * deallocating its shadow. 604 */ 605 if ((robject->flags & OBJ_DEAD) == 0 && 606 (robject->handle == NULL) && 607 (robject->type == OBJT_DEFAULT || 608 robject->type == OBJT_SWAP)) { 609 610 robject->ref_count++; 611retry: 612 if (robject->paging_in_progress) { 613 VM_OBJECT_WUNLOCK(object); 614 vm_object_pip_wait(robject, 615 "objde1"); 616 temp = robject->backing_object; 617 if (object == temp) { 618 VM_OBJECT_WLOCK(object); 619 goto retry; 620 } 621 } else if (object->paging_in_progress) { 622 VM_OBJECT_WUNLOCK(robject); 623 object->flags |= OBJ_PIPWNT; 624 VM_OBJECT_SLEEP(object, object, 625 PDROP | PVM, "objde2", 0); 626 VM_OBJECT_WLOCK(robject); 627 temp = robject->backing_object; 628 if (object == temp) { 629 VM_OBJECT_WLOCK(object); 630 goto retry; 631 } 632 } else 633 VM_OBJECT_WUNLOCK(object); 634 635 if (robject->ref_count == 1) { 636 robject->ref_count--; 637 object = robject; 638 goto doterm; 639 } 640 object = robject; 641 vm_object_collapse(object); 642 VM_OBJECT_WUNLOCK(object); 643 continue; 644 } 645 VM_OBJECT_WUNLOCK(robject); 646 } 647 VM_OBJECT_WUNLOCK(object); 648 return; 649 } 650doterm: 651 temp = object->backing_object; 652 if (temp != NULL) { 653 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 654 ("shadowed tmpfs v_object 2 %p", object)); 655 VM_OBJECT_WLOCK(temp); 656 LIST_REMOVE(object, shadow_list); 657 temp->shadow_count--; 658 VM_OBJECT_WUNLOCK(temp); 659 object->backing_object = NULL; 660 } 661 /* 662 * Don't double-terminate, we could be in a termination 663 * recursion due to the terminate having to sync data 664 * to disk. 665 */ 666 if ((object->flags & OBJ_DEAD) == 0) 667 vm_object_terminate(object); 668 else 669 VM_OBJECT_WUNLOCK(object); 670 object = temp; 671 } 672} 673 674/* 675 * vm_object_destroy removes the object from the global object list 676 * and frees the space for the object. 677 */ 678void 679vm_object_destroy(vm_object_t object) 680{ 681 682 /* 683 * Release the allocation charge. 684 */ 685 if (object->cred != NULL) { 686 swap_release_by_cred(object->charge, object->cred); 687 object->charge = 0; 688 crfree(object->cred); 689 object->cred = NULL; 690 } 691 692 /* 693 * Free the space for the object. 694 */ 695 uma_zfree(obj_zone, object); 696} 697 698/* 699 * vm_object_terminate actually destroys the specified object, freeing 700 * up all previously used resources. 701 * 702 * The object must be locked. 703 * This routine may block. 704 */ 705void 706vm_object_terminate(vm_object_t object) 707{ 708 vm_page_t p, p_next; 709 710 VM_OBJECT_ASSERT_WLOCKED(object); 711 712 /* 713 * Make sure no one uses us. 714 */ 715 vm_object_set_flag(object, OBJ_DEAD); 716 717 /* 718 * wait for the pageout daemon to be done with the object 719 */ 720 vm_object_pip_wait(object, "objtrm"); 721 722 KASSERT(!object->paging_in_progress, 723 ("vm_object_terminate: pageout in progress")); 724 725 /* 726 * Clean and free the pages, as appropriate. All references to the 727 * object are gone, so we don't need to lock it. 728 */ 729 if (object->type == OBJT_VNODE) { 730 struct vnode *vp = (struct vnode *)object->handle; 731 732 /* 733 * Clean pages and flush buffers. 734 */ 735 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 736 VM_OBJECT_WUNLOCK(object); 737 738 vinvalbuf(vp, V_SAVE, 0, 0); 739 740 BO_LOCK(&vp->v_bufobj); 741 vp->v_bufobj.bo_flag |= BO_DEAD; 742 BO_UNLOCK(&vp->v_bufobj); 743 744 VM_OBJECT_WLOCK(object); 745 } 746 747 KASSERT(object->ref_count == 0, 748 ("vm_object_terminate: object with references, ref_count=%d", 749 object->ref_count)); 750 751 /* 752 * Free any remaining pageable pages. This also removes them from the 753 * paging queues. However, don't free wired pages, just remove them 754 * from the object. Rather than incrementally removing each page from 755 * the object, the page and object are reset to any empty state. 756 */ 757 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 758 vm_page_assert_unbusied(p); 759 vm_page_lock(p); 760 /* 761 * Optimize the page's removal from the object by resetting 762 * its "object" field. Specifically, if the page is not 763 * wired, then the effect of this assignment is that 764 * vm_page_free()'s call to vm_page_remove() will return 765 * immediately without modifying the page or the object. 766 */ 767 p->object = NULL; 768 if (p->wire_count == 0) { 769 vm_page_free(p); 770 PCPU_INC(cnt.v_pfree); 771 } 772 vm_page_unlock(p); 773 } 774 /* 775 * If the object contained any pages, then reset it to an empty state. 776 * None of the object's fields, including "resident_page_count", were 777 * modified by the preceding loop. 778 */ 779 if (object->resident_page_count != 0) { 780 vm_radix_reclaim_allnodes(&object->rtree); 781 TAILQ_INIT(&object->memq); 782 object->resident_page_count = 0; 783 if (object->type == OBJT_VNODE) 784 vdrop(object->handle); 785 } 786 787#if VM_NRESERVLEVEL > 0 788 if (__predict_false(!LIST_EMPTY(&object->rvq))) 789 vm_reserv_break_all(object); 790#endif 791 if (__predict_false(!vm_object_cache_is_empty(object))) 792 vm_page_cache_free(object, 0, 0); 793 794 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 795 object->type == OBJT_SWAP, 796 ("%s: non-swap obj %p has cred", __func__, object)); 797 798 /* 799 * Let the pager know object is dead. 800 */ 801 vm_pager_deallocate(object); 802 VM_OBJECT_WUNLOCK(object); 803 804 vm_object_destroy(object); 805} 806 807/* 808 * Make the page read-only so that we can clear the object flags. However, if 809 * this is a nosync mmap then the object is likely to stay dirty so do not 810 * mess with the page and do not clear the object flags. Returns TRUE if the 811 * page should be flushed, and FALSE otherwise. 812 */ 813static boolean_t 814vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 815{ 816 817 /* 818 * If we have been asked to skip nosync pages and this is a 819 * nosync page, skip it. Note that the object flags were not 820 * cleared in this case so we do not have to set them. 821 */ 822 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 823 *clearobjflags = FALSE; 824 return (FALSE); 825 } else { 826 pmap_remove_write(p); 827 return (p->dirty != 0); 828 } 829} 830 831/* 832 * vm_object_page_clean 833 * 834 * Clean all dirty pages in the specified range of object. Leaves page 835 * on whatever queue it is currently on. If NOSYNC is set then do not 836 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 837 * leaving the object dirty. 838 * 839 * When stuffing pages asynchronously, allow clustering. XXX we need a 840 * synchronous clustering mode implementation. 841 * 842 * Odd semantics: if start == end, we clean everything. 843 * 844 * The object must be locked. 845 * 846 * Returns FALSE if some page from the range was not written, as 847 * reported by the pager, and TRUE otherwise. 848 */ 849boolean_t 850vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 851 int flags) 852{ 853 vm_page_t np, p; 854 vm_pindex_t pi, tend, tstart; 855 int curgeneration, n, pagerflags; 856 boolean_t clearobjflags, eio, res; 857 858 VM_OBJECT_ASSERT_WLOCKED(object); 859 860 /* 861 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 862 * objects. The check below prevents the function from 863 * operating on non-vnode objects. 864 */ 865 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 866 object->resident_page_count == 0) 867 return (TRUE); 868 869 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 870 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 871 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 872 873 tstart = OFF_TO_IDX(start); 874 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 875 clearobjflags = tstart == 0 && tend >= object->size; 876 res = TRUE; 877 878rescan: 879 curgeneration = object->generation; 880 881 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 882 pi = p->pindex; 883 if (pi >= tend) 884 break; 885 np = TAILQ_NEXT(p, listq); 886 if (p->valid == 0) 887 continue; 888 if (vm_page_sleep_if_busy(p, "vpcwai")) { 889 if (object->generation != curgeneration) { 890 if ((flags & OBJPC_SYNC) != 0) 891 goto rescan; 892 else 893 clearobjflags = FALSE; 894 } 895 np = vm_page_find_least(object, pi); 896 continue; 897 } 898 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 899 continue; 900 901 n = vm_object_page_collect_flush(object, p, pagerflags, 902 flags, &clearobjflags, &eio); 903 if (eio) { 904 res = FALSE; 905 clearobjflags = FALSE; 906 } 907 if (object->generation != curgeneration) { 908 if ((flags & OBJPC_SYNC) != 0) 909 goto rescan; 910 else 911 clearobjflags = FALSE; 912 } 913 914 /* 915 * If the VOP_PUTPAGES() did a truncated write, so 916 * that even the first page of the run is not fully 917 * written, vm_pageout_flush() returns 0 as the run 918 * length. Since the condition that caused truncated 919 * write may be permanent, e.g. exhausted free space, 920 * accepting n == 0 would cause an infinite loop. 921 * 922 * Forwarding the iterator leaves the unwritten page 923 * behind, but there is not much we can do there if 924 * filesystem refuses to write it. 925 */ 926 if (n == 0) { 927 n = 1; 928 clearobjflags = FALSE; 929 } 930 np = vm_page_find_least(object, pi + n); 931 } 932#if 0 933 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 934#endif 935 936 if (clearobjflags) 937 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 938 return (res); 939} 940 941static int 942vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 943 int flags, boolean_t *clearobjflags, boolean_t *eio) 944{ 945 vm_page_t ma[vm_pageout_page_count], p_first, tp; 946 int count, i, mreq, runlen; 947 948 vm_page_lock_assert(p, MA_NOTOWNED); 949 VM_OBJECT_ASSERT_WLOCKED(object); 950 951 count = 1; 952 mreq = 0; 953 954 for (tp = p; count < vm_pageout_page_count; count++) { 955 tp = vm_page_next(tp); 956 if (tp == NULL || vm_page_busied(tp)) 957 break; 958 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 959 break; 960 } 961 962 for (p_first = p; count < vm_pageout_page_count; count++) { 963 tp = vm_page_prev(p_first); 964 if (tp == NULL || vm_page_busied(tp)) 965 break; 966 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 967 break; 968 p_first = tp; 969 mreq++; 970 } 971 972 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 973 ma[i] = tp; 974 975 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 976 return (runlen); 977} 978 979/* 980 * Note that there is absolutely no sense in writing out 981 * anonymous objects, so we track down the vnode object 982 * to write out. 983 * We invalidate (remove) all pages from the address space 984 * for semantic correctness. 985 * 986 * If the backing object is a device object with unmanaged pages, then any 987 * mappings to the specified range of pages must be removed before this 988 * function is called. 989 * 990 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 991 * may start out with a NULL object. 992 */ 993boolean_t 994vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 995 boolean_t syncio, boolean_t invalidate) 996{ 997 vm_object_t backing_object; 998 struct vnode *vp; 999 struct mount *mp; 1000 int error, flags, fsync_after; 1001 boolean_t res; 1002 1003 if (object == NULL) 1004 return (TRUE); 1005 res = TRUE; 1006 error = 0; 1007 VM_OBJECT_WLOCK(object); 1008 while ((backing_object = object->backing_object) != NULL) { 1009 VM_OBJECT_WLOCK(backing_object); 1010 offset += object->backing_object_offset; 1011 VM_OBJECT_WUNLOCK(object); 1012 object = backing_object; 1013 if (object->size < OFF_TO_IDX(offset + size)) 1014 size = IDX_TO_OFF(object->size) - offset; 1015 } 1016 /* 1017 * Flush pages if writing is allowed, invalidate them 1018 * if invalidation requested. Pages undergoing I/O 1019 * will be ignored by vm_object_page_remove(). 1020 * 1021 * We cannot lock the vnode and then wait for paging 1022 * to complete without deadlocking against vm_fault. 1023 * Instead we simply call vm_object_page_remove() and 1024 * allow it to block internally on a page-by-page 1025 * basis when it encounters pages undergoing async 1026 * I/O. 1027 */ 1028 if (object->type == OBJT_VNODE && 1029 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1030 vp = object->handle; 1031 VM_OBJECT_WUNLOCK(object); 1032 (void) vn_start_write(vp, &mp, V_WAIT); 1033 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1034 if (syncio && !invalidate && offset == 0 && 1035 OFF_TO_IDX(size) == object->size) { 1036 /* 1037 * If syncing the whole mapping of the file, 1038 * it is faster to schedule all the writes in 1039 * async mode, also allowing the clustering, 1040 * and then wait for i/o to complete. 1041 */ 1042 flags = 0; 1043 fsync_after = TRUE; 1044 } else { 1045 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1046 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1047 fsync_after = FALSE; 1048 } 1049 VM_OBJECT_WLOCK(object); 1050 res = vm_object_page_clean(object, offset, offset + size, 1051 flags); 1052 VM_OBJECT_WUNLOCK(object); 1053 if (fsync_after) 1054 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1055 VOP_UNLOCK(vp, 0); 1056 vn_finished_write(mp); 1057 if (error != 0) 1058 res = FALSE; 1059 VM_OBJECT_WLOCK(object); 1060 } 1061 if ((object->type == OBJT_VNODE || 1062 object->type == OBJT_DEVICE) && invalidate) { 1063 if (object->type == OBJT_DEVICE) 1064 /* 1065 * The option OBJPR_NOTMAPPED must be passed here 1066 * because vm_object_page_remove() cannot remove 1067 * unmanaged mappings. 1068 */ 1069 flags = OBJPR_NOTMAPPED; 1070 else if (old_msync) 1071 flags = 0; 1072 else 1073 flags = OBJPR_CLEANONLY; 1074 vm_object_page_remove(object, OFF_TO_IDX(offset), 1075 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1076 } 1077 VM_OBJECT_WUNLOCK(object); 1078 return (res); 1079} 1080 1081/* 1082 * vm_object_madvise: 1083 * 1084 * Implements the madvise function at the object/page level. 1085 * 1086 * MADV_WILLNEED (any object) 1087 * 1088 * Activate the specified pages if they are resident. 1089 * 1090 * MADV_DONTNEED (any object) 1091 * 1092 * Deactivate the specified pages if they are resident. 1093 * 1094 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1095 * OBJ_ONEMAPPING only) 1096 * 1097 * Deactivate and clean the specified pages if they are 1098 * resident. This permits the process to reuse the pages 1099 * without faulting or the kernel to reclaim the pages 1100 * without I/O. 1101 */ 1102void 1103vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1104 int advise) 1105{ 1106 vm_pindex_t tpindex; 1107 vm_object_t backing_object, tobject; 1108 vm_page_t m; 1109 1110 if (object == NULL) 1111 return; 1112 VM_OBJECT_WLOCK(object); 1113 /* 1114 * Locate and adjust resident pages 1115 */ 1116 for (; pindex < end; pindex += 1) { 1117relookup: 1118 tobject = object; 1119 tpindex = pindex; 1120shadowlookup: 1121 /* 1122 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1123 * and those pages must be OBJ_ONEMAPPING. 1124 */ 1125 if (advise == MADV_FREE) { 1126 if ((tobject->type != OBJT_DEFAULT && 1127 tobject->type != OBJT_SWAP) || 1128 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1129 goto unlock_tobject; 1130 } 1131 } else if ((tobject->flags & OBJ_UNMANAGED) != 0) 1132 goto unlock_tobject; 1133 m = vm_page_lookup(tobject, tpindex); 1134 if (m == NULL && advise == MADV_WILLNEED) { 1135 /* 1136 * If the page is cached, reactivate it. 1137 */ 1138 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1139 VM_ALLOC_NOBUSY); 1140 } 1141 if (m == NULL) { 1142 /* 1143 * There may be swap even if there is no backing page 1144 */ 1145 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1146 swap_pager_freespace(tobject, tpindex, 1); 1147 /* 1148 * next object 1149 */ 1150 backing_object = tobject->backing_object; 1151 if (backing_object == NULL) 1152 goto unlock_tobject; 1153 VM_OBJECT_WLOCK(backing_object); 1154 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1155 if (tobject != object) 1156 VM_OBJECT_WUNLOCK(tobject); 1157 tobject = backing_object; 1158 goto shadowlookup; 1159 } else if (m->valid != VM_PAGE_BITS_ALL) 1160 goto unlock_tobject; 1161 /* 1162 * If the page is not in a normal state, skip it. 1163 */ 1164 vm_page_lock(m); 1165 if (m->hold_count != 0 || m->wire_count != 0) { 1166 vm_page_unlock(m); 1167 goto unlock_tobject; 1168 } 1169 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1170 ("vm_object_madvise: page %p is fictitious", m)); 1171 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1172 ("vm_object_madvise: page %p is not managed", m)); 1173 if (vm_page_busied(m)) { 1174 if (advise == MADV_WILLNEED) { 1175 /* 1176 * Reference the page before unlocking and 1177 * sleeping so that the page daemon is less 1178 * likely to reclaim it. 1179 */ 1180 vm_page_aflag_set(m, PGA_REFERENCED); 1181 } 1182 if (object != tobject) 1183 VM_OBJECT_WUNLOCK(object); 1184 VM_OBJECT_WUNLOCK(tobject); 1185 vm_page_busy_sleep(m, "madvpo", false); 1186 VM_OBJECT_WLOCK(object); 1187 goto relookup; 1188 } 1189 if (advise == MADV_WILLNEED) { 1190 vm_page_activate(m); 1191 } else { 1192 vm_page_advise(m, advise); 1193 } 1194 vm_page_unlock(m); 1195 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1196 swap_pager_freespace(tobject, tpindex, 1); 1197unlock_tobject: 1198 if (tobject != object) 1199 VM_OBJECT_WUNLOCK(tobject); 1200 } 1201 VM_OBJECT_WUNLOCK(object); 1202} 1203 1204/* 1205 * vm_object_shadow: 1206 * 1207 * Create a new object which is backed by the 1208 * specified existing object range. The source 1209 * object reference is deallocated. 1210 * 1211 * The new object and offset into that object 1212 * are returned in the source parameters. 1213 */ 1214void 1215vm_object_shadow( 1216 vm_object_t *object, /* IN/OUT */ 1217 vm_ooffset_t *offset, /* IN/OUT */ 1218 vm_size_t length) 1219{ 1220 vm_object_t source; 1221 vm_object_t result; 1222 1223 source = *object; 1224 1225 /* 1226 * Don't create the new object if the old object isn't shared. 1227 */ 1228 if (source != NULL) { 1229 VM_OBJECT_WLOCK(source); 1230 if (source->ref_count == 1 && 1231 source->handle == NULL && 1232 (source->type == OBJT_DEFAULT || 1233 source->type == OBJT_SWAP)) { 1234 VM_OBJECT_WUNLOCK(source); 1235 return; 1236 } 1237 VM_OBJECT_WUNLOCK(source); 1238 } 1239 1240 /* 1241 * Allocate a new object with the given length. 1242 */ 1243 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1244 1245 /* 1246 * The new object shadows the source object, adding a reference to it. 1247 * Our caller changes his reference to point to the new object, 1248 * removing a reference to the source object. Net result: no change 1249 * of reference count. 1250 * 1251 * Try to optimize the result object's page color when shadowing 1252 * in order to maintain page coloring consistency in the combined 1253 * shadowed object. 1254 */ 1255 result->backing_object = source; 1256 /* 1257 * Store the offset into the source object, and fix up the offset into 1258 * the new object. 1259 */ 1260 result->backing_object_offset = *offset; 1261 if (source != NULL) { 1262 VM_OBJECT_WLOCK(source); 1263 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1264 source->shadow_count++; 1265#if VM_NRESERVLEVEL > 0 1266 result->flags |= source->flags & OBJ_COLORED; 1267 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1268 ((1 << (VM_NFREEORDER - 1)) - 1); 1269#endif 1270 VM_OBJECT_WUNLOCK(source); 1271 } 1272 1273 1274 /* 1275 * Return the new things 1276 */ 1277 *offset = 0; 1278 *object = result; 1279} 1280 1281/* 1282 * vm_object_split: 1283 * 1284 * Split the pages in a map entry into a new object. This affords 1285 * easier removal of unused pages, and keeps object inheritance from 1286 * being a negative impact on memory usage. 1287 */ 1288void 1289vm_object_split(vm_map_entry_t entry) 1290{ 1291 vm_page_t m, m_next; 1292 vm_object_t orig_object, new_object, source; 1293 vm_pindex_t idx, offidxstart; 1294 vm_size_t size; 1295 1296 orig_object = entry->object.vm_object; 1297 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1298 return; 1299 if (orig_object->ref_count <= 1) 1300 return; 1301 VM_OBJECT_WUNLOCK(orig_object); 1302 1303 offidxstart = OFF_TO_IDX(entry->offset); 1304 size = atop(entry->end - entry->start); 1305 1306 /* 1307 * If swap_pager_copy() is later called, it will convert new_object 1308 * into a swap object. 1309 */ 1310 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1311 1312 /* 1313 * At this point, the new object is still private, so the order in 1314 * which the original and new objects are locked does not matter. 1315 */ 1316 VM_OBJECT_WLOCK(new_object); 1317 VM_OBJECT_WLOCK(orig_object); 1318 source = orig_object->backing_object; 1319 if (source != NULL) { 1320 VM_OBJECT_WLOCK(source); 1321 if ((source->flags & OBJ_DEAD) != 0) { 1322 VM_OBJECT_WUNLOCK(source); 1323 VM_OBJECT_WUNLOCK(orig_object); 1324 VM_OBJECT_WUNLOCK(new_object); 1325 vm_object_deallocate(new_object); 1326 VM_OBJECT_WLOCK(orig_object); 1327 return; 1328 } 1329 LIST_INSERT_HEAD(&source->shadow_head, 1330 new_object, shadow_list); 1331 source->shadow_count++; 1332 vm_object_reference_locked(source); /* for new_object */ 1333 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1334 VM_OBJECT_WUNLOCK(source); 1335 new_object->backing_object_offset = 1336 orig_object->backing_object_offset + entry->offset; 1337 new_object->backing_object = source; 1338 } 1339 if (orig_object->cred != NULL) { 1340 new_object->cred = orig_object->cred; 1341 crhold(orig_object->cred); 1342 new_object->charge = ptoa(size); 1343 KASSERT(orig_object->charge >= ptoa(size), 1344 ("orig_object->charge < 0")); 1345 orig_object->charge -= ptoa(size); 1346 } 1347retry: 1348 m = vm_page_find_least(orig_object, offidxstart); 1349 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1350 m = m_next) { 1351 m_next = TAILQ_NEXT(m, listq); 1352 1353 /* 1354 * We must wait for pending I/O to complete before we can 1355 * rename the page. 1356 * 1357 * We do not have to VM_PROT_NONE the page as mappings should 1358 * not be changed by this operation. 1359 */ 1360 if (vm_page_busied(m)) { 1361 VM_OBJECT_WUNLOCK(new_object); 1362 vm_page_lock(m); 1363 VM_OBJECT_WUNLOCK(orig_object); 1364 vm_page_busy_sleep(m, "spltwt", false); 1365 VM_OBJECT_WLOCK(orig_object); 1366 VM_OBJECT_WLOCK(new_object); 1367 goto retry; 1368 } 1369 1370 /* vm_page_rename() will handle dirty and cache. */ 1371 if (vm_page_rename(m, new_object, idx)) { 1372 VM_OBJECT_WUNLOCK(new_object); 1373 VM_OBJECT_WUNLOCK(orig_object); 1374 VM_WAIT; 1375 VM_OBJECT_WLOCK(orig_object); 1376 VM_OBJECT_WLOCK(new_object); 1377 goto retry; 1378 } 1379#if VM_NRESERVLEVEL > 0 1380 /* 1381 * If some of the reservation's allocated pages remain with 1382 * the original object, then transferring the reservation to 1383 * the new object is neither particularly beneficial nor 1384 * particularly harmful as compared to leaving the reservation 1385 * with the original object. If, however, all of the 1386 * reservation's allocated pages are transferred to the new 1387 * object, then transferring the reservation is typically 1388 * beneficial. Determining which of these two cases applies 1389 * would be more costly than unconditionally renaming the 1390 * reservation. 1391 */ 1392 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1393#endif 1394 if (orig_object->type == OBJT_SWAP) 1395 vm_page_xbusy(m); 1396 } 1397 if (orig_object->type == OBJT_SWAP) { 1398 /* 1399 * swap_pager_copy() can sleep, in which case the orig_object's 1400 * and new_object's locks are released and reacquired. 1401 */ 1402 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1403 TAILQ_FOREACH(m, &new_object->memq, listq) 1404 vm_page_xunbusy(m); 1405 1406 /* 1407 * Transfer any cached pages from orig_object to new_object. 1408 * If swap_pager_copy() found swapped out pages within the 1409 * specified range of orig_object, then it changed 1410 * new_object's type to OBJT_SWAP when it transferred those 1411 * pages to new_object. Otherwise, new_object's type 1412 * should still be OBJT_DEFAULT and orig_object should not 1413 * contain any cached pages within the specified range. 1414 */ 1415 if (__predict_false(!vm_object_cache_is_empty(orig_object))) 1416 vm_page_cache_transfer(orig_object, offidxstart, 1417 new_object); 1418 } 1419 VM_OBJECT_WUNLOCK(orig_object); 1420 VM_OBJECT_WUNLOCK(new_object); 1421 entry->object.vm_object = new_object; 1422 entry->offset = 0LL; 1423 vm_object_deallocate(orig_object); 1424 VM_OBJECT_WLOCK(new_object); 1425} 1426 1427#define OBSC_TEST_ALL_SHADOWED 0x0001 1428#define OBSC_COLLAPSE_NOWAIT 0x0002 1429#define OBSC_COLLAPSE_WAIT 0x0004 1430 1431static vm_page_t 1432vm_object_backing_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1433 int op) 1434{ 1435 vm_object_t backing_object; 1436 1437 VM_OBJECT_ASSERT_WLOCKED(object); 1438 backing_object = object->backing_object; 1439 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1440 1441 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1442 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1443 ("invalid ownership %p %p %p", p, object, backing_object)); 1444 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1445 return (next); 1446 if (p != NULL) 1447 vm_page_lock(p); 1448 VM_OBJECT_WUNLOCK(object); 1449 VM_OBJECT_WUNLOCK(backing_object); 1450 if (p == NULL) 1451 VM_WAIT; 1452 else 1453 vm_page_busy_sleep(p, "vmocol", false); 1454 VM_OBJECT_WLOCK(object); 1455 VM_OBJECT_WLOCK(backing_object); 1456 return (TAILQ_FIRST(&backing_object->memq)); 1457} 1458 1459static bool 1460vm_object_backing_scan(vm_object_t object, int op) 1461{ 1462 vm_object_t backing_object; 1463 vm_page_t next, p, pp; 1464 vm_pindex_t backing_offset_index, new_pindex; 1465 1466 VM_OBJECT_ASSERT_WLOCKED(object); 1467 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1468 1469 backing_object = object->backing_object; 1470 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1471 1472 /* 1473 * Initial conditions 1474 */ 1475 if (op & OBSC_TEST_ALL_SHADOWED) { 1476 /* 1477 * We do not want to have to test for the existence of cache 1478 * or swap pages in the backing object. XXX but with the 1479 * new swapper this would be pretty easy to do. 1480 * 1481 * XXX what about anonymous MAP_SHARED memory that hasn't 1482 * been ZFOD faulted yet? If we do not test for this, the 1483 * shadow test may succeed! XXX 1484 */ 1485 if (backing_object->type != OBJT_DEFAULT) { 1486 return (false); 1487 } 1488 } 1489 if (op & OBSC_COLLAPSE_WAIT) { 1490 vm_object_set_flag(backing_object, OBJ_DEAD); 1491 } 1492 1493 /* 1494 * Our scan 1495 */ 1496 p = TAILQ_FIRST(&backing_object->memq); 1497 while (p) { 1498 next = TAILQ_NEXT(p, listq); 1499 new_pindex = p->pindex - backing_offset_index; 1500 if (op & OBSC_TEST_ALL_SHADOWED) { 1501 /* 1502 * Ignore pages outside the parent object's range 1503 * and outside the parent object's mapping of the 1504 * backing object. 1505 * 1506 * Note that we do not busy the backing object's 1507 * page. 1508 */ 1509 if (p->pindex < backing_offset_index || 1510 new_pindex >= object->size) { 1511 p = next; 1512 continue; 1513 } 1514 1515 /* 1516 * See if the parent has the page or if the parent's 1517 * object pager has the page. If the parent has the 1518 * page but the page is not valid, the parent's 1519 * object pager must have the page. 1520 * 1521 * If this fails, the parent does not completely shadow 1522 * the object and we might as well give up now. 1523 */ 1524 1525 pp = vm_page_lookup(object, new_pindex); 1526 if ((pp == NULL || pp->valid == 0) && 1527 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1528 return (false); 1529 } 1530 1531 /* 1532 * Check for busy page 1533 */ 1534 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1535 if (vm_page_busied(p)) { 1536 p = vm_object_backing_scan_wait(object, p, 1537 next, op); 1538 continue; 1539 } 1540 1541 KASSERT(p->object == backing_object, 1542 ("vm_object_backing_scan: object mismatch")); 1543 1544 if (p->pindex < backing_offset_index || 1545 new_pindex >= object->size) { 1546 if (backing_object->type == OBJT_SWAP) 1547 swap_pager_freespace(backing_object, 1548 p->pindex, 1); 1549 1550 /* 1551 * Page is out of the parent object's range, we 1552 * can simply destroy it. 1553 */ 1554 vm_page_lock(p); 1555 KASSERT(!pmap_page_is_mapped(p), 1556 ("freeing mapped page %p", p)); 1557 if (p->wire_count == 0) 1558 vm_page_free(p); 1559 else 1560 vm_page_remove(p); 1561 vm_page_unlock(p); 1562 p = next; 1563 continue; 1564 } 1565 1566 pp = vm_page_lookup(object, new_pindex); 1567 if (pp != NULL && vm_page_busied(pp)) { 1568 /* 1569 * The page in the parent is busy and 1570 * possibly not (yet) valid. Until 1571 * its state is finalized by the busy 1572 * bit owner, we can't tell whether it 1573 * shadows the original page. 1574 * Therefore, we must either skip it 1575 * and the original (backing_object) 1576 * page or wait for its state to be 1577 * finalized. 1578 * 1579 * This is due to a race with vm_fault() 1580 * where we must unbusy the original 1581 * (backing_obj) page before we can 1582 * (re)lock the parent. Hence we can 1583 * get here. 1584 */ 1585 p = vm_object_backing_scan_wait(object, pp, 1586 next, op); 1587 continue; 1588 } 1589 1590 KASSERT(pp == NULL || pp->valid != 0, 1591 ("unbusy invalid page %p", pp)); 1592 1593 if (pp != NULL || vm_pager_has_page(object, 1594 new_pindex, NULL, NULL)) { 1595 /* 1596 * The page already exists in the 1597 * parent OR swap exists for this 1598 * location in the parent. Leave the 1599 * parent's page alone. Destroy the 1600 * original page from the backing 1601 * object. 1602 */ 1603 if (backing_object->type == OBJT_SWAP) 1604 swap_pager_freespace(backing_object, 1605 p->pindex, 1); 1606 vm_page_lock(p); 1607 KASSERT(!pmap_page_is_mapped(p), 1608 ("freeing mapped page %p", p)); 1609 if (p->wire_count == 0) 1610 vm_page_free(p); 1611 else 1612 vm_page_remove(p); 1613 vm_page_unlock(p); 1614 p = next; 1615 continue; 1616 } 1617 1618 /* 1619 * Page does not exist in parent, rename the 1620 * page from the backing object to the main object. 1621 * 1622 * If the page was mapped to a process, it can remain 1623 * mapped through the rename. 1624 * vm_page_rename() will handle dirty and cache. 1625 */ 1626 if (vm_page_rename(p, object, new_pindex)) { 1627 p = vm_object_backing_scan_wait(object, NULL, 1628 next, op); 1629 continue; 1630 } 1631 1632 /* Use the old pindex to free the right page. */ 1633 if (backing_object->type == OBJT_SWAP) 1634 swap_pager_freespace(backing_object, 1635 new_pindex + backing_offset_index, 1); 1636 1637#if VM_NRESERVLEVEL > 0 1638 /* 1639 * Rename the reservation. 1640 */ 1641 vm_reserv_rename(p, object, backing_object, 1642 backing_offset_index); 1643#endif 1644 } 1645 p = next; 1646 } 1647 return (true); 1648} 1649 1650 1651/* 1652 * this version of collapse allows the operation to occur earlier and 1653 * when paging_in_progress is true for an object... This is not a complete 1654 * operation, but should plug 99.9% of the rest of the leaks. 1655 */ 1656static void 1657vm_object_qcollapse(vm_object_t object) 1658{ 1659 vm_object_t backing_object = object->backing_object; 1660 1661 VM_OBJECT_ASSERT_WLOCKED(object); 1662 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1663 1664 if (backing_object->ref_count != 1) 1665 return; 1666 1667 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1668} 1669 1670/* 1671 * vm_object_collapse: 1672 * 1673 * Collapse an object with the object backing it. 1674 * Pages in the backing object are moved into the 1675 * parent, and the backing object is deallocated. 1676 */ 1677void 1678vm_object_collapse(vm_object_t object) 1679{ 1680 vm_object_t backing_object, new_backing_object; 1681 1682 VM_OBJECT_ASSERT_WLOCKED(object); 1683 1684 while (TRUE) { 1685 /* 1686 * Verify that the conditions are right for collapse: 1687 * 1688 * The object exists and the backing object exists. 1689 */ 1690 if ((backing_object = object->backing_object) == NULL) 1691 break; 1692 1693 /* 1694 * we check the backing object first, because it is most likely 1695 * not collapsable. 1696 */ 1697 VM_OBJECT_WLOCK(backing_object); 1698 if (backing_object->handle != NULL || 1699 (backing_object->type != OBJT_DEFAULT && 1700 backing_object->type != OBJT_SWAP) || 1701 (backing_object->flags & OBJ_DEAD) || 1702 object->handle != NULL || 1703 (object->type != OBJT_DEFAULT && 1704 object->type != OBJT_SWAP) || 1705 (object->flags & OBJ_DEAD)) { 1706 VM_OBJECT_WUNLOCK(backing_object); 1707 break; 1708 } 1709 1710 if (object->paging_in_progress != 0 || 1711 backing_object->paging_in_progress != 0) { 1712 vm_object_qcollapse(object); 1713 VM_OBJECT_WUNLOCK(backing_object); 1714 break; 1715 } 1716 1717 /* 1718 * We know that we can either collapse the backing object (if 1719 * the parent is the only reference to it) or (perhaps) have 1720 * the parent bypass the object if the parent happens to shadow 1721 * all the resident pages in the entire backing object. 1722 * 1723 * This is ignoring pager-backed pages such as swap pages. 1724 * vm_object_backing_scan fails the shadowing test in this 1725 * case. 1726 */ 1727 if (backing_object->ref_count == 1) { 1728 vm_object_pip_add(object, 1); 1729 vm_object_pip_add(backing_object, 1); 1730 1731 /* 1732 * If there is exactly one reference to the backing 1733 * object, we can collapse it into the parent. 1734 */ 1735 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1736 1737#if VM_NRESERVLEVEL > 0 1738 /* 1739 * Break any reservations from backing_object. 1740 */ 1741 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1742 vm_reserv_break_all(backing_object); 1743#endif 1744 1745 /* 1746 * Move the pager from backing_object to object. 1747 */ 1748 if (backing_object->type == OBJT_SWAP) { 1749 /* 1750 * swap_pager_copy() can sleep, in which case 1751 * the backing_object's and object's locks are 1752 * released and reacquired. 1753 * Since swap_pager_copy() is being asked to 1754 * destroy the source, it will change the 1755 * backing_object's type to OBJT_DEFAULT. 1756 */ 1757 swap_pager_copy( 1758 backing_object, 1759 object, 1760 OFF_TO_IDX(object->backing_object_offset), TRUE); 1761 1762 /* 1763 * Free any cached pages from backing_object. 1764 */ 1765 if (__predict_false( 1766 !vm_object_cache_is_empty(backing_object))) 1767 vm_page_cache_free(backing_object, 0, 0); 1768 } 1769 /* 1770 * Object now shadows whatever backing_object did. 1771 * Note that the reference to 1772 * backing_object->backing_object moves from within 1773 * backing_object to within object. 1774 */ 1775 LIST_REMOVE(object, shadow_list); 1776 backing_object->shadow_count--; 1777 if (backing_object->backing_object) { 1778 VM_OBJECT_WLOCK(backing_object->backing_object); 1779 LIST_REMOVE(backing_object, shadow_list); 1780 LIST_INSERT_HEAD( 1781 &backing_object->backing_object->shadow_head, 1782 object, shadow_list); 1783 /* 1784 * The shadow_count has not changed. 1785 */ 1786 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1787 } 1788 object->backing_object = backing_object->backing_object; 1789 object->backing_object_offset += 1790 backing_object->backing_object_offset; 1791 1792 /* 1793 * Discard backing_object. 1794 * 1795 * Since the backing object has no pages, no pager left, 1796 * and no object references within it, all that is 1797 * necessary is to dispose of it. 1798 */ 1799 KASSERT(backing_object->ref_count == 1, ( 1800"backing_object %p was somehow re-referenced during collapse!", 1801 backing_object)); 1802 vm_object_pip_wakeup(backing_object); 1803 backing_object->type = OBJT_DEAD; 1804 backing_object->ref_count = 0; 1805 VM_OBJECT_WUNLOCK(backing_object); 1806 vm_object_destroy(backing_object); 1807 1808 vm_object_pip_wakeup(object); 1809 object_collapses++; 1810 } else { 1811 /* 1812 * If we do not entirely shadow the backing object, 1813 * there is nothing we can do so we give up. 1814 */ 1815 if (object->resident_page_count != object->size && 1816 !vm_object_backing_scan(object, 1817 OBSC_TEST_ALL_SHADOWED)) { 1818 VM_OBJECT_WUNLOCK(backing_object); 1819 break; 1820 } 1821 1822 /* 1823 * Make the parent shadow the next object in the 1824 * chain. Deallocating backing_object will not remove 1825 * it, since its reference count is at least 2. 1826 */ 1827 LIST_REMOVE(object, shadow_list); 1828 backing_object->shadow_count--; 1829 1830 new_backing_object = backing_object->backing_object; 1831 if ((object->backing_object = new_backing_object) != NULL) { 1832 VM_OBJECT_WLOCK(new_backing_object); 1833 LIST_INSERT_HEAD( 1834 &new_backing_object->shadow_head, 1835 object, 1836 shadow_list 1837 ); 1838 new_backing_object->shadow_count++; 1839 vm_object_reference_locked(new_backing_object); 1840 VM_OBJECT_WUNLOCK(new_backing_object); 1841 object->backing_object_offset += 1842 backing_object->backing_object_offset; 1843 } 1844 1845 /* 1846 * Drop the reference count on backing_object. Since 1847 * its ref_count was at least 2, it will not vanish. 1848 */ 1849 backing_object->ref_count--; 1850 VM_OBJECT_WUNLOCK(backing_object); 1851 object_bypasses++; 1852 } 1853 1854 /* 1855 * Try again with this object's new backing object. 1856 */ 1857 } 1858} 1859 1860/* 1861 * vm_object_page_remove: 1862 * 1863 * For the given object, either frees or invalidates each of the 1864 * specified pages. In general, a page is freed. However, if a page is 1865 * wired for any reason other than the existence of a managed, wired 1866 * mapping, then it may be invalidated but not removed from the object. 1867 * Pages are specified by the given range ["start", "end") and the option 1868 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1869 * extends from "start" to the end of the object. If the option 1870 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1871 * specified range are affected. If the option OBJPR_NOTMAPPED is 1872 * specified, then the pages within the specified range must have no 1873 * mappings. Otherwise, if this option is not specified, any mappings to 1874 * the specified pages are removed before the pages are freed or 1875 * invalidated. 1876 * 1877 * In general, this operation should only be performed on objects that 1878 * contain managed pages. There are, however, two exceptions. First, it 1879 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1880 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1881 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1882 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1883 * 1884 * The object must be locked. 1885 */ 1886void 1887vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1888 int options) 1889{ 1890 vm_page_t p, next; 1891 1892 VM_OBJECT_ASSERT_WLOCKED(object); 1893 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1894 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1895 ("vm_object_page_remove: illegal options for object %p", object)); 1896 if (object->resident_page_count == 0) 1897 goto skipmemq; 1898 vm_object_pip_add(object, 1); 1899again: 1900 p = vm_page_find_least(object, start); 1901 1902 /* 1903 * Here, the variable "p" is either (1) the page with the least pindex 1904 * greater than or equal to the parameter "start" or (2) NULL. 1905 */ 1906 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1907 next = TAILQ_NEXT(p, listq); 1908 1909 /* 1910 * If the page is wired for any reason besides the existence 1911 * of managed, wired mappings, then it cannot be freed. For 1912 * example, fictitious pages, which represent device memory, 1913 * are inherently wired and cannot be freed. They can, 1914 * however, be invalidated if the option OBJPR_CLEANONLY is 1915 * not specified. 1916 */ 1917 vm_page_lock(p); 1918 if (vm_page_xbusied(p)) { 1919 VM_OBJECT_WUNLOCK(object); 1920 vm_page_busy_sleep(p, "vmopax", true); 1921 VM_OBJECT_WLOCK(object); 1922 goto again; 1923 } 1924 if (p->wire_count != 0) { 1925 if ((options & OBJPR_NOTMAPPED) == 0) 1926 pmap_remove_all(p); 1927 if ((options & OBJPR_CLEANONLY) == 0) { 1928 p->valid = 0; 1929 vm_page_undirty(p); 1930 } 1931 goto next; 1932 } 1933 if (vm_page_busied(p)) { 1934 VM_OBJECT_WUNLOCK(object); 1935 vm_page_busy_sleep(p, "vmopar", false); 1936 VM_OBJECT_WLOCK(object); 1937 goto again; 1938 } 1939 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1940 ("vm_object_page_remove: page %p is fictitious", p)); 1941 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1942 if ((options & OBJPR_NOTMAPPED) == 0) 1943 pmap_remove_write(p); 1944 if (p->dirty) 1945 goto next; 1946 } 1947 if ((options & OBJPR_NOTMAPPED) == 0) 1948 pmap_remove_all(p); 1949 vm_page_free(p); 1950next: 1951 vm_page_unlock(p); 1952 } 1953 vm_object_pip_wakeup(object); 1954skipmemq: 1955 if (__predict_false(!vm_object_cache_is_empty(object))) 1956 vm_page_cache_free(object, start, end); 1957} 1958 1959/* 1960 * vm_object_page_cache: 1961 * 1962 * For the given object, attempt to move the specified clean 1963 * pages to the cache queue. If a page is wired for any reason, 1964 * then it will not be changed. Pages are specified by the given 1965 * range ["start", "end"). As a special case, if "end" is zero, 1966 * then the range extends from "start" to the end of the object. 1967 * Any mappings to the specified pages are removed before the 1968 * pages are moved to the cache queue. 1969 * 1970 * This operation should only be performed on objects that 1971 * contain non-fictitious, managed pages. 1972 * 1973 * The object must be locked. 1974 */ 1975void 1976vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1977{ 1978 struct mtx *mtx, *new_mtx; 1979 vm_page_t p, next; 1980 1981 VM_OBJECT_ASSERT_WLOCKED(object); 1982 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 1983 ("vm_object_page_cache: illegal object %p", object)); 1984 if (object->resident_page_count == 0) 1985 return; 1986 p = vm_page_find_least(object, start); 1987 1988 /* 1989 * Here, the variable "p" is either (1) the page with the least pindex 1990 * greater than or equal to the parameter "start" or (2) NULL. 1991 */ 1992 mtx = NULL; 1993 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1994 next = TAILQ_NEXT(p, listq); 1995 1996 /* 1997 * Avoid releasing and reacquiring the same page lock. 1998 */ 1999 new_mtx = vm_page_lockptr(p); 2000 if (mtx != new_mtx) { 2001 if (mtx != NULL) 2002 mtx_unlock(mtx); 2003 mtx = new_mtx; 2004 mtx_lock(mtx); 2005 } 2006 vm_page_try_to_cache(p); 2007 } 2008 if (mtx != NULL) 2009 mtx_unlock(mtx); 2010} 2011 2012/* 2013 * Populate the specified range of the object with valid pages. Returns 2014 * TRUE if the range is successfully populated and FALSE otherwise. 2015 * 2016 * Note: This function should be optimized to pass a larger array of 2017 * pages to vm_pager_get_pages() before it is applied to a non- 2018 * OBJT_DEVICE object. 2019 * 2020 * The object must be locked. 2021 */ 2022boolean_t 2023vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2024{ 2025 vm_page_t m, ma[1]; 2026 vm_pindex_t pindex; 2027 int rv; 2028 2029 VM_OBJECT_ASSERT_WLOCKED(object); 2030 for (pindex = start; pindex < end; pindex++) { 2031 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2032 if (m->valid != VM_PAGE_BITS_ALL) { 2033 ma[0] = m; 2034 rv = vm_pager_get_pages(object, ma, 1, 0); 2035 m = vm_page_lookup(object, pindex); 2036 if (m == NULL) 2037 break; 2038 if (rv != VM_PAGER_OK) { 2039 vm_page_lock(m); 2040 vm_page_free(m); 2041 vm_page_unlock(m); 2042 break; 2043 } 2044 } 2045 /* 2046 * Keep "m" busy because a subsequent iteration may unlock 2047 * the object. 2048 */ 2049 } 2050 if (pindex > start) { 2051 m = vm_page_lookup(object, start); 2052 while (m != NULL && m->pindex < pindex) { 2053 vm_page_xunbusy(m); 2054 m = TAILQ_NEXT(m, listq); 2055 } 2056 } 2057 return (pindex == end); 2058} 2059 2060/* 2061 * Routine: vm_object_coalesce 2062 * Function: Coalesces two objects backing up adjoining 2063 * regions of memory into a single object. 2064 * 2065 * returns TRUE if objects were combined. 2066 * 2067 * NOTE: Only works at the moment if the second object is NULL - 2068 * if it's not, which object do we lock first? 2069 * 2070 * Parameters: 2071 * prev_object First object to coalesce 2072 * prev_offset Offset into prev_object 2073 * prev_size Size of reference to prev_object 2074 * next_size Size of reference to the second object 2075 * reserved Indicator that extension region has 2076 * swap accounted for 2077 * 2078 * Conditions: 2079 * The object must *not* be locked. 2080 */ 2081boolean_t 2082vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2083 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2084{ 2085 vm_pindex_t next_pindex; 2086 2087 if (prev_object == NULL) 2088 return (TRUE); 2089 VM_OBJECT_WLOCK(prev_object); 2090 if ((prev_object->type != OBJT_DEFAULT && 2091 prev_object->type != OBJT_SWAP) || 2092 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2093 VM_OBJECT_WUNLOCK(prev_object); 2094 return (FALSE); 2095 } 2096 2097 /* 2098 * Try to collapse the object first 2099 */ 2100 vm_object_collapse(prev_object); 2101 2102 /* 2103 * Can't coalesce if: . more than one reference . paged out . shadows 2104 * another object . has a copy elsewhere (any of which mean that the 2105 * pages not mapped to prev_entry may be in use anyway) 2106 */ 2107 if (prev_object->backing_object != NULL) { 2108 VM_OBJECT_WUNLOCK(prev_object); 2109 return (FALSE); 2110 } 2111 2112 prev_size >>= PAGE_SHIFT; 2113 next_size >>= PAGE_SHIFT; 2114 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2115 2116 if ((prev_object->ref_count > 1) && 2117 (prev_object->size != next_pindex)) { 2118 VM_OBJECT_WUNLOCK(prev_object); 2119 return (FALSE); 2120 } 2121 2122 /* 2123 * Account for the charge. 2124 */ 2125 if (prev_object->cred != NULL) { 2126 2127 /* 2128 * If prev_object was charged, then this mapping, 2129 * althought not charged now, may become writable 2130 * later. Non-NULL cred in the object would prevent 2131 * swap reservation during enabling of the write 2132 * access, so reserve swap now. Failed reservation 2133 * cause allocation of the separate object for the map 2134 * entry, and swap reservation for this entry is 2135 * managed in appropriate time. 2136 */ 2137 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2138 prev_object->cred)) { 2139 VM_OBJECT_WUNLOCK(prev_object); 2140 return (FALSE); 2141 } 2142 prev_object->charge += ptoa(next_size); 2143 } 2144 2145 /* 2146 * Remove any pages that may still be in the object from a previous 2147 * deallocation. 2148 */ 2149 if (next_pindex < prev_object->size) { 2150 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2151 next_size, 0); 2152 if (prev_object->type == OBJT_SWAP) 2153 swap_pager_freespace(prev_object, 2154 next_pindex, next_size); 2155#if 0 2156 if (prev_object->cred != NULL) { 2157 KASSERT(prev_object->charge >= 2158 ptoa(prev_object->size - next_pindex), 2159 ("object %p overcharged 1 %jx %jx", prev_object, 2160 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2161 prev_object->charge -= ptoa(prev_object->size - 2162 next_pindex); 2163 } 2164#endif 2165 } 2166 2167 /* 2168 * Extend the object if necessary. 2169 */ 2170 if (next_pindex + next_size > prev_object->size) 2171 prev_object->size = next_pindex + next_size; 2172 2173 VM_OBJECT_WUNLOCK(prev_object); 2174 return (TRUE); 2175} 2176 2177void 2178vm_object_set_writeable_dirty(vm_object_t object) 2179{ 2180 2181 VM_OBJECT_ASSERT_WLOCKED(object); 2182 if (object->type != OBJT_VNODE) { 2183 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2184 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2185 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2186 } 2187 return; 2188 } 2189 object->generation++; 2190 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2191 return; 2192 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2193} 2194 2195/* 2196 * vm_object_unwire: 2197 * 2198 * For each page offset within the specified range of the given object, 2199 * find the highest-level page in the shadow chain and unwire it. A page 2200 * must exist at every page offset, and the highest-level page must be 2201 * wired. 2202 */ 2203void 2204vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2205 uint8_t queue) 2206{ 2207 vm_object_t tobject; 2208 vm_page_t m, tm; 2209 vm_pindex_t end_pindex, pindex, tpindex; 2210 int depth, locked_depth; 2211 2212 KASSERT((offset & PAGE_MASK) == 0, 2213 ("vm_object_unwire: offset is not page aligned")); 2214 KASSERT((length & PAGE_MASK) == 0, 2215 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2216 /* The wired count of a fictitious page never changes. */ 2217 if ((object->flags & OBJ_FICTITIOUS) != 0) 2218 return; 2219 pindex = OFF_TO_IDX(offset); 2220 end_pindex = pindex + atop(length); 2221 locked_depth = 1; 2222 VM_OBJECT_RLOCK(object); 2223 m = vm_page_find_least(object, pindex); 2224 while (pindex < end_pindex) { 2225 if (m == NULL || pindex < m->pindex) { 2226 /* 2227 * The first object in the shadow chain doesn't 2228 * contain a page at the current index. Therefore, 2229 * the page must exist in a backing object. 2230 */ 2231 tobject = object; 2232 tpindex = pindex; 2233 depth = 0; 2234 do { 2235 tpindex += 2236 OFF_TO_IDX(tobject->backing_object_offset); 2237 tobject = tobject->backing_object; 2238 KASSERT(tobject != NULL, 2239 ("vm_object_unwire: missing page")); 2240 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2241 goto next_page; 2242 depth++; 2243 if (depth == locked_depth) { 2244 locked_depth++; 2245 VM_OBJECT_RLOCK(tobject); 2246 } 2247 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2248 NULL); 2249 } else { 2250 tm = m; 2251 m = TAILQ_NEXT(m, listq); 2252 } 2253 vm_page_lock(tm); 2254 vm_page_unwire(tm, queue); 2255 vm_page_unlock(tm); 2256next_page: 2257 pindex++; 2258 } 2259 /* Release the accumulated object locks. */ 2260 for (depth = 0; depth < locked_depth; depth++) { 2261 tobject = object->backing_object; 2262 VM_OBJECT_RUNLOCK(object); 2263 object = tobject; 2264 } 2265} 2266 2267struct vnode * 2268vm_object_vnode(vm_object_t object) 2269{ 2270 2271 VM_OBJECT_ASSERT_LOCKED(object); 2272 if (object->type == OBJT_VNODE) 2273 return (object->handle); 2274 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) 2275 return (object->un_pager.swp.swp_tmpfs); 2276 return (NULL); 2277} 2278 2279static int 2280sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2281{ 2282 struct kinfo_vmobject *kvo; 2283 char *fullpath, *freepath; 2284 struct vnode *vp; 2285 struct vattr va; 2286 vm_object_t obj; 2287 vm_page_t m; 2288 int count, error; 2289 2290 if (req->oldptr == NULL) { 2291 /* 2292 * If an old buffer has not been provided, generate an 2293 * estimate of the space needed for a subsequent call. 2294 */ 2295 mtx_lock(&vm_object_list_mtx); 2296 count = 0; 2297 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2298 if (obj->type == OBJT_DEAD) 2299 continue; 2300 count++; 2301 } 2302 mtx_unlock(&vm_object_list_mtx); 2303 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2304 count * 11 / 10)); 2305 } 2306 2307 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2308 error = 0; 2309 2310 /* 2311 * VM objects are type stable and are never removed from the 2312 * list once added. This allows us to safely read obj->object_list 2313 * after reacquiring the VM object lock. 2314 */ 2315 mtx_lock(&vm_object_list_mtx); 2316 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2317 if (obj->type == OBJT_DEAD) 2318 continue; 2319 VM_OBJECT_RLOCK(obj); 2320 if (obj->type == OBJT_DEAD) { 2321 VM_OBJECT_RUNLOCK(obj); 2322 continue; 2323 } 2324 mtx_unlock(&vm_object_list_mtx); 2325 kvo->kvo_size = ptoa(obj->size); 2326 kvo->kvo_resident = obj->resident_page_count; 2327 kvo->kvo_ref_count = obj->ref_count; 2328 kvo->kvo_shadow_count = obj->shadow_count; 2329 kvo->kvo_memattr = obj->memattr; 2330 kvo->kvo_active = 0; 2331 kvo->kvo_inactive = 0; 2332 TAILQ_FOREACH(m, &obj->memq, listq) { 2333 /* 2334 * A page may belong to the object but be 2335 * dequeued and set to PQ_NONE while the 2336 * object lock is not held. This makes the 2337 * reads of m->queue below racy, and we do not 2338 * count pages set to PQ_NONE. However, this 2339 * sysctl is only meant to give an 2340 * approximation of the system anyway. 2341 */ 2342 if (m->queue == PQ_ACTIVE) 2343 kvo->kvo_active++; 2344 else if (m->queue == PQ_INACTIVE) 2345 kvo->kvo_inactive++; 2346 } 2347 2348 kvo->kvo_vn_fileid = 0; 2349 kvo->kvo_vn_fsid = 0; 2350 freepath = NULL; 2351 fullpath = ""; 2352 vp = NULL; 2353 switch (obj->type) { 2354 case OBJT_DEFAULT: 2355 kvo->kvo_type = KVME_TYPE_DEFAULT; 2356 break; 2357 case OBJT_VNODE: 2358 kvo->kvo_type = KVME_TYPE_VNODE; 2359 vp = obj->handle; 2360 vref(vp); 2361 break; 2362 case OBJT_SWAP: 2363 kvo->kvo_type = KVME_TYPE_SWAP; 2364 break; 2365 case OBJT_DEVICE: 2366 kvo->kvo_type = KVME_TYPE_DEVICE; 2367 break; 2368 case OBJT_PHYS: 2369 kvo->kvo_type = KVME_TYPE_PHYS; 2370 break; 2371 case OBJT_DEAD: 2372 kvo->kvo_type = KVME_TYPE_DEAD; 2373 break; 2374 case OBJT_SG: 2375 kvo->kvo_type = KVME_TYPE_SG; 2376 break; 2377 case OBJT_MGTDEVICE: 2378 kvo->kvo_type = KVME_TYPE_MGTDEVICE; 2379 break; 2380 default: 2381 kvo->kvo_type = KVME_TYPE_UNKNOWN; 2382 break; 2383 } 2384 VM_OBJECT_RUNLOCK(obj); 2385 if (vp != NULL) { 2386 vn_fullpath(curthread, vp, &fullpath, &freepath); 2387 vn_lock(vp, LK_SHARED | LK_RETRY); 2388 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2389 kvo->kvo_vn_fileid = va.va_fileid; 2390 kvo->kvo_vn_fsid = va.va_fsid; 2391 } 2392 vput(vp); 2393 } 2394 2395 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2396 if (freepath != NULL) 2397 free(freepath, M_TEMP); 2398 2399 /* Pack record size down */ 2400 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2401 + strlen(kvo->kvo_path) + 1; 2402 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2403 sizeof(uint64_t)); 2404 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2405 mtx_lock(&vm_object_list_mtx); 2406 if (error) 2407 break; 2408 } 2409 mtx_unlock(&vm_object_list_mtx); 2410 free(kvo, M_TEMP); 2411 return (error); 2412} 2413SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2414 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2415 "List of VM objects"); 2416 2417#include "opt_ddb.h" 2418#ifdef DDB 2419#include <sys/kernel.h> 2420 2421#include <sys/cons.h> 2422 2423#include <ddb/ddb.h> 2424 2425static int 2426_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2427{ 2428 vm_map_t tmpm; 2429 vm_map_entry_t tmpe; 2430 vm_object_t obj; 2431 int entcount; 2432 2433 if (map == 0) 2434 return 0; 2435 2436 if (entry == 0) { 2437 tmpe = map->header.next; 2438 entcount = map->nentries; 2439 while (entcount-- && (tmpe != &map->header)) { 2440 if (_vm_object_in_map(map, object, tmpe)) { 2441 return 1; 2442 } 2443 tmpe = tmpe->next; 2444 } 2445 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2446 tmpm = entry->object.sub_map; 2447 tmpe = tmpm->header.next; 2448 entcount = tmpm->nentries; 2449 while (entcount-- && tmpe != &tmpm->header) { 2450 if (_vm_object_in_map(tmpm, object, tmpe)) { 2451 return 1; 2452 } 2453 tmpe = tmpe->next; 2454 } 2455 } else if ((obj = entry->object.vm_object) != NULL) { 2456 for (; obj; obj = obj->backing_object) 2457 if (obj == object) { 2458 return 1; 2459 } 2460 } 2461 return 0; 2462} 2463 2464static int 2465vm_object_in_map(vm_object_t object) 2466{ 2467 struct proc *p; 2468 2469 /* sx_slock(&allproc_lock); */ 2470 FOREACH_PROC_IN_SYSTEM(p) { 2471 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2472 continue; 2473 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2474 /* sx_sunlock(&allproc_lock); */ 2475 return 1; 2476 } 2477 } 2478 /* sx_sunlock(&allproc_lock); */ 2479 if (_vm_object_in_map(kernel_map, object, 0)) 2480 return 1; 2481 return 0; 2482} 2483 2484DB_SHOW_COMMAND(vmochk, vm_object_check) 2485{ 2486 vm_object_t object; 2487 2488 /* 2489 * make sure that internal objs are in a map somewhere 2490 * and none have zero ref counts. 2491 */ 2492 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2493 if (object->handle == NULL && 2494 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2495 if (object->ref_count == 0) { 2496 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2497 (long)object->size); 2498 } 2499 if (!vm_object_in_map(object)) { 2500 db_printf( 2501 "vmochk: internal obj is not in a map: " 2502 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2503 object->ref_count, (u_long)object->size, 2504 (u_long)object->size, 2505 (void *)object->backing_object); 2506 } 2507 } 2508 } 2509} 2510 2511/* 2512 * vm_object_print: [ debug ] 2513 */ 2514DB_SHOW_COMMAND(object, vm_object_print_static) 2515{ 2516 /* XXX convert args. */ 2517 vm_object_t object = (vm_object_t)addr; 2518 boolean_t full = have_addr; 2519 2520 vm_page_t p; 2521 2522 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2523#define count was_count 2524 2525 int count; 2526 2527 if (object == NULL) 2528 return; 2529 2530 db_iprintf( 2531 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2532 object, (int)object->type, (uintmax_t)object->size, 2533 object->resident_page_count, object->ref_count, object->flags, 2534 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2535 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2536 object->shadow_count, 2537 object->backing_object ? object->backing_object->ref_count : 0, 2538 object->backing_object, (uintmax_t)object->backing_object_offset); 2539 2540 if (!full) 2541 return; 2542 2543 db_indent += 2; 2544 count = 0; 2545 TAILQ_FOREACH(p, &object->memq, listq) { 2546 if (count == 0) 2547 db_iprintf("memory:="); 2548 else if (count == 6) { 2549 db_printf("\n"); 2550 db_iprintf(" ..."); 2551 count = 0; 2552 } else 2553 db_printf(","); 2554 count++; 2555 2556 db_printf("(off=0x%jx,page=0x%jx)", 2557 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2558 } 2559 if (count != 0) 2560 db_printf("\n"); 2561 db_indent -= 2; 2562} 2563 2564/* XXX. */ 2565#undef count 2566 2567/* XXX need this non-static entry for calling from vm_map_print. */ 2568void 2569vm_object_print( 2570 /* db_expr_t */ long addr, 2571 boolean_t have_addr, 2572 /* db_expr_t */ long count, 2573 char *modif) 2574{ 2575 vm_object_print_static(addr, have_addr, count, modif); 2576} 2577 2578DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2579{ 2580 vm_object_t object; 2581 vm_pindex_t fidx; 2582 vm_paddr_t pa; 2583 vm_page_t m, prev_m; 2584 int rcount, nl, c; 2585 2586 nl = 0; 2587 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2588 db_printf("new object: %p\n", (void *)object); 2589 if (nl > 18) { 2590 c = cngetc(); 2591 if (c != ' ') 2592 return; 2593 nl = 0; 2594 } 2595 nl++; 2596 rcount = 0; 2597 fidx = 0; 2598 pa = -1; 2599 TAILQ_FOREACH(m, &object->memq, listq) { 2600 if (m->pindex > 128) 2601 break; 2602 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2603 prev_m->pindex + 1 != m->pindex) { 2604 if (rcount) { 2605 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2606 (long)fidx, rcount, (long)pa); 2607 if (nl > 18) { 2608 c = cngetc(); 2609 if (c != ' ') 2610 return; 2611 nl = 0; 2612 } 2613 nl++; 2614 rcount = 0; 2615 } 2616 } 2617 if (rcount && 2618 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2619 ++rcount; 2620 continue; 2621 } 2622 if (rcount) { 2623 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2624 (long)fidx, rcount, (long)pa); 2625 if (nl > 18) { 2626 c = cngetc(); 2627 if (c != ' ') 2628 return; 2629 nl = 0; 2630 } 2631 nl++; 2632 } 2633 fidx = m->pindex; 2634 pa = VM_PAGE_TO_PHYS(m); 2635 rcount = 1; 2636 } 2637 if (rcount) { 2638 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2639 (long)fidx, rcount, (long)pa); 2640 if (nl > 18) { 2641 c = cngetc(); 2642 if (c != ' ') 2643 return; 2644 nl = 0; 2645 } 2646 nl++; 2647 } 2648 } 2649} 2650#endif /* DDB */ 2651