vm_object.c revision 263359
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 263359 2014-03-19 13:04:16Z kib $"); 67 68#include "opt_vm.h" 69 70#include <sys/param.h> 71#include <sys/systm.h> 72#include <sys/lock.h> 73#include <sys/mman.h> 74#include <sys/mount.h> 75#include <sys/kernel.h> 76#include <sys/sysctl.h> 77#include <sys/mutex.h> 78#include <sys/proc.h> /* for curproc, pageproc */ 79#include <sys/socket.h> 80#include <sys/resourcevar.h> 81#include <sys/rwlock.h> 82#include <sys/vnode.h> 83#include <sys/vmmeter.h> 84#include <sys/sx.h> 85 86#include <vm/vm.h> 87#include <vm/vm_param.h> 88#include <vm/pmap.h> 89#include <vm/vm_map.h> 90#include <vm/vm_object.h> 91#include <vm/vm_page.h> 92#include <vm/vm_pageout.h> 93#include <vm/vm_pager.h> 94#include <vm/swap_pager.h> 95#include <vm/vm_kern.h> 96#include <vm/vm_extern.h> 97#include <vm/vm_radix.h> 98#include <vm/vm_reserv.h> 99#include <vm/uma.h> 100 101static int old_msync; 102SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 103 "Use old (insecure) msync behavior"); 104 105static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 106 int pagerflags, int flags, boolean_t *clearobjflags, 107 boolean_t *eio); 108static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 109 boolean_t *clearobjflags); 110static void vm_object_qcollapse(vm_object_t object); 111static void vm_object_vndeallocate(vm_object_t object); 112 113/* 114 * Virtual memory objects maintain the actual data 115 * associated with allocated virtual memory. A given 116 * page of memory exists within exactly one object. 117 * 118 * An object is only deallocated when all "references" 119 * are given up. Only one "reference" to a given 120 * region of an object should be writeable. 121 * 122 * Associated with each object is a list of all resident 123 * memory pages belonging to that object; this list is 124 * maintained by the "vm_page" module, and locked by the object's 125 * lock. 126 * 127 * Each object also records a "pager" routine which is 128 * used to retrieve (and store) pages to the proper backing 129 * storage. In addition, objects may be backed by other 130 * objects from which they were virtual-copied. 131 * 132 * The only items within the object structure which are 133 * modified after time of creation are: 134 * reference count locked by object's lock 135 * pager routine locked by object's lock 136 * 137 */ 138 139struct object_q vm_object_list; 140struct mtx vm_object_list_mtx; /* lock for object list and count */ 141 142struct vm_object kernel_object_store; 143struct vm_object kmem_object_store; 144 145static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 146 "VM object stats"); 147 148static long object_collapses; 149SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 150 &object_collapses, 0, "VM object collapses"); 151 152static long object_bypasses; 153SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 154 &object_bypasses, 0, "VM object bypasses"); 155 156static uma_zone_t obj_zone; 157 158static int vm_object_zinit(void *mem, int size, int flags); 159 160#ifdef INVARIANTS 161static void vm_object_zdtor(void *mem, int size, void *arg); 162 163static void 164vm_object_zdtor(void *mem, int size, void *arg) 165{ 166 vm_object_t object; 167 168 object = (vm_object_t)mem; 169 KASSERT(TAILQ_EMPTY(&object->memq), 170 ("object %p has resident pages in its memq", object)); 171 KASSERT(vm_radix_is_empty(&object->rtree), 172 ("object %p has resident pages in its trie", object)); 173#if VM_NRESERVLEVEL > 0 174 KASSERT(LIST_EMPTY(&object->rvq), 175 ("object %p has reservations", 176 object)); 177#endif 178 KASSERT(vm_object_cache_is_empty(object), 179 ("object %p has cached pages", 180 object)); 181 KASSERT(object->paging_in_progress == 0, 182 ("object %p paging_in_progress = %d", 183 object, object->paging_in_progress)); 184 KASSERT(object->resident_page_count == 0, 185 ("object %p resident_page_count = %d", 186 object, object->resident_page_count)); 187 KASSERT(object->shadow_count == 0, 188 ("object %p shadow_count = %d", 189 object, object->shadow_count)); 190} 191#endif 192 193static int 194vm_object_zinit(void *mem, int size, int flags) 195{ 196 vm_object_t object; 197 198 object = (vm_object_t)mem; 199 bzero(&object->lock, sizeof(object->lock)); 200 rw_init_flags(&object->lock, "vm object", RW_DUPOK); 201 202 /* These are true for any object that has been freed */ 203 object->rtree.rt_root = 0; 204 object->rtree.rt_flags = 0; 205 object->paging_in_progress = 0; 206 object->resident_page_count = 0; 207 object->shadow_count = 0; 208 object->cache.rt_root = 0; 209 object->cache.rt_flags = 0; 210 return (0); 211} 212 213static void 214_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 215{ 216 217 TAILQ_INIT(&object->memq); 218 LIST_INIT(&object->shadow_head); 219 220 object->type = type; 221 switch (type) { 222 case OBJT_DEAD: 223 panic("_vm_object_allocate: can't create OBJT_DEAD"); 224 case OBJT_DEFAULT: 225 case OBJT_SWAP: 226 object->flags = OBJ_ONEMAPPING; 227 break; 228 case OBJT_DEVICE: 229 case OBJT_SG: 230 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 231 break; 232 case OBJT_MGTDEVICE: 233 object->flags = OBJ_FICTITIOUS; 234 break; 235 case OBJT_PHYS: 236 object->flags = OBJ_UNMANAGED; 237 break; 238 case OBJT_VNODE: 239 object->flags = 0; 240 break; 241 default: 242 panic("_vm_object_allocate: type %d is undefined", type); 243 } 244 object->size = size; 245 object->generation = 1; 246 object->ref_count = 1; 247 object->memattr = VM_MEMATTR_DEFAULT; 248 object->cred = NULL; 249 object->charge = 0; 250 object->handle = NULL; 251 object->backing_object = NULL; 252 object->backing_object_offset = (vm_ooffset_t) 0; 253#if VM_NRESERVLEVEL > 0 254 LIST_INIT(&object->rvq); 255#endif 256 257 mtx_lock(&vm_object_list_mtx); 258 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 259 mtx_unlock(&vm_object_list_mtx); 260} 261 262/* 263 * vm_object_init: 264 * 265 * Initialize the VM objects module. 266 */ 267void 268vm_object_init(void) 269{ 270 TAILQ_INIT(&vm_object_list); 271 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 272 273 rw_init(&kernel_object->lock, "kernel vm object"); 274 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 275 kernel_object); 276#if VM_NRESERVLEVEL > 0 277 kernel_object->flags |= OBJ_COLORED; 278 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 279#endif 280 281 rw_init(&kmem_object->lock, "kmem vm object"); 282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 283 kmem_object); 284#if VM_NRESERVLEVEL > 0 285 kmem_object->flags |= OBJ_COLORED; 286 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 287#endif 288 289 /* 290 * The lock portion of struct vm_object must be type stable due 291 * to vm_pageout_fallback_object_lock locking a vm object 292 * without holding any references to it. 293 */ 294 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 295#ifdef INVARIANTS 296 vm_object_zdtor, 297#else 298 NULL, 299#endif 300 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 301 302 vm_radix_init(); 303} 304 305void 306vm_object_clear_flag(vm_object_t object, u_short bits) 307{ 308 309 VM_OBJECT_ASSERT_WLOCKED(object); 310 object->flags &= ~bits; 311} 312 313/* 314 * Sets the default memory attribute for the specified object. Pages 315 * that are allocated to this object are by default assigned this memory 316 * attribute. 317 * 318 * Presently, this function must be called before any pages are allocated 319 * to the object. In the future, this requirement may be relaxed for 320 * "default" and "swap" objects. 321 */ 322int 323vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 324{ 325 326 VM_OBJECT_ASSERT_WLOCKED(object); 327 switch (object->type) { 328 case OBJT_DEFAULT: 329 case OBJT_DEVICE: 330 case OBJT_MGTDEVICE: 331 case OBJT_PHYS: 332 case OBJT_SG: 333 case OBJT_SWAP: 334 case OBJT_VNODE: 335 if (!TAILQ_EMPTY(&object->memq)) 336 return (KERN_FAILURE); 337 break; 338 case OBJT_DEAD: 339 return (KERN_INVALID_ARGUMENT); 340 default: 341 panic("vm_object_set_memattr: object %p is of undefined type", 342 object); 343 } 344 object->memattr = memattr; 345 return (KERN_SUCCESS); 346} 347 348void 349vm_object_pip_add(vm_object_t object, short i) 350{ 351 352 VM_OBJECT_ASSERT_WLOCKED(object); 353 object->paging_in_progress += i; 354} 355 356void 357vm_object_pip_subtract(vm_object_t object, short i) 358{ 359 360 VM_OBJECT_ASSERT_WLOCKED(object); 361 object->paging_in_progress -= i; 362} 363 364void 365vm_object_pip_wakeup(vm_object_t object) 366{ 367 368 VM_OBJECT_ASSERT_WLOCKED(object); 369 object->paging_in_progress--; 370 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 371 vm_object_clear_flag(object, OBJ_PIPWNT); 372 wakeup(object); 373 } 374} 375 376void 377vm_object_pip_wakeupn(vm_object_t object, short i) 378{ 379 380 VM_OBJECT_ASSERT_WLOCKED(object); 381 if (i) 382 object->paging_in_progress -= i; 383 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 384 vm_object_clear_flag(object, OBJ_PIPWNT); 385 wakeup(object); 386 } 387} 388 389void 390vm_object_pip_wait(vm_object_t object, char *waitid) 391{ 392 393 VM_OBJECT_ASSERT_WLOCKED(object); 394 while (object->paging_in_progress) { 395 object->flags |= OBJ_PIPWNT; 396 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 397 } 398} 399 400/* 401 * vm_object_allocate: 402 * 403 * Returns a new object with the given size. 404 */ 405vm_object_t 406vm_object_allocate(objtype_t type, vm_pindex_t size) 407{ 408 vm_object_t object; 409 410 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 411 _vm_object_allocate(type, size, object); 412 return (object); 413} 414 415 416/* 417 * vm_object_reference: 418 * 419 * Gets another reference to the given object. Note: OBJ_DEAD 420 * objects can be referenced during final cleaning. 421 */ 422void 423vm_object_reference(vm_object_t object) 424{ 425 if (object == NULL) 426 return; 427 VM_OBJECT_WLOCK(object); 428 vm_object_reference_locked(object); 429 VM_OBJECT_WUNLOCK(object); 430} 431 432/* 433 * vm_object_reference_locked: 434 * 435 * Gets another reference to the given object. 436 * 437 * The object must be locked. 438 */ 439void 440vm_object_reference_locked(vm_object_t object) 441{ 442 struct vnode *vp; 443 444 VM_OBJECT_ASSERT_WLOCKED(object); 445 object->ref_count++; 446 if (object->type == OBJT_VNODE) { 447 vp = object->handle; 448 vref(vp); 449 } 450} 451 452/* 453 * Handle deallocating an object of type OBJT_VNODE. 454 */ 455static void 456vm_object_vndeallocate(vm_object_t object) 457{ 458 struct vnode *vp = (struct vnode *) object->handle; 459 460 VM_OBJECT_ASSERT_WLOCKED(object); 461 KASSERT(object->type == OBJT_VNODE, 462 ("vm_object_vndeallocate: not a vnode object")); 463 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 464#ifdef INVARIANTS 465 if (object->ref_count == 0) { 466 vprint("vm_object_vndeallocate", vp); 467 panic("vm_object_vndeallocate: bad object reference count"); 468 } 469#endif 470 471 if (object->ref_count > 1) { 472 object->ref_count--; 473 VM_OBJECT_WUNLOCK(object); 474 /* vrele may need the vnode lock. */ 475 vrele(vp); 476 } else { 477 vhold(vp); 478 VM_OBJECT_WUNLOCK(object); 479 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 480 vdrop(vp); 481 VM_OBJECT_WLOCK(object); 482 object->ref_count--; 483 if (object->type == OBJT_DEAD) { 484 VM_OBJECT_WUNLOCK(object); 485 VOP_UNLOCK(vp, 0); 486 } else { 487 if (object->ref_count == 0) 488 VOP_UNSET_TEXT(vp); 489 VM_OBJECT_WUNLOCK(object); 490 vput(vp); 491 } 492 } 493} 494 495/* 496 * vm_object_deallocate: 497 * 498 * Release a reference to the specified object, 499 * gained either through a vm_object_allocate 500 * or a vm_object_reference call. When all references 501 * are gone, storage associated with this object 502 * may be relinquished. 503 * 504 * No object may be locked. 505 */ 506void 507vm_object_deallocate(vm_object_t object) 508{ 509 vm_object_t temp; 510 struct vnode *vp; 511 512 while (object != NULL) { 513 VM_OBJECT_WLOCK(object); 514 if (object->type == OBJT_VNODE) { 515 vm_object_vndeallocate(object); 516 return; 517 } 518 519 KASSERT(object->ref_count != 0, 520 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 521 522 /* 523 * If the reference count goes to 0 we start calling 524 * vm_object_terminate() on the object chain. 525 * A ref count of 1 may be a special case depending on the 526 * shadow count being 0 or 1. 527 */ 528 object->ref_count--; 529 if (object->ref_count > 1) { 530 VM_OBJECT_WUNLOCK(object); 531 return; 532 } else if (object->ref_count == 1) { 533 if (object->type == OBJT_SWAP && 534 (object->flags & OBJ_TMPFS) != 0) { 535 vp = object->un_pager.swp.swp_tmpfs; 536 vhold(vp); 537 VM_OBJECT_WUNLOCK(object); 538 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 539 VM_OBJECT_WLOCK(object); 540 if (object->type == OBJT_DEAD || 541 object->ref_count != 1) { 542 VM_OBJECT_WUNLOCK(object); 543 VOP_UNLOCK(vp, 0); 544 vdrop(vp); 545 return; 546 } 547 if ((object->flags & OBJ_TMPFS) != 0) 548 VOP_UNSET_TEXT(vp); 549 VOP_UNLOCK(vp, 0); 550 vdrop(vp); 551 } 552 if (object->shadow_count == 0 && 553 object->handle == NULL && 554 (object->type == OBJT_DEFAULT || 555 (object->type == OBJT_SWAP && 556 (object->flags & OBJ_TMPFS) == 0))) { 557 vm_object_set_flag(object, OBJ_ONEMAPPING); 558 } else if ((object->shadow_count == 1) && 559 (object->handle == NULL) && 560 (object->type == OBJT_DEFAULT || 561 object->type == OBJT_SWAP)) { 562 KASSERT((object->flags & OBJ_TMPFS) == 0, 563 ("shadowed tmpfs v_object %p", object)); 564 vm_object_t robject; 565 566 robject = LIST_FIRST(&object->shadow_head); 567 KASSERT(robject != NULL, 568 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 569 object->ref_count, 570 object->shadow_count)); 571 if (!VM_OBJECT_TRYWLOCK(robject)) { 572 /* 573 * Avoid a potential deadlock. 574 */ 575 object->ref_count++; 576 VM_OBJECT_WUNLOCK(object); 577 /* 578 * More likely than not the thread 579 * holding robject's lock has lower 580 * priority than the current thread. 581 * Let the lower priority thread run. 582 */ 583 pause("vmo_de", 1); 584 continue; 585 } 586 /* 587 * Collapse object into its shadow unless its 588 * shadow is dead. In that case, object will 589 * be deallocated by the thread that is 590 * deallocating its shadow. 591 */ 592 if ((robject->flags & OBJ_DEAD) == 0 && 593 (robject->handle == NULL) && 594 (robject->type == OBJT_DEFAULT || 595 robject->type == OBJT_SWAP)) { 596 597 robject->ref_count++; 598retry: 599 if (robject->paging_in_progress) { 600 VM_OBJECT_WUNLOCK(object); 601 vm_object_pip_wait(robject, 602 "objde1"); 603 temp = robject->backing_object; 604 if (object == temp) { 605 VM_OBJECT_WLOCK(object); 606 goto retry; 607 } 608 } else if (object->paging_in_progress) { 609 VM_OBJECT_WUNLOCK(robject); 610 object->flags |= OBJ_PIPWNT; 611 VM_OBJECT_SLEEP(object, object, 612 PDROP | PVM, "objde2", 0); 613 VM_OBJECT_WLOCK(robject); 614 temp = robject->backing_object; 615 if (object == temp) { 616 VM_OBJECT_WLOCK(object); 617 goto retry; 618 } 619 } else 620 VM_OBJECT_WUNLOCK(object); 621 622 if (robject->ref_count == 1) { 623 robject->ref_count--; 624 object = robject; 625 goto doterm; 626 } 627 object = robject; 628 vm_object_collapse(object); 629 VM_OBJECT_WUNLOCK(object); 630 continue; 631 } 632 VM_OBJECT_WUNLOCK(robject); 633 } 634 VM_OBJECT_WUNLOCK(object); 635 return; 636 } 637doterm: 638 temp = object->backing_object; 639 if (temp != NULL) { 640 VM_OBJECT_WLOCK(temp); 641 LIST_REMOVE(object, shadow_list); 642 temp->shadow_count--; 643 VM_OBJECT_WUNLOCK(temp); 644 object->backing_object = NULL; 645 } 646 /* 647 * Don't double-terminate, we could be in a termination 648 * recursion due to the terminate having to sync data 649 * to disk. 650 */ 651 if ((object->flags & OBJ_DEAD) == 0) 652 vm_object_terminate(object); 653 else 654 VM_OBJECT_WUNLOCK(object); 655 object = temp; 656 } 657} 658 659/* 660 * vm_object_destroy removes the object from the global object list 661 * and frees the space for the object. 662 */ 663void 664vm_object_destroy(vm_object_t object) 665{ 666 667 /* 668 * Remove the object from the global object list. 669 */ 670 mtx_lock(&vm_object_list_mtx); 671 TAILQ_REMOVE(&vm_object_list, object, object_list); 672 mtx_unlock(&vm_object_list_mtx); 673 674 /* 675 * Release the allocation charge. 676 */ 677 if (object->cred != NULL) { 678 KASSERT(object->type == OBJT_DEFAULT || 679 object->type == OBJT_SWAP, 680 ("vm_object_terminate: non-swap obj %p has cred", 681 object)); 682 swap_release_by_cred(object->charge, object->cred); 683 object->charge = 0; 684 crfree(object->cred); 685 object->cred = NULL; 686 } 687 688 /* 689 * Free the space for the object. 690 */ 691 uma_zfree(obj_zone, object); 692} 693 694/* 695 * vm_object_terminate actually destroys the specified object, freeing 696 * up all previously used resources. 697 * 698 * The object must be locked. 699 * This routine may block. 700 */ 701void 702vm_object_terminate(vm_object_t object) 703{ 704 vm_page_t p, p_next; 705 706 VM_OBJECT_ASSERT_WLOCKED(object); 707 708 /* 709 * Make sure no one uses us. 710 */ 711 vm_object_set_flag(object, OBJ_DEAD); 712 713 /* 714 * wait for the pageout daemon to be done with the object 715 */ 716 vm_object_pip_wait(object, "objtrm"); 717 718 KASSERT(!object->paging_in_progress, 719 ("vm_object_terminate: pageout in progress")); 720 721 /* 722 * Clean and free the pages, as appropriate. All references to the 723 * object are gone, so we don't need to lock it. 724 */ 725 if (object->type == OBJT_VNODE) { 726 struct vnode *vp = (struct vnode *)object->handle; 727 728 /* 729 * Clean pages and flush buffers. 730 */ 731 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 732 VM_OBJECT_WUNLOCK(object); 733 734 vinvalbuf(vp, V_SAVE, 0, 0); 735 736 VM_OBJECT_WLOCK(object); 737 } 738 739 KASSERT(object->ref_count == 0, 740 ("vm_object_terminate: object with references, ref_count=%d", 741 object->ref_count)); 742 743 /* 744 * Free any remaining pageable pages. This also removes them from the 745 * paging queues. However, don't free wired pages, just remove them 746 * from the object. Rather than incrementally removing each page from 747 * the object, the page and object are reset to any empty state. 748 */ 749 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 750 vm_page_assert_unbusied(p); 751 vm_page_lock(p); 752 /* 753 * Optimize the page's removal from the object by resetting 754 * its "object" field. Specifically, if the page is not 755 * wired, then the effect of this assignment is that 756 * vm_page_free()'s call to vm_page_remove() will return 757 * immediately without modifying the page or the object. 758 */ 759 p->object = NULL; 760 if (p->wire_count == 0) { 761 vm_page_free(p); 762 PCPU_INC(cnt.v_pfree); 763 } 764 vm_page_unlock(p); 765 } 766 /* 767 * If the object contained any pages, then reset it to an empty state. 768 * None of the object's fields, including "resident_page_count", were 769 * modified by the preceding loop. 770 */ 771 if (object->resident_page_count != 0) { 772 vm_radix_reclaim_allnodes(&object->rtree); 773 TAILQ_INIT(&object->memq); 774 object->resident_page_count = 0; 775 if (object->type == OBJT_VNODE) 776 vdrop(object->handle); 777 } 778 779#if VM_NRESERVLEVEL > 0 780 if (__predict_false(!LIST_EMPTY(&object->rvq))) 781 vm_reserv_break_all(object); 782#endif 783 if (__predict_false(!vm_object_cache_is_empty(object))) 784 vm_page_cache_free(object, 0, 0); 785 786 /* 787 * Let the pager know object is dead. 788 */ 789 vm_pager_deallocate(object); 790 VM_OBJECT_WUNLOCK(object); 791 792 vm_object_destroy(object); 793} 794 795/* 796 * Make the page read-only so that we can clear the object flags. However, if 797 * this is a nosync mmap then the object is likely to stay dirty so do not 798 * mess with the page and do not clear the object flags. Returns TRUE if the 799 * page should be flushed, and FALSE otherwise. 800 */ 801static boolean_t 802vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 803{ 804 805 /* 806 * If we have been asked to skip nosync pages and this is a 807 * nosync page, skip it. Note that the object flags were not 808 * cleared in this case so we do not have to set them. 809 */ 810 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 811 *clearobjflags = FALSE; 812 return (FALSE); 813 } else { 814 pmap_remove_write(p); 815 return (p->dirty != 0); 816 } 817} 818 819/* 820 * vm_object_page_clean 821 * 822 * Clean all dirty pages in the specified range of object. Leaves page 823 * on whatever queue it is currently on. If NOSYNC is set then do not 824 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 825 * leaving the object dirty. 826 * 827 * When stuffing pages asynchronously, allow clustering. XXX we need a 828 * synchronous clustering mode implementation. 829 * 830 * Odd semantics: if start == end, we clean everything. 831 * 832 * The object must be locked. 833 * 834 * Returns FALSE if some page from the range was not written, as 835 * reported by the pager, and TRUE otherwise. 836 */ 837boolean_t 838vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 839 int flags) 840{ 841 vm_page_t np, p; 842 vm_pindex_t pi, tend, tstart; 843 int curgeneration, n, pagerflags; 844 boolean_t clearobjflags, eio, res; 845 846 VM_OBJECT_ASSERT_WLOCKED(object); 847 848 /* 849 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 850 * objects. The check below prevents the function from 851 * operating on non-vnode objects. 852 */ 853 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 854 object->resident_page_count == 0) 855 return (TRUE); 856 857 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 858 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 859 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 860 861 tstart = OFF_TO_IDX(start); 862 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 863 clearobjflags = tstart == 0 && tend >= object->size; 864 res = TRUE; 865 866rescan: 867 curgeneration = object->generation; 868 869 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 870 pi = p->pindex; 871 if (pi >= tend) 872 break; 873 np = TAILQ_NEXT(p, listq); 874 if (p->valid == 0) 875 continue; 876 if (vm_page_sleep_if_busy(p, "vpcwai")) { 877 if (object->generation != curgeneration) { 878 if ((flags & OBJPC_SYNC) != 0) 879 goto rescan; 880 else 881 clearobjflags = FALSE; 882 } 883 np = vm_page_find_least(object, pi); 884 continue; 885 } 886 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 887 continue; 888 889 n = vm_object_page_collect_flush(object, p, pagerflags, 890 flags, &clearobjflags, &eio); 891 if (eio) { 892 res = FALSE; 893 clearobjflags = FALSE; 894 } 895 if (object->generation != curgeneration) { 896 if ((flags & OBJPC_SYNC) != 0) 897 goto rescan; 898 else 899 clearobjflags = FALSE; 900 } 901 902 /* 903 * If the VOP_PUTPAGES() did a truncated write, so 904 * that even the first page of the run is not fully 905 * written, vm_pageout_flush() returns 0 as the run 906 * length. Since the condition that caused truncated 907 * write may be permanent, e.g. exhausted free space, 908 * accepting n == 0 would cause an infinite loop. 909 * 910 * Forwarding the iterator leaves the unwritten page 911 * behind, but there is not much we can do there if 912 * filesystem refuses to write it. 913 */ 914 if (n == 0) { 915 n = 1; 916 clearobjflags = FALSE; 917 } 918 np = vm_page_find_least(object, pi + n); 919 } 920#if 0 921 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 922#endif 923 924 if (clearobjflags) 925 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 926 return (res); 927} 928 929static int 930vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 931 int flags, boolean_t *clearobjflags, boolean_t *eio) 932{ 933 vm_page_t ma[vm_pageout_page_count], p_first, tp; 934 int count, i, mreq, runlen; 935 936 vm_page_lock_assert(p, MA_NOTOWNED); 937 VM_OBJECT_ASSERT_WLOCKED(object); 938 939 count = 1; 940 mreq = 0; 941 942 for (tp = p; count < vm_pageout_page_count; count++) { 943 tp = vm_page_next(tp); 944 if (tp == NULL || vm_page_busied(tp)) 945 break; 946 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 947 break; 948 } 949 950 for (p_first = p; count < vm_pageout_page_count; count++) { 951 tp = vm_page_prev(p_first); 952 if (tp == NULL || vm_page_busied(tp)) 953 break; 954 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 955 break; 956 p_first = tp; 957 mreq++; 958 } 959 960 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 961 ma[i] = tp; 962 963 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 964 return (runlen); 965} 966 967/* 968 * Note that there is absolutely no sense in writing out 969 * anonymous objects, so we track down the vnode object 970 * to write out. 971 * We invalidate (remove) all pages from the address space 972 * for semantic correctness. 973 * 974 * If the backing object is a device object with unmanaged pages, then any 975 * mappings to the specified range of pages must be removed before this 976 * function is called. 977 * 978 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 979 * may start out with a NULL object. 980 */ 981boolean_t 982vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 983 boolean_t syncio, boolean_t invalidate) 984{ 985 vm_object_t backing_object; 986 struct vnode *vp; 987 struct mount *mp; 988 int error, flags, fsync_after; 989 boolean_t res; 990 991 if (object == NULL) 992 return (TRUE); 993 res = TRUE; 994 error = 0; 995 VM_OBJECT_WLOCK(object); 996 while ((backing_object = object->backing_object) != NULL) { 997 VM_OBJECT_WLOCK(backing_object); 998 offset += object->backing_object_offset; 999 VM_OBJECT_WUNLOCK(object); 1000 object = backing_object; 1001 if (object->size < OFF_TO_IDX(offset + size)) 1002 size = IDX_TO_OFF(object->size) - offset; 1003 } 1004 /* 1005 * Flush pages if writing is allowed, invalidate them 1006 * if invalidation requested. Pages undergoing I/O 1007 * will be ignored by vm_object_page_remove(). 1008 * 1009 * We cannot lock the vnode and then wait for paging 1010 * to complete without deadlocking against vm_fault. 1011 * Instead we simply call vm_object_page_remove() and 1012 * allow it to block internally on a page-by-page 1013 * basis when it encounters pages undergoing async 1014 * I/O. 1015 */ 1016 if (object->type == OBJT_VNODE && 1017 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1018 vp = object->handle; 1019 VM_OBJECT_WUNLOCK(object); 1020 (void) vn_start_write(vp, &mp, V_WAIT); 1021 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1022 if (syncio && !invalidate && offset == 0 && 1023 OFF_TO_IDX(size) == object->size) { 1024 /* 1025 * If syncing the whole mapping of the file, 1026 * it is faster to schedule all the writes in 1027 * async mode, also allowing the clustering, 1028 * and then wait for i/o to complete. 1029 */ 1030 flags = 0; 1031 fsync_after = TRUE; 1032 } else { 1033 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1034 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1035 fsync_after = FALSE; 1036 } 1037 VM_OBJECT_WLOCK(object); 1038 res = vm_object_page_clean(object, offset, offset + size, 1039 flags); 1040 VM_OBJECT_WUNLOCK(object); 1041 if (fsync_after) 1042 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1043 VOP_UNLOCK(vp, 0); 1044 vn_finished_write(mp); 1045 if (error != 0) 1046 res = FALSE; 1047 VM_OBJECT_WLOCK(object); 1048 } 1049 if ((object->type == OBJT_VNODE || 1050 object->type == OBJT_DEVICE) && invalidate) { 1051 if (object->type == OBJT_DEVICE) 1052 /* 1053 * The option OBJPR_NOTMAPPED must be passed here 1054 * because vm_object_page_remove() cannot remove 1055 * unmanaged mappings. 1056 */ 1057 flags = OBJPR_NOTMAPPED; 1058 else if (old_msync) 1059 flags = OBJPR_NOTWIRED; 1060 else 1061 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED; 1062 vm_object_page_remove(object, OFF_TO_IDX(offset), 1063 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1064 } 1065 VM_OBJECT_WUNLOCK(object); 1066 return (res); 1067} 1068 1069/* 1070 * vm_object_madvise: 1071 * 1072 * Implements the madvise function at the object/page level. 1073 * 1074 * MADV_WILLNEED (any object) 1075 * 1076 * Activate the specified pages if they are resident. 1077 * 1078 * MADV_DONTNEED (any object) 1079 * 1080 * Deactivate the specified pages if they are resident. 1081 * 1082 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1083 * OBJ_ONEMAPPING only) 1084 * 1085 * Deactivate and clean the specified pages if they are 1086 * resident. This permits the process to reuse the pages 1087 * without faulting or the kernel to reclaim the pages 1088 * without I/O. 1089 */ 1090void 1091vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1092 int advise) 1093{ 1094 vm_pindex_t tpindex; 1095 vm_object_t backing_object, tobject; 1096 vm_page_t m; 1097 1098 if (object == NULL) 1099 return; 1100 VM_OBJECT_WLOCK(object); 1101 /* 1102 * Locate and adjust resident pages 1103 */ 1104 for (; pindex < end; pindex += 1) { 1105relookup: 1106 tobject = object; 1107 tpindex = pindex; 1108shadowlookup: 1109 /* 1110 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1111 * and those pages must be OBJ_ONEMAPPING. 1112 */ 1113 if (advise == MADV_FREE) { 1114 if ((tobject->type != OBJT_DEFAULT && 1115 tobject->type != OBJT_SWAP) || 1116 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1117 goto unlock_tobject; 1118 } 1119 } else if ((tobject->flags & OBJ_UNMANAGED) != 0) 1120 goto unlock_tobject; 1121 m = vm_page_lookup(tobject, tpindex); 1122 if (m == NULL && advise == MADV_WILLNEED) { 1123 /* 1124 * If the page is cached, reactivate it. 1125 */ 1126 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1127 VM_ALLOC_NOBUSY); 1128 } 1129 if (m == NULL) { 1130 /* 1131 * There may be swap even if there is no backing page 1132 */ 1133 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1134 swap_pager_freespace(tobject, tpindex, 1); 1135 /* 1136 * next object 1137 */ 1138 backing_object = tobject->backing_object; 1139 if (backing_object == NULL) 1140 goto unlock_tobject; 1141 VM_OBJECT_WLOCK(backing_object); 1142 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1143 if (tobject != object) 1144 VM_OBJECT_WUNLOCK(tobject); 1145 tobject = backing_object; 1146 goto shadowlookup; 1147 } else if (m->valid != VM_PAGE_BITS_ALL) 1148 goto unlock_tobject; 1149 /* 1150 * If the page is not in a normal state, skip it. 1151 */ 1152 vm_page_lock(m); 1153 if (m->hold_count != 0 || m->wire_count != 0) { 1154 vm_page_unlock(m); 1155 goto unlock_tobject; 1156 } 1157 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1158 ("vm_object_madvise: page %p is fictitious", m)); 1159 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1160 ("vm_object_madvise: page %p is not managed", m)); 1161 if (vm_page_busied(m)) { 1162 if (advise == MADV_WILLNEED) { 1163 /* 1164 * Reference the page before unlocking and 1165 * sleeping so that the page daemon is less 1166 * likely to reclaim it. 1167 */ 1168 vm_page_aflag_set(m, PGA_REFERENCED); 1169 } 1170 if (object != tobject) 1171 VM_OBJECT_WUNLOCK(object); 1172 VM_OBJECT_WUNLOCK(tobject); 1173 vm_page_busy_sleep(m, "madvpo"); 1174 VM_OBJECT_WLOCK(object); 1175 goto relookup; 1176 } 1177 if (advise == MADV_WILLNEED) { 1178 vm_page_activate(m); 1179 } else { 1180 vm_page_advise(m, advise); 1181 } 1182 vm_page_unlock(m); 1183 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1184 swap_pager_freespace(tobject, tpindex, 1); 1185unlock_tobject: 1186 if (tobject != object) 1187 VM_OBJECT_WUNLOCK(tobject); 1188 } 1189 VM_OBJECT_WUNLOCK(object); 1190} 1191 1192/* 1193 * vm_object_shadow: 1194 * 1195 * Create a new object which is backed by the 1196 * specified existing object range. The source 1197 * object reference is deallocated. 1198 * 1199 * The new object and offset into that object 1200 * are returned in the source parameters. 1201 */ 1202void 1203vm_object_shadow( 1204 vm_object_t *object, /* IN/OUT */ 1205 vm_ooffset_t *offset, /* IN/OUT */ 1206 vm_size_t length) 1207{ 1208 vm_object_t source; 1209 vm_object_t result; 1210 1211 source = *object; 1212 1213 /* 1214 * Don't create the new object if the old object isn't shared. 1215 */ 1216 if (source != NULL) { 1217 VM_OBJECT_WLOCK(source); 1218 if (source->ref_count == 1 && 1219 source->handle == NULL && 1220 (source->type == OBJT_DEFAULT || 1221 source->type == OBJT_SWAP)) { 1222 VM_OBJECT_WUNLOCK(source); 1223 return; 1224 } 1225 VM_OBJECT_WUNLOCK(source); 1226 } 1227 1228 /* 1229 * Allocate a new object with the given length. 1230 */ 1231 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1232 1233 /* 1234 * The new object shadows the source object, adding a reference to it. 1235 * Our caller changes his reference to point to the new object, 1236 * removing a reference to the source object. Net result: no change 1237 * of reference count. 1238 * 1239 * Try to optimize the result object's page color when shadowing 1240 * in order to maintain page coloring consistency in the combined 1241 * shadowed object. 1242 */ 1243 result->backing_object = source; 1244 /* 1245 * Store the offset into the source object, and fix up the offset into 1246 * the new object. 1247 */ 1248 result->backing_object_offset = *offset; 1249 if (source != NULL) { 1250 VM_OBJECT_WLOCK(source); 1251 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1252 source->shadow_count++; 1253#if VM_NRESERVLEVEL > 0 1254 result->flags |= source->flags & OBJ_COLORED; 1255 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1256 ((1 << (VM_NFREEORDER - 1)) - 1); 1257#endif 1258 VM_OBJECT_WUNLOCK(source); 1259 } 1260 1261 1262 /* 1263 * Return the new things 1264 */ 1265 *offset = 0; 1266 *object = result; 1267} 1268 1269/* 1270 * vm_object_split: 1271 * 1272 * Split the pages in a map entry into a new object. This affords 1273 * easier removal of unused pages, and keeps object inheritance from 1274 * being a negative impact on memory usage. 1275 */ 1276void 1277vm_object_split(vm_map_entry_t entry) 1278{ 1279 vm_page_t m, m_next; 1280 vm_object_t orig_object, new_object, source; 1281 vm_pindex_t idx, offidxstart; 1282 vm_size_t size; 1283 1284 orig_object = entry->object.vm_object; 1285 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1286 return; 1287 if (orig_object->ref_count <= 1) 1288 return; 1289 VM_OBJECT_WUNLOCK(orig_object); 1290 1291 offidxstart = OFF_TO_IDX(entry->offset); 1292 size = atop(entry->end - entry->start); 1293 1294 /* 1295 * If swap_pager_copy() is later called, it will convert new_object 1296 * into a swap object. 1297 */ 1298 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1299 1300 /* 1301 * At this point, the new object is still private, so the order in 1302 * which the original and new objects are locked does not matter. 1303 */ 1304 VM_OBJECT_WLOCK(new_object); 1305 VM_OBJECT_WLOCK(orig_object); 1306 source = orig_object->backing_object; 1307 if (source != NULL) { 1308 VM_OBJECT_WLOCK(source); 1309 if ((source->flags & OBJ_DEAD) != 0) { 1310 VM_OBJECT_WUNLOCK(source); 1311 VM_OBJECT_WUNLOCK(orig_object); 1312 VM_OBJECT_WUNLOCK(new_object); 1313 vm_object_deallocate(new_object); 1314 VM_OBJECT_WLOCK(orig_object); 1315 return; 1316 } 1317 LIST_INSERT_HEAD(&source->shadow_head, 1318 new_object, shadow_list); 1319 source->shadow_count++; 1320 vm_object_reference_locked(source); /* for new_object */ 1321 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1322 VM_OBJECT_WUNLOCK(source); 1323 new_object->backing_object_offset = 1324 orig_object->backing_object_offset + entry->offset; 1325 new_object->backing_object = source; 1326 } 1327 if (orig_object->cred != NULL) { 1328 new_object->cred = orig_object->cred; 1329 crhold(orig_object->cred); 1330 new_object->charge = ptoa(size); 1331 KASSERT(orig_object->charge >= ptoa(size), 1332 ("orig_object->charge < 0")); 1333 orig_object->charge -= ptoa(size); 1334 } 1335retry: 1336 m = vm_page_find_least(orig_object, offidxstart); 1337 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1338 m = m_next) { 1339 m_next = TAILQ_NEXT(m, listq); 1340 1341 /* 1342 * We must wait for pending I/O to complete before we can 1343 * rename the page. 1344 * 1345 * We do not have to VM_PROT_NONE the page as mappings should 1346 * not be changed by this operation. 1347 */ 1348 if (vm_page_busied(m)) { 1349 VM_OBJECT_WUNLOCK(new_object); 1350 vm_page_lock(m); 1351 VM_OBJECT_WUNLOCK(orig_object); 1352 vm_page_busy_sleep(m, "spltwt"); 1353 VM_OBJECT_WLOCK(orig_object); 1354 VM_OBJECT_WLOCK(new_object); 1355 goto retry; 1356 } 1357 1358 /* vm_page_rename() will handle dirty and cache. */ 1359 if (vm_page_rename(m, new_object, idx)) { 1360 VM_OBJECT_WUNLOCK(new_object); 1361 VM_OBJECT_WUNLOCK(orig_object); 1362 VM_WAIT; 1363 VM_OBJECT_WLOCK(orig_object); 1364 VM_OBJECT_WLOCK(new_object); 1365 goto retry; 1366 } 1367#if VM_NRESERVLEVEL > 0 1368 /* 1369 * If some of the reservation's allocated pages remain with 1370 * the original object, then transferring the reservation to 1371 * the new object is neither particularly beneficial nor 1372 * particularly harmful as compared to leaving the reservation 1373 * with the original object. If, however, all of the 1374 * reservation's allocated pages are transferred to the new 1375 * object, then transferring the reservation is typically 1376 * beneficial. Determining which of these two cases applies 1377 * would be more costly than unconditionally renaming the 1378 * reservation. 1379 */ 1380 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1381#endif 1382 if (orig_object->type == OBJT_SWAP) 1383 vm_page_xbusy(m); 1384 } 1385 if (orig_object->type == OBJT_SWAP) { 1386 /* 1387 * swap_pager_copy() can sleep, in which case the orig_object's 1388 * and new_object's locks are released and reacquired. 1389 */ 1390 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1391 TAILQ_FOREACH(m, &new_object->memq, listq) 1392 vm_page_xunbusy(m); 1393 1394 /* 1395 * Transfer any cached pages from orig_object to new_object. 1396 * If swap_pager_copy() found swapped out pages within the 1397 * specified range of orig_object, then it changed 1398 * new_object's type to OBJT_SWAP when it transferred those 1399 * pages to new_object. Otherwise, new_object's type 1400 * should still be OBJT_DEFAULT and orig_object should not 1401 * contain any cached pages within the specified range. 1402 */ 1403 if (__predict_false(!vm_object_cache_is_empty(orig_object))) 1404 vm_page_cache_transfer(orig_object, offidxstart, 1405 new_object); 1406 } 1407 VM_OBJECT_WUNLOCK(orig_object); 1408 VM_OBJECT_WUNLOCK(new_object); 1409 entry->object.vm_object = new_object; 1410 entry->offset = 0LL; 1411 vm_object_deallocate(orig_object); 1412 VM_OBJECT_WLOCK(new_object); 1413} 1414 1415#define OBSC_TEST_ALL_SHADOWED 0x0001 1416#define OBSC_COLLAPSE_NOWAIT 0x0002 1417#define OBSC_COLLAPSE_WAIT 0x0004 1418 1419static int 1420vm_object_backing_scan(vm_object_t object, int op) 1421{ 1422 int r = 1; 1423 vm_page_t p; 1424 vm_object_t backing_object; 1425 vm_pindex_t backing_offset_index; 1426 1427 VM_OBJECT_ASSERT_WLOCKED(object); 1428 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1429 1430 backing_object = object->backing_object; 1431 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1432 1433 /* 1434 * Initial conditions 1435 */ 1436 if (op & OBSC_TEST_ALL_SHADOWED) { 1437 /* 1438 * We do not want to have to test for the existence of cache 1439 * or swap pages in the backing object. XXX but with the 1440 * new swapper this would be pretty easy to do. 1441 * 1442 * XXX what about anonymous MAP_SHARED memory that hasn't 1443 * been ZFOD faulted yet? If we do not test for this, the 1444 * shadow test may succeed! XXX 1445 */ 1446 if (backing_object->type != OBJT_DEFAULT) { 1447 return (0); 1448 } 1449 } 1450 if (op & OBSC_COLLAPSE_WAIT) { 1451 vm_object_set_flag(backing_object, OBJ_DEAD); 1452 } 1453 1454 /* 1455 * Our scan 1456 */ 1457 p = TAILQ_FIRST(&backing_object->memq); 1458 while (p) { 1459 vm_page_t next = TAILQ_NEXT(p, listq); 1460 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1461 1462 if (op & OBSC_TEST_ALL_SHADOWED) { 1463 vm_page_t pp; 1464 1465 /* 1466 * Ignore pages outside the parent object's range 1467 * and outside the parent object's mapping of the 1468 * backing object. 1469 * 1470 * note that we do not busy the backing object's 1471 * page. 1472 */ 1473 if ( 1474 p->pindex < backing_offset_index || 1475 new_pindex >= object->size 1476 ) { 1477 p = next; 1478 continue; 1479 } 1480 1481 /* 1482 * See if the parent has the page or if the parent's 1483 * object pager has the page. If the parent has the 1484 * page but the page is not valid, the parent's 1485 * object pager must have the page. 1486 * 1487 * If this fails, the parent does not completely shadow 1488 * the object and we might as well give up now. 1489 */ 1490 1491 pp = vm_page_lookup(object, new_pindex); 1492 if ( 1493 (pp == NULL || pp->valid == 0) && 1494 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1495 ) { 1496 r = 0; 1497 break; 1498 } 1499 } 1500 1501 /* 1502 * Check for busy page 1503 */ 1504 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1505 vm_page_t pp; 1506 1507 if (op & OBSC_COLLAPSE_NOWAIT) { 1508 if (!p->valid || vm_page_busied(p)) { 1509 p = next; 1510 continue; 1511 } 1512 } else if (op & OBSC_COLLAPSE_WAIT) { 1513 if (vm_page_busied(p)) { 1514 VM_OBJECT_WUNLOCK(object); 1515 vm_page_lock(p); 1516 VM_OBJECT_WUNLOCK(backing_object); 1517 vm_page_busy_sleep(p, "vmocol"); 1518 VM_OBJECT_WLOCK(object); 1519 VM_OBJECT_WLOCK(backing_object); 1520 /* 1521 * If we slept, anything could have 1522 * happened. Since the object is 1523 * marked dead, the backing offset 1524 * should not have changed so we 1525 * just restart our scan. 1526 */ 1527 p = TAILQ_FIRST(&backing_object->memq); 1528 continue; 1529 } 1530 } 1531 1532 KASSERT( 1533 p->object == backing_object, 1534 ("vm_object_backing_scan: object mismatch") 1535 ); 1536 1537 if ( 1538 p->pindex < backing_offset_index || 1539 new_pindex >= object->size 1540 ) { 1541 if (backing_object->type == OBJT_SWAP) 1542 swap_pager_freespace(backing_object, 1543 p->pindex, 1); 1544 1545 /* 1546 * Page is out of the parent object's range, we 1547 * can simply destroy it. 1548 */ 1549 vm_page_lock(p); 1550 KASSERT(!pmap_page_is_mapped(p), 1551 ("freeing mapped page %p", p)); 1552 if (p->wire_count == 0) 1553 vm_page_free(p); 1554 else 1555 vm_page_remove(p); 1556 vm_page_unlock(p); 1557 p = next; 1558 continue; 1559 } 1560 1561 pp = vm_page_lookup(object, new_pindex); 1562 if ( 1563 (op & OBSC_COLLAPSE_NOWAIT) != 0 && 1564 (pp != NULL && pp->valid == 0) 1565 ) { 1566 if (backing_object->type == OBJT_SWAP) 1567 swap_pager_freespace(backing_object, 1568 p->pindex, 1); 1569 1570 /* 1571 * The page in the parent is not (yet) valid. 1572 * We don't know anything about the state of 1573 * the original page. It might be mapped, 1574 * so we must avoid the next if here. 1575 * 1576 * This is due to a race in vm_fault() where 1577 * we must unbusy the original (backing_obj) 1578 * page before we can (re)lock the parent. 1579 * Hence we can get here. 1580 */ 1581 p = next; 1582 continue; 1583 } 1584 if ( 1585 pp != NULL || 1586 vm_pager_has_page(object, new_pindex, NULL, NULL) 1587 ) { 1588 if (backing_object->type == OBJT_SWAP) 1589 swap_pager_freespace(backing_object, 1590 p->pindex, 1); 1591 1592 /* 1593 * page already exists in parent OR swap exists 1594 * for this location in the parent. Destroy 1595 * the original page from the backing object. 1596 * 1597 * Leave the parent's page alone 1598 */ 1599 vm_page_lock(p); 1600 KASSERT(!pmap_page_is_mapped(p), 1601 ("freeing mapped page %p", p)); 1602 if (p->wire_count == 0) 1603 vm_page_free(p); 1604 else 1605 vm_page_remove(p); 1606 vm_page_unlock(p); 1607 p = next; 1608 continue; 1609 } 1610 1611 /* 1612 * Page does not exist in parent, rename the 1613 * page from the backing object to the main object. 1614 * 1615 * If the page was mapped to a process, it can remain 1616 * mapped through the rename. 1617 * vm_page_rename() will handle dirty and cache. 1618 */ 1619 if (vm_page_rename(p, object, new_pindex)) { 1620 if (op & OBSC_COLLAPSE_NOWAIT) { 1621 p = next; 1622 continue; 1623 } 1624 VM_OBJECT_WLOCK(backing_object); 1625 VM_OBJECT_WUNLOCK(object); 1626 VM_WAIT; 1627 VM_OBJECT_WLOCK(object); 1628 VM_OBJECT_WLOCK(backing_object); 1629 p = TAILQ_FIRST(&backing_object->memq); 1630 continue; 1631 } 1632 1633 /* Use the old pindex to free the right page. */ 1634 if (backing_object->type == OBJT_SWAP) 1635 swap_pager_freespace(backing_object, 1636 new_pindex + backing_offset_index, 1); 1637 1638#if VM_NRESERVLEVEL > 0 1639 /* 1640 * Rename the reservation. 1641 */ 1642 vm_reserv_rename(p, object, backing_object, 1643 backing_offset_index); 1644#endif 1645 } 1646 p = next; 1647 } 1648 return (r); 1649} 1650 1651 1652/* 1653 * this version of collapse allows the operation to occur earlier and 1654 * when paging_in_progress is true for an object... This is not a complete 1655 * operation, but should plug 99.9% of the rest of the leaks. 1656 */ 1657static void 1658vm_object_qcollapse(vm_object_t object) 1659{ 1660 vm_object_t backing_object = object->backing_object; 1661 1662 VM_OBJECT_ASSERT_WLOCKED(object); 1663 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1664 1665 if (backing_object->ref_count != 1) 1666 return; 1667 1668 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1669} 1670 1671/* 1672 * vm_object_collapse: 1673 * 1674 * Collapse an object with the object backing it. 1675 * Pages in the backing object are moved into the 1676 * parent, and the backing object is deallocated. 1677 */ 1678void 1679vm_object_collapse(vm_object_t object) 1680{ 1681 VM_OBJECT_ASSERT_WLOCKED(object); 1682 1683 while (TRUE) { 1684 vm_object_t backing_object; 1685 1686 /* 1687 * Verify that the conditions are right for collapse: 1688 * 1689 * The object exists and the backing object exists. 1690 */ 1691 if ((backing_object = object->backing_object) == NULL) 1692 break; 1693 1694 /* 1695 * we check the backing object first, because it is most likely 1696 * not collapsable. 1697 */ 1698 VM_OBJECT_WLOCK(backing_object); 1699 if (backing_object->handle != NULL || 1700 (backing_object->type != OBJT_DEFAULT && 1701 backing_object->type != OBJT_SWAP) || 1702 (backing_object->flags & OBJ_DEAD) || 1703 object->handle != NULL || 1704 (object->type != OBJT_DEFAULT && 1705 object->type != OBJT_SWAP) || 1706 (object->flags & OBJ_DEAD)) { 1707 VM_OBJECT_WUNLOCK(backing_object); 1708 break; 1709 } 1710 1711 if ( 1712 object->paging_in_progress != 0 || 1713 backing_object->paging_in_progress != 0 1714 ) { 1715 vm_object_qcollapse(object); 1716 VM_OBJECT_WUNLOCK(backing_object); 1717 break; 1718 } 1719 /* 1720 * We know that we can either collapse the backing object (if 1721 * the parent is the only reference to it) or (perhaps) have 1722 * the parent bypass the object if the parent happens to shadow 1723 * all the resident pages in the entire backing object. 1724 * 1725 * This is ignoring pager-backed pages such as swap pages. 1726 * vm_object_backing_scan fails the shadowing test in this 1727 * case. 1728 */ 1729 if (backing_object->ref_count == 1) { 1730 /* 1731 * If there is exactly one reference to the backing 1732 * object, we can collapse it into the parent. 1733 */ 1734 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1735 1736#if VM_NRESERVLEVEL > 0 1737 /* 1738 * Break any reservations from backing_object. 1739 */ 1740 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1741 vm_reserv_break_all(backing_object); 1742#endif 1743 1744 /* 1745 * Move the pager from backing_object to object. 1746 */ 1747 if (backing_object->type == OBJT_SWAP) { 1748 /* 1749 * swap_pager_copy() can sleep, in which case 1750 * the backing_object's and object's locks are 1751 * released and reacquired. 1752 * Since swap_pager_copy() is being asked to 1753 * destroy the source, it will change the 1754 * backing_object's type to OBJT_DEFAULT. 1755 */ 1756 swap_pager_copy( 1757 backing_object, 1758 object, 1759 OFF_TO_IDX(object->backing_object_offset), TRUE); 1760 1761 /* 1762 * Free any cached pages from backing_object. 1763 */ 1764 if (__predict_false( 1765 !vm_object_cache_is_empty(backing_object))) 1766 vm_page_cache_free(backing_object, 0, 0); 1767 } 1768 /* 1769 * Object now shadows whatever backing_object did. 1770 * Note that the reference to 1771 * backing_object->backing_object moves from within 1772 * backing_object to within object. 1773 */ 1774 LIST_REMOVE(object, shadow_list); 1775 backing_object->shadow_count--; 1776 if (backing_object->backing_object) { 1777 VM_OBJECT_WLOCK(backing_object->backing_object); 1778 LIST_REMOVE(backing_object, shadow_list); 1779 LIST_INSERT_HEAD( 1780 &backing_object->backing_object->shadow_head, 1781 object, shadow_list); 1782 /* 1783 * The shadow_count has not changed. 1784 */ 1785 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1786 } 1787 object->backing_object = backing_object->backing_object; 1788 object->backing_object_offset += 1789 backing_object->backing_object_offset; 1790 1791 /* 1792 * Discard backing_object. 1793 * 1794 * Since the backing object has no pages, no pager left, 1795 * and no object references within it, all that is 1796 * necessary is to dispose of it. 1797 */ 1798 KASSERT(backing_object->ref_count == 1, ( 1799"backing_object %p was somehow re-referenced during collapse!", 1800 backing_object)); 1801 VM_OBJECT_WUNLOCK(backing_object); 1802 vm_object_destroy(backing_object); 1803 1804 object_collapses++; 1805 } else { 1806 vm_object_t new_backing_object; 1807 1808 /* 1809 * If we do not entirely shadow the backing object, 1810 * there is nothing we can do so we give up. 1811 */ 1812 if (object->resident_page_count != object->size && 1813 vm_object_backing_scan(object, 1814 OBSC_TEST_ALL_SHADOWED) == 0) { 1815 VM_OBJECT_WUNLOCK(backing_object); 1816 break; 1817 } 1818 1819 /* 1820 * Make the parent shadow the next object in the 1821 * chain. Deallocating backing_object will not remove 1822 * it, since its reference count is at least 2. 1823 */ 1824 LIST_REMOVE(object, shadow_list); 1825 backing_object->shadow_count--; 1826 1827 new_backing_object = backing_object->backing_object; 1828 if ((object->backing_object = new_backing_object) != NULL) { 1829 VM_OBJECT_WLOCK(new_backing_object); 1830 LIST_INSERT_HEAD( 1831 &new_backing_object->shadow_head, 1832 object, 1833 shadow_list 1834 ); 1835 new_backing_object->shadow_count++; 1836 vm_object_reference_locked(new_backing_object); 1837 VM_OBJECT_WUNLOCK(new_backing_object); 1838 object->backing_object_offset += 1839 backing_object->backing_object_offset; 1840 } 1841 1842 /* 1843 * Drop the reference count on backing_object. Since 1844 * its ref_count was at least 2, it will not vanish. 1845 */ 1846 backing_object->ref_count--; 1847 VM_OBJECT_WUNLOCK(backing_object); 1848 object_bypasses++; 1849 } 1850 1851 /* 1852 * Try again with this object's new backing object. 1853 */ 1854 } 1855} 1856 1857/* 1858 * vm_object_page_remove: 1859 * 1860 * For the given object, either frees or invalidates each of the 1861 * specified pages. In general, a page is freed. However, if a page is 1862 * wired for any reason other than the existence of a managed, wired 1863 * mapping, then it may be invalidated but not removed from the object. 1864 * Pages are specified by the given range ["start", "end") and the option 1865 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1866 * extends from "start" to the end of the object. If the option 1867 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1868 * specified range are affected. If the option OBJPR_NOTMAPPED is 1869 * specified, then the pages within the specified range must have no 1870 * mappings. Otherwise, if this option is not specified, any mappings to 1871 * the specified pages are removed before the pages are freed or 1872 * invalidated. 1873 * 1874 * In general, this operation should only be performed on objects that 1875 * contain managed pages. There are, however, two exceptions. First, it 1876 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1877 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1878 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1879 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1880 * 1881 * The object must be locked. 1882 */ 1883void 1884vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1885 int options) 1886{ 1887 vm_page_t p, next; 1888 int wirings; 1889 1890 VM_OBJECT_ASSERT_WLOCKED(object); 1891 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1892 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1893 ("vm_object_page_remove: illegal options for object %p", object)); 1894 if (object->resident_page_count == 0) 1895 goto skipmemq; 1896 vm_object_pip_add(object, 1); 1897again: 1898 p = vm_page_find_least(object, start); 1899 1900 /* 1901 * Here, the variable "p" is either (1) the page with the least pindex 1902 * greater than or equal to the parameter "start" or (2) NULL. 1903 */ 1904 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1905 next = TAILQ_NEXT(p, listq); 1906 1907 /* 1908 * If the page is wired for any reason besides the existence 1909 * of managed, wired mappings, then it cannot be freed. For 1910 * example, fictitious pages, which represent device memory, 1911 * are inherently wired and cannot be freed. They can, 1912 * however, be invalidated if the option OBJPR_CLEANONLY is 1913 * not specified. 1914 */ 1915 vm_page_lock(p); 1916 if (vm_page_xbusied(p)) { 1917 VM_OBJECT_WUNLOCK(object); 1918 vm_page_busy_sleep(p, "vmopax"); 1919 VM_OBJECT_WLOCK(object); 1920 goto again; 1921 } 1922 if ((wirings = p->wire_count) != 0 && 1923 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) { 1924 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) == 1925 0) { 1926 pmap_remove_all(p); 1927 /* Account for removal of wired mappings. */ 1928 if (wirings != 0) 1929 p->wire_count -= wirings; 1930 } 1931 if ((options & OBJPR_CLEANONLY) == 0) { 1932 p->valid = 0; 1933 vm_page_undirty(p); 1934 } 1935 goto next; 1936 } 1937 if (vm_page_busied(p)) { 1938 VM_OBJECT_WUNLOCK(object); 1939 vm_page_busy_sleep(p, "vmopar"); 1940 VM_OBJECT_WLOCK(object); 1941 goto again; 1942 } 1943 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1944 ("vm_object_page_remove: page %p is fictitious", p)); 1945 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1946 if ((options & OBJPR_NOTMAPPED) == 0) 1947 pmap_remove_write(p); 1948 if (p->dirty) 1949 goto next; 1950 } 1951 if ((options & OBJPR_NOTMAPPED) == 0) { 1952 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0) 1953 goto next; 1954 pmap_remove_all(p); 1955 /* Account for removal of wired mappings. */ 1956 if (wirings != 0) { 1957 KASSERT(p->wire_count == wirings, 1958 ("inconsistent wire count %d %d %p", 1959 p->wire_count, wirings, p)); 1960 p->wire_count = 0; 1961 atomic_subtract_int(&cnt.v_wire_count, 1); 1962 } 1963 } 1964 vm_page_free(p); 1965next: 1966 vm_page_unlock(p); 1967 } 1968 vm_object_pip_wakeup(object); 1969skipmemq: 1970 if (__predict_false(!vm_object_cache_is_empty(object))) 1971 vm_page_cache_free(object, start, end); 1972} 1973 1974/* 1975 * vm_object_page_cache: 1976 * 1977 * For the given object, attempt to move the specified clean 1978 * pages to the cache queue. If a page is wired for any reason, 1979 * then it will not be changed. Pages are specified by the given 1980 * range ["start", "end"). As a special case, if "end" is zero, 1981 * then the range extends from "start" to the end of the object. 1982 * Any mappings to the specified pages are removed before the 1983 * pages are moved to the cache queue. 1984 * 1985 * This operation should only be performed on objects that 1986 * contain non-fictitious, managed pages. 1987 * 1988 * The object must be locked. 1989 */ 1990void 1991vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1992{ 1993 struct mtx *mtx, *new_mtx; 1994 vm_page_t p, next; 1995 1996 VM_OBJECT_ASSERT_WLOCKED(object); 1997 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 1998 ("vm_object_page_cache: illegal object %p", object)); 1999 if (object->resident_page_count == 0) 2000 return; 2001 p = vm_page_find_least(object, start); 2002 2003 /* 2004 * Here, the variable "p" is either (1) the page with the least pindex 2005 * greater than or equal to the parameter "start" or (2) NULL. 2006 */ 2007 mtx = NULL; 2008 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2009 next = TAILQ_NEXT(p, listq); 2010 2011 /* 2012 * Avoid releasing and reacquiring the same page lock. 2013 */ 2014 new_mtx = vm_page_lockptr(p); 2015 if (mtx != new_mtx) { 2016 if (mtx != NULL) 2017 mtx_unlock(mtx); 2018 mtx = new_mtx; 2019 mtx_lock(mtx); 2020 } 2021 vm_page_try_to_cache(p); 2022 } 2023 if (mtx != NULL) 2024 mtx_unlock(mtx); 2025} 2026 2027/* 2028 * Populate the specified range of the object with valid pages. Returns 2029 * TRUE if the range is successfully populated and FALSE otherwise. 2030 * 2031 * Note: This function should be optimized to pass a larger array of 2032 * pages to vm_pager_get_pages() before it is applied to a non- 2033 * OBJT_DEVICE object. 2034 * 2035 * The object must be locked. 2036 */ 2037boolean_t 2038vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2039{ 2040 vm_page_t m, ma[1]; 2041 vm_pindex_t pindex; 2042 int rv; 2043 2044 VM_OBJECT_ASSERT_WLOCKED(object); 2045 for (pindex = start; pindex < end; pindex++) { 2046 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2047 if (m->valid != VM_PAGE_BITS_ALL) { 2048 ma[0] = m; 2049 rv = vm_pager_get_pages(object, ma, 1, 0); 2050 m = vm_page_lookup(object, pindex); 2051 if (m == NULL) 2052 break; 2053 if (rv != VM_PAGER_OK) { 2054 vm_page_lock(m); 2055 vm_page_free(m); 2056 vm_page_unlock(m); 2057 break; 2058 } 2059 } 2060 /* 2061 * Keep "m" busy because a subsequent iteration may unlock 2062 * the object. 2063 */ 2064 } 2065 if (pindex > start) { 2066 m = vm_page_lookup(object, start); 2067 while (m != NULL && m->pindex < pindex) { 2068 vm_page_xunbusy(m); 2069 m = TAILQ_NEXT(m, listq); 2070 } 2071 } 2072 return (pindex == end); 2073} 2074 2075/* 2076 * Routine: vm_object_coalesce 2077 * Function: Coalesces two objects backing up adjoining 2078 * regions of memory into a single object. 2079 * 2080 * returns TRUE if objects were combined. 2081 * 2082 * NOTE: Only works at the moment if the second object is NULL - 2083 * if it's not, which object do we lock first? 2084 * 2085 * Parameters: 2086 * prev_object First object to coalesce 2087 * prev_offset Offset into prev_object 2088 * prev_size Size of reference to prev_object 2089 * next_size Size of reference to the second object 2090 * reserved Indicator that extension region has 2091 * swap accounted for 2092 * 2093 * Conditions: 2094 * The object must *not* be locked. 2095 */ 2096boolean_t 2097vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2098 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2099{ 2100 vm_pindex_t next_pindex; 2101 2102 if (prev_object == NULL) 2103 return (TRUE); 2104 VM_OBJECT_WLOCK(prev_object); 2105 if ((prev_object->type != OBJT_DEFAULT && 2106 prev_object->type != OBJT_SWAP) || 2107 (prev_object->flags & OBJ_TMPFS) != 0) { 2108 VM_OBJECT_WUNLOCK(prev_object); 2109 return (FALSE); 2110 } 2111 2112 /* 2113 * Try to collapse the object first 2114 */ 2115 vm_object_collapse(prev_object); 2116 2117 /* 2118 * Can't coalesce if: . more than one reference . paged out . shadows 2119 * another object . has a copy elsewhere (any of which mean that the 2120 * pages not mapped to prev_entry may be in use anyway) 2121 */ 2122 if (prev_object->backing_object != NULL) { 2123 VM_OBJECT_WUNLOCK(prev_object); 2124 return (FALSE); 2125 } 2126 2127 prev_size >>= PAGE_SHIFT; 2128 next_size >>= PAGE_SHIFT; 2129 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2130 2131 if ((prev_object->ref_count > 1) && 2132 (prev_object->size != next_pindex)) { 2133 VM_OBJECT_WUNLOCK(prev_object); 2134 return (FALSE); 2135 } 2136 2137 /* 2138 * Account for the charge. 2139 */ 2140 if (prev_object->cred != NULL) { 2141 2142 /* 2143 * If prev_object was charged, then this mapping, 2144 * althought not charged now, may become writable 2145 * later. Non-NULL cred in the object would prevent 2146 * swap reservation during enabling of the write 2147 * access, so reserve swap now. Failed reservation 2148 * cause allocation of the separate object for the map 2149 * entry, and swap reservation for this entry is 2150 * managed in appropriate time. 2151 */ 2152 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2153 prev_object->cred)) { 2154 return (FALSE); 2155 } 2156 prev_object->charge += ptoa(next_size); 2157 } 2158 2159 /* 2160 * Remove any pages that may still be in the object from a previous 2161 * deallocation. 2162 */ 2163 if (next_pindex < prev_object->size) { 2164 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2165 next_size, 0); 2166 if (prev_object->type == OBJT_SWAP) 2167 swap_pager_freespace(prev_object, 2168 next_pindex, next_size); 2169#if 0 2170 if (prev_object->cred != NULL) { 2171 KASSERT(prev_object->charge >= 2172 ptoa(prev_object->size - next_pindex), 2173 ("object %p overcharged 1 %jx %jx", prev_object, 2174 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2175 prev_object->charge -= ptoa(prev_object->size - 2176 next_pindex); 2177 } 2178#endif 2179 } 2180 2181 /* 2182 * Extend the object if necessary. 2183 */ 2184 if (next_pindex + next_size > prev_object->size) 2185 prev_object->size = next_pindex + next_size; 2186 2187 VM_OBJECT_WUNLOCK(prev_object); 2188 return (TRUE); 2189} 2190 2191void 2192vm_object_set_writeable_dirty(vm_object_t object) 2193{ 2194 2195 VM_OBJECT_ASSERT_WLOCKED(object); 2196 if (object->type != OBJT_VNODE) 2197 return; 2198 object->generation++; 2199 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2200 return; 2201 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2202} 2203 2204#include "opt_ddb.h" 2205#ifdef DDB 2206#include <sys/kernel.h> 2207 2208#include <sys/cons.h> 2209 2210#include <ddb/ddb.h> 2211 2212static int 2213_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2214{ 2215 vm_map_t tmpm; 2216 vm_map_entry_t tmpe; 2217 vm_object_t obj; 2218 int entcount; 2219 2220 if (map == 0) 2221 return 0; 2222 2223 if (entry == 0) { 2224 tmpe = map->header.next; 2225 entcount = map->nentries; 2226 while (entcount-- && (tmpe != &map->header)) { 2227 if (_vm_object_in_map(map, object, tmpe)) { 2228 return 1; 2229 } 2230 tmpe = tmpe->next; 2231 } 2232 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2233 tmpm = entry->object.sub_map; 2234 tmpe = tmpm->header.next; 2235 entcount = tmpm->nentries; 2236 while (entcount-- && tmpe != &tmpm->header) { 2237 if (_vm_object_in_map(tmpm, object, tmpe)) { 2238 return 1; 2239 } 2240 tmpe = tmpe->next; 2241 } 2242 } else if ((obj = entry->object.vm_object) != NULL) { 2243 for (; obj; obj = obj->backing_object) 2244 if (obj == object) { 2245 return 1; 2246 } 2247 } 2248 return 0; 2249} 2250 2251static int 2252vm_object_in_map(vm_object_t object) 2253{ 2254 struct proc *p; 2255 2256 /* sx_slock(&allproc_lock); */ 2257 FOREACH_PROC_IN_SYSTEM(p) { 2258 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2259 continue; 2260 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2261 /* sx_sunlock(&allproc_lock); */ 2262 return 1; 2263 } 2264 } 2265 /* sx_sunlock(&allproc_lock); */ 2266 if (_vm_object_in_map(kernel_map, object, 0)) 2267 return 1; 2268 return 0; 2269} 2270 2271DB_SHOW_COMMAND(vmochk, vm_object_check) 2272{ 2273 vm_object_t object; 2274 2275 /* 2276 * make sure that internal objs are in a map somewhere 2277 * and none have zero ref counts. 2278 */ 2279 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2280 if (object->handle == NULL && 2281 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2282 if (object->ref_count == 0) { 2283 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2284 (long)object->size); 2285 } 2286 if (!vm_object_in_map(object)) { 2287 db_printf( 2288 "vmochk: internal obj is not in a map: " 2289 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2290 object->ref_count, (u_long)object->size, 2291 (u_long)object->size, 2292 (void *)object->backing_object); 2293 } 2294 } 2295 } 2296} 2297 2298/* 2299 * vm_object_print: [ debug ] 2300 */ 2301DB_SHOW_COMMAND(object, vm_object_print_static) 2302{ 2303 /* XXX convert args. */ 2304 vm_object_t object = (vm_object_t)addr; 2305 boolean_t full = have_addr; 2306 2307 vm_page_t p; 2308 2309 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2310#define count was_count 2311 2312 int count; 2313 2314 if (object == NULL) 2315 return; 2316 2317 db_iprintf( 2318 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2319 object, (int)object->type, (uintmax_t)object->size, 2320 object->resident_page_count, object->ref_count, object->flags, 2321 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2322 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2323 object->shadow_count, 2324 object->backing_object ? object->backing_object->ref_count : 0, 2325 object->backing_object, (uintmax_t)object->backing_object_offset); 2326 2327 if (!full) 2328 return; 2329 2330 db_indent += 2; 2331 count = 0; 2332 TAILQ_FOREACH(p, &object->memq, listq) { 2333 if (count == 0) 2334 db_iprintf("memory:="); 2335 else if (count == 6) { 2336 db_printf("\n"); 2337 db_iprintf(" ..."); 2338 count = 0; 2339 } else 2340 db_printf(","); 2341 count++; 2342 2343 db_printf("(off=0x%jx,page=0x%jx)", 2344 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2345 } 2346 if (count != 0) 2347 db_printf("\n"); 2348 db_indent -= 2; 2349} 2350 2351/* XXX. */ 2352#undef count 2353 2354/* XXX need this non-static entry for calling from vm_map_print. */ 2355void 2356vm_object_print( 2357 /* db_expr_t */ long addr, 2358 boolean_t have_addr, 2359 /* db_expr_t */ long count, 2360 char *modif) 2361{ 2362 vm_object_print_static(addr, have_addr, count, modif); 2363} 2364 2365DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2366{ 2367 vm_object_t object; 2368 vm_pindex_t fidx; 2369 vm_paddr_t pa; 2370 vm_page_t m, prev_m; 2371 int rcount, nl, c; 2372 2373 nl = 0; 2374 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2375 db_printf("new object: %p\n", (void *)object); 2376 if (nl > 18) { 2377 c = cngetc(); 2378 if (c != ' ') 2379 return; 2380 nl = 0; 2381 } 2382 nl++; 2383 rcount = 0; 2384 fidx = 0; 2385 pa = -1; 2386 TAILQ_FOREACH(m, &object->memq, listq) { 2387 if (m->pindex > 128) 2388 break; 2389 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2390 prev_m->pindex + 1 != m->pindex) { 2391 if (rcount) { 2392 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2393 (long)fidx, rcount, (long)pa); 2394 if (nl > 18) { 2395 c = cngetc(); 2396 if (c != ' ') 2397 return; 2398 nl = 0; 2399 } 2400 nl++; 2401 rcount = 0; 2402 } 2403 } 2404 if (rcount && 2405 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2406 ++rcount; 2407 continue; 2408 } 2409 if (rcount) { 2410 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2411 (long)fidx, rcount, (long)pa); 2412 if (nl > 18) { 2413 c = cngetc(); 2414 if (c != ' ') 2415 return; 2416 nl = 0; 2417 } 2418 nl++; 2419 } 2420 fidx = m->pindex; 2421 pa = VM_PAGE_TO_PHYS(m); 2422 rcount = 1; 2423 } 2424 if (rcount) { 2425 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2426 (long)fidx, rcount, (long)pa); 2427 if (nl > 18) { 2428 c = cngetc(); 2429 if (c != ' ') 2430 return; 2431 nl = 0; 2432 } 2433 nl++; 2434 } 2435 } 2436} 2437#endif /* DDB */ 2438