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 * 3. 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 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $ 61 */ 62 63 /* 64 * Virtual memory object module. 65 */ 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/proc.h> /* for curproc, pageproc */ 70 #include <sys/thread.h> 71 #include <sys/vnode.h> 72 #include <sys/vmmeter.h> 73 #include <sys/mman.h> 74 #include <sys/mount.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/refcount.h> 78 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <vm/pmap.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_object.h> 84 #include <vm/vm_page.h> 85 #include <vm/vm_pageout.h> 86 #include <vm/vm_pager.h> 87 #include <vm/swap_pager.h> 88 #include <vm/vm_kern.h> 89 #include <vm/vm_extern.h> 90 #include <vm/vm_zone.h> 91 92 #define EASY_SCAN_FACTOR 8 93 94 static void vm_object_qcollapse(vm_object_t object, 95 vm_object_t backing_object); 96 static void vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 97 int pagerflags); 98 static void vm_object_lock_init(vm_object_t); 99 100 101 /* 102 * Virtual memory objects maintain the actual data 103 * associated with allocated virtual memory. A given 104 * page of memory exists within exactly one object. 105 * 106 * An object is only deallocated when all "references" 107 * are given up. Only one "reference" to a given 108 * region of an object should be writeable. 109 * 110 * Associated with each object is a list of all resident 111 * memory pages belonging to that object; this list is 112 * maintained by the "vm_page" module, and locked by the object's 113 * lock. 114 * 115 * Each object also records a "pager" routine which is 116 * used to retrieve (and store) pages to the proper backing 117 * storage. In addition, objects may be backed by other 118 * objects from which they were virtual-copied. 119 * 120 * The only items within the object structure which are 121 * modified after time of creation are: 122 * reference count locked by object's lock 123 * pager routine locked by object's lock 124 * 125 */ 126 127 struct object_q vm_object_list; /* locked by vmobj_token */ 128 struct vm_object kernel_object; 129 130 static long vm_object_count; /* locked by vmobj_token */ 131 extern int vm_pageout_page_count; 132 133 static long object_collapses; 134 static long object_bypasses; 135 static int next_index; 136 static vm_zone_t obj_zone; 137 static struct vm_zone obj_zone_store; 138 #define VM_OBJECTS_INIT 256 139 static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 140 141 /* 142 * Misc low level routines 143 */ 144 static void 145 vm_object_lock_init(vm_object_t obj) 146 { 147 #if defined(DEBUG_LOCKS) 148 int i; 149 150 obj->debug_hold_bitmap = 0; 151 obj->debug_hold_ovfl = 0; 152 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 153 obj->debug_hold_thrs[i] = NULL; 154 obj->debug_hold_file[i] = NULL; 155 obj->debug_hold_line[i] = 0; 156 } 157 #endif 158 } 159 160 void 161 vm_object_lock_swap(void) 162 { 163 lwkt_token_swap(); 164 } 165 166 void 167 vm_object_lock(vm_object_t obj) 168 { 169 lwkt_gettoken(&obj->token); 170 } 171 172 /* 173 * Returns TRUE on sucesss 174 */ 175 static int 176 vm_object_lock_try(vm_object_t obj) 177 { 178 return(lwkt_trytoken(&obj->token)); 179 } 180 181 void 182 vm_object_lock_shared(vm_object_t obj) 183 { 184 lwkt_gettoken_shared(&obj->token); 185 } 186 187 void 188 vm_object_unlock(vm_object_t obj) 189 { 190 lwkt_reltoken(&obj->token); 191 } 192 193 static __inline void 194 vm_object_assert_held(vm_object_t obj) 195 { 196 ASSERT_LWKT_TOKEN_HELD(&obj->token); 197 } 198 199 void 200 #ifndef DEBUG_LOCKS 201 vm_object_hold(vm_object_t obj) 202 #else 203 debugvm_object_hold(vm_object_t obj, char *file, int line) 204 #endif 205 { 206 KKASSERT(obj != NULL); 207 208 /* 209 * Object must be held (object allocation is stable due to callers 210 * context, typically already holding the token on a parent object) 211 * prior to potentially blocking on the lock, otherwise the object 212 * can get ripped away from us. 213 */ 214 refcount_acquire(&obj->hold_count); 215 vm_object_lock(obj); 216 217 #if defined(DEBUG_LOCKS) 218 int i; 219 u_int mask; 220 221 for (;;) { 222 mask = ~obj->debug_hold_bitmap; 223 cpu_ccfence(); 224 if (mask == 0xFFFFFFFFU) { 225 if (obj->debug_hold_ovfl == 0) 226 obj->debug_hold_ovfl = 1; 227 break; 228 } 229 i = ffs(mask) - 1; 230 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 231 ~mask | (1 << i))) { 232 obj->debug_hold_bitmap |= (1 << i); 233 obj->debug_hold_thrs[i] = curthread; 234 obj->debug_hold_file[i] = file; 235 obj->debug_hold_line[i] = line; 236 break; 237 } 238 } 239 #endif 240 } 241 242 int 243 #ifndef DEBUG_LOCKS 244 vm_object_hold_try(vm_object_t obj) 245 #else 246 debugvm_object_hold_try(vm_object_t obj, char *file, int line) 247 #endif 248 { 249 KKASSERT(obj != NULL); 250 251 /* 252 * Object must be held (object allocation is stable due to callers 253 * context, typically already holding the token on a parent object) 254 * prior to potentially blocking on the lock, otherwise the object 255 * can get ripped away from us. 256 */ 257 refcount_acquire(&obj->hold_count); 258 if (vm_object_lock_try(obj) == 0) { 259 if (refcount_release(&obj->hold_count)) { 260 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) 261 zfree(obj_zone, obj); 262 } 263 return(0); 264 } 265 266 #if defined(DEBUG_LOCKS) 267 int i; 268 u_int mask; 269 270 for (;;) { 271 mask = ~obj->debug_hold_bitmap; 272 cpu_ccfence(); 273 if (mask == 0xFFFFFFFFU) { 274 if (obj->debug_hold_ovfl == 0) 275 obj->debug_hold_ovfl = 1; 276 break; 277 } 278 i = ffs(mask) - 1; 279 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 280 ~mask | (1 << i))) { 281 obj->debug_hold_bitmap |= (1 << i); 282 obj->debug_hold_thrs[i] = curthread; 283 obj->debug_hold_file[i] = file; 284 obj->debug_hold_line[i] = line; 285 break; 286 } 287 } 288 #endif 289 return(1); 290 } 291 292 void 293 #ifndef DEBUG_LOCKS 294 vm_object_hold_shared(vm_object_t obj) 295 #else 296 debugvm_object_hold_shared(vm_object_t obj, char *file, int line) 297 #endif 298 { 299 KKASSERT(obj != NULL); 300 301 /* 302 * Object must be held (object allocation is stable due to callers 303 * context, typically already holding the token on a parent object) 304 * prior to potentially blocking on the lock, otherwise the object 305 * can get ripped away from us. 306 */ 307 refcount_acquire(&obj->hold_count); 308 vm_object_lock_shared(obj); 309 310 #if defined(DEBUG_LOCKS) 311 int i; 312 u_int mask; 313 314 for (;;) { 315 mask = ~obj->debug_hold_bitmap; 316 cpu_ccfence(); 317 if (mask == 0xFFFFFFFFU) { 318 if (obj->debug_hold_ovfl == 0) 319 obj->debug_hold_ovfl = 1; 320 break; 321 } 322 i = ffs(mask) - 1; 323 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 324 ~mask | (1 << i))) { 325 obj->debug_hold_bitmap |= (1 << i); 326 obj->debug_hold_thrs[i] = curthread; 327 obj->debug_hold_file[i] = file; 328 obj->debug_hold_line[i] = line; 329 break; 330 } 331 } 332 #endif 333 } 334 335 /* 336 * Obtain either a shared or exclusive lock on VM object 337 * based on whether this is a terminal vnode object or not. 338 */ 339 int 340 #ifndef DEBUG_LOCKS 341 vm_object_hold_maybe_shared(vm_object_t obj) 342 #else 343 debugvm_object_hold_maybe_shared(vm_object_t obj, char *file, int line) 344 #endif 345 { 346 if (vm_shared_fault && 347 obj->type == OBJT_VNODE && 348 obj->backing_object == NULL) { 349 vm_object_hold_shared(obj); 350 return(1); 351 } else { 352 vm_object_hold(obj); 353 return(0); 354 } 355 } 356 357 /* 358 * Drop the token and hold_count on the object. 359 */ 360 void 361 vm_object_drop(vm_object_t obj) 362 { 363 if (obj == NULL) 364 return; 365 366 #if defined(DEBUG_LOCKS) 367 int found = 0; 368 int i; 369 370 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 371 if ((obj->debug_hold_bitmap & (1 << i)) && 372 (obj->debug_hold_thrs[i] == curthread)) { 373 obj->debug_hold_bitmap &= ~(1 << i); 374 obj->debug_hold_thrs[i] = NULL; 375 obj->debug_hold_file[i] = NULL; 376 obj->debug_hold_line[i] = 0; 377 found = 1; 378 break; 379 } 380 } 381 382 if (found == 0 && obj->debug_hold_ovfl == 0) 383 panic("vm_object: attempt to drop hold on non-self-held obj"); 384 #endif 385 386 /* 387 * No new holders should be possible once we drop hold_count 1->0 as 388 * there is no longer any way to reference the object. 389 */ 390 KKASSERT(obj->hold_count > 0); 391 if (refcount_release(&obj->hold_count)) { 392 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) { 393 vm_object_unlock(obj); 394 zfree(obj_zone, obj); 395 } else { 396 vm_object_unlock(obj); 397 } 398 } else { 399 vm_object_unlock(obj); 400 } 401 } 402 403 /* 404 * Initialize a freshly allocated object, returning a held object. 405 * 406 * Used only by vm_object_allocate() and zinitna(). 407 * 408 * No requirements. 409 */ 410 void 411 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 412 { 413 int incr; 414 415 RB_INIT(&object->rb_memq); 416 LIST_INIT(&object->shadow_head); 417 lwkt_token_init(&object->token, "vmobj"); 418 419 object->type = type; 420 object->size = size; 421 object->ref_count = 1; 422 object->hold_count = 0; 423 object->flags = 0; 424 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 425 vm_object_set_flag(object, OBJ_ONEMAPPING); 426 object->paging_in_progress = 0; 427 object->resident_page_count = 0; 428 object->agg_pv_list_count = 0; 429 object->shadow_count = 0; 430 /* cpu localization twist */ 431 object->pg_color = (int)(intptr_t)curthread; 432 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 433 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 434 else 435 incr = size; 436 next_index = (next_index + incr) & PQ_L2_MASK; 437 object->handle = NULL; 438 object->backing_object = NULL; 439 object->backing_object_offset = (vm_ooffset_t)0; 440 441 object->generation++; 442 object->swblock_count = 0; 443 RB_INIT(&object->swblock_root); 444 vm_object_lock_init(object); 445 pmap_object_init(object); 446 447 vm_object_hold(object); 448 lwkt_gettoken(&vmobj_token); 449 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 450 vm_object_count++; 451 lwkt_reltoken(&vmobj_token); 452 } 453 454 /* 455 * Initialize the VM objects module. 456 * 457 * Called from the low level boot code only. 458 */ 459 void 460 vm_object_init(void) 461 { 462 TAILQ_INIT(&vm_object_list); 463 464 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 465 &kernel_object); 466 vm_object_drop(&kernel_object); 467 468 obj_zone = &obj_zone_store; 469 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 470 vm_objects_init, VM_OBJECTS_INIT); 471 } 472 473 void 474 vm_object_init2(void) 475 { 476 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1); 477 } 478 479 /* 480 * Allocate and return a new object of the specified type and size. 481 * 482 * No requirements. 483 */ 484 vm_object_t 485 vm_object_allocate(objtype_t type, vm_pindex_t size) 486 { 487 vm_object_t result; 488 489 result = (vm_object_t) zalloc(obj_zone); 490 491 _vm_object_allocate(type, size, result); 492 vm_object_drop(result); 493 494 return (result); 495 } 496 497 /* 498 * This version returns a held object, allowing further atomic initialization 499 * of the object. 500 */ 501 vm_object_t 502 vm_object_allocate_hold(objtype_t type, vm_pindex_t size) 503 { 504 vm_object_t result; 505 506 result = (vm_object_t) zalloc(obj_zone); 507 508 _vm_object_allocate(type, size, result); 509 510 return (result); 511 } 512 513 /* 514 * Add an additional reference to a vm_object. The object must already be 515 * held. The original non-lock version is no longer supported. The object 516 * must NOT be chain locked by anyone at the time the reference is added. 517 * 518 * Referencing a chain-locked object can blow up the fairly sensitive 519 * ref_count and shadow_count tests in the deallocator. Most callers 520 * will call vm_object_chain_wait() prior to calling 521 * vm_object_reference_locked() to avoid the case. 522 * 523 * The object must be held, but may be held shared if desired (hence why 524 * we use an atomic op). 525 */ 526 void 527 vm_object_reference_locked(vm_object_t object) 528 { 529 KKASSERT(object != NULL); 530 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 531 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0); 532 atomic_add_int(&object->ref_count, 1); 533 if (object->type == OBJT_VNODE) { 534 vref(object->handle); 535 /* XXX what if the vnode is being destroyed? */ 536 } 537 } 538 539 /* 540 * Object OBJ_CHAINLOCK lock handling. 541 * 542 * The caller can chain-lock backing objects recursively and then 543 * use vm_object_chain_release_all() to undo the whole chain. 544 * 545 * Chain locks are used to prevent collapses and are only applicable 546 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations 547 * on other object types are ignored. This is also important because 548 * it allows e.g. the vnode underlying a memory mapping to take concurrent 549 * faults. 550 * 551 * The object must usually be held on entry, though intermediate 552 * objects need not be held on release. 553 */ 554 void 555 vm_object_chain_wait(vm_object_t object) 556 { 557 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 558 while (object->flags & OBJ_CHAINLOCK) { 559 vm_object_set_flag(object, OBJ_CHAINWANT); 560 tsleep(object, 0, "objchain", 0); 561 } 562 } 563 564 void 565 vm_object_chain_acquire(vm_object_t object) 566 { 567 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) { 568 vm_object_chain_wait(object); 569 vm_object_set_flag(object, OBJ_CHAINLOCK); 570 } 571 } 572 573 void 574 vm_object_chain_release(vm_object_t object) 575 { 576 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 577 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) { 578 KKASSERT(object->flags & OBJ_CHAINLOCK); 579 if (object->flags & OBJ_CHAINWANT) { 580 vm_object_clear_flag(object, 581 OBJ_CHAINLOCK | OBJ_CHAINWANT); 582 wakeup(object); 583 } else { 584 vm_object_clear_flag(object, OBJ_CHAINLOCK); 585 } 586 } 587 } 588 589 /* 590 * This releases the entire chain of objects from first_object to and 591 * including stopobj, flowing through object->backing_object. 592 * 593 * We release stopobj first as an optimization as this object is most 594 * likely to be shared across multiple processes. 595 */ 596 void 597 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj) 598 { 599 vm_object_t backing_object; 600 vm_object_t object; 601 602 vm_object_chain_release(stopobj); 603 object = first_object; 604 605 while (object != stopobj) { 606 KKASSERT(object); 607 if (object != first_object) 608 vm_object_hold(object); 609 backing_object = object->backing_object; 610 vm_object_chain_release(object); 611 if (object != first_object) 612 vm_object_drop(object); 613 object = backing_object; 614 } 615 } 616 617 /* 618 * Dereference an object and its underlying vnode. 619 * 620 * The object must be held exclusively and will remain held on return. 621 * (We don't need an atomic op due to the exclusivity). 622 */ 623 static void 624 vm_object_vndeallocate(vm_object_t object) 625 { 626 struct vnode *vp = (struct vnode *) object->handle; 627 628 KASSERT(object->type == OBJT_VNODE, 629 ("vm_object_vndeallocate: not a vnode object")); 630 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 631 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 632 #ifdef INVARIANTS 633 if (object->ref_count == 0) { 634 vprint("vm_object_vndeallocate", vp); 635 panic("vm_object_vndeallocate: bad object reference count"); 636 } 637 #endif 638 object->ref_count--; 639 if (object->ref_count == 0) 640 vclrflags(vp, VTEXT); 641 vrele(vp); 642 } 643 644 /* 645 * Release a reference to the specified object, gained either through a 646 * vm_object_allocate or a vm_object_reference call. When all references 647 * are gone, storage associated with this object may be relinquished. 648 * 649 * The caller does not have to hold the object locked but must have control 650 * over the reference in question in order to guarantee that the object 651 * does not get ripped out from under us. 652 * 653 * XXX Currently all deallocations require an exclusive lock. 654 */ 655 void 656 vm_object_deallocate(vm_object_t object) 657 { 658 if (object) { 659 vm_object_hold(object); 660 vm_object_deallocate_locked(object); 661 vm_object_drop(object); 662 } 663 } 664 665 void 666 vm_object_deallocate_locked(vm_object_t object) 667 { 668 struct vm_object_dealloc_list *dlist = NULL; 669 struct vm_object_dealloc_list *dtmp; 670 vm_object_t temp; 671 int must_drop = 0; 672 673 /* 674 * We may chain deallocate object, but additional objects may 675 * collect on the dlist which also have to be deallocated. We 676 * must avoid a recursion, vm_object chains can get deep. 677 */ 678 again: 679 while (object != NULL) { 680 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token); 681 #if 0 682 /* 683 * Don't rip a ref_count out from under an object undergoing 684 * collapse, it will confuse the collapse code. 685 */ 686 vm_object_chain_wait(object); 687 #endif 688 if (object->type == OBJT_VNODE) { 689 vm_object_vndeallocate(object); 690 break; 691 } 692 693 if (object->ref_count == 0) { 694 panic("vm_object_deallocate: object deallocated " 695 "too many times: %d", object->type); 696 } 697 if (object->ref_count > 2) { 698 object->ref_count--; 699 break; 700 } 701 702 /* 703 * Here on ref_count of one or two, which are special cases for 704 * objects. 705 * 706 * Nominal ref_count > 1 case if the second ref is not from 707 * a shadow. 708 * 709 * (ONEMAPPING only applies to DEFAULT AND SWAP objects) 710 */ 711 if (object->ref_count == 2 && object->shadow_count == 0) { 712 if (object->type == OBJT_DEFAULT || 713 object->type == OBJT_SWAP) { 714 vm_object_set_flag(object, OBJ_ONEMAPPING); 715 } 716 object->ref_count--; 717 break; 718 } 719 720 /* 721 * If the second ref is from a shadow we chain along it 722 * upwards if object's handle is exhausted. 723 * 724 * We have to decrement object->ref_count before potentially 725 * collapsing the first shadow object or the collapse code 726 * will not be able to handle the degenerate case to remove 727 * object. However, if we do it too early the object can 728 * get ripped out from under us. 729 */ 730 if (object->ref_count == 2 && object->shadow_count == 1 && 731 object->handle == NULL && (object->type == OBJT_DEFAULT || 732 object->type == OBJT_SWAP)) { 733 temp = LIST_FIRST(&object->shadow_head); 734 KKASSERT(temp != NULL); 735 vm_object_hold(temp); 736 737 /* 738 * Wait for any paging to complete so the collapse 739 * doesn't (or isn't likely to) qcollapse. pip 740 * waiting must occur before we acquire the 741 * chainlock. 742 */ 743 while ( 744 temp->paging_in_progress || 745 object->paging_in_progress 746 ) { 747 vm_object_pip_wait(temp, "objde1"); 748 vm_object_pip_wait(object, "objde2"); 749 } 750 751 /* 752 * If the parent is locked we have to give up, as 753 * otherwise we would be acquiring locks in the 754 * wrong order and potentially deadlock. 755 */ 756 if (temp->flags & OBJ_CHAINLOCK) { 757 vm_object_drop(temp); 758 goto skip; 759 } 760 vm_object_chain_acquire(temp); 761 762 /* 763 * Recheck/retry after the hold and the paging 764 * wait, both of which can block us. 765 */ 766 if (object->ref_count != 2 || 767 object->shadow_count != 1 || 768 object->handle || 769 LIST_FIRST(&object->shadow_head) != temp || 770 (object->type != OBJT_DEFAULT && 771 object->type != OBJT_SWAP)) { 772 vm_object_chain_release(temp); 773 vm_object_drop(temp); 774 continue; 775 } 776 777 /* 778 * We can safely drop object's ref_count now. 779 */ 780 KKASSERT(object->ref_count == 2); 781 object->ref_count--; 782 783 /* 784 * If our single parent is not collapseable just 785 * decrement ref_count (2->1) and stop. 786 */ 787 if (temp->handle || (temp->type != OBJT_DEFAULT && 788 temp->type != OBJT_SWAP)) { 789 vm_object_chain_release(temp); 790 vm_object_drop(temp); 791 break; 792 } 793 794 /* 795 * At this point we have already dropped object's 796 * ref_count so it is possible for a race to 797 * deallocate obj out from under us. Any collapse 798 * will re-check the situation. We must not block 799 * until we are able to collapse. 800 * 801 * Bump temp's ref_count to avoid an unwanted 802 * degenerate recursion (can't call 803 * vm_object_reference_locked() because it asserts 804 * that CHAINLOCK is not set). 805 */ 806 temp->ref_count++; 807 KKASSERT(temp->ref_count > 1); 808 809 /* 810 * Collapse temp, then deallocate the extra ref 811 * formally. 812 */ 813 vm_object_collapse(temp, &dlist); 814 vm_object_chain_release(temp); 815 if (must_drop) { 816 vm_object_lock_swap(); 817 vm_object_drop(object); 818 } 819 object = temp; 820 must_drop = 1; 821 continue; 822 } 823 824 /* 825 * Drop the ref and handle termination on the 1->0 transition. 826 * We may have blocked above so we have to recheck. 827 */ 828 skip: 829 KKASSERT(object->ref_count != 0); 830 if (object->ref_count >= 2) { 831 object->ref_count--; 832 break; 833 } 834 KKASSERT(object->ref_count == 1); 835 836 /* 837 * 1->0 transition. Chain through the backing_object. 838 * Maintain the ref until we've located the backing object, 839 * then re-check. 840 */ 841 while ((temp = object->backing_object) != NULL) { 842 vm_object_hold(temp); 843 if (temp == object->backing_object) 844 break; 845 vm_object_drop(temp); 846 } 847 848 /* 849 * 1->0 transition verified, retry if ref_count is no longer 850 * 1. Otherwise disconnect the backing_object (temp) and 851 * clean up. 852 */ 853 if (object->ref_count != 1) { 854 vm_object_drop(temp); 855 continue; 856 } 857 858 /* 859 * It shouldn't be possible for the object to be chain locked 860 * if we're removing the last ref on it. 861 */ 862 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0); 863 864 if (temp) { 865 LIST_REMOVE(object, shadow_list); 866 temp->shadow_count--; 867 temp->generation++; 868 object->backing_object = NULL; 869 } 870 871 --object->ref_count; 872 if ((object->flags & OBJ_DEAD) == 0) 873 vm_object_terminate(object); 874 if (must_drop && temp) 875 vm_object_lock_swap(); 876 if (must_drop) 877 vm_object_drop(object); 878 object = temp; 879 must_drop = 1; 880 } 881 if (must_drop && object) 882 vm_object_drop(object); 883 884 /* 885 * Additional tail recursion on dlist. Avoid a recursion. Objects 886 * on the dlist have a hold count but are not locked. 887 */ 888 if ((dtmp = dlist) != NULL) { 889 dlist = dtmp->next; 890 object = dtmp->object; 891 kfree(dtmp, M_TEMP); 892 893 vm_object_lock(object); /* already held, add lock */ 894 must_drop = 1; /* and we're responsible for it */ 895 goto again; 896 } 897 } 898 899 /* 900 * Destroy the specified object, freeing up related resources. 901 * 902 * The object must have zero references. 903 * 904 * The object must held. The caller is responsible for dropping the object 905 * after terminate returns. Terminate does NOT drop the object. 906 */ 907 static int vm_object_terminate_callback(vm_page_t p, void *data); 908 909 void 910 vm_object_terminate(vm_object_t object) 911 { 912 /* 913 * Make sure no one uses us. Once we set OBJ_DEAD we should be 914 * able to safely block. 915 */ 916 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 917 KKASSERT((object->flags & OBJ_DEAD) == 0); 918 vm_object_set_flag(object, OBJ_DEAD); 919 920 /* 921 * Wait for the pageout daemon to be done with the object 922 */ 923 vm_object_pip_wait(object, "objtrm1"); 924 925 KASSERT(!object->paging_in_progress, 926 ("vm_object_terminate: pageout in progress")); 927 928 /* 929 * Clean and free the pages, as appropriate. All references to the 930 * object are gone, so we don't need to lock it. 931 */ 932 if (object->type == OBJT_VNODE) { 933 struct vnode *vp; 934 935 /* 936 * Clean pages and flush buffers. 937 * 938 * NOTE! TMPFS buffer flushes do not typically flush the 939 * actual page to swap as this would be highly 940 * inefficient, and normal filesystems usually wrap 941 * page flushes with buffer cache buffers. 942 * 943 * To deal with this we have to call vinvalbuf() both 944 * before and after the vm_object_page_clean(). 945 */ 946 vp = (struct vnode *) object->handle; 947 vinvalbuf(vp, V_SAVE, 0, 0); 948 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 949 vinvalbuf(vp, V_SAVE, 0, 0); 950 } 951 952 /* 953 * Wait for any I/O to complete, after which there had better not 954 * be any references left on the object. 955 */ 956 vm_object_pip_wait(object, "objtrm2"); 957 958 if (object->ref_count != 0) { 959 panic("vm_object_terminate: object with references, " 960 "ref_count=%d", object->ref_count); 961 } 962 963 /* 964 * Cleanup any shared pmaps associated with this object. 965 */ 966 pmap_object_free(object); 967 968 /* 969 * Now free any remaining pages. For internal objects, this also 970 * removes them from paging queues. Don't free wired pages, just 971 * remove them from the object. 972 */ 973 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 974 vm_object_terminate_callback, NULL); 975 976 /* 977 * Let the pager know object is dead. 978 */ 979 vm_pager_deallocate(object); 980 981 /* 982 * Wait for the object hold count to hit 1, clean out pages as 983 * we go. vmobj_token interlocks any race conditions that might 984 * pick the object up from the vm_object_list after we have cleared 985 * rb_memq. 986 */ 987 for (;;) { 988 if (RB_ROOT(&object->rb_memq) == NULL) 989 break; 990 kprintf("vm_object_terminate: Warning, object %p " 991 "still has %d pages\n", 992 object, object->resident_page_count); 993 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 994 vm_object_terminate_callback, NULL); 995 } 996 997 /* 998 * There had better not be any pages left 999 */ 1000 KKASSERT(object->resident_page_count == 0); 1001 1002 /* 1003 * Remove the object from the global object list. 1004 */ 1005 lwkt_gettoken(&vmobj_token); 1006 TAILQ_REMOVE(&vm_object_list, object, object_list); 1007 vm_object_count--; 1008 lwkt_reltoken(&vmobj_token); 1009 vm_object_dead_wakeup(object); 1010 1011 if (object->ref_count != 0) { 1012 panic("vm_object_terminate2: object with references, " 1013 "ref_count=%d", object->ref_count); 1014 } 1015 1016 /* 1017 * NOTE: The object hold_count is at least 1, so we cannot zfree() 1018 * the object here. See vm_object_drop(). 1019 */ 1020 } 1021 1022 /* 1023 * The caller must hold the object. 1024 */ 1025 static int 1026 vm_object_terminate_callback(vm_page_t p, void *data __unused) 1027 { 1028 vm_object_t object; 1029 1030 object = p->object; 1031 vm_page_busy_wait(p, TRUE, "vmpgtrm"); 1032 if (object != p->object) { 1033 kprintf("vm_object_terminate: Warning: Encountered " 1034 "busied page %p on queue %d\n", p, p->queue); 1035 vm_page_wakeup(p); 1036 } else if (p->wire_count == 0) { 1037 /* 1038 * NOTE: p->dirty and PG_NEED_COMMIT are ignored. 1039 */ 1040 vm_page_free(p); 1041 mycpu->gd_cnt.v_pfree++; 1042 } else { 1043 if (p->queue != PQ_NONE) 1044 kprintf("vm_object_terminate: Warning: Encountered " 1045 "wired page %p on queue %d\n", p, p->queue); 1046 vm_page_remove(p); 1047 vm_page_wakeup(p); 1048 } 1049 lwkt_yield(); 1050 return(0); 1051 } 1052 1053 /* 1054 * The object is dead but still has an object<->pager association. Sleep 1055 * and return. The caller typically retests the association in a loop. 1056 * 1057 * The caller must hold the object. 1058 */ 1059 void 1060 vm_object_dead_sleep(vm_object_t object, const char *wmesg) 1061 { 1062 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1063 if (object->handle) { 1064 vm_object_set_flag(object, OBJ_DEADWNT); 1065 tsleep(object, 0, wmesg, 0); 1066 /* object may be invalid after this point */ 1067 } 1068 } 1069 1070 /* 1071 * Wakeup anyone waiting for the object<->pager disassociation on 1072 * a dead object. 1073 * 1074 * The caller must hold the object. 1075 */ 1076 void 1077 vm_object_dead_wakeup(vm_object_t object) 1078 { 1079 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1080 if (object->flags & OBJ_DEADWNT) { 1081 vm_object_clear_flag(object, OBJ_DEADWNT); 1082 wakeup(object); 1083 } 1084 } 1085 1086 /* 1087 * Clean all dirty pages in the specified range of object. Leaves page 1088 * on whatever queue it is currently on. If NOSYNC is set then do not 1089 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 1090 * leaving the object dirty. 1091 * 1092 * When stuffing pages asynchronously, allow clustering. XXX we need a 1093 * synchronous clustering mode implementation. 1094 * 1095 * Odd semantics: if start == end, we clean everything. 1096 * 1097 * The object must be locked? XXX 1098 */ 1099 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 1100 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 1101 1102 void 1103 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1104 int flags) 1105 { 1106 struct rb_vm_page_scan_info info; 1107 struct vnode *vp; 1108 int wholescan; 1109 int pagerflags; 1110 int generation; 1111 1112 vm_object_hold(object); 1113 if (object->type != OBJT_VNODE || 1114 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 1115 vm_object_drop(object); 1116 return; 1117 } 1118 1119 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 1120 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1121 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 1122 1123 vp = object->handle; 1124 1125 /* 1126 * Interlock other major object operations. This allows us to 1127 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 1128 */ 1129 vm_object_set_flag(object, OBJ_CLEANING); 1130 1131 /* 1132 * Handle 'entire object' case 1133 */ 1134 info.start_pindex = start; 1135 if (end == 0) { 1136 info.end_pindex = object->size - 1; 1137 } else { 1138 info.end_pindex = end - 1; 1139 } 1140 wholescan = (start == 0 && info.end_pindex == object->size - 1); 1141 info.limit = flags; 1142 info.pagerflags = pagerflags; 1143 info.object = object; 1144 1145 /* 1146 * If cleaning the entire object do a pass to mark the pages read-only. 1147 * If everything worked out ok, clear OBJ_WRITEABLE and 1148 * OBJ_MIGHTBEDIRTY. 1149 */ 1150 if (wholescan) { 1151 info.error = 0; 1152 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1153 vm_object_page_clean_pass1, &info); 1154 if (info.error == 0) { 1155 vm_object_clear_flag(object, 1156 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1157 if (object->type == OBJT_VNODE && 1158 (vp = (struct vnode *)object->handle) != NULL) { 1159 if (vp->v_flag & VOBJDIRTY) 1160 vclrflags(vp, VOBJDIRTY); 1161 } 1162 } 1163 } 1164 1165 /* 1166 * Do a pass to clean all the dirty pages we find. 1167 */ 1168 do { 1169 info.error = 0; 1170 generation = object->generation; 1171 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1172 vm_object_page_clean_pass2, &info); 1173 } while (info.error || generation != object->generation); 1174 1175 vm_object_clear_flag(object, OBJ_CLEANING); 1176 vm_object_drop(object); 1177 } 1178 1179 /* 1180 * The caller must hold the object. 1181 */ 1182 static 1183 int 1184 vm_object_page_clean_pass1(struct vm_page *p, void *data) 1185 { 1186 struct rb_vm_page_scan_info *info = data; 1187 1188 vm_page_flag_set(p, PG_CLEANCHK); 1189 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1190 info->error = 1; 1191 } else if (vm_page_busy_try(p, FALSE) == 0) { 1192 vm_page_protect(p, VM_PROT_READ); /* must not block */ 1193 vm_page_wakeup(p); 1194 } else { 1195 info->error = 1; 1196 } 1197 lwkt_yield(); 1198 return(0); 1199 } 1200 1201 /* 1202 * The caller must hold the object 1203 */ 1204 static 1205 int 1206 vm_object_page_clean_pass2(struct vm_page *p, void *data) 1207 { 1208 struct rb_vm_page_scan_info *info = data; 1209 int generation; 1210 1211 /* 1212 * Do not mess with pages that were inserted after we started 1213 * the cleaning pass. 1214 */ 1215 if ((p->flags & PG_CLEANCHK) == 0) 1216 goto done; 1217 1218 generation = info->object->generation; 1219 vm_page_busy_wait(p, TRUE, "vpcwai"); 1220 if (p->object != info->object || 1221 info->object->generation != generation) { 1222 info->error = 1; 1223 vm_page_wakeup(p); 1224 goto done; 1225 } 1226 1227 /* 1228 * Before wasting time traversing the pmaps, check for trivial 1229 * cases where the page cannot be dirty. 1230 */ 1231 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 1232 KKASSERT((p->dirty & p->valid) == 0 && 1233 (p->flags & PG_NEED_COMMIT) == 0); 1234 vm_page_wakeup(p); 1235 goto done; 1236 } 1237 1238 /* 1239 * Check whether the page is dirty or not. The page has been set 1240 * to be read-only so the check will not race a user dirtying the 1241 * page. 1242 */ 1243 vm_page_test_dirty(p); 1244 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) { 1245 vm_page_flag_clear(p, PG_CLEANCHK); 1246 vm_page_wakeup(p); 1247 goto done; 1248 } 1249 1250 /* 1251 * If we have been asked to skip nosync pages and this is a 1252 * nosync page, skip it. Note that the object flags were 1253 * not cleared in this case (because pass1 will have returned an 1254 * error), so we do not have to set them. 1255 */ 1256 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1257 vm_page_flag_clear(p, PG_CLEANCHK); 1258 vm_page_wakeup(p); 1259 goto done; 1260 } 1261 1262 /* 1263 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 1264 * the pages that get successfully flushed. Set info->error if 1265 * we raced an object modification. 1266 */ 1267 vm_object_page_collect_flush(info->object, p, info->pagerflags); 1268 vm_wait_nominal(); 1269 done: 1270 lwkt_yield(); 1271 return(0); 1272 } 1273 1274 /* 1275 * Collect the specified page and nearby pages and flush them out. 1276 * The number of pages flushed is returned. The passed page is busied 1277 * by the caller and we are responsible for its disposition. 1278 * 1279 * The caller must hold the object. 1280 */ 1281 static void 1282 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 1283 { 1284 int runlen; 1285 int error; 1286 int maxf; 1287 int chkb; 1288 int maxb; 1289 int i; 1290 vm_pindex_t pi; 1291 vm_page_t maf[vm_pageout_page_count]; 1292 vm_page_t mab[vm_pageout_page_count]; 1293 vm_page_t ma[vm_pageout_page_count]; 1294 1295 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1296 1297 pi = p->pindex; 1298 1299 maxf = 0; 1300 for(i = 1; i < vm_pageout_page_count; i++) { 1301 vm_page_t tp; 1302 1303 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error); 1304 if (error) 1305 break; 1306 if (tp == NULL) 1307 break; 1308 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1309 (tp->flags & PG_CLEANCHK) == 0) { 1310 vm_page_wakeup(tp); 1311 break; 1312 } 1313 if ((tp->queue - tp->pc) == PQ_CACHE) { 1314 vm_page_flag_clear(tp, PG_CLEANCHK); 1315 vm_page_wakeup(tp); 1316 break; 1317 } 1318 vm_page_test_dirty(tp); 1319 if ((tp->dirty & tp->valid) == 0 && 1320 (tp->flags & PG_NEED_COMMIT) == 0) { 1321 vm_page_flag_clear(tp, PG_CLEANCHK); 1322 vm_page_wakeup(tp); 1323 break; 1324 } 1325 maf[i - 1] = tp; 1326 maxf++; 1327 } 1328 1329 maxb = 0; 1330 chkb = vm_pageout_page_count - maxf; 1331 /* 1332 * NOTE: chkb can be 0 1333 */ 1334 for(i = 1; chkb && i < chkb; i++) { 1335 vm_page_t tp; 1336 1337 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error); 1338 if (error) 1339 break; 1340 if (tp == NULL) 1341 break; 1342 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1343 (tp->flags & PG_CLEANCHK) == 0) { 1344 vm_page_wakeup(tp); 1345 break; 1346 } 1347 if ((tp->queue - tp->pc) == PQ_CACHE) { 1348 vm_page_flag_clear(tp, PG_CLEANCHK); 1349 vm_page_wakeup(tp); 1350 break; 1351 } 1352 vm_page_test_dirty(tp); 1353 if ((tp->dirty & tp->valid) == 0 && 1354 (tp->flags & PG_NEED_COMMIT) == 0) { 1355 vm_page_flag_clear(tp, PG_CLEANCHK); 1356 vm_page_wakeup(tp); 1357 break; 1358 } 1359 mab[i - 1] = tp; 1360 maxb++; 1361 } 1362 1363 /* 1364 * All pages in the maf[] and mab[] array are busied. 1365 */ 1366 for (i = 0; i < maxb; i++) { 1367 int index = (maxb - i) - 1; 1368 ma[index] = mab[i]; 1369 vm_page_flag_clear(ma[index], PG_CLEANCHK); 1370 } 1371 vm_page_flag_clear(p, PG_CLEANCHK); 1372 ma[maxb] = p; 1373 for(i = 0; i < maxf; i++) { 1374 int index = (maxb + i) + 1; 1375 ma[index] = maf[i]; 1376 vm_page_flag_clear(ma[index], PG_CLEANCHK); 1377 } 1378 runlen = maxb + maxf + 1; 1379 1380 for (i = 0; i < runlen; i++) /* XXX need this any more? */ 1381 vm_page_hold(ma[i]); 1382 1383 vm_pageout_flush(ma, runlen, pagerflags); 1384 1385 for (i = 0; i < runlen; i++) /* XXX need this any more? */ 1386 vm_page_unhold(ma[i]); 1387 } 1388 1389 /* 1390 * Same as vm_object_pmap_copy, except range checking really 1391 * works, and is meant for small sections of an object. 1392 * 1393 * This code protects resident pages by making them read-only 1394 * and is typically called on a fork or split when a page 1395 * is converted to copy-on-write. 1396 * 1397 * NOTE: If the page is already at VM_PROT_NONE, calling 1398 * vm_page_protect will have no effect. 1399 */ 1400 void 1401 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1402 { 1403 vm_pindex_t idx; 1404 vm_page_t p; 1405 1406 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 1407 return; 1408 1409 vm_object_hold(object); 1410 for (idx = start; idx < end; idx++) { 1411 p = vm_page_lookup(object, idx); 1412 if (p == NULL) 1413 continue; 1414 vm_page_protect(p, VM_PROT_READ); 1415 } 1416 vm_object_drop(object); 1417 } 1418 1419 /* 1420 * Removes all physical pages in the specified object range from all 1421 * physical maps. 1422 * 1423 * The object must *not* be locked. 1424 */ 1425 1426 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 1427 1428 void 1429 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1430 { 1431 struct rb_vm_page_scan_info info; 1432 1433 if (object == NULL) 1434 return; 1435 info.start_pindex = start; 1436 info.end_pindex = end - 1; 1437 1438 vm_object_hold(object); 1439 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1440 vm_object_pmap_remove_callback, &info); 1441 if (start == 0 && end == object->size) 1442 vm_object_clear_flag(object, OBJ_WRITEABLE); 1443 vm_object_drop(object); 1444 } 1445 1446 /* 1447 * The caller must hold the object 1448 */ 1449 static int 1450 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 1451 { 1452 vm_page_protect(p, VM_PROT_NONE); 1453 return(0); 1454 } 1455 1456 /* 1457 * Implements the madvise function at the object/page level. 1458 * 1459 * MADV_WILLNEED (any object) 1460 * 1461 * Activate the specified pages if they are resident. 1462 * 1463 * MADV_DONTNEED (any object) 1464 * 1465 * Deactivate the specified pages if they are resident. 1466 * 1467 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 1468 * 1469 * Deactivate and clean the specified pages if they are 1470 * resident. This permits the process to reuse the pages 1471 * without faulting or the kernel to reclaim the pages 1472 * without I/O. 1473 * 1474 * No requirements. 1475 */ 1476 void 1477 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1478 { 1479 vm_pindex_t end, tpindex; 1480 vm_object_t tobject; 1481 vm_object_t xobj; 1482 vm_page_t m; 1483 int error; 1484 1485 if (object == NULL) 1486 return; 1487 1488 end = pindex + count; 1489 1490 vm_object_hold(object); 1491 tobject = object; 1492 1493 /* 1494 * Locate and adjust resident pages 1495 */ 1496 for (; pindex < end; pindex += 1) { 1497 relookup: 1498 if (tobject != object) 1499 vm_object_drop(tobject); 1500 tobject = object; 1501 tpindex = pindex; 1502 shadowlookup: 1503 /* 1504 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1505 * and those pages must be OBJ_ONEMAPPING. 1506 */ 1507 if (advise == MADV_FREE) { 1508 if ((tobject->type != OBJT_DEFAULT && 1509 tobject->type != OBJT_SWAP) || 1510 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1511 continue; 1512 } 1513 } 1514 1515 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error); 1516 1517 if (error) { 1518 vm_page_sleep_busy(m, TRUE, "madvpo"); 1519 goto relookup; 1520 } 1521 if (m == NULL) { 1522 /* 1523 * There may be swap even if there is no backing page 1524 */ 1525 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1526 swap_pager_freespace(tobject, tpindex, 1); 1527 1528 /* 1529 * next object 1530 */ 1531 while ((xobj = tobject->backing_object) != NULL) { 1532 KKASSERT(xobj != object); 1533 vm_object_hold(xobj); 1534 if (xobj == tobject->backing_object) 1535 break; 1536 vm_object_drop(xobj); 1537 } 1538 if (xobj == NULL) 1539 continue; 1540 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1541 if (tobject != object) { 1542 vm_object_lock_swap(); 1543 vm_object_drop(tobject); 1544 } 1545 tobject = xobj; 1546 goto shadowlookup; 1547 } 1548 1549 /* 1550 * If the page is not in a normal active state, we skip it. 1551 * If the page is not managed there are no page queues to 1552 * mess with. Things can break if we mess with pages in 1553 * any of the below states. 1554 */ 1555 if (m->wire_count || 1556 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 1557 m->valid != VM_PAGE_BITS_ALL 1558 ) { 1559 vm_page_wakeup(m); 1560 continue; 1561 } 1562 1563 /* 1564 * Theoretically once a page is known not to be busy, an 1565 * interrupt cannot come along and rip it out from under us. 1566 */ 1567 1568 if (advise == MADV_WILLNEED) { 1569 vm_page_activate(m); 1570 } else if (advise == MADV_DONTNEED) { 1571 vm_page_dontneed(m); 1572 } else if (advise == MADV_FREE) { 1573 /* 1574 * Mark the page clean. This will allow the page 1575 * to be freed up by the system. However, such pages 1576 * are often reused quickly by malloc()/free() 1577 * so we do not do anything that would cause 1578 * a page fault if we can help it. 1579 * 1580 * Specifically, we do not try to actually free 1581 * the page now nor do we try to put it in the 1582 * cache (which would cause a page fault on reuse). 1583 * 1584 * But we do make the page is freeable as we 1585 * can without actually taking the step of unmapping 1586 * it. 1587 */ 1588 pmap_clear_modify(m); 1589 m->dirty = 0; 1590 m->act_count = 0; 1591 vm_page_dontneed(m); 1592 if (tobject->type == OBJT_SWAP) 1593 swap_pager_freespace(tobject, tpindex, 1); 1594 } 1595 vm_page_wakeup(m); 1596 } 1597 if (tobject != object) 1598 vm_object_drop(tobject); 1599 vm_object_drop(object); 1600 } 1601 1602 /* 1603 * Create a new object which is backed by the specified existing object 1604 * range. Replace the pointer and offset that was pointing at the existing 1605 * object with the pointer/offset for the new object. 1606 * 1607 * No other requirements. 1608 */ 1609 void 1610 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length, 1611 int addref) 1612 { 1613 vm_object_t source; 1614 vm_object_t result; 1615 1616 source = *objectp; 1617 1618 /* 1619 * Don't create the new object if the old object isn't shared. 1620 * We have to chain wait before adding the reference to avoid 1621 * racing a collapse or deallocation. 1622 * 1623 * Add the additional ref to source here to avoid racing a later 1624 * collapse or deallocation. Clear the ONEMAPPING flag whether 1625 * addref is TRUE or not in this case because the original object 1626 * will be shadowed. 1627 */ 1628 if (source) { 1629 vm_object_hold(source); 1630 vm_object_chain_wait(source); 1631 if (source->ref_count == 1 && 1632 source->handle == NULL && 1633 (source->type == OBJT_DEFAULT || 1634 source->type == OBJT_SWAP)) { 1635 vm_object_drop(source); 1636 if (addref) { 1637 vm_object_reference_locked(source); 1638 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1639 } 1640 return; 1641 } 1642 vm_object_reference_locked(source); 1643 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1644 } 1645 1646 /* 1647 * Allocate a new object with the given length. The new object 1648 * is returned referenced but we may have to add another one. 1649 * If we are adding a second reference we must clear OBJ_ONEMAPPING. 1650 * (typically because the caller is about to clone a vm_map_entry). 1651 * 1652 * The source object currently has an extra reference to prevent 1653 * collapses into it while we mess with its shadow list, which 1654 * we will remove later in this routine. 1655 */ 1656 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL) 1657 panic("vm_object_shadow: no object for shadowing"); 1658 vm_object_hold(result); 1659 if (addref) { 1660 vm_object_reference_locked(result); 1661 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1662 } 1663 1664 /* 1665 * The new object shadows the source object. Chain wait before 1666 * adjusting shadow_count or the shadow list to avoid races. 1667 * 1668 * Try to optimize the result object's page color when shadowing 1669 * in order to maintain page coloring consistency in the combined 1670 * shadowed object. 1671 */ 1672 KKASSERT(result->backing_object == NULL); 1673 result->backing_object = source; 1674 if (source) { 1675 vm_object_chain_wait(source); 1676 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1677 source->shadow_count++; 1678 source->generation++; 1679 /* cpu localization twist */ 1680 result->pg_color = (int)(intptr_t)curthread; 1681 } 1682 1683 /* 1684 * Adjust the return storage. Drop the ref on source before 1685 * returning. 1686 */ 1687 result->backing_object_offset = *offset; 1688 vm_object_drop(result); 1689 *offset = 0; 1690 if (source) { 1691 vm_object_deallocate_locked(source); 1692 vm_object_drop(source); 1693 } 1694 1695 /* 1696 * Return the new things 1697 */ 1698 *objectp = result; 1699 } 1700 1701 #define OBSC_TEST_ALL_SHADOWED 0x0001 1702 #define OBSC_COLLAPSE_NOWAIT 0x0002 1703 #define OBSC_COLLAPSE_WAIT 0x0004 1704 1705 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1706 1707 /* 1708 * The caller must hold the object. 1709 */ 1710 static __inline int 1711 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op) 1712 { 1713 struct rb_vm_page_scan_info info; 1714 1715 vm_object_assert_held(object); 1716 vm_object_assert_held(backing_object); 1717 1718 KKASSERT(backing_object == object->backing_object); 1719 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1720 1721 /* 1722 * Initial conditions 1723 */ 1724 if (op & OBSC_TEST_ALL_SHADOWED) { 1725 /* 1726 * We do not want to have to test for the existence of 1727 * swap pages in the backing object. XXX but with the 1728 * new swapper this would be pretty easy to do. 1729 * 1730 * XXX what about anonymous MAP_SHARED memory that hasn't 1731 * been ZFOD faulted yet? If we do not test for this, the 1732 * shadow test may succeed! XXX 1733 */ 1734 if (backing_object->type != OBJT_DEFAULT) 1735 return(0); 1736 } 1737 if (op & OBSC_COLLAPSE_WAIT) { 1738 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1739 vm_object_set_flag(backing_object, OBJ_DEAD); 1740 lwkt_gettoken(&vmobj_token); 1741 TAILQ_REMOVE(&vm_object_list, backing_object, object_list); 1742 vm_object_count--; 1743 lwkt_reltoken(&vmobj_token); 1744 vm_object_dead_wakeup(backing_object); 1745 } 1746 1747 /* 1748 * Our scan. We have to retry if a negative error code is returned, 1749 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1750 * the scan had to be stopped because the parent does not completely 1751 * shadow the child. 1752 */ 1753 info.object = object; 1754 info.backing_object = backing_object; 1755 info.limit = op; 1756 do { 1757 info.error = 1; 1758 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1759 vm_object_backing_scan_callback, 1760 &info); 1761 } while (info.error < 0); 1762 1763 return(info.error); 1764 } 1765 1766 /* 1767 * The caller must hold the object. 1768 */ 1769 static int 1770 vm_object_backing_scan_callback(vm_page_t p, void *data) 1771 { 1772 struct rb_vm_page_scan_info *info = data; 1773 vm_object_t backing_object; 1774 vm_object_t object; 1775 vm_pindex_t pindex; 1776 vm_pindex_t new_pindex; 1777 vm_pindex_t backing_offset_index; 1778 int op; 1779 1780 pindex = p->pindex; 1781 new_pindex = pindex - info->backing_offset_index; 1782 op = info->limit; 1783 object = info->object; 1784 backing_object = info->backing_object; 1785 backing_offset_index = info->backing_offset_index; 1786 1787 if (op & OBSC_TEST_ALL_SHADOWED) { 1788 vm_page_t pp; 1789 1790 /* 1791 * Ignore pages outside the parent object's range 1792 * and outside the parent object's mapping of the 1793 * backing object. 1794 * 1795 * note that we do not busy the backing object's 1796 * page. 1797 */ 1798 if (pindex < backing_offset_index || 1799 new_pindex >= object->size 1800 ) { 1801 return(0); 1802 } 1803 1804 /* 1805 * See if the parent has the page or if the parent's 1806 * object pager has the page. If the parent has the 1807 * page but the page is not valid, the parent's 1808 * object pager must have the page. 1809 * 1810 * If this fails, the parent does not completely shadow 1811 * the object and we might as well give up now. 1812 */ 1813 pp = vm_page_lookup(object, new_pindex); 1814 if ((pp == NULL || pp->valid == 0) && 1815 !vm_pager_has_page(object, new_pindex) 1816 ) { 1817 info->error = 0; /* problemo */ 1818 return(-1); /* stop the scan */ 1819 } 1820 } 1821 1822 /* 1823 * Check for busy page. Note that we may have lost (p) when we 1824 * possibly blocked above. 1825 */ 1826 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1827 vm_page_t pp; 1828 1829 if (vm_page_busy_try(p, TRUE)) { 1830 if (op & OBSC_COLLAPSE_NOWAIT) { 1831 return(0); 1832 } else { 1833 /* 1834 * If we slept, anything could have 1835 * happened. Ask that the scan be restarted. 1836 * 1837 * Since the object is marked dead, the 1838 * backing offset should not have changed. 1839 */ 1840 vm_page_sleep_busy(p, TRUE, "vmocol"); 1841 info->error = -1; 1842 return(-1); 1843 } 1844 } 1845 1846 /* 1847 * If (p) is no longer valid restart the scan. 1848 */ 1849 if (p->object != backing_object || p->pindex != pindex) { 1850 kprintf("vm_object_backing_scan: Warning: page " 1851 "%p ripped out from under us\n", p); 1852 vm_page_wakeup(p); 1853 info->error = -1; 1854 return(-1); 1855 } 1856 1857 if (op & OBSC_COLLAPSE_NOWAIT) { 1858 if (p->valid == 0 || 1859 p->wire_count || 1860 (p->flags & PG_NEED_COMMIT)) { 1861 vm_page_wakeup(p); 1862 return(0); 1863 } 1864 } else { 1865 /* XXX what if p->valid == 0 , hold_count, etc? */ 1866 } 1867 1868 KASSERT( 1869 p->object == backing_object, 1870 ("vm_object_qcollapse(): object mismatch") 1871 ); 1872 1873 /* 1874 * Destroy any associated swap 1875 */ 1876 if (backing_object->type == OBJT_SWAP) 1877 swap_pager_freespace(backing_object, p->pindex, 1); 1878 1879 if ( 1880 p->pindex < backing_offset_index || 1881 new_pindex >= object->size 1882 ) { 1883 /* 1884 * Page is out of the parent object's range, we 1885 * can simply destroy it. 1886 */ 1887 vm_page_protect(p, VM_PROT_NONE); 1888 vm_page_free(p); 1889 return(0); 1890 } 1891 1892 pp = vm_page_lookup(object, new_pindex); 1893 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 1894 /* 1895 * page already exists in parent OR swap exists 1896 * for this location in the parent. Destroy 1897 * the original page from the backing object. 1898 * 1899 * Leave the parent's page alone 1900 */ 1901 vm_page_protect(p, VM_PROT_NONE); 1902 vm_page_free(p); 1903 return(0); 1904 } 1905 1906 /* 1907 * Page does not exist in parent, rename the 1908 * page from the backing object to the main object. 1909 * 1910 * If the page was mapped to a process, it can remain 1911 * mapped through the rename. 1912 */ 1913 if ((p->queue - p->pc) == PQ_CACHE) 1914 vm_page_deactivate(p); 1915 1916 vm_page_rename(p, object, new_pindex); 1917 vm_page_wakeup(p); 1918 /* page automatically made dirty by rename */ 1919 } 1920 return(0); 1921 } 1922 1923 /* 1924 * This version of collapse allows the operation to occur earlier and 1925 * when paging_in_progress is true for an object... This is not a complete 1926 * operation, but should plug 99.9% of the rest of the leaks. 1927 * 1928 * The caller must hold the object and backing_object and both must be 1929 * chainlocked. 1930 * 1931 * (only called from vm_object_collapse) 1932 */ 1933 static void 1934 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object) 1935 { 1936 if (backing_object->ref_count == 1) { 1937 backing_object->ref_count += 2; 1938 vm_object_backing_scan(object, backing_object, 1939 OBSC_COLLAPSE_NOWAIT); 1940 backing_object->ref_count -= 2; 1941 } 1942 } 1943 1944 /* 1945 * Collapse an object with the object backing it. Pages in the backing 1946 * object are moved into the parent, and the backing object is deallocated. 1947 * Any conflict is resolved in favor of the parent's existing pages. 1948 * 1949 * object must be held and chain-locked on call. 1950 * 1951 * The caller must have an extra ref on object to prevent a race from 1952 * destroying it during the collapse. 1953 */ 1954 void 1955 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp) 1956 { 1957 struct vm_object_dealloc_list *dlist = NULL; 1958 vm_object_t backing_object; 1959 1960 /* 1961 * Only one thread is attempting a collapse at any given moment. 1962 * There are few restrictions for (object) that callers of this 1963 * function check so reentrancy is likely. 1964 */ 1965 KKASSERT(object != NULL); 1966 vm_object_assert_held(object); 1967 KKASSERT(object->flags & OBJ_CHAINLOCK); 1968 1969 for (;;) { 1970 vm_object_t bbobj; 1971 int dodealloc; 1972 1973 /* 1974 * We have to hold the backing object, check races. 1975 */ 1976 while ((backing_object = object->backing_object) != NULL) { 1977 vm_object_hold(backing_object); 1978 if (backing_object == object->backing_object) 1979 break; 1980 vm_object_drop(backing_object); 1981 } 1982 1983 /* 1984 * No backing object? Nothing to collapse then. 1985 */ 1986 if (backing_object == NULL) 1987 break; 1988 1989 /* 1990 * You can't collapse with a non-default/non-swap object. 1991 */ 1992 if (backing_object->type != OBJT_DEFAULT && 1993 backing_object->type != OBJT_SWAP) { 1994 vm_object_drop(backing_object); 1995 backing_object = NULL; 1996 break; 1997 } 1998 1999 /* 2000 * Chain-lock the backing object too because if we 2001 * successfully merge its pages into the top object we 2002 * will collapse backing_object->backing_object as the 2003 * new backing_object. Re-check that it is still our 2004 * backing object. 2005 */ 2006 vm_object_chain_acquire(backing_object); 2007 if (backing_object != object->backing_object) { 2008 vm_object_chain_release(backing_object); 2009 vm_object_drop(backing_object); 2010 continue; 2011 } 2012 2013 /* 2014 * we check the backing object first, because it is most likely 2015 * not collapsable. 2016 */ 2017 if (backing_object->handle != NULL || 2018 (backing_object->type != OBJT_DEFAULT && 2019 backing_object->type != OBJT_SWAP) || 2020 (backing_object->flags & OBJ_DEAD) || 2021 object->handle != NULL || 2022 (object->type != OBJT_DEFAULT && 2023 object->type != OBJT_SWAP) || 2024 (object->flags & OBJ_DEAD)) { 2025 break; 2026 } 2027 2028 /* 2029 * If paging is in progress we can't do a normal collapse. 2030 */ 2031 if ( 2032 object->paging_in_progress != 0 || 2033 backing_object->paging_in_progress != 0 2034 ) { 2035 vm_object_qcollapse(object, backing_object); 2036 break; 2037 } 2038 2039 /* 2040 * We know that we can either collapse the backing object (if 2041 * the parent is the only reference to it) or (perhaps) have 2042 * the parent bypass the object if the parent happens to shadow 2043 * all the resident pages in the entire backing object. 2044 * 2045 * This is ignoring pager-backed pages such as swap pages. 2046 * vm_object_backing_scan fails the shadowing test in this 2047 * case. 2048 */ 2049 if (backing_object->ref_count == 1) { 2050 /* 2051 * If there is exactly one reference to the backing 2052 * object, we can collapse it into the parent. 2053 */ 2054 KKASSERT(object->backing_object == backing_object); 2055 vm_object_backing_scan(object, backing_object, 2056 OBSC_COLLAPSE_WAIT); 2057 2058 /* 2059 * Move the pager from backing_object to object. 2060 */ 2061 if (backing_object->type == OBJT_SWAP) { 2062 vm_object_pip_add(backing_object, 1); 2063 2064 /* 2065 * scrap the paging_offset junk and do a 2066 * discrete copy. This also removes major 2067 * assumptions about how the swap-pager 2068 * works from where it doesn't belong. The 2069 * new swapper is able to optimize the 2070 * destroy-source case. 2071 */ 2072 vm_object_pip_add(object, 1); 2073 swap_pager_copy(backing_object, object, 2074 OFF_TO_IDX(object->backing_object_offset), 2075 TRUE); 2076 vm_object_pip_wakeup(object); 2077 vm_object_pip_wakeup(backing_object); 2078 } 2079 2080 /* 2081 * Object now shadows whatever backing_object did. 2082 * Remove object from backing_object's shadow_list. 2083 */ 2084 LIST_REMOVE(object, shadow_list); 2085 KKASSERT(object->backing_object == backing_object); 2086 backing_object->shadow_count--; 2087 backing_object->generation++; 2088 2089 /* 2090 * backing_object->backing_object moves from within 2091 * backing_object to within object. 2092 */ 2093 while ((bbobj = backing_object->backing_object) != NULL) { 2094 vm_object_hold(bbobj); 2095 if (bbobj == backing_object->backing_object) 2096 break; 2097 vm_object_drop(bbobj); 2098 } 2099 if (bbobj) { 2100 LIST_REMOVE(backing_object, shadow_list); 2101 bbobj->shadow_count--; 2102 bbobj->generation++; 2103 backing_object->backing_object = NULL; 2104 } 2105 object->backing_object = bbobj; 2106 if (bbobj) { 2107 LIST_INSERT_HEAD(&bbobj->shadow_head, 2108 object, shadow_list); 2109 bbobj->shadow_count++; 2110 bbobj->generation++; 2111 } 2112 2113 object->backing_object_offset += 2114 backing_object->backing_object_offset; 2115 2116 vm_object_drop(bbobj); 2117 2118 /* 2119 * Discard the old backing_object. Nothing should be 2120 * able to ref it, other than a vm_map_split(), 2121 * and vm_map_split() will stall on our chain lock. 2122 * And we control the parent so it shouldn't be 2123 * possible for it to go away either. 2124 * 2125 * Since the backing object has no pages, no pager 2126 * left, and no object references within it, all 2127 * that is necessary is to dispose of it. 2128 */ 2129 KASSERT(backing_object->ref_count == 1, 2130 ("backing_object %p was somehow " 2131 "re-referenced during collapse!", 2132 backing_object)); 2133 KASSERT(RB_EMPTY(&backing_object->rb_memq), 2134 ("backing_object %p somehow has left " 2135 "over pages during collapse!", 2136 backing_object)); 2137 2138 /* 2139 * The object can be destroyed. 2140 * 2141 * XXX just fall through and dodealloc instead 2142 * of forcing destruction? 2143 */ 2144 --backing_object->ref_count; 2145 if ((backing_object->flags & OBJ_DEAD) == 0) 2146 vm_object_terminate(backing_object); 2147 object_collapses++; 2148 dodealloc = 0; 2149 } else { 2150 /* 2151 * If we do not entirely shadow the backing object, 2152 * there is nothing we can do so we give up. 2153 */ 2154 if (vm_object_backing_scan(object, backing_object, 2155 OBSC_TEST_ALL_SHADOWED) == 0) { 2156 break; 2157 } 2158 2159 /* 2160 * bbobj is backing_object->backing_object. Since 2161 * object completely shadows backing_object we can 2162 * bypass it and become backed by bbobj instead. 2163 */ 2164 while ((bbobj = backing_object->backing_object) != NULL) { 2165 vm_object_hold(bbobj); 2166 if (bbobj == backing_object->backing_object) 2167 break; 2168 vm_object_drop(bbobj); 2169 } 2170 2171 /* 2172 * Make object shadow bbobj instead of backing_object. 2173 * Remove object from backing_object's shadow list. 2174 * 2175 * Deallocating backing_object will not remove 2176 * it, since its reference count is at least 2. 2177 */ 2178 KKASSERT(object->backing_object == backing_object); 2179 LIST_REMOVE(object, shadow_list); 2180 backing_object->shadow_count--; 2181 backing_object->generation++; 2182 2183 /* 2184 * Add a ref to bbobj, bbobj now shadows object. 2185 * 2186 * NOTE: backing_object->backing_object still points 2187 * to bbobj. That relationship remains intact 2188 * because backing_object has > 1 ref, so 2189 * someone else is pointing to it (hence why 2190 * we can't collapse it into object and can 2191 * only handle the all-shadowed bypass case). 2192 */ 2193 if (bbobj) { 2194 vm_object_chain_wait(bbobj); 2195 vm_object_reference_locked(bbobj); 2196 LIST_INSERT_HEAD(&bbobj->shadow_head, 2197 object, shadow_list); 2198 bbobj->shadow_count++; 2199 bbobj->generation++; 2200 object->backing_object_offset += 2201 backing_object->backing_object_offset; 2202 object->backing_object = bbobj; 2203 vm_object_drop(bbobj); 2204 } else { 2205 object->backing_object = NULL; 2206 } 2207 2208 /* 2209 * Drop the reference count on backing_object. To 2210 * handle ref_count races properly we can't assume 2211 * that the ref_count is still at least 2 so we 2212 * have to actually call vm_object_deallocate() 2213 * (after clearing the chainlock). 2214 */ 2215 object_bypasses++; 2216 dodealloc = 1; 2217 } 2218 2219 /* 2220 * Ok, we want to loop on the new object->bbobj association, 2221 * possibly collapsing it further. However if dodealloc is 2222 * non-zero we have to deallocate the backing_object which 2223 * itself can potentially undergo a collapse, creating a 2224 * recursion depth issue with the LWKT token subsystem. 2225 * 2226 * In the case where we must deallocate the backing_object 2227 * it is possible now that the backing_object has a single 2228 * shadow count on some other object (not represented here 2229 * as yet), since it no longer shadows us. Thus when we 2230 * call vm_object_deallocate() it may attempt to collapse 2231 * itself into its remaining parent. 2232 */ 2233 if (dodealloc) { 2234 struct vm_object_dealloc_list *dtmp; 2235 2236 vm_object_chain_release(backing_object); 2237 vm_object_unlock(backing_object); 2238 /* backing_object remains held */ 2239 2240 /* 2241 * Auto-deallocation list for caller convenience. 2242 */ 2243 if (dlistp == NULL) 2244 dlistp = &dlist; 2245 2246 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK); 2247 dtmp->object = backing_object; 2248 dtmp->next = *dlistp; 2249 *dlistp = dtmp; 2250 } else { 2251 vm_object_chain_release(backing_object); 2252 vm_object_drop(backing_object); 2253 } 2254 /* backing_object = NULL; not needed */ 2255 /* loop */ 2256 } 2257 2258 /* 2259 * Clean up any left over backing_object 2260 */ 2261 if (backing_object) { 2262 vm_object_chain_release(backing_object); 2263 vm_object_drop(backing_object); 2264 } 2265 2266 /* 2267 * Clean up any auto-deallocation list. This is a convenience 2268 * for top-level callers so they don't have to pass &dlist. 2269 * Do not clean up any caller-passed dlistp, the caller will 2270 * do that. 2271 */ 2272 if (dlist) 2273 vm_object_deallocate_list(&dlist); 2274 2275 } 2276 2277 /* 2278 * vm_object_collapse() may collect additional objects in need of 2279 * deallocation. This routine deallocates these objects. The 2280 * deallocation itself can trigger additional collapses (which the 2281 * deallocate function takes care of). This procedure is used to 2282 * reduce procedural recursion since these vm_object shadow chains 2283 * can become quite long. 2284 */ 2285 void 2286 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp) 2287 { 2288 struct vm_object_dealloc_list *dlist; 2289 2290 while ((dlist = *dlistp) != NULL) { 2291 *dlistp = dlist->next; 2292 vm_object_lock(dlist->object); 2293 vm_object_deallocate_locked(dlist->object); 2294 vm_object_drop(dlist->object); 2295 kfree(dlist, M_TEMP); 2296 } 2297 } 2298 2299 /* 2300 * Removes all physical pages in the specified object range from the 2301 * object's list of pages. 2302 * 2303 * No requirements. 2304 */ 2305 static int vm_object_page_remove_callback(vm_page_t p, void *data); 2306 2307 void 2308 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2309 boolean_t clean_only) 2310 { 2311 struct rb_vm_page_scan_info info; 2312 int all; 2313 2314 /* 2315 * Degenerate cases and assertions 2316 */ 2317 vm_object_hold(object); 2318 if (object == NULL || 2319 (object->resident_page_count == 0 && object->swblock_count == 0)) { 2320 vm_object_drop(object); 2321 return; 2322 } 2323 KASSERT(object->type != OBJT_PHYS, 2324 ("attempt to remove pages from a physical object")); 2325 2326 /* 2327 * Indicate that paging is occuring on the object 2328 */ 2329 vm_object_pip_add(object, 1); 2330 2331 /* 2332 * Figure out the actual removal range and whether we are removing 2333 * the entire contents of the object or not. If removing the entire 2334 * contents, be sure to get all pages, even those that might be 2335 * beyond the end of the object. 2336 */ 2337 info.start_pindex = start; 2338 if (end == 0) 2339 info.end_pindex = (vm_pindex_t)-1; 2340 else 2341 info.end_pindex = end - 1; 2342 info.limit = clean_only; 2343 all = (start == 0 && info.end_pindex >= object->size - 1); 2344 2345 /* 2346 * Loop until we are sure we have gotten them all. 2347 */ 2348 do { 2349 info.error = 0; 2350 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2351 vm_object_page_remove_callback, &info); 2352 } while (info.error); 2353 2354 /* 2355 * Remove any related swap if throwing away pages, or for 2356 * non-swap objects (the swap is a clean copy in that case). 2357 */ 2358 if (object->type != OBJT_SWAP || clean_only == FALSE) { 2359 if (all) 2360 swap_pager_freespace_all(object); 2361 else 2362 swap_pager_freespace(object, info.start_pindex, 2363 info.end_pindex - info.start_pindex + 1); 2364 } 2365 2366 /* 2367 * Cleanup 2368 */ 2369 vm_object_pip_wakeup(object); 2370 vm_object_drop(object); 2371 } 2372 2373 /* 2374 * The caller must hold the object 2375 */ 2376 static int 2377 vm_object_page_remove_callback(vm_page_t p, void *data) 2378 { 2379 struct rb_vm_page_scan_info *info = data; 2380 2381 if (vm_page_busy_try(p, TRUE)) { 2382 vm_page_sleep_busy(p, TRUE, "vmopar"); 2383 info->error = 1; 2384 return(0); 2385 } 2386 2387 /* 2388 * Wired pages cannot be destroyed, but they can be invalidated 2389 * and we do so if clean_only (limit) is not set. 2390 * 2391 * WARNING! The page may be wired due to being part of a buffer 2392 * cache buffer, and the buffer might be marked B_CACHE. 2393 * This is fine as part of a truncation but VFSs must be 2394 * sure to fix the buffer up when re-extending the file. 2395 * 2396 * NOTE! PG_NEED_COMMIT is ignored. 2397 */ 2398 if (p->wire_count != 0) { 2399 vm_page_protect(p, VM_PROT_NONE); 2400 if (info->limit == 0) 2401 p->valid = 0; 2402 vm_page_wakeup(p); 2403 return(0); 2404 } 2405 2406 /* 2407 * limit is our clean_only flag. If set and the page is dirty or 2408 * requires a commit, do not free it. If set and the page is being 2409 * held by someone, do not free it. 2410 */ 2411 if (info->limit && p->valid) { 2412 vm_page_test_dirty(p); 2413 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) { 2414 vm_page_wakeup(p); 2415 return(0); 2416 } 2417 #if 0 2418 if (p->hold_count) { 2419 vm_page_wakeup(p); 2420 return(0); 2421 } 2422 #endif 2423 } 2424 2425 /* 2426 * Destroy the page 2427 */ 2428 vm_page_protect(p, VM_PROT_NONE); 2429 vm_page_free(p); 2430 return(0); 2431 } 2432 2433 /* 2434 * Coalesces two objects backing up adjoining regions of memory into a 2435 * single object. 2436 * 2437 * returns TRUE if objects were combined. 2438 * 2439 * NOTE: Only works at the moment if the second object is NULL - 2440 * if it's not, which object do we lock first? 2441 * 2442 * Parameters: 2443 * prev_object First object to coalesce 2444 * prev_offset Offset into prev_object 2445 * next_object Second object into coalesce 2446 * next_offset Offset into next_object 2447 * 2448 * prev_size Size of reference to prev_object 2449 * next_size Size of reference to next_object 2450 * 2451 * The caller does not need to hold (prev_object) but must have a stable 2452 * pointer to it (typically by holding the vm_map locked). 2453 */ 2454 boolean_t 2455 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 2456 vm_size_t prev_size, vm_size_t next_size) 2457 { 2458 vm_pindex_t next_pindex; 2459 2460 if (prev_object == NULL) 2461 return (TRUE); 2462 2463 vm_object_hold(prev_object); 2464 2465 if (prev_object->type != OBJT_DEFAULT && 2466 prev_object->type != OBJT_SWAP) { 2467 vm_object_drop(prev_object); 2468 return (FALSE); 2469 } 2470 2471 /* 2472 * Try to collapse the object first 2473 */ 2474 vm_object_chain_acquire(prev_object); 2475 vm_object_collapse(prev_object, NULL); 2476 2477 /* 2478 * Can't coalesce if: . more than one reference . paged out . shadows 2479 * another object . has a copy elsewhere (any of which mean that the 2480 * pages not mapped to prev_entry may be in use anyway) 2481 */ 2482 2483 if (prev_object->backing_object != NULL) { 2484 vm_object_chain_release(prev_object); 2485 vm_object_drop(prev_object); 2486 return (FALSE); 2487 } 2488 2489 prev_size >>= PAGE_SHIFT; 2490 next_size >>= PAGE_SHIFT; 2491 next_pindex = prev_pindex + prev_size; 2492 2493 if ((prev_object->ref_count > 1) && 2494 (prev_object->size != next_pindex)) { 2495 vm_object_chain_release(prev_object); 2496 vm_object_drop(prev_object); 2497 return (FALSE); 2498 } 2499 2500 /* 2501 * Remove any pages that may still be in the object from a previous 2502 * deallocation. 2503 */ 2504 if (next_pindex < prev_object->size) { 2505 vm_object_page_remove(prev_object, 2506 next_pindex, 2507 next_pindex + next_size, FALSE); 2508 if (prev_object->type == OBJT_SWAP) 2509 swap_pager_freespace(prev_object, 2510 next_pindex, next_size); 2511 } 2512 2513 /* 2514 * Extend the object if necessary. 2515 */ 2516 if (next_pindex + next_size > prev_object->size) 2517 prev_object->size = next_pindex + next_size; 2518 2519 vm_object_chain_release(prev_object); 2520 vm_object_drop(prev_object); 2521 return (TRUE); 2522 } 2523 2524 /* 2525 * Make the object writable and flag is being possibly dirty. 2526 * 2527 * The caller must hold the object. XXX called from vm_page_dirty(), 2528 * There is currently no requirement to hold the object. 2529 */ 2530 void 2531 vm_object_set_writeable_dirty(vm_object_t object) 2532 { 2533 struct vnode *vp; 2534 2535 /*vm_object_assert_held(object);*/ 2536 /* 2537 * Avoid contention in vm fault path by checking the state before 2538 * issuing an atomic op on it. 2539 */ 2540 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) != 2541 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) { 2542 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 2543 } 2544 if (object->type == OBJT_VNODE && 2545 (vp = (struct vnode *)object->handle) != NULL) { 2546 if ((vp->v_flag & VOBJDIRTY) == 0) { 2547 vsetflags(vp, VOBJDIRTY); 2548 } 2549 } 2550 } 2551 2552 #include "opt_ddb.h" 2553 #ifdef DDB 2554 #include <sys/kernel.h> 2555 2556 #include <sys/cons.h> 2557 2558 #include <ddb/ddb.h> 2559 2560 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 2561 vm_map_entry_t entry); 2562 static int vm_object_in_map (vm_object_t object); 2563 2564 /* 2565 * The caller must hold the object. 2566 */ 2567 static int 2568 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2569 { 2570 vm_map_t tmpm; 2571 vm_map_entry_t tmpe; 2572 vm_object_t obj, nobj; 2573 int entcount; 2574 2575 if (map == 0) 2576 return 0; 2577 if (entry == 0) { 2578 tmpe = map->header.next; 2579 entcount = map->nentries; 2580 while (entcount-- && (tmpe != &map->header)) { 2581 if( _vm_object_in_map(map, object, tmpe)) { 2582 return 1; 2583 } 2584 tmpe = tmpe->next; 2585 } 2586 return (0); 2587 } 2588 switch(entry->maptype) { 2589 case VM_MAPTYPE_SUBMAP: 2590 tmpm = entry->object.sub_map; 2591 tmpe = tmpm->header.next; 2592 entcount = tmpm->nentries; 2593 while (entcount-- && tmpe != &tmpm->header) { 2594 if( _vm_object_in_map(tmpm, object, tmpe)) { 2595 return 1; 2596 } 2597 tmpe = tmpe->next; 2598 } 2599 break; 2600 case VM_MAPTYPE_NORMAL: 2601 case VM_MAPTYPE_VPAGETABLE: 2602 obj = entry->object.vm_object; 2603 while (obj) { 2604 if (obj == object) { 2605 if (obj != entry->object.vm_object) 2606 vm_object_drop(obj); 2607 return 1; 2608 } 2609 while ((nobj = obj->backing_object) != NULL) { 2610 vm_object_hold(nobj); 2611 if (nobj == obj->backing_object) 2612 break; 2613 vm_object_drop(nobj); 2614 } 2615 if (obj != entry->object.vm_object) { 2616 if (nobj) 2617 vm_object_lock_swap(); 2618 vm_object_drop(obj); 2619 } 2620 obj = nobj; 2621 } 2622 break; 2623 default: 2624 break; 2625 } 2626 return 0; 2627 } 2628 2629 static int vm_object_in_map_callback(struct proc *p, void *data); 2630 2631 struct vm_object_in_map_info { 2632 vm_object_t object; 2633 int rv; 2634 }; 2635 2636 /* 2637 * Debugging only 2638 */ 2639 static int 2640 vm_object_in_map(vm_object_t object) 2641 { 2642 struct vm_object_in_map_info info; 2643 2644 info.rv = 0; 2645 info.object = object; 2646 2647 allproc_scan(vm_object_in_map_callback, &info); 2648 if (info.rv) 2649 return 1; 2650 if( _vm_object_in_map(&kernel_map, object, 0)) 2651 return 1; 2652 if( _vm_object_in_map(&pager_map, object, 0)) 2653 return 1; 2654 if( _vm_object_in_map(&buffer_map, object, 0)) 2655 return 1; 2656 return 0; 2657 } 2658 2659 /* 2660 * Debugging only 2661 */ 2662 static int 2663 vm_object_in_map_callback(struct proc *p, void *data) 2664 { 2665 struct vm_object_in_map_info *info = data; 2666 2667 if (p->p_vmspace) { 2668 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 2669 info->rv = 1; 2670 return -1; 2671 } 2672 } 2673 return (0); 2674 } 2675 2676 DB_SHOW_COMMAND(vmochk, vm_object_check) 2677 { 2678 vm_object_t object; 2679 2680 /* 2681 * make sure that internal objs are in a map somewhere 2682 * and none have zero ref counts. 2683 */ 2684 for (object = TAILQ_FIRST(&vm_object_list); 2685 object != NULL; 2686 object = TAILQ_NEXT(object, object_list)) { 2687 if (object->type == OBJT_MARKER) 2688 continue; 2689 if (object->handle == NULL && 2690 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2691 if (object->ref_count == 0) { 2692 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2693 (long)object->size); 2694 } 2695 if (!vm_object_in_map(object)) { 2696 db_printf( 2697 "vmochk: internal obj is not in a map: " 2698 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2699 object->ref_count, (u_long)object->size, 2700 (u_long)object->size, 2701 (void *)object->backing_object); 2702 } 2703 } 2704 } 2705 } 2706 2707 /* 2708 * Debugging only 2709 */ 2710 DB_SHOW_COMMAND(object, vm_object_print_static) 2711 { 2712 /* XXX convert args. */ 2713 vm_object_t object = (vm_object_t)addr; 2714 boolean_t full = have_addr; 2715 2716 vm_page_t p; 2717 2718 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2719 #define count was_count 2720 2721 int count; 2722 2723 if (object == NULL) 2724 return; 2725 2726 db_iprintf( 2727 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2728 object, (int)object->type, (u_long)object->size, 2729 object->resident_page_count, object->ref_count, object->flags); 2730 /* 2731 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2732 */ 2733 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2734 object->shadow_count, 2735 object->backing_object ? object->backing_object->ref_count : 0, 2736 object->backing_object, (long)object->backing_object_offset); 2737 2738 if (!full) 2739 return; 2740 2741 db_indent += 2; 2742 count = 0; 2743 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 2744 if (count == 0) 2745 db_iprintf("memory:="); 2746 else if (count == 6) { 2747 db_printf("\n"); 2748 db_iprintf(" ..."); 2749 count = 0; 2750 } else 2751 db_printf(","); 2752 count++; 2753 2754 db_printf("(off=0x%lx,page=0x%lx)", 2755 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2756 } 2757 if (count != 0) 2758 db_printf("\n"); 2759 db_indent -= 2; 2760 } 2761 2762 /* XXX. */ 2763 #undef count 2764 2765 /* 2766 * XXX need this non-static entry for calling from vm_map_print. 2767 * 2768 * Debugging only 2769 */ 2770 void 2771 vm_object_print(/* db_expr_t */ long addr, 2772 boolean_t have_addr, 2773 /* db_expr_t */ long count, 2774 char *modif) 2775 { 2776 vm_object_print_static(addr, have_addr, count, modif); 2777 } 2778 2779 /* 2780 * Debugging only 2781 */ 2782 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2783 { 2784 vm_object_t object; 2785 int nl = 0; 2786 int c; 2787 for (object = TAILQ_FIRST(&vm_object_list); 2788 object != NULL; 2789 object = TAILQ_NEXT(object, object_list)) { 2790 vm_pindex_t idx, fidx; 2791 vm_pindex_t osize; 2792 vm_paddr_t pa = -1, padiff; 2793 int rcount; 2794 vm_page_t m; 2795 2796 if (object->type == OBJT_MARKER) 2797 continue; 2798 db_printf("new object: %p\n", (void *)object); 2799 if ( nl > 18) { 2800 c = cngetc(); 2801 if (c != ' ') 2802 return; 2803 nl = 0; 2804 } 2805 nl++; 2806 rcount = 0; 2807 fidx = 0; 2808 osize = object->size; 2809 if (osize > 128) 2810 osize = 128; 2811 for (idx = 0; idx < osize; idx++) { 2812 m = vm_page_lookup(object, idx); 2813 if (m == NULL) { 2814 if (rcount) { 2815 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2816 (long)fidx, rcount, (long)pa); 2817 if ( nl > 18) { 2818 c = cngetc(); 2819 if (c != ' ') 2820 return; 2821 nl = 0; 2822 } 2823 nl++; 2824 rcount = 0; 2825 } 2826 continue; 2827 } 2828 2829 2830 if (rcount && 2831 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2832 ++rcount; 2833 continue; 2834 } 2835 if (rcount) { 2836 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2837 padiff >>= PAGE_SHIFT; 2838 padiff &= PQ_L2_MASK; 2839 if (padiff == 0) { 2840 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2841 ++rcount; 2842 continue; 2843 } 2844 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2845 (long)fidx, rcount, (long)pa); 2846 db_printf("pd(%ld)\n", (long)padiff); 2847 if ( nl > 18) { 2848 c = cngetc(); 2849 if (c != ' ') 2850 return; 2851 nl = 0; 2852 } 2853 nl++; 2854 } 2855 fidx = idx; 2856 pa = VM_PAGE_TO_PHYS(m); 2857 rcount = 1; 2858 } 2859 if (rcount) { 2860 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2861 (long)fidx, rcount, (long)pa); 2862 if ( nl > 18) { 2863 c = cngetc(); 2864 if (c != ' ') 2865 return; 2866 nl = 0; 2867 } 2868 nl++; 2869 } 2870 } 2871 } 2872 #endif /* DDB */ 2873