1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 39 * 40 * 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 * 66 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $ 67 */ 68 69 /* 70 * Virtual memory object module. 71 */ 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/proc.h> /* for curproc, pageproc */ 76 #include <sys/thread.h> 77 #include <sys/vnode.h> 78 #include <sys/vmmeter.h> 79 #include <sys/mman.h> 80 #include <sys/mount.h> 81 #include <sys/kernel.h> 82 #include <sys/sysctl.h> 83 #include <sys/refcount.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pageout.h> 92 #include <vm/vm_pager.h> 93 #include <vm/swap_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/vm_zone.h> 97 98 #define EASY_SCAN_FACTOR 8 99 100 static void vm_object_qcollapse(vm_object_t object, 101 vm_object_t backing_object); 102 static void vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 103 int pagerflags); 104 static void vm_object_lock_init(vm_object_t); 105 106 107 /* 108 * Virtual memory objects maintain the actual data 109 * associated with allocated virtual memory. A given 110 * page of memory exists within exactly one object. 111 * 112 * An object is only deallocated when all "references" 113 * are given up. Only one "reference" to a given 114 * region of an object should be writeable. 115 * 116 * Associated with each object is a list of all resident 117 * memory pages belonging to that object; this list is 118 * maintained by the "vm_page" module, and locked by the object's 119 * lock. 120 * 121 * Each object also records a "pager" routine which is 122 * used to retrieve (and store) pages to the proper backing 123 * storage. In addition, objects may be backed by other 124 * objects from which they were virtual-copied. 125 * 126 * The only items within the object structure which are 127 * modified after time of creation are: 128 * reference count locked by object's lock 129 * pager routine locked by object's lock 130 * 131 */ 132 133 struct object_q vm_object_list; /* locked by vmobj_token */ 134 struct vm_object kernel_object; 135 136 static long vm_object_count; /* locked by vmobj_token */ 137 extern int vm_pageout_page_count; 138 139 static long object_collapses; 140 static long object_bypasses; 141 static int next_index; 142 static vm_zone_t obj_zone; 143 static struct vm_zone obj_zone_store; 144 #define VM_OBJECTS_INIT 256 145 static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 146 147 /* 148 * Misc low level routines 149 */ 150 static void 151 vm_object_lock_init(vm_object_t obj) 152 { 153 #if defined(DEBUG_LOCKS) 154 int i; 155 156 obj->debug_hold_bitmap = 0; 157 obj->debug_hold_ovfl = 0; 158 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 159 obj->debug_hold_thrs[i] = NULL; 160 obj->debug_hold_file[i] = NULL; 161 obj->debug_hold_line[i] = 0; 162 } 163 #endif 164 } 165 166 void 167 vm_object_lock_swap(void) 168 { 169 lwkt_token_swap(); 170 } 171 172 void 173 vm_object_lock(vm_object_t obj) 174 { 175 lwkt_gettoken(&obj->token); 176 } 177 178 /* 179 * Returns TRUE on sucesss 180 */ 181 static int 182 vm_object_lock_try(vm_object_t obj) 183 { 184 return(lwkt_trytoken(&obj->token)); 185 } 186 187 void 188 vm_object_lock_shared(vm_object_t obj) 189 { 190 lwkt_gettoken_shared(&obj->token); 191 } 192 193 void 194 vm_object_unlock(vm_object_t obj) 195 { 196 lwkt_reltoken(&obj->token); 197 } 198 199 static __inline void 200 vm_object_assert_held(vm_object_t obj) 201 { 202 ASSERT_LWKT_TOKEN_HELD(&obj->token); 203 } 204 205 void 206 #ifndef DEBUG_LOCKS 207 vm_object_hold(vm_object_t obj) 208 #else 209 debugvm_object_hold(vm_object_t obj, char *file, int line) 210 #endif 211 { 212 KKASSERT(obj != NULL); 213 214 /* 215 * Object must be held (object allocation is stable due to callers 216 * context, typically already holding the token on a parent object) 217 * prior to potentially blocking on the lock, otherwise the object 218 * can get ripped away from us. 219 */ 220 refcount_acquire(&obj->hold_count); 221 vm_object_lock(obj); 222 223 #if defined(DEBUG_LOCKS) 224 int i; 225 u_int mask; 226 227 for (;;) { 228 mask = ~obj->debug_hold_bitmap; 229 cpu_ccfence(); 230 if (mask == 0xFFFFFFFFU) { 231 if (obj->debug_hold_ovfl == 0) 232 obj->debug_hold_ovfl = 1; 233 break; 234 } 235 i = ffs(mask) - 1; 236 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 237 ~mask | (1 << i))) { 238 obj->debug_hold_bitmap |= (1 << i); 239 obj->debug_hold_thrs[i] = curthread; 240 obj->debug_hold_file[i] = file; 241 obj->debug_hold_line[i] = line; 242 break; 243 } 244 } 245 #endif 246 } 247 248 int 249 #ifndef DEBUG_LOCKS 250 vm_object_hold_try(vm_object_t obj) 251 #else 252 debugvm_object_hold_try(vm_object_t obj, char *file, int line) 253 #endif 254 { 255 KKASSERT(obj != NULL); 256 257 /* 258 * Object must be held (object allocation is stable due to callers 259 * context, typically already holding the token on a parent object) 260 * prior to potentially blocking on the lock, otherwise the object 261 * can get ripped away from us. 262 */ 263 refcount_acquire(&obj->hold_count); 264 if (vm_object_lock_try(obj) == 0) { 265 if (refcount_release(&obj->hold_count)) { 266 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) 267 zfree(obj_zone, obj); 268 } 269 return(0); 270 } 271 272 #if defined(DEBUG_LOCKS) 273 int i; 274 u_int mask; 275 276 for (;;) { 277 mask = ~obj->debug_hold_bitmap; 278 cpu_ccfence(); 279 if (mask == 0xFFFFFFFFU) { 280 if (obj->debug_hold_ovfl == 0) 281 obj->debug_hold_ovfl = 1; 282 break; 283 } 284 i = ffs(mask) - 1; 285 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 286 ~mask | (1 << i))) { 287 obj->debug_hold_bitmap |= (1 << i); 288 obj->debug_hold_thrs[i] = curthread; 289 obj->debug_hold_file[i] = file; 290 obj->debug_hold_line[i] = line; 291 break; 292 } 293 } 294 #endif 295 return(1); 296 } 297 298 void 299 #ifndef DEBUG_LOCKS 300 vm_object_hold_shared(vm_object_t obj) 301 #else 302 debugvm_object_hold_shared(vm_object_t obj, char *file, int line) 303 #endif 304 { 305 KKASSERT(obj != NULL); 306 307 /* 308 * Object must be held (object allocation is stable due to callers 309 * context, typically already holding the token on a parent object) 310 * prior to potentially blocking on the lock, otherwise the object 311 * can get ripped away from us. 312 */ 313 refcount_acquire(&obj->hold_count); 314 vm_object_lock_shared(obj); 315 316 #if defined(DEBUG_LOCKS) 317 int i; 318 u_int mask; 319 320 for (;;) { 321 mask = ~obj->debug_hold_bitmap; 322 cpu_ccfence(); 323 if (mask == 0xFFFFFFFFU) { 324 if (obj->debug_hold_ovfl == 0) 325 obj->debug_hold_ovfl = 1; 326 break; 327 } 328 i = ffs(mask) - 1; 329 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask, 330 ~mask | (1 << i))) { 331 obj->debug_hold_bitmap |= (1 << i); 332 obj->debug_hold_thrs[i] = curthread; 333 obj->debug_hold_file[i] = file; 334 obj->debug_hold_line[i] = line; 335 break; 336 } 337 } 338 #endif 339 } 340 341 /* 342 * Drop the token and hold_count on the object. 343 */ 344 void 345 vm_object_drop(vm_object_t obj) 346 { 347 if (obj == NULL) 348 return; 349 350 #if defined(DEBUG_LOCKS) 351 int found = 0; 352 int i; 353 354 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 355 if ((obj->debug_hold_bitmap & (1 << i)) && 356 (obj->debug_hold_thrs[i] == curthread)) { 357 obj->debug_hold_bitmap &= ~(1 << i); 358 obj->debug_hold_thrs[i] = NULL; 359 obj->debug_hold_file[i] = NULL; 360 obj->debug_hold_line[i] = 0; 361 found = 1; 362 break; 363 } 364 } 365 366 if (found == 0 && obj->debug_hold_ovfl == 0) 367 panic("vm_object: attempt to drop hold on non-self-held obj"); 368 #endif 369 370 /* 371 * No new holders should be possible once we drop hold_count 1->0 as 372 * there is no longer any way to reference the object. 373 */ 374 KKASSERT(obj->hold_count > 0); 375 if (refcount_release(&obj->hold_count)) { 376 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) { 377 vm_object_unlock(obj); 378 zfree(obj_zone, obj); 379 } else { 380 vm_object_unlock(obj); 381 } 382 } else { 383 vm_object_unlock(obj); 384 } 385 } 386 387 /* 388 * Initialize a freshly allocated object, returning a held object. 389 * 390 * Used only by vm_object_allocate() and zinitna(). 391 * 392 * No requirements. 393 */ 394 void 395 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 396 { 397 int incr; 398 399 RB_INIT(&object->rb_memq); 400 LIST_INIT(&object->shadow_head); 401 lwkt_token_init(&object->token, "vmobj"); 402 403 object->type = type; 404 object->size = size; 405 object->ref_count = 1; 406 object->hold_count = 0; 407 object->flags = 0; 408 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 409 vm_object_set_flag(object, OBJ_ONEMAPPING); 410 object->paging_in_progress = 0; 411 object->resident_page_count = 0; 412 object->agg_pv_list_count = 0; 413 object->shadow_count = 0; 414 /* cpu localization twist */ 415 object->pg_color = (int)(intptr_t)curthread; 416 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 417 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 418 else 419 incr = size; 420 next_index = (next_index + incr) & PQ_L2_MASK; 421 object->handle = NULL; 422 object->backing_object = NULL; 423 object->backing_object_offset = (vm_ooffset_t)0; 424 425 object->generation++; 426 object->swblock_count = 0; 427 RB_INIT(&object->swblock_root); 428 vm_object_lock_init(object); 429 pmap_object_init(object); 430 431 vm_object_hold(object); 432 lwkt_gettoken(&vmobj_token); 433 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 434 vm_object_count++; 435 lwkt_reltoken(&vmobj_token); 436 } 437 438 /* 439 * Initialize the VM objects module. 440 * 441 * Called from the low level boot code only. 442 */ 443 void 444 vm_object_init(void) 445 { 446 TAILQ_INIT(&vm_object_list); 447 448 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 449 &kernel_object); 450 vm_object_drop(&kernel_object); 451 452 obj_zone = &obj_zone_store; 453 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 454 vm_objects_init, VM_OBJECTS_INIT); 455 } 456 457 void 458 vm_object_init2(void) 459 { 460 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1); 461 } 462 463 /* 464 * Allocate and return a new object of the specified type and size. 465 * 466 * No requirements. 467 */ 468 vm_object_t 469 vm_object_allocate(objtype_t type, vm_pindex_t size) 470 { 471 vm_object_t result; 472 473 result = (vm_object_t) zalloc(obj_zone); 474 475 _vm_object_allocate(type, size, result); 476 vm_object_drop(result); 477 478 return (result); 479 } 480 481 /* 482 * This version returns a held object, allowing further atomic initialization 483 * of the object. 484 */ 485 vm_object_t 486 vm_object_allocate_hold(objtype_t type, vm_pindex_t size) 487 { 488 vm_object_t result; 489 490 result = (vm_object_t) zalloc(obj_zone); 491 492 _vm_object_allocate(type, size, result); 493 494 return (result); 495 } 496 497 /* 498 * Add an additional reference to a vm_object. The object must already be 499 * held. The original non-lock version is no longer supported. The object 500 * must NOT be chain locked by anyone at the time the reference is added. 501 * 502 * Referencing a chain-locked object can blow up the fairly sensitive 503 * ref_count and shadow_count tests in the deallocator. Most callers 504 * will call vm_object_chain_wait() prior to calling 505 * vm_object_reference_locked() to avoid the case. 506 * 507 * The object must be held. 508 */ 509 void 510 vm_object_reference_locked(vm_object_t object) 511 { 512 KKASSERT(object != NULL); 513 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 514 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0); 515 object->ref_count++; 516 if (object->type == OBJT_VNODE) { 517 vref(object->handle); 518 /* XXX what if the vnode is being destroyed? */ 519 } 520 } 521 522 /* 523 * Object OBJ_CHAINLOCK lock handling. 524 * 525 * The caller can chain-lock backing objects recursively and then 526 * use vm_object_chain_release_all() to undo the whole chain. 527 * 528 * Chain locks are used to prevent collapses and are only applicable 529 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations 530 * on other object types are ignored. This is also important because 531 * it allows e.g. the vnode underlying a memory mapping to take concurrent 532 * faults. 533 * 534 * The object must usually be held on entry, though intermediate 535 * objects need not be held on release. 536 */ 537 void 538 vm_object_chain_wait(vm_object_t object) 539 { 540 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 541 while (object->flags & OBJ_CHAINLOCK) { 542 vm_object_set_flag(object, OBJ_CHAINWANT); 543 tsleep(object, 0, "objchain", 0); 544 } 545 } 546 547 void 548 vm_object_chain_acquire(vm_object_t object) 549 { 550 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) { 551 vm_object_chain_wait(object); 552 vm_object_set_flag(object, OBJ_CHAINLOCK); 553 } 554 } 555 556 void 557 vm_object_chain_release(vm_object_t object) 558 { 559 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 560 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) { 561 KKASSERT(object->flags & OBJ_CHAINLOCK); 562 if (object->flags & OBJ_CHAINWANT) { 563 vm_object_clear_flag(object, 564 OBJ_CHAINLOCK | OBJ_CHAINWANT); 565 wakeup(object); 566 } else { 567 vm_object_clear_flag(object, OBJ_CHAINLOCK); 568 } 569 } 570 } 571 572 /* 573 * This releases the entire chain of objects from first_object to and 574 * including stopobj, flowing through object->backing_object. 575 * 576 * We release stopobj first as an optimization as this object is most 577 * likely to be shared across multiple processes. 578 */ 579 void 580 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj) 581 { 582 vm_object_t backing_object; 583 vm_object_t object; 584 585 vm_object_chain_release(stopobj); 586 object = first_object; 587 588 while (object != stopobj) { 589 KKASSERT(object); 590 if (object != first_object) 591 vm_object_hold(object); 592 backing_object = object->backing_object; 593 vm_object_chain_release(object); 594 if (object != first_object) 595 vm_object_drop(object); 596 object = backing_object; 597 } 598 } 599 600 /* 601 * Dereference an object and its underlying vnode. 602 * 603 * The object must be held and will be held on return. 604 */ 605 static void 606 vm_object_vndeallocate(vm_object_t object) 607 { 608 struct vnode *vp = (struct vnode *) object->handle; 609 610 KASSERT(object->type == OBJT_VNODE, 611 ("vm_object_vndeallocate: not a vnode object")); 612 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 613 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 614 #ifdef INVARIANTS 615 if (object->ref_count == 0) { 616 vprint("vm_object_vndeallocate", vp); 617 panic("vm_object_vndeallocate: bad object reference count"); 618 } 619 #endif 620 object->ref_count--; 621 if (object->ref_count == 0) 622 vclrflags(vp, VTEXT); 623 vrele(vp); 624 } 625 626 /* 627 * Release a reference to the specified object, gained either through a 628 * vm_object_allocate or a vm_object_reference call. When all references 629 * are gone, storage associated with this object may be relinquished. 630 * 631 * The caller does not have to hold the object locked but must have control 632 * over the reference in question in order to guarantee that the object 633 * does not get ripped out from under us. 634 */ 635 void 636 vm_object_deallocate(vm_object_t object) 637 { 638 if (object) { 639 vm_object_hold(object); 640 vm_object_deallocate_locked(object); 641 vm_object_drop(object); 642 } 643 } 644 645 void 646 vm_object_deallocate_locked(vm_object_t object) 647 { 648 struct vm_object_dealloc_list *dlist = NULL; 649 struct vm_object_dealloc_list *dtmp; 650 vm_object_t temp; 651 int must_drop = 0; 652 653 /* 654 * We may chain deallocate object, but additional objects may 655 * collect on the dlist which also have to be deallocated. We 656 * must avoid a recursion, vm_object chains can get deep. 657 */ 658 again: 659 while (object != NULL) { 660 #if 0 661 /* 662 * Don't rip a ref_count out from under an object undergoing 663 * collapse, it will confuse the collapse code. 664 */ 665 vm_object_chain_wait(object); 666 #endif 667 if (object->type == OBJT_VNODE) { 668 vm_object_vndeallocate(object); 669 break; 670 } 671 672 if (object->ref_count == 0) { 673 panic("vm_object_deallocate: object deallocated " 674 "too many times: %d", object->type); 675 } 676 if (object->ref_count > 2) { 677 object->ref_count--; 678 break; 679 } 680 681 /* 682 * Here on ref_count of one or two, which are special cases for 683 * objects. 684 * 685 * Nominal ref_count > 1 case if the second ref is not from 686 * a shadow. 687 */ 688 if (object->ref_count == 2 && object->shadow_count == 0) { 689 vm_object_set_flag(object, OBJ_ONEMAPPING); 690 object->ref_count--; 691 break; 692 } 693 694 /* 695 * If the second ref is from a shadow we chain along it 696 * upwards if object's handle is exhausted. 697 * 698 * We have to decrement object->ref_count before potentially 699 * collapsing the first shadow object or the collapse code 700 * will not be able to handle the degenerate case to remove 701 * object. However, if we do it too early the object can 702 * get ripped out from under us. 703 */ 704 if (object->ref_count == 2 && object->shadow_count == 1 && 705 object->handle == NULL && (object->type == OBJT_DEFAULT || 706 object->type == OBJT_SWAP)) { 707 temp = LIST_FIRST(&object->shadow_head); 708 KKASSERT(temp != NULL); 709 vm_object_hold(temp); 710 711 /* 712 * Wait for any paging to complete so the collapse 713 * doesn't (or isn't likely to) qcollapse. pip 714 * waiting must occur before we acquire the 715 * chainlock. 716 */ 717 while ( 718 temp->paging_in_progress || 719 object->paging_in_progress 720 ) { 721 vm_object_pip_wait(temp, "objde1"); 722 vm_object_pip_wait(object, "objde2"); 723 } 724 725 /* 726 * If the parent is locked we have to give up, as 727 * otherwise we would be acquiring locks in the 728 * wrong order and potentially deadlock. 729 */ 730 if (temp->flags & OBJ_CHAINLOCK) { 731 vm_object_drop(temp); 732 goto skip; 733 } 734 vm_object_chain_acquire(temp); 735 736 /* 737 * Recheck/retry after the hold and the paging 738 * wait, both of which can block us. 739 */ 740 if (object->ref_count != 2 || 741 object->shadow_count != 1 || 742 object->handle || 743 LIST_FIRST(&object->shadow_head) != temp || 744 (object->type != OBJT_DEFAULT && 745 object->type != OBJT_SWAP)) { 746 vm_object_chain_release(temp); 747 vm_object_drop(temp); 748 continue; 749 } 750 751 /* 752 * We can safely drop object's ref_count now. 753 */ 754 KKASSERT(object->ref_count == 2); 755 object->ref_count--; 756 757 /* 758 * If our single parent is not collapseable just 759 * decrement ref_count (2->1) and stop. 760 */ 761 if (temp->handle || (temp->type != OBJT_DEFAULT && 762 temp->type != OBJT_SWAP)) { 763 vm_object_chain_release(temp); 764 vm_object_drop(temp); 765 break; 766 } 767 768 /* 769 * At this point we have already dropped object's 770 * ref_count so it is possible for a race to 771 * deallocate obj out from under us. Any collapse 772 * will re-check the situation. We must not block 773 * until we are able to collapse. 774 * 775 * Bump temp's ref_count to avoid an unwanted 776 * degenerate recursion (can't call 777 * vm_object_reference_locked() because it asserts 778 * that CHAINLOCK is not set). 779 */ 780 temp->ref_count++; 781 KKASSERT(temp->ref_count > 1); 782 783 /* 784 * Collapse temp, then deallocate the extra ref 785 * formally. 786 */ 787 vm_object_collapse(temp, &dlist); 788 vm_object_chain_release(temp); 789 if (must_drop) { 790 vm_object_lock_swap(); 791 vm_object_drop(object); 792 } 793 object = temp; 794 must_drop = 1; 795 continue; 796 } 797 798 /* 799 * Drop the ref and handle termination on the 1->0 transition. 800 * We may have blocked above so we have to recheck. 801 */ 802 skip: 803 KKASSERT(object->ref_count != 0); 804 if (object->ref_count >= 2) { 805 object->ref_count--; 806 break; 807 } 808 KKASSERT(object->ref_count == 1); 809 810 /* 811 * 1->0 transition. Chain through the backing_object. 812 * Maintain the ref until we've located the backing object, 813 * then re-check. 814 */ 815 while ((temp = object->backing_object) != NULL) { 816 vm_object_hold(temp); 817 if (temp == object->backing_object) 818 break; 819 vm_object_drop(temp); 820 } 821 822 /* 823 * 1->0 transition verified, retry if ref_count is no longer 824 * 1. Otherwise disconnect the backing_object (temp) and 825 * clean up. 826 */ 827 if (object->ref_count != 1) { 828 vm_object_drop(temp); 829 continue; 830 } 831 832 /* 833 * It shouldn't be possible for the object to be chain locked 834 * if we're removing the last ref on it. 835 */ 836 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0); 837 838 if (temp) { 839 LIST_REMOVE(object, shadow_list); 840 temp->shadow_count--; 841 temp->generation++; 842 object->backing_object = NULL; 843 } 844 845 --object->ref_count; 846 if ((object->flags & OBJ_DEAD) == 0) 847 vm_object_terminate(object); 848 if (must_drop && temp) 849 vm_object_lock_swap(); 850 if (must_drop) 851 vm_object_drop(object); 852 object = temp; 853 must_drop = 1; 854 } 855 if (must_drop && object) 856 vm_object_drop(object); 857 858 /* 859 * Additional tail recursion on dlist. Avoid a recursion. Objects 860 * on the dlist have a hold count but are not locked. 861 */ 862 if ((dtmp = dlist) != NULL) { 863 dlist = dtmp->next; 864 object = dtmp->object; 865 kfree(dtmp, M_TEMP); 866 867 vm_object_lock(object); /* already held, add lock */ 868 must_drop = 1; /* and we're responsible for it */ 869 goto again; 870 } 871 } 872 873 /* 874 * Destroy the specified object, freeing up related resources. 875 * 876 * The object must have zero references. 877 * 878 * The object must held. The caller is responsible for dropping the object 879 * after terminate returns. Terminate does NOT drop the object. 880 */ 881 static int vm_object_terminate_callback(vm_page_t p, void *data); 882 883 void 884 vm_object_terminate(vm_object_t object) 885 { 886 /* 887 * Make sure no one uses us. Once we set OBJ_DEAD we should be 888 * able to safely block. 889 */ 890 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 891 KKASSERT((object->flags & OBJ_DEAD) == 0); 892 vm_object_set_flag(object, OBJ_DEAD); 893 894 /* 895 * Wait for the pageout daemon to be done with the object 896 */ 897 vm_object_pip_wait(object, "objtrm1"); 898 899 KASSERT(!object->paging_in_progress, 900 ("vm_object_terminate: pageout in progress")); 901 902 /* 903 * Clean and free the pages, as appropriate. All references to the 904 * object are gone, so we don't need to lock it. 905 */ 906 if (object->type == OBJT_VNODE) { 907 struct vnode *vp; 908 909 /* 910 * Clean pages and flush buffers. 911 * 912 * NOTE! TMPFS buffer flushes do not typically flush the 913 * actual page to swap as this would be highly 914 * inefficient, and normal filesystems usually wrap 915 * page flushes with buffer cache buffers. 916 * 917 * To deal with this we have to call vinvalbuf() both 918 * before and after the vm_object_page_clean(). 919 */ 920 vp = (struct vnode *) object->handle; 921 vinvalbuf(vp, V_SAVE, 0, 0); 922 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 923 vinvalbuf(vp, V_SAVE, 0, 0); 924 } 925 926 /* 927 * Wait for any I/O to complete, after which there had better not 928 * be any references left on the object. 929 */ 930 vm_object_pip_wait(object, "objtrm2"); 931 932 if (object->ref_count != 0) { 933 panic("vm_object_terminate: object with references, " 934 "ref_count=%d", object->ref_count); 935 } 936 937 /* 938 * Cleanup any shared pmaps associated with this object. 939 */ 940 pmap_object_free(object); 941 942 /* 943 * Now free any remaining pages. For internal objects, this also 944 * removes them from paging queues. Don't free wired pages, just 945 * remove them from the object. 946 */ 947 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 948 vm_object_terminate_callback, NULL); 949 950 /* 951 * Let the pager know object is dead. 952 */ 953 vm_pager_deallocate(object); 954 955 /* 956 * Wait for the object hold count to hit 1, clean out pages as 957 * we go. vmobj_token interlocks any race conditions that might 958 * pick the object up from the vm_object_list after we have cleared 959 * rb_memq. 960 */ 961 for (;;) { 962 if (RB_ROOT(&object->rb_memq) == NULL) 963 break; 964 kprintf("vm_object_terminate: Warning, object %p " 965 "still has %d pages\n", 966 object, object->resident_page_count); 967 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 968 vm_object_terminate_callback, NULL); 969 } 970 971 /* 972 * There had better not be any pages left 973 */ 974 KKASSERT(object->resident_page_count == 0); 975 976 /* 977 * Remove the object from the global object list. 978 */ 979 lwkt_gettoken(&vmobj_token); 980 TAILQ_REMOVE(&vm_object_list, object, object_list); 981 vm_object_count--; 982 lwkt_reltoken(&vmobj_token); 983 vm_object_dead_wakeup(object); 984 985 if (object->ref_count != 0) { 986 panic("vm_object_terminate2: object with references, " 987 "ref_count=%d", object->ref_count); 988 } 989 990 /* 991 * NOTE: The object hold_count is at least 1, so we cannot zfree() 992 * the object here. See vm_object_drop(). 993 */ 994 } 995 996 /* 997 * The caller must hold the object. 998 */ 999 static int 1000 vm_object_terminate_callback(vm_page_t p, void *data __unused) 1001 { 1002 vm_object_t object; 1003 1004 object = p->object; 1005 vm_page_busy_wait(p, TRUE, "vmpgtrm"); 1006 if (object != p->object) { 1007 kprintf("vm_object_terminate: Warning: Encountered " 1008 "busied page %p on queue %d\n", p, p->queue); 1009 vm_page_wakeup(p); 1010 } else if (p->wire_count == 0) { 1011 /* 1012 * NOTE: p->dirty and PG_NEED_COMMIT are ignored. 1013 */ 1014 vm_page_free(p); 1015 mycpu->gd_cnt.v_pfree++; 1016 } else { 1017 if (p->queue != PQ_NONE) 1018 kprintf("vm_object_terminate: Warning: Encountered " 1019 "wired page %p on queue %d\n", p, p->queue); 1020 vm_page_remove(p); 1021 vm_page_wakeup(p); 1022 } 1023 lwkt_yield(); 1024 return(0); 1025 } 1026 1027 /* 1028 * The object is dead but still has an object<->pager association. Sleep 1029 * and return. The caller typically retests the association in a loop. 1030 * 1031 * The caller must hold the object. 1032 */ 1033 void 1034 vm_object_dead_sleep(vm_object_t object, const char *wmesg) 1035 { 1036 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1037 if (object->handle) { 1038 vm_object_set_flag(object, OBJ_DEADWNT); 1039 tsleep(object, 0, wmesg, 0); 1040 /* object may be invalid after this point */ 1041 } 1042 } 1043 1044 /* 1045 * Wakeup anyone waiting for the object<->pager disassociation on 1046 * a dead object. 1047 * 1048 * The caller must hold the object. 1049 */ 1050 void 1051 vm_object_dead_wakeup(vm_object_t object) 1052 { 1053 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1054 if (object->flags & OBJ_DEADWNT) { 1055 vm_object_clear_flag(object, OBJ_DEADWNT); 1056 wakeup(object); 1057 } 1058 } 1059 1060 /* 1061 * Clean all dirty pages in the specified range of object. Leaves page 1062 * on whatever queue it is currently on. If NOSYNC is set then do not 1063 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 1064 * leaving the object dirty. 1065 * 1066 * When stuffing pages asynchronously, allow clustering. XXX we need a 1067 * synchronous clustering mode implementation. 1068 * 1069 * Odd semantics: if start == end, we clean everything. 1070 * 1071 * The object must be locked? XXX 1072 */ 1073 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 1074 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 1075 1076 void 1077 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1078 int flags) 1079 { 1080 struct rb_vm_page_scan_info info; 1081 struct vnode *vp; 1082 int wholescan; 1083 int pagerflags; 1084 int generation; 1085 1086 vm_object_hold(object); 1087 if (object->type != OBJT_VNODE || 1088 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 1089 vm_object_drop(object); 1090 return; 1091 } 1092 1093 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 1094 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1095 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 1096 1097 vp = object->handle; 1098 1099 /* 1100 * Interlock other major object operations. This allows us to 1101 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 1102 */ 1103 vm_object_set_flag(object, OBJ_CLEANING); 1104 1105 /* 1106 * Handle 'entire object' case 1107 */ 1108 info.start_pindex = start; 1109 if (end == 0) { 1110 info.end_pindex = object->size - 1; 1111 } else { 1112 info.end_pindex = end - 1; 1113 } 1114 wholescan = (start == 0 && info.end_pindex == object->size - 1); 1115 info.limit = flags; 1116 info.pagerflags = pagerflags; 1117 info.object = object; 1118 1119 /* 1120 * If cleaning the entire object do a pass to mark the pages read-only. 1121 * If everything worked out ok, clear OBJ_WRITEABLE and 1122 * OBJ_MIGHTBEDIRTY. 1123 */ 1124 if (wholescan) { 1125 info.error = 0; 1126 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1127 vm_object_page_clean_pass1, &info); 1128 if (info.error == 0) { 1129 vm_object_clear_flag(object, 1130 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1131 if (object->type == OBJT_VNODE && 1132 (vp = (struct vnode *)object->handle) != NULL) { 1133 if (vp->v_flag & VOBJDIRTY) 1134 vclrflags(vp, VOBJDIRTY); 1135 } 1136 } 1137 } 1138 1139 /* 1140 * Do a pass to clean all the dirty pages we find. 1141 */ 1142 do { 1143 info.error = 0; 1144 generation = object->generation; 1145 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1146 vm_object_page_clean_pass2, &info); 1147 } while (info.error || generation != object->generation); 1148 1149 vm_object_clear_flag(object, OBJ_CLEANING); 1150 vm_object_drop(object); 1151 } 1152 1153 /* 1154 * The caller must hold the object. 1155 */ 1156 static 1157 int 1158 vm_object_page_clean_pass1(struct vm_page *p, void *data) 1159 { 1160 struct rb_vm_page_scan_info *info = data; 1161 1162 vm_page_flag_set(p, PG_CLEANCHK); 1163 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1164 info->error = 1; 1165 } else if (vm_page_busy_try(p, FALSE) == 0) { 1166 vm_page_protect(p, VM_PROT_READ); /* must not block */ 1167 vm_page_wakeup(p); 1168 } else { 1169 info->error = 1; 1170 } 1171 lwkt_yield(); 1172 return(0); 1173 } 1174 1175 /* 1176 * The caller must hold the object 1177 */ 1178 static 1179 int 1180 vm_object_page_clean_pass2(struct vm_page *p, void *data) 1181 { 1182 struct rb_vm_page_scan_info *info = data; 1183 int generation; 1184 1185 /* 1186 * Do not mess with pages that were inserted after we started 1187 * the cleaning pass. 1188 */ 1189 if ((p->flags & PG_CLEANCHK) == 0) 1190 goto done; 1191 1192 generation = info->object->generation; 1193 vm_page_busy_wait(p, TRUE, "vpcwai"); 1194 if (p->object != info->object || 1195 info->object->generation != generation) { 1196 info->error = 1; 1197 vm_page_wakeup(p); 1198 goto done; 1199 } 1200 1201 /* 1202 * Before wasting time traversing the pmaps, check for trivial 1203 * cases where the page cannot be dirty. 1204 */ 1205 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 1206 KKASSERT((p->dirty & p->valid) == 0 && 1207 (p->flags & PG_NEED_COMMIT) == 0); 1208 vm_page_wakeup(p); 1209 goto done; 1210 } 1211 1212 /* 1213 * Check whether the page is dirty or not. The page has been set 1214 * to be read-only so the check will not race a user dirtying the 1215 * page. 1216 */ 1217 vm_page_test_dirty(p); 1218 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) { 1219 vm_page_flag_clear(p, PG_CLEANCHK); 1220 vm_page_wakeup(p); 1221 goto done; 1222 } 1223 1224 /* 1225 * If we have been asked to skip nosync pages and this is a 1226 * nosync page, skip it. Note that the object flags were 1227 * not cleared in this case (because pass1 will have returned an 1228 * error), so we do not have to set them. 1229 */ 1230 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1231 vm_page_flag_clear(p, PG_CLEANCHK); 1232 vm_page_wakeup(p); 1233 goto done; 1234 } 1235 1236 /* 1237 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 1238 * the pages that get successfully flushed. Set info->error if 1239 * we raced an object modification. 1240 */ 1241 vm_object_page_collect_flush(info->object, p, info->pagerflags); 1242 vm_wait_nominal(); 1243 done: 1244 lwkt_yield(); 1245 return(0); 1246 } 1247 1248 /* 1249 * Collect the specified page and nearby pages and flush them out. 1250 * The number of pages flushed is returned. The passed page is busied 1251 * by the caller and we are responsible for its disposition. 1252 * 1253 * The caller must hold the object. 1254 */ 1255 static void 1256 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 1257 { 1258 int runlen; 1259 int error; 1260 int maxf; 1261 int chkb; 1262 int maxb; 1263 int i; 1264 vm_pindex_t pi; 1265 vm_page_t maf[vm_pageout_page_count]; 1266 vm_page_t mab[vm_pageout_page_count]; 1267 vm_page_t ma[vm_pageout_page_count]; 1268 1269 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1270 1271 pi = p->pindex; 1272 1273 maxf = 0; 1274 for(i = 1; i < vm_pageout_page_count; i++) { 1275 vm_page_t tp; 1276 1277 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error); 1278 if (error) 1279 break; 1280 if (tp == NULL) 1281 break; 1282 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1283 (tp->flags & PG_CLEANCHK) == 0) { 1284 vm_page_wakeup(tp); 1285 break; 1286 } 1287 if ((tp->queue - tp->pc) == PQ_CACHE) { 1288 vm_page_flag_clear(tp, PG_CLEANCHK); 1289 vm_page_wakeup(tp); 1290 break; 1291 } 1292 vm_page_test_dirty(tp); 1293 if ((tp->dirty & tp->valid) == 0 && 1294 (tp->flags & PG_NEED_COMMIT) == 0) { 1295 vm_page_flag_clear(tp, PG_CLEANCHK); 1296 vm_page_wakeup(tp); 1297 break; 1298 } 1299 maf[i - 1] = tp; 1300 maxf++; 1301 } 1302 1303 maxb = 0; 1304 chkb = vm_pageout_page_count - maxf; 1305 /* 1306 * NOTE: chkb can be 0 1307 */ 1308 for(i = 1; chkb && i < chkb; i++) { 1309 vm_page_t tp; 1310 1311 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error); 1312 if (error) 1313 break; 1314 if (tp == NULL) 1315 break; 1316 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1317 (tp->flags & PG_CLEANCHK) == 0) { 1318 vm_page_wakeup(tp); 1319 break; 1320 } 1321 if ((tp->queue - tp->pc) == PQ_CACHE) { 1322 vm_page_flag_clear(tp, PG_CLEANCHK); 1323 vm_page_wakeup(tp); 1324 break; 1325 } 1326 vm_page_test_dirty(tp); 1327 if ((tp->dirty & tp->valid) == 0 && 1328 (tp->flags & PG_NEED_COMMIT) == 0) { 1329 vm_page_flag_clear(tp, PG_CLEANCHK); 1330 vm_page_wakeup(tp); 1331 break; 1332 } 1333 mab[i - 1] = tp; 1334 maxb++; 1335 } 1336 1337 /* 1338 * All pages in the maf[] and mab[] array are busied. 1339 */ 1340 for (i = 0; i < maxb; i++) { 1341 int index = (maxb - i) - 1; 1342 ma[index] = mab[i]; 1343 vm_page_flag_clear(ma[index], PG_CLEANCHK); 1344 } 1345 vm_page_flag_clear(p, PG_CLEANCHK); 1346 ma[maxb] = p; 1347 for(i = 0; i < maxf; i++) { 1348 int index = (maxb + i) + 1; 1349 ma[index] = maf[i]; 1350 vm_page_flag_clear(ma[index], PG_CLEANCHK); 1351 } 1352 runlen = maxb + maxf + 1; 1353 1354 for (i = 0; i < runlen; i++) /* XXX need this any more? */ 1355 vm_page_hold(ma[i]); 1356 1357 vm_pageout_flush(ma, runlen, pagerflags); 1358 1359 for (i = 0; i < runlen; i++) /* XXX need this any more? */ 1360 vm_page_unhold(ma[i]); 1361 } 1362 1363 /* 1364 * Same as vm_object_pmap_copy, except range checking really 1365 * works, and is meant for small sections of an object. 1366 * 1367 * This code protects resident pages by making them read-only 1368 * and is typically called on a fork or split when a page 1369 * is converted to copy-on-write. 1370 * 1371 * NOTE: If the page is already at VM_PROT_NONE, calling 1372 * vm_page_protect will have no effect. 1373 */ 1374 void 1375 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1376 { 1377 vm_pindex_t idx; 1378 vm_page_t p; 1379 1380 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 1381 return; 1382 1383 vm_object_hold(object); 1384 for (idx = start; idx < end; idx++) { 1385 p = vm_page_lookup(object, idx); 1386 if (p == NULL) 1387 continue; 1388 vm_page_protect(p, VM_PROT_READ); 1389 } 1390 vm_object_drop(object); 1391 } 1392 1393 /* 1394 * Removes all physical pages in the specified object range from all 1395 * physical maps. 1396 * 1397 * The object must *not* be locked. 1398 */ 1399 1400 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 1401 1402 void 1403 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1404 { 1405 struct rb_vm_page_scan_info info; 1406 1407 if (object == NULL) 1408 return; 1409 info.start_pindex = start; 1410 info.end_pindex = end - 1; 1411 1412 vm_object_hold(object); 1413 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1414 vm_object_pmap_remove_callback, &info); 1415 if (start == 0 && end == object->size) 1416 vm_object_clear_flag(object, OBJ_WRITEABLE); 1417 vm_object_drop(object); 1418 } 1419 1420 /* 1421 * The caller must hold the object 1422 */ 1423 static int 1424 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 1425 { 1426 vm_page_protect(p, VM_PROT_NONE); 1427 return(0); 1428 } 1429 1430 /* 1431 * Implements the madvise function at the object/page level. 1432 * 1433 * MADV_WILLNEED (any object) 1434 * 1435 * Activate the specified pages if they are resident. 1436 * 1437 * MADV_DONTNEED (any object) 1438 * 1439 * Deactivate the specified pages if they are resident. 1440 * 1441 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 1442 * 1443 * Deactivate and clean the specified pages if they are 1444 * resident. This permits the process to reuse the pages 1445 * without faulting or the kernel to reclaim the pages 1446 * without I/O. 1447 * 1448 * No requirements. 1449 */ 1450 void 1451 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1452 { 1453 vm_pindex_t end, tpindex; 1454 vm_object_t tobject; 1455 vm_object_t xobj; 1456 vm_page_t m; 1457 int error; 1458 1459 if (object == NULL) 1460 return; 1461 1462 end = pindex + count; 1463 1464 vm_object_hold(object); 1465 tobject = object; 1466 1467 /* 1468 * Locate and adjust resident pages 1469 */ 1470 for (; pindex < end; pindex += 1) { 1471 relookup: 1472 if (tobject != object) 1473 vm_object_drop(tobject); 1474 tobject = object; 1475 tpindex = pindex; 1476 shadowlookup: 1477 /* 1478 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1479 * and those pages must be OBJ_ONEMAPPING. 1480 */ 1481 if (advise == MADV_FREE) { 1482 if ((tobject->type != OBJT_DEFAULT && 1483 tobject->type != OBJT_SWAP) || 1484 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1485 continue; 1486 } 1487 } 1488 1489 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error); 1490 1491 if (error) { 1492 vm_page_sleep_busy(m, TRUE, "madvpo"); 1493 goto relookup; 1494 } 1495 if (m == NULL) { 1496 /* 1497 * There may be swap even if there is no backing page 1498 */ 1499 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1500 swap_pager_freespace(tobject, tpindex, 1); 1501 1502 /* 1503 * next object 1504 */ 1505 while ((xobj = tobject->backing_object) != NULL) { 1506 KKASSERT(xobj != object); 1507 vm_object_hold(xobj); 1508 if (xobj == tobject->backing_object) 1509 break; 1510 vm_object_drop(xobj); 1511 } 1512 if (xobj == NULL) 1513 continue; 1514 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1515 if (tobject != object) { 1516 vm_object_lock_swap(); 1517 vm_object_drop(tobject); 1518 } 1519 tobject = xobj; 1520 goto shadowlookup; 1521 } 1522 1523 /* 1524 * If the page is not in a normal active state, we skip it. 1525 * If the page is not managed there are no page queues to 1526 * mess with. Things can break if we mess with pages in 1527 * any of the below states. 1528 */ 1529 if (m->wire_count || 1530 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 1531 m->valid != VM_PAGE_BITS_ALL 1532 ) { 1533 vm_page_wakeup(m); 1534 continue; 1535 } 1536 1537 /* 1538 * Theoretically once a page is known not to be busy, an 1539 * interrupt cannot come along and rip it out from under us. 1540 */ 1541 1542 if (advise == MADV_WILLNEED) { 1543 vm_page_activate(m); 1544 } else if (advise == MADV_DONTNEED) { 1545 vm_page_dontneed(m); 1546 } else if (advise == MADV_FREE) { 1547 /* 1548 * Mark the page clean. This will allow the page 1549 * to be freed up by the system. However, such pages 1550 * are often reused quickly by malloc()/free() 1551 * so we do not do anything that would cause 1552 * a page fault if we can help it. 1553 * 1554 * Specifically, we do not try to actually free 1555 * the page now nor do we try to put it in the 1556 * cache (which would cause a page fault on reuse). 1557 * 1558 * But we do make the page is freeable as we 1559 * can without actually taking the step of unmapping 1560 * it. 1561 */ 1562 pmap_clear_modify(m); 1563 m->dirty = 0; 1564 m->act_count = 0; 1565 vm_page_dontneed(m); 1566 if (tobject->type == OBJT_SWAP) 1567 swap_pager_freespace(tobject, tpindex, 1); 1568 } 1569 vm_page_wakeup(m); 1570 } 1571 if (tobject != object) 1572 vm_object_drop(tobject); 1573 vm_object_drop(object); 1574 } 1575 1576 /* 1577 * Create a new object which is backed by the specified existing object 1578 * range. Replace the pointer and offset that was pointing at the existing 1579 * object with the pointer/offset for the new object. 1580 * 1581 * No other requirements. 1582 */ 1583 void 1584 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length, 1585 int addref) 1586 { 1587 vm_object_t source; 1588 vm_object_t result; 1589 1590 source = *objectp; 1591 1592 /* 1593 * Don't create the new object if the old object isn't shared. 1594 * We have to chain wait before adding the reference to avoid 1595 * racing a collapse or deallocation. 1596 * 1597 * Add the additional ref to source here to avoid racing a later 1598 * collapse or deallocation. Clear the ONEMAPPING flag whether 1599 * addref is TRUE or not in this case because the original object 1600 * will be shadowed. 1601 */ 1602 if (source) { 1603 vm_object_hold(source); 1604 vm_object_chain_wait(source); 1605 if (source->ref_count == 1 && 1606 source->handle == NULL && 1607 (source->type == OBJT_DEFAULT || 1608 source->type == OBJT_SWAP)) { 1609 vm_object_drop(source); 1610 if (addref) { 1611 vm_object_reference_locked(source); 1612 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1613 } 1614 return; 1615 } 1616 vm_object_reference_locked(source); 1617 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1618 } 1619 1620 /* 1621 * Allocate a new object with the given length. The new object 1622 * is returned referenced but we may have to add another one. 1623 * If we are adding a second reference we must clear OBJ_ONEMAPPING. 1624 * (typically because the caller is about to clone a vm_map_entry). 1625 * 1626 * The source object currently has an extra reference to prevent 1627 * collapses into it while we mess with its shadow list, which 1628 * we will remove later in this routine. 1629 */ 1630 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL) 1631 panic("vm_object_shadow: no object for shadowing"); 1632 vm_object_hold(result); 1633 if (addref) { 1634 vm_object_reference_locked(result); 1635 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1636 } 1637 1638 /* 1639 * The new object shadows the source object. Chain wait before 1640 * adjusting shadow_count or the shadow list to avoid races. 1641 * 1642 * Try to optimize the result object's page color when shadowing 1643 * in order to maintain page coloring consistency in the combined 1644 * shadowed object. 1645 */ 1646 KKASSERT(result->backing_object == NULL); 1647 result->backing_object = source; 1648 if (source) { 1649 vm_object_chain_wait(source); 1650 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1651 source->shadow_count++; 1652 source->generation++; 1653 /* cpu localization twist */ 1654 result->pg_color = (int)(intptr_t)curthread; 1655 } 1656 1657 /* 1658 * Adjust the return storage. Drop the ref on source before 1659 * returning. 1660 */ 1661 result->backing_object_offset = *offset; 1662 vm_object_drop(result); 1663 *offset = 0; 1664 if (source) { 1665 vm_object_deallocate_locked(source); 1666 vm_object_drop(source); 1667 } 1668 1669 /* 1670 * Return the new things 1671 */ 1672 *objectp = result; 1673 } 1674 1675 #define OBSC_TEST_ALL_SHADOWED 0x0001 1676 #define OBSC_COLLAPSE_NOWAIT 0x0002 1677 #define OBSC_COLLAPSE_WAIT 0x0004 1678 1679 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1680 1681 /* 1682 * The caller must hold the object. 1683 */ 1684 static __inline int 1685 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op) 1686 { 1687 struct rb_vm_page_scan_info info; 1688 1689 vm_object_assert_held(object); 1690 vm_object_assert_held(backing_object); 1691 1692 KKASSERT(backing_object == object->backing_object); 1693 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1694 1695 /* 1696 * Initial conditions 1697 */ 1698 if (op & OBSC_TEST_ALL_SHADOWED) { 1699 /* 1700 * We do not want to have to test for the existence of 1701 * swap pages in the backing object. XXX but with the 1702 * new swapper this would be pretty easy to do. 1703 * 1704 * XXX what about anonymous MAP_SHARED memory that hasn't 1705 * been ZFOD faulted yet? If we do not test for this, the 1706 * shadow test may succeed! XXX 1707 */ 1708 if (backing_object->type != OBJT_DEFAULT) 1709 return(0); 1710 } 1711 if (op & OBSC_COLLAPSE_WAIT) { 1712 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1713 vm_object_set_flag(backing_object, OBJ_DEAD); 1714 lwkt_gettoken(&vmobj_token); 1715 TAILQ_REMOVE(&vm_object_list, backing_object, object_list); 1716 vm_object_count--; 1717 lwkt_reltoken(&vmobj_token); 1718 vm_object_dead_wakeup(backing_object); 1719 } 1720 1721 /* 1722 * Our scan. We have to retry if a negative error code is returned, 1723 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1724 * the scan had to be stopped because the parent does not completely 1725 * shadow the child. 1726 */ 1727 info.object = object; 1728 info.backing_object = backing_object; 1729 info.limit = op; 1730 do { 1731 info.error = 1; 1732 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1733 vm_object_backing_scan_callback, 1734 &info); 1735 } while (info.error < 0); 1736 1737 return(info.error); 1738 } 1739 1740 /* 1741 * The caller must hold the object. 1742 */ 1743 static int 1744 vm_object_backing_scan_callback(vm_page_t p, void *data) 1745 { 1746 struct rb_vm_page_scan_info *info = data; 1747 vm_object_t backing_object; 1748 vm_object_t object; 1749 vm_pindex_t pindex; 1750 vm_pindex_t new_pindex; 1751 vm_pindex_t backing_offset_index; 1752 int op; 1753 1754 pindex = p->pindex; 1755 new_pindex = pindex - info->backing_offset_index; 1756 op = info->limit; 1757 object = info->object; 1758 backing_object = info->backing_object; 1759 backing_offset_index = info->backing_offset_index; 1760 1761 if (op & OBSC_TEST_ALL_SHADOWED) { 1762 vm_page_t pp; 1763 1764 /* 1765 * Ignore pages outside the parent object's range 1766 * and outside the parent object's mapping of the 1767 * backing object. 1768 * 1769 * note that we do not busy the backing object's 1770 * page. 1771 */ 1772 if (pindex < backing_offset_index || 1773 new_pindex >= object->size 1774 ) { 1775 return(0); 1776 } 1777 1778 /* 1779 * See if the parent has the page or if the parent's 1780 * object pager has the page. If the parent has the 1781 * page but the page is not valid, the parent's 1782 * object pager must have the page. 1783 * 1784 * If this fails, the parent does not completely shadow 1785 * the object and we might as well give up now. 1786 */ 1787 pp = vm_page_lookup(object, new_pindex); 1788 if ((pp == NULL || pp->valid == 0) && 1789 !vm_pager_has_page(object, new_pindex) 1790 ) { 1791 info->error = 0; /* problemo */ 1792 return(-1); /* stop the scan */ 1793 } 1794 } 1795 1796 /* 1797 * Check for busy page. Note that we may have lost (p) when we 1798 * possibly blocked above. 1799 */ 1800 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1801 vm_page_t pp; 1802 1803 if (vm_page_busy_try(p, TRUE)) { 1804 if (op & OBSC_COLLAPSE_NOWAIT) { 1805 return(0); 1806 } else { 1807 /* 1808 * If we slept, anything could have 1809 * happened. Ask that the scan be restarted. 1810 * 1811 * Since the object is marked dead, the 1812 * backing offset should not have changed. 1813 */ 1814 vm_page_sleep_busy(p, TRUE, "vmocol"); 1815 info->error = -1; 1816 return(-1); 1817 } 1818 } 1819 1820 /* 1821 * If (p) is no longer valid restart the scan. 1822 */ 1823 if (p->object != backing_object || p->pindex != pindex) { 1824 kprintf("vm_object_backing_scan: Warning: page " 1825 "%p ripped out from under us\n", p); 1826 vm_page_wakeup(p); 1827 info->error = -1; 1828 return(-1); 1829 } 1830 1831 if (op & OBSC_COLLAPSE_NOWAIT) { 1832 if (p->valid == 0 || 1833 p->wire_count || 1834 (p->flags & PG_NEED_COMMIT)) { 1835 vm_page_wakeup(p); 1836 return(0); 1837 } 1838 } else { 1839 /* XXX what if p->valid == 0 , hold_count, etc? */ 1840 } 1841 1842 KASSERT( 1843 p->object == backing_object, 1844 ("vm_object_qcollapse(): object mismatch") 1845 ); 1846 1847 /* 1848 * Destroy any associated swap 1849 */ 1850 if (backing_object->type == OBJT_SWAP) 1851 swap_pager_freespace(backing_object, p->pindex, 1); 1852 1853 if ( 1854 p->pindex < backing_offset_index || 1855 new_pindex >= object->size 1856 ) { 1857 /* 1858 * Page is out of the parent object's range, we 1859 * can simply destroy it. 1860 */ 1861 vm_page_protect(p, VM_PROT_NONE); 1862 vm_page_free(p); 1863 return(0); 1864 } 1865 1866 pp = vm_page_lookup(object, new_pindex); 1867 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 1868 /* 1869 * page already exists in parent OR swap exists 1870 * for this location in the parent. Destroy 1871 * the original page from the backing object. 1872 * 1873 * Leave the parent's page alone 1874 */ 1875 vm_page_protect(p, VM_PROT_NONE); 1876 vm_page_free(p); 1877 return(0); 1878 } 1879 1880 /* 1881 * Page does not exist in parent, rename the 1882 * page from the backing object to the main object. 1883 * 1884 * If the page was mapped to a process, it can remain 1885 * mapped through the rename. 1886 */ 1887 if ((p->queue - p->pc) == PQ_CACHE) 1888 vm_page_deactivate(p); 1889 1890 vm_page_rename(p, object, new_pindex); 1891 vm_page_wakeup(p); 1892 /* page automatically made dirty by rename */ 1893 } 1894 return(0); 1895 } 1896 1897 /* 1898 * This version of collapse allows the operation to occur earlier and 1899 * when paging_in_progress is true for an object... This is not a complete 1900 * operation, but should plug 99.9% of the rest of the leaks. 1901 * 1902 * The caller must hold the object and backing_object and both must be 1903 * chainlocked. 1904 * 1905 * (only called from vm_object_collapse) 1906 */ 1907 static void 1908 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object) 1909 { 1910 if (backing_object->ref_count == 1) { 1911 backing_object->ref_count += 2; 1912 vm_object_backing_scan(object, backing_object, 1913 OBSC_COLLAPSE_NOWAIT); 1914 backing_object->ref_count -= 2; 1915 } 1916 } 1917 1918 /* 1919 * Collapse an object with the object backing it. Pages in the backing 1920 * object are moved into the parent, and the backing object is deallocated. 1921 * Any conflict is resolved in favor of the parent's existing pages. 1922 * 1923 * object must be held and chain-locked on call. 1924 * 1925 * The caller must have an extra ref on object to prevent a race from 1926 * destroying it during the collapse. 1927 */ 1928 void 1929 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp) 1930 { 1931 struct vm_object_dealloc_list *dlist = NULL; 1932 vm_object_t backing_object; 1933 1934 /* 1935 * Only one thread is attempting a collapse at any given moment. 1936 * There are few restrictions for (object) that callers of this 1937 * function check so reentrancy is likely. 1938 */ 1939 KKASSERT(object != NULL); 1940 vm_object_assert_held(object); 1941 KKASSERT(object->flags & OBJ_CHAINLOCK); 1942 1943 for (;;) { 1944 vm_object_t bbobj; 1945 int dodealloc; 1946 1947 /* 1948 * We have to hold the backing object, check races. 1949 */ 1950 while ((backing_object = object->backing_object) != NULL) { 1951 vm_object_hold(backing_object); 1952 if (backing_object == object->backing_object) 1953 break; 1954 vm_object_drop(backing_object); 1955 } 1956 1957 /* 1958 * No backing object? Nothing to collapse then. 1959 */ 1960 if (backing_object == NULL) 1961 break; 1962 1963 /* 1964 * You can't collapse with a non-default/non-swap object. 1965 */ 1966 if (backing_object->type != OBJT_DEFAULT && 1967 backing_object->type != OBJT_SWAP) { 1968 vm_object_drop(backing_object); 1969 backing_object = NULL; 1970 break; 1971 } 1972 1973 /* 1974 * Chain-lock the backing object too because if we 1975 * successfully merge its pages into the top object we 1976 * will collapse backing_object->backing_object as the 1977 * new backing_object. Re-check that it is still our 1978 * backing object. 1979 */ 1980 vm_object_chain_acquire(backing_object); 1981 if (backing_object != object->backing_object) { 1982 vm_object_chain_release(backing_object); 1983 vm_object_drop(backing_object); 1984 continue; 1985 } 1986 1987 /* 1988 * we check the backing object first, because it is most likely 1989 * not collapsable. 1990 */ 1991 if (backing_object->handle != NULL || 1992 (backing_object->type != OBJT_DEFAULT && 1993 backing_object->type != OBJT_SWAP) || 1994 (backing_object->flags & OBJ_DEAD) || 1995 object->handle != NULL || 1996 (object->type != OBJT_DEFAULT && 1997 object->type != OBJT_SWAP) || 1998 (object->flags & OBJ_DEAD)) { 1999 break; 2000 } 2001 2002 /* 2003 * If paging is in progress we can't do a normal collapse. 2004 */ 2005 if ( 2006 object->paging_in_progress != 0 || 2007 backing_object->paging_in_progress != 0 2008 ) { 2009 vm_object_qcollapse(object, backing_object); 2010 break; 2011 } 2012 2013 /* 2014 * We know that we can either collapse the backing object (if 2015 * the parent is the only reference to it) or (perhaps) have 2016 * the parent bypass the object if the parent happens to shadow 2017 * all the resident pages in the entire backing object. 2018 * 2019 * This is ignoring pager-backed pages such as swap pages. 2020 * vm_object_backing_scan fails the shadowing test in this 2021 * case. 2022 */ 2023 if (backing_object->ref_count == 1) { 2024 /* 2025 * If there is exactly one reference to the backing 2026 * object, we can collapse it into the parent. 2027 */ 2028 KKASSERT(object->backing_object == backing_object); 2029 vm_object_backing_scan(object, backing_object, 2030 OBSC_COLLAPSE_WAIT); 2031 2032 /* 2033 * Move the pager from backing_object to object. 2034 */ 2035 if (backing_object->type == OBJT_SWAP) { 2036 vm_object_pip_add(backing_object, 1); 2037 2038 /* 2039 * scrap the paging_offset junk and do a 2040 * discrete copy. This also removes major 2041 * assumptions about how the swap-pager 2042 * works from where it doesn't belong. The 2043 * new swapper is able to optimize the 2044 * destroy-source case. 2045 */ 2046 vm_object_pip_add(object, 1); 2047 swap_pager_copy(backing_object, object, 2048 OFF_TO_IDX(object->backing_object_offset), 2049 TRUE); 2050 vm_object_pip_wakeup(object); 2051 vm_object_pip_wakeup(backing_object); 2052 } 2053 2054 /* 2055 * Object now shadows whatever backing_object did. 2056 * Remove object from backing_object's shadow_list. 2057 */ 2058 LIST_REMOVE(object, shadow_list); 2059 KKASSERT(object->backing_object == backing_object); 2060 backing_object->shadow_count--; 2061 backing_object->generation++; 2062 2063 /* 2064 * backing_object->backing_object moves from within 2065 * backing_object to within object. 2066 */ 2067 while ((bbobj = backing_object->backing_object) != NULL) { 2068 vm_object_hold(bbobj); 2069 if (bbobj == backing_object->backing_object) 2070 break; 2071 vm_object_drop(bbobj); 2072 } 2073 if (bbobj) { 2074 LIST_REMOVE(backing_object, shadow_list); 2075 bbobj->shadow_count--; 2076 bbobj->generation++; 2077 backing_object->backing_object = NULL; 2078 } 2079 object->backing_object = bbobj; 2080 if (bbobj) { 2081 LIST_INSERT_HEAD(&bbobj->shadow_head, 2082 object, shadow_list); 2083 bbobj->shadow_count++; 2084 bbobj->generation++; 2085 } 2086 2087 object->backing_object_offset += 2088 backing_object->backing_object_offset; 2089 2090 vm_object_drop(bbobj); 2091 2092 /* 2093 * Discard the old backing_object. Nothing should be 2094 * able to ref it, other than a vm_map_split(), 2095 * and vm_map_split() will stall on our chain lock. 2096 * And we control the parent so it shouldn't be 2097 * possible for it to go away either. 2098 * 2099 * Since the backing object has no pages, no pager 2100 * left, and no object references within it, all 2101 * that is necessary is to dispose of it. 2102 */ 2103 KASSERT(backing_object->ref_count == 1, 2104 ("backing_object %p was somehow " 2105 "re-referenced during collapse!", 2106 backing_object)); 2107 KASSERT(RB_EMPTY(&backing_object->rb_memq), 2108 ("backing_object %p somehow has left " 2109 "over pages during collapse!", 2110 backing_object)); 2111 2112 /* 2113 * The object can be destroyed. 2114 * 2115 * XXX just fall through and dodealloc instead 2116 * of forcing destruction? 2117 */ 2118 --backing_object->ref_count; 2119 if ((backing_object->flags & OBJ_DEAD) == 0) 2120 vm_object_terminate(backing_object); 2121 object_collapses++; 2122 dodealloc = 0; 2123 } else { 2124 /* 2125 * If we do not entirely shadow the backing object, 2126 * there is nothing we can do so we give up. 2127 */ 2128 if (vm_object_backing_scan(object, backing_object, 2129 OBSC_TEST_ALL_SHADOWED) == 0) { 2130 break; 2131 } 2132 2133 /* 2134 * bbobj is backing_object->backing_object. Since 2135 * object completely shadows backing_object we can 2136 * bypass it and become backed by bbobj instead. 2137 */ 2138 while ((bbobj = backing_object->backing_object) != NULL) { 2139 vm_object_hold(bbobj); 2140 if (bbobj == backing_object->backing_object) 2141 break; 2142 vm_object_drop(bbobj); 2143 } 2144 2145 /* 2146 * Make object shadow bbobj instead of backing_object. 2147 * Remove object from backing_object's shadow list. 2148 * 2149 * Deallocating backing_object will not remove 2150 * it, since its reference count is at least 2. 2151 */ 2152 KKASSERT(object->backing_object == backing_object); 2153 LIST_REMOVE(object, shadow_list); 2154 backing_object->shadow_count--; 2155 backing_object->generation++; 2156 2157 /* 2158 * Add a ref to bbobj, bbobj now shadows object. 2159 * 2160 * NOTE: backing_object->backing_object still points 2161 * to bbobj. That relationship remains intact 2162 * because backing_object has > 1 ref, so 2163 * someone else is pointing to it (hence why 2164 * we can't collapse it into object and can 2165 * only handle the all-shadowed bypass case). 2166 */ 2167 if (bbobj) { 2168 vm_object_chain_wait(bbobj); 2169 vm_object_reference_locked(bbobj); 2170 LIST_INSERT_HEAD(&bbobj->shadow_head, 2171 object, shadow_list); 2172 bbobj->shadow_count++; 2173 bbobj->generation++; 2174 object->backing_object_offset += 2175 backing_object->backing_object_offset; 2176 object->backing_object = bbobj; 2177 vm_object_drop(bbobj); 2178 } else { 2179 object->backing_object = NULL; 2180 } 2181 2182 /* 2183 * Drop the reference count on backing_object. To 2184 * handle ref_count races properly we can't assume 2185 * that the ref_count is still at least 2 so we 2186 * have to actually call vm_object_deallocate() 2187 * (after clearing the chainlock). 2188 */ 2189 object_bypasses++; 2190 dodealloc = 1; 2191 } 2192 2193 /* 2194 * Ok, we want to loop on the new object->bbobj association, 2195 * possibly collapsing it further. However if dodealloc is 2196 * non-zero we have to deallocate the backing_object which 2197 * itself can potentially undergo a collapse, creating a 2198 * recursion depth issue with the LWKT token subsystem. 2199 * 2200 * In the case where we must deallocate the backing_object 2201 * it is possible now that the backing_object has a single 2202 * shadow count on some other object (not represented here 2203 * as yet), since it no longer shadows us. Thus when we 2204 * call vm_object_deallocate() it may attempt to collapse 2205 * itself into its remaining parent. 2206 */ 2207 if (dodealloc) { 2208 struct vm_object_dealloc_list *dtmp; 2209 2210 vm_object_chain_release(backing_object); 2211 vm_object_unlock(backing_object); 2212 /* backing_object remains held */ 2213 2214 /* 2215 * Auto-deallocation list for caller convenience. 2216 */ 2217 if (dlistp == NULL) 2218 dlistp = &dlist; 2219 2220 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK); 2221 dtmp->object = backing_object; 2222 dtmp->next = *dlistp; 2223 *dlistp = dtmp; 2224 } else { 2225 vm_object_chain_release(backing_object); 2226 vm_object_drop(backing_object); 2227 } 2228 /* backing_object = NULL; not needed */ 2229 /* loop */ 2230 } 2231 2232 /* 2233 * Clean up any left over backing_object 2234 */ 2235 if (backing_object) { 2236 vm_object_chain_release(backing_object); 2237 vm_object_drop(backing_object); 2238 } 2239 2240 /* 2241 * Clean up any auto-deallocation list. This is a convenience 2242 * for top-level callers so they don't have to pass &dlist. 2243 * Do not clean up any caller-passed dlistp, the caller will 2244 * do that. 2245 */ 2246 if (dlist) 2247 vm_object_deallocate_list(&dlist); 2248 2249 } 2250 2251 /* 2252 * vm_object_collapse() may collect additional objects in need of 2253 * deallocation. This routine deallocates these objects. The 2254 * deallocation itself can trigger additional collapses (which the 2255 * deallocate function takes care of). This procedure is used to 2256 * reduce procedural recursion since these vm_object shadow chains 2257 * can become quite long. 2258 */ 2259 void 2260 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp) 2261 { 2262 struct vm_object_dealloc_list *dlist; 2263 2264 while ((dlist = *dlistp) != NULL) { 2265 *dlistp = dlist->next; 2266 vm_object_lock(dlist->object); 2267 vm_object_deallocate_locked(dlist->object); 2268 vm_object_drop(dlist->object); 2269 kfree(dlist, M_TEMP); 2270 } 2271 } 2272 2273 /* 2274 * Removes all physical pages in the specified object range from the 2275 * object's list of pages. 2276 * 2277 * No requirements. 2278 */ 2279 static int vm_object_page_remove_callback(vm_page_t p, void *data); 2280 2281 void 2282 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2283 boolean_t clean_only) 2284 { 2285 struct rb_vm_page_scan_info info; 2286 int all; 2287 2288 /* 2289 * Degenerate cases and assertions 2290 */ 2291 vm_object_hold(object); 2292 if (object == NULL || 2293 (object->resident_page_count == 0 && object->swblock_count == 0)) { 2294 vm_object_drop(object); 2295 return; 2296 } 2297 KASSERT(object->type != OBJT_PHYS, 2298 ("attempt to remove pages from a physical object")); 2299 2300 /* 2301 * Indicate that paging is occuring on the object 2302 */ 2303 vm_object_pip_add(object, 1); 2304 2305 /* 2306 * Figure out the actual removal range and whether we are removing 2307 * the entire contents of the object or not. If removing the entire 2308 * contents, be sure to get all pages, even those that might be 2309 * beyond the end of the object. 2310 */ 2311 info.start_pindex = start; 2312 if (end == 0) 2313 info.end_pindex = (vm_pindex_t)-1; 2314 else 2315 info.end_pindex = end - 1; 2316 info.limit = clean_only; 2317 all = (start == 0 && info.end_pindex >= object->size - 1); 2318 2319 /* 2320 * Loop until we are sure we have gotten them all. 2321 */ 2322 do { 2323 info.error = 0; 2324 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2325 vm_object_page_remove_callback, &info); 2326 } while (info.error); 2327 2328 /* 2329 * Remove any related swap if throwing away pages, or for 2330 * non-swap objects (the swap is a clean copy in that case). 2331 */ 2332 if (object->type != OBJT_SWAP || clean_only == FALSE) { 2333 if (all) 2334 swap_pager_freespace_all(object); 2335 else 2336 swap_pager_freespace(object, info.start_pindex, 2337 info.end_pindex - info.start_pindex + 1); 2338 } 2339 2340 /* 2341 * Cleanup 2342 */ 2343 vm_object_pip_wakeup(object); 2344 vm_object_drop(object); 2345 } 2346 2347 /* 2348 * The caller must hold the object 2349 */ 2350 static int 2351 vm_object_page_remove_callback(vm_page_t p, void *data) 2352 { 2353 struct rb_vm_page_scan_info *info = data; 2354 2355 if (vm_page_busy_try(p, TRUE)) { 2356 vm_page_sleep_busy(p, TRUE, "vmopar"); 2357 info->error = 1; 2358 return(0); 2359 } 2360 2361 /* 2362 * Wired pages cannot be destroyed, but they can be invalidated 2363 * and we do so if clean_only (limit) is not set. 2364 * 2365 * WARNING! The page may be wired due to being part of a buffer 2366 * cache buffer, and the buffer might be marked B_CACHE. 2367 * This is fine as part of a truncation but VFSs must be 2368 * sure to fix the buffer up when re-extending the file. 2369 * 2370 * NOTE! PG_NEED_COMMIT is ignored. 2371 */ 2372 if (p->wire_count != 0) { 2373 vm_page_protect(p, VM_PROT_NONE); 2374 if (info->limit == 0) 2375 p->valid = 0; 2376 vm_page_wakeup(p); 2377 return(0); 2378 } 2379 2380 /* 2381 * limit is our clean_only flag. If set and the page is dirty or 2382 * requires a commit, do not free it. If set and the page is being 2383 * held by someone, do not free it. 2384 */ 2385 if (info->limit && p->valid) { 2386 vm_page_test_dirty(p); 2387 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) { 2388 vm_page_wakeup(p); 2389 return(0); 2390 } 2391 #if 0 2392 if (p->hold_count) { 2393 vm_page_wakeup(p); 2394 return(0); 2395 } 2396 #endif 2397 } 2398 2399 /* 2400 * Destroy the page 2401 */ 2402 vm_page_protect(p, VM_PROT_NONE); 2403 vm_page_free(p); 2404 return(0); 2405 } 2406 2407 /* 2408 * Coalesces two objects backing up adjoining regions of memory into a 2409 * single object. 2410 * 2411 * returns TRUE if objects were combined. 2412 * 2413 * NOTE: Only works at the moment if the second object is NULL - 2414 * if it's not, which object do we lock first? 2415 * 2416 * Parameters: 2417 * prev_object First object to coalesce 2418 * prev_offset Offset into prev_object 2419 * next_object Second object into coalesce 2420 * next_offset Offset into next_object 2421 * 2422 * prev_size Size of reference to prev_object 2423 * next_size Size of reference to next_object 2424 * 2425 * The caller does not need to hold (prev_object) but must have a stable 2426 * pointer to it (typically by holding the vm_map locked). 2427 */ 2428 boolean_t 2429 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 2430 vm_size_t prev_size, vm_size_t next_size) 2431 { 2432 vm_pindex_t next_pindex; 2433 2434 if (prev_object == NULL) 2435 return (TRUE); 2436 2437 vm_object_hold(prev_object); 2438 2439 if (prev_object->type != OBJT_DEFAULT && 2440 prev_object->type != OBJT_SWAP) { 2441 vm_object_drop(prev_object); 2442 return (FALSE); 2443 } 2444 2445 /* 2446 * Try to collapse the object first 2447 */ 2448 vm_object_chain_acquire(prev_object); 2449 vm_object_collapse(prev_object, NULL); 2450 2451 /* 2452 * Can't coalesce if: . more than one reference . paged out . shadows 2453 * another object . has a copy elsewhere (any of which mean that the 2454 * pages not mapped to prev_entry may be in use anyway) 2455 */ 2456 2457 if (prev_object->backing_object != NULL) { 2458 vm_object_chain_release(prev_object); 2459 vm_object_drop(prev_object); 2460 return (FALSE); 2461 } 2462 2463 prev_size >>= PAGE_SHIFT; 2464 next_size >>= PAGE_SHIFT; 2465 next_pindex = prev_pindex + prev_size; 2466 2467 if ((prev_object->ref_count > 1) && 2468 (prev_object->size != next_pindex)) { 2469 vm_object_chain_release(prev_object); 2470 vm_object_drop(prev_object); 2471 return (FALSE); 2472 } 2473 2474 /* 2475 * Remove any pages that may still be in the object from a previous 2476 * deallocation. 2477 */ 2478 if (next_pindex < prev_object->size) { 2479 vm_object_page_remove(prev_object, 2480 next_pindex, 2481 next_pindex + next_size, FALSE); 2482 if (prev_object->type == OBJT_SWAP) 2483 swap_pager_freespace(prev_object, 2484 next_pindex, next_size); 2485 } 2486 2487 /* 2488 * Extend the object if necessary. 2489 */ 2490 if (next_pindex + next_size > prev_object->size) 2491 prev_object->size = next_pindex + next_size; 2492 2493 vm_object_chain_release(prev_object); 2494 vm_object_drop(prev_object); 2495 return (TRUE); 2496 } 2497 2498 /* 2499 * Make the object writable and flag is being possibly dirty. 2500 * 2501 * The caller must hold the object. XXX called from vm_page_dirty(), 2502 * There is currently no requirement to hold the object. 2503 */ 2504 void 2505 vm_object_set_writeable_dirty(vm_object_t object) 2506 { 2507 struct vnode *vp; 2508 2509 /*vm_object_assert_held(object);*/ 2510 /* 2511 * Avoid contention in vm fault path by checking the state before 2512 * issuing an atomic op on it. 2513 */ 2514 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) != 2515 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) { 2516 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 2517 } 2518 if (object->type == OBJT_VNODE && 2519 (vp = (struct vnode *)object->handle) != NULL) { 2520 if ((vp->v_flag & VOBJDIRTY) == 0) { 2521 vsetflags(vp, VOBJDIRTY); 2522 } 2523 } 2524 } 2525 2526 #include "opt_ddb.h" 2527 #ifdef DDB 2528 #include <sys/kernel.h> 2529 2530 #include <sys/cons.h> 2531 2532 #include <ddb/ddb.h> 2533 2534 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 2535 vm_map_entry_t entry); 2536 static int vm_object_in_map (vm_object_t object); 2537 2538 /* 2539 * The caller must hold the object. 2540 */ 2541 static int 2542 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2543 { 2544 vm_map_t tmpm; 2545 vm_map_entry_t tmpe; 2546 vm_object_t obj, nobj; 2547 int entcount; 2548 2549 if (map == 0) 2550 return 0; 2551 if (entry == 0) { 2552 tmpe = map->header.next; 2553 entcount = map->nentries; 2554 while (entcount-- && (tmpe != &map->header)) { 2555 if( _vm_object_in_map(map, object, tmpe)) { 2556 return 1; 2557 } 2558 tmpe = tmpe->next; 2559 } 2560 return (0); 2561 } 2562 switch(entry->maptype) { 2563 case VM_MAPTYPE_SUBMAP: 2564 tmpm = entry->object.sub_map; 2565 tmpe = tmpm->header.next; 2566 entcount = tmpm->nentries; 2567 while (entcount-- && tmpe != &tmpm->header) { 2568 if( _vm_object_in_map(tmpm, object, tmpe)) { 2569 return 1; 2570 } 2571 tmpe = tmpe->next; 2572 } 2573 break; 2574 case VM_MAPTYPE_NORMAL: 2575 case VM_MAPTYPE_VPAGETABLE: 2576 obj = entry->object.vm_object; 2577 while (obj) { 2578 if (obj == object) { 2579 if (obj != entry->object.vm_object) 2580 vm_object_drop(obj); 2581 return 1; 2582 } 2583 while ((nobj = obj->backing_object) != NULL) { 2584 vm_object_hold(nobj); 2585 if (nobj == obj->backing_object) 2586 break; 2587 vm_object_drop(nobj); 2588 } 2589 if (obj != entry->object.vm_object) { 2590 if (nobj) 2591 vm_object_lock_swap(); 2592 vm_object_drop(obj); 2593 } 2594 obj = nobj; 2595 } 2596 break; 2597 default: 2598 break; 2599 } 2600 return 0; 2601 } 2602 2603 static int vm_object_in_map_callback(struct proc *p, void *data); 2604 2605 struct vm_object_in_map_info { 2606 vm_object_t object; 2607 int rv; 2608 }; 2609 2610 /* 2611 * Debugging only 2612 */ 2613 static int 2614 vm_object_in_map(vm_object_t object) 2615 { 2616 struct vm_object_in_map_info info; 2617 2618 info.rv = 0; 2619 info.object = object; 2620 2621 allproc_scan(vm_object_in_map_callback, &info); 2622 if (info.rv) 2623 return 1; 2624 if( _vm_object_in_map(&kernel_map, object, 0)) 2625 return 1; 2626 if( _vm_object_in_map(&pager_map, object, 0)) 2627 return 1; 2628 if( _vm_object_in_map(&buffer_map, object, 0)) 2629 return 1; 2630 return 0; 2631 } 2632 2633 /* 2634 * Debugging only 2635 */ 2636 static int 2637 vm_object_in_map_callback(struct proc *p, void *data) 2638 { 2639 struct vm_object_in_map_info *info = data; 2640 2641 if (p->p_vmspace) { 2642 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 2643 info->rv = 1; 2644 return -1; 2645 } 2646 } 2647 return (0); 2648 } 2649 2650 DB_SHOW_COMMAND(vmochk, vm_object_check) 2651 { 2652 vm_object_t object; 2653 2654 /* 2655 * make sure that internal objs are in a map somewhere 2656 * and none have zero ref counts. 2657 */ 2658 for (object = TAILQ_FIRST(&vm_object_list); 2659 object != NULL; 2660 object = TAILQ_NEXT(object, object_list)) { 2661 if (object->type == OBJT_MARKER) 2662 continue; 2663 if (object->handle == NULL && 2664 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2665 if (object->ref_count == 0) { 2666 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2667 (long)object->size); 2668 } 2669 if (!vm_object_in_map(object)) { 2670 db_printf( 2671 "vmochk: internal obj is not in a map: " 2672 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2673 object->ref_count, (u_long)object->size, 2674 (u_long)object->size, 2675 (void *)object->backing_object); 2676 } 2677 } 2678 } 2679 } 2680 2681 /* 2682 * Debugging only 2683 */ 2684 DB_SHOW_COMMAND(object, vm_object_print_static) 2685 { 2686 /* XXX convert args. */ 2687 vm_object_t object = (vm_object_t)addr; 2688 boolean_t full = have_addr; 2689 2690 vm_page_t p; 2691 2692 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2693 #define count was_count 2694 2695 int count; 2696 2697 if (object == NULL) 2698 return; 2699 2700 db_iprintf( 2701 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2702 object, (int)object->type, (u_long)object->size, 2703 object->resident_page_count, object->ref_count, object->flags); 2704 /* 2705 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2706 */ 2707 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2708 object->shadow_count, 2709 object->backing_object ? object->backing_object->ref_count : 0, 2710 object->backing_object, (long)object->backing_object_offset); 2711 2712 if (!full) 2713 return; 2714 2715 db_indent += 2; 2716 count = 0; 2717 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 2718 if (count == 0) 2719 db_iprintf("memory:="); 2720 else if (count == 6) { 2721 db_printf("\n"); 2722 db_iprintf(" ..."); 2723 count = 0; 2724 } else 2725 db_printf(","); 2726 count++; 2727 2728 db_printf("(off=0x%lx,page=0x%lx)", 2729 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2730 } 2731 if (count != 0) 2732 db_printf("\n"); 2733 db_indent -= 2; 2734 } 2735 2736 /* XXX. */ 2737 #undef count 2738 2739 /* 2740 * XXX need this non-static entry for calling from vm_map_print. 2741 * 2742 * Debugging only 2743 */ 2744 void 2745 vm_object_print(/* db_expr_t */ long addr, 2746 boolean_t have_addr, 2747 /* db_expr_t */ long count, 2748 char *modif) 2749 { 2750 vm_object_print_static(addr, have_addr, count, modif); 2751 } 2752 2753 /* 2754 * Debugging only 2755 */ 2756 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2757 { 2758 vm_object_t object; 2759 int nl = 0; 2760 int c; 2761 for (object = TAILQ_FIRST(&vm_object_list); 2762 object != NULL; 2763 object = TAILQ_NEXT(object, object_list)) { 2764 vm_pindex_t idx, fidx; 2765 vm_pindex_t osize; 2766 vm_paddr_t pa = -1, padiff; 2767 int rcount; 2768 vm_page_t m; 2769 2770 if (object->type == OBJT_MARKER) 2771 continue; 2772 db_printf("new object: %p\n", (void *)object); 2773 if ( nl > 18) { 2774 c = cngetc(); 2775 if (c != ' ') 2776 return; 2777 nl = 0; 2778 } 2779 nl++; 2780 rcount = 0; 2781 fidx = 0; 2782 osize = object->size; 2783 if (osize > 128) 2784 osize = 128; 2785 for (idx = 0; idx < osize; idx++) { 2786 m = vm_page_lookup(object, idx); 2787 if (m == NULL) { 2788 if (rcount) { 2789 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2790 (long)fidx, rcount, (long)pa); 2791 if ( nl > 18) { 2792 c = cngetc(); 2793 if (c != ' ') 2794 return; 2795 nl = 0; 2796 } 2797 nl++; 2798 rcount = 0; 2799 } 2800 continue; 2801 } 2802 2803 2804 if (rcount && 2805 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2806 ++rcount; 2807 continue; 2808 } 2809 if (rcount) { 2810 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2811 padiff >>= PAGE_SHIFT; 2812 padiff &= PQ_L2_MASK; 2813 if (padiff == 0) { 2814 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2815 ++rcount; 2816 continue; 2817 } 2818 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2819 (long)fidx, rcount, (long)pa); 2820 db_printf("pd(%ld)\n", (long)padiff); 2821 if ( nl > 18) { 2822 c = cngetc(); 2823 if (c != ' ') 2824 return; 2825 nl = 0; 2826 } 2827 nl++; 2828 } 2829 fidx = idx; 2830 pa = VM_PAGE_TO_PHYS(m); 2831 rcount = 1; 2832 } 2833 if (rcount) { 2834 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2835 (long)fidx, rcount, (long)pa); 2836 if ( nl > 18) { 2837 c = cngetc(); 2838 if (c != ' ') 2839 return; 2840 nl = 0; 2841 } 2842 nl++; 2843 } 2844 } 2845 } 2846 #endif /* DDB */ 2847