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