1 /* 2 * Copyright (c) 1991, 1993, 2013 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 * 60 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $ 61 */ 62 63 /* 64 * Virtual memory object module. 65 */ 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/proc.h> /* for curproc, pageproc */ 70 #include <sys/thread.h> 71 #include <sys/vnode.h> 72 #include <sys/vmmeter.h> 73 #include <sys/mman.h> 74 #include <sys/mount.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/refcount.h> 78 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <vm/pmap.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_object.h> 84 #include <vm/vm_page.h> 85 #include <vm/vm_pageout.h> 86 #include <vm/vm_pager.h> 87 #include <vm/swap_pager.h> 88 #include <vm/vm_kern.h> 89 #include <vm/vm_extern.h> 90 #include <vm/vm_zone.h> 91 92 #include <vm/vm_page2.h> 93 94 #include <machine/specialreg.h> 95 96 #define EASY_SCAN_FACTOR 8 97 98 static void vm_object_qcollapse(vm_object_t object, 99 vm_object_t backing_object); 100 static void vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 101 int pagerflags); 102 static void vm_object_lock_init(vm_object_t); 103 104 105 /* 106 * Virtual memory objects maintain the actual data 107 * associated with allocated virtual memory. A given 108 * page of memory exists within exactly one object. 109 * 110 * An object is only deallocated when all "references" 111 * are given up. Only one "reference" to a given 112 * region of an object should be writeable. 113 * 114 * Associated with each object is a list of all resident 115 * memory pages belonging to that object; this list is 116 * maintained by the "vm_page" module, and locked by the object's 117 * lock. 118 * 119 * Each object also records a "pager" routine which is 120 * used to retrieve (and store) pages to the proper backing 121 * storage. In addition, objects may be backed by other 122 * objects from which they were virtual-copied. 123 * 124 * The only items within the object structure which are 125 * modified after time of creation are: 126 * reference count locked by object's lock 127 * pager routine locked by object's lock 128 * 129 */ 130 131 struct vm_object kernel_object; 132 133 static long vm_object_count; 134 135 static long object_collapses; 136 static long object_bypasses; 137 static int next_index; 138 static vm_zone_t obj_zone; 139 static struct vm_zone obj_zone_store; 140 #define VM_OBJECTS_INIT 256 141 static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 142 143 struct object_q vm_object_lists[VMOBJ_HSIZE]; 144 struct lwkt_token vmobj_tokens[VMOBJ_HSIZE]; 145 146 #if defined(DEBUG_LOCKS) 147 148 #define vm_object_vndeallocate(obj, vpp) \ 149 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__) 150 151 /* 152 * Debug helper to track hold/drop/ref/deallocate calls. 153 */ 154 static void 155 debugvm_object_add(vm_object_t obj, char *file, int line, int addrem) 156 { 157 int i; 158 159 i = atomic_fetchadd_int(&obj->debug_index, 1); 160 i = i & (VMOBJ_DEBUG_ARRAY_SIZE - 1); 161 ksnprintf(obj->debug_hold_thrs[i], 162 sizeof(obj->debug_hold_thrs[i]), 163 "%c%d:(%d):%s", 164 (addrem == -1 ? '-' : (addrem == 1 ? '+' : '=')), 165 (curthread->td_proc ? curthread->td_proc->p_pid : -1), 166 obj->ref_count, 167 curthread->td_comm); 168 obj->debug_hold_file[i] = file; 169 obj->debug_hold_line[i] = line; 170 #if 0 171 /* Uncomment for debugging obj refs/derefs in reproducable cases */ 172 if (strcmp(curthread->td_comm, "sshd") == 0) { 173 kprintf("%d %p refs=%d ar=%d file: %s/%d\n", 174 (curthread->td_proc ? curthread->td_proc->p_pid : -1), 175 obj, obj->ref_count, addrem, file, line); 176 } 177 #endif 178 } 179 180 #endif 181 182 /* 183 * Misc low level routines 184 */ 185 static void 186 vm_object_lock_init(vm_object_t obj) 187 { 188 #if defined(DEBUG_LOCKS) 189 int i; 190 191 obj->debug_index = 0; 192 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 193 obj->debug_hold_thrs[i][0] = 0; 194 obj->debug_hold_file[i] = NULL; 195 obj->debug_hold_line[i] = 0; 196 } 197 #endif 198 } 199 200 void 201 vm_object_lock_swap(void) 202 { 203 lwkt_token_swap(); 204 } 205 206 void 207 vm_object_lock(vm_object_t obj) 208 { 209 lwkt_gettoken(&obj->token); 210 } 211 212 /* 213 * Returns TRUE on sucesss 214 */ 215 static int 216 vm_object_lock_try(vm_object_t obj) 217 { 218 return(lwkt_trytoken(&obj->token)); 219 } 220 221 void 222 vm_object_lock_shared(vm_object_t obj) 223 { 224 lwkt_gettoken_shared(&obj->token); 225 } 226 227 void 228 vm_object_unlock(vm_object_t obj) 229 { 230 lwkt_reltoken(&obj->token); 231 } 232 233 void 234 vm_object_upgrade(vm_object_t obj) 235 { 236 lwkt_reltoken(&obj->token); 237 lwkt_gettoken(&obj->token); 238 } 239 240 void 241 vm_object_downgrade(vm_object_t obj) 242 { 243 lwkt_reltoken(&obj->token); 244 lwkt_gettoken_shared(&obj->token); 245 } 246 247 static __inline void 248 vm_object_assert_held(vm_object_t obj) 249 { 250 ASSERT_LWKT_TOKEN_HELD(&obj->token); 251 } 252 253 void 254 VMOBJDEBUG(vm_object_hold)(vm_object_t obj VMOBJDBARGS) 255 { 256 KKASSERT(obj != NULL); 257 258 /* 259 * Object must be held (object allocation is stable due to callers 260 * context, typically already holding the token on a parent object) 261 * prior to potentially blocking on the lock, otherwise the object 262 * can get ripped away from us. 263 */ 264 refcount_acquire(&obj->hold_count); 265 vm_object_lock(obj); 266 267 #if defined(DEBUG_LOCKS) 268 debugvm_object_add(obj, file, line, 1); 269 #endif 270 } 271 272 int 273 VMOBJDEBUG(vm_object_hold_try)(vm_object_t obj VMOBJDBARGS) 274 { 275 KKASSERT(obj != NULL); 276 277 /* 278 * Object must be held (object allocation is stable due to callers 279 * context, typically already holding the token on a parent object) 280 * prior to potentially blocking on the lock, otherwise the object 281 * can get ripped away from us. 282 */ 283 refcount_acquire(&obj->hold_count); 284 if (vm_object_lock_try(obj) == 0) { 285 if (refcount_release(&obj->hold_count)) { 286 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) 287 zfree(obj_zone, obj); 288 } 289 return(0); 290 } 291 292 #if defined(DEBUG_LOCKS) 293 debugvm_object_add(obj, file, line, 1); 294 #endif 295 return(1); 296 } 297 298 void 299 VMOBJDEBUG(vm_object_hold_shared)(vm_object_t obj VMOBJDBARGS) 300 { 301 KKASSERT(obj != NULL); 302 303 /* 304 * Object must be held (object allocation is stable due to callers 305 * context, typically already holding the token on a parent object) 306 * prior to potentially blocking on the lock, otherwise the object 307 * can get ripped away from us. 308 */ 309 refcount_acquire(&obj->hold_count); 310 vm_object_lock_shared(obj); 311 312 #if defined(DEBUG_LOCKS) 313 debugvm_object_add(obj, file, line, 1); 314 #endif 315 } 316 317 /* 318 * Drop the token and hold_count on the object. 319 * 320 * WARNING! Token might be shared. 321 */ 322 void 323 VMOBJDEBUG(vm_object_drop)(vm_object_t obj VMOBJDBARGS) 324 { 325 if (obj == NULL) 326 return; 327 328 /* 329 * No new holders should be possible once we drop hold_count 1->0 as 330 * there is no longer any way to reference the object. 331 */ 332 KKASSERT(obj->hold_count > 0); 333 if (refcount_release(&obj->hold_count)) { 334 #if defined(DEBUG_LOCKS) 335 debugvm_object_add(obj, file, line, -1); 336 #endif 337 338 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) { 339 vm_object_unlock(obj); 340 zfree(obj_zone, obj); 341 } else { 342 vm_object_unlock(obj); 343 } 344 } else { 345 #if defined(DEBUG_LOCKS) 346 debugvm_object_add(obj, file, line, -1); 347 #endif 348 vm_object_unlock(obj); 349 } 350 } 351 352 /* 353 * Initialize a freshly allocated object, returning a held object. 354 * 355 * Used only by vm_object_allocate() and zinitna(). 356 * 357 * No requirements. 358 */ 359 void 360 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 361 { 362 int incr; 363 int n; 364 365 RB_INIT(&object->rb_memq); 366 LIST_INIT(&object->shadow_head); 367 lwkt_token_init(&object->token, "vmobj"); 368 369 object->type = type; 370 object->size = size; 371 object->ref_count = 1; 372 object->memattr = VM_MEMATTR_DEFAULT; 373 object->hold_count = 0; 374 object->flags = 0; 375 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 376 vm_object_set_flag(object, OBJ_ONEMAPPING); 377 object->paging_in_progress = 0; 378 object->resident_page_count = 0; 379 object->agg_pv_list_count = 0; 380 object->shadow_count = 0; 381 /* cpu localization twist */ 382 object->pg_color = (int)(intptr_t)curthread; 383 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 384 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 385 else 386 incr = size; 387 next_index = (next_index + incr) & PQ_L2_MASK; 388 object->handle = NULL; 389 object->backing_object = NULL; 390 object->backing_object_offset = (vm_ooffset_t)0; 391 392 object->generation++; 393 object->swblock_count = 0; 394 RB_INIT(&object->swblock_root); 395 vm_object_lock_init(object); 396 pmap_object_init(object); 397 398 vm_object_hold(object); 399 400 n = VMOBJ_HASH(object); 401 atomic_add_long(&vm_object_count, 1); 402 lwkt_gettoken(&vmobj_tokens[n]); 403 TAILQ_INSERT_TAIL(&vm_object_lists[n], object, object_list); 404 lwkt_reltoken(&vmobj_tokens[n]); 405 } 406 407 /* 408 * Initialize the VM objects module. 409 * 410 * Called from the low level boot code only. 411 */ 412 void 413 vm_object_init(void) 414 { 415 int i; 416 417 for (i = 0; i < VMOBJ_HSIZE; ++i) { 418 TAILQ_INIT(&vm_object_lists[i]); 419 lwkt_token_init(&vmobj_tokens[i], "vmobjlst"); 420 } 421 422 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 423 &kernel_object); 424 vm_object_drop(&kernel_object); 425 426 obj_zone = &obj_zone_store; 427 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 428 vm_objects_init, VM_OBJECTS_INIT); 429 } 430 431 void 432 vm_object_init2(void) 433 { 434 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1); 435 } 436 437 /* 438 * Allocate and return a new object of the specified type and size. 439 * 440 * No requirements. 441 */ 442 vm_object_t 443 vm_object_allocate(objtype_t type, vm_pindex_t size) 444 { 445 vm_object_t result; 446 447 result = (vm_object_t) zalloc(obj_zone); 448 449 _vm_object_allocate(type, size, result); 450 vm_object_drop(result); 451 452 return (result); 453 } 454 455 /* 456 * This version returns a held object, allowing further atomic initialization 457 * of the object. 458 */ 459 vm_object_t 460 vm_object_allocate_hold(objtype_t type, vm_pindex_t size) 461 { 462 vm_object_t result; 463 464 result = (vm_object_t) zalloc(obj_zone); 465 466 _vm_object_allocate(type, size, result); 467 468 return (result); 469 } 470 471 /* 472 * Add an additional reference to a vm_object. The object must already be 473 * held. The original non-lock version is no longer supported. The object 474 * must NOT be chain locked by anyone at the time the reference is added. 475 * 476 * Referencing a chain-locked object can blow up the fairly sensitive 477 * ref_count and shadow_count tests in the deallocator. Most callers 478 * will call vm_object_chain_wait() prior to calling 479 * vm_object_reference_locked() to avoid the case. 480 * 481 * The object must be held, but may be held shared if desired (hence why 482 * we use an atomic op). 483 */ 484 void 485 VMOBJDEBUG(vm_object_reference_locked)(vm_object_t object VMOBJDBARGS) 486 { 487 KKASSERT(object != NULL); 488 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 489 KKASSERT((object->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) == 0); 490 atomic_add_int(&object->ref_count, 1); 491 if (object->type == OBJT_VNODE) { 492 vref(object->handle); 493 /* XXX what if the vnode is being destroyed? */ 494 } 495 #if defined(DEBUG_LOCKS) 496 debugvm_object_add(object, file, line, 1); 497 #endif 498 } 499 500 /* 501 * This version is only allowed for vnode objects. 502 */ 503 void 504 VMOBJDEBUG(vm_object_reference_quick)(vm_object_t object VMOBJDBARGS) 505 { 506 KKASSERT(object->type == OBJT_VNODE); 507 atomic_add_int(&object->ref_count, 1); 508 vref(object->handle); 509 #if defined(DEBUG_LOCKS) 510 debugvm_object_add(object, file, line, 1); 511 #endif 512 } 513 514 /* 515 * Object OBJ_CHAINLOCK lock handling. 516 * 517 * The caller can chain-lock backing objects recursively and then 518 * use vm_object_chain_release_all() to undo the whole chain. 519 * 520 * Chain locks are used to prevent collapses and are only applicable 521 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations 522 * on other object types are ignored. This is also important because 523 * it allows e.g. the vnode underlying a memory mapping to take concurrent 524 * faults. 525 * 526 * The object must usually be held on entry, though intermediate 527 * objects need not be held on release. The object must be held exclusively, 528 * NOT shared. Note that the prefault path checks the shared state and 529 * avoids using the chain functions. 530 */ 531 void 532 vm_object_chain_wait(vm_object_t object, int shared) 533 { 534 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 535 for (;;) { 536 uint32_t chainlk = object->chainlk; 537 538 cpu_ccfence(); 539 if (shared) { 540 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) { 541 tsleep_interlock(object, 0); 542 if (atomic_cmpset_int(&object->chainlk, 543 chainlk, 544 chainlk | CHAINLK_WAIT)) { 545 tsleep(object, PINTERLOCKED, 546 "objchns", 0); 547 } 548 /* retry */ 549 } else { 550 break; 551 } 552 /* retry */ 553 } else { 554 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) { 555 tsleep_interlock(object, 0); 556 if (atomic_cmpset_int(&object->chainlk, 557 chainlk, 558 chainlk | CHAINLK_WAIT)) 559 { 560 tsleep(object, PINTERLOCKED, 561 "objchnx", 0); 562 } 563 /* retry */ 564 } else { 565 if (atomic_cmpset_int(&object->chainlk, 566 chainlk, 567 chainlk & ~CHAINLK_WAIT)) 568 { 569 if (chainlk & CHAINLK_WAIT) 570 wakeup(object); 571 break; 572 } 573 /* retry */ 574 } 575 } 576 /* retry */ 577 } 578 } 579 580 void 581 vm_object_chain_acquire(vm_object_t object, int shared) 582 { 583 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) 584 return; 585 if (vm_shared_fault == 0) 586 shared = 0; 587 588 for (;;) { 589 uint32_t chainlk = object->chainlk; 590 591 cpu_ccfence(); 592 if (shared) { 593 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) { 594 tsleep_interlock(object, 0); 595 if (atomic_cmpset_int(&object->chainlk, 596 chainlk, 597 chainlk | CHAINLK_WAIT)) { 598 tsleep(object, PINTERLOCKED, 599 "objchns", 0); 600 } 601 /* retry */ 602 } else if (atomic_cmpset_int(&object->chainlk, 603 chainlk, chainlk + 1)) { 604 break; 605 } 606 /* retry */ 607 } else { 608 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) { 609 tsleep_interlock(object, 0); 610 if (atomic_cmpset_int(&object->chainlk, 611 chainlk, 612 chainlk | 613 CHAINLK_WAIT | 614 CHAINLK_EXCLREQ)) { 615 tsleep(object, PINTERLOCKED, 616 "objchnx", 0); 617 } 618 /* retry */ 619 } else { 620 if (atomic_cmpset_int(&object->chainlk, 621 chainlk, 622 (chainlk | CHAINLK_EXCL) & 623 ~(CHAINLK_EXCLREQ | 624 CHAINLK_WAIT))) { 625 if (chainlk & CHAINLK_WAIT) 626 wakeup(object); 627 break; 628 } 629 /* retry */ 630 } 631 } 632 /* retry */ 633 } 634 } 635 636 void 637 vm_object_chain_release(vm_object_t object) 638 { 639 /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/ 640 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) 641 return; 642 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL)); 643 for (;;) { 644 uint32_t chainlk = object->chainlk; 645 646 cpu_ccfence(); 647 if (chainlk & CHAINLK_MASK) { 648 if ((chainlk & CHAINLK_MASK) == 1 && 649 atomic_cmpset_int(&object->chainlk, 650 chainlk, 651 (chainlk - 1) & ~CHAINLK_WAIT)) { 652 if (chainlk & CHAINLK_WAIT) 653 wakeup(object); 654 break; 655 } 656 if ((chainlk & CHAINLK_MASK) > 1 && 657 atomic_cmpset_int(&object->chainlk, 658 chainlk, chainlk - 1)) { 659 break; 660 } 661 /* retry */ 662 } else { 663 KKASSERT(chainlk & CHAINLK_EXCL); 664 if (atomic_cmpset_int(&object->chainlk, 665 chainlk, 666 chainlk & ~(CHAINLK_EXCL | 667 CHAINLK_WAIT))) { 668 if (chainlk & CHAINLK_WAIT) 669 wakeup(object); 670 break; 671 } 672 } 673 } 674 } 675 676 /* 677 * Release the chain from first_object through and including stopobj. 678 * The caller is typically holding the first and last object locked 679 * (shared or exclusive) to prevent destruction races. 680 * 681 * We release stopobj first as an optimization as this object is most 682 * likely to be shared across multiple processes. 683 */ 684 void 685 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj) 686 { 687 vm_object_t backing_object; 688 vm_object_t object; 689 690 vm_object_chain_release(stopobj); 691 object = first_object; 692 693 while (object != stopobj) { 694 KKASSERT(object); 695 backing_object = object->backing_object; 696 vm_object_chain_release(object); 697 object = backing_object; 698 } 699 } 700 701 /* 702 * Dereference an object and its underlying vnode. The object may be 703 * held shared. On return the object will remain held. 704 * 705 * This function may return a vnode in *vpp which the caller must release 706 * after the caller drops its own lock. If vpp is NULL, we assume that 707 * the caller was holding an exclusive lock on the object and we vrele() 708 * the vp ourselves. 709 */ 710 static void 711 VMOBJDEBUG(vm_object_vndeallocate)(vm_object_t object, struct vnode **vpp 712 VMOBJDBARGS) 713 { 714 struct vnode *vp = (struct vnode *) object->handle; 715 716 KASSERT(object->type == OBJT_VNODE, 717 ("vm_object_vndeallocate: not a vnode object")); 718 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 719 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 720 #ifdef INVARIANTS 721 if (object->ref_count == 0) { 722 vprint("vm_object_vndeallocate", vp); 723 panic("vm_object_vndeallocate: bad object reference count"); 724 } 725 #endif 726 for (;;) { 727 int count = object->ref_count; 728 cpu_ccfence(); 729 if (count == 1) { 730 vm_object_upgrade(object); 731 if (atomic_cmpset_int(&object->ref_count, count, 0)) { 732 vclrflags(vp, VTEXT); 733 break; 734 } 735 } else { 736 if (atomic_cmpset_int(&object->ref_count, 737 count, count - 1)) { 738 break; 739 } 740 } 741 /* retry */ 742 } 743 #if defined(DEBUG_LOCKS) 744 debugvm_object_add(object, file, line, -1); 745 #endif 746 747 /* 748 * vrele or return the vp to vrele. We can only safely vrele(vp) 749 * if the object was locked exclusively. But there are two races 750 * here. 751 * 752 * We had to upgrade the object above to safely clear VTEXT 753 * but the alternative path where the shared lock is retained 754 * can STILL race to 0 in other paths and cause our own vrele() 755 * to terminate the vnode. We can't allow that if the VM object 756 * is still locked shared. 757 */ 758 if (vpp) 759 *vpp = vp; 760 else 761 vrele(vp); 762 } 763 764 /* 765 * Release a reference to the specified object, gained either through a 766 * vm_object_allocate or a vm_object_reference call. When all references 767 * are gone, storage associated with this object may be relinquished. 768 * 769 * The caller does not have to hold the object locked but must have control 770 * over the reference in question in order to guarantee that the object 771 * does not get ripped out from under us. 772 * 773 * XXX Currently all deallocations require an exclusive lock. 774 */ 775 void 776 VMOBJDEBUG(vm_object_deallocate)(vm_object_t object VMOBJDBARGS) 777 { 778 struct vnode *vp; 779 int count; 780 781 if (object == NULL) 782 return; 783 784 for (;;) { 785 count = object->ref_count; 786 cpu_ccfence(); 787 788 /* 789 * If decrementing the count enters into special handling 790 * territory (0, 1, or 2) we have to do it the hard way. 791 * Fortunate though, objects with only a few refs like this 792 * are not likely to be heavily contended anyway. 793 * 794 * For vnode objects we only care about 1->0 transitions. 795 */ 796 if (count <= 3 || (object->type == OBJT_VNODE && count <= 1)) { 797 #if defined(DEBUG_LOCKS) 798 debugvm_object_add(object, file, line, 0); 799 #endif 800 vm_object_hold(object); 801 vm_object_deallocate_locked(object); 802 vm_object_drop(object); 803 break; 804 } 805 806 /* 807 * Try to decrement ref_count without acquiring a hold on 808 * the object. This is particularly important for the exec*() 809 * and exit*() code paths because the program binary may 810 * have a great deal of sharing and an exclusive lock will 811 * crowbar performance in those circumstances. 812 */ 813 if (object->type == OBJT_VNODE) { 814 vp = (struct vnode *)object->handle; 815 if (atomic_cmpset_int(&object->ref_count, 816 count, count - 1)) { 817 #if defined(DEBUG_LOCKS) 818 debugvm_object_add(object, file, line, -1); 819 #endif 820 821 vrele(vp); 822 break; 823 } 824 /* retry */ 825 } else { 826 if (atomic_cmpset_int(&object->ref_count, 827 count, count - 1)) { 828 #if defined(DEBUG_LOCKS) 829 debugvm_object_add(object, file, line, -1); 830 #endif 831 break; 832 } 833 /* retry */ 834 } 835 /* retry */ 836 } 837 } 838 839 void 840 VMOBJDEBUG(vm_object_deallocate_locked)(vm_object_t object VMOBJDBARGS) 841 { 842 struct vm_object_dealloc_list *dlist = NULL; 843 struct vm_object_dealloc_list *dtmp; 844 vm_object_t temp; 845 int must_drop = 0; 846 847 /* 848 * We may chain deallocate object, but additional objects may 849 * collect on the dlist which also have to be deallocated. We 850 * must avoid a recursion, vm_object chains can get deep. 851 */ 852 853 again: 854 while (object != NULL) { 855 /* 856 * vnode case, caller either locked the object exclusively 857 * or this is a recursion with must_drop != 0 and the vnode 858 * object will be locked shared. 859 * 860 * If locked shared we have to drop the object before we can 861 * call vrele() or risk a shared/exclusive livelock. 862 */ 863 if (object->type == OBJT_VNODE) { 864 ASSERT_LWKT_TOKEN_HELD(&object->token); 865 if (must_drop) { 866 struct vnode *tmp_vp; 867 868 vm_object_vndeallocate(object, &tmp_vp); 869 vm_object_drop(object); 870 must_drop = 0; 871 object = NULL; 872 vrele(tmp_vp); 873 } else { 874 vm_object_vndeallocate(object, NULL); 875 } 876 break; 877 } 878 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token); 879 880 /* 881 * Normal case (object is locked exclusively) 882 */ 883 if (object->ref_count == 0) { 884 panic("vm_object_deallocate: object deallocated " 885 "too many times: %d", object->type); 886 } 887 if (object->ref_count > 2) { 888 atomic_add_int(&object->ref_count, -1); 889 #if defined(DEBUG_LOCKS) 890 debugvm_object_add(object, file, line, -1); 891 #endif 892 break; 893 } 894 895 /* 896 * Here on ref_count of one or two, which are special cases for 897 * objects. 898 * 899 * Nominal ref_count > 1 case if the second ref is not from 900 * a shadow. 901 * 902 * (ONEMAPPING only applies to DEFAULT AND SWAP objects) 903 */ 904 if (object->ref_count == 2 && object->shadow_count == 0) { 905 if (object->type == OBJT_DEFAULT || 906 object->type == OBJT_SWAP) { 907 vm_object_set_flag(object, OBJ_ONEMAPPING); 908 } 909 atomic_add_int(&object->ref_count, -1); 910 #if defined(DEBUG_LOCKS) 911 debugvm_object_add(object, file, line, -1); 912 #endif 913 break; 914 } 915 916 /* 917 * If the second ref is from a shadow we chain along it 918 * upwards if object's handle is exhausted. 919 * 920 * We have to decrement object->ref_count before potentially 921 * collapsing the first shadow object or the collapse code 922 * will not be able to handle the degenerate case to remove 923 * object. However, if we do it too early the object can 924 * get ripped out from under us. 925 */ 926 if (object->ref_count == 2 && object->shadow_count == 1 && 927 object->handle == NULL && (object->type == OBJT_DEFAULT || 928 object->type == OBJT_SWAP)) { 929 temp = LIST_FIRST(&object->shadow_head); 930 KKASSERT(temp != NULL); 931 vm_object_hold(temp); 932 933 /* 934 * Wait for any paging to complete so the collapse 935 * doesn't (or isn't likely to) qcollapse. pip 936 * waiting must occur before we acquire the 937 * chainlock. 938 */ 939 while ( 940 temp->paging_in_progress || 941 object->paging_in_progress 942 ) { 943 vm_object_pip_wait(temp, "objde1"); 944 vm_object_pip_wait(object, "objde2"); 945 } 946 947 /* 948 * If the parent is locked we have to give up, as 949 * otherwise we would be acquiring locks in the 950 * wrong order and potentially deadlock. 951 */ 952 if (temp->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) { 953 vm_object_drop(temp); 954 goto skip; 955 } 956 vm_object_chain_acquire(temp, 0); 957 958 /* 959 * Recheck/retry after the hold and the paging 960 * wait, both of which can block us. 961 */ 962 if (object->ref_count != 2 || 963 object->shadow_count != 1 || 964 object->handle || 965 LIST_FIRST(&object->shadow_head) != temp || 966 (object->type != OBJT_DEFAULT && 967 object->type != OBJT_SWAP)) { 968 vm_object_chain_release(temp); 969 vm_object_drop(temp); 970 continue; 971 } 972 973 /* 974 * We can safely drop object's ref_count now. 975 */ 976 KKASSERT(object->ref_count == 2); 977 atomic_add_int(&object->ref_count, -1); 978 #if defined(DEBUG_LOCKS) 979 debugvm_object_add(object, file, line, -1); 980 #endif 981 982 /* 983 * If our single parent is not collapseable just 984 * decrement ref_count (2->1) and stop. 985 */ 986 if (temp->handle || (temp->type != OBJT_DEFAULT && 987 temp->type != OBJT_SWAP)) { 988 vm_object_chain_release(temp); 989 vm_object_drop(temp); 990 break; 991 } 992 993 /* 994 * At this point we have already dropped object's 995 * ref_count so it is possible for a race to 996 * deallocate obj out from under us. Any collapse 997 * will re-check the situation. We must not block 998 * until we are able to collapse. 999 * 1000 * Bump temp's ref_count to avoid an unwanted 1001 * degenerate recursion (can't call 1002 * vm_object_reference_locked() because it asserts 1003 * that CHAINLOCK is not set). 1004 */ 1005 atomic_add_int(&temp->ref_count, 1); 1006 KKASSERT(temp->ref_count > 1); 1007 1008 /* 1009 * Collapse temp, then deallocate the extra ref 1010 * formally. 1011 */ 1012 vm_object_collapse(temp, &dlist); 1013 vm_object_chain_release(temp); 1014 if (must_drop) { 1015 vm_object_lock_swap(); 1016 vm_object_drop(object); 1017 } 1018 object = temp; 1019 must_drop = 1; 1020 continue; 1021 } 1022 1023 /* 1024 * Drop the ref and handle termination on the 1->0 transition. 1025 * We may have blocked above so we have to recheck. 1026 */ 1027 skip: 1028 KKASSERT(object->ref_count != 0); 1029 if (object->ref_count >= 2) { 1030 atomic_add_int(&object->ref_count, -1); 1031 #if defined(DEBUG_LOCKS) 1032 debugvm_object_add(object, file, line, -1); 1033 #endif 1034 break; 1035 } 1036 KKASSERT(object->ref_count == 1); 1037 1038 /* 1039 * 1->0 transition. Chain through the backing_object. 1040 * Maintain the ref until we've located the backing object, 1041 * then re-check. 1042 */ 1043 while ((temp = object->backing_object) != NULL) { 1044 if (temp->type == OBJT_VNODE) 1045 vm_object_hold_shared(temp); 1046 else 1047 vm_object_hold(temp); 1048 if (temp == object->backing_object) 1049 break; 1050 vm_object_drop(temp); 1051 } 1052 1053 /* 1054 * 1->0 transition verified, retry if ref_count is no longer 1055 * 1. Otherwise disconnect the backing_object (temp) and 1056 * clean up. 1057 */ 1058 if (object->ref_count != 1) { 1059 vm_object_drop(temp); 1060 continue; 1061 } 1062 1063 /* 1064 * It shouldn't be possible for the object to be chain locked 1065 * if we're removing the last ref on it. 1066 * 1067 * Removing object from temp's shadow list requires dropping 1068 * temp, which we will do on loop. 1069 * 1070 * NOTE! vnodes do not use the shadow list, but still have 1071 * the backing_object reference. 1072 */ 1073 KKASSERT((object->chainlk & (CHAINLK_EXCL|CHAINLK_MASK)) == 0); 1074 1075 if (temp) { 1076 if (object->flags & OBJ_ONSHADOW) { 1077 LIST_REMOVE(object, shadow_list); 1078 temp->shadow_count--; 1079 temp->generation++; 1080 vm_object_clear_flag(object, OBJ_ONSHADOW); 1081 } 1082 object->backing_object = NULL; 1083 } 1084 1085 atomic_add_int(&object->ref_count, -1); 1086 if ((object->flags & OBJ_DEAD) == 0) 1087 vm_object_terminate(object); 1088 if (must_drop && temp) 1089 vm_object_lock_swap(); 1090 if (must_drop) 1091 vm_object_drop(object); 1092 object = temp; 1093 must_drop = 1; 1094 } 1095 1096 if (must_drop && object) 1097 vm_object_drop(object); 1098 1099 /* 1100 * Additional tail recursion on dlist. Avoid a recursion. Objects 1101 * on the dlist have a hold count but are not locked. 1102 */ 1103 if ((dtmp = dlist) != NULL) { 1104 dlist = dtmp->next; 1105 object = dtmp->object; 1106 kfree(dtmp, M_TEMP); 1107 1108 vm_object_lock(object); /* already held, add lock */ 1109 must_drop = 1; /* and we're responsible for it */ 1110 goto again; 1111 } 1112 } 1113 1114 /* 1115 * Destroy the specified object, freeing up related resources. 1116 * 1117 * The object must have zero references. 1118 * 1119 * The object must held. The caller is responsible for dropping the object 1120 * after terminate returns. Terminate does NOT drop the object. 1121 */ 1122 static int vm_object_terminate_callback(vm_page_t p, void *data); 1123 1124 void 1125 vm_object_terminate(vm_object_t object) 1126 { 1127 int n; 1128 1129 /* 1130 * Make sure no one uses us. Once we set OBJ_DEAD we should be 1131 * able to safely block. 1132 */ 1133 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1134 KKASSERT((object->flags & OBJ_DEAD) == 0); 1135 vm_object_set_flag(object, OBJ_DEAD); 1136 1137 /* 1138 * Wait for the pageout daemon to be done with the object 1139 */ 1140 vm_object_pip_wait(object, "objtrm1"); 1141 1142 KASSERT(!object->paging_in_progress, 1143 ("vm_object_terminate: pageout in progress")); 1144 1145 /* 1146 * Clean and free the pages, as appropriate. All references to the 1147 * object are gone, so we don't need to lock it. 1148 */ 1149 if (object->type == OBJT_VNODE) { 1150 struct vnode *vp; 1151 1152 /* 1153 * Clean pages and flush buffers. 1154 * 1155 * NOTE! TMPFS buffer flushes do not typically flush the 1156 * actual page to swap as this would be highly 1157 * inefficient, and normal filesystems usually wrap 1158 * page flushes with buffer cache buffers. 1159 * 1160 * To deal with this we have to call vinvalbuf() both 1161 * before and after the vm_object_page_clean(). 1162 */ 1163 vp = (struct vnode *) object->handle; 1164 vinvalbuf(vp, V_SAVE, 0, 0); 1165 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 1166 vinvalbuf(vp, V_SAVE, 0, 0); 1167 } 1168 1169 /* 1170 * Wait for any I/O to complete, after which there had better not 1171 * be any references left on the object. 1172 */ 1173 vm_object_pip_wait(object, "objtrm2"); 1174 1175 if (object->ref_count != 0) { 1176 panic("vm_object_terminate: object with references, " 1177 "ref_count=%d", object->ref_count); 1178 } 1179 1180 /* 1181 * Cleanup any shared pmaps associated with this object. 1182 */ 1183 pmap_object_free(object); 1184 1185 /* 1186 * Now free any remaining pages. For internal objects, this also 1187 * removes them from paging queues. Don't free wired pages, just 1188 * remove them from the object. 1189 */ 1190 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1191 vm_object_terminate_callback, NULL); 1192 1193 /* 1194 * Let the pager know object is dead. 1195 */ 1196 vm_pager_deallocate(object); 1197 1198 /* 1199 * Wait for the object hold count to hit 1, clean out pages as 1200 * we go. vmobj_token interlocks any race conditions that might 1201 * pick the object up from the vm_object_list after we have cleared 1202 * rb_memq. 1203 */ 1204 for (;;) { 1205 if (RB_ROOT(&object->rb_memq) == NULL) 1206 break; 1207 kprintf("vm_object_terminate: Warning, object %p " 1208 "still has %d pages\n", 1209 object, object->resident_page_count); 1210 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1211 vm_object_terminate_callback, NULL); 1212 } 1213 1214 /* 1215 * There had better not be any pages left 1216 */ 1217 KKASSERT(object->resident_page_count == 0); 1218 1219 /* 1220 * Remove the object from the global object list. 1221 */ 1222 n = VMOBJ_HASH(object); 1223 lwkt_gettoken(&vmobj_tokens[n]); 1224 TAILQ_REMOVE(&vm_object_lists[n], object, object_list); 1225 lwkt_reltoken(&vmobj_tokens[n]); 1226 atomic_add_long(&vm_object_count, -1); 1227 1228 if (object->ref_count != 0) { 1229 panic("vm_object_terminate2: object with references, " 1230 "ref_count=%d", object->ref_count); 1231 } 1232 1233 /* 1234 * NOTE: The object hold_count is at least 1, so we cannot zfree() 1235 * the object here. See vm_object_drop(). 1236 */ 1237 } 1238 1239 /* 1240 * The caller must hold the object. 1241 */ 1242 static int 1243 vm_object_terminate_callback(vm_page_t p, void *data __unused) 1244 { 1245 vm_object_t object; 1246 1247 object = p->object; 1248 vm_page_busy_wait(p, TRUE, "vmpgtrm"); 1249 if (object != p->object) { 1250 kprintf("vm_object_terminate: Warning: Encountered " 1251 "busied page %p on queue %d\n", p, p->queue); 1252 vm_page_wakeup(p); 1253 } else if (p->wire_count == 0) { 1254 /* 1255 * NOTE: p->dirty and PG_NEED_COMMIT are ignored. 1256 */ 1257 vm_page_free(p); 1258 mycpu->gd_cnt.v_pfree++; 1259 } else { 1260 if (p->queue != PQ_NONE) 1261 kprintf("vm_object_terminate: Warning: Encountered " 1262 "wired page %p on queue %d\n", p, p->queue); 1263 vm_page_remove(p); 1264 vm_page_wakeup(p); 1265 } 1266 lwkt_yield(); 1267 return(0); 1268 } 1269 1270 /* 1271 * Clean all dirty pages in the specified range of object. Leaves page 1272 * on whatever queue it is currently on. If NOSYNC is set then do not 1273 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 1274 * leaving the object dirty. 1275 * 1276 * When stuffing pages asynchronously, allow clustering. XXX we need a 1277 * synchronous clustering mode implementation. 1278 * 1279 * Odd semantics: if start == end, we clean everything. 1280 * 1281 * The object must be locked? XXX 1282 */ 1283 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 1284 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 1285 1286 void 1287 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1288 int flags) 1289 { 1290 struct rb_vm_page_scan_info info; 1291 struct vnode *vp; 1292 int wholescan; 1293 int pagerflags; 1294 int generation; 1295 1296 vm_object_hold(object); 1297 if (object->type != OBJT_VNODE || 1298 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 1299 vm_object_drop(object); 1300 return; 1301 } 1302 1303 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 1304 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1305 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 1306 1307 vp = object->handle; 1308 1309 /* 1310 * Interlock other major object operations. This allows us to 1311 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 1312 */ 1313 vm_object_set_flag(object, OBJ_CLEANING); 1314 1315 /* 1316 * Handle 'entire object' case 1317 */ 1318 info.start_pindex = start; 1319 if (end == 0) { 1320 info.end_pindex = object->size - 1; 1321 } else { 1322 info.end_pindex = end - 1; 1323 } 1324 wholescan = (start == 0 && info.end_pindex == object->size - 1); 1325 info.limit = flags; 1326 info.pagerflags = pagerflags; 1327 info.object = object; 1328 1329 /* 1330 * If cleaning the entire object do a pass to mark the pages read-only. 1331 * If everything worked out ok, clear OBJ_WRITEABLE and 1332 * OBJ_MIGHTBEDIRTY. 1333 */ 1334 if (wholescan) { 1335 info.error = 0; 1336 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1337 vm_object_page_clean_pass1, &info); 1338 if (info.error == 0) { 1339 vm_object_clear_flag(object, 1340 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1341 if (object->type == OBJT_VNODE && 1342 (vp = (struct vnode *)object->handle) != NULL) { 1343 /* 1344 * Use new-style interface to clear VISDIRTY 1345 * because the vnode is not necessarily removed 1346 * from the syncer list(s) as often as it was 1347 * under the old interface, which can leave 1348 * the vnode on the syncer list after reclaim. 1349 */ 1350 vclrobjdirty(vp); 1351 } 1352 } 1353 } 1354 1355 /* 1356 * Do a pass to clean all the dirty pages we find. 1357 */ 1358 do { 1359 info.error = 0; 1360 generation = object->generation; 1361 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1362 vm_object_page_clean_pass2, &info); 1363 } while (info.error || generation != object->generation); 1364 1365 vm_object_clear_flag(object, OBJ_CLEANING); 1366 vm_object_drop(object); 1367 } 1368 1369 /* 1370 * The caller must hold the object. 1371 */ 1372 static 1373 int 1374 vm_object_page_clean_pass1(struct vm_page *p, void *data) 1375 { 1376 struct rb_vm_page_scan_info *info = data; 1377 1378 vm_page_flag_set(p, PG_CLEANCHK); 1379 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1380 info->error = 1; 1381 } else if (vm_page_busy_try(p, FALSE) == 0) { 1382 vm_page_protect(p, VM_PROT_READ); /* must not block */ 1383 vm_page_wakeup(p); 1384 } else { 1385 info->error = 1; 1386 } 1387 lwkt_yield(); 1388 return(0); 1389 } 1390 1391 /* 1392 * The caller must hold the object 1393 */ 1394 static 1395 int 1396 vm_object_page_clean_pass2(struct vm_page *p, void *data) 1397 { 1398 struct rb_vm_page_scan_info *info = data; 1399 int generation; 1400 1401 /* 1402 * Do not mess with pages that were inserted after we started 1403 * the cleaning pass. 1404 */ 1405 if ((p->flags & PG_CLEANCHK) == 0) 1406 goto done; 1407 1408 generation = info->object->generation; 1409 vm_page_busy_wait(p, TRUE, "vpcwai"); 1410 if (p->object != info->object || 1411 info->object->generation != generation) { 1412 info->error = 1; 1413 vm_page_wakeup(p); 1414 goto done; 1415 } 1416 1417 /* 1418 * Before wasting time traversing the pmaps, check for trivial 1419 * cases where the page cannot be dirty. 1420 */ 1421 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 1422 KKASSERT((p->dirty & p->valid) == 0 && 1423 (p->flags & PG_NEED_COMMIT) == 0); 1424 vm_page_wakeup(p); 1425 goto done; 1426 } 1427 1428 /* 1429 * Check whether the page is dirty or not. The page has been set 1430 * to be read-only so the check will not race a user dirtying the 1431 * page. 1432 */ 1433 vm_page_test_dirty(p); 1434 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) { 1435 vm_page_flag_clear(p, PG_CLEANCHK); 1436 vm_page_wakeup(p); 1437 goto done; 1438 } 1439 1440 /* 1441 * If we have been asked to skip nosync pages and this is a 1442 * nosync page, skip it. Note that the object flags were 1443 * not cleared in this case (because pass1 will have returned an 1444 * error), so we do not have to set them. 1445 */ 1446 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1447 vm_page_flag_clear(p, PG_CLEANCHK); 1448 vm_page_wakeup(p); 1449 goto done; 1450 } 1451 1452 /* 1453 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 1454 * the pages that get successfully flushed. Set info->error if 1455 * we raced an object modification. 1456 */ 1457 vm_object_page_collect_flush(info->object, p, info->pagerflags); 1458 /* vm_wait_nominal(); this can deadlock the system in syncer/pageout */ 1459 done: 1460 lwkt_yield(); 1461 return(0); 1462 } 1463 1464 /* 1465 * Collect the specified page and nearby pages and flush them out. 1466 * The number of pages flushed is returned. The passed page is busied 1467 * by the caller and we are responsible for its disposition. 1468 * 1469 * The caller must hold the object. 1470 */ 1471 static void 1472 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 1473 { 1474 int error; 1475 int is; 1476 int ib; 1477 int i; 1478 int page_base; 1479 vm_pindex_t pi; 1480 vm_page_t ma[BLIST_MAX_ALLOC]; 1481 1482 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1483 1484 pi = p->pindex; 1485 page_base = pi % BLIST_MAX_ALLOC; 1486 ma[page_base] = p; 1487 ib = page_base - 1; 1488 is = page_base + 1; 1489 1490 while (ib >= 0) { 1491 vm_page_t tp; 1492 1493 tp = vm_page_lookup_busy_try(object, pi - page_base + ib, 1494 TRUE, &error); 1495 if (error) 1496 break; 1497 if (tp == NULL) 1498 break; 1499 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1500 (tp->flags & PG_CLEANCHK) == 0) { 1501 vm_page_wakeup(tp); 1502 break; 1503 } 1504 if ((tp->queue - tp->pc) == PQ_CACHE) { 1505 vm_page_flag_clear(tp, PG_CLEANCHK); 1506 vm_page_wakeup(tp); 1507 break; 1508 } 1509 vm_page_test_dirty(tp); 1510 if ((tp->dirty & tp->valid) == 0 && 1511 (tp->flags & PG_NEED_COMMIT) == 0) { 1512 vm_page_flag_clear(tp, PG_CLEANCHK); 1513 vm_page_wakeup(tp); 1514 break; 1515 } 1516 ma[ib] = tp; 1517 --ib; 1518 } 1519 ++ib; /* fixup */ 1520 1521 while (is < BLIST_MAX_ALLOC && 1522 pi - page_base + is < object->size) { 1523 vm_page_t tp; 1524 1525 tp = vm_page_lookup_busy_try(object, pi - page_base + is, 1526 TRUE, &error); 1527 if (error) 1528 break; 1529 if (tp == NULL) 1530 break; 1531 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1532 (tp->flags & PG_CLEANCHK) == 0) { 1533 vm_page_wakeup(tp); 1534 break; 1535 } 1536 if ((tp->queue - tp->pc) == PQ_CACHE) { 1537 vm_page_flag_clear(tp, PG_CLEANCHK); 1538 vm_page_wakeup(tp); 1539 break; 1540 } 1541 vm_page_test_dirty(tp); 1542 if ((tp->dirty & tp->valid) == 0 && 1543 (tp->flags & PG_NEED_COMMIT) == 0) { 1544 vm_page_flag_clear(tp, PG_CLEANCHK); 1545 vm_page_wakeup(tp); 1546 break; 1547 } 1548 ma[is] = tp; 1549 ++is; 1550 } 1551 1552 /* 1553 * All pages in the ma[] array are busied now 1554 */ 1555 for (i = ib; i < is; ++i) { 1556 vm_page_flag_clear(ma[i], PG_CLEANCHK); 1557 vm_page_hold(ma[i]); /* XXX need this any more? */ 1558 } 1559 vm_pageout_flush(&ma[ib], is - ib, pagerflags); 1560 for (i = ib; i < is; ++i) /* XXX need this any more? */ 1561 vm_page_unhold(ma[i]); 1562 } 1563 1564 /* 1565 * Same as vm_object_pmap_copy, except range checking really 1566 * works, and is meant for small sections of an object. 1567 * 1568 * This code protects resident pages by making them read-only 1569 * and is typically called on a fork or split when a page 1570 * is converted to copy-on-write. 1571 * 1572 * NOTE: If the page is already at VM_PROT_NONE, calling 1573 * vm_page_protect will have no effect. 1574 */ 1575 void 1576 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1577 { 1578 vm_pindex_t idx; 1579 vm_page_t p; 1580 1581 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 1582 return; 1583 1584 vm_object_hold(object); 1585 for (idx = start; idx < end; idx++) { 1586 p = vm_page_lookup(object, idx); 1587 if (p == NULL) 1588 continue; 1589 vm_page_protect(p, VM_PROT_READ); 1590 } 1591 vm_object_drop(object); 1592 } 1593 1594 /* 1595 * Removes all physical pages in the specified object range from all 1596 * physical maps. 1597 * 1598 * The object must *not* be locked. 1599 */ 1600 1601 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 1602 1603 void 1604 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1605 { 1606 struct rb_vm_page_scan_info info; 1607 1608 if (object == NULL) 1609 return; 1610 info.start_pindex = start; 1611 info.end_pindex = end - 1; 1612 1613 vm_object_hold(object); 1614 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1615 vm_object_pmap_remove_callback, &info); 1616 if (start == 0 && end == object->size) 1617 vm_object_clear_flag(object, OBJ_WRITEABLE); 1618 vm_object_drop(object); 1619 } 1620 1621 /* 1622 * The caller must hold the object 1623 */ 1624 static int 1625 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 1626 { 1627 vm_page_protect(p, VM_PROT_NONE); 1628 return(0); 1629 } 1630 1631 /* 1632 * Implements the madvise function at the object/page level. 1633 * 1634 * MADV_WILLNEED (any object) 1635 * 1636 * Activate the specified pages if they are resident. 1637 * 1638 * MADV_DONTNEED (any object) 1639 * 1640 * Deactivate the specified pages if they are resident. 1641 * 1642 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 1643 * 1644 * Deactivate and clean the specified pages if they are 1645 * resident. This permits the process to reuse the pages 1646 * without faulting or the kernel to reclaim the pages 1647 * without I/O. 1648 * 1649 * No requirements. 1650 */ 1651 void 1652 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1653 { 1654 vm_pindex_t end, tpindex; 1655 vm_object_t tobject; 1656 vm_object_t xobj; 1657 vm_page_t m; 1658 int error; 1659 1660 if (object == NULL) 1661 return; 1662 1663 end = pindex + count; 1664 1665 vm_object_hold(object); 1666 tobject = object; 1667 1668 /* 1669 * Locate and adjust resident pages 1670 */ 1671 for (; pindex < end; pindex += 1) { 1672 relookup: 1673 if (tobject != object) 1674 vm_object_drop(tobject); 1675 tobject = object; 1676 tpindex = pindex; 1677 shadowlookup: 1678 /* 1679 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1680 * and those pages must be OBJ_ONEMAPPING. 1681 */ 1682 if (advise == MADV_FREE) { 1683 if ((tobject->type != OBJT_DEFAULT && 1684 tobject->type != OBJT_SWAP) || 1685 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1686 continue; 1687 } 1688 } 1689 1690 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error); 1691 1692 if (error) { 1693 vm_page_sleep_busy(m, TRUE, "madvpo"); 1694 goto relookup; 1695 } 1696 if (m == NULL) { 1697 /* 1698 * There may be swap even if there is no backing page 1699 */ 1700 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1701 swap_pager_freespace(tobject, tpindex, 1); 1702 1703 /* 1704 * next object 1705 */ 1706 while ((xobj = tobject->backing_object) != NULL) { 1707 KKASSERT(xobj != object); 1708 vm_object_hold(xobj); 1709 if (xobj == tobject->backing_object) 1710 break; 1711 vm_object_drop(xobj); 1712 } 1713 if (xobj == NULL) 1714 continue; 1715 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1716 if (tobject != object) { 1717 vm_object_lock_swap(); 1718 vm_object_drop(tobject); 1719 } 1720 tobject = xobj; 1721 goto shadowlookup; 1722 } 1723 1724 /* 1725 * If the page is not in a normal active state, we skip it. 1726 * If the page is not managed there are no page queues to 1727 * mess with. Things can break if we mess with pages in 1728 * any of the below states. 1729 */ 1730 if (m->wire_count || 1731 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 1732 m->valid != VM_PAGE_BITS_ALL 1733 ) { 1734 vm_page_wakeup(m); 1735 continue; 1736 } 1737 1738 /* 1739 * Theoretically once a page is known not to be busy, an 1740 * interrupt cannot come along and rip it out from under us. 1741 */ 1742 1743 if (advise == MADV_WILLNEED) { 1744 vm_page_activate(m); 1745 } else if (advise == MADV_DONTNEED) { 1746 vm_page_dontneed(m); 1747 } else if (advise == MADV_FREE) { 1748 /* 1749 * Mark the page clean. This will allow the page 1750 * to be freed up by the system. However, such pages 1751 * are often reused quickly by malloc()/free() 1752 * so we do not do anything that would cause 1753 * a page fault if we can help it. 1754 * 1755 * Specifically, we do not try to actually free 1756 * the page now nor do we try to put it in the 1757 * cache (which would cause a page fault on reuse). 1758 * 1759 * But we do make the page is freeable as we 1760 * can without actually taking the step of unmapping 1761 * it. 1762 */ 1763 pmap_clear_modify(m); 1764 m->dirty = 0; 1765 m->act_count = 0; 1766 vm_page_dontneed(m); 1767 if (tobject->type == OBJT_SWAP) 1768 swap_pager_freespace(tobject, tpindex, 1); 1769 } 1770 vm_page_wakeup(m); 1771 } 1772 if (tobject != object) 1773 vm_object_drop(tobject); 1774 vm_object_drop(object); 1775 } 1776 1777 /* 1778 * Create a new object which is backed by the specified existing object 1779 * range. Replace the pointer and offset that was pointing at the existing 1780 * object with the pointer/offset for the new object. 1781 * 1782 * If addref is non-zero the returned object is given an additional reference. 1783 * This mechanic exists to avoid the situation where refs might be 1 and 1784 * race against a collapse when the caller intends to bump it. So the 1785 * caller cannot add the ref after the fact. Used when the caller is 1786 * duplicating a vm_map_entry. 1787 * 1788 * No other requirements. 1789 */ 1790 void 1791 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length, 1792 int addref) 1793 { 1794 vm_object_t source; 1795 vm_object_t result; 1796 int useshadowlist; 1797 1798 source = *objectp; 1799 1800 /* 1801 * Don't create the new object if the old object isn't shared. 1802 * We have to chain wait before adding the reference to avoid 1803 * racing a collapse or deallocation. 1804 * 1805 * Clear OBJ_ONEMAPPING flag when shadowing. 1806 * 1807 * The caller owns a ref on source via *objectp which we are going 1808 * to replace. This ref is inherited by the backing_object assignment. 1809 * from nobject and does not need to be incremented here. 1810 * 1811 * However, we add a temporary extra reference to the original source 1812 * prior to holding nobject in case we block, to avoid races where 1813 * someone else might believe that the source can be collapsed. 1814 */ 1815 useshadowlist = 0; 1816 if (source) { 1817 if (source->type != OBJT_VNODE) { 1818 useshadowlist = 1; 1819 vm_object_hold(source); 1820 vm_object_chain_wait(source, 0); 1821 if (source->ref_count == 1 && 1822 source->handle == NULL && 1823 (source->type == OBJT_DEFAULT || 1824 source->type == OBJT_SWAP)) { 1825 if (addref) { 1826 vm_object_reference_locked(source); 1827 vm_object_clear_flag(source, 1828 OBJ_ONEMAPPING); 1829 } 1830 vm_object_drop(source); 1831 return; 1832 } 1833 vm_object_reference_locked(source); 1834 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1835 } else { 1836 vm_object_reference_quick(source); 1837 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1838 } 1839 } 1840 1841 /* 1842 * Allocate a new object with the given length. The new object 1843 * is returned referenced but we may have to add another one. 1844 * If we are adding a second reference we must clear OBJ_ONEMAPPING. 1845 * (typically because the caller is about to clone a vm_map_entry). 1846 * 1847 * The source object currently has an extra reference to prevent 1848 * collapses into it while we mess with its shadow list, which 1849 * we will remove later in this routine. 1850 * 1851 * The target object may require a second reference if asked for one 1852 * by the caller. 1853 */ 1854 result = vm_object_allocate(OBJT_DEFAULT, length); 1855 if (result == NULL) 1856 panic("vm_object_shadow: no object for shadowing"); 1857 vm_object_hold(result); 1858 if (addref) { 1859 vm_object_reference_locked(result); 1860 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1861 } 1862 1863 /* 1864 * The new object shadows the source object. Chain wait before 1865 * adjusting shadow_count or the shadow list to avoid races. 1866 * 1867 * Try to optimize the result object's page color when shadowing 1868 * in order to maintain page coloring consistency in the combined 1869 * shadowed object. 1870 * 1871 * The backing_object reference to source requires adding a ref to 1872 * source. We simply inherit the ref from the original *objectp 1873 * (which we are replacing) so no additional refs need to be added. 1874 * (we must still clean up the extra ref we had to prevent collapse 1875 * races). 1876 * 1877 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS 1878 */ 1879 KKASSERT(result->backing_object == NULL); 1880 result->backing_object = source; 1881 if (source) { 1882 if (useshadowlist) { 1883 vm_object_chain_wait(source, 0); 1884 LIST_INSERT_HEAD(&source->shadow_head, 1885 result, shadow_list); 1886 source->shadow_count++; 1887 source->generation++; 1888 vm_object_set_flag(result, OBJ_ONSHADOW); 1889 } 1890 /* cpu localization twist */ 1891 result->pg_color = (int)(intptr_t)curthread; 1892 } 1893 1894 /* 1895 * Adjust the return storage. Drop the ref on source before 1896 * returning. 1897 */ 1898 result->backing_object_offset = *offset; 1899 vm_object_drop(result); 1900 *offset = 0; 1901 if (source) { 1902 if (useshadowlist) { 1903 vm_object_deallocate_locked(source); 1904 vm_object_drop(source); 1905 } else { 1906 vm_object_deallocate(source); 1907 } 1908 } 1909 1910 /* 1911 * Return the new things 1912 */ 1913 *objectp = result; 1914 } 1915 1916 #define OBSC_TEST_ALL_SHADOWED 0x0001 1917 #define OBSC_COLLAPSE_NOWAIT 0x0002 1918 #define OBSC_COLLAPSE_WAIT 0x0004 1919 1920 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1921 1922 /* 1923 * The caller must hold the object. 1924 */ 1925 static __inline int 1926 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op) 1927 { 1928 struct rb_vm_page_scan_info info; 1929 int n; 1930 1931 vm_object_assert_held(object); 1932 vm_object_assert_held(backing_object); 1933 1934 KKASSERT(backing_object == object->backing_object); 1935 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1936 1937 /* 1938 * Initial conditions 1939 */ 1940 if (op & OBSC_TEST_ALL_SHADOWED) { 1941 /* 1942 * We do not want to have to test for the existence of 1943 * swap pages in the backing object. XXX but with the 1944 * new swapper this would be pretty easy to do. 1945 * 1946 * XXX what about anonymous MAP_SHARED memory that hasn't 1947 * been ZFOD faulted yet? If we do not test for this, the 1948 * shadow test may succeed! XXX 1949 */ 1950 if (backing_object->type != OBJT_DEFAULT) 1951 return(0); 1952 } 1953 if (op & OBSC_COLLAPSE_WAIT) { 1954 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1955 vm_object_set_flag(backing_object, OBJ_DEAD); 1956 1957 n = VMOBJ_HASH(backing_object); 1958 lwkt_gettoken(&vmobj_tokens[n]); 1959 TAILQ_REMOVE(&vm_object_lists[n], backing_object, object_list); 1960 lwkt_reltoken(&vmobj_tokens[n]); 1961 atomic_add_long(&vm_object_count, -1); 1962 } 1963 1964 /* 1965 * Our scan. We have to retry if a negative error code is returned, 1966 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1967 * the scan had to be stopped because the parent does not completely 1968 * shadow the child. 1969 */ 1970 info.object = object; 1971 info.backing_object = backing_object; 1972 info.limit = op; 1973 do { 1974 info.error = 1; 1975 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1976 vm_object_backing_scan_callback, 1977 &info); 1978 } while (info.error < 0); 1979 1980 return(info.error); 1981 } 1982 1983 /* 1984 * The caller must hold the object. 1985 */ 1986 static int 1987 vm_object_backing_scan_callback(vm_page_t p, void *data) 1988 { 1989 struct rb_vm_page_scan_info *info = data; 1990 vm_object_t backing_object; 1991 vm_object_t object; 1992 vm_pindex_t pindex; 1993 vm_pindex_t new_pindex; 1994 vm_pindex_t backing_offset_index; 1995 int op; 1996 1997 pindex = p->pindex; 1998 new_pindex = pindex - info->backing_offset_index; 1999 op = info->limit; 2000 object = info->object; 2001 backing_object = info->backing_object; 2002 backing_offset_index = info->backing_offset_index; 2003 2004 if (op & OBSC_TEST_ALL_SHADOWED) { 2005 vm_page_t pp; 2006 2007 /* 2008 * Ignore pages outside the parent object's range 2009 * and outside the parent object's mapping of the 2010 * backing object. 2011 * 2012 * note that we do not busy the backing object's 2013 * page. 2014 */ 2015 if (pindex < backing_offset_index || 2016 new_pindex >= object->size 2017 ) { 2018 return(0); 2019 } 2020 2021 /* 2022 * See if the parent has the page or if the parent's 2023 * object pager has the page. If the parent has the 2024 * page but the page is not valid, the parent's 2025 * object pager must have the page. 2026 * 2027 * If this fails, the parent does not completely shadow 2028 * the object and we might as well give up now. 2029 */ 2030 pp = vm_page_lookup(object, new_pindex); 2031 if ((pp == NULL || pp->valid == 0) && 2032 !vm_pager_has_page(object, new_pindex) 2033 ) { 2034 info->error = 0; /* problemo */ 2035 return(-1); /* stop the scan */ 2036 } 2037 } 2038 2039 /* 2040 * Check for busy page. Note that we may have lost (p) when we 2041 * possibly blocked above. 2042 */ 2043 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 2044 vm_page_t pp; 2045 2046 if (vm_page_busy_try(p, TRUE)) { 2047 if (op & OBSC_COLLAPSE_NOWAIT) { 2048 return(0); 2049 } else { 2050 /* 2051 * If we slept, anything could have 2052 * happened. Ask that the scan be restarted. 2053 * 2054 * Since the object is marked dead, the 2055 * backing offset should not have changed. 2056 */ 2057 vm_page_sleep_busy(p, TRUE, "vmocol"); 2058 info->error = -1; 2059 return(-1); 2060 } 2061 } 2062 2063 /* 2064 * If (p) is no longer valid restart the scan. 2065 */ 2066 if (p->object != backing_object || p->pindex != pindex) { 2067 kprintf("vm_object_backing_scan: Warning: page " 2068 "%p ripped out from under us\n", p); 2069 vm_page_wakeup(p); 2070 info->error = -1; 2071 return(-1); 2072 } 2073 2074 if (op & OBSC_COLLAPSE_NOWAIT) { 2075 if (p->valid == 0 || 2076 p->wire_count || 2077 (p->flags & PG_NEED_COMMIT)) { 2078 vm_page_wakeup(p); 2079 return(0); 2080 } 2081 } else { 2082 /* XXX what if p->valid == 0 , hold_count, etc? */ 2083 } 2084 2085 KASSERT( 2086 p->object == backing_object, 2087 ("vm_object_qcollapse(): object mismatch") 2088 ); 2089 2090 /* 2091 * Destroy any associated swap 2092 */ 2093 if (backing_object->type == OBJT_SWAP) 2094 swap_pager_freespace(backing_object, p->pindex, 1); 2095 2096 if ( 2097 p->pindex < backing_offset_index || 2098 new_pindex >= object->size 2099 ) { 2100 /* 2101 * Page is out of the parent object's range, we 2102 * can simply destroy it. 2103 */ 2104 vm_page_protect(p, VM_PROT_NONE); 2105 vm_page_free(p); 2106 return(0); 2107 } 2108 2109 pp = vm_page_lookup(object, new_pindex); 2110 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 2111 /* 2112 * page already exists in parent OR swap exists 2113 * for this location in the parent. Destroy 2114 * the original page from the backing object. 2115 * 2116 * Leave the parent's page alone 2117 */ 2118 vm_page_protect(p, VM_PROT_NONE); 2119 vm_page_free(p); 2120 return(0); 2121 } 2122 2123 /* 2124 * Page does not exist in parent, rename the 2125 * page from the backing object to the main object. 2126 * 2127 * If the page was mapped to a process, it can remain 2128 * mapped through the rename. 2129 */ 2130 if ((p->queue - p->pc) == PQ_CACHE) 2131 vm_page_deactivate(p); 2132 2133 vm_page_rename(p, object, new_pindex); 2134 vm_page_wakeup(p); 2135 /* page automatically made dirty by rename */ 2136 } 2137 return(0); 2138 } 2139 2140 /* 2141 * This version of collapse allows the operation to occur earlier and 2142 * when paging_in_progress is true for an object... This is not a complete 2143 * operation, but should plug 99.9% of the rest of the leaks. 2144 * 2145 * The caller must hold the object and backing_object and both must be 2146 * chainlocked. 2147 * 2148 * (only called from vm_object_collapse) 2149 */ 2150 static void 2151 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object) 2152 { 2153 if (backing_object->ref_count == 1) { 2154 atomic_add_int(&backing_object->ref_count, 2); 2155 #if defined(DEBUG_LOCKS) 2156 debugvm_object_add(backing_object, "qcollapse", 1, 2); 2157 #endif 2158 vm_object_backing_scan(object, backing_object, 2159 OBSC_COLLAPSE_NOWAIT); 2160 atomic_add_int(&backing_object->ref_count, -2); 2161 #if defined(DEBUG_LOCKS) 2162 debugvm_object_add(backing_object, "qcollapse", 2, -2); 2163 #endif 2164 } 2165 } 2166 2167 /* 2168 * Collapse an object with the object backing it. Pages in the backing 2169 * object are moved into the parent, and the backing object is deallocated. 2170 * Any conflict is resolved in favor of the parent's existing pages. 2171 * 2172 * object must be held and chain-locked on call. 2173 * 2174 * The caller must have an extra ref on object to prevent a race from 2175 * destroying it during the collapse. 2176 */ 2177 void 2178 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp) 2179 { 2180 struct vm_object_dealloc_list *dlist = NULL; 2181 vm_object_t backing_object; 2182 2183 /* 2184 * Only one thread is attempting a collapse at any given moment. 2185 * There are few restrictions for (object) that callers of this 2186 * function check so reentrancy is likely. 2187 */ 2188 KKASSERT(object != NULL); 2189 vm_object_assert_held(object); 2190 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL)); 2191 2192 for (;;) { 2193 vm_object_t bbobj; 2194 int dodealloc; 2195 2196 /* 2197 * We can only collapse a DEFAULT/SWAP object with a 2198 * DEFAULT/SWAP object. 2199 */ 2200 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) { 2201 backing_object = NULL; 2202 break; 2203 } 2204 2205 backing_object = object->backing_object; 2206 if (backing_object == NULL) 2207 break; 2208 if (backing_object->type != OBJT_DEFAULT && 2209 backing_object->type != OBJT_SWAP) { 2210 backing_object = NULL; 2211 break; 2212 } 2213 2214 /* 2215 * Hold the backing_object and check for races 2216 */ 2217 vm_object_hold(backing_object); 2218 if (backing_object != object->backing_object || 2219 (backing_object->type != OBJT_DEFAULT && 2220 backing_object->type != OBJT_SWAP)) { 2221 vm_object_drop(backing_object); 2222 continue; 2223 } 2224 2225 /* 2226 * Chain-lock the backing object too because if we 2227 * successfully merge its pages into the top object we 2228 * will collapse backing_object->backing_object as the 2229 * new backing_object. Re-check that it is still our 2230 * backing object. 2231 */ 2232 vm_object_chain_acquire(backing_object, 0); 2233 if (backing_object != object->backing_object) { 2234 vm_object_chain_release(backing_object); 2235 vm_object_drop(backing_object); 2236 continue; 2237 } 2238 2239 /* 2240 * we check the backing object first, because it is most likely 2241 * not collapsable. 2242 */ 2243 if (backing_object->handle != NULL || 2244 (backing_object->type != OBJT_DEFAULT && 2245 backing_object->type != OBJT_SWAP) || 2246 (backing_object->flags & OBJ_DEAD) || 2247 object->handle != NULL || 2248 (object->type != OBJT_DEFAULT && 2249 object->type != OBJT_SWAP) || 2250 (object->flags & OBJ_DEAD)) { 2251 break; 2252 } 2253 2254 /* 2255 * If paging is in progress we can't do a normal collapse. 2256 */ 2257 if ( 2258 object->paging_in_progress != 0 || 2259 backing_object->paging_in_progress != 0 2260 ) { 2261 vm_object_qcollapse(object, backing_object); 2262 break; 2263 } 2264 2265 /* 2266 * We know that we can either collapse the backing object (if 2267 * the parent is the only reference to it) or (perhaps) have 2268 * the parent bypass the object if the parent happens to shadow 2269 * all the resident pages in the entire backing object. 2270 * 2271 * This is ignoring pager-backed pages such as swap pages. 2272 * vm_object_backing_scan fails the shadowing test in this 2273 * case. 2274 */ 2275 if (backing_object->ref_count == 1) { 2276 /* 2277 * If there is exactly one reference to the backing 2278 * object, we can collapse it into the parent. 2279 */ 2280 KKASSERT(object->backing_object == backing_object); 2281 vm_object_backing_scan(object, backing_object, 2282 OBSC_COLLAPSE_WAIT); 2283 2284 /* 2285 * Move the pager from backing_object to object. 2286 */ 2287 if (backing_object->type == OBJT_SWAP) { 2288 vm_object_pip_add(backing_object, 1); 2289 2290 /* 2291 * scrap the paging_offset junk and do a 2292 * discrete copy. This also removes major 2293 * assumptions about how the swap-pager 2294 * works from where it doesn't belong. The 2295 * new swapper is able to optimize the 2296 * destroy-source case. 2297 */ 2298 vm_object_pip_add(object, 1); 2299 swap_pager_copy(backing_object, object, 2300 OFF_TO_IDX(object->backing_object_offset), 2301 TRUE); 2302 vm_object_pip_wakeup(object); 2303 vm_object_pip_wakeup(backing_object); 2304 } 2305 2306 /* 2307 * Object now shadows whatever backing_object did. 2308 * Remove object from backing_object's shadow_list. 2309 * 2310 * Removing object from backing_objects shadow list 2311 * requires releasing object, which we will do below. 2312 */ 2313 KKASSERT(object->backing_object == backing_object); 2314 if (object->flags & OBJ_ONSHADOW) { 2315 LIST_REMOVE(object, shadow_list); 2316 backing_object->shadow_count--; 2317 backing_object->generation++; 2318 vm_object_clear_flag(object, OBJ_ONSHADOW); 2319 } 2320 2321 /* 2322 * backing_object->backing_object moves from within 2323 * backing_object to within object. 2324 * 2325 * OBJT_VNODE bbobj's should have empty shadow lists. 2326 */ 2327 while ((bbobj = backing_object->backing_object) != NULL) { 2328 if (bbobj->type == OBJT_VNODE) 2329 vm_object_hold_shared(bbobj); 2330 else 2331 vm_object_hold(bbobj); 2332 if (bbobj == backing_object->backing_object) 2333 break; 2334 vm_object_drop(bbobj); 2335 } 2336 2337 /* 2338 * We are removing backing_object from bbobj's 2339 * shadow list and adding object to bbobj's shadow 2340 * list, so the ref_count on bbobj is unchanged. 2341 */ 2342 if (bbobj) { 2343 if (backing_object->flags & OBJ_ONSHADOW) { 2344 /* not locked exclusively if vnode */ 2345 KKASSERT(bbobj->type != OBJT_VNODE); 2346 LIST_REMOVE(backing_object, 2347 shadow_list); 2348 bbobj->shadow_count--; 2349 bbobj->generation++; 2350 vm_object_clear_flag(backing_object, 2351 OBJ_ONSHADOW); 2352 } 2353 backing_object->backing_object = NULL; 2354 } 2355 object->backing_object = bbobj; 2356 if (bbobj) { 2357 if (bbobj->type != OBJT_VNODE) { 2358 LIST_INSERT_HEAD(&bbobj->shadow_head, 2359 object, shadow_list); 2360 bbobj->shadow_count++; 2361 bbobj->generation++; 2362 vm_object_set_flag(object, 2363 OBJ_ONSHADOW); 2364 } 2365 } 2366 2367 object->backing_object_offset += 2368 backing_object->backing_object_offset; 2369 2370 vm_object_drop(bbobj); 2371 2372 /* 2373 * Discard the old backing_object. Nothing should be 2374 * able to ref it, other than a vm_map_split(), 2375 * and vm_map_split() will stall on our chain lock. 2376 * And we control the parent so it shouldn't be 2377 * possible for it to go away either. 2378 * 2379 * Since the backing object has no pages, no pager 2380 * left, and no object references within it, all 2381 * that is necessary is to dispose of it. 2382 */ 2383 KASSERT(backing_object->ref_count == 1, 2384 ("backing_object %p was somehow " 2385 "re-referenced during collapse!", 2386 backing_object)); 2387 KASSERT(RB_EMPTY(&backing_object->rb_memq), 2388 ("backing_object %p somehow has left " 2389 "over pages during collapse!", 2390 backing_object)); 2391 2392 /* 2393 * The object can be destroyed. 2394 * 2395 * XXX just fall through and dodealloc instead 2396 * of forcing destruction? 2397 */ 2398 atomic_add_int(&backing_object->ref_count, -1); 2399 #if defined(DEBUG_LOCKS) 2400 debugvm_object_add(backing_object, "collapse", 1, -1); 2401 #endif 2402 if ((backing_object->flags & OBJ_DEAD) == 0) 2403 vm_object_terminate(backing_object); 2404 object_collapses++; 2405 dodealloc = 0; 2406 } else { 2407 /* 2408 * If we do not entirely shadow the backing object, 2409 * there is nothing we can do so we give up. 2410 */ 2411 if (vm_object_backing_scan(object, backing_object, 2412 OBSC_TEST_ALL_SHADOWED) == 0) { 2413 break; 2414 } 2415 2416 /* 2417 * bbobj is backing_object->backing_object. Since 2418 * object completely shadows backing_object we can 2419 * bypass it and become backed by bbobj instead. 2420 * 2421 * The shadow list for vnode backing objects is not 2422 * used and a shared hold is allowed. 2423 */ 2424 while ((bbobj = backing_object->backing_object) != NULL) { 2425 if (bbobj->type == OBJT_VNODE) 2426 vm_object_hold_shared(bbobj); 2427 else 2428 vm_object_hold(bbobj); 2429 if (bbobj == backing_object->backing_object) 2430 break; 2431 vm_object_drop(bbobj); 2432 } 2433 2434 /* 2435 * Make object shadow bbobj instead of backing_object. 2436 * Remove object from backing_object's shadow list. 2437 * 2438 * Deallocating backing_object will not remove 2439 * it, since its reference count is at least 2. 2440 * 2441 * Removing object from backing_object's shadow 2442 * list requires releasing a ref, which we do 2443 * below by setting dodealloc to 1. 2444 */ 2445 KKASSERT(object->backing_object == backing_object); 2446 if (object->flags & OBJ_ONSHADOW) { 2447 LIST_REMOVE(object, shadow_list); 2448 backing_object->shadow_count--; 2449 backing_object->generation++; 2450 vm_object_clear_flag(object, OBJ_ONSHADOW); 2451 } 2452 2453 /* 2454 * Add a ref to bbobj, bbobj now shadows object. 2455 * 2456 * NOTE: backing_object->backing_object still points 2457 * to bbobj. That relationship remains intact 2458 * because backing_object has > 1 ref, so 2459 * someone else is pointing to it (hence why 2460 * we can't collapse it into object and can 2461 * only handle the all-shadowed bypass case). 2462 */ 2463 if (bbobj) { 2464 if (bbobj->type != OBJT_VNODE) { 2465 vm_object_chain_wait(bbobj, 0); 2466 vm_object_reference_locked(bbobj); 2467 LIST_INSERT_HEAD(&bbobj->shadow_head, 2468 object, shadow_list); 2469 bbobj->shadow_count++; 2470 bbobj->generation++; 2471 vm_object_set_flag(object, 2472 OBJ_ONSHADOW); 2473 } else { 2474 vm_object_reference_quick(bbobj); 2475 } 2476 object->backing_object_offset += 2477 backing_object->backing_object_offset; 2478 object->backing_object = bbobj; 2479 vm_object_drop(bbobj); 2480 } else { 2481 object->backing_object = NULL; 2482 } 2483 2484 /* 2485 * Drop the reference count on backing_object. To 2486 * handle ref_count races properly we can't assume 2487 * that the ref_count is still at least 2 so we 2488 * have to actually call vm_object_deallocate() 2489 * (after clearing the chainlock). 2490 */ 2491 object_bypasses++; 2492 dodealloc = 1; 2493 } 2494 2495 /* 2496 * Ok, we want to loop on the new object->bbobj association, 2497 * possibly collapsing it further. However if dodealloc is 2498 * non-zero we have to deallocate the backing_object which 2499 * itself can potentially undergo a collapse, creating a 2500 * recursion depth issue with the LWKT token subsystem. 2501 * 2502 * In the case where we must deallocate the backing_object 2503 * it is possible now that the backing_object has a single 2504 * shadow count on some other object (not represented here 2505 * as yet), since it no longer shadows us. Thus when we 2506 * call vm_object_deallocate() it may attempt to collapse 2507 * itself into its remaining parent. 2508 */ 2509 if (dodealloc) { 2510 struct vm_object_dealloc_list *dtmp; 2511 2512 vm_object_chain_release(backing_object); 2513 vm_object_unlock(backing_object); 2514 /* backing_object remains held */ 2515 2516 /* 2517 * Auto-deallocation list for caller convenience. 2518 */ 2519 if (dlistp == NULL) 2520 dlistp = &dlist; 2521 2522 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK); 2523 dtmp->object = backing_object; 2524 dtmp->next = *dlistp; 2525 *dlistp = dtmp; 2526 } else { 2527 vm_object_chain_release(backing_object); 2528 vm_object_drop(backing_object); 2529 } 2530 /* backing_object = NULL; not needed */ 2531 /* loop */ 2532 } 2533 2534 /* 2535 * Clean up any left over backing_object 2536 */ 2537 if (backing_object) { 2538 vm_object_chain_release(backing_object); 2539 vm_object_drop(backing_object); 2540 } 2541 2542 /* 2543 * Clean up any auto-deallocation list. This is a convenience 2544 * for top-level callers so they don't have to pass &dlist. 2545 * Do not clean up any caller-passed dlistp, the caller will 2546 * do that. 2547 */ 2548 if (dlist) 2549 vm_object_deallocate_list(&dlist); 2550 2551 } 2552 2553 /* 2554 * vm_object_collapse() may collect additional objects in need of 2555 * deallocation. This routine deallocates these objects. The 2556 * deallocation itself can trigger additional collapses (which the 2557 * deallocate function takes care of). This procedure is used to 2558 * reduce procedural recursion since these vm_object shadow chains 2559 * can become quite long. 2560 */ 2561 void 2562 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp) 2563 { 2564 struct vm_object_dealloc_list *dlist; 2565 2566 while ((dlist = *dlistp) != NULL) { 2567 *dlistp = dlist->next; 2568 vm_object_lock(dlist->object); 2569 vm_object_deallocate_locked(dlist->object); 2570 vm_object_drop(dlist->object); 2571 kfree(dlist, M_TEMP); 2572 } 2573 } 2574 2575 /* 2576 * Removes all physical pages in the specified object range from the 2577 * object's list of pages. 2578 * 2579 * No requirements. 2580 */ 2581 static int vm_object_page_remove_callback(vm_page_t p, void *data); 2582 2583 void 2584 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2585 boolean_t clean_only) 2586 { 2587 struct rb_vm_page_scan_info info; 2588 int all; 2589 2590 /* 2591 * Degenerate cases and assertions 2592 */ 2593 vm_object_hold(object); 2594 if (object == NULL || 2595 (object->resident_page_count == 0 && object->swblock_count == 0)) { 2596 vm_object_drop(object); 2597 return; 2598 } 2599 KASSERT(object->type != OBJT_PHYS, 2600 ("attempt to remove pages from a physical object")); 2601 2602 /* 2603 * Indicate that paging is occuring on the object 2604 */ 2605 vm_object_pip_add(object, 1); 2606 2607 /* 2608 * Figure out the actual removal range and whether we are removing 2609 * the entire contents of the object or not. If removing the entire 2610 * contents, be sure to get all pages, even those that might be 2611 * beyond the end of the object. 2612 */ 2613 info.start_pindex = start; 2614 if (end == 0) 2615 info.end_pindex = (vm_pindex_t)-1; 2616 else 2617 info.end_pindex = end - 1; 2618 info.limit = clean_only; 2619 all = (start == 0 && info.end_pindex >= object->size - 1); 2620 2621 /* 2622 * Loop until we are sure we have gotten them all. 2623 */ 2624 do { 2625 info.error = 0; 2626 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2627 vm_object_page_remove_callback, &info); 2628 } while (info.error); 2629 2630 /* 2631 * Remove any related swap if throwing away pages, or for 2632 * non-swap objects (the swap is a clean copy in that case). 2633 */ 2634 if (object->type != OBJT_SWAP || clean_only == FALSE) { 2635 if (all) 2636 swap_pager_freespace_all(object); 2637 else 2638 swap_pager_freespace(object, info.start_pindex, 2639 info.end_pindex - info.start_pindex + 1); 2640 } 2641 2642 /* 2643 * Cleanup 2644 */ 2645 vm_object_pip_wakeup(object); 2646 vm_object_drop(object); 2647 } 2648 2649 /* 2650 * The caller must hold the object 2651 */ 2652 static int 2653 vm_object_page_remove_callback(vm_page_t p, void *data) 2654 { 2655 struct rb_vm_page_scan_info *info = data; 2656 2657 if (vm_page_busy_try(p, TRUE)) { 2658 vm_page_sleep_busy(p, TRUE, "vmopar"); 2659 info->error = 1; 2660 return(0); 2661 } 2662 2663 /* 2664 * Wired pages cannot be destroyed, but they can be invalidated 2665 * and we do so if clean_only (limit) is not set. 2666 * 2667 * WARNING! The page may be wired due to being part of a buffer 2668 * cache buffer, and the buffer might be marked B_CACHE. 2669 * This is fine as part of a truncation but VFSs must be 2670 * sure to fix the buffer up when re-extending the file. 2671 * 2672 * NOTE! PG_NEED_COMMIT is ignored. 2673 */ 2674 if (p->wire_count != 0) { 2675 vm_page_protect(p, VM_PROT_NONE); 2676 if (info->limit == 0) 2677 p->valid = 0; 2678 vm_page_wakeup(p); 2679 return(0); 2680 } 2681 2682 /* 2683 * limit is our clean_only flag. If set and the page is dirty or 2684 * requires a commit, do not free it. If set and the page is being 2685 * held by someone, do not free it. 2686 */ 2687 if (info->limit && p->valid) { 2688 vm_page_test_dirty(p); 2689 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) { 2690 vm_page_wakeup(p); 2691 return(0); 2692 } 2693 } 2694 2695 /* 2696 * Destroy the page 2697 */ 2698 vm_page_protect(p, VM_PROT_NONE); 2699 vm_page_free(p); 2700 return(0); 2701 } 2702 2703 /* 2704 * Coalesces two objects backing up adjoining regions of memory into a 2705 * single object. 2706 * 2707 * returns TRUE if objects were combined. 2708 * 2709 * NOTE: Only works at the moment if the second object is NULL - 2710 * if it's not, which object do we lock first? 2711 * 2712 * Parameters: 2713 * prev_object First object to coalesce 2714 * prev_offset Offset into prev_object 2715 * next_object Second object into coalesce 2716 * next_offset Offset into next_object 2717 * 2718 * prev_size Size of reference to prev_object 2719 * next_size Size of reference to next_object 2720 * 2721 * The caller does not need to hold (prev_object) but must have a stable 2722 * pointer to it (typically by holding the vm_map locked). 2723 */ 2724 boolean_t 2725 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 2726 vm_size_t prev_size, vm_size_t next_size) 2727 { 2728 vm_pindex_t next_pindex; 2729 2730 if (prev_object == NULL) 2731 return (TRUE); 2732 2733 vm_object_hold(prev_object); 2734 2735 if (prev_object->type != OBJT_DEFAULT && 2736 prev_object->type != OBJT_SWAP) { 2737 vm_object_drop(prev_object); 2738 return (FALSE); 2739 } 2740 2741 /* 2742 * Try to collapse the object first 2743 */ 2744 vm_object_chain_acquire(prev_object, 0); 2745 vm_object_collapse(prev_object, NULL); 2746 2747 /* 2748 * Can't coalesce if: . more than one reference . paged out . shadows 2749 * another object . has a copy elsewhere (any of which mean that the 2750 * pages not mapped to prev_entry may be in use anyway) 2751 */ 2752 2753 if (prev_object->backing_object != NULL) { 2754 vm_object_chain_release(prev_object); 2755 vm_object_drop(prev_object); 2756 return (FALSE); 2757 } 2758 2759 prev_size >>= PAGE_SHIFT; 2760 next_size >>= PAGE_SHIFT; 2761 next_pindex = prev_pindex + prev_size; 2762 2763 if ((prev_object->ref_count > 1) && 2764 (prev_object->size != next_pindex)) { 2765 vm_object_chain_release(prev_object); 2766 vm_object_drop(prev_object); 2767 return (FALSE); 2768 } 2769 2770 /* 2771 * Remove any pages that may still be in the object from a previous 2772 * deallocation. 2773 */ 2774 if (next_pindex < prev_object->size) { 2775 vm_object_page_remove(prev_object, 2776 next_pindex, 2777 next_pindex + next_size, FALSE); 2778 if (prev_object->type == OBJT_SWAP) 2779 swap_pager_freespace(prev_object, 2780 next_pindex, next_size); 2781 } 2782 2783 /* 2784 * Extend the object if necessary. 2785 */ 2786 if (next_pindex + next_size > prev_object->size) 2787 prev_object->size = next_pindex + next_size; 2788 2789 vm_object_chain_release(prev_object); 2790 vm_object_drop(prev_object); 2791 return (TRUE); 2792 } 2793 2794 /* 2795 * Make the object writable and flag is being possibly dirty. 2796 * 2797 * The object might not be held (or might be held but held shared), 2798 * the related vnode is probably not held either. Object and vnode are 2799 * stable by virtue of the vm_page busied by the caller preventing 2800 * destruction. 2801 * 2802 * If the related mount is flagged MNTK_THR_SYNC we need to call 2803 * vsetobjdirty(). Filesystems using this option usually shortcut 2804 * synchronization by only scanning the syncer list. 2805 */ 2806 void 2807 vm_object_set_writeable_dirty(vm_object_t object) 2808 { 2809 struct vnode *vp; 2810 2811 /*vm_object_assert_held(object);*/ 2812 /* 2813 * Avoid contention in vm fault path by checking the state before 2814 * issuing an atomic op on it. 2815 */ 2816 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) != 2817 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) { 2818 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 2819 } 2820 if (object->type == OBJT_VNODE && 2821 (vp = (struct vnode *)object->handle) != NULL) { 2822 if ((vp->v_flag & VOBJDIRTY) == 0) { 2823 if (vp->v_mount && 2824 (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) { 2825 /* 2826 * New style THR_SYNC places vnodes on the 2827 * syncer list more deterministically. 2828 */ 2829 vsetobjdirty(vp); 2830 } else { 2831 /* 2832 * Old style scan would not necessarily place 2833 * a vnode on the syncer list when possibly 2834 * modified via mmap. 2835 */ 2836 vsetflags(vp, VOBJDIRTY); 2837 } 2838 } 2839 } 2840 } 2841 2842 #include "opt_ddb.h" 2843 #ifdef DDB 2844 #include <sys/kernel.h> 2845 2846 #include <sys/cons.h> 2847 2848 #include <ddb/ddb.h> 2849 2850 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 2851 vm_map_entry_t entry); 2852 static int vm_object_in_map (vm_object_t object); 2853 2854 /* 2855 * The caller must hold the object. 2856 */ 2857 static int 2858 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2859 { 2860 vm_map_t tmpm; 2861 vm_map_entry_t tmpe; 2862 vm_object_t obj, nobj; 2863 int entcount; 2864 2865 if (map == 0) 2866 return 0; 2867 if (entry == 0) { 2868 tmpe = map->header.next; 2869 entcount = map->nentries; 2870 while (entcount-- && (tmpe != &map->header)) { 2871 if( _vm_object_in_map(map, object, tmpe)) { 2872 return 1; 2873 } 2874 tmpe = tmpe->next; 2875 } 2876 return (0); 2877 } 2878 switch(entry->maptype) { 2879 case VM_MAPTYPE_SUBMAP: 2880 tmpm = entry->object.sub_map; 2881 tmpe = tmpm->header.next; 2882 entcount = tmpm->nentries; 2883 while (entcount-- && tmpe != &tmpm->header) { 2884 if( _vm_object_in_map(tmpm, object, tmpe)) { 2885 return 1; 2886 } 2887 tmpe = tmpe->next; 2888 } 2889 break; 2890 case VM_MAPTYPE_NORMAL: 2891 case VM_MAPTYPE_VPAGETABLE: 2892 obj = entry->object.vm_object; 2893 while (obj) { 2894 if (obj == object) { 2895 if (obj != entry->object.vm_object) 2896 vm_object_drop(obj); 2897 return 1; 2898 } 2899 while ((nobj = obj->backing_object) != NULL) { 2900 vm_object_hold(nobj); 2901 if (nobj == obj->backing_object) 2902 break; 2903 vm_object_drop(nobj); 2904 } 2905 if (obj != entry->object.vm_object) { 2906 if (nobj) 2907 vm_object_lock_swap(); 2908 vm_object_drop(obj); 2909 } 2910 obj = nobj; 2911 } 2912 break; 2913 default: 2914 break; 2915 } 2916 return 0; 2917 } 2918 2919 static int vm_object_in_map_callback(struct proc *p, void *data); 2920 2921 struct vm_object_in_map_info { 2922 vm_object_t object; 2923 int rv; 2924 }; 2925 2926 /* 2927 * Debugging only 2928 */ 2929 static int 2930 vm_object_in_map(vm_object_t object) 2931 { 2932 struct vm_object_in_map_info info; 2933 2934 info.rv = 0; 2935 info.object = object; 2936 2937 allproc_scan(vm_object_in_map_callback, &info); 2938 if (info.rv) 2939 return 1; 2940 if( _vm_object_in_map(&kernel_map, object, 0)) 2941 return 1; 2942 if( _vm_object_in_map(&pager_map, object, 0)) 2943 return 1; 2944 if( _vm_object_in_map(&buffer_map, object, 0)) 2945 return 1; 2946 return 0; 2947 } 2948 2949 /* 2950 * Debugging only 2951 */ 2952 static int 2953 vm_object_in_map_callback(struct proc *p, void *data) 2954 { 2955 struct vm_object_in_map_info *info = data; 2956 2957 if (p->p_vmspace) { 2958 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 2959 info->rv = 1; 2960 return -1; 2961 } 2962 } 2963 return (0); 2964 } 2965 2966 DB_SHOW_COMMAND(vmochk, vm_object_check) 2967 { 2968 vm_object_t object; 2969 int n; 2970 2971 /* 2972 * make sure that internal objs are in a map somewhere 2973 * and none have zero ref counts. 2974 */ 2975 for (n = 0; n < VMOBJ_HSIZE; ++n) { 2976 for (object = TAILQ_FIRST(&vm_object_lists[n]); 2977 object != NULL; 2978 object = TAILQ_NEXT(object, object_list)) { 2979 if (object->type == OBJT_MARKER) 2980 continue; 2981 if (object->handle != NULL || 2982 (object->type != OBJT_DEFAULT && 2983 object->type != OBJT_SWAP)) { 2984 continue; 2985 } 2986 if (object->ref_count == 0) { 2987 db_printf("vmochk: internal obj has " 2988 "zero ref count: %ld\n", 2989 (long)object->size); 2990 } 2991 if (vm_object_in_map(object)) 2992 continue; 2993 db_printf("vmochk: internal obj is not in a map: " 2994 "ref: %d, size: %lu: 0x%lx, " 2995 "backing_object: %p\n", 2996 object->ref_count, (u_long)object->size, 2997 (u_long)object->size, 2998 (void *)object->backing_object); 2999 } 3000 } 3001 } 3002 3003 /* 3004 * Debugging only 3005 */ 3006 DB_SHOW_COMMAND(object, vm_object_print_static) 3007 { 3008 /* XXX convert args. */ 3009 vm_object_t object = (vm_object_t)addr; 3010 boolean_t full = have_addr; 3011 3012 vm_page_t p; 3013 3014 /* XXX count is an (unused) arg. Avoid shadowing it. */ 3015 #define count was_count 3016 3017 int count; 3018 3019 if (object == NULL) 3020 return; 3021 3022 db_iprintf( 3023 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 3024 object, (int)object->type, (u_long)object->size, 3025 object->resident_page_count, object->ref_count, object->flags); 3026 /* 3027 * XXX no %qd in kernel. Truncate object->backing_object_offset. 3028 */ 3029 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 3030 object->shadow_count, 3031 object->backing_object ? object->backing_object->ref_count : 0, 3032 object->backing_object, (long)object->backing_object_offset); 3033 3034 if (!full) 3035 return; 3036 3037 db_indent += 2; 3038 count = 0; 3039 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 3040 if (count == 0) 3041 db_iprintf("memory:="); 3042 else if (count == 6) { 3043 db_printf("\n"); 3044 db_iprintf(" ..."); 3045 count = 0; 3046 } else 3047 db_printf(","); 3048 count++; 3049 3050 db_printf("(off=0x%lx,page=0x%lx)", 3051 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 3052 } 3053 if (count != 0) 3054 db_printf("\n"); 3055 db_indent -= 2; 3056 } 3057 3058 /* XXX. */ 3059 #undef count 3060 3061 /* 3062 * XXX need this non-static entry for calling from vm_map_print. 3063 * 3064 * Debugging only 3065 */ 3066 void 3067 vm_object_print(/* db_expr_t */ long addr, 3068 boolean_t have_addr, 3069 /* db_expr_t */ long count, 3070 char *modif) 3071 { 3072 vm_object_print_static(addr, have_addr, count, modif); 3073 } 3074 3075 /* 3076 * Debugging only 3077 */ 3078 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 3079 { 3080 vm_object_t object; 3081 int nl = 0; 3082 int c; 3083 int n; 3084 3085 for (n = 0; n < VMOBJ_HSIZE; ++n) { 3086 for (object = TAILQ_FIRST(&vm_object_lists[n]); 3087 object != NULL; 3088 object = TAILQ_NEXT(object, object_list)) { 3089 vm_pindex_t idx, fidx; 3090 vm_pindex_t osize; 3091 vm_paddr_t pa = -1, padiff; 3092 int rcount; 3093 vm_page_t m; 3094 3095 if (object->type == OBJT_MARKER) 3096 continue; 3097 db_printf("new object: %p\n", (void *)object); 3098 if ( nl > 18) { 3099 c = cngetc(); 3100 if (c != ' ') 3101 return; 3102 nl = 0; 3103 } 3104 nl++; 3105 rcount = 0; 3106 fidx = 0; 3107 osize = object->size; 3108 if (osize > 128) 3109 osize = 128; 3110 for (idx = 0; idx < osize; idx++) { 3111 m = vm_page_lookup(object, idx); 3112 if (m == NULL) { 3113 if (rcount) { 3114 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 3115 (long)fidx, rcount, (long)pa); 3116 if ( nl > 18) { 3117 c = cngetc(); 3118 if (c != ' ') 3119 return; 3120 nl = 0; 3121 } 3122 nl++; 3123 rcount = 0; 3124 } 3125 continue; 3126 } 3127 3128 if (rcount && 3129 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 3130 ++rcount; 3131 continue; 3132 } 3133 if (rcount) { 3134 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 3135 padiff >>= PAGE_SHIFT; 3136 padiff &= PQ_L2_MASK; 3137 if (padiff == 0) { 3138 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 3139 ++rcount; 3140 continue; 3141 } 3142 db_printf(" index(%ld)run(%d)pa(0x%lx)", 3143 (long)fidx, rcount, (long)pa); 3144 db_printf("pd(%ld)\n", (long)padiff); 3145 if ( nl > 18) { 3146 c = cngetc(); 3147 if (c != ' ') 3148 return; 3149 nl = 0; 3150 } 3151 nl++; 3152 } 3153 fidx = idx; 3154 pa = VM_PAGE_TO_PHYS(m); 3155 rcount = 1; 3156 } 3157 if (rcount) { 3158 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 3159 (long)fidx, rcount, (long)pa); 3160 if ( nl > 18) { 3161 c = cngetc(); 3162 if (c != ' ') 3163 return; 3164 nl = 0; 3165 } 3166 nl++; 3167 } 3168 } 3169 } 3170 } 3171 #endif /* DDB */ 3172