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 object_collapses; 134 static long object_bypasses; 135 136 struct vm_object_hash vm_object_hash[VMOBJ_HSIZE]; 137 138 MALLOC_DEFINE(M_VM_OBJECT, "vm_object", "vm_object structures"); 139 140 #if defined(DEBUG_LOCKS) 141 142 #define vm_object_vndeallocate(obj, vpp) \ 143 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__) 144 145 /* 146 * Debug helper to track hold/drop/ref/deallocate calls. 147 */ 148 static void 149 debugvm_object_add(vm_object_t obj, char *file, int line, int addrem) 150 { 151 int i; 152 153 i = atomic_fetchadd_int(&obj->debug_index, 1); 154 i = i & (VMOBJ_DEBUG_ARRAY_SIZE - 1); 155 ksnprintf(obj->debug_hold_thrs[i], 156 sizeof(obj->debug_hold_thrs[i]), 157 "%c%d:(%d):%s", 158 (addrem == -1 ? '-' : (addrem == 1 ? '+' : '=')), 159 (curthread->td_proc ? curthread->td_proc->p_pid : -1), 160 obj->ref_count, 161 curthread->td_comm); 162 obj->debug_hold_file[i] = file; 163 obj->debug_hold_line[i] = line; 164 #if 0 165 /* Uncomment for debugging obj refs/derefs in reproducable cases */ 166 if (strcmp(curthread->td_comm, "sshd") == 0) { 167 kprintf("%d %p refs=%d ar=%d file: %s/%d\n", 168 (curthread->td_proc ? curthread->td_proc->p_pid : -1), 169 obj, obj->ref_count, addrem, file, line); 170 } 171 #endif 172 } 173 174 #endif 175 176 /* 177 * Misc low level routines 178 */ 179 static void 180 vm_object_lock_init(vm_object_t obj) 181 { 182 #if defined(DEBUG_LOCKS) 183 int i; 184 185 obj->debug_index = 0; 186 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 187 obj->debug_hold_thrs[i][0] = 0; 188 obj->debug_hold_file[i] = NULL; 189 obj->debug_hold_line[i] = 0; 190 } 191 #endif 192 } 193 194 void 195 vm_object_lock_swap(void) 196 { 197 lwkt_token_swap(); 198 } 199 200 void 201 vm_object_lock(vm_object_t obj) 202 { 203 lwkt_gettoken(&obj->token); 204 } 205 206 /* 207 * Returns TRUE on sucesss 208 */ 209 static int 210 vm_object_lock_try(vm_object_t obj) 211 { 212 return(lwkt_trytoken(&obj->token)); 213 } 214 215 void 216 vm_object_lock_shared(vm_object_t obj) 217 { 218 lwkt_gettoken_shared(&obj->token); 219 } 220 221 void 222 vm_object_unlock(vm_object_t obj) 223 { 224 lwkt_reltoken(&obj->token); 225 } 226 227 void 228 vm_object_upgrade(vm_object_t obj) 229 { 230 lwkt_reltoken(&obj->token); 231 lwkt_gettoken(&obj->token); 232 } 233 234 void 235 vm_object_downgrade(vm_object_t obj) 236 { 237 lwkt_reltoken(&obj->token); 238 lwkt_gettoken_shared(&obj->token); 239 } 240 241 static __inline void 242 vm_object_assert_held(vm_object_t obj) 243 { 244 ASSERT_LWKT_TOKEN_HELD(&obj->token); 245 } 246 247 static __inline int 248 vm_quickcolor(void) 249 { 250 globaldata_t gd = mycpu; 251 int pg_color; 252 253 pg_color = (int)(intptr_t)gd->gd_curthread >> 10; 254 pg_color += gd->gd_quick_color; 255 gd->gd_quick_color += PQ_PRIME2; 256 257 return pg_color; 258 } 259 260 void 261 VMOBJDEBUG(vm_object_hold)(vm_object_t obj VMOBJDBARGS) 262 { 263 KKASSERT(obj != NULL); 264 265 /* 266 * Object must be held (object allocation is stable due to callers 267 * context, typically already holding the token on a parent object) 268 * prior to potentially blocking on the lock, otherwise the object 269 * can get ripped away from us. 270 */ 271 refcount_acquire(&obj->hold_count); 272 vm_object_lock(obj); 273 274 #if defined(DEBUG_LOCKS) 275 debugvm_object_add(obj, file, line, 1); 276 #endif 277 } 278 279 int 280 VMOBJDEBUG(vm_object_hold_try)(vm_object_t obj VMOBJDBARGS) 281 { 282 KKASSERT(obj != NULL); 283 284 /* 285 * Object must be held (object allocation is stable due to callers 286 * context, typically already holding the token on a parent object) 287 * prior to potentially blocking on the lock, otherwise the object 288 * can get ripped away from us. 289 */ 290 refcount_acquire(&obj->hold_count); 291 if (vm_object_lock_try(obj) == 0) { 292 if (refcount_release(&obj->hold_count)) { 293 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) 294 kfree(obj, M_VM_OBJECT); 295 } 296 return(0); 297 } 298 299 #if defined(DEBUG_LOCKS) 300 debugvm_object_add(obj, file, line, 1); 301 #endif 302 return(1); 303 } 304 305 void 306 VMOBJDEBUG(vm_object_hold_shared)(vm_object_t obj VMOBJDBARGS) 307 { 308 KKASSERT(obj != NULL); 309 310 /* 311 * Object must be held (object allocation is stable due to callers 312 * context, typically already holding the token on a parent object) 313 * prior to potentially blocking on the lock, otherwise the object 314 * can get ripped away from us. 315 */ 316 refcount_acquire(&obj->hold_count); 317 vm_object_lock_shared(obj); 318 319 #if defined(DEBUG_LOCKS) 320 debugvm_object_add(obj, file, line, 1); 321 #endif 322 } 323 324 /* 325 * Drop the token and hold_count on the object. 326 * 327 * WARNING! Token might be shared. 328 */ 329 void 330 VMOBJDEBUG(vm_object_drop)(vm_object_t obj VMOBJDBARGS) 331 { 332 if (obj == NULL) 333 return; 334 335 /* 336 * No new holders should be possible once we drop hold_count 1->0 as 337 * there is no longer any way to reference the object. 338 */ 339 KKASSERT(obj->hold_count > 0); 340 if (refcount_release(&obj->hold_count)) { 341 #if defined(DEBUG_LOCKS) 342 debugvm_object_add(obj, file, line, -1); 343 #endif 344 345 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) { 346 vm_object_unlock(obj); 347 kfree(obj, M_VM_OBJECT); 348 } else { 349 vm_object_unlock(obj); 350 } 351 } else { 352 #if defined(DEBUG_LOCKS) 353 debugvm_object_add(obj, file, line, -1); 354 #endif 355 vm_object_unlock(obj); 356 } 357 } 358 359 /* 360 * Initialize a freshly allocated object, returning a held object. 361 * 362 * Used only by vm_object_allocate() and zinitna(). 363 * 364 * No requirements. 365 */ 366 void 367 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 368 { 369 struct vm_object_hash *hash; 370 371 RB_INIT(&object->rb_memq); 372 LIST_INIT(&object->shadow_head); 373 lwkt_token_init(&object->token, "vmobj"); 374 375 object->type = type; 376 object->size = size; 377 object->ref_count = 1; 378 object->memattr = VM_MEMATTR_DEFAULT; 379 object->hold_count = 0; 380 object->flags = 0; 381 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 382 vm_object_set_flag(object, OBJ_ONEMAPPING); 383 object->paging_in_progress = 0; 384 object->resident_page_count = 0; 385 object->agg_pv_list_count = 0; 386 object->shadow_count = 0; 387 /* cpu localization twist */ 388 object->pg_color = vm_quickcolor(); 389 object->handle = NULL; 390 object->backing_object = NULL; 391 object->backing_object_offset = (vm_ooffset_t)0; 392 393 object->generation++; 394 object->swblock_count = 0; 395 RB_INIT(&object->swblock_root); 396 vm_object_lock_init(object); 397 pmap_object_init(object); 398 399 vm_object_hold(object); 400 401 hash = VMOBJ_HASH(object); 402 lwkt_gettoken(&hash->token); 403 TAILQ_INSERT_TAIL(&hash->list, object, object_list); 404 lwkt_reltoken(&hash->token); 405 } 406 407 /* 408 * Initialize the VM objects module. 409 * 410 * Called from the low level boot code only. Note that this occurs before 411 * kmalloc is initialized so we cannot allocate any VM objects. 412 */ 413 void 414 vm_object_init(void) 415 { 416 int i; 417 418 for (i = 0; i < VMOBJ_HSIZE; ++i) { 419 TAILQ_INIT(&vm_object_hash[i].list); 420 lwkt_token_init(&vm_object_hash[i].token, "vmobjlst"); 421 } 422 423 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 424 &kernel_object); 425 vm_object_drop(&kernel_object); 426 } 427 428 void 429 vm_object_init2(void) 430 { 431 kmalloc_set_unlimited(M_VM_OBJECT); 432 } 433 434 /* 435 * Allocate and return a new object of the specified type and size. 436 * 437 * No requirements. 438 */ 439 vm_object_t 440 vm_object_allocate(objtype_t type, vm_pindex_t size) 441 { 442 vm_object_t obj; 443 444 obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO); 445 _vm_object_allocate(type, size, obj); 446 vm_object_drop(obj); 447 448 return (obj); 449 } 450 451 /* 452 * This version returns a held object, allowing further atomic initialization 453 * of the object. 454 */ 455 vm_object_t 456 vm_object_allocate_hold(objtype_t type, vm_pindex_t size) 457 { 458 vm_object_t obj; 459 460 obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO); 461 _vm_object_allocate(type, size, obj); 462 463 return (obj); 464 } 465 466 /* 467 * Add an additional reference to a vm_object. The object must already be 468 * held. The original non-lock version is no longer supported. The object 469 * must NOT be chain locked by anyone at the time the reference is added. 470 * 471 * Referencing a chain-locked object can blow up the fairly sensitive 472 * ref_count and shadow_count tests in the deallocator. Most callers 473 * will call vm_object_chain_wait() prior to calling 474 * vm_object_reference_locked() to avoid the case. 475 * 476 * The object must be held, but may be held shared if desired (hence why 477 * we use an atomic op). 478 */ 479 void 480 VMOBJDEBUG(vm_object_reference_locked)(vm_object_t object VMOBJDBARGS) 481 { 482 KKASSERT(object != NULL); 483 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 484 KKASSERT((object->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) == 0); 485 atomic_add_int(&object->ref_count, 1); 486 if (object->type == OBJT_VNODE) { 487 vref(object->handle); 488 /* XXX what if the vnode is being destroyed? */ 489 } 490 #if defined(DEBUG_LOCKS) 491 debugvm_object_add(object, file, line, 1); 492 #endif 493 } 494 495 /* 496 * This version is only allowed for vnode objects. 497 */ 498 void 499 VMOBJDEBUG(vm_object_reference_quick)(vm_object_t object VMOBJDBARGS) 500 { 501 KKASSERT(object->type == OBJT_VNODE); 502 atomic_add_int(&object->ref_count, 1); 503 vref(object->handle); 504 #if defined(DEBUG_LOCKS) 505 debugvm_object_add(object, file, line, 1); 506 #endif 507 } 508 509 /* 510 * Object OBJ_CHAINLOCK lock handling. 511 * 512 * The caller can chain-lock backing objects recursively and then 513 * use vm_object_chain_release_all() to undo the whole chain. 514 * 515 * Chain locks are used to prevent collapses and are only applicable 516 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations 517 * on other object types are ignored. This is also important because 518 * it allows e.g. the vnode underlying a memory mapping to take concurrent 519 * faults. 520 * 521 * The object must usually be held on entry, though intermediate 522 * objects need not be held on release. The object must be held exclusively, 523 * NOT shared. Note that the prefault path checks the shared state and 524 * avoids using the chain functions. 525 */ 526 void 527 vm_object_chain_wait(vm_object_t object, int shared) 528 { 529 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 530 for (;;) { 531 uint32_t chainlk = object->chainlk; 532 533 cpu_ccfence(); 534 if (shared) { 535 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) { 536 tsleep_interlock(object, 0); 537 if (atomic_cmpset_int(&object->chainlk, 538 chainlk, 539 chainlk | CHAINLK_WAIT)) { 540 tsleep(object, PINTERLOCKED, 541 "objchns", 0); 542 } 543 /* retry */ 544 } else { 545 break; 546 } 547 /* retry */ 548 } else { 549 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) { 550 tsleep_interlock(object, 0); 551 if (atomic_cmpset_int(&object->chainlk, 552 chainlk, 553 chainlk | CHAINLK_WAIT)) 554 { 555 tsleep(object, PINTERLOCKED, 556 "objchnx", 0); 557 } 558 /* retry */ 559 } else { 560 if (atomic_cmpset_int(&object->chainlk, 561 chainlk, 562 chainlk & ~CHAINLK_WAIT)) 563 { 564 if (chainlk & CHAINLK_WAIT) 565 wakeup(object); 566 break; 567 } 568 /* retry */ 569 } 570 } 571 /* retry */ 572 } 573 } 574 575 void 576 vm_object_chain_acquire(vm_object_t object, int shared) 577 { 578 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) 579 return; 580 if (vm_shared_fault == 0) 581 shared = 0; 582 583 for (;;) { 584 uint32_t chainlk = object->chainlk; 585 586 cpu_ccfence(); 587 if (shared) { 588 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) { 589 tsleep_interlock(object, 0); 590 if (atomic_cmpset_int(&object->chainlk, 591 chainlk, 592 chainlk | CHAINLK_WAIT)) { 593 tsleep(object, PINTERLOCKED, 594 "objchns", 0); 595 } 596 /* retry */ 597 } else if (atomic_cmpset_int(&object->chainlk, 598 chainlk, chainlk + 1)) { 599 break; 600 } 601 /* retry */ 602 } else { 603 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) { 604 tsleep_interlock(object, 0); 605 if (atomic_cmpset_int(&object->chainlk, 606 chainlk, 607 chainlk | 608 CHAINLK_WAIT | 609 CHAINLK_EXCLREQ)) { 610 tsleep(object, PINTERLOCKED, 611 "objchnx", 0); 612 } 613 /* retry */ 614 } else { 615 if (atomic_cmpset_int(&object->chainlk, 616 chainlk, 617 (chainlk | CHAINLK_EXCL) & 618 ~(CHAINLK_EXCLREQ | 619 CHAINLK_WAIT))) { 620 if (chainlk & CHAINLK_WAIT) 621 wakeup(object); 622 break; 623 } 624 /* retry */ 625 } 626 } 627 /* retry */ 628 } 629 } 630 631 void 632 vm_object_chain_release(vm_object_t object) 633 { 634 /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/ 635 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) 636 return; 637 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL)); 638 for (;;) { 639 uint32_t chainlk = object->chainlk; 640 641 cpu_ccfence(); 642 if (chainlk & CHAINLK_MASK) { 643 if ((chainlk & CHAINLK_MASK) == 1 && 644 atomic_cmpset_int(&object->chainlk, 645 chainlk, 646 (chainlk - 1) & ~CHAINLK_WAIT)) { 647 if (chainlk & CHAINLK_WAIT) 648 wakeup(object); 649 break; 650 } 651 if ((chainlk & CHAINLK_MASK) > 1 && 652 atomic_cmpset_int(&object->chainlk, 653 chainlk, chainlk - 1)) { 654 break; 655 } 656 /* retry */ 657 } else { 658 KKASSERT(chainlk & CHAINLK_EXCL); 659 if (atomic_cmpset_int(&object->chainlk, 660 chainlk, 661 chainlk & ~(CHAINLK_EXCL | 662 CHAINLK_WAIT))) { 663 if (chainlk & CHAINLK_WAIT) 664 wakeup(object); 665 break; 666 } 667 } 668 } 669 } 670 671 /* 672 * Release the chain from first_object through and including stopobj. 673 * The caller is typically holding the first and last object locked 674 * (shared or exclusive) to prevent destruction races. 675 * 676 * We release stopobj first as an optimization as this object is most 677 * likely to be shared across multiple processes. 678 */ 679 void 680 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj) 681 { 682 vm_object_t backing_object; 683 vm_object_t object; 684 685 vm_object_chain_release(stopobj); 686 object = first_object; 687 688 while (object != stopobj) { 689 KKASSERT(object); 690 backing_object = object->backing_object; 691 vm_object_chain_release(object); 692 object = backing_object; 693 } 694 } 695 696 /* 697 * Dereference an object and its underlying vnode. The object may be 698 * held shared. On return the object will remain held. 699 * 700 * This function may return a vnode in *vpp which the caller must release 701 * after the caller drops its own lock. If vpp is NULL, we assume that 702 * the caller was holding an exclusive lock on the object and we vrele() 703 * the vp ourselves. 704 */ 705 static void 706 VMOBJDEBUG(vm_object_vndeallocate)(vm_object_t object, struct vnode **vpp 707 VMOBJDBARGS) 708 { 709 struct vnode *vp = (struct vnode *) object->handle; 710 711 KASSERT(object->type == OBJT_VNODE, 712 ("vm_object_vndeallocate: not a vnode object")); 713 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 714 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 715 #ifdef INVARIANTS 716 if (object->ref_count == 0) { 717 vprint("vm_object_vndeallocate", vp); 718 panic("vm_object_vndeallocate: bad object reference count"); 719 } 720 #endif 721 for (;;) { 722 int count = object->ref_count; 723 cpu_ccfence(); 724 if (count == 1) { 725 vm_object_upgrade(object); 726 if (atomic_cmpset_int(&object->ref_count, count, 0)) { 727 vclrflags(vp, VTEXT); 728 break; 729 } 730 } else { 731 if (atomic_cmpset_int(&object->ref_count, 732 count, count - 1)) { 733 break; 734 } 735 } 736 /* retry */ 737 } 738 #if defined(DEBUG_LOCKS) 739 debugvm_object_add(object, file, line, -1); 740 #endif 741 742 /* 743 * vrele or return the vp to vrele. We can only safely vrele(vp) 744 * if the object was locked exclusively. But there are two races 745 * here. 746 * 747 * We had to upgrade the object above to safely clear VTEXT 748 * but the alternative path where the shared lock is retained 749 * can STILL race to 0 in other paths and cause our own vrele() 750 * to terminate the vnode. We can't allow that if the VM object 751 * is still locked shared. 752 */ 753 if (vpp) 754 *vpp = vp; 755 else 756 vrele(vp); 757 } 758 759 /* 760 * Release a reference to the specified object, gained either through a 761 * vm_object_allocate or a vm_object_reference call. When all references 762 * are gone, storage associated with this object may be relinquished. 763 * 764 * The caller does not have to hold the object locked but must have control 765 * over the reference in question in order to guarantee that the object 766 * does not get ripped out from under us. 767 * 768 * XXX Currently all deallocations require an exclusive lock. 769 */ 770 void 771 VMOBJDEBUG(vm_object_deallocate)(vm_object_t object VMOBJDBARGS) 772 { 773 struct vnode *vp; 774 int count; 775 776 if (object == NULL) 777 return; 778 779 for (;;) { 780 count = object->ref_count; 781 cpu_ccfence(); 782 783 /* 784 * If decrementing the count enters into special handling 785 * territory (0, 1, or 2) we have to do it the hard way. 786 * Fortunate though, objects with only a few refs like this 787 * are not likely to be heavily contended anyway. 788 * 789 * For vnode objects we only care about 1->0 transitions. 790 */ 791 if (count <= 3 || (object->type == OBJT_VNODE && count <= 1)) { 792 #if defined(DEBUG_LOCKS) 793 debugvm_object_add(object, file, line, 0); 794 #endif 795 vm_object_hold(object); 796 vm_object_deallocate_locked(object); 797 vm_object_drop(object); 798 break; 799 } 800 801 /* 802 * Try to decrement ref_count without acquiring a hold on 803 * the object. This is particularly important for the exec*() 804 * and exit*() code paths because the program binary may 805 * have a great deal of sharing and an exclusive lock will 806 * crowbar performance in those circumstances. 807 */ 808 if (object->type == OBJT_VNODE) { 809 vp = (struct vnode *)object->handle; 810 if (atomic_cmpset_int(&object->ref_count, 811 count, count - 1)) { 812 #if defined(DEBUG_LOCKS) 813 debugvm_object_add(object, file, line, -1); 814 #endif 815 816 vrele(vp); 817 break; 818 } 819 /* retry */ 820 } else { 821 if (atomic_cmpset_int(&object->ref_count, 822 count, count - 1)) { 823 #if defined(DEBUG_LOCKS) 824 debugvm_object_add(object, file, line, -1); 825 #endif 826 break; 827 } 828 /* retry */ 829 } 830 /* retry */ 831 } 832 } 833 834 void 835 VMOBJDEBUG(vm_object_deallocate_locked)(vm_object_t object VMOBJDBARGS) 836 { 837 struct vm_object_dealloc_list *dlist = NULL; 838 struct vm_object_dealloc_list *dtmp; 839 vm_object_t temp; 840 int must_drop = 0; 841 842 /* 843 * We may chain deallocate object, but additional objects may 844 * collect on the dlist which also have to be deallocated. We 845 * must avoid a recursion, vm_object chains can get deep. 846 */ 847 848 again: 849 while (object != NULL) { 850 /* 851 * vnode case, caller either locked the object exclusively 852 * or this is a recursion with must_drop != 0 and the vnode 853 * object will be locked shared. 854 * 855 * If locked shared we have to drop the object before we can 856 * call vrele() or risk a shared/exclusive livelock. 857 */ 858 if (object->type == OBJT_VNODE) { 859 ASSERT_LWKT_TOKEN_HELD(&object->token); 860 if (must_drop) { 861 struct vnode *tmp_vp; 862 863 vm_object_vndeallocate(object, &tmp_vp); 864 vm_object_drop(object); 865 must_drop = 0; 866 object = NULL; 867 vrele(tmp_vp); 868 } else { 869 vm_object_vndeallocate(object, NULL); 870 } 871 break; 872 } 873 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token); 874 875 /* 876 * Normal case (object is locked exclusively) 877 */ 878 if (object->ref_count == 0) { 879 panic("vm_object_deallocate: object deallocated " 880 "too many times: %d", object->type); 881 } 882 if (object->ref_count > 2) { 883 atomic_add_int(&object->ref_count, -1); 884 #if defined(DEBUG_LOCKS) 885 debugvm_object_add(object, file, line, -1); 886 #endif 887 break; 888 } 889 890 /* 891 * Here on ref_count of one or two, which are special cases for 892 * objects. 893 * 894 * Nominal ref_count > 1 case if the second ref is not from 895 * a shadow. 896 * 897 * (ONEMAPPING only applies to DEFAULT AND SWAP objects) 898 */ 899 if (object->ref_count == 2 && object->shadow_count == 0) { 900 if (object->type == OBJT_DEFAULT || 901 object->type == OBJT_SWAP) { 902 vm_object_set_flag(object, OBJ_ONEMAPPING); 903 } 904 atomic_add_int(&object->ref_count, -1); 905 #if defined(DEBUG_LOCKS) 906 debugvm_object_add(object, file, line, -1); 907 #endif 908 break; 909 } 910 911 /* 912 * If the second ref is from a shadow we chain along it 913 * upwards if object's handle is exhausted. 914 * 915 * We have to decrement object->ref_count before potentially 916 * collapsing the first shadow object or the collapse code 917 * will not be able to handle the degenerate case to remove 918 * object. However, if we do it too early the object can 919 * get ripped out from under us. 920 */ 921 if (object->ref_count == 2 && object->shadow_count == 1 && 922 object->handle == NULL && (object->type == OBJT_DEFAULT || 923 object->type == OBJT_SWAP)) { 924 temp = LIST_FIRST(&object->shadow_head); 925 KKASSERT(temp != NULL); 926 vm_object_hold(temp); 927 928 /* 929 * Wait for any paging to complete so the collapse 930 * doesn't (or isn't likely to) qcollapse. pip 931 * waiting must occur before we acquire the 932 * chainlock. 933 */ 934 while ( 935 temp->paging_in_progress || 936 object->paging_in_progress 937 ) { 938 vm_object_pip_wait(temp, "objde1"); 939 vm_object_pip_wait(object, "objde2"); 940 } 941 942 /* 943 * If the parent is locked we have to give up, as 944 * otherwise we would be acquiring locks in the 945 * wrong order and potentially deadlock. 946 */ 947 if (temp->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) { 948 vm_object_drop(temp); 949 goto skip; 950 } 951 vm_object_chain_acquire(temp, 0); 952 953 /* 954 * Recheck/retry after the hold and the paging 955 * wait, both of which can block us. 956 */ 957 if (object->ref_count != 2 || 958 object->shadow_count != 1 || 959 object->handle || 960 LIST_FIRST(&object->shadow_head) != temp || 961 (object->type != OBJT_DEFAULT && 962 object->type != OBJT_SWAP)) { 963 vm_object_chain_release(temp); 964 vm_object_drop(temp); 965 continue; 966 } 967 968 /* 969 * We can safely drop object's ref_count now. 970 */ 971 KKASSERT(object->ref_count == 2); 972 atomic_add_int(&object->ref_count, -1); 973 #if defined(DEBUG_LOCKS) 974 debugvm_object_add(object, file, line, -1); 975 #endif 976 977 /* 978 * If our single parent is not collapseable just 979 * decrement ref_count (2->1) and stop. 980 */ 981 if (temp->handle || (temp->type != OBJT_DEFAULT && 982 temp->type != OBJT_SWAP)) { 983 vm_object_chain_release(temp); 984 vm_object_drop(temp); 985 break; 986 } 987 988 /* 989 * At this point we have already dropped object's 990 * ref_count so it is possible for a race to 991 * deallocate obj out from under us. Any collapse 992 * will re-check the situation. We must not block 993 * until we are able to collapse. 994 * 995 * Bump temp's ref_count to avoid an unwanted 996 * degenerate recursion (can't call 997 * vm_object_reference_locked() because it asserts 998 * that CHAINLOCK is not set). 999 */ 1000 atomic_add_int(&temp->ref_count, 1); 1001 KKASSERT(temp->ref_count > 1); 1002 1003 /* 1004 * Collapse temp, then deallocate the extra ref 1005 * formally. 1006 */ 1007 vm_object_collapse(temp, &dlist); 1008 vm_object_chain_release(temp); 1009 if (must_drop) { 1010 vm_object_lock_swap(); 1011 vm_object_drop(object); 1012 } 1013 object = temp; 1014 must_drop = 1; 1015 continue; 1016 } 1017 1018 /* 1019 * Drop the ref and handle termination on the 1->0 transition. 1020 * We may have blocked above so we have to recheck. 1021 */ 1022 skip: 1023 KKASSERT(object->ref_count != 0); 1024 if (object->ref_count >= 2) { 1025 atomic_add_int(&object->ref_count, -1); 1026 #if defined(DEBUG_LOCKS) 1027 debugvm_object_add(object, file, line, -1); 1028 #endif 1029 break; 1030 } 1031 KKASSERT(object->ref_count == 1); 1032 1033 /* 1034 * 1->0 transition. Chain through the backing_object. 1035 * Maintain the ref until we've located the backing object, 1036 * then re-check. 1037 */ 1038 while ((temp = object->backing_object) != NULL) { 1039 if (temp->type == OBJT_VNODE) 1040 vm_object_hold_shared(temp); 1041 else 1042 vm_object_hold(temp); 1043 if (temp == object->backing_object) 1044 break; 1045 vm_object_drop(temp); 1046 } 1047 1048 /* 1049 * 1->0 transition verified, retry if ref_count is no longer 1050 * 1. Otherwise disconnect the backing_object (temp) and 1051 * clean up. 1052 */ 1053 if (object->ref_count != 1) { 1054 vm_object_drop(temp); 1055 continue; 1056 } 1057 1058 /* 1059 * It shouldn't be possible for the object to be chain locked 1060 * if we're removing the last ref on it. 1061 * 1062 * Removing object from temp's shadow list requires dropping 1063 * temp, which we will do on loop. 1064 * 1065 * NOTE! vnodes do not use the shadow list, but still have 1066 * the backing_object reference. 1067 */ 1068 KKASSERT((object->chainlk & (CHAINLK_EXCL|CHAINLK_MASK)) == 0); 1069 1070 if (temp) { 1071 if (object->flags & OBJ_ONSHADOW) { 1072 LIST_REMOVE(object, shadow_list); 1073 temp->shadow_count--; 1074 temp->generation++; 1075 vm_object_clear_flag(object, OBJ_ONSHADOW); 1076 } 1077 object->backing_object = NULL; 1078 } 1079 1080 atomic_add_int(&object->ref_count, -1); 1081 if ((object->flags & OBJ_DEAD) == 0) 1082 vm_object_terminate(object); 1083 if (must_drop && temp) 1084 vm_object_lock_swap(); 1085 if (must_drop) 1086 vm_object_drop(object); 1087 object = temp; 1088 must_drop = 1; 1089 } 1090 1091 if (must_drop && object) 1092 vm_object_drop(object); 1093 1094 /* 1095 * Additional tail recursion on dlist. Avoid a recursion. Objects 1096 * on the dlist have a hold count but are not locked. 1097 */ 1098 if ((dtmp = dlist) != NULL) { 1099 dlist = dtmp->next; 1100 object = dtmp->object; 1101 kfree(dtmp, M_TEMP); 1102 1103 vm_object_lock(object); /* already held, add lock */ 1104 must_drop = 1; /* and we're responsible for it */ 1105 goto again; 1106 } 1107 } 1108 1109 /* 1110 * Destroy the specified object, freeing up related resources. 1111 * 1112 * The object must have zero references. 1113 * 1114 * The object must held. The caller is responsible for dropping the object 1115 * after terminate returns. Terminate does NOT drop the object. 1116 */ 1117 static int vm_object_terminate_callback(vm_page_t p, void *data); 1118 1119 void 1120 vm_object_terminate(vm_object_t object) 1121 { 1122 struct rb_vm_page_scan_info info; 1123 struct vm_object_hash *hash; 1124 1125 /* 1126 * Make sure no one uses us. Once we set OBJ_DEAD we should be 1127 * able to safely block. 1128 */ 1129 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1130 KKASSERT((object->flags & OBJ_DEAD) == 0); 1131 vm_object_set_flag(object, OBJ_DEAD); 1132 1133 /* 1134 * Wait for the pageout daemon to be done with the object 1135 */ 1136 vm_object_pip_wait(object, "objtrm1"); 1137 1138 KASSERT(!object->paging_in_progress, 1139 ("vm_object_terminate: pageout in progress")); 1140 1141 /* 1142 * Clean and free the pages, as appropriate. All references to the 1143 * object are gone, so we don't need to lock it. 1144 */ 1145 if (object->type == OBJT_VNODE) { 1146 struct vnode *vp; 1147 1148 /* 1149 * Clean pages and flush buffers. 1150 * 1151 * NOTE! TMPFS buffer flushes do not typically flush the 1152 * actual page to swap as this would be highly 1153 * inefficient, and normal filesystems usually wrap 1154 * page flushes with buffer cache buffers. 1155 * 1156 * To deal with this we have to call vinvalbuf() both 1157 * before and after the vm_object_page_clean(). 1158 */ 1159 vp = (struct vnode *) object->handle; 1160 vinvalbuf(vp, V_SAVE, 0, 0); 1161 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 1162 vinvalbuf(vp, V_SAVE, 0, 0); 1163 } 1164 1165 /* 1166 * Wait for any I/O to complete, after which there had better not 1167 * be any references left on the object. 1168 */ 1169 vm_object_pip_wait(object, "objtrm2"); 1170 1171 if (object->ref_count != 0) { 1172 panic("vm_object_terminate: object with references, " 1173 "ref_count=%d", object->ref_count); 1174 } 1175 1176 /* 1177 * Cleanup any shared pmaps associated with this object. 1178 */ 1179 pmap_object_free(object); 1180 1181 /* 1182 * Now free any remaining pages. For internal objects, this also 1183 * removes them from paging queues. Don't free wired pages, just 1184 * remove them from the object. 1185 */ 1186 info.count = 0; 1187 info.object = object; 1188 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1189 vm_object_terminate_callback, &info); 1190 1191 /* 1192 * Let the pager know object is dead. 1193 */ 1194 vm_pager_deallocate(object); 1195 1196 /* 1197 * Wait for the object hold count to hit 1, clean out pages as 1198 * we go. vmobj_token interlocks any race conditions that might 1199 * pick the object up from the vm_object_list after we have cleared 1200 * rb_memq. 1201 */ 1202 for (;;) { 1203 if (RB_ROOT(&object->rb_memq) == NULL) 1204 break; 1205 kprintf("vm_object_terminate: Warning, object %p " 1206 "still has %ld pages\n", 1207 object, object->resident_page_count); 1208 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 1209 vm_object_terminate_callback, &info); 1210 } 1211 1212 /* 1213 * There had better not be any pages left 1214 */ 1215 KKASSERT(object->resident_page_count == 0); 1216 1217 /* 1218 * Remove the object from the global object list. 1219 */ 1220 hash = VMOBJ_HASH(object); 1221 lwkt_gettoken(&hash->token); 1222 TAILQ_REMOVE(&hash->list, object, object_list); 1223 lwkt_reltoken(&hash->token); 1224 1225 if (object->ref_count != 0) { 1226 panic("vm_object_terminate2: object with references, " 1227 "ref_count=%d", object->ref_count); 1228 } 1229 1230 /* 1231 * NOTE: The object hold_count is at least 1, so we cannot kfree() 1232 * the object here. See vm_object_drop(). 1233 */ 1234 } 1235 1236 /* 1237 * The caller must hold the object. 1238 */ 1239 static int 1240 vm_object_terminate_callback(vm_page_t p, void *data) 1241 { 1242 struct rb_vm_page_scan_info *info = data; 1243 vm_object_t object; 1244 1245 if ((++info->count & 63) == 0) 1246 lwkt_user_yield(); 1247 object = p->object; 1248 if (object != info->object) { 1249 kprintf("vm_object_terminate_callback: obj/pg race %p/%p\n", 1250 info->object, p); 1251 return(0); 1252 } 1253 vm_page_busy_wait(p, TRUE, "vmpgtrm"); 1254 if (object != p->object) { 1255 kprintf("vm_object_terminate: Warning: Encountered " 1256 "busied page %p on queue %d\n", p, p->queue); 1257 vm_page_wakeup(p); 1258 } else if (p->wire_count == 0) { 1259 /* 1260 * NOTE: p->dirty and PG_NEED_COMMIT are ignored. 1261 */ 1262 vm_page_free(p); 1263 mycpu->gd_cnt.v_pfree++; 1264 } else { 1265 if (p->queue != PQ_NONE) 1266 kprintf("vm_object_terminate: Warning: Encountered " 1267 "wired page %p on queue %d\n", p, p->queue); 1268 vm_page_remove(p); 1269 vm_page_wakeup(p); 1270 } 1271 return(0); 1272 } 1273 1274 /* 1275 * Clean all dirty pages in the specified range of object. Leaves page 1276 * on whatever queue it is currently on. If NOSYNC is set then do not 1277 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 1278 * leaving the object dirty. 1279 * 1280 * When stuffing pages asynchronously, allow clustering. XXX we need a 1281 * synchronous clustering mode implementation. 1282 * 1283 * Odd semantics: if start == end, we clean everything. 1284 * 1285 * The object must be locked? XXX 1286 */ 1287 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 1288 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 1289 1290 void 1291 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1292 int flags) 1293 { 1294 struct rb_vm_page_scan_info info; 1295 struct vnode *vp; 1296 int wholescan; 1297 int pagerflags; 1298 int generation; 1299 1300 vm_object_hold(object); 1301 if (object->type != OBJT_VNODE || 1302 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 1303 vm_object_drop(object); 1304 return; 1305 } 1306 1307 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 1308 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1309 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 1310 1311 vp = object->handle; 1312 1313 /* 1314 * Interlock other major object operations. This allows us to 1315 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 1316 */ 1317 vm_object_set_flag(object, OBJ_CLEANING); 1318 1319 /* 1320 * Handle 'entire object' case 1321 */ 1322 info.start_pindex = start; 1323 if (end == 0) { 1324 info.end_pindex = object->size - 1; 1325 } else { 1326 info.end_pindex = end - 1; 1327 } 1328 wholescan = (start == 0 && info.end_pindex == object->size - 1); 1329 info.limit = flags; 1330 info.pagerflags = pagerflags; 1331 info.object = object; 1332 1333 /* 1334 * If cleaning the entire object do a pass to mark the pages read-only. 1335 * If everything worked out ok, clear OBJ_WRITEABLE and 1336 * OBJ_MIGHTBEDIRTY. 1337 */ 1338 if (wholescan) { 1339 info.error = 0; 1340 info.count = 0; 1341 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1342 vm_object_page_clean_pass1, &info); 1343 if (info.error == 0) { 1344 vm_object_clear_flag(object, 1345 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1346 if (object->type == OBJT_VNODE && 1347 (vp = (struct vnode *)object->handle) != NULL) { 1348 /* 1349 * Use new-style interface to clear VISDIRTY 1350 * because the vnode is not necessarily removed 1351 * from the syncer list(s) as often as it was 1352 * under the old interface, which can leave 1353 * the vnode on the syncer list after reclaim. 1354 */ 1355 vclrobjdirty(vp); 1356 } 1357 } 1358 } 1359 1360 /* 1361 * Do a pass to clean all the dirty pages we find. 1362 */ 1363 do { 1364 info.error = 0; 1365 info.count = 0; 1366 generation = object->generation; 1367 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1368 vm_object_page_clean_pass2, &info); 1369 } while (info.error || generation != object->generation); 1370 1371 vm_object_clear_flag(object, OBJ_CLEANING); 1372 vm_object_drop(object); 1373 } 1374 1375 /* 1376 * The caller must hold the object. 1377 */ 1378 static 1379 int 1380 vm_object_page_clean_pass1(struct vm_page *p, void *data) 1381 { 1382 struct rb_vm_page_scan_info *info = data; 1383 1384 if ((++info->count & 63) == 0) 1385 lwkt_user_yield(); 1386 if (p->object != info->object || 1387 p->pindex < info->start_pindex || 1388 p->pindex > info->end_pindex) { 1389 kprintf("vm_object_page_clean_pass1: obj/pg race %p/%p\n", 1390 info->object, p); 1391 return(0); 1392 } 1393 vm_page_flag_set(p, PG_CLEANCHK); 1394 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1395 info->error = 1; 1396 } else if (vm_page_busy_try(p, FALSE) == 0) { 1397 if (p->object == info->object) 1398 vm_page_protect(p, VM_PROT_READ); 1399 vm_page_wakeup(p); 1400 } else { 1401 info->error = 1; 1402 } 1403 return(0); 1404 } 1405 1406 /* 1407 * The caller must hold the object 1408 */ 1409 static 1410 int 1411 vm_object_page_clean_pass2(struct vm_page *p, void *data) 1412 { 1413 struct rb_vm_page_scan_info *info = data; 1414 int generation; 1415 1416 if (p->object != info->object || 1417 p->pindex < info->start_pindex || 1418 p->pindex > info->end_pindex) { 1419 kprintf("vm_object_page_clean_pass2: obj/pg race %p/%p\n", 1420 info->object, p); 1421 return(0); 1422 } 1423 1424 /* 1425 * Do not mess with pages that were inserted after we started 1426 * the cleaning pass. 1427 */ 1428 if ((p->flags & PG_CLEANCHK) == 0) 1429 goto done; 1430 1431 generation = info->object->generation; 1432 vm_page_busy_wait(p, TRUE, "vpcwai"); 1433 1434 if (p->object != info->object || 1435 p->pindex < info->start_pindex || 1436 p->pindex > info->end_pindex || 1437 info->object->generation != generation) { 1438 info->error = 1; 1439 vm_page_wakeup(p); 1440 goto done; 1441 } 1442 1443 /* 1444 * Before wasting time traversing the pmaps, check for trivial 1445 * cases where the page cannot be dirty. 1446 */ 1447 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 1448 KKASSERT((p->dirty & p->valid) == 0 && 1449 (p->flags & PG_NEED_COMMIT) == 0); 1450 vm_page_wakeup(p); 1451 goto done; 1452 } 1453 1454 /* 1455 * Check whether the page is dirty or not. The page has been set 1456 * to be read-only so the check will not race a user dirtying the 1457 * page. 1458 */ 1459 vm_page_test_dirty(p); 1460 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) { 1461 vm_page_flag_clear(p, PG_CLEANCHK); 1462 vm_page_wakeup(p); 1463 goto done; 1464 } 1465 1466 /* 1467 * If we have been asked to skip nosync pages and this is a 1468 * nosync page, skip it. Note that the object flags were 1469 * not cleared in this case (because pass1 will have returned an 1470 * error), so we do not have to set them. 1471 */ 1472 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 1473 vm_page_flag_clear(p, PG_CLEANCHK); 1474 vm_page_wakeup(p); 1475 goto done; 1476 } 1477 1478 /* 1479 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 1480 * the pages that get successfully flushed. Set info->error if 1481 * we raced an object modification. 1482 */ 1483 vm_object_page_collect_flush(info->object, p, info->pagerflags); 1484 /* vm_wait_nominal(); this can deadlock the system in syncer/pageout */ 1485 done: 1486 if ((++info->count & 63) == 0) 1487 lwkt_user_yield(); 1488 1489 return(0); 1490 } 1491 1492 /* 1493 * Collect the specified page and nearby pages and flush them out. 1494 * The number of pages flushed is returned. The passed page is busied 1495 * by the caller and we are responsible for its disposition. 1496 * 1497 * The caller must hold the object. 1498 */ 1499 static void 1500 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 1501 { 1502 int error; 1503 int is; 1504 int ib; 1505 int i; 1506 int page_base; 1507 vm_pindex_t pi; 1508 vm_page_t ma[BLIST_MAX_ALLOC]; 1509 1510 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1511 1512 pi = p->pindex; 1513 page_base = pi % BLIST_MAX_ALLOC; 1514 ma[page_base] = p; 1515 ib = page_base - 1; 1516 is = page_base + 1; 1517 1518 while (ib >= 0) { 1519 vm_page_t tp; 1520 1521 tp = vm_page_lookup_busy_try(object, pi - page_base + ib, 1522 TRUE, &error); 1523 if (error) 1524 break; 1525 if (tp == NULL) 1526 break; 1527 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1528 (tp->flags & PG_CLEANCHK) == 0) { 1529 vm_page_wakeup(tp); 1530 break; 1531 } 1532 if ((tp->queue - tp->pc) == PQ_CACHE) { 1533 vm_page_flag_clear(tp, PG_CLEANCHK); 1534 vm_page_wakeup(tp); 1535 break; 1536 } 1537 vm_page_test_dirty(tp); 1538 if ((tp->dirty & tp->valid) == 0 && 1539 (tp->flags & PG_NEED_COMMIT) == 0) { 1540 vm_page_flag_clear(tp, PG_CLEANCHK); 1541 vm_page_wakeup(tp); 1542 break; 1543 } 1544 ma[ib] = tp; 1545 --ib; 1546 } 1547 ++ib; /* fixup */ 1548 1549 while (is < BLIST_MAX_ALLOC && 1550 pi - page_base + is < object->size) { 1551 vm_page_t tp; 1552 1553 tp = vm_page_lookup_busy_try(object, pi - page_base + is, 1554 TRUE, &error); 1555 if (error) 1556 break; 1557 if (tp == NULL) 1558 break; 1559 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1560 (tp->flags & PG_CLEANCHK) == 0) { 1561 vm_page_wakeup(tp); 1562 break; 1563 } 1564 if ((tp->queue - tp->pc) == PQ_CACHE) { 1565 vm_page_flag_clear(tp, PG_CLEANCHK); 1566 vm_page_wakeup(tp); 1567 break; 1568 } 1569 vm_page_test_dirty(tp); 1570 if ((tp->dirty & tp->valid) == 0 && 1571 (tp->flags & PG_NEED_COMMIT) == 0) { 1572 vm_page_flag_clear(tp, PG_CLEANCHK); 1573 vm_page_wakeup(tp); 1574 break; 1575 } 1576 ma[is] = tp; 1577 ++is; 1578 } 1579 1580 /* 1581 * All pages in the ma[] array are busied now 1582 */ 1583 for (i = ib; i < is; ++i) { 1584 vm_page_flag_clear(ma[i], PG_CLEANCHK); 1585 vm_page_hold(ma[i]); /* XXX need this any more? */ 1586 } 1587 vm_pageout_flush(&ma[ib], is - ib, pagerflags); 1588 for (i = ib; i < is; ++i) /* XXX need this any more? */ 1589 vm_page_unhold(ma[i]); 1590 } 1591 1592 /* 1593 * Same as vm_object_pmap_copy, except range checking really 1594 * works, and is meant for small sections of an object. 1595 * 1596 * This code protects resident pages by making them read-only 1597 * and is typically called on a fork or split when a page 1598 * is converted to copy-on-write. 1599 * 1600 * NOTE: If the page is already at VM_PROT_NONE, calling 1601 * vm_page_protect will have no effect. 1602 */ 1603 void 1604 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1605 { 1606 vm_pindex_t idx; 1607 vm_page_t p; 1608 1609 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 1610 return; 1611 1612 vm_object_hold(object); 1613 for (idx = start; idx < end; idx++) { 1614 p = vm_page_lookup(object, idx); 1615 if (p == NULL) 1616 continue; 1617 vm_page_protect(p, VM_PROT_READ); 1618 } 1619 vm_object_drop(object); 1620 } 1621 1622 /* 1623 * Removes all physical pages in the specified object range from all 1624 * physical maps. 1625 * 1626 * The object must *not* be locked. 1627 */ 1628 1629 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 1630 1631 void 1632 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1633 { 1634 struct rb_vm_page_scan_info info; 1635 1636 if (object == NULL) 1637 return; 1638 info.start_pindex = start; 1639 info.end_pindex = end - 1; 1640 info.count = 0; 1641 info.object = object; 1642 1643 vm_object_hold(object); 1644 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1645 vm_object_pmap_remove_callback, &info); 1646 if (start == 0 && end == object->size) 1647 vm_object_clear_flag(object, OBJ_WRITEABLE); 1648 vm_object_drop(object); 1649 } 1650 1651 /* 1652 * The caller must hold the object 1653 */ 1654 static int 1655 vm_object_pmap_remove_callback(vm_page_t p, void *data) 1656 { 1657 struct rb_vm_page_scan_info *info = data; 1658 1659 if ((++info->count & 63) == 0) 1660 lwkt_user_yield(); 1661 1662 if (info->object != p->object || 1663 p->pindex < info->start_pindex || 1664 p->pindex > info->end_pindex) { 1665 kprintf("vm_object_pmap_remove_callback: obj/pg race %p/%p\n", 1666 info->object, p); 1667 return(0); 1668 } 1669 1670 vm_page_protect(p, VM_PROT_NONE); 1671 1672 return(0); 1673 } 1674 1675 /* 1676 * Implements the madvise function at the object/page level. 1677 * 1678 * MADV_WILLNEED (any object) 1679 * 1680 * Activate the specified pages if they are resident. 1681 * 1682 * MADV_DONTNEED (any object) 1683 * 1684 * Deactivate the specified pages if they are resident. 1685 * 1686 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 1687 * 1688 * Deactivate and clean the specified pages if they are 1689 * resident. This permits the process to reuse the pages 1690 * without faulting or the kernel to reclaim the pages 1691 * without I/O. 1692 * 1693 * No requirements. 1694 */ 1695 void 1696 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1697 { 1698 vm_pindex_t end, tpindex; 1699 vm_object_t tobject; 1700 vm_object_t xobj; 1701 vm_page_t m; 1702 int error; 1703 1704 if (object == NULL) 1705 return; 1706 1707 end = pindex + count; 1708 1709 vm_object_hold(object); 1710 tobject = object; 1711 1712 /* 1713 * Locate and adjust resident pages 1714 */ 1715 for (; pindex < end; pindex += 1) { 1716 relookup: 1717 if (tobject != object) 1718 vm_object_drop(tobject); 1719 tobject = object; 1720 tpindex = pindex; 1721 shadowlookup: 1722 /* 1723 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1724 * and those pages must be OBJ_ONEMAPPING. 1725 */ 1726 if (advise == MADV_FREE) { 1727 if ((tobject->type != OBJT_DEFAULT && 1728 tobject->type != OBJT_SWAP) || 1729 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1730 continue; 1731 } 1732 } 1733 1734 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error); 1735 1736 if (error) { 1737 vm_page_sleep_busy(m, TRUE, "madvpo"); 1738 goto relookup; 1739 } 1740 if (m == NULL) { 1741 /* 1742 * There may be swap even if there is no backing page 1743 */ 1744 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1745 swap_pager_freespace(tobject, tpindex, 1); 1746 1747 /* 1748 * next object 1749 */ 1750 while ((xobj = tobject->backing_object) != NULL) { 1751 KKASSERT(xobj != object); 1752 vm_object_hold(xobj); 1753 if (xobj == tobject->backing_object) 1754 break; 1755 vm_object_drop(xobj); 1756 } 1757 if (xobj == NULL) 1758 continue; 1759 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1760 if (tobject != object) { 1761 vm_object_lock_swap(); 1762 vm_object_drop(tobject); 1763 } 1764 tobject = xobj; 1765 goto shadowlookup; 1766 } 1767 1768 /* 1769 * If the page is not in a normal active state, we skip it. 1770 * If the page is not managed there are no page queues to 1771 * mess with. Things can break if we mess with pages in 1772 * any of the below states. 1773 */ 1774 if (m->wire_count || 1775 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 1776 m->valid != VM_PAGE_BITS_ALL 1777 ) { 1778 vm_page_wakeup(m); 1779 continue; 1780 } 1781 1782 /* 1783 * Theoretically once a page is known not to be busy, an 1784 * interrupt cannot come along and rip it out from under us. 1785 */ 1786 1787 if (advise == MADV_WILLNEED) { 1788 vm_page_activate(m); 1789 } else if (advise == MADV_DONTNEED) { 1790 vm_page_dontneed(m); 1791 } else if (advise == MADV_FREE) { 1792 /* 1793 * Mark the page clean. This will allow the page 1794 * to be freed up by the system. However, such pages 1795 * are often reused quickly by malloc()/free() 1796 * so we do not do anything that would cause 1797 * a page fault if we can help it. 1798 * 1799 * Specifically, we do not try to actually free 1800 * the page now nor do we try to put it in the 1801 * cache (which would cause a page fault on reuse). 1802 * 1803 * But we do make the page is freeable as we 1804 * can without actually taking the step of unmapping 1805 * it. 1806 */ 1807 pmap_clear_modify(m); 1808 m->dirty = 0; 1809 m->act_count = 0; 1810 vm_page_dontneed(m); 1811 if (tobject->type == OBJT_SWAP) 1812 swap_pager_freespace(tobject, tpindex, 1); 1813 } 1814 vm_page_wakeup(m); 1815 } 1816 if (tobject != object) 1817 vm_object_drop(tobject); 1818 vm_object_drop(object); 1819 } 1820 1821 /* 1822 * Create a new object which is backed by the specified existing object 1823 * range. Replace the pointer and offset that was pointing at the existing 1824 * object with the pointer/offset for the new object. 1825 * 1826 * If addref is non-zero the returned object is given an additional reference. 1827 * This mechanic exists to avoid the situation where refs might be 1 and 1828 * race against a collapse when the caller intends to bump it. So the 1829 * caller cannot add the ref after the fact. Used when the caller is 1830 * duplicating a vm_map_entry. 1831 * 1832 * No other requirements. 1833 */ 1834 void 1835 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length, 1836 int addref) 1837 { 1838 vm_object_t source; 1839 vm_object_t result; 1840 int useshadowlist; 1841 1842 source = *objectp; 1843 1844 /* 1845 * Don't create the new object if the old object isn't shared. 1846 * We have to chain wait before adding the reference to avoid 1847 * racing a collapse or deallocation. 1848 * 1849 * Clear OBJ_ONEMAPPING flag when shadowing. 1850 * 1851 * The caller owns a ref on source via *objectp which we are going 1852 * to replace. This ref is inherited by the backing_object assignment. 1853 * from nobject and does not need to be incremented here. 1854 * 1855 * However, we add a temporary extra reference to the original source 1856 * prior to holding nobject in case we block, to avoid races where 1857 * someone else might believe that the source can be collapsed. 1858 */ 1859 useshadowlist = 0; 1860 if (source) { 1861 if (source->type != OBJT_VNODE) { 1862 useshadowlist = 1; 1863 vm_object_hold(source); 1864 vm_object_chain_wait(source, 0); 1865 if (source->ref_count == 1 && 1866 source->handle == NULL && 1867 (source->type == OBJT_DEFAULT || 1868 source->type == OBJT_SWAP)) { 1869 if (addref) { 1870 vm_object_reference_locked(source); 1871 vm_object_clear_flag(source, 1872 OBJ_ONEMAPPING); 1873 } 1874 vm_object_drop(source); 1875 return; 1876 } 1877 vm_object_reference_locked(source); 1878 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1879 } else { 1880 vm_object_reference_quick(source); 1881 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1882 } 1883 } 1884 1885 /* 1886 * Allocate a new object with the given length. The new object 1887 * is returned referenced but we may have to add another one. 1888 * If we are adding a second reference we must clear OBJ_ONEMAPPING. 1889 * (typically because the caller is about to clone a vm_map_entry). 1890 * 1891 * The source object currently has an extra reference to prevent 1892 * collapses into it while we mess with its shadow list, which 1893 * we will remove later in this routine. 1894 * 1895 * The target object may require a second reference if asked for one 1896 * by the caller. 1897 */ 1898 result = vm_object_allocate(OBJT_DEFAULT, length); 1899 if (result == NULL) 1900 panic("vm_object_shadow: no object for shadowing"); 1901 vm_object_hold(result); 1902 if (addref) { 1903 vm_object_reference_locked(result); 1904 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1905 } 1906 1907 /* 1908 * The new object shadows the source object. Chain wait before 1909 * adjusting shadow_count or the shadow list to avoid races. 1910 * 1911 * Try to optimize the result object's page color when shadowing 1912 * in order to maintain page coloring consistency in the combined 1913 * shadowed object. 1914 * 1915 * The backing_object reference to source requires adding a ref to 1916 * source. We simply inherit the ref from the original *objectp 1917 * (which we are replacing) so no additional refs need to be added. 1918 * (we must still clean up the extra ref we had to prevent collapse 1919 * races). 1920 * 1921 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS 1922 */ 1923 KKASSERT(result->backing_object == NULL); 1924 result->backing_object = source; 1925 if (source) { 1926 if (useshadowlist) { 1927 vm_object_chain_wait(source, 0); 1928 LIST_INSERT_HEAD(&source->shadow_head, 1929 result, shadow_list); 1930 source->shadow_count++; 1931 source->generation++; 1932 vm_object_set_flag(result, OBJ_ONSHADOW); 1933 } 1934 /* cpu localization twist */ 1935 result->pg_color = vm_quickcolor(); 1936 } 1937 1938 /* 1939 * Adjust the return storage. Drop the ref on source before 1940 * returning. 1941 */ 1942 result->backing_object_offset = *offset; 1943 vm_object_drop(result); 1944 *offset = 0; 1945 if (source) { 1946 if (useshadowlist) { 1947 vm_object_deallocate_locked(source); 1948 vm_object_drop(source); 1949 } else { 1950 vm_object_deallocate(source); 1951 } 1952 } 1953 1954 /* 1955 * Return the new things 1956 */ 1957 *objectp = result; 1958 } 1959 1960 #define OBSC_TEST_ALL_SHADOWED 0x0001 1961 #define OBSC_COLLAPSE_NOWAIT 0x0002 1962 #define OBSC_COLLAPSE_WAIT 0x0004 1963 1964 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1965 1966 /* 1967 * The caller must hold the object. 1968 */ 1969 static __inline int 1970 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op) 1971 { 1972 struct rb_vm_page_scan_info info; 1973 struct vm_object_hash *hash; 1974 1975 vm_object_assert_held(object); 1976 vm_object_assert_held(backing_object); 1977 1978 KKASSERT(backing_object == object->backing_object); 1979 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1980 1981 /* 1982 * Initial conditions 1983 */ 1984 if (op & OBSC_TEST_ALL_SHADOWED) { 1985 /* 1986 * We do not want to have to test for the existence of 1987 * swap pages in the backing object. XXX but with the 1988 * new swapper this would be pretty easy to do. 1989 * 1990 * XXX what about anonymous MAP_SHARED memory that hasn't 1991 * been ZFOD faulted yet? If we do not test for this, the 1992 * shadow test may succeed! XXX 1993 */ 1994 if (backing_object->type != OBJT_DEFAULT) 1995 return(0); 1996 } 1997 if (op & OBSC_COLLAPSE_WAIT) { 1998 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1999 vm_object_set_flag(backing_object, OBJ_DEAD); 2000 2001 hash = VMOBJ_HASH(backing_object); 2002 lwkt_gettoken(&hash->token); 2003 TAILQ_REMOVE(&hash->list, backing_object, object_list); 2004 lwkt_reltoken(&hash->token); 2005 } 2006 2007 /* 2008 * Our scan. We have to retry if a negative error code is returned, 2009 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 2010 * the scan had to be stopped because the parent does not completely 2011 * shadow the child. 2012 */ 2013 info.object = object; 2014 info.backing_object = backing_object; 2015 info.limit = op; 2016 info.count = 0; 2017 do { 2018 info.error = 1; 2019 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 2020 vm_object_backing_scan_callback, 2021 &info); 2022 } while (info.error < 0); 2023 2024 return(info.error); 2025 } 2026 2027 /* 2028 * The caller must hold the object. 2029 */ 2030 static int 2031 vm_object_backing_scan_callback(vm_page_t p, void *data) 2032 { 2033 struct rb_vm_page_scan_info *info = data; 2034 vm_object_t backing_object; 2035 vm_object_t object; 2036 vm_pindex_t pindex; 2037 vm_pindex_t new_pindex; 2038 vm_pindex_t backing_offset_index; 2039 int op; 2040 2041 pindex = p->pindex; 2042 new_pindex = pindex - info->backing_offset_index; 2043 op = info->limit; 2044 object = info->object; 2045 backing_object = info->backing_object; 2046 backing_offset_index = info->backing_offset_index; 2047 2048 if (op & OBSC_TEST_ALL_SHADOWED) { 2049 vm_page_t pp; 2050 2051 /* 2052 * Ignore pages outside the parent object's range 2053 * and outside the parent object's mapping of the 2054 * backing object. 2055 * 2056 * note that we do not busy the backing object's 2057 * page. 2058 */ 2059 if (pindex < backing_offset_index || 2060 new_pindex >= object->size 2061 ) { 2062 return(0); 2063 } 2064 2065 /* 2066 * See if the parent has the page or if the parent's 2067 * object pager has the page. If the parent has the 2068 * page but the page is not valid, the parent's 2069 * object pager must have the page. 2070 * 2071 * If this fails, the parent does not completely shadow 2072 * the object and we might as well give up now. 2073 */ 2074 pp = vm_page_lookup(object, new_pindex); 2075 if ((pp == NULL || pp->valid == 0) && 2076 !vm_pager_has_page(object, new_pindex) 2077 ) { 2078 info->error = 0; /* problemo */ 2079 return(-1); /* stop the scan */ 2080 } 2081 } 2082 2083 /* 2084 * Check for busy page. Note that we may have lost (p) when we 2085 * possibly blocked above. 2086 */ 2087 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 2088 vm_page_t pp; 2089 2090 if (vm_page_busy_try(p, TRUE)) { 2091 if (op & OBSC_COLLAPSE_NOWAIT) { 2092 return(0); 2093 } else { 2094 /* 2095 * If we slept, anything could have 2096 * happened. Ask that the scan be restarted. 2097 * 2098 * Since the object is marked dead, the 2099 * backing offset should not have changed. 2100 */ 2101 vm_page_sleep_busy(p, TRUE, "vmocol"); 2102 info->error = -1; 2103 return(-1); 2104 } 2105 } 2106 2107 /* 2108 * If (p) is no longer valid restart the scan. 2109 */ 2110 if (p->object != backing_object || p->pindex != pindex) { 2111 kprintf("vm_object_backing_scan: Warning: page " 2112 "%p ripped out from under us\n", p); 2113 vm_page_wakeup(p); 2114 info->error = -1; 2115 return(-1); 2116 } 2117 2118 if (op & OBSC_COLLAPSE_NOWAIT) { 2119 if (p->valid == 0 || 2120 p->wire_count || 2121 (p->flags & PG_NEED_COMMIT)) { 2122 vm_page_wakeup(p); 2123 return(0); 2124 } 2125 } else { 2126 /* XXX what if p->valid == 0 , hold_count, etc? */ 2127 } 2128 2129 KASSERT( 2130 p->object == backing_object, 2131 ("vm_object_qcollapse(): object mismatch") 2132 ); 2133 2134 /* 2135 * Destroy any associated swap 2136 */ 2137 if (backing_object->type == OBJT_SWAP) 2138 swap_pager_freespace(backing_object, p->pindex, 1); 2139 2140 if ( 2141 p->pindex < backing_offset_index || 2142 new_pindex >= object->size 2143 ) { 2144 /* 2145 * Page is out of the parent object's range, we 2146 * can simply destroy it. 2147 */ 2148 vm_page_protect(p, VM_PROT_NONE); 2149 vm_page_free(p); 2150 return(0); 2151 } 2152 2153 pp = vm_page_lookup(object, new_pindex); 2154 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 2155 /* 2156 * page already exists in parent OR swap exists 2157 * for this location in the parent. Destroy 2158 * the original page from the backing object. 2159 * 2160 * Leave the parent's page alone 2161 */ 2162 vm_page_protect(p, VM_PROT_NONE); 2163 vm_page_free(p); 2164 return(0); 2165 } 2166 2167 /* 2168 * Page does not exist in parent, rename the 2169 * page from the backing object to the main object. 2170 * 2171 * If the page was mapped to a process, it can remain 2172 * mapped through the rename. 2173 */ 2174 if ((p->queue - p->pc) == PQ_CACHE) 2175 vm_page_deactivate(p); 2176 2177 vm_page_rename(p, object, new_pindex); 2178 vm_page_wakeup(p); 2179 /* page automatically made dirty by rename */ 2180 } 2181 return(0); 2182 } 2183 2184 /* 2185 * This version of collapse allows the operation to occur earlier and 2186 * when paging_in_progress is true for an object... This is not a complete 2187 * operation, but should plug 99.9% of the rest of the leaks. 2188 * 2189 * The caller must hold the object and backing_object and both must be 2190 * chainlocked. 2191 * 2192 * (only called from vm_object_collapse) 2193 */ 2194 static void 2195 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object) 2196 { 2197 if (backing_object->ref_count == 1) { 2198 atomic_add_int(&backing_object->ref_count, 2); 2199 #if defined(DEBUG_LOCKS) 2200 debugvm_object_add(backing_object, "qcollapse", 1, 2); 2201 #endif 2202 vm_object_backing_scan(object, backing_object, 2203 OBSC_COLLAPSE_NOWAIT); 2204 atomic_add_int(&backing_object->ref_count, -2); 2205 #if defined(DEBUG_LOCKS) 2206 debugvm_object_add(backing_object, "qcollapse", 2, -2); 2207 #endif 2208 } 2209 } 2210 2211 /* 2212 * Collapse an object with the object backing it. Pages in the backing 2213 * object are moved into the parent, and the backing object is deallocated. 2214 * Any conflict is resolved in favor of the parent's existing pages. 2215 * 2216 * object must be held and chain-locked on call. 2217 * 2218 * The caller must have an extra ref on object to prevent a race from 2219 * destroying it during the collapse. 2220 */ 2221 void 2222 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp) 2223 { 2224 struct vm_object_dealloc_list *dlist = NULL; 2225 vm_object_t backing_object; 2226 2227 /* 2228 * Only one thread is attempting a collapse at any given moment. 2229 * There are few restrictions for (object) that callers of this 2230 * function check so reentrancy is likely. 2231 */ 2232 KKASSERT(object != NULL); 2233 vm_object_assert_held(object); 2234 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL)); 2235 2236 for (;;) { 2237 vm_object_t bbobj; 2238 int dodealloc; 2239 2240 /* 2241 * We can only collapse a DEFAULT/SWAP object with a 2242 * DEFAULT/SWAP object. 2243 */ 2244 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) { 2245 backing_object = NULL; 2246 break; 2247 } 2248 2249 backing_object = object->backing_object; 2250 if (backing_object == NULL) 2251 break; 2252 if (backing_object->type != OBJT_DEFAULT && 2253 backing_object->type != OBJT_SWAP) { 2254 backing_object = NULL; 2255 break; 2256 } 2257 2258 /* 2259 * Hold the backing_object and check for races 2260 */ 2261 vm_object_hold(backing_object); 2262 if (backing_object != object->backing_object || 2263 (backing_object->type != OBJT_DEFAULT && 2264 backing_object->type != OBJT_SWAP)) { 2265 vm_object_drop(backing_object); 2266 continue; 2267 } 2268 2269 /* 2270 * Chain-lock the backing object too because if we 2271 * successfully merge its pages into the top object we 2272 * will collapse backing_object->backing_object as the 2273 * new backing_object. Re-check that it is still our 2274 * backing object. 2275 */ 2276 vm_object_chain_acquire(backing_object, 0); 2277 if (backing_object != object->backing_object) { 2278 vm_object_chain_release(backing_object); 2279 vm_object_drop(backing_object); 2280 continue; 2281 } 2282 2283 /* 2284 * we check the backing object first, because it is most likely 2285 * not collapsable. 2286 */ 2287 if (backing_object->handle != NULL || 2288 (backing_object->type != OBJT_DEFAULT && 2289 backing_object->type != OBJT_SWAP) || 2290 (backing_object->flags & OBJ_DEAD) || 2291 object->handle != NULL || 2292 (object->type != OBJT_DEFAULT && 2293 object->type != OBJT_SWAP) || 2294 (object->flags & OBJ_DEAD)) { 2295 break; 2296 } 2297 2298 /* 2299 * If paging is in progress we can't do a normal collapse. 2300 */ 2301 if ( 2302 object->paging_in_progress != 0 || 2303 backing_object->paging_in_progress != 0 2304 ) { 2305 vm_object_qcollapse(object, backing_object); 2306 break; 2307 } 2308 2309 /* 2310 * We know that we can either collapse the backing object (if 2311 * the parent is the only reference to it) or (perhaps) have 2312 * the parent bypass the object if the parent happens to shadow 2313 * all the resident pages in the entire backing object. 2314 * 2315 * This is ignoring pager-backed pages such as swap pages. 2316 * vm_object_backing_scan fails the shadowing test in this 2317 * case. 2318 */ 2319 if (backing_object->ref_count == 1) { 2320 /* 2321 * If there is exactly one reference to the backing 2322 * object, we can collapse it into the parent. 2323 */ 2324 KKASSERT(object->backing_object == backing_object); 2325 vm_object_backing_scan(object, backing_object, 2326 OBSC_COLLAPSE_WAIT); 2327 2328 /* 2329 * Move the pager from backing_object to object. 2330 */ 2331 if (backing_object->type == OBJT_SWAP) { 2332 vm_object_pip_add(backing_object, 1); 2333 2334 /* 2335 * scrap the paging_offset junk and do a 2336 * discrete copy. This also removes major 2337 * assumptions about how the swap-pager 2338 * works from where it doesn't belong. The 2339 * new swapper is able to optimize the 2340 * destroy-source case. 2341 */ 2342 vm_object_pip_add(object, 1); 2343 swap_pager_copy(backing_object, object, 2344 OFF_TO_IDX(object->backing_object_offset), 2345 TRUE); 2346 vm_object_pip_wakeup(object); 2347 vm_object_pip_wakeup(backing_object); 2348 } 2349 2350 /* 2351 * Object now shadows whatever backing_object did. 2352 * Remove object from backing_object's shadow_list. 2353 * 2354 * Removing object from backing_objects shadow list 2355 * requires releasing object, which we will do below. 2356 */ 2357 KKASSERT(object->backing_object == backing_object); 2358 if (object->flags & OBJ_ONSHADOW) { 2359 LIST_REMOVE(object, shadow_list); 2360 backing_object->shadow_count--; 2361 backing_object->generation++; 2362 vm_object_clear_flag(object, OBJ_ONSHADOW); 2363 } 2364 2365 /* 2366 * backing_object->backing_object moves from within 2367 * backing_object to within object. 2368 * 2369 * OBJT_VNODE bbobj's should have empty shadow lists. 2370 */ 2371 while ((bbobj = backing_object->backing_object) != NULL) { 2372 if (bbobj->type == OBJT_VNODE) 2373 vm_object_hold_shared(bbobj); 2374 else 2375 vm_object_hold(bbobj); 2376 if (bbobj == backing_object->backing_object) 2377 break; 2378 vm_object_drop(bbobj); 2379 } 2380 2381 /* 2382 * We are removing backing_object from bbobj's 2383 * shadow list and adding object to bbobj's shadow 2384 * list, so the ref_count on bbobj is unchanged. 2385 */ 2386 if (bbobj) { 2387 if (backing_object->flags & OBJ_ONSHADOW) { 2388 /* not locked exclusively if vnode */ 2389 KKASSERT(bbobj->type != OBJT_VNODE); 2390 LIST_REMOVE(backing_object, 2391 shadow_list); 2392 bbobj->shadow_count--; 2393 bbobj->generation++; 2394 vm_object_clear_flag(backing_object, 2395 OBJ_ONSHADOW); 2396 } 2397 backing_object->backing_object = NULL; 2398 } 2399 object->backing_object = bbobj; 2400 if (bbobj) { 2401 if (bbobj->type != OBJT_VNODE) { 2402 LIST_INSERT_HEAD(&bbobj->shadow_head, 2403 object, shadow_list); 2404 bbobj->shadow_count++; 2405 bbobj->generation++; 2406 vm_object_set_flag(object, 2407 OBJ_ONSHADOW); 2408 } 2409 } 2410 2411 object->backing_object_offset += 2412 backing_object->backing_object_offset; 2413 2414 vm_object_drop(bbobj); 2415 2416 /* 2417 * Discard the old backing_object. Nothing should be 2418 * able to ref it, other than a vm_map_split(), 2419 * and vm_map_split() will stall on our chain lock. 2420 * And we control the parent so it shouldn't be 2421 * possible for it to go away either. 2422 * 2423 * Since the backing object has no pages, no pager 2424 * left, and no object references within it, all 2425 * that is necessary is to dispose of it. 2426 */ 2427 KASSERT(backing_object->ref_count == 1, 2428 ("backing_object %p was somehow " 2429 "re-referenced during collapse!", 2430 backing_object)); 2431 KASSERT(RB_EMPTY(&backing_object->rb_memq), 2432 ("backing_object %p somehow has left " 2433 "over pages during collapse!", 2434 backing_object)); 2435 2436 /* 2437 * The object can be destroyed. 2438 * 2439 * XXX just fall through and dodealloc instead 2440 * of forcing destruction? 2441 */ 2442 atomic_add_int(&backing_object->ref_count, -1); 2443 #if defined(DEBUG_LOCKS) 2444 debugvm_object_add(backing_object, "collapse", 1, -1); 2445 #endif 2446 if ((backing_object->flags & OBJ_DEAD) == 0) 2447 vm_object_terminate(backing_object); 2448 object_collapses++; 2449 dodealloc = 0; 2450 } else { 2451 /* 2452 * If we do not entirely shadow the backing object, 2453 * there is nothing we can do so we give up. 2454 */ 2455 if (vm_object_backing_scan(object, backing_object, 2456 OBSC_TEST_ALL_SHADOWED) == 0) { 2457 break; 2458 } 2459 2460 /* 2461 * bbobj is backing_object->backing_object. Since 2462 * object completely shadows backing_object we can 2463 * bypass it and become backed by bbobj instead. 2464 * 2465 * The shadow list for vnode backing objects is not 2466 * used and a shared hold is allowed. 2467 */ 2468 while ((bbobj = backing_object->backing_object) != NULL) { 2469 if (bbobj->type == OBJT_VNODE) 2470 vm_object_hold_shared(bbobj); 2471 else 2472 vm_object_hold(bbobj); 2473 if (bbobj == backing_object->backing_object) 2474 break; 2475 vm_object_drop(bbobj); 2476 } 2477 2478 /* 2479 * Make object shadow bbobj instead of backing_object. 2480 * Remove object from backing_object's shadow list. 2481 * 2482 * Deallocating backing_object will not remove 2483 * it, since its reference count is at least 2. 2484 * 2485 * Removing object from backing_object's shadow 2486 * list requires releasing a ref, which we do 2487 * below by setting dodealloc to 1. 2488 */ 2489 KKASSERT(object->backing_object == backing_object); 2490 if (object->flags & OBJ_ONSHADOW) { 2491 LIST_REMOVE(object, shadow_list); 2492 backing_object->shadow_count--; 2493 backing_object->generation++; 2494 vm_object_clear_flag(object, OBJ_ONSHADOW); 2495 } 2496 2497 /* 2498 * Add a ref to bbobj, bbobj now shadows object. 2499 * 2500 * NOTE: backing_object->backing_object still points 2501 * to bbobj. That relationship remains intact 2502 * because backing_object has > 1 ref, so 2503 * someone else is pointing to it (hence why 2504 * we can't collapse it into object and can 2505 * only handle the all-shadowed bypass case). 2506 */ 2507 if (bbobj) { 2508 if (bbobj->type != OBJT_VNODE) { 2509 vm_object_chain_wait(bbobj, 0); 2510 vm_object_reference_locked(bbobj); 2511 LIST_INSERT_HEAD(&bbobj->shadow_head, 2512 object, shadow_list); 2513 bbobj->shadow_count++; 2514 bbobj->generation++; 2515 vm_object_set_flag(object, 2516 OBJ_ONSHADOW); 2517 } else { 2518 vm_object_reference_quick(bbobj); 2519 } 2520 object->backing_object_offset += 2521 backing_object->backing_object_offset; 2522 object->backing_object = bbobj; 2523 vm_object_drop(bbobj); 2524 } else { 2525 object->backing_object = NULL; 2526 } 2527 2528 /* 2529 * Drop the reference count on backing_object. To 2530 * handle ref_count races properly we can't assume 2531 * that the ref_count is still at least 2 so we 2532 * have to actually call vm_object_deallocate() 2533 * (after clearing the chainlock). 2534 */ 2535 object_bypasses++; 2536 dodealloc = 1; 2537 } 2538 2539 /* 2540 * Ok, we want to loop on the new object->bbobj association, 2541 * possibly collapsing it further. However if dodealloc is 2542 * non-zero we have to deallocate the backing_object which 2543 * itself can potentially undergo a collapse, creating a 2544 * recursion depth issue with the LWKT token subsystem. 2545 * 2546 * In the case where we must deallocate the backing_object 2547 * it is possible now that the backing_object has a single 2548 * shadow count on some other object (not represented here 2549 * as yet), since it no longer shadows us. Thus when we 2550 * call vm_object_deallocate() it may attempt to collapse 2551 * itself into its remaining parent. 2552 */ 2553 if (dodealloc) { 2554 struct vm_object_dealloc_list *dtmp; 2555 2556 vm_object_chain_release(backing_object); 2557 vm_object_unlock(backing_object); 2558 /* backing_object remains held */ 2559 2560 /* 2561 * Auto-deallocation list for caller convenience. 2562 */ 2563 if (dlistp == NULL) 2564 dlistp = &dlist; 2565 2566 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK); 2567 dtmp->object = backing_object; 2568 dtmp->next = *dlistp; 2569 *dlistp = dtmp; 2570 } else { 2571 vm_object_chain_release(backing_object); 2572 vm_object_drop(backing_object); 2573 } 2574 /* backing_object = NULL; not needed */ 2575 /* loop */ 2576 } 2577 2578 /* 2579 * Clean up any left over backing_object 2580 */ 2581 if (backing_object) { 2582 vm_object_chain_release(backing_object); 2583 vm_object_drop(backing_object); 2584 } 2585 2586 /* 2587 * Clean up any auto-deallocation list. This is a convenience 2588 * for top-level callers so they don't have to pass &dlist. 2589 * Do not clean up any caller-passed dlistp, the caller will 2590 * do that. 2591 */ 2592 if (dlist) 2593 vm_object_deallocate_list(&dlist); 2594 2595 } 2596 2597 /* 2598 * vm_object_collapse() may collect additional objects in need of 2599 * deallocation. This routine deallocates these objects. The 2600 * deallocation itself can trigger additional collapses (which the 2601 * deallocate function takes care of). This procedure is used to 2602 * reduce procedural recursion since these vm_object shadow chains 2603 * can become quite long. 2604 */ 2605 void 2606 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp) 2607 { 2608 struct vm_object_dealloc_list *dlist; 2609 2610 while ((dlist = *dlistp) != NULL) { 2611 *dlistp = dlist->next; 2612 vm_object_lock(dlist->object); 2613 vm_object_deallocate_locked(dlist->object); 2614 vm_object_drop(dlist->object); 2615 kfree(dlist, M_TEMP); 2616 } 2617 } 2618 2619 /* 2620 * Removes all physical pages in the specified object range from the 2621 * object's list of pages. 2622 * 2623 * No requirements. 2624 */ 2625 static int vm_object_page_remove_callback(vm_page_t p, void *data); 2626 2627 void 2628 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2629 boolean_t clean_only) 2630 { 2631 struct rb_vm_page_scan_info info; 2632 int all; 2633 2634 /* 2635 * Degenerate cases and assertions 2636 */ 2637 vm_object_hold(object); 2638 if (object == NULL || 2639 (object->resident_page_count == 0 && object->swblock_count == 0)) { 2640 vm_object_drop(object); 2641 return; 2642 } 2643 KASSERT(object->type != OBJT_PHYS, 2644 ("attempt to remove pages from a physical object")); 2645 2646 /* 2647 * Indicate that paging is occuring on the object 2648 */ 2649 vm_object_pip_add(object, 1); 2650 2651 /* 2652 * Figure out the actual removal range and whether we are removing 2653 * the entire contents of the object or not. If removing the entire 2654 * contents, be sure to get all pages, even those that might be 2655 * beyond the end of the object. 2656 */ 2657 info.object = object; 2658 info.start_pindex = start; 2659 if (end == 0) 2660 info.end_pindex = (vm_pindex_t)-1; 2661 else 2662 info.end_pindex = end - 1; 2663 info.limit = clean_only; 2664 info.count = 0; 2665 all = (start == 0 && info.end_pindex >= object->size - 1); 2666 2667 /* 2668 * Loop until we are sure we have gotten them all. 2669 */ 2670 do { 2671 info.error = 0; 2672 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 2673 vm_object_page_remove_callback, &info); 2674 } while (info.error); 2675 2676 /* 2677 * Remove any related swap if throwing away pages, or for 2678 * non-swap objects (the swap is a clean copy in that case). 2679 */ 2680 if (object->type != OBJT_SWAP || clean_only == FALSE) { 2681 if (all) 2682 swap_pager_freespace_all(object); 2683 else 2684 swap_pager_freespace(object, info.start_pindex, 2685 info.end_pindex - info.start_pindex + 1); 2686 } 2687 2688 /* 2689 * Cleanup 2690 */ 2691 vm_object_pip_wakeup(object); 2692 vm_object_drop(object); 2693 } 2694 2695 /* 2696 * The caller must hold the object. 2697 * 2698 * NOTE: User yields are allowed when removing more than one page, but not 2699 * allowed if only removing one page (the path for single page removals 2700 * might hold a spinlock). 2701 */ 2702 static int 2703 vm_object_page_remove_callback(vm_page_t p, void *data) 2704 { 2705 struct rb_vm_page_scan_info *info = data; 2706 2707 if ((++info->count & 63) == 0) 2708 lwkt_user_yield(); 2709 2710 if (info->object != p->object || 2711 p->pindex < info->start_pindex || 2712 p->pindex > info->end_pindex) { 2713 kprintf("vm_object_page_remove_callbackA: obj/pg race %p/%p\n", 2714 info->object, p); 2715 return(0); 2716 } 2717 if (vm_page_busy_try(p, TRUE)) { 2718 vm_page_sleep_busy(p, TRUE, "vmopar"); 2719 info->error = 1; 2720 return(0); 2721 } 2722 if (info->object != p->object) { 2723 /* this should never happen */ 2724 kprintf("vm_object_page_remove_callbackB: obj/pg race %p/%p\n", 2725 info->object, p); 2726 vm_page_wakeup(p); 2727 return(0); 2728 } 2729 2730 /* 2731 * Wired pages cannot be destroyed, but they can be invalidated 2732 * and we do so if clean_only (limit) is not set. 2733 * 2734 * WARNING! The page may be wired due to being part of a buffer 2735 * cache buffer, and the buffer might be marked B_CACHE. 2736 * This is fine as part of a truncation but VFSs must be 2737 * sure to fix the buffer up when re-extending the file. 2738 * 2739 * NOTE! PG_NEED_COMMIT is ignored. 2740 */ 2741 if (p->wire_count != 0) { 2742 vm_page_protect(p, VM_PROT_NONE); 2743 if (info->limit == 0) 2744 p->valid = 0; 2745 vm_page_wakeup(p); 2746 return(0); 2747 } 2748 2749 /* 2750 * limit is our clean_only flag. If set and the page is dirty or 2751 * requires a commit, do not free it. If set and the page is being 2752 * held by someone, do not free it. 2753 */ 2754 if (info->limit && p->valid) { 2755 vm_page_test_dirty(p); 2756 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) { 2757 vm_page_wakeup(p); 2758 return(0); 2759 } 2760 } 2761 2762 /* 2763 * Destroy the page 2764 */ 2765 vm_page_protect(p, VM_PROT_NONE); 2766 vm_page_free(p); 2767 2768 return(0); 2769 } 2770 2771 /* 2772 * Coalesces two objects backing up adjoining regions of memory into a 2773 * single object. 2774 * 2775 * returns TRUE if objects were combined. 2776 * 2777 * NOTE: Only works at the moment if the second object is NULL - 2778 * if it's not, which object do we lock first? 2779 * 2780 * Parameters: 2781 * prev_object First object to coalesce 2782 * prev_offset Offset into prev_object 2783 * next_object Second object into coalesce 2784 * next_offset Offset into next_object 2785 * 2786 * prev_size Size of reference to prev_object 2787 * next_size Size of reference to next_object 2788 * 2789 * The caller does not need to hold (prev_object) but must have a stable 2790 * pointer to it (typically by holding the vm_map locked). 2791 */ 2792 boolean_t 2793 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 2794 vm_size_t prev_size, vm_size_t next_size) 2795 { 2796 vm_pindex_t next_pindex; 2797 2798 if (prev_object == NULL) 2799 return (TRUE); 2800 2801 vm_object_hold(prev_object); 2802 2803 if (prev_object->type != OBJT_DEFAULT && 2804 prev_object->type != OBJT_SWAP) { 2805 vm_object_drop(prev_object); 2806 return (FALSE); 2807 } 2808 2809 /* 2810 * Try to collapse the object first 2811 */ 2812 vm_object_chain_acquire(prev_object, 0); 2813 vm_object_collapse(prev_object, NULL); 2814 2815 /* 2816 * Can't coalesce if: . more than one reference . paged out . shadows 2817 * another object . has a copy elsewhere (any of which mean that the 2818 * pages not mapped to prev_entry may be in use anyway) 2819 */ 2820 2821 if (prev_object->backing_object != NULL) { 2822 vm_object_chain_release(prev_object); 2823 vm_object_drop(prev_object); 2824 return (FALSE); 2825 } 2826 2827 prev_size >>= PAGE_SHIFT; 2828 next_size >>= PAGE_SHIFT; 2829 next_pindex = prev_pindex + prev_size; 2830 2831 if ((prev_object->ref_count > 1) && 2832 (prev_object->size != next_pindex)) { 2833 vm_object_chain_release(prev_object); 2834 vm_object_drop(prev_object); 2835 return (FALSE); 2836 } 2837 2838 /* 2839 * Remove any pages that may still be in the object from a previous 2840 * deallocation. 2841 */ 2842 if (next_pindex < prev_object->size) { 2843 vm_object_page_remove(prev_object, 2844 next_pindex, 2845 next_pindex + next_size, FALSE); 2846 if (prev_object->type == OBJT_SWAP) 2847 swap_pager_freespace(prev_object, 2848 next_pindex, next_size); 2849 } 2850 2851 /* 2852 * Extend the object if necessary. 2853 */ 2854 if (next_pindex + next_size > prev_object->size) 2855 prev_object->size = next_pindex + next_size; 2856 2857 vm_object_chain_release(prev_object); 2858 vm_object_drop(prev_object); 2859 return (TRUE); 2860 } 2861 2862 /* 2863 * Make the object writable and flag is being possibly dirty. 2864 * 2865 * The object might not be held (or might be held but held shared), 2866 * the related vnode is probably not held either. Object and vnode are 2867 * stable by virtue of the vm_page busied by the caller preventing 2868 * destruction. 2869 * 2870 * If the related mount is flagged MNTK_THR_SYNC we need to call 2871 * vsetobjdirty(). Filesystems using this option usually shortcut 2872 * synchronization by only scanning the syncer list. 2873 */ 2874 void 2875 vm_object_set_writeable_dirty(vm_object_t object) 2876 { 2877 struct vnode *vp; 2878 2879 /*vm_object_assert_held(object);*/ 2880 /* 2881 * Avoid contention in vm fault path by checking the state before 2882 * issuing an atomic op on it. 2883 */ 2884 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) != 2885 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) { 2886 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 2887 } 2888 if (object->type == OBJT_VNODE && 2889 (vp = (struct vnode *)object->handle) != NULL) { 2890 if ((vp->v_flag & VOBJDIRTY) == 0) { 2891 if (vp->v_mount && 2892 (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) { 2893 /* 2894 * New style THR_SYNC places vnodes on the 2895 * syncer list more deterministically. 2896 */ 2897 vsetobjdirty(vp); 2898 } else { 2899 /* 2900 * Old style scan would not necessarily place 2901 * a vnode on the syncer list when possibly 2902 * modified via mmap. 2903 */ 2904 vsetflags(vp, VOBJDIRTY); 2905 } 2906 } 2907 } 2908 } 2909 2910 #include "opt_ddb.h" 2911 #ifdef DDB 2912 #include <sys/kernel.h> 2913 2914 #include <sys/cons.h> 2915 2916 #include <ddb/ddb.h> 2917 2918 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 2919 vm_map_entry_t entry); 2920 static int vm_object_in_map (vm_object_t object); 2921 2922 /* 2923 * The caller must hold the object. 2924 */ 2925 static int 2926 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2927 { 2928 vm_map_t tmpm; 2929 vm_map_entry_t tmpe; 2930 vm_object_t obj, nobj; 2931 int entcount; 2932 2933 if (map == 0) 2934 return 0; 2935 if (entry == 0) { 2936 tmpe = map->header.next; 2937 entcount = map->nentries; 2938 while (entcount-- && (tmpe != &map->header)) { 2939 if( _vm_object_in_map(map, object, tmpe)) { 2940 return 1; 2941 } 2942 tmpe = tmpe->next; 2943 } 2944 return (0); 2945 } 2946 switch(entry->maptype) { 2947 case VM_MAPTYPE_SUBMAP: 2948 tmpm = entry->object.sub_map; 2949 tmpe = tmpm->header.next; 2950 entcount = tmpm->nentries; 2951 while (entcount-- && tmpe != &tmpm->header) { 2952 if( _vm_object_in_map(tmpm, object, tmpe)) { 2953 return 1; 2954 } 2955 tmpe = tmpe->next; 2956 } 2957 break; 2958 case VM_MAPTYPE_NORMAL: 2959 case VM_MAPTYPE_VPAGETABLE: 2960 obj = entry->object.vm_object; 2961 while (obj) { 2962 if (obj == object) { 2963 if (obj != entry->object.vm_object) 2964 vm_object_drop(obj); 2965 return 1; 2966 } 2967 while ((nobj = obj->backing_object) != NULL) { 2968 vm_object_hold(nobj); 2969 if (nobj == obj->backing_object) 2970 break; 2971 vm_object_drop(nobj); 2972 } 2973 if (obj != entry->object.vm_object) { 2974 if (nobj) 2975 vm_object_lock_swap(); 2976 vm_object_drop(obj); 2977 } 2978 obj = nobj; 2979 } 2980 break; 2981 default: 2982 break; 2983 } 2984 return 0; 2985 } 2986 2987 static int vm_object_in_map_callback(struct proc *p, void *data); 2988 2989 struct vm_object_in_map_info { 2990 vm_object_t object; 2991 int rv; 2992 }; 2993 2994 /* 2995 * Debugging only 2996 */ 2997 static int 2998 vm_object_in_map(vm_object_t object) 2999 { 3000 struct vm_object_in_map_info info; 3001 3002 info.rv = 0; 3003 info.object = object; 3004 3005 allproc_scan(vm_object_in_map_callback, &info); 3006 if (info.rv) 3007 return 1; 3008 if( _vm_object_in_map(&kernel_map, object, 0)) 3009 return 1; 3010 if( _vm_object_in_map(&pager_map, object, 0)) 3011 return 1; 3012 if( _vm_object_in_map(&buffer_map, object, 0)) 3013 return 1; 3014 return 0; 3015 } 3016 3017 /* 3018 * Debugging only 3019 */ 3020 static int 3021 vm_object_in_map_callback(struct proc *p, void *data) 3022 { 3023 struct vm_object_in_map_info *info = data; 3024 3025 if (p->p_vmspace) { 3026 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 3027 info->rv = 1; 3028 return -1; 3029 } 3030 } 3031 return (0); 3032 } 3033 3034 DB_SHOW_COMMAND(vmochk, vm_object_check) 3035 { 3036 struct vm_object_hash *hash; 3037 vm_object_t object; 3038 int n; 3039 3040 /* 3041 * make sure that internal objs are in a map somewhere 3042 * and none have zero ref counts. 3043 */ 3044 for (n = 0; n < VMOBJ_HSIZE; ++n) { 3045 hash = &vm_object_hash[n]; 3046 for (object = TAILQ_FIRST(&hash->list); 3047 object != NULL; 3048 object = TAILQ_NEXT(object, object_list)) { 3049 if (object->type == OBJT_MARKER) 3050 continue; 3051 if (object->handle != NULL || 3052 (object->type != OBJT_DEFAULT && 3053 object->type != OBJT_SWAP)) { 3054 continue; 3055 } 3056 if (object->ref_count == 0) { 3057 db_printf("vmochk: internal obj has " 3058 "zero ref count: %ld\n", 3059 (long)object->size); 3060 } 3061 if (vm_object_in_map(object)) 3062 continue; 3063 db_printf("vmochk: internal obj is not in a map: " 3064 "ref: %d, size: %lu: 0x%lx, " 3065 "backing_object: %p\n", 3066 object->ref_count, (u_long)object->size, 3067 (u_long)object->size, 3068 (void *)object->backing_object); 3069 } 3070 } 3071 } 3072 3073 /* 3074 * Debugging only 3075 */ 3076 DB_SHOW_COMMAND(object, vm_object_print_static) 3077 { 3078 /* XXX convert args. */ 3079 vm_object_t object = (vm_object_t)addr; 3080 boolean_t full = have_addr; 3081 3082 vm_page_t p; 3083 3084 /* XXX count is an (unused) arg. Avoid shadowing it. */ 3085 #define count was_count 3086 3087 int count; 3088 3089 if (object == NULL) 3090 return; 3091 3092 db_iprintf( 3093 "Object %p: type=%d, size=0x%lx, res=%ld, ref=%d, flags=0x%x\n", 3094 object, (int)object->type, (u_long)object->size, 3095 object->resident_page_count, object->ref_count, object->flags); 3096 /* 3097 * XXX no %qd in kernel. Truncate object->backing_object_offset. 3098 */ 3099 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 3100 object->shadow_count, 3101 object->backing_object ? object->backing_object->ref_count : 0, 3102 object->backing_object, (long)object->backing_object_offset); 3103 3104 if (!full) 3105 return; 3106 3107 db_indent += 2; 3108 count = 0; 3109 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 3110 if (count == 0) 3111 db_iprintf("memory:="); 3112 else if (count == 6) { 3113 db_printf("\n"); 3114 db_iprintf(" ..."); 3115 count = 0; 3116 } else 3117 db_printf(","); 3118 count++; 3119 3120 db_printf("(off=0x%lx,page=0x%lx)", 3121 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 3122 } 3123 if (count != 0) 3124 db_printf("\n"); 3125 db_indent -= 2; 3126 } 3127 3128 /* XXX. */ 3129 #undef count 3130 3131 /* 3132 * XXX need this non-static entry for calling from vm_map_print. 3133 * 3134 * Debugging only 3135 */ 3136 void 3137 vm_object_print(/* db_expr_t */ long addr, 3138 boolean_t have_addr, 3139 /* db_expr_t */ long count, 3140 char *modif) 3141 { 3142 vm_object_print_static(addr, have_addr, count, modif); 3143 } 3144 3145 /* 3146 * Debugging only 3147 */ 3148 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 3149 { 3150 struct vm_object_hash *hash; 3151 vm_object_t object; 3152 int nl = 0; 3153 int c; 3154 int n; 3155 3156 for (n = 0; n < VMOBJ_HSIZE; ++n) { 3157 hash = &vm_object_hash[n]; 3158 for (object = TAILQ_FIRST(&hash->list); 3159 object != NULL; 3160 object = TAILQ_NEXT(object, object_list)) { 3161 vm_pindex_t idx, fidx; 3162 vm_pindex_t osize; 3163 vm_paddr_t pa = -1, padiff; 3164 int rcount; 3165 vm_page_t m; 3166 3167 if (object->type == OBJT_MARKER) 3168 continue; 3169 db_printf("new object: %p\n", (void *)object); 3170 if ( nl > 18) { 3171 c = cngetc(); 3172 if (c != ' ') 3173 return; 3174 nl = 0; 3175 } 3176 nl++; 3177 rcount = 0; 3178 fidx = 0; 3179 osize = object->size; 3180 if (osize > 128) 3181 osize = 128; 3182 for (idx = 0; idx < osize; idx++) { 3183 m = vm_page_lookup(object, idx); 3184 if (m == NULL) { 3185 if (rcount) { 3186 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 3187 (long)fidx, rcount, (long)pa); 3188 if ( nl > 18) { 3189 c = cngetc(); 3190 if (c != ' ') 3191 return; 3192 nl = 0; 3193 } 3194 nl++; 3195 rcount = 0; 3196 } 3197 continue; 3198 } 3199 3200 if (rcount && 3201 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 3202 ++rcount; 3203 continue; 3204 } 3205 if (rcount) { 3206 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 3207 padiff >>= PAGE_SHIFT; 3208 padiff &= PQ_L2_MASK; 3209 if (padiff == 0) { 3210 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 3211 ++rcount; 3212 continue; 3213 } 3214 db_printf(" index(%ld)run(%d)pa(0x%lx)", 3215 (long)fidx, rcount, (long)pa); 3216 db_printf("pd(%ld)\n", (long)padiff); 3217 if ( nl > 18) { 3218 c = cngetc(); 3219 if (c != ' ') 3220 return; 3221 nl = 0; 3222 } 3223 nl++; 3224 } 3225 fidx = idx; 3226 pa = VM_PAGE_TO_PHYS(m); 3227 rcount = 1; 3228 } 3229 if (rcount) { 3230 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 3231 (long)fidx, rcount, (long)pa); 3232 if ( nl > 18) { 3233 c = cngetc(); 3234 if (c != ' ') 3235 return; 3236 nl = 0; 3237 } 3238 nl++; 3239 } 3240 } 3241 } 3242 } 3243 #endif /* DDB */ 3244