vm_object.c revision 92511
1/* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD: head/sys/vm/vm_object.c 92511 2002-03-17 18:37:37Z alc $ 65 */ 66 67/* 68 * Virtual memory object module. 69 */ 70 71#include <sys/param.h> 72#include <sys/systm.h> 73#include <sys/lock.h> 74#include <sys/mman.h> 75#include <sys/mount.h> 76#include <sys/kernel.h> 77#include <sys/sysctl.h> 78#include <sys/mutex.h> 79#include <sys/proc.h> /* for curproc, pageproc */ 80#include <sys/socket.h> 81#include <sys/vnode.h> 82#include <sys/vmmeter.h> 83#include <sys/sx.h> 84 85#include <vm/vm.h> 86#include <vm/vm_param.h> 87#include <vm/pmap.h> 88#include <vm/vm_map.h> 89#include <vm/vm_object.h> 90#include <vm/vm_page.h> 91#include <vm/vm_pageout.h> 92#include <vm/vm_pager.h> 93#include <vm/vm_zone.h> 94#include <vm/swap_pager.h> 95#include <vm/vm_kern.h> 96#include <vm/vm_extern.h> 97 98#define EASY_SCAN_FACTOR 8 99 100#define MSYNC_FLUSH_HARDSEQ 0x01 101#define MSYNC_FLUSH_SOFTSEQ 0x02 102 103/* 104 * msync / VM object flushing optimizations 105 */ 106static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 107SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 108 CTLFLAG_RW, &msync_flush_flags, 0, ""); 109 110static void vm_object_qcollapse(vm_object_t object); 111static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 112 113/* 114 * Virtual memory objects maintain the actual data 115 * associated with allocated virtual memory. A given 116 * page of memory exists within exactly one object. 117 * 118 * An object is only deallocated when all "references" 119 * are given up. Only one "reference" to a given 120 * region of an object should be writeable. 121 * 122 * Associated with each object is a list of all resident 123 * memory pages belonging to that object; this list is 124 * maintained by the "vm_page" module, and locked by the object's 125 * lock. 126 * 127 * Each object also records a "pager" routine which is 128 * used to retrieve (and store) pages to the proper backing 129 * storage. In addition, objects may be backed by other 130 * objects from which they were virtual-copied. 131 * 132 * The only items within the object structure which are 133 * modified after time of creation are: 134 * reference count locked by object's lock 135 * pager routine locked by object's lock 136 * 137 */ 138 139struct object_q vm_object_list; 140static struct mtx vm_object_list_mtx; /* lock for object list and count */ 141vm_object_t kernel_object; 142vm_object_t kmem_object; 143static struct vm_object kernel_object_store; 144static struct vm_object kmem_object_store; 145extern int vm_pageout_page_count; 146 147static long object_collapses; 148static long object_bypasses; 149static int next_index; 150static vm_zone_t obj_zone; 151static struct vm_zone obj_zone_store; 152static int object_hash_rand; 153#define VM_OBJECTS_INIT 256 154static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 155 156void 157_vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object) 158{ 159 int incr; 160 161 GIANT_REQUIRED; 162 163 TAILQ_INIT(&object->memq); 164 TAILQ_INIT(&object->shadow_head); 165 166 object->type = type; 167 object->size = size; 168 object->ref_count = 1; 169 object->flags = 0; 170 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 171 vm_object_set_flag(object, OBJ_ONEMAPPING); 172 object->paging_in_progress = 0; 173 object->resident_page_count = 0; 174 object->shadow_count = 0; 175 object->pg_color = next_index; 176 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 177 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 178 else 179 incr = size; 180 next_index = (next_index + incr) & PQ_L2_MASK; 181 object->handle = NULL; 182 object->backing_object = NULL; 183 object->backing_object_offset = (vm_ooffset_t) 0; 184 /* 185 * Try to generate a number that will spread objects out in the 186 * hash table. We 'wipe' new objects across the hash in 128 page 187 * increments plus 1 more to offset it a little more by the time 188 * it wraps around. 189 */ 190 object->hash_rand = object_hash_rand - 129; 191 192 object->generation++; 193 194 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 195 196 object_hash_rand = object->hash_rand; 197} 198 199/* 200 * vm_object_init: 201 * 202 * Initialize the VM objects module. 203 */ 204void 205vm_object_init(void) 206{ 207 GIANT_REQUIRED; 208 209 TAILQ_INIT(&vm_object_list); 210 mtx_init(&vm_object_list_mtx, "vm object_list", MTX_DEF); 211 212 kernel_object = &kernel_object_store; 213 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 214 kernel_object); 215 216 kmem_object = &kmem_object_store; 217 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 218 kmem_object); 219 220 obj_zone = &obj_zone_store; 221 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 222 vm_objects_init, VM_OBJECTS_INIT); 223} 224 225void 226vm_object_init2(void) 227{ 228 zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1); 229} 230 231void 232vm_object_set_flag(vm_object_t object, u_short bits) 233{ 234 GIANT_REQUIRED; 235 object->flags |= bits; 236} 237 238void 239vm_object_clear_flag(vm_object_t object, u_short bits) 240{ 241 GIANT_REQUIRED; 242 object->flags &= ~bits; 243} 244 245void 246vm_object_pip_add(vm_object_t object, short i) 247{ 248 GIANT_REQUIRED; 249 object->paging_in_progress += i; 250} 251 252void 253vm_object_pip_subtract(vm_object_t object, short i) 254{ 255 GIANT_REQUIRED; 256 object->paging_in_progress -= i; 257} 258 259void 260vm_object_pip_wakeup(vm_object_t object) 261{ 262 GIANT_REQUIRED; 263 object->paging_in_progress--; 264 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 265 vm_object_clear_flag(object, OBJ_PIPWNT); 266 wakeup(object); 267 } 268} 269 270void 271vm_object_pip_wakeupn(vm_object_t object, short i) 272{ 273 GIANT_REQUIRED; 274 if (i) 275 object->paging_in_progress -= i; 276 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 277 vm_object_clear_flag(object, OBJ_PIPWNT); 278 wakeup(object); 279 } 280} 281 282void 283vm_object_pip_sleep(vm_object_t object, char *waitid) 284{ 285 GIANT_REQUIRED; 286 if (object->paging_in_progress) { 287 int s = splvm(); 288 if (object->paging_in_progress) { 289 vm_object_set_flag(object, OBJ_PIPWNT); 290 tsleep(object, PVM, waitid, 0); 291 } 292 splx(s); 293 } 294} 295 296void 297vm_object_pip_wait(vm_object_t object, char *waitid) 298{ 299 GIANT_REQUIRED; 300 while (object->paging_in_progress) 301 vm_object_pip_sleep(object, waitid); 302} 303 304/* 305 * vm_object_allocate: 306 * 307 * Returns a new object with the given size. 308 */ 309vm_object_t 310vm_object_allocate(objtype_t type, vm_size_t size) 311{ 312 vm_object_t result; 313 314 GIANT_REQUIRED; 315 316 result = (vm_object_t) zalloc(obj_zone); 317 _vm_object_allocate(type, size, result); 318 319 return (result); 320} 321 322 323/* 324 * vm_object_reference: 325 * 326 * Gets another reference to the given object. 327 */ 328void 329vm_object_reference(vm_object_t object) 330{ 331 GIANT_REQUIRED; 332 333 if (object == NULL) 334 return; 335 336#if 0 337 /* object can be re-referenced during final cleaning */ 338 KASSERT(!(object->flags & OBJ_DEAD), 339 ("vm_object_reference: attempting to reference dead obj")); 340#endif 341 342 object->ref_count++; 343 if (object->type == OBJT_VNODE) { 344 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) { 345 printf("vm_object_reference: delay in getting object\n"); 346 } 347 } 348} 349 350/* 351 * handle deallocating a object of type OBJT_VNODE 352 */ 353void 354vm_object_vndeallocate(vm_object_t object) 355{ 356 struct vnode *vp = (struct vnode *) object->handle; 357 358 GIANT_REQUIRED; 359 KASSERT(object->type == OBJT_VNODE, 360 ("vm_object_vndeallocate: not a vnode object")); 361 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 362#ifdef INVARIANTS 363 if (object->ref_count == 0) { 364 vprint("vm_object_vndeallocate", vp); 365 panic("vm_object_vndeallocate: bad object reference count"); 366 } 367#endif 368 369 object->ref_count--; 370 if (object->ref_count == 0) { 371 vp->v_flag &= ~VTEXT; 372 vm_object_clear_flag(object, OBJ_OPT); 373 } 374 /* 375 * vrele may need a vop lock 376 */ 377 vrele(vp); 378} 379 380/* 381 * vm_object_deallocate: 382 * 383 * Release a reference to the specified object, 384 * gained either through a vm_object_allocate 385 * or a vm_object_reference call. When all references 386 * are gone, storage associated with this object 387 * may be relinquished. 388 * 389 * No object may be locked. 390 */ 391void 392vm_object_deallocate(vm_object_t object) 393{ 394 vm_object_t temp; 395 396 GIANT_REQUIRED; 397 398 while (object != NULL) { 399 400 if (object->type == OBJT_VNODE) { 401 vm_object_vndeallocate(object); 402 return; 403 } 404 405 KASSERT(object->ref_count != 0, 406 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 407 408 /* 409 * If the reference count goes to 0 we start calling 410 * vm_object_terminate() on the object chain. 411 * A ref count of 1 may be a special case depending on the 412 * shadow count being 0 or 1. 413 */ 414 object->ref_count--; 415 if (object->ref_count > 1) { 416 return; 417 } else if (object->ref_count == 1) { 418 if (object->shadow_count == 0) { 419 vm_object_set_flag(object, OBJ_ONEMAPPING); 420 } else if ((object->shadow_count == 1) && 421 (object->handle == NULL) && 422 (object->type == OBJT_DEFAULT || 423 object->type == OBJT_SWAP)) { 424 vm_object_t robject; 425 426 robject = TAILQ_FIRST(&object->shadow_head); 427 KASSERT(robject != NULL, 428 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 429 object->ref_count, 430 object->shadow_count)); 431 if ((robject->handle == NULL) && 432 (robject->type == OBJT_DEFAULT || 433 robject->type == OBJT_SWAP)) { 434 435 robject->ref_count++; 436 437 while ( 438 robject->paging_in_progress || 439 object->paging_in_progress 440 ) { 441 vm_object_pip_sleep(robject, "objde1"); 442 vm_object_pip_sleep(object, "objde2"); 443 } 444 445 if (robject->ref_count == 1) { 446 robject->ref_count--; 447 object = robject; 448 goto doterm; 449 } 450 451 object = robject; 452 vm_object_collapse(object); 453 continue; 454 } 455 } 456 457 return; 458 459 } 460 461doterm: 462 463 temp = object->backing_object; 464 if (temp) { 465 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 466 temp->shadow_count--; 467 if (temp->ref_count == 0) 468 vm_object_clear_flag(temp, OBJ_OPT); 469 temp->generation++; 470 object->backing_object = NULL; 471 } 472 /* 473 * Don't double-terminate, we could be in a termination 474 * recursion due to the terminate having to sync data 475 * to disk. 476 */ 477 if ((object->flags & OBJ_DEAD) == 0) 478 vm_object_terminate(object); 479 object = temp; 480 } 481} 482 483/* 484 * vm_object_terminate actually destroys the specified object, freeing 485 * up all previously used resources. 486 * 487 * The object must be locked. 488 * This routine may block. 489 */ 490void 491vm_object_terminate(vm_object_t object) 492{ 493 vm_page_t p; 494 int s; 495 496 GIANT_REQUIRED; 497 498 /* 499 * Make sure no one uses us. 500 */ 501 vm_object_set_flag(object, OBJ_DEAD); 502 503 /* 504 * wait for the pageout daemon to be done with the object 505 */ 506 vm_object_pip_wait(object, "objtrm"); 507 508 KASSERT(!object->paging_in_progress, 509 ("vm_object_terminate: pageout in progress")); 510 511 /* 512 * Clean and free the pages, as appropriate. All references to the 513 * object are gone, so we don't need to lock it. 514 */ 515 if (object->type == OBJT_VNODE) { 516 struct vnode *vp; 517 518 /* 519 * Freeze optimized copies. 520 */ 521 vm_freeze_copyopts(object, 0, object->size); 522 523 /* 524 * Clean pages and flush buffers. 525 */ 526 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 527 528 vp = (struct vnode *) object->handle; 529 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 530 } 531 532 KASSERT(object->ref_count == 0, 533 ("vm_object_terminate: object with references, ref_count=%d", 534 object->ref_count)); 535 536 /* 537 * Now free any remaining pages. For internal objects, this also 538 * removes them from paging queues. Don't free wired pages, just 539 * remove them from the object. 540 */ 541 s = splvm(); 542 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 543 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 544 ("vm_object_terminate: freeing busy page %p " 545 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 546 if (p->wire_count == 0) { 547 vm_page_busy(p); 548 vm_page_free(p); 549 cnt.v_pfree++; 550 } else { 551 vm_page_busy(p); 552 vm_page_remove(p); 553 } 554 } 555 splx(s); 556 557 /* 558 * Let the pager know object is dead. 559 */ 560 vm_pager_deallocate(object); 561 562 /* 563 * Remove the object from the global object list. 564 */ 565 mtx_lock(&vm_object_list_mtx); 566 TAILQ_REMOVE(&vm_object_list, object, object_list); 567 mtx_unlock(&vm_object_list_mtx); 568 569 wakeup(object); 570 571 /* 572 * Free the space for the object. 573 */ 574 zfree(obj_zone, object); 575} 576 577/* 578 * vm_object_page_clean 579 * 580 * Clean all dirty pages in the specified range of object. Leaves page 581 * on whatever queue it is currently on. If NOSYNC is set then do not 582 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 583 * leaving the object dirty. 584 * 585 * Odd semantics: if start == end, we clean everything. 586 * 587 * The object must be locked. 588 */ 589void 590vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 591{ 592 vm_page_t p, np; 593 vm_offset_t tstart, tend; 594 vm_pindex_t pi; 595 struct vnode *vp; 596 int clearobjflags; 597 int pagerflags; 598 int curgeneration; 599 600 GIANT_REQUIRED; 601 602 if (object->type != OBJT_VNODE || 603 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 604 return; 605 606 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0; 607 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 608 609 vp = object->handle; 610 611 vm_object_set_flag(object, OBJ_CLEANING); 612 613 tstart = start; 614 if (end == 0) { 615 tend = object->size; 616 } else { 617 tend = end; 618 } 619 620 /* 621 * If the caller is smart and only msync()s a range he knows is 622 * dirty, we may be able to avoid an object scan. This results in 623 * a phenominal improvement in performance. We cannot do this 624 * as a matter of course because the object may be huge - e.g. 625 * the size might be in the gigabytes or terrabytes. 626 */ 627 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 628 vm_offset_t tscan; 629 int scanlimit; 630 int scanreset; 631 632 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 633 if (scanreset < 16) 634 scanreset = 16; 635 636 scanlimit = scanreset; 637 tscan = tstart; 638 while (tscan < tend) { 639 curgeneration = object->generation; 640 p = vm_page_lookup(object, tscan); 641 if (p == NULL || p->valid == 0 || 642 (p->queue - p->pc) == PQ_CACHE) { 643 if (--scanlimit == 0) 644 break; 645 ++tscan; 646 continue; 647 } 648 vm_page_test_dirty(p); 649 if ((p->dirty & p->valid) == 0) { 650 if (--scanlimit == 0) 651 break; 652 ++tscan; 653 continue; 654 } 655 /* 656 * If we have been asked to skip nosync pages and 657 * this is a nosync page, we can't continue. 658 */ 659 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 660 if (--scanlimit == 0) 661 break; 662 ++tscan; 663 continue; 664 } 665 scanlimit = scanreset; 666 667 /* 668 * This returns 0 if it was unable to busy the first 669 * page (i.e. had to sleep). 670 */ 671 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 672 } 673 674 /* 675 * If everything was dirty and we flushed it successfully, 676 * and the requested range is not the entire object, we 677 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 678 * return immediately. 679 */ 680 if (tscan >= tend && (tstart || tend < object->size)) { 681 vm_object_clear_flag(object, OBJ_CLEANING); 682 return; 683 } 684 } 685 686 /* 687 * Generally set CLEANCHK interlock and make the page read-only so 688 * we can then clear the object flags. 689 * 690 * However, if this is a nosync mmap then the object is likely to 691 * stay dirty so do not mess with the page and do not clear the 692 * object flags. 693 */ 694 clearobjflags = 1; 695 696 TAILQ_FOREACH(p, &object->memq, listq) { 697 vm_page_flag_set(p, PG_CLEANCHK); 698 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 699 clearobjflags = 0; 700 else 701 vm_page_protect(p, VM_PROT_READ); 702 } 703 704 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 705 struct vnode *vp; 706 707 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 708 if (object->type == OBJT_VNODE && 709 (vp = (struct vnode *)object->handle) != NULL) { 710 if (vp->v_flag & VOBJDIRTY) { 711 mtx_lock(&vp->v_interlock); 712 vp->v_flag &= ~VOBJDIRTY; 713 mtx_unlock(&vp->v_interlock); 714 } 715 } 716 } 717 718rescan: 719 curgeneration = object->generation; 720 721 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 722 int n; 723 724 np = TAILQ_NEXT(p, listq); 725 726again: 727 pi = p->pindex; 728 if (((p->flags & PG_CLEANCHK) == 0) || 729 (pi < tstart) || (pi >= tend) || 730 (p->valid == 0) || 731 ((p->queue - p->pc) == PQ_CACHE)) { 732 vm_page_flag_clear(p, PG_CLEANCHK); 733 continue; 734 } 735 736 vm_page_test_dirty(p); 737 if ((p->dirty & p->valid) == 0) { 738 vm_page_flag_clear(p, PG_CLEANCHK); 739 continue; 740 } 741 742 /* 743 * If we have been asked to skip nosync pages and this is a 744 * nosync page, skip it. Note that the object flags were 745 * not cleared in this case so we do not have to set them. 746 */ 747 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 748 vm_page_flag_clear(p, PG_CLEANCHK); 749 continue; 750 } 751 752 n = vm_object_page_collect_flush(object, p, 753 curgeneration, pagerflags); 754 if (n == 0) 755 goto rescan; 756 757 if (object->generation != curgeneration) 758 goto rescan; 759 760 /* 761 * Try to optimize the next page. If we can't we pick up 762 * our (random) scan where we left off. 763 */ 764 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 765 if ((p = vm_page_lookup(object, pi + n)) != NULL) 766 goto again; 767 } 768 } 769 770#if 0 771 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 772#endif 773 774 vm_object_clear_flag(object, OBJ_CLEANING); 775 return; 776} 777 778static int 779vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 780{ 781 int runlen; 782 int s; 783 int maxf; 784 int chkb; 785 int maxb; 786 int i; 787 vm_pindex_t pi; 788 vm_page_t maf[vm_pageout_page_count]; 789 vm_page_t mab[vm_pageout_page_count]; 790 vm_page_t ma[vm_pageout_page_count]; 791 792 s = splvm(); 793 pi = p->pindex; 794 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 795 if (object->generation != curgeneration) { 796 splx(s); 797 return(0); 798 } 799 } 800 801 maxf = 0; 802 for(i = 1; i < vm_pageout_page_count; i++) { 803 vm_page_t tp; 804 805 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 806 if ((tp->flags & PG_BUSY) || 807 (tp->flags & PG_CLEANCHK) == 0 || 808 (tp->busy != 0)) 809 break; 810 if((tp->queue - tp->pc) == PQ_CACHE) { 811 vm_page_flag_clear(tp, PG_CLEANCHK); 812 break; 813 } 814 vm_page_test_dirty(tp); 815 if ((tp->dirty & tp->valid) == 0) { 816 vm_page_flag_clear(tp, PG_CLEANCHK); 817 break; 818 } 819 maf[ i - 1 ] = tp; 820 maxf++; 821 continue; 822 } 823 break; 824 } 825 826 maxb = 0; 827 chkb = vm_pageout_page_count - maxf; 828 if (chkb) { 829 for(i = 1; i < chkb;i++) { 830 vm_page_t tp; 831 832 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 833 if ((tp->flags & PG_BUSY) || 834 (tp->flags & PG_CLEANCHK) == 0 || 835 (tp->busy != 0)) 836 break; 837 if ((tp->queue - tp->pc) == PQ_CACHE) { 838 vm_page_flag_clear(tp, PG_CLEANCHK); 839 break; 840 } 841 vm_page_test_dirty(tp); 842 if ((tp->dirty & tp->valid) == 0) { 843 vm_page_flag_clear(tp, PG_CLEANCHK); 844 break; 845 } 846 mab[ i - 1 ] = tp; 847 maxb++; 848 continue; 849 } 850 break; 851 } 852 } 853 854 for(i = 0; i < maxb; i++) { 855 int index = (maxb - i) - 1; 856 ma[index] = mab[i]; 857 vm_page_flag_clear(ma[index], PG_CLEANCHK); 858 } 859 vm_page_flag_clear(p, PG_CLEANCHK); 860 ma[maxb] = p; 861 for(i = 0; i < maxf; i++) { 862 int index = (maxb + i) + 1; 863 ma[index] = maf[i]; 864 vm_page_flag_clear(ma[index], PG_CLEANCHK); 865 } 866 runlen = maxb + maxf + 1; 867 868 splx(s); 869 vm_pageout_flush(ma, runlen, pagerflags); 870 for (i = 0; i < runlen; i++) { 871 if (ma[i]->valid & ma[i]->dirty) { 872 vm_page_protect(ma[i], VM_PROT_READ); 873 vm_page_flag_set(ma[i], PG_CLEANCHK); 874 875 /* 876 * maxf will end up being the actual number of pages 877 * we wrote out contiguously, non-inclusive of the 878 * first page. We do not count look-behind pages. 879 */ 880 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 881 maxf = i - maxb - 1; 882 } 883 } 884 return(maxf + 1); 885} 886 887/* 888 * Same as vm_object_pmap_copy, except range checking really 889 * works, and is meant for small sections of an object. 890 * 891 * This code protects resident pages by making them read-only 892 * and is typically called on a fork or split when a page 893 * is converted to copy-on-write. 894 * 895 * NOTE: If the page is already at VM_PROT_NONE, calling 896 * vm_page_protect will have no effect. 897 */ 898void 899vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 900{ 901 vm_pindex_t idx; 902 vm_page_t p; 903 904 GIANT_REQUIRED; 905 906 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 907 return; 908 909 for (idx = start; idx < end; idx++) { 910 p = vm_page_lookup(object, idx); 911 if (p == NULL) 912 continue; 913 vm_page_protect(p, VM_PROT_READ); 914 } 915} 916 917/* 918 * vm_object_pmap_remove: 919 * 920 * Removes all physical pages in the specified 921 * object range from all physical maps. 922 * 923 * The object must *not* be locked. 924 */ 925void 926vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 927{ 928 vm_page_t p; 929 930 GIANT_REQUIRED; 931 if (object == NULL) 932 return; 933 TAILQ_FOREACH(p, &object->memq, listq) { 934 if (p->pindex >= start && p->pindex < end) 935 vm_page_protect(p, VM_PROT_NONE); 936 } 937 if ((start == 0) && (object->size == end)) 938 vm_object_clear_flag(object, OBJ_WRITEABLE); 939} 940 941/* 942 * vm_object_madvise: 943 * 944 * Implements the madvise function at the object/page level. 945 * 946 * MADV_WILLNEED (any object) 947 * 948 * Activate the specified pages if they are resident. 949 * 950 * MADV_DONTNEED (any object) 951 * 952 * Deactivate the specified pages if they are resident. 953 * 954 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 955 * OBJ_ONEMAPPING only) 956 * 957 * Deactivate and clean the specified pages if they are 958 * resident. This permits the process to reuse the pages 959 * without faulting or the kernel to reclaim the pages 960 * without I/O. 961 */ 962void 963vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 964{ 965 vm_pindex_t end, tpindex; 966 vm_object_t tobject; 967 vm_page_t m; 968 969 GIANT_REQUIRED; 970 if (object == NULL) 971 return; 972 973 end = pindex + count; 974 975 /* 976 * Locate and adjust resident pages 977 */ 978 for (; pindex < end; pindex += 1) { 979relookup: 980 tobject = object; 981 tpindex = pindex; 982shadowlookup: 983 /* 984 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 985 * and those pages must be OBJ_ONEMAPPING. 986 */ 987 if (advise == MADV_FREE) { 988 if ((tobject->type != OBJT_DEFAULT && 989 tobject->type != OBJT_SWAP) || 990 (tobject->flags & OBJ_ONEMAPPING) == 0) { 991 continue; 992 } 993 } 994 995 m = vm_page_lookup(tobject, tpindex); 996 997 if (m == NULL) { 998 /* 999 * There may be swap even if there is no backing page 1000 */ 1001 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1002 swap_pager_freespace(tobject, tpindex, 1); 1003 1004 /* 1005 * next object 1006 */ 1007 tobject = tobject->backing_object; 1008 if (tobject == NULL) 1009 continue; 1010 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1011 goto shadowlookup; 1012 } 1013 1014 /* 1015 * If the page is busy or not in a normal active state, 1016 * we skip it. If the page is not managed there are no 1017 * page queues to mess with. Things can break if we mess 1018 * with pages in any of the below states. 1019 */ 1020 if ( 1021 m->hold_count || 1022 m->wire_count || 1023 (m->flags & PG_UNMANAGED) || 1024 m->valid != VM_PAGE_BITS_ALL 1025 ) { 1026 continue; 1027 } 1028 1029 if (vm_page_sleep_busy(m, TRUE, "madvpo")) 1030 goto relookup; 1031 1032 if (advise == MADV_WILLNEED) { 1033 vm_page_activate(m); 1034 } else if (advise == MADV_DONTNEED) { 1035 vm_page_dontneed(m); 1036 } else if (advise == MADV_FREE) { 1037 /* 1038 * Mark the page clean. This will allow the page 1039 * to be freed up by the system. However, such pages 1040 * are often reused quickly by malloc()/free() 1041 * so we do not do anything that would cause 1042 * a page fault if we can help it. 1043 * 1044 * Specifically, we do not try to actually free 1045 * the page now nor do we try to put it in the 1046 * cache (which would cause a page fault on reuse). 1047 * 1048 * But we do make the page is freeable as we 1049 * can without actually taking the step of unmapping 1050 * it. 1051 */ 1052 pmap_clear_modify(m); 1053 m->dirty = 0; 1054 m->act_count = 0; 1055 vm_page_dontneed(m); 1056 if (tobject->type == OBJT_SWAP) 1057 swap_pager_freespace(tobject, tpindex, 1); 1058 } 1059 } 1060} 1061 1062/* 1063 * vm_object_shadow: 1064 * 1065 * Create a new object which is backed by the 1066 * specified existing object range. The source 1067 * object reference is deallocated. 1068 * 1069 * The new object and offset into that object 1070 * are returned in the source parameters. 1071 */ 1072void 1073vm_object_shadow( 1074 vm_object_t *object, /* IN/OUT */ 1075 vm_ooffset_t *offset, /* IN/OUT */ 1076 vm_size_t length) 1077{ 1078 vm_object_t source; 1079 vm_object_t result; 1080 1081 GIANT_REQUIRED; 1082 source = *object; 1083 1084 /* 1085 * Don't create the new object if the old object isn't shared. 1086 */ 1087 if (source != NULL && 1088 source->ref_count == 1 && 1089 source->handle == NULL && 1090 (source->type == OBJT_DEFAULT || 1091 source->type == OBJT_SWAP)) 1092 return; 1093 1094 /* 1095 * Allocate a new object with the given length 1096 */ 1097 result = vm_object_allocate(OBJT_DEFAULT, length); 1098 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1099 1100 /* 1101 * The new object shadows the source object, adding a reference to it. 1102 * Our caller changes his reference to point to the new object, 1103 * removing a reference to the source object. Net result: no change 1104 * of reference count. 1105 * 1106 * Try to optimize the result object's page color when shadowing 1107 * in order to maintain page coloring consistency in the combined 1108 * shadowed object. 1109 */ 1110 result->backing_object = source; 1111 if (source) { 1112 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1113 source->shadow_count++; 1114 source->generation++; 1115 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK; 1116 } 1117 1118 /* 1119 * Store the offset into the source object, and fix up the offset into 1120 * the new object. 1121 */ 1122 result->backing_object_offset = *offset; 1123 1124 /* 1125 * Return the new things 1126 */ 1127 *offset = 0; 1128 *object = result; 1129} 1130 1131#define OBSC_TEST_ALL_SHADOWED 0x0001 1132#define OBSC_COLLAPSE_NOWAIT 0x0002 1133#define OBSC_COLLAPSE_WAIT 0x0004 1134 1135static __inline int 1136vm_object_backing_scan(vm_object_t object, int op) 1137{ 1138 int s; 1139 int r = 1; 1140 vm_page_t p; 1141 vm_object_t backing_object; 1142 vm_pindex_t backing_offset_index; 1143 1144 s = splvm(); 1145 GIANT_REQUIRED; 1146 1147 backing_object = object->backing_object; 1148 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1149 1150 /* 1151 * Initial conditions 1152 */ 1153 if (op & OBSC_TEST_ALL_SHADOWED) { 1154 /* 1155 * We do not want to have to test for the existence of 1156 * swap pages in the backing object. XXX but with the 1157 * new swapper this would be pretty easy to do. 1158 * 1159 * XXX what about anonymous MAP_SHARED memory that hasn't 1160 * been ZFOD faulted yet? If we do not test for this, the 1161 * shadow test may succeed! XXX 1162 */ 1163 if (backing_object->type != OBJT_DEFAULT) { 1164 splx(s); 1165 return (0); 1166 } 1167 } 1168 if (op & OBSC_COLLAPSE_WAIT) { 1169 vm_object_set_flag(backing_object, OBJ_DEAD); 1170 } 1171 1172 /* 1173 * Our scan 1174 */ 1175 p = TAILQ_FIRST(&backing_object->memq); 1176 while (p) { 1177 vm_page_t next = TAILQ_NEXT(p, listq); 1178 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1179 1180 if (op & OBSC_TEST_ALL_SHADOWED) { 1181 vm_page_t pp; 1182 1183 /* 1184 * Ignore pages outside the parent object's range 1185 * and outside the parent object's mapping of the 1186 * backing object. 1187 * 1188 * note that we do not busy the backing object's 1189 * page. 1190 */ 1191 if ( 1192 p->pindex < backing_offset_index || 1193 new_pindex >= object->size 1194 ) { 1195 p = next; 1196 continue; 1197 } 1198 1199 /* 1200 * See if the parent has the page or if the parent's 1201 * object pager has the page. If the parent has the 1202 * page but the page is not valid, the parent's 1203 * object pager must have the page. 1204 * 1205 * If this fails, the parent does not completely shadow 1206 * the object and we might as well give up now. 1207 */ 1208 1209 pp = vm_page_lookup(object, new_pindex); 1210 if ( 1211 (pp == NULL || pp->valid == 0) && 1212 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1213 ) { 1214 r = 0; 1215 break; 1216 } 1217 } 1218 1219 /* 1220 * Check for busy page 1221 */ 1222 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1223 vm_page_t pp; 1224 1225 if (op & OBSC_COLLAPSE_NOWAIT) { 1226 if ( 1227 (p->flags & PG_BUSY) || 1228 !p->valid || 1229 p->hold_count || 1230 p->wire_count || 1231 p->busy 1232 ) { 1233 p = next; 1234 continue; 1235 } 1236 } else if (op & OBSC_COLLAPSE_WAIT) { 1237 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1238 /* 1239 * If we slept, anything could have 1240 * happened. Since the object is 1241 * marked dead, the backing offset 1242 * should not have changed so we 1243 * just restart our scan. 1244 */ 1245 p = TAILQ_FIRST(&backing_object->memq); 1246 continue; 1247 } 1248 } 1249 1250 /* 1251 * Busy the page 1252 */ 1253 vm_page_busy(p); 1254 1255 KASSERT( 1256 p->object == backing_object, 1257 ("vm_object_qcollapse(): object mismatch") 1258 ); 1259 1260 /* 1261 * Destroy any associated swap 1262 */ 1263 if (backing_object->type == OBJT_SWAP) { 1264 swap_pager_freespace( 1265 backing_object, 1266 p->pindex, 1267 1 1268 ); 1269 } 1270 1271 if ( 1272 p->pindex < backing_offset_index || 1273 new_pindex >= object->size 1274 ) { 1275 /* 1276 * Page is out of the parent object's range, we 1277 * can simply destroy it. 1278 */ 1279 vm_page_protect(p, VM_PROT_NONE); 1280 vm_page_free(p); 1281 p = next; 1282 continue; 1283 } 1284 1285 pp = vm_page_lookup(object, new_pindex); 1286 if ( 1287 pp != NULL || 1288 vm_pager_has_page(object, new_pindex, NULL, NULL) 1289 ) { 1290 /* 1291 * page already exists in parent OR swap exists 1292 * for this location in the parent. Destroy 1293 * the original page from the backing object. 1294 * 1295 * Leave the parent's page alone 1296 */ 1297 vm_page_protect(p, VM_PROT_NONE); 1298 vm_page_free(p); 1299 p = next; 1300 continue; 1301 } 1302 1303 /* 1304 * Page does not exist in parent, rename the 1305 * page from the backing object to the main object. 1306 * 1307 * If the page was mapped to a process, it can remain 1308 * mapped through the rename. 1309 */ 1310 if ((p->queue - p->pc) == PQ_CACHE) 1311 vm_page_deactivate(p); 1312 1313 vm_page_rename(p, object, new_pindex); 1314 /* page automatically made dirty by rename */ 1315 } 1316 p = next; 1317 } 1318 splx(s); 1319 return (r); 1320} 1321 1322 1323/* 1324 * this version of collapse allows the operation to occur earlier and 1325 * when paging_in_progress is true for an object... This is not a complete 1326 * operation, but should plug 99.9% of the rest of the leaks. 1327 */ 1328static void 1329vm_object_qcollapse(vm_object_t object) 1330{ 1331 vm_object_t backing_object = object->backing_object; 1332 1333 GIANT_REQUIRED; 1334 1335 if (backing_object->ref_count != 1) 1336 return; 1337 1338 backing_object->ref_count += 2; 1339 1340 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1341 1342 backing_object->ref_count -= 2; 1343} 1344 1345/* 1346 * vm_object_collapse: 1347 * 1348 * Collapse an object with the object backing it. 1349 * Pages in the backing object are moved into the 1350 * parent, and the backing object is deallocated. 1351 */ 1352void 1353vm_object_collapse(vm_object_t object) 1354{ 1355 GIANT_REQUIRED; 1356 1357 while (TRUE) { 1358 vm_object_t backing_object; 1359 1360 /* 1361 * Verify that the conditions are right for collapse: 1362 * 1363 * The object exists and the backing object exists. 1364 */ 1365 if (object == NULL) 1366 break; 1367 1368 if ((backing_object = object->backing_object) == NULL) 1369 break; 1370 1371 /* 1372 * we check the backing object first, because it is most likely 1373 * not collapsable. 1374 */ 1375 if (backing_object->handle != NULL || 1376 (backing_object->type != OBJT_DEFAULT && 1377 backing_object->type != OBJT_SWAP) || 1378 (backing_object->flags & OBJ_DEAD) || 1379 object->handle != NULL || 1380 (object->type != OBJT_DEFAULT && 1381 object->type != OBJT_SWAP) || 1382 (object->flags & OBJ_DEAD)) { 1383 break; 1384 } 1385 1386 if ( 1387 object->paging_in_progress != 0 || 1388 backing_object->paging_in_progress != 0 1389 ) { 1390 vm_object_qcollapse(object); 1391 break; 1392 } 1393 1394 /* 1395 * We know that we can either collapse the backing object (if 1396 * the parent is the only reference to it) or (perhaps) have 1397 * the parent bypass the object if the parent happens to shadow 1398 * all the resident pages in the entire backing object. 1399 * 1400 * This is ignoring pager-backed pages such as swap pages. 1401 * vm_object_backing_scan fails the shadowing test in this 1402 * case. 1403 */ 1404 if (backing_object->ref_count == 1) { 1405 /* 1406 * If there is exactly one reference to the backing 1407 * object, we can collapse it into the parent. 1408 */ 1409 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1410 1411 /* 1412 * Move the pager from backing_object to object. 1413 */ 1414 if (backing_object->type == OBJT_SWAP) { 1415 vm_object_pip_add(backing_object, 1); 1416 1417 /* 1418 * scrap the paging_offset junk and do a 1419 * discrete copy. This also removes major 1420 * assumptions about how the swap-pager 1421 * works from where it doesn't belong. The 1422 * new swapper is able to optimize the 1423 * destroy-source case. 1424 */ 1425 vm_object_pip_add(object, 1); 1426 swap_pager_copy( 1427 backing_object, 1428 object, 1429 OFF_TO_IDX(object->backing_object_offset), TRUE); 1430 vm_object_pip_wakeup(object); 1431 1432 vm_object_pip_wakeup(backing_object); 1433 } 1434 /* 1435 * Object now shadows whatever backing_object did. 1436 * Note that the reference to 1437 * backing_object->backing_object moves from within 1438 * backing_object to within object. 1439 */ 1440 TAILQ_REMOVE( 1441 &object->backing_object->shadow_head, 1442 object, 1443 shadow_list 1444 ); 1445 object->backing_object->shadow_count--; 1446 object->backing_object->generation++; 1447 if (backing_object->backing_object) { 1448 TAILQ_REMOVE( 1449 &backing_object->backing_object->shadow_head, 1450 backing_object, 1451 shadow_list 1452 ); 1453 backing_object->backing_object->shadow_count--; 1454 backing_object->backing_object->generation++; 1455 } 1456 object->backing_object = backing_object->backing_object; 1457 if (object->backing_object) { 1458 TAILQ_INSERT_TAIL( 1459 &object->backing_object->shadow_head, 1460 object, 1461 shadow_list 1462 ); 1463 object->backing_object->shadow_count++; 1464 object->backing_object->generation++; 1465 } 1466 1467 object->backing_object_offset += 1468 backing_object->backing_object_offset; 1469 1470 /* 1471 * Discard backing_object. 1472 * 1473 * Since the backing object has no pages, no pager left, 1474 * and no object references within it, all that is 1475 * necessary is to dispose of it. 1476 */ 1477 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1478 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1479 1480 TAILQ_REMOVE( 1481 &vm_object_list, 1482 backing_object, 1483 object_list 1484 ); 1485 1486 zfree(obj_zone, backing_object); 1487 1488 object_collapses++; 1489 } else { 1490 vm_object_t new_backing_object; 1491 1492 /* 1493 * If we do not entirely shadow the backing object, 1494 * there is nothing we can do so we give up. 1495 */ 1496 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1497 break; 1498 } 1499 1500 /* 1501 * Make the parent shadow the next object in the 1502 * chain. Deallocating backing_object will not remove 1503 * it, since its reference count is at least 2. 1504 */ 1505 TAILQ_REMOVE( 1506 &backing_object->shadow_head, 1507 object, 1508 shadow_list 1509 ); 1510 backing_object->shadow_count--; 1511 backing_object->generation++; 1512 1513 new_backing_object = backing_object->backing_object; 1514 if ((object->backing_object = new_backing_object) != NULL) { 1515 vm_object_reference(new_backing_object); 1516 TAILQ_INSERT_TAIL( 1517 &new_backing_object->shadow_head, 1518 object, 1519 shadow_list 1520 ); 1521 new_backing_object->shadow_count++; 1522 new_backing_object->generation++; 1523 object->backing_object_offset += 1524 backing_object->backing_object_offset; 1525 } 1526 1527 /* 1528 * Drop the reference count on backing_object. Since 1529 * its ref_count was at least 2, it will not vanish; 1530 * so we don't need to call vm_object_deallocate, but 1531 * we do anyway. 1532 */ 1533 vm_object_deallocate(backing_object); 1534 object_bypasses++; 1535 } 1536 1537 /* 1538 * Try again with this object's new backing object. 1539 */ 1540 } 1541} 1542 1543/* 1544 * vm_object_page_remove: [internal] 1545 * 1546 * Removes all physical pages in the specified 1547 * object range from the object's list of pages. 1548 * 1549 * The object must be locked. 1550 */ 1551void 1552vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1553{ 1554 vm_page_t p, next; 1555 unsigned int size; 1556 int all; 1557 1558 GIANT_REQUIRED; 1559 1560 if (object == NULL || 1561 object->resident_page_count == 0) 1562 return; 1563 1564 all = ((end == 0) && (start == 0)); 1565 1566 /* 1567 * Since physically-backed objects do not use managed pages, we can't 1568 * remove pages from the object (we must instead remove the page 1569 * references, and then destroy the object). 1570 */ 1571 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1572 1573 vm_object_pip_add(object, 1); 1574again: 1575 size = end - start; 1576 if (all || size > object->resident_page_count / 4) { 1577 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) { 1578 next = TAILQ_NEXT(p, listq); 1579 if (all || ((start <= p->pindex) && (p->pindex < end))) { 1580 if (p->wire_count != 0) { 1581 vm_page_protect(p, VM_PROT_NONE); 1582 if (!clean_only) 1583 p->valid = 0; 1584 continue; 1585 } 1586 1587 /* 1588 * The busy flags are only cleared at 1589 * interrupt -- minimize the spl transitions 1590 */ 1591 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1592 goto again; 1593 1594 if (clean_only && p->valid) { 1595 vm_page_test_dirty(p); 1596 if (p->valid & p->dirty) 1597 continue; 1598 } 1599 1600 vm_page_busy(p); 1601 vm_page_protect(p, VM_PROT_NONE); 1602 vm_page_free(p); 1603 } 1604 } 1605 } else { 1606 while (size > 0) { 1607 if ((p = vm_page_lookup(object, start)) != 0) { 1608 1609 if (p->wire_count != 0) { 1610 vm_page_protect(p, VM_PROT_NONE); 1611 if (!clean_only) 1612 p->valid = 0; 1613 start += 1; 1614 size -= 1; 1615 continue; 1616 } 1617 1618 /* 1619 * The busy flags are only cleared at 1620 * interrupt -- minimize the spl transitions 1621 */ 1622 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1623 goto again; 1624 1625 if (clean_only && p->valid) { 1626 vm_page_test_dirty(p); 1627 if (p->valid & p->dirty) { 1628 start += 1; 1629 size -= 1; 1630 continue; 1631 } 1632 } 1633 1634 vm_page_busy(p); 1635 vm_page_protect(p, VM_PROT_NONE); 1636 vm_page_free(p); 1637 } 1638 start += 1; 1639 size -= 1; 1640 } 1641 } 1642 vm_object_pip_wakeup(object); 1643} 1644 1645/* 1646 * Routine: vm_object_coalesce 1647 * Function: Coalesces two objects backing up adjoining 1648 * regions of memory into a single object. 1649 * 1650 * returns TRUE if objects were combined. 1651 * 1652 * NOTE: Only works at the moment if the second object is NULL - 1653 * if it's not, which object do we lock first? 1654 * 1655 * Parameters: 1656 * prev_object First object to coalesce 1657 * prev_offset Offset into prev_object 1658 * next_object Second object into coalesce 1659 * next_offset Offset into next_object 1660 * 1661 * prev_size Size of reference to prev_object 1662 * next_size Size of reference to next_object 1663 * 1664 * Conditions: 1665 * The object must *not* be locked. 1666 */ 1667boolean_t 1668vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size) 1669{ 1670 vm_pindex_t next_pindex; 1671 1672 GIANT_REQUIRED; 1673 1674 if (prev_object == NULL) { 1675 return (TRUE); 1676 } 1677 1678 if (prev_object->type != OBJT_DEFAULT && 1679 prev_object->type != OBJT_SWAP) { 1680 return (FALSE); 1681 } 1682 1683 /* 1684 * Try to collapse the object first 1685 */ 1686 vm_object_collapse(prev_object); 1687 1688 /* 1689 * Can't coalesce if: . more than one reference . paged out . shadows 1690 * another object . has a copy elsewhere (any of which mean that the 1691 * pages not mapped to prev_entry may be in use anyway) 1692 */ 1693 if (prev_object->backing_object != NULL) { 1694 return (FALSE); 1695 } 1696 1697 prev_size >>= PAGE_SHIFT; 1698 next_size >>= PAGE_SHIFT; 1699 next_pindex = prev_pindex + prev_size; 1700 1701 if ((prev_object->ref_count > 1) && 1702 (prev_object->size != next_pindex)) { 1703 return (FALSE); 1704 } 1705 1706 /* 1707 * Remove any pages that may still be in the object from a previous 1708 * deallocation. 1709 */ 1710 if (next_pindex < prev_object->size) { 1711 vm_object_page_remove(prev_object, 1712 next_pindex, 1713 next_pindex + next_size, FALSE); 1714 if (prev_object->type == OBJT_SWAP) 1715 swap_pager_freespace(prev_object, 1716 next_pindex, next_size); 1717 } 1718 1719 /* 1720 * Extend the object if necessary. 1721 */ 1722 if (next_pindex + next_size > prev_object->size) 1723 prev_object->size = next_pindex + next_size; 1724 1725 return (TRUE); 1726} 1727 1728void 1729vm_object_set_writeable_dirty(vm_object_t object) 1730{ 1731 struct vnode *vp; 1732 1733 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1734 if (object->type == OBJT_VNODE && 1735 (vp = (struct vnode *)object->handle) != NULL) { 1736 if ((vp->v_flag & VOBJDIRTY) == 0) { 1737 mtx_lock(&vp->v_interlock); 1738 vp->v_flag |= VOBJDIRTY; 1739 mtx_unlock(&vp->v_interlock); 1740 } 1741 } 1742} 1743 1744#include "opt_ddb.h" 1745#ifdef DDB 1746#include <sys/kernel.h> 1747 1748#include <sys/cons.h> 1749 1750#include <ddb/ddb.h> 1751 1752static int 1753_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1754{ 1755 vm_map_t tmpm; 1756 vm_map_entry_t tmpe; 1757 vm_object_t obj; 1758 int entcount; 1759 1760 if (map == 0) 1761 return 0; 1762 1763 if (entry == 0) { 1764 tmpe = map->header.next; 1765 entcount = map->nentries; 1766 while (entcount-- && (tmpe != &map->header)) { 1767 if (_vm_object_in_map(map, object, tmpe)) { 1768 return 1; 1769 } 1770 tmpe = tmpe->next; 1771 } 1772 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1773 tmpm = entry->object.sub_map; 1774 tmpe = tmpm->header.next; 1775 entcount = tmpm->nentries; 1776 while (entcount-- && tmpe != &tmpm->header) { 1777 if (_vm_object_in_map(tmpm, object, tmpe)) { 1778 return 1; 1779 } 1780 tmpe = tmpe->next; 1781 } 1782 } else if ((obj = entry->object.vm_object) != NULL) { 1783 for (; obj; obj = obj->backing_object) 1784 if (obj == object) { 1785 return 1; 1786 } 1787 } 1788 return 0; 1789} 1790 1791static int 1792vm_object_in_map(vm_object_t object) 1793{ 1794 struct proc *p; 1795 1796 /* sx_slock(&allproc_lock); */ 1797 LIST_FOREACH(p, &allproc, p_list) { 1798 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1799 continue; 1800 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1801 /* sx_sunlock(&allproc_lock); */ 1802 return 1; 1803 } 1804 } 1805 /* sx_sunlock(&allproc_lock); */ 1806 if (_vm_object_in_map(kernel_map, object, 0)) 1807 return 1; 1808 if (_vm_object_in_map(kmem_map, object, 0)) 1809 return 1; 1810 if (_vm_object_in_map(pager_map, object, 0)) 1811 return 1; 1812 if (_vm_object_in_map(buffer_map, object, 0)) 1813 return 1; 1814 return 0; 1815} 1816 1817DB_SHOW_COMMAND(vmochk, vm_object_check) 1818{ 1819 vm_object_t object; 1820 1821 /* 1822 * make sure that internal objs are in a map somewhere 1823 * and none have zero ref counts. 1824 */ 1825 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1826 if (object->handle == NULL && 1827 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1828 if (object->ref_count == 0) { 1829 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1830 (long)object->size); 1831 } 1832 if (!vm_object_in_map(object)) { 1833 db_printf( 1834 "vmochk: internal obj is not in a map: " 1835 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1836 object->ref_count, (u_long)object->size, 1837 (u_long)object->size, 1838 (void *)object->backing_object); 1839 } 1840 } 1841 } 1842} 1843 1844/* 1845 * vm_object_print: [ debug ] 1846 */ 1847DB_SHOW_COMMAND(object, vm_object_print_static) 1848{ 1849 /* XXX convert args. */ 1850 vm_object_t object = (vm_object_t)addr; 1851 boolean_t full = have_addr; 1852 1853 vm_page_t p; 1854 1855 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1856#define count was_count 1857 1858 int count; 1859 1860 if (object == NULL) 1861 return; 1862 1863 db_iprintf( 1864 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 1865 object, (int)object->type, (u_long)object->size, 1866 object->resident_page_count, object->ref_count, object->flags); 1867 /* 1868 * XXX no %qd in kernel. Truncate object->backing_object_offset. 1869 */ 1870 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 1871 object->shadow_count, 1872 object->backing_object ? object->backing_object->ref_count : 0, 1873 object->backing_object, (long)object->backing_object_offset); 1874 1875 if (!full) 1876 return; 1877 1878 db_indent += 2; 1879 count = 0; 1880 TAILQ_FOREACH(p, &object->memq, listq) { 1881 if (count == 0) 1882 db_iprintf("memory:="); 1883 else if (count == 6) { 1884 db_printf("\n"); 1885 db_iprintf(" ..."); 1886 count = 0; 1887 } else 1888 db_printf(","); 1889 count++; 1890 1891 db_printf("(off=0x%lx,page=0x%lx)", 1892 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 1893 } 1894 if (count != 0) 1895 db_printf("\n"); 1896 db_indent -= 2; 1897} 1898 1899/* XXX. */ 1900#undef count 1901 1902/* XXX need this non-static entry for calling from vm_map_print. */ 1903void 1904vm_object_print( 1905 /* db_expr_t */ long addr, 1906 boolean_t have_addr, 1907 /* db_expr_t */ long count, 1908 char *modif) 1909{ 1910 vm_object_print_static(addr, have_addr, count, modif); 1911} 1912 1913DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 1914{ 1915 vm_object_t object; 1916 int nl = 0; 1917 int c; 1918 1919 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1920 vm_pindex_t idx, fidx; 1921 vm_pindex_t osize; 1922 vm_offset_t pa = -1, padiff; 1923 int rcount; 1924 vm_page_t m; 1925 1926 db_printf("new object: %p\n", (void *)object); 1927 if (nl > 18) { 1928 c = cngetc(); 1929 if (c != ' ') 1930 return; 1931 nl = 0; 1932 } 1933 nl++; 1934 rcount = 0; 1935 fidx = 0; 1936 osize = object->size; 1937 if (osize > 128) 1938 osize = 128; 1939 for (idx = 0; idx < osize; idx++) { 1940 m = vm_page_lookup(object, idx); 1941 if (m == NULL) { 1942 if (rcount) { 1943 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 1944 (long)fidx, rcount, (long)pa); 1945 if (nl > 18) { 1946 c = cngetc(); 1947 if (c != ' ') 1948 return; 1949 nl = 0; 1950 } 1951 nl++; 1952 rcount = 0; 1953 } 1954 continue; 1955 } 1956 1957 1958 if (rcount && 1959 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 1960 ++rcount; 1961 continue; 1962 } 1963 if (rcount) { 1964 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 1965 padiff >>= PAGE_SHIFT; 1966 padiff &= PQ_L2_MASK; 1967 if (padiff == 0) { 1968 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 1969 ++rcount; 1970 continue; 1971 } 1972 db_printf(" index(%ld)run(%d)pa(0x%lx)", 1973 (long)fidx, rcount, (long)pa); 1974 db_printf("pd(%ld)\n", (long)padiff); 1975 if (nl > 18) { 1976 c = cngetc(); 1977 if (c != ' ') 1978 return; 1979 nl = 0; 1980 } 1981 nl++; 1982 } 1983 fidx = idx; 1984 pa = VM_PAGE_TO_PHYS(m); 1985 rcount = 1; 1986 } 1987 if (rcount) { 1988 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 1989 (long)fidx, rcount, (long)pa); 1990 if (nl > 18) { 1991 c = cngetc(); 1992 if (c != ' ') 1993 return; 1994 nl = 0; 1995 } 1996 nl++; 1997 } 1998 } 1999} 2000#endif /* DDB */ 2001