vm_object.c revision 116079
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 116079 2003-06-09 06:50:02Z 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/swap_pager.h> 94#include <vm/vm_kern.h> 95#include <vm/vm_extern.h> 96#include <vm/uma.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; 140struct mtx vm_object_list_mtx; /* lock for object list and count */ 141 142struct vm_object kernel_object_store; 143struct vm_object kmem_object_store; 144 145static long object_collapses; 146static long object_bypasses; 147static int next_index; 148static uma_zone_t obj_zone; 149#define VM_OBJECTS_INIT 256 150 151static void vm_object_zinit(void *mem, int size); 152 153#ifdef INVARIANTS 154static void vm_object_zdtor(void *mem, int size, void *arg); 155 156static void 157vm_object_zdtor(void *mem, int size, void *arg) 158{ 159 vm_object_t object; 160 161 object = (vm_object_t)mem; 162 KASSERT(object->paging_in_progress == 0, 163 ("object %p paging_in_progress = %d", 164 object, object->paging_in_progress)); 165 KASSERT(object->resident_page_count == 0, 166 ("object %p resident_page_count = %d", 167 object, object->resident_page_count)); 168 KASSERT(object->shadow_count == 0, 169 ("object %p shadow_count = %d", 170 object, object->shadow_count)); 171} 172#endif 173 174static void 175vm_object_zinit(void *mem, int size) 176{ 177 vm_object_t object; 178 179 object = (vm_object_t)mem; 180 bzero(&object->mtx, sizeof(object->mtx)); 181 VM_OBJECT_LOCK_INIT(object); 182 183 /* These are true for any object that has been freed */ 184 object->paging_in_progress = 0; 185 object->resident_page_count = 0; 186 object->shadow_count = 0; 187} 188 189void 190_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 191{ 192 int incr; 193 194 TAILQ_INIT(&object->memq); 195 LIST_INIT(&object->shadow_head); 196 197 object->root = NULL; 198 object->type = type; 199 object->size = size; 200 object->ref_count = 1; 201 object->flags = 0; 202 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 203 object->flags = OBJ_ONEMAPPING; 204 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 205 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 206 else 207 incr = size; 208 do 209 object->pg_color = next_index; 210 while (!atomic_cmpset_int(&next_index, object->pg_color, 211 (object->pg_color + incr) & PQ_L2_MASK)); 212 object->handle = NULL; 213 object->backing_object = NULL; 214 object->backing_object_offset = (vm_ooffset_t) 0; 215 216 atomic_add_int(&object->generation, 1); 217 218 mtx_lock(&vm_object_list_mtx); 219 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 220 mtx_unlock(&vm_object_list_mtx); 221} 222 223/* 224 * vm_object_init: 225 * 226 * Initialize the VM objects module. 227 */ 228void 229vm_object_init(void) 230{ 231 TAILQ_INIT(&vm_object_list); 232 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 233 234 VM_OBJECT_LOCK_INIT(&kernel_object_store); 235 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 236 kernel_object); 237 238 /* 239 * The kmem object's mutex is given a unique name, instead of 240 * "vm object", to avoid false reports of lock-order reversal 241 * with a system map mutex. 242 */ 243 mtx_init(VM_OBJECT_MTX(kmem_object), "kmem object", NULL, MTX_DEF); 244 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 245 kmem_object); 246 247 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 248#ifdef INVARIANTS 249 vm_object_zdtor, 250#else 251 NULL, 252#endif 253 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 254 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 255} 256 257void 258vm_object_set_flag(vm_object_t object, u_short bits) 259{ 260 object->flags |= bits; 261} 262 263void 264vm_object_clear_flag(vm_object_t object, u_short bits) 265{ 266 267 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 268 object->flags &= ~bits; 269} 270 271void 272vm_object_pip_add(vm_object_t object, short i) 273{ 274 275 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 276 object->paging_in_progress += i; 277} 278 279void 280vm_object_pip_subtract(vm_object_t object, short i) 281{ 282 283 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 284 object->paging_in_progress -= i; 285} 286 287void 288vm_object_pip_wakeup(vm_object_t object) 289{ 290 291 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 292 object->paging_in_progress--; 293 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 294 vm_object_clear_flag(object, OBJ_PIPWNT); 295 wakeup(object); 296 } 297} 298 299void 300vm_object_pip_wakeupn(vm_object_t object, short i) 301{ 302 303 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 304 if (i) 305 object->paging_in_progress -= i; 306 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 307 vm_object_clear_flag(object, OBJ_PIPWNT); 308 wakeup(object); 309 } 310} 311 312void 313vm_object_pip_wait(vm_object_t object, char *waitid) 314{ 315 316 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 317 while (object->paging_in_progress) { 318 object->flags |= OBJ_PIPWNT; 319 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 320 } 321} 322 323/* 324 * vm_object_allocate_wait 325 * 326 * Return a new object with the given size, and give the user the 327 * option of waiting for it to complete or failing if the needed 328 * memory isn't available. 329 */ 330vm_object_t 331vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 332{ 333 vm_object_t result; 334 335 result = (vm_object_t) uma_zalloc(obj_zone, flags); 336 337 if (result != NULL) 338 _vm_object_allocate(type, size, result); 339 340 return (result); 341} 342 343/* 344 * vm_object_allocate: 345 * 346 * Returns a new object with the given size. 347 */ 348vm_object_t 349vm_object_allocate(objtype_t type, vm_pindex_t size) 350{ 351 return(vm_object_allocate_wait(type, size, M_WAITOK)); 352} 353 354 355/* 356 * vm_object_reference: 357 * 358 * Gets another reference to the given object. Note: OBJ_DEAD 359 * objects can be referenced during final cleaning. 360 */ 361void 362vm_object_reference(vm_object_t object) 363{ 364 if (object == NULL) 365 return; 366 if (object != kmem_object) 367 mtx_lock(&Giant); 368 VM_OBJECT_LOCK(object); 369 object->ref_count++; 370 VM_OBJECT_UNLOCK(object); 371 if (object->type == OBJT_VNODE) { 372 while (vget((struct vnode *) object->handle, LK_RETRY, curthread)) { 373 printf("vm_object_reference: delay in getting object\n"); 374 } 375 } 376 if (object != kmem_object) 377 mtx_unlock(&Giant); 378} 379 380/* 381 * Handle deallocating an object of type OBJT_VNODE. 382 */ 383void 384vm_object_vndeallocate(vm_object_t object) 385{ 386 struct vnode *vp = (struct vnode *) object->handle; 387 388 GIANT_REQUIRED; 389 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 390 KASSERT(object->type == OBJT_VNODE, 391 ("vm_object_vndeallocate: not a vnode object")); 392 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 393#ifdef INVARIANTS 394 if (object->ref_count == 0) { 395 vprint("vm_object_vndeallocate", vp); 396 panic("vm_object_vndeallocate: bad object reference count"); 397 } 398#endif 399 400 object->ref_count--; 401 if (object->ref_count == 0) { 402 mp_fixme("Unlocked vflag access."); 403 vp->v_vflag &= ~VV_TEXT; 404 } 405 VM_OBJECT_UNLOCK(object); 406 /* 407 * vrele may need a vop lock 408 */ 409 vrele(vp); 410} 411 412/* 413 * vm_object_deallocate: 414 * 415 * Release a reference to the specified object, 416 * gained either through a vm_object_allocate 417 * or a vm_object_reference call. When all references 418 * are gone, storage associated with this object 419 * may be relinquished. 420 * 421 * No object may be locked. 422 */ 423void 424vm_object_deallocate(vm_object_t object) 425{ 426 vm_object_t temp; 427 428 if (object != kmem_object) 429 mtx_lock(&Giant); 430 while (object != NULL) { 431 VM_OBJECT_LOCK(object); 432 if (object->type == OBJT_VNODE) { 433 vm_object_vndeallocate(object); 434 goto done; 435 } 436 437 KASSERT(object->ref_count != 0, 438 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 439 440 /* 441 * If the reference count goes to 0 we start calling 442 * vm_object_terminate() on the object chain. 443 * A ref count of 1 may be a special case depending on the 444 * shadow count being 0 or 1. 445 */ 446 object->ref_count--; 447 if (object->ref_count > 1) { 448 VM_OBJECT_UNLOCK(object); 449 goto done; 450 } else if (object->ref_count == 1) { 451 if (object->shadow_count == 0) { 452 vm_object_set_flag(object, OBJ_ONEMAPPING); 453 } else if ((object->shadow_count == 1) && 454 (object->handle == NULL) && 455 (object->type == OBJT_DEFAULT || 456 object->type == OBJT_SWAP)) { 457 vm_object_t robject; 458 459 robject = LIST_FIRST(&object->shadow_head); 460 KASSERT(robject != NULL, 461 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 462 object->ref_count, 463 object->shadow_count)); 464 if (!VM_OBJECT_TRYLOCK(robject)) { 465 /* 466 * Avoid a potential deadlock. 467 */ 468 object->ref_count++; 469 VM_OBJECT_UNLOCK(object); 470 continue; 471 } 472 if ((robject->handle == NULL) && 473 (robject->type == OBJT_DEFAULT || 474 robject->type == OBJT_SWAP)) { 475 476 robject->ref_count++; 477retry: 478 if (robject->paging_in_progress) { 479 VM_OBJECT_UNLOCK(object); 480 vm_object_pip_wait(robject, 481 "objde1"); 482 VM_OBJECT_LOCK(object); 483 goto retry; 484 } else if (object->paging_in_progress) { 485 VM_OBJECT_UNLOCK(robject); 486 object->flags |= OBJ_PIPWNT; 487 msleep(object, 488 VM_OBJECT_MTX(object), 489 PDROP | PVM, "objde2", 0); 490 VM_OBJECT_LOCK(robject); 491 VM_OBJECT_LOCK(object); 492 goto retry; 493 } 494 VM_OBJECT_UNLOCK(object); 495 if (robject->ref_count == 1) { 496 robject->ref_count--; 497 object = robject; 498 goto doterm; 499 } 500 object = robject; 501 vm_object_collapse(object); 502 VM_OBJECT_UNLOCK(object); 503 continue; 504 } 505 VM_OBJECT_UNLOCK(robject); 506 } 507 VM_OBJECT_UNLOCK(object); 508 goto done; 509 } 510doterm: 511 temp = object->backing_object; 512 if (temp != NULL) { 513 VM_OBJECT_LOCK(temp); 514 LIST_REMOVE(object, shadow_list); 515 temp->shadow_count--; 516 temp->generation++; 517 VM_OBJECT_UNLOCK(temp); 518 object->backing_object = NULL; 519 } 520 /* 521 * Don't double-terminate, we could be in a termination 522 * recursion due to the terminate having to sync data 523 * to disk. 524 */ 525 if ((object->flags & OBJ_DEAD) == 0) 526 vm_object_terminate(object); 527 else 528 VM_OBJECT_UNLOCK(object); 529 object = temp; 530 } 531done: 532 if (object != kmem_object) 533 mtx_unlock(&Giant); 534} 535 536/* 537 * vm_object_terminate actually destroys the specified object, freeing 538 * up all previously used resources. 539 * 540 * The object must be locked. 541 * This routine may block. 542 */ 543void 544vm_object_terminate(vm_object_t object) 545{ 546 vm_page_t p; 547 int s; 548 549 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 550 551 /* 552 * Make sure no one uses us. 553 */ 554 vm_object_set_flag(object, OBJ_DEAD); 555 556 /* 557 * wait for the pageout daemon to be done with the object 558 */ 559 vm_object_pip_wait(object, "objtrm"); 560 561 KASSERT(!object->paging_in_progress, 562 ("vm_object_terminate: pageout in progress")); 563 564 /* 565 * Clean and free the pages, as appropriate. All references to the 566 * object are gone, so we don't need to lock it. 567 */ 568 if (object->type == OBJT_VNODE) { 569 struct vnode *vp = (struct vnode *)object->handle; 570 571 /* 572 * Clean pages and flush buffers. 573 */ 574 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 575 VM_OBJECT_UNLOCK(object); 576 577 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 578 579 VM_OBJECT_LOCK(object); 580 } 581 582 KASSERT(object->ref_count == 0, 583 ("vm_object_terminate: object with references, ref_count=%d", 584 object->ref_count)); 585 586 /* 587 * Now free any remaining pages. For internal objects, this also 588 * removes them from paging queues. Don't free wired pages, just 589 * remove them from the object. 590 */ 591 s = splvm(); 592 vm_page_lock_queues(); 593 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 594 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 595 ("vm_object_terminate: freeing busy page %p " 596 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 597 if (p->wire_count == 0) { 598 vm_page_busy(p); 599 vm_page_free(p); 600 cnt.v_pfree++; 601 } else { 602 vm_page_busy(p); 603 vm_page_remove(p); 604 } 605 } 606 vm_page_unlock_queues(); 607 splx(s); 608 609 /* 610 * Let the pager know object is dead. 611 */ 612 vm_pager_deallocate(object); 613 VM_OBJECT_UNLOCK(object); 614 615 /* 616 * Remove the object from the global object list. 617 */ 618 mtx_lock(&vm_object_list_mtx); 619 TAILQ_REMOVE(&vm_object_list, object, object_list); 620 mtx_unlock(&vm_object_list_mtx); 621 622 wakeup(object); 623 624 /* 625 * Free the space for the object. 626 */ 627 uma_zfree(obj_zone, object); 628} 629 630/* 631 * vm_object_page_clean 632 * 633 * Clean all dirty pages in the specified range of object. Leaves page 634 * on whatever queue it is currently on. If NOSYNC is set then do not 635 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 636 * leaving the object dirty. 637 * 638 * When stuffing pages asynchronously, allow clustering. XXX we need a 639 * synchronous clustering mode implementation. 640 * 641 * Odd semantics: if start == end, we clean everything. 642 * 643 * The object must be locked. 644 */ 645void 646vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 647{ 648 vm_page_t p, np; 649 vm_pindex_t tstart, tend; 650 vm_pindex_t pi; 651 struct vnode *vp; 652 int clearobjflags; 653 int pagerflags; 654 int curgeneration; 655 656 GIANT_REQUIRED; 657 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 658 if (object->type != OBJT_VNODE || 659 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 660 return; 661 662 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 663 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 664 665 vp = object->handle; 666 667 vm_object_set_flag(object, OBJ_CLEANING); 668 669 tstart = start; 670 if (end == 0) { 671 tend = object->size; 672 } else { 673 tend = end; 674 } 675 676 vm_page_lock_queues(); 677 /* 678 * If the caller is smart and only msync()s a range he knows is 679 * dirty, we may be able to avoid an object scan. This results in 680 * a phenominal improvement in performance. We cannot do this 681 * as a matter of course because the object may be huge - e.g. 682 * the size might be in the gigabytes or terrabytes. 683 */ 684 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 685 vm_pindex_t tscan; 686 int scanlimit; 687 int scanreset; 688 689 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 690 if (scanreset < 16) 691 scanreset = 16; 692 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 693 694 scanlimit = scanreset; 695 tscan = tstart; 696 while (tscan < tend) { 697 curgeneration = object->generation; 698 p = vm_page_lookup(object, tscan); 699 if (p == NULL || p->valid == 0 || 700 (p->queue - p->pc) == PQ_CACHE) { 701 if (--scanlimit == 0) 702 break; 703 ++tscan; 704 continue; 705 } 706 vm_page_test_dirty(p); 707 if ((p->dirty & p->valid) == 0) { 708 if (--scanlimit == 0) 709 break; 710 ++tscan; 711 continue; 712 } 713 /* 714 * If we have been asked to skip nosync pages and 715 * this is a nosync page, we can't continue. 716 */ 717 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 718 if (--scanlimit == 0) 719 break; 720 ++tscan; 721 continue; 722 } 723 scanlimit = scanreset; 724 725 /* 726 * This returns 0 if it was unable to busy the first 727 * page (i.e. had to sleep). 728 */ 729 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 730 } 731 732 /* 733 * If everything was dirty and we flushed it successfully, 734 * and the requested range is not the entire object, we 735 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 736 * return immediately. 737 */ 738 if (tscan >= tend && (tstart || tend < object->size)) { 739 vm_page_unlock_queues(); 740 vm_object_clear_flag(object, OBJ_CLEANING); 741 return; 742 } 743 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 744 } 745 746 /* 747 * Generally set CLEANCHK interlock and make the page read-only so 748 * we can then clear the object flags. 749 * 750 * However, if this is a nosync mmap then the object is likely to 751 * stay dirty so do not mess with the page and do not clear the 752 * object flags. 753 */ 754 clearobjflags = 1; 755 TAILQ_FOREACH(p, &object->memq, listq) { 756 vm_page_flag_set(p, PG_CLEANCHK); 757 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 758 clearobjflags = 0; 759 else 760 pmap_page_protect(p, VM_PROT_READ); 761 } 762 763 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 764 struct vnode *vp; 765 766 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 767 if (object->type == OBJT_VNODE && 768 (vp = (struct vnode *)object->handle) != NULL) { 769 VI_LOCK(vp); 770 if (vp->v_iflag & VI_OBJDIRTY) 771 vp->v_iflag &= ~VI_OBJDIRTY; 772 VI_UNLOCK(vp); 773 } 774 } 775 776rescan: 777 curgeneration = object->generation; 778 779 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 780 int n; 781 782 np = TAILQ_NEXT(p, listq); 783 784again: 785 pi = p->pindex; 786 if (((p->flags & PG_CLEANCHK) == 0) || 787 (pi < tstart) || (pi >= tend) || 788 (p->valid == 0) || 789 ((p->queue - p->pc) == PQ_CACHE)) { 790 vm_page_flag_clear(p, PG_CLEANCHK); 791 continue; 792 } 793 794 vm_page_test_dirty(p); 795 if ((p->dirty & p->valid) == 0) { 796 vm_page_flag_clear(p, PG_CLEANCHK); 797 continue; 798 } 799 800 /* 801 * If we have been asked to skip nosync pages and this is a 802 * nosync page, skip it. Note that the object flags were 803 * not cleared in this case so we do not have to set them. 804 */ 805 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 806 vm_page_flag_clear(p, PG_CLEANCHK); 807 continue; 808 } 809 810 n = vm_object_page_collect_flush(object, p, 811 curgeneration, pagerflags); 812 if (n == 0) 813 goto rescan; 814 815 if (object->generation != curgeneration) 816 goto rescan; 817 818 /* 819 * Try to optimize the next page. If we can't we pick up 820 * our (random) scan where we left off. 821 */ 822 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 823 if ((p = vm_page_lookup(object, pi + n)) != NULL) 824 goto again; 825 } 826 } 827 vm_page_unlock_queues(); 828#if 0 829 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 830#endif 831 832 vm_object_clear_flag(object, OBJ_CLEANING); 833 return; 834} 835 836static int 837vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 838{ 839 int runlen; 840 int s; 841 int maxf; 842 int chkb; 843 int maxb; 844 int i; 845 vm_pindex_t pi; 846 vm_page_t maf[vm_pageout_page_count]; 847 vm_page_t mab[vm_pageout_page_count]; 848 vm_page_t ma[vm_pageout_page_count]; 849 850 s = splvm(); 851 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 852 pi = p->pindex; 853 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 854 vm_page_lock_queues(); 855 if (object->generation != curgeneration) { 856 splx(s); 857 return(0); 858 } 859 } 860 maxf = 0; 861 for(i = 1; i < vm_pageout_page_count; i++) { 862 vm_page_t tp; 863 864 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 865 if ((tp->flags & PG_BUSY) || 866 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 867 (tp->flags & PG_CLEANCHK) == 0) || 868 (tp->busy != 0)) 869 break; 870 if((tp->queue - tp->pc) == PQ_CACHE) { 871 vm_page_flag_clear(tp, PG_CLEANCHK); 872 break; 873 } 874 vm_page_test_dirty(tp); 875 if ((tp->dirty & tp->valid) == 0) { 876 vm_page_flag_clear(tp, PG_CLEANCHK); 877 break; 878 } 879 maf[ i - 1 ] = tp; 880 maxf++; 881 continue; 882 } 883 break; 884 } 885 886 maxb = 0; 887 chkb = vm_pageout_page_count - maxf; 888 if (chkb) { 889 for(i = 1; i < chkb;i++) { 890 vm_page_t tp; 891 892 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 893 if ((tp->flags & PG_BUSY) || 894 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 895 (tp->flags & PG_CLEANCHK) == 0) || 896 (tp->busy != 0)) 897 break; 898 if ((tp->queue - tp->pc) == PQ_CACHE) { 899 vm_page_flag_clear(tp, PG_CLEANCHK); 900 break; 901 } 902 vm_page_test_dirty(tp); 903 if ((tp->dirty & tp->valid) == 0) { 904 vm_page_flag_clear(tp, PG_CLEANCHK); 905 break; 906 } 907 mab[ i - 1 ] = tp; 908 maxb++; 909 continue; 910 } 911 break; 912 } 913 } 914 915 for(i = 0; i < maxb; i++) { 916 int index = (maxb - i) - 1; 917 ma[index] = mab[i]; 918 vm_page_flag_clear(ma[index], PG_CLEANCHK); 919 } 920 vm_page_flag_clear(p, PG_CLEANCHK); 921 ma[maxb] = p; 922 for(i = 0; i < maxf; i++) { 923 int index = (maxb + i) + 1; 924 ma[index] = maf[i]; 925 vm_page_flag_clear(ma[index], PG_CLEANCHK); 926 } 927 runlen = maxb + maxf + 1; 928 929 splx(s); 930 vm_pageout_flush(ma, runlen, pagerflags, TRUE); 931 for (i = 0; i < runlen; i++) { 932 if (ma[i]->valid & ma[i]->dirty) { 933 pmap_page_protect(ma[i], VM_PROT_READ); 934 vm_page_flag_set(ma[i], PG_CLEANCHK); 935 936 /* 937 * maxf will end up being the actual number of pages 938 * we wrote out contiguously, non-inclusive of the 939 * first page. We do not count look-behind pages. 940 */ 941 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 942 maxf = i - maxb - 1; 943 } 944 } 945 return(maxf + 1); 946} 947 948/* 949 * vm_object_madvise: 950 * 951 * Implements the madvise function at the object/page level. 952 * 953 * MADV_WILLNEED (any object) 954 * 955 * Activate the specified pages if they are resident. 956 * 957 * MADV_DONTNEED (any object) 958 * 959 * Deactivate the specified pages if they are resident. 960 * 961 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 962 * OBJ_ONEMAPPING only) 963 * 964 * Deactivate and clean the specified pages if they are 965 * resident. This permits the process to reuse the pages 966 * without faulting or the kernel to reclaim the pages 967 * without I/O. 968 */ 969void 970vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 971{ 972 vm_pindex_t end, tpindex; 973 vm_object_t backing_object, tobject; 974 vm_page_t m; 975 976 if (object == NULL) 977 return; 978 979 mtx_lock(&Giant); 980 981 end = pindex + count; 982 983 /* 984 * Locate and adjust resident pages 985 */ 986 for (; pindex < end; pindex += 1) { 987relookup: 988 tobject = object; 989 tpindex = pindex; 990 VM_OBJECT_LOCK(tobject); 991shadowlookup: 992 /* 993 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 994 * and those pages must be OBJ_ONEMAPPING. 995 */ 996 if (advise == MADV_FREE) { 997 if ((tobject->type != OBJT_DEFAULT && 998 tobject->type != OBJT_SWAP) || 999 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1000 goto unlock_tobject; 1001 } 1002 } 1003 1004 m = vm_page_lookup(tobject, tpindex); 1005 1006 if (m == NULL) { 1007 /* 1008 * There may be swap even if there is no backing page 1009 */ 1010 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1011 swap_pager_freespace(tobject, tpindex, 1); 1012 1013 /* 1014 * next object 1015 */ 1016 backing_object = tobject->backing_object; 1017 if (backing_object == NULL) 1018 goto unlock_tobject; 1019 VM_OBJECT_LOCK(backing_object); 1020 VM_OBJECT_UNLOCK(tobject); 1021 tobject = backing_object; 1022 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1023 goto shadowlookup; 1024 } 1025 1026 /* 1027 * If the page is busy or not in a normal active state, 1028 * we skip it. If the page is not managed there are no 1029 * page queues to mess with. Things can break if we mess 1030 * with pages in any of the below states. 1031 */ 1032 vm_page_lock_queues(); 1033 if (m->hold_count || 1034 m->wire_count || 1035 (m->flags & PG_UNMANAGED) || 1036 m->valid != VM_PAGE_BITS_ALL) { 1037 vm_page_unlock_queues(); 1038 goto unlock_tobject; 1039 } 1040 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) { 1041 VM_OBJECT_UNLOCK(tobject); 1042 goto relookup; 1043 } 1044 if (advise == MADV_WILLNEED) { 1045 vm_page_activate(m); 1046 } else if (advise == MADV_DONTNEED) { 1047 vm_page_dontneed(m); 1048 } else if (advise == MADV_FREE) { 1049 /* 1050 * Mark the page clean. This will allow the page 1051 * to be freed up by the system. However, such pages 1052 * are often reused quickly by malloc()/free() 1053 * so we do not do anything that would cause 1054 * a page fault if we can help it. 1055 * 1056 * Specifically, we do not try to actually free 1057 * the page now nor do we try to put it in the 1058 * cache (which would cause a page fault on reuse). 1059 * 1060 * But we do make the page is freeable as we 1061 * can without actually taking the step of unmapping 1062 * it. 1063 */ 1064 pmap_clear_modify(m); 1065 m->dirty = 0; 1066 m->act_count = 0; 1067 vm_page_dontneed(m); 1068 } 1069 vm_page_unlock_queues(); 1070 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1071 swap_pager_freespace(tobject, tpindex, 1); 1072unlock_tobject: 1073 VM_OBJECT_UNLOCK(tobject); 1074 } 1075 mtx_unlock(&Giant); 1076} 1077 1078/* 1079 * vm_object_shadow: 1080 * 1081 * Create a new object which is backed by the 1082 * specified existing object range. The source 1083 * object reference is deallocated. 1084 * 1085 * The new object and offset into that object 1086 * are returned in the source parameters. 1087 */ 1088void 1089vm_object_shadow( 1090 vm_object_t *object, /* IN/OUT */ 1091 vm_ooffset_t *offset, /* IN/OUT */ 1092 vm_size_t length) 1093{ 1094 vm_object_t source; 1095 vm_object_t result; 1096 1097 GIANT_REQUIRED; 1098 1099 source = *object; 1100 1101 /* 1102 * Don't create the new object if the old object isn't shared. 1103 */ 1104 if (source != NULL) { 1105 VM_OBJECT_LOCK(source); 1106 if (source->ref_count == 1 && 1107 source->handle == NULL && 1108 (source->type == OBJT_DEFAULT || 1109 source->type == OBJT_SWAP)) { 1110 VM_OBJECT_UNLOCK(source); 1111 return; 1112 } 1113 VM_OBJECT_UNLOCK(source); 1114 } 1115 1116 /* 1117 * Allocate a new object with the given length. 1118 */ 1119 result = vm_object_allocate(OBJT_DEFAULT, length); 1120 1121 /* 1122 * The new object shadows the source object, adding a reference to it. 1123 * Our caller changes his reference to point to the new object, 1124 * removing a reference to the source object. Net result: no change 1125 * of reference count. 1126 * 1127 * Try to optimize the result object's page color when shadowing 1128 * in order to maintain page coloring consistency in the combined 1129 * shadowed object. 1130 */ 1131 result->backing_object = source; 1132 if (source != NULL) { 1133 VM_OBJECT_LOCK(source); 1134 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1135 source->shadow_count++; 1136 source->generation++; 1137 if (length < source->size) 1138 length = source->size; 1139 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1140 source->generation > 1) 1141 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1142 result->pg_color = (source->pg_color + 1143 length * source->generation) & PQ_L2_MASK; 1144 VM_OBJECT_UNLOCK(source); 1145 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1146 PQ_L2_MASK; 1147 } 1148 1149 /* 1150 * Store the offset into the source object, and fix up the offset into 1151 * the new object. 1152 */ 1153 result->backing_object_offset = *offset; 1154 1155 /* 1156 * Return the new things 1157 */ 1158 *offset = 0; 1159 *object = result; 1160} 1161 1162/* 1163 * vm_object_split: 1164 * 1165 * Split the pages in a map entry into a new object. This affords 1166 * easier removal of unused pages, and keeps object inheritance from 1167 * being a negative impact on memory usage. 1168 */ 1169void 1170vm_object_split(vm_map_entry_t entry) 1171{ 1172 vm_page_t m; 1173 vm_object_t orig_object, new_object, source; 1174 vm_offset_t s, e; 1175 vm_pindex_t offidxstart, offidxend; 1176 vm_size_t idx, size; 1177 vm_ooffset_t offset; 1178 1179 GIANT_REQUIRED; 1180 1181 orig_object = entry->object.vm_object; 1182 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1183 return; 1184 if (orig_object->ref_count <= 1) 1185 return; 1186 1187 offset = entry->offset; 1188 s = entry->start; 1189 e = entry->end; 1190 1191 offidxstart = OFF_TO_IDX(offset); 1192 offidxend = offidxstart + OFF_TO_IDX(e - s); 1193 size = offidxend - offidxstart; 1194 1195 new_object = vm_pager_allocate(orig_object->type, 1196 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1197 if (new_object == NULL) 1198 return; 1199 1200 source = orig_object->backing_object; 1201 if (source != NULL) { 1202 vm_object_reference(source); /* Referenced by new_object */ 1203 VM_OBJECT_LOCK(source); 1204 LIST_INSERT_HEAD(&source->shadow_head, 1205 new_object, shadow_list); 1206 source->shadow_count++; 1207 source->generation++; 1208 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1209 VM_OBJECT_UNLOCK(source); 1210 new_object->backing_object_offset = 1211 orig_object->backing_object_offset + offset; 1212 new_object->backing_object = source; 1213 } 1214 VM_OBJECT_LOCK(orig_object); 1215 for (idx = 0; idx < size; idx++) { 1216 retry: 1217 m = vm_page_lookup(orig_object, offidxstart + idx); 1218 if (m == NULL) 1219 continue; 1220 1221 /* 1222 * We must wait for pending I/O to complete before we can 1223 * rename the page. 1224 * 1225 * We do not have to VM_PROT_NONE the page as mappings should 1226 * not be changed by this operation. 1227 */ 1228 vm_page_lock_queues(); 1229 if (vm_page_sleep_if_busy(m, TRUE, "spltwt")) 1230 goto retry; 1231 1232 vm_page_busy(m); 1233 vm_page_rename(m, new_object, idx); 1234 /* page automatically made dirty by rename and cache handled */ 1235 vm_page_busy(m); 1236 vm_page_unlock_queues(); 1237 } 1238 if (orig_object->type == OBJT_SWAP) { 1239 vm_object_pip_add(orig_object, 1); 1240 VM_OBJECT_UNLOCK(orig_object); 1241 /* 1242 * copy orig_object pages into new_object 1243 * and destroy unneeded pages in 1244 * shadow object. 1245 */ 1246 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1247 VM_OBJECT_LOCK(orig_object); 1248 vm_object_pip_wakeup(orig_object); 1249 } 1250 VM_OBJECT_UNLOCK(orig_object); 1251 vm_page_lock_queues(); 1252 TAILQ_FOREACH(m, &new_object->memq, listq) 1253 vm_page_wakeup(m); 1254 vm_page_unlock_queues(); 1255 entry->object.vm_object = new_object; 1256 entry->offset = 0LL; 1257 vm_object_deallocate(orig_object); 1258} 1259 1260#define OBSC_TEST_ALL_SHADOWED 0x0001 1261#define OBSC_COLLAPSE_NOWAIT 0x0002 1262#define OBSC_COLLAPSE_WAIT 0x0004 1263 1264static __inline int 1265vm_object_backing_scan(vm_object_t object, int op) 1266{ 1267 int s; 1268 int r = 1; 1269 vm_page_t p; 1270 vm_object_t backing_object; 1271 vm_pindex_t backing_offset_index; 1272 1273 s = splvm(); 1274 GIANT_REQUIRED; 1275 1276 backing_object = object->backing_object; 1277 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1278 1279 /* 1280 * Initial conditions 1281 */ 1282 if (op & OBSC_TEST_ALL_SHADOWED) { 1283 /* 1284 * We do not want to have to test for the existence of 1285 * swap pages in the backing object. XXX but with the 1286 * new swapper this would be pretty easy to do. 1287 * 1288 * XXX what about anonymous MAP_SHARED memory that hasn't 1289 * been ZFOD faulted yet? If we do not test for this, the 1290 * shadow test may succeed! XXX 1291 */ 1292 if (backing_object->type != OBJT_DEFAULT) { 1293 splx(s); 1294 return (0); 1295 } 1296 } 1297 if (op & OBSC_COLLAPSE_WAIT) { 1298 vm_object_set_flag(backing_object, OBJ_DEAD); 1299 } 1300 1301 /* 1302 * Our scan 1303 */ 1304 p = TAILQ_FIRST(&backing_object->memq); 1305 while (p) { 1306 vm_page_t next = TAILQ_NEXT(p, listq); 1307 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1308 1309 if (op & OBSC_TEST_ALL_SHADOWED) { 1310 vm_page_t pp; 1311 1312 /* 1313 * Ignore pages outside the parent object's range 1314 * and outside the parent object's mapping of the 1315 * backing object. 1316 * 1317 * note that we do not busy the backing object's 1318 * page. 1319 */ 1320 if ( 1321 p->pindex < backing_offset_index || 1322 new_pindex >= object->size 1323 ) { 1324 p = next; 1325 continue; 1326 } 1327 1328 /* 1329 * See if the parent has the page or if the parent's 1330 * object pager has the page. If the parent has the 1331 * page but the page is not valid, the parent's 1332 * object pager must have the page. 1333 * 1334 * If this fails, the parent does not completely shadow 1335 * the object and we might as well give up now. 1336 */ 1337 1338 pp = vm_page_lookup(object, new_pindex); 1339 if ( 1340 (pp == NULL || pp->valid == 0) && 1341 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1342 ) { 1343 r = 0; 1344 break; 1345 } 1346 } 1347 1348 /* 1349 * Check for busy page 1350 */ 1351 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1352 vm_page_t pp; 1353 1354 vm_page_lock_queues(); 1355 if (op & OBSC_COLLAPSE_NOWAIT) { 1356 if ((p->flags & PG_BUSY) || 1357 !p->valid || 1358 p->hold_count || 1359 p->wire_count || 1360 p->busy) { 1361 vm_page_unlock_queues(); 1362 p = next; 1363 continue; 1364 } 1365 } else if (op & OBSC_COLLAPSE_WAIT) { 1366 if (vm_page_sleep_if_busy(p, TRUE, "vmocol")) { 1367 /* 1368 * If we slept, anything could have 1369 * happened. Since the object is 1370 * marked dead, the backing offset 1371 * should not have changed so we 1372 * just restart our scan. 1373 */ 1374 p = TAILQ_FIRST(&backing_object->memq); 1375 continue; 1376 } 1377 } 1378 1379 /* 1380 * Busy the page 1381 */ 1382 vm_page_busy(p); 1383 vm_page_unlock_queues(); 1384 1385 KASSERT( 1386 p->object == backing_object, 1387 ("vm_object_qcollapse(): object mismatch") 1388 ); 1389 1390 /* 1391 * Destroy any associated swap 1392 */ 1393 if (backing_object->type == OBJT_SWAP) { 1394 swap_pager_freespace( 1395 backing_object, 1396 p->pindex, 1397 1 1398 ); 1399 } 1400 1401 if ( 1402 p->pindex < backing_offset_index || 1403 new_pindex >= object->size 1404 ) { 1405 /* 1406 * Page is out of the parent object's range, we 1407 * can simply destroy it. 1408 */ 1409 vm_page_lock_queues(); 1410 pmap_remove_all(p); 1411 vm_page_free(p); 1412 vm_page_unlock_queues(); 1413 p = next; 1414 continue; 1415 } 1416 1417 pp = vm_page_lookup(object, new_pindex); 1418 if ( 1419 pp != NULL || 1420 vm_pager_has_page(object, new_pindex, NULL, NULL) 1421 ) { 1422 /* 1423 * page already exists in parent OR swap exists 1424 * for this location in the parent. Destroy 1425 * the original page from the backing object. 1426 * 1427 * Leave the parent's page alone 1428 */ 1429 vm_page_lock_queues(); 1430 pmap_remove_all(p); 1431 vm_page_free(p); 1432 vm_page_unlock_queues(); 1433 p = next; 1434 continue; 1435 } 1436 1437 /* 1438 * Page does not exist in parent, rename the 1439 * page from the backing object to the main object. 1440 * 1441 * If the page was mapped to a process, it can remain 1442 * mapped through the rename. 1443 */ 1444 vm_page_lock_queues(); 1445 vm_page_rename(p, object, new_pindex); 1446 vm_page_unlock_queues(); 1447 /* page automatically made dirty by rename */ 1448 } 1449 p = next; 1450 } 1451 splx(s); 1452 return (r); 1453} 1454 1455 1456/* 1457 * this version of collapse allows the operation to occur earlier and 1458 * when paging_in_progress is true for an object... This is not a complete 1459 * operation, but should plug 99.9% of the rest of the leaks. 1460 */ 1461static void 1462vm_object_qcollapse(vm_object_t object) 1463{ 1464 vm_object_t backing_object = object->backing_object; 1465 1466 GIANT_REQUIRED; 1467 1468 if (backing_object->ref_count != 1) 1469 return; 1470 1471 backing_object->ref_count += 2; 1472 1473 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1474 1475 backing_object->ref_count -= 2; 1476} 1477 1478/* 1479 * vm_object_collapse: 1480 * 1481 * Collapse an object with the object backing it. 1482 * Pages in the backing object are moved into the 1483 * parent, and the backing object is deallocated. 1484 */ 1485void 1486vm_object_collapse(vm_object_t object) 1487{ 1488 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1489 1490 while (TRUE) { 1491 vm_object_t backing_object; 1492 1493 /* 1494 * Verify that the conditions are right for collapse: 1495 * 1496 * The object exists and the backing object exists. 1497 */ 1498 if ((backing_object = object->backing_object) == NULL) 1499 break; 1500 1501 /* 1502 * we check the backing object first, because it is most likely 1503 * not collapsable. 1504 */ 1505 VM_OBJECT_LOCK(backing_object); 1506 if (backing_object->handle != NULL || 1507 (backing_object->type != OBJT_DEFAULT && 1508 backing_object->type != OBJT_SWAP) || 1509 (backing_object->flags & OBJ_DEAD) || 1510 object->handle != NULL || 1511 (object->type != OBJT_DEFAULT && 1512 object->type != OBJT_SWAP) || 1513 (object->flags & OBJ_DEAD)) { 1514 VM_OBJECT_UNLOCK(backing_object); 1515 break; 1516 } 1517 1518 if ( 1519 object->paging_in_progress != 0 || 1520 backing_object->paging_in_progress != 0 1521 ) { 1522 vm_object_qcollapse(object); 1523 VM_OBJECT_UNLOCK(backing_object); 1524 break; 1525 } 1526/* XXX */ VM_OBJECT_UNLOCK(object); 1527 /* 1528 * We know that we can either collapse the backing object (if 1529 * the parent is the only reference to it) or (perhaps) have 1530 * the parent bypass the object if the parent happens to shadow 1531 * all the resident pages in the entire backing object. 1532 * 1533 * This is ignoring pager-backed pages such as swap pages. 1534 * vm_object_backing_scan fails the shadowing test in this 1535 * case. 1536 */ 1537 if (backing_object->ref_count == 1) { 1538 /* 1539 * If there is exactly one reference to the backing 1540 * object, we can collapse it into the parent. 1541 */ 1542 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1543 1544 /* 1545 * Move the pager from backing_object to object. 1546 */ 1547 if (backing_object->type == OBJT_SWAP) { 1548 vm_object_pip_add(backing_object, 1); 1549 VM_OBJECT_UNLOCK(backing_object); 1550 /* 1551 * scrap the paging_offset junk and do a 1552 * discrete copy. This also removes major 1553 * assumptions about how the swap-pager 1554 * works from where it doesn't belong. The 1555 * new swapper is able to optimize the 1556 * destroy-source case. 1557 */ 1558 VM_OBJECT_LOCK(object); 1559 vm_object_pip_add(object, 1); 1560 VM_OBJECT_UNLOCK(object); 1561 swap_pager_copy( 1562 backing_object, 1563 object, 1564 OFF_TO_IDX(object->backing_object_offset), TRUE); 1565 VM_OBJECT_LOCK(object); 1566 vm_object_pip_wakeup(object); 1567 VM_OBJECT_UNLOCK(object); 1568 1569 VM_OBJECT_LOCK(backing_object); 1570 vm_object_pip_wakeup(backing_object); 1571 } 1572 /* 1573 * Object now shadows whatever backing_object did. 1574 * Note that the reference to 1575 * backing_object->backing_object moves from within 1576 * backing_object to within object. 1577 */ 1578 LIST_REMOVE(object, shadow_list); 1579 backing_object->shadow_count--; 1580 backing_object->generation++; 1581 if (backing_object->backing_object) { 1582 VM_OBJECT_LOCK(backing_object->backing_object); 1583 LIST_REMOVE(backing_object, shadow_list); 1584 backing_object->backing_object->shadow_count--; 1585 backing_object->backing_object->generation++; 1586 VM_OBJECT_UNLOCK(backing_object->backing_object); 1587 } 1588 object->backing_object = backing_object->backing_object; 1589 if (object->backing_object) { 1590 VM_OBJECT_LOCK(object->backing_object); 1591 LIST_INSERT_HEAD( 1592 &object->backing_object->shadow_head, 1593 object, 1594 shadow_list 1595 ); 1596 object->backing_object->shadow_count++; 1597 object->backing_object->generation++; 1598 VM_OBJECT_UNLOCK(object->backing_object); 1599 } 1600 1601 object->backing_object_offset += 1602 backing_object->backing_object_offset; 1603 1604 /* 1605 * Discard backing_object. 1606 * 1607 * Since the backing object has no pages, no pager left, 1608 * and no object references within it, all that is 1609 * necessary is to dispose of it. 1610 */ 1611 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1612 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1613 VM_OBJECT_UNLOCK(backing_object); 1614 1615 mtx_lock(&vm_object_list_mtx); 1616 TAILQ_REMOVE( 1617 &vm_object_list, 1618 backing_object, 1619 object_list 1620 ); 1621 mtx_unlock(&vm_object_list_mtx); 1622 1623 uma_zfree(obj_zone, backing_object); 1624 1625 object_collapses++; 1626 } else { 1627 vm_object_t new_backing_object; 1628 1629 /* 1630 * If we do not entirely shadow the backing object, 1631 * there is nothing we can do so we give up. 1632 */ 1633 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1634 VM_OBJECT_UNLOCK(backing_object); 1635/* XXX */ VM_OBJECT_LOCK(object); 1636 break; 1637 } 1638 1639 /* 1640 * Make the parent shadow the next object in the 1641 * chain. Deallocating backing_object will not remove 1642 * it, since its reference count is at least 2. 1643 */ 1644 LIST_REMOVE(object, shadow_list); 1645 backing_object->shadow_count--; 1646 backing_object->generation++; 1647 VM_OBJECT_UNLOCK(backing_object); 1648 1649 new_backing_object = backing_object->backing_object; 1650 if ((object->backing_object = new_backing_object) != NULL) { 1651 vm_object_reference(new_backing_object); 1652 VM_OBJECT_LOCK(new_backing_object); 1653 LIST_INSERT_HEAD( 1654 &new_backing_object->shadow_head, 1655 object, 1656 shadow_list 1657 ); 1658 new_backing_object->shadow_count++; 1659 new_backing_object->generation++; 1660 VM_OBJECT_UNLOCK(new_backing_object); 1661 object->backing_object_offset += 1662 backing_object->backing_object_offset; 1663 } 1664 1665 /* 1666 * Drop the reference count on backing_object. Since 1667 * its ref_count was at least 2, it will not vanish; 1668 * so we don't need to call vm_object_deallocate, but 1669 * we do anyway. 1670 */ 1671 vm_object_deallocate(backing_object); 1672 object_bypasses++; 1673 } 1674 1675 /* 1676 * Try again with this object's new backing object. 1677 */ 1678/* XXX */ VM_OBJECT_LOCK(object); 1679 } 1680} 1681 1682/* 1683 * vm_object_page_remove: 1684 * 1685 * Removes all physical pages in the given range from the 1686 * object's list of pages. If the range's end is zero, all 1687 * physical pages from the range's start to the end of the object 1688 * are deleted. 1689 * 1690 * The object must be locked. 1691 */ 1692void 1693vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1694 boolean_t clean_only) 1695{ 1696 vm_page_t p, next; 1697 1698 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1699 if (object->resident_page_count == 0) 1700 return; 1701 1702 /* 1703 * Since physically-backed objects do not use managed pages, we can't 1704 * remove pages from the object (we must instead remove the page 1705 * references, and then destroy the object). 1706 */ 1707 KASSERT(object->type != OBJT_PHYS, 1708 ("attempt to remove pages from a physical object")); 1709 1710 vm_object_pip_add(object, 1); 1711again: 1712 vm_page_lock_queues(); 1713 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1714 if (p->pindex < start) { 1715 p = vm_page_splay(start, object->root); 1716 if ((object->root = p)->pindex < start) 1717 p = TAILQ_NEXT(p, listq); 1718 } 1719 } 1720 /* 1721 * Assert: the variable p is either (1) the page with the 1722 * least pindex greater than or equal to the parameter pindex 1723 * or (2) NULL. 1724 */ 1725 for (; 1726 p != NULL && (p->pindex < end || end == 0); 1727 p = next) { 1728 next = TAILQ_NEXT(p, listq); 1729 1730 if (p->wire_count != 0) { 1731 pmap_remove_all(p); 1732 if (!clean_only) 1733 p->valid = 0; 1734 continue; 1735 } 1736 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1737 goto again; 1738 if (clean_only && p->valid) { 1739 vm_page_test_dirty(p); 1740 if (p->valid & p->dirty) 1741 continue; 1742 } 1743 vm_page_busy(p); 1744 pmap_remove_all(p); 1745 vm_page_free(p); 1746 } 1747 vm_page_unlock_queues(); 1748 vm_object_pip_wakeup(object); 1749} 1750 1751/* 1752 * Routine: vm_object_coalesce 1753 * Function: Coalesces two objects backing up adjoining 1754 * regions of memory into a single object. 1755 * 1756 * returns TRUE if objects were combined. 1757 * 1758 * NOTE: Only works at the moment if the second object is NULL - 1759 * if it's not, which object do we lock first? 1760 * 1761 * Parameters: 1762 * prev_object First object to coalesce 1763 * prev_offset Offset into prev_object 1764 * next_object Second object into coalesce 1765 * next_offset Offset into next_object 1766 * 1767 * prev_size Size of reference to prev_object 1768 * next_size Size of reference to next_object 1769 * 1770 * Conditions: 1771 * The object must *not* be locked. 1772 */ 1773boolean_t 1774vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1775 vm_size_t prev_size, vm_size_t next_size) 1776{ 1777 vm_pindex_t next_pindex; 1778 1779 if (prev_object == NULL) 1780 return (TRUE); 1781 mtx_lock(&Giant); 1782 VM_OBJECT_LOCK(prev_object); 1783 if (prev_object->type != OBJT_DEFAULT && 1784 prev_object->type != OBJT_SWAP) { 1785 VM_OBJECT_UNLOCK(prev_object); 1786 mtx_unlock(&Giant); 1787 return (FALSE); 1788 } 1789 1790 /* 1791 * Try to collapse the object first 1792 */ 1793 vm_object_collapse(prev_object); 1794 1795 /* 1796 * Can't coalesce if: . more than one reference . paged out . shadows 1797 * another object . has a copy elsewhere (any of which mean that the 1798 * pages not mapped to prev_entry may be in use anyway) 1799 */ 1800 if (prev_object->backing_object != NULL) { 1801 VM_OBJECT_UNLOCK(prev_object); 1802 mtx_unlock(&Giant); 1803 return (FALSE); 1804 } 1805 1806 prev_size >>= PAGE_SHIFT; 1807 next_size >>= PAGE_SHIFT; 1808 next_pindex = prev_pindex + prev_size; 1809 1810 if ((prev_object->ref_count > 1) && 1811 (prev_object->size != next_pindex)) { 1812 VM_OBJECT_UNLOCK(prev_object); 1813 mtx_unlock(&Giant); 1814 return (FALSE); 1815 } 1816 1817 /* 1818 * Remove any pages that may still be in the object from a previous 1819 * deallocation. 1820 */ 1821 if (next_pindex < prev_object->size) { 1822 vm_object_page_remove(prev_object, 1823 next_pindex, 1824 next_pindex + next_size, FALSE); 1825 if (prev_object->type == OBJT_SWAP) 1826 swap_pager_freespace(prev_object, 1827 next_pindex, next_size); 1828 } 1829 1830 /* 1831 * Extend the object if necessary. 1832 */ 1833 if (next_pindex + next_size > prev_object->size) 1834 prev_object->size = next_pindex + next_size; 1835 1836 VM_OBJECT_UNLOCK(prev_object); 1837 mtx_unlock(&Giant); 1838 return (TRUE); 1839} 1840 1841void 1842vm_object_set_writeable_dirty(vm_object_t object) 1843{ 1844 struct vnode *vp; 1845 1846 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1847 if (object->type == OBJT_VNODE && 1848 (vp = (struct vnode *)object->handle) != NULL) { 1849 VI_LOCK(vp); 1850 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1851 vp->v_iflag |= VI_OBJDIRTY; 1852 VI_UNLOCK(vp); 1853 } 1854} 1855 1856#include "opt_ddb.h" 1857#ifdef DDB 1858#include <sys/kernel.h> 1859 1860#include <sys/cons.h> 1861 1862#include <ddb/ddb.h> 1863 1864static int 1865_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1866{ 1867 vm_map_t tmpm; 1868 vm_map_entry_t tmpe; 1869 vm_object_t obj; 1870 int entcount; 1871 1872 if (map == 0) 1873 return 0; 1874 1875 if (entry == 0) { 1876 tmpe = map->header.next; 1877 entcount = map->nentries; 1878 while (entcount-- && (tmpe != &map->header)) { 1879 if (_vm_object_in_map(map, object, tmpe)) { 1880 return 1; 1881 } 1882 tmpe = tmpe->next; 1883 } 1884 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1885 tmpm = entry->object.sub_map; 1886 tmpe = tmpm->header.next; 1887 entcount = tmpm->nentries; 1888 while (entcount-- && tmpe != &tmpm->header) { 1889 if (_vm_object_in_map(tmpm, object, tmpe)) { 1890 return 1; 1891 } 1892 tmpe = tmpe->next; 1893 } 1894 } else if ((obj = entry->object.vm_object) != NULL) { 1895 for (; obj; obj = obj->backing_object) 1896 if (obj == object) { 1897 return 1; 1898 } 1899 } 1900 return 0; 1901} 1902 1903static int 1904vm_object_in_map(vm_object_t object) 1905{ 1906 struct proc *p; 1907 1908 /* sx_slock(&allproc_lock); */ 1909 LIST_FOREACH(p, &allproc, p_list) { 1910 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1911 continue; 1912 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1913 /* sx_sunlock(&allproc_lock); */ 1914 return 1; 1915 } 1916 } 1917 /* sx_sunlock(&allproc_lock); */ 1918 if (_vm_object_in_map(kernel_map, object, 0)) 1919 return 1; 1920 if (_vm_object_in_map(kmem_map, object, 0)) 1921 return 1; 1922 if (_vm_object_in_map(pager_map, object, 0)) 1923 return 1; 1924 if (_vm_object_in_map(buffer_map, object, 0)) 1925 return 1; 1926 return 0; 1927} 1928 1929DB_SHOW_COMMAND(vmochk, vm_object_check) 1930{ 1931 vm_object_t object; 1932 1933 /* 1934 * make sure that internal objs are in a map somewhere 1935 * and none have zero ref counts. 1936 */ 1937 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1938 if (object->handle == NULL && 1939 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1940 if (object->ref_count == 0) { 1941 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1942 (long)object->size); 1943 } 1944 if (!vm_object_in_map(object)) { 1945 db_printf( 1946 "vmochk: internal obj is not in a map: " 1947 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1948 object->ref_count, (u_long)object->size, 1949 (u_long)object->size, 1950 (void *)object->backing_object); 1951 } 1952 } 1953 } 1954} 1955 1956/* 1957 * vm_object_print: [ debug ] 1958 */ 1959DB_SHOW_COMMAND(object, vm_object_print_static) 1960{ 1961 /* XXX convert args. */ 1962 vm_object_t object = (vm_object_t)addr; 1963 boolean_t full = have_addr; 1964 1965 vm_page_t p; 1966 1967 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1968#define count was_count 1969 1970 int count; 1971 1972 if (object == NULL) 1973 return; 1974 1975 db_iprintf( 1976 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 1977 object, (int)object->type, (uintmax_t)object->size, 1978 object->resident_page_count, object->ref_count, object->flags); 1979 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 1980 object->shadow_count, 1981 object->backing_object ? object->backing_object->ref_count : 0, 1982 object->backing_object, (uintmax_t)object->backing_object_offset); 1983 1984 if (!full) 1985 return; 1986 1987 db_indent += 2; 1988 count = 0; 1989 TAILQ_FOREACH(p, &object->memq, listq) { 1990 if (count == 0) 1991 db_iprintf("memory:="); 1992 else if (count == 6) { 1993 db_printf("\n"); 1994 db_iprintf(" ..."); 1995 count = 0; 1996 } else 1997 db_printf(","); 1998 count++; 1999 2000 db_printf("(off=0x%jx,page=0x%jx)", 2001 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2002 } 2003 if (count != 0) 2004 db_printf("\n"); 2005 db_indent -= 2; 2006} 2007 2008/* XXX. */ 2009#undef count 2010 2011/* XXX need this non-static entry for calling from vm_map_print. */ 2012void 2013vm_object_print( 2014 /* db_expr_t */ long addr, 2015 boolean_t have_addr, 2016 /* db_expr_t */ long count, 2017 char *modif) 2018{ 2019 vm_object_print_static(addr, have_addr, count, modif); 2020} 2021 2022DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2023{ 2024 vm_object_t object; 2025 int nl = 0; 2026 int c; 2027 2028 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2029 vm_pindex_t idx, fidx; 2030 vm_pindex_t osize; 2031 vm_paddr_t pa = -1, padiff; 2032 int rcount; 2033 vm_page_t m; 2034 2035 db_printf("new object: %p\n", (void *)object); 2036 if (nl > 18) { 2037 c = cngetc(); 2038 if (c != ' ') 2039 return; 2040 nl = 0; 2041 } 2042 nl++; 2043 rcount = 0; 2044 fidx = 0; 2045 osize = object->size; 2046 if (osize > 128) 2047 osize = 128; 2048 for (idx = 0; idx < osize; idx++) { 2049 m = vm_page_lookup(object, idx); 2050 if (m == NULL) { 2051 if (rcount) { 2052 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2053 (long)fidx, rcount, (long)pa); 2054 if (nl > 18) { 2055 c = cngetc(); 2056 if (c != ' ') 2057 return; 2058 nl = 0; 2059 } 2060 nl++; 2061 rcount = 0; 2062 } 2063 continue; 2064 } 2065 2066 2067 if (rcount && 2068 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2069 ++rcount; 2070 continue; 2071 } 2072 if (rcount) { 2073 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2074 padiff >>= PAGE_SHIFT; 2075 padiff &= PQ_L2_MASK; 2076 if (padiff == 0) { 2077 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2078 ++rcount; 2079 continue; 2080 } 2081 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2082 (long)fidx, rcount, (long)pa); 2083 db_printf("pd(%ld)\n", (long)padiff); 2084 if (nl > 18) { 2085 c = cngetc(); 2086 if (c != ' ') 2087 return; 2088 nl = 0; 2089 } 2090 nl++; 2091 } 2092 fidx = idx; 2093 pa = VM_PAGE_TO_PHYS(m); 2094 rcount = 1; 2095 } 2096 if (rcount) { 2097 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2098 (long)fidx, rcount, (long)pa); 2099 if (nl > 18) { 2100 c = cngetc(); 2101 if (c != ' ') 2102 return; 2103 nl = 0; 2104 } 2105 nl++; 2106 } 2107 } 2108} 2109#endif /* DDB */ 2110