vm_object.c revision 132636
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 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61/* 62 * Virtual memory object module. 63 */ 64 65#include <sys/cdefs.h> 66__FBSDID("$FreeBSD: head/sys/vm/vm_object.c 132636 2004-07-25 19:28:10Z alc $"); 67 68#include <sys/param.h> 69#include <sys/systm.h> 70#include <sys/lock.h> 71#include <sys/mman.h> 72#include <sys/mount.h> 73#include <sys/kernel.h> 74#include <sys/sysctl.h> 75#include <sys/mutex.h> 76#include <sys/proc.h> /* for curproc, pageproc */ 77#include <sys/socket.h> 78#include <sys/vnode.h> 79#include <sys/vmmeter.h> 80#include <sys/sx.h> 81 82#include <vm/vm.h> 83#include <vm/vm_param.h> 84#include <vm/pmap.h> 85#include <vm/vm_map.h> 86#include <vm/vm_object.h> 87#include <vm/vm_page.h> 88#include <vm/vm_pageout.h> 89#include <vm/vm_pager.h> 90#include <vm/swap_pager.h> 91#include <vm/vm_kern.h> 92#include <vm/vm_extern.h> 93#include <vm/uma.h> 94 95#define EASY_SCAN_FACTOR 8 96 97#define MSYNC_FLUSH_HARDSEQ 0x01 98#define MSYNC_FLUSH_SOFTSEQ 0x02 99 100/* 101 * msync / VM object flushing optimizations 102 */ 103static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 104SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 105 CTLFLAG_RW, &msync_flush_flags, 0, ""); 106 107static int old_msync; 108SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 109 "Use old (insecure) msync behavior"); 110 111static void vm_object_qcollapse(vm_object_t object); 112static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 113 114/* 115 * Virtual memory objects maintain the actual data 116 * associated with allocated virtual memory. A given 117 * page of memory exists within exactly one object. 118 * 119 * An object is only deallocated when all "references" 120 * are given up. Only one "reference" to a given 121 * region of an object should be writeable. 122 * 123 * Associated with each object is a list of all resident 124 * memory pages belonging to that object; this list is 125 * maintained by the "vm_page" module, and locked by the object's 126 * lock. 127 * 128 * Each object also records a "pager" routine which is 129 * used to retrieve (and store) pages to the proper backing 130 * storage. In addition, objects may be backed by other 131 * objects from which they were virtual-copied. 132 * 133 * The only items within the object structure which are 134 * modified after time of creation are: 135 * reference count locked by object's lock 136 * pager routine locked by object's lock 137 * 138 */ 139 140struct object_q vm_object_list; 141struct mtx vm_object_list_mtx; /* lock for object list and count */ 142 143struct vm_object kernel_object_store; 144struct vm_object kmem_object_store; 145 146static long object_collapses; 147static long object_bypasses; 148static int next_index; 149static uma_zone_t obj_zone; 150#define VM_OBJECTS_INIT 256 151 152static void vm_object_zinit(void *mem, int size); 153 154#ifdef INVARIANTS 155static void vm_object_zdtor(void *mem, int size, void *arg); 156 157static void 158vm_object_zdtor(void *mem, int size, void *arg) 159{ 160 vm_object_t object; 161 162 object = (vm_object_t)mem; 163 KASSERT(TAILQ_EMPTY(&object->memq), 164 ("object %p has resident pages", 165 object)); 166 KASSERT(object->paging_in_progress == 0, 167 ("object %p paging_in_progress = %d", 168 object, object->paging_in_progress)); 169 KASSERT(object->resident_page_count == 0, 170 ("object %p resident_page_count = %d", 171 object, object->resident_page_count)); 172 KASSERT(object->shadow_count == 0, 173 ("object %p shadow_count = %d", 174 object, object->shadow_count)); 175} 176#endif 177 178static void 179vm_object_zinit(void *mem, int size) 180{ 181 vm_object_t object; 182 183 object = (vm_object_t)mem; 184 bzero(&object->mtx, sizeof(object->mtx)); 185 VM_OBJECT_LOCK_INIT(object, "standard object"); 186 187 /* These are true for any object that has been freed */ 188 object->paging_in_progress = 0; 189 object->resident_page_count = 0; 190 object->shadow_count = 0; 191} 192 193void 194_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 195{ 196 int incr; 197 198 TAILQ_INIT(&object->memq); 199 LIST_INIT(&object->shadow_head); 200 201 object->root = NULL; 202 object->type = type; 203 object->size = size; 204 object->generation = 1; 205 object->ref_count = 1; 206 object->flags = 0; 207 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 208 object->flags = OBJ_ONEMAPPING; 209 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 210 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 211 else 212 incr = size; 213 do 214 object->pg_color = next_index; 215 while (!atomic_cmpset_int(&next_index, object->pg_color, 216 (object->pg_color + incr) & PQ_L2_MASK)); 217 object->handle = NULL; 218 object->backing_object = NULL; 219 object->backing_object_offset = (vm_ooffset_t) 0; 220 221 mtx_lock(&vm_object_list_mtx); 222 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 223 mtx_unlock(&vm_object_list_mtx); 224} 225 226/* 227 * vm_object_init: 228 * 229 * Initialize the VM objects module. 230 */ 231void 232vm_object_init(void) 233{ 234 TAILQ_INIT(&vm_object_list); 235 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 236 237 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object"); 238 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 239 kernel_object); 240 241 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object"); 242 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 243 kmem_object); 244 245 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 246#ifdef INVARIANTS 247 vm_object_zdtor, 248#else 249 NULL, 250#endif 251 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); 252 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 253} 254 255void 256vm_object_clear_flag(vm_object_t object, u_short bits) 257{ 258 259 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 260 object->flags &= ~bits; 261} 262 263void 264vm_object_pip_add(vm_object_t object, short i) 265{ 266 267 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 268 object->paging_in_progress += i; 269} 270 271void 272vm_object_pip_subtract(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_wakeup(vm_object_t object) 281{ 282 283 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 284 object->paging_in_progress--; 285 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 286 vm_object_clear_flag(object, OBJ_PIPWNT); 287 wakeup(object); 288 } 289} 290 291void 292vm_object_pip_wakeupn(vm_object_t object, short i) 293{ 294 295 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 296 if (i) 297 object->paging_in_progress -= i; 298 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 299 vm_object_clear_flag(object, OBJ_PIPWNT); 300 wakeup(object); 301 } 302} 303 304void 305vm_object_pip_wait(vm_object_t object, char *waitid) 306{ 307 308 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 309 while (object->paging_in_progress) { 310 object->flags |= OBJ_PIPWNT; 311 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 312 } 313} 314 315/* 316 * vm_object_allocate_wait 317 * 318 * Return a new object with the given size, and give the user the 319 * option of waiting for it to complete or failing if the needed 320 * memory isn't available. 321 */ 322vm_object_t 323vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 324{ 325 vm_object_t result; 326 327 result = (vm_object_t) uma_zalloc(obj_zone, flags); 328 329 if (result != NULL) 330 _vm_object_allocate(type, size, result); 331 332 return (result); 333} 334 335/* 336 * vm_object_allocate: 337 * 338 * Returns a new object with the given size. 339 */ 340vm_object_t 341vm_object_allocate(objtype_t type, vm_pindex_t size) 342{ 343 return(vm_object_allocate_wait(type, size, M_WAITOK)); 344} 345 346 347/* 348 * vm_object_reference: 349 * 350 * Gets another reference to the given object. Note: OBJ_DEAD 351 * objects can be referenced during final cleaning. 352 */ 353void 354vm_object_reference(vm_object_t object) 355{ 356 struct vnode *vp; 357 int flags; 358 359 if (object == NULL) 360 return; 361 VM_OBJECT_LOCK(object); 362 object->ref_count++; 363 if (object->type == OBJT_VNODE) { 364 vp = object->handle; 365 VI_LOCK(vp); 366 VM_OBJECT_UNLOCK(object); 367 for (flags = LK_INTERLOCK; vget(vp, flags, curthread); 368 flags = 0) 369 printf("vm_object_reference: delay in vget\n"); 370 } else 371 VM_OBJECT_UNLOCK(object); 372} 373 374/* 375 * vm_object_reference_locked: 376 * 377 * Gets another reference to the given object. 378 * 379 * The object must be locked. 380 */ 381void 382vm_object_reference_locked(vm_object_t object) 383{ 384 struct vnode *vp; 385 386 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 387 KASSERT((object->flags & OBJ_DEAD) == 0, 388 ("vm_object_reference_locked: dead object referenced")); 389 object->ref_count++; 390 if (object->type == OBJT_VNODE) { 391 vp = object->handle; 392 vref(vp); 393 } 394} 395 396/* 397 * Handle deallocating an object of type OBJT_VNODE. 398 */ 399void 400vm_object_vndeallocate(vm_object_t object) 401{ 402 struct vnode *vp = (struct vnode *) object->handle; 403 404 GIANT_REQUIRED; 405 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 406 KASSERT(object->type == OBJT_VNODE, 407 ("vm_object_vndeallocate: not a vnode object")); 408 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 409#ifdef INVARIANTS 410 if (object->ref_count == 0) { 411 vprint("vm_object_vndeallocate", vp); 412 panic("vm_object_vndeallocate: bad object reference count"); 413 } 414#endif 415 416 object->ref_count--; 417 if (object->ref_count == 0) { 418 mp_fixme("Unlocked vflag access."); 419 vp->v_vflag &= ~VV_TEXT; 420 } 421 VM_OBJECT_UNLOCK(object); 422 /* 423 * vrele may need a vop lock 424 */ 425 vrele(vp); 426} 427 428/* 429 * vm_object_deallocate: 430 * 431 * Release a reference to the specified object, 432 * gained either through a vm_object_allocate 433 * or a vm_object_reference call. When all references 434 * are gone, storage associated with this object 435 * may be relinquished. 436 * 437 * No object may be locked. 438 */ 439void 440vm_object_deallocate(vm_object_t object) 441{ 442 vm_object_t temp; 443 444 while (object != NULL) { 445 /* 446 * In general, the object should be locked when working with 447 * its type. In this case, in order to maintain proper lock 448 * ordering, an exception is possible because a vnode-backed 449 * object never changes its type. 450 */ 451 if (object->type == OBJT_VNODE) 452 mtx_lock(&Giant); 453 VM_OBJECT_LOCK(object); 454 if (object->type == OBJT_VNODE) { 455 vm_object_vndeallocate(object); 456 mtx_unlock(&Giant); 457 return; 458 } 459 460 KASSERT(object->ref_count != 0, 461 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 462 463 /* 464 * If the reference count goes to 0 we start calling 465 * vm_object_terminate() on the object chain. 466 * A ref count of 1 may be a special case depending on the 467 * shadow count being 0 or 1. 468 */ 469 object->ref_count--; 470 if (object->ref_count > 1) { 471 VM_OBJECT_UNLOCK(object); 472 return; 473 } else if (object->ref_count == 1) { 474 if (object->shadow_count == 0) { 475 vm_object_set_flag(object, OBJ_ONEMAPPING); 476 } else if ((object->shadow_count == 1) && 477 (object->handle == NULL) && 478 (object->type == OBJT_DEFAULT || 479 object->type == OBJT_SWAP)) { 480 vm_object_t robject; 481 482 robject = LIST_FIRST(&object->shadow_head); 483 KASSERT(robject != NULL, 484 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 485 object->ref_count, 486 object->shadow_count)); 487 if (!VM_OBJECT_TRYLOCK(robject)) { 488 /* 489 * Avoid a potential deadlock. 490 */ 491 object->ref_count++; 492 VM_OBJECT_UNLOCK(object); 493 /* 494 * More likely than not the thread 495 * holding robject's lock has lower 496 * priority than the current thread. 497 * Let the lower priority thread run. 498 */ 499 tsleep(&proc0, PVM, "vmo_de", 1); 500 continue; 501 } 502 if ((robject->handle == NULL) && 503 (robject->type == OBJT_DEFAULT || 504 robject->type == OBJT_SWAP)) { 505 506 robject->ref_count++; 507retry: 508 if (robject->paging_in_progress) { 509 VM_OBJECT_UNLOCK(object); 510 vm_object_pip_wait(robject, 511 "objde1"); 512 VM_OBJECT_LOCK(object); 513 goto retry; 514 } else if (object->paging_in_progress) { 515 VM_OBJECT_UNLOCK(robject); 516 object->flags |= OBJ_PIPWNT; 517 msleep(object, 518 VM_OBJECT_MTX(object), 519 PDROP | PVM, "objde2", 0); 520 VM_OBJECT_LOCK(robject); 521 VM_OBJECT_LOCK(object); 522 goto retry; 523 } 524 VM_OBJECT_UNLOCK(object); 525 if (robject->ref_count == 1) { 526 robject->ref_count--; 527 object = robject; 528 goto doterm; 529 } 530 object = robject; 531 vm_object_collapse(object); 532 VM_OBJECT_UNLOCK(object); 533 continue; 534 } 535 VM_OBJECT_UNLOCK(robject); 536 } 537 VM_OBJECT_UNLOCK(object); 538 return; 539 } 540doterm: 541 temp = object->backing_object; 542 if (temp != NULL) { 543 VM_OBJECT_LOCK(temp); 544 LIST_REMOVE(object, shadow_list); 545 temp->shadow_count--; 546 temp->generation++; 547 VM_OBJECT_UNLOCK(temp); 548 object->backing_object = NULL; 549 } 550 /* 551 * Don't double-terminate, we could be in a termination 552 * recursion due to the terminate having to sync data 553 * to disk. 554 */ 555 if ((object->flags & OBJ_DEAD) == 0) 556 vm_object_terminate(object); 557 else 558 VM_OBJECT_UNLOCK(object); 559 object = temp; 560 } 561} 562 563/* 564 * vm_object_terminate actually destroys the specified object, freeing 565 * up all previously used resources. 566 * 567 * The object must be locked. 568 * This routine may block. 569 */ 570void 571vm_object_terminate(vm_object_t object) 572{ 573 vm_page_t p; 574 575 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 576 577 /* 578 * Make sure no one uses us. 579 */ 580 vm_object_set_flag(object, OBJ_DEAD); 581 582 /* 583 * wait for the pageout daemon to be done with the object 584 */ 585 vm_object_pip_wait(object, "objtrm"); 586 587 KASSERT(!object->paging_in_progress, 588 ("vm_object_terminate: pageout in progress")); 589 590 /* 591 * Clean and free the pages, as appropriate. All references to the 592 * object are gone, so we don't need to lock it. 593 */ 594 if (object->type == OBJT_VNODE) { 595 struct vnode *vp = (struct vnode *)object->handle; 596 597 /* 598 * Clean pages and flush buffers. 599 */ 600 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 601 VM_OBJECT_UNLOCK(object); 602 603 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 604 605 VM_OBJECT_LOCK(object); 606 } 607 608 KASSERT(object->ref_count == 0, 609 ("vm_object_terminate: object with references, ref_count=%d", 610 object->ref_count)); 611 612 /* 613 * Now free any remaining pages. For internal objects, this also 614 * removes them from paging queues. Don't free wired pages, just 615 * remove them from the object. 616 */ 617 vm_page_lock_queues(); 618 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 619 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 620 ("vm_object_terminate: freeing busy page %p " 621 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 622 if (p->wire_count == 0) { 623 vm_page_busy(p); 624 vm_page_free(p); 625 cnt.v_pfree++; 626 } else { 627 vm_page_busy(p); 628 vm_page_remove(p); 629 } 630 } 631 vm_page_unlock_queues(); 632 633 /* 634 * Let the pager know object is dead. 635 */ 636 vm_pager_deallocate(object); 637 VM_OBJECT_UNLOCK(object); 638 639 /* 640 * Remove the object from the global object list. 641 */ 642 mtx_lock(&vm_object_list_mtx); 643 TAILQ_REMOVE(&vm_object_list, object, object_list); 644 mtx_unlock(&vm_object_list_mtx); 645 646 wakeup(object); 647 648 /* 649 * Free the space for the object. 650 */ 651 uma_zfree(obj_zone, object); 652} 653 654/* 655 * vm_object_page_clean 656 * 657 * Clean all dirty pages in the specified range of object. Leaves page 658 * on whatever queue it is currently on. If NOSYNC is set then do not 659 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 660 * leaving the object dirty. 661 * 662 * When stuffing pages asynchronously, allow clustering. XXX we need a 663 * synchronous clustering mode implementation. 664 * 665 * Odd semantics: if start == end, we clean everything. 666 * 667 * The object must be locked. 668 */ 669void 670vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 671{ 672 vm_page_t p, np; 673 vm_pindex_t tstart, tend; 674 vm_pindex_t pi; 675 int clearobjflags; 676 int pagerflags; 677 int curgeneration; 678 679 GIANT_REQUIRED; 680 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 681 if (object->type != OBJT_VNODE || 682 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 683 return; 684 685 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 686 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 687 688 vm_object_set_flag(object, OBJ_CLEANING); 689 690 tstart = start; 691 if (end == 0) { 692 tend = object->size; 693 } else { 694 tend = end; 695 } 696 697 vm_page_lock_queues(); 698 /* 699 * If the caller is smart and only msync()s a range he knows is 700 * dirty, we may be able to avoid an object scan. This results in 701 * a phenominal improvement in performance. We cannot do this 702 * as a matter of course because the object may be huge - e.g. 703 * the size might be in the gigabytes or terrabytes. 704 */ 705 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 706 vm_pindex_t tscan; 707 int scanlimit; 708 int scanreset; 709 710 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 711 if (scanreset < 16) 712 scanreset = 16; 713 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 714 715 scanlimit = scanreset; 716 tscan = tstart; 717 while (tscan < tend) { 718 curgeneration = object->generation; 719 p = vm_page_lookup(object, tscan); 720 if (p == NULL || p->valid == 0 || 721 (p->queue - p->pc) == PQ_CACHE) { 722 if (--scanlimit == 0) 723 break; 724 ++tscan; 725 continue; 726 } 727 vm_page_test_dirty(p); 728 if ((p->dirty & p->valid) == 0) { 729 if (--scanlimit == 0) 730 break; 731 ++tscan; 732 continue; 733 } 734 /* 735 * If we have been asked to skip nosync pages and 736 * this is a nosync page, we can't continue. 737 */ 738 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 739 if (--scanlimit == 0) 740 break; 741 ++tscan; 742 continue; 743 } 744 scanlimit = scanreset; 745 746 /* 747 * This returns 0 if it was unable to busy the first 748 * page (i.e. had to sleep). 749 */ 750 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 751 } 752 753 /* 754 * If everything was dirty and we flushed it successfully, 755 * and the requested range is not the entire object, we 756 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 757 * return immediately. 758 */ 759 if (tscan >= tend && (tstart || tend < object->size)) { 760 vm_page_unlock_queues(); 761 vm_object_clear_flag(object, OBJ_CLEANING); 762 return; 763 } 764 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 765 } 766 767 /* 768 * Generally set CLEANCHK interlock and make the page read-only so 769 * we can then clear the object flags. 770 * 771 * However, if this is a nosync mmap then the object is likely to 772 * stay dirty so do not mess with the page and do not clear the 773 * object flags. 774 */ 775 clearobjflags = 1; 776 TAILQ_FOREACH(p, &object->memq, listq) { 777 vm_page_flag_set(p, PG_CLEANCHK); 778 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 779 clearobjflags = 0; 780 else 781 pmap_page_protect(p, VM_PROT_READ); 782 } 783 784 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 785 struct vnode *vp; 786 787 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 788 if (object->type == OBJT_VNODE && 789 (vp = (struct vnode *)object->handle) != NULL) { 790 VI_LOCK(vp); 791 if (vp->v_iflag & VI_OBJDIRTY) 792 vp->v_iflag &= ~VI_OBJDIRTY; 793 VI_UNLOCK(vp); 794 } 795 } 796 797rescan: 798 curgeneration = object->generation; 799 800 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 801 int n; 802 803 np = TAILQ_NEXT(p, listq); 804 805again: 806 pi = p->pindex; 807 if (((p->flags & PG_CLEANCHK) == 0) || 808 (pi < tstart) || (pi >= tend) || 809 (p->valid == 0) || 810 ((p->queue - p->pc) == PQ_CACHE)) { 811 vm_page_flag_clear(p, PG_CLEANCHK); 812 continue; 813 } 814 815 vm_page_test_dirty(p); 816 if ((p->dirty & p->valid) == 0) { 817 vm_page_flag_clear(p, PG_CLEANCHK); 818 continue; 819 } 820 821 /* 822 * If we have been asked to skip nosync pages and this is a 823 * nosync page, skip it. Note that the object flags were 824 * not cleared in this case so we do not have to set them. 825 */ 826 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 827 vm_page_flag_clear(p, PG_CLEANCHK); 828 continue; 829 } 830 831 n = vm_object_page_collect_flush(object, p, 832 curgeneration, pagerflags); 833 if (n == 0) 834 goto rescan; 835 836 if (object->generation != curgeneration) 837 goto rescan; 838 839 /* 840 * Try to optimize the next page. If we can't we pick up 841 * our (random) scan where we left off. 842 */ 843 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 844 if ((p = vm_page_lookup(object, pi + n)) != NULL) 845 goto again; 846 } 847 } 848 vm_page_unlock_queues(); 849#if 0 850 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 851#endif 852 853 vm_object_clear_flag(object, OBJ_CLEANING); 854 return; 855} 856 857static int 858vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 859{ 860 int runlen; 861 int maxf; 862 int chkb; 863 int maxb; 864 int i; 865 vm_pindex_t pi; 866 vm_page_t maf[vm_pageout_page_count]; 867 vm_page_t mab[vm_pageout_page_count]; 868 vm_page_t ma[vm_pageout_page_count]; 869 870 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 871 pi = p->pindex; 872 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 873 vm_page_lock_queues(); 874 if (object->generation != curgeneration) { 875 return(0); 876 } 877 } 878 maxf = 0; 879 for(i = 1; i < vm_pageout_page_count; i++) { 880 vm_page_t tp; 881 882 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 883 if ((tp->flags & PG_BUSY) || 884 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 885 (tp->flags & PG_CLEANCHK) == 0) || 886 (tp->busy != 0)) 887 break; 888 if((tp->queue - tp->pc) == PQ_CACHE) { 889 vm_page_flag_clear(tp, PG_CLEANCHK); 890 break; 891 } 892 vm_page_test_dirty(tp); 893 if ((tp->dirty & tp->valid) == 0) { 894 vm_page_flag_clear(tp, PG_CLEANCHK); 895 break; 896 } 897 maf[ i - 1 ] = tp; 898 maxf++; 899 continue; 900 } 901 break; 902 } 903 904 maxb = 0; 905 chkb = vm_pageout_page_count - maxf; 906 if (chkb) { 907 for(i = 1; i < chkb;i++) { 908 vm_page_t tp; 909 910 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 911 if ((tp->flags & PG_BUSY) || 912 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 913 (tp->flags & PG_CLEANCHK) == 0) || 914 (tp->busy != 0)) 915 break; 916 if ((tp->queue - tp->pc) == PQ_CACHE) { 917 vm_page_flag_clear(tp, PG_CLEANCHK); 918 break; 919 } 920 vm_page_test_dirty(tp); 921 if ((tp->dirty & tp->valid) == 0) { 922 vm_page_flag_clear(tp, PG_CLEANCHK); 923 break; 924 } 925 mab[ i - 1 ] = tp; 926 maxb++; 927 continue; 928 } 929 break; 930 } 931 } 932 933 for(i = 0; i < maxb; i++) { 934 int index = (maxb - i) - 1; 935 ma[index] = mab[i]; 936 vm_page_flag_clear(ma[index], PG_CLEANCHK); 937 } 938 vm_page_flag_clear(p, PG_CLEANCHK); 939 ma[maxb] = p; 940 for(i = 0; i < maxf; i++) { 941 int index = (maxb + i) + 1; 942 ma[index] = maf[i]; 943 vm_page_flag_clear(ma[index], PG_CLEANCHK); 944 } 945 runlen = maxb + maxf + 1; 946 947 vm_pageout_flush(ma, runlen, pagerflags); 948 for (i = 0; i < runlen; i++) { 949 if (ma[i]->valid & ma[i]->dirty) { 950 pmap_page_protect(ma[i], VM_PROT_READ); 951 vm_page_flag_set(ma[i], PG_CLEANCHK); 952 953 /* 954 * maxf will end up being the actual number of pages 955 * we wrote out contiguously, non-inclusive of the 956 * first page. We do not count look-behind pages. 957 */ 958 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 959 maxf = i - maxb - 1; 960 } 961 } 962 return(maxf + 1); 963} 964 965/* 966 * Note that there is absolutely no sense in writing out 967 * anonymous objects, so we track down the vnode object 968 * to write out. 969 * We invalidate (remove) all pages from the address space 970 * for semantic correctness. 971 * 972 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 973 * may start out with a NULL object. 974 */ 975void 976vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 977 boolean_t syncio, boolean_t invalidate) 978{ 979 vm_object_t backing_object; 980 struct vnode *vp; 981 int flags; 982 983 if (object == NULL) 984 return; 985 VM_OBJECT_LOCK(object); 986 while ((backing_object = object->backing_object) != NULL) { 987 VM_OBJECT_LOCK(backing_object); 988 VM_OBJECT_UNLOCK(object); 989 object = backing_object; 990 offset += object->backing_object_offset; 991 if (object->size < OFF_TO_IDX(offset + size)) 992 size = IDX_TO_OFF(object->size) - offset; 993 } 994 /* 995 * Flush pages if writing is allowed, invalidate them 996 * if invalidation requested. Pages undergoing I/O 997 * will be ignored by vm_object_page_remove(). 998 * 999 * We cannot lock the vnode and then wait for paging 1000 * to complete without deadlocking against vm_fault. 1001 * Instead we simply call vm_object_page_remove() and 1002 * allow it to block internally on a page-by-page 1003 * basis when it encounters pages undergoing async 1004 * I/O. 1005 */ 1006 if (object->type == OBJT_VNODE && 1007 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1008 vp = object->handle; 1009 VM_OBJECT_UNLOCK(object); 1010 mtx_lock(&Giant); 1011 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); 1012 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1013 flags |= invalidate ? OBJPC_INVAL : 0; 1014 VM_OBJECT_LOCK(object); 1015 vm_object_page_clean(object, 1016 OFF_TO_IDX(offset), 1017 OFF_TO_IDX(offset + size + PAGE_MASK), 1018 flags); 1019 VM_OBJECT_UNLOCK(object); 1020 VOP_UNLOCK(vp, 0, curthread); 1021 mtx_unlock(&Giant); 1022 VM_OBJECT_LOCK(object); 1023 } 1024 if ((object->type == OBJT_VNODE || 1025 object->type == OBJT_DEVICE) && invalidate) { 1026 vm_object_page_remove(object, 1027 OFF_TO_IDX(offset), 1028 OFF_TO_IDX(offset + size + PAGE_MASK), 1029 old_msync ? FALSE : TRUE); 1030 } 1031 VM_OBJECT_UNLOCK(object); 1032} 1033 1034/* 1035 * vm_object_madvise: 1036 * 1037 * Implements the madvise function at the object/page level. 1038 * 1039 * MADV_WILLNEED (any object) 1040 * 1041 * Activate the specified pages if they are resident. 1042 * 1043 * MADV_DONTNEED (any object) 1044 * 1045 * Deactivate the specified pages if they are resident. 1046 * 1047 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1048 * OBJ_ONEMAPPING only) 1049 * 1050 * Deactivate and clean the specified pages if they are 1051 * resident. This permits the process to reuse the pages 1052 * without faulting or the kernel to reclaim the pages 1053 * without I/O. 1054 */ 1055void 1056vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1057{ 1058 vm_pindex_t end, tpindex; 1059 vm_object_t backing_object, tobject; 1060 vm_page_t m; 1061 1062 if (object == NULL) 1063 return; 1064 end = pindex + count; 1065 /* 1066 * Locate and adjust resident pages 1067 */ 1068 for (; pindex < end; pindex += 1) { 1069relookup: 1070 tobject = object; 1071 tpindex = pindex; 1072 VM_OBJECT_LOCK(tobject); 1073shadowlookup: 1074 /* 1075 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1076 * and those pages must be OBJ_ONEMAPPING. 1077 */ 1078 if (advise == MADV_FREE) { 1079 if ((tobject->type != OBJT_DEFAULT && 1080 tobject->type != OBJT_SWAP) || 1081 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1082 goto unlock_tobject; 1083 } 1084 } 1085 m = vm_page_lookup(tobject, tpindex); 1086 if (m == NULL) { 1087 /* 1088 * There may be swap even if there is no backing page 1089 */ 1090 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1091 swap_pager_freespace(tobject, tpindex, 1); 1092 /* 1093 * next object 1094 */ 1095 backing_object = tobject->backing_object; 1096 if (backing_object == NULL) 1097 goto unlock_tobject; 1098 VM_OBJECT_LOCK(backing_object); 1099 VM_OBJECT_UNLOCK(tobject); 1100 tobject = backing_object; 1101 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1102 goto shadowlookup; 1103 } 1104 /* 1105 * If the page is busy or not in a normal active state, 1106 * we skip it. If the page is not managed there are no 1107 * page queues to mess with. Things can break if we mess 1108 * with pages in any of the below states. 1109 */ 1110 vm_page_lock_queues(); 1111 if (m->hold_count || 1112 m->wire_count || 1113 (m->flags & PG_UNMANAGED) || 1114 m->valid != VM_PAGE_BITS_ALL) { 1115 vm_page_unlock_queues(); 1116 goto unlock_tobject; 1117 } 1118 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) { 1119 VM_OBJECT_UNLOCK(tobject); 1120 goto relookup; 1121 } 1122 if (advise == MADV_WILLNEED) { 1123 vm_page_activate(m); 1124 } else if (advise == MADV_DONTNEED) { 1125 vm_page_dontneed(m); 1126 } else if (advise == MADV_FREE) { 1127 /* 1128 * Mark the page clean. This will allow the page 1129 * to be freed up by the system. However, such pages 1130 * are often reused quickly by malloc()/free() 1131 * so we do not do anything that would cause 1132 * a page fault if we can help it. 1133 * 1134 * Specifically, we do not try to actually free 1135 * the page now nor do we try to put it in the 1136 * cache (which would cause a page fault on reuse). 1137 * 1138 * But we do make the page is freeable as we 1139 * can without actually taking the step of unmapping 1140 * it. 1141 */ 1142 pmap_clear_modify(m); 1143 m->dirty = 0; 1144 m->act_count = 0; 1145 vm_page_dontneed(m); 1146 } 1147 vm_page_unlock_queues(); 1148 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1149 swap_pager_freespace(tobject, tpindex, 1); 1150unlock_tobject: 1151 VM_OBJECT_UNLOCK(tobject); 1152 } 1153} 1154 1155/* 1156 * vm_object_shadow: 1157 * 1158 * Create a new object which is backed by the 1159 * specified existing object range. The source 1160 * object reference is deallocated. 1161 * 1162 * The new object and offset into that object 1163 * are returned in the source parameters. 1164 */ 1165void 1166vm_object_shadow( 1167 vm_object_t *object, /* IN/OUT */ 1168 vm_ooffset_t *offset, /* IN/OUT */ 1169 vm_size_t length) 1170{ 1171 vm_object_t source; 1172 vm_object_t result; 1173 1174 source = *object; 1175 1176 /* 1177 * Don't create the new object if the old object isn't shared. 1178 */ 1179 if (source != NULL) { 1180 VM_OBJECT_LOCK(source); 1181 if (source->ref_count == 1 && 1182 source->handle == NULL && 1183 (source->type == OBJT_DEFAULT || 1184 source->type == OBJT_SWAP)) { 1185 VM_OBJECT_UNLOCK(source); 1186 return; 1187 } 1188 VM_OBJECT_UNLOCK(source); 1189 } 1190 1191 /* 1192 * Allocate a new object with the given length. 1193 */ 1194 result = vm_object_allocate(OBJT_DEFAULT, length); 1195 1196 /* 1197 * The new object shadows the source object, adding a reference to it. 1198 * Our caller changes his reference to point to the new object, 1199 * removing a reference to the source object. Net result: no change 1200 * of reference count. 1201 * 1202 * Try to optimize the result object's page color when shadowing 1203 * in order to maintain page coloring consistency in the combined 1204 * shadowed object. 1205 */ 1206 result->backing_object = source; 1207 /* 1208 * Store the offset into the source object, and fix up the offset into 1209 * the new object. 1210 */ 1211 result->backing_object_offset = *offset; 1212 if (source != NULL) { 1213 VM_OBJECT_LOCK(source); 1214 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1215 source->shadow_count++; 1216 source->generation++; 1217 if (length < source->size) 1218 length = source->size; 1219 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1220 source->generation > 1) 1221 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1222 result->pg_color = (source->pg_color + 1223 length * source->generation) & PQ_L2_MASK; 1224 VM_OBJECT_UNLOCK(source); 1225 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1226 PQ_L2_MASK; 1227 } 1228 1229 1230 /* 1231 * Return the new things 1232 */ 1233 *offset = 0; 1234 *object = result; 1235} 1236 1237/* 1238 * vm_object_split: 1239 * 1240 * Split the pages in a map entry into a new object. This affords 1241 * easier removal of unused pages, and keeps object inheritance from 1242 * being a negative impact on memory usage. 1243 */ 1244void 1245vm_object_split(vm_map_entry_t entry) 1246{ 1247 vm_page_t m; 1248 vm_object_t orig_object, new_object, source; 1249 vm_pindex_t offidxstart, offidxend; 1250 vm_size_t idx, size; 1251 1252 orig_object = entry->object.vm_object; 1253 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1254 return; 1255 if (orig_object->ref_count <= 1) 1256 return; 1257 VM_OBJECT_UNLOCK(orig_object); 1258 1259 offidxstart = OFF_TO_IDX(entry->offset); 1260 offidxend = offidxstart + OFF_TO_IDX(entry->end - entry->start); 1261 size = offidxend - offidxstart; 1262 1263 /* 1264 * If swap_pager_copy() is later called, it will convert new_object 1265 * into a swap object. 1266 */ 1267 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1268 1269 VM_OBJECT_LOCK(new_object); 1270 VM_OBJECT_LOCK(orig_object); 1271 source = orig_object->backing_object; 1272 if (source != NULL) { 1273 VM_OBJECT_LOCK(source); 1274 LIST_INSERT_HEAD(&source->shadow_head, 1275 new_object, shadow_list); 1276 source->shadow_count++; 1277 source->generation++; 1278 vm_object_reference_locked(source); /* for new_object */ 1279 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1280 VM_OBJECT_UNLOCK(source); 1281 new_object->backing_object_offset = 1282 orig_object->backing_object_offset + entry->offset; 1283 new_object->backing_object = source; 1284 } 1285 for (idx = 0; idx < size; idx++) { 1286 retry: 1287 m = vm_page_lookup(orig_object, offidxstart + idx); 1288 if (m == NULL) 1289 continue; 1290 1291 /* 1292 * We must wait for pending I/O to complete before we can 1293 * rename the page. 1294 * 1295 * We do not have to VM_PROT_NONE the page as mappings should 1296 * not be changed by this operation. 1297 */ 1298 vm_page_lock_queues(); 1299 if ((m->flags & PG_BUSY) || m->busy) { 1300 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1301 VM_OBJECT_UNLOCK(orig_object); 1302 VM_OBJECT_UNLOCK(new_object); 1303 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "spltwt", 0); 1304 VM_OBJECT_LOCK(new_object); 1305 VM_OBJECT_LOCK(orig_object); 1306 goto retry; 1307 } 1308 vm_page_busy(m); 1309 vm_page_rename(m, new_object, idx); 1310 /* page automatically made dirty by rename and cache handled */ 1311 vm_page_busy(m); 1312 vm_page_unlock_queues(); 1313 } 1314 if (orig_object->type == OBJT_SWAP) { 1315 /* 1316 * swap_pager_copy() can sleep, in which case the orig_object's 1317 * and new_object's locks are released and reacquired. 1318 */ 1319 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1320 } 1321 VM_OBJECT_UNLOCK(orig_object); 1322 vm_page_lock_queues(); 1323 TAILQ_FOREACH(m, &new_object->memq, listq) 1324 vm_page_wakeup(m); 1325 vm_page_unlock_queues(); 1326 VM_OBJECT_UNLOCK(new_object); 1327 entry->object.vm_object = new_object; 1328 entry->offset = 0LL; 1329 vm_object_deallocate(orig_object); 1330 VM_OBJECT_LOCK(new_object); 1331} 1332 1333#define OBSC_TEST_ALL_SHADOWED 0x0001 1334#define OBSC_COLLAPSE_NOWAIT 0x0002 1335#define OBSC_COLLAPSE_WAIT 0x0004 1336 1337static int 1338vm_object_backing_scan(vm_object_t object, int op) 1339{ 1340 int r = 1; 1341 vm_page_t p; 1342 vm_object_t backing_object; 1343 vm_pindex_t backing_offset_index; 1344 1345 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1346 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); 1347 1348 backing_object = object->backing_object; 1349 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1350 1351 /* 1352 * Initial conditions 1353 */ 1354 if (op & OBSC_TEST_ALL_SHADOWED) { 1355 /* 1356 * We do not want to have to test for the existence of 1357 * swap pages in the backing object. XXX but with the 1358 * new swapper this would be pretty easy to do. 1359 * 1360 * XXX what about anonymous MAP_SHARED memory that hasn't 1361 * been ZFOD faulted yet? If we do not test for this, the 1362 * shadow test may succeed! XXX 1363 */ 1364 if (backing_object->type != OBJT_DEFAULT) { 1365 return (0); 1366 } 1367 } 1368 if (op & OBSC_COLLAPSE_WAIT) { 1369 vm_object_set_flag(backing_object, OBJ_DEAD); 1370 } 1371 1372 /* 1373 * Our scan 1374 */ 1375 p = TAILQ_FIRST(&backing_object->memq); 1376 while (p) { 1377 vm_page_t next = TAILQ_NEXT(p, listq); 1378 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1379 1380 if (op & OBSC_TEST_ALL_SHADOWED) { 1381 vm_page_t pp; 1382 1383 /* 1384 * Ignore pages outside the parent object's range 1385 * and outside the parent object's mapping of the 1386 * backing object. 1387 * 1388 * note that we do not busy the backing object's 1389 * page. 1390 */ 1391 if ( 1392 p->pindex < backing_offset_index || 1393 new_pindex >= object->size 1394 ) { 1395 p = next; 1396 continue; 1397 } 1398 1399 /* 1400 * See if the parent has the page or if the parent's 1401 * object pager has the page. If the parent has the 1402 * page but the page is not valid, the parent's 1403 * object pager must have the page. 1404 * 1405 * If this fails, the parent does not completely shadow 1406 * the object and we might as well give up now. 1407 */ 1408 1409 pp = vm_page_lookup(object, new_pindex); 1410 if ( 1411 (pp == NULL || pp->valid == 0) && 1412 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1413 ) { 1414 r = 0; 1415 break; 1416 } 1417 } 1418 1419 /* 1420 * Check for busy page 1421 */ 1422 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1423 vm_page_t pp; 1424 1425 vm_page_lock_queues(); 1426 if (op & OBSC_COLLAPSE_NOWAIT) { 1427 if ((p->flags & PG_BUSY) || 1428 !p->valid || 1429 p->hold_count || 1430 p->wire_count || 1431 p->busy) { 1432 vm_page_unlock_queues(); 1433 p = next; 1434 continue; 1435 } 1436 } else if (op & OBSC_COLLAPSE_WAIT) { 1437 if ((p->flags & PG_BUSY) || p->busy) { 1438 vm_page_flag_set(p, 1439 PG_WANTED | PG_REFERENCED); 1440 VM_OBJECT_UNLOCK(backing_object); 1441 VM_OBJECT_UNLOCK(object); 1442 msleep(p, &vm_page_queue_mtx, 1443 PDROP | PVM, "vmocol", 0); 1444 VM_OBJECT_LOCK(object); 1445 VM_OBJECT_LOCK(backing_object); 1446 /* 1447 * If we slept, anything could have 1448 * happened. Since the object is 1449 * marked dead, the backing offset 1450 * should not have changed so we 1451 * just restart our scan. 1452 */ 1453 p = TAILQ_FIRST(&backing_object->memq); 1454 continue; 1455 } 1456 } 1457 1458 /* 1459 * Busy the page 1460 */ 1461 vm_page_busy(p); 1462 vm_page_unlock_queues(); 1463 1464 KASSERT( 1465 p->object == backing_object, 1466 ("vm_object_qcollapse(): object mismatch") 1467 ); 1468 1469 /* 1470 * Destroy any associated swap 1471 */ 1472 if (backing_object->type == OBJT_SWAP) { 1473 swap_pager_freespace( 1474 backing_object, 1475 p->pindex, 1476 1 1477 ); 1478 } 1479 1480 if ( 1481 p->pindex < backing_offset_index || 1482 new_pindex >= object->size 1483 ) { 1484 /* 1485 * Page is out of the parent object's range, we 1486 * can simply destroy it. 1487 */ 1488 vm_page_lock_queues(); 1489 pmap_remove_all(p); 1490 vm_page_free(p); 1491 vm_page_unlock_queues(); 1492 p = next; 1493 continue; 1494 } 1495 1496 pp = vm_page_lookup(object, new_pindex); 1497 if ( 1498 pp != NULL || 1499 vm_pager_has_page(object, new_pindex, NULL, NULL) 1500 ) { 1501 /* 1502 * page already exists in parent OR swap exists 1503 * for this location in the parent. Destroy 1504 * the original page from the backing object. 1505 * 1506 * Leave the parent's page alone 1507 */ 1508 vm_page_lock_queues(); 1509 pmap_remove_all(p); 1510 vm_page_free(p); 1511 vm_page_unlock_queues(); 1512 p = next; 1513 continue; 1514 } 1515 1516 /* 1517 * Page does not exist in parent, rename the 1518 * page from the backing object to the main object. 1519 * 1520 * If the page was mapped to a process, it can remain 1521 * mapped through the rename. 1522 */ 1523 vm_page_lock_queues(); 1524 vm_page_rename(p, object, new_pindex); 1525 vm_page_unlock_queues(); 1526 /* page automatically made dirty by rename */ 1527 } 1528 p = next; 1529 } 1530 return (r); 1531} 1532 1533 1534/* 1535 * this version of collapse allows the operation to occur earlier and 1536 * when paging_in_progress is true for an object... This is not a complete 1537 * operation, but should plug 99.9% of the rest of the leaks. 1538 */ 1539static void 1540vm_object_qcollapse(vm_object_t object) 1541{ 1542 vm_object_t backing_object = object->backing_object; 1543 1544 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1545 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1546 1547 if (backing_object->ref_count != 1) 1548 return; 1549 1550 backing_object->ref_count += 2; 1551 1552 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1553 1554 backing_object->ref_count -= 2; 1555} 1556 1557/* 1558 * vm_object_collapse: 1559 * 1560 * Collapse an object with the object backing it. 1561 * Pages in the backing object are moved into the 1562 * parent, and the backing object is deallocated. 1563 */ 1564void 1565vm_object_collapse(vm_object_t object) 1566{ 1567 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1568 1569 while (TRUE) { 1570 vm_object_t backing_object; 1571 1572 /* 1573 * Verify that the conditions are right for collapse: 1574 * 1575 * The object exists and the backing object exists. 1576 */ 1577 if ((backing_object = object->backing_object) == NULL) 1578 break; 1579 1580 /* 1581 * we check the backing object first, because it is most likely 1582 * not collapsable. 1583 */ 1584 VM_OBJECT_LOCK(backing_object); 1585 if (backing_object->handle != NULL || 1586 (backing_object->type != OBJT_DEFAULT && 1587 backing_object->type != OBJT_SWAP) || 1588 (backing_object->flags & OBJ_DEAD) || 1589 object->handle != NULL || 1590 (object->type != OBJT_DEFAULT && 1591 object->type != OBJT_SWAP) || 1592 (object->flags & OBJ_DEAD)) { 1593 VM_OBJECT_UNLOCK(backing_object); 1594 break; 1595 } 1596 1597 if ( 1598 object->paging_in_progress != 0 || 1599 backing_object->paging_in_progress != 0 1600 ) { 1601 vm_object_qcollapse(object); 1602 VM_OBJECT_UNLOCK(backing_object); 1603 break; 1604 } 1605 /* 1606 * We know that we can either collapse the backing object (if 1607 * the parent is the only reference to it) or (perhaps) have 1608 * the parent bypass the object if the parent happens to shadow 1609 * all the resident pages in the entire backing object. 1610 * 1611 * This is ignoring pager-backed pages such as swap pages. 1612 * vm_object_backing_scan fails the shadowing test in this 1613 * case. 1614 */ 1615 if (backing_object->ref_count == 1) { 1616 /* 1617 * If there is exactly one reference to the backing 1618 * object, we can collapse it into the parent. 1619 */ 1620 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1621 1622 /* 1623 * Move the pager from backing_object to object. 1624 */ 1625 if (backing_object->type == OBJT_SWAP) { 1626 /* 1627 * swap_pager_copy() can sleep, in which case 1628 * the backing_object's and object's locks are 1629 * released and reacquired. 1630 */ 1631 swap_pager_copy( 1632 backing_object, 1633 object, 1634 OFF_TO_IDX(object->backing_object_offset), TRUE); 1635 } 1636 /* 1637 * Object now shadows whatever backing_object did. 1638 * Note that the reference to 1639 * backing_object->backing_object moves from within 1640 * backing_object to within object. 1641 */ 1642 LIST_REMOVE(object, shadow_list); 1643 backing_object->shadow_count--; 1644 backing_object->generation++; 1645 if (backing_object->backing_object) { 1646 VM_OBJECT_LOCK(backing_object->backing_object); 1647 LIST_REMOVE(backing_object, shadow_list); 1648 LIST_INSERT_HEAD( 1649 &backing_object->backing_object->shadow_head, 1650 object, shadow_list); 1651 /* 1652 * The shadow_count has not changed. 1653 */ 1654 backing_object->backing_object->generation++; 1655 VM_OBJECT_UNLOCK(backing_object->backing_object); 1656 } 1657 object->backing_object = backing_object->backing_object; 1658 object->backing_object_offset += 1659 backing_object->backing_object_offset; 1660 1661 /* 1662 * Discard backing_object. 1663 * 1664 * Since the backing object has no pages, no pager left, 1665 * and no object references within it, all that is 1666 * necessary is to dispose of it. 1667 */ 1668 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1669 VM_OBJECT_UNLOCK(backing_object); 1670 1671 mtx_lock(&vm_object_list_mtx); 1672 TAILQ_REMOVE( 1673 &vm_object_list, 1674 backing_object, 1675 object_list 1676 ); 1677 mtx_unlock(&vm_object_list_mtx); 1678 1679 uma_zfree(obj_zone, backing_object); 1680 1681 object_collapses++; 1682 } else { 1683 vm_object_t new_backing_object; 1684 1685 /* 1686 * If we do not entirely shadow the backing object, 1687 * there is nothing we can do so we give up. 1688 */ 1689 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1690 VM_OBJECT_UNLOCK(backing_object); 1691 break; 1692 } 1693 1694 /* 1695 * Make the parent shadow the next object in the 1696 * chain. Deallocating backing_object will not remove 1697 * it, since its reference count is at least 2. 1698 */ 1699 LIST_REMOVE(object, shadow_list); 1700 backing_object->shadow_count--; 1701 backing_object->generation++; 1702 1703 new_backing_object = backing_object->backing_object; 1704 if ((object->backing_object = new_backing_object) != NULL) { 1705 VM_OBJECT_LOCK(new_backing_object); 1706 LIST_INSERT_HEAD( 1707 &new_backing_object->shadow_head, 1708 object, 1709 shadow_list 1710 ); 1711 new_backing_object->shadow_count++; 1712 new_backing_object->generation++; 1713 vm_object_reference_locked(new_backing_object); 1714 VM_OBJECT_UNLOCK(new_backing_object); 1715 object->backing_object_offset += 1716 backing_object->backing_object_offset; 1717 } 1718 1719 /* 1720 * Drop the reference count on backing_object. Since 1721 * its ref_count was at least 2, it will not vanish. 1722 */ 1723 backing_object->ref_count--; 1724 VM_OBJECT_UNLOCK(backing_object); 1725 object_bypasses++; 1726 } 1727 1728 /* 1729 * Try again with this object's new backing object. 1730 */ 1731 } 1732} 1733 1734/* 1735 * vm_object_page_remove: 1736 * 1737 * Removes all physical pages in the given range from the 1738 * object's list of pages. If the range's end is zero, all 1739 * physical pages from the range's start to the end of the object 1740 * are deleted. 1741 * 1742 * The object must be locked. 1743 */ 1744void 1745vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1746 boolean_t clean_only) 1747{ 1748 vm_page_t p, next; 1749 1750 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1751 if (object->resident_page_count == 0) 1752 return; 1753 1754 /* 1755 * Since physically-backed objects do not use managed pages, we can't 1756 * remove pages from the object (we must instead remove the page 1757 * references, and then destroy the object). 1758 */ 1759 KASSERT(object->type != OBJT_PHYS, 1760 ("attempt to remove pages from a physical object")); 1761 1762 vm_object_pip_add(object, 1); 1763again: 1764 vm_page_lock_queues(); 1765 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1766 if (p->pindex < start) { 1767 p = vm_page_splay(start, object->root); 1768 if ((object->root = p)->pindex < start) 1769 p = TAILQ_NEXT(p, listq); 1770 } 1771 } 1772 /* 1773 * Assert: the variable p is either (1) the page with the 1774 * least pindex greater than or equal to the parameter pindex 1775 * or (2) NULL. 1776 */ 1777 for (; 1778 p != NULL && (p->pindex < end || end == 0); 1779 p = next) { 1780 next = TAILQ_NEXT(p, listq); 1781 1782 if (p->wire_count != 0) { 1783 pmap_remove_all(p); 1784 if (!clean_only) 1785 p->valid = 0; 1786 continue; 1787 } 1788 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1789 goto again; 1790 if (clean_only && p->valid) { 1791 pmap_page_protect(p, VM_PROT_READ | VM_PROT_EXECUTE); 1792 if (p->valid & p->dirty) 1793 continue; 1794 } 1795 vm_page_busy(p); 1796 pmap_remove_all(p); 1797 vm_page_free(p); 1798 } 1799 vm_page_unlock_queues(); 1800 vm_object_pip_wakeup(object); 1801} 1802 1803/* 1804 * Routine: vm_object_coalesce 1805 * Function: Coalesces two objects backing up adjoining 1806 * regions of memory into a single object. 1807 * 1808 * returns TRUE if objects were combined. 1809 * 1810 * NOTE: Only works at the moment if the second object is NULL - 1811 * if it's not, which object do we lock first? 1812 * 1813 * Parameters: 1814 * prev_object First object to coalesce 1815 * prev_offset Offset into prev_object 1816 * prev_size Size of reference to prev_object 1817 * next_size Size of reference to the second object 1818 * 1819 * Conditions: 1820 * The object must *not* be locked. 1821 */ 1822boolean_t 1823vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 1824 vm_size_t prev_size, vm_size_t next_size) 1825{ 1826 vm_pindex_t next_pindex; 1827 1828 if (prev_object == NULL) 1829 return (TRUE); 1830 VM_OBJECT_LOCK(prev_object); 1831 if (prev_object->type != OBJT_DEFAULT && 1832 prev_object->type != OBJT_SWAP) { 1833 VM_OBJECT_UNLOCK(prev_object); 1834 return (FALSE); 1835 } 1836 1837 /* 1838 * Try to collapse the object first 1839 */ 1840 vm_object_collapse(prev_object); 1841 1842 /* 1843 * Can't coalesce if: . more than one reference . paged out . shadows 1844 * another object . has a copy elsewhere (any of which mean that the 1845 * pages not mapped to prev_entry may be in use anyway) 1846 */ 1847 if (prev_object->backing_object != NULL) { 1848 VM_OBJECT_UNLOCK(prev_object); 1849 return (FALSE); 1850 } 1851 1852 prev_size >>= PAGE_SHIFT; 1853 next_size >>= PAGE_SHIFT; 1854 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 1855 1856 if ((prev_object->ref_count > 1) && 1857 (prev_object->size != next_pindex)) { 1858 VM_OBJECT_UNLOCK(prev_object); 1859 return (FALSE); 1860 } 1861 1862 /* 1863 * Remove any pages that may still be in the object from a previous 1864 * deallocation. 1865 */ 1866 if (next_pindex < prev_object->size) { 1867 vm_object_page_remove(prev_object, 1868 next_pindex, 1869 next_pindex + next_size, FALSE); 1870 if (prev_object->type == OBJT_SWAP) 1871 swap_pager_freespace(prev_object, 1872 next_pindex, next_size); 1873 } 1874 1875 /* 1876 * Extend the object if necessary. 1877 */ 1878 if (next_pindex + next_size > prev_object->size) 1879 prev_object->size = next_pindex + next_size; 1880 1881 VM_OBJECT_UNLOCK(prev_object); 1882 return (TRUE); 1883} 1884 1885void 1886vm_object_set_writeable_dirty(vm_object_t object) 1887{ 1888 struct vnode *vp; 1889 1890 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1891 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1892 if (object->type == OBJT_VNODE && 1893 (vp = (struct vnode *)object->handle) != NULL) { 1894 VI_LOCK(vp); 1895 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1896 vp->v_iflag |= VI_OBJDIRTY; 1897 VI_UNLOCK(vp); 1898 } 1899} 1900 1901#include "opt_ddb.h" 1902#ifdef DDB 1903#include <sys/kernel.h> 1904 1905#include <sys/cons.h> 1906 1907#include <ddb/ddb.h> 1908 1909static int 1910_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1911{ 1912 vm_map_t tmpm; 1913 vm_map_entry_t tmpe; 1914 vm_object_t obj; 1915 int entcount; 1916 1917 if (map == 0) 1918 return 0; 1919 1920 if (entry == 0) { 1921 tmpe = map->header.next; 1922 entcount = map->nentries; 1923 while (entcount-- && (tmpe != &map->header)) { 1924 if (_vm_object_in_map(map, object, tmpe)) { 1925 return 1; 1926 } 1927 tmpe = tmpe->next; 1928 } 1929 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1930 tmpm = entry->object.sub_map; 1931 tmpe = tmpm->header.next; 1932 entcount = tmpm->nentries; 1933 while (entcount-- && tmpe != &tmpm->header) { 1934 if (_vm_object_in_map(tmpm, object, tmpe)) { 1935 return 1; 1936 } 1937 tmpe = tmpe->next; 1938 } 1939 } else if ((obj = entry->object.vm_object) != NULL) { 1940 for (; obj; obj = obj->backing_object) 1941 if (obj == object) { 1942 return 1; 1943 } 1944 } 1945 return 0; 1946} 1947 1948static int 1949vm_object_in_map(vm_object_t object) 1950{ 1951 struct proc *p; 1952 1953 /* sx_slock(&allproc_lock); */ 1954 LIST_FOREACH(p, &allproc, p_list) { 1955 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1956 continue; 1957 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1958 /* sx_sunlock(&allproc_lock); */ 1959 return 1; 1960 } 1961 } 1962 /* sx_sunlock(&allproc_lock); */ 1963 if (_vm_object_in_map(kernel_map, object, 0)) 1964 return 1; 1965 if (_vm_object_in_map(kmem_map, object, 0)) 1966 return 1; 1967 if (_vm_object_in_map(pager_map, object, 0)) 1968 return 1; 1969 if (_vm_object_in_map(buffer_map, object, 0)) 1970 return 1; 1971 return 0; 1972} 1973 1974DB_SHOW_COMMAND(vmochk, vm_object_check) 1975{ 1976 vm_object_t object; 1977 1978 /* 1979 * make sure that internal objs are in a map somewhere 1980 * and none have zero ref counts. 1981 */ 1982 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1983 if (object->handle == NULL && 1984 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1985 if (object->ref_count == 0) { 1986 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1987 (long)object->size); 1988 } 1989 if (!vm_object_in_map(object)) { 1990 db_printf( 1991 "vmochk: internal obj is not in a map: " 1992 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1993 object->ref_count, (u_long)object->size, 1994 (u_long)object->size, 1995 (void *)object->backing_object); 1996 } 1997 } 1998 } 1999} 2000 2001/* 2002 * vm_object_print: [ debug ] 2003 */ 2004DB_SHOW_COMMAND(object, vm_object_print_static) 2005{ 2006 /* XXX convert args. */ 2007 vm_object_t object = (vm_object_t)addr; 2008 boolean_t full = have_addr; 2009 2010 vm_page_t p; 2011 2012 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2013#define count was_count 2014 2015 int count; 2016 2017 if (object == NULL) 2018 return; 2019 2020 db_iprintf( 2021 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 2022 object, (int)object->type, (uintmax_t)object->size, 2023 object->resident_page_count, object->ref_count, object->flags); 2024 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2025 object->shadow_count, 2026 object->backing_object ? object->backing_object->ref_count : 0, 2027 object->backing_object, (uintmax_t)object->backing_object_offset); 2028 2029 if (!full) 2030 return; 2031 2032 db_indent += 2; 2033 count = 0; 2034 TAILQ_FOREACH(p, &object->memq, listq) { 2035 if (count == 0) 2036 db_iprintf("memory:="); 2037 else if (count == 6) { 2038 db_printf("\n"); 2039 db_iprintf(" ..."); 2040 count = 0; 2041 } else 2042 db_printf(","); 2043 count++; 2044 2045 db_printf("(off=0x%jx,page=0x%jx)", 2046 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2047 } 2048 if (count != 0) 2049 db_printf("\n"); 2050 db_indent -= 2; 2051} 2052 2053/* XXX. */ 2054#undef count 2055 2056/* XXX need this non-static entry for calling from vm_map_print. */ 2057void 2058vm_object_print( 2059 /* db_expr_t */ long addr, 2060 boolean_t have_addr, 2061 /* db_expr_t */ long count, 2062 char *modif) 2063{ 2064 vm_object_print_static(addr, have_addr, count, modif); 2065} 2066 2067DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2068{ 2069 vm_object_t object; 2070 int nl = 0; 2071 int c; 2072 2073 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2074 vm_pindex_t idx, fidx; 2075 vm_pindex_t osize; 2076 vm_paddr_t pa = -1, padiff; 2077 int rcount; 2078 vm_page_t m; 2079 2080 db_printf("new object: %p\n", (void *)object); 2081 if (nl > 18) { 2082 c = cngetc(); 2083 if (c != ' ') 2084 return; 2085 nl = 0; 2086 } 2087 nl++; 2088 rcount = 0; 2089 fidx = 0; 2090 osize = object->size; 2091 if (osize > 128) 2092 osize = 128; 2093 for (idx = 0; idx < osize; idx++) { 2094 m = vm_page_lookup(object, idx); 2095 if (m == NULL) { 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 rcount = 0; 2107 } 2108 continue; 2109 } 2110 2111 2112 if (rcount && 2113 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2114 ++rcount; 2115 continue; 2116 } 2117 if (rcount) { 2118 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2119 padiff >>= PAGE_SHIFT; 2120 padiff &= PQ_L2_MASK; 2121 if (padiff == 0) { 2122 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2123 ++rcount; 2124 continue; 2125 } 2126 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2127 (long)fidx, rcount, (long)pa); 2128 db_printf("pd(%ld)\n", (long)padiff); 2129 if (nl > 18) { 2130 c = cngetc(); 2131 if (c != ' ') 2132 return; 2133 nl = 0; 2134 } 2135 nl++; 2136 } 2137 fidx = idx; 2138 pa = VM_PAGE_TO_PHYS(m); 2139 rcount = 1; 2140 } 2141 if (rcount) { 2142 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2143 (long)fidx, rcount, (long)pa); 2144 if (nl > 18) { 2145 c = cngetc(); 2146 if (c != ' ') 2147 return; 2148 nl = 0; 2149 } 2150 nl++; 2151 } 2152 } 2153} 2154#endif /* DDB */ 2155