vm_object.c revision 139825
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 139825 2005-01-07 02:29:27Z imp $"); 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; 148 149/* 150 * next_index determines the page color that is assigned to the next 151 * allocated object. Accesses to next_index are not synchronized 152 * because the effects of two or more object allocations using 153 * next_index simultaneously are inconsequential. At any given time, 154 * numerous objects have the same page color. 155 */ 156static int next_index; 157 158static uma_zone_t obj_zone; 159#define VM_OBJECTS_INIT 256 160 161static int vm_object_zinit(void *mem, int size, int flags); 162 163#ifdef INVARIANTS 164static void vm_object_zdtor(void *mem, int size, void *arg); 165 166static void 167vm_object_zdtor(void *mem, int size, void *arg) 168{ 169 vm_object_t object; 170 171 object = (vm_object_t)mem; 172 KASSERT(TAILQ_EMPTY(&object->memq), 173 ("object %p has resident pages", 174 object)); 175 KASSERT(object->paging_in_progress == 0, 176 ("object %p paging_in_progress = %d", 177 object, object->paging_in_progress)); 178 KASSERT(object->resident_page_count == 0, 179 ("object %p resident_page_count = %d", 180 object, object->resident_page_count)); 181 KASSERT(object->shadow_count == 0, 182 ("object %p shadow_count = %d", 183 object, object->shadow_count)); 184} 185#endif 186 187static int 188vm_object_zinit(void *mem, int size, int flags) 189{ 190 vm_object_t object; 191 192 object = (vm_object_t)mem; 193 bzero(&object->mtx, sizeof(object->mtx)); 194 VM_OBJECT_LOCK_INIT(object, "standard object"); 195 196 /* These are true for any object that has been freed */ 197 object->paging_in_progress = 0; 198 object->resident_page_count = 0; 199 object->shadow_count = 0; 200 return (0); 201} 202 203void 204_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 205{ 206 int incr; 207 208 TAILQ_INIT(&object->memq); 209 LIST_INIT(&object->shadow_head); 210 211 object->root = NULL; 212 object->type = type; 213 object->size = size; 214 object->generation = 1; 215 object->ref_count = 1; 216 object->flags = 0; 217 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 218 object->flags = OBJ_ONEMAPPING; 219 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 220 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 221 else 222 incr = size; 223 object->pg_color = next_index; 224 next_index = (object->pg_color + incr) & PQ_L2_MASK; 225 object->handle = NULL; 226 object->backing_object = NULL; 227 object->backing_object_offset = (vm_ooffset_t) 0; 228 229 mtx_lock(&vm_object_list_mtx); 230 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 231 mtx_unlock(&vm_object_list_mtx); 232} 233 234/* 235 * vm_object_init: 236 * 237 * Initialize the VM objects module. 238 */ 239void 240vm_object_init(void) 241{ 242 TAILQ_INIT(&vm_object_list); 243 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 244 245 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object"); 246 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 247 kernel_object); 248 249 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object"); 250 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 251 kmem_object); 252 253 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 254#ifdef INVARIANTS 255 vm_object_zdtor, 256#else 257 NULL, 258#endif 259 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); 260 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 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: 325 * 326 * Returns a new object with the given size. 327 */ 328vm_object_t 329vm_object_allocate(objtype_t type, vm_pindex_t size) 330{ 331 vm_object_t object; 332 333 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 334 _vm_object_allocate(type, size, object); 335 return (object); 336} 337 338 339/* 340 * vm_object_reference: 341 * 342 * Gets another reference to the given object. Note: OBJ_DEAD 343 * objects can be referenced during final cleaning. 344 */ 345void 346vm_object_reference(vm_object_t object) 347{ 348 struct vnode *vp; 349 int flags; 350 351 if (object == NULL) 352 return; 353 VM_OBJECT_LOCK(object); 354 object->ref_count++; 355 if (object->type == OBJT_VNODE) { 356 vp = object->handle; 357 VI_LOCK(vp); 358 VM_OBJECT_UNLOCK(object); 359 for (flags = LK_INTERLOCK; vget(vp, flags, curthread); 360 flags = 0) 361 printf("vm_object_reference: delay in vget\n"); 362 } else 363 VM_OBJECT_UNLOCK(object); 364} 365 366/* 367 * vm_object_reference_locked: 368 * 369 * Gets another reference to the given object. 370 * 371 * The object must be locked. 372 */ 373void 374vm_object_reference_locked(vm_object_t object) 375{ 376 struct vnode *vp; 377 378 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 379 KASSERT((object->flags & OBJ_DEAD) == 0, 380 ("vm_object_reference_locked: dead object referenced")); 381 object->ref_count++; 382 if (object->type == OBJT_VNODE) { 383 vp = object->handle; 384 vref(vp); 385 } 386} 387 388/* 389 * Handle deallocating an object of type OBJT_VNODE. 390 */ 391void 392vm_object_vndeallocate(vm_object_t object) 393{ 394 struct vnode *vp = (struct vnode *) object->handle; 395 396 GIANT_REQUIRED; 397 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 398 KASSERT(object->type == OBJT_VNODE, 399 ("vm_object_vndeallocate: not a vnode object")); 400 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 401#ifdef INVARIANTS 402 if (object->ref_count == 0) { 403 vprint("vm_object_vndeallocate", vp); 404 panic("vm_object_vndeallocate: bad object reference count"); 405 } 406#endif 407 408 object->ref_count--; 409 if (object->ref_count == 0) { 410 mp_fixme("Unlocked vflag access."); 411 vp->v_vflag &= ~VV_TEXT; 412 } 413 VM_OBJECT_UNLOCK(object); 414 /* 415 * vrele may need a vop lock 416 */ 417 vrele(vp); 418} 419 420/* 421 * vm_object_deallocate: 422 * 423 * Release a reference to the specified object, 424 * gained either through a vm_object_allocate 425 * or a vm_object_reference call. When all references 426 * are gone, storage associated with this object 427 * may be relinquished. 428 * 429 * No object may be locked. 430 */ 431void 432vm_object_deallocate(vm_object_t object) 433{ 434 vm_object_t temp; 435 436 while (object != NULL) { 437 /* 438 * In general, the object should be locked when working with 439 * its type. In this case, in order to maintain proper lock 440 * ordering, an exception is possible because a vnode-backed 441 * object never changes its type. 442 */ 443 if (object->type == OBJT_VNODE) 444 mtx_lock(&Giant); 445 VM_OBJECT_LOCK(object); 446 if (object->type == OBJT_VNODE) { 447 vm_object_vndeallocate(object); 448 mtx_unlock(&Giant); 449 return; 450 } 451 452 KASSERT(object->ref_count != 0, 453 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 454 455 /* 456 * If the reference count goes to 0 we start calling 457 * vm_object_terminate() on the object chain. 458 * A ref count of 1 may be a special case depending on the 459 * shadow count being 0 or 1. 460 */ 461 object->ref_count--; 462 if (object->ref_count > 1) { 463 VM_OBJECT_UNLOCK(object); 464 return; 465 } else if (object->ref_count == 1) { 466 if (object->shadow_count == 0) { 467 vm_object_set_flag(object, OBJ_ONEMAPPING); 468 } else if ((object->shadow_count == 1) && 469 (object->handle == NULL) && 470 (object->type == OBJT_DEFAULT || 471 object->type == OBJT_SWAP)) { 472 vm_object_t robject; 473 474 robject = LIST_FIRST(&object->shadow_head); 475 KASSERT(robject != NULL, 476 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 477 object->ref_count, 478 object->shadow_count)); 479 if (!VM_OBJECT_TRYLOCK(robject)) { 480 /* 481 * Avoid a potential deadlock. 482 */ 483 object->ref_count++; 484 VM_OBJECT_UNLOCK(object); 485 /* 486 * More likely than not the thread 487 * holding robject's lock has lower 488 * priority than the current thread. 489 * Let the lower priority thread run. 490 */ 491 tsleep(&proc0, PVM, "vmo_de", 1); 492 continue; 493 } 494 if ((robject->handle == NULL) && 495 (robject->type == OBJT_DEFAULT || 496 robject->type == OBJT_SWAP)) { 497 498 robject->ref_count++; 499retry: 500 if (robject->paging_in_progress) { 501 VM_OBJECT_UNLOCK(object); 502 vm_object_pip_wait(robject, 503 "objde1"); 504 VM_OBJECT_LOCK(object); 505 goto retry; 506 } else if (object->paging_in_progress) { 507 VM_OBJECT_UNLOCK(robject); 508 object->flags |= OBJ_PIPWNT; 509 msleep(object, 510 VM_OBJECT_MTX(object), 511 PDROP | PVM, "objde2", 0); 512 VM_OBJECT_LOCK(robject); 513 VM_OBJECT_LOCK(object); 514 goto retry; 515 } 516 VM_OBJECT_UNLOCK(object); 517 if (robject->ref_count == 1) { 518 robject->ref_count--; 519 object = robject; 520 goto doterm; 521 } 522 object = robject; 523 vm_object_collapse(object); 524 VM_OBJECT_UNLOCK(object); 525 continue; 526 } 527 VM_OBJECT_UNLOCK(robject); 528 } 529 VM_OBJECT_UNLOCK(object); 530 return; 531 } 532doterm: 533 temp = object->backing_object; 534 if (temp != NULL) { 535 VM_OBJECT_LOCK(temp); 536 LIST_REMOVE(object, shadow_list); 537 temp->shadow_count--; 538 temp->generation++; 539 VM_OBJECT_UNLOCK(temp); 540 object->backing_object = NULL; 541 } 542 /* 543 * Don't double-terminate, we could be in a termination 544 * recursion due to the terminate having to sync data 545 * to disk. 546 */ 547 if ((object->flags & OBJ_DEAD) == 0) 548 vm_object_terminate(object); 549 else 550 VM_OBJECT_UNLOCK(object); 551 object = temp; 552 } 553} 554 555/* 556 * vm_object_terminate actually destroys the specified object, freeing 557 * up all previously used resources. 558 * 559 * The object must be locked. 560 * This routine may block. 561 */ 562void 563vm_object_terminate(vm_object_t object) 564{ 565 vm_page_t p; 566 567 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 568 569 /* 570 * Make sure no one uses us. 571 */ 572 vm_object_set_flag(object, OBJ_DEAD); 573 574 /* 575 * wait for the pageout daemon to be done with the object 576 */ 577 vm_object_pip_wait(object, "objtrm"); 578 579 KASSERT(!object->paging_in_progress, 580 ("vm_object_terminate: pageout in progress")); 581 582 /* 583 * Clean and free the pages, as appropriate. All references to the 584 * object are gone, so we don't need to lock it. 585 */ 586 if (object->type == OBJT_VNODE) { 587 struct vnode *vp = (struct vnode *)object->handle; 588 589 /* 590 * Clean pages and flush buffers. 591 */ 592 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 593 VM_OBJECT_UNLOCK(object); 594 595 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 596 597 VM_OBJECT_LOCK(object); 598 } 599 600 KASSERT(object->ref_count == 0, 601 ("vm_object_terminate: object with references, ref_count=%d", 602 object->ref_count)); 603 604 /* 605 * Now free any remaining pages. For internal objects, this also 606 * removes them from paging queues. Don't free wired pages, just 607 * remove them from the object. 608 */ 609 vm_page_lock_queues(); 610 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 611 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 612 ("vm_object_terminate: freeing busy page %p " 613 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 614 if (p->wire_count == 0) { 615 vm_page_free(p); 616 cnt.v_pfree++; 617 } else { 618 vm_page_remove(p); 619 } 620 } 621 vm_page_unlock_queues(); 622 623 /* 624 * Let the pager know object is dead. 625 */ 626 vm_pager_deallocate(object); 627 VM_OBJECT_UNLOCK(object); 628 629 /* 630 * Remove the object from the global object list. 631 */ 632 mtx_lock(&vm_object_list_mtx); 633 TAILQ_REMOVE(&vm_object_list, object, object_list); 634 mtx_unlock(&vm_object_list_mtx); 635 636 /* 637 * Free the space for the object. 638 */ 639 uma_zfree(obj_zone, object); 640} 641 642/* 643 * vm_object_page_clean 644 * 645 * Clean all dirty pages in the specified range of object. Leaves page 646 * on whatever queue it is currently on. If NOSYNC is set then do not 647 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 648 * leaving the object dirty. 649 * 650 * When stuffing pages asynchronously, allow clustering. XXX we need a 651 * synchronous clustering mode implementation. 652 * 653 * Odd semantics: if start == end, we clean everything. 654 * 655 * The object must be locked. 656 */ 657void 658vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 659{ 660 vm_page_t p, np; 661 vm_pindex_t tstart, tend; 662 vm_pindex_t pi; 663 int clearobjflags; 664 int pagerflags; 665 int curgeneration; 666 667 GIANT_REQUIRED; 668 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 669 if (object->type != OBJT_VNODE || 670 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 671 return; 672 673 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 674 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 675 676 vm_object_set_flag(object, OBJ_CLEANING); 677 678 tstart = start; 679 if (end == 0) { 680 tend = object->size; 681 } else { 682 tend = end; 683 } 684 685 vm_page_lock_queues(); 686 /* 687 * If the caller is smart and only msync()s a range he knows is 688 * dirty, we may be able to avoid an object scan. This results in 689 * a phenominal improvement in performance. We cannot do this 690 * as a matter of course because the object may be huge - e.g. 691 * the size might be in the gigabytes or terrabytes. 692 */ 693 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 694 vm_pindex_t tscan; 695 int scanlimit; 696 int scanreset; 697 698 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 699 if (scanreset < 16) 700 scanreset = 16; 701 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 702 703 scanlimit = scanreset; 704 tscan = tstart; 705 while (tscan < tend) { 706 curgeneration = object->generation; 707 p = vm_page_lookup(object, tscan); 708 if (p == NULL || p->valid == 0 || 709 (p->queue - p->pc) == PQ_CACHE) { 710 if (--scanlimit == 0) 711 break; 712 ++tscan; 713 continue; 714 } 715 vm_page_test_dirty(p); 716 if ((p->dirty & p->valid) == 0) { 717 if (--scanlimit == 0) 718 break; 719 ++tscan; 720 continue; 721 } 722 /* 723 * If we have been asked to skip nosync pages and 724 * this is a nosync page, we can't continue. 725 */ 726 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 727 if (--scanlimit == 0) 728 break; 729 ++tscan; 730 continue; 731 } 732 scanlimit = scanreset; 733 734 /* 735 * This returns 0 if it was unable to busy the first 736 * page (i.e. had to sleep). 737 */ 738 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 739 } 740 741 /* 742 * If everything was dirty and we flushed it successfully, 743 * and the requested range is not the entire object, we 744 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 745 * return immediately. 746 */ 747 if (tscan >= tend && (tstart || tend < object->size)) { 748 vm_page_unlock_queues(); 749 vm_object_clear_flag(object, OBJ_CLEANING); 750 return; 751 } 752 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 753 } 754 755 /* 756 * Generally set CLEANCHK interlock and make the page read-only so 757 * we can then clear the object flags. 758 * 759 * However, if this is a nosync mmap then the object is likely to 760 * stay dirty so do not mess with the page and do not clear the 761 * object flags. 762 */ 763 clearobjflags = 1; 764 TAILQ_FOREACH(p, &object->memq, listq) { 765 vm_page_flag_set(p, PG_CLEANCHK); 766 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 767 clearobjflags = 0; 768 else 769 pmap_page_protect(p, VM_PROT_READ); 770 } 771 772 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 773 struct vnode *vp; 774 775 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 776 if (object->type == OBJT_VNODE && 777 (vp = (struct vnode *)object->handle) != NULL) { 778 VI_LOCK(vp); 779 if (vp->v_iflag & VI_OBJDIRTY) 780 vp->v_iflag &= ~VI_OBJDIRTY; 781 VI_UNLOCK(vp); 782 } 783 } 784 785rescan: 786 curgeneration = object->generation; 787 788 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 789 int n; 790 791 np = TAILQ_NEXT(p, listq); 792 793again: 794 pi = p->pindex; 795 if (((p->flags & PG_CLEANCHK) == 0) || 796 (pi < tstart) || (pi >= tend) || 797 (p->valid == 0) || 798 ((p->queue - p->pc) == PQ_CACHE)) { 799 vm_page_flag_clear(p, PG_CLEANCHK); 800 continue; 801 } 802 803 vm_page_test_dirty(p); 804 if ((p->dirty & p->valid) == 0) { 805 vm_page_flag_clear(p, PG_CLEANCHK); 806 continue; 807 } 808 809 /* 810 * If we have been asked to skip nosync pages and this is a 811 * nosync page, skip it. Note that the object flags were 812 * not cleared in this case so we do not have to set them. 813 */ 814 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 815 vm_page_flag_clear(p, PG_CLEANCHK); 816 continue; 817 } 818 819 n = vm_object_page_collect_flush(object, p, 820 curgeneration, pagerflags); 821 if (n == 0) 822 goto rescan; 823 824 if (object->generation != curgeneration) 825 goto rescan; 826 827 /* 828 * Try to optimize the next page. If we can't we pick up 829 * our (random) scan where we left off. 830 */ 831 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 832 if ((p = vm_page_lookup(object, pi + n)) != NULL) 833 goto again; 834 } 835 } 836 vm_page_unlock_queues(); 837#if 0 838 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 839#endif 840 841 vm_object_clear_flag(object, OBJ_CLEANING); 842 return; 843} 844 845static int 846vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 847{ 848 int runlen; 849 int maxf; 850 int chkb; 851 int maxb; 852 int i; 853 vm_pindex_t pi; 854 vm_page_t maf[vm_pageout_page_count]; 855 vm_page_t mab[vm_pageout_page_count]; 856 vm_page_t ma[vm_pageout_page_count]; 857 858 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 859 pi = p->pindex; 860 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 861 vm_page_lock_queues(); 862 if (object->generation != curgeneration) { 863 return(0); 864 } 865 } 866 maxf = 0; 867 for(i = 1; i < vm_pageout_page_count; i++) { 868 vm_page_t tp; 869 870 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 871 if ((tp->flags & PG_BUSY) || 872 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 873 (tp->flags & PG_CLEANCHK) == 0) || 874 (tp->busy != 0)) 875 break; 876 if((tp->queue - tp->pc) == PQ_CACHE) { 877 vm_page_flag_clear(tp, PG_CLEANCHK); 878 break; 879 } 880 vm_page_test_dirty(tp); 881 if ((tp->dirty & tp->valid) == 0) { 882 vm_page_flag_clear(tp, PG_CLEANCHK); 883 break; 884 } 885 maf[ i - 1 ] = tp; 886 maxf++; 887 continue; 888 } 889 break; 890 } 891 892 maxb = 0; 893 chkb = vm_pageout_page_count - maxf; 894 if (chkb) { 895 for(i = 1; i < chkb;i++) { 896 vm_page_t tp; 897 898 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 899 if ((tp->flags & PG_BUSY) || 900 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 901 (tp->flags & PG_CLEANCHK) == 0) || 902 (tp->busy != 0)) 903 break; 904 if ((tp->queue - tp->pc) == PQ_CACHE) { 905 vm_page_flag_clear(tp, PG_CLEANCHK); 906 break; 907 } 908 vm_page_test_dirty(tp); 909 if ((tp->dirty & tp->valid) == 0) { 910 vm_page_flag_clear(tp, PG_CLEANCHK); 911 break; 912 } 913 mab[ i - 1 ] = tp; 914 maxb++; 915 continue; 916 } 917 break; 918 } 919 } 920 921 for(i = 0; i < maxb; i++) { 922 int index = (maxb - i) - 1; 923 ma[index] = mab[i]; 924 vm_page_flag_clear(ma[index], PG_CLEANCHK); 925 } 926 vm_page_flag_clear(p, PG_CLEANCHK); 927 ma[maxb] = p; 928 for(i = 0; i < maxf; i++) { 929 int index = (maxb + i) + 1; 930 ma[index] = maf[i]; 931 vm_page_flag_clear(ma[index], PG_CLEANCHK); 932 } 933 runlen = maxb + maxf + 1; 934 935 vm_pageout_flush(ma, runlen, pagerflags); 936 for (i = 0; i < runlen; i++) { 937 if (ma[i]->valid & ma[i]->dirty) { 938 pmap_page_protect(ma[i], VM_PROT_READ); 939 vm_page_flag_set(ma[i], PG_CLEANCHK); 940 941 /* 942 * maxf will end up being the actual number of pages 943 * we wrote out contiguously, non-inclusive of the 944 * first page. We do not count look-behind pages. 945 */ 946 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 947 maxf = i - maxb - 1; 948 } 949 } 950 return(maxf + 1); 951} 952 953/* 954 * Note that there is absolutely no sense in writing out 955 * anonymous objects, so we track down the vnode object 956 * to write out. 957 * We invalidate (remove) all pages from the address space 958 * for semantic correctness. 959 * 960 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 961 * may start out with a NULL object. 962 */ 963void 964vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 965 boolean_t syncio, boolean_t invalidate) 966{ 967 vm_object_t backing_object; 968 struct vnode *vp; 969 int flags; 970 971 if (object == NULL) 972 return; 973 VM_OBJECT_LOCK(object); 974 while ((backing_object = object->backing_object) != NULL) { 975 VM_OBJECT_LOCK(backing_object); 976 offset += object->backing_object_offset; 977 VM_OBJECT_UNLOCK(object); 978 object = backing_object; 979 if (object->size < OFF_TO_IDX(offset + size)) 980 size = IDX_TO_OFF(object->size) - offset; 981 } 982 /* 983 * Flush pages if writing is allowed, invalidate them 984 * if invalidation requested. Pages undergoing I/O 985 * will be ignored by vm_object_page_remove(). 986 * 987 * We cannot lock the vnode and then wait for paging 988 * to complete without deadlocking against vm_fault. 989 * Instead we simply call vm_object_page_remove() and 990 * allow it to block internally on a page-by-page 991 * basis when it encounters pages undergoing async 992 * I/O. 993 */ 994 if (object->type == OBJT_VNODE && 995 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 996 vp = object->handle; 997 VM_OBJECT_UNLOCK(object); 998 mtx_lock(&Giant); 999 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); 1000 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1001 flags |= invalidate ? OBJPC_INVAL : 0; 1002 VM_OBJECT_LOCK(object); 1003 vm_object_page_clean(object, 1004 OFF_TO_IDX(offset), 1005 OFF_TO_IDX(offset + size + PAGE_MASK), 1006 flags); 1007 VM_OBJECT_UNLOCK(object); 1008 VOP_UNLOCK(vp, 0, curthread); 1009 mtx_unlock(&Giant); 1010 VM_OBJECT_LOCK(object); 1011 } 1012 if ((object->type == OBJT_VNODE || 1013 object->type == OBJT_DEVICE) && invalidate) { 1014 boolean_t purge; 1015 purge = old_msync || (object->type == OBJT_DEVICE); 1016 vm_object_page_remove(object, 1017 OFF_TO_IDX(offset), 1018 OFF_TO_IDX(offset + size + PAGE_MASK), 1019 purge ? FALSE : TRUE); 1020 } 1021 VM_OBJECT_UNLOCK(object); 1022} 1023 1024/* 1025 * vm_object_madvise: 1026 * 1027 * Implements the madvise function at the object/page level. 1028 * 1029 * MADV_WILLNEED (any object) 1030 * 1031 * Activate the specified pages if they are resident. 1032 * 1033 * MADV_DONTNEED (any object) 1034 * 1035 * Deactivate the specified pages if they are resident. 1036 * 1037 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1038 * OBJ_ONEMAPPING only) 1039 * 1040 * Deactivate and clean the specified pages if they are 1041 * resident. This permits the process to reuse the pages 1042 * without faulting or the kernel to reclaim the pages 1043 * without I/O. 1044 */ 1045void 1046vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1047{ 1048 vm_pindex_t end, tpindex; 1049 vm_object_t backing_object, tobject; 1050 vm_page_t m; 1051 1052 if (object == NULL) 1053 return; 1054 VM_OBJECT_LOCK(object); 1055 end = pindex + count; 1056 /* 1057 * Locate and adjust resident pages 1058 */ 1059 for (; pindex < end; pindex += 1) { 1060relookup: 1061 tobject = object; 1062 tpindex = pindex; 1063shadowlookup: 1064 /* 1065 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1066 * and those pages must be OBJ_ONEMAPPING. 1067 */ 1068 if (advise == MADV_FREE) { 1069 if ((tobject->type != OBJT_DEFAULT && 1070 tobject->type != OBJT_SWAP) || 1071 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1072 goto unlock_tobject; 1073 } 1074 } 1075 m = vm_page_lookup(tobject, tpindex); 1076 if (m == NULL) { 1077 /* 1078 * There may be swap even if there is no backing page 1079 */ 1080 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1081 swap_pager_freespace(tobject, tpindex, 1); 1082 /* 1083 * next object 1084 */ 1085 backing_object = tobject->backing_object; 1086 if (backing_object == NULL) 1087 goto unlock_tobject; 1088 VM_OBJECT_LOCK(backing_object); 1089 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1090 if (tobject != object) 1091 VM_OBJECT_UNLOCK(tobject); 1092 tobject = backing_object; 1093 goto shadowlookup; 1094 } 1095 /* 1096 * If the page is busy or not in a normal active state, 1097 * we skip it. If the page is not managed there are no 1098 * page queues to mess with. Things can break if we mess 1099 * with pages in any of the below states. 1100 */ 1101 vm_page_lock_queues(); 1102 if (m->hold_count || 1103 m->wire_count || 1104 (m->flags & PG_UNMANAGED) || 1105 m->valid != VM_PAGE_BITS_ALL) { 1106 vm_page_unlock_queues(); 1107 goto unlock_tobject; 1108 } 1109 if ((m->flags & PG_BUSY) || m->busy) { 1110 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1111 if (object != tobject) 1112 VM_OBJECT_UNLOCK(object); 1113 VM_OBJECT_UNLOCK(tobject); 1114 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "madvpo", 0); 1115 VM_OBJECT_LOCK(object); 1116 goto relookup; 1117 } 1118 if (advise == MADV_WILLNEED) { 1119 vm_page_activate(m); 1120 } else if (advise == MADV_DONTNEED) { 1121 vm_page_dontneed(m); 1122 } else if (advise == MADV_FREE) { 1123 /* 1124 * Mark the page clean. This will allow the page 1125 * to be freed up by the system. However, such pages 1126 * are often reused quickly by malloc()/free() 1127 * so we do not do anything that would cause 1128 * a page fault if we can help it. 1129 * 1130 * Specifically, we do not try to actually free 1131 * the page now nor do we try to put it in the 1132 * cache (which would cause a page fault on reuse). 1133 * 1134 * But we do make the page is freeable as we 1135 * can without actually taking the step of unmapping 1136 * it. 1137 */ 1138 pmap_clear_modify(m); 1139 m->dirty = 0; 1140 m->act_count = 0; 1141 vm_page_dontneed(m); 1142 } 1143 vm_page_unlock_queues(); 1144 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1145 swap_pager_freespace(tobject, tpindex, 1); 1146unlock_tobject: 1147 if (tobject != object) 1148 VM_OBJECT_UNLOCK(tobject); 1149 } 1150 VM_OBJECT_UNLOCK(object); 1151} 1152 1153/* 1154 * vm_object_shadow: 1155 * 1156 * Create a new object which is backed by the 1157 * specified existing object range. The source 1158 * object reference is deallocated. 1159 * 1160 * The new object and offset into that object 1161 * are returned in the source parameters. 1162 */ 1163void 1164vm_object_shadow( 1165 vm_object_t *object, /* IN/OUT */ 1166 vm_ooffset_t *offset, /* IN/OUT */ 1167 vm_size_t length) 1168{ 1169 vm_object_t source; 1170 vm_object_t result; 1171 1172 source = *object; 1173 1174 /* 1175 * Don't create the new object if the old object isn't shared. 1176 */ 1177 if (source != NULL) { 1178 VM_OBJECT_LOCK(source); 1179 if (source->ref_count == 1 && 1180 source->handle == NULL && 1181 (source->type == OBJT_DEFAULT || 1182 source->type == OBJT_SWAP)) { 1183 VM_OBJECT_UNLOCK(source); 1184 return; 1185 } 1186 VM_OBJECT_UNLOCK(source); 1187 } 1188 1189 /* 1190 * Allocate a new object with the given length. 1191 */ 1192 result = vm_object_allocate(OBJT_DEFAULT, length); 1193 1194 /* 1195 * The new object shadows the source object, adding a reference to it. 1196 * Our caller changes his reference to point to the new object, 1197 * removing a reference to the source object. Net result: no change 1198 * of reference count. 1199 * 1200 * Try to optimize the result object's page color when shadowing 1201 * in order to maintain page coloring consistency in the combined 1202 * shadowed object. 1203 */ 1204 result->backing_object = source; 1205 /* 1206 * Store the offset into the source object, and fix up the offset into 1207 * the new object. 1208 */ 1209 result->backing_object_offset = *offset; 1210 if (source != NULL) { 1211 VM_OBJECT_LOCK(source); 1212 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1213 source->shadow_count++; 1214 source->generation++; 1215 if (length < source->size) 1216 length = source->size; 1217 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1218 source->generation > 1) 1219 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1220 result->pg_color = (source->pg_color + 1221 length * source->generation) & PQ_L2_MASK; 1222 VM_OBJECT_UNLOCK(source); 1223 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1224 PQ_L2_MASK; 1225 } 1226 1227 1228 /* 1229 * Return the new things 1230 */ 1231 *offset = 0; 1232 *object = result; 1233} 1234 1235/* 1236 * vm_object_split: 1237 * 1238 * Split the pages in a map entry into a new object. This affords 1239 * easier removal of unused pages, and keeps object inheritance from 1240 * being a negative impact on memory usage. 1241 */ 1242void 1243vm_object_split(vm_map_entry_t entry) 1244{ 1245 vm_page_t m; 1246 vm_object_t orig_object, new_object, source; 1247 vm_pindex_t offidxstart, offidxend; 1248 vm_size_t idx, size; 1249 1250 orig_object = entry->object.vm_object; 1251 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1252 return; 1253 if (orig_object->ref_count <= 1) 1254 return; 1255 VM_OBJECT_UNLOCK(orig_object); 1256 1257 offidxstart = OFF_TO_IDX(entry->offset); 1258 offidxend = offidxstart + OFF_TO_IDX(entry->end - entry->start); 1259 size = offidxend - offidxstart; 1260 1261 /* 1262 * If swap_pager_copy() is later called, it will convert new_object 1263 * into a swap object. 1264 */ 1265 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1266 1267 VM_OBJECT_LOCK(new_object); 1268 VM_OBJECT_LOCK(orig_object); 1269 source = orig_object->backing_object; 1270 if (source != NULL) { 1271 VM_OBJECT_LOCK(source); 1272 LIST_INSERT_HEAD(&source->shadow_head, 1273 new_object, shadow_list); 1274 source->shadow_count++; 1275 source->generation++; 1276 vm_object_reference_locked(source); /* for new_object */ 1277 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1278 VM_OBJECT_UNLOCK(source); 1279 new_object->backing_object_offset = 1280 orig_object->backing_object_offset + entry->offset; 1281 new_object->backing_object = source; 1282 } 1283 for (idx = 0; idx < size; idx++) { 1284 retry: 1285 m = vm_page_lookup(orig_object, offidxstart + idx); 1286 if (m == NULL) 1287 continue; 1288 1289 /* 1290 * We must wait for pending I/O to complete before we can 1291 * rename the page. 1292 * 1293 * We do not have to VM_PROT_NONE the page as mappings should 1294 * not be changed by this operation. 1295 */ 1296 vm_page_lock_queues(); 1297 if ((m->flags & PG_BUSY) || m->busy) { 1298 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1299 VM_OBJECT_UNLOCK(orig_object); 1300 VM_OBJECT_UNLOCK(new_object); 1301 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "spltwt", 0); 1302 VM_OBJECT_LOCK(new_object); 1303 VM_OBJECT_LOCK(orig_object); 1304 goto retry; 1305 } 1306 vm_page_rename(m, new_object, idx); 1307 /* page automatically made dirty by rename and cache handled */ 1308 vm_page_busy(m); 1309 vm_page_unlock_queues(); 1310 } 1311 if (orig_object->type == OBJT_SWAP) { 1312 /* 1313 * swap_pager_copy() can sleep, in which case the orig_object's 1314 * and new_object's locks are released and reacquired. 1315 */ 1316 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1317 } 1318 VM_OBJECT_UNLOCK(orig_object); 1319 vm_page_lock_queues(); 1320 TAILQ_FOREACH(m, &new_object->memq, listq) 1321 vm_page_wakeup(m); 1322 vm_page_unlock_queues(); 1323 VM_OBJECT_UNLOCK(new_object); 1324 entry->object.vm_object = new_object; 1325 entry->offset = 0LL; 1326 vm_object_deallocate(orig_object); 1327 VM_OBJECT_LOCK(new_object); 1328} 1329 1330#define OBSC_TEST_ALL_SHADOWED 0x0001 1331#define OBSC_COLLAPSE_NOWAIT 0x0002 1332#define OBSC_COLLAPSE_WAIT 0x0004 1333 1334static int 1335vm_object_backing_scan(vm_object_t object, int op) 1336{ 1337 int r = 1; 1338 vm_page_t p; 1339 vm_object_t backing_object; 1340 vm_pindex_t backing_offset_index; 1341 1342 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1343 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); 1344 1345 backing_object = object->backing_object; 1346 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1347 1348 /* 1349 * Initial conditions 1350 */ 1351 if (op & OBSC_TEST_ALL_SHADOWED) { 1352 /* 1353 * We do not want to have to test for the existence of 1354 * swap pages in the backing object. XXX but with the 1355 * new swapper this would be pretty easy to do. 1356 * 1357 * XXX what about anonymous MAP_SHARED memory that hasn't 1358 * been ZFOD faulted yet? If we do not test for this, the 1359 * shadow test may succeed! XXX 1360 */ 1361 if (backing_object->type != OBJT_DEFAULT) { 1362 return (0); 1363 } 1364 } 1365 if (op & OBSC_COLLAPSE_WAIT) { 1366 vm_object_set_flag(backing_object, OBJ_DEAD); 1367 } 1368 1369 /* 1370 * Our scan 1371 */ 1372 p = TAILQ_FIRST(&backing_object->memq); 1373 while (p) { 1374 vm_page_t next = TAILQ_NEXT(p, listq); 1375 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1376 1377 if (op & OBSC_TEST_ALL_SHADOWED) { 1378 vm_page_t pp; 1379 1380 /* 1381 * Ignore pages outside the parent object's range 1382 * and outside the parent object's mapping of the 1383 * backing object. 1384 * 1385 * note that we do not busy the backing object's 1386 * page. 1387 */ 1388 if ( 1389 p->pindex < backing_offset_index || 1390 new_pindex >= object->size 1391 ) { 1392 p = next; 1393 continue; 1394 } 1395 1396 /* 1397 * See if the parent has the page or if the parent's 1398 * object pager has the page. If the parent has the 1399 * page but the page is not valid, the parent's 1400 * object pager must have the page. 1401 * 1402 * If this fails, the parent does not completely shadow 1403 * the object and we might as well give up now. 1404 */ 1405 1406 pp = vm_page_lookup(object, new_pindex); 1407 if ( 1408 (pp == NULL || pp->valid == 0) && 1409 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1410 ) { 1411 r = 0; 1412 break; 1413 } 1414 } 1415 1416 /* 1417 * Check for busy page 1418 */ 1419 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1420 vm_page_t pp; 1421 1422 vm_page_lock_queues(); 1423 if (op & OBSC_COLLAPSE_NOWAIT) { 1424 if ((p->flags & PG_BUSY) || 1425 !p->valid || 1426 p->hold_count || 1427 p->wire_count || 1428 p->busy) { 1429 vm_page_unlock_queues(); 1430 p = next; 1431 continue; 1432 } 1433 } else if (op & OBSC_COLLAPSE_WAIT) { 1434 if ((p->flags & PG_BUSY) || p->busy) { 1435 vm_page_flag_set(p, 1436 PG_WANTED | PG_REFERENCED); 1437 VM_OBJECT_UNLOCK(backing_object); 1438 VM_OBJECT_UNLOCK(object); 1439 msleep(p, &vm_page_queue_mtx, 1440 PDROP | PVM, "vmocol", 0); 1441 VM_OBJECT_LOCK(object); 1442 VM_OBJECT_LOCK(backing_object); 1443 /* 1444 * If we slept, anything could have 1445 * happened. Since the object is 1446 * marked dead, the backing offset 1447 * should not have changed so we 1448 * just restart our scan. 1449 */ 1450 p = TAILQ_FIRST(&backing_object->memq); 1451 continue; 1452 } 1453 } 1454 vm_page_unlock_queues(); 1455 1456 KASSERT( 1457 p->object == backing_object, 1458 ("vm_object_qcollapse(): object mismatch") 1459 ); 1460 1461 /* 1462 * Destroy any associated swap 1463 */ 1464 if (backing_object->type == OBJT_SWAP) { 1465 swap_pager_freespace( 1466 backing_object, 1467 p->pindex, 1468 1 1469 ); 1470 } 1471 1472 if ( 1473 p->pindex < backing_offset_index || 1474 new_pindex >= object->size 1475 ) { 1476 /* 1477 * Page is out of the parent object's range, we 1478 * can simply destroy it. 1479 */ 1480 vm_page_lock_queues(); 1481 pmap_remove_all(p); 1482 vm_page_free(p); 1483 vm_page_unlock_queues(); 1484 p = next; 1485 continue; 1486 } 1487 1488 pp = vm_page_lookup(object, new_pindex); 1489 if ( 1490 pp != NULL || 1491 vm_pager_has_page(object, new_pindex, NULL, NULL) 1492 ) { 1493 /* 1494 * page already exists in parent OR swap exists 1495 * for this location in the parent. Destroy 1496 * the original page from the backing object. 1497 * 1498 * Leave the parent's page alone 1499 */ 1500 vm_page_lock_queues(); 1501 pmap_remove_all(p); 1502 vm_page_free(p); 1503 vm_page_unlock_queues(); 1504 p = next; 1505 continue; 1506 } 1507 1508 /* 1509 * Page does not exist in parent, rename the 1510 * page from the backing object to the main object. 1511 * 1512 * If the page was mapped to a process, it can remain 1513 * mapped through the rename. 1514 */ 1515 vm_page_lock_queues(); 1516 vm_page_rename(p, object, new_pindex); 1517 vm_page_unlock_queues(); 1518 /* page automatically made dirty by rename */ 1519 } 1520 p = next; 1521 } 1522 return (r); 1523} 1524 1525 1526/* 1527 * this version of collapse allows the operation to occur earlier and 1528 * when paging_in_progress is true for an object... This is not a complete 1529 * operation, but should plug 99.9% of the rest of the leaks. 1530 */ 1531static void 1532vm_object_qcollapse(vm_object_t object) 1533{ 1534 vm_object_t backing_object = object->backing_object; 1535 1536 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1537 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1538 1539 if (backing_object->ref_count != 1) 1540 return; 1541 1542 backing_object->ref_count += 2; 1543 1544 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1545 1546 backing_object->ref_count -= 2; 1547} 1548 1549/* 1550 * vm_object_collapse: 1551 * 1552 * Collapse an object with the object backing it. 1553 * Pages in the backing object are moved into the 1554 * parent, and the backing object is deallocated. 1555 */ 1556void 1557vm_object_collapse(vm_object_t object) 1558{ 1559 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1560 1561 while (TRUE) { 1562 vm_object_t backing_object; 1563 1564 /* 1565 * Verify that the conditions are right for collapse: 1566 * 1567 * The object exists and the backing object exists. 1568 */ 1569 if ((backing_object = object->backing_object) == NULL) 1570 break; 1571 1572 /* 1573 * we check the backing object first, because it is most likely 1574 * not collapsable. 1575 */ 1576 VM_OBJECT_LOCK(backing_object); 1577 if (backing_object->handle != NULL || 1578 (backing_object->type != OBJT_DEFAULT && 1579 backing_object->type != OBJT_SWAP) || 1580 (backing_object->flags & OBJ_DEAD) || 1581 object->handle != NULL || 1582 (object->type != OBJT_DEFAULT && 1583 object->type != OBJT_SWAP) || 1584 (object->flags & OBJ_DEAD)) { 1585 VM_OBJECT_UNLOCK(backing_object); 1586 break; 1587 } 1588 1589 if ( 1590 object->paging_in_progress != 0 || 1591 backing_object->paging_in_progress != 0 1592 ) { 1593 vm_object_qcollapse(object); 1594 VM_OBJECT_UNLOCK(backing_object); 1595 break; 1596 } 1597 /* 1598 * We know that we can either collapse the backing object (if 1599 * the parent is the only reference to it) or (perhaps) have 1600 * the parent bypass the object if the parent happens to shadow 1601 * all the resident pages in the entire backing object. 1602 * 1603 * This is ignoring pager-backed pages such as swap pages. 1604 * vm_object_backing_scan fails the shadowing test in this 1605 * case. 1606 */ 1607 if (backing_object->ref_count == 1) { 1608 /* 1609 * If there is exactly one reference to the backing 1610 * object, we can collapse it into the parent. 1611 */ 1612 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1613 1614 /* 1615 * Move the pager from backing_object to object. 1616 */ 1617 if (backing_object->type == OBJT_SWAP) { 1618 /* 1619 * swap_pager_copy() can sleep, in which case 1620 * the backing_object's and object's locks are 1621 * released and reacquired. 1622 */ 1623 swap_pager_copy( 1624 backing_object, 1625 object, 1626 OFF_TO_IDX(object->backing_object_offset), TRUE); 1627 } 1628 /* 1629 * Object now shadows whatever backing_object did. 1630 * Note that the reference to 1631 * backing_object->backing_object moves from within 1632 * backing_object to within object. 1633 */ 1634 LIST_REMOVE(object, shadow_list); 1635 backing_object->shadow_count--; 1636 backing_object->generation++; 1637 if (backing_object->backing_object) { 1638 VM_OBJECT_LOCK(backing_object->backing_object); 1639 LIST_REMOVE(backing_object, shadow_list); 1640 LIST_INSERT_HEAD( 1641 &backing_object->backing_object->shadow_head, 1642 object, shadow_list); 1643 /* 1644 * The shadow_count has not changed. 1645 */ 1646 backing_object->backing_object->generation++; 1647 VM_OBJECT_UNLOCK(backing_object->backing_object); 1648 } 1649 object->backing_object = backing_object->backing_object; 1650 object->backing_object_offset += 1651 backing_object->backing_object_offset; 1652 1653 /* 1654 * Discard backing_object. 1655 * 1656 * Since the backing object has no pages, no pager left, 1657 * and no object references within it, all that is 1658 * necessary is to dispose of it. 1659 */ 1660 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1661 VM_OBJECT_UNLOCK(backing_object); 1662 1663 mtx_lock(&vm_object_list_mtx); 1664 TAILQ_REMOVE( 1665 &vm_object_list, 1666 backing_object, 1667 object_list 1668 ); 1669 mtx_unlock(&vm_object_list_mtx); 1670 1671 uma_zfree(obj_zone, backing_object); 1672 1673 object_collapses++; 1674 } else { 1675 vm_object_t new_backing_object; 1676 1677 /* 1678 * If we do not entirely shadow the backing object, 1679 * there is nothing we can do so we give up. 1680 */ 1681 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1682 VM_OBJECT_UNLOCK(backing_object); 1683 break; 1684 } 1685 1686 /* 1687 * Make the parent shadow the next object in the 1688 * chain. Deallocating backing_object will not remove 1689 * it, since its reference count is at least 2. 1690 */ 1691 LIST_REMOVE(object, shadow_list); 1692 backing_object->shadow_count--; 1693 backing_object->generation++; 1694 1695 new_backing_object = backing_object->backing_object; 1696 if ((object->backing_object = new_backing_object) != NULL) { 1697 VM_OBJECT_LOCK(new_backing_object); 1698 LIST_INSERT_HEAD( 1699 &new_backing_object->shadow_head, 1700 object, 1701 shadow_list 1702 ); 1703 new_backing_object->shadow_count++; 1704 new_backing_object->generation++; 1705 vm_object_reference_locked(new_backing_object); 1706 VM_OBJECT_UNLOCK(new_backing_object); 1707 object->backing_object_offset += 1708 backing_object->backing_object_offset; 1709 } 1710 1711 /* 1712 * Drop the reference count on backing_object. Since 1713 * its ref_count was at least 2, it will not vanish. 1714 */ 1715 backing_object->ref_count--; 1716 VM_OBJECT_UNLOCK(backing_object); 1717 object_bypasses++; 1718 } 1719 1720 /* 1721 * Try again with this object's new backing object. 1722 */ 1723 } 1724} 1725 1726/* 1727 * vm_object_page_remove: 1728 * 1729 * Removes all physical pages in the given range from the 1730 * object's list of pages. If the range's end is zero, all 1731 * physical pages from the range's start to the end of the object 1732 * are deleted. 1733 * 1734 * The object must be locked. 1735 */ 1736void 1737vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1738 boolean_t clean_only) 1739{ 1740 vm_page_t p, next; 1741 1742 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1743 if (object->resident_page_count == 0) 1744 return; 1745 1746 /* 1747 * Since physically-backed objects do not use managed pages, we can't 1748 * remove pages from the object (we must instead remove the page 1749 * references, and then destroy the object). 1750 */ 1751 KASSERT(object->type != OBJT_PHYS, 1752 ("attempt to remove pages from a physical object")); 1753 1754 vm_object_pip_add(object, 1); 1755again: 1756 vm_page_lock_queues(); 1757 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1758 if (p->pindex < start) { 1759 p = vm_page_splay(start, object->root); 1760 if ((object->root = p)->pindex < start) 1761 p = TAILQ_NEXT(p, listq); 1762 } 1763 } 1764 /* 1765 * Assert: the variable p is either (1) the page with the 1766 * least pindex greater than or equal to the parameter pindex 1767 * or (2) NULL. 1768 */ 1769 for (; 1770 p != NULL && (p->pindex < end || end == 0); 1771 p = next) { 1772 next = TAILQ_NEXT(p, listq); 1773 1774 if (p->wire_count != 0) { 1775 pmap_remove_all(p); 1776 if (!clean_only) 1777 p->valid = 0; 1778 continue; 1779 } 1780 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1781 goto again; 1782 if (clean_only && p->valid) { 1783 pmap_page_protect(p, VM_PROT_READ | VM_PROT_EXECUTE); 1784 if (p->valid & p->dirty) 1785 continue; 1786 } 1787 pmap_remove_all(p); 1788 vm_page_free(p); 1789 } 1790 vm_page_unlock_queues(); 1791 vm_object_pip_wakeup(object); 1792} 1793 1794/* 1795 * Routine: vm_object_coalesce 1796 * Function: Coalesces two objects backing up adjoining 1797 * regions of memory into a single object. 1798 * 1799 * returns TRUE if objects were combined. 1800 * 1801 * NOTE: Only works at the moment if the second object is NULL - 1802 * if it's not, which object do we lock first? 1803 * 1804 * Parameters: 1805 * prev_object First object to coalesce 1806 * prev_offset Offset into prev_object 1807 * prev_size Size of reference to prev_object 1808 * next_size Size of reference to the second object 1809 * 1810 * Conditions: 1811 * The object must *not* be locked. 1812 */ 1813boolean_t 1814vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 1815 vm_size_t prev_size, vm_size_t next_size) 1816{ 1817 vm_pindex_t next_pindex; 1818 1819 if (prev_object == NULL) 1820 return (TRUE); 1821 VM_OBJECT_LOCK(prev_object); 1822 if (prev_object->type != OBJT_DEFAULT && 1823 prev_object->type != OBJT_SWAP) { 1824 VM_OBJECT_UNLOCK(prev_object); 1825 return (FALSE); 1826 } 1827 1828 /* 1829 * Try to collapse the object first 1830 */ 1831 vm_object_collapse(prev_object); 1832 1833 /* 1834 * Can't coalesce if: . more than one reference . paged out . shadows 1835 * another object . has a copy elsewhere (any of which mean that the 1836 * pages not mapped to prev_entry may be in use anyway) 1837 */ 1838 if (prev_object->backing_object != NULL) { 1839 VM_OBJECT_UNLOCK(prev_object); 1840 return (FALSE); 1841 } 1842 1843 prev_size >>= PAGE_SHIFT; 1844 next_size >>= PAGE_SHIFT; 1845 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 1846 1847 if ((prev_object->ref_count > 1) && 1848 (prev_object->size != next_pindex)) { 1849 VM_OBJECT_UNLOCK(prev_object); 1850 return (FALSE); 1851 } 1852 1853 /* 1854 * Remove any pages that may still be in the object from a previous 1855 * deallocation. 1856 */ 1857 if (next_pindex < prev_object->size) { 1858 vm_object_page_remove(prev_object, 1859 next_pindex, 1860 next_pindex + next_size, FALSE); 1861 if (prev_object->type == OBJT_SWAP) 1862 swap_pager_freespace(prev_object, 1863 next_pindex, next_size); 1864 } 1865 1866 /* 1867 * Extend the object if necessary. 1868 */ 1869 if (next_pindex + next_size > prev_object->size) 1870 prev_object->size = next_pindex + next_size; 1871 1872 VM_OBJECT_UNLOCK(prev_object); 1873 return (TRUE); 1874} 1875 1876void 1877vm_object_set_writeable_dirty(vm_object_t object) 1878{ 1879 struct vnode *vp; 1880 1881 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1882 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1883 if (object->type == OBJT_VNODE && 1884 (vp = (struct vnode *)object->handle) != NULL) { 1885 VI_LOCK(vp); 1886 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1887 vp->v_iflag |= VI_OBJDIRTY; 1888 VI_UNLOCK(vp); 1889 } 1890} 1891 1892#include "opt_ddb.h" 1893#ifdef DDB 1894#include <sys/kernel.h> 1895 1896#include <sys/cons.h> 1897 1898#include <ddb/ddb.h> 1899 1900static int 1901_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1902{ 1903 vm_map_t tmpm; 1904 vm_map_entry_t tmpe; 1905 vm_object_t obj; 1906 int entcount; 1907 1908 if (map == 0) 1909 return 0; 1910 1911 if (entry == 0) { 1912 tmpe = map->header.next; 1913 entcount = map->nentries; 1914 while (entcount-- && (tmpe != &map->header)) { 1915 if (_vm_object_in_map(map, object, tmpe)) { 1916 return 1; 1917 } 1918 tmpe = tmpe->next; 1919 } 1920 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1921 tmpm = entry->object.sub_map; 1922 tmpe = tmpm->header.next; 1923 entcount = tmpm->nentries; 1924 while (entcount-- && tmpe != &tmpm->header) { 1925 if (_vm_object_in_map(tmpm, object, tmpe)) { 1926 return 1; 1927 } 1928 tmpe = tmpe->next; 1929 } 1930 } else if ((obj = entry->object.vm_object) != NULL) { 1931 for (; obj; obj = obj->backing_object) 1932 if (obj == object) { 1933 return 1; 1934 } 1935 } 1936 return 0; 1937} 1938 1939static int 1940vm_object_in_map(vm_object_t object) 1941{ 1942 struct proc *p; 1943 1944 /* sx_slock(&allproc_lock); */ 1945 LIST_FOREACH(p, &allproc, p_list) { 1946 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1947 continue; 1948 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1949 /* sx_sunlock(&allproc_lock); */ 1950 return 1; 1951 } 1952 } 1953 /* sx_sunlock(&allproc_lock); */ 1954 if (_vm_object_in_map(kernel_map, object, 0)) 1955 return 1; 1956 if (_vm_object_in_map(kmem_map, object, 0)) 1957 return 1; 1958 if (_vm_object_in_map(pager_map, object, 0)) 1959 return 1; 1960 if (_vm_object_in_map(buffer_map, object, 0)) 1961 return 1; 1962 return 0; 1963} 1964 1965DB_SHOW_COMMAND(vmochk, vm_object_check) 1966{ 1967 vm_object_t object; 1968 1969 /* 1970 * make sure that internal objs are in a map somewhere 1971 * and none have zero ref counts. 1972 */ 1973 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1974 if (object->handle == NULL && 1975 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1976 if (object->ref_count == 0) { 1977 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1978 (long)object->size); 1979 } 1980 if (!vm_object_in_map(object)) { 1981 db_printf( 1982 "vmochk: internal obj is not in a map: " 1983 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1984 object->ref_count, (u_long)object->size, 1985 (u_long)object->size, 1986 (void *)object->backing_object); 1987 } 1988 } 1989 } 1990} 1991 1992/* 1993 * vm_object_print: [ debug ] 1994 */ 1995DB_SHOW_COMMAND(object, vm_object_print_static) 1996{ 1997 /* XXX convert args. */ 1998 vm_object_t object = (vm_object_t)addr; 1999 boolean_t full = have_addr; 2000 2001 vm_page_t p; 2002 2003 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2004#define count was_count 2005 2006 int count; 2007 2008 if (object == NULL) 2009 return; 2010 2011 db_iprintf( 2012 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 2013 object, (int)object->type, (uintmax_t)object->size, 2014 object->resident_page_count, object->ref_count, object->flags); 2015 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2016 object->shadow_count, 2017 object->backing_object ? object->backing_object->ref_count : 0, 2018 object->backing_object, (uintmax_t)object->backing_object_offset); 2019 2020 if (!full) 2021 return; 2022 2023 db_indent += 2; 2024 count = 0; 2025 TAILQ_FOREACH(p, &object->memq, listq) { 2026 if (count == 0) 2027 db_iprintf("memory:="); 2028 else if (count == 6) { 2029 db_printf("\n"); 2030 db_iprintf(" ..."); 2031 count = 0; 2032 } else 2033 db_printf(","); 2034 count++; 2035 2036 db_printf("(off=0x%jx,page=0x%jx)", 2037 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2038 } 2039 if (count != 0) 2040 db_printf("\n"); 2041 db_indent -= 2; 2042} 2043 2044/* XXX. */ 2045#undef count 2046 2047/* XXX need this non-static entry for calling from vm_map_print. */ 2048void 2049vm_object_print( 2050 /* db_expr_t */ long addr, 2051 boolean_t have_addr, 2052 /* db_expr_t */ long count, 2053 char *modif) 2054{ 2055 vm_object_print_static(addr, have_addr, count, modif); 2056} 2057 2058DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2059{ 2060 vm_object_t object; 2061 int nl = 0; 2062 int c; 2063 2064 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2065 vm_pindex_t idx, fidx; 2066 vm_pindex_t osize; 2067 vm_paddr_t pa = -1, padiff; 2068 int rcount; 2069 vm_page_t m; 2070 2071 db_printf("new object: %p\n", (void *)object); 2072 if (nl > 18) { 2073 c = cngetc(); 2074 if (c != ' ') 2075 return; 2076 nl = 0; 2077 } 2078 nl++; 2079 rcount = 0; 2080 fidx = 0; 2081 osize = object->size; 2082 if (osize > 128) 2083 osize = 128; 2084 for (idx = 0; idx < osize; idx++) { 2085 m = vm_page_lookup(object, idx); 2086 if (m == NULL) { 2087 if (rcount) { 2088 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2089 (long)fidx, rcount, (long)pa); 2090 if (nl > 18) { 2091 c = cngetc(); 2092 if (c != ' ') 2093 return; 2094 nl = 0; 2095 } 2096 nl++; 2097 rcount = 0; 2098 } 2099 continue; 2100 } 2101 2102 2103 if (rcount && 2104 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2105 ++rcount; 2106 continue; 2107 } 2108 if (rcount) { 2109 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2110 padiff >>= PAGE_SHIFT; 2111 padiff &= PQ_L2_MASK; 2112 if (padiff == 0) { 2113 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2114 ++rcount; 2115 continue; 2116 } 2117 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2118 (long)fidx, rcount, (long)pa); 2119 db_printf("pd(%ld)\n", (long)padiff); 2120 if (nl > 18) { 2121 c = cngetc(); 2122 if (c != ' ') 2123 return; 2124 nl = 0; 2125 } 2126 nl++; 2127 } 2128 fidx = idx; 2129 pa = VM_PAGE_TO_PHYS(m); 2130 rcount = 1; 2131 } 2132 if (rcount) { 2133 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2134 (long)fidx, rcount, (long)pa); 2135 if (nl > 18) { 2136 c = cngetc(); 2137 if (c != ' ') 2138 return; 2139 nl = 0; 2140 } 2141 nl++; 2142 } 2143 } 2144} 2145#endif /* DDB */ 2146