1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $ 65 * $DragonFly: src/sys/vm/vm_object.c,v 1.30 2007/03/20 00:55:10 dillon Exp $ 66 */ 67 68 /* 69 * Virtual memory object module. 70 */ 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/proc.h> /* for curproc, pageproc */ 75 #include <sys/vnode.h> 76 #include <sys/vmmeter.h> 77 #include <sys/mman.h> 78 #include <sys/mount.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.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/vm_zone.h> 94 95 #define EASY_SCAN_FACTOR 8 96 97 static void vm_object_qcollapse(vm_object_t object); 98 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 99 int pagerflags); 100 101 /* 102 * Virtual memory objects maintain the actual data 103 * associated with allocated virtual memory. A given 104 * page of memory exists within exactly one object. 105 * 106 * An object is only deallocated when all "references" 107 * are given up. Only one "reference" to a given 108 * region of an object should be writeable. 109 * 110 * Associated with each object is a list of all resident 111 * memory pages belonging to that object; this list is 112 * maintained by the "vm_page" module, and locked by the object's 113 * lock. 114 * 115 * Each object also records a "pager" routine which is 116 * used to retrieve (and store) pages to the proper backing 117 * storage. In addition, objects may be backed by other 118 * objects from which they were virtual-copied. 119 * 120 * The only items within the object structure which are 121 * modified after time of creation are: 122 * reference count locked by object's lock 123 * pager routine locked by object's lock 124 * 125 */ 126 127 struct object_q vm_object_list; 128 struct vm_object kernel_object; 129 130 static long vm_object_count; /* count of all objects */ 131 extern int vm_pageout_page_count; 132 133 static long object_collapses; 134 static long object_bypasses; 135 static int next_index; 136 static vm_zone_t obj_zone; 137 static struct vm_zone obj_zone_store; 138 static int object_hash_rand; 139 #define VM_OBJECTS_INIT 256 140 static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 141 142 void 143 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object) 144 { 145 int incr; 146 RB_INIT(&object->rb_memq); 147 LIST_INIT(&object->shadow_head); 148 149 object->type = type; 150 object->size = size; 151 object->ref_count = 1; 152 object->flags = 0; 153 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 154 vm_object_set_flag(object, OBJ_ONEMAPPING); 155 object->paging_in_progress = 0; 156 object->resident_page_count = 0; 157 object->shadow_count = 0; 158 object->pg_color = next_index; 159 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 160 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 161 else 162 incr = size; 163 next_index = (next_index + incr) & PQ_L2_MASK; 164 object->handle = NULL; 165 object->backing_object = NULL; 166 object->backing_object_offset = (vm_ooffset_t) 0; 167 /* 168 * Try to generate a number that will spread objects out in the 169 * hash table. We 'wipe' new objects across the hash in 128 page 170 * increments plus 1 more to offset it a little more by the time 171 * it wraps around. 172 */ 173 object->hash_rand = object_hash_rand - 129; 174 175 object->generation++; 176 177 crit_enter(); 178 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 179 vm_object_count++; 180 object_hash_rand = object->hash_rand; 181 crit_exit(); 182 } 183 184 /* 185 * vm_object_init: 186 * 187 * Initialize the VM objects module. 188 */ 189 void 190 vm_object_init(void) 191 { 192 TAILQ_INIT(&vm_object_list); 193 194 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 195 &kernel_object); 196 197 obj_zone = &obj_zone_store; 198 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 199 vm_objects_init, VM_OBJECTS_INIT); 200 } 201 202 void 203 vm_object_init2(void) 204 { 205 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1); 206 } 207 208 /* 209 * vm_object_allocate: 210 * 211 * Returns a new object with the given size. 212 */ 213 214 vm_object_t 215 vm_object_allocate(objtype_t type, vm_size_t size) 216 { 217 vm_object_t result; 218 219 result = (vm_object_t) zalloc(obj_zone); 220 221 _vm_object_allocate(type, size, result); 222 223 return (result); 224 } 225 226 227 /* 228 * vm_object_reference: 229 * 230 * Gets another reference to the given object. 231 */ 232 void 233 vm_object_reference(vm_object_t object) 234 { 235 if (object == NULL) 236 return; 237 238 object->ref_count++; 239 if (object->type == OBJT_VNODE) { 240 vref(object->handle); 241 /* XXX what if the vnode is being destroyed? */ 242 } 243 } 244 245 static void 246 vm_object_vndeallocate(vm_object_t object) 247 { 248 struct vnode *vp = (struct vnode *) object->handle; 249 250 KASSERT(object->type == OBJT_VNODE, 251 ("vm_object_vndeallocate: not a vnode object")); 252 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 253 #ifdef INVARIANTS 254 if (object->ref_count == 0) { 255 vprint("vm_object_vndeallocate", vp); 256 panic("vm_object_vndeallocate: bad object reference count"); 257 } 258 #endif 259 260 object->ref_count--; 261 if (object->ref_count == 0) 262 vp->v_flag &= ~VTEXT; 263 vrele(vp); 264 } 265 266 /* 267 * vm_object_deallocate: 268 * 269 * Release a reference to the specified object, 270 * gained either through a vm_object_allocate 271 * or a vm_object_reference call. When all references 272 * are gone, storage associated with this object 273 * may be relinquished. 274 * 275 * No object may be locked. 276 */ 277 void 278 vm_object_deallocate(vm_object_t object) 279 { 280 vm_object_t temp; 281 282 while (object != NULL) { 283 if (object->type == OBJT_VNODE) { 284 vm_object_vndeallocate(object); 285 return; 286 } 287 288 if (object->ref_count == 0) { 289 panic("vm_object_deallocate: object deallocated too many times: %d", object->type); 290 } else if (object->ref_count > 2) { 291 object->ref_count--; 292 return; 293 } 294 295 /* 296 * Here on ref_count of one or two, which are special cases for 297 * objects. 298 */ 299 if ((object->ref_count == 2) && (object->shadow_count == 0)) { 300 vm_object_set_flag(object, OBJ_ONEMAPPING); 301 object->ref_count--; 302 return; 303 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) { 304 object->ref_count--; 305 if ((object->handle == NULL) && 306 (object->type == OBJT_DEFAULT || 307 object->type == OBJT_SWAP)) { 308 vm_object_t robject; 309 310 robject = LIST_FIRST(&object->shadow_head); 311 KASSERT(robject != NULL, 312 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 313 object->ref_count, 314 object->shadow_count)); 315 if ((robject->handle == NULL) && 316 (robject->type == OBJT_DEFAULT || 317 robject->type == OBJT_SWAP)) { 318 319 robject->ref_count++; 320 321 while ( 322 robject->paging_in_progress || 323 object->paging_in_progress 324 ) { 325 vm_object_pip_sleep(robject, "objde1"); 326 vm_object_pip_sleep(object, "objde2"); 327 } 328 329 if (robject->ref_count == 1) { 330 robject->ref_count--; 331 object = robject; 332 goto doterm; 333 } 334 335 object = robject; 336 vm_object_collapse(object); 337 continue; 338 } 339 } 340 341 return; 342 343 } else { 344 object->ref_count--; 345 if (object->ref_count != 0) 346 return; 347 } 348 349 doterm: 350 351 temp = object->backing_object; 352 if (temp) { 353 LIST_REMOVE(object, shadow_list); 354 temp->shadow_count--; 355 temp->generation++; 356 object->backing_object = NULL; 357 } 358 359 /* 360 * Don't double-terminate, we could be in a termination 361 * recursion due to the terminate having to sync data 362 * to disk. 363 */ 364 if ((object->flags & OBJ_DEAD) == 0) 365 vm_object_terminate(object); 366 object = temp; 367 } 368 } 369 370 /* 371 * vm_object_terminate actually destroys the specified object, freeing 372 * up all previously used resources. 373 * 374 * The object must be locked. 375 * This routine may block. 376 */ 377 static int vm_object_terminate_callback(vm_page_t p, void *data); 378 379 void 380 vm_object_terminate(vm_object_t object) 381 { 382 /* 383 * Make sure no one uses us. 384 */ 385 vm_object_set_flag(object, OBJ_DEAD); 386 387 /* 388 * wait for the pageout daemon to be done with the object 389 */ 390 vm_object_pip_wait(object, "objtrm"); 391 392 KASSERT(!object->paging_in_progress, 393 ("vm_object_terminate: pageout in progress")); 394 395 /* 396 * Clean and free the pages, as appropriate. All references to the 397 * object are gone, so we don't need to lock it. 398 */ 399 if (object->type == OBJT_VNODE) { 400 struct vnode *vp; 401 402 /* 403 * Clean pages and flush buffers. 404 */ 405 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 406 407 vp = (struct vnode *) object->handle; 408 vinvalbuf(vp, V_SAVE, 0, 0); 409 } 410 411 /* 412 * Wait for any I/O to complete, after which there had better not 413 * be any references left on the object. 414 */ 415 vm_object_pip_wait(object, "objtrm"); 416 417 if (object->ref_count != 0) 418 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count); 419 420 /* 421 * Now free any remaining pages. For internal objects, this also 422 * removes them from paging queues. Don't free wired pages, just 423 * remove them from the object. 424 */ 425 crit_enter(); 426 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 427 vm_object_terminate_callback, NULL); 428 crit_exit(); 429 430 /* 431 * Let the pager know object is dead. 432 */ 433 vm_pager_deallocate(object); 434 435 /* 436 * Remove the object from the global object list. 437 */ 438 crit_enter(); 439 TAILQ_REMOVE(&vm_object_list, object, object_list); 440 vm_object_count--; 441 crit_exit(); 442 443 wakeup(object); 444 if (object->ref_count != 0) 445 panic("vm_object_terminate2: object with references, ref_count=%d", object->ref_count); 446 447 /* 448 * Free the space for the object. 449 */ 450 zfree(obj_zone, object); 451 } 452 453 static int 454 vm_object_terminate_callback(vm_page_t p, void *data __unused) 455 { 456 if (p->busy || (p->flags & PG_BUSY)) 457 panic("vm_object_terminate: freeing busy page %p", p); 458 if (p->wire_count == 0) { 459 vm_page_busy(p); 460 vm_page_free(p); 461 mycpu->gd_cnt.v_pfree++; 462 } else { 463 if (p->queue != PQ_NONE) 464 kprintf("vm_object_terminate: Warning: Encountered wired page %p on queue %d\n", p, p->queue); 465 vm_page_busy(p); 466 vm_page_remove(p); 467 vm_page_wakeup(p); 468 } 469 return(0); 470 } 471 472 /* 473 * vm_object_page_clean 474 * 475 * Clean all dirty pages in the specified range of object. Leaves page 476 * on whatever queue it is currently on. If NOSYNC is set then do not 477 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 478 * leaving the object dirty. 479 * 480 * When stuffing pages asynchronously, allow clustering. XXX we need a 481 * synchronous clustering mode implementation. 482 * 483 * Odd semantics: if start == end, we clean everything. 484 */ 485 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 486 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 487 488 void 489 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 490 int flags) 491 { 492 struct rb_vm_page_scan_info info; 493 struct vnode *vp; 494 int wholescan; 495 int pagerflags; 496 int curgeneration; 497 498 if (object->type != OBJT_VNODE || 499 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 500 return; 501 502 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 503 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 504 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 505 506 vp = object->handle; 507 508 /* 509 * Interlock other major object operations. This allows us to 510 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 511 */ 512 crit_enter(); 513 vm_object_set_flag(object, OBJ_CLEANING); 514 515 /* 516 * Handle 'entire object' case 517 */ 518 info.start_pindex = start; 519 if (end == 0) { 520 info.end_pindex = object->size - 1; 521 } else { 522 info.end_pindex = end - 1; 523 } 524 wholescan = (start == 0 && info.end_pindex == object->size - 1); 525 info.limit = flags; 526 info.pagerflags = pagerflags; 527 info.object = object; 528 529 /* 530 * If cleaning the entire object do a pass to mark the pages read-only. 531 * If everything worked out ok, clear OBJ_WRITEABLE and 532 * OBJ_MIGHTBEDIRTY. 533 */ 534 if (wholescan) { 535 info.error = 0; 536 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 537 vm_object_page_clean_pass1, &info); 538 if (info.error == 0) { 539 vm_object_clear_flag(object, 540 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 541 if (object->type == OBJT_VNODE && 542 (vp = (struct vnode *)object->handle) != NULL) { 543 if (vp->v_flag & VOBJDIRTY) 544 vclrflags(vp, VOBJDIRTY); 545 } 546 } 547 } 548 549 /* 550 * Do a pass to clean all the dirty pages we find. 551 */ 552 do { 553 info.error = 0; 554 curgeneration = object->generation; 555 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 556 vm_object_page_clean_pass2, &info); 557 } while (info.error || curgeneration != object->generation); 558 559 vm_object_clear_flag(object, OBJ_CLEANING); 560 crit_exit(); 561 } 562 563 static 564 int 565 vm_object_page_clean_pass1(struct vm_page *p, void *data) 566 { 567 struct rb_vm_page_scan_info *info = data; 568 569 vm_page_flag_set(p, PG_CLEANCHK); 570 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 571 info->error = 1; 572 else 573 vm_page_protect(p, VM_PROT_READ); 574 return(0); 575 } 576 577 static 578 int 579 vm_object_page_clean_pass2(struct vm_page *p, void *data) 580 { 581 struct rb_vm_page_scan_info *info = data; 582 int n; 583 584 /* 585 * Do not mess with pages that were inserted after we started 586 * the cleaning pass. 587 */ 588 if ((p->flags & PG_CLEANCHK) == 0) 589 return(0); 590 591 /* 592 * Before wasting time traversing the pmaps, check for trivial 593 * cases where the page cannot be dirty. 594 */ 595 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 596 KKASSERT((p->dirty & p->valid) == 0); 597 return(0); 598 } 599 600 /* 601 * Check whether the page is dirty or not. The page has been set 602 * to be read-only so the check will not race a user dirtying the 603 * page. 604 */ 605 vm_page_test_dirty(p); 606 if ((p->dirty & p->valid) == 0) { 607 vm_page_flag_clear(p, PG_CLEANCHK); 608 return(0); 609 } 610 611 /* 612 * If we have been asked to skip nosync pages and this is a 613 * nosync page, skip it. Note that the object flags were 614 * not cleared in this case (because pass1 will have returned an 615 * error), so we do not have to set them. 616 */ 617 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 618 vm_page_flag_clear(p, PG_CLEANCHK); 619 return(0); 620 } 621 622 /* 623 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 624 * the pages that get successfully flushed. Set info->error if 625 * we raced an object modification. 626 */ 627 n = vm_object_page_collect_flush(info->object, p, info->pagerflags); 628 if (n == 0) 629 info->error = 1; 630 return(0); 631 } 632 633 /* 634 * This routine must be called within a critical section to properly avoid 635 * an interrupt unbusy/free race that can occur prior to the busy check. 636 * 637 * Using the object generation number here to detect page ripout is not 638 * the best idea in the world. XXX 639 * 640 * NOTE: we operate under the assumption that a page found to not be busy 641 * will not be ripped out from under us by an interrupt. XXX we should 642 * recode this to explicitly busy the pages. 643 */ 644 static int 645 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 646 { 647 int runlen; 648 int maxf; 649 int chkb; 650 int maxb; 651 int i; 652 int curgeneration; 653 vm_pindex_t pi; 654 vm_page_t maf[vm_pageout_page_count]; 655 vm_page_t mab[vm_pageout_page_count]; 656 vm_page_t ma[vm_pageout_page_count]; 657 658 curgeneration = object->generation; 659 660 pi = p->pindex; 661 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 662 if (object->generation != curgeneration) { 663 return(0); 664 } 665 } 666 KKASSERT(p->object == object && p->pindex == pi); 667 668 maxf = 0; 669 for(i = 1; i < vm_pageout_page_count; i++) { 670 vm_page_t tp; 671 672 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 673 if ((tp->flags & PG_BUSY) || 674 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 675 (tp->flags & PG_CLEANCHK) == 0) || 676 (tp->busy != 0)) 677 break; 678 if((tp->queue - tp->pc) == PQ_CACHE) { 679 vm_page_flag_clear(tp, PG_CLEANCHK); 680 break; 681 } 682 vm_page_test_dirty(tp); 683 if ((tp->dirty & tp->valid) == 0) { 684 vm_page_flag_clear(tp, PG_CLEANCHK); 685 break; 686 } 687 maf[ i - 1 ] = tp; 688 maxf++; 689 continue; 690 } 691 break; 692 } 693 694 maxb = 0; 695 chkb = vm_pageout_page_count - maxf; 696 if (chkb) { 697 for(i = 1; i < chkb;i++) { 698 vm_page_t tp; 699 700 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 701 if ((tp->flags & PG_BUSY) || 702 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 703 (tp->flags & PG_CLEANCHK) == 0) || 704 (tp->busy != 0)) 705 break; 706 if((tp->queue - tp->pc) == PQ_CACHE) { 707 vm_page_flag_clear(tp, PG_CLEANCHK); 708 break; 709 } 710 vm_page_test_dirty(tp); 711 if ((tp->dirty & tp->valid) == 0) { 712 vm_page_flag_clear(tp, PG_CLEANCHK); 713 break; 714 } 715 mab[ i - 1 ] = tp; 716 maxb++; 717 continue; 718 } 719 break; 720 } 721 } 722 723 for(i = 0; i < maxb; i++) { 724 int index = (maxb - i) - 1; 725 ma[index] = mab[i]; 726 vm_page_flag_clear(ma[index], PG_CLEANCHK); 727 } 728 vm_page_flag_clear(p, PG_CLEANCHK); 729 ma[maxb] = p; 730 for(i = 0; i < maxf; i++) { 731 int index = (maxb + i) + 1; 732 ma[index] = maf[i]; 733 vm_page_flag_clear(ma[index], PG_CLEANCHK); 734 } 735 runlen = maxb + maxf + 1; 736 737 vm_pageout_flush(ma, runlen, pagerflags); 738 for (i = 0; i < runlen; i++) { 739 if (ma[i]->valid & ma[i]->dirty) { 740 vm_page_protect(ma[i], VM_PROT_READ); 741 vm_page_flag_set(ma[i], PG_CLEANCHK); 742 743 /* 744 * maxf will end up being the actual number of pages 745 * we wrote out contiguously, non-inclusive of the 746 * first page. We do not count look-behind pages. 747 */ 748 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 749 maxf = i - maxb - 1; 750 } 751 } 752 return(maxf + 1); 753 } 754 755 #ifdef not_used 756 /* XXX I cannot tell if this should be an exported symbol */ 757 /* 758 * vm_object_deactivate_pages 759 * 760 * Deactivate all pages in the specified object. (Keep its pages 761 * in memory even though it is no longer referenced.) 762 * 763 * The object must be locked. 764 */ 765 static int vm_object_deactivate_pages_callback(vm_page_t p, void *data); 766 767 static void 768 vm_object_deactivate_pages(vm_object_t object) 769 { 770 crit_enter(); 771 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 772 vm_object_deactivate_pages_callback, NULL); 773 crit_exit(); 774 } 775 776 static int 777 vm_object_deactivate_pages_callback(vm_page_t p, void *data __unused) 778 { 779 vm_page_deactivate(p); 780 return(0); 781 } 782 783 #endif 784 785 /* 786 * Same as vm_object_pmap_copy, except range checking really 787 * works, and is meant for small sections of an object. 788 * 789 * This code protects resident pages by making them read-only 790 * and is typically called on a fork or split when a page 791 * is converted to copy-on-write. 792 * 793 * NOTE: If the page is already at VM_PROT_NONE, calling 794 * vm_page_protect will have no effect. 795 */ 796 void 797 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 798 { 799 vm_pindex_t idx; 800 vm_page_t p; 801 802 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 803 return; 804 805 /* 806 * spl protection needed to prevent races between the lookup, 807 * an interrupt unbusy/free, and our protect call. 808 */ 809 crit_enter(); 810 for (idx = start; idx < end; idx++) { 811 p = vm_page_lookup(object, idx); 812 if (p == NULL) 813 continue; 814 vm_page_protect(p, VM_PROT_READ); 815 } 816 crit_exit(); 817 } 818 819 /* 820 * vm_object_pmap_remove: 821 * 822 * Removes all physical pages in the specified 823 * object range from all physical maps. 824 * 825 * The object must *not* be locked. 826 */ 827 828 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 829 830 void 831 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 832 { 833 struct rb_vm_page_scan_info info; 834 835 if (object == NULL) 836 return; 837 info.start_pindex = start; 838 info.end_pindex = end - 1; 839 crit_enter(); 840 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 841 vm_object_pmap_remove_callback, &info); 842 if (start == 0 && end == object->size) 843 vm_object_clear_flag(object, OBJ_WRITEABLE); 844 crit_exit(); 845 } 846 847 static int 848 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 849 { 850 vm_page_protect(p, VM_PROT_NONE); 851 return(0); 852 } 853 854 /* 855 * vm_object_madvise: 856 * 857 * Implements the madvise function at the object/page level. 858 * 859 * MADV_WILLNEED (any object) 860 * 861 * Activate the specified pages if they are resident. 862 * 863 * MADV_DONTNEED (any object) 864 * 865 * Deactivate the specified pages if they are resident. 866 * 867 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 868 * OBJ_ONEMAPPING only) 869 * 870 * Deactivate and clean the specified pages if they are 871 * resident. This permits the process to reuse the pages 872 * without faulting or the kernel to reclaim the pages 873 * without I/O. 874 */ 875 void 876 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 877 { 878 vm_pindex_t end, tpindex; 879 vm_object_t tobject; 880 vm_page_t m; 881 882 if (object == NULL) 883 return; 884 885 end = pindex + count; 886 887 /* 888 * Locate and adjust resident pages 889 */ 890 891 for (; pindex < end; pindex += 1) { 892 relookup: 893 tobject = object; 894 tpindex = pindex; 895 shadowlookup: 896 /* 897 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 898 * and those pages must be OBJ_ONEMAPPING. 899 */ 900 if (advise == MADV_FREE) { 901 if ((tobject->type != OBJT_DEFAULT && 902 tobject->type != OBJT_SWAP) || 903 (tobject->flags & OBJ_ONEMAPPING) == 0) { 904 continue; 905 } 906 } 907 908 /* 909 * spl protection is required to avoid a race between the 910 * lookup, an interrupt unbusy/free, and our busy check. 911 */ 912 913 crit_enter(); 914 m = vm_page_lookup(tobject, tpindex); 915 916 if (m == NULL) { 917 /* 918 * There may be swap even if there is no backing page 919 */ 920 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 921 swap_pager_freespace(tobject, tpindex, 1); 922 923 /* 924 * next object 925 */ 926 crit_exit(); 927 if (tobject->backing_object == NULL) 928 continue; 929 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 930 tobject = tobject->backing_object; 931 goto shadowlookup; 932 } 933 934 /* 935 * If the page is busy or not in a normal active state, 936 * we skip it. If the page is not managed there are no 937 * page queues to mess with. Things can break if we mess 938 * with pages in any of the below states. 939 */ 940 if ( 941 m->hold_count || 942 m->wire_count || 943 (m->flags & PG_UNMANAGED) || 944 m->valid != VM_PAGE_BITS_ALL 945 ) { 946 crit_exit(); 947 continue; 948 } 949 950 if (vm_page_sleep_busy(m, TRUE, "madvpo")) { 951 crit_exit(); 952 goto relookup; 953 } 954 crit_exit(); 955 956 /* 957 * Theoretically once a page is known not to be busy, an 958 * interrupt cannot come along and rip it out from under us. 959 */ 960 961 if (advise == MADV_WILLNEED) { 962 vm_page_activate(m); 963 } else if (advise == MADV_DONTNEED) { 964 vm_page_dontneed(m); 965 } else if (advise == MADV_FREE) { 966 /* 967 * Mark the page clean. This will allow the page 968 * to be freed up by the system. However, such pages 969 * are often reused quickly by malloc()/free() 970 * so we do not do anything that would cause 971 * a page fault if we can help it. 972 * 973 * Specifically, we do not try to actually free 974 * the page now nor do we try to put it in the 975 * cache (which would cause a page fault on reuse). 976 * 977 * But we do make the page is freeable as we 978 * can without actually taking the step of unmapping 979 * it. 980 */ 981 pmap_clear_modify(m); 982 m->dirty = 0; 983 m->act_count = 0; 984 vm_page_dontneed(m); 985 if (tobject->type == OBJT_SWAP) 986 swap_pager_freespace(tobject, tpindex, 1); 987 } 988 } 989 } 990 991 /* 992 * vm_object_shadow: 993 * 994 * Create a new object which is backed by the 995 * specified existing object range. The source 996 * object reference is deallocated. 997 * 998 * The new object and offset into that object 999 * are returned in the source parameters. 1000 */ 1001 1002 void 1003 vm_object_shadow(vm_object_t *object, /* IN/OUT */ 1004 vm_ooffset_t *offset, /* IN/OUT */ 1005 vm_size_t length) 1006 { 1007 vm_object_t source; 1008 vm_object_t result; 1009 1010 source = *object; 1011 1012 /* 1013 * Don't create the new object if the old object isn't shared. 1014 */ 1015 1016 if (source != NULL && 1017 source->ref_count == 1 && 1018 source->handle == NULL && 1019 (source->type == OBJT_DEFAULT || 1020 source->type == OBJT_SWAP)) 1021 return; 1022 1023 /* 1024 * Allocate a new object with the given length 1025 */ 1026 1027 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL) 1028 panic("vm_object_shadow: no object for shadowing"); 1029 1030 /* 1031 * The new object shadows the source object, adding a reference to it. 1032 * Our caller changes his reference to point to the new object, 1033 * removing a reference to the source object. Net result: no change 1034 * of reference count. 1035 * 1036 * Try to optimize the result object's page color when shadowing 1037 * in order to maintain page coloring consistency in the combined 1038 * shadowed object. 1039 */ 1040 result->backing_object = source; 1041 if (source) { 1042 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1043 source->shadow_count++; 1044 source->generation++; 1045 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK; 1046 } 1047 1048 /* 1049 * Store the offset into the source object, and fix up the offset into 1050 * the new object. 1051 */ 1052 1053 result->backing_object_offset = *offset; 1054 1055 /* 1056 * Return the new things 1057 */ 1058 1059 *offset = 0; 1060 *object = result; 1061 } 1062 1063 #define OBSC_TEST_ALL_SHADOWED 0x0001 1064 #define OBSC_COLLAPSE_NOWAIT 0x0002 1065 #define OBSC_COLLAPSE_WAIT 0x0004 1066 1067 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1068 1069 static __inline int 1070 vm_object_backing_scan(vm_object_t object, int op) 1071 { 1072 struct rb_vm_page_scan_info info; 1073 vm_object_t backing_object; 1074 1075 /* 1076 * spl protection is required to avoid races between the memq/lookup, 1077 * an interrupt doing an unbusy/free, and our busy check. Amoung 1078 * other things. 1079 */ 1080 crit_enter(); 1081 1082 backing_object = object->backing_object; 1083 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1084 1085 /* 1086 * Initial conditions 1087 */ 1088 1089 if (op & OBSC_TEST_ALL_SHADOWED) { 1090 /* 1091 * We do not want to have to test for the existence of 1092 * swap pages in the backing object. XXX but with the 1093 * new swapper this would be pretty easy to do. 1094 * 1095 * XXX what about anonymous MAP_SHARED memory that hasn't 1096 * been ZFOD faulted yet? If we do not test for this, the 1097 * shadow test may succeed! XXX 1098 */ 1099 if (backing_object->type != OBJT_DEFAULT) { 1100 crit_exit(); 1101 return(0); 1102 } 1103 } 1104 if (op & OBSC_COLLAPSE_WAIT) { 1105 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1106 vm_object_set_flag(backing_object, OBJ_DEAD); 1107 } 1108 1109 /* 1110 * Our scan. We have to retry if a negative error code is returned, 1111 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1112 * the scan had to be stopped because the parent does not completely 1113 * shadow the child. 1114 */ 1115 info.object = object; 1116 info.backing_object = backing_object; 1117 info.limit = op; 1118 do { 1119 info.error = 1; 1120 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1121 vm_object_backing_scan_callback, 1122 &info); 1123 } while (info.error < 0); 1124 crit_exit(); 1125 return(info.error); 1126 } 1127 1128 static int 1129 vm_object_backing_scan_callback(vm_page_t p, void *data) 1130 { 1131 struct rb_vm_page_scan_info *info = data; 1132 vm_object_t backing_object; 1133 vm_object_t object; 1134 vm_pindex_t new_pindex; 1135 vm_pindex_t backing_offset_index; 1136 int op; 1137 1138 new_pindex = p->pindex - info->backing_offset_index; 1139 op = info->limit; 1140 object = info->object; 1141 backing_object = info->backing_object; 1142 backing_offset_index = info->backing_offset_index; 1143 1144 if (op & OBSC_TEST_ALL_SHADOWED) { 1145 vm_page_t pp; 1146 1147 /* 1148 * Ignore pages outside the parent object's range 1149 * and outside the parent object's mapping of the 1150 * backing object. 1151 * 1152 * note that we do not busy the backing object's 1153 * page. 1154 */ 1155 if ( 1156 p->pindex < backing_offset_index || 1157 new_pindex >= object->size 1158 ) { 1159 return(0); 1160 } 1161 1162 /* 1163 * See if the parent has the page or if the parent's 1164 * object pager has the page. If the parent has the 1165 * page but the page is not valid, the parent's 1166 * object pager must have the page. 1167 * 1168 * If this fails, the parent does not completely shadow 1169 * the object and we might as well give up now. 1170 */ 1171 1172 pp = vm_page_lookup(object, new_pindex); 1173 if ( 1174 (pp == NULL || pp->valid == 0) && 1175 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1176 ) { 1177 info->error = 0; /* problemo */ 1178 return(-1); /* stop the scan */ 1179 } 1180 } 1181 1182 /* 1183 * Check for busy page 1184 */ 1185 1186 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1187 vm_page_t pp; 1188 1189 if (op & OBSC_COLLAPSE_NOWAIT) { 1190 if ( 1191 (p->flags & PG_BUSY) || 1192 !p->valid || 1193 p->hold_count || 1194 p->wire_count || 1195 p->busy 1196 ) { 1197 return(0); 1198 } 1199 } else if (op & OBSC_COLLAPSE_WAIT) { 1200 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1201 /* 1202 * If we slept, anything could have 1203 * happened. Ask that the scan be restarted. 1204 * 1205 * Since the object is marked dead, the 1206 * backing offset should not have changed. 1207 */ 1208 info->error = -1; 1209 return(-1); 1210 } 1211 } 1212 1213 /* 1214 * Busy the page 1215 */ 1216 vm_page_busy(p); 1217 1218 KASSERT( 1219 p->object == backing_object, 1220 ("vm_object_qcollapse(): object mismatch") 1221 ); 1222 1223 /* 1224 * Destroy any associated swap 1225 */ 1226 if (backing_object->type == OBJT_SWAP) { 1227 swap_pager_freespace( 1228 backing_object, 1229 p->pindex, 1230 1 1231 ); 1232 } 1233 1234 if ( 1235 p->pindex < backing_offset_index || 1236 new_pindex >= object->size 1237 ) { 1238 /* 1239 * Page is out of the parent object's range, we 1240 * can simply destroy it. 1241 */ 1242 vm_page_protect(p, VM_PROT_NONE); 1243 vm_page_free(p); 1244 return(0); 1245 } 1246 1247 pp = vm_page_lookup(object, new_pindex); 1248 if ( 1249 pp != NULL || 1250 vm_pager_has_page(object, new_pindex, NULL, NULL) 1251 ) { 1252 /* 1253 * page already exists in parent OR swap exists 1254 * for this location in the parent. Destroy 1255 * the original page from the backing object. 1256 * 1257 * Leave the parent's page alone 1258 */ 1259 vm_page_protect(p, VM_PROT_NONE); 1260 vm_page_free(p); 1261 return(0); 1262 } 1263 1264 /* 1265 * Page does not exist in parent, rename the 1266 * page from the backing object to the main object. 1267 * 1268 * If the page was mapped to a process, it can remain 1269 * mapped through the rename. 1270 */ 1271 if ((p->queue - p->pc) == PQ_CACHE) 1272 vm_page_deactivate(p); 1273 1274 vm_page_rename(p, object, new_pindex); 1275 /* page automatically made dirty by rename */ 1276 } 1277 return(0); 1278 } 1279 1280 /* 1281 * this version of collapse allows the operation to occur earlier and 1282 * when paging_in_progress is true for an object... This is not a complete 1283 * operation, but should plug 99.9% of the rest of the leaks. 1284 */ 1285 static void 1286 vm_object_qcollapse(vm_object_t object) 1287 { 1288 vm_object_t backing_object = object->backing_object; 1289 1290 if (backing_object->ref_count != 1) 1291 return; 1292 1293 backing_object->ref_count += 2; 1294 1295 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1296 1297 backing_object->ref_count -= 2; 1298 } 1299 1300 /* 1301 * vm_object_collapse: 1302 * 1303 * Collapse an object with the object backing it. 1304 * Pages in the backing object are moved into the 1305 * parent, and the backing object is deallocated. 1306 */ 1307 void 1308 vm_object_collapse(vm_object_t object) 1309 { 1310 while (TRUE) { 1311 vm_object_t backing_object; 1312 1313 /* 1314 * Verify that the conditions are right for collapse: 1315 * 1316 * The object exists and the backing object exists. 1317 */ 1318 if (object == NULL) 1319 break; 1320 1321 if ((backing_object = object->backing_object) == NULL) 1322 break; 1323 1324 /* 1325 * we check the backing object first, because it is most likely 1326 * not collapsable. 1327 */ 1328 if (backing_object->handle != NULL || 1329 (backing_object->type != OBJT_DEFAULT && 1330 backing_object->type != OBJT_SWAP) || 1331 (backing_object->flags & OBJ_DEAD) || 1332 object->handle != NULL || 1333 (object->type != OBJT_DEFAULT && 1334 object->type != OBJT_SWAP) || 1335 (object->flags & OBJ_DEAD)) { 1336 break; 1337 } 1338 1339 if ( 1340 object->paging_in_progress != 0 || 1341 backing_object->paging_in_progress != 0 1342 ) { 1343 vm_object_qcollapse(object); 1344 break; 1345 } 1346 1347 /* 1348 * We know that we can either collapse the backing object (if 1349 * the parent is the only reference to it) or (perhaps) have 1350 * the parent bypass the object if the parent happens to shadow 1351 * all the resident pages in the entire backing object. 1352 * 1353 * This is ignoring pager-backed pages such as swap pages. 1354 * vm_object_backing_scan fails the shadowing test in this 1355 * case. 1356 */ 1357 1358 if (backing_object->ref_count == 1) { 1359 /* 1360 * If there is exactly one reference to the backing 1361 * object, we can collapse it into the parent. 1362 */ 1363 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1364 1365 /* 1366 * Move the pager from backing_object to object. 1367 */ 1368 1369 if (backing_object->type == OBJT_SWAP) { 1370 vm_object_pip_add(backing_object, 1); 1371 1372 /* 1373 * scrap the paging_offset junk and do a 1374 * discrete copy. This also removes major 1375 * assumptions about how the swap-pager 1376 * works from where it doesn't belong. The 1377 * new swapper is able to optimize the 1378 * destroy-source case. 1379 */ 1380 1381 vm_object_pip_add(object, 1); 1382 swap_pager_copy( 1383 backing_object, 1384 object, 1385 OFF_TO_IDX(object->backing_object_offset), TRUE); 1386 vm_object_pip_wakeup(object); 1387 1388 vm_object_pip_wakeup(backing_object); 1389 } 1390 /* 1391 * Object now shadows whatever backing_object did. 1392 * Note that the reference to 1393 * backing_object->backing_object moves from within 1394 * backing_object to within object. 1395 */ 1396 1397 LIST_REMOVE(object, shadow_list); 1398 object->backing_object->shadow_count--; 1399 object->backing_object->generation++; 1400 if (backing_object->backing_object) { 1401 LIST_REMOVE(backing_object, shadow_list); 1402 backing_object->backing_object->shadow_count--; 1403 backing_object->backing_object->generation++; 1404 } 1405 object->backing_object = backing_object->backing_object; 1406 if (object->backing_object) { 1407 LIST_INSERT_HEAD( 1408 &object->backing_object->shadow_head, 1409 object, 1410 shadow_list 1411 ); 1412 object->backing_object->shadow_count++; 1413 object->backing_object->generation++; 1414 } 1415 1416 object->backing_object_offset += 1417 backing_object->backing_object_offset; 1418 1419 /* 1420 * Discard backing_object. 1421 * 1422 * Since the backing object has no pages, no pager left, 1423 * and no object references within it, all that is 1424 * necessary is to dispose of it. 1425 */ 1426 1427 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1428 KASSERT(RB_EMPTY(&backing_object->rb_memq), ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1429 crit_enter(); 1430 TAILQ_REMOVE( 1431 &vm_object_list, 1432 backing_object, 1433 object_list 1434 ); 1435 vm_object_count--; 1436 crit_exit(); 1437 1438 zfree(obj_zone, backing_object); 1439 1440 object_collapses++; 1441 } else { 1442 vm_object_t new_backing_object; 1443 1444 /* 1445 * If we do not entirely shadow the backing object, 1446 * there is nothing we can do so we give up. 1447 */ 1448 1449 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1450 break; 1451 } 1452 1453 /* 1454 * Make the parent shadow the next object in the 1455 * chain. Deallocating backing_object will not remove 1456 * it, since its reference count is at least 2. 1457 */ 1458 1459 LIST_REMOVE(object, shadow_list); 1460 backing_object->shadow_count--; 1461 backing_object->generation++; 1462 1463 new_backing_object = backing_object->backing_object; 1464 if ((object->backing_object = new_backing_object) != NULL) { 1465 vm_object_reference(new_backing_object); 1466 LIST_INSERT_HEAD( 1467 &new_backing_object->shadow_head, 1468 object, 1469 shadow_list 1470 ); 1471 new_backing_object->shadow_count++; 1472 new_backing_object->generation++; 1473 object->backing_object_offset += 1474 backing_object->backing_object_offset; 1475 } 1476 1477 /* 1478 * Drop the reference count on backing_object. Since 1479 * its ref_count was at least 2, it will not vanish; 1480 * so we don't need to call vm_object_deallocate, but 1481 * we do anyway. 1482 */ 1483 vm_object_deallocate(backing_object); 1484 object_bypasses++; 1485 } 1486 1487 /* 1488 * Try again with this object's new backing object. 1489 */ 1490 } 1491 } 1492 1493 /* 1494 * vm_object_page_remove: [internal] 1495 * 1496 * Removes all physical pages in the specified 1497 * object range from the object's list of pages. 1498 */ 1499 static int vm_object_page_remove_callback(vm_page_t p, void *data); 1500 1501 void 1502 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1503 boolean_t clean_only) 1504 { 1505 struct rb_vm_page_scan_info info; 1506 int all; 1507 1508 /* 1509 * Degenerate cases and assertions 1510 */ 1511 if (object == NULL || object->resident_page_count == 0) 1512 return; 1513 KASSERT(object->type != OBJT_PHYS, 1514 ("attempt to remove pages from a physical object")); 1515 1516 /* 1517 * Indicate that paging is occuring on the object 1518 */ 1519 crit_enter(); 1520 vm_object_pip_add(object, 1); 1521 1522 /* 1523 * Figure out the actual removal range and whether we are removing 1524 * the entire contents of the object or not. If removing the entire 1525 * contents, be sure to get all pages, even those that might be 1526 * beyond the end of the object. 1527 */ 1528 info.start_pindex = start; 1529 if (end == 0) 1530 info.end_pindex = (vm_pindex_t)-1; 1531 else 1532 info.end_pindex = end - 1; 1533 info.limit = clean_only; 1534 all = (start == 0 && info.end_pindex >= object->size - 1); 1535 1536 /* 1537 * Loop until we are sure we have gotten them all. 1538 */ 1539 do { 1540 info.error = 0; 1541 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1542 vm_object_page_remove_callback, &info); 1543 } while (info.error); 1544 1545 /* 1546 * Cleanup 1547 */ 1548 vm_object_pip_wakeup(object); 1549 crit_exit(); 1550 } 1551 1552 static int 1553 vm_object_page_remove_callback(vm_page_t p, void *data) 1554 { 1555 struct rb_vm_page_scan_info *info = data; 1556 1557 /* 1558 * Wired pages cannot be destroyed, but they can be invalidated 1559 * and we do so if clean_only (limit) is not set. 1560 */ 1561 if (p->wire_count != 0) { 1562 vm_page_protect(p, VM_PROT_NONE); 1563 if (info->limit == 0) 1564 p->valid = 0; 1565 return(0); 1566 } 1567 1568 /* 1569 * The busy flags are only cleared at 1570 * interrupt -- minimize the spl transitions 1571 */ 1572 1573 if (vm_page_sleep_busy(p, TRUE, "vmopar")) { 1574 info->error = 1; 1575 return(0); 1576 } 1577 1578 /* 1579 * limit is our clean_only flag. If set and the page is dirty, do 1580 * not free it. 1581 */ 1582 if (info->limit && p->valid) { 1583 vm_page_test_dirty(p); 1584 if (p->valid & p->dirty) 1585 return(0); 1586 } 1587 1588 /* 1589 * Destroy the page 1590 */ 1591 vm_page_busy(p); 1592 vm_page_protect(p, VM_PROT_NONE); 1593 vm_page_free(p); 1594 return(0); 1595 } 1596 1597 /* 1598 * Routine: vm_object_coalesce 1599 * Function: Coalesces two objects backing up adjoining 1600 * regions of memory into a single object. 1601 * 1602 * returns TRUE if objects were combined. 1603 * 1604 * NOTE: Only works at the moment if the second object is NULL - 1605 * if it's not, which object do we lock first? 1606 * 1607 * Parameters: 1608 * prev_object First object to coalesce 1609 * prev_offset Offset into prev_object 1610 * next_object Second object into coalesce 1611 * next_offset Offset into next_object 1612 * 1613 * prev_size Size of reference to prev_object 1614 * next_size Size of reference to next_object 1615 * 1616 * Conditions: 1617 * The object must *not* be locked. 1618 */ 1619 boolean_t 1620 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1621 vm_size_t prev_size, vm_size_t next_size) 1622 { 1623 vm_pindex_t next_pindex; 1624 1625 if (prev_object == NULL) { 1626 return (TRUE); 1627 } 1628 1629 if (prev_object->type != OBJT_DEFAULT && 1630 prev_object->type != OBJT_SWAP) { 1631 return (FALSE); 1632 } 1633 1634 /* 1635 * Try to collapse the object first 1636 */ 1637 vm_object_collapse(prev_object); 1638 1639 /* 1640 * Can't coalesce if: . more than one reference . paged out . shadows 1641 * another object . has a copy elsewhere (any of which mean that the 1642 * pages not mapped to prev_entry may be in use anyway) 1643 */ 1644 1645 if (prev_object->backing_object != NULL) { 1646 return (FALSE); 1647 } 1648 1649 prev_size >>= PAGE_SHIFT; 1650 next_size >>= PAGE_SHIFT; 1651 next_pindex = prev_pindex + prev_size; 1652 1653 if ((prev_object->ref_count > 1) && 1654 (prev_object->size != next_pindex)) { 1655 return (FALSE); 1656 } 1657 1658 /* 1659 * Remove any pages that may still be in the object from a previous 1660 * deallocation. 1661 */ 1662 if (next_pindex < prev_object->size) { 1663 vm_object_page_remove(prev_object, 1664 next_pindex, 1665 next_pindex + next_size, FALSE); 1666 if (prev_object->type == OBJT_SWAP) 1667 swap_pager_freespace(prev_object, 1668 next_pindex, next_size); 1669 } 1670 1671 /* 1672 * Extend the object if necessary. 1673 */ 1674 if (next_pindex + next_size > prev_object->size) 1675 prev_object->size = next_pindex + next_size; 1676 1677 return (TRUE); 1678 } 1679 1680 void 1681 vm_object_set_writeable_dirty(vm_object_t object) 1682 { 1683 struct vnode *vp; 1684 1685 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1686 if (object->type == OBJT_VNODE && 1687 (vp = (struct vnode *)object->handle) != NULL) { 1688 if ((vp->v_flag & VOBJDIRTY) == 0) { 1689 vsetflags(vp, VOBJDIRTY); 1690 } 1691 } 1692 } 1693 1694 1695 1696 #include "opt_ddb.h" 1697 #ifdef DDB 1698 #include <sys/kernel.h> 1699 1700 #include <sys/cons.h> 1701 1702 #include <ddb/ddb.h> 1703 1704 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 1705 vm_map_entry_t entry); 1706 static int vm_object_in_map (vm_object_t object); 1707 1708 static int 1709 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1710 { 1711 vm_map_t tmpm; 1712 vm_map_entry_t tmpe; 1713 vm_object_t obj; 1714 int entcount; 1715 1716 if (map == 0) 1717 return 0; 1718 if (entry == 0) { 1719 tmpe = map->header.next; 1720 entcount = map->nentries; 1721 while (entcount-- && (tmpe != &map->header)) { 1722 if( _vm_object_in_map(map, object, tmpe)) { 1723 return 1; 1724 } 1725 tmpe = tmpe->next; 1726 } 1727 return (0); 1728 } 1729 switch(entry->maptype) { 1730 case VM_MAPTYPE_SUBMAP: 1731 tmpm = entry->object.sub_map; 1732 tmpe = tmpm->header.next; 1733 entcount = tmpm->nentries; 1734 while (entcount-- && tmpe != &tmpm->header) { 1735 if( _vm_object_in_map(tmpm, object, tmpe)) { 1736 return 1; 1737 } 1738 tmpe = tmpe->next; 1739 } 1740 break; 1741 case VM_MAPTYPE_NORMAL: 1742 case VM_MAPTYPE_VPAGETABLE: 1743 obj = entry->object.vm_object; 1744 while (obj) { 1745 if (obj == object) 1746 return 1; 1747 obj = obj->backing_object; 1748 } 1749 break; 1750 default: 1751 break; 1752 } 1753 return 0; 1754 } 1755 1756 static int vm_object_in_map_callback(struct proc *p, void *data); 1757 1758 struct vm_object_in_map_info { 1759 vm_object_t object; 1760 int rv; 1761 }; 1762 1763 static int 1764 vm_object_in_map(vm_object_t object) 1765 { 1766 struct vm_object_in_map_info info; 1767 1768 info.rv = 0; 1769 info.object = object; 1770 1771 allproc_scan(vm_object_in_map_callback, &info); 1772 if (info.rv) 1773 return 1; 1774 if( _vm_object_in_map(&kernel_map, object, 0)) 1775 return 1; 1776 if( _vm_object_in_map(&pager_map, object, 0)) 1777 return 1; 1778 if( _vm_object_in_map(&buffer_map, object, 0)) 1779 return 1; 1780 return 0; 1781 } 1782 1783 static int 1784 vm_object_in_map_callback(struct proc *p, void *data) 1785 { 1786 struct vm_object_in_map_info *info = data; 1787 1788 if (p->p_vmspace) { 1789 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 1790 info->rv = 1; 1791 return -1; 1792 } 1793 } 1794 return (0); 1795 } 1796 1797 DB_SHOW_COMMAND(vmochk, vm_object_check) 1798 { 1799 vm_object_t object; 1800 1801 /* 1802 * make sure that internal objs are in a map somewhere 1803 * and none have zero ref counts. 1804 */ 1805 for (object = TAILQ_FIRST(&vm_object_list); 1806 object != NULL; 1807 object = TAILQ_NEXT(object, object_list)) { 1808 if (object->handle == NULL && 1809 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1810 if (object->ref_count == 0) { 1811 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1812 (long)object->size); 1813 } 1814 if (!vm_object_in_map(object)) { 1815 db_printf( 1816 "vmochk: internal obj is not in a map: " 1817 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1818 object->ref_count, (u_long)object->size, 1819 (u_long)object->size, 1820 (void *)object->backing_object); 1821 } 1822 } 1823 } 1824 } 1825 1826 /* 1827 * vm_object_print: [ debug ] 1828 */ 1829 DB_SHOW_COMMAND(object, vm_object_print_static) 1830 { 1831 /* XXX convert args. */ 1832 vm_object_t object = (vm_object_t)addr; 1833 boolean_t full = have_addr; 1834 1835 vm_page_t p; 1836 1837 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1838 #define count was_count 1839 1840 int count; 1841 1842 if (object == NULL) 1843 return; 1844 1845 db_iprintf( 1846 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 1847 object, (int)object->type, (u_long)object->size, 1848 object->resident_page_count, object->ref_count, object->flags); 1849 /* 1850 * XXX no %qd in kernel. Truncate object->backing_object_offset. 1851 */ 1852 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 1853 object->shadow_count, 1854 object->backing_object ? object->backing_object->ref_count : 0, 1855 object->backing_object, (long)object->backing_object_offset); 1856 1857 if (!full) 1858 return; 1859 1860 db_indent += 2; 1861 count = 0; 1862 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 1863 if (count == 0) 1864 db_iprintf("memory:="); 1865 else if (count == 6) { 1866 db_printf("\n"); 1867 db_iprintf(" ..."); 1868 count = 0; 1869 } else 1870 db_printf(","); 1871 count++; 1872 1873 db_printf("(off=0x%lx,page=0x%lx)", 1874 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 1875 } 1876 if (count != 0) 1877 db_printf("\n"); 1878 db_indent -= 2; 1879 } 1880 1881 /* XXX. */ 1882 #undef count 1883 1884 /* XXX need this non-static entry for calling from vm_map_print. */ 1885 void 1886 vm_object_print(/* db_expr_t */ long addr, 1887 boolean_t have_addr, 1888 /* db_expr_t */ long count, 1889 char *modif) 1890 { 1891 vm_object_print_static(addr, have_addr, count, modif); 1892 } 1893 1894 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 1895 { 1896 vm_object_t object; 1897 int nl = 0; 1898 int c; 1899 for (object = TAILQ_FIRST(&vm_object_list); 1900 object != NULL; 1901 object = TAILQ_NEXT(object, object_list)) { 1902 vm_pindex_t idx, fidx; 1903 vm_pindex_t osize; 1904 vm_paddr_t pa = -1, padiff; 1905 int rcount; 1906 vm_page_t m; 1907 1908 db_printf("new object: %p\n", (void *)object); 1909 if ( nl > 18) { 1910 c = cngetc(); 1911 if (c != ' ') 1912 return; 1913 nl = 0; 1914 } 1915 nl++; 1916 rcount = 0; 1917 fidx = 0; 1918 osize = object->size; 1919 if (osize > 128) 1920 osize = 128; 1921 for (idx = 0; idx < osize; idx++) { 1922 m = vm_page_lookup(object, idx); 1923 if (m == NULL) { 1924 if (rcount) { 1925 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 1926 (long)fidx, rcount, (long)pa); 1927 if ( nl > 18) { 1928 c = cngetc(); 1929 if (c != ' ') 1930 return; 1931 nl = 0; 1932 } 1933 nl++; 1934 rcount = 0; 1935 } 1936 continue; 1937 } 1938 1939 1940 if (rcount && 1941 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 1942 ++rcount; 1943 continue; 1944 } 1945 if (rcount) { 1946 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 1947 padiff >>= PAGE_SHIFT; 1948 padiff &= PQ_L2_MASK; 1949 if (padiff == 0) { 1950 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 1951 ++rcount; 1952 continue; 1953 } 1954 db_printf(" index(%ld)run(%d)pa(0x%lx)", 1955 (long)fidx, rcount, (long)pa); 1956 db_printf("pd(%ld)\n", (long)padiff); 1957 if ( nl > 18) { 1958 c = cngetc(); 1959 if (c != ' ') 1960 return; 1961 nl = 0; 1962 } 1963 nl++; 1964 } 1965 fidx = idx; 1966 pa = VM_PAGE_TO_PHYS(m); 1967 rcount = 1; 1968 } 1969 if (rcount) { 1970 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 1971 (long)fidx, rcount, (long)pa); 1972 if ( nl > 18) { 1973 c = cngetc(); 1974 if (c != ' ') 1975 return; 1976 nl = 0; 1977 } 1978 nl++; 1979 } 1980 } 1981 } 1982 #endif /* DDB */ 1983