1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 39 * 40 * 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 * 66 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $ 67 */ 68 69 /* 70 * Virtual memory object module. 71 */ 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/proc.h> /* for curproc, pageproc */ 76 #include <sys/thread.h> 77 #include <sys/vnode.h> 78 #include <sys/vmmeter.h> 79 #include <sys/mman.h> 80 #include <sys/mount.h> 81 #include <sys/kernel.h> 82 #include <sys/sysctl.h> 83 #include <sys/refcount.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pageout.h> 92 #include <vm/vm_pager.h> 93 #include <vm/swap_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/vm_zone.h> 97 98 #define EASY_SCAN_FACTOR 8 99 100 static void vm_object_qcollapse(vm_object_t object); 101 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 102 int pagerflags); 103 static void vm_object_lock_init(vm_object_t); 104 static void vm_object_hold_wake(vm_object_t); 105 static void vm_object_hold_wait(vm_object_t); 106 107 108 /* 109 * Virtual memory objects maintain the actual data 110 * associated with allocated virtual memory. A given 111 * page of memory exists within exactly one object. 112 * 113 * An object is only deallocated when all "references" 114 * are given up. Only one "reference" to a given 115 * region of an object should be writeable. 116 * 117 * Associated with each object is a list of all resident 118 * memory pages belonging to that object; this list is 119 * maintained by the "vm_page" module, and locked by the object's 120 * lock. 121 * 122 * Each object also records a "pager" routine which is 123 * used to retrieve (and store) pages to the proper backing 124 * storage. In addition, objects may be backed by other 125 * objects from which they were virtual-copied. 126 * 127 * The only items within the object structure which are 128 * modified after time of creation are: 129 * reference count locked by object's lock 130 * pager routine locked by object's lock 131 * 132 */ 133 134 struct object_q vm_object_list; /* locked by vmobj_token */ 135 struct vm_object kernel_object; 136 137 static long vm_object_count; /* locked by vmobj_token */ 138 extern int vm_pageout_page_count; 139 140 static long object_collapses; 141 static long object_bypasses; 142 static int next_index; 143 static vm_zone_t obj_zone; 144 static struct vm_zone obj_zone_store; 145 #define VM_OBJECTS_INIT 256 146 static struct vm_object vm_objects_init[VM_OBJECTS_INIT]; 147 148 /* 149 * Initialize a freshly allocated object 150 * 151 * Used only by vm_object_allocate() and zinitna(). 152 * 153 * No requirements. 154 */ 155 void 156 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 157 { 158 int incr; 159 160 RB_INIT(&object->rb_memq); 161 LIST_INIT(&object->shadow_head); 162 163 object->type = type; 164 object->size = size; 165 object->ref_count = 1; 166 object->hold_count = 0; 167 object->flags = 0; 168 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 169 vm_object_set_flag(object, OBJ_ONEMAPPING); 170 object->paging_in_progress = 0; 171 object->resident_page_count = 0; 172 object->agg_pv_list_count = 0; 173 object->shadow_count = 0; 174 object->pg_color = next_index; 175 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 176 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 177 else 178 incr = size; 179 next_index = (next_index + incr) & PQ_L2_MASK; 180 object->handle = NULL; 181 object->backing_object = NULL; 182 object->backing_object_offset = (vm_ooffset_t) 0; 183 184 object->generation++; 185 object->swblock_count = 0; 186 RB_INIT(&object->swblock_root); 187 vm_object_lock_init(object); 188 189 lwkt_gettoken(&vmobj_token); 190 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 191 vm_object_count++; 192 lwkt_reltoken(&vmobj_token); 193 } 194 195 /* 196 * Initialize the VM objects module. 197 * 198 * Called from the low level boot code only. 199 */ 200 void 201 vm_object_init(void) 202 { 203 TAILQ_INIT(&vm_object_list); 204 205 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd), 206 &kernel_object); 207 208 obj_zone = &obj_zone_store; 209 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object), 210 vm_objects_init, VM_OBJECTS_INIT); 211 } 212 213 void 214 vm_object_init2(void) 215 { 216 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1); 217 } 218 219 /* 220 * Allocate and return a new object of the specified type and size. 221 * 222 * No requirements. 223 */ 224 vm_object_t 225 vm_object_allocate(objtype_t type, vm_pindex_t size) 226 { 227 vm_object_t result; 228 229 result = (vm_object_t) zalloc(obj_zone); 230 231 _vm_object_allocate(type, size, result); 232 233 return (result); 234 } 235 236 /* 237 * Add an additional reference to a vm_object. 238 * 239 * Object passed by caller must be stable or caller must already 240 * hold vmobj_token to avoid races. 241 */ 242 void 243 vm_object_reference(vm_object_t object) 244 { 245 lwkt_gettoken(&vmobj_token); 246 vm_object_reference_locked(object); 247 lwkt_reltoken(&vmobj_token); 248 } 249 250 void 251 vm_object_reference_locked(vm_object_t object) 252 { 253 if (object) { 254 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 255 object->ref_count++; 256 if (object->type == OBJT_VNODE) { 257 vref(object->handle); 258 /* XXX what if the vnode is being destroyed? */ 259 } 260 } 261 } 262 263 /* 264 * Dereference an object and its underlying vnode. 265 * 266 * The caller must hold vmobj_token. 267 */ 268 static void 269 vm_object_vndeallocate(vm_object_t object) 270 { 271 struct vnode *vp = (struct vnode *) object->handle; 272 273 KASSERT(object->type == OBJT_VNODE, 274 ("vm_object_vndeallocate: not a vnode object")); 275 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 276 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 277 #ifdef INVARIANTS 278 if (object->ref_count == 0) { 279 vprint("vm_object_vndeallocate", vp); 280 panic("vm_object_vndeallocate: bad object reference count"); 281 } 282 #endif 283 284 object->ref_count--; 285 if (object->ref_count == 0) 286 vclrflags(vp, VTEXT); 287 vrele(vp); 288 } 289 290 /* 291 * Release a reference to the specified object, gained either through a 292 * vm_object_allocate or a vm_object_reference call. When all references 293 * are gone, storage associated with this object may be relinquished. 294 */ 295 void 296 vm_object_deallocate(vm_object_t object) 297 { 298 lwkt_gettoken(&vmobj_token); 299 vm_object_deallocate_locked(object); 300 lwkt_reltoken(&vmobj_token); 301 } 302 303 void 304 vm_object_deallocate_locked(vm_object_t object) 305 { 306 vm_object_t temp; 307 308 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 309 310 while (object != NULL) { 311 if (object->type == OBJT_VNODE) { 312 vm_object_vndeallocate(object); 313 break; 314 } 315 316 if (object->ref_count == 0) { 317 panic("vm_object_deallocate: object deallocated " 318 "too many times: %d", object->type); 319 } 320 if (object->ref_count > 2) { 321 object->ref_count--; 322 break; 323 } 324 325 /* 326 * We currently need the vm_token from this point on, and 327 * we must recheck ref_count after acquiring it. 328 */ 329 lwkt_gettoken(&vm_token); 330 331 if (object->ref_count > 2) { 332 object->ref_count--; 333 lwkt_reltoken(&vm_token); 334 break; 335 } 336 337 /* 338 * Here on ref_count of one or two, which are special cases for 339 * objects. 340 */ 341 if ((object->ref_count == 2) && (object->shadow_count == 0)) { 342 vm_object_set_flag(object, OBJ_ONEMAPPING); 343 object->ref_count--; 344 lwkt_reltoken(&vm_token); 345 break; 346 } 347 if ((object->ref_count == 2) && (object->shadow_count == 1)) { 348 object->ref_count--; 349 if ((object->handle == NULL) && 350 (object->type == OBJT_DEFAULT || 351 object->type == OBJT_SWAP)) { 352 vm_object_t robject; 353 354 robject = LIST_FIRST(&object->shadow_head); 355 KASSERT(robject != NULL, 356 ("vm_object_deallocate: ref_count: " 357 "%d, shadow_count: %d", 358 object->ref_count, 359 object->shadow_count)); 360 361 if ((robject->handle == NULL) && 362 (robject->type == OBJT_DEFAULT || 363 robject->type == OBJT_SWAP)) { 364 365 robject->ref_count++; 366 367 while ( 368 robject->paging_in_progress || 369 object->paging_in_progress 370 ) { 371 vm_object_pip_sleep(robject, "objde1"); 372 vm_object_pip_sleep(object, "objde2"); 373 } 374 375 if (robject->ref_count == 1) { 376 robject->ref_count--; 377 object = robject; 378 goto doterm; 379 } 380 381 object = robject; 382 vm_object_collapse(object); 383 lwkt_reltoken(&vm_token); 384 continue; 385 } 386 } 387 lwkt_reltoken(&vm_token); 388 break; 389 } 390 391 /* 392 * Normal dereferencing path 393 */ 394 object->ref_count--; 395 if (object->ref_count != 0) { 396 lwkt_reltoken(&vm_token); 397 break; 398 } 399 400 /* 401 * Termination path 402 */ 403 doterm: 404 temp = object->backing_object; 405 if (temp) { 406 LIST_REMOVE(object, shadow_list); 407 temp->shadow_count--; 408 temp->generation++; 409 object->backing_object = NULL; 410 } 411 lwkt_reltoken(&vm_token); 412 413 /* 414 * Don't double-terminate, we could be in a termination 415 * recursion due to the terminate having to sync data 416 * to disk. 417 */ 418 if ((object->flags & OBJ_DEAD) == 0) 419 vm_object_terminate(object); 420 object = temp; 421 } 422 } 423 424 /* 425 * Destroy the specified object, freeing up related resources. 426 * 427 * The object must have zero references. 428 * 429 * The caller must be holding vmobj_token and properly interlock with 430 * OBJ_DEAD. 431 */ 432 static int vm_object_terminate_callback(vm_page_t p, void *data); 433 434 void 435 vm_object_terminate(vm_object_t object) 436 { 437 /* 438 * Make sure no one uses us. Once we set OBJ_DEAD we should be 439 * able to safely block. 440 */ 441 KKASSERT((object->flags & OBJ_DEAD) == 0); 442 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 443 vm_object_set_flag(object, OBJ_DEAD); 444 445 /* 446 * Wait for the pageout daemon to be done with the object 447 */ 448 vm_object_pip_wait(object, "objtrm1"); 449 450 KASSERT(!object->paging_in_progress, 451 ("vm_object_terminate: pageout in progress")); 452 453 /* 454 * Clean and free the pages, as appropriate. All references to the 455 * object are gone, so we don't need to lock it. 456 */ 457 if (object->type == OBJT_VNODE) { 458 struct vnode *vp; 459 460 /* 461 * Clean pages and flush buffers. 462 */ 463 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 464 465 vp = (struct vnode *) object->handle; 466 vinvalbuf(vp, V_SAVE, 0, 0); 467 } 468 469 /* 470 * Wait for any I/O to complete, after which there had better not 471 * be any references left on the object. 472 */ 473 vm_object_pip_wait(object, "objtrm2"); 474 475 if (object->ref_count != 0) { 476 panic("vm_object_terminate: object with references, " 477 "ref_count=%d", object->ref_count); 478 } 479 480 /* 481 * Now free any remaining pages. For internal objects, this also 482 * removes them from paging queues. Don't free wired pages, just 483 * remove them from the object. 484 */ 485 lwkt_gettoken(&vm_token); 486 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 487 vm_object_terminate_callback, NULL); 488 lwkt_reltoken(&vm_token); 489 490 /* 491 * Let the pager know object is dead. 492 */ 493 vm_pager_deallocate(object); 494 495 /* 496 * Wait for the object hold count to hit zero, clean out pages as 497 * we go. 498 */ 499 lwkt_gettoken(&vm_token); 500 for (;;) { 501 vm_object_hold_wait(object); 502 if (RB_ROOT(&object->rb_memq) == NULL) 503 break; 504 kprintf("vm_object_terminate: Warning, object %p " 505 "still has %d pages\n", 506 object, object->resident_page_count); 507 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 508 vm_object_terminate_callback, NULL); 509 } 510 lwkt_reltoken(&vm_token); 511 512 /* 513 * There had better not be any pages left 514 */ 515 KKASSERT(object->resident_page_count == 0); 516 517 /* 518 * Remove the object from the global object list. 519 * 520 * (we are holding vmobj_token) 521 */ 522 TAILQ_REMOVE(&vm_object_list, object, object_list); 523 vm_object_count--; 524 vm_object_dead_wakeup(object); 525 526 if (object->ref_count != 0) { 527 panic("vm_object_terminate2: object with references, " 528 "ref_count=%d", object->ref_count); 529 } 530 531 /* 532 * Free the space for the object. 533 */ 534 zfree(obj_zone, object); 535 } 536 537 /* 538 * The caller must hold vm_token. 539 */ 540 static int 541 vm_object_terminate_callback(vm_page_t p, void *data __unused) 542 { 543 if (p->busy || (p->flags & PG_BUSY)) 544 panic("vm_object_terminate: freeing busy page %p", p); 545 if (p->wire_count == 0) { 546 vm_page_busy(p); 547 vm_page_free(p); 548 mycpu->gd_cnt.v_pfree++; 549 } else { 550 if (p->queue != PQ_NONE) 551 kprintf("vm_object_terminate: Warning: Encountered wired page %p on queue %d\n", p, p->queue); 552 vm_page_busy(p); 553 vm_page_remove(p); 554 vm_page_wakeup(p); 555 } 556 return(0); 557 } 558 559 /* 560 * The object is dead but still has an object<->pager association. Sleep 561 * and return. The caller typically retests the association in a loop. 562 * 563 * Must be called with the vmobj_token held. 564 */ 565 void 566 vm_object_dead_sleep(vm_object_t object, const char *wmesg) 567 { 568 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 569 if (object->handle) { 570 vm_object_set_flag(object, OBJ_DEADWNT); 571 tsleep(object, 0, wmesg, 0); 572 /* object may be invalid after this point */ 573 } 574 } 575 576 /* 577 * Wakeup anyone waiting for the object<->pager disassociation on 578 * a dead object. 579 * 580 * Must be called with the vmobj_token held. 581 */ 582 void 583 vm_object_dead_wakeup(vm_object_t object) 584 { 585 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 586 if (object->flags & OBJ_DEADWNT) { 587 vm_object_clear_flag(object, OBJ_DEADWNT); 588 wakeup(object); 589 } 590 } 591 592 /* 593 * Clean all dirty pages in the specified range of object. Leaves page 594 * on whatever queue it is currently on. If NOSYNC is set then do not 595 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 596 * leaving the object dirty. 597 * 598 * When stuffing pages asynchronously, allow clustering. XXX we need a 599 * synchronous clustering mode implementation. 600 * 601 * Odd semantics: if start == end, we clean everything. 602 * 603 * The object must be locked? XXX 604 */ 605 static int vm_object_page_clean_pass1(struct vm_page *p, void *data); 606 static int vm_object_page_clean_pass2(struct vm_page *p, void *data); 607 608 void 609 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 610 int flags) 611 { 612 struct rb_vm_page_scan_info info; 613 struct vnode *vp; 614 int wholescan; 615 int pagerflags; 616 int curgeneration; 617 618 lwkt_gettoken(&vm_token); 619 if (object->type != OBJT_VNODE || 620 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 621 lwkt_reltoken(&vm_token); 622 return; 623 } 624 625 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 626 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 627 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 628 629 vp = object->handle; 630 631 /* 632 * Interlock other major object operations. This allows us to 633 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY. 634 */ 635 crit_enter(); 636 vm_object_set_flag(object, OBJ_CLEANING); 637 638 /* 639 * Handle 'entire object' case 640 */ 641 info.start_pindex = start; 642 if (end == 0) { 643 info.end_pindex = object->size - 1; 644 } else { 645 info.end_pindex = end - 1; 646 } 647 wholescan = (start == 0 && info.end_pindex == object->size - 1); 648 info.limit = flags; 649 info.pagerflags = pagerflags; 650 info.object = object; 651 652 /* 653 * If cleaning the entire object do a pass to mark the pages read-only. 654 * If everything worked out ok, clear OBJ_WRITEABLE and 655 * OBJ_MIGHTBEDIRTY. 656 */ 657 if (wholescan) { 658 info.error = 0; 659 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 660 vm_object_page_clean_pass1, &info); 661 if (info.error == 0) { 662 vm_object_clear_flag(object, 663 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 664 if (object->type == OBJT_VNODE && 665 (vp = (struct vnode *)object->handle) != NULL) { 666 if (vp->v_flag & VOBJDIRTY) 667 vclrflags(vp, VOBJDIRTY); 668 } 669 } 670 } 671 672 /* 673 * Do a pass to clean all the dirty pages we find. 674 */ 675 do { 676 info.error = 0; 677 curgeneration = object->generation; 678 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 679 vm_object_page_clean_pass2, &info); 680 } while (info.error || curgeneration != object->generation); 681 682 vm_object_clear_flag(object, OBJ_CLEANING); 683 crit_exit(); 684 lwkt_reltoken(&vm_token); 685 } 686 687 /* 688 * The caller must hold vm_token. 689 */ 690 static 691 int 692 vm_object_page_clean_pass1(struct vm_page *p, void *data) 693 { 694 struct rb_vm_page_scan_info *info = data; 695 696 vm_page_flag_set(p, PG_CLEANCHK); 697 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 698 info->error = 1; 699 else 700 vm_page_protect(p, VM_PROT_READ); /* must not block */ 701 return(0); 702 } 703 704 /* 705 * The caller must hold vm_token. 706 */ 707 static 708 int 709 vm_object_page_clean_pass2(struct vm_page *p, void *data) 710 { 711 struct rb_vm_page_scan_info *info = data; 712 int n; 713 714 /* 715 * Do not mess with pages that were inserted after we started 716 * the cleaning pass. 717 */ 718 if ((p->flags & PG_CLEANCHK) == 0) 719 return(0); 720 721 /* 722 * Before wasting time traversing the pmaps, check for trivial 723 * cases where the page cannot be dirty. 724 */ 725 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 726 KKASSERT((p->dirty & p->valid) == 0); 727 return(0); 728 } 729 730 /* 731 * Check whether the page is dirty or not. The page has been set 732 * to be read-only so the check will not race a user dirtying the 733 * page. 734 */ 735 vm_page_test_dirty(p); 736 if ((p->dirty & p->valid) == 0) { 737 vm_page_flag_clear(p, PG_CLEANCHK); 738 return(0); 739 } 740 741 /* 742 * If we have been asked to skip nosync pages and this is a 743 * nosync page, skip it. Note that the object flags were 744 * not cleared in this case (because pass1 will have returned an 745 * error), so we do not have to set them. 746 */ 747 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 748 vm_page_flag_clear(p, PG_CLEANCHK); 749 return(0); 750 } 751 752 /* 753 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 754 * the pages that get successfully flushed. Set info->error if 755 * we raced an object modification. 756 */ 757 n = vm_object_page_collect_flush(info->object, p, info->pagerflags); 758 if (n == 0) 759 info->error = 1; 760 return(0); 761 } 762 763 /* 764 * Collect the specified page and nearby pages and flush them out. 765 * The number of pages flushed is returned. 766 * 767 * The caller must hold vm_token. 768 */ 769 static int 770 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 771 { 772 int runlen; 773 int maxf; 774 int chkb; 775 int maxb; 776 int i; 777 int curgeneration; 778 vm_pindex_t pi; 779 vm_page_t maf[vm_pageout_page_count]; 780 vm_page_t mab[vm_pageout_page_count]; 781 vm_page_t ma[vm_pageout_page_count]; 782 783 curgeneration = object->generation; 784 785 pi = p->pindex; 786 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 787 if (object->generation != curgeneration) { 788 return(0); 789 } 790 } 791 KKASSERT(p->object == object && p->pindex == pi); 792 793 maxf = 0; 794 for(i = 1; i < vm_pageout_page_count; i++) { 795 vm_page_t tp; 796 797 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 798 if ((tp->flags & PG_BUSY) || 799 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 800 (tp->flags & PG_CLEANCHK) == 0) || 801 (tp->busy != 0)) 802 break; 803 if((tp->queue - tp->pc) == PQ_CACHE) { 804 vm_page_flag_clear(tp, PG_CLEANCHK); 805 break; 806 } 807 vm_page_test_dirty(tp); 808 if ((tp->dirty & tp->valid) == 0) { 809 vm_page_flag_clear(tp, PG_CLEANCHK); 810 break; 811 } 812 maf[ i - 1 ] = tp; 813 maxf++; 814 continue; 815 } 816 break; 817 } 818 819 maxb = 0; 820 chkb = vm_pageout_page_count - maxf; 821 if (chkb) { 822 for(i = 1; i < chkb;i++) { 823 vm_page_t tp; 824 825 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 826 if ((tp->flags & PG_BUSY) || 827 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 828 (tp->flags & PG_CLEANCHK) == 0) || 829 (tp->busy != 0)) 830 break; 831 if((tp->queue - tp->pc) == PQ_CACHE) { 832 vm_page_flag_clear(tp, PG_CLEANCHK); 833 break; 834 } 835 vm_page_test_dirty(tp); 836 if ((tp->dirty & tp->valid) == 0) { 837 vm_page_flag_clear(tp, PG_CLEANCHK); 838 break; 839 } 840 mab[ i - 1 ] = tp; 841 maxb++; 842 continue; 843 } 844 break; 845 } 846 } 847 848 for(i = 0; i < maxb; i++) { 849 int index = (maxb - i) - 1; 850 ma[index] = mab[i]; 851 vm_page_flag_clear(ma[index], PG_CLEANCHK); 852 } 853 vm_page_flag_clear(p, PG_CLEANCHK); 854 ma[maxb] = p; 855 for(i = 0; i < maxf; i++) { 856 int index = (maxb + i) + 1; 857 ma[index] = maf[i]; 858 vm_page_flag_clear(ma[index], PG_CLEANCHK); 859 } 860 runlen = maxb + maxf + 1; 861 862 vm_pageout_flush(ma, runlen, pagerflags); 863 for (i = 0; i < runlen; i++) { 864 if (ma[i]->valid & ma[i]->dirty) { 865 vm_page_protect(ma[i], VM_PROT_READ); 866 vm_page_flag_set(ma[i], PG_CLEANCHK); 867 868 /* 869 * maxf will end up being the actual number of pages 870 * we wrote out contiguously, non-inclusive of the 871 * first page. We do not count look-behind pages. 872 */ 873 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 874 maxf = i - maxb - 1; 875 } 876 } 877 return(maxf + 1); 878 } 879 880 /* 881 * Same as vm_object_pmap_copy, except range checking really 882 * works, and is meant for small sections of an object. 883 * 884 * This code protects resident pages by making them read-only 885 * and is typically called on a fork or split when a page 886 * is converted to copy-on-write. 887 * 888 * NOTE: If the page is already at VM_PROT_NONE, calling 889 * vm_page_protect will have no effect. 890 */ 891 void 892 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 893 { 894 vm_pindex_t idx; 895 vm_page_t p; 896 897 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 898 return; 899 900 /* 901 * spl protection needed to prevent races between the lookup, 902 * an interrupt unbusy/free, and our protect call. 903 */ 904 crit_enter(); 905 lwkt_gettoken(&vm_token); 906 for (idx = start; idx < end; idx++) { 907 p = vm_page_lookup(object, idx); 908 if (p == NULL) 909 continue; 910 vm_page_protect(p, VM_PROT_READ); 911 } 912 lwkt_reltoken(&vm_token); 913 crit_exit(); 914 } 915 916 /* 917 * Removes all physical pages in the specified object range from all 918 * physical maps. 919 * 920 * The object must *not* be locked. 921 */ 922 923 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 924 925 void 926 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 927 { 928 struct rb_vm_page_scan_info info; 929 930 if (object == NULL) 931 return; 932 info.start_pindex = start; 933 info.end_pindex = end - 1; 934 935 crit_enter(); 936 lwkt_gettoken(&vm_token); 937 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 938 vm_object_pmap_remove_callback, &info); 939 if (start == 0 && end == object->size) 940 vm_object_clear_flag(object, OBJ_WRITEABLE); 941 lwkt_reltoken(&vm_token); 942 crit_exit(); 943 } 944 945 /* 946 * The caller must hold vm_token. 947 */ 948 static int 949 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 950 { 951 vm_page_protect(p, VM_PROT_NONE); 952 return(0); 953 } 954 955 /* 956 * Implements the madvise function at the object/page level. 957 * 958 * MADV_WILLNEED (any object) 959 * 960 * Activate the specified pages if they are resident. 961 * 962 * MADV_DONTNEED (any object) 963 * 964 * Deactivate the specified pages if they are resident. 965 * 966 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 967 * 968 * Deactivate and clean the specified pages if they are 969 * resident. This permits the process to reuse the pages 970 * without faulting or the kernel to reclaim the pages 971 * without I/O. 972 * 973 * No requirements. 974 */ 975 void 976 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 977 { 978 vm_pindex_t end, tpindex; 979 vm_object_t tobject; 980 vm_page_t m; 981 982 if (object == NULL) 983 return; 984 985 end = pindex + count; 986 987 lwkt_gettoken(&vm_token); 988 989 /* 990 * Locate and adjust resident pages 991 */ 992 for (; pindex < end; pindex += 1) { 993 relookup: 994 tobject = object; 995 tpindex = pindex; 996 shadowlookup: 997 /* 998 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 999 * and those pages must be OBJ_ONEMAPPING. 1000 */ 1001 if (advise == MADV_FREE) { 1002 if ((tobject->type != OBJT_DEFAULT && 1003 tobject->type != OBJT_SWAP) || 1004 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1005 continue; 1006 } 1007 } 1008 1009 /* 1010 * spl protection is required to avoid a race between the 1011 * lookup, an interrupt unbusy/free, and our busy check. 1012 */ 1013 1014 crit_enter(); 1015 m = vm_page_lookup(tobject, tpindex); 1016 1017 if (m == NULL) { 1018 /* 1019 * There may be swap even if there is no backing page 1020 */ 1021 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1022 swap_pager_freespace(tobject, tpindex, 1); 1023 1024 /* 1025 * next object 1026 */ 1027 crit_exit(); 1028 if (tobject->backing_object == NULL) 1029 continue; 1030 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1031 tobject = tobject->backing_object; 1032 goto shadowlookup; 1033 } 1034 1035 /* 1036 * If the page is busy or not in a normal active state, 1037 * we skip it. If the page is not managed there are no 1038 * page queues to mess with. Things can break if we mess 1039 * with pages in any of the below states. 1040 */ 1041 if ( 1042 m->hold_count || 1043 m->wire_count || 1044 (m->flags & PG_UNMANAGED) || 1045 m->valid != VM_PAGE_BITS_ALL 1046 ) { 1047 crit_exit(); 1048 continue; 1049 } 1050 1051 if (vm_page_sleep_busy(m, TRUE, "madvpo")) { 1052 crit_exit(); 1053 goto relookup; 1054 } 1055 vm_page_busy(m); 1056 crit_exit(); 1057 1058 /* 1059 * Theoretically once a page is known not to be busy, an 1060 * interrupt cannot come along and rip it out from under us. 1061 */ 1062 1063 if (advise == MADV_WILLNEED) { 1064 vm_page_activate(m); 1065 } else if (advise == MADV_DONTNEED) { 1066 vm_page_dontneed(m); 1067 } else if (advise == MADV_FREE) { 1068 /* 1069 * Mark the page clean. This will allow the page 1070 * to be freed up by the system. However, such pages 1071 * are often reused quickly by malloc()/free() 1072 * so we do not do anything that would cause 1073 * a page fault if we can help it. 1074 * 1075 * Specifically, we do not try to actually free 1076 * the page now nor do we try to put it in the 1077 * cache (which would cause a page fault on reuse). 1078 * 1079 * But we do make the page is freeable as we 1080 * can without actually taking the step of unmapping 1081 * it. 1082 */ 1083 pmap_clear_modify(m); 1084 m->dirty = 0; 1085 m->act_count = 0; 1086 vm_page_dontneed(m); 1087 if (tobject->type == OBJT_SWAP) 1088 swap_pager_freespace(tobject, tpindex, 1); 1089 } 1090 vm_page_wakeup(m); 1091 } 1092 lwkt_reltoken(&vm_token); 1093 } 1094 1095 /* 1096 * Create a new object which is backed by the specified existing object 1097 * range. The source object reference is deallocated. 1098 * 1099 * The new object and offset into that object are returned in the source 1100 * parameters. 1101 * 1102 * No other requirements. 1103 */ 1104 void 1105 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length) 1106 { 1107 vm_object_t source; 1108 vm_object_t result; 1109 1110 source = *object; 1111 1112 /* 1113 * Don't create the new object if the old object isn't shared. 1114 */ 1115 lwkt_gettoken(&vm_token); 1116 1117 if (source != NULL && 1118 source->ref_count == 1 && 1119 source->handle == NULL && 1120 (source->type == OBJT_DEFAULT || 1121 source->type == OBJT_SWAP)) { 1122 lwkt_reltoken(&vm_token); 1123 return; 1124 } 1125 1126 /* 1127 * Allocate a new object with the given length 1128 */ 1129 1130 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL) 1131 panic("vm_object_shadow: no object for shadowing"); 1132 1133 /* 1134 * The new object shadows the source object, adding a reference to it. 1135 * Our caller changes his reference to point to the new object, 1136 * removing a reference to the source object. Net result: no change 1137 * of reference count. 1138 * 1139 * Try to optimize the result object's page color when shadowing 1140 * in order to maintain page coloring consistency in the combined 1141 * shadowed object. 1142 */ 1143 result->backing_object = source; 1144 if (source) { 1145 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1146 source->shadow_count++; 1147 source->generation++; 1148 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK; 1149 } 1150 1151 /* 1152 * Store the offset into the source object, and fix up the offset into 1153 * the new object. 1154 */ 1155 result->backing_object_offset = *offset; 1156 lwkt_reltoken(&vm_token); 1157 1158 /* 1159 * Return the new things 1160 */ 1161 *offset = 0; 1162 *object = result; 1163 } 1164 1165 #define OBSC_TEST_ALL_SHADOWED 0x0001 1166 #define OBSC_COLLAPSE_NOWAIT 0x0002 1167 #define OBSC_COLLAPSE_WAIT 0x0004 1168 1169 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1170 1171 /* 1172 * The caller must hold vm_token. 1173 */ 1174 static __inline int 1175 vm_object_backing_scan(vm_object_t object, int op) 1176 { 1177 struct rb_vm_page_scan_info info; 1178 vm_object_t backing_object; 1179 1180 crit_enter(); 1181 1182 backing_object = object->backing_object; 1183 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1184 1185 /* 1186 * Initial conditions 1187 */ 1188 1189 if (op & OBSC_TEST_ALL_SHADOWED) { 1190 /* 1191 * We do not want to have to test for the existence of 1192 * swap pages in the backing object. XXX but with the 1193 * new swapper this would be pretty easy to do. 1194 * 1195 * XXX what about anonymous MAP_SHARED memory that hasn't 1196 * been ZFOD faulted yet? If we do not test for this, the 1197 * shadow test may succeed! XXX 1198 */ 1199 if (backing_object->type != OBJT_DEFAULT) { 1200 crit_exit(); 1201 return(0); 1202 } 1203 } 1204 if (op & OBSC_COLLAPSE_WAIT) { 1205 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1206 vm_object_set_flag(backing_object, OBJ_DEAD); 1207 } 1208 1209 /* 1210 * Our scan. We have to retry if a negative error code is returned, 1211 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1212 * the scan had to be stopped because the parent does not completely 1213 * shadow the child. 1214 */ 1215 info.object = object; 1216 info.backing_object = backing_object; 1217 info.limit = op; 1218 do { 1219 info.error = 1; 1220 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1221 vm_object_backing_scan_callback, 1222 &info); 1223 } while (info.error < 0); 1224 crit_exit(); 1225 return(info.error); 1226 } 1227 1228 /* 1229 * The caller must hold vm_token. 1230 */ 1231 static int 1232 vm_object_backing_scan_callback(vm_page_t p, void *data) 1233 { 1234 struct rb_vm_page_scan_info *info = data; 1235 vm_object_t backing_object; 1236 vm_object_t object; 1237 vm_pindex_t new_pindex; 1238 vm_pindex_t backing_offset_index; 1239 int op; 1240 1241 new_pindex = p->pindex - info->backing_offset_index; 1242 op = info->limit; 1243 object = info->object; 1244 backing_object = info->backing_object; 1245 backing_offset_index = info->backing_offset_index; 1246 1247 if (op & OBSC_TEST_ALL_SHADOWED) { 1248 vm_page_t pp; 1249 1250 /* 1251 * Ignore pages outside the parent object's range 1252 * and outside the parent object's mapping of the 1253 * backing object. 1254 * 1255 * note that we do not busy the backing object's 1256 * page. 1257 */ 1258 if ( 1259 p->pindex < backing_offset_index || 1260 new_pindex >= object->size 1261 ) { 1262 return(0); 1263 } 1264 1265 /* 1266 * See if the parent has the page or if the parent's 1267 * object pager has the page. If the parent has the 1268 * page but the page is not valid, the parent's 1269 * object pager must have the page. 1270 * 1271 * If this fails, the parent does not completely shadow 1272 * the object and we might as well give up now. 1273 */ 1274 1275 pp = vm_page_lookup(object, new_pindex); 1276 if ((pp == NULL || pp->valid == 0) && 1277 !vm_pager_has_page(object, new_pindex) 1278 ) { 1279 info->error = 0; /* problemo */ 1280 return(-1); /* stop the scan */ 1281 } 1282 } 1283 1284 /* 1285 * Check for busy page 1286 */ 1287 1288 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1289 vm_page_t pp; 1290 1291 if (op & OBSC_COLLAPSE_NOWAIT) { 1292 if ( 1293 (p->flags & PG_BUSY) || 1294 !p->valid || 1295 p->hold_count || 1296 p->wire_count || 1297 p->busy 1298 ) { 1299 return(0); 1300 } 1301 } else if (op & OBSC_COLLAPSE_WAIT) { 1302 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1303 /* 1304 * If we slept, anything could have 1305 * happened. Ask that the scan be restarted. 1306 * 1307 * Since the object is marked dead, the 1308 * backing offset should not have changed. 1309 */ 1310 info->error = -1; 1311 return(-1); 1312 } 1313 } 1314 1315 /* 1316 * Busy the page 1317 */ 1318 vm_page_busy(p); 1319 1320 KASSERT( 1321 p->object == backing_object, 1322 ("vm_object_qcollapse(): object mismatch") 1323 ); 1324 1325 /* 1326 * Destroy any associated swap 1327 */ 1328 if (backing_object->type == OBJT_SWAP) 1329 swap_pager_freespace(backing_object, p->pindex, 1); 1330 1331 if ( 1332 p->pindex < backing_offset_index || 1333 new_pindex >= object->size 1334 ) { 1335 /* 1336 * Page is out of the parent object's range, we 1337 * can simply destroy it. 1338 */ 1339 vm_page_protect(p, VM_PROT_NONE); 1340 vm_page_free(p); 1341 return(0); 1342 } 1343 1344 pp = vm_page_lookup(object, new_pindex); 1345 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 1346 /* 1347 * page already exists in parent OR swap exists 1348 * for this location in the parent. Destroy 1349 * the original page from the backing object. 1350 * 1351 * Leave the parent's page alone 1352 */ 1353 vm_page_protect(p, VM_PROT_NONE); 1354 vm_page_free(p); 1355 return(0); 1356 } 1357 1358 /* 1359 * Page does not exist in parent, rename the 1360 * page from the backing object to the main object. 1361 * 1362 * If the page was mapped to a process, it can remain 1363 * mapped through the rename. 1364 */ 1365 if ((p->queue - p->pc) == PQ_CACHE) 1366 vm_page_deactivate(p); 1367 1368 vm_page_rename(p, object, new_pindex); 1369 /* page automatically made dirty by rename */ 1370 } 1371 return(0); 1372 } 1373 1374 /* 1375 * This version of collapse allows the operation to occur earlier and 1376 * when paging_in_progress is true for an object... This is not a complete 1377 * operation, but should plug 99.9% of the rest of the leaks. 1378 * 1379 * The caller must hold vm_token and vmobj_token. 1380 * (only called from vm_object_collapse) 1381 */ 1382 static void 1383 vm_object_qcollapse(vm_object_t object) 1384 { 1385 vm_object_t backing_object = object->backing_object; 1386 1387 if (backing_object->ref_count != 1) 1388 return; 1389 1390 backing_object->ref_count += 2; 1391 1392 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1393 1394 backing_object->ref_count -= 2; 1395 } 1396 1397 /* 1398 * Collapse an object with the object backing it. Pages in the backing 1399 * object are moved into the parent, and the backing object is deallocated. 1400 */ 1401 void 1402 vm_object_collapse(vm_object_t object) 1403 { 1404 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1405 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 1406 1407 while (TRUE) { 1408 vm_object_t backing_object; 1409 1410 /* 1411 * Verify that the conditions are right for collapse: 1412 * 1413 * The object exists and the backing object exists. 1414 */ 1415 if (object == NULL) 1416 break; 1417 1418 if ((backing_object = object->backing_object) == NULL) 1419 break; 1420 1421 /* 1422 * we check the backing object first, because it is most likely 1423 * not collapsable. 1424 */ 1425 if (backing_object->handle != NULL || 1426 (backing_object->type != OBJT_DEFAULT && 1427 backing_object->type != OBJT_SWAP) || 1428 (backing_object->flags & OBJ_DEAD) || 1429 object->handle != NULL || 1430 (object->type != OBJT_DEFAULT && 1431 object->type != OBJT_SWAP) || 1432 (object->flags & OBJ_DEAD)) { 1433 break; 1434 } 1435 1436 if ( 1437 object->paging_in_progress != 0 || 1438 backing_object->paging_in_progress != 0 1439 ) { 1440 vm_object_qcollapse(object); 1441 break; 1442 } 1443 1444 /* 1445 * We know that we can either collapse the backing object (if 1446 * the parent is the only reference to it) or (perhaps) have 1447 * the parent bypass the object if the parent happens to shadow 1448 * all the resident pages in the entire backing object. 1449 * 1450 * This is ignoring pager-backed pages such as swap pages. 1451 * vm_object_backing_scan fails the shadowing test in this 1452 * case. 1453 */ 1454 1455 if (backing_object->ref_count == 1) { 1456 /* 1457 * If there is exactly one reference to the backing 1458 * object, we can collapse it into the parent. 1459 */ 1460 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1461 1462 /* 1463 * Move the pager from backing_object to object. 1464 */ 1465 1466 if (backing_object->type == OBJT_SWAP) { 1467 vm_object_pip_add(backing_object, 1); 1468 1469 /* 1470 * scrap the paging_offset junk and do a 1471 * discrete copy. This also removes major 1472 * assumptions about how the swap-pager 1473 * works from where it doesn't belong. The 1474 * new swapper is able to optimize the 1475 * destroy-source case. 1476 */ 1477 1478 vm_object_pip_add(object, 1); 1479 swap_pager_copy( 1480 backing_object, 1481 object, 1482 OFF_TO_IDX(object->backing_object_offset), TRUE); 1483 vm_object_pip_wakeup(object); 1484 1485 vm_object_pip_wakeup(backing_object); 1486 } 1487 /* 1488 * Object now shadows whatever backing_object did. 1489 * Note that the reference to 1490 * backing_object->backing_object moves from within 1491 * backing_object to within object. 1492 */ 1493 1494 LIST_REMOVE(object, shadow_list); 1495 object->backing_object->shadow_count--; 1496 object->backing_object->generation++; 1497 if (backing_object->backing_object) { 1498 LIST_REMOVE(backing_object, shadow_list); 1499 backing_object->backing_object->shadow_count--; 1500 backing_object->backing_object->generation++; 1501 } 1502 object->backing_object = backing_object->backing_object; 1503 if (object->backing_object) { 1504 LIST_INSERT_HEAD( 1505 &object->backing_object->shadow_head, 1506 object, 1507 shadow_list 1508 ); 1509 object->backing_object->shadow_count++; 1510 object->backing_object->generation++; 1511 } 1512 1513 object->backing_object_offset += 1514 backing_object->backing_object_offset; 1515 1516 /* 1517 * Discard backing_object. 1518 * 1519 * Since the backing object has no pages, no pager left, 1520 * and no object references within it, all that is 1521 * necessary is to dispose of it. 1522 */ 1523 1524 KASSERT(backing_object->ref_count == 1, 1525 ("backing_object %p was somehow " 1526 "re-referenced during collapse!", 1527 backing_object)); 1528 KASSERT(RB_EMPTY(&backing_object->rb_memq), 1529 ("backing_object %p somehow has left " 1530 "over pages during collapse!", 1531 backing_object)); 1532 1533 /* 1534 * Wait for hold count to hit zero 1535 */ 1536 vm_object_hold_wait(backing_object); 1537 1538 /* (we are holding vmobj_token) */ 1539 TAILQ_REMOVE(&vm_object_list, backing_object, 1540 object_list); 1541 vm_object_count--; 1542 1543 zfree(obj_zone, backing_object); 1544 1545 object_collapses++; 1546 } else { 1547 vm_object_t new_backing_object; 1548 1549 /* 1550 * If we do not entirely shadow the backing object, 1551 * there is nothing we can do so we give up. 1552 */ 1553 1554 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1555 break; 1556 } 1557 1558 /* 1559 * Make the parent shadow the next object in the 1560 * chain. Deallocating backing_object will not remove 1561 * it, since its reference count is at least 2. 1562 */ 1563 1564 LIST_REMOVE(object, shadow_list); 1565 backing_object->shadow_count--; 1566 backing_object->generation++; 1567 1568 new_backing_object = backing_object->backing_object; 1569 if ((object->backing_object = new_backing_object) != NULL) { 1570 vm_object_reference(new_backing_object); 1571 LIST_INSERT_HEAD( 1572 &new_backing_object->shadow_head, 1573 object, 1574 shadow_list 1575 ); 1576 new_backing_object->shadow_count++; 1577 new_backing_object->generation++; 1578 object->backing_object_offset += 1579 backing_object->backing_object_offset; 1580 } 1581 1582 /* 1583 * Drop the reference count on backing_object. Since 1584 * its ref_count was at least 2, it will not vanish; 1585 * so we don't need to call vm_object_deallocate, but 1586 * we do anyway. 1587 */ 1588 vm_object_deallocate_locked(backing_object); 1589 object_bypasses++; 1590 } 1591 1592 /* 1593 * Try again with this object's new backing object. 1594 */ 1595 } 1596 } 1597 1598 /* 1599 * Removes all physical pages in the specified object range from the 1600 * object's list of pages. 1601 * 1602 * No requirements. 1603 */ 1604 static int vm_object_page_remove_callback(vm_page_t p, void *data); 1605 1606 void 1607 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1608 boolean_t clean_only) 1609 { 1610 struct rb_vm_page_scan_info info; 1611 int all; 1612 1613 /* 1614 * Degenerate cases and assertions 1615 */ 1616 lwkt_gettoken(&vm_token); 1617 if (object == NULL || 1618 (object->resident_page_count == 0 && object->swblock_count == 0)) { 1619 lwkt_reltoken(&vm_token); 1620 return; 1621 } 1622 KASSERT(object->type != OBJT_PHYS, 1623 ("attempt to remove pages from a physical object")); 1624 1625 /* 1626 * Indicate that paging is occuring on the object 1627 */ 1628 crit_enter(); 1629 vm_object_pip_add(object, 1); 1630 1631 /* 1632 * Figure out the actual removal range and whether we are removing 1633 * the entire contents of the object or not. If removing the entire 1634 * contents, be sure to get all pages, even those that might be 1635 * beyond the end of the object. 1636 */ 1637 info.start_pindex = start; 1638 if (end == 0) 1639 info.end_pindex = (vm_pindex_t)-1; 1640 else 1641 info.end_pindex = end - 1; 1642 info.limit = clean_only; 1643 all = (start == 0 && info.end_pindex >= object->size - 1); 1644 1645 /* 1646 * Loop until we are sure we have gotten them all. 1647 */ 1648 do { 1649 info.error = 0; 1650 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1651 vm_object_page_remove_callback, &info); 1652 } while (info.error); 1653 1654 /* 1655 * Remove any related swap if throwing away pages, or for 1656 * non-swap objects (the swap is a clean copy in that case). 1657 */ 1658 if (object->type != OBJT_SWAP || clean_only == FALSE) { 1659 if (all) 1660 swap_pager_freespace_all(object); 1661 else 1662 swap_pager_freespace(object, info.start_pindex, 1663 info.end_pindex - info.start_pindex + 1); 1664 } 1665 1666 /* 1667 * Cleanup 1668 */ 1669 vm_object_pip_wakeup(object); 1670 crit_exit(); 1671 lwkt_reltoken(&vm_token); 1672 } 1673 1674 /* 1675 * The caller must hold vm_token. 1676 */ 1677 static int 1678 vm_object_page_remove_callback(vm_page_t p, void *data) 1679 { 1680 struct rb_vm_page_scan_info *info = data; 1681 1682 /* 1683 * Wired pages cannot be destroyed, but they can be invalidated 1684 * and we do so if clean_only (limit) is not set. 1685 * 1686 * WARNING! The page may be wired due to being part of a buffer 1687 * cache buffer, and the buffer might be marked B_CACHE. 1688 * This is fine as part of a truncation but VFSs must be 1689 * sure to fix the buffer up when re-extending the file. 1690 */ 1691 if (p->wire_count != 0) { 1692 vm_page_protect(p, VM_PROT_NONE); 1693 if (info->limit == 0) 1694 p->valid = 0; 1695 return(0); 1696 } 1697 1698 /* 1699 * The busy flags are only cleared at 1700 * interrupt -- minimize the spl transitions 1701 */ 1702 1703 if (vm_page_sleep_busy(p, TRUE, "vmopar")) { 1704 info->error = 1; 1705 return(0); 1706 } 1707 1708 /* 1709 * limit is our clean_only flag. If set and the page is dirty, do 1710 * not free it. If set and the page is being held by someone, do 1711 * not free it. 1712 */ 1713 if (info->limit && p->valid) { 1714 vm_page_test_dirty(p); 1715 if (p->valid & p->dirty) 1716 return(0); 1717 if (p->hold_count) 1718 return(0); 1719 } 1720 1721 /* 1722 * Destroy the page 1723 */ 1724 vm_page_busy(p); 1725 vm_page_protect(p, VM_PROT_NONE); 1726 vm_page_free(p); 1727 return(0); 1728 } 1729 1730 /* 1731 * Coalesces two objects backing up adjoining regions of memory into a 1732 * single object. 1733 * 1734 * returns TRUE if objects were combined. 1735 * 1736 * NOTE: Only works at the moment if the second object is NULL - 1737 * if it's not, which object do we lock first? 1738 * 1739 * Parameters: 1740 * prev_object First object to coalesce 1741 * prev_offset Offset into prev_object 1742 * next_object Second object into coalesce 1743 * next_offset Offset into next_object 1744 * 1745 * prev_size Size of reference to prev_object 1746 * next_size Size of reference to next_object 1747 * 1748 * The object must not be locked. 1749 * The caller must hold vm_token and vmobj_token. 1750 */ 1751 boolean_t 1752 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1753 vm_size_t prev_size, vm_size_t next_size) 1754 { 1755 vm_pindex_t next_pindex; 1756 1757 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1758 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 1759 1760 if (prev_object == NULL) { 1761 return (TRUE); 1762 } 1763 1764 if (prev_object->type != OBJT_DEFAULT && 1765 prev_object->type != OBJT_SWAP) { 1766 return (FALSE); 1767 } 1768 1769 /* 1770 * Try to collapse the object first 1771 */ 1772 vm_object_collapse(prev_object); 1773 1774 /* 1775 * Can't coalesce if: . more than one reference . paged out . shadows 1776 * another object . has a copy elsewhere (any of which mean that the 1777 * pages not mapped to prev_entry may be in use anyway) 1778 */ 1779 1780 if (prev_object->backing_object != NULL) { 1781 return (FALSE); 1782 } 1783 1784 prev_size >>= PAGE_SHIFT; 1785 next_size >>= PAGE_SHIFT; 1786 next_pindex = prev_pindex + prev_size; 1787 1788 if ((prev_object->ref_count > 1) && 1789 (prev_object->size != next_pindex)) { 1790 return (FALSE); 1791 } 1792 1793 /* 1794 * Remove any pages that may still be in the object from a previous 1795 * deallocation. 1796 */ 1797 if (next_pindex < prev_object->size) { 1798 vm_object_page_remove(prev_object, 1799 next_pindex, 1800 next_pindex + next_size, FALSE); 1801 if (prev_object->type == OBJT_SWAP) 1802 swap_pager_freespace(prev_object, 1803 next_pindex, next_size); 1804 } 1805 1806 /* 1807 * Extend the object if necessary. 1808 */ 1809 if (next_pindex + next_size > prev_object->size) 1810 prev_object->size = next_pindex + next_size; 1811 1812 return (TRUE); 1813 } 1814 1815 /* 1816 * Make the object writable and flag is being possibly dirty. 1817 * 1818 * No requirements. 1819 */ 1820 void 1821 vm_object_set_writeable_dirty(vm_object_t object) 1822 { 1823 struct vnode *vp; 1824 1825 lwkt_gettoken(&vm_token); 1826 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1827 if (object->type == OBJT_VNODE && 1828 (vp = (struct vnode *)object->handle) != NULL) { 1829 if ((vp->v_flag & VOBJDIRTY) == 0) { 1830 vsetflags(vp, VOBJDIRTY); 1831 } 1832 } 1833 lwkt_reltoken(&vm_token); 1834 } 1835 1836 static void 1837 vm_object_lock_init(vm_object_t obj) 1838 { 1839 #if defined(DEBUG_LOCKS) 1840 int i; 1841 1842 obj->debug_hold_bitmap = 0; 1843 obj->debug_hold_ovfl = 0; 1844 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1845 obj->debug_hold_thrs[i] = NULL; 1846 } 1847 #endif 1848 } 1849 1850 void 1851 vm_object_lock(vm_object_t obj) 1852 { 1853 lwkt_getpooltoken(obj); 1854 } 1855 1856 void 1857 vm_object_unlock(vm_object_t obj) 1858 { 1859 lwkt_relpooltoken(obj); 1860 } 1861 1862 void 1863 vm_object_hold(vm_object_t obj) 1864 { 1865 vm_object_lock(obj); 1866 1867 refcount_acquire(&obj->hold_count); 1868 1869 #if defined(DEBUG_LOCKS) 1870 int i; 1871 1872 i = ffs(~obj->debug_hold_bitmap) - 1; 1873 if (i == -1) { 1874 kprintf("vm_object hold count > VMOBJ_DEBUG_ARRAY_SIZE"); 1875 obj->debug_hold_ovfl = 1; 1876 } 1877 1878 obj->debug_hold_bitmap |= (1 << i); 1879 obj->debug_hold_thrs[i] = curthread; 1880 #endif 1881 } 1882 1883 void 1884 vm_object_drop(vm_object_t obj) 1885 { 1886 int rc; 1887 1888 #if defined(DEBUG_LOCKS) 1889 int found = 0; 1890 int i; 1891 1892 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1893 if ((obj->debug_hold_bitmap & (1 << i)) && 1894 (obj->debug_hold_thrs[i] == curthread)) { 1895 obj->debug_hold_bitmap &= ~(1 << i); 1896 obj->debug_hold_thrs[i] = NULL; 1897 found = 1; 1898 break; 1899 } 1900 } 1901 1902 if (found == 0 && obj->debug_hold_ovfl == 0) 1903 panic("vm_object: attempt to drop hold on non-self-held obj"); 1904 #endif 1905 1906 rc = refcount_release(&obj->hold_count); 1907 vm_object_unlock(obj); 1908 1909 if (rc) 1910 vm_object_hold_wake(obj); 1911 } 1912 1913 static void 1914 vm_object_hold_wake(vm_object_t obj) 1915 { 1916 wakeup(obj); 1917 } 1918 1919 static void 1920 vm_object_hold_wait(vm_object_t obj) 1921 { 1922 vm_object_lock(obj); 1923 1924 #if defined(DEBUG_LOCKS) 1925 int i; 1926 1927 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1928 if ((obj->debug_hold_bitmap & (1 << i)) && 1929 (obj->debug_hold_thrs[i] == curthread)) 1930 panic("vm_object: self-hold in terminate or collapse"); 1931 } 1932 #endif 1933 1934 while (obj->hold_count) 1935 tsleep(obj, 0, "vmobjhld", 0); 1936 1937 vm_object_unlock(obj); 1938 } 1939 1940 #include "opt_ddb.h" 1941 #ifdef DDB 1942 #include <sys/kernel.h> 1943 1944 #include <sys/cons.h> 1945 1946 #include <ddb/ddb.h> 1947 1948 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 1949 vm_map_entry_t entry); 1950 static int vm_object_in_map (vm_object_t object); 1951 1952 /* 1953 * The caller must hold vm_token. 1954 */ 1955 static int 1956 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1957 { 1958 vm_map_t tmpm; 1959 vm_map_entry_t tmpe; 1960 vm_object_t obj; 1961 int entcount; 1962 1963 if (map == 0) 1964 return 0; 1965 if (entry == 0) { 1966 tmpe = map->header.next; 1967 entcount = map->nentries; 1968 while (entcount-- && (tmpe != &map->header)) { 1969 if( _vm_object_in_map(map, object, tmpe)) { 1970 return 1; 1971 } 1972 tmpe = tmpe->next; 1973 } 1974 return (0); 1975 } 1976 switch(entry->maptype) { 1977 case VM_MAPTYPE_SUBMAP: 1978 tmpm = entry->object.sub_map; 1979 tmpe = tmpm->header.next; 1980 entcount = tmpm->nentries; 1981 while (entcount-- && tmpe != &tmpm->header) { 1982 if( _vm_object_in_map(tmpm, object, tmpe)) { 1983 return 1; 1984 } 1985 tmpe = tmpe->next; 1986 } 1987 break; 1988 case VM_MAPTYPE_NORMAL: 1989 case VM_MAPTYPE_VPAGETABLE: 1990 obj = entry->object.vm_object; 1991 while (obj) { 1992 if (obj == object) 1993 return 1; 1994 obj = obj->backing_object; 1995 } 1996 break; 1997 default: 1998 break; 1999 } 2000 return 0; 2001 } 2002 2003 static int vm_object_in_map_callback(struct proc *p, void *data); 2004 2005 struct vm_object_in_map_info { 2006 vm_object_t object; 2007 int rv; 2008 }; 2009 2010 /* 2011 * Debugging only 2012 */ 2013 static int 2014 vm_object_in_map(vm_object_t object) 2015 { 2016 struct vm_object_in_map_info info; 2017 2018 info.rv = 0; 2019 info.object = object; 2020 2021 allproc_scan(vm_object_in_map_callback, &info); 2022 if (info.rv) 2023 return 1; 2024 if( _vm_object_in_map(&kernel_map, object, 0)) 2025 return 1; 2026 if( _vm_object_in_map(&pager_map, object, 0)) 2027 return 1; 2028 if( _vm_object_in_map(&buffer_map, object, 0)) 2029 return 1; 2030 return 0; 2031 } 2032 2033 /* 2034 * Debugging only 2035 */ 2036 static int 2037 vm_object_in_map_callback(struct proc *p, void *data) 2038 { 2039 struct vm_object_in_map_info *info = data; 2040 2041 if (p->p_vmspace) { 2042 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 2043 info->rv = 1; 2044 return -1; 2045 } 2046 } 2047 return (0); 2048 } 2049 2050 DB_SHOW_COMMAND(vmochk, vm_object_check) 2051 { 2052 vm_object_t object; 2053 2054 /* 2055 * make sure that internal objs are in a map somewhere 2056 * and none have zero ref counts. 2057 */ 2058 for (object = TAILQ_FIRST(&vm_object_list); 2059 object != NULL; 2060 object = TAILQ_NEXT(object, object_list)) { 2061 if (object->type == OBJT_MARKER) 2062 continue; 2063 if (object->handle == NULL && 2064 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2065 if (object->ref_count == 0) { 2066 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2067 (long)object->size); 2068 } 2069 if (!vm_object_in_map(object)) { 2070 db_printf( 2071 "vmochk: internal obj is not in a map: " 2072 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2073 object->ref_count, (u_long)object->size, 2074 (u_long)object->size, 2075 (void *)object->backing_object); 2076 } 2077 } 2078 } 2079 } 2080 2081 /* 2082 * Debugging only 2083 */ 2084 DB_SHOW_COMMAND(object, vm_object_print_static) 2085 { 2086 /* XXX convert args. */ 2087 vm_object_t object = (vm_object_t)addr; 2088 boolean_t full = have_addr; 2089 2090 vm_page_t p; 2091 2092 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2093 #define count was_count 2094 2095 int count; 2096 2097 if (object == NULL) 2098 return; 2099 2100 db_iprintf( 2101 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2102 object, (int)object->type, (u_long)object->size, 2103 object->resident_page_count, object->ref_count, object->flags); 2104 /* 2105 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2106 */ 2107 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2108 object->shadow_count, 2109 object->backing_object ? object->backing_object->ref_count : 0, 2110 object->backing_object, (long)object->backing_object_offset); 2111 2112 if (!full) 2113 return; 2114 2115 db_indent += 2; 2116 count = 0; 2117 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 2118 if (count == 0) 2119 db_iprintf("memory:="); 2120 else if (count == 6) { 2121 db_printf("\n"); 2122 db_iprintf(" ..."); 2123 count = 0; 2124 } else 2125 db_printf(","); 2126 count++; 2127 2128 db_printf("(off=0x%lx,page=0x%lx)", 2129 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2130 } 2131 if (count != 0) 2132 db_printf("\n"); 2133 db_indent -= 2; 2134 } 2135 2136 /* XXX. */ 2137 #undef count 2138 2139 /* 2140 * XXX need this non-static entry for calling from vm_map_print. 2141 * 2142 * Debugging only 2143 */ 2144 void 2145 vm_object_print(/* db_expr_t */ long addr, 2146 boolean_t have_addr, 2147 /* db_expr_t */ long count, 2148 char *modif) 2149 { 2150 vm_object_print_static(addr, have_addr, count, modif); 2151 } 2152 2153 /* 2154 * Debugging only 2155 */ 2156 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2157 { 2158 vm_object_t object; 2159 int nl = 0; 2160 int c; 2161 for (object = TAILQ_FIRST(&vm_object_list); 2162 object != NULL; 2163 object = TAILQ_NEXT(object, object_list)) { 2164 vm_pindex_t idx, fidx; 2165 vm_pindex_t osize; 2166 vm_paddr_t pa = -1, padiff; 2167 int rcount; 2168 vm_page_t m; 2169 2170 if (object->type == OBJT_MARKER) 2171 continue; 2172 db_printf("new object: %p\n", (void *)object); 2173 if ( nl > 18) { 2174 c = cngetc(); 2175 if (c != ' ') 2176 return; 2177 nl = 0; 2178 } 2179 nl++; 2180 rcount = 0; 2181 fidx = 0; 2182 osize = object->size; 2183 if (osize > 128) 2184 osize = 128; 2185 for (idx = 0; idx < osize; idx++) { 2186 m = vm_page_lookup(object, idx); 2187 if (m == NULL) { 2188 if (rcount) { 2189 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2190 (long)fidx, rcount, (long)pa); 2191 if ( nl > 18) { 2192 c = cngetc(); 2193 if (c != ' ') 2194 return; 2195 nl = 0; 2196 } 2197 nl++; 2198 rcount = 0; 2199 } 2200 continue; 2201 } 2202 2203 2204 if (rcount && 2205 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2206 ++rcount; 2207 continue; 2208 } 2209 if (rcount) { 2210 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2211 padiff >>= PAGE_SHIFT; 2212 padiff &= PQ_L2_MASK; 2213 if (padiff == 0) { 2214 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2215 ++rcount; 2216 continue; 2217 } 2218 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2219 (long)fidx, rcount, (long)pa); 2220 db_printf("pd(%ld)\n", (long)padiff); 2221 if ( nl > 18) { 2222 c = cngetc(); 2223 if (c != ' ') 2224 return; 2225 nl = 0; 2226 } 2227 nl++; 2228 } 2229 fidx = idx; 2230 pa = VM_PAGE_TO_PHYS(m); 2231 rcount = 1; 2232 } 2233 if (rcount) { 2234 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2235 (long)fidx, rcount, (long)pa); 2236 if ( nl > 18) { 2237 c = cngetc(); 2238 if (c != ' ') 2239 return; 2240 nl = 0; 2241 } 2242 nl++; 2243 } 2244 } 2245 } 2246 #endif /* DDB */ 2247