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 vm_object_hold(object); 619 if (object->type != OBJT_VNODE || 620 (object->flags & OBJ_MIGHTBEDIRTY) == 0) { 621 vm_object_drop(object); 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 lwkt_gettoken(&vm_token); 660 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 661 vm_object_page_clean_pass1, &info); 662 lwkt_reltoken(&vm_token); 663 if (info.error == 0) { 664 vm_object_clear_flag(object, 665 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 666 if (object->type == OBJT_VNODE && 667 (vp = (struct vnode *)object->handle) != NULL) { 668 if (vp->v_flag & VOBJDIRTY) 669 vclrflags(vp, VOBJDIRTY); 670 } 671 } 672 } 673 674 /* 675 * Do a pass to clean all the dirty pages we find. 676 */ 677 do { 678 info.error = 0; 679 curgeneration = object->generation; 680 lwkt_gettoken(&vm_token); 681 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 682 vm_object_page_clean_pass2, &info); 683 lwkt_reltoken(&vm_token); 684 } while (info.error || curgeneration != object->generation); 685 686 vm_object_clear_flag(object, OBJ_CLEANING); 687 crit_exit(); 688 vm_object_drop(object); 689 } 690 691 /* 692 * The caller must hold vm_token. 693 */ 694 static 695 int 696 vm_object_page_clean_pass1(struct vm_page *p, void *data) 697 { 698 struct rb_vm_page_scan_info *info = data; 699 700 vm_page_flag_set(p, PG_CLEANCHK); 701 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 702 info->error = 1; 703 else 704 vm_page_protect(p, VM_PROT_READ); /* must not block */ 705 return(0); 706 } 707 708 /* 709 * The caller must hold vm_token. 710 */ 711 static 712 int 713 vm_object_page_clean_pass2(struct vm_page *p, void *data) 714 { 715 struct rb_vm_page_scan_info *info = data; 716 int n; 717 718 /* 719 * Do not mess with pages that were inserted after we started 720 * the cleaning pass. 721 */ 722 if ((p->flags & PG_CLEANCHK) == 0) 723 return(0); 724 725 /* 726 * Before wasting time traversing the pmaps, check for trivial 727 * cases where the page cannot be dirty. 728 */ 729 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) { 730 KKASSERT((p->dirty & p->valid) == 0); 731 return(0); 732 } 733 734 /* 735 * Check whether the page is dirty or not. The page has been set 736 * to be read-only so the check will not race a user dirtying the 737 * page. 738 */ 739 vm_page_test_dirty(p); 740 if ((p->dirty & p->valid) == 0) { 741 vm_page_flag_clear(p, PG_CLEANCHK); 742 return(0); 743 } 744 745 /* 746 * If we have been asked to skip nosync pages and this is a 747 * nosync page, skip it. Note that the object flags were 748 * not cleared in this case (because pass1 will have returned an 749 * error), so we do not have to set them. 750 */ 751 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 752 vm_page_flag_clear(p, PG_CLEANCHK); 753 return(0); 754 } 755 756 /* 757 * Flush as many pages as we can. PG_CLEANCHK will be cleared on 758 * the pages that get successfully flushed. Set info->error if 759 * we raced an object modification. 760 */ 761 n = vm_object_page_collect_flush(info->object, p, info->pagerflags); 762 if (n == 0) 763 info->error = 1; 764 return(0); 765 } 766 767 /* 768 * Collect the specified page and nearby pages and flush them out. 769 * The number of pages flushed is returned. 770 * 771 * The caller must hold vm_token. 772 */ 773 static int 774 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags) 775 { 776 int runlen; 777 int maxf; 778 int chkb; 779 int maxb; 780 int i; 781 int curgeneration; 782 vm_pindex_t pi; 783 vm_page_t maf[vm_pageout_page_count]; 784 vm_page_t mab[vm_pageout_page_count]; 785 vm_page_t ma[vm_pageout_page_count]; 786 787 curgeneration = object->generation; 788 789 pi = p->pindex; 790 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 791 if (object->generation != curgeneration) { 792 return(0); 793 } 794 } 795 KKASSERT(p->object == object && p->pindex == pi); 796 797 maxf = 0; 798 for(i = 1; i < vm_pageout_page_count; i++) { 799 vm_page_t tp; 800 801 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 802 if ((tp->flags & PG_BUSY) || 803 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 804 (tp->flags & PG_CLEANCHK) == 0) || 805 (tp->busy != 0)) 806 break; 807 if((tp->queue - tp->pc) == PQ_CACHE) { 808 vm_page_flag_clear(tp, PG_CLEANCHK); 809 break; 810 } 811 vm_page_test_dirty(tp); 812 if ((tp->dirty & tp->valid) == 0) { 813 vm_page_flag_clear(tp, PG_CLEANCHK); 814 break; 815 } 816 maf[ i - 1 ] = tp; 817 maxf++; 818 continue; 819 } 820 break; 821 } 822 823 maxb = 0; 824 chkb = vm_pageout_page_count - maxf; 825 if (chkb) { 826 for(i = 1; i < chkb;i++) { 827 vm_page_t tp; 828 829 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 830 if ((tp->flags & PG_BUSY) || 831 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 832 (tp->flags & PG_CLEANCHK) == 0) || 833 (tp->busy != 0)) 834 break; 835 if((tp->queue - tp->pc) == PQ_CACHE) { 836 vm_page_flag_clear(tp, PG_CLEANCHK); 837 break; 838 } 839 vm_page_test_dirty(tp); 840 if ((tp->dirty & tp->valid) == 0) { 841 vm_page_flag_clear(tp, PG_CLEANCHK); 842 break; 843 } 844 mab[ i - 1 ] = tp; 845 maxb++; 846 continue; 847 } 848 break; 849 } 850 } 851 852 for(i = 0; i < maxb; i++) { 853 int index = (maxb - i) - 1; 854 ma[index] = mab[i]; 855 vm_page_flag_clear(ma[index], PG_CLEANCHK); 856 } 857 vm_page_flag_clear(p, PG_CLEANCHK); 858 ma[maxb] = p; 859 for(i = 0; i < maxf; i++) { 860 int index = (maxb + i) + 1; 861 ma[index] = maf[i]; 862 vm_page_flag_clear(ma[index], PG_CLEANCHK); 863 } 864 runlen = maxb + maxf + 1; 865 866 vm_pageout_flush(ma, runlen, pagerflags); 867 for (i = 0; i < runlen; i++) { 868 if (ma[i]->valid & ma[i]->dirty) { 869 vm_page_protect(ma[i], VM_PROT_READ); 870 vm_page_flag_set(ma[i], PG_CLEANCHK); 871 872 /* 873 * maxf will end up being the actual number of pages 874 * we wrote out contiguously, non-inclusive of the 875 * first page. We do not count look-behind pages. 876 */ 877 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 878 maxf = i - maxb - 1; 879 } 880 } 881 return(maxf + 1); 882 } 883 884 /* 885 * Same as vm_object_pmap_copy, except range checking really 886 * works, and is meant for small sections of an object. 887 * 888 * This code protects resident pages by making them read-only 889 * and is typically called on a fork or split when a page 890 * is converted to copy-on-write. 891 * 892 * NOTE: If the page is already at VM_PROT_NONE, calling 893 * vm_page_protect will have no effect. 894 */ 895 void 896 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 897 { 898 vm_pindex_t idx; 899 vm_page_t p; 900 901 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 902 return; 903 904 /* 905 * spl protection needed to prevent races between the lookup, 906 * an interrupt unbusy/free, and our protect call. 907 */ 908 crit_enter(); 909 lwkt_gettoken(&vm_token); 910 for (idx = start; idx < end; idx++) { 911 p = vm_page_lookup(object, idx); 912 if (p == NULL) 913 continue; 914 vm_page_protect(p, VM_PROT_READ); 915 } 916 lwkt_reltoken(&vm_token); 917 crit_exit(); 918 } 919 920 /* 921 * Removes all physical pages in the specified object range from all 922 * physical maps. 923 * 924 * The object must *not* be locked. 925 */ 926 927 static int vm_object_pmap_remove_callback(vm_page_t p, void *data); 928 929 void 930 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 931 { 932 struct rb_vm_page_scan_info info; 933 934 if (object == NULL) 935 return; 936 info.start_pindex = start; 937 info.end_pindex = end - 1; 938 939 crit_enter(); 940 lwkt_gettoken(&vm_token); 941 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 942 vm_object_pmap_remove_callback, &info); 943 if (start == 0 && end == object->size) 944 vm_object_clear_flag(object, OBJ_WRITEABLE); 945 lwkt_reltoken(&vm_token); 946 crit_exit(); 947 } 948 949 /* 950 * The caller must hold vm_token. 951 */ 952 static int 953 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused) 954 { 955 vm_page_protect(p, VM_PROT_NONE); 956 return(0); 957 } 958 959 /* 960 * Implements the madvise function at the object/page level. 961 * 962 * MADV_WILLNEED (any object) 963 * 964 * Activate the specified pages if they are resident. 965 * 966 * MADV_DONTNEED (any object) 967 * 968 * Deactivate the specified pages if they are resident. 969 * 970 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only) 971 * 972 * Deactivate and clean the specified pages if they are 973 * resident. This permits the process to reuse the pages 974 * without faulting or the kernel to reclaim the pages 975 * without I/O. 976 * 977 * No requirements. 978 */ 979 void 980 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 981 { 982 vm_pindex_t end, tpindex; 983 vm_object_t tobject; 984 vm_page_t m; 985 986 if (object == NULL) 987 return; 988 989 end = pindex + count; 990 991 lwkt_gettoken(&vm_token); 992 993 /* 994 * Locate and adjust resident pages 995 */ 996 for (; pindex < end; pindex += 1) { 997 relookup: 998 tobject = object; 999 tpindex = pindex; 1000 shadowlookup: 1001 /* 1002 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1003 * and those pages must be OBJ_ONEMAPPING. 1004 */ 1005 if (advise == MADV_FREE) { 1006 if ((tobject->type != OBJT_DEFAULT && 1007 tobject->type != OBJT_SWAP) || 1008 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1009 continue; 1010 } 1011 } 1012 1013 /* 1014 * spl protection is required to avoid a race between the 1015 * lookup, an interrupt unbusy/free, and our busy check. 1016 */ 1017 1018 crit_enter(); 1019 m = vm_page_lookup(tobject, tpindex); 1020 1021 if (m == NULL) { 1022 /* 1023 * There may be swap even if there is no backing page 1024 */ 1025 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1026 swap_pager_freespace(tobject, tpindex, 1); 1027 1028 /* 1029 * next object 1030 */ 1031 crit_exit(); 1032 if (tobject->backing_object == NULL) 1033 continue; 1034 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1035 tobject = tobject->backing_object; 1036 goto shadowlookup; 1037 } 1038 1039 /* 1040 * If the page is busy or not in a normal active state, 1041 * we skip it. If the page is not managed there are no 1042 * page queues to mess with. Things can break if we mess 1043 * with pages in any of the below states. 1044 */ 1045 if ( 1046 m->hold_count || 1047 m->wire_count || 1048 (m->flags & PG_UNMANAGED) || 1049 m->valid != VM_PAGE_BITS_ALL 1050 ) { 1051 crit_exit(); 1052 continue; 1053 } 1054 1055 if (vm_page_sleep_busy(m, TRUE, "madvpo")) { 1056 crit_exit(); 1057 goto relookup; 1058 } 1059 vm_page_busy(m); 1060 crit_exit(); 1061 1062 /* 1063 * Theoretically once a page is known not to be busy, an 1064 * interrupt cannot come along and rip it out from under us. 1065 */ 1066 1067 if (advise == MADV_WILLNEED) { 1068 vm_page_activate(m); 1069 } else if (advise == MADV_DONTNEED) { 1070 vm_page_dontneed(m); 1071 } else if (advise == MADV_FREE) { 1072 /* 1073 * Mark the page clean. This will allow the page 1074 * to be freed up by the system. However, such pages 1075 * are often reused quickly by malloc()/free() 1076 * so we do not do anything that would cause 1077 * a page fault if we can help it. 1078 * 1079 * Specifically, we do not try to actually free 1080 * the page now nor do we try to put it in the 1081 * cache (which would cause a page fault on reuse). 1082 * 1083 * But we do make the page is freeable as we 1084 * can without actually taking the step of unmapping 1085 * it. 1086 */ 1087 pmap_clear_modify(m); 1088 m->dirty = 0; 1089 m->act_count = 0; 1090 vm_page_dontneed(m); 1091 if (tobject->type == OBJT_SWAP) 1092 swap_pager_freespace(tobject, tpindex, 1); 1093 } 1094 vm_page_wakeup(m); 1095 } 1096 lwkt_reltoken(&vm_token); 1097 } 1098 1099 /* 1100 * Create a new object which is backed by the specified existing object 1101 * range. The source object reference is deallocated. 1102 * 1103 * The new object and offset into that object are returned in the source 1104 * parameters. 1105 * 1106 * No other requirements. 1107 */ 1108 void 1109 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length) 1110 { 1111 vm_object_t source; 1112 vm_object_t result; 1113 1114 source = *object; 1115 1116 /* 1117 * Don't create the new object if the old object isn't shared. 1118 */ 1119 lwkt_gettoken(&vm_token); 1120 1121 if (source != NULL && 1122 source->ref_count == 1 && 1123 source->handle == NULL && 1124 (source->type == OBJT_DEFAULT || 1125 source->type == OBJT_SWAP)) { 1126 lwkt_reltoken(&vm_token); 1127 return; 1128 } 1129 1130 /* 1131 * Allocate a new object with the given length 1132 */ 1133 1134 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL) 1135 panic("vm_object_shadow: no object for shadowing"); 1136 1137 /* 1138 * The new object shadows the source object, adding a reference to it. 1139 * Our caller changes his reference to point to the new object, 1140 * removing a reference to the source object. Net result: no change 1141 * of reference count. 1142 * 1143 * Try to optimize the result object's page color when shadowing 1144 * in order to maintain page coloring consistency in the combined 1145 * shadowed object. 1146 */ 1147 result->backing_object = source; 1148 if (source) { 1149 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1150 source->shadow_count++; 1151 source->generation++; 1152 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK; 1153 } 1154 1155 /* 1156 * Store the offset into the source object, and fix up the offset into 1157 * the new object. 1158 */ 1159 result->backing_object_offset = *offset; 1160 lwkt_reltoken(&vm_token); 1161 1162 /* 1163 * Return the new things 1164 */ 1165 *offset = 0; 1166 *object = result; 1167 } 1168 1169 #define OBSC_TEST_ALL_SHADOWED 0x0001 1170 #define OBSC_COLLAPSE_NOWAIT 0x0002 1171 #define OBSC_COLLAPSE_WAIT 0x0004 1172 1173 static int vm_object_backing_scan_callback(vm_page_t p, void *data); 1174 1175 /* 1176 * The caller must hold vm_token. 1177 */ 1178 static __inline int 1179 vm_object_backing_scan(vm_object_t object, int op) 1180 { 1181 struct rb_vm_page_scan_info info; 1182 vm_object_t backing_object; 1183 1184 crit_enter(); 1185 1186 backing_object = object->backing_object; 1187 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1188 1189 /* 1190 * Initial conditions 1191 */ 1192 1193 if (op & OBSC_TEST_ALL_SHADOWED) { 1194 /* 1195 * We do not want to have to test for the existence of 1196 * swap pages in the backing object. XXX but with the 1197 * new swapper this would be pretty easy to do. 1198 * 1199 * XXX what about anonymous MAP_SHARED memory that hasn't 1200 * been ZFOD faulted yet? If we do not test for this, the 1201 * shadow test may succeed! XXX 1202 */ 1203 if (backing_object->type != OBJT_DEFAULT) { 1204 crit_exit(); 1205 return(0); 1206 } 1207 } 1208 if (op & OBSC_COLLAPSE_WAIT) { 1209 KKASSERT((backing_object->flags & OBJ_DEAD) == 0); 1210 vm_object_set_flag(backing_object, OBJ_DEAD); 1211 } 1212 1213 /* 1214 * Our scan. We have to retry if a negative error code is returned, 1215 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that 1216 * the scan had to be stopped because the parent does not completely 1217 * shadow the child. 1218 */ 1219 info.object = object; 1220 info.backing_object = backing_object; 1221 info.limit = op; 1222 do { 1223 info.error = 1; 1224 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL, 1225 vm_object_backing_scan_callback, 1226 &info); 1227 } while (info.error < 0); 1228 crit_exit(); 1229 return(info.error); 1230 } 1231 1232 /* 1233 * The caller must hold vm_token. 1234 */ 1235 static int 1236 vm_object_backing_scan_callback(vm_page_t p, void *data) 1237 { 1238 struct rb_vm_page_scan_info *info = data; 1239 vm_object_t backing_object; 1240 vm_object_t object; 1241 vm_pindex_t new_pindex; 1242 vm_pindex_t backing_offset_index; 1243 int op; 1244 1245 new_pindex = p->pindex - info->backing_offset_index; 1246 op = info->limit; 1247 object = info->object; 1248 backing_object = info->backing_object; 1249 backing_offset_index = info->backing_offset_index; 1250 1251 if (op & OBSC_TEST_ALL_SHADOWED) { 1252 vm_page_t pp; 1253 1254 /* 1255 * Ignore pages outside the parent object's range 1256 * and outside the parent object's mapping of the 1257 * backing object. 1258 * 1259 * note that we do not busy the backing object's 1260 * page. 1261 */ 1262 if ( 1263 p->pindex < backing_offset_index || 1264 new_pindex >= object->size 1265 ) { 1266 return(0); 1267 } 1268 1269 /* 1270 * See if the parent has the page or if the parent's 1271 * object pager has the page. If the parent has the 1272 * page but the page is not valid, the parent's 1273 * object pager must have the page. 1274 * 1275 * If this fails, the parent does not completely shadow 1276 * the object and we might as well give up now. 1277 */ 1278 1279 pp = vm_page_lookup(object, new_pindex); 1280 if ((pp == NULL || pp->valid == 0) && 1281 !vm_pager_has_page(object, new_pindex) 1282 ) { 1283 info->error = 0; /* problemo */ 1284 return(-1); /* stop the scan */ 1285 } 1286 } 1287 1288 /* 1289 * Check for busy page 1290 */ 1291 1292 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1293 vm_page_t pp; 1294 1295 if (op & OBSC_COLLAPSE_NOWAIT) { 1296 if ( 1297 (p->flags & PG_BUSY) || 1298 !p->valid || 1299 p->hold_count || 1300 p->wire_count || 1301 p->busy 1302 ) { 1303 return(0); 1304 } 1305 } else if (op & OBSC_COLLAPSE_WAIT) { 1306 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1307 /* 1308 * If we slept, anything could have 1309 * happened. Ask that the scan be restarted. 1310 * 1311 * Since the object is marked dead, the 1312 * backing offset should not have changed. 1313 */ 1314 info->error = -1; 1315 return(-1); 1316 } 1317 } 1318 1319 /* 1320 * Busy the page 1321 */ 1322 vm_page_busy(p); 1323 1324 KASSERT( 1325 p->object == backing_object, 1326 ("vm_object_qcollapse(): object mismatch") 1327 ); 1328 1329 /* 1330 * Destroy any associated swap 1331 */ 1332 if (backing_object->type == OBJT_SWAP) 1333 swap_pager_freespace(backing_object, p->pindex, 1); 1334 1335 if ( 1336 p->pindex < backing_offset_index || 1337 new_pindex >= object->size 1338 ) { 1339 /* 1340 * Page is out of the parent object's range, we 1341 * can simply destroy it. 1342 */ 1343 vm_page_protect(p, VM_PROT_NONE); 1344 vm_page_free(p); 1345 return(0); 1346 } 1347 1348 pp = vm_page_lookup(object, new_pindex); 1349 if (pp != NULL || vm_pager_has_page(object, new_pindex)) { 1350 /* 1351 * page already exists in parent OR swap exists 1352 * for this location in the parent. Destroy 1353 * the original page from the backing object. 1354 * 1355 * Leave the parent's page alone 1356 */ 1357 vm_page_protect(p, VM_PROT_NONE); 1358 vm_page_free(p); 1359 return(0); 1360 } 1361 1362 /* 1363 * Page does not exist in parent, rename the 1364 * page from the backing object to the main object. 1365 * 1366 * If the page was mapped to a process, it can remain 1367 * mapped through the rename. 1368 */ 1369 if ((p->queue - p->pc) == PQ_CACHE) 1370 vm_page_deactivate(p); 1371 1372 vm_page_rename(p, object, new_pindex); 1373 /* page automatically made dirty by rename */ 1374 } 1375 return(0); 1376 } 1377 1378 /* 1379 * This version of collapse allows the operation to occur earlier and 1380 * when paging_in_progress is true for an object... This is not a complete 1381 * operation, but should plug 99.9% of the rest of the leaks. 1382 * 1383 * The caller must hold vm_token and vmobj_token. 1384 * (only called from vm_object_collapse) 1385 */ 1386 static void 1387 vm_object_qcollapse(vm_object_t object) 1388 { 1389 vm_object_t backing_object = object->backing_object; 1390 1391 if (backing_object->ref_count != 1) 1392 return; 1393 1394 backing_object->ref_count += 2; 1395 1396 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1397 1398 backing_object->ref_count -= 2; 1399 } 1400 1401 /* 1402 * Collapse an object with the object backing it. Pages in the backing 1403 * object are moved into the parent, and the backing object is deallocated. 1404 */ 1405 void 1406 vm_object_collapse(vm_object_t object) 1407 { 1408 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1409 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 1410 1411 while (TRUE) { 1412 vm_object_t backing_object; 1413 1414 /* 1415 * Verify that the conditions are right for collapse: 1416 * 1417 * The object exists and the backing object exists. 1418 */ 1419 if (object == NULL) 1420 break; 1421 1422 if ((backing_object = object->backing_object) == NULL) 1423 break; 1424 1425 /* 1426 * we check the backing object first, because it is most likely 1427 * not collapsable. 1428 */ 1429 if (backing_object->handle != NULL || 1430 (backing_object->type != OBJT_DEFAULT && 1431 backing_object->type != OBJT_SWAP) || 1432 (backing_object->flags & OBJ_DEAD) || 1433 object->handle != NULL || 1434 (object->type != OBJT_DEFAULT && 1435 object->type != OBJT_SWAP) || 1436 (object->flags & OBJ_DEAD)) { 1437 break; 1438 } 1439 1440 if ( 1441 object->paging_in_progress != 0 || 1442 backing_object->paging_in_progress != 0 1443 ) { 1444 vm_object_qcollapse(object); 1445 break; 1446 } 1447 1448 /* 1449 * We know that we can either collapse the backing object (if 1450 * the parent is the only reference to it) or (perhaps) have 1451 * the parent bypass the object if the parent happens to shadow 1452 * all the resident pages in the entire backing object. 1453 * 1454 * This is ignoring pager-backed pages such as swap pages. 1455 * vm_object_backing_scan fails the shadowing test in this 1456 * case. 1457 */ 1458 1459 if (backing_object->ref_count == 1) { 1460 /* 1461 * If there is exactly one reference to the backing 1462 * object, we can collapse it into the parent. 1463 */ 1464 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1465 1466 /* 1467 * Move the pager from backing_object to object. 1468 */ 1469 1470 if (backing_object->type == OBJT_SWAP) { 1471 vm_object_pip_add(backing_object, 1); 1472 1473 /* 1474 * scrap the paging_offset junk and do a 1475 * discrete copy. This also removes major 1476 * assumptions about how the swap-pager 1477 * works from where it doesn't belong. The 1478 * new swapper is able to optimize the 1479 * destroy-source case. 1480 */ 1481 1482 vm_object_pip_add(object, 1); 1483 swap_pager_copy( 1484 backing_object, 1485 object, 1486 OFF_TO_IDX(object->backing_object_offset), TRUE); 1487 vm_object_pip_wakeup(object); 1488 1489 vm_object_pip_wakeup(backing_object); 1490 } 1491 /* 1492 * Object now shadows whatever backing_object did. 1493 * Note that the reference to 1494 * backing_object->backing_object moves from within 1495 * backing_object to within object. 1496 */ 1497 1498 LIST_REMOVE(object, shadow_list); 1499 object->backing_object->shadow_count--; 1500 object->backing_object->generation++; 1501 if (backing_object->backing_object) { 1502 LIST_REMOVE(backing_object, shadow_list); 1503 backing_object->backing_object->shadow_count--; 1504 backing_object->backing_object->generation++; 1505 } 1506 object->backing_object = backing_object->backing_object; 1507 if (object->backing_object) { 1508 LIST_INSERT_HEAD( 1509 &object->backing_object->shadow_head, 1510 object, 1511 shadow_list 1512 ); 1513 object->backing_object->shadow_count++; 1514 object->backing_object->generation++; 1515 } 1516 1517 object->backing_object_offset += 1518 backing_object->backing_object_offset; 1519 1520 /* 1521 * Discard backing_object. 1522 * 1523 * Since the backing object has no pages, no pager left, 1524 * and no object references within it, all that is 1525 * necessary is to dispose of it. 1526 */ 1527 1528 KASSERT(backing_object->ref_count == 1, 1529 ("backing_object %p was somehow " 1530 "re-referenced during collapse!", 1531 backing_object)); 1532 KASSERT(RB_EMPTY(&backing_object->rb_memq), 1533 ("backing_object %p somehow has left " 1534 "over pages during collapse!", 1535 backing_object)); 1536 1537 /* 1538 * Wait for hold count to hit zero 1539 */ 1540 vm_object_hold_wait(backing_object); 1541 1542 /* (we are holding vmobj_token) */ 1543 TAILQ_REMOVE(&vm_object_list, backing_object, 1544 object_list); 1545 vm_object_count--; 1546 1547 zfree(obj_zone, backing_object); 1548 1549 object_collapses++; 1550 } else { 1551 vm_object_t new_backing_object; 1552 1553 /* 1554 * If we do not entirely shadow the backing object, 1555 * there is nothing we can do so we give up. 1556 */ 1557 1558 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1559 break; 1560 } 1561 1562 /* 1563 * Make the parent shadow the next object in the 1564 * chain. Deallocating backing_object will not remove 1565 * it, since its reference count is at least 2. 1566 */ 1567 1568 LIST_REMOVE(object, shadow_list); 1569 backing_object->shadow_count--; 1570 backing_object->generation++; 1571 1572 new_backing_object = backing_object->backing_object; 1573 if ((object->backing_object = new_backing_object) != NULL) { 1574 vm_object_reference(new_backing_object); 1575 LIST_INSERT_HEAD( 1576 &new_backing_object->shadow_head, 1577 object, 1578 shadow_list 1579 ); 1580 new_backing_object->shadow_count++; 1581 new_backing_object->generation++; 1582 object->backing_object_offset += 1583 backing_object->backing_object_offset; 1584 } 1585 1586 /* 1587 * Drop the reference count on backing_object. Since 1588 * its ref_count was at least 2, it will not vanish; 1589 * so we don't need to call vm_object_deallocate, but 1590 * we do anyway. 1591 */ 1592 vm_object_deallocate_locked(backing_object); 1593 object_bypasses++; 1594 } 1595 1596 /* 1597 * Try again with this object's new backing object. 1598 */ 1599 } 1600 } 1601 1602 /* 1603 * Removes all physical pages in the specified object range from the 1604 * object's list of pages. 1605 * 1606 * No requirements. 1607 */ 1608 static int vm_object_page_remove_callback(vm_page_t p, void *data); 1609 1610 void 1611 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1612 boolean_t clean_only) 1613 { 1614 struct rb_vm_page_scan_info info; 1615 int all; 1616 1617 /* 1618 * Degenerate cases and assertions 1619 */ 1620 lwkt_gettoken(&vm_token); 1621 if (object == NULL || 1622 (object->resident_page_count == 0 && object->swblock_count == 0)) { 1623 lwkt_reltoken(&vm_token); 1624 return; 1625 } 1626 KASSERT(object->type != OBJT_PHYS, 1627 ("attempt to remove pages from a physical object")); 1628 1629 /* 1630 * Indicate that paging is occuring on the object 1631 */ 1632 crit_enter(); 1633 vm_object_pip_add(object, 1); 1634 1635 /* 1636 * Figure out the actual removal range and whether we are removing 1637 * the entire contents of the object or not. If removing the entire 1638 * contents, be sure to get all pages, even those that might be 1639 * beyond the end of the object. 1640 */ 1641 info.start_pindex = start; 1642 if (end == 0) 1643 info.end_pindex = (vm_pindex_t)-1; 1644 else 1645 info.end_pindex = end - 1; 1646 info.limit = clean_only; 1647 all = (start == 0 && info.end_pindex >= object->size - 1); 1648 1649 /* 1650 * Loop until we are sure we have gotten them all. 1651 */ 1652 do { 1653 info.error = 0; 1654 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp, 1655 vm_object_page_remove_callback, &info); 1656 } while (info.error); 1657 1658 /* 1659 * Remove any related swap if throwing away pages, or for 1660 * non-swap objects (the swap is a clean copy in that case). 1661 */ 1662 if (object->type != OBJT_SWAP || clean_only == FALSE) { 1663 if (all) 1664 swap_pager_freespace_all(object); 1665 else 1666 swap_pager_freespace(object, info.start_pindex, 1667 info.end_pindex - info.start_pindex + 1); 1668 } 1669 1670 /* 1671 * Cleanup 1672 */ 1673 vm_object_pip_wakeup(object); 1674 crit_exit(); 1675 lwkt_reltoken(&vm_token); 1676 } 1677 1678 /* 1679 * The caller must hold vm_token. 1680 */ 1681 static int 1682 vm_object_page_remove_callback(vm_page_t p, void *data) 1683 { 1684 struct rb_vm_page_scan_info *info = data; 1685 1686 /* 1687 * Wired pages cannot be destroyed, but they can be invalidated 1688 * and we do so if clean_only (limit) is not set. 1689 * 1690 * WARNING! The page may be wired due to being part of a buffer 1691 * cache buffer, and the buffer might be marked B_CACHE. 1692 * This is fine as part of a truncation but VFSs must be 1693 * sure to fix the buffer up when re-extending the file. 1694 */ 1695 if (p->wire_count != 0) { 1696 vm_page_protect(p, VM_PROT_NONE); 1697 if (info->limit == 0) 1698 p->valid = 0; 1699 return(0); 1700 } 1701 1702 /* 1703 * The busy flags are only cleared at 1704 * interrupt -- minimize the spl transitions 1705 */ 1706 1707 if (vm_page_sleep_busy(p, TRUE, "vmopar")) { 1708 info->error = 1; 1709 return(0); 1710 } 1711 1712 /* 1713 * limit is our clean_only flag. If set and the page is dirty, do 1714 * not free it. If set and the page is being held by someone, do 1715 * not free it. 1716 */ 1717 if (info->limit && p->valid) { 1718 vm_page_test_dirty(p); 1719 if (p->valid & p->dirty) 1720 return(0); 1721 if (p->hold_count) 1722 return(0); 1723 } 1724 1725 /* 1726 * Destroy the page 1727 */ 1728 vm_page_busy(p); 1729 vm_page_protect(p, VM_PROT_NONE); 1730 vm_page_free(p); 1731 return(0); 1732 } 1733 1734 /* 1735 * Coalesces two objects backing up adjoining regions of memory into a 1736 * single object. 1737 * 1738 * returns TRUE if objects were combined. 1739 * 1740 * NOTE: Only works at the moment if the second object is NULL - 1741 * if it's not, which object do we lock first? 1742 * 1743 * Parameters: 1744 * prev_object First object to coalesce 1745 * prev_offset Offset into prev_object 1746 * next_object Second object into coalesce 1747 * next_offset Offset into next_object 1748 * 1749 * prev_size Size of reference to prev_object 1750 * next_size Size of reference to next_object 1751 * 1752 * The object must not be locked. 1753 * The caller must hold vm_token and vmobj_token. 1754 */ 1755 boolean_t 1756 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1757 vm_size_t prev_size, vm_size_t next_size) 1758 { 1759 vm_pindex_t next_pindex; 1760 1761 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1762 ASSERT_LWKT_TOKEN_HELD(&vmobj_token); 1763 1764 if (prev_object == NULL) { 1765 return (TRUE); 1766 } 1767 1768 if (prev_object->type != OBJT_DEFAULT && 1769 prev_object->type != OBJT_SWAP) { 1770 return (FALSE); 1771 } 1772 1773 /* 1774 * Try to collapse the object first 1775 */ 1776 vm_object_collapse(prev_object); 1777 1778 /* 1779 * Can't coalesce if: . more than one reference . paged out . shadows 1780 * another object . has a copy elsewhere (any of which mean that the 1781 * pages not mapped to prev_entry may be in use anyway) 1782 */ 1783 1784 if (prev_object->backing_object != NULL) { 1785 return (FALSE); 1786 } 1787 1788 prev_size >>= PAGE_SHIFT; 1789 next_size >>= PAGE_SHIFT; 1790 next_pindex = prev_pindex + prev_size; 1791 1792 if ((prev_object->ref_count > 1) && 1793 (prev_object->size != next_pindex)) { 1794 return (FALSE); 1795 } 1796 1797 /* 1798 * Remove any pages that may still be in the object from a previous 1799 * deallocation. 1800 */ 1801 if (next_pindex < prev_object->size) { 1802 vm_object_page_remove(prev_object, 1803 next_pindex, 1804 next_pindex + next_size, FALSE); 1805 if (prev_object->type == OBJT_SWAP) 1806 swap_pager_freespace(prev_object, 1807 next_pindex, next_size); 1808 } 1809 1810 /* 1811 * Extend the object if necessary. 1812 */ 1813 if (next_pindex + next_size > prev_object->size) 1814 prev_object->size = next_pindex + next_size; 1815 1816 return (TRUE); 1817 } 1818 1819 /* 1820 * Make the object writable and flag is being possibly dirty. 1821 * 1822 * No requirements. 1823 */ 1824 void 1825 vm_object_set_writeable_dirty(vm_object_t object) 1826 { 1827 struct vnode *vp; 1828 1829 lwkt_gettoken(&vm_token); 1830 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1831 if (object->type == OBJT_VNODE && 1832 (vp = (struct vnode *)object->handle) != NULL) { 1833 if ((vp->v_flag & VOBJDIRTY) == 0) { 1834 vsetflags(vp, VOBJDIRTY); 1835 } 1836 } 1837 lwkt_reltoken(&vm_token); 1838 } 1839 1840 static void 1841 vm_object_lock_init(vm_object_t obj) 1842 { 1843 #if defined(DEBUG_LOCKS) 1844 int i; 1845 1846 obj->debug_hold_bitmap = 0; 1847 obj->debug_hold_ovfl = 0; 1848 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1849 obj->debug_hold_thrs[i] = NULL; 1850 } 1851 #endif 1852 } 1853 1854 void 1855 vm_object_lock(vm_object_t obj) 1856 { 1857 lwkt_getpooltoken(obj); 1858 } 1859 1860 void 1861 vm_object_unlock(vm_object_t obj) 1862 { 1863 lwkt_relpooltoken(obj); 1864 } 1865 1866 void 1867 vm_object_hold(vm_object_t obj) 1868 { 1869 vm_object_lock(obj); 1870 1871 refcount_acquire(&obj->hold_count); 1872 1873 #if defined(DEBUG_LOCKS) 1874 int i; 1875 1876 i = ffs(~obj->debug_hold_bitmap) - 1; 1877 if (i == -1) { 1878 kprintf("vm_object hold count > VMOBJ_DEBUG_ARRAY_SIZE"); 1879 obj->debug_hold_ovfl = 1; 1880 } 1881 1882 obj->debug_hold_bitmap |= (1 << i); 1883 obj->debug_hold_thrs[i] = curthread; 1884 #endif 1885 } 1886 1887 void 1888 vm_object_drop(vm_object_t obj) 1889 { 1890 int rc; 1891 1892 #if defined(DEBUG_LOCKS) 1893 int found = 0; 1894 int i; 1895 1896 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1897 if ((obj->debug_hold_bitmap & (1 << i)) && 1898 (obj->debug_hold_thrs[i] == curthread)) { 1899 obj->debug_hold_bitmap &= ~(1 << i); 1900 obj->debug_hold_thrs[i] = NULL; 1901 found = 1; 1902 break; 1903 } 1904 } 1905 1906 if (found == 0 && obj->debug_hold_ovfl == 0) 1907 panic("vm_object: attempt to drop hold on non-self-held obj"); 1908 #endif 1909 1910 rc = refcount_release(&obj->hold_count); 1911 vm_object_unlock(obj); 1912 1913 if (rc) 1914 vm_object_hold_wake(obj); 1915 } 1916 1917 static void 1918 vm_object_hold_wake(vm_object_t obj) 1919 { 1920 wakeup(obj); 1921 } 1922 1923 static void 1924 vm_object_hold_wait(vm_object_t obj) 1925 { 1926 vm_object_lock(obj); 1927 1928 #if defined(DEBUG_LOCKS) 1929 int i; 1930 1931 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) { 1932 if ((obj->debug_hold_bitmap & (1 << i)) && 1933 (obj->debug_hold_thrs[i] == curthread)) 1934 panic("vm_object: self-hold in terminate or collapse"); 1935 } 1936 #endif 1937 1938 while (obj->hold_count) 1939 tsleep(obj, 0, "vmobjhld", 0); 1940 1941 vm_object_unlock(obj); 1942 } 1943 1944 #include "opt_ddb.h" 1945 #ifdef DDB 1946 #include <sys/kernel.h> 1947 1948 #include <sys/cons.h> 1949 1950 #include <ddb/ddb.h> 1951 1952 static int _vm_object_in_map (vm_map_t map, vm_object_t object, 1953 vm_map_entry_t entry); 1954 static int vm_object_in_map (vm_object_t object); 1955 1956 /* 1957 * The caller must hold vm_token. 1958 */ 1959 static int 1960 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1961 { 1962 vm_map_t tmpm; 1963 vm_map_entry_t tmpe; 1964 vm_object_t obj; 1965 int entcount; 1966 1967 if (map == 0) 1968 return 0; 1969 if (entry == 0) { 1970 tmpe = map->header.next; 1971 entcount = map->nentries; 1972 while (entcount-- && (tmpe != &map->header)) { 1973 if( _vm_object_in_map(map, object, tmpe)) { 1974 return 1; 1975 } 1976 tmpe = tmpe->next; 1977 } 1978 return (0); 1979 } 1980 switch(entry->maptype) { 1981 case VM_MAPTYPE_SUBMAP: 1982 tmpm = entry->object.sub_map; 1983 tmpe = tmpm->header.next; 1984 entcount = tmpm->nentries; 1985 while (entcount-- && tmpe != &tmpm->header) { 1986 if( _vm_object_in_map(tmpm, object, tmpe)) { 1987 return 1; 1988 } 1989 tmpe = tmpe->next; 1990 } 1991 break; 1992 case VM_MAPTYPE_NORMAL: 1993 case VM_MAPTYPE_VPAGETABLE: 1994 obj = entry->object.vm_object; 1995 while (obj) { 1996 if (obj == object) 1997 return 1; 1998 obj = obj->backing_object; 1999 } 2000 break; 2001 default: 2002 break; 2003 } 2004 return 0; 2005 } 2006 2007 static int vm_object_in_map_callback(struct proc *p, void *data); 2008 2009 struct vm_object_in_map_info { 2010 vm_object_t object; 2011 int rv; 2012 }; 2013 2014 /* 2015 * Debugging only 2016 */ 2017 static int 2018 vm_object_in_map(vm_object_t object) 2019 { 2020 struct vm_object_in_map_info info; 2021 2022 info.rv = 0; 2023 info.object = object; 2024 2025 allproc_scan(vm_object_in_map_callback, &info); 2026 if (info.rv) 2027 return 1; 2028 if( _vm_object_in_map(&kernel_map, object, 0)) 2029 return 1; 2030 if( _vm_object_in_map(&pager_map, object, 0)) 2031 return 1; 2032 if( _vm_object_in_map(&buffer_map, object, 0)) 2033 return 1; 2034 return 0; 2035 } 2036 2037 /* 2038 * Debugging only 2039 */ 2040 static int 2041 vm_object_in_map_callback(struct proc *p, void *data) 2042 { 2043 struct vm_object_in_map_info *info = data; 2044 2045 if (p->p_vmspace) { 2046 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) { 2047 info->rv = 1; 2048 return -1; 2049 } 2050 } 2051 return (0); 2052 } 2053 2054 DB_SHOW_COMMAND(vmochk, vm_object_check) 2055 { 2056 vm_object_t object; 2057 2058 /* 2059 * make sure that internal objs are in a map somewhere 2060 * and none have zero ref counts. 2061 */ 2062 for (object = TAILQ_FIRST(&vm_object_list); 2063 object != NULL; 2064 object = TAILQ_NEXT(object, object_list)) { 2065 if (object->type == OBJT_MARKER) 2066 continue; 2067 if (object->handle == NULL && 2068 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2069 if (object->ref_count == 0) { 2070 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2071 (long)object->size); 2072 } 2073 if (!vm_object_in_map(object)) { 2074 db_printf( 2075 "vmochk: internal obj is not in a map: " 2076 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2077 object->ref_count, (u_long)object->size, 2078 (u_long)object->size, 2079 (void *)object->backing_object); 2080 } 2081 } 2082 } 2083 } 2084 2085 /* 2086 * Debugging only 2087 */ 2088 DB_SHOW_COMMAND(object, vm_object_print_static) 2089 { 2090 /* XXX convert args. */ 2091 vm_object_t object = (vm_object_t)addr; 2092 boolean_t full = have_addr; 2093 2094 vm_page_t p; 2095 2096 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2097 #define count was_count 2098 2099 int count; 2100 2101 if (object == NULL) 2102 return; 2103 2104 db_iprintf( 2105 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2106 object, (int)object->type, (u_long)object->size, 2107 object->resident_page_count, object->ref_count, object->flags); 2108 /* 2109 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2110 */ 2111 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2112 object->shadow_count, 2113 object->backing_object ? object->backing_object->ref_count : 0, 2114 object->backing_object, (long)object->backing_object_offset); 2115 2116 if (!full) 2117 return; 2118 2119 db_indent += 2; 2120 count = 0; 2121 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) { 2122 if (count == 0) 2123 db_iprintf("memory:="); 2124 else if (count == 6) { 2125 db_printf("\n"); 2126 db_iprintf(" ..."); 2127 count = 0; 2128 } else 2129 db_printf(","); 2130 count++; 2131 2132 db_printf("(off=0x%lx,page=0x%lx)", 2133 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2134 } 2135 if (count != 0) 2136 db_printf("\n"); 2137 db_indent -= 2; 2138 } 2139 2140 /* XXX. */ 2141 #undef count 2142 2143 /* 2144 * XXX need this non-static entry for calling from vm_map_print. 2145 * 2146 * Debugging only 2147 */ 2148 void 2149 vm_object_print(/* db_expr_t */ long addr, 2150 boolean_t have_addr, 2151 /* db_expr_t */ long count, 2152 char *modif) 2153 { 2154 vm_object_print_static(addr, have_addr, count, modif); 2155 } 2156 2157 /* 2158 * Debugging only 2159 */ 2160 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2161 { 2162 vm_object_t object; 2163 int nl = 0; 2164 int c; 2165 for (object = TAILQ_FIRST(&vm_object_list); 2166 object != NULL; 2167 object = TAILQ_NEXT(object, object_list)) { 2168 vm_pindex_t idx, fidx; 2169 vm_pindex_t osize; 2170 vm_paddr_t pa = -1, padiff; 2171 int rcount; 2172 vm_page_t m; 2173 2174 if (object->type == OBJT_MARKER) 2175 continue; 2176 db_printf("new object: %p\n", (void *)object); 2177 if ( nl > 18) { 2178 c = cngetc(); 2179 if (c != ' ') 2180 return; 2181 nl = 0; 2182 } 2183 nl++; 2184 rcount = 0; 2185 fidx = 0; 2186 osize = object->size; 2187 if (osize > 128) 2188 osize = 128; 2189 for (idx = 0; idx < osize; idx++) { 2190 m = vm_page_lookup(object, idx); 2191 if (m == NULL) { 2192 if (rcount) { 2193 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2194 (long)fidx, rcount, (long)pa); 2195 if ( nl > 18) { 2196 c = cngetc(); 2197 if (c != ' ') 2198 return; 2199 nl = 0; 2200 } 2201 nl++; 2202 rcount = 0; 2203 } 2204 continue; 2205 } 2206 2207 2208 if (rcount && 2209 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2210 ++rcount; 2211 continue; 2212 } 2213 if (rcount) { 2214 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2215 padiff >>= PAGE_SHIFT; 2216 padiff &= PQ_L2_MASK; 2217 if (padiff == 0) { 2218 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2219 ++rcount; 2220 continue; 2221 } 2222 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2223 (long)fidx, rcount, (long)pa); 2224 db_printf("pd(%ld)\n", (long)padiff); 2225 if ( nl > 18) { 2226 c = cngetc(); 2227 if (c != ' ') 2228 return; 2229 nl = 0; 2230 } 2231 nl++; 2232 } 2233 fidx = idx; 2234 pa = VM_PAGE_TO_PHYS(m); 2235 rcount = 1; 2236 } 2237 if (rcount) { 2238 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2239 (long)fidx, rcount, (long)pa); 2240 if ( nl > 18) { 2241 c = cngetc(); 2242 if (c != ' ') 2243 return; 2244 nl = 0; 2245 } 2246 nl++; 2247 } 2248 } 2249 } 2250 #endif /* DDB */ 2251