1 /*- 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61 /* 62 * Virtual memory object module. 63 */ 64 65 #include <sys/cdefs.h> 66 __FBSDID("$FreeBSD$"); 67 68 #include <sys/param.h> 69 #include <sys/systm.h> 70 #include <sys/lock.h> 71 #include <sys/mman.h> 72 #include <sys/mount.h> 73 #include <sys/kernel.h> 74 #include <sys/sysctl.h> 75 #include <sys/mutex.h> 76 #include <sys/proc.h> /* for curproc, pageproc */ 77 #include <sys/socket.h> 78 #include <sys/vnode.h> 79 #include <sys/vmmeter.h> 80 #include <sys/sx.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <vm/pmap.h> 85 #include <vm/vm_map.h> 86 #include <vm/vm_object.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_pageout.h> 89 #include <vm/vm_pager.h> 90 #include <vm/swap_pager.h> 91 #include <vm/vm_kern.h> 92 #include <vm/vm_extern.h> 93 #include <vm/uma.h> 94 95 #define EASY_SCAN_FACTOR 8 96 97 #define MSYNC_FLUSH_HARDSEQ 0x01 98 #define MSYNC_FLUSH_SOFTSEQ 0x02 99 100 /* 101 * msync / VM object flushing optimizations 102 */ 103 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 104 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 105 CTLFLAG_RW, &msync_flush_flags, 0, ""); 106 107 static int old_msync; 108 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 109 "Use old (insecure) msync behavior"); 110 111 static void vm_object_qcollapse(vm_object_t object); 112 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 113 static void vm_object_vndeallocate(vm_object_t object); 114 115 /* 116 * Virtual memory objects maintain the actual data 117 * associated with allocated virtual memory. A given 118 * page of memory exists within exactly one object. 119 * 120 * An object is only deallocated when all "references" 121 * are given up. Only one "reference" to a given 122 * region of an object should be writeable. 123 * 124 * Associated with each object is a list of all resident 125 * memory pages belonging to that object; this list is 126 * maintained by the "vm_page" module, and locked by the object's 127 * lock. 128 * 129 * Each object also records a "pager" routine which is 130 * used to retrieve (and store) pages to the proper backing 131 * storage. In addition, objects may be backed by other 132 * objects from which they were virtual-copied. 133 * 134 * The only items within the object structure which are 135 * modified after time of creation are: 136 * reference count locked by object's lock 137 * pager routine locked by object's lock 138 * 139 */ 140 141 struct object_q vm_object_list; 142 struct mtx vm_object_list_mtx; /* lock for object list and count */ 143 144 struct vm_object kernel_object_store; 145 struct vm_object kmem_object_store; 146 147 SYSCTL_DECL(_vm_stats); 148 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); 149 150 static long object_collapses; 151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 152 &object_collapses, 0, "VM object collapses"); 153 154 static long object_bypasses; 155 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 156 &object_bypasses, 0, "VM object bypasses"); 157 158 /* 159 * next_index determines the page color that is assigned to the next 160 * allocated object. Accesses to next_index are not synchronized 161 * because the effects of two or more object allocations using 162 * next_index simultaneously are inconsequential. At any given time, 163 * numerous objects have the same page color. 164 */ 165 static int next_index; 166 167 static uma_zone_t obj_zone; 168 169 static int vm_object_zinit(void *mem, int size, int flags); 170 171 #ifdef INVARIANTS 172 static void vm_object_zdtor(void *mem, int size, void *arg); 173 174 static void 175 vm_object_zdtor(void *mem, int size, void *arg) 176 { 177 vm_object_t object; 178 179 object = (vm_object_t)mem; 180 KASSERT(TAILQ_EMPTY(&object->memq), 181 ("object %p has resident pages", 182 object)); 183 KASSERT(object->paging_in_progress == 0, 184 ("object %p paging_in_progress = %d", 185 object, object->paging_in_progress)); 186 KASSERT(object->resident_page_count == 0, 187 ("object %p resident_page_count = %d", 188 object, object->resident_page_count)); 189 KASSERT(object->shadow_count == 0, 190 ("object %p shadow_count = %d", 191 object, object->shadow_count)); 192 } 193 #endif 194 195 static int 196 vm_object_zinit(void *mem, int size, int flags) 197 { 198 vm_object_t object; 199 200 object = (vm_object_t)mem; 201 bzero(&object->mtx, sizeof(object->mtx)); 202 VM_OBJECT_LOCK_INIT(object, "standard object"); 203 204 /* These are true for any object that has been freed */ 205 object->paging_in_progress = 0; 206 object->resident_page_count = 0; 207 object->shadow_count = 0; 208 return (0); 209 } 210 211 void 212 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 213 { 214 int incr; 215 216 TAILQ_INIT(&object->memq); 217 LIST_INIT(&object->shadow_head); 218 219 object->root = NULL; 220 object->type = type; 221 object->size = size; 222 object->generation = 1; 223 object->ref_count = 1; 224 object->flags = 0; 225 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 226 object->flags = OBJ_ONEMAPPING; 227 incr = PQ_MAXLENGTH; 228 if (size <= incr) 229 incr = size; 230 object->pg_color = next_index; 231 next_index = (object->pg_color + incr) & PQ_COLORMASK; 232 object->handle = NULL; 233 object->backing_object = NULL; 234 object->backing_object_offset = (vm_ooffset_t) 0; 235 236 mtx_lock(&vm_object_list_mtx); 237 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 238 mtx_unlock(&vm_object_list_mtx); 239 } 240 241 /* 242 * vm_object_init: 243 * 244 * Initialize the VM objects module. 245 */ 246 void 247 vm_object_init(void) 248 { 249 TAILQ_INIT(&vm_object_list); 250 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 251 252 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object"); 253 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 254 kernel_object); 255 256 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object"); 257 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 258 kmem_object); 259 260 /* 261 * The lock portion of struct vm_object must be type stable due 262 * to vm_pageout_fallback_object_lock locking a vm object 263 * without holding any references to it. 264 */ 265 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 266 #ifdef INVARIANTS 267 vm_object_zdtor, 268 #else 269 NULL, 270 #endif 271 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); 272 } 273 274 void 275 vm_object_clear_flag(vm_object_t object, u_short bits) 276 { 277 278 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 279 object->flags &= ~bits; 280 } 281 282 void 283 vm_object_pip_add(vm_object_t object, short i) 284 { 285 286 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 287 object->paging_in_progress += i; 288 } 289 290 void 291 vm_object_pip_subtract(vm_object_t object, short i) 292 { 293 294 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 295 object->paging_in_progress -= i; 296 } 297 298 void 299 vm_object_pip_wakeup(vm_object_t object) 300 { 301 302 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 303 object->paging_in_progress--; 304 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 305 vm_object_clear_flag(object, OBJ_PIPWNT); 306 wakeup(object); 307 } 308 } 309 310 void 311 vm_object_pip_wakeupn(vm_object_t object, short i) 312 { 313 314 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 315 if (i) 316 object->paging_in_progress -= i; 317 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 318 vm_object_clear_flag(object, OBJ_PIPWNT); 319 wakeup(object); 320 } 321 } 322 323 void 324 vm_object_pip_wait(vm_object_t object, char *waitid) 325 { 326 327 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 328 while (object->paging_in_progress) { 329 object->flags |= OBJ_PIPWNT; 330 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 331 } 332 } 333 334 /* 335 * vm_object_allocate: 336 * 337 * Returns a new object with the given size. 338 */ 339 vm_object_t 340 vm_object_allocate(objtype_t type, vm_pindex_t size) 341 { 342 vm_object_t object; 343 344 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 345 _vm_object_allocate(type, size, object); 346 return (object); 347 } 348 349 350 /* 351 * vm_object_reference: 352 * 353 * Gets another reference to the given object. Note: OBJ_DEAD 354 * objects can be referenced during final cleaning. 355 */ 356 void 357 vm_object_reference(vm_object_t object) 358 { 359 struct vnode *vp; 360 361 if (object == NULL) 362 return; 363 VM_OBJECT_LOCK(object); 364 object->ref_count++; 365 if (object->type == OBJT_VNODE) { 366 int vfslocked; 367 368 vp = object->handle; 369 VM_OBJECT_UNLOCK(object); 370 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 371 vget(vp, LK_RETRY, curthread); 372 VFS_UNLOCK_GIANT(vfslocked); 373 } else 374 VM_OBJECT_UNLOCK(object); 375 } 376 377 /* 378 * vm_object_reference_locked: 379 * 380 * Gets another reference to the given object. 381 * 382 * The object must be locked. 383 */ 384 void 385 vm_object_reference_locked(vm_object_t object) 386 { 387 struct vnode *vp; 388 389 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 390 KASSERT((object->flags & OBJ_DEAD) == 0, 391 ("vm_object_reference_locked: dead object referenced")); 392 object->ref_count++; 393 if (object->type == OBJT_VNODE) { 394 vp = object->handle; 395 vref(vp); 396 } 397 } 398 399 /* 400 * Handle deallocating an object of type OBJT_VNODE. 401 */ 402 static void 403 vm_object_vndeallocate(vm_object_t object) 404 { 405 struct vnode *vp = (struct vnode *) object->handle; 406 407 VFS_ASSERT_GIANT(vp->v_mount); 408 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 409 KASSERT(object->type == OBJT_VNODE, 410 ("vm_object_vndeallocate: not a vnode object")); 411 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 412 #ifdef INVARIANTS 413 if (object->ref_count == 0) { 414 vprint("vm_object_vndeallocate", vp); 415 panic("vm_object_vndeallocate: bad object reference count"); 416 } 417 #endif 418 419 object->ref_count--; 420 if (object->ref_count == 0) { 421 mp_fixme("Unlocked vflag access."); 422 vp->v_vflag &= ~VV_TEXT; 423 } 424 VM_OBJECT_UNLOCK(object); 425 /* 426 * vrele may need a vop lock 427 */ 428 vrele(vp); 429 } 430 431 /* 432 * vm_object_deallocate: 433 * 434 * Release a reference to the specified object, 435 * gained either through a vm_object_allocate 436 * or a vm_object_reference call. When all references 437 * are gone, storage associated with this object 438 * may be relinquished. 439 * 440 * No object may be locked. 441 */ 442 void 443 vm_object_deallocate(vm_object_t object) 444 { 445 vm_object_t temp; 446 447 while (object != NULL) { 448 int vfslocked; 449 450 vfslocked = 0; 451 restart: 452 VM_OBJECT_LOCK(object); 453 if (object->type == OBJT_VNODE) { 454 struct vnode *vp = (struct vnode *) object->handle; 455 456 /* 457 * Conditionally acquire Giant for a vnode-backed 458 * object. We have to be careful since the type of 459 * a vnode object can change while the object is 460 * unlocked. 461 */ 462 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) { 463 vfslocked = 1; 464 if (!mtx_trylock(&Giant)) { 465 VM_OBJECT_UNLOCK(object); 466 mtx_lock(&Giant); 467 goto restart; 468 } 469 } 470 vm_object_vndeallocate(object); 471 VFS_UNLOCK_GIANT(vfslocked); 472 return; 473 } else 474 /* 475 * This is to handle the case that the object 476 * changed type while we dropped its lock to 477 * obtain Giant. 478 */ 479 VFS_UNLOCK_GIANT(vfslocked); 480 481 KASSERT(object->ref_count != 0, 482 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 483 484 /* 485 * If the reference count goes to 0 we start calling 486 * vm_object_terminate() on the object chain. 487 * A ref count of 1 may be a special case depending on the 488 * shadow count being 0 or 1. 489 */ 490 object->ref_count--; 491 if (object->ref_count > 1) { 492 VM_OBJECT_UNLOCK(object); 493 return; 494 } else if (object->ref_count == 1) { 495 if (object->shadow_count == 0) { 496 vm_object_set_flag(object, OBJ_ONEMAPPING); 497 } else if ((object->shadow_count == 1) && 498 (object->handle == NULL) && 499 (object->type == OBJT_DEFAULT || 500 object->type == OBJT_SWAP)) { 501 vm_object_t robject; 502 503 robject = LIST_FIRST(&object->shadow_head); 504 KASSERT(robject != NULL, 505 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 506 object->ref_count, 507 object->shadow_count)); 508 if (!VM_OBJECT_TRYLOCK(robject)) { 509 /* 510 * Avoid a potential deadlock. 511 */ 512 object->ref_count++; 513 VM_OBJECT_UNLOCK(object); 514 /* 515 * More likely than not the thread 516 * holding robject's lock has lower 517 * priority than the current thread. 518 * Let the lower priority thread run. 519 */ 520 tsleep(&proc0, PVM, "vmo_de", 1); 521 continue; 522 } 523 /* 524 * Collapse object into its shadow unless its 525 * shadow is dead. In that case, object will 526 * be deallocated by the thread that is 527 * deallocating its shadow. 528 */ 529 if ((robject->flags & OBJ_DEAD) == 0 && 530 (robject->handle == NULL) && 531 (robject->type == OBJT_DEFAULT || 532 robject->type == OBJT_SWAP)) { 533 534 robject->ref_count++; 535 retry: 536 if (robject->paging_in_progress) { 537 VM_OBJECT_UNLOCK(object); 538 vm_object_pip_wait(robject, 539 "objde1"); 540 temp = robject->backing_object; 541 if (object == temp) { 542 VM_OBJECT_LOCK(object); 543 goto retry; 544 } 545 } else if (object->paging_in_progress) { 546 VM_OBJECT_UNLOCK(robject); 547 object->flags |= OBJ_PIPWNT; 548 msleep(object, 549 VM_OBJECT_MTX(object), 550 PDROP | PVM, "objde2", 0); 551 VM_OBJECT_LOCK(robject); 552 temp = robject->backing_object; 553 if (object == temp) { 554 VM_OBJECT_LOCK(object); 555 goto retry; 556 } 557 } else 558 VM_OBJECT_UNLOCK(object); 559 560 if (robject->ref_count == 1) { 561 robject->ref_count--; 562 object = robject; 563 goto doterm; 564 } 565 object = robject; 566 vm_object_collapse(object); 567 VM_OBJECT_UNLOCK(object); 568 continue; 569 } 570 VM_OBJECT_UNLOCK(robject); 571 } 572 VM_OBJECT_UNLOCK(object); 573 return; 574 } 575 doterm: 576 temp = object->backing_object; 577 if (temp != NULL) { 578 VM_OBJECT_LOCK(temp); 579 LIST_REMOVE(object, shadow_list); 580 temp->shadow_count--; 581 temp->generation++; 582 VM_OBJECT_UNLOCK(temp); 583 object->backing_object = NULL; 584 } 585 /* 586 * Don't double-terminate, we could be in a termination 587 * recursion due to the terminate having to sync data 588 * to disk. 589 */ 590 if ((object->flags & OBJ_DEAD) == 0) 591 vm_object_terminate(object); 592 else 593 VM_OBJECT_UNLOCK(object); 594 object = temp; 595 } 596 } 597 598 /* 599 * vm_object_terminate actually destroys the specified object, freeing 600 * up all previously used resources. 601 * 602 * The object must be locked. 603 * This routine may block. 604 */ 605 void 606 vm_object_terminate(vm_object_t object) 607 { 608 vm_page_t p; 609 610 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 611 612 /* 613 * Make sure no one uses us. 614 */ 615 vm_object_set_flag(object, OBJ_DEAD); 616 617 /* 618 * wait for the pageout daemon to be done with the object 619 */ 620 vm_object_pip_wait(object, "objtrm"); 621 622 KASSERT(!object->paging_in_progress, 623 ("vm_object_terminate: pageout in progress")); 624 625 /* 626 * Clean and free the pages, as appropriate. All references to the 627 * object are gone, so we don't need to lock it. 628 */ 629 if (object->type == OBJT_VNODE) { 630 struct vnode *vp = (struct vnode *)object->handle; 631 632 /* 633 * Clean pages and flush buffers. 634 */ 635 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 636 VM_OBJECT_UNLOCK(object); 637 638 vinvalbuf(vp, V_SAVE, NULL, 0, 0); 639 640 VM_OBJECT_LOCK(object); 641 } 642 643 KASSERT(object->ref_count == 0, 644 ("vm_object_terminate: object with references, ref_count=%d", 645 object->ref_count)); 646 647 /* 648 * Now free any remaining pages. For internal objects, this also 649 * removes them from paging queues. Don't free wired pages, just 650 * remove them from the object. 651 */ 652 vm_page_lock_queues(); 653 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 654 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 655 ("vm_object_terminate: freeing busy page %p " 656 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 657 if (p->wire_count == 0) { 658 vm_page_free(p); 659 cnt.v_pfree++; 660 } else { 661 vm_page_remove(p); 662 } 663 } 664 vm_page_unlock_queues(); 665 666 /* 667 * Let the pager know object is dead. 668 */ 669 vm_pager_deallocate(object); 670 VM_OBJECT_UNLOCK(object); 671 672 /* 673 * Remove the object from the global object list. 674 */ 675 mtx_lock(&vm_object_list_mtx); 676 TAILQ_REMOVE(&vm_object_list, object, object_list); 677 mtx_unlock(&vm_object_list_mtx); 678 679 /* 680 * Free the space for the object. 681 */ 682 uma_zfree(obj_zone, object); 683 } 684 685 /* 686 * vm_object_page_clean 687 * 688 * Clean all dirty pages in the specified range of object. Leaves page 689 * on whatever queue it is currently on. If NOSYNC is set then do not 690 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 691 * leaving the object dirty. 692 * 693 * When stuffing pages asynchronously, allow clustering. XXX we need a 694 * synchronous clustering mode implementation. 695 * 696 * Odd semantics: if start == end, we clean everything. 697 * 698 * The object must be locked. 699 */ 700 void 701 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 702 { 703 vm_page_t p, np; 704 vm_pindex_t tstart, tend; 705 vm_pindex_t pi; 706 int clearobjflags; 707 int pagerflags; 708 int curgeneration; 709 710 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 711 if (object->type != OBJT_VNODE || 712 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 713 return; 714 715 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 716 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 717 718 vm_object_set_flag(object, OBJ_CLEANING); 719 720 tstart = start; 721 if (end == 0) { 722 tend = object->size; 723 } else { 724 tend = end; 725 } 726 727 vm_page_lock_queues(); 728 /* 729 * If the caller is smart and only msync()s a range he knows is 730 * dirty, we may be able to avoid an object scan. This results in 731 * a phenominal improvement in performance. We cannot do this 732 * as a matter of course because the object may be huge - e.g. 733 * the size might be in the gigabytes or terrabytes. 734 */ 735 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 736 vm_pindex_t tscan; 737 int scanlimit; 738 int scanreset; 739 740 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 741 if (scanreset < 16) 742 scanreset = 16; 743 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 744 745 scanlimit = scanreset; 746 tscan = tstart; 747 while (tscan < tend) { 748 curgeneration = object->generation; 749 p = vm_page_lookup(object, tscan); 750 if (p == NULL || p->valid == 0 || 751 VM_PAGE_INQUEUE1(p, PQ_CACHE)) { 752 if (--scanlimit == 0) 753 break; 754 ++tscan; 755 continue; 756 } 757 vm_page_test_dirty(p); 758 if ((p->dirty & p->valid) == 0) { 759 if (--scanlimit == 0) 760 break; 761 ++tscan; 762 continue; 763 } 764 /* 765 * If we have been asked to skip nosync pages and 766 * this is a nosync page, we can't continue. 767 */ 768 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 769 if (--scanlimit == 0) 770 break; 771 ++tscan; 772 continue; 773 } 774 scanlimit = scanreset; 775 776 /* 777 * This returns 0 if it was unable to busy the first 778 * page (i.e. had to sleep). 779 */ 780 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 781 } 782 783 /* 784 * If everything was dirty and we flushed it successfully, 785 * and the requested range is not the entire object, we 786 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 787 * return immediately. 788 */ 789 if (tscan >= tend && (tstart || tend < object->size)) { 790 vm_page_unlock_queues(); 791 vm_object_clear_flag(object, OBJ_CLEANING); 792 return; 793 } 794 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 795 } 796 797 /* 798 * Generally set CLEANCHK interlock and make the page read-only so 799 * we can then clear the object flags. 800 * 801 * However, if this is a nosync mmap then the object is likely to 802 * stay dirty so do not mess with the page and do not clear the 803 * object flags. 804 */ 805 clearobjflags = 1; 806 TAILQ_FOREACH(p, &object->memq, listq) { 807 vm_page_flag_set(p, PG_CLEANCHK); 808 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 809 clearobjflags = 0; 810 else 811 pmap_remove_write(p); 812 } 813 814 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 815 struct vnode *vp; 816 817 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 818 if (object->type == OBJT_VNODE && 819 (vp = (struct vnode *)object->handle) != NULL) { 820 VI_LOCK(vp); 821 if (vp->v_iflag & VI_OBJDIRTY) 822 vp->v_iflag &= ~VI_OBJDIRTY; 823 VI_UNLOCK(vp); 824 } 825 } 826 827 rescan: 828 curgeneration = object->generation; 829 830 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 831 int n; 832 833 np = TAILQ_NEXT(p, listq); 834 835 again: 836 pi = p->pindex; 837 if (((p->flags & PG_CLEANCHK) == 0) || 838 (pi < tstart) || (pi >= tend) || 839 (p->valid == 0) || 840 VM_PAGE_INQUEUE1(p, PQ_CACHE)) { 841 vm_page_flag_clear(p, PG_CLEANCHK); 842 continue; 843 } 844 845 vm_page_test_dirty(p); 846 if ((p->dirty & p->valid) == 0) { 847 vm_page_flag_clear(p, PG_CLEANCHK); 848 continue; 849 } 850 851 /* 852 * If we have been asked to skip nosync pages and this is a 853 * nosync page, skip it. Note that the object flags were 854 * not cleared in this case so we do not have to set them. 855 */ 856 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 857 vm_page_flag_clear(p, PG_CLEANCHK); 858 continue; 859 } 860 861 n = vm_object_page_collect_flush(object, p, 862 curgeneration, pagerflags); 863 if (n == 0) 864 goto rescan; 865 866 if (object->generation != curgeneration) 867 goto rescan; 868 869 /* 870 * Try to optimize the next page. If we can't we pick up 871 * our (random) scan where we left off. 872 */ 873 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 874 if ((p = vm_page_lookup(object, pi + n)) != NULL) 875 goto again; 876 } 877 } 878 vm_page_unlock_queues(); 879 #if 0 880 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 881 #endif 882 883 vm_object_clear_flag(object, OBJ_CLEANING); 884 return; 885 } 886 887 static int 888 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 889 { 890 int runlen; 891 int maxf; 892 int chkb; 893 int maxb; 894 int i; 895 vm_pindex_t pi; 896 vm_page_t maf[vm_pageout_page_count]; 897 vm_page_t mab[vm_pageout_page_count]; 898 vm_page_t ma[vm_pageout_page_count]; 899 900 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 901 pi = p->pindex; 902 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 903 vm_page_lock_queues(); 904 if (object->generation != curgeneration) { 905 return(0); 906 } 907 } 908 maxf = 0; 909 for(i = 1; i < vm_pageout_page_count; i++) { 910 vm_page_t tp; 911 912 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 913 if ((tp->flags & PG_BUSY) || 914 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 915 (tp->flags & PG_CLEANCHK) == 0) || 916 (tp->busy != 0)) 917 break; 918 if (VM_PAGE_INQUEUE1(tp, PQ_CACHE)) { 919 vm_page_flag_clear(tp, PG_CLEANCHK); 920 break; 921 } 922 vm_page_test_dirty(tp); 923 if ((tp->dirty & tp->valid) == 0) { 924 vm_page_flag_clear(tp, PG_CLEANCHK); 925 break; 926 } 927 maf[ i - 1 ] = tp; 928 maxf++; 929 continue; 930 } 931 break; 932 } 933 934 maxb = 0; 935 chkb = vm_pageout_page_count - maxf; 936 if (chkb) { 937 for(i = 1; i < chkb;i++) { 938 vm_page_t tp; 939 940 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 941 if ((tp->flags & PG_BUSY) || 942 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 943 (tp->flags & PG_CLEANCHK) == 0) || 944 (tp->busy != 0)) 945 break; 946 if (VM_PAGE_INQUEUE1(tp, PQ_CACHE)) { 947 vm_page_flag_clear(tp, PG_CLEANCHK); 948 break; 949 } 950 vm_page_test_dirty(tp); 951 if ((tp->dirty & tp->valid) == 0) { 952 vm_page_flag_clear(tp, PG_CLEANCHK); 953 break; 954 } 955 mab[ i - 1 ] = tp; 956 maxb++; 957 continue; 958 } 959 break; 960 } 961 } 962 963 for(i = 0; i < maxb; i++) { 964 int index = (maxb - i) - 1; 965 ma[index] = mab[i]; 966 vm_page_flag_clear(ma[index], PG_CLEANCHK); 967 } 968 vm_page_flag_clear(p, PG_CLEANCHK); 969 ma[maxb] = p; 970 for(i = 0; i < maxf; i++) { 971 int index = (maxb + i) + 1; 972 ma[index] = maf[i]; 973 vm_page_flag_clear(ma[index], PG_CLEANCHK); 974 } 975 runlen = maxb + maxf + 1; 976 977 vm_pageout_flush(ma, runlen, pagerflags); 978 for (i = 0; i < runlen; i++) { 979 if (ma[i]->valid & ma[i]->dirty) { 980 pmap_remove_write(ma[i]); 981 vm_page_flag_set(ma[i], PG_CLEANCHK); 982 983 /* 984 * maxf will end up being the actual number of pages 985 * we wrote out contiguously, non-inclusive of the 986 * first page. We do not count look-behind pages. 987 */ 988 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 989 maxf = i - maxb - 1; 990 } 991 } 992 return(maxf + 1); 993 } 994 995 /* 996 * Note that there is absolutely no sense in writing out 997 * anonymous objects, so we track down the vnode object 998 * to write out. 999 * We invalidate (remove) all pages from the address space 1000 * for semantic correctness. 1001 * 1002 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1003 * may start out with a NULL object. 1004 */ 1005 void 1006 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1007 boolean_t syncio, boolean_t invalidate) 1008 { 1009 vm_object_t backing_object; 1010 struct vnode *vp; 1011 struct mount *mp; 1012 int flags; 1013 1014 if (object == NULL) 1015 return; 1016 VM_OBJECT_LOCK(object); 1017 while ((backing_object = object->backing_object) != NULL) { 1018 VM_OBJECT_LOCK(backing_object); 1019 offset += object->backing_object_offset; 1020 VM_OBJECT_UNLOCK(object); 1021 object = backing_object; 1022 if (object->size < OFF_TO_IDX(offset + size)) 1023 size = IDX_TO_OFF(object->size) - offset; 1024 } 1025 /* 1026 * Flush pages if writing is allowed, invalidate them 1027 * if invalidation requested. Pages undergoing I/O 1028 * will be ignored by vm_object_page_remove(). 1029 * 1030 * We cannot lock the vnode and then wait for paging 1031 * to complete without deadlocking against vm_fault. 1032 * Instead we simply call vm_object_page_remove() and 1033 * allow it to block internally on a page-by-page 1034 * basis when it encounters pages undergoing async 1035 * I/O. 1036 */ 1037 if (object->type == OBJT_VNODE && 1038 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1039 int vfslocked; 1040 vp = object->handle; 1041 VM_OBJECT_UNLOCK(object); 1042 (void) vn_start_write(vp, &mp, V_WAIT); 1043 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1044 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread); 1045 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1046 flags |= invalidate ? OBJPC_INVAL : 0; 1047 VM_OBJECT_LOCK(object); 1048 vm_object_page_clean(object, 1049 OFF_TO_IDX(offset), 1050 OFF_TO_IDX(offset + size + PAGE_MASK), 1051 flags); 1052 VM_OBJECT_UNLOCK(object); 1053 VOP_UNLOCK(vp, 0, curthread); 1054 VFS_UNLOCK_GIANT(vfslocked); 1055 vn_finished_write(mp); 1056 VM_OBJECT_LOCK(object); 1057 } 1058 if ((object->type == OBJT_VNODE || 1059 object->type == OBJT_DEVICE) && invalidate) { 1060 boolean_t purge; 1061 purge = old_msync || (object->type == OBJT_DEVICE); 1062 vm_object_page_remove(object, 1063 OFF_TO_IDX(offset), 1064 OFF_TO_IDX(offset + size + PAGE_MASK), 1065 purge ? FALSE : TRUE); 1066 } 1067 VM_OBJECT_UNLOCK(object); 1068 } 1069 1070 /* 1071 * vm_object_madvise: 1072 * 1073 * Implements the madvise function at the object/page level. 1074 * 1075 * MADV_WILLNEED (any object) 1076 * 1077 * Activate the specified pages if they are resident. 1078 * 1079 * MADV_DONTNEED (any object) 1080 * 1081 * Deactivate the specified pages if they are resident. 1082 * 1083 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1084 * OBJ_ONEMAPPING only) 1085 * 1086 * Deactivate and clean the specified pages if they are 1087 * resident. This permits the process to reuse the pages 1088 * without faulting or the kernel to reclaim the pages 1089 * without I/O. 1090 */ 1091 void 1092 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1093 { 1094 vm_pindex_t end, tpindex; 1095 vm_object_t backing_object, tobject; 1096 vm_page_t m; 1097 1098 if (object == NULL) 1099 return; 1100 VM_OBJECT_LOCK(object); 1101 end = pindex + count; 1102 /* 1103 * Locate and adjust resident pages 1104 */ 1105 for (; pindex < end; pindex += 1) { 1106 relookup: 1107 tobject = object; 1108 tpindex = pindex; 1109 shadowlookup: 1110 /* 1111 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1112 * and those pages must be OBJ_ONEMAPPING. 1113 */ 1114 if (advise == MADV_FREE) { 1115 if ((tobject->type != OBJT_DEFAULT && 1116 tobject->type != OBJT_SWAP) || 1117 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1118 goto unlock_tobject; 1119 } 1120 } 1121 m = vm_page_lookup(tobject, tpindex); 1122 if (m == NULL) { 1123 /* 1124 * There may be swap even if there is no backing page 1125 */ 1126 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1127 swap_pager_freespace(tobject, tpindex, 1); 1128 /* 1129 * next object 1130 */ 1131 backing_object = tobject->backing_object; 1132 if (backing_object == NULL) 1133 goto unlock_tobject; 1134 VM_OBJECT_LOCK(backing_object); 1135 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1136 if (tobject != object) 1137 VM_OBJECT_UNLOCK(tobject); 1138 tobject = backing_object; 1139 goto shadowlookup; 1140 } 1141 /* 1142 * If the page is busy or not in a normal active state, 1143 * we skip it. If the page is not managed there are no 1144 * page queues to mess with. Things can break if we mess 1145 * with pages in any of the below states. 1146 */ 1147 vm_page_lock_queues(); 1148 if (m->hold_count || 1149 m->wire_count || 1150 (m->flags & PG_UNMANAGED) || 1151 m->valid != VM_PAGE_BITS_ALL) { 1152 vm_page_unlock_queues(); 1153 goto unlock_tobject; 1154 } 1155 if ((m->flags & PG_BUSY) || m->busy) { 1156 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1157 vm_page_unlock_queues(); 1158 if (object != tobject) 1159 VM_OBJECT_UNLOCK(object); 1160 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0); 1161 VM_OBJECT_LOCK(object); 1162 goto relookup; 1163 } 1164 if (advise == MADV_WILLNEED) { 1165 vm_page_activate(m); 1166 } else if (advise == MADV_DONTNEED) { 1167 vm_page_dontneed(m); 1168 } else if (advise == MADV_FREE) { 1169 /* 1170 * Mark the page clean. This will allow the page 1171 * to be freed up by the system. However, such pages 1172 * are often reused quickly by malloc()/free() 1173 * so we do not do anything that would cause 1174 * a page fault if we can help it. 1175 * 1176 * Specifically, we do not try to actually free 1177 * the page now nor do we try to put it in the 1178 * cache (which would cause a page fault on reuse). 1179 * 1180 * But we do make the page is freeable as we 1181 * can without actually taking the step of unmapping 1182 * it. 1183 */ 1184 pmap_clear_modify(m); 1185 m->dirty = 0; 1186 m->act_count = 0; 1187 vm_page_dontneed(m); 1188 } 1189 vm_page_unlock_queues(); 1190 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1191 swap_pager_freespace(tobject, tpindex, 1); 1192 unlock_tobject: 1193 if (tobject != object) 1194 VM_OBJECT_UNLOCK(tobject); 1195 } 1196 VM_OBJECT_UNLOCK(object); 1197 } 1198 1199 /* 1200 * vm_object_shadow: 1201 * 1202 * Create a new object which is backed by the 1203 * specified existing object range. The source 1204 * object reference is deallocated. 1205 * 1206 * The new object and offset into that object 1207 * are returned in the source parameters. 1208 */ 1209 void 1210 vm_object_shadow( 1211 vm_object_t *object, /* IN/OUT */ 1212 vm_ooffset_t *offset, /* IN/OUT */ 1213 vm_size_t length) 1214 { 1215 vm_object_t source; 1216 vm_object_t result; 1217 1218 source = *object; 1219 1220 /* 1221 * Don't create the new object if the old object isn't shared. 1222 */ 1223 if (source != NULL) { 1224 VM_OBJECT_LOCK(source); 1225 if (source->ref_count == 1 && 1226 source->handle == NULL && 1227 (source->type == OBJT_DEFAULT || 1228 source->type == OBJT_SWAP)) { 1229 VM_OBJECT_UNLOCK(source); 1230 return; 1231 } 1232 VM_OBJECT_UNLOCK(source); 1233 } 1234 1235 /* 1236 * Allocate a new object with the given length. 1237 */ 1238 result = vm_object_allocate(OBJT_DEFAULT, length); 1239 1240 /* 1241 * The new object shadows the source object, adding a reference to it. 1242 * Our caller changes his reference to point to the new object, 1243 * removing a reference to the source object. Net result: no change 1244 * of reference count. 1245 * 1246 * Try to optimize the result object's page color when shadowing 1247 * in order to maintain page coloring consistency in the combined 1248 * shadowed object. 1249 */ 1250 result->backing_object = source; 1251 /* 1252 * Store the offset into the source object, and fix up the offset into 1253 * the new object. 1254 */ 1255 result->backing_object_offset = *offset; 1256 if (source != NULL) { 1257 VM_OBJECT_LOCK(source); 1258 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1259 source->shadow_count++; 1260 source->generation++; 1261 if (length < source->size) 1262 length = source->size; 1263 if (length > PQ_MAXLENGTH || source->generation > 1) 1264 length = PQ_MAXLENGTH; 1265 result->pg_color = (source->pg_color + 1266 length * source->generation) & PQ_COLORMASK; 1267 result->flags |= source->flags & OBJ_NEEDGIANT; 1268 VM_OBJECT_UNLOCK(source); 1269 next_index = (result->pg_color + PQ_MAXLENGTH) & PQ_COLORMASK; 1270 } 1271 1272 1273 /* 1274 * Return the new things 1275 */ 1276 *offset = 0; 1277 *object = result; 1278 } 1279 1280 /* 1281 * vm_object_split: 1282 * 1283 * Split the pages in a map entry into a new object. This affords 1284 * easier removal of unused pages, and keeps object inheritance from 1285 * being a negative impact on memory usage. 1286 */ 1287 void 1288 vm_object_split(vm_map_entry_t entry) 1289 { 1290 vm_page_t m; 1291 vm_object_t orig_object, new_object, source; 1292 vm_pindex_t offidxstart, offidxend; 1293 vm_size_t idx, size; 1294 1295 orig_object = entry->object.vm_object; 1296 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1297 return; 1298 if (orig_object->ref_count <= 1) 1299 return; 1300 VM_OBJECT_UNLOCK(orig_object); 1301 1302 offidxstart = OFF_TO_IDX(entry->offset); 1303 offidxend = offidxstart + OFF_TO_IDX(entry->end - entry->start); 1304 size = offidxend - offidxstart; 1305 1306 /* 1307 * If swap_pager_copy() is later called, it will convert new_object 1308 * into a swap object. 1309 */ 1310 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1311 1312 VM_OBJECT_LOCK(new_object); 1313 VM_OBJECT_LOCK(orig_object); 1314 source = orig_object->backing_object; 1315 if (source != NULL) { 1316 VM_OBJECT_LOCK(source); 1317 LIST_INSERT_HEAD(&source->shadow_head, 1318 new_object, shadow_list); 1319 source->shadow_count++; 1320 source->generation++; 1321 vm_object_reference_locked(source); /* for new_object */ 1322 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1323 VM_OBJECT_UNLOCK(source); 1324 new_object->backing_object_offset = 1325 orig_object->backing_object_offset + entry->offset; 1326 new_object->backing_object = source; 1327 } 1328 new_object->flags |= orig_object->flags & OBJ_NEEDGIANT; 1329 vm_page_lock_queues(); 1330 for (idx = 0; idx < size; idx++) { 1331 retry: 1332 m = vm_page_lookup(orig_object, offidxstart + idx); 1333 if (m == NULL) 1334 continue; 1335 1336 /* 1337 * We must wait for pending I/O to complete before we can 1338 * rename the page. 1339 * 1340 * We do not have to VM_PROT_NONE the page as mappings should 1341 * not be changed by this operation. 1342 */ 1343 if ((m->flags & PG_BUSY) || m->busy) { 1344 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1345 vm_page_unlock_queues(); 1346 VM_OBJECT_UNLOCK(new_object); 1347 msleep(m, VM_OBJECT_MTX(orig_object), PDROP | PVM, "spltwt", 0); 1348 VM_OBJECT_LOCK(new_object); 1349 VM_OBJECT_LOCK(orig_object); 1350 vm_page_lock_queues(); 1351 goto retry; 1352 } 1353 vm_page_rename(m, new_object, idx); 1354 /* page automatically made dirty by rename and cache handled */ 1355 vm_page_busy(m); 1356 } 1357 vm_page_unlock_queues(); 1358 if (orig_object->type == OBJT_SWAP) { 1359 /* 1360 * swap_pager_copy() can sleep, in which case the orig_object's 1361 * and new_object's locks are released and reacquired. 1362 */ 1363 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1364 } 1365 VM_OBJECT_UNLOCK(orig_object); 1366 vm_page_lock_queues(); 1367 TAILQ_FOREACH(m, &new_object->memq, listq) 1368 vm_page_wakeup(m); 1369 vm_page_unlock_queues(); 1370 VM_OBJECT_UNLOCK(new_object); 1371 entry->object.vm_object = new_object; 1372 entry->offset = 0LL; 1373 vm_object_deallocate(orig_object); 1374 VM_OBJECT_LOCK(new_object); 1375 } 1376 1377 #define OBSC_TEST_ALL_SHADOWED 0x0001 1378 #define OBSC_COLLAPSE_NOWAIT 0x0002 1379 #define OBSC_COLLAPSE_WAIT 0x0004 1380 1381 static int 1382 vm_object_backing_scan(vm_object_t object, int op) 1383 { 1384 int r = 1; 1385 vm_page_t p; 1386 vm_object_t backing_object; 1387 vm_pindex_t backing_offset_index; 1388 1389 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1390 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); 1391 1392 backing_object = object->backing_object; 1393 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1394 1395 /* 1396 * Initial conditions 1397 */ 1398 if (op & OBSC_TEST_ALL_SHADOWED) { 1399 /* 1400 * We do not want to have to test for the existence of 1401 * swap pages in the backing object. XXX but with the 1402 * new swapper this would be pretty easy to do. 1403 * 1404 * XXX what about anonymous MAP_SHARED memory that hasn't 1405 * been ZFOD faulted yet? If we do not test for this, the 1406 * shadow test may succeed! XXX 1407 */ 1408 if (backing_object->type != OBJT_DEFAULT) { 1409 return (0); 1410 } 1411 } 1412 if (op & OBSC_COLLAPSE_WAIT) { 1413 vm_object_set_flag(backing_object, OBJ_DEAD); 1414 } 1415 1416 /* 1417 * Our scan 1418 */ 1419 p = TAILQ_FIRST(&backing_object->memq); 1420 while (p) { 1421 vm_page_t next = TAILQ_NEXT(p, listq); 1422 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1423 1424 if (op & OBSC_TEST_ALL_SHADOWED) { 1425 vm_page_t pp; 1426 1427 /* 1428 * Ignore pages outside the parent object's range 1429 * and outside the parent object's mapping of the 1430 * backing object. 1431 * 1432 * note that we do not busy the backing object's 1433 * page. 1434 */ 1435 if ( 1436 p->pindex < backing_offset_index || 1437 new_pindex >= object->size 1438 ) { 1439 p = next; 1440 continue; 1441 } 1442 1443 /* 1444 * See if the parent has the page or if the parent's 1445 * object pager has the page. If the parent has the 1446 * page but the page is not valid, the parent's 1447 * object pager must have the page. 1448 * 1449 * If this fails, the parent does not completely shadow 1450 * the object and we might as well give up now. 1451 */ 1452 1453 pp = vm_page_lookup(object, new_pindex); 1454 if ( 1455 (pp == NULL || pp->valid == 0) && 1456 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1457 ) { 1458 r = 0; 1459 break; 1460 } 1461 } 1462 1463 /* 1464 * Check for busy page 1465 */ 1466 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1467 vm_page_t pp; 1468 1469 if (op & OBSC_COLLAPSE_NOWAIT) { 1470 if ((p->flags & PG_BUSY) || 1471 !p->valid || 1472 p->busy) { 1473 p = next; 1474 continue; 1475 } 1476 } else if (op & OBSC_COLLAPSE_WAIT) { 1477 if ((p->flags & PG_BUSY) || p->busy) { 1478 vm_page_lock_queues(); 1479 vm_page_flag_set(p, 1480 PG_WANTED | PG_REFERENCED); 1481 vm_page_unlock_queues(); 1482 VM_OBJECT_UNLOCK(object); 1483 msleep(p, VM_OBJECT_MTX(backing_object), 1484 PDROP | PVM, "vmocol", 0); 1485 VM_OBJECT_LOCK(object); 1486 VM_OBJECT_LOCK(backing_object); 1487 /* 1488 * If we slept, anything could have 1489 * happened. Since the object is 1490 * marked dead, the backing offset 1491 * should not have changed so we 1492 * just restart our scan. 1493 */ 1494 p = TAILQ_FIRST(&backing_object->memq); 1495 continue; 1496 } 1497 } 1498 1499 KASSERT( 1500 p->object == backing_object, 1501 ("vm_object_backing_scan: object mismatch") 1502 ); 1503 1504 /* 1505 * Destroy any associated swap 1506 */ 1507 if (backing_object->type == OBJT_SWAP) { 1508 swap_pager_freespace( 1509 backing_object, 1510 p->pindex, 1511 1 1512 ); 1513 } 1514 1515 if ( 1516 p->pindex < backing_offset_index || 1517 new_pindex >= object->size 1518 ) { 1519 /* 1520 * Page is out of the parent object's range, we 1521 * can simply destroy it. 1522 */ 1523 vm_page_lock_queues(); 1524 KASSERT(!pmap_page_is_mapped(p), 1525 ("freeing mapped page %p", p)); 1526 if (p->wire_count == 0) 1527 vm_page_free(p); 1528 else 1529 vm_page_remove(p); 1530 vm_page_unlock_queues(); 1531 p = next; 1532 continue; 1533 } 1534 1535 pp = vm_page_lookup(object, new_pindex); 1536 if ( 1537 pp != NULL || 1538 vm_pager_has_page(object, new_pindex, NULL, NULL) 1539 ) { 1540 /* 1541 * page already exists in parent OR swap exists 1542 * for this location in the parent. Destroy 1543 * the original page from the backing object. 1544 * 1545 * Leave the parent's page alone 1546 */ 1547 vm_page_lock_queues(); 1548 KASSERT(!pmap_page_is_mapped(p), 1549 ("freeing mapped page %p", p)); 1550 if (p->wire_count == 0) 1551 vm_page_free(p); 1552 else 1553 vm_page_remove(p); 1554 vm_page_unlock_queues(); 1555 p = next; 1556 continue; 1557 } 1558 1559 /* 1560 * Page does not exist in parent, rename the 1561 * page from the backing object to the main object. 1562 * 1563 * If the page was mapped to a process, it can remain 1564 * mapped through the rename. 1565 */ 1566 vm_page_lock_queues(); 1567 vm_page_rename(p, object, new_pindex); 1568 vm_page_unlock_queues(); 1569 /* page automatically made dirty by rename */ 1570 } 1571 p = next; 1572 } 1573 return (r); 1574 } 1575 1576 1577 /* 1578 * this version of collapse allows the operation to occur earlier and 1579 * when paging_in_progress is true for an object... This is not a complete 1580 * operation, but should plug 99.9% of the rest of the leaks. 1581 */ 1582 static void 1583 vm_object_qcollapse(vm_object_t object) 1584 { 1585 vm_object_t backing_object = object->backing_object; 1586 1587 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1588 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1589 1590 if (backing_object->ref_count != 1) 1591 return; 1592 1593 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1594 } 1595 1596 /* 1597 * vm_object_collapse: 1598 * 1599 * Collapse an object with the object backing it. 1600 * Pages in the backing object are moved into the 1601 * parent, and the backing object is deallocated. 1602 */ 1603 void 1604 vm_object_collapse(vm_object_t object) 1605 { 1606 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1607 1608 while (TRUE) { 1609 vm_object_t backing_object; 1610 1611 /* 1612 * Verify that the conditions are right for collapse: 1613 * 1614 * The object exists and the backing object exists. 1615 */ 1616 if ((backing_object = object->backing_object) == NULL) 1617 break; 1618 1619 /* 1620 * we check the backing object first, because it is most likely 1621 * not collapsable. 1622 */ 1623 VM_OBJECT_LOCK(backing_object); 1624 if (backing_object->handle != NULL || 1625 (backing_object->type != OBJT_DEFAULT && 1626 backing_object->type != OBJT_SWAP) || 1627 (backing_object->flags & OBJ_DEAD) || 1628 object->handle != NULL || 1629 (object->type != OBJT_DEFAULT && 1630 object->type != OBJT_SWAP) || 1631 (object->flags & OBJ_DEAD)) { 1632 VM_OBJECT_UNLOCK(backing_object); 1633 break; 1634 } 1635 1636 if ( 1637 object->paging_in_progress != 0 || 1638 backing_object->paging_in_progress != 0 1639 ) { 1640 vm_object_qcollapse(object); 1641 VM_OBJECT_UNLOCK(backing_object); 1642 break; 1643 } 1644 /* 1645 * We know that we can either collapse the backing object (if 1646 * the parent is the only reference to it) or (perhaps) have 1647 * the parent bypass the object if the parent happens to shadow 1648 * all the resident pages in the entire backing object. 1649 * 1650 * This is ignoring pager-backed pages such as swap pages. 1651 * vm_object_backing_scan fails the shadowing test in this 1652 * case. 1653 */ 1654 if (backing_object->ref_count == 1) { 1655 /* 1656 * If there is exactly one reference to the backing 1657 * object, we can collapse it into the parent. 1658 */ 1659 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1660 1661 /* 1662 * Move the pager from backing_object to object. 1663 */ 1664 if (backing_object->type == OBJT_SWAP) { 1665 /* 1666 * swap_pager_copy() can sleep, in which case 1667 * the backing_object's and object's locks are 1668 * released and reacquired. 1669 */ 1670 swap_pager_copy( 1671 backing_object, 1672 object, 1673 OFF_TO_IDX(object->backing_object_offset), TRUE); 1674 } 1675 /* 1676 * Object now shadows whatever backing_object did. 1677 * Note that the reference to 1678 * backing_object->backing_object moves from within 1679 * backing_object to within object. 1680 */ 1681 LIST_REMOVE(object, shadow_list); 1682 backing_object->shadow_count--; 1683 backing_object->generation++; 1684 if (backing_object->backing_object) { 1685 VM_OBJECT_LOCK(backing_object->backing_object); 1686 LIST_REMOVE(backing_object, shadow_list); 1687 LIST_INSERT_HEAD( 1688 &backing_object->backing_object->shadow_head, 1689 object, shadow_list); 1690 /* 1691 * The shadow_count has not changed. 1692 */ 1693 backing_object->backing_object->generation++; 1694 VM_OBJECT_UNLOCK(backing_object->backing_object); 1695 } 1696 object->backing_object = backing_object->backing_object; 1697 object->backing_object_offset += 1698 backing_object->backing_object_offset; 1699 1700 /* 1701 * Discard backing_object. 1702 * 1703 * Since the backing object has no pages, no pager left, 1704 * and no object references within it, all that is 1705 * necessary is to dispose of it. 1706 */ 1707 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1708 VM_OBJECT_UNLOCK(backing_object); 1709 1710 mtx_lock(&vm_object_list_mtx); 1711 TAILQ_REMOVE( 1712 &vm_object_list, 1713 backing_object, 1714 object_list 1715 ); 1716 mtx_unlock(&vm_object_list_mtx); 1717 1718 uma_zfree(obj_zone, backing_object); 1719 1720 object_collapses++; 1721 } else { 1722 vm_object_t new_backing_object; 1723 1724 /* 1725 * If we do not entirely shadow the backing object, 1726 * there is nothing we can do so we give up. 1727 */ 1728 if (object->resident_page_count != object->size && 1729 vm_object_backing_scan(object, 1730 OBSC_TEST_ALL_SHADOWED) == 0) { 1731 VM_OBJECT_UNLOCK(backing_object); 1732 break; 1733 } 1734 1735 /* 1736 * Make the parent shadow the next object in the 1737 * chain. Deallocating backing_object will not remove 1738 * it, since its reference count is at least 2. 1739 */ 1740 LIST_REMOVE(object, shadow_list); 1741 backing_object->shadow_count--; 1742 backing_object->generation++; 1743 1744 new_backing_object = backing_object->backing_object; 1745 if ((object->backing_object = new_backing_object) != NULL) { 1746 VM_OBJECT_LOCK(new_backing_object); 1747 LIST_INSERT_HEAD( 1748 &new_backing_object->shadow_head, 1749 object, 1750 shadow_list 1751 ); 1752 new_backing_object->shadow_count++; 1753 new_backing_object->generation++; 1754 vm_object_reference_locked(new_backing_object); 1755 VM_OBJECT_UNLOCK(new_backing_object); 1756 object->backing_object_offset += 1757 backing_object->backing_object_offset; 1758 } 1759 1760 /* 1761 * Drop the reference count on backing_object. Since 1762 * its ref_count was at least 2, it will not vanish. 1763 */ 1764 backing_object->ref_count--; 1765 VM_OBJECT_UNLOCK(backing_object); 1766 object_bypasses++; 1767 } 1768 1769 /* 1770 * Try again with this object's new backing object. 1771 */ 1772 } 1773 } 1774 1775 /* 1776 * vm_object_page_remove: 1777 * 1778 * Removes all physical pages in the given range from the 1779 * object's list of pages. If the range's end is zero, all 1780 * physical pages from the range's start to the end of the object 1781 * are deleted. 1782 * 1783 * The object must be locked. 1784 */ 1785 void 1786 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1787 boolean_t clean_only) 1788 { 1789 vm_page_t p, next; 1790 1791 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1792 if (object->resident_page_count == 0) 1793 return; 1794 1795 /* 1796 * Since physically-backed objects do not use managed pages, we can't 1797 * remove pages from the object (we must instead remove the page 1798 * references, and then destroy the object). 1799 */ 1800 KASSERT(object->type != OBJT_PHYS, 1801 ("attempt to remove pages from a physical object")); 1802 1803 vm_object_pip_add(object, 1); 1804 again: 1805 vm_page_lock_queues(); 1806 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1807 if (p->pindex < start) { 1808 p = vm_page_splay(start, object->root); 1809 if ((object->root = p)->pindex < start) 1810 p = TAILQ_NEXT(p, listq); 1811 } 1812 } 1813 /* 1814 * Assert: the variable p is either (1) the page with the 1815 * least pindex greater than or equal to the parameter pindex 1816 * or (2) NULL. 1817 */ 1818 for (; 1819 p != NULL && (p->pindex < end || end == 0); 1820 p = next) { 1821 next = TAILQ_NEXT(p, listq); 1822 1823 if (p->wire_count != 0) { 1824 pmap_remove_all(p); 1825 if (!clean_only) 1826 p->valid = 0; 1827 continue; 1828 } 1829 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1830 goto again; 1831 if (clean_only && p->valid) { 1832 pmap_remove_write(p); 1833 if (p->valid & p->dirty) 1834 continue; 1835 } 1836 pmap_remove_all(p); 1837 vm_page_free(p); 1838 } 1839 vm_page_unlock_queues(); 1840 vm_object_pip_wakeup(object); 1841 } 1842 1843 /* 1844 * Routine: vm_object_coalesce 1845 * Function: Coalesces two objects backing up adjoining 1846 * regions of memory into a single object. 1847 * 1848 * returns TRUE if objects were combined. 1849 * 1850 * NOTE: Only works at the moment if the second object is NULL - 1851 * if it's not, which object do we lock first? 1852 * 1853 * Parameters: 1854 * prev_object First object to coalesce 1855 * prev_offset Offset into prev_object 1856 * prev_size Size of reference to prev_object 1857 * next_size Size of reference to the second object 1858 * 1859 * Conditions: 1860 * The object must *not* be locked. 1861 */ 1862 boolean_t 1863 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 1864 vm_size_t prev_size, vm_size_t next_size) 1865 { 1866 vm_pindex_t next_pindex; 1867 1868 if (prev_object == NULL) 1869 return (TRUE); 1870 VM_OBJECT_LOCK(prev_object); 1871 if (prev_object->type != OBJT_DEFAULT && 1872 prev_object->type != OBJT_SWAP) { 1873 VM_OBJECT_UNLOCK(prev_object); 1874 return (FALSE); 1875 } 1876 1877 /* 1878 * Try to collapse the object first 1879 */ 1880 vm_object_collapse(prev_object); 1881 1882 /* 1883 * Can't coalesce if: . more than one reference . paged out . shadows 1884 * another object . has a copy elsewhere (any of which mean that the 1885 * pages not mapped to prev_entry may be in use anyway) 1886 */ 1887 if (prev_object->backing_object != NULL) { 1888 VM_OBJECT_UNLOCK(prev_object); 1889 return (FALSE); 1890 } 1891 1892 prev_size >>= PAGE_SHIFT; 1893 next_size >>= PAGE_SHIFT; 1894 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 1895 1896 if ((prev_object->ref_count > 1) && 1897 (prev_object->size != next_pindex)) { 1898 VM_OBJECT_UNLOCK(prev_object); 1899 return (FALSE); 1900 } 1901 1902 /* 1903 * Remove any pages that may still be in the object from a previous 1904 * deallocation. 1905 */ 1906 if (next_pindex < prev_object->size) { 1907 vm_object_page_remove(prev_object, 1908 next_pindex, 1909 next_pindex + next_size, FALSE); 1910 if (prev_object->type == OBJT_SWAP) 1911 swap_pager_freespace(prev_object, 1912 next_pindex, next_size); 1913 } 1914 1915 /* 1916 * Extend the object if necessary. 1917 */ 1918 if (next_pindex + next_size > prev_object->size) 1919 prev_object->size = next_pindex + next_size; 1920 1921 VM_OBJECT_UNLOCK(prev_object); 1922 return (TRUE); 1923 } 1924 1925 void 1926 vm_object_set_writeable_dirty(vm_object_t object) 1927 { 1928 struct vnode *vp; 1929 1930 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1931 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 1932 return; 1933 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 1934 if (object->type == OBJT_VNODE && 1935 (vp = (struct vnode *)object->handle) != NULL) { 1936 VI_LOCK(vp); 1937 vp->v_iflag |= VI_OBJDIRTY; 1938 VI_UNLOCK(vp); 1939 } 1940 } 1941 1942 #include "opt_ddb.h" 1943 #ifdef DDB 1944 #include <sys/kernel.h> 1945 1946 #include <sys/cons.h> 1947 1948 #include <ddb/ddb.h> 1949 1950 static int 1951 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1952 { 1953 vm_map_t tmpm; 1954 vm_map_entry_t tmpe; 1955 vm_object_t obj; 1956 int entcount; 1957 1958 if (map == 0) 1959 return 0; 1960 1961 if (entry == 0) { 1962 tmpe = map->header.next; 1963 entcount = map->nentries; 1964 while (entcount-- && (tmpe != &map->header)) { 1965 if (_vm_object_in_map(map, object, tmpe)) { 1966 return 1; 1967 } 1968 tmpe = tmpe->next; 1969 } 1970 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1971 tmpm = entry->object.sub_map; 1972 tmpe = tmpm->header.next; 1973 entcount = tmpm->nentries; 1974 while (entcount-- && tmpe != &tmpm->header) { 1975 if (_vm_object_in_map(tmpm, object, tmpe)) { 1976 return 1; 1977 } 1978 tmpe = tmpe->next; 1979 } 1980 } else if ((obj = entry->object.vm_object) != NULL) { 1981 for (; obj; obj = obj->backing_object) 1982 if (obj == object) { 1983 return 1; 1984 } 1985 } 1986 return 0; 1987 } 1988 1989 static int 1990 vm_object_in_map(vm_object_t object) 1991 { 1992 struct proc *p; 1993 1994 /* sx_slock(&allproc_lock); */ 1995 LIST_FOREACH(p, &allproc, p_list) { 1996 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1997 continue; 1998 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1999 /* sx_sunlock(&allproc_lock); */ 2000 return 1; 2001 } 2002 } 2003 /* sx_sunlock(&allproc_lock); */ 2004 if (_vm_object_in_map(kernel_map, object, 0)) 2005 return 1; 2006 if (_vm_object_in_map(kmem_map, object, 0)) 2007 return 1; 2008 if (_vm_object_in_map(pager_map, object, 0)) 2009 return 1; 2010 if (_vm_object_in_map(buffer_map, object, 0)) 2011 return 1; 2012 return 0; 2013 } 2014 2015 DB_SHOW_COMMAND(vmochk, vm_object_check) 2016 { 2017 vm_object_t object; 2018 2019 /* 2020 * make sure that internal objs are in a map somewhere 2021 * and none have zero ref counts. 2022 */ 2023 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2024 if (object->handle == NULL && 2025 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2026 if (object->ref_count == 0) { 2027 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2028 (long)object->size); 2029 } 2030 if (!vm_object_in_map(object)) { 2031 db_printf( 2032 "vmochk: internal obj is not in a map: " 2033 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2034 object->ref_count, (u_long)object->size, 2035 (u_long)object->size, 2036 (void *)object->backing_object); 2037 } 2038 } 2039 } 2040 } 2041 2042 /* 2043 * vm_object_print: [ debug ] 2044 */ 2045 DB_SHOW_COMMAND(object, vm_object_print_static) 2046 { 2047 /* XXX convert args. */ 2048 vm_object_t object = (vm_object_t)addr; 2049 boolean_t full = have_addr; 2050 2051 vm_page_t p; 2052 2053 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2054 #define count was_count 2055 2056 int count; 2057 2058 if (object == NULL) 2059 return; 2060 2061 db_iprintf( 2062 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 2063 object, (int)object->type, (uintmax_t)object->size, 2064 object->resident_page_count, object->ref_count, object->flags); 2065 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2066 object->shadow_count, 2067 object->backing_object ? object->backing_object->ref_count : 0, 2068 object->backing_object, (uintmax_t)object->backing_object_offset); 2069 2070 if (!full) 2071 return; 2072 2073 db_indent += 2; 2074 count = 0; 2075 TAILQ_FOREACH(p, &object->memq, listq) { 2076 if (count == 0) 2077 db_iprintf("memory:="); 2078 else if (count == 6) { 2079 db_printf("\n"); 2080 db_iprintf(" ..."); 2081 count = 0; 2082 } else 2083 db_printf(","); 2084 count++; 2085 2086 db_printf("(off=0x%jx,page=0x%jx)", 2087 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2088 } 2089 if (count != 0) 2090 db_printf("\n"); 2091 db_indent -= 2; 2092 } 2093 2094 /* XXX. */ 2095 #undef count 2096 2097 /* XXX need this non-static entry for calling from vm_map_print. */ 2098 void 2099 vm_object_print( 2100 /* db_expr_t */ long addr, 2101 boolean_t have_addr, 2102 /* db_expr_t */ long count, 2103 char *modif) 2104 { 2105 vm_object_print_static(addr, have_addr, count, modif); 2106 } 2107 2108 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2109 { 2110 vm_object_t object; 2111 int nl = 0; 2112 int c; 2113 2114 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2115 vm_pindex_t idx, fidx; 2116 vm_pindex_t osize; 2117 vm_paddr_t pa = -1, padiff; 2118 int rcount; 2119 vm_page_t m; 2120 2121 db_printf("new object: %p\n", (void *)object); 2122 if (nl > 18) { 2123 c = cngetc(); 2124 if (c != ' ') 2125 return; 2126 nl = 0; 2127 } 2128 nl++; 2129 rcount = 0; 2130 fidx = 0; 2131 osize = object->size; 2132 if (osize > 128) 2133 osize = 128; 2134 for (idx = 0; idx < osize; idx++) { 2135 m = vm_page_lookup(object, idx); 2136 if (m == NULL) { 2137 if (rcount) { 2138 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2139 (long)fidx, rcount, (long)pa); 2140 if (nl > 18) { 2141 c = cngetc(); 2142 if (c != ' ') 2143 return; 2144 nl = 0; 2145 } 2146 nl++; 2147 rcount = 0; 2148 } 2149 continue; 2150 } 2151 2152 2153 if (rcount && 2154 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2155 ++rcount; 2156 continue; 2157 } 2158 if (rcount) { 2159 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2160 padiff >>= PAGE_SHIFT; 2161 padiff &= PQ_COLORMASK; 2162 if (padiff == 0) { 2163 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2164 ++rcount; 2165 continue; 2166 } 2167 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2168 (long)fidx, rcount, (long)pa); 2169 db_printf("pd(%ld)\n", (long)padiff); 2170 if (nl > 18) { 2171 c = cngetc(); 2172 if (c != ' ') 2173 return; 2174 nl = 0; 2175 } 2176 nl++; 2177 } 2178 fidx = idx; 2179 pa = VM_PAGE_TO_PHYS(m); 2180 rcount = 1; 2181 } 2182 if (rcount) { 2183 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2184 (long)fidx, rcount, (long)pa); 2185 if (nl > 18) { 2186 c = cngetc(); 2187 if (c != ' ') 2188 return; 2189 nl = 0; 2190 } 2191 nl++; 2192 } 2193 } 2194 } 2195 #endif /* DDB */ 2196