1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD$ 65 */ 66 67 /* 68 * Virtual memory object module. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/lock.h> 74 #include <sys/mman.h> 75 #include <sys/mount.h> 76 #include <sys/kernel.h> 77 #include <sys/sysctl.h> 78 #include <sys/mutex.h> 79 #include <sys/proc.h> /* for curproc, pageproc */ 80 #include <sys/socket.h> 81 #include <sys/vnode.h> 82 #include <sys/vmmeter.h> 83 #include <sys/sx.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/uma.h> 97 98 #define EASY_SCAN_FACTOR 8 99 100 #define MSYNC_FLUSH_HARDSEQ 0x01 101 #define MSYNC_FLUSH_SOFTSEQ 0x02 102 103 /* 104 * msync / VM object flushing optimizations 105 */ 106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 108 CTLFLAG_RW, &msync_flush_flags, 0, ""); 109 110 static void vm_object_qcollapse(vm_object_t object); 111 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 112 static void vm_object_pip_sleep(vm_object_t object, char *waitid); 113 114 /* 115 * Virtual memory objects maintain the actual data 116 * associated with allocated virtual memory. A given 117 * page of memory exists within exactly one object. 118 * 119 * An object is only deallocated when all "references" 120 * are given up. Only one "reference" to a given 121 * region of an object should be writeable. 122 * 123 * Associated with each object is a list of all resident 124 * memory pages belonging to that object; this list is 125 * maintained by the "vm_page" module, and locked by the object's 126 * lock. 127 * 128 * Each object also records a "pager" routine which is 129 * used to retrieve (and store) pages to the proper backing 130 * storage. In addition, objects may be backed by other 131 * objects from which they were virtual-copied. 132 * 133 * The only items within the object structure which are 134 * modified after time of creation are: 135 * reference count locked by object's lock 136 * pager routine locked by object's lock 137 * 138 */ 139 140 struct object_q vm_object_list; 141 struct mtx vm_object_list_mtx; /* lock for object list and count */ 142 vm_object_t kernel_object; 143 vm_object_t kmem_object; 144 static struct vm_object kernel_object_store; 145 static struct vm_object kmem_object_store; 146 147 static long object_collapses; 148 static long object_bypasses; 149 static int next_index; 150 static uma_zone_t obj_zone; 151 #define VM_OBJECTS_INIT 256 152 153 static void vm_object_zinit(void *mem, int size); 154 155 #ifdef INVARIANTS 156 static void vm_object_zdtor(void *mem, int size, void *arg); 157 158 static void 159 vm_object_zdtor(void *mem, int size, void *arg) 160 { 161 vm_object_t object; 162 163 object = (vm_object_t)mem; 164 KASSERT(object->paging_in_progress == 0, 165 ("object %p paging_in_progress = %d", 166 object, object->paging_in_progress)); 167 KASSERT(object->resident_page_count == 0, 168 ("object %p resident_page_count = %d", 169 object, object->resident_page_count)); 170 KASSERT(object->shadow_count == 0, 171 ("object %p shadow_count = %d", 172 object, object->shadow_count)); 173 } 174 #endif 175 176 static void 177 vm_object_zinit(void *mem, int size) 178 { 179 vm_object_t object; 180 181 object = (vm_object_t)mem; 182 bzero(&object->mtx, sizeof(object->mtx)); 183 VM_OBJECT_LOCK_INIT(object); 184 185 /* These are true for any object that has been freed */ 186 object->paging_in_progress = 0; 187 object->resident_page_count = 0; 188 object->shadow_count = 0; 189 } 190 191 void 192 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 193 { 194 int incr; 195 196 TAILQ_INIT(&object->memq); 197 TAILQ_INIT(&object->shadow_head); 198 199 object->root = NULL; 200 object->type = type; 201 object->size = size; 202 object->ref_count = 1; 203 object->flags = 0; 204 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 205 vm_object_set_flag(object, OBJ_ONEMAPPING); 206 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 207 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 208 else 209 incr = size; 210 do 211 object->pg_color = next_index; 212 while (!atomic_cmpset_int(&next_index, object->pg_color, 213 (object->pg_color + incr) & PQ_L2_MASK)); 214 object->handle = NULL; 215 object->backing_object = NULL; 216 object->backing_object_offset = (vm_ooffset_t) 0; 217 218 atomic_add_int(&object->generation, 1); 219 220 mtx_lock(&vm_object_list_mtx); 221 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 222 mtx_unlock(&vm_object_list_mtx); 223 } 224 225 /* 226 * vm_object_init: 227 * 228 * Initialize the VM objects module. 229 */ 230 void 231 vm_object_init(void) 232 { 233 TAILQ_INIT(&vm_object_list); 234 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 235 236 kernel_object = &kernel_object_store; 237 VM_OBJECT_LOCK_INIT(&kernel_object_store); 238 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 239 kernel_object); 240 241 kmem_object = &kmem_object_store; 242 VM_OBJECT_LOCK_INIT(&kmem_object_store); 243 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 244 kmem_object); 245 246 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 247 #ifdef INVARIANTS 248 vm_object_zdtor, 249 #else 250 NULL, 251 #endif 252 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 253 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 254 } 255 256 void 257 vm_object_set_flag(vm_object_t object, u_short bits) 258 { 259 object->flags |= bits; 260 } 261 262 void 263 vm_object_clear_flag(vm_object_t object, u_short bits) 264 { 265 266 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 267 object->flags &= ~bits; 268 } 269 270 void 271 vm_object_pip_add(vm_object_t object, short i) 272 { 273 274 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 275 object->paging_in_progress += i; 276 } 277 278 void 279 vm_object_pip_subtract(vm_object_t object, short i) 280 { 281 282 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 283 object->paging_in_progress -= i; 284 } 285 286 void 287 vm_object_pip_wakeup(vm_object_t object) 288 { 289 290 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 291 object->paging_in_progress--; 292 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 293 vm_object_clear_flag(object, OBJ_PIPWNT); 294 wakeup(object); 295 } 296 } 297 298 void 299 vm_object_pip_wakeupn(vm_object_t object, short i) 300 { 301 302 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 303 if (i) 304 object->paging_in_progress -= i; 305 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 306 vm_object_clear_flag(object, OBJ_PIPWNT); 307 wakeup(object); 308 } 309 } 310 311 static void 312 vm_object_pip_sleep(vm_object_t object, char *waitid) 313 { 314 GIANT_REQUIRED; 315 if (object->paging_in_progress) { 316 int s = splvm(); 317 if (object->paging_in_progress) { 318 vm_object_set_flag(object, OBJ_PIPWNT); 319 tsleep(object, PVM, waitid, 0); 320 } 321 splx(s); 322 } 323 } 324 325 void 326 vm_object_pip_wait(vm_object_t object, char *waitid) 327 { 328 329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 330 while (object->paging_in_progress) { 331 object->flags |= OBJ_PIPWNT; 332 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 333 } 334 } 335 336 /* 337 * vm_object_allocate_wait 338 * 339 * Return a new object with the given size, and give the user the 340 * option of waiting for it to complete or failing if the needed 341 * memory isn't available. 342 */ 343 vm_object_t 344 vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 345 { 346 vm_object_t result; 347 348 result = (vm_object_t) uma_zalloc(obj_zone, flags); 349 350 if (result != NULL) 351 _vm_object_allocate(type, size, result); 352 353 return (result); 354 } 355 356 /* 357 * vm_object_allocate: 358 * 359 * Returns a new object with the given size. 360 */ 361 vm_object_t 362 vm_object_allocate(objtype_t type, vm_pindex_t size) 363 { 364 return(vm_object_allocate_wait(type, size, M_WAITOK)); 365 } 366 367 368 /* 369 * vm_object_reference: 370 * 371 * Gets another reference to the given object. Note: OBJ_DEAD 372 * objects can be referenced during final cleaning. 373 */ 374 void 375 vm_object_reference(vm_object_t object) 376 { 377 if (object == NULL) 378 return; 379 if (object != kmem_object) 380 mtx_lock(&Giant); 381 VM_OBJECT_LOCK(object); 382 object->ref_count++; 383 VM_OBJECT_UNLOCK(object); 384 if (object->type == OBJT_VNODE) { 385 while (vget((struct vnode *) object->handle, LK_RETRY, curthread)) { 386 printf("vm_object_reference: delay in getting object\n"); 387 } 388 } 389 if (object != kmem_object) 390 mtx_unlock(&Giant); 391 } 392 393 /* 394 * Handle deallocating an object of type OBJT_VNODE. 395 */ 396 void 397 vm_object_vndeallocate(vm_object_t object) 398 { 399 struct vnode *vp = (struct vnode *) object->handle; 400 401 GIANT_REQUIRED; 402 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 403 KASSERT(object->type == OBJT_VNODE, 404 ("vm_object_vndeallocate: not a vnode object")); 405 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 406 #ifdef INVARIANTS 407 if (object->ref_count == 0) { 408 vprint("vm_object_vndeallocate", vp); 409 panic("vm_object_vndeallocate: bad object reference count"); 410 } 411 #endif 412 413 object->ref_count--; 414 if (object->ref_count == 0) { 415 mp_fixme("Unlocked vflag access."); 416 vp->v_vflag &= ~VV_TEXT; 417 } 418 VM_OBJECT_UNLOCK(object); 419 /* 420 * vrele may need a vop lock 421 */ 422 vrele(vp); 423 } 424 425 /* 426 * vm_object_deallocate: 427 * 428 * Release a reference to the specified object, 429 * gained either through a vm_object_allocate 430 * or a vm_object_reference call. When all references 431 * are gone, storage associated with this object 432 * may be relinquished. 433 * 434 * No object may be locked. 435 */ 436 void 437 vm_object_deallocate(vm_object_t object) 438 { 439 vm_object_t temp; 440 441 vm_object_lock(object); 442 while (object != NULL) { 443 444 if (object->type == OBJT_VNODE) { 445 VM_OBJECT_LOCK(object); 446 vm_object_vndeallocate(object); 447 mtx_unlock(&Giant); 448 return; 449 } 450 451 KASSERT(object->ref_count != 0, 452 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 453 454 /* 455 * If the reference count goes to 0 we start calling 456 * vm_object_terminate() on the object chain. 457 * A ref count of 1 may be a special case depending on the 458 * shadow count being 0 or 1. 459 */ 460 object->ref_count--; 461 if (object->ref_count > 1) { 462 vm_object_unlock(object); 463 return; 464 } else if (object->ref_count == 1) { 465 if (object->shadow_count == 0) { 466 vm_object_set_flag(object, OBJ_ONEMAPPING); 467 } else if ((object->shadow_count == 1) && 468 (object->handle == NULL) && 469 (object->type == OBJT_DEFAULT || 470 object->type == OBJT_SWAP)) { 471 vm_object_t robject; 472 473 robject = TAILQ_FIRST(&object->shadow_head); 474 KASSERT(robject != NULL, 475 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 476 object->ref_count, 477 object->shadow_count)); 478 if ((robject->handle == NULL) && 479 (robject->type == OBJT_DEFAULT || 480 robject->type == OBJT_SWAP)) { 481 482 robject->ref_count++; 483 484 while ( 485 robject->paging_in_progress || 486 object->paging_in_progress 487 ) { 488 vm_object_pip_sleep(robject, "objde1"); 489 vm_object_pip_sleep(object, "objde2"); 490 } 491 492 if (robject->ref_count == 1) { 493 robject->ref_count--; 494 object = robject; 495 goto doterm; 496 } 497 498 object = robject; 499 vm_object_collapse(object); 500 continue; 501 } 502 } 503 vm_object_unlock(object); 504 return; 505 } 506 doterm: 507 VM_OBJECT_LOCK(object); 508 temp = object->backing_object; 509 if (temp != NULL) { 510 VM_OBJECT_LOCK(temp); 511 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 512 temp->shadow_count--; 513 temp->generation++; 514 VM_OBJECT_UNLOCK(temp); 515 object->backing_object = NULL; 516 } 517 /* 518 * Don't double-terminate, we could be in a termination 519 * recursion due to the terminate having to sync data 520 * to disk. 521 */ 522 if ((object->flags & OBJ_DEAD) == 0) 523 vm_object_terminate(object); 524 else 525 VM_OBJECT_UNLOCK(object); 526 object = temp; 527 } 528 vm_object_unlock(object); 529 } 530 531 /* 532 * vm_object_terminate actually destroys the specified object, freeing 533 * up all previously used resources. 534 * 535 * The object must be locked. 536 * This routine may block. 537 */ 538 void 539 vm_object_terminate(vm_object_t object) 540 { 541 vm_page_t p; 542 int s; 543 544 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 545 546 /* 547 * Make sure no one uses us. 548 */ 549 vm_object_set_flag(object, OBJ_DEAD); 550 551 /* 552 * wait for the pageout daemon to be done with the object 553 */ 554 vm_object_pip_wait(object, "objtrm"); 555 556 KASSERT(!object->paging_in_progress, 557 ("vm_object_terminate: pageout in progress")); 558 559 /* 560 * Clean and free the pages, as appropriate. All references to the 561 * object are gone, so we don't need to lock it. 562 */ 563 if (object->type == OBJT_VNODE) { 564 struct vnode *vp = (struct vnode *)object->handle; 565 566 /* 567 * Clean pages and flush buffers. 568 */ 569 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 570 VM_OBJECT_UNLOCK(object); 571 572 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 573 574 VM_OBJECT_LOCK(object); 575 } 576 577 KASSERT(object->ref_count == 0, 578 ("vm_object_terminate: object with references, ref_count=%d", 579 object->ref_count)); 580 581 /* 582 * Now free any remaining pages. For internal objects, this also 583 * removes them from paging queues. Don't free wired pages, just 584 * remove them from the object. 585 */ 586 s = splvm(); 587 vm_page_lock_queues(); 588 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 589 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 590 ("vm_object_terminate: freeing busy page %p " 591 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 592 if (p->wire_count == 0) { 593 vm_page_busy(p); 594 vm_page_free(p); 595 cnt.v_pfree++; 596 } else { 597 vm_page_busy(p); 598 vm_page_remove(p); 599 } 600 } 601 vm_page_unlock_queues(); 602 splx(s); 603 604 /* 605 * Let the pager know object is dead. 606 */ 607 vm_pager_deallocate(object); 608 VM_OBJECT_UNLOCK(object); 609 610 /* 611 * Remove the object from the global object list. 612 */ 613 mtx_lock(&vm_object_list_mtx); 614 TAILQ_REMOVE(&vm_object_list, object, object_list); 615 mtx_unlock(&vm_object_list_mtx); 616 617 wakeup(object); 618 619 /* 620 * Free the space for the object. 621 */ 622 uma_zfree(obj_zone, object); 623 } 624 625 /* 626 * vm_object_page_clean 627 * 628 * Clean all dirty pages in the specified range of object. Leaves page 629 * on whatever queue it is currently on. If NOSYNC is set then do not 630 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 631 * leaving the object dirty. 632 * 633 * When stuffing pages asynchronously, allow clustering. XXX we need a 634 * synchronous clustering mode implementation. 635 * 636 * Odd semantics: if start == end, we clean everything. 637 * 638 * The object must be locked. 639 */ 640 void 641 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 642 { 643 vm_page_t p, np; 644 vm_pindex_t tstart, tend; 645 vm_pindex_t pi; 646 struct vnode *vp; 647 int clearobjflags; 648 int pagerflags; 649 int curgeneration; 650 651 GIANT_REQUIRED; 652 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 653 if (object->type != OBJT_VNODE || 654 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 655 return; 656 657 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 658 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 659 660 vp = object->handle; 661 662 vm_object_set_flag(object, OBJ_CLEANING); 663 664 tstart = start; 665 if (end == 0) { 666 tend = object->size; 667 } else { 668 tend = end; 669 } 670 671 vm_page_lock_queues(); 672 /* 673 * If the caller is smart and only msync()s a range he knows is 674 * dirty, we may be able to avoid an object scan. This results in 675 * a phenominal improvement in performance. We cannot do this 676 * as a matter of course because the object may be huge - e.g. 677 * the size might be in the gigabytes or terrabytes. 678 */ 679 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 680 vm_pindex_t tscan; 681 int scanlimit; 682 int scanreset; 683 684 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 685 if (scanreset < 16) 686 scanreset = 16; 687 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 688 689 scanlimit = scanreset; 690 tscan = tstart; 691 while (tscan < tend) { 692 curgeneration = object->generation; 693 p = vm_page_lookup(object, tscan); 694 if (p == NULL || p->valid == 0 || 695 (p->queue - p->pc) == PQ_CACHE) { 696 if (--scanlimit == 0) 697 break; 698 ++tscan; 699 continue; 700 } 701 vm_page_test_dirty(p); 702 if ((p->dirty & p->valid) == 0) { 703 if (--scanlimit == 0) 704 break; 705 ++tscan; 706 continue; 707 } 708 /* 709 * If we have been asked to skip nosync pages and 710 * this is a nosync page, we can't continue. 711 */ 712 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 713 if (--scanlimit == 0) 714 break; 715 ++tscan; 716 continue; 717 } 718 scanlimit = scanreset; 719 720 /* 721 * This returns 0 if it was unable to busy the first 722 * page (i.e. had to sleep). 723 */ 724 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 725 } 726 727 /* 728 * If everything was dirty and we flushed it successfully, 729 * and the requested range is not the entire object, we 730 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 731 * return immediately. 732 */ 733 if (tscan >= tend && (tstart || tend < object->size)) { 734 vm_page_unlock_queues(); 735 vm_object_clear_flag(object, OBJ_CLEANING); 736 return; 737 } 738 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 739 } 740 741 /* 742 * Generally set CLEANCHK interlock and make the page read-only so 743 * we can then clear the object flags. 744 * 745 * However, if this is a nosync mmap then the object is likely to 746 * stay dirty so do not mess with the page and do not clear the 747 * object flags. 748 */ 749 clearobjflags = 1; 750 TAILQ_FOREACH(p, &object->memq, listq) { 751 vm_page_flag_set(p, PG_CLEANCHK); 752 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 753 clearobjflags = 0; 754 else 755 pmap_page_protect(p, VM_PROT_READ); 756 } 757 758 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 759 struct vnode *vp; 760 761 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 762 if (object->type == OBJT_VNODE && 763 (vp = (struct vnode *)object->handle) != NULL) { 764 VI_LOCK(vp); 765 if (vp->v_iflag & VI_OBJDIRTY) 766 vp->v_iflag &= ~VI_OBJDIRTY; 767 VI_UNLOCK(vp); 768 } 769 } 770 771 rescan: 772 curgeneration = object->generation; 773 774 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 775 int n; 776 777 np = TAILQ_NEXT(p, listq); 778 779 again: 780 pi = p->pindex; 781 if (((p->flags & PG_CLEANCHK) == 0) || 782 (pi < tstart) || (pi >= tend) || 783 (p->valid == 0) || 784 ((p->queue - p->pc) == PQ_CACHE)) { 785 vm_page_flag_clear(p, PG_CLEANCHK); 786 continue; 787 } 788 789 vm_page_test_dirty(p); 790 if ((p->dirty & p->valid) == 0) { 791 vm_page_flag_clear(p, PG_CLEANCHK); 792 continue; 793 } 794 795 /* 796 * If we have been asked to skip nosync pages and this is a 797 * nosync page, skip it. Note that the object flags were 798 * not cleared in this case so we do not have to set them. 799 */ 800 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 801 vm_page_flag_clear(p, PG_CLEANCHK); 802 continue; 803 } 804 805 n = vm_object_page_collect_flush(object, p, 806 curgeneration, pagerflags); 807 if (n == 0) 808 goto rescan; 809 810 if (object->generation != curgeneration) 811 goto rescan; 812 813 /* 814 * Try to optimize the next page. If we can't we pick up 815 * our (random) scan where we left off. 816 */ 817 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 818 if ((p = vm_page_lookup(object, pi + n)) != NULL) 819 goto again; 820 } 821 } 822 vm_page_unlock_queues(); 823 #if 0 824 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 825 #endif 826 827 vm_object_clear_flag(object, OBJ_CLEANING); 828 return; 829 } 830 831 static int 832 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 833 { 834 int runlen; 835 int s; 836 int maxf; 837 int chkb; 838 int maxb; 839 int i; 840 vm_pindex_t pi; 841 vm_page_t maf[vm_pageout_page_count]; 842 vm_page_t mab[vm_pageout_page_count]; 843 vm_page_t ma[vm_pageout_page_count]; 844 845 s = splvm(); 846 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 847 pi = p->pindex; 848 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 849 vm_page_lock_queues(); 850 if (object->generation != curgeneration) { 851 splx(s); 852 return(0); 853 } 854 } 855 maxf = 0; 856 for(i = 1; i < vm_pageout_page_count; i++) { 857 vm_page_t tp; 858 859 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 860 if ((tp->flags & PG_BUSY) || 861 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 862 (tp->flags & PG_CLEANCHK) == 0) || 863 (tp->busy != 0)) 864 break; 865 if((tp->queue - tp->pc) == PQ_CACHE) { 866 vm_page_flag_clear(tp, PG_CLEANCHK); 867 break; 868 } 869 vm_page_test_dirty(tp); 870 if ((tp->dirty & tp->valid) == 0) { 871 vm_page_flag_clear(tp, PG_CLEANCHK); 872 break; 873 } 874 maf[ i - 1 ] = tp; 875 maxf++; 876 continue; 877 } 878 break; 879 } 880 881 maxb = 0; 882 chkb = vm_pageout_page_count - maxf; 883 if (chkb) { 884 for(i = 1; i < chkb;i++) { 885 vm_page_t tp; 886 887 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 888 if ((tp->flags & PG_BUSY) || 889 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 890 (tp->flags & PG_CLEANCHK) == 0) || 891 (tp->busy != 0)) 892 break; 893 if ((tp->queue - tp->pc) == PQ_CACHE) { 894 vm_page_flag_clear(tp, PG_CLEANCHK); 895 break; 896 } 897 vm_page_test_dirty(tp); 898 if ((tp->dirty & tp->valid) == 0) { 899 vm_page_flag_clear(tp, PG_CLEANCHK); 900 break; 901 } 902 mab[ i - 1 ] = tp; 903 maxb++; 904 continue; 905 } 906 break; 907 } 908 } 909 910 for(i = 0; i < maxb; i++) { 911 int index = (maxb - i) - 1; 912 ma[index] = mab[i]; 913 vm_page_flag_clear(ma[index], PG_CLEANCHK); 914 } 915 vm_page_flag_clear(p, PG_CLEANCHK); 916 ma[maxb] = p; 917 for(i = 0; i < maxf; i++) { 918 int index = (maxb + i) + 1; 919 ma[index] = maf[i]; 920 vm_page_flag_clear(ma[index], PG_CLEANCHK); 921 } 922 runlen = maxb + maxf + 1; 923 924 splx(s); 925 vm_pageout_flush(ma, runlen, pagerflags, TRUE); 926 for (i = 0; i < runlen; i++) { 927 if (ma[i]->valid & ma[i]->dirty) { 928 pmap_page_protect(ma[i], VM_PROT_READ); 929 vm_page_flag_set(ma[i], PG_CLEANCHK); 930 931 /* 932 * maxf will end up being the actual number of pages 933 * we wrote out contiguously, non-inclusive of the 934 * first page. We do not count look-behind pages. 935 */ 936 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 937 maxf = i - maxb - 1; 938 } 939 } 940 return(maxf + 1); 941 } 942 943 /* 944 * vm_object_madvise: 945 * 946 * Implements the madvise function at the object/page level. 947 * 948 * MADV_WILLNEED (any object) 949 * 950 * Activate the specified pages if they are resident. 951 * 952 * MADV_DONTNEED (any object) 953 * 954 * Deactivate the specified pages if they are resident. 955 * 956 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 957 * OBJ_ONEMAPPING only) 958 * 959 * Deactivate and clean the specified pages if they are 960 * resident. This permits the process to reuse the pages 961 * without faulting or the kernel to reclaim the pages 962 * without I/O. 963 */ 964 void 965 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 966 { 967 vm_pindex_t end, tpindex; 968 vm_object_t tobject; 969 vm_page_t m; 970 971 if (object == NULL) 972 return; 973 974 vm_object_lock(object); 975 976 end = pindex + count; 977 978 /* 979 * Locate and adjust resident pages 980 */ 981 for (; pindex < end; pindex += 1) { 982 relookup: 983 tobject = object; 984 tpindex = pindex; 985 shadowlookup: 986 /* 987 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 988 * and those pages must be OBJ_ONEMAPPING. 989 */ 990 if (advise == MADV_FREE) { 991 if ((tobject->type != OBJT_DEFAULT && 992 tobject->type != OBJT_SWAP) || 993 (tobject->flags & OBJ_ONEMAPPING) == 0) { 994 continue; 995 } 996 } 997 998 m = vm_page_lookup(tobject, tpindex); 999 1000 if (m == NULL) { 1001 /* 1002 * There may be swap even if there is no backing page 1003 */ 1004 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1005 swap_pager_freespace(tobject, tpindex, 1); 1006 1007 /* 1008 * next object 1009 */ 1010 tobject = tobject->backing_object; 1011 if (tobject == NULL) 1012 continue; 1013 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1014 goto shadowlookup; 1015 } 1016 1017 /* 1018 * If the page is busy or not in a normal active state, 1019 * we skip it. If the page is not managed there are no 1020 * page queues to mess with. Things can break if we mess 1021 * with pages in any of the below states. 1022 */ 1023 vm_page_lock_queues(); 1024 if (m->hold_count || 1025 m->wire_count || 1026 (m->flags & PG_UNMANAGED) || 1027 m->valid != VM_PAGE_BITS_ALL) { 1028 vm_page_unlock_queues(); 1029 continue; 1030 } 1031 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) 1032 goto relookup; 1033 if (advise == MADV_WILLNEED) { 1034 vm_page_activate(m); 1035 } else if (advise == MADV_DONTNEED) { 1036 vm_page_dontneed(m); 1037 } else if (advise == MADV_FREE) { 1038 /* 1039 * Mark the page clean. This will allow the page 1040 * to be freed up by the system. However, such pages 1041 * are often reused quickly by malloc()/free() 1042 * so we do not do anything that would cause 1043 * a page fault if we can help it. 1044 * 1045 * Specifically, we do not try to actually free 1046 * the page now nor do we try to put it in the 1047 * cache (which would cause a page fault on reuse). 1048 * 1049 * But we do make the page is freeable as we 1050 * can without actually taking the step of unmapping 1051 * it. 1052 */ 1053 pmap_clear_modify(m); 1054 m->dirty = 0; 1055 m->act_count = 0; 1056 vm_page_dontneed(m); 1057 } 1058 vm_page_unlock_queues(); 1059 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1060 swap_pager_freespace(tobject, tpindex, 1); 1061 } 1062 vm_object_unlock(object); 1063 } 1064 1065 /* 1066 * vm_object_shadow: 1067 * 1068 * Create a new object which is backed by the 1069 * specified existing object range. The source 1070 * object reference is deallocated. 1071 * 1072 * The new object and offset into that object 1073 * are returned in the source parameters. 1074 */ 1075 void 1076 vm_object_shadow( 1077 vm_object_t *object, /* IN/OUT */ 1078 vm_ooffset_t *offset, /* IN/OUT */ 1079 vm_size_t length) 1080 { 1081 vm_object_t source; 1082 vm_object_t result; 1083 1084 GIANT_REQUIRED; 1085 1086 source = *object; 1087 1088 /* 1089 * Don't create the new object if the old object isn't shared. 1090 */ 1091 if (source != NULL) { 1092 VM_OBJECT_LOCK(source); 1093 if (source->ref_count == 1 && 1094 source->handle == NULL && 1095 (source->type == OBJT_DEFAULT || 1096 source->type == OBJT_SWAP)) { 1097 VM_OBJECT_UNLOCK(source); 1098 return; 1099 } 1100 VM_OBJECT_UNLOCK(source); 1101 } 1102 1103 /* 1104 * Allocate a new object with the given length. 1105 */ 1106 result = vm_object_allocate(OBJT_DEFAULT, length); 1107 1108 /* 1109 * The new object shadows the source object, adding a reference to it. 1110 * Our caller changes his reference to point to the new object, 1111 * removing a reference to the source object. Net result: no change 1112 * of reference count. 1113 * 1114 * Try to optimize the result object's page color when shadowing 1115 * in order to maintain page coloring consistency in the combined 1116 * shadowed object. 1117 */ 1118 result->backing_object = source; 1119 if (source != NULL) { 1120 VM_OBJECT_LOCK(source); 1121 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1122 source->shadow_count++; 1123 source->generation++; 1124 if (length < source->size) 1125 length = source->size; 1126 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1127 source->generation > 1) 1128 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1129 result->pg_color = (source->pg_color + 1130 length * source->generation) & PQ_L2_MASK; 1131 VM_OBJECT_UNLOCK(source); 1132 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1133 PQ_L2_MASK; 1134 } 1135 1136 /* 1137 * Store the offset into the source object, and fix up the offset into 1138 * the new object. 1139 */ 1140 result->backing_object_offset = *offset; 1141 1142 /* 1143 * Return the new things 1144 */ 1145 *offset = 0; 1146 *object = result; 1147 } 1148 1149 /* 1150 * vm_object_split: 1151 * 1152 * Split the pages in a map entry into a new object. This affords 1153 * easier removal of unused pages, and keeps object inheritance from 1154 * being a negative impact on memory usage. 1155 */ 1156 void 1157 vm_object_split(vm_map_entry_t entry) 1158 { 1159 vm_page_t m; 1160 vm_object_t orig_object, new_object, source; 1161 vm_offset_t s, e; 1162 vm_pindex_t offidxstart, offidxend; 1163 vm_size_t idx, size; 1164 vm_ooffset_t offset; 1165 1166 GIANT_REQUIRED; 1167 1168 orig_object = entry->object.vm_object; 1169 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1170 return; 1171 if (orig_object->ref_count <= 1) 1172 return; 1173 1174 offset = entry->offset; 1175 s = entry->start; 1176 e = entry->end; 1177 1178 offidxstart = OFF_TO_IDX(offset); 1179 offidxend = offidxstart + OFF_TO_IDX(e - s); 1180 size = offidxend - offidxstart; 1181 1182 new_object = vm_pager_allocate(orig_object->type, 1183 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1184 if (new_object == NULL) 1185 return; 1186 1187 source = orig_object->backing_object; 1188 if (source != NULL) { 1189 vm_object_reference(source); /* Referenced by new_object */ 1190 VM_OBJECT_LOCK(source); 1191 TAILQ_INSERT_TAIL(&source->shadow_head, 1192 new_object, shadow_list); 1193 source->shadow_count++; 1194 source->generation++; 1195 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1196 VM_OBJECT_UNLOCK(source); 1197 new_object->backing_object_offset = 1198 orig_object->backing_object_offset + offset; 1199 new_object->backing_object = source; 1200 } 1201 VM_OBJECT_LOCK(orig_object); 1202 for (idx = 0; idx < size; idx++) { 1203 retry: 1204 m = vm_page_lookup(orig_object, offidxstart + idx); 1205 if (m == NULL) 1206 continue; 1207 1208 /* 1209 * We must wait for pending I/O to complete before we can 1210 * rename the page. 1211 * 1212 * We do not have to VM_PROT_NONE the page as mappings should 1213 * not be changed by this operation. 1214 */ 1215 vm_page_lock_queues(); 1216 if (vm_page_sleep_if_busy(m, TRUE, "spltwt")) 1217 goto retry; 1218 1219 vm_page_busy(m); 1220 vm_page_rename(m, new_object, idx); 1221 /* page automatically made dirty by rename and cache handled */ 1222 vm_page_busy(m); 1223 vm_page_unlock_queues(); 1224 } 1225 if (orig_object->type == OBJT_SWAP) { 1226 vm_object_pip_add(orig_object, 1); 1227 VM_OBJECT_UNLOCK(orig_object); 1228 /* 1229 * copy orig_object pages into new_object 1230 * and destroy unneeded pages in 1231 * shadow object. 1232 */ 1233 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1234 VM_OBJECT_LOCK(orig_object); 1235 vm_object_pip_wakeup(orig_object); 1236 } 1237 VM_OBJECT_UNLOCK(orig_object); 1238 vm_page_lock_queues(); 1239 TAILQ_FOREACH(m, &new_object->memq, listq) 1240 vm_page_wakeup(m); 1241 vm_page_unlock_queues(); 1242 entry->object.vm_object = new_object; 1243 entry->offset = 0LL; 1244 vm_object_deallocate(orig_object); 1245 } 1246 1247 #define OBSC_TEST_ALL_SHADOWED 0x0001 1248 #define OBSC_COLLAPSE_NOWAIT 0x0002 1249 #define OBSC_COLLAPSE_WAIT 0x0004 1250 1251 static __inline int 1252 vm_object_backing_scan(vm_object_t object, int op) 1253 { 1254 int s; 1255 int r = 1; 1256 vm_page_t p; 1257 vm_object_t backing_object; 1258 vm_pindex_t backing_offset_index; 1259 1260 s = splvm(); 1261 GIANT_REQUIRED; 1262 1263 backing_object = object->backing_object; 1264 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1265 1266 /* 1267 * Initial conditions 1268 */ 1269 if (op & OBSC_TEST_ALL_SHADOWED) { 1270 /* 1271 * We do not want to have to test for the existence of 1272 * swap pages in the backing object. XXX but with the 1273 * new swapper this would be pretty easy to do. 1274 * 1275 * XXX what about anonymous MAP_SHARED memory that hasn't 1276 * been ZFOD faulted yet? If we do not test for this, the 1277 * shadow test may succeed! XXX 1278 */ 1279 if (backing_object->type != OBJT_DEFAULT) { 1280 splx(s); 1281 return (0); 1282 } 1283 } 1284 if (op & OBSC_COLLAPSE_WAIT) { 1285 vm_object_set_flag(backing_object, OBJ_DEAD); 1286 } 1287 1288 /* 1289 * Our scan 1290 */ 1291 p = TAILQ_FIRST(&backing_object->memq); 1292 while (p) { 1293 vm_page_t next = TAILQ_NEXT(p, listq); 1294 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1295 1296 if (op & OBSC_TEST_ALL_SHADOWED) { 1297 vm_page_t pp; 1298 1299 /* 1300 * Ignore pages outside the parent object's range 1301 * and outside the parent object's mapping of the 1302 * backing object. 1303 * 1304 * note that we do not busy the backing object's 1305 * page. 1306 */ 1307 if ( 1308 p->pindex < backing_offset_index || 1309 new_pindex >= object->size 1310 ) { 1311 p = next; 1312 continue; 1313 } 1314 1315 /* 1316 * See if the parent has the page or if the parent's 1317 * object pager has the page. If the parent has the 1318 * page but the page is not valid, the parent's 1319 * object pager must have the page. 1320 * 1321 * If this fails, the parent does not completely shadow 1322 * the object and we might as well give up now. 1323 */ 1324 1325 pp = vm_page_lookup(object, new_pindex); 1326 if ( 1327 (pp == NULL || pp->valid == 0) && 1328 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1329 ) { 1330 r = 0; 1331 break; 1332 } 1333 } 1334 1335 /* 1336 * Check for busy page 1337 */ 1338 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1339 vm_page_t pp; 1340 1341 vm_page_lock_queues(); 1342 if (op & OBSC_COLLAPSE_NOWAIT) { 1343 if ((p->flags & PG_BUSY) || 1344 !p->valid || 1345 p->hold_count || 1346 p->wire_count || 1347 p->busy) { 1348 vm_page_unlock_queues(); 1349 p = next; 1350 continue; 1351 } 1352 } else if (op & OBSC_COLLAPSE_WAIT) { 1353 if (vm_page_sleep_if_busy(p, TRUE, "vmocol")) { 1354 /* 1355 * If we slept, anything could have 1356 * happened. Since the object is 1357 * marked dead, the backing offset 1358 * should not have changed so we 1359 * just restart our scan. 1360 */ 1361 p = TAILQ_FIRST(&backing_object->memq); 1362 continue; 1363 } 1364 } 1365 1366 /* 1367 * Busy the page 1368 */ 1369 vm_page_busy(p); 1370 vm_page_unlock_queues(); 1371 1372 KASSERT( 1373 p->object == backing_object, 1374 ("vm_object_qcollapse(): object mismatch") 1375 ); 1376 1377 /* 1378 * Destroy any associated swap 1379 */ 1380 if (backing_object->type == OBJT_SWAP) { 1381 swap_pager_freespace( 1382 backing_object, 1383 p->pindex, 1384 1 1385 ); 1386 } 1387 1388 if ( 1389 p->pindex < backing_offset_index || 1390 new_pindex >= object->size 1391 ) { 1392 /* 1393 * Page is out of the parent object's range, we 1394 * can simply destroy it. 1395 */ 1396 vm_page_lock_queues(); 1397 pmap_remove_all(p); 1398 vm_page_free(p); 1399 vm_page_unlock_queues(); 1400 p = next; 1401 continue; 1402 } 1403 1404 pp = vm_page_lookup(object, new_pindex); 1405 if ( 1406 pp != NULL || 1407 vm_pager_has_page(object, new_pindex, NULL, NULL) 1408 ) { 1409 /* 1410 * page already exists in parent OR swap exists 1411 * for this location in the parent. Destroy 1412 * the original page from the backing object. 1413 * 1414 * Leave the parent's page alone 1415 */ 1416 vm_page_lock_queues(); 1417 pmap_remove_all(p); 1418 vm_page_free(p); 1419 vm_page_unlock_queues(); 1420 p = next; 1421 continue; 1422 } 1423 1424 /* 1425 * Page does not exist in parent, rename the 1426 * page from the backing object to the main object. 1427 * 1428 * If the page was mapped to a process, it can remain 1429 * mapped through the rename. 1430 */ 1431 vm_page_lock_queues(); 1432 vm_page_rename(p, object, new_pindex); 1433 vm_page_unlock_queues(); 1434 /* page automatically made dirty by rename */ 1435 } 1436 p = next; 1437 } 1438 splx(s); 1439 return (r); 1440 } 1441 1442 1443 /* 1444 * this version of collapse allows the operation to occur earlier and 1445 * when paging_in_progress is true for an object... This is not a complete 1446 * operation, but should plug 99.9% of the rest of the leaks. 1447 */ 1448 static void 1449 vm_object_qcollapse(vm_object_t object) 1450 { 1451 vm_object_t backing_object = object->backing_object; 1452 1453 GIANT_REQUIRED; 1454 1455 if (backing_object->ref_count != 1) 1456 return; 1457 1458 backing_object->ref_count += 2; 1459 1460 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1461 1462 backing_object->ref_count -= 2; 1463 } 1464 1465 /* 1466 * vm_object_collapse: 1467 * 1468 * Collapse an object with the object backing it. 1469 * Pages in the backing object are moved into the 1470 * parent, and the backing object is deallocated. 1471 */ 1472 void 1473 vm_object_collapse(vm_object_t object) 1474 { 1475 GIANT_REQUIRED; 1476 1477 while (TRUE) { 1478 vm_object_t backing_object; 1479 1480 /* 1481 * Verify that the conditions are right for collapse: 1482 * 1483 * The object exists and the backing object exists. 1484 */ 1485 if (object == NULL) 1486 break; 1487 1488 if ((backing_object = object->backing_object) == NULL) 1489 break; 1490 1491 /* 1492 * we check the backing object first, because it is most likely 1493 * not collapsable. 1494 */ 1495 if (backing_object->handle != NULL || 1496 (backing_object->type != OBJT_DEFAULT && 1497 backing_object->type != OBJT_SWAP) || 1498 (backing_object->flags & OBJ_DEAD) || 1499 object->handle != NULL || 1500 (object->type != OBJT_DEFAULT && 1501 object->type != OBJT_SWAP) || 1502 (object->flags & OBJ_DEAD)) { 1503 break; 1504 } 1505 1506 if ( 1507 object->paging_in_progress != 0 || 1508 backing_object->paging_in_progress != 0 1509 ) { 1510 vm_object_qcollapse(object); 1511 break; 1512 } 1513 1514 /* 1515 * We know that we can either collapse the backing object (if 1516 * the parent is the only reference to it) or (perhaps) have 1517 * the parent bypass the object if the parent happens to shadow 1518 * all the resident pages in the entire backing object. 1519 * 1520 * This is ignoring pager-backed pages such as swap pages. 1521 * vm_object_backing_scan fails the shadowing test in this 1522 * case. 1523 */ 1524 if (backing_object->ref_count == 1) { 1525 /* 1526 * If there is exactly one reference to the backing 1527 * object, we can collapse it into the parent. 1528 */ 1529 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1530 1531 /* 1532 * Move the pager from backing_object to object. 1533 */ 1534 VM_OBJECT_LOCK(backing_object); 1535 if (backing_object->type == OBJT_SWAP) { 1536 vm_object_pip_add(backing_object, 1); 1537 VM_OBJECT_UNLOCK(backing_object); 1538 /* 1539 * scrap the paging_offset junk and do a 1540 * discrete copy. This also removes major 1541 * assumptions about how the swap-pager 1542 * works from where it doesn't belong. The 1543 * new swapper is able to optimize the 1544 * destroy-source case. 1545 */ 1546 VM_OBJECT_LOCK(object); 1547 vm_object_pip_add(object, 1); 1548 VM_OBJECT_UNLOCK(object); 1549 swap_pager_copy( 1550 backing_object, 1551 object, 1552 OFF_TO_IDX(object->backing_object_offset), TRUE); 1553 VM_OBJECT_LOCK(object); 1554 vm_object_pip_wakeup(object); 1555 VM_OBJECT_UNLOCK(object); 1556 1557 VM_OBJECT_LOCK(backing_object); 1558 vm_object_pip_wakeup(backing_object); 1559 } 1560 /* 1561 * Object now shadows whatever backing_object did. 1562 * Note that the reference to 1563 * backing_object->backing_object moves from within 1564 * backing_object to within object. 1565 */ 1566 TAILQ_REMOVE( 1567 &backing_object->shadow_head, 1568 object, 1569 shadow_list 1570 ); 1571 backing_object->shadow_count--; 1572 backing_object->generation++; 1573 if (backing_object->backing_object) { 1574 VM_OBJECT_LOCK(backing_object->backing_object); 1575 TAILQ_REMOVE( 1576 &backing_object->backing_object->shadow_head, 1577 backing_object, 1578 shadow_list 1579 ); 1580 backing_object->backing_object->shadow_count--; 1581 backing_object->backing_object->generation++; 1582 VM_OBJECT_UNLOCK(backing_object->backing_object); 1583 } 1584 object->backing_object = backing_object->backing_object; 1585 if (object->backing_object) { 1586 VM_OBJECT_LOCK(object->backing_object); 1587 TAILQ_INSERT_TAIL( 1588 &object->backing_object->shadow_head, 1589 object, 1590 shadow_list 1591 ); 1592 object->backing_object->shadow_count++; 1593 object->backing_object->generation++; 1594 VM_OBJECT_UNLOCK(object->backing_object); 1595 } 1596 1597 object->backing_object_offset += 1598 backing_object->backing_object_offset; 1599 1600 /* 1601 * Discard backing_object. 1602 * 1603 * Since the backing object has no pages, no pager left, 1604 * and no object references within it, all that is 1605 * necessary is to dispose of it. 1606 */ 1607 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1608 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1609 VM_OBJECT_UNLOCK(backing_object); 1610 1611 mtx_lock(&vm_object_list_mtx); 1612 TAILQ_REMOVE( 1613 &vm_object_list, 1614 backing_object, 1615 object_list 1616 ); 1617 mtx_unlock(&vm_object_list_mtx); 1618 1619 uma_zfree(obj_zone, backing_object); 1620 1621 object_collapses++; 1622 } else { 1623 vm_object_t new_backing_object; 1624 1625 /* 1626 * If we do not entirely shadow the backing object, 1627 * there is nothing we can do so we give up. 1628 */ 1629 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1630 break; 1631 } 1632 1633 /* 1634 * Make the parent shadow the next object in the 1635 * chain. Deallocating backing_object will not remove 1636 * it, since its reference count is at least 2. 1637 */ 1638 VM_OBJECT_LOCK(backing_object); 1639 TAILQ_REMOVE( 1640 &backing_object->shadow_head, 1641 object, 1642 shadow_list 1643 ); 1644 backing_object->shadow_count--; 1645 backing_object->generation++; 1646 VM_OBJECT_UNLOCK(backing_object); 1647 1648 new_backing_object = backing_object->backing_object; 1649 if ((object->backing_object = new_backing_object) != NULL) { 1650 vm_object_reference(new_backing_object); 1651 VM_OBJECT_LOCK(new_backing_object); 1652 TAILQ_INSERT_TAIL( 1653 &new_backing_object->shadow_head, 1654 object, 1655 shadow_list 1656 ); 1657 new_backing_object->shadow_count++; 1658 new_backing_object->generation++; 1659 VM_OBJECT_UNLOCK(new_backing_object); 1660 object->backing_object_offset += 1661 backing_object->backing_object_offset; 1662 } 1663 1664 /* 1665 * Drop the reference count on backing_object. Since 1666 * its ref_count was at least 2, it will not vanish; 1667 * so we don't need to call vm_object_deallocate, but 1668 * we do anyway. 1669 */ 1670 vm_object_deallocate(backing_object); 1671 object_bypasses++; 1672 } 1673 1674 /* 1675 * Try again with this object's new backing object. 1676 */ 1677 } 1678 } 1679 1680 /* 1681 * vm_object_page_remove: 1682 * 1683 * Removes all physical pages in the given range from the 1684 * object's list of pages. If the range's end is zero, all 1685 * physical pages from the range's start to the end of the object 1686 * are deleted. 1687 * 1688 * The object must be locked. 1689 */ 1690 void 1691 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1692 boolean_t clean_only) 1693 { 1694 vm_page_t p, next; 1695 1696 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1697 if (object->resident_page_count == 0) 1698 return; 1699 1700 /* 1701 * Since physically-backed objects do not use managed pages, we can't 1702 * remove pages from the object (we must instead remove the page 1703 * references, and then destroy the object). 1704 */ 1705 KASSERT(object->type != OBJT_PHYS, 1706 ("attempt to remove pages from a physical object")); 1707 1708 vm_object_pip_add(object, 1); 1709 again: 1710 vm_page_lock_queues(); 1711 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1712 if (p->pindex < start) { 1713 p = vm_page_splay(start, object->root); 1714 if ((object->root = p)->pindex < start) 1715 p = TAILQ_NEXT(p, listq); 1716 } 1717 } 1718 /* 1719 * Assert: the variable p is either (1) the page with the 1720 * least pindex greater than or equal to the parameter pindex 1721 * or (2) NULL. 1722 */ 1723 for (; 1724 p != NULL && (p->pindex < end || end == 0); 1725 p = next) { 1726 next = TAILQ_NEXT(p, listq); 1727 1728 if (p->wire_count != 0) { 1729 pmap_remove_all(p); 1730 if (!clean_only) 1731 p->valid = 0; 1732 continue; 1733 } 1734 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1735 goto again; 1736 if (clean_only && p->valid) { 1737 vm_page_test_dirty(p); 1738 if (p->valid & p->dirty) 1739 continue; 1740 } 1741 vm_page_busy(p); 1742 pmap_remove_all(p); 1743 vm_page_free(p); 1744 } 1745 vm_page_unlock_queues(); 1746 vm_object_pip_wakeup(object); 1747 } 1748 1749 /* 1750 * Routine: vm_object_coalesce 1751 * Function: Coalesces two objects backing up adjoining 1752 * regions of memory into a single object. 1753 * 1754 * returns TRUE if objects were combined. 1755 * 1756 * NOTE: Only works at the moment if the second object is NULL - 1757 * if it's not, which object do we lock first? 1758 * 1759 * Parameters: 1760 * prev_object First object to coalesce 1761 * prev_offset Offset into prev_object 1762 * next_object Second object into coalesce 1763 * next_offset Offset into next_object 1764 * 1765 * prev_size Size of reference to prev_object 1766 * next_size Size of reference to next_object 1767 * 1768 * Conditions: 1769 * The object must *not* be locked. 1770 */ 1771 boolean_t 1772 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1773 vm_size_t prev_size, vm_size_t next_size) 1774 { 1775 vm_pindex_t next_pindex; 1776 1777 if (prev_object == NULL) 1778 return (TRUE); 1779 vm_object_lock(prev_object); 1780 if (prev_object->type != OBJT_DEFAULT && 1781 prev_object->type != OBJT_SWAP) { 1782 vm_object_unlock(prev_object); 1783 return (FALSE); 1784 } 1785 1786 /* 1787 * Try to collapse the object first 1788 */ 1789 vm_object_collapse(prev_object); 1790 1791 /* 1792 * Can't coalesce if: . more than one reference . paged out . shadows 1793 * another object . has a copy elsewhere (any of which mean that the 1794 * pages not mapped to prev_entry may be in use anyway) 1795 */ 1796 if (prev_object->backing_object != NULL) { 1797 vm_object_unlock(prev_object); 1798 return (FALSE); 1799 } 1800 1801 prev_size >>= PAGE_SHIFT; 1802 next_size >>= PAGE_SHIFT; 1803 next_pindex = prev_pindex + prev_size; 1804 1805 if ((prev_object->ref_count > 1) && 1806 (prev_object->size != next_pindex)) { 1807 vm_object_unlock(prev_object); 1808 return (FALSE); 1809 } 1810 1811 /* 1812 * Remove any pages that may still be in the object from a previous 1813 * deallocation. 1814 */ 1815 if (next_pindex < prev_object->size) { 1816 VM_OBJECT_LOCK(prev_object); 1817 vm_object_page_remove(prev_object, 1818 next_pindex, 1819 next_pindex + next_size, FALSE); 1820 if (prev_object->type == OBJT_SWAP) 1821 swap_pager_freespace(prev_object, 1822 next_pindex, next_size); 1823 VM_OBJECT_UNLOCK(prev_object); 1824 } 1825 1826 /* 1827 * Extend the object if necessary. 1828 */ 1829 if (next_pindex + next_size > prev_object->size) 1830 prev_object->size = next_pindex + next_size; 1831 1832 vm_object_unlock(prev_object); 1833 return (TRUE); 1834 } 1835 1836 void 1837 vm_object_set_writeable_dirty(vm_object_t object) 1838 { 1839 struct vnode *vp; 1840 1841 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1842 if (object->type == OBJT_VNODE && 1843 (vp = (struct vnode *)object->handle) != NULL) { 1844 VI_LOCK(vp); 1845 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1846 vp->v_iflag |= VI_OBJDIRTY; 1847 VI_UNLOCK(vp); 1848 } 1849 } 1850 1851 #include "opt_ddb.h" 1852 #ifdef DDB 1853 #include <sys/kernel.h> 1854 1855 #include <sys/cons.h> 1856 1857 #include <ddb/ddb.h> 1858 1859 static int 1860 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1861 { 1862 vm_map_t tmpm; 1863 vm_map_entry_t tmpe; 1864 vm_object_t obj; 1865 int entcount; 1866 1867 if (map == 0) 1868 return 0; 1869 1870 if (entry == 0) { 1871 tmpe = map->header.next; 1872 entcount = map->nentries; 1873 while (entcount-- && (tmpe != &map->header)) { 1874 if (_vm_object_in_map(map, object, tmpe)) { 1875 return 1; 1876 } 1877 tmpe = tmpe->next; 1878 } 1879 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1880 tmpm = entry->object.sub_map; 1881 tmpe = tmpm->header.next; 1882 entcount = tmpm->nentries; 1883 while (entcount-- && tmpe != &tmpm->header) { 1884 if (_vm_object_in_map(tmpm, object, tmpe)) { 1885 return 1; 1886 } 1887 tmpe = tmpe->next; 1888 } 1889 } else if ((obj = entry->object.vm_object) != NULL) { 1890 for (; obj; obj = obj->backing_object) 1891 if (obj == object) { 1892 return 1; 1893 } 1894 } 1895 return 0; 1896 } 1897 1898 static int 1899 vm_object_in_map(vm_object_t object) 1900 { 1901 struct proc *p; 1902 1903 /* sx_slock(&allproc_lock); */ 1904 LIST_FOREACH(p, &allproc, p_list) { 1905 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1906 continue; 1907 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1908 /* sx_sunlock(&allproc_lock); */ 1909 return 1; 1910 } 1911 } 1912 /* sx_sunlock(&allproc_lock); */ 1913 if (_vm_object_in_map(kernel_map, object, 0)) 1914 return 1; 1915 if (_vm_object_in_map(kmem_map, object, 0)) 1916 return 1; 1917 if (_vm_object_in_map(pager_map, object, 0)) 1918 return 1; 1919 if (_vm_object_in_map(buffer_map, object, 0)) 1920 return 1; 1921 return 0; 1922 } 1923 1924 DB_SHOW_COMMAND(vmochk, vm_object_check) 1925 { 1926 vm_object_t object; 1927 1928 /* 1929 * make sure that internal objs are in a map somewhere 1930 * and none have zero ref counts. 1931 */ 1932 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1933 if (object->handle == NULL && 1934 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1935 if (object->ref_count == 0) { 1936 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1937 (long)object->size); 1938 } 1939 if (!vm_object_in_map(object)) { 1940 db_printf( 1941 "vmochk: internal obj is not in a map: " 1942 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1943 object->ref_count, (u_long)object->size, 1944 (u_long)object->size, 1945 (void *)object->backing_object); 1946 } 1947 } 1948 } 1949 } 1950 1951 /* 1952 * vm_object_print: [ debug ] 1953 */ 1954 DB_SHOW_COMMAND(object, vm_object_print_static) 1955 { 1956 /* XXX convert args. */ 1957 vm_object_t object = (vm_object_t)addr; 1958 boolean_t full = have_addr; 1959 1960 vm_page_t p; 1961 1962 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1963 #define count was_count 1964 1965 int count; 1966 1967 if (object == NULL) 1968 return; 1969 1970 db_iprintf( 1971 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 1972 object, (int)object->type, (uintmax_t)object->size, 1973 object->resident_page_count, object->ref_count, object->flags); 1974 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 1975 object->shadow_count, 1976 object->backing_object ? object->backing_object->ref_count : 0, 1977 object->backing_object, (uintmax_t)object->backing_object_offset); 1978 1979 if (!full) 1980 return; 1981 1982 db_indent += 2; 1983 count = 0; 1984 TAILQ_FOREACH(p, &object->memq, listq) { 1985 if (count == 0) 1986 db_iprintf("memory:="); 1987 else if (count == 6) { 1988 db_printf("\n"); 1989 db_iprintf(" ..."); 1990 count = 0; 1991 } else 1992 db_printf(","); 1993 count++; 1994 1995 db_printf("(off=0x%jx,page=0x%jx)", 1996 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 1997 } 1998 if (count != 0) 1999 db_printf("\n"); 2000 db_indent -= 2; 2001 } 2002 2003 /* XXX. */ 2004 #undef count 2005 2006 /* XXX need this non-static entry for calling from vm_map_print. */ 2007 void 2008 vm_object_print( 2009 /* db_expr_t */ long addr, 2010 boolean_t have_addr, 2011 /* db_expr_t */ long count, 2012 char *modif) 2013 { 2014 vm_object_print_static(addr, have_addr, count, modif); 2015 } 2016 2017 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2018 { 2019 vm_object_t object; 2020 int nl = 0; 2021 int c; 2022 2023 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2024 vm_pindex_t idx, fidx; 2025 vm_pindex_t osize; 2026 vm_paddr_t pa = -1, padiff; 2027 int rcount; 2028 vm_page_t m; 2029 2030 db_printf("new object: %p\n", (void *)object); 2031 if (nl > 18) { 2032 c = cngetc(); 2033 if (c != ' ') 2034 return; 2035 nl = 0; 2036 } 2037 nl++; 2038 rcount = 0; 2039 fidx = 0; 2040 osize = object->size; 2041 if (osize > 128) 2042 osize = 128; 2043 for (idx = 0; idx < osize; idx++) { 2044 m = vm_page_lookup(object, idx); 2045 if (m == NULL) { 2046 if (rcount) { 2047 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2048 (long)fidx, rcount, (long)pa); 2049 if (nl > 18) { 2050 c = cngetc(); 2051 if (c != ' ') 2052 return; 2053 nl = 0; 2054 } 2055 nl++; 2056 rcount = 0; 2057 } 2058 continue; 2059 } 2060 2061 2062 if (rcount && 2063 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2064 ++rcount; 2065 continue; 2066 } 2067 if (rcount) { 2068 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2069 padiff >>= PAGE_SHIFT; 2070 padiff &= PQ_L2_MASK; 2071 if (padiff == 0) { 2072 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2073 ++rcount; 2074 continue; 2075 } 2076 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2077 (long)fidx, rcount, (long)pa); 2078 db_printf("pd(%ld)\n", (long)padiff); 2079 if (nl > 18) { 2080 c = cngetc(); 2081 if (c != ' ') 2082 return; 2083 nl = 0; 2084 } 2085 nl++; 2086 } 2087 fidx = idx; 2088 pa = VM_PAGE_TO_PHYS(m); 2089 rcount = 1; 2090 } 2091 if (rcount) { 2092 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2093 (long)fidx, rcount, (long)pa); 2094 if (nl > 18) { 2095 c = cngetc(); 2096 if (c != ' ') 2097 return; 2098 nl = 0; 2099 } 2100 nl++; 2101 } 2102 } 2103 } 2104 #endif /* DDB */ 2105