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/stdint.h> 82 #include <sys/vnode.h> 83 #include <sys/vmmeter.h> 84 #include <sys/sx.h> 85 86 #include <vm/vm.h> 87 #include <vm/vm_param.h> 88 #include <vm/pmap.h> 89 #include <vm/vm_map.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_pageout.h> 93 #include <vm/vm_pager.h> 94 #include <vm/swap_pager.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_extern.h> 97 #include <vm/uma.h> 98 99 #define EASY_SCAN_FACTOR 8 100 101 #define MSYNC_FLUSH_HARDSEQ 0x01 102 #define MSYNC_FLUSH_SOFTSEQ 0x02 103 104 /* 105 * msync / VM object flushing optimizations 106 */ 107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 109 CTLFLAG_RW, &msync_flush_flags, 0, ""); 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); 113 114 /* 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 extern int vm_pageout_page_count; 147 148 static long object_collapses; 149 static long object_bypasses; 150 static int next_index; 151 static uma_zone_t obj_zone; 152 #define VM_OBJECTS_INIT 256 153 154 static void vm_object_zinit(void *mem, int size); 155 156 #ifdef INVARIANTS 157 static void vm_object_zdtor(void *mem, int size, void *arg); 158 159 static void 160 vm_object_zdtor(void *mem, int size, void *arg) 161 { 162 vm_object_t object; 163 164 object = (vm_object_t)mem; 165 KASSERT(object->paging_in_progress == 0, 166 ("object %p paging_in_progress = %d", 167 object, object->paging_in_progress)); 168 KASSERT(object->resident_page_count == 0, 169 ("object %p resident_page_count = %d", 170 object, object->resident_page_count)); 171 KASSERT(object->shadow_count == 0, 172 ("object %p shadow_count = %d", 173 object, object->shadow_count)); 174 } 175 #endif 176 177 static void 178 vm_object_zinit(void *mem, int size) 179 { 180 vm_object_t object; 181 182 object = (vm_object_t)mem; 183 bzero(&object->mtx, sizeof(object->mtx)); 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 mtx_init(&object->mtx, "vm object", NULL, MTX_DEF); 197 198 TAILQ_INIT(&object->memq); 199 TAILQ_INIT(&object->shadow_head); 200 201 object->root = NULL; 202 object->type = type; 203 object->size = size; 204 object->ref_count = 1; 205 object->flags = 0; 206 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 207 vm_object_set_flag(object, OBJ_ONEMAPPING); 208 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 209 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 210 else 211 incr = size; 212 do 213 object->pg_color = next_index; 214 while (!atomic_cmpset_int(&next_index, object->pg_color, 215 (object->pg_color + incr) & PQ_L2_MASK)); 216 object->handle = NULL; 217 object->backing_object = NULL; 218 object->backing_object_offset = (vm_ooffset_t) 0; 219 220 atomic_add_int(&object->generation, 1); 221 222 mtx_lock(&vm_object_list_mtx); 223 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 224 mtx_unlock(&vm_object_list_mtx); 225 } 226 227 /* 228 * vm_object_init: 229 * 230 * Initialize the VM objects module. 231 */ 232 void 233 vm_object_init(void) 234 { 235 TAILQ_INIT(&vm_object_list); 236 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 237 238 kernel_object = &kernel_object_store; 239 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 240 kernel_object); 241 242 kmem_object = &kmem_object_store; 243 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 244 kmem_object); 245 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 246 #ifdef INVARIANTS 247 vm_object_zdtor, 248 #else 249 NULL, 250 #endif 251 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 252 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 253 } 254 255 void 256 vm_object_set_flag(vm_object_t object, u_short bits) 257 { 258 object->flags |= bits; 259 } 260 261 void 262 vm_object_clear_flag(vm_object_t object, u_short bits) 263 { 264 265 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED); 266 object->flags &= ~bits; 267 } 268 269 void 270 vm_object_pip_add(vm_object_t object, short i) 271 { 272 GIANT_REQUIRED; 273 object->paging_in_progress += i; 274 } 275 276 void 277 vm_object_pip_subtract(vm_object_t object, short i) 278 { 279 GIANT_REQUIRED; 280 object->paging_in_progress -= i; 281 } 282 283 void 284 vm_object_pip_wakeup(vm_object_t object) 285 { 286 GIANT_REQUIRED; 287 object->paging_in_progress--; 288 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 289 vm_object_clear_flag(object, OBJ_PIPWNT); 290 wakeup(object); 291 } 292 } 293 294 void 295 vm_object_pip_wakeupn(vm_object_t object, short i) 296 { 297 GIANT_REQUIRED; 298 if (i) 299 object->paging_in_progress -= i; 300 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 301 vm_object_clear_flag(object, OBJ_PIPWNT); 302 wakeup(object); 303 } 304 } 305 306 void 307 vm_object_pip_sleep(vm_object_t object, char *waitid) 308 { 309 GIANT_REQUIRED; 310 if (object->paging_in_progress) { 311 int s = splvm(); 312 if (object->paging_in_progress) { 313 vm_object_set_flag(object, OBJ_PIPWNT); 314 tsleep(object, PVM, waitid, 0); 315 } 316 splx(s); 317 } 318 } 319 320 void 321 vm_object_pip_wait(vm_object_t object, char *waitid) 322 { 323 GIANT_REQUIRED; 324 while (object->paging_in_progress) 325 vm_object_pip_sleep(object, waitid); 326 } 327 328 /* 329 * vm_object_allocate_wait 330 * 331 * Return a new object with the given size, and give the user the 332 * option of waiting for it to complete or failing if the needed 333 * memory isn't available. 334 */ 335 vm_object_t 336 vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 337 { 338 vm_object_t result; 339 340 result = (vm_object_t) uma_zalloc(obj_zone, flags); 341 342 if (result != NULL) 343 _vm_object_allocate(type, size, result); 344 345 return (result); 346 } 347 348 /* 349 * vm_object_allocate: 350 * 351 * Returns a new object with the given size. 352 */ 353 vm_object_t 354 vm_object_allocate(objtype_t type, vm_pindex_t size) 355 { 356 return(vm_object_allocate_wait(type, size, M_WAITOK)); 357 } 358 359 360 /* 361 * vm_object_reference: 362 * 363 * Gets another reference to the given object. 364 */ 365 void 366 vm_object_reference(vm_object_t object) 367 { 368 if (object == NULL) 369 return; 370 371 vm_object_lock(object); 372 #if 0 373 /* object can be re-referenced during final cleaning */ 374 KASSERT(!(object->flags & OBJ_DEAD), 375 ("vm_object_reference: attempting to reference dead obj")); 376 #endif 377 378 object->ref_count++; 379 if (object->type == OBJT_VNODE) { 380 while (vget((struct vnode *) object->handle, LK_RETRY, curthread)) { 381 printf("vm_object_reference: delay in getting object\n"); 382 } 383 } 384 vm_object_unlock(object); 385 } 386 387 /* 388 * Handle deallocating an object of type OBJT_VNODE. 389 */ 390 void 391 vm_object_vndeallocate(vm_object_t object) 392 { 393 struct vnode *vp = (struct vnode *) object->handle; 394 395 GIANT_REQUIRED; 396 KASSERT(object->type == OBJT_VNODE, 397 ("vm_object_vndeallocate: not a vnode object")); 398 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 399 #ifdef INVARIANTS 400 if (object->ref_count == 0) { 401 vprint("vm_object_vndeallocate", vp); 402 panic("vm_object_vndeallocate: bad object reference count"); 403 } 404 #endif 405 406 object->ref_count--; 407 if (object->ref_count == 0) { 408 mp_fixme("Unlocked vflag access."); 409 vp->v_vflag &= ~VV_TEXT; 410 #ifdef ENABLE_VFS_IOOPT 411 vm_object_clear_flag(object, OBJ_OPT); 412 #endif 413 } 414 /* 415 * vrele may need a vop lock 416 */ 417 vrele(vp); 418 } 419 420 /* 421 * vm_object_deallocate: 422 * 423 * Release a reference to the specified object, 424 * gained either through a vm_object_allocate 425 * or a vm_object_reference call. When all references 426 * are gone, storage associated with this object 427 * may be relinquished. 428 * 429 * No object may be locked. 430 */ 431 void 432 vm_object_deallocate(vm_object_t object) 433 { 434 vm_object_t temp; 435 436 vm_object_lock(object); 437 while (object != NULL) { 438 439 if (object->type == OBJT_VNODE) { 440 vm_object_vndeallocate(object); 441 vm_object_unlock(object); 442 return; 443 } 444 445 KASSERT(object->ref_count != 0, 446 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 447 448 /* 449 * If the reference count goes to 0 we start calling 450 * vm_object_terminate() on the object chain. 451 * A ref count of 1 may be a special case depending on the 452 * shadow count being 0 or 1. 453 */ 454 object->ref_count--; 455 if (object->ref_count > 1) { 456 vm_object_unlock(object); 457 return; 458 } else if (object->ref_count == 1) { 459 if (object->shadow_count == 0) { 460 vm_object_set_flag(object, OBJ_ONEMAPPING); 461 } else if ((object->shadow_count == 1) && 462 (object->handle == NULL) && 463 (object->type == OBJT_DEFAULT || 464 object->type == OBJT_SWAP)) { 465 vm_object_t robject; 466 467 robject = TAILQ_FIRST(&object->shadow_head); 468 KASSERT(robject != NULL, 469 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 470 object->ref_count, 471 object->shadow_count)); 472 if ((robject->handle == NULL) && 473 (robject->type == OBJT_DEFAULT || 474 robject->type == OBJT_SWAP)) { 475 476 robject->ref_count++; 477 478 while ( 479 robject->paging_in_progress || 480 object->paging_in_progress 481 ) { 482 vm_object_pip_sleep(robject, "objde1"); 483 vm_object_pip_sleep(object, "objde2"); 484 } 485 486 if (robject->ref_count == 1) { 487 robject->ref_count--; 488 object = robject; 489 goto doterm; 490 } 491 492 object = robject; 493 vm_object_collapse(object); 494 continue; 495 } 496 } 497 vm_object_unlock(object); 498 return; 499 } 500 doterm: 501 temp = object->backing_object; 502 if (temp) { 503 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 504 temp->shadow_count--; 505 #ifdef ENABLE_VFS_IOOPT 506 if (temp->ref_count == 0) 507 vm_object_clear_flag(temp, OBJ_OPT); 508 #endif 509 temp->generation++; 510 object->backing_object = NULL; 511 } 512 /* 513 * Don't double-terminate, we could be in a termination 514 * recursion due to the terminate having to sync data 515 * to disk. 516 */ 517 if ((object->flags & OBJ_DEAD) == 0) 518 vm_object_terminate(object); 519 object = temp; 520 } 521 vm_object_unlock(object); 522 } 523 524 /* 525 * vm_object_terminate actually destroys the specified object, freeing 526 * up all previously used resources. 527 * 528 * The object must be locked. 529 * This routine may block. 530 */ 531 void 532 vm_object_terminate(vm_object_t object) 533 { 534 vm_page_t p; 535 int s; 536 537 GIANT_REQUIRED; 538 539 /* 540 * Make sure no one uses us. 541 */ 542 vm_object_set_flag(object, OBJ_DEAD); 543 544 /* 545 * wait for the pageout daemon to be done with the object 546 */ 547 vm_object_pip_wait(object, "objtrm"); 548 549 KASSERT(!object->paging_in_progress, 550 ("vm_object_terminate: pageout in progress")); 551 552 /* 553 * Clean and free the pages, as appropriate. All references to the 554 * object are gone, so we don't need to lock it. 555 */ 556 if (object->type == OBJT_VNODE) { 557 struct vnode *vp; 558 559 #ifdef ENABLE_VFS_IOOPT 560 /* 561 * Freeze optimized copies. 562 */ 563 vm_freeze_copyopts(object, 0, object->size); 564 #endif 565 /* 566 * Clean pages and flush buffers. 567 */ 568 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 569 570 vp = (struct vnode *) object->handle; 571 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 572 } 573 574 KASSERT(object->ref_count == 0, 575 ("vm_object_terminate: object with references, ref_count=%d", 576 object->ref_count)); 577 578 /* 579 * Now free any remaining pages. For internal objects, this also 580 * removes them from paging queues. Don't free wired pages, just 581 * remove them from the object. 582 */ 583 s = splvm(); 584 vm_page_lock_queues(); 585 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 586 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 587 ("vm_object_terminate: freeing busy page %p " 588 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 589 if (p->wire_count == 0) { 590 vm_page_busy(p); 591 vm_page_free(p); 592 cnt.v_pfree++; 593 } else { 594 vm_page_busy(p); 595 vm_page_remove(p); 596 } 597 } 598 vm_page_unlock_queues(); 599 splx(s); 600 601 /* 602 * Let the pager know object is dead. 603 */ 604 vm_pager_deallocate(object); 605 606 /* 607 * Remove the object from the global object list. 608 */ 609 mtx_lock(&vm_object_list_mtx); 610 TAILQ_REMOVE(&vm_object_list, object, object_list); 611 mtx_unlock(&vm_object_list_mtx); 612 613 mtx_destroy(&object->mtx); 614 wakeup(object); 615 616 /* 617 * Free the space for the object. 618 */ 619 uma_zfree(obj_zone, object); 620 } 621 622 /* 623 * vm_object_page_clean 624 * 625 * Clean all dirty pages in the specified range of object. Leaves page 626 * on whatever queue it is currently on. If NOSYNC is set then do not 627 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 628 * leaving the object dirty. 629 * 630 * When stuffing pages asynchronously, allow clustering. XXX we need a 631 * synchronous clustering mode implementation. 632 * 633 * Odd semantics: if start == end, we clean everything. 634 * 635 * The object must be locked. 636 */ 637 void 638 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 639 { 640 vm_page_t p, np; 641 vm_pindex_t tstart, tend; 642 vm_pindex_t pi; 643 struct vnode *vp; 644 int clearobjflags; 645 int pagerflags; 646 int curgeneration; 647 648 GIANT_REQUIRED; 649 650 if (object->type != OBJT_VNODE || 651 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 652 return; 653 654 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 655 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 656 657 vp = object->handle; 658 659 vm_object_set_flag(object, OBJ_CLEANING); 660 661 tstart = start; 662 if (end == 0) { 663 tend = object->size; 664 } else { 665 tend = end; 666 } 667 668 vm_page_lock_queues(); 669 /* 670 * If the caller is smart and only msync()s a range he knows is 671 * dirty, we may be able to avoid an object scan. This results in 672 * a phenominal improvement in performance. We cannot do this 673 * as a matter of course because the object may be huge - e.g. 674 * the size might be in the gigabytes or terrabytes. 675 */ 676 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 677 vm_pindex_t tscan; 678 int scanlimit; 679 int scanreset; 680 681 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 682 if (scanreset < 16) 683 scanreset = 16; 684 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 685 686 scanlimit = scanreset; 687 tscan = tstart; 688 while (tscan < tend) { 689 curgeneration = object->generation; 690 p = vm_page_lookup(object, tscan); 691 if (p == NULL || p->valid == 0 || 692 (p->queue - p->pc) == PQ_CACHE) { 693 if (--scanlimit == 0) 694 break; 695 ++tscan; 696 continue; 697 } 698 vm_page_test_dirty(p); 699 if ((p->dirty & p->valid) == 0) { 700 if (--scanlimit == 0) 701 break; 702 ++tscan; 703 continue; 704 } 705 /* 706 * If we have been asked to skip nosync pages and 707 * this is a nosync page, we can't continue. 708 */ 709 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 710 if (--scanlimit == 0) 711 break; 712 ++tscan; 713 continue; 714 } 715 scanlimit = scanreset; 716 717 /* 718 * This returns 0 if it was unable to busy the first 719 * page (i.e. had to sleep). 720 */ 721 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 722 } 723 724 /* 725 * If everything was dirty and we flushed it successfully, 726 * and the requested range is not the entire object, we 727 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 728 * return immediately. 729 */ 730 if (tscan >= tend && (tstart || tend < object->size)) { 731 vm_page_unlock_queues(); 732 vm_object_clear_flag(object, OBJ_CLEANING); 733 return; 734 } 735 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 736 } 737 738 /* 739 * Generally set CLEANCHK interlock and make the page read-only so 740 * we can then clear the object flags. 741 * 742 * However, if this is a nosync mmap then the object is likely to 743 * stay dirty so do not mess with the page and do not clear the 744 * object flags. 745 */ 746 clearobjflags = 1; 747 TAILQ_FOREACH(p, &object->memq, listq) { 748 vm_page_flag_set(p, PG_CLEANCHK); 749 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 750 clearobjflags = 0; 751 else 752 pmap_page_protect(p, VM_PROT_READ); 753 } 754 755 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 756 struct vnode *vp; 757 758 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 759 if (object->type == OBJT_VNODE && 760 (vp = (struct vnode *)object->handle) != NULL) { 761 VI_LOCK(vp); 762 if (vp->v_iflag & VI_OBJDIRTY) 763 vp->v_iflag &= ~VI_OBJDIRTY; 764 VI_UNLOCK(vp); 765 } 766 } 767 768 rescan: 769 curgeneration = object->generation; 770 771 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 772 int n; 773 774 np = TAILQ_NEXT(p, listq); 775 776 again: 777 pi = p->pindex; 778 if (((p->flags & PG_CLEANCHK) == 0) || 779 (pi < tstart) || (pi >= tend) || 780 (p->valid == 0) || 781 ((p->queue - p->pc) == PQ_CACHE)) { 782 vm_page_flag_clear(p, PG_CLEANCHK); 783 continue; 784 } 785 786 vm_page_test_dirty(p); 787 if ((p->dirty & p->valid) == 0) { 788 vm_page_flag_clear(p, PG_CLEANCHK); 789 continue; 790 } 791 792 /* 793 * If we have been asked to skip nosync pages and this is a 794 * nosync page, skip it. Note that the object flags were 795 * not cleared in this case so we do not have to set them. 796 */ 797 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 798 vm_page_flag_clear(p, PG_CLEANCHK); 799 continue; 800 } 801 802 n = vm_object_page_collect_flush(object, p, 803 curgeneration, pagerflags); 804 if (n == 0) 805 goto rescan; 806 807 if (object->generation != curgeneration) 808 goto rescan; 809 810 /* 811 * Try to optimize the next page. If we can't we pick up 812 * our (random) scan where we left off. 813 */ 814 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 815 if ((p = vm_page_lookup(object, pi + n)) != NULL) 816 goto again; 817 } 818 } 819 vm_page_unlock_queues(); 820 #if 0 821 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 822 #endif 823 824 vm_object_clear_flag(object, OBJ_CLEANING); 825 return; 826 } 827 828 static int 829 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 830 { 831 int runlen; 832 int s; 833 int maxf; 834 int chkb; 835 int maxb; 836 int i; 837 vm_pindex_t pi; 838 vm_page_t maf[vm_pageout_page_count]; 839 vm_page_t mab[vm_pageout_page_count]; 840 vm_page_t ma[vm_pageout_page_count]; 841 842 s = splvm(); 843 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 844 pi = p->pindex; 845 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 846 vm_page_lock_queues(); 847 if (object->generation != curgeneration) { 848 splx(s); 849 return(0); 850 } 851 } 852 maxf = 0; 853 for(i = 1; i < vm_pageout_page_count; i++) { 854 vm_page_t tp; 855 856 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 857 if ((tp->flags & PG_BUSY) || 858 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 859 (tp->flags & PG_CLEANCHK) == 0) || 860 (tp->busy != 0)) 861 break; 862 if((tp->queue - tp->pc) == PQ_CACHE) { 863 vm_page_flag_clear(tp, PG_CLEANCHK); 864 break; 865 } 866 vm_page_test_dirty(tp); 867 if ((tp->dirty & tp->valid) == 0) { 868 vm_page_flag_clear(tp, PG_CLEANCHK); 869 break; 870 } 871 maf[ i - 1 ] = tp; 872 maxf++; 873 continue; 874 } 875 break; 876 } 877 878 maxb = 0; 879 chkb = vm_pageout_page_count - maxf; 880 if (chkb) { 881 for(i = 1; i < chkb;i++) { 882 vm_page_t tp; 883 884 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 885 if ((tp->flags & PG_BUSY) || 886 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 887 (tp->flags & PG_CLEANCHK) == 0) || 888 (tp->busy != 0)) 889 break; 890 if ((tp->queue - tp->pc) == PQ_CACHE) { 891 vm_page_flag_clear(tp, PG_CLEANCHK); 892 break; 893 } 894 vm_page_test_dirty(tp); 895 if ((tp->dirty & tp->valid) == 0) { 896 vm_page_flag_clear(tp, PG_CLEANCHK); 897 break; 898 } 899 mab[ i - 1 ] = tp; 900 maxb++; 901 continue; 902 } 903 break; 904 } 905 } 906 907 for(i = 0; i < maxb; i++) { 908 int index = (maxb - i) - 1; 909 ma[index] = mab[i]; 910 vm_page_flag_clear(ma[index], PG_CLEANCHK); 911 } 912 vm_page_flag_clear(p, PG_CLEANCHK); 913 ma[maxb] = p; 914 for(i = 0; i < maxf; i++) { 915 int index = (maxb + i) + 1; 916 ma[index] = maf[i]; 917 vm_page_flag_clear(ma[index], PG_CLEANCHK); 918 } 919 runlen = maxb + maxf + 1; 920 921 splx(s); 922 vm_pageout_flush(ma, runlen, pagerflags); 923 for (i = 0; i < runlen; i++) { 924 if (ma[i]->valid & ma[i]->dirty) { 925 pmap_page_protect(ma[i], VM_PROT_READ); 926 vm_page_flag_set(ma[i], PG_CLEANCHK); 927 928 /* 929 * maxf will end up being the actual number of pages 930 * we wrote out contiguously, non-inclusive of the 931 * first page. We do not count look-behind pages. 932 */ 933 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 934 maxf = i - maxb - 1; 935 } 936 } 937 return(maxf + 1); 938 } 939 940 #ifdef ENABLE_VFS_IOOPT 941 /* 942 * Same as vm_object_pmap_copy, except range checking really 943 * works, and is meant for small sections of an object. 944 * 945 * This code protects resident pages by making them read-only 946 * and is typically called on a fork or split when a page 947 * is converted to copy-on-write. 948 * 949 * NOTE: If the page is already at VM_PROT_NONE, calling 950 * pmap_page_protect will have no effect. 951 */ 952 void 953 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 954 { 955 vm_pindex_t idx; 956 vm_page_t p; 957 958 GIANT_REQUIRED; 959 960 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 961 return; 962 vm_page_lock_queues(); 963 for (idx = start; idx < end; idx++) { 964 p = vm_page_lookup(object, idx); 965 if (p == NULL) 966 continue; 967 pmap_page_protect(p, VM_PROT_READ); 968 } 969 vm_page_unlock_queues(); 970 } 971 #endif 972 973 /* 974 * vm_object_madvise: 975 * 976 * Implements the madvise function at the object/page level. 977 * 978 * MADV_WILLNEED (any object) 979 * 980 * Activate the specified pages if they are resident. 981 * 982 * MADV_DONTNEED (any object) 983 * 984 * Deactivate the specified pages if they are resident. 985 * 986 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 987 * OBJ_ONEMAPPING only) 988 * 989 * Deactivate and clean the specified pages if they are 990 * resident. This permits the process to reuse the pages 991 * without faulting or the kernel to reclaim the pages 992 * without I/O. 993 */ 994 void 995 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 996 { 997 vm_pindex_t end, tpindex; 998 vm_object_t tobject; 999 vm_page_t m; 1000 1001 if (object == NULL) 1002 return; 1003 1004 vm_object_lock(object); 1005 1006 end = pindex + count; 1007 1008 /* 1009 * Locate and adjust resident pages 1010 */ 1011 for (; pindex < end; pindex += 1) { 1012 relookup: 1013 tobject = object; 1014 tpindex = pindex; 1015 shadowlookup: 1016 /* 1017 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1018 * and those pages must be OBJ_ONEMAPPING. 1019 */ 1020 if (advise == MADV_FREE) { 1021 if ((tobject->type != OBJT_DEFAULT && 1022 tobject->type != OBJT_SWAP) || 1023 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1024 continue; 1025 } 1026 } 1027 1028 m = vm_page_lookup(tobject, tpindex); 1029 1030 if (m == NULL) { 1031 /* 1032 * There may be swap even if there is no backing page 1033 */ 1034 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1035 swap_pager_freespace(tobject, tpindex, 1); 1036 1037 /* 1038 * next object 1039 */ 1040 tobject = tobject->backing_object; 1041 if (tobject == NULL) 1042 continue; 1043 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1044 goto shadowlookup; 1045 } 1046 1047 /* 1048 * If the page is busy or not in a normal active state, 1049 * we skip it. If the page is not managed there are no 1050 * page queues to mess with. Things can break if we mess 1051 * with pages in any of the below states. 1052 */ 1053 vm_page_lock_queues(); 1054 if (m->hold_count || 1055 m->wire_count || 1056 (m->flags & PG_UNMANAGED) || 1057 m->valid != VM_PAGE_BITS_ALL) { 1058 vm_page_unlock_queues(); 1059 continue; 1060 } 1061 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) 1062 goto relookup; 1063 if (advise == MADV_WILLNEED) { 1064 vm_page_activate(m); 1065 } else if (advise == MADV_DONTNEED) { 1066 vm_page_dontneed(m); 1067 } else if (advise == MADV_FREE) { 1068 /* 1069 * Mark the page clean. This will allow the page 1070 * to be freed up by the system. However, such pages 1071 * are often reused quickly by malloc()/free() 1072 * so we do not do anything that would cause 1073 * a page fault if we can help it. 1074 * 1075 * Specifically, we do not try to actually free 1076 * the page now nor do we try to put it in the 1077 * cache (which would cause a page fault on reuse). 1078 * 1079 * But we do make the page is freeable as we 1080 * can without actually taking the step of unmapping 1081 * it. 1082 */ 1083 pmap_clear_modify(m); 1084 m->dirty = 0; 1085 m->act_count = 0; 1086 vm_page_dontneed(m); 1087 } 1088 vm_page_unlock_queues(); 1089 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1090 swap_pager_freespace(tobject, tpindex, 1); 1091 } 1092 vm_object_unlock(object); 1093 } 1094 1095 /* 1096 * vm_object_shadow: 1097 * 1098 * Create a new object which is backed by the 1099 * specified existing object range. The source 1100 * object reference is deallocated. 1101 * 1102 * The new object and offset into that object 1103 * are returned in the source parameters. 1104 */ 1105 void 1106 vm_object_shadow( 1107 vm_object_t *object, /* IN/OUT */ 1108 vm_ooffset_t *offset, /* IN/OUT */ 1109 vm_size_t length) 1110 { 1111 vm_object_t source; 1112 vm_object_t result; 1113 1114 source = *object; 1115 1116 vm_object_lock(source); 1117 /* 1118 * Don't create the new object if the old object isn't shared. 1119 */ 1120 if (source != NULL && 1121 source->ref_count == 1 && 1122 source->handle == NULL && 1123 (source->type == OBJT_DEFAULT || 1124 source->type == OBJT_SWAP)) { 1125 vm_object_unlock(source); 1126 return; 1127 } 1128 1129 /* 1130 * Allocate a new object with the given length 1131 */ 1132 result = vm_object_allocate(OBJT_DEFAULT, length); 1133 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1134 1135 /* 1136 * The new object shadows the source object, adding a reference to it. 1137 * Our caller changes his reference to point to the new object, 1138 * removing a reference to the source object. Net result: no change 1139 * of reference count. 1140 * 1141 * Try to optimize the result object's page color when shadowing 1142 * in order to maintain page coloring consistency in the combined 1143 * shadowed object. 1144 */ 1145 result->backing_object = source; 1146 if (source) { 1147 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1148 source->shadow_count++; 1149 source->generation++; 1150 if (length < source->size) 1151 length = source->size; 1152 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1153 source->generation > 1) 1154 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1155 result->pg_color = (source->pg_color + 1156 length * source->generation) & PQ_L2_MASK; 1157 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1158 PQ_L2_MASK; 1159 } 1160 1161 /* 1162 * Store the offset into the source object, and fix up the offset into 1163 * the new object. 1164 */ 1165 result->backing_object_offset = *offset; 1166 1167 /* 1168 * Return the new things 1169 */ 1170 *offset = 0; 1171 *object = result; 1172 1173 vm_object_unlock(source); 1174 } 1175 1176 /* 1177 * vm_object_split: 1178 * 1179 * Split the pages in a map entry into a new object. This affords 1180 * easier removal of unused pages, and keeps object inheritance from 1181 * being a negative impact on memory usage. 1182 */ 1183 void 1184 vm_object_split(vm_map_entry_t entry) 1185 { 1186 vm_page_t m; 1187 vm_object_t orig_object, new_object, source; 1188 vm_offset_t s, e; 1189 vm_pindex_t offidxstart, offidxend; 1190 vm_size_t idx, size; 1191 vm_ooffset_t offset; 1192 1193 GIANT_REQUIRED; 1194 1195 orig_object = entry->object.vm_object; 1196 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1197 return; 1198 if (orig_object->ref_count <= 1) 1199 return; 1200 1201 offset = entry->offset; 1202 s = entry->start; 1203 e = entry->end; 1204 1205 offidxstart = OFF_TO_IDX(offset); 1206 offidxend = offidxstart + OFF_TO_IDX(e - s); 1207 size = offidxend - offidxstart; 1208 1209 new_object = vm_pager_allocate(orig_object->type, 1210 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1211 if (new_object == NULL) 1212 return; 1213 1214 source = orig_object->backing_object; 1215 if (source != NULL) { 1216 vm_object_reference(source); /* Referenced by new_object */ 1217 TAILQ_INSERT_TAIL(&source->shadow_head, 1218 new_object, shadow_list); 1219 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1220 new_object->backing_object_offset = 1221 orig_object->backing_object_offset + offset; 1222 new_object->backing_object = source; 1223 source->shadow_count++; 1224 source->generation++; 1225 } 1226 for (idx = 0; idx < size; idx++) { 1227 retry: 1228 m = vm_page_lookup(orig_object, offidxstart + idx); 1229 if (m == NULL) 1230 continue; 1231 1232 /* 1233 * We must wait for pending I/O to complete before we can 1234 * rename the page. 1235 * 1236 * We do not have to VM_PROT_NONE the page as mappings should 1237 * not be changed by this operation. 1238 */ 1239 vm_page_lock_queues(); 1240 if (vm_page_sleep_if_busy(m, TRUE, "spltwt")) 1241 goto retry; 1242 1243 vm_page_busy(m); 1244 vm_page_rename(m, new_object, idx); 1245 /* page automatically made dirty by rename and cache handled */ 1246 vm_page_busy(m); 1247 vm_page_unlock_queues(); 1248 } 1249 if (orig_object->type == OBJT_SWAP) { 1250 vm_object_pip_add(orig_object, 1); 1251 /* 1252 * copy orig_object pages into new_object 1253 * and destroy unneeded pages in 1254 * shadow object. 1255 */ 1256 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1257 vm_object_pip_wakeup(orig_object); 1258 } 1259 vm_page_lock_queues(); 1260 TAILQ_FOREACH(m, &new_object->memq, listq) 1261 vm_page_wakeup(m); 1262 vm_page_unlock_queues(); 1263 entry->object.vm_object = new_object; 1264 entry->offset = 0LL; 1265 vm_object_deallocate(orig_object); 1266 } 1267 1268 #define OBSC_TEST_ALL_SHADOWED 0x0001 1269 #define OBSC_COLLAPSE_NOWAIT 0x0002 1270 #define OBSC_COLLAPSE_WAIT 0x0004 1271 1272 static __inline int 1273 vm_object_backing_scan(vm_object_t object, int op) 1274 { 1275 int s; 1276 int r = 1; 1277 vm_page_t p; 1278 vm_object_t backing_object; 1279 vm_pindex_t backing_offset_index; 1280 1281 s = splvm(); 1282 GIANT_REQUIRED; 1283 1284 backing_object = object->backing_object; 1285 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1286 1287 /* 1288 * Initial conditions 1289 */ 1290 if (op & OBSC_TEST_ALL_SHADOWED) { 1291 /* 1292 * We do not want to have to test for the existence of 1293 * swap pages in the backing object. XXX but with the 1294 * new swapper this would be pretty easy to do. 1295 * 1296 * XXX what about anonymous MAP_SHARED memory that hasn't 1297 * been ZFOD faulted yet? If we do not test for this, the 1298 * shadow test may succeed! XXX 1299 */ 1300 if (backing_object->type != OBJT_DEFAULT) { 1301 splx(s); 1302 return (0); 1303 } 1304 } 1305 if (op & OBSC_COLLAPSE_WAIT) { 1306 vm_object_set_flag(backing_object, OBJ_DEAD); 1307 } 1308 1309 /* 1310 * Our scan 1311 */ 1312 p = TAILQ_FIRST(&backing_object->memq); 1313 while (p) { 1314 vm_page_t next = TAILQ_NEXT(p, listq); 1315 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1316 1317 if (op & OBSC_TEST_ALL_SHADOWED) { 1318 vm_page_t pp; 1319 1320 /* 1321 * Ignore pages outside the parent object's range 1322 * and outside the parent object's mapping of the 1323 * backing object. 1324 * 1325 * note that we do not busy the backing object's 1326 * page. 1327 */ 1328 if ( 1329 p->pindex < backing_offset_index || 1330 new_pindex >= object->size 1331 ) { 1332 p = next; 1333 continue; 1334 } 1335 1336 /* 1337 * See if the parent has the page or if the parent's 1338 * object pager has the page. If the parent has the 1339 * page but the page is not valid, the parent's 1340 * object pager must have the page. 1341 * 1342 * If this fails, the parent does not completely shadow 1343 * the object and we might as well give up now. 1344 */ 1345 1346 pp = vm_page_lookup(object, new_pindex); 1347 if ( 1348 (pp == NULL || pp->valid == 0) && 1349 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1350 ) { 1351 r = 0; 1352 break; 1353 } 1354 } 1355 1356 /* 1357 * Check for busy page 1358 */ 1359 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1360 vm_page_t pp; 1361 1362 vm_page_lock_queues(); 1363 if (op & OBSC_COLLAPSE_NOWAIT) { 1364 if ((p->flags & PG_BUSY) || 1365 !p->valid || 1366 p->hold_count || 1367 p->wire_count || 1368 p->busy) { 1369 vm_page_unlock_queues(); 1370 p = next; 1371 continue; 1372 } 1373 } else if (op & OBSC_COLLAPSE_WAIT) { 1374 if (vm_page_sleep_if_busy(p, TRUE, "vmocol")) { 1375 /* 1376 * If we slept, anything could have 1377 * happened. Since the object is 1378 * marked dead, the backing offset 1379 * should not have changed so we 1380 * just restart our scan. 1381 */ 1382 p = TAILQ_FIRST(&backing_object->memq); 1383 continue; 1384 } 1385 } 1386 1387 /* 1388 * Busy the page 1389 */ 1390 vm_page_busy(p); 1391 vm_page_unlock_queues(); 1392 1393 KASSERT( 1394 p->object == backing_object, 1395 ("vm_object_qcollapse(): object mismatch") 1396 ); 1397 1398 /* 1399 * Destroy any associated swap 1400 */ 1401 if (backing_object->type == OBJT_SWAP) { 1402 swap_pager_freespace( 1403 backing_object, 1404 p->pindex, 1405 1 1406 ); 1407 } 1408 1409 if ( 1410 p->pindex < backing_offset_index || 1411 new_pindex >= object->size 1412 ) { 1413 /* 1414 * Page is out of the parent object's range, we 1415 * can simply destroy it. 1416 */ 1417 vm_page_lock_queues(); 1418 pmap_remove_all(p); 1419 vm_page_free(p); 1420 vm_page_unlock_queues(); 1421 p = next; 1422 continue; 1423 } 1424 1425 pp = vm_page_lookup(object, new_pindex); 1426 if ( 1427 pp != NULL || 1428 vm_pager_has_page(object, new_pindex, NULL, NULL) 1429 ) { 1430 /* 1431 * page already exists in parent OR swap exists 1432 * for this location in the parent. Destroy 1433 * the original page from the backing object. 1434 * 1435 * Leave the parent's page alone 1436 */ 1437 vm_page_lock_queues(); 1438 pmap_remove_all(p); 1439 vm_page_free(p); 1440 vm_page_unlock_queues(); 1441 p = next; 1442 continue; 1443 } 1444 1445 /* 1446 * Page does not exist in parent, rename the 1447 * page from the backing object to the main object. 1448 * 1449 * If the page was mapped to a process, it can remain 1450 * mapped through the rename. 1451 */ 1452 vm_page_lock_queues(); 1453 vm_page_rename(p, object, new_pindex); 1454 vm_page_unlock_queues(); 1455 /* page automatically made dirty by rename */ 1456 } 1457 p = next; 1458 } 1459 splx(s); 1460 return (r); 1461 } 1462 1463 1464 /* 1465 * this version of collapse allows the operation to occur earlier and 1466 * when paging_in_progress is true for an object... This is not a complete 1467 * operation, but should plug 99.9% of the rest of the leaks. 1468 */ 1469 static void 1470 vm_object_qcollapse(vm_object_t object) 1471 { 1472 vm_object_t backing_object = object->backing_object; 1473 1474 GIANT_REQUIRED; 1475 1476 if (backing_object->ref_count != 1) 1477 return; 1478 1479 backing_object->ref_count += 2; 1480 1481 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1482 1483 backing_object->ref_count -= 2; 1484 } 1485 1486 /* 1487 * vm_object_collapse: 1488 * 1489 * Collapse an object with the object backing it. 1490 * Pages in the backing object are moved into the 1491 * parent, and the backing object is deallocated. 1492 */ 1493 void 1494 vm_object_collapse(vm_object_t object) 1495 { 1496 GIANT_REQUIRED; 1497 1498 while (TRUE) { 1499 vm_object_t backing_object; 1500 1501 /* 1502 * Verify that the conditions are right for collapse: 1503 * 1504 * The object exists and the backing object exists. 1505 */ 1506 if (object == NULL) 1507 break; 1508 1509 if ((backing_object = object->backing_object) == NULL) 1510 break; 1511 1512 /* 1513 * we check the backing object first, because it is most likely 1514 * not collapsable. 1515 */ 1516 if (backing_object->handle != NULL || 1517 (backing_object->type != OBJT_DEFAULT && 1518 backing_object->type != OBJT_SWAP) || 1519 (backing_object->flags & OBJ_DEAD) || 1520 object->handle != NULL || 1521 (object->type != OBJT_DEFAULT && 1522 object->type != OBJT_SWAP) || 1523 (object->flags & OBJ_DEAD)) { 1524 break; 1525 } 1526 1527 if ( 1528 object->paging_in_progress != 0 || 1529 backing_object->paging_in_progress != 0 1530 ) { 1531 vm_object_qcollapse(object); 1532 break; 1533 } 1534 1535 /* 1536 * We know that we can either collapse the backing object (if 1537 * the parent is the only reference to it) or (perhaps) have 1538 * the parent bypass the object if the parent happens to shadow 1539 * all the resident pages in the entire backing object. 1540 * 1541 * This is ignoring pager-backed pages such as swap pages. 1542 * vm_object_backing_scan fails the shadowing test in this 1543 * case. 1544 */ 1545 if (backing_object->ref_count == 1) { 1546 /* 1547 * If there is exactly one reference to the backing 1548 * object, we can collapse it into the parent. 1549 */ 1550 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1551 1552 /* 1553 * Move the pager from backing_object to object. 1554 */ 1555 if (backing_object->type == OBJT_SWAP) { 1556 vm_object_pip_add(backing_object, 1); 1557 1558 /* 1559 * scrap the paging_offset junk and do a 1560 * discrete copy. This also removes major 1561 * assumptions about how the swap-pager 1562 * works from where it doesn't belong. The 1563 * new swapper is able to optimize the 1564 * destroy-source case. 1565 */ 1566 vm_object_pip_add(object, 1); 1567 swap_pager_copy( 1568 backing_object, 1569 object, 1570 OFF_TO_IDX(object->backing_object_offset), TRUE); 1571 vm_object_pip_wakeup(object); 1572 1573 vm_object_pip_wakeup(backing_object); 1574 } 1575 /* 1576 * Object now shadows whatever backing_object did. 1577 * Note that the reference to 1578 * backing_object->backing_object moves from within 1579 * backing_object to within object. 1580 */ 1581 TAILQ_REMOVE( 1582 &object->backing_object->shadow_head, 1583 object, 1584 shadow_list 1585 ); 1586 object->backing_object->shadow_count--; 1587 object->backing_object->generation++; 1588 if (backing_object->backing_object) { 1589 TAILQ_REMOVE( 1590 &backing_object->backing_object->shadow_head, 1591 backing_object, 1592 shadow_list 1593 ); 1594 backing_object->backing_object->shadow_count--; 1595 backing_object->backing_object->generation++; 1596 } 1597 object->backing_object = backing_object->backing_object; 1598 if (object->backing_object) { 1599 TAILQ_INSERT_TAIL( 1600 &object->backing_object->shadow_head, 1601 object, 1602 shadow_list 1603 ); 1604 object->backing_object->shadow_count++; 1605 object->backing_object->generation++; 1606 } 1607 1608 object->backing_object_offset += 1609 backing_object->backing_object_offset; 1610 1611 /* 1612 * Discard backing_object. 1613 * 1614 * Since the backing object has no pages, no pager left, 1615 * and no object references within it, all that is 1616 * necessary is to dispose of it. 1617 */ 1618 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1619 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1620 1621 mtx_lock(&vm_object_list_mtx); 1622 TAILQ_REMOVE( 1623 &vm_object_list, 1624 backing_object, 1625 object_list 1626 ); 1627 mtx_unlock(&vm_object_list_mtx); 1628 1629 mtx_destroy(&backing_object->mtx); 1630 1631 uma_zfree(obj_zone, backing_object); 1632 1633 object_collapses++; 1634 } else { 1635 vm_object_t new_backing_object; 1636 1637 /* 1638 * If we do not entirely shadow the backing object, 1639 * there is nothing we can do so we give up. 1640 */ 1641 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1642 break; 1643 } 1644 1645 /* 1646 * Make the parent shadow the next object in the 1647 * chain. Deallocating backing_object will not remove 1648 * it, since its reference count is at least 2. 1649 */ 1650 TAILQ_REMOVE( 1651 &backing_object->shadow_head, 1652 object, 1653 shadow_list 1654 ); 1655 backing_object->shadow_count--; 1656 backing_object->generation++; 1657 1658 new_backing_object = backing_object->backing_object; 1659 if ((object->backing_object = new_backing_object) != NULL) { 1660 vm_object_reference(new_backing_object); 1661 TAILQ_INSERT_TAIL( 1662 &new_backing_object->shadow_head, 1663 object, 1664 shadow_list 1665 ); 1666 new_backing_object->shadow_count++; 1667 new_backing_object->generation++; 1668 object->backing_object_offset += 1669 backing_object->backing_object_offset; 1670 } 1671 1672 /* 1673 * Drop the reference count on backing_object. Since 1674 * its ref_count was at least 2, it will not vanish; 1675 * so we don't need to call vm_object_deallocate, but 1676 * we do anyway. 1677 */ 1678 vm_object_deallocate(backing_object); 1679 object_bypasses++; 1680 } 1681 1682 /* 1683 * Try again with this object's new backing object. 1684 */ 1685 } 1686 } 1687 1688 /* 1689 * vm_object_page_remove: [internal] 1690 * 1691 * Removes all physical pages in the specified 1692 * object range from the object's list of pages. 1693 * 1694 * The object must be locked. 1695 */ 1696 void 1697 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1698 { 1699 vm_page_t p, next; 1700 int all; 1701 1702 if (object == NULL || 1703 object->resident_page_count == 0) 1704 return; 1705 all = ((end == 0) && (start == 0)); 1706 1707 /* 1708 * Since physically-backed objects do not use managed pages, we can't 1709 * remove pages from the object (we must instead remove the page 1710 * references, and then destroy the object). 1711 */ 1712 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1713 1714 vm_object_pip_add(object, 1); 1715 again: 1716 vm_page_lock_queues(); 1717 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1718 if (p->pindex < start) { 1719 p = vm_page_splay(start, object->root); 1720 if ((object->root = p)->pindex < start) 1721 p = TAILQ_NEXT(p, listq); 1722 } 1723 } 1724 /* 1725 * Assert: the variable p is either (1) the page with the 1726 * least pindex greater than or equal to the parameter pindex 1727 * or (2) NULL. 1728 */ 1729 for (; 1730 p != NULL && (all || p->pindex < end); 1731 p = next) { 1732 next = TAILQ_NEXT(p, listq); 1733 1734 if (p->wire_count != 0) { 1735 pmap_remove_all(p); 1736 if (!clean_only) 1737 p->valid = 0; 1738 continue; 1739 } 1740 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1741 goto again; 1742 if (clean_only && p->valid) { 1743 vm_page_test_dirty(p); 1744 if (p->valid & p->dirty) 1745 continue; 1746 } 1747 vm_page_busy(p); 1748 pmap_remove_all(p); 1749 vm_page_free(p); 1750 } 1751 vm_page_unlock_queues(); 1752 vm_object_pip_wakeup(object); 1753 } 1754 1755 /* 1756 * Routine: vm_object_coalesce 1757 * Function: Coalesces two objects backing up adjoining 1758 * regions of memory into a single object. 1759 * 1760 * returns TRUE if objects were combined. 1761 * 1762 * NOTE: Only works at the moment if the second object is NULL - 1763 * if it's not, which object do we lock first? 1764 * 1765 * Parameters: 1766 * prev_object First object to coalesce 1767 * prev_offset Offset into prev_object 1768 * next_object Second object into coalesce 1769 * next_offset Offset into next_object 1770 * 1771 * prev_size Size of reference to prev_object 1772 * next_size Size of reference to next_object 1773 * 1774 * Conditions: 1775 * The object must *not* be locked. 1776 */ 1777 boolean_t 1778 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1779 vm_size_t prev_size, vm_size_t next_size) 1780 { 1781 vm_pindex_t next_pindex; 1782 1783 if (prev_object == NULL) 1784 return (TRUE); 1785 vm_object_lock(prev_object); 1786 if (prev_object->type != OBJT_DEFAULT && 1787 prev_object->type != OBJT_SWAP) { 1788 vm_object_unlock(prev_object); 1789 return (FALSE); 1790 } 1791 1792 /* 1793 * Try to collapse the object first 1794 */ 1795 vm_object_collapse(prev_object); 1796 1797 /* 1798 * Can't coalesce if: . more than one reference . paged out . shadows 1799 * another object . has a copy elsewhere (any of which mean that the 1800 * pages not mapped to prev_entry may be in use anyway) 1801 */ 1802 if (prev_object->backing_object != NULL) { 1803 vm_object_unlock(prev_object); 1804 return (FALSE); 1805 } 1806 1807 prev_size >>= PAGE_SHIFT; 1808 next_size >>= PAGE_SHIFT; 1809 next_pindex = prev_pindex + prev_size; 1810 1811 if ((prev_object->ref_count > 1) && 1812 (prev_object->size != next_pindex)) { 1813 vm_object_unlock(prev_object); 1814 return (FALSE); 1815 } 1816 1817 /* 1818 * Remove any pages that may still be in the object from a previous 1819 * deallocation. 1820 */ 1821 if (next_pindex < prev_object->size) { 1822 vm_object_page_remove(prev_object, 1823 next_pindex, 1824 next_pindex + next_size, FALSE); 1825 if (prev_object->type == OBJT_SWAP) 1826 swap_pager_freespace(prev_object, 1827 next_pindex, next_size); 1828 } 1829 1830 /* 1831 * Extend the object if necessary. 1832 */ 1833 if (next_pindex + next_size > prev_object->size) 1834 prev_object->size = next_pindex + next_size; 1835 1836 vm_object_unlock(prev_object); 1837 return (TRUE); 1838 } 1839 1840 void 1841 vm_object_set_writeable_dirty(vm_object_t object) 1842 { 1843 struct vnode *vp; 1844 1845 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1846 if (object->type == OBJT_VNODE && 1847 (vp = (struct vnode *)object->handle) != NULL) { 1848 VI_LOCK(vp); 1849 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1850 vp->v_iflag |= VI_OBJDIRTY; 1851 VI_UNLOCK(vp); 1852 } 1853 } 1854 1855 #ifdef ENABLE_VFS_IOOPT 1856 /* 1857 * Experimental support for zero-copy I/O 1858 * 1859 * Performs the copy_on_write operations necessary to allow the virtual copies 1860 * into user space to work. This has to be called for write(2) system calls 1861 * from other processes, file unlinking, and file size shrinkage. 1862 */ 1863 void 1864 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 1865 { 1866 int rv; 1867 vm_object_t robject; 1868 vm_pindex_t idx; 1869 1870 GIANT_REQUIRED; 1871 if ((object == NULL) || 1872 ((object->flags & OBJ_OPT) == 0)) 1873 return; 1874 1875 if (object->shadow_count > object->ref_count) 1876 panic("vm_freeze_copyopts: sc > rc"); 1877 1878 while ((robject = TAILQ_FIRST(&object->shadow_head)) != NULL) { 1879 vm_pindex_t bo_pindex; 1880 vm_page_t m_in, m_out; 1881 1882 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 1883 1884 vm_object_reference(robject); 1885 1886 vm_object_pip_wait(robject, "objfrz"); 1887 1888 if (robject->ref_count == 1) { 1889 vm_object_deallocate(robject); 1890 continue; 1891 } 1892 1893 vm_object_pip_add(robject, 1); 1894 1895 for (idx = 0; idx < robject->size; idx++) { 1896 1897 m_out = vm_page_grab(robject, idx, 1898 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1899 1900 if (m_out->valid == 0) { 1901 m_in = vm_page_grab(object, bo_pindex + idx, 1902 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1903 vm_page_lock_queues(); 1904 if (m_in->valid == 0) { 1905 vm_page_unlock_queues(); 1906 rv = vm_pager_get_pages(object, &m_in, 1, 0); 1907 if (rv != VM_PAGER_OK) { 1908 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 1909 continue; 1910 } 1911 vm_page_lock_queues(); 1912 vm_page_deactivate(m_in); 1913 } 1914 1915 pmap_remove_all(m_in); 1916 vm_page_unlock_queues(); 1917 pmap_copy_page(m_in, m_out); 1918 vm_page_lock_queues(); 1919 m_out->valid = m_in->valid; 1920 vm_page_dirty(m_out); 1921 vm_page_activate(m_out); 1922 vm_page_wakeup(m_in); 1923 } else 1924 vm_page_lock_queues(); 1925 vm_page_wakeup(m_out); 1926 vm_page_unlock_queues(); 1927 } 1928 1929 object->shadow_count--; 1930 object->ref_count--; 1931 TAILQ_REMOVE(&object->shadow_head, robject, shadow_list); 1932 robject->backing_object = NULL; 1933 robject->backing_object_offset = 0; 1934 1935 vm_object_pip_wakeup(robject); 1936 vm_object_deallocate(robject); 1937 } 1938 1939 vm_object_clear_flag(object, OBJ_OPT); 1940 } 1941 #endif 1942 1943 #include "opt_ddb.h" 1944 #ifdef DDB 1945 #include <sys/kernel.h> 1946 1947 #include <sys/cons.h> 1948 1949 #include <ddb/ddb.h> 1950 1951 static int 1952 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1953 { 1954 vm_map_t tmpm; 1955 vm_map_entry_t tmpe; 1956 vm_object_t obj; 1957 int entcount; 1958 1959 if (map == 0) 1960 return 0; 1961 1962 if (entry == 0) { 1963 tmpe = map->header.next; 1964 entcount = map->nentries; 1965 while (entcount-- && (tmpe != &map->header)) { 1966 if (_vm_object_in_map(map, object, tmpe)) { 1967 return 1; 1968 } 1969 tmpe = tmpe->next; 1970 } 1971 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1972 tmpm = entry->object.sub_map; 1973 tmpe = tmpm->header.next; 1974 entcount = tmpm->nentries; 1975 while (entcount-- && tmpe != &tmpm->header) { 1976 if (_vm_object_in_map(tmpm, object, tmpe)) { 1977 return 1; 1978 } 1979 tmpe = tmpe->next; 1980 } 1981 } else if ((obj = entry->object.vm_object) != NULL) { 1982 for (; obj; obj = obj->backing_object) 1983 if (obj == object) { 1984 return 1; 1985 } 1986 } 1987 return 0; 1988 } 1989 1990 static int 1991 vm_object_in_map(vm_object_t object) 1992 { 1993 struct proc *p; 1994 1995 /* sx_slock(&allproc_lock); */ 1996 LIST_FOREACH(p, &allproc, p_list) { 1997 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1998 continue; 1999 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2000 /* sx_sunlock(&allproc_lock); */ 2001 return 1; 2002 } 2003 } 2004 /* sx_sunlock(&allproc_lock); */ 2005 if (_vm_object_in_map(kernel_map, object, 0)) 2006 return 1; 2007 if (_vm_object_in_map(kmem_map, object, 0)) 2008 return 1; 2009 if (_vm_object_in_map(pager_map, object, 0)) 2010 return 1; 2011 if (_vm_object_in_map(buffer_map, object, 0)) 2012 return 1; 2013 return 0; 2014 } 2015 2016 DB_SHOW_COMMAND(vmochk, vm_object_check) 2017 { 2018 vm_object_t object; 2019 2020 /* 2021 * make sure that internal objs are in a map somewhere 2022 * and none have zero ref counts. 2023 */ 2024 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2025 if (object->handle == NULL && 2026 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2027 if (object->ref_count == 0) { 2028 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2029 (long)object->size); 2030 } 2031 if (!vm_object_in_map(object)) { 2032 db_printf( 2033 "vmochk: internal obj is not in a map: " 2034 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2035 object->ref_count, (u_long)object->size, 2036 (u_long)object->size, 2037 (void *)object->backing_object); 2038 } 2039 } 2040 } 2041 } 2042 2043 /* 2044 * vm_object_print: [ debug ] 2045 */ 2046 DB_SHOW_COMMAND(object, vm_object_print_static) 2047 { 2048 /* XXX convert args. */ 2049 vm_object_t object = (vm_object_t)addr; 2050 boolean_t full = have_addr; 2051 2052 vm_page_t p; 2053 2054 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2055 #define count was_count 2056 2057 int count; 2058 2059 if (object == NULL) 2060 return; 2061 2062 db_iprintf( 2063 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 2064 object, (int)object->type, (uintmax_t)object->size, 2065 object->resident_page_count, object->ref_count, object->flags); 2066 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2067 object->shadow_count, 2068 object->backing_object ? object->backing_object->ref_count : 0, 2069 object->backing_object, (uintmax_t)object->backing_object_offset); 2070 2071 if (!full) 2072 return; 2073 2074 db_indent += 2; 2075 count = 0; 2076 TAILQ_FOREACH(p, &object->memq, listq) { 2077 if (count == 0) 2078 db_iprintf("memory:="); 2079 else if (count == 6) { 2080 db_printf("\n"); 2081 db_iprintf(" ..."); 2082 count = 0; 2083 } else 2084 db_printf(","); 2085 count++; 2086 2087 db_printf("(off=0x%jx,page=0x%jx)", 2088 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2089 } 2090 if (count != 0) 2091 db_printf("\n"); 2092 db_indent -= 2; 2093 } 2094 2095 /* XXX. */ 2096 #undef count 2097 2098 /* XXX need this non-static entry for calling from vm_map_print. */ 2099 void 2100 vm_object_print( 2101 /* db_expr_t */ long addr, 2102 boolean_t have_addr, 2103 /* db_expr_t */ long count, 2104 char *modif) 2105 { 2106 vm_object_print_static(addr, have_addr, count, modif); 2107 } 2108 2109 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2110 { 2111 vm_object_t object; 2112 int nl = 0; 2113 int c; 2114 2115 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2116 vm_pindex_t idx, fidx; 2117 vm_pindex_t osize; 2118 vm_offset_t pa = -1, padiff; 2119 int rcount; 2120 vm_page_t m; 2121 2122 db_printf("new object: %p\n", (void *)object); 2123 if (nl > 18) { 2124 c = cngetc(); 2125 if (c != ' ') 2126 return; 2127 nl = 0; 2128 } 2129 nl++; 2130 rcount = 0; 2131 fidx = 0; 2132 osize = object->size; 2133 if (osize > 128) 2134 osize = 128; 2135 for (idx = 0; idx < osize; idx++) { 2136 m = vm_page_lookup(object, idx); 2137 if (m == NULL) { 2138 if (rcount) { 2139 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2140 (long)fidx, rcount, (long)pa); 2141 if (nl > 18) { 2142 c = cngetc(); 2143 if (c != ' ') 2144 return; 2145 nl = 0; 2146 } 2147 nl++; 2148 rcount = 0; 2149 } 2150 continue; 2151 } 2152 2153 2154 if (rcount && 2155 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2156 ++rcount; 2157 continue; 2158 } 2159 if (rcount) { 2160 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2161 padiff >>= PAGE_SHIFT; 2162 padiff &= PQ_L2_MASK; 2163 if (padiff == 0) { 2164 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2165 ++rcount; 2166 continue; 2167 } 2168 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2169 (long)fidx, rcount, (long)pa); 2170 db_printf("pd(%ld)\n", (long)padiff); 2171 if (nl > 18) { 2172 c = cngetc(); 2173 if (c != ' ') 2174 return; 2175 nl = 0; 2176 } 2177 nl++; 2178 } 2179 fidx = idx; 2180 pa = VM_PAGE_TO_PHYS(m); 2181 rcount = 1; 2182 } 2183 if (rcount) { 2184 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2185 (long)fidx, rcount, (long)pa); 2186 if (nl > 18) { 2187 c = cngetc(); 2188 if (c != ' ') 2189 return; 2190 nl = 0; 2191 } 2192 nl++; 2193 } 2194 } 2195 } 2196 #endif /* DDB */ 2197