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); 112 113 /* 114 * Virtual memory objects maintain the actual data 115 * associated with allocated virtual memory. A given 116 * page of memory exists within exactly one object. 117 * 118 * An object is only deallocated when all "references" 119 * are given up. Only one "reference" to a given 120 * region of an object should be writeable. 121 * 122 * Associated with each object is a list of all resident 123 * memory pages belonging to that object; this list is 124 * maintained by the "vm_page" module, and locked by the object's 125 * lock. 126 * 127 * Each object also records a "pager" routine which is 128 * used to retrieve (and store) pages to the proper backing 129 * storage. In addition, objects may be backed by other 130 * objects from which they were virtual-copied. 131 * 132 * The only items within the object structure which are 133 * modified after time of creation are: 134 * reference count locked by object's lock 135 * pager routine locked by object's lock 136 * 137 */ 138 139 struct object_q vm_object_list; 140 struct mtx vm_object_list_mtx; /* lock for object list and count */ 141 vm_object_t kernel_object; 142 vm_object_t kmem_object; 143 static struct vm_object kernel_object_store; 144 static struct vm_object kmem_object_store; 145 extern int vm_pageout_page_count; 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; 182 183 /* These are true for any object that has been freed */ 184 object->paging_in_progress = 0; 185 object->resident_page_count = 0; 186 object->shadow_count = 0; 187 } 188 189 void 190 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object) 191 { 192 static int object_hash_rand; 193 int exp, incr; 194 195 TAILQ_INIT(&object->memq); 196 TAILQ_INIT(&object->shadow_head); 197 198 object->type = type; 199 object->size = size; 200 object->ref_count = 1; 201 object->flags = 0; 202 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 203 vm_object_set_flag(object, OBJ_ONEMAPPING); 204 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 205 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 206 else 207 incr = size; 208 do 209 object->pg_color = next_index; 210 while (!atomic_cmpset_int(&next_index, object->pg_color, 211 (object->pg_color + incr) & PQ_L2_MASK)); 212 object->handle = NULL; 213 object->backing_object = NULL; 214 object->backing_object_offset = (vm_ooffset_t) 0; 215 /* 216 * Try to generate a number that will spread objects out in the 217 * hash table. We 'wipe' new objects across the hash in 128 page 218 * increments plus 1 more to offset it a little more by the time 219 * it wraps around. 220 */ 221 do { 222 exp = object_hash_rand; 223 object->hash_rand = exp - 129; 224 } while (!atomic_cmpset_int(&object_hash_rand, exp, object->hash_rand)); 225 226 object->generation++; /* atomicity needed? XXX */ 227 228 mtx_lock(&vm_object_list_mtx); 229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 230 mtx_unlock(&vm_object_list_mtx); 231 } 232 233 /* 234 * vm_object_init: 235 * 236 * Initialize the VM objects module. 237 */ 238 void 239 vm_object_init(void) 240 { 241 TAILQ_INIT(&vm_object_list); 242 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 243 244 kernel_object = &kernel_object_store; 245 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 246 kernel_object); 247 248 kmem_object = &kmem_object_store; 249 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 250 kmem_object); 251 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 252 #ifdef INVARIANTS 253 vm_object_zdtor, 254 #else 255 NULL, 256 #endif 257 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 258 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 259 } 260 261 void 262 vm_object_init2(void) 263 { 264 } 265 266 void 267 vm_object_set_flag(vm_object_t object, u_short bits) 268 { 269 object->flags |= bits; 270 } 271 272 void 273 vm_object_clear_flag(vm_object_t object, u_short bits) 274 { 275 GIANT_REQUIRED; 276 object->flags &= ~bits; 277 } 278 279 void 280 vm_object_pip_add(vm_object_t object, short i) 281 { 282 GIANT_REQUIRED; 283 object->paging_in_progress += i; 284 } 285 286 void 287 vm_object_pip_subtract(vm_object_t object, short i) 288 { 289 GIANT_REQUIRED; 290 object->paging_in_progress -= i; 291 } 292 293 void 294 vm_object_pip_wakeup(vm_object_t object) 295 { 296 GIANT_REQUIRED; 297 object->paging_in_progress--; 298 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 299 vm_object_clear_flag(object, OBJ_PIPWNT); 300 wakeup(object); 301 } 302 } 303 304 void 305 vm_object_pip_wakeupn(vm_object_t object, short i) 306 { 307 GIANT_REQUIRED; 308 if (i) 309 object->paging_in_progress -= i; 310 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 311 vm_object_clear_flag(object, OBJ_PIPWNT); 312 wakeup(object); 313 } 314 } 315 316 void 317 vm_object_pip_sleep(vm_object_t object, char *waitid) 318 { 319 GIANT_REQUIRED; 320 if (object->paging_in_progress) { 321 int s = splvm(); 322 if (object->paging_in_progress) { 323 vm_object_set_flag(object, OBJ_PIPWNT); 324 tsleep(object, PVM, waitid, 0); 325 } 326 splx(s); 327 } 328 } 329 330 void 331 vm_object_pip_wait(vm_object_t object, char *waitid) 332 { 333 GIANT_REQUIRED; 334 while (object->paging_in_progress) 335 vm_object_pip_sleep(object, waitid); 336 } 337 338 /* 339 * vm_object_allocate: 340 * 341 * Returns a new object with the given size. 342 */ 343 vm_object_t 344 vm_object_allocate(objtype_t type, vm_size_t size) 345 { 346 vm_object_t result; 347 348 result = (vm_object_t) uma_zalloc(obj_zone, M_WAITOK); 349 _vm_object_allocate(type, size, result); 350 351 return (result); 352 } 353 354 355 /* 356 * vm_object_reference: 357 * 358 * Gets another reference to the given object. 359 */ 360 void 361 vm_object_reference(vm_object_t object) 362 { 363 GIANT_REQUIRED; 364 365 if (object == NULL) 366 return; 367 368 #if 0 369 /* object can be re-referenced during final cleaning */ 370 KASSERT(!(object->flags & OBJ_DEAD), 371 ("vm_object_reference: attempting to reference dead obj")); 372 #endif 373 374 object->ref_count++; 375 if (object->type == OBJT_VNODE) { 376 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) { 377 printf("vm_object_reference: delay in getting object\n"); 378 } 379 } 380 } 381 382 /* 383 * handle deallocating a object of type OBJT_VNODE 384 */ 385 void 386 vm_object_vndeallocate(vm_object_t object) 387 { 388 struct vnode *vp = (struct vnode *) object->handle; 389 390 GIANT_REQUIRED; 391 KASSERT(object->type == OBJT_VNODE, 392 ("vm_object_vndeallocate: not a vnode object")); 393 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 394 #ifdef INVARIANTS 395 if (object->ref_count == 0) { 396 vprint("vm_object_vndeallocate", vp); 397 panic("vm_object_vndeallocate: bad object reference count"); 398 } 399 #endif 400 401 object->ref_count--; 402 if (object->ref_count == 0) { 403 vp->v_flag &= ~VTEXT; 404 #ifdef ENABLE_VFS_IOOPT 405 vm_object_clear_flag(object, OBJ_OPT); 406 #endif 407 } 408 /* 409 * vrele may need a vop lock 410 */ 411 vrele(vp); 412 } 413 414 /* 415 * vm_object_deallocate: 416 * 417 * Release a reference to the specified object, 418 * gained either through a vm_object_allocate 419 * or a vm_object_reference call. When all references 420 * are gone, storage associated with this object 421 * may be relinquished. 422 * 423 * No object may be locked. 424 */ 425 void 426 vm_object_deallocate(vm_object_t object) 427 { 428 vm_object_t temp; 429 430 GIANT_REQUIRED; 431 432 while (object != NULL) { 433 434 if (object->type == OBJT_VNODE) { 435 vm_object_vndeallocate(object); 436 return; 437 } 438 439 KASSERT(object->ref_count != 0, 440 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 441 442 /* 443 * If the reference count goes to 0 we start calling 444 * vm_object_terminate() on the object chain. 445 * A ref count of 1 may be a special case depending on the 446 * shadow count being 0 or 1. 447 */ 448 object->ref_count--; 449 if (object->ref_count > 1) { 450 return; 451 } else if (object->ref_count == 1) { 452 if (object->shadow_count == 0) { 453 vm_object_set_flag(object, OBJ_ONEMAPPING); 454 } else if ((object->shadow_count == 1) && 455 (object->handle == NULL) && 456 (object->type == OBJT_DEFAULT || 457 object->type == OBJT_SWAP)) { 458 vm_object_t robject; 459 460 robject = TAILQ_FIRST(&object->shadow_head); 461 KASSERT(robject != NULL, 462 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 463 object->ref_count, 464 object->shadow_count)); 465 if ((robject->handle == NULL) && 466 (robject->type == OBJT_DEFAULT || 467 robject->type == OBJT_SWAP)) { 468 469 robject->ref_count++; 470 471 while ( 472 robject->paging_in_progress || 473 object->paging_in_progress 474 ) { 475 vm_object_pip_sleep(robject, "objde1"); 476 vm_object_pip_sleep(object, "objde2"); 477 } 478 479 if (robject->ref_count == 1) { 480 robject->ref_count--; 481 object = robject; 482 goto doterm; 483 } 484 485 object = robject; 486 vm_object_collapse(object); 487 continue; 488 } 489 } 490 491 return; 492 493 } 494 495 doterm: 496 497 temp = object->backing_object; 498 if (temp) { 499 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 500 temp->shadow_count--; 501 #ifdef ENABLE_VFS_IOOPT 502 if (temp->ref_count == 0) 503 vm_object_clear_flag(temp, OBJ_OPT); 504 #endif 505 temp->generation++; 506 object->backing_object = NULL; 507 } 508 /* 509 * Don't double-terminate, we could be in a termination 510 * recursion due to the terminate having to sync data 511 * to disk. 512 */ 513 if ((object->flags & OBJ_DEAD) == 0) 514 vm_object_terminate(object); 515 object = temp; 516 } 517 } 518 519 /* 520 * vm_object_terminate actually destroys the specified object, freeing 521 * up all previously used resources. 522 * 523 * The object must be locked. 524 * This routine may block. 525 */ 526 void 527 vm_object_terminate(vm_object_t object) 528 { 529 vm_page_t p; 530 int s; 531 532 GIANT_REQUIRED; 533 534 /* 535 * Make sure no one uses us. 536 */ 537 vm_object_set_flag(object, OBJ_DEAD); 538 539 /* 540 * wait for the pageout daemon to be done with the object 541 */ 542 vm_object_pip_wait(object, "objtrm"); 543 544 KASSERT(!object->paging_in_progress, 545 ("vm_object_terminate: pageout in progress")); 546 547 /* 548 * Clean and free the pages, as appropriate. All references to the 549 * object are gone, so we don't need to lock it. 550 */ 551 if (object->type == OBJT_VNODE) { 552 struct vnode *vp; 553 554 #ifdef ENABLE_VFS_IOOPT 555 /* 556 * Freeze optimized copies. 557 */ 558 vm_freeze_copyopts(object, 0, object->size); 559 #endif 560 /* 561 * Clean pages and flush buffers. 562 */ 563 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 564 565 vp = (struct vnode *) object->handle; 566 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 567 } 568 569 KASSERT(object->ref_count == 0, 570 ("vm_object_terminate: object with references, ref_count=%d", 571 object->ref_count)); 572 573 /* 574 * Now free any remaining pages. For internal objects, this also 575 * removes them from paging queues. Don't free wired pages, just 576 * remove them from the object. 577 */ 578 s = splvm(); 579 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 580 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 581 ("vm_object_terminate: freeing busy page %p " 582 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 583 if (p->wire_count == 0) { 584 vm_page_busy(p); 585 vm_page_free(p); 586 cnt.v_pfree++; 587 } else { 588 vm_page_busy(p); 589 vm_page_remove(p); 590 } 591 } 592 splx(s); 593 594 /* 595 * Let the pager know object is dead. 596 */ 597 vm_pager_deallocate(object); 598 599 /* 600 * Remove the object from the global object list. 601 */ 602 mtx_lock(&vm_object_list_mtx); 603 TAILQ_REMOVE(&vm_object_list, object, object_list); 604 mtx_unlock(&vm_object_list_mtx); 605 606 wakeup(object); 607 608 /* 609 * Free the space for the object. 610 */ 611 uma_zfree(obj_zone, object); 612 } 613 614 /* 615 * vm_object_page_clean 616 * 617 * Clean all dirty pages in the specified range of object. Leaves page 618 * on whatever queue it is currently on. If NOSYNC is set then do not 619 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 620 * leaving the object dirty. 621 * 622 * Odd semantics: if start == end, we clean everything. 623 * 624 * The object must be locked. 625 */ 626 void 627 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 628 { 629 vm_page_t p, np; 630 vm_offset_t tstart, tend; 631 vm_pindex_t pi; 632 struct vnode *vp; 633 int clearobjflags; 634 int pagerflags; 635 int curgeneration; 636 637 GIANT_REQUIRED; 638 639 if (object->type != OBJT_VNODE || 640 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 641 return; 642 643 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0; 644 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 645 646 vp = object->handle; 647 648 vm_object_set_flag(object, OBJ_CLEANING); 649 650 tstart = start; 651 if (end == 0) { 652 tend = object->size; 653 } else { 654 tend = end; 655 } 656 657 /* 658 * If the caller is smart and only msync()s a range he knows is 659 * dirty, we may be able to avoid an object scan. This results in 660 * a phenominal improvement in performance. We cannot do this 661 * as a matter of course because the object may be huge - e.g. 662 * the size might be in the gigabytes or terrabytes. 663 */ 664 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 665 vm_offset_t tscan; 666 int scanlimit; 667 int scanreset; 668 669 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 670 if (scanreset < 16) 671 scanreset = 16; 672 673 scanlimit = scanreset; 674 tscan = tstart; 675 while (tscan < tend) { 676 curgeneration = object->generation; 677 p = vm_page_lookup(object, tscan); 678 if (p == NULL || p->valid == 0 || 679 (p->queue - p->pc) == PQ_CACHE) { 680 if (--scanlimit == 0) 681 break; 682 ++tscan; 683 continue; 684 } 685 vm_page_test_dirty(p); 686 if ((p->dirty & p->valid) == 0) { 687 if (--scanlimit == 0) 688 break; 689 ++tscan; 690 continue; 691 } 692 /* 693 * If we have been asked to skip nosync pages and 694 * this is a nosync page, we can't continue. 695 */ 696 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 697 if (--scanlimit == 0) 698 break; 699 ++tscan; 700 continue; 701 } 702 scanlimit = scanreset; 703 704 /* 705 * This returns 0 if it was unable to busy the first 706 * page (i.e. had to sleep). 707 */ 708 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 709 } 710 711 /* 712 * If everything was dirty and we flushed it successfully, 713 * and the requested range is not the entire object, we 714 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 715 * return immediately. 716 */ 717 if (tscan >= tend && (tstart || tend < object->size)) { 718 vm_object_clear_flag(object, OBJ_CLEANING); 719 return; 720 } 721 } 722 723 /* 724 * Generally set CLEANCHK interlock and make the page read-only so 725 * we can then clear the object flags. 726 * 727 * However, if this is a nosync mmap then the object is likely to 728 * stay dirty so do not mess with the page and do not clear the 729 * object flags. 730 */ 731 clearobjflags = 1; 732 733 TAILQ_FOREACH(p, &object->memq, listq) { 734 vm_page_flag_set(p, PG_CLEANCHK); 735 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 736 clearobjflags = 0; 737 else 738 vm_page_protect(p, VM_PROT_READ); 739 } 740 741 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 742 struct vnode *vp; 743 744 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 745 if (object->type == OBJT_VNODE && 746 (vp = (struct vnode *)object->handle) != NULL) { 747 if (vp->v_flag & VOBJDIRTY) { 748 mtx_lock(&vp->v_interlock); 749 vp->v_flag &= ~VOBJDIRTY; 750 mtx_unlock(&vp->v_interlock); 751 } 752 } 753 } 754 755 rescan: 756 curgeneration = object->generation; 757 758 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 759 int n; 760 761 np = TAILQ_NEXT(p, listq); 762 763 again: 764 pi = p->pindex; 765 if (((p->flags & PG_CLEANCHK) == 0) || 766 (pi < tstart) || (pi >= tend) || 767 (p->valid == 0) || 768 ((p->queue - p->pc) == PQ_CACHE)) { 769 vm_page_flag_clear(p, PG_CLEANCHK); 770 continue; 771 } 772 773 vm_page_test_dirty(p); 774 if ((p->dirty & p->valid) == 0) { 775 vm_page_flag_clear(p, PG_CLEANCHK); 776 continue; 777 } 778 779 /* 780 * If we have been asked to skip nosync pages and this is a 781 * nosync page, skip it. Note that the object flags were 782 * not cleared in this case so we do not have to set them. 783 */ 784 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 785 vm_page_flag_clear(p, PG_CLEANCHK); 786 continue; 787 } 788 789 n = vm_object_page_collect_flush(object, p, 790 curgeneration, pagerflags); 791 if (n == 0) 792 goto rescan; 793 794 if (object->generation != curgeneration) 795 goto rescan; 796 797 /* 798 * Try to optimize the next page. If we can't we pick up 799 * our (random) scan where we left off. 800 */ 801 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 802 if ((p = vm_page_lookup(object, pi + n)) != NULL) 803 goto again; 804 } 805 } 806 807 #if 0 808 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 809 #endif 810 811 vm_object_clear_flag(object, OBJ_CLEANING); 812 return; 813 } 814 815 static int 816 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 817 { 818 int runlen; 819 int s; 820 int maxf; 821 int chkb; 822 int maxb; 823 int i; 824 vm_pindex_t pi; 825 vm_page_t maf[vm_pageout_page_count]; 826 vm_page_t mab[vm_pageout_page_count]; 827 vm_page_t ma[vm_pageout_page_count]; 828 829 s = splvm(); 830 pi = p->pindex; 831 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 832 if (object->generation != curgeneration) { 833 splx(s); 834 return(0); 835 } 836 } 837 838 maxf = 0; 839 for(i = 1; i < vm_pageout_page_count; i++) { 840 vm_page_t tp; 841 842 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 843 if ((tp->flags & PG_BUSY) || 844 (tp->flags & PG_CLEANCHK) == 0 || 845 (tp->busy != 0)) 846 break; 847 if((tp->queue - tp->pc) == PQ_CACHE) { 848 vm_page_flag_clear(tp, PG_CLEANCHK); 849 break; 850 } 851 vm_page_test_dirty(tp); 852 if ((tp->dirty & tp->valid) == 0) { 853 vm_page_flag_clear(tp, PG_CLEANCHK); 854 break; 855 } 856 maf[ i - 1 ] = tp; 857 maxf++; 858 continue; 859 } 860 break; 861 } 862 863 maxb = 0; 864 chkb = vm_pageout_page_count - maxf; 865 if (chkb) { 866 for(i = 1; i < chkb;i++) { 867 vm_page_t tp; 868 869 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 870 if ((tp->flags & PG_BUSY) || 871 (tp->flags & PG_CLEANCHK) == 0 || 872 (tp->busy != 0)) 873 break; 874 if ((tp->queue - tp->pc) == PQ_CACHE) { 875 vm_page_flag_clear(tp, PG_CLEANCHK); 876 break; 877 } 878 vm_page_test_dirty(tp); 879 if ((tp->dirty & tp->valid) == 0) { 880 vm_page_flag_clear(tp, PG_CLEANCHK); 881 break; 882 } 883 mab[ i - 1 ] = tp; 884 maxb++; 885 continue; 886 } 887 break; 888 } 889 } 890 891 for(i = 0; i < maxb; i++) { 892 int index = (maxb - i) - 1; 893 ma[index] = mab[i]; 894 vm_page_flag_clear(ma[index], PG_CLEANCHK); 895 } 896 vm_page_flag_clear(p, PG_CLEANCHK); 897 ma[maxb] = p; 898 for(i = 0; i < maxf; i++) { 899 int index = (maxb + i) + 1; 900 ma[index] = maf[i]; 901 vm_page_flag_clear(ma[index], PG_CLEANCHK); 902 } 903 runlen = maxb + maxf + 1; 904 905 splx(s); 906 vm_pageout_flush(ma, runlen, pagerflags); 907 for (i = 0; i < runlen; i++) { 908 if (ma[i]->valid & ma[i]->dirty) { 909 vm_page_protect(ma[i], VM_PROT_READ); 910 vm_page_flag_set(ma[i], PG_CLEANCHK); 911 912 /* 913 * maxf will end up being the actual number of pages 914 * we wrote out contiguously, non-inclusive of the 915 * first page. We do not count look-behind pages. 916 */ 917 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 918 maxf = i - maxb - 1; 919 } 920 } 921 return(maxf + 1); 922 } 923 924 /* 925 * Same as vm_object_pmap_copy, except range checking really 926 * works, and is meant for small sections of an object. 927 * 928 * This code protects resident pages by making them read-only 929 * and is typically called on a fork or split when a page 930 * is converted to copy-on-write. 931 * 932 * NOTE: If the page is already at VM_PROT_NONE, calling 933 * vm_page_protect will have no effect. 934 */ 935 void 936 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 937 { 938 vm_pindex_t idx; 939 vm_page_t p; 940 941 GIANT_REQUIRED; 942 943 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 944 return; 945 946 for (idx = start; idx < end; idx++) { 947 p = vm_page_lookup(object, idx); 948 if (p == NULL) 949 continue; 950 vm_page_protect(p, VM_PROT_READ); 951 } 952 } 953 954 /* 955 * vm_object_pmap_remove: 956 * 957 * Removes all physical pages in the specified 958 * object range from all physical maps. 959 * 960 * The object must *not* be locked. 961 */ 962 void 963 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 964 { 965 vm_page_t p; 966 967 GIANT_REQUIRED; 968 if (object == NULL) 969 return; 970 TAILQ_FOREACH(p, &object->memq, listq) { 971 if (p->pindex >= start && p->pindex < end) 972 vm_page_protect(p, VM_PROT_NONE); 973 } 974 if ((start == 0) && (object->size == end)) 975 vm_object_clear_flag(object, OBJ_WRITEABLE); 976 } 977 978 /* 979 * vm_object_madvise: 980 * 981 * Implements the madvise function at the object/page level. 982 * 983 * MADV_WILLNEED (any object) 984 * 985 * Activate the specified pages if they are resident. 986 * 987 * MADV_DONTNEED (any object) 988 * 989 * Deactivate the specified pages if they are resident. 990 * 991 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 992 * OBJ_ONEMAPPING only) 993 * 994 * Deactivate and clean the specified pages if they are 995 * resident. This permits the process to reuse the pages 996 * without faulting or the kernel to reclaim the pages 997 * without I/O. 998 */ 999 void 1000 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1001 { 1002 vm_pindex_t end, tpindex; 1003 vm_object_t tobject; 1004 vm_page_t m; 1005 1006 GIANT_REQUIRED; 1007 if (object == NULL) 1008 return; 1009 1010 end = pindex + count; 1011 1012 /* 1013 * Locate and adjust resident pages 1014 */ 1015 for (; pindex < end; pindex += 1) { 1016 relookup: 1017 tobject = object; 1018 tpindex = pindex; 1019 shadowlookup: 1020 /* 1021 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1022 * and those pages must be OBJ_ONEMAPPING. 1023 */ 1024 if (advise == MADV_FREE) { 1025 if ((tobject->type != OBJT_DEFAULT && 1026 tobject->type != OBJT_SWAP) || 1027 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1028 continue; 1029 } 1030 } 1031 1032 m = vm_page_lookup(tobject, tpindex); 1033 1034 if (m == NULL) { 1035 /* 1036 * There may be swap even if there is no backing page 1037 */ 1038 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1039 swap_pager_freespace(tobject, tpindex, 1); 1040 1041 /* 1042 * next object 1043 */ 1044 tobject = tobject->backing_object; 1045 if (tobject == NULL) 1046 continue; 1047 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1048 goto shadowlookup; 1049 } 1050 1051 /* 1052 * If the page is busy or not in a normal active state, 1053 * we skip it. If the page is not managed there are no 1054 * page queues to mess with. Things can break if we mess 1055 * with pages in any of the below states. 1056 */ 1057 if ( 1058 m->hold_count || 1059 m->wire_count || 1060 (m->flags & PG_UNMANAGED) || 1061 m->valid != VM_PAGE_BITS_ALL 1062 ) { 1063 continue; 1064 } 1065 1066 if (vm_page_sleep_busy(m, TRUE, "madvpo")) 1067 goto relookup; 1068 1069 if (advise == MADV_WILLNEED) { 1070 vm_page_activate(m); 1071 } else if (advise == MADV_DONTNEED) { 1072 vm_page_dontneed(m); 1073 } else if (advise == MADV_FREE) { 1074 /* 1075 * Mark the page clean. This will allow the page 1076 * to be freed up by the system. However, such pages 1077 * are often reused quickly by malloc()/free() 1078 * so we do not do anything that would cause 1079 * a page fault if we can help it. 1080 * 1081 * Specifically, we do not try to actually free 1082 * the page now nor do we try to put it in the 1083 * cache (which would cause a page fault on reuse). 1084 * 1085 * But we do make the page is freeable as we 1086 * can without actually taking the step of unmapping 1087 * it. 1088 */ 1089 pmap_clear_modify(m); 1090 m->dirty = 0; 1091 m->act_count = 0; 1092 vm_page_dontneed(m); 1093 if (tobject->type == OBJT_SWAP) 1094 swap_pager_freespace(tobject, tpindex, 1); 1095 } 1096 } 1097 } 1098 1099 /* 1100 * vm_object_shadow: 1101 * 1102 * Create a new object which is backed by the 1103 * specified existing object range. The source 1104 * object reference is deallocated. 1105 * 1106 * The new object and offset into that object 1107 * are returned in the source parameters. 1108 */ 1109 void 1110 vm_object_shadow( 1111 vm_object_t *object, /* IN/OUT */ 1112 vm_ooffset_t *offset, /* IN/OUT */ 1113 vm_size_t length) 1114 { 1115 vm_object_t source; 1116 vm_object_t result; 1117 1118 GIANT_REQUIRED; 1119 source = *object; 1120 1121 /* 1122 * Don't create the new object if the old object isn't shared. 1123 */ 1124 if (source != NULL && 1125 source->ref_count == 1 && 1126 source->handle == NULL && 1127 (source->type == OBJT_DEFAULT || 1128 source->type == OBJT_SWAP)) 1129 return; 1130 1131 /* 1132 * Allocate a new object with the given length 1133 */ 1134 result = vm_object_allocate(OBJT_DEFAULT, length); 1135 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1136 1137 /* 1138 * The new object shadows the source object, adding a reference to it. 1139 * Our caller changes his reference to point to the new object, 1140 * removing a reference to the source object. Net result: no change 1141 * of reference count. 1142 * 1143 * Try to optimize the result object's page color when shadowing 1144 * in order to maintain page coloring consistency in the combined 1145 * shadowed object. 1146 */ 1147 result->backing_object = source; 1148 if (source) { 1149 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1150 source->shadow_count++; 1151 source->generation++; 1152 if (length < source->size) 1153 length = source->size; 1154 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1155 source->generation > 1) 1156 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1157 result->pg_color = (source->pg_color + 1158 length * source->generation) & PQ_L2_MASK; 1159 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1160 PQ_L2_MASK; 1161 } 1162 1163 /* 1164 * Store the offset into the source object, and fix up the offset into 1165 * the new object. 1166 */ 1167 result->backing_object_offset = *offset; 1168 1169 /* 1170 * Return the new things 1171 */ 1172 *offset = 0; 1173 *object = result; 1174 } 1175 1176 #define OBSC_TEST_ALL_SHADOWED 0x0001 1177 #define OBSC_COLLAPSE_NOWAIT 0x0002 1178 #define OBSC_COLLAPSE_WAIT 0x0004 1179 1180 static __inline int 1181 vm_object_backing_scan(vm_object_t object, int op) 1182 { 1183 int s; 1184 int r = 1; 1185 vm_page_t p; 1186 vm_object_t backing_object; 1187 vm_pindex_t backing_offset_index; 1188 1189 s = splvm(); 1190 GIANT_REQUIRED; 1191 1192 backing_object = object->backing_object; 1193 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1194 1195 /* 1196 * Initial conditions 1197 */ 1198 if (op & OBSC_TEST_ALL_SHADOWED) { 1199 /* 1200 * We do not want to have to test for the existence of 1201 * swap pages in the backing object. XXX but with the 1202 * new swapper this would be pretty easy to do. 1203 * 1204 * XXX what about anonymous MAP_SHARED memory that hasn't 1205 * been ZFOD faulted yet? If we do not test for this, the 1206 * shadow test may succeed! XXX 1207 */ 1208 if (backing_object->type != OBJT_DEFAULT) { 1209 splx(s); 1210 return (0); 1211 } 1212 } 1213 if (op & OBSC_COLLAPSE_WAIT) { 1214 vm_object_set_flag(backing_object, OBJ_DEAD); 1215 } 1216 1217 /* 1218 * Our scan 1219 */ 1220 p = TAILQ_FIRST(&backing_object->memq); 1221 while (p) { 1222 vm_page_t next = TAILQ_NEXT(p, listq); 1223 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1224 1225 if (op & OBSC_TEST_ALL_SHADOWED) { 1226 vm_page_t pp; 1227 1228 /* 1229 * Ignore pages outside the parent object's range 1230 * and outside the parent object's mapping of the 1231 * backing object. 1232 * 1233 * note that we do not busy the backing object's 1234 * page. 1235 */ 1236 if ( 1237 p->pindex < backing_offset_index || 1238 new_pindex >= object->size 1239 ) { 1240 p = next; 1241 continue; 1242 } 1243 1244 /* 1245 * See if the parent has the page or if the parent's 1246 * object pager has the page. If the parent has the 1247 * page but the page is not valid, the parent's 1248 * object pager must have the page. 1249 * 1250 * If this fails, the parent does not completely shadow 1251 * the object and we might as well give up now. 1252 */ 1253 1254 pp = vm_page_lookup(object, new_pindex); 1255 if ( 1256 (pp == NULL || pp->valid == 0) && 1257 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1258 ) { 1259 r = 0; 1260 break; 1261 } 1262 } 1263 1264 /* 1265 * Check for busy page 1266 */ 1267 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1268 vm_page_t pp; 1269 1270 if (op & OBSC_COLLAPSE_NOWAIT) { 1271 if ( 1272 (p->flags & PG_BUSY) || 1273 !p->valid || 1274 p->hold_count || 1275 p->wire_count || 1276 p->busy 1277 ) { 1278 p = next; 1279 continue; 1280 } 1281 } else if (op & OBSC_COLLAPSE_WAIT) { 1282 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1283 /* 1284 * If we slept, anything could have 1285 * happened. Since the object is 1286 * marked dead, the backing offset 1287 * should not have changed so we 1288 * just restart our scan. 1289 */ 1290 p = TAILQ_FIRST(&backing_object->memq); 1291 continue; 1292 } 1293 } 1294 1295 /* 1296 * Busy the page 1297 */ 1298 vm_page_busy(p); 1299 1300 KASSERT( 1301 p->object == backing_object, 1302 ("vm_object_qcollapse(): object mismatch") 1303 ); 1304 1305 /* 1306 * Destroy any associated swap 1307 */ 1308 if (backing_object->type == OBJT_SWAP) { 1309 swap_pager_freespace( 1310 backing_object, 1311 p->pindex, 1312 1 1313 ); 1314 } 1315 1316 if ( 1317 p->pindex < backing_offset_index || 1318 new_pindex >= object->size 1319 ) { 1320 /* 1321 * Page is out of the parent object's range, we 1322 * can simply destroy it. 1323 */ 1324 vm_page_protect(p, VM_PROT_NONE); 1325 vm_page_free(p); 1326 p = next; 1327 continue; 1328 } 1329 1330 pp = vm_page_lookup(object, new_pindex); 1331 if ( 1332 pp != NULL || 1333 vm_pager_has_page(object, new_pindex, NULL, NULL) 1334 ) { 1335 /* 1336 * page already exists in parent OR swap exists 1337 * for this location in the parent. Destroy 1338 * the original page from the backing object. 1339 * 1340 * Leave the parent's page alone 1341 */ 1342 vm_page_protect(p, VM_PROT_NONE); 1343 vm_page_free(p); 1344 p = next; 1345 continue; 1346 } 1347 1348 /* 1349 * Page does not exist in parent, rename the 1350 * page from the backing object to the main object. 1351 * 1352 * If the page was mapped to a process, it can remain 1353 * mapped through the rename. 1354 */ 1355 if ((p->queue - p->pc) == PQ_CACHE) 1356 vm_page_deactivate(p); 1357 1358 vm_page_rename(p, object, new_pindex); 1359 /* page automatically made dirty by rename */ 1360 } 1361 p = next; 1362 } 1363 splx(s); 1364 return (r); 1365 } 1366 1367 1368 /* 1369 * this version of collapse allows the operation to occur earlier and 1370 * when paging_in_progress is true for an object... This is not a complete 1371 * operation, but should plug 99.9% of the rest of the leaks. 1372 */ 1373 static void 1374 vm_object_qcollapse(vm_object_t object) 1375 { 1376 vm_object_t backing_object = object->backing_object; 1377 1378 GIANT_REQUIRED; 1379 1380 if (backing_object->ref_count != 1) 1381 return; 1382 1383 backing_object->ref_count += 2; 1384 1385 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1386 1387 backing_object->ref_count -= 2; 1388 } 1389 1390 /* 1391 * vm_object_collapse: 1392 * 1393 * Collapse an object with the object backing it. 1394 * Pages in the backing object are moved into the 1395 * parent, and the backing object is deallocated. 1396 */ 1397 void 1398 vm_object_collapse(vm_object_t object) 1399 { 1400 GIANT_REQUIRED; 1401 1402 while (TRUE) { 1403 vm_object_t backing_object; 1404 1405 /* 1406 * Verify that the conditions are right for collapse: 1407 * 1408 * The object exists and the backing object exists. 1409 */ 1410 if (object == NULL) 1411 break; 1412 1413 if ((backing_object = object->backing_object) == NULL) 1414 break; 1415 1416 /* 1417 * we check the backing object first, because it is most likely 1418 * not collapsable. 1419 */ 1420 if (backing_object->handle != NULL || 1421 (backing_object->type != OBJT_DEFAULT && 1422 backing_object->type != OBJT_SWAP) || 1423 (backing_object->flags & OBJ_DEAD) || 1424 object->handle != NULL || 1425 (object->type != OBJT_DEFAULT && 1426 object->type != OBJT_SWAP) || 1427 (object->flags & OBJ_DEAD)) { 1428 break; 1429 } 1430 1431 if ( 1432 object->paging_in_progress != 0 || 1433 backing_object->paging_in_progress != 0 1434 ) { 1435 vm_object_qcollapse(object); 1436 break; 1437 } 1438 1439 /* 1440 * We know that we can either collapse the backing object (if 1441 * the parent is the only reference to it) or (perhaps) have 1442 * the parent bypass the object if the parent happens to shadow 1443 * all the resident pages in the entire backing object. 1444 * 1445 * This is ignoring pager-backed pages such as swap pages. 1446 * vm_object_backing_scan fails the shadowing test in this 1447 * case. 1448 */ 1449 if (backing_object->ref_count == 1) { 1450 /* 1451 * If there is exactly one reference to the backing 1452 * object, we can collapse it into the parent. 1453 */ 1454 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1455 1456 /* 1457 * Move the pager from backing_object to object. 1458 */ 1459 if (backing_object->type == OBJT_SWAP) { 1460 vm_object_pip_add(backing_object, 1); 1461 1462 /* 1463 * scrap the paging_offset junk and do a 1464 * discrete copy. This also removes major 1465 * assumptions about how the swap-pager 1466 * works from where it doesn't belong. The 1467 * new swapper is able to optimize the 1468 * destroy-source case. 1469 */ 1470 vm_object_pip_add(object, 1); 1471 swap_pager_copy( 1472 backing_object, 1473 object, 1474 OFF_TO_IDX(object->backing_object_offset), TRUE); 1475 vm_object_pip_wakeup(object); 1476 1477 vm_object_pip_wakeup(backing_object); 1478 } 1479 /* 1480 * Object now shadows whatever backing_object did. 1481 * Note that the reference to 1482 * backing_object->backing_object moves from within 1483 * backing_object to within object. 1484 */ 1485 TAILQ_REMOVE( 1486 &object->backing_object->shadow_head, 1487 object, 1488 shadow_list 1489 ); 1490 object->backing_object->shadow_count--; 1491 object->backing_object->generation++; 1492 if (backing_object->backing_object) { 1493 TAILQ_REMOVE( 1494 &backing_object->backing_object->shadow_head, 1495 backing_object, 1496 shadow_list 1497 ); 1498 backing_object->backing_object->shadow_count--; 1499 backing_object->backing_object->generation++; 1500 } 1501 object->backing_object = backing_object->backing_object; 1502 if (object->backing_object) { 1503 TAILQ_INSERT_TAIL( 1504 &object->backing_object->shadow_head, 1505 object, 1506 shadow_list 1507 ); 1508 object->backing_object->shadow_count++; 1509 object->backing_object->generation++; 1510 } 1511 1512 object->backing_object_offset += 1513 backing_object->backing_object_offset; 1514 1515 /* 1516 * Discard backing_object. 1517 * 1518 * Since the backing object has no pages, no pager left, 1519 * and no object references within it, all that is 1520 * necessary is to dispose of it. 1521 */ 1522 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1523 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1524 1525 mtx_lock(&vm_object_list_mtx); 1526 TAILQ_REMOVE( 1527 &vm_object_list, 1528 backing_object, 1529 object_list 1530 ); 1531 mtx_unlock(&vm_object_list_mtx); 1532 1533 uma_zfree(obj_zone, backing_object); 1534 1535 object_collapses++; 1536 } else { 1537 vm_object_t new_backing_object; 1538 1539 /* 1540 * If we do not entirely shadow the backing object, 1541 * there is nothing we can do so we give up. 1542 */ 1543 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1544 break; 1545 } 1546 1547 /* 1548 * Make the parent shadow the next object in the 1549 * chain. Deallocating backing_object will not remove 1550 * it, since its reference count is at least 2. 1551 */ 1552 TAILQ_REMOVE( 1553 &backing_object->shadow_head, 1554 object, 1555 shadow_list 1556 ); 1557 backing_object->shadow_count--; 1558 backing_object->generation++; 1559 1560 new_backing_object = backing_object->backing_object; 1561 if ((object->backing_object = new_backing_object) != NULL) { 1562 vm_object_reference(new_backing_object); 1563 TAILQ_INSERT_TAIL( 1564 &new_backing_object->shadow_head, 1565 object, 1566 shadow_list 1567 ); 1568 new_backing_object->shadow_count++; 1569 new_backing_object->generation++; 1570 object->backing_object_offset += 1571 backing_object->backing_object_offset; 1572 } 1573 1574 /* 1575 * Drop the reference count on backing_object. Since 1576 * its ref_count was at least 2, it will not vanish; 1577 * so we don't need to call vm_object_deallocate, but 1578 * we do anyway. 1579 */ 1580 vm_object_deallocate(backing_object); 1581 object_bypasses++; 1582 } 1583 1584 /* 1585 * Try again with this object's new backing object. 1586 */ 1587 } 1588 } 1589 1590 /* 1591 * vm_object_page_remove: [internal] 1592 * 1593 * Removes all physical pages in the specified 1594 * object range from the object's list of pages. 1595 * 1596 * The object must be locked. 1597 */ 1598 void 1599 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1600 { 1601 vm_page_t p, next; 1602 unsigned int size; 1603 int all; 1604 1605 GIANT_REQUIRED; 1606 1607 if (object == NULL || 1608 object->resident_page_count == 0) 1609 return; 1610 1611 all = ((end == 0) && (start == 0)); 1612 1613 /* 1614 * Since physically-backed objects do not use managed pages, we can't 1615 * remove pages from the object (we must instead remove the page 1616 * references, and then destroy the object). 1617 */ 1618 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1619 1620 vm_object_pip_add(object, 1); 1621 again: 1622 size = end - start; 1623 if (all || size > object->resident_page_count / 4) { 1624 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) { 1625 next = TAILQ_NEXT(p, listq); 1626 if (all || ((start <= p->pindex) && (p->pindex < end))) { 1627 if (p->wire_count != 0) { 1628 vm_page_protect(p, VM_PROT_NONE); 1629 if (!clean_only) 1630 p->valid = 0; 1631 continue; 1632 } 1633 1634 /* 1635 * The busy flags are only cleared at 1636 * interrupt -- minimize the spl transitions 1637 */ 1638 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1639 goto again; 1640 1641 if (clean_only && p->valid) { 1642 vm_page_test_dirty(p); 1643 if (p->valid & p->dirty) 1644 continue; 1645 } 1646 1647 vm_page_busy(p); 1648 vm_page_protect(p, VM_PROT_NONE); 1649 vm_page_free(p); 1650 } 1651 } 1652 } else { 1653 while (size > 0) { 1654 if ((p = vm_page_lookup(object, start)) != 0) { 1655 1656 if (p->wire_count != 0) { 1657 vm_page_protect(p, VM_PROT_NONE); 1658 if (!clean_only) 1659 p->valid = 0; 1660 start += 1; 1661 size -= 1; 1662 continue; 1663 } 1664 1665 /* 1666 * The busy flags are only cleared at 1667 * interrupt -- minimize the spl transitions 1668 */ 1669 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1670 goto again; 1671 1672 if (clean_only && p->valid) { 1673 vm_page_test_dirty(p); 1674 if (p->valid & p->dirty) { 1675 start += 1; 1676 size -= 1; 1677 continue; 1678 } 1679 } 1680 1681 vm_page_busy(p); 1682 vm_page_protect(p, VM_PROT_NONE); 1683 vm_page_free(p); 1684 } 1685 start += 1; 1686 size -= 1; 1687 } 1688 } 1689 vm_object_pip_wakeup(object); 1690 } 1691 1692 /* 1693 * Routine: vm_object_coalesce 1694 * Function: Coalesces two objects backing up adjoining 1695 * regions of memory into a single object. 1696 * 1697 * returns TRUE if objects were combined. 1698 * 1699 * NOTE: Only works at the moment if the second object is NULL - 1700 * if it's not, which object do we lock first? 1701 * 1702 * Parameters: 1703 * prev_object First object to coalesce 1704 * prev_offset Offset into prev_object 1705 * next_object Second object into coalesce 1706 * next_offset Offset into next_object 1707 * 1708 * prev_size Size of reference to prev_object 1709 * next_size Size of reference to next_object 1710 * 1711 * Conditions: 1712 * The object must *not* be locked. 1713 */ 1714 boolean_t 1715 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size) 1716 { 1717 vm_pindex_t next_pindex; 1718 1719 GIANT_REQUIRED; 1720 1721 if (prev_object == NULL) { 1722 return (TRUE); 1723 } 1724 1725 if (prev_object->type != OBJT_DEFAULT && 1726 prev_object->type != OBJT_SWAP) { 1727 return (FALSE); 1728 } 1729 1730 /* 1731 * Try to collapse the object first 1732 */ 1733 vm_object_collapse(prev_object); 1734 1735 /* 1736 * Can't coalesce if: . more than one reference . paged out . shadows 1737 * another object . has a copy elsewhere (any of which mean that the 1738 * pages not mapped to prev_entry may be in use anyway) 1739 */ 1740 if (prev_object->backing_object != NULL) { 1741 return (FALSE); 1742 } 1743 1744 prev_size >>= PAGE_SHIFT; 1745 next_size >>= PAGE_SHIFT; 1746 next_pindex = prev_pindex + prev_size; 1747 1748 if ((prev_object->ref_count > 1) && 1749 (prev_object->size != next_pindex)) { 1750 return (FALSE); 1751 } 1752 1753 /* 1754 * Remove any pages that may still be in the object from a previous 1755 * deallocation. 1756 */ 1757 if (next_pindex < prev_object->size) { 1758 vm_object_page_remove(prev_object, 1759 next_pindex, 1760 next_pindex + next_size, FALSE); 1761 if (prev_object->type == OBJT_SWAP) 1762 swap_pager_freespace(prev_object, 1763 next_pindex, next_size); 1764 } 1765 1766 /* 1767 * Extend the object if necessary. 1768 */ 1769 if (next_pindex + next_size > prev_object->size) 1770 prev_object->size = next_pindex + next_size; 1771 1772 return (TRUE); 1773 } 1774 1775 void 1776 vm_object_set_writeable_dirty(vm_object_t object) 1777 { 1778 struct vnode *vp; 1779 1780 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1781 if (object->type == OBJT_VNODE && 1782 (vp = (struct vnode *)object->handle) != NULL) { 1783 if ((vp->v_flag & VOBJDIRTY) == 0) { 1784 mtx_lock(&vp->v_interlock); 1785 vp->v_flag |= VOBJDIRTY; 1786 mtx_unlock(&vp->v_interlock); 1787 } 1788 } 1789 } 1790 1791 #ifdef ENABLE_VFS_IOOPT 1792 /* 1793 * Experimental support for zero-copy I/O 1794 * 1795 * Performs the copy_on_write operations necessary to allow the virtual copies 1796 * into user space to work. This has to be called for write(2) system calls 1797 * from other processes, file unlinking, and file size shrinkage. 1798 */ 1799 void 1800 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 1801 { 1802 int rv; 1803 vm_object_t robject; 1804 vm_pindex_t idx; 1805 1806 GIANT_REQUIRED; 1807 if ((object == NULL) || 1808 ((object->flags & OBJ_OPT) == 0)) 1809 return; 1810 1811 if (object->shadow_count > object->ref_count) 1812 panic("vm_freeze_copyopts: sc > rc"); 1813 1814 while ((robject = TAILQ_FIRST(&object->shadow_head)) != NULL) { 1815 vm_pindex_t bo_pindex; 1816 vm_page_t m_in, m_out; 1817 1818 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 1819 1820 vm_object_reference(robject); 1821 1822 vm_object_pip_wait(robject, "objfrz"); 1823 1824 if (robject->ref_count == 1) { 1825 vm_object_deallocate(robject); 1826 continue; 1827 } 1828 1829 vm_object_pip_add(robject, 1); 1830 1831 for (idx = 0; idx < robject->size; idx++) { 1832 1833 m_out = vm_page_grab(robject, idx, 1834 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1835 1836 if (m_out->valid == 0) { 1837 m_in = vm_page_grab(object, bo_pindex + idx, 1838 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1839 if (m_in->valid == 0) { 1840 rv = vm_pager_get_pages(object, &m_in, 1, 0); 1841 if (rv != VM_PAGER_OK) { 1842 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 1843 continue; 1844 } 1845 vm_page_deactivate(m_in); 1846 } 1847 1848 vm_page_protect(m_in, VM_PROT_NONE); 1849 pmap_copy_page(m_in, m_out); 1850 m_out->valid = m_in->valid; 1851 vm_page_dirty(m_out); 1852 vm_page_activate(m_out); 1853 vm_page_wakeup(m_in); 1854 } 1855 vm_page_wakeup(m_out); 1856 } 1857 1858 object->shadow_count--; 1859 object->ref_count--; 1860 TAILQ_REMOVE(&object->shadow_head, robject, shadow_list); 1861 robject->backing_object = NULL; 1862 robject->backing_object_offset = 0; 1863 1864 vm_object_pip_wakeup(robject); 1865 vm_object_deallocate(robject); 1866 } 1867 1868 vm_object_clear_flag(object, OBJ_OPT); 1869 } 1870 #endif 1871 1872 #include "opt_ddb.h" 1873 #ifdef DDB 1874 #include <sys/kernel.h> 1875 1876 #include <sys/cons.h> 1877 1878 #include <ddb/ddb.h> 1879 1880 static int 1881 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1882 { 1883 vm_map_t tmpm; 1884 vm_map_entry_t tmpe; 1885 vm_object_t obj; 1886 int entcount; 1887 1888 if (map == 0) 1889 return 0; 1890 1891 if (entry == 0) { 1892 tmpe = map->header.next; 1893 entcount = map->nentries; 1894 while (entcount-- && (tmpe != &map->header)) { 1895 if (_vm_object_in_map(map, object, tmpe)) { 1896 return 1; 1897 } 1898 tmpe = tmpe->next; 1899 } 1900 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1901 tmpm = entry->object.sub_map; 1902 tmpe = tmpm->header.next; 1903 entcount = tmpm->nentries; 1904 while (entcount-- && tmpe != &tmpm->header) { 1905 if (_vm_object_in_map(tmpm, object, tmpe)) { 1906 return 1; 1907 } 1908 tmpe = tmpe->next; 1909 } 1910 } else if ((obj = entry->object.vm_object) != NULL) { 1911 for (; obj; obj = obj->backing_object) 1912 if (obj == object) { 1913 return 1; 1914 } 1915 } 1916 return 0; 1917 } 1918 1919 static int 1920 vm_object_in_map(vm_object_t object) 1921 { 1922 struct proc *p; 1923 1924 /* sx_slock(&allproc_lock); */ 1925 LIST_FOREACH(p, &allproc, p_list) { 1926 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1927 continue; 1928 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1929 /* sx_sunlock(&allproc_lock); */ 1930 return 1; 1931 } 1932 } 1933 /* sx_sunlock(&allproc_lock); */ 1934 if (_vm_object_in_map(kernel_map, object, 0)) 1935 return 1; 1936 if (_vm_object_in_map(kmem_map, object, 0)) 1937 return 1; 1938 if (_vm_object_in_map(pager_map, object, 0)) 1939 return 1; 1940 if (_vm_object_in_map(buffer_map, object, 0)) 1941 return 1; 1942 return 0; 1943 } 1944 1945 DB_SHOW_COMMAND(vmochk, vm_object_check) 1946 { 1947 vm_object_t object; 1948 1949 /* 1950 * make sure that internal objs are in a map somewhere 1951 * and none have zero ref counts. 1952 */ 1953 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1954 if (object->handle == NULL && 1955 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1956 if (object->ref_count == 0) { 1957 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1958 (long)object->size); 1959 } 1960 if (!vm_object_in_map(object)) { 1961 db_printf( 1962 "vmochk: internal obj is not in a map: " 1963 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1964 object->ref_count, (u_long)object->size, 1965 (u_long)object->size, 1966 (void *)object->backing_object); 1967 } 1968 } 1969 } 1970 } 1971 1972 /* 1973 * vm_object_print: [ debug ] 1974 */ 1975 DB_SHOW_COMMAND(object, vm_object_print_static) 1976 { 1977 /* XXX convert args. */ 1978 vm_object_t object = (vm_object_t)addr; 1979 boolean_t full = have_addr; 1980 1981 vm_page_t p; 1982 1983 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1984 #define count was_count 1985 1986 int count; 1987 1988 if (object == NULL) 1989 return; 1990 1991 db_iprintf( 1992 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 1993 object, (int)object->type, (u_long)object->size, 1994 object->resident_page_count, object->ref_count, object->flags); 1995 /* 1996 * XXX no %qd in kernel. Truncate object->backing_object_offset. 1997 */ 1998 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 1999 object->shadow_count, 2000 object->backing_object ? object->backing_object->ref_count : 0, 2001 object->backing_object, (long)object->backing_object_offset); 2002 2003 if (!full) 2004 return; 2005 2006 db_indent += 2; 2007 count = 0; 2008 TAILQ_FOREACH(p, &object->memq, listq) { 2009 if (count == 0) 2010 db_iprintf("memory:="); 2011 else if (count == 6) { 2012 db_printf("\n"); 2013 db_iprintf(" ..."); 2014 count = 0; 2015 } else 2016 db_printf(","); 2017 count++; 2018 2019 db_printf("(off=0x%lx,page=0x%lx)", 2020 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2021 } 2022 if (count != 0) 2023 db_printf("\n"); 2024 db_indent -= 2; 2025 } 2026 2027 /* XXX. */ 2028 #undef count 2029 2030 /* XXX need this non-static entry for calling from vm_map_print. */ 2031 void 2032 vm_object_print( 2033 /* db_expr_t */ long addr, 2034 boolean_t have_addr, 2035 /* db_expr_t */ long count, 2036 char *modif) 2037 { 2038 vm_object_print_static(addr, have_addr, count, modif); 2039 } 2040 2041 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2042 { 2043 vm_object_t object; 2044 int nl = 0; 2045 int c; 2046 2047 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2048 vm_pindex_t idx, fidx; 2049 vm_pindex_t osize; 2050 vm_offset_t pa = -1, padiff; 2051 int rcount; 2052 vm_page_t m; 2053 2054 db_printf("new object: %p\n", (void *)object); 2055 if (nl > 18) { 2056 c = cngetc(); 2057 if (c != ' ') 2058 return; 2059 nl = 0; 2060 } 2061 nl++; 2062 rcount = 0; 2063 fidx = 0; 2064 osize = object->size; 2065 if (osize > 128) 2066 osize = 128; 2067 for (idx = 0; idx < osize; idx++) { 2068 m = vm_page_lookup(object, idx); 2069 if (m == NULL) { 2070 if (rcount) { 2071 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2072 (long)fidx, rcount, (long)pa); 2073 if (nl > 18) { 2074 c = cngetc(); 2075 if (c != ' ') 2076 return; 2077 nl = 0; 2078 } 2079 nl++; 2080 rcount = 0; 2081 } 2082 continue; 2083 } 2084 2085 2086 if (rcount && 2087 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2088 ++rcount; 2089 continue; 2090 } 2091 if (rcount) { 2092 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2093 padiff >>= PAGE_SHIFT; 2094 padiff &= PQ_L2_MASK; 2095 if (padiff == 0) { 2096 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2097 ++rcount; 2098 continue; 2099 } 2100 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2101 (long)fidx, rcount, (long)pa); 2102 db_printf("pd(%ld)\n", (long)padiff); 2103 if (nl > 18) { 2104 c = cngetc(); 2105 if (c != ' ') 2106 return; 2107 nl = 0; 2108 } 2109 nl++; 2110 } 2111 fidx = idx; 2112 pa = VM_PAGE_TO_PHYS(m); 2113 rcount = 1; 2114 } 2115 if (rcount) { 2116 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2117 (long)fidx, rcount, (long)pa); 2118 if (nl > 18) { 2119 c = cngetc(); 2120 if (c != ' ') 2121 return; 2122 nl = 0; 2123 } 2124 nl++; 2125 } 2126 } 2127 } 2128 #endif /* DDB */ 2129