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_pindex_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->root = NULL; 199 object->type = type; 200 object->size = size; 201 object->ref_count = 1; 202 object->flags = 0; 203 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 204 vm_object_set_flag(object, OBJ_ONEMAPPING); 205 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 206 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 207 else 208 incr = size; 209 do 210 object->pg_color = next_index; 211 while (!atomic_cmpset_int(&next_index, object->pg_color, 212 (object->pg_color + incr) & PQ_L2_MASK)); 213 object->handle = NULL; 214 object->backing_object = NULL; 215 object->backing_object_offset = (vm_ooffset_t) 0; 216 /* 217 * Try to generate a number that will spread objects out in the 218 * hash table. We 'wipe' new objects across the hash in 128 page 219 * increments plus 1 more to offset it a little more by the time 220 * it wraps around. 221 */ 222 do { 223 exp = object_hash_rand; 224 object->hash_rand = exp - 129; 225 } while (!atomic_cmpset_int(&object_hash_rand, exp, object->hash_rand)); 226 227 object->generation++; /* atomicity needed? XXX */ 228 229 mtx_lock(&vm_object_list_mtx); 230 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 231 mtx_unlock(&vm_object_list_mtx); 232 } 233 234 /* 235 * vm_object_init: 236 * 237 * Initialize the VM objects module. 238 */ 239 void 240 vm_object_init(void) 241 { 242 TAILQ_INIT(&vm_object_list); 243 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 244 245 kernel_object = &kernel_object_store; 246 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 247 kernel_object); 248 249 kmem_object = &kmem_object_store; 250 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 251 kmem_object); 252 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 253 #ifdef INVARIANTS 254 vm_object_zdtor, 255 #else 256 NULL, 257 #endif 258 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 259 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 260 } 261 262 void 263 vm_object_init2(void) 264 { 265 } 266 267 void 268 vm_object_set_flag(vm_object_t object, u_short bits) 269 { 270 object->flags |= bits; 271 } 272 273 void 274 vm_object_clear_flag(vm_object_t object, u_short bits) 275 { 276 GIANT_REQUIRED; 277 object->flags &= ~bits; 278 } 279 280 void 281 vm_object_pip_add(vm_object_t object, short i) 282 { 283 GIANT_REQUIRED; 284 object->paging_in_progress += i; 285 } 286 287 void 288 vm_object_pip_subtract(vm_object_t object, short i) 289 { 290 GIANT_REQUIRED; 291 object->paging_in_progress -= i; 292 } 293 294 void 295 vm_object_pip_wakeup(vm_object_t object) 296 { 297 GIANT_REQUIRED; 298 object->paging_in_progress--; 299 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 300 vm_object_clear_flag(object, OBJ_PIPWNT); 301 wakeup(object); 302 } 303 } 304 305 void 306 vm_object_pip_wakeupn(vm_object_t object, short i) 307 { 308 GIANT_REQUIRED; 309 if (i) 310 object->paging_in_progress -= i; 311 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 312 vm_object_clear_flag(object, OBJ_PIPWNT); 313 wakeup(object); 314 } 315 } 316 317 void 318 vm_object_pip_sleep(vm_object_t object, char *waitid) 319 { 320 GIANT_REQUIRED; 321 if (object->paging_in_progress) { 322 int s = splvm(); 323 if (object->paging_in_progress) { 324 vm_object_set_flag(object, OBJ_PIPWNT); 325 tsleep(object, PVM, waitid, 0); 326 } 327 splx(s); 328 } 329 } 330 331 void 332 vm_object_pip_wait(vm_object_t object, char *waitid) 333 { 334 GIANT_REQUIRED; 335 while (object->paging_in_progress) 336 vm_object_pip_sleep(object, waitid); 337 } 338 339 /* 340 * vm_object_allocate_wait 341 * 342 * Return a new object with the given size, and give the user the 343 * option of waiting for it to complete or failing if the needed 344 * memory isn't available. 345 */ 346 vm_object_t 347 vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 348 { 349 vm_object_t result; 350 351 result = (vm_object_t) uma_zalloc(obj_zone, flags); 352 353 if (result != NULL) 354 _vm_object_allocate(type, size, result); 355 356 return (result); 357 } 358 359 /* 360 * vm_object_allocate: 361 * 362 * Returns a new object with the given size. 363 */ 364 vm_object_t 365 vm_object_allocate(objtype_t type, vm_pindex_t size) 366 { 367 return(vm_object_allocate_wait(type, size, M_WAITOK)); 368 } 369 370 371 /* 372 * vm_object_reference: 373 * 374 * Gets another reference to the given object. 375 */ 376 void 377 vm_object_reference(vm_object_t object) 378 { 379 if (object == NULL) 380 return; 381 382 vm_object_lock(object); 383 #if 0 384 /* object can be re-referenced during final cleaning */ 385 KASSERT(!(object->flags & OBJ_DEAD), 386 ("vm_object_reference: attempting to reference dead obj")); 387 #endif 388 389 object->ref_count++; 390 if (object->type == OBJT_VNODE) { 391 while (vget((struct vnode *) object->handle, LK_RETRY, curthread)) { 392 printf("vm_object_reference: delay in getting object\n"); 393 } 394 } 395 vm_object_unlock(object); 396 } 397 398 /* 399 * handle deallocating a object of type OBJT_VNODE 400 */ 401 void 402 vm_object_vndeallocate(vm_object_t object) 403 { 404 struct vnode *vp = (struct vnode *) object->handle; 405 406 GIANT_REQUIRED; 407 KASSERT(object->type == OBJT_VNODE, 408 ("vm_object_vndeallocate: not a vnode object")); 409 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 410 #ifdef INVARIANTS 411 if (object->ref_count == 0) { 412 vprint("vm_object_vndeallocate", vp); 413 panic("vm_object_vndeallocate: bad object reference count"); 414 } 415 #endif 416 417 object->ref_count--; 418 if (object->ref_count == 0) { 419 mp_fixme("Unlocked vflag access."); 420 vp->v_vflag &= ~VV_TEXT; 421 #ifdef ENABLE_VFS_IOOPT 422 vm_object_clear_flag(object, OBJ_OPT); 423 #endif 424 } 425 /* 426 * vrele may need a vop lock 427 */ 428 vrele(vp); 429 } 430 431 /* 432 * vm_object_deallocate: 433 * 434 * Release a reference to the specified object, 435 * gained either through a vm_object_allocate 436 * or a vm_object_reference call. When all references 437 * are gone, storage associated with this object 438 * may be relinquished. 439 * 440 * No object may be locked. 441 */ 442 void 443 vm_object_deallocate(vm_object_t object) 444 { 445 vm_object_t temp; 446 447 mtx_lock(&Giant); 448 while (object != NULL) { 449 450 if (object->type == OBJT_VNODE) { 451 vm_object_vndeallocate(object); 452 mtx_unlock(&Giant); 453 return; 454 } 455 456 KASSERT(object->ref_count != 0, 457 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 458 459 /* 460 * If the reference count goes to 0 we start calling 461 * vm_object_terminate() on the object chain. 462 * A ref count of 1 may be a special case depending on the 463 * shadow count being 0 or 1. 464 */ 465 object->ref_count--; 466 if (object->ref_count > 1) { 467 mtx_unlock(&Giant); 468 return; 469 } else if (object->ref_count == 1) { 470 if (object->shadow_count == 0) { 471 vm_object_set_flag(object, OBJ_ONEMAPPING); 472 } else if ((object->shadow_count == 1) && 473 (object->handle == NULL) && 474 (object->type == OBJT_DEFAULT || 475 object->type == OBJT_SWAP)) { 476 vm_object_t robject; 477 478 robject = TAILQ_FIRST(&object->shadow_head); 479 KASSERT(robject != NULL, 480 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 481 object->ref_count, 482 object->shadow_count)); 483 if ((robject->handle == NULL) && 484 (robject->type == OBJT_DEFAULT || 485 robject->type == OBJT_SWAP)) { 486 487 robject->ref_count++; 488 489 while ( 490 robject->paging_in_progress || 491 object->paging_in_progress 492 ) { 493 vm_object_pip_sleep(robject, "objde1"); 494 vm_object_pip_sleep(object, "objde2"); 495 } 496 497 if (robject->ref_count == 1) { 498 robject->ref_count--; 499 object = robject; 500 goto doterm; 501 } 502 503 object = robject; 504 vm_object_collapse(object); 505 continue; 506 } 507 } 508 mtx_unlock(&Giant); 509 return; 510 } 511 doterm: 512 temp = object->backing_object; 513 if (temp) { 514 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 515 temp->shadow_count--; 516 #ifdef ENABLE_VFS_IOOPT 517 if (temp->ref_count == 0) 518 vm_object_clear_flag(temp, OBJ_OPT); 519 #endif 520 temp->generation++; 521 object->backing_object = NULL; 522 } 523 /* 524 * Don't double-terminate, we could be in a termination 525 * recursion due to the terminate having to sync data 526 * to disk. 527 */ 528 if ((object->flags & OBJ_DEAD) == 0) 529 vm_object_terminate(object); 530 object = temp; 531 } 532 mtx_unlock(&Giant); 533 } 534 535 /* 536 * vm_object_terminate actually destroys the specified object, freeing 537 * up all previously used resources. 538 * 539 * The object must be locked. 540 * This routine may block. 541 */ 542 void 543 vm_object_terminate(vm_object_t object) 544 { 545 vm_page_t p; 546 int s; 547 548 GIANT_REQUIRED; 549 550 /* 551 * Make sure no one uses us. 552 */ 553 vm_object_set_flag(object, OBJ_DEAD); 554 555 /* 556 * wait for the pageout daemon to be done with the object 557 */ 558 vm_object_pip_wait(object, "objtrm"); 559 560 KASSERT(!object->paging_in_progress, 561 ("vm_object_terminate: pageout in progress")); 562 563 /* 564 * Clean and free the pages, as appropriate. All references to the 565 * object are gone, so we don't need to lock it. 566 */ 567 if (object->type == OBJT_VNODE) { 568 struct vnode *vp; 569 570 #ifdef ENABLE_VFS_IOOPT 571 /* 572 * Freeze optimized copies. 573 */ 574 vm_freeze_copyopts(object, 0, object->size); 575 #endif 576 /* 577 * Clean pages and flush buffers. 578 */ 579 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 580 581 vp = (struct vnode *) object->handle; 582 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 583 } 584 585 KASSERT(object->ref_count == 0, 586 ("vm_object_terminate: object with references, ref_count=%d", 587 object->ref_count)); 588 589 /* 590 * Now free any remaining pages. For internal objects, this also 591 * removes them from paging queues. Don't free wired pages, just 592 * remove them from the object. 593 */ 594 s = splvm(); 595 vm_page_lock_queues(); 596 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 597 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 598 ("vm_object_terminate: freeing busy page %p " 599 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 600 if (p->wire_count == 0) { 601 vm_page_busy(p); 602 vm_page_free(p); 603 cnt.v_pfree++; 604 } else { 605 vm_page_busy(p); 606 vm_page_remove(p); 607 } 608 } 609 vm_page_unlock_queues(); 610 splx(s); 611 612 /* 613 * Let the pager know object is dead. 614 */ 615 vm_pager_deallocate(object); 616 617 /* 618 * Remove the object from the global object list. 619 */ 620 mtx_lock(&vm_object_list_mtx); 621 TAILQ_REMOVE(&vm_object_list, object, object_list); 622 mtx_unlock(&vm_object_list_mtx); 623 624 wakeup(object); 625 626 /* 627 * Free the space for the object. 628 */ 629 uma_zfree(obj_zone, object); 630 } 631 632 /* 633 * vm_object_page_clean 634 * 635 * Clean all dirty pages in the specified range of object. Leaves page 636 * on whatever queue it is currently on. If NOSYNC is set then do not 637 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 638 * leaving the object dirty. 639 * 640 * Odd semantics: if start == end, we clean everything. 641 * 642 * The object must be locked. 643 */ 644 void 645 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 646 { 647 vm_page_t p, np; 648 vm_pindex_t tstart, tend; 649 vm_pindex_t pi; 650 struct vnode *vp; 651 int clearobjflags; 652 int pagerflags; 653 int curgeneration; 654 655 GIANT_REQUIRED; 656 657 if (object->type != OBJT_VNODE || 658 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 659 return; 660 661 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0; 662 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 663 664 vp = object->handle; 665 666 vm_object_set_flag(object, OBJ_CLEANING); 667 668 tstart = start; 669 if (end == 0) { 670 tend = object->size; 671 } else { 672 tend = end; 673 } 674 675 /* 676 * If the caller is smart and only msync()s a range he knows is 677 * dirty, we may be able to avoid an object scan. This results in 678 * a phenominal improvement in performance. We cannot do this 679 * as a matter of course because the object may be huge - e.g. 680 * the size might be in the gigabytes or terrabytes. 681 */ 682 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 683 vm_pindex_t tscan; 684 int scanlimit; 685 int scanreset; 686 687 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 688 if (scanreset < 16) 689 scanreset = 16; 690 691 scanlimit = scanreset; 692 tscan = tstart; 693 while (tscan < tend) { 694 curgeneration = object->generation; 695 p = vm_page_lookup(object, tscan); 696 if (p == NULL || p->valid == 0 || 697 (p->queue - p->pc) == PQ_CACHE) { 698 if (--scanlimit == 0) 699 break; 700 ++tscan; 701 continue; 702 } 703 vm_page_test_dirty(p); 704 if ((p->dirty & p->valid) == 0) { 705 if (--scanlimit == 0) 706 break; 707 ++tscan; 708 continue; 709 } 710 /* 711 * If we have been asked to skip nosync pages and 712 * this is a nosync page, we can't continue. 713 */ 714 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 715 if (--scanlimit == 0) 716 break; 717 ++tscan; 718 continue; 719 } 720 scanlimit = scanreset; 721 722 /* 723 * This returns 0 if it was unable to busy the first 724 * page (i.e. had to sleep). 725 */ 726 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 727 } 728 729 /* 730 * If everything was dirty and we flushed it successfully, 731 * and the requested range is not the entire object, we 732 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 733 * return immediately. 734 */ 735 if (tscan >= tend && (tstart || tend < object->size)) { 736 vm_object_clear_flag(object, OBJ_CLEANING); 737 return; 738 } 739 } 740 741 /* 742 * Generally set CLEANCHK interlock and make the page read-only so 743 * we can then clear the object flags. 744 * 745 * However, if this is a nosync mmap then the object is likely to 746 * stay dirty so do not mess with the page and do not clear the 747 * object flags. 748 */ 749 clearobjflags = 1; 750 751 TAILQ_FOREACH(p, &object->memq, listq) { 752 vm_page_flag_set(p, PG_CLEANCHK); 753 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 754 clearobjflags = 0; 755 else 756 vm_page_protect(p, VM_PROT_READ); 757 } 758 759 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 760 struct vnode *vp; 761 762 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 763 if (object->type == OBJT_VNODE && 764 (vp = (struct vnode *)object->handle) != NULL) { 765 VI_LOCK(vp); 766 if (vp->v_iflag & VI_OBJDIRTY) 767 vp->v_iflag &= ~VI_OBJDIRTY; 768 VI_UNLOCK(vp); 769 } 770 } 771 772 rescan: 773 curgeneration = object->generation; 774 775 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 776 int n; 777 778 np = TAILQ_NEXT(p, listq); 779 780 again: 781 pi = p->pindex; 782 if (((p->flags & PG_CLEANCHK) == 0) || 783 (pi < tstart) || (pi >= tend) || 784 (p->valid == 0) || 785 ((p->queue - p->pc) == PQ_CACHE)) { 786 vm_page_flag_clear(p, PG_CLEANCHK); 787 continue; 788 } 789 790 vm_page_test_dirty(p); 791 if ((p->dirty & p->valid) == 0) { 792 vm_page_flag_clear(p, PG_CLEANCHK); 793 continue; 794 } 795 796 /* 797 * If we have been asked to skip nosync pages and this is a 798 * nosync page, skip it. Note that the object flags were 799 * not cleared in this case so we do not have to set them. 800 */ 801 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 802 vm_page_flag_clear(p, PG_CLEANCHK); 803 continue; 804 } 805 806 n = vm_object_page_collect_flush(object, p, 807 curgeneration, pagerflags); 808 if (n == 0) 809 goto rescan; 810 811 if (object->generation != curgeneration) 812 goto rescan; 813 814 /* 815 * Try to optimize the next page. If we can't we pick up 816 * our (random) scan where we left off. 817 */ 818 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 819 if ((p = vm_page_lookup(object, pi + n)) != NULL) 820 goto again; 821 } 822 } 823 824 #if 0 825 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 826 #endif 827 828 vm_object_clear_flag(object, OBJ_CLEANING); 829 return; 830 } 831 832 static int 833 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 834 { 835 int runlen; 836 int s; 837 int maxf; 838 int chkb; 839 int maxb; 840 int i; 841 vm_pindex_t pi; 842 vm_page_t maf[vm_pageout_page_count]; 843 vm_page_t mab[vm_pageout_page_count]; 844 vm_page_t ma[vm_pageout_page_count]; 845 846 s = splvm(); 847 pi = p->pindex; 848 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 849 if (object->generation != curgeneration) { 850 splx(s); 851 return(0); 852 } 853 } 854 vm_page_lock_queues(); 855 maxf = 0; 856 for(i = 1; i < vm_pageout_page_count; i++) { 857 vm_page_t tp; 858 859 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 860 if ((tp->flags & PG_BUSY) || 861 (tp->flags & PG_CLEANCHK) == 0 || 862 (tp->busy != 0)) 863 break; 864 if((tp->queue - tp->pc) == PQ_CACHE) { 865 vm_page_flag_clear(tp, PG_CLEANCHK); 866 break; 867 } 868 vm_page_test_dirty(tp); 869 if ((tp->dirty & tp->valid) == 0) { 870 vm_page_flag_clear(tp, PG_CLEANCHK); 871 break; 872 } 873 maf[ i - 1 ] = tp; 874 maxf++; 875 continue; 876 } 877 break; 878 } 879 880 maxb = 0; 881 chkb = vm_pageout_page_count - maxf; 882 if (chkb) { 883 for(i = 1; i < chkb;i++) { 884 vm_page_t tp; 885 886 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 887 if ((tp->flags & PG_BUSY) || 888 (tp->flags & PG_CLEANCHK) == 0 || 889 (tp->busy != 0)) 890 break; 891 if ((tp->queue - tp->pc) == PQ_CACHE) { 892 vm_page_flag_clear(tp, PG_CLEANCHK); 893 break; 894 } 895 vm_page_test_dirty(tp); 896 if ((tp->dirty & tp->valid) == 0) { 897 vm_page_flag_clear(tp, PG_CLEANCHK); 898 break; 899 } 900 mab[ i - 1 ] = tp; 901 maxb++; 902 continue; 903 } 904 break; 905 } 906 } 907 908 for(i = 0; i < maxb; i++) { 909 int index = (maxb - i) - 1; 910 ma[index] = mab[i]; 911 vm_page_flag_clear(ma[index], PG_CLEANCHK); 912 } 913 vm_page_flag_clear(p, PG_CLEANCHK); 914 ma[maxb] = p; 915 for(i = 0; i < maxf; i++) { 916 int index = (maxb + i) + 1; 917 ma[index] = maf[i]; 918 vm_page_flag_clear(ma[index], PG_CLEANCHK); 919 } 920 runlen = maxb + maxf + 1; 921 922 splx(s); 923 vm_pageout_flush(ma, runlen, pagerflags); 924 for (i = 0; i < runlen; i++) { 925 if (ma[i]->valid & ma[i]->dirty) { 926 vm_page_protect(ma[i], VM_PROT_READ); 927 vm_page_flag_set(ma[i], PG_CLEANCHK); 928 929 /* 930 * maxf will end up being the actual number of pages 931 * we wrote out contiguously, non-inclusive of the 932 * first page. We do not count look-behind pages. 933 */ 934 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 935 maxf = i - maxb - 1; 936 } 937 } 938 vm_page_unlock_queues(); 939 return(maxf + 1); 940 } 941 942 #ifdef ENABLE_VFS_IOOPT 943 /* 944 * Same as vm_object_pmap_copy, except range checking really 945 * works, and is meant for small sections of an object. 946 * 947 * This code protects resident pages by making them read-only 948 * and is typically called on a fork or split when a page 949 * is converted to copy-on-write. 950 * 951 * NOTE: If the page is already at VM_PROT_NONE, calling 952 * vm_page_protect will have no effect. 953 */ 954 void 955 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 956 { 957 vm_pindex_t idx; 958 vm_page_t p; 959 960 GIANT_REQUIRED; 961 962 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 963 return; 964 965 for (idx = start; idx < end; idx++) { 966 p = vm_page_lookup(object, idx); 967 if (p == NULL) 968 continue; 969 vm_page_protect(p, VM_PROT_READ); 970 } 971 } 972 #endif 973 974 /* 975 * vm_object_pmap_remove: 976 * 977 * Removes all physical pages in the specified 978 * object range from all physical maps. 979 * 980 * The object must *not* be locked. 981 */ 982 void 983 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 984 { 985 vm_page_t p; 986 987 GIANT_REQUIRED; 988 if (object == NULL) 989 return; 990 TAILQ_FOREACH(p, &object->memq, listq) { 991 if (p->pindex >= start && p->pindex < end) 992 vm_page_protect(p, VM_PROT_NONE); 993 } 994 if ((start == 0) && (object->size == end)) 995 vm_object_clear_flag(object, OBJ_WRITEABLE); 996 } 997 998 /* 999 * vm_object_madvise: 1000 * 1001 * Implements the madvise function at the object/page level. 1002 * 1003 * MADV_WILLNEED (any object) 1004 * 1005 * Activate the specified pages if they are resident. 1006 * 1007 * MADV_DONTNEED (any object) 1008 * 1009 * Deactivate the specified pages if they are resident. 1010 * 1011 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1012 * OBJ_ONEMAPPING only) 1013 * 1014 * Deactivate and clean the specified pages if they are 1015 * resident. This permits the process to reuse the pages 1016 * without faulting or the kernel to reclaim the pages 1017 * without I/O. 1018 */ 1019 void 1020 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1021 { 1022 vm_pindex_t end, tpindex; 1023 vm_object_t tobject; 1024 vm_page_t m; 1025 1026 if (object == NULL) 1027 return; 1028 1029 vm_object_lock(object); 1030 1031 end = pindex + count; 1032 1033 /* 1034 * Locate and adjust resident pages 1035 */ 1036 for (; pindex < end; pindex += 1) { 1037 relookup: 1038 tobject = object; 1039 tpindex = pindex; 1040 shadowlookup: 1041 /* 1042 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1043 * and those pages must be OBJ_ONEMAPPING. 1044 */ 1045 if (advise == MADV_FREE) { 1046 if ((tobject->type != OBJT_DEFAULT && 1047 tobject->type != OBJT_SWAP) || 1048 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1049 continue; 1050 } 1051 } 1052 1053 m = vm_page_lookup(tobject, tpindex); 1054 1055 if (m == NULL) { 1056 /* 1057 * There may be swap even if there is no backing page 1058 */ 1059 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1060 swap_pager_freespace(tobject, tpindex, 1); 1061 1062 /* 1063 * next object 1064 */ 1065 tobject = tobject->backing_object; 1066 if (tobject == NULL) 1067 continue; 1068 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1069 goto shadowlookup; 1070 } 1071 1072 /* 1073 * If the page is busy or not in a normal active state, 1074 * we skip it. If the page is not managed there are no 1075 * page queues to mess with. Things can break if we mess 1076 * with pages in any of the below states. 1077 */ 1078 vm_page_lock_queues(); 1079 if (m->hold_count || 1080 m->wire_count || 1081 (m->flags & PG_UNMANAGED) || 1082 m->valid != VM_PAGE_BITS_ALL) { 1083 vm_page_unlock_queues(); 1084 continue; 1085 } 1086 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) 1087 goto relookup; 1088 if (advise == MADV_WILLNEED) { 1089 vm_page_activate(m); 1090 } else if (advise == MADV_DONTNEED) { 1091 vm_page_dontneed(m); 1092 } else if (advise == MADV_FREE) { 1093 /* 1094 * Mark the page clean. This will allow the page 1095 * to be freed up by the system. However, such pages 1096 * are often reused quickly by malloc()/free() 1097 * so we do not do anything that would cause 1098 * a page fault if we can help it. 1099 * 1100 * Specifically, we do not try to actually free 1101 * the page now nor do we try to put it in the 1102 * cache (which would cause a page fault on reuse). 1103 * 1104 * But we do make the page is freeable as we 1105 * can without actually taking the step of unmapping 1106 * it. 1107 */ 1108 pmap_clear_modify(m); 1109 m->dirty = 0; 1110 m->act_count = 0; 1111 vm_page_dontneed(m); 1112 } 1113 vm_page_unlock_queues(); 1114 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1115 swap_pager_freespace(tobject, tpindex, 1); 1116 } 1117 vm_object_unlock(object); 1118 } 1119 1120 /* 1121 * vm_object_shadow: 1122 * 1123 * Create a new object which is backed by the 1124 * specified existing object range. The source 1125 * object reference is deallocated. 1126 * 1127 * The new object and offset into that object 1128 * are returned in the source parameters. 1129 */ 1130 void 1131 vm_object_shadow( 1132 vm_object_t *object, /* IN/OUT */ 1133 vm_ooffset_t *offset, /* IN/OUT */ 1134 vm_size_t length) 1135 { 1136 vm_object_t source; 1137 vm_object_t result; 1138 1139 source = *object; 1140 1141 vm_object_lock(source); 1142 /* 1143 * Don't create the new object if the old object isn't shared. 1144 */ 1145 if (source != NULL && 1146 source->ref_count == 1 && 1147 source->handle == NULL && 1148 (source->type == OBJT_DEFAULT || 1149 source->type == OBJT_SWAP)) { 1150 vm_object_unlock(source); 1151 return; 1152 } 1153 1154 /* 1155 * Allocate a new object with the given length 1156 */ 1157 result = vm_object_allocate(OBJT_DEFAULT, length); 1158 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1159 1160 /* 1161 * The new object shadows the source object, adding a reference to it. 1162 * Our caller changes his reference to point to the new object, 1163 * removing a reference to the source object. Net result: no change 1164 * of reference count. 1165 * 1166 * Try to optimize the result object's page color when shadowing 1167 * in order to maintain page coloring consistency in the combined 1168 * shadowed object. 1169 */ 1170 result->backing_object = source; 1171 if (source) { 1172 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1173 source->shadow_count++; 1174 source->generation++; 1175 if (length < source->size) 1176 length = source->size; 1177 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1178 source->generation > 1) 1179 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1180 result->pg_color = (source->pg_color + 1181 length * source->generation) & PQ_L2_MASK; 1182 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1183 PQ_L2_MASK; 1184 } 1185 1186 /* 1187 * Store the offset into the source object, and fix up the offset into 1188 * the new object. 1189 */ 1190 result->backing_object_offset = *offset; 1191 1192 /* 1193 * Return the new things 1194 */ 1195 *offset = 0; 1196 *object = result; 1197 1198 vm_object_unlock(source); 1199 } 1200 1201 /* 1202 * vm_object_split: 1203 * 1204 * Split the pages in a map entry into a new object. This affords 1205 * easier removal of unused pages, and keeps object inheritance from 1206 * being a negative impact on memory usage. 1207 */ 1208 void 1209 vm_object_split(vm_map_entry_t entry) 1210 { 1211 vm_page_t m; 1212 vm_object_t orig_object, new_object, source; 1213 vm_offset_t s, e; 1214 vm_pindex_t offidxstart, offidxend; 1215 vm_size_t idx, size; 1216 vm_ooffset_t offset; 1217 1218 GIANT_REQUIRED; 1219 1220 orig_object = entry->object.vm_object; 1221 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1222 return; 1223 if (orig_object->ref_count <= 1) 1224 return; 1225 1226 offset = entry->offset; 1227 s = entry->start; 1228 e = entry->end; 1229 1230 offidxstart = OFF_TO_IDX(offset); 1231 offidxend = offidxstart + OFF_TO_IDX(e - s); 1232 size = offidxend - offidxstart; 1233 1234 new_object = vm_pager_allocate(orig_object->type, 1235 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1236 if (new_object == NULL) 1237 return; 1238 1239 source = orig_object->backing_object; 1240 if (source != NULL) { 1241 vm_object_reference(source); /* Referenced by new_object */ 1242 TAILQ_INSERT_TAIL(&source->shadow_head, 1243 new_object, shadow_list); 1244 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1245 new_object->backing_object_offset = 1246 orig_object->backing_object_offset + offset; 1247 new_object->backing_object = source; 1248 source->shadow_count++; 1249 source->generation++; 1250 } 1251 for (idx = 0; idx < size; idx++) { 1252 retry: 1253 m = vm_page_lookup(orig_object, offidxstart + idx); 1254 if (m == NULL) 1255 continue; 1256 1257 /* 1258 * We must wait for pending I/O to complete before we can 1259 * rename the page. 1260 * 1261 * We do not have to VM_PROT_NONE the page as mappings should 1262 * not be changed by this operation. 1263 */ 1264 vm_page_lock_queues(); 1265 if (vm_page_sleep_if_busy(m, TRUE, "spltwt")) 1266 goto retry; 1267 1268 vm_page_busy(m); 1269 vm_page_unlock_queues(); 1270 vm_page_rename(m, new_object, idx); 1271 /* page automatically made dirty by rename and cache handled */ 1272 vm_page_busy(m); 1273 } 1274 if (orig_object->type == OBJT_SWAP) { 1275 vm_object_pip_add(orig_object, 1); 1276 /* 1277 * copy orig_object pages into new_object 1278 * and destroy unneeded pages in 1279 * shadow object. 1280 */ 1281 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1282 vm_object_pip_wakeup(orig_object); 1283 } 1284 TAILQ_FOREACH(m, &new_object->memq, listq) 1285 vm_page_wakeup(m); 1286 entry->object.vm_object = new_object; 1287 entry->offset = 0LL; 1288 vm_object_deallocate(orig_object); 1289 } 1290 1291 #define OBSC_TEST_ALL_SHADOWED 0x0001 1292 #define OBSC_COLLAPSE_NOWAIT 0x0002 1293 #define OBSC_COLLAPSE_WAIT 0x0004 1294 1295 static __inline int 1296 vm_object_backing_scan(vm_object_t object, int op) 1297 { 1298 int s; 1299 int r = 1; 1300 vm_page_t p; 1301 vm_object_t backing_object; 1302 vm_pindex_t backing_offset_index; 1303 1304 s = splvm(); 1305 GIANT_REQUIRED; 1306 1307 backing_object = object->backing_object; 1308 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1309 1310 /* 1311 * Initial conditions 1312 */ 1313 if (op & OBSC_TEST_ALL_SHADOWED) { 1314 /* 1315 * We do not want to have to test for the existence of 1316 * swap pages in the backing object. XXX but with the 1317 * new swapper this would be pretty easy to do. 1318 * 1319 * XXX what about anonymous MAP_SHARED memory that hasn't 1320 * been ZFOD faulted yet? If we do not test for this, the 1321 * shadow test may succeed! XXX 1322 */ 1323 if (backing_object->type != OBJT_DEFAULT) { 1324 splx(s); 1325 return (0); 1326 } 1327 } 1328 if (op & OBSC_COLLAPSE_WAIT) { 1329 vm_object_set_flag(backing_object, OBJ_DEAD); 1330 } 1331 1332 /* 1333 * Our scan 1334 */ 1335 p = TAILQ_FIRST(&backing_object->memq); 1336 while (p) { 1337 vm_page_t next = TAILQ_NEXT(p, listq); 1338 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1339 1340 if (op & OBSC_TEST_ALL_SHADOWED) { 1341 vm_page_t pp; 1342 1343 /* 1344 * Ignore pages outside the parent object's range 1345 * and outside the parent object's mapping of the 1346 * backing object. 1347 * 1348 * note that we do not busy the backing object's 1349 * page. 1350 */ 1351 if ( 1352 p->pindex < backing_offset_index || 1353 new_pindex >= object->size 1354 ) { 1355 p = next; 1356 continue; 1357 } 1358 1359 /* 1360 * See if the parent has the page or if the parent's 1361 * object pager has the page. If the parent has the 1362 * page but the page is not valid, the parent's 1363 * object pager must have the page. 1364 * 1365 * If this fails, the parent does not completely shadow 1366 * the object and we might as well give up now. 1367 */ 1368 1369 pp = vm_page_lookup(object, new_pindex); 1370 if ( 1371 (pp == NULL || pp->valid == 0) && 1372 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1373 ) { 1374 r = 0; 1375 break; 1376 } 1377 } 1378 1379 /* 1380 * Check for busy page 1381 */ 1382 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1383 vm_page_t pp; 1384 1385 vm_page_lock_queues(); 1386 if (op & OBSC_COLLAPSE_NOWAIT) { 1387 if ((p->flags & PG_BUSY) || 1388 !p->valid || 1389 p->hold_count || 1390 p->wire_count || 1391 p->busy) { 1392 vm_page_unlock_queues(); 1393 p = next; 1394 continue; 1395 } 1396 } else if (op & OBSC_COLLAPSE_WAIT) { 1397 if (vm_page_sleep_if_busy(p, TRUE, "vmocol")) { 1398 /* 1399 * If we slept, anything could have 1400 * happened. Since the object is 1401 * marked dead, the backing offset 1402 * should not have changed so we 1403 * just restart our scan. 1404 */ 1405 p = TAILQ_FIRST(&backing_object->memq); 1406 continue; 1407 } 1408 } 1409 1410 /* 1411 * Busy the page 1412 */ 1413 vm_page_busy(p); 1414 vm_page_unlock_queues(); 1415 1416 KASSERT( 1417 p->object == backing_object, 1418 ("vm_object_qcollapse(): object mismatch") 1419 ); 1420 1421 /* 1422 * Destroy any associated swap 1423 */ 1424 if (backing_object->type == OBJT_SWAP) { 1425 swap_pager_freespace( 1426 backing_object, 1427 p->pindex, 1428 1 1429 ); 1430 } 1431 1432 if ( 1433 p->pindex < backing_offset_index || 1434 new_pindex >= object->size 1435 ) { 1436 /* 1437 * Page is out of the parent object's range, we 1438 * can simply destroy it. 1439 */ 1440 vm_page_lock_queues(); 1441 vm_page_protect(p, VM_PROT_NONE); 1442 vm_page_free(p); 1443 vm_page_unlock_queues(); 1444 p = next; 1445 continue; 1446 } 1447 1448 pp = vm_page_lookup(object, new_pindex); 1449 if ( 1450 pp != NULL || 1451 vm_pager_has_page(object, new_pindex, NULL, NULL) 1452 ) { 1453 /* 1454 * page already exists in parent OR swap exists 1455 * for this location in the parent. Destroy 1456 * the original page from the backing object. 1457 * 1458 * Leave the parent's page alone 1459 */ 1460 vm_page_lock_queues(); 1461 vm_page_protect(p, VM_PROT_NONE); 1462 vm_page_free(p); 1463 vm_page_unlock_queues(); 1464 p = next; 1465 continue; 1466 } 1467 1468 /* 1469 * Page does not exist in parent, rename the 1470 * page from the backing object to the main object. 1471 * 1472 * If the page was mapped to a process, it can remain 1473 * mapped through the rename. 1474 */ 1475 vm_page_rename(p, object, new_pindex); 1476 /* page automatically made dirty by rename */ 1477 } 1478 p = next; 1479 } 1480 splx(s); 1481 return (r); 1482 } 1483 1484 1485 /* 1486 * this version of collapse allows the operation to occur earlier and 1487 * when paging_in_progress is true for an object... This is not a complete 1488 * operation, but should plug 99.9% of the rest of the leaks. 1489 */ 1490 static void 1491 vm_object_qcollapse(vm_object_t object) 1492 { 1493 vm_object_t backing_object = object->backing_object; 1494 1495 GIANT_REQUIRED; 1496 1497 if (backing_object->ref_count != 1) 1498 return; 1499 1500 backing_object->ref_count += 2; 1501 1502 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1503 1504 backing_object->ref_count -= 2; 1505 } 1506 1507 /* 1508 * vm_object_collapse: 1509 * 1510 * Collapse an object with the object backing it. 1511 * Pages in the backing object are moved into the 1512 * parent, and the backing object is deallocated. 1513 */ 1514 void 1515 vm_object_collapse(vm_object_t object) 1516 { 1517 GIANT_REQUIRED; 1518 1519 while (TRUE) { 1520 vm_object_t backing_object; 1521 1522 /* 1523 * Verify that the conditions are right for collapse: 1524 * 1525 * The object exists and the backing object exists. 1526 */ 1527 if (object == NULL) 1528 break; 1529 1530 if ((backing_object = object->backing_object) == NULL) 1531 break; 1532 1533 /* 1534 * we check the backing object first, because it is most likely 1535 * not collapsable. 1536 */ 1537 if (backing_object->handle != NULL || 1538 (backing_object->type != OBJT_DEFAULT && 1539 backing_object->type != OBJT_SWAP) || 1540 (backing_object->flags & OBJ_DEAD) || 1541 object->handle != NULL || 1542 (object->type != OBJT_DEFAULT && 1543 object->type != OBJT_SWAP) || 1544 (object->flags & OBJ_DEAD)) { 1545 break; 1546 } 1547 1548 if ( 1549 object->paging_in_progress != 0 || 1550 backing_object->paging_in_progress != 0 1551 ) { 1552 vm_object_qcollapse(object); 1553 break; 1554 } 1555 1556 /* 1557 * We know that we can either collapse the backing object (if 1558 * the parent is the only reference to it) or (perhaps) have 1559 * the parent bypass the object if the parent happens to shadow 1560 * all the resident pages in the entire backing object. 1561 * 1562 * This is ignoring pager-backed pages such as swap pages. 1563 * vm_object_backing_scan fails the shadowing test in this 1564 * case. 1565 */ 1566 if (backing_object->ref_count == 1) { 1567 /* 1568 * If there is exactly one reference to the backing 1569 * object, we can collapse it into the parent. 1570 */ 1571 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1572 1573 /* 1574 * Move the pager from backing_object to object. 1575 */ 1576 if (backing_object->type == OBJT_SWAP) { 1577 vm_object_pip_add(backing_object, 1); 1578 1579 /* 1580 * scrap the paging_offset junk and do a 1581 * discrete copy. This also removes major 1582 * assumptions about how the swap-pager 1583 * works from where it doesn't belong. The 1584 * new swapper is able to optimize the 1585 * destroy-source case. 1586 */ 1587 vm_object_pip_add(object, 1); 1588 swap_pager_copy( 1589 backing_object, 1590 object, 1591 OFF_TO_IDX(object->backing_object_offset), TRUE); 1592 vm_object_pip_wakeup(object); 1593 1594 vm_object_pip_wakeup(backing_object); 1595 } 1596 /* 1597 * Object now shadows whatever backing_object did. 1598 * Note that the reference to 1599 * backing_object->backing_object moves from within 1600 * backing_object to within object. 1601 */ 1602 TAILQ_REMOVE( 1603 &object->backing_object->shadow_head, 1604 object, 1605 shadow_list 1606 ); 1607 object->backing_object->shadow_count--; 1608 object->backing_object->generation++; 1609 if (backing_object->backing_object) { 1610 TAILQ_REMOVE( 1611 &backing_object->backing_object->shadow_head, 1612 backing_object, 1613 shadow_list 1614 ); 1615 backing_object->backing_object->shadow_count--; 1616 backing_object->backing_object->generation++; 1617 } 1618 object->backing_object = backing_object->backing_object; 1619 if (object->backing_object) { 1620 TAILQ_INSERT_TAIL( 1621 &object->backing_object->shadow_head, 1622 object, 1623 shadow_list 1624 ); 1625 object->backing_object->shadow_count++; 1626 object->backing_object->generation++; 1627 } 1628 1629 object->backing_object_offset += 1630 backing_object->backing_object_offset; 1631 1632 /* 1633 * Discard backing_object. 1634 * 1635 * Since the backing object has no pages, no pager left, 1636 * and no object references within it, all that is 1637 * necessary is to dispose of it. 1638 */ 1639 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1640 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1641 1642 mtx_lock(&vm_object_list_mtx); 1643 TAILQ_REMOVE( 1644 &vm_object_list, 1645 backing_object, 1646 object_list 1647 ); 1648 mtx_unlock(&vm_object_list_mtx); 1649 1650 uma_zfree(obj_zone, backing_object); 1651 1652 object_collapses++; 1653 } else { 1654 vm_object_t new_backing_object; 1655 1656 /* 1657 * If we do not entirely shadow the backing object, 1658 * there is nothing we can do so we give up. 1659 */ 1660 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1661 break; 1662 } 1663 1664 /* 1665 * Make the parent shadow the next object in the 1666 * chain. Deallocating backing_object will not remove 1667 * it, since its reference count is at least 2. 1668 */ 1669 TAILQ_REMOVE( 1670 &backing_object->shadow_head, 1671 object, 1672 shadow_list 1673 ); 1674 backing_object->shadow_count--; 1675 backing_object->generation++; 1676 1677 new_backing_object = backing_object->backing_object; 1678 if ((object->backing_object = new_backing_object) != NULL) { 1679 vm_object_reference(new_backing_object); 1680 TAILQ_INSERT_TAIL( 1681 &new_backing_object->shadow_head, 1682 object, 1683 shadow_list 1684 ); 1685 new_backing_object->shadow_count++; 1686 new_backing_object->generation++; 1687 object->backing_object_offset += 1688 backing_object->backing_object_offset; 1689 } 1690 1691 /* 1692 * Drop the reference count on backing_object. Since 1693 * its ref_count was at least 2, it will not vanish; 1694 * so we don't need to call vm_object_deallocate, but 1695 * we do anyway. 1696 */ 1697 vm_object_deallocate(backing_object); 1698 object_bypasses++; 1699 } 1700 1701 /* 1702 * Try again with this object's new backing object. 1703 */ 1704 } 1705 } 1706 1707 /* 1708 * vm_object_page_remove: [internal] 1709 * 1710 * Removes all physical pages in the specified 1711 * object range from the object's list of pages. 1712 * 1713 * The object must be locked. 1714 */ 1715 void 1716 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1717 { 1718 vm_page_t p, next; 1719 vm_pindex_t size; 1720 int all; 1721 1722 if (object == NULL) 1723 return; 1724 1725 mtx_lock(&Giant); 1726 if (object->resident_page_count == 0) { 1727 mtx_unlock(&Giant); 1728 return; 1729 } 1730 all = ((end == 0) && (start == 0)); 1731 1732 /* 1733 * Since physically-backed objects do not use managed pages, we can't 1734 * remove pages from the object (we must instead remove the page 1735 * references, and then destroy the object). 1736 */ 1737 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1738 1739 vm_object_pip_add(object, 1); 1740 again: 1741 vm_page_lock_queues(); 1742 size = end - start; 1743 if (all || size > object->resident_page_count / 4) { 1744 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) { 1745 next = TAILQ_NEXT(p, listq); 1746 if (all || ((start <= p->pindex) && (p->pindex < end))) { 1747 if (p->wire_count != 0) { 1748 vm_page_protect(p, VM_PROT_NONE); 1749 if (!clean_only) 1750 p->valid = 0; 1751 continue; 1752 } 1753 1754 /* 1755 * The busy flags are only cleared at 1756 * interrupt -- minimize the spl transitions 1757 */ 1758 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1759 goto again; 1760 1761 if (clean_only && p->valid) { 1762 vm_page_test_dirty(p); 1763 if (p->valid & p->dirty) 1764 continue; 1765 } 1766 vm_page_busy(p); 1767 vm_page_protect(p, VM_PROT_NONE); 1768 vm_page_free(p); 1769 } 1770 } 1771 } else { 1772 while (size > 0) { 1773 if ((p = vm_page_lookup(object, start)) != NULL) { 1774 if (p->wire_count != 0) { 1775 vm_page_protect(p, VM_PROT_NONE); 1776 if (!clean_only) 1777 p->valid = 0; 1778 start += 1; 1779 size -= 1; 1780 continue; 1781 } 1782 1783 /* 1784 * The busy flags are only cleared at 1785 * interrupt -- minimize the spl transitions 1786 */ 1787 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1788 goto again; 1789 1790 if (clean_only && p->valid) { 1791 vm_page_test_dirty(p); 1792 if (p->valid & p->dirty) { 1793 start += 1; 1794 size -= 1; 1795 continue; 1796 } 1797 } 1798 vm_page_busy(p); 1799 vm_page_protect(p, VM_PROT_NONE); 1800 vm_page_free(p); 1801 } 1802 start += 1; 1803 size -= 1; 1804 } 1805 } 1806 vm_page_unlock_queues(); 1807 vm_object_pip_wakeup(object); 1808 mtx_unlock(&Giant); 1809 } 1810 1811 /* 1812 * Routine: vm_object_coalesce 1813 * Function: Coalesces two objects backing up adjoining 1814 * regions of memory into a single object. 1815 * 1816 * returns TRUE if objects were combined. 1817 * 1818 * NOTE: Only works at the moment if the second object is NULL - 1819 * if it's not, which object do we lock first? 1820 * 1821 * Parameters: 1822 * prev_object First object to coalesce 1823 * prev_offset Offset into prev_object 1824 * next_object Second object into coalesce 1825 * next_offset Offset into next_object 1826 * 1827 * prev_size Size of reference to prev_object 1828 * next_size Size of reference to next_object 1829 * 1830 * Conditions: 1831 * The object must *not* be locked. 1832 */ 1833 boolean_t 1834 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1835 vm_size_t prev_size, vm_size_t next_size) 1836 { 1837 vm_pindex_t next_pindex; 1838 1839 if (prev_object == NULL) 1840 return (TRUE); 1841 vm_object_lock(prev_object); 1842 if (prev_object->type != OBJT_DEFAULT && 1843 prev_object->type != OBJT_SWAP) { 1844 vm_object_unlock(prev_object); 1845 return (FALSE); 1846 } 1847 1848 /* 1849 * Try to collapse the object first 1850 */ 1851 vm_object_collapse(prev_object); 1852 1853 /* 1854 * Can't coalesce if: . more than one reference . paged out . shadows 1855 * another object . has a copy elsewhere (any of which mean that the 1856 * pages not mapped to prev_entry may be in use anyway) 1857 */ 1858 if (prev_object->backing_object != NULL) { 1859 vm_object_unlock(prev_object); 1860 return (FALSE); 1861 } 1862 1863 prev_size >>= PAGE_SHIFT; 1864 next_size >>= PAGE_SHIFT; 1865 next_pindex = prev_pindex + prev_size; 1866 1867 if ((prev_object->ref_count > 1) && 1868 (prev_object->size != next_pindex)) { 1869 vm_object_unlock(prev_object); 1870 return (FALSE); 1871 } 1872 1873 /* 1874 * Remove any pages that may still be in the object from a previous 1875 * deallocation. 1876 */ 1877 if (next_pindex < prev_object->size) { 1878 vm_object_page_remove(prev_object, 1879 next_pindex, 1880 next_pindex + next_size, FALSE); 1881 if (prev_object->type == OBJT_SWAP) 1882 swap_pager_freespace(prev_object, 1883 next_pindex, next_size); 1884 } 1885 1886 /* 1887 * Extend the object if necessary. 1888 */ 1889 if (next_pindex + next_size > prev_object->size) 1890 prev_object->size = next_pindex + next_size; 1891 1892 vm_object_unlock(prev_object); 1893 return (TRUE); 1894 } 1895 1896 void 1897 vm_object_set_writeable_dirty(vm_object_t object) 1898 { 1899 struct vnode *vp; 1900 1901 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1902 if (object->type == OBJT_VNODE && 1903 (vp = (struct vnode *)object->handle) != NULL) { 1904 VI_LOCK(vp); 1905 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1906 vp->v_iflag |= VI_OBJDIRTY; 1907 VI_UNLOCK(vp); 1908 } 1909 } 1910 1911 #ifdef ENABLE_VFS_IOOPT 1912 /* 1913 * Experimental support for zero-copy I/O 1914 * 1915 * Performs the copy_on_write operations necessary to allow the virtual copies 1916 * into user space to work. This has to be called for write(2) system calls 1917 * from other processes, file unlinking, and file size shrinkage. 1918 */ 1919 void 1920 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 1921 { 1922 int rv; 1923 vm_object_t robject; 1924 vm_pindex_t idx; 1925 1926 GIANT_REQUIRED; 1927 if ((object == NULL) || 1928 ((object->flags & OBJ_OPT) == 0)) 1929 return; 1930 1931 if (object->shadow_count > object->ref_count) 1932 panic("vm_freeze_copyopts: sc > rc"); 1933 1934 while ((robject = TAILQ_FIRST(&object->shadow_head)) != NULL) { 1935 vm_pindex_t bo_pindex; 1936 vm_page_t m_in, m_out; 1937 1938 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 1939 1940 vm_object_reference(robject); 1941 1942 vm_object_pip_wait(robject, "objfrz"); 1943 1944 if (robject->ref_count == 1) { 1945 vm_object_deallocate(robject); 1946 continue; 1947 } 1948 1949 vm_object_pip_add(robject, 1); 1950 1951 for (idx = 0; idx < robject->size; idx++) { 1952 1953 m_out = vm_page_grab(robject, idx, 1954 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1955 1956 if (m_out->valid == 0) { 1957 m_in = vm_page_grab(object, bo_pindex + idx, 1958 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1959 if (m_in->valid == 0) { 1960 rv = vm_pager_get_pages(object, &m_in, 1, 0); 1961 if (rv != VM_PAGER_OK) { 1962 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 1963 continue; 1964 } 1965 vm_page_lock_queues(); 1966 vm_page_deactivate(m_in); 1967 vm_page_unlock_queues(); 1968 } 1969 1970 vm_page_protect(m_in, VM_PROT_NONE); 1971 pmap_copy_page(m_in, m_out); 1972 m_out->valid = m_in->valid; 1973 vm_page_dirty(m_out); 1974 vm_page_lock_queues(); 1975 vm_page_activate(m_out); 1976 vm_page_unlock_queues(); 1977 vm_page_wakeup(m_in); 1978 } 1979 vm_page_wakeup(m_out); 1980 } 1981 1982 object->shadow_count--; 1983 object->ref_count--; 1984 TAILQ_REMOVE(&object->shadow_head, robject, shadow_list); 1985 robject->backing_object = NULL; 1986 robject->backing_object_offset = 0; 1987 1988 vm_object_pip_wakeup(robject); 1989 vm_object_deallocate(robject); 1990 } 1991 1992 vm_object_clear_flag(object, OBJ_OPT); 1993 } 1994 #endif 1995 1996 #include "opt_ddb.h" 1997 #ifdef DDB 1998 #include <sys/kernel.h> 1999 2000 #include <sys/cons.h> 2001 2002 #include <ddb/ddb.h> 2003 2004 static int 2005 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2006 { 2007 vm_map_t tmpm; 2008 vm_map_entry_t tmpe; 2009 vm_object_t obj; 2010 int entcount; 2011 2012 if (map == 0) 2013 return 0; 2014 2015 if (entry == 0) { 2016 tmpe = map->header.next; 2017 entcount = map->nentries; 2018 while (entcount-- && (tmpe != &map->header)) { 2019 if (_vm_object_in_map(map, object, tmpe)) { 2020 return 1; 2021 } 2022 tmpe = tmpe->next; 2023 } 2024 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2025 tmpm = entry->object.sub_map; 2026 tmpe = tmpm->header.next; 2027 entcount = tmpm->nentries; 2028 while (entcount-- && tmpe != &tmpm->header) { 2029 if (_vm_object_in_map(tmpm, object, tmpe)) { 2030 return 1; 2031 } 2032 tmpe = tmpe->next; 2033 } 2034 } else if ((obj = entry->object.vm_object) != NULL) { 2035 for (; obj; obj = obj->backing_object) 2036 if (obj == object) { 2037 return 1; 2038 } 2039 } 2040 return 0; 2041 } 2042 2043 static int 2044 vm_object_in_map(vm_object_t object) 2045 { 2046 struct proc *p; 2047 2048 /* sx_slock(&allproc_lock); */ 2049 LIST_FOREACH(p, &allproc, p_list) { 2050 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2051 continue; 2052 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2053 /* sx_sunlock(&allproc_lock); */ 2054 return 1; 2055 } 2056 } 2057 /* sx_sunlock(&allproc_lock); */ 2058 if (_vm_object_in_map(kernel_map, object, 0)) 2059 return 1; 2060 if (_vm_object_in_map(kmem_map, object, 0)) 2061 return 1; 2062 if (_vm_object_in_map(pager_map, object, 0)) 2063 return 1; 2064 if (_vm_object_in_map(buffer_map, object, 0)) 2065 return 1; 2066 return 0; 2067 } 2068 2069 DB_SHOW_COMMAND(vmochk, vm_object_check) 2070 { 2071 vm_object_t object; 2072 2073 /* 2074 * make sure that internal objs are in a map somewhere 2075 * and none have zero ref counts. 2076 */ 2077 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2078 if (object->handle == NULL && 2079 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2080 if (object->ref_count == 0) { 2081 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2082 (long)object->size); 2083 } 2084 if (!vm_object_in_map(object)) { 2085 db_printf( 2086 "vmochk: internal obj is not in a map: " 2087 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2088 object->ref_count, (u_long)object->size, 2089 (u_long)object->size, 2090 (void *)object->backing_object); 2091 } 2092 } 2093 } 2094 } 2095 2096 /* 2097 * vm_object_print: [ debug ] 2098 */ 2099 DB_SHOW_COMMAND(object, vm_object_print_static) 2100 { 2101 /* XXX convert args. */ 2102 vm_object_t object = (vm_object_t)addr; 2103 boolean_t full = have_addr; 2104 2105 vm_page_t p; 2106 2107 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2108 #define count was_count 2109 2110 int count; 2111 2112 if (object == NULL) 2113 return; 2114 2115 db_iprintf( 2116 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2117 object, (int)object->type, (u_long)object->size, 2118 object->resident_page_count, object->ref_count, object->flags); 2119 /* 2120 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2121 */ 2122 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2123 object->shadow_count, 2124 object->backing_object ? object->backing_object->ref_count : 0, 2125 object->backing_object, (long)object->backing_object_offset); 2126 2127 if (!full) 2128 return; 2129 2130 db_indent += 2; 2131 count = 0; 2132 TAILQ_FOREACH(p, &object->memq, listq) { 2133 if (count == 0) 2134 db_iprintf("memory:="); 2135 else if (count == 6) { 2136 db_printf("\n"); 2137 db_iprintf(" ..."); 2138 count = 0; 2139 } else 2140 db_printf(","); 2141 count++; 2142 2143 db_printf("(off=0x%lx,page=0x%lx)", 2144 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2145 } 2146 if (count != 0) 2147 db_printf("\n"); 2148 db_indent -= 2; 2149 } 2150 2151 /* XXX. */ 2152 #undef count 2153 2154 /* XXX need this non-static entry for calling from vm_map_print. */ 2155 void 2156 vm_object_print( 2157 /* db_expr_t */ long addr, 2158 boolean_t have_addr, 2159 /* db_expr_t */ long count, 2160 char *modif) 2161 { 2162 vm_object_print_static(addr, have_addr, count, modif); 2163 } 2164 2165 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2166 { 2167 vm_object_t object; 2168 int nl = 0; 2169 int c; 2170 2171 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2172 vm_pindex_t idx, fidx; 2173 vm_pindex_t osize; 2174 vm_offset_t pa = -1, padiff; 2175 int rcount; 2176 vm_page_t m; 2177 2178 db_printf("new object: %p\n", (void *)object); 2179 if (nl > 18) { 2180 c = cngetc(); 2181 if (c != ' ') 2182 return; 2183 nl = 0; 2184 } 2185 nl++; 2186 rcount = 0; 2187 fidx = 0; 2188 osize = object->size; 2189 if (osize > 128) 2190 osize = 128; 2191 for (idx = 0; idx < osize; idx++) { 2192 m = vm_page_lookup(object, idx); 2193 if (m == NULL) { 2194 if (rcount) { 2195 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2196 (long)fidx, rcount, (long)pa); 2197 if (nl > 18) { 2198 c = cngetc(); 2199 if (c != ' ') 2200 return; 2201 nl = 0; 2202 } 2203 nl++; 2204 rcount = 0; 2205 } 2206 continue; 2207 } 2208 2209 2210 if (rcount && 2211 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2212 ++rcount; 2213 continue; 2214 } 2215 if (rcount) { 2216 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2217 padiff >>= PAGE_SHIFT; 2218 padiff &= PQ_L2_MASK; 2219 if (padiff == 0) { 2220 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2221 ++rcount; 2222 continue; 2223 } 2224 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2225 (long)fidx, rcount, (long)pa); 2226 db_printf("pd(%ld)\n", (long)padiff); 2227 if (nl > 18) { 2228 c = cngetc(); 2229 if (c != ' ') 2230 return; 2231 nl = 0; 2232 } 2233 nl++; 2234 } 2235 fidx = idx; 2236 pa = VM_PAGE_TO_PHYS(m); 2237 rcount = 1; 2238 } 2239 if (rcount) { 2240 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2241 (long)fidx, rcount, (long)pa); 2242 if (nl > 18) { 2243 c = cngetc(); 2244 if (c != ' ') 2245 return; 2246 nl = 0; 2247 } 2248 nl++; 2249 } 2250 } 2251 } 2252 #endif /* DDB */ 2253