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