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 vp->v_flag &= ~VTEXT; 419 #ifdef ENABLE_VFS_IOOPT 420 vm_object_clear_flag(object, OBJ_OPT); 421 #endif 422 } 423 /* 424 * vrele may need a vop lock 425 */ 426 vrele(vp); 427 } 428 429 /* 430 * vm_object_deallocate: 431 * 432 * Release a reference to the specified object, 433 * gained either through a vm_object_allocate 434 * or a vm_object_reference call. When all references 435 * are gone, storage associated with this object 436 * may be relinquished. 437 * 438 * No object may be locked. 439 */ 440 void 441 vm_object_deallocate(vm_object_t object) 442 { 443 vm_object_t temp; 444 445 mtx_lock(&Giant); 446 while (object != NULL) { 447 448 if (object->type == OBJT_VNODE) { 449 vm_object_vndeallocate(object); 450 mtx_unlock(&Giant); 451 return; 452 } 453 454 KASSERT(object->ref_count != 0, 455 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 456 457 /* 458 * If the reference count goes to 0 we start calling 459 * vm_object_terminate() on the object chain. 460 * A ref count of 1 may be a special case depending on the 461 * shadow count being 0 or 1. 462 */ 463 object->ref_count--; 464 if (object->ref_count > 1) { 465 mtx_unlock(&Giant); 466 return; 467 } else if (object->ref_count == 1) { 468 if (object->shadow_count == 0) { 469 vm_object_set_flag(object, OBJ_ONEMAPPING); 470 } else if ((object->shadow_count == 1) && 471 (object->handle == NULL) && 472 (object->type == OBJT_DEFAULT || 473 object->type == OBJT_SWAP)) { 474 vm_object_t robject; 475 476 robject = TAILQ_FIRST(&object->shadow_head); 477 KASSERT(robject != NULL, 478 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 479 object->ref_count, 480 object->shadow_count)); 481 if ((robject->handle == NULL) && 482 (robject->type == OBJT_DEFAULT || 483 robject->type == OBJT_SWAP)) { 484 485 robject->ref_count++; 486 487 while ( 488 robject->paging_in_progress || 489 object->paging_in_progress 490 ) { 491 vm_object_pip_sleep(robject, "objde1"); 492 vm_object_pip_sleep(object, "objde2"); 493 } 494 495 if (robject->ref_count == 1) { 496 robject->ref_count--; 497 object = robject; 498 goto doterm; 499 } 500 501 object = robject; 502 vm_object_collapse(object); 503 continue; 504 } 505 } 506 mtx_unlock(&Giant); 507 return; 508 } 509 doterm: 510 temp = object->backing_object; 511 if (temp) { 512 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 513 temp->shadow_count--; 514 #ifdef ENABLE_VFS_IOOPT 515 if (temp->ref_count == 0) 516 vm_object_clear_flag(temp, OBJ_OPT); 517 #endif 518 temp->generation++; 519 object->backing_object = NULL; 520 } 521 /* 522 * Don't double-terminate, we could be in a termination 523 * recursion due to the terminate having to sync data 524 * to disk. 525 */ 526 if ((object->flags & OBJ_DEAD) == 0) 527 vm_object_terminate(object); 528 object = temp; 529 } 530 mtx_unlock(&Giant); 531 } 532 533 /* 534 * vm_object_terminate actually destroys the specified object, freeing 535 * up all previously used resources. 536 * 537 * The object must be locked. 538 * This routine may block. 539 */ 540 void 541 vm_object_terminate(vm_object_t object) 542 { 543 vm_page_t p; 544 int s; 545 546 GIANT_REQUIRED; 547 548 /* 549 * Make sure no one uses us. 550 */ 551 vm_object_set_flag(object, OBJ_DEAD); 552 553 /* 554 * wait for the pageout daemon to be done with the object 555 */ 556 vm_object_pip_wait(object, "objtrm"); 557 558 KASSERT(!object->paging_in_progress, 559 ("vm_object_terminate: pageout in progress")); 560 561 /* 562 * Clean and free the pages, as appropriate. All references to the 563 * object are gone, so we don't need to lock it. 564 */ 565 if (object->type == OBJT_VNODE) { 566 struct vnode *vp; 567 568 #ifdef ENABLE_VFS_IOOPT 569 /* 570 * Freeze optimized copies. 571 */ 572 vm_freeze_copyopts(object, 0, object->size); 573 #endif 574 /* 575 * Clean pages and flush buffers. 576 */ 577 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 578 579 vp = (struct vnode *) object->handle; 580 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 581 } 582 583 KASSERT(object->ref_count == 0, 584 ("vm_object_terminate: object with references, ref_count=%d", 585 object->ref_count)); 586 587 /* 588 * Now free any remaining pages. For internal objects, this also 589 * removes them from paging queues. Don't free wired pages, just 590 * remove them from the object. 591 */ 592 s = splvm(); 593 vm_page_lock_queues(); 594 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 595 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 596 ("vm_object_terminate: freeing busy page %p " 597 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 598 if (p->wire_count == 0) { 599 vm_page_busy(p); 600 vm_page_free(p); 601 cnt.v_pfree++; 602 } else { 603 vm_page_busy(p); 604 vm_page_remove(p); 605 } 606 } 607 vm_page_unlock_queues(); 608 splx(s); 609 610 /* 611 * Let the pager know object is dead. 612 */ 613 vm_pager_deallocate(object); 614 615 /* 616 * Remove the object from the global object list. 617 */ 618 mtx_lock(&vm_object_list_mtx); 619 TAILQ_REMOVE(&vm_object_list, object, object_list); 620 mtx_unlock(&vm_object_list_mtx); 621 622 wakeup(object); 623 624 /* 625 * Free the space for the object. 626 */ 627 uma_zfree(obj_zone, object); 628 } 629 630 /* 631 * vm_object_page_clean 632 * 633 * Clean all dirty pages in the specified range of object. Leaves page 634 * on whatever queue it is currently on. If NOSYNC is set then do not 635 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 636 * leaving the object dirty. 637 * 638 * Odd semantics: if start == end, we clean everything. 639 * 640 * The object must be locked. 641 */ 642 void 643 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 644 { 645 vm_page_t p, np; 646 vm_pindex_t tstart, tend; 647 vm_pindex_t pi; 648 struct vnode *vp; 649 int clearobjflags; 650 int pagerflags; 651 int curgeneration; 652 653 GIANT_REQUIRED; 654 655 if (object->type != OBJT_VNODE || 656 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 657 return; 658 659 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0; 660 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 661 662 vp = object->handle; 663 664 vm_object_set_flag(object, OBJ_CLEANING); 665 666 tstart = start; 667 if (end == 0) { 668 tend = object->size; 669 } else { 670 tend = end; 671 } 672 673 /* 674 * If the caller is smart and only msync()s a range he knows is 675 * dirty, we may be able to avoid an object scan. This results in 676 * a phenominal improvement in performance. We cannot do this 677 * as a matter of course because the object may be huge - e.g. 678 * the size might be in the gigabytes or terrabytes. 679 */ 680 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 681 vm_pindex_t tscan; 682 int scanlimit; 683 int scanreset; 684 685 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 686 if (scanreset < 16) 687 scanreset = 16; 688 689 scanlimit = scanreset; 690 tscan = tstart; 691 while (tscan < tend) { 692 curgeneration = object->generation; 693 p = vm_page_lookup(object, tscan); 694 if (p == NULL || p->valid == 0 || 695 (p->queue - p->pc) == PQ_CACHE) { 696 if (--scanlimit == 0) 697 break; 698 ++tscan; 699 continue; 700 } 701 vm_page_test_dirty(p); 702 if ((p->dirty & p->valid) == 0) { 703 if (--scanlimit == 0) 704 break; 705 ++tscan; 706 continue; 707 } 708 /* 709 * If we have been asked to skip nosync pages and 710 * this is a nosync page, we can't continue. 711 */ 712 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 713 if (--scanlimit == 0) 714 break; 715 ++tscan; 716 continue; 717 } 718 scanlimit = scanreset; 719 720 /* 721 * This returns 0 if it was unable to busy the first 722 * page (i.e. had to sleep). 723 */ 724 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 725 } 726 727 /* 728 * If everything was dirty and we flushed it successfully, 729 * and the requested range is not the entire object, we 730 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 731 * return immediately. 732 */ 733 if (tscan >= tend && (tstart || tend < object->size)) { 734 vm_object_clear_flag(object, OBJ_CLEANING); 735 return; 736 } 737 } 738 739 /* 740 * Generally set CLEANCHK interlock and make the page read-only so 741 * we can then clear the object flags. 742 * 743 * However, if this is a nosync mmap then the object is likely to 744 * stay dirty so do not mess with the page and do not clear the 745 * object flags. 746 */ 747 clearobjflags = 1; 748 749 TAILQ_FOREACH(p, &object->memq, listq) { 750 vm_page_flag_set(p, PG_CLEANCHK); 751 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 752 clearobjflags = 0; 753 else 754 vm_page_protect(p, VM_PROT_READ); 755 } 756 757 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 758 struct vnode *vp; 759 760 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 761 if (object->type == OBJT_VNODE && 762 (vp = (struct vnode *)object->handle) != NULL) { 763 if (vp->v_flag & VOBJDIRTY) { 764 mtx_lock(&vp->v_interlock); 765 vp->v_flag &= ~VOBJDIRTY; 766 mtx_unlock(&vp->v_interlock); 767 } 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 if ( 1078 m->hold_count || 1079 m->wire_count || 1080 (m->flags & PG_UNMANAGED) || 1081 m->valid != VM_PAGE_BITS_ALL 1082 ) { 1083 continue; 1084 } 1085 1086 if (vm_page_sleep_busy(m, TRUE, "madvpo")) 1087 goto relookup; 1088 vm_page_lock_queues(); 1089 if (advise == MADV_WILLNEED) { 1090 vm_page_activate(m); 1091 } else if (advise == MADV_DONTNEED) { 1092 vm_page_dontneed(m); 1093 } else if (advise == MADV_FREE) { 1094 /* 1095 * Mark the page clean. This will allow the page 1096 * to be freed up by the system. However, such pages 1097 * are often reused quickly by malloc()/free() 1098 * so we do not do anything that would cause 1099 * a page fault if we can help it. 1100 * 1101 * Specifically, we do not try to actually free 1102 * the page now nor do we try to put it in the 1103 * cache (which would cause a page fault on reuse). 1104 * 1105 * But we do make the page is freeable as we 1106 * can without actually taking the step of unmapping 1107 * it. 1108 */ 1109 pmap_clear_modify(m); 1110 m->dirty = 0; 1111 m->act_count = 0; 1112 vm_page_dontneed(m); 1113 } 1114 vm_page_unlock_queues(); 1115 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1116 swap_pager_freespace(tobject, tpindex, 1); 1117 } 1118 mtx_unlock(&Giant); 1119 } 1120 1121 /* 1122 * vm_object_shadow: 1123 * 1124 * Create a new object which is backed by the 1125 * specified existing object range. The source 1126 * object reference is deallocated. 1127 * 1128 * The new object and offset into that object 1129 * are returned in the source parameters. 1130 */ 1131 void 1132 vm_object_shadow( 1133 vm_object_t *object, /* IN/OUT */ 1134 vm_ooffset_t *offset, /* IN/OUT */ 1135 vm_size_t length) 1136 { 1137 vm_object_t source; 1138 vm_object_t result; 1139 1140 source = *object; 1141 1142 mtx_lock(&Giant); 1143 /* 1144 * Don't create the new object if the old object isn't shared. 1145 */ 1146 if (source != NULL && 1147 source->ref_count == 1 && 1148 source->handle == NULL && 1149 (source->type == OBJT_DEFAULT || 1150 source->type == OBJT_SWAP)) { 1151 mtx_unlock(&Giant); 1152 return; 1153 } 1154 1155 /* 1156 * Allocate a new object with the given length 1157 */ 1158 result = vm_object_allocate(OBJT_DEFAULT, length); 1159 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1160 1161 /* 1162 * The new object shadows the source object, adding a reference to it. 1163 * Our caller changes his reference to point to the new object, 1164 * removing a reference to the source object. Net result: no change 1165 * of reference count. 1166 * 1167 * Try to optimize the result object's page color when shadowing 1168 * in order to maintain page coloring consistency in the combined 1169 * shadowed object. 1170 */ 1171 result->backing_object = source; 1172 if (source) { 1173 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1174 source->shadow_count++; 1175 source->generation++; 1176 if (length < source->size) 1177 length = source->size; 1178 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1179 source->generation > 1) 1180 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1181 result->pg_color = (source->pg_color + 1182 length * source->generation) & PQ_L2_MASK; 1183 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1184 PQ_L2_MASK; 1185 } 1186 1187 /* 1188 * Store the offset into the source object, and fix up the offset into 1189 * the new object. 1190 */ 1191 result->backing_object_offset = *offset; 1192 1193 /* 1194 * Return the new things 1195 */ 1196 *offset = 0; 1197 *object = result; 1198 1199 mtx_unlock(&Giant); 1200 } 1201 1202 /* 1203 * vm_object_split: 1204 * 1205 * Split the pages in a map entry into a new object. This affords 1206 * easier removal of unused pages, and keeps object inheritance from 1207 * being a negative impact on memory usage. 1208 */ 1209 void 1210 vm_object_split(vm_map_entry_t entry) 1211 { 1212 vm_page_t m; 1213 vm_object_t orig_object, new_object, source; 1214 vm_offset_t s, e; 1215 vm_pindex_t offidxstart, offidxend; 1216 vm_size_t idx, size; 1217 vm_ooffset_t offset; 1218 1219 GIANT_REQUIRED; 1220 1221 orig_object = entry->object.vm_object; 1222 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1223 return; 1224 if (orig_object->ref_count <= 1) 1225 return; 1226 1227 offset = entry->offset; 1228 s = entry->start; 1229 e = entry->end; 1230 1231 offidxstart = OFF_TO_IDX(offset); 1232 offidxend = offidxstart + OFF_TO_IDX(e - s); 1233 size = offidxend - offidxstart; 1234 1235 new_object = vm_pager_allocate(orig_object->type, 1236 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1237 if (new_object == NULL) 1238 return; 1239 1240 source = orig_object->backing_object; 1241 if (source != NULL) { 1242 vm_object_reference(source); /* Referenced by new_object */ 1243 TAILQ_INSERT_TAIL(&source->shadow_head, 1244 new_object, shadow_list); 1245 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1246 new_object->backing_object_offset = 1247 orig_object->backing_object_offset + offset; 1248 new_object->backing_object = source; 1249 source->shadow_count++; 1250 source->generation++; 1251 } 1252 for (idx = 0; idx < size; idx++) { 1253 retry: 1254 m = vm_page_lookup(orig_object, offidxstart + idx); 1255 if (m == NULL) 1256 continue; 1257 1258 /* 1259 * We must wait for pending I/O to complete before we can 1260 * rename the page. 1261 * 1262 * We do not have to VM_PROT_NONE the page as mappings should 1263 * not be changed by this operation. 1264 */ 1265 if (vm_page_sleep_busy(m, TRUE, "spltwt")) 1266 goto retry; 1267 1268 vm_page_busy(m); 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 if (op & OBSC_COLLAPSE_NOWAIT) { 1385 if ( 1386 (p->flags & PG_BUSY) || 1387 !p->valid || 1388 p->hold_count || 1389 p->wire_count || 1390 p->busy 1391 ) { 1392 p = next; 1393 continue; 1394 } 1395 } else if (op & OBSC_COLLAPSE_WAIT) { 1396 if (vm_page_sleep_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 1414 KASSERT( 1415 p->object == backing_object, 1416 ("vm_object_qcollapse(): object mismatch") 1417 ); 1418 1419 /* 1420 * Destroy any associated swap 1421 */ 1422 if (backing_object->type == OBJT_SWAP) { 1423 swap_pager_freespace( 1424 backing_object, 1425 p->pindex, 1426 1 1427 ); 1428 } 1429 1430 if ( 1431 p->pindex < backing_offset_index || 1432 new_pindex >= object->size 1433 ) { 1434 /* 1435 * Page is out of the parent object's range, we 1436 * can simply destroy it. 1437 */ 1438 vm_page_lock_queues(); 1439 vm_page_protect(p, VM_PROT_NONE); 1440 vm_page_free(p); 1441 vm_page_unlock_queues(); 1442 p = next; 1443 continue; 1444 } 1445 1446 pp = vm_page_lookup(object, new_pindex); 1447 if ( 1448 pp != NULL || 1449 vm_pager_has_page(object, new_pindex, NULL, NULL) 1450 ) { 1451 /* 1452 * page already exists in parent OR swap exists 1453 * for this location in the parent. Destroy 1454 * the original page from the backing object. 1455 * 1456 * Leave the parent's page alone 1457 */ 1458 vm_page_lock_queues(); 1459 vm_page_protect(p, VM_PROT_NONE); 1460 vm_page_free(p); 1461 vm_page_unlock_queues(); 1462 p = next; 1463 continue; 1464 } 1465 1466 /* 1467 * Page does not exist in parent, rename the 1468 * page from the backing object to the main object. 1469 * 1470 * If the page was mapped to a process, it can remain 1471 * mapped through the rename. 1472 */ 1473 vm_page_rename(p, object, new_pindex); 1474 /* page automatically made dirty by rename */ 1475 } 1476 p = next; 1477 } 1478 splx(s); 1479 return (r); 1480 } 1481 1482 1483 /* 1484 * this version of collapse allows the operation to occur earlier and 1485 * when paging_in_progress is true for an object... This is not a complete 1486 * operation, but should plug 99.9% of the rest of the leaks. 1487 */ 1488 static void 1489 vm_object_qcollapse(vm_object_t object) 1490 { 1491 vm_object_t backing_object = object->backing_object; 1492 1493 GIANT_REQUIRED; 1494 1495 if (backing_object->ref_count != 1) 1496 return; 1497 1498 backing_object->ref_count += 2; 1499 1500 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1501 1502 backing_object->ref_count -= 2; 1503 } 1504 1505 /* 1506 * vm_object_collapse: 1507 * 1508 * Collapse an object with the object backing it. 1509 * Pages in the backing object are moved into the 1510 * parent, and the backing object is deallocated. 1511 */ 1512 void 1513 vm_object_collapse(vm_object_t object) 1514 { 1515 GIANT_REQUIRED; 1516 1517 while (TRUE) { 1518 vm_object_t backing_object; 1519 1520 /* 1521 * Verify that the conditions are right for collapse: 1522 * 1523 * The object exists and the backing object exists. 1524 */ 1525 if (object == NULL) 1526 break; 1527 1528 if ((backing_object = object->backing_object) == NULL) 1529 break; 1530 1531 /* 1532 * we check the backing object first, because it is most likely 1533 * not collapsable. 1534 */ 1535 if (backing_object->handle != NULL || 1536 (backing_object->type != OBJT_DEFAULT && 1537 backing_object->type != OBJT_SWAP) || 1538 (backing_object->flags & OBJ_DEAD) || 1539 object->handle != NULL || 1540 (object->type != OBJT_DEFAULT && 1541 object->type != OBJT_SWAP) || 1542 (object->flags & OBJ_DEAD)) { 1543 break; 1544 } 1545 1546 if ( 1547 object->paging_in_progress != 0 || 1548 backing_object->paging_in_progress != 0 1549 ) { 1550 vm_object_qcollapse(object); 1551 break; 1552 } 1553 1554 /* 1555 * We know that we can either collapse the backing object (if 1556 * the parent is the only reference to it) or (perhaps) have 1557 * the parent bypass the object if the parent happens to shadow 1558 * all the resident pages in the entire backing object. 1559 * 1560 * This is ignoring pager-backed pages such as swap pages. 1561 * vm_object_backing_scan fails the shadowing test in this 1562 * case. 1563 */ 1564 if (backing_object->ref_count == 1) { 1565 /* 1566 * If there is exactly one reference to the backing 1567 * object, we can collapse it into the parent. 1568 */ 1569 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1570 1571 /* 1572 * Move the pager from backing_object to object. 1573 */ 1574 if (backing_object->type == OBJT_SWAP) { 1575 vm_object_pip_add(backing_object, 1); 1576 1577 /* 1578 * scrap the paging_offset junk and do a 1579 * discrete copy. This also removes major 1580 * assumptions about how the swap-pager 1581 * works from where it doesn't belong. The 1582 * new swapper is able to optimize the 1583 * destroy-source case. 1584 */ 1585 vm_object_pip_add(object, 1); 1586 swap_pager_copy( 1587 backing_object, 1588 object, 1589 OFF_TO_IDX(object->backing_object_offset), TRUE); 1590 vm_object_pip_wakeup(object); 1591 1592 vm_object_pip_wakeup(backing_object); 1593 } 1594 /* 1595 * Object now shadows whatever backing_object did. 1596 * Note that the reference to 1597 * backing_object->backing_object moves from within 1598 * backing_object to within object. 1599 */ 1600 TAILQ_REMOVE( 1601 &object->backing_object->shadow_head, 1602 object, 1603 shadow_list 1604 ); 1605 object->backing_object->shadow_count--; 1606 object->backing_object->generation++; 1607 if (backing_object->backing_object) { 1608 TAILQ_REMOVE( 1609 &backing_object->backing_object->shadow_head, 1610 backing_object, 1611 shadow_list 1612 ); 1613 backing_object->backing_object->shadow_count--; 1614 backing_object->backing_object->generation++; 1615 } 1616 object->backing_object = backing_object->backing_object; 1617 if (object->backing_object) { 1618 TAILQ_INSERT_TAIL( 1619 &object->backing_object->shadow_head, 1620 object, 1621 shadow_list 1622 ); 1623 object->backing_object->shadow_count++; 1624 object->backing_object->generation++; 1625 } 1626 1627 object->backing_object_offset += 1628 backing_object->backing_object_offset; 1629 1630 /* 1631 * Discard backing_object. 1632 * 1633 * Since the backing object has no pages, no pager left, 1634 * and no object references within it, all that is 1635 * necessary is to dispose of it. 1636 */ 1637 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1638 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1639 1640 mtx_lock(&vm_object_list_mtx); 1641 TAILQ_REMOVE( 1642 &vm_object_list, 1643 backing_object, 1644 object_list 1645 ); 1646 mtx_unlock(&vm_object_list_mtx); 1647 1648 uma_zfree(obj_zone, backing_object); 1649 1650 object_collapses++; 1651 } else { 1652 vm_object_t new_backing_object; 1653 1654 /* 1655 * If we do not entirely shadow the backing object, 1656 * there is nothing we can do so we give up. 1657 */ 1658 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1659 break; 1660 } 1661 1662 /* 1663 * Make the parent shadow the next object in the 1664 * chain. Deallocating backing_object will not remove 1665 * it, since its reference count is at least 2. 1666 */ 1667 TAILQ_REMOVE( 1668 &backing_object->shadow_head, 1669 object, 1670 shadow_list 1671 ); 1672 backing_object->shadow_count--; 1673 backing_object->generation++; 1674 1675 new_backing_object = backing_object->backing_object; 1676 if ((object->backing_object = new_backing_object) != NULL) { 1677 vm_object_reference(new_backing_object); 1678 TAILQ_INSERT_TAIL( 1679 &new_backing_object->shadow_head, 1680 object, 1681 shadow_list 1682 ); 1683 new_backing_object->shadow_count++; 1684 new_backing_object->generation++; 1685 object->backing_object_offset += 1686 backing_object->backing_object_offset; 1687 } 1688 1689 /* 1690 * Drop the reference count on backing_object. Since 1691 * its ref_count was at least 2, it will not vanish; 1692 * so we don't need to call vm_object_deallocate, but 1693 * we do anyway. 1694 */ 1695 vm_object_deallocate(backing_object); 1696 object_bypasses++; 1697 } 1698 1699 /* 1700 * Try again with this object's new backing object. 1701 */ 1702 } 1703 } 1704 1705 /* 1706 * vm_object_page_remove: [internal] 1707 * 1708 * Removes all physical pages in the specified 1709 * object range from the object's list of pages. 1710 * 1711 * The object must be locked. 1712 */ 1713 void 1714 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1715 { 1716 vm_page_t p, next; 1717 vm_pindex_t size; 1718 int all; 1719 1720 if (object == NULL) 1721 return; 1722 1723 mtx_lock(&Giant); 1724 if (object->resident_page_count == 0) { 1725 mtx_unlock(&Giant); 1726 return; 1727 } 1728 all = ((end == 0) && (start == 0)); 1729 1730 /* 1731 * Since physically-backed objects do not use managed pages, we can't 1732 * remove pages from the object (we must instead remove the page 1733 * references, and then destroy the object). 1734 */ 1735 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1736 1737 vm_object_pip_add(object, 1); 1738 again: 1739 size = end - start; 1740 if (all || size > object->resident_page_count / 4) { 1741 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) { 1742 next = TAILQ_NEXT(p, listq); 1743 if (all || ((start <= p->pindex) && (p->pindex < end))) { 1744 if (p->wire_count != 0) { 1745 vm_page_protect(p, VM_PROT_NONE); 1746 if (!clean_only) 1747 p->valid = 0; 1748 continue; 1749 } 1750 1751 /* 1752 * The busy flags are only cleared at 1753 * interrupt -- minimize the spl transitions 1754 */ 1755 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1756 goto again; 1757 1758 if (clean_only && p->valid) { 1759 vm_page_test_dirty(p); 1760 if (p->valid & p->dirty) 1761 continue; 1762 } 1763 vm_page_lock_queues(); 1764 vm_page_busy(p); 1765 vm_page_protect(p, VM_PROT_NONE); 1766 vm_page_free(p); 1767 vm_page_unlock_queues(); 1768 } 1769 } 1770 } else { 1771 while (size > 0) { 1772 if ((p = vm_page_lookup(object, start)) != 0) { 1773 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_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_lock_queues(); 1799 vm_page_busy(p); 1800 vm_page_protect(p, VM_PROT_NONE); 1801 vm_page_free(p); 1802 vm_page_unlock_queues(); 1803 } 1804 start += 1; 1805 size -= 1; 1806 } 1807 } 1808 vm_object_pip_wakeup(object); 1809 mtx_unlock(&Giant); 1810 } 1811 1812 /* 1813 * Routine: vm_object_coalesce 1814 * Function: Coalesces two objects backing up adjoining 1815 * regions of memory into a single object. 1816 * 1817 * returns TRUE if objects were combined. 1818 * 1819 * NOTE: Only works at the moment if the second object is NULL - 1820 * if it's not, which object do we lock first? 1821 * 1822 * Parameters: 1823 * prev_object First object to coalesce 1824 * prev_offset Offset into prev_object 1825 * next_object Second object into coalesce 1826 * next_offset Offset into next_object 1827 * 1828 * prev_size Size of reference to prev_object 1829 * next_size Size of reference to next_object 1830 * 1831 * Conditions: 1832 * The object must *not* be locked. 1833 */ 1834 boolean_t 1835 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1836 vm_size_t prev_size, vm_size_t next_size) 1837 { 1838 vm_pindex_t next_pindex; 1839 1840 if (prev_object == NULL) 1841 return (TRUE); 1842 mtx_lock(&Giant); 1843 if (prev_object->type != OBJT_DEFAULT && 1844 prev_object->type != OBJT_SWAP) { 1845 mtx_unlock(&Giant); 1846 return (FALSE); 1847 } 1848 1849 /* 1850 * Try to collapse the object first 1851 */ 1852 vm_object_collapse(prev_object); 1853 1854 /* 1855 * Can't coalesce if: . more than one reference . paged out . shadows 1856 * another object . has a copy elsewhere (any of which mean that the 1857 * pages not mapped to prev_entry may be in use anyway) 1858 */ 1859 if (prev_object->backing_object != NULL) { 1860 mtx_unlock(&Giant); 1861 return (FALSE); 1862 } 1863 1864 prev_size >>= PAGE_SHIFT; 1865 next_size >>= PAGE_SHIFT; 1866 next_pindex = prev_pindex + prev_size; 1867 1868 if ((prev_object->ref_count > 1) && 1869 (prev_object->size != next_pindex)) { 1870 mtx_unlock(&Giant); 1871 return (FALSE); 1872 } 1873 1874 /* 1875 * Remove any pages that may still be in the object from a previous 1876 * deallocation. 1877 */ 1878 if (next_pindex < prev_object->size) { 1879 vm_object_page_remove(prev_object, 1880 next_pindex, 1881 next_pindex + next_size, FALSE); 1882 if (prev_object->type == OBJT_SWAP) 1883 swap_pager_freespace(prev_object, 1884 next_pindex, next_size); 1885 } 1886 1887 /* 1888 * Extend the object if necessary. 1889 */ 1890 if (next_pindex + next_size > prev_object->size) 1891 prev_object->size = next_pindex + next_size; 1892 1893 mtx_unlock(&Giant); 1894 return (TRUE); 1895 } 1896 1897 void 1898 vm_object_set_writeable_dirty(vm_object_t object) 1899 { 1900 struct vnode *vp; 1901 1902 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1903 if (object->type == OBJT_VNODE && 1904 (vp = (struct vnode *)object->handle) != NULL) { 1905 if ((vp->v_flag & VOBJDIRTY) == 0) { 1906 mtx_lock(&vp->v_interlock); 1907 vp->v_flag |= VOBJDIRTY; 1908 mtx_unlock(&vp->v_interlock); 1909 } 1910 } 1911 } 1912 1913 #ifdef ENABLE_VFS_IOOPT 1914 /* 1915 * Experimental support for zero-copy I/O 1916 * 1917 * Performs the copy_on_write operations necessary to allow the virtual copies 1918 * into user space to work. This has to be called for write(2) system calls 1919 * from other processes, file unlinking, and file size shrinkage. 1920 */ 1921 void 1922 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 1923 { 1924 int rv; 1925 vm_object_t robject; 1926 vm_pindex_t idx; 1927 1928 GIANT_REQUIRED; 1929 if ((object == NULL) || 1930 ((object->flags & OBJ_OPT) == 0)) 1931 return; 1932 1933 if (object->shadow_count > object->ref_count) 1934 panic("vm_freeze_copyopts: sc > rc"); 1935 1936 while ((robject = TAILQ_FIRST(&object->shadow_head)) != NULL) { 1937 vm_pindex_t bo_pindex; 1938 vm_page_t m_in, m_out; 1939 1940 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 1941 1942 vm_object_reference(robject); 1943 1944 vm_object_pip_wait(robject, "objfrz"); 1945 1946 if (robject->ref_count == 1) { 1947 vm_object_deallocate(robject); 1948 continue; 1949 } 1950 1951 vm_object_pip_add(robject, 1); 1952 1953 for (idx = 0; idx < robject->size; idx++) { 1954 1955 m_out = vm_page_grab(robject, idx, 1956 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1957 1958 if (m_out->valid == 0) { 1959 m_in = vm_page_grab(object, bo_pindex + idx, 1960 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1961 if (m_in->valid == 0) { 1962 rv = vm_pager_get_pages(object, &m_in, 1, 0); 1963 if (rv != VM_PAGER_OK) { 1964 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 1965 continue; 1966 } 1967 vm_page_deactivate(m_in); 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_activate(m_out); 1975 vm_page_wakeup(m_in); 1976 } 1977 vm_page_wakeup(m_out); 1978 } 1979 1980 object->shadow_count--; 1981 object->ref_count--; 1982 TAILQ_REMOVE(&object->shadow_head, robject, shadow_list); 1983 robject->backing_object = NULL; 1984 robject->backing_object_offset = 0; 1985 1986 vm_object_pip_wakeup(robject); 1987 vm_object_deallocate(robject); 1988 } 1989 1990 vm_object_clear_flag(object, OBJ_OPT); 1991 } 1992 #endif 1993 1994 #include "opt_ddb.h" 1995 #ifdef DDB 1996 #include <sys/kernel.h> 1997 1998 #include <sys/cons.h> 1999 2000 #include <ddb/ddb.h> 2001 2002 static int 2003 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2004 { 2005 vm_map_t tmpm; 2006 vm_map_entry_t tmpe; 2007 vm_object_t obj; 2008 int entcount; 2009 2010 if (map == 0) 2011 return 0; 2012 2013 if (entry == 0) { 2014 tmpe = map->header.next; 2015 entcount = map->nentries; 2016 while (entcount-- && (tmpe != &map->header)) { 2017 if (_vm_object_in_map(map, object, tmpe)) { 2018 return 1; 2019 } 2020 tmpe = tmpe->next; 2021 } 2022 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2023 tmpm = entry->object.sub_map; 2024 tmpe = tmpm->header.next; 2025 entcount = tmpm->nentries; 2026 while (entcount-- && tmpe != &tmpm->header) { 2027 if (_vm_object_in_map(tmpm, object, tmpe)) { 2028 return 1; 2029 } 2030 tmpe = tmpe->next; 2031 } 2032 } else if ((obj = entry->object.vm_object) != NULL) { 2033 for (; obj; obj = obj->backing_object) 2034 if (obj == object) { 2035 return 1; 2036 } 2037 } 2038 return 0; 2039 } 2040 2041 static int 2042 vm_object_in_map(vm_object_t object) 2043 { 2044 struct proc *p; 2045 2046 /* sx_slock(&allproc_lock); */ 2047 LIST_FOREACH(p, &allproc, p_list) { 2048 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2049 continue; 2050 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2051 /* sx_sunlock(&allproc_lock); */ 2052 return 1; 2053 } 2054 } 2055 /* sx_sunlock(&allproc_lock); */ 2056 if (_vm_object_in_map(kernel_map, object, 0)) 2057 return 1; 2058 if (_vm_object_in_map(kmem_map, object, 0)) 2059 return 1; 2060 if (_vm_object_in_map(pager_map, object, 0)) 2061 return 1; 2062 if (_vm_object_in_map(buffer_map, object, 0)) 2063 return 1; 2064 return 0; 2065 } 2066 2067 DB_SHOW_COMMAND(vmochk, vm_object_check) 2068 { 2069 vm_object_t object; 2070 2071 /* 2072 * make sure that internal objs are in a map somewhere 2073 * and none have zero ref counts. 2074 */ 2075 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2076 if (object->handle == NULL && 2077 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2078 if (object->ref_count == 0) { 2079 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2080 (long)object->size); 2081 } 2082 if (!vm_object_in_map(object)) { 2083 db_printf( 2084 "vmochk: internal obj is not in a map: " 2085 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2086 object->ref_count, (u_long)object->size, 2087 (u_long)object->size, 2088 (void *)object->backing_object); 2089 } 2090 } 2091 } 2092 } 2093 2094 /* 2095 * vm_object_print: [ debug ] 2096 */ 2097 DB_SHOW_COMMAND(object, vm_object_print_static) 2098 { 2099 /* XXX convert args. */ 2100 vm_object_t object = (vm_object_t)addr; 2101 boolean_t full = have_addr; 2102 2103 vm_page_t p; 2104 2105 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2106 #define count was_count 2107 2108 int count; 2109 2110 if (object == NULL) 2111 return; 2112 2113 db_iprintf( 2114 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 2115 object, (int)object->type, (u_long)object->size, 2116 object->resident_page_count, object->ref_count, object->flags); 2117 /* 2118 * XXX no %qd in kernel. Truncate object->backing_object_offset. 2119 */ 2120 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2121 object->shadow_count, 2122 object->backing_object ? object->backing_object->ref_count : 0, 2123 object->backing_object, (long)object->backing_object_offset); 2124 2125 if (!full) 2126 return; 2127 2128 db_indent += 2; 2129 count = 0; 2130 TAILQ_FOREACH(p, &object->memq, listq) { 2131 if (count == 0) 2132 db_iprintf("memory:="); 2133 else if (count == 6) { 2134 db_printf("\n"); 2135 db_iprintf(" ..."); 2136 count = 0; 2137 } else 2138 db_printf(","); 2139 count++; 2140 2141 db_printf("(off=0x%lx,page=0x%lx)", 2142 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2143 } 2144 if (count != 0) 2145 db_printf("\n"); 2146 db_indent -= 2; 2147 } 2148 2149 /* XXX. */ 2150 #undef count 2151 2152 /* XXX need this non-static entry for calling from vm_map_print. */ 2153 void 2154 vm_object_print( 2155 /* db_expr_t */ long addr, 2156 boolean_t have_addr, 2157 /* db_expr_t */ long count, 2158 char *modif) 2159 { 2160 vm_object_print_static(addr, have_addr, count, modif); 2161 } 2162 2163 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2164 { 2165 vm_object_t object; 2166 int nl = 0; 2167 int c; 2168 2169 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2170 vm_pindex_t idx, fidx; 2171 vm_pindex_t osize; 2172 vm_offset_t pa = -1, padiff; 2173 int rcount; 2174 vm_page_t m; 2175 2176 db_printf("new object: %p\n", (void *)object); 2177 if (nl > 18) { 2178 c = cngetc(); 2179 if (c != ' ') 2180 return; 2181 nl = 0; 2182 } 2183 nl++; 2184 rcount = 0; 2185 fidx = 0; 2186 osize = object->size; 2187 if (osize > 128) 2188 osize = 128; 2189 for (idx = 0; idx < osize; idx++) { 2190 m = vm_page_lookup(object, idx); 2191 if (m == NULL) { 2192 if (rcount) { 2193 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2194 (long)fidx, rcount, (long)pa); 2195 if (nl > 18) { 2196 c = cngetc(); 2197 if (c != ' ') 2198 return; 2199 nl = 0; 2200 } 2201 nl++; 2202 rcount = 0; 2203 } 2204 continue; 2205 } 2206 2207 2208 if (rcount && 2209 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2210 ++rcount; 2211 continue; 2212 } 2213 if (rcount) { 2214 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2215 padiff >>= PAGE_SHIFT; 2216 padiff &= PQ_L2_MASK; 2217 if (padiff == 0) { 2218 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2219 ++rcount; 2220 continue; 2221 } 2222 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2223 (long)fidx, rcount, (long)pa); 2224 db_printf("pd(%ld)\n", (long)padiff); 2225 if (nl > 18) { 2226 c = cngetc(); 2227 if (c != ' ') 2228 return; 2229 nl = 0; 2230 } 2231 nl++; 2232 } 2233 fidx = idx; 2234 pa = VM_PAGE_TO_PHYS(m); 2235 rcount = 1; 2236 } 2237 if (rcount) { 2238 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2239 (long)fidx, rcount, (long)pa); 2240 if (nl > 18) { 2241 c = cngetc(); 2242 if (c != ' ') 2243 return; 2244 nl = 0; 2245 } 2246 nl++; 2247 } 2248 } 2249 } 2250 #endif /* DDB */ 2251