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