1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD$ 65 */ 66 67 /* 68 * Virtual memory object module. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/lock.h> 74 #include <sys/mman.h> 75 #include <sys/mount.h> 76 #include <sys/kernel.h> 77 #include <sys/sysctl.h> 78 #include <sys/mutex.h> 79 #include <sys/proc.h> /* for curproc, pageproc */ 80 #include <sys/socket.h> 81 #include <sys/vnode.h> 82 #include <sys/vmmeter.h> 83 #include <sys/sx.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pageout.h> 92 #include <vm/vm_pager.h> 93 #include <vm/swap_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/uma.h> 97 98 #define EASY_SCAN_FACTOR 8 99 100 #define MSYNC_FLUSH_HARDSEQ 0x01 101 #define MSYNC_FLUSH_SOFTSEQ 0x02 102 103 /* 104 * msync / VM object flushing optimizations 105 */ 106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 108 CTLFLAG_RW, &msync_flush_flags, 0, ""); 109 110 static void vm_object_qcollapse(vm_object_t object); 111 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 112 113 /* 114 * Virtual memory objects maintain the actual data 115 * associated with allocated virtual memory. A given 116 * page of memory exists within exactly one object. 117 * 118 * An object is only deallocated when all "references" 119 * are given up. Only one "reference" to a given 120 * region of an object should be writeable. 121 * 122 * Associated with each object is a list of all resident 123 * memory pages belonging to that object; this list is 124 * maintained by the "vm_page" module, and locked by the object's 125 * lock. 126 * 127 * Each object also records a "pager" routine which is 128 * used to retrieve (and store) pages to the proper backing 129 * storage. In addition, objects may be backed by other 130 * objects from which they were virtual-copied. 131 * 132 * The only items within the object structure which are 133 * modified after time of creation are: 134 * reference count locked by object's lock 135 * pager routine locked by object's lock 136 * 137 */ 138 139 struct object_q vm_object_list; 140 struct mtx vm_object_list_mtx; /* lock for object list and count */ 141 vm_object_t kernel_object; 142 vm_object_t kmem_object; 143 static struct vm_object kernel_object_store; 144 static struct vm_object kmem_object_store; 145 extern int vm_pageout_page_count; 146 147 static long object_collapses; 148 static long object_bypasses; 149 static int next_index; 150 static uma_zone_t obj_zone; 151 #define VM_OBJECTS_INIT 256 152 153 static void vm_object_zinit(void *mem, int size); 154 155 #ifdef INVARIANTS 156 static void vm_object_zdtor(void *mem, int size, void *arg); 157 158 static void 159 vm_object_zdtor(void *mem, int size, void *arg) 160 { 161 vm_object_t object; 162 163 object = (vm_object_t)mem; 164 KASSERT(object->paging_in_progress == 0, 165 ("object %p paging_in_progress = %d", 166 object, object->paging_in_progress)); 167 KASSERT(object->resident_page_count == 0, 168 ("object %p resident_page_count = %d", 169 object, object->resident_page_count)); 170 KASSERT(object->shadow_count == 0, 171 ("object %p shadow_count = %d", 172 object, object->shadow_count)); 173 } 174 #endif 175 176 static void 177 vm_object_zinit(void *mem, int size) 178 { 179 vm_object_t object; 180 181 object = (vm_object_t)mem; 182 183 /* These are true for any object that has been freed */ 184 object->paging_in_progress = 0; 185 object->resident_page_count = 0; 186 object->shadow_count = 0; 187 } 188 189 void 190 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object) 191 { 192 static int object_hash_rand; 193 int exp, incr; 194 195 TAILQ_INIT(&object->memq); 196 TAILQ_INIT(&object->shadow_head); 197 198 object->type = type; 199 object->size = size; 200 object->ref_count = 1; 201 object->flags = 0; 202 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 203 vm_object_set_flag(object, OBJ_ONEMAPPING); 204 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1)) 205 incr = PQ_L2_SIZE / 3 + PQ_PRIME1; 206 else 207 incr = size; 208 do 209 object->pg_color = next_index; 210 while (!atomic_cmpset_int(&next_index, object->pg_color, 211 (object->pg_color + incr) & PQ_L2_MASK)); 212 object->handle = NULL; 213 object->backing_object = NULL; 214 object->backing_object_offset = (vm_ooffset_t) 0; 215 /* 216 * Try to generate a number that will spread objects out in the 217 * hash table. We 'wipe' new objects across the hash in 128 page 218 * increments plus 1 more to offset it a little more by the time 219 * it wraps around. 220 */ 221 do { 222 exp = object_hash_rand; 223 object->hash_rand = exp - 129; 224 } while (!atomic_cmpset_int(&object_hash_rand, exp, object->hash_rand)); 225 226 object->generation++; /* atomicity needed? XXX */ 227 228 mtx_lock(&vm_object_list_mtx); 229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 230 mtx_unlock(&vm_object_list_mtx); 231 } 232 233 /* 234 * vm_object_init: 235 * 236 * Initialize the VM objects module. 237 */ 238 void 239 vm_object_init(void) 240 { 241 TAILQ_INIT(&vm_object_list); 242 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 243 244 kernel_object = &kernel_object_store; 245 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 246 kernel_object); 247 248 kmem_object = &kmem_object_store; 249 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 250 kmem_object); 251 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 252 #ifdef INVARIANTS 253 vm_object_zdtor, 254 #else 255 NULL, 256 #endif 257 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 258 uma_prealloc(obj_zone, VM_OBJECTS_INIT); 259 } 260 261 void 262 vm_object_init2(void) 263 { 264 } 265 266 void 267 vm_object_set_flag(vm_object_t object, u_short bits) 268 { 269 object->flags |= bits; 270 } 271 272 void 273 vm_object_clear_flag(vm_object_t object, u_short bits) 274 { 275 GIANT_REQUIRED; 276 object->flags &= ~bits; 277 } 278 279 void 280 vm_object_pip_add(vm_object_t object, short i) 281 { 282 GIANT_REQUIRED; 283 object->paging_in_progress += i; 284 } 285 286 void 287 vm_object_pip_subtract(vm_object_t object, short i) 288 { 289 GIANT_REQUIRED; 290 object->paging_in_progress -= i; 291 } 292 293 void 294 vm_object_pip_wakeup(vm_object_t object) 295 { 296 GIANT_REQUIRED; 297 object->paging_in_progress--; 298 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 299 vm_object_clear_flag(object, OBJ_PIPWNT); 300 wakeup(object); 301 } 302 } 303 304 void 305 vm_object_pip_wakeupn(vm_object_t object, short i) 306 { 307 GIANT_REQUIRED; 308 if (i) 309 object->paging_in_progress -= i; 310 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 311 vm_object_clear_flag(object, OBJ_PIPWNT); 312 wakeup(object); 313 } 314 } 315 316 void 317 vm_object_pip_sleep(vm_object_t object, char *waitid) 318 { 319 GIANT_REQUIRED; 320 if (object->paging_in_progress) { 321 int s = splvm(); 322 if (object->paging_in_progress) { 323 vm_object_set_flag(object, OBJ_PIPWNT); 324 tsleep(object, PVM, waitid, 0); 325 } 326 splx(s); 327 } 328 } 329 330 void 331 vm_object_pip_wait(vm_object_t object, char *waitid) 332 { 333 GIANT_REQUIRED; 334 while (object->paging_in_progress) 335 vm_object_pip_sleep(object, waitid); 336 } 337 338 /* 339 * vm_object_allocate: 340 * 341 * Returns a new object with the given size. 342 */ 343 vm_object_t 344 vm_object_allocate(objtype_t type, vm_size_t size) 345 { 346 vm_object_t result; 347 348 result = (vm_object_t) uma_zalloc(obj_zone, M_WAITOK); 349 _vm_object_allocate(type, size, result); 350 351 return (result); 352 } 353 354 355 /* 356 * vm_object_reference: 357 * 358 * Gets another reference to the given object. 359 */ 360 void 361 vm_object_reference(vm_object_t object) 362 { 363 if (object == NULL) 364 return; 365 366 mtx_lock(&Giant); 367 #if 0 368 /* object can be re-referenced during final cleaning */ 369 KASSERT(!(object->flags & OBJ_DEAD), 370 ("vm_object_reference: attempting to reference dead obj")); 371 #endif 372 373 object->ref_count++; 374 if (object->type == OBJT_VNODE) { 375 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) { 376 printf("vm_object_reference: delay in getting object\n"); 377 } 378 } 379 mtx_unlock(&Giant); 380 } 381 382 /* 383 * handle deallocating a object of type OBJT_VNODE 384 */ 385 void 386 vm_object_vndeallocate(vm_object_t object) 387 { 388 struct vnode *vp = (struct vnode *) object->handle; 389 390 GIANT_REQUIRED; 391 KASSERT(object->type == OBJT_VNODE, 392 ("vm_object_vndeallocate: not a vnode object")); 393 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 394 #ifdef INVARIANTS 395 if (object->ref_count == 0) { 396 vprint("vm_object_vndeallocate", vp); 397 panic("vm_object_vndeallocate: bad object reference count"); 398 } 399 #endif 400 401 object->ref_count--; 402 if (object->ref_count == 0) { 403 vp->v_flag &= ~VTEXT; 404 #ifdef ENABLE_VFS_IOOPT 405 vm_object_clear_flag(object, OBJ_OPT); 406 #endif 407 } 408 /* 409 * vrele may need a vop lock 410 */ 411 vrele(vp); 412 } 413 414 /* 415 * vm_object_deallocate: 416 * 417 * Release a reference to the specified object, 418 * gained either through a vm_object_allocate 419 * or a vm_object_reference call. When all references 420 * are gone, storage associated with this object 421 * may be relinquished. 422 * 423 * No object may be locked. 424 */ 425 void 426 vm_object_deallocate(vm_object_t object) 427 { 428 vm_object_t temp; 429 430 mtx_lock(&Giant); 431 while (object != NULL) { 432 433 if (object->type == OBJT_VNODE) { 434 vm_object_vndeallocate(object); 435 mtx_unlock(&Giant); 436 return; 437 } 438 439 KASSERT(object->ref_count != 0, 440 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 441 442 /* 443 * If the reference count goes to 0 we start calling 444 * vm_object_terminate() on the object chain. 445 * A ref count of 1 may be a special case depending on the 446 * shadow count being 0 or 1. 447 */ 448 object->ref_count--; 449 if (object->ref_count > 1) { 450 mtx_unlock(&Giant); 451 return; 452 } else if (object->ref_count == 1) { 453 if (object->shadow_count == 0) { 454 vm_object_set_flag(object, OBJ_ONEMAPPING); 455 } else if ((object->shadow_count == 1) && 456 (object->handle == NULL) && 457 (object->type == OBJT_DEFAULT || 458 object->type == OBJT_SWAP)) { 459 vm_object_t robject; 460 461 robject = TAILQ_FIRST(&object->shadow_head); 462 KASSERT(robject != NULL, 463 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 464 object->ref_count, 465 object->shadow_count)); 466 if ((robject->handle == NULL) && 467 (robject->type == OBJT_DEFAULT || 468 robject->type == OBJT_SWAP)) { 469 470 robject->ref_count++; 471 472 while ( 473 robject->paging_in_progress || 474 object->paging_in_progress 475 ) { 476 vm_object_pip_sleep(robject, "objde1"); 477 vm_object_pip_sleep(object, "objde2"); 478 } 479 480 if (robject->ref_count == 1) { 481 robject->ref_count--; 482 object = robject; 483 goto doterm; 484 } 485 486 object = robject; 487 vm_object_collapse(object); 488 continue; 489 } 490 } 491 mtx_unlock(&Giant); 492 return; 493 } 494 doterm: 495 temp = object->backing_object; 496 if (temp) { 497 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list); 498 temp->shadow_count--; 499 #ifdef ENABLE_VFS_IOOPT 500 if (temp->ref_count == 0) 501 vm_object_clear_flag(temp, OBJ_OPT); 502 #endif 503 temp->generation++; 504 object->backing_object = NULL; 505 } 506 /* 507 * Don't double-terminate, we could be in a termination 508 * recursion due to the terminate having to sync data 509 * to disk. 510 */ 511 if ((object->flags & OBJ_DEAD) == 0) 512 vm_object_terminate(object); 513 object = temp; 514 } 515 mtx_unlock(&Giant); 516 } 517 518 /* 519 * vm_object_terminate actually destroys the specified object, freeing 520 * up all previously used resources. 521 * 522 * The object must be locked. 523 * This routine may block. 524 */ 525 void 526 vm_object_terminate(vm_object_t object) 527 { 528 vm_page_t p; 529 int s; 530 531 GIANT_REQUIRED; 532 533 /* 534 * Make sure no one uses us. 535 */ 536 vm_object_set_flag(object, OBJ_DEAD); 537 538 /* 539 * wait for the pageout daemon to be done with the object 540 */ 541 vm_object_pip_wait(object, "objtrm"); 542 543 KASSERT(!object->paging_in_progress, 544 ("vm_object_terminate: pageout in progress")); 545 546 /* 547 * Clean and free the pages, as appropriate. All references to the 548 * object are gone, so we don't need to lock it. 549 */ 550 if (object->type == OBJT_VNODE) { 551 struct vnode *vp; 552 553 #ifdef ENABLE_VFS_IOOPT 554 /* 555 * Freeze optimized copies. 556 */ 557 vm_freeze_copyopts(object, 0, object->size); 558 #endif 559 /* 560 * Clean pages and flush buffers. 561 */ 562 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 563 564 vp = (struct vnode *) object->handle; 565 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 566 } 567 568 KASSERT(object->ref_count == 0, 569 ("vm_object_terminate: object with references, ref_count=%d", 570 object->ref_count)); 571 572 /* 573 * Now free any remaining pages. For internal objects, this also 574 * removes them from paging queues. Don't free wired pages, just 575 * remove them from the object. 576 */ 577 s = splvm(); 578 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 579 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 580 ("vm_object_terminate: freeing busy page %p " 581 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 582 if (p->wire_count == 0) { 583 vm_page_busy(p); 584 vm_page_free(p); 585 cnt.v_pfree++; 586 } else { 587 vm_page_busy(p); 588 vm_page_remove(p); 589 } 590 } 591 splx(s); 592 593 /* 594 * Let the pager know object is dead. 595 */ 596 vm_pager_deallocate(object); 597 598 /* 599 * Remove the object from the global object list. 600 */ 601 mtx_lock(&vm_object_list_mtx); 602 TAILQ_REMOVE(&vm_object_list, object, object_list); 603 mtx_unlock(&vm_object_list_mtx); 604 605 wakeup(object); 606 607 /* 608 * Free the space for the object. 609 */ 610 uma_zfree(obj_zone, object); 611 } 612 613 /* 614 * vm_object_page_clean 615 * 616 * Clean all dirty pages in the specified range of object. Leaves page 617 * on whatever queue it is currently on. If NOSYNC is set then do not 618 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 619 * leaving the object dirty. 620 * 621 * Odd semantics: if start == end, we clean everything. 622 * 623 * The object must be locked. 624 */ 625 void 626 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 627 { 628 vm_page_t p, np; 629 vm_offset_t tstart, tend; 630 vm_pindex_t pi; 631 struct vnode *vp; 632 int clearobjflags; 633 int pagerflags; 634 int curgeneration; 635 636 GIANT_REQUIRED; 637 638 if (object->type != OBJT_VNODE || 639 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 640 return; 641 642 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0; 643 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 644 645 vp = object->handle; 646 647 vm_object_set_flag(object, OBJ_CLEANING); 648 649 tstart = start; 650 if (end == 0) { 651 tend = object->size; 652 } else { 653 tend = end; 654 } 655 656 /* 657 * If the caller is smart and only msync()s a range he knows is 658 * dirty, we may be able to avoid an object scan. This results in 659 * a phenominal improvement in performance. We cannot do this 660 * as a matter of course because the object may be huge - e.g. 661 * the size might be in the gigabytes or terrabytes. 662 */ 663 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 664 vm_offset_t tscan; 665 int scanlimit; 666 int scanreset; 667 668 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 669 if (scanreset < 16) 670 scanreset = 16; 671 672 scanlimit = scanreset; 673 tscan = tstart; 674 while (tscan < tend) { 675 curgeneration = object->generation; 676 p = vm_page_lookup(object, tscan); 677 if (p == NULL || p->valid == 0 || 678 (p->queue - p->pc) == PQ_CACHE) { 679 if (--scanlimit == 0) 680 break; 681 ++tscan; 682 continue; 683 } 684 vm_page_test_dirty(p); 685 if ((p->dirty & p->valid) == 0) { 686 if (--scanlimit == 0) 687 break; 688 ++tscan; 689 continue; 690 } 691 /* 692 * If we have been asked to skip nosync pages and 693 * this is a nosync page, we can't continue. 694 */ 695 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 696 if (--scanlimit == 0) 697 break; 698 ++tscan; 699 continue; 700 } 701 scanlimit = scanreset; 702 703 /* 704 * This returns 0 if it was unable to busy the first 705 * page (i.e. had to sleep). 706 */ 707 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 708 } 709 710 /* 711 * If everything was dirty and we flushed it successfully, 712 * and the requested range is not the entire object, we 713 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 714 * return immediately. 715 */ 716 if (tscan >= tend && (tstart || tend < object->size)) { 717 vm_object_clear_flag(object, OBJ_CLEANING); 718 return; 719 } 720 } 721 722 /* 723 * Generally set CLEANCHK interlock and make the page read-only so 724 * we can then clear the object flags. 725 * 726 * However, if this is a nosync mmap then the object is likely to 727 * stay dirty so do not mess with the page and do not clear the 728 * object flags. 729 */ 730 clearobjflags = 1; 731 732 TAILQ_FOREACH(p, &object->memq, listq) { 733 vm_page_flag_set(p, PG_CLEANCHK); 734 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 735 clearobjflags = 0; 736 else 737 vm_page_protect(p, VM_PROT_READ); 738 } 739 740 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 741 struct vnode *vp; 742 743 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 744 if (object->type == OBJT_VNODE && 745 (vp = (struct vnode *)object->handle) != NULL) { 746 if (vp->v_flag & VOBJDIRTY) { 747 mtx_lock(&vp->v_interlock); 748 vp->v_flag &= ~VOBJDIRTY; 749 mtx_unlock(&vp->v_interlock); 750 } 751 } 752 } 753 754 rescan: 755 curgeneration = object->generation; 756 757 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 758 int n; 759 760 np = TAILQ_NEXT(p, listq); 761 762 again: 763 pi = p->pindex; 764 if (((p->flags & PG_CLEANCHK) == 0) || 765 (pi < tstart) || (pi >= tend) || 766 (p->valid == 0) || 767 ((p->queue - p->pc) == PQ_CACHE)) { 768 vm_page_flag_clear(p, PG_CLEANCHK); 769 continue; 770 } 771 772 vm_page_test_dirty(p); 773 if ((p->dirty & p->valid) == 0) { 774 vm_page_flag_clear(p, PG_CLEANCHK); 775 continue; 776 } 777 778 /* 779 * If we have been asked to skip nosync pages and this is a 780 * nosync page, skip it. Note that the object flags were 781 * not cleared in this case so we do not have to set them. 782 */ 783 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 784 vm_page_flag_clear(p, PG_CLEANCHK); 785 continue; 786 } 787 788 n = vm_object_page_collect_flush(object, p, 789 curgeneration, pagerflags); 790 if (n == 0) 791 goto rescan; 792 793 if (object->generation != curgeneration) 794 goto rescan; 795 796 /* 797 * Try to optimize the next page. If we can't we pick up 798 * our (random) scan where we left off. 799 */ 800 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 801 if ((p = vm_page_lookup(object, pi + n)) != NULL) 802 goto again; 803 } 804 } 805 806 #if 0 807 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 808 #endif 809 810 vm_object_clear_flag(object, OBJ_CLEANING); 811 return; 812 } 813 814 static int 815 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 816 { 817 int runlen; 818 int s; 819 int maxf; 820 int chkb; 821 int maxb; 822 int i; 823 vm_pindex_t pi; 824 vm_page_t maf[vm_pageout_page_count]; 825 vm_page_t mab[vm_pageout_page_count]; 826 vm_page_t ma[vm_pageout_page_count]; 827 828 s = splvm(); 829 pi = p->pindex; 830 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) { 831 if (object->generation != curgeneration) { 832 splx(s); 833 return(0); 834 } 835 } 836 837 maxf = 0; 838 for(i = 1; i < vm_pageout_page_count; i++) { 839 vm_page_t tp; 840 841 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 842 if ((tp->flags & PG_BUSY) || 843 (tp->flags & PG_CLEANCHK) == 0 || 844 (tp->busy != 0)) 845 break; 846 if((tp->queue - tp->pc) == PQ_CACHE) { 847 vm_page_flag_clear(tp, PG_CLEANCHK); 848 break; 849 } 850 vm_page_test_dirty(tp); 851 if ((tp->dirty & tp->valid) == 0) { 852 vm_page_flag_clear(tp, PG_CLEANCHK); 853 break; 854 } 855 maf[ i - 1 ] = tp; 856 maxf++; 857 continue; 858 } 859 break; 860 } 861 862 maxb = 0; 863 chkb = vm_pageout_page_count - maxf; 864 if (chkb) { 865 for(i = 1; i < chkb;i++) { 866 vm_page_t tp; 867 868 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 869 if ((tp->flags & PG_BUSY) || 870 (tp->flags & PG_CLEANCHK) == 0 || 871 (tp->busy != 0)) 872 break; 873 if ((tp->queue - tp->pc) == PQ_CACHE) { 874 vm_page_flag_clear(tp, PG_CLEANCHK); 875 break; 876 } 877 vm_page_test_dirty(tp); 878 if ((tp->dirty & tp->valid) == 0) { 879 vm_page_flag_clear(tp, PG_CLEANCHK); 880 break; 881 } 882 mab[ i - 1 ] = tp; 883 maxb++; 884 continue; 885 } 886 break; 887 } 888 } 889 890 for(i = 0; i < maxb; i++) { 891 int index = (maxb - i) - 1; 892 ma[index] = mab[i]; 893 vm_page_flag_clear(ma[index], PG_CLEANCHK); 894 } 895 vm_page_flag_clear(p, PG_CLEANCHK); 896 ma[maxb] = p; 897 for(i = 0; i < maxf; i++) { 898 int index = (maxb + i) + 1; 899 ma[index] = maf[i]; 900 vm_page_flag_clear(ma[index], PG_CLEANCHK); 901 } 902 runlen = maxb + maxf + 1; 903 904 splx(s); 905 vm_pageout_flush(ma, runlen, pagerflags); 906 for (i = 0; i < runlen; i++) { 907 if (ma[i]->valid & ma[i]->dirty) { 908 vm_page_protect(ma[i], VM_PROT_READ); 909 vm_page_flag_set(ma[i], PG_CLEANCHK); 910 911 /* 912 * maxf will end up being the actual number of pages 913 * we wrote out contiguously, non-inclusive of the 914 * first page. We do not count look-behind pages. 915 */ 916 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 917 maxf = i - maxb - 1; 918 } 919 } 920 return(maxf + 1); 921 } 922 923 /* 924 * Same as vm_object_pmap_copy, except range checking really 925 * works, and is meant for small sections of an object. 926 * 927 * This code protects resident pages by making them read-only 928 * and is typically called on a fork or split when a page 929 * is converted to copy-on-write. 930 * 931 * NOTE: If the page is already at VM_PROT_NONE, calling 932 * vm_page_protect will have no effect. 933 */ 934 void 935 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 936 { 937 vm_pindex_t idx; 938 vm_page_t p; 939 940 GIANT_REQUIRED; 941 942 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0) 943 return; 944 945 for (idx = start; idx < end; idx++) { 946 p = vm_page_lookup(object, idx); 947 if (p == NULL) 948 continue; 949 vm_page_protect(p, VM_PROT_READ); 950 } 951 } 952 953 /* 954 * vm_object_pmap_remove: 955 * 956 * Removes all physical pages in the specified 957 * object range from all physical maps. 958 * 959 * The object must *not* be locked. 960 */ 961 void 962 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 963 { 964 vm_page_t p; 965 966 GIANT_REQUIRED; 967 if (object == NULL) 968 return; 969 TAILQ_FOREACH(p, &object->memq, listq) { 970 if (p->pindex >= start && p->pindex < end) 971 vm_page_protect(p, VM_PROT_NONE); 972 } 973 if ((start == 0) && (object->size == end)) 974 vm_object_clear_flag(object, OBJ_WRITEABLE); 975 } 976 977 /* 978 * vm_object_madvise: 979 * 980 * Implements the madvise function at the object/page level. 981 * 982 * MADV_WILLNEED (any object) 983 * 984 * Activate the specified pages if they are resident. 985 * 986 * MADV_DONTNEED (any object) 987 * 988 * Deactivate the specified pages if they are resident. 989 * 990 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 991 * OBJ_ONEMAPPING only) 992 * 993 * Deactivate and clean the specified pages if they are 994 * resident. This permits the process to reuse the pages 995 * without faulting or the kernel to reclaim the pages 996 * without I/O. 997 */ 998 void 999 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1000 { 1001 vm_pindex_t end, tpindex; 1002 vm_object_t tobject; 1003 vm_page_t m; 1004 1005 if (object == NULL) 1006 return; 1007 1008 mtx_lock(&Giant); 1009 1010 end = pindex + count; 1011 1012 /* 1013 * Locate and adjust resident pages 1014 */ 1015 for (; pindex < end; pindex += 1) { 1016 relookup: 1017 tobject = object; 1018 tpindex = pindex; 1019 shadowlookup: 1020 /* 1021 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1022 * and those pages must be OBJ_ONEMAPPING. 1023 */ 1024 if (advise == MADV_FREE) { 1025 if ((tobject->type != OBJT_DEFAULT && 1026 tobject->type != OBJT_SWAP) || 1027 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1028 continue; 1029 } 1030 } 1031 1032 m = vm_page_lookup(tobject, tpindex); 1033 1034 if (m == NULL) { 1035 /* 1036 * There may be swap even if there is no backing page 1037 */ 1038 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1039 swap_pager_freespace(tobject, tpindex, 1); 1040 1041 /* 1042 * next object 1043 */ 1044 tobject = tobject->backing_object; 1045 if (tobject == NULL) 1046 continue; 1047 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1048 goto shadowlookup; 1049 } 1050 1051 /* 1052 * If the page is busy or not in a normal active state, 1053 * we skip it. If the page is not managed there are no 1054 * page queues to mess with. Things can break if we mess 1055 * with pages in any of the below states. 1056 */ 1057 if ( 1058 m->hold_count || 1059 m->wire_count || 1060 (m->flags & PG_UNMANAGED) || 1061 m->valid != VM_PAGE_BITS_ALL 1062 ) { 1063 continue; 1064 } 1065 1066 if (vm_page_sleep_busy(m, TRUE, "madvpo")) 1067 goto relookup; 1068 1069 if (advise == MADV_WILLNEED) { 1070 vm_page_activate(m); 1071 } else if (advise == MADV_DONTNEED) { 1072 vm_page_dontneed(m); 1073 } else if (advise == MADV_FREE) { 1074 /* 1075 * Mark the page clean. This will allow the page 1076 * to be freed up by the system. However, such pages 1077 * are often reused quickly by malloc()/free() 1078 * so we do not do anything that would cause 1079 * a page fault if we can help it. 1080 * 1081 * Specifically, we do not try to actually free 1082 * the page now nor do we try to put it in the 1083 * cache (which would cause a page fault on reuse). 1084 * 1085 * But we do make the page is freeable as we 1086 * can without actually taking the step of unmapping 1087 * it. 1088 */ 1089 pmap_clear_modify(m); 1090 m->dirty = 0; 1091 m->act_count = 0; 1092 vm_page_dontneed(m); 1093 if (tobject->type == OBJT_SWAP) 1094 swap_pager_freespace(tobject, tpindex, 1); 1095 } 1096 } 1097 mtx_unlock(&Giant); 1098 } 1099 1100 /* 1101 * vm_object_shadow: 1102 * 1103 * Create a new object which is backed by the 1104 * specified existing object range. The source 1105 * object reference is deallocated. 1106 * 1107 * The new object and offset into that object 1108 * are returned in the source parameters. 1109 */ 1110 void 1111 vm_object_shadow( 1112 vm_object_t *object, /* IN/OUT */ 1113 vm_ooffset_t *offset, /* IN/OUT */ 1114 vm_size_t length) 1115 { 1116 vm_object_t source; 1117 vm_object_t result; 1118 1119 GIANT_REQUIRED; 1120 source = *object; 1121 1122 /* 1123 * Don't create the new object if the old object isn't shared. 1124 */ 1125 if (source != NULL && 1126 source->ref_count == 1 && 1127 source->handle == NULL && 1128 (source->type == OBJT_DEFAULT || 1129 source->type == OBJT_SWAP)) 1130 return; 1131 1132 /* 1133 * Allocate a new object with the given length 1134 */ 1135 result = vm_object_allocate(OBJT_DEFAULT, length); 1136 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing")); 1137 1138 /* 1139 * The new object shadows the source object, adding a reference to it. 1140 * Our caller changes his reference to point to the new object, 1141 * removing a reference to the source object. Net result: no change 1142 * of reference count. 1143 * 1144 * Try to optimize the result object's page color when shadowing 1145 * in order to maintain page coloring consistency in the combined 1146 * shadowed object. 1147 */ 1148 result->backing_object = source; 1149 if (source) { 1150 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list); 1151 source->shadow_count++; 1152 source->generation++; 1153 if (length < source->size) 1154 length = source->size; 1155 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1156 source->generation > 1) 1157 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1158 result->pg_color = (source->pg_color + 1159 length * source->generation) & PQ_L2_MASK; 1160 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1161 PQ_L2_MASK; 1162 } 1163 1164 /* 1165 * Store the offset into the source object, and fix up the offset into 1166 * the new object. 1167 */ 1168 result->backing_object_offset = *offset; 1169 1170 /* 1171 * Return the new things 1172 */ 1173 *offset = 0; 1174 *object = result; 1175 } 1176 1177 #define OBSC_TEST_ALL_SHADOWED 0x0001 1178 #define OBSC_COLLAPSE_NOWAIT 0x0002 1179 #define OBSC_COLLAPSE_WAIT 0x0004 1180 1181 static __inline int 1182 vm_object_backing_scan(vm_object_t object, int op) 1183 { 1184 int s; 1185 int r = 1; 1186 vm_page_t p; 1187 vm_object_t backing_object; 1188 vm_pindex_t backing_offset_index; 1189 1190 s = splvm(); 1191 GIANT_REQUIRED; 1192 1193 backing_object = object->backing_object; 1194 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1195 1196 /* 1197 * Initial conditions 1198 */ 1199 if (op & OBSC_TEST_ALL_SHADOWED) { 1200 /* 1201 * We do not want to have to test for the existence of 1202 * swap pages in the backing object. XXX but with the 1203 * new swapper this would be pretty easy to do. 1204 * 1205 * XXX what about anonymous MAP_SHARED memory that hasn't 1206 * been ZFOD faulted yet? If we do not test for this, the 1207 * shadow test may succeed! XXX 1208 */ 1209 if (backing_object->type != OBJT_DEFAULT) { 1210 splx(s); 1211 return (0); 1212 } 1213 } 1214 if (op & OBSC_COLLAPSE_WAIT) { 1215 vm_object_set_flag(backing_object, OBJ_DEAD); 1216 } 1217 1218 /* 1219 * Our scan 1220 */ 1221 p = TAILQ_FIRST(&backing_object->memq); 1222 while (p) { 1223 vm_page_t next = TAILQ_NEXT(p, listq); 1224 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1225 1226 if (op & OBSC_TEST_ALL_SHADOWED) { 1227 vm_page_t pp; 1228 1229 /* 1230 * Ignore pages outside the parent object's range 1231 * and outside the parent object's mapping of the 1232 * backing object. 1233 * 1234 * note that we do not busy the backing object's 1235 * page. 1236 */ 1237 if ( 1238 p->pindex < backing_offset_index || 1239 new_pindex >= object->size 1240 ) { 1241 p = next; 1242 continue; 1243 } 1244 1245 /* 1246 * See if the parent has the page or if the parent's 1247 * object pager has the page. If the parent has the 1248 * page but the page is not valid, the parent's 1249 * object pager must have the page. 1250 * 1251 * If this fails, the parent does not completely shadow 1252 * the object and we might as well give up now. 1253 */ 1254 1255 pp = vm_page_lookup(object, new_pindex); 1256 if ( 1257 (pp == NULL || pp->valid == 0) && 1258 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1259 ) { 1260 r = 0; 1261 break; 1262 } 1263 } 1264 1265 /* 1266 * Check for busy page 1267 */ 1268 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1269 vm_page_t pp; 1270 1271 if (op & OBSC_COLLAPSE_NOWAIT) { 1272 if ( 1273 (p->flags & PG_BUSY) || 1274 !p->valid || 1275 p->hold_count || 1276 p->wire_count || 1277 p->busy 1278 ) { 1279 p = next; 1280 continue; 1281 } 1282 } else if (op & OBSC_COLLAPSE_WAIT) { 1283 if (vm_page_sleep_busy(p, TRUE, "vmocol")) { 1284 /* 1285 * If we slept, anything could have 1286 * happened. Since the object is 1287 * marked dead, the backing offset 1288 * should not have changed so we 1289 * just restart our scan. 1290 */ 1291 p = TAILQ_FIRST(&backing_object->memq); 1292 continue; 1293 } 1294 } 1295 1296 /* 1297 * Busy the page 1298 */ 1299 vm_page_busy(p); 1300 1301 KASSERT( 1302 p->object == backing_object, 1303 ("vm_object_qcollapse(): object mismatch") 1304 ); 1305 1306 /* 1307 * Destroy any associated swap 1308 */ 1309 if (backing_object->type == OBJT_SWAP) { 1310 swap_pager_freespace( 1311 backing_object, 1312 p->pindex, 1313 1 1314 ); 1315 } 1316 1317 if ( 1318 p->pindex < backing_offset_index || 1319 new_pindex >= object->size 1320 ) { 1321 /* 1322 * Page is out of the parent object's range, we 1323 * can simply destroy it. 1324 */ 1325 vm_page_protect(p, VM_PROT_NONE); 1326 vm_page_free(p); 1327 p = next; 1328 continue; 1329 } 1330 1331 pp = vm_page_lookup(object, new_pindex); 1332 if ( 1333 pp != NULL || 1334 vm_pager_has_page(object, new_pindex, NULL, NULL) 1335 ) { 1336 /* 1337 * page already exists in parent OR swap exists 1338 * for this location in the parent. Destroy 1339 * the original page from the backing object. 1340 * 1341 * Leave the parent's page alone 1342 */ 1343 vm_page_protect(p, VM_PROT_NONE); 1344 vm_page_free(p); 1345 p = next; 1346 continue; 1347 } 1348 1349 /* 1350 * Page does not exist in parent, rename the 1351 * page from the backing object to the main object. 1352 * 1353 * If the page was mapped to a process, it can remain 1354 * mapped through the rename. 1355 */ 1356 if ((p->queue - p->pc) == PQ_CACHE) 1357 vm_page_deactivate(p); 1358 1359 vm_page_rename(p, object, new_pindex); 1360 /* page automatically made dirty by rename */ 1361 } 1362 p = next; 1363 } 1364 splx(s); 1365 return (r); 1366 } 1367 1368 1369 /* 1370 * this version of collapse allows the operation to occur earlier and 1371 * when paging_in_progress is true for an object... This is not a complete 1372 * operation, but should plug 99.9% of the rest of the leaks. 1373 */ 1374 static void 1375 vm_object_qcollapse(vm_object_t object) 1376 { 1377 vm_object_t backing_object = object->backing_object; 1378 1379 GIANT_REQUIRED; 1380 1381 if (backing_object->ref_count != 1) 1382 return; 1383 1384 backing_object->ref_count += 2; 1385 1386 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1387 1388 backing_object->ref_count -= 2; 1389 } 1390 1391 /* 1392 * vm_object_collapse: 1393 * 1394 * Collapse an object with the object backing it. 1395 * Pages in the backing object are moved into the 1396 * parent, and the backing object is deallocated. 1397 */ 1398 void 1399 vm_object_collapse(vm_object_t object) 1400 { 1401 GIANT_REQUIRED; 1402 1403 while (TRUE) { 1404 vm_object_t backing_object; 1405 1406 /* 1407 * Verify that the conditions are right for collapse: 1408 * 1409 * The object exists and the backing object exists. 1410 */ 1411 if (object == NULL) 1412 break; 1413 1414 if ((backing_object = object->backing_object) == NULL) 1415 break; 1416 1417 /* 1418 * we check the backing object first, because it is most likely 1419 * not collapsable. 1420 */ 1421 if (backing_object->handle != NULL || 1422 (backing_object->type != OBJT_DEFAULT && 1423 backing_object->type != OBJT_SWAP) || 1424 (backing_object->flags & OBJ_DEAD) || 1425 object->handle != NULL || 1426 (object->type != OBJT_DEFAULT && 1427 object->type != OBJT_SWAP) || 1428 (object->flags & OBJ_DEAD)) { 1429 break; 1430 } 1431 1432 if ( 1433 object->paging_in_progress != 0 || 1434 backing_object->paging_in_progress != 0 1435 ) { 1436 vm_object_qcollapse(object); 1437 break; 1438 } 1439 1440 /* 1441 * We know that we can either collapse the backing object (if 1442 * the parent is the only reference to it) or (perhaps) have 1443 * the parent bypass the object if the parent happens to shadow 1444 * all the resident pages in the entire backing object. 1445 * 1446 * This is ignoring pager-backed pages such as swap pages. 1447 * vm_object_backing_scan fails the shadowing test in this 1448 * case. 1449 */ 1450 if (backing_object->ref_count == 1) { 1451 /* 1452 * If there is exactly one reference to the backing 1453 * object, we can collapse it into the parent. 1454 */ 1455 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1456 1457 /* 1458 * Move the pager from backing_object to object. 1459 */ 1460 if (backing_object->type == OBJT_SWAP) { 1461 vm_object_pip_add(backing_object, 1); 1462 1463 /* 1464 * scrap the paging_offset junk and do a 1465 * discrete copy. This also removes major 1466 * assumptions about how the swap-pager 1467 * works from where it doesn't belong. The 1468 * new swapper is able to optimize the 1469 * destroy-source case. 1470 */ 1471 vm_object_pip_add(object, 1); 1472 swap_pager_copy( 1473 backing_object, 1474 object, 1475 OFF_TO_IDX(object->backing_object_offset), TRUE); 1476 vm_object_pip_wakeup(object); 1477 1478 vm_object_pip_wakeup(backing_object); 1479 } 1480 /* 1481 * Object now shadows whatever backing_object did. 1482 * Note that the reference to 1483 * backing_object->backing_object moves from within 1484 * backing_object to within object. 1485 */ 1486 TAILQ_REMOVE( 1487 &object->backing_object->shadow_head, 1488 object, 1489 shadow_list 1490 ); 1491 object->backing_object->shadow_count--; 1492 object->backing_object->generation++; 1493 if (backing_object->backing_object) { 1494 TAILQ_REMOVE( 1495 &backing_object->backing_object->shadow_head, 1496 backing_object, 1497 shadow_list 1498 ); 1499 backing_object->backing_object->shadow_count--; 1500 backing_object->backing_object->generation++; 1501 } 1502 object->backing_object = backing_object->backing_object; 1503 if (object->backing_object) { 1504 TAILQ_INSERT_TAIL( 1505 &object->backing_object->shadow_head, 1506 object, 1507 shadow_list 1508 ); 1509 object->backing_object->shadow_count++; 1510 object->backing_object->generation++; 1511 } 1512 1513 object->backing_object_offset += 1514 backing_object->backing_object_offset; 1515 1516 /* 1517 * Discard backing_object. 1518 * 1519 * Since the backing object has no pages, no pager left, 1520 * and no object references within it, all that is 1521 * necessary is to dispose of it. 1522 */ 1523 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1524 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1525 1526 mtx_lock(&vm_object_list_mtx); 1527 TAILQ_REMOVE( 1528 &vm_object_list, 1529 backing_object, 1530 object_list 1531 ); 1532 mtx_unlock(&vm_object_list_mtx); 1533 1534 uma_zfree(obj_zone, backing_object); 1535 1536 object_collapses++; 1537 } else { 1538 vm_object_t new_backing_object; 1539 1540 /* 1541 * If we do not entirely shadow the backing object, 1542 * there is nothing we can do so we give up. 1543 */ 1544 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1545 break; 1546 } 1547 1548 /* 1549 * Make the parent shadow the next object in the 1550 * chain. Deallocating backing_object will not remove 1551 * it, since its reference count is at least 2. 1552 */ 1553 TAILQ_REMOVE( 1554 &backing_object->shadow_head, 1555 object, 1556 shadow_list 1557 ); 1558 backing_object->shadow_count--; 1559 backing_object->generation++; 1560 1561 new_backing_object = backing_object->backing_object; 1562 if ((object->backing_object = new_backing_object) != NULL) { 1563 vm_object_reference(new_backing_object); 1564 TAILQ_INSERT_TAIL( 1565 &new_backing_object->shadow_head, 1566 object, 1567 shadow_list 1568 ); 1569 new_backing_object->shadow_count++; 1570 new_backing_object->generation++; 1571 object->backing_object_offset += 1572 backing_object->backing_object_offset; 1573 } 1574 1575 /* 1576 * Drop the reference count on backing_object. Since 1577 * its ref_count was at least 2, it will not vanish; 1578 * so we don't need to call vm_object_deallocate, but 1579 * we do anyway. 1580 */ 1581 vm_object_deallocate(backing_object); 1582 object_bypasses++; 1583 } 1584 1585 /* 1586 * Try again with this object's new backing object. 1587 */ 1588 } 1589 } 1590 1591 /* 1592 * vm_object_page_remove: [internal] 1593 * 1594 * Removes all physical pages in the specified 1595 * object range from the object's list of pages. 1596 * 1597 * The object must be locked. 1598 */ 1599 void 1600 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) 1601 { 1602 vm_page_t p, next; 1603 unsigned int size; 1604 int all; 1605 1606 GIANT_REQUIRED; 1607 1608 if (object == NULL || 1609 object->resident_page_count == 0) 1610 return; 1611 1612 all = ((end == 0) && (start == 0)); 1613 1614 /* 1615 * Since physically-backed objects do not use managed pages, we can't 1616 * remove pages from the object (we must instead remove the page 1617 * references, and then destroy the object). 1618 */ 1619 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object")); 1620 1621 vm_object_pip_add(object, 1); 1622 again: 1623 size = end - start; 1624 if (all || size > object->resident_page_count / 4) { 1625 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) { 1626 next = TAILQ_NEXT(p, listq); 1627 if (all || ((start <= p->pindex) && (p->pindex < end))) { 1628 if (p->wire_count != 0) { 1629 vm_page_protect(p, VM_PROT_NONE); 1630 if (!clean_only) 1631 p->valid = 0; 1632 continue; 1633 } 1634 1635 /* 1636 * The busy flags are only cleared at 1637 * interrupt -- minimize the spl transitions 1638 */ 1639 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1640 goto again; 1641 1642 if (clean_only && p->valid) { 1643 vm_page_test_dirty(p); 1644 if (p->valid & p->dirty) 1645 continue; 1646 } 1647 1648 vm_page_busy(p); 1649 vm_page_protect(p, VM_PROT_NONE); 1650 vm_page_free(p); 1651 } 1652 } 1653 } else { 1654 while (size > 0) { 1655 if ((p = vm_page_lookup(object, start)) != 0) { 1656 1657 if (p->wire_count != 0) { 1658 vm_page_protect(p, VM_PROT_NONE); 1659 if (!clean_only) 1660 p->valid = 0; 1661 start += 1; 1662 size -= 1; 1663 continue; 1664 } 1665 1666 /* 1667 * The busy flags are only cleared at 1668 * interrupt -- minimize the spl transitions 1669 */ 1670 if (vm_page_sleep_busy(p, TRUE, "vmopar")) 1671 goto again; 1672 1673 if (clean_only && p->valid) { 1674 vm_page_test_dirty(p); 1675 if (p->valid & p->dirty) { 1676 start += 1; 1677 size -= 1; 1678 continue; 1679 } 1680 } 1681 1682 vm_page_busy(p); 1683 vm_page_protect(p, VM_PROT_NONE); 1684 vm_page_free(p); 1685 } 1686 start += 1; 1687 size -= 1; 1688 } 1689 } 1690 vm_object_pip_wakeup(object); 1691 } 1692 1693 /* 1694 * Routine: vm_object_coalesce 1695 * Function: Coalesces two objects backing up adjoining 1696 * regions of memory into a single object. 1697 * 1698 * returns TRUE if objects were combined. 1699 * 1700 * NOTE: Only works at the moment if the second object is NULL - 1701 * if it's not, which object do we lock first? 1702 * 1703 * Parameters: 1704 * prev_object First object to coalesce 1705 * prev_offset Offset into prev_object 1706 * next_object Second object into coalesce 1707 * next_offset Offset into next_object 1708 * 1709 * prev_size Size of reference to prev_object 1710 * next_size Size of reference to next_object 1711 * 1712 * Conditions: 1713 * The object must *not* be locked. 1714 */ 1715 boolean_t 1716 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size) 1717 { 1718 vm_pindex_t next_pindex; 1719 1720 GIANT_REQUIRED; 1721 1722 if (prev_object == NULL) { 1723 return (TRUE); 1724 } 1725 1726 if (prev_object->type != OBJT_DEFAULT && 1727 prev_object->type != OBJT_SWAP) { 1728 return (FALSE); 1729 } 1730 1731 /* 1732 * Try to collapse the object first 1733 */ 1734 vm_object_collapse(prev_object); 1735 1736 /* 1737 * Can't coalesce if: . more than one reference . paged out . shadows 1738 * another object . has a copy elsewhere (any of which mean that the 1739 * pages not mapped to prev_entry may be in use anyway) 1740 */ 1741 if (prev_object->backing_object != NULL) { 1742 return (FALSE); 1743 } 1744 1745 prev_size >>= PAGE_SHIFT; 1746 next_size >>= PAGE_SHIFT; 1747 next_pindex = prev_pindex + prev_size; 1748 1749 if ((prev_object->ref_count > 1) && 1750 (prev_object->size != next_pindex)) { 1751 return (FALSE); 1752 } 1753 1754 /* 1755 * Remove any pages that may still be in the object from a previous 1756 * deallocation. 1757 */ 1758 if (next_pindex < prev_object->size) { 1759 vm_object_page_remove(prev_object, 1760 next_pindex, 1761 next_pindex + next_size, FALSE); 1762 if (prev_object->type == OBJT_SWAP) 1763 swap_pager_freespace(prev_object, 1764 next_pindex, next_size); 1765 } 1766 1767 /* 1768 * Extend the object if necessary. 1769 */ 1770 if (next_pindex + next_size > prev_object->size) 1771 prev_object->size = next_pindex + next_size; 1772 1773 return (TRUE); 1774 } 1775 1776 void 1777 vm_object_set_writeable_dirty(vm_object_t object) 1778 { 1779 struct vnode *vp; 1780 1781 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1782 if (object->type == OBJT_VNODE && 1783 (vp = (struct vnode *)object->handle) != NULL) { 1784 if ((vp->v_flag & VOBJDIRTY) == 0) { 1785 mtx_lock(&vp->v_interlock); 1786 vp->v_flag |= VOBJDIRTY; 1787 mtx_unlock(&vp->v_interlock); 1788 } 1789 } 1790 } 1791 1792 #ifdef ENABLE_VFS_IOOPT 1793 /* 1794 * Experimental support for zero-copy I/O 1795 * 1796 * Performs the copy_on_write operations necessary to allow the virtual copies 1797 * into user space to work. This has to be called for write(2) system calls 1798 * from other processes, file unlinking, and file size shrinkage. 1799 */ 1800 void 1801 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 1802 { 1803 int rv; 1804 vm_object_t robject; 1805 vm_pindex_t idx; 1806 1807 GIANT_REQUIRED; 1808 if ((object == NULL) || 1809 ((object->flags & OBJ_OPT) == 0)) 1810 return; 1811 1812 if (object->shadow_count > object->ref_count) 1813 panic("vm_freeze_copyopts: sc > rc"); 1814 1815 while ((robject = TAILQ_FIRST(&object->shadow_head)) != NULL) { 1816 vm_pindex_t bo_pindex; 1817 vm_page_t m_in, m_out; 1818 1819 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 1820 1821 vm_object_reference(robject); 1822 1823 vm_object_pip_wait(robject, "objfrz"); 1824 1825 if (robject->ref_count == 1) { 1826 vm_object_deallocate(robject); 1827 continue; 1828 } 1829 1830 vm_object_pip_add(robject, 1); 1831 1832 for (idx = 0; idx < robject->size; idx++) { 1833 1834 m_out = vm_page_grab(robject, idx, 1835 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1836 1837 if (m_out->valid == 0) { 1838 m_in = vm_page_grab(object, bo_pindex + idx, 1839 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 1840 if (m_in->valid == 0) { 1841 rv = vm_pager_get_pages(object, &m_in, 1, 0); 1842 if (rv != VM_PAGER_OK) { 1843 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 1844 continue; 1845 } 1846 vm_page_deactivate(m_in); 1847 } 1848 1849 vm_page_protect(m_in, VM_PROT_NONE); 1850 pmap_copy_page(m_in, m_out); 1851 m_out->valid = m_in->valid; 1852 vm_page_dirty(m_out); 1853 vm_page_activate(m_out); 1854 vm_page_wakeup(m_in); 1855 } 1856 vm_page_wakeup(m_out); 1857 } 1858 1859 object->shadow_count--; 1860 object->ref_count--; 1861 TAILQ_REMOVE(&object->shadow_head, robject, shadow_list); 1862 robject->backing_object = NULL; 1863 robject->backing_object_offset = 0; 1864 1865 vm_object_pip_wakeup(robject); 1866 vm_object_deallocate(robject); 1867 } 1868 1869 vm_object_clear_flag(object, OBJ_OPT); 1870 } 1871 #endif 1872 1873 #include "opt_ddb.h" 1874 #ifdef DDB 1875 #include <sys/kernel.h> 1876 1877 #include <sys/cons.h> 1878 1879 #include <ddb/ddb.h> 1880 1881 static int 1882 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1883 { 1884 vm_map_t tmpm; 1885 vm_map_entry_t tmpe; 1886 vm_object_t obj; 1887 int entcount; 1888 1889 if (map == 0) 1890 return 0; 1891 1892 if (entry == 0) { 1893 tmpe = map->header.next; 1894 entcount = map->nentries; 1895 while (entcount-- && (tmpe != &map->header)) { 1896 if (_vm_object_in_map(map, object, tmpe)) { 1897 return 1; 1898 } 1899 tmpe = tmpe->next; 1900 } 1901 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1902 tmpm = entry->object.sub_map; 1903 tmpe = tmpm->header.next; 1904 entcount = tmpm->nentries; 1905 while (entcount-- && tmpe != &tmpm->header) { 1906 if (_vm_object_in_map(tmpm, object, tmpe)) { 1907 return 1; 1908 } 1909 tmpe = tmpe->next; 1910 } 1911 } else if ((obj = entry->object.vm_object) != NULL) { 1912 for (; obj; obj = obj->backing_object) 1913 if (obj == object) { 1914 return 1; 1915 } 1916 } 1917 return 0; 1918 } 1919 1920 static int 1921 vm_object_in_map(vm_object_t object) 1922 { 1923 struct proc *p; 1924 1925 /* sx_slock(&allproc_lock); */ 1926 LIST_FOREACH(p, &allproc, p_list) { 1927 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1928 continue; 1929 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1930 /* sx_sunlock(&allproc_lock); */ 1931 return 1; 1932 } 1933 } 1934 /* sx_sunlock(&allproc_lock); */ 1935 if (_vm_object_in_map(kernel_map, object, 0)) 1936 return 1; 1937 if (_vm_object_in_map(kmem_map, object, 0)) 1938 return 1; 1939 if (_vm_object_in_map(pager_map, object, 0)) 1940 return 1; 1941 if (_vm_object_in_map(buffer_map, object, 0)) 1942 return 1; 1943 return 0; 1944 } 1945 1946 DB_SHOW_COMMAND(vmochk, vm_object_check) 1947 { 1948 vm_object_t object; 1949 1950 /* 1951 * make sure that internal objs are in a map somewhere 1952 * and none have zero ref counts. 1953 */ 1954 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1955 if (object->handle == NULL && 1956 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1957 if (object->ref_count == 0) { 1958 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1959 (long)object->size); 1960 } 1961 if (!vm_object_in_map(object)) { 1962 db_printf( 1963 "vmochk: internal obj is not in a map: " 1964 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1965 object->ref_count, (u_long)object->size, 1966 (u_long)object->size, 1967 (void *)object->backing_object); 1968 } 1969 } 1970 } 1971 } 1972 1973 /* 1974 * vm_object_print: [ debug ] 1975 */ 1976 DB_SHOW_COMMAND(object, vm_object_print_static) 1977 { 1978 /* XXX convert args. */ 1979 vm_object_t object = (vm_object_t)addr; 1980 boolean_t full = have_addr; 1981 1982 vm_page_t p; 1983 1984 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1985 #define count was_count 1986 1987 int count; 1988 1989 if (object == NULL) 1990 return; 1991 1992 db_iprintf( 1993 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n", 1994 object, (int)object->type, (u_long)object->size, 1995 object->resident_page_count, object->ref_count, object->flags); 1996 /* 1997 * XXX no %qd in kernel. Truncate object->backing_object_offset. 1998 */ 1999 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n", 2000 object->shadow_count, 2001 object->backing_object ? object->backing_object->ref_count : 0, 2002 object->backing_object, (long)object->backing_object_offset); 2003 2004 if (!full) 2005 return; 2006 2007 db_indent += 2; 2008 count = 0; 2009 TAILQ_FOREACH(p, &object->memq, listq) { 2010 if (count == 0) 2011 db_iprintf("memory:="); 2012 else if (count == 6) { 2013 db_printf("\n"); 2014 db_iprintf(" ..."); 2015 count = 0; 2016 } else 2017 db_printf(","); 2018 count++; 2019 2020 db_printf("(off=0x%lx,page=0x%lx)", 2021 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p)); 2022 } 2023 if (count != 0) 2024 db_printf("\n"); 2025 db_indent -= 2; 2026 } 2027 2028 /* XXX. */ 2029 #undef count 2030 2031 /* XXX need this non-static entry for calling from vm_map_print. */ 2032 void 2033 vm_object_print( 2034 /* db_expr_t */ long addr, 2035 boolean_t have_addr, 2036 /* db_expr_t */ long count, 2037 char *modif) 2038 { 2039 vm_object_print_static(addr, have_addr, count, modif); 2040 } 2041 2042 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2043 { 2044 vm_object_t object; 2045 int nl = 0; 2046 int c; 2047 2048 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2049 vm_pindex_t idx, fidx; 2050 vm_pindex_t osize; 2051 vm_offset_t pa = -1, padiff; 2052 int rcount; 2053 vm_page_t m; 2054 2055 db_printf("new object: %p\n", (void *)object); 2056 if (nl > 18) { 2057 c = cngetc(); 2058 if (c != ' ') 2059 return; 2060 nl = 0; 2061 } 2062 nl++; 2063 rcount = 0; 2064 fidx = 0; 2065 osize = object->size; 2066 if (osize > 128) 2067 osize = 128; 2068 for (idx = 0; idx < osize; idx++) { 2069 m = vm_page_lookup(object, idx); 2070 if (m == NULL) { 2071 if (rcount) { 2072 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2073 (long)fidx, rcount, (long)pa); 2074 if (nl > 18) { 2075 c = cngetc(); 2076 if (c != ' ') 2077 return; 2078 nl = 0; 2079 } 2080 nl++; 2081 rcount = 0; 2082 } 2083 continue; 2084 } 2085 2086 2087 if (rcount && 2088 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2089 ++rcount; 2090 continue; 2091 } 2092 if (rcount) { 2093 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2094 padiff >>= PAGE_SHIFT; 2095 padiff &= PQ_L2_MASK; 2096 if (padiff == 0) { 2097 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2098 ++rcount; 2099 continue; 2100 } 2101 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2102 (long)fidx, rcount, (long)pa); 2103 db_printf("pd(%ld)\n", (long)padiff); 2104 if (nl > 18) { 2105 c = cngetc(); 2106 if (c != ' ') 2107 return; 2108 nl = 0; 2109 } 2110 nl++; 2111 } 2112 fidx = idx; 2113 pa = VM_PAGE_TO_PHYS(m); 2114 rcount = 1; 2115 } 2116 if (rcount) { 2117 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2118 (long)fidx, rcount, (long)pa); 2119 if (nl > 18) { 2120 c = cngetc(); 2121 if (c != ' ') 2122 return; 2123 nl = 0; 2124 } 2125 nl++; 2126 } 2127 } 2128 } 2129 #endif /* DDB */ 2130