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