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 65 /* 66 * Virtual memory object module. 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/lock.h> 75 #include <sys/mman.h> 76 #include <sys/mount.h> 77 #include <sys/kernel.h> 78 #include <sys/sysctl.h> 79 #include <sys/mutex.h> 80 #include <sys/proc.h> /* for curproc, pageproc */ 81 #include <sys/socket.h> 82 #include <sys/vnode.h> 83 #include <sys/vmmeter.h> 84 #include <sys/sx.h> 85 86 #include <vm/vm.h> 87 #include <vm/vm_param.h> 88 #include <vm/pmap.h> 89 #include <vm/vm_map.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_pageout.h> 93 #include <vm/vm_pager.h> 94 #include <vm/swap_pager.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_extern.h> 97 #include <vm/uma.h> 98 99 #define EASY_SCAN_FACTOR 8 100 101 #define MSYNC_FLUSH_HARDSEQ 0x01 102 #define MSYNC_FLUSH_SOFTSEQ 0x02 103 104 /* 105 * msync / VM object flushing optimizations 106 */ 107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, 109 CTLFLAG_RW, &msync_flush_flags, 0, ""); 110 111 static void vm_object_qcollapse(vm_object_t object); 112 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 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 object->flags = 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 void 314 vm_object_pip_wait(vm_object_t object, char *waitid) 315 { 316 317 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 318 while (object->paging_in_progress) { 319 object->flags |= OBJ_PIPWNT; 320 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 321 } 322 } 323 324 /* 325 * vm_object_allocate_wait 326 * 327 * Return a new object with the given size, and give the user the 328 * option of waiting for it to complete or failing if the needed 329 * memory isn't available. 330 */ 331 vm_object_t 332 vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags) 333 { 334 vm_object_t result; 335 336 result = (vm_object_t) uma_zalloc(obj_zone, flags); 337 338 if (result != NULL) 339 _vm_object_allocate(type, size, result); 340 341 return (result); 342 } 343 344 /* 345 * vm_object_allocate: 346 * 347 * Returns a new object with the given size. 348 */ 349 vm_object_t 350 vm_object_allocate(objtype_t type, vm_pindex_t size) 351 { 352 return(vm_object_allocate_wait(type, size, M_WAITOK)); 353 } 354 355 356 /* 357 * vm_object_reference: 358 * 359 * Gets another reference to the given object. Note: OBJ_DEAD 360 * objects can be referenced during final cleaning. 361 */ 362 void 363 vm_object_reference(vm_object_t object) 364 { 365 if (object == NULL) 366 return; 367 if (object != kmem_object) 368 mtx_lock(&Giant); 369 VM_OBJECT_LOCK(object); 370 object->ref_count++; 371 VM_OBJECT_UNLOCK(object); 372 if (object->type == OBJT_VNODE) { 373 while (vget((struct vnode *) object->handle, LK_RETRY, curthread)) { 374 printf("vm_object_reference: delay in getting object\n"); 375 } 376 } 377 if (object != kmem_object) 378 mtx_unlock(&Giant); 379 } 380 381 /* 382 * Handle deallocating an object of type OBJT_VNODE. 383 */ 384 void 385 vm_object_vndeallocate(vm_object_t object) 386 { 387 struct vnode *vp = (struct vnode *) object->handle; 388 389 GIANT_REQUIRED; 390 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 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 mp_fixme("Unlocked vflag access."); 404 vp->v_vflag &= ~VV_TEXT; 405 } 406 VM_OBJECT_UNLOCK(object); 407 /* 408 * vrele may need a vop lock 409 */ 410 vrele(vp); 411 } 412 413 /* 414 * vm_object_deallocate: 415 * 416 * Release a reference to the specified object, 417 * gained either through a vm_object_allocate 418 * or a vm_object_reference call. When all references 419 * are gone, storage associated with this object 420 * may be relinquished. 421 * 422 * No object may be locked. 423 */ 424 void 425 vm_object_deallocate(vm_object_t object) 426 { 427 vm_object_t temp; 428 429 if (object != kmem_object) 430 mtx_lock(&Giant); 431 while (object != NULL) { 432 VM_OBJECT_LOCK(object); 433 if (object->type == OBJT_VNODE) { 434 vm_object_vndeallocate(object); 435 goto done; 436 } 437 438 KASSERT(object->ref_count != 0, 439 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 440 441 /* 442 * If the reference count goes to 0 we start calling 443 * vm_object_terminate() on the object chain. 444 * A ref count of 1 may be a special case depending on the 445 * shadow count being 0 or 1. 446 */ 447 object->ref_count--; 448 if (object->ref_count > 1) { 449 VM_OBJECT_UNLOCK(object); 450 goto done; 451 } else if (object->ref_count == 1) { 452 if (object->shadow_count == 0) { 453 vm_object_set_flag(object, OBJ_ONEMAPPING); 454 } else if ((object->shadow_count == 1) && 455 (object->handle == NULL) && 456 (object->type == OBJT_DEFAULT || 457 object->type == OBJT_SWAP)) { 458 vm_object_t robject; 459 460 robject = LIST_FIRST(&object->shadow_head); 461 KASSERT(robject != NULL, 462 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 463 object->ref_count, 464 object->shadow_count)); 465 if (!VM_OBJECT_TRYLOCK(robject)) { 466 /* 467 * Avoid a potential deadlock. 468 */ 469 object->ref_count++; 470 VM_OBJECT_UNLOCK(object); 471 continue; 472 } 473 if ((robject->handle == NULL) && 474 (robject->type == OBJT_DEFAULT || 475 robject->type == OBJT_SWAP)) { 476 477 robject->ref_count++; 478 retry: 479 if (robject->paging_in_progress) { 480 VM_OBJECT_UNLOCK(object); 481 vm_object_pip_wait(robject, 482 "objde1"); 483 VM_OBJECT_LOCK(object); 484 goto retry; 485 } else if (object->paging_in_progress) { 486 VM_OBJECT_UNLOCK(robject); 487 object->flags |= OBJ_PIPWNT; 488 msleep(object, 489 VM_OBJECT_MTX(object), 490 PDROP | PVM, "objde2", 0); 491 VM_OBJECT_LOCK(robject); 492 VM_OBJECT_LOCK(object); 493 goto retry; 494 } 495 VM_OBJECT_UNLOCK(object); 496 if (robject->ref_count == 1) { 497 robject->ref_count--; 498 object = robject; 499 goto doterm; 500 } 501 object = robject; 502 vm_object_collapse(object); 503 VM_OBJECT_UNLOCK(object); 504 continue; 505 } 506 VM_OBJECT_UNLOCK(robject); 507 } 508 VM_OBJECT_UNLOCK(object); 509 goto done; 510 } 511 doterm: 512 temp = object->backing_object; 513 if (temp != NULL) { 514 VM_OBJECT_LOCK(temp); 515 LIST_REMOVE(object, shadow_list); 516 temp->shadow_count--; 517 temp->generation++; 518 VM_OBJECT_UNLOCK(temp); 519 object->backing_object = NULL; 520 } 521 /* 522 * Don't double-terminate, we could be in a termination 523 * recursion due to the terminate having to sync data 524 * to disk. 525 */ 526 if ((object->flags & OBJ_DEAD) == 0) 527 vm_object_terminate(object); 528 else 529 VM_OBJECT_UNLOCK(object); 530 object = temp; 531 } 532 done: 533 if (object != kmem_object) 534 mtx_unlock(&Giant); 535 } 536 537 /* 538 * vm_object_terminate actually destroys the specified object, freeing 539 * up all previously used resources. 540 * 541 * The object must be locked. 542 * This routine may block. 543 */ 544 void 545 vm_object_terminate(vm_object_t object) 546 { 547 vm_page_t p; 548 int s; 549 550 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 551 552 /* 553 * Make sure no one uses us. 554 */ 555 vm_object_set_flag(object, OBJ_DEAD); 556 557 /* 558 * wait for the pageout daemon to be done with the object 559 */ 560 vm_object_pip_wait(object, "objtrm"); 561 562 KASSERT(!object->paging_in_progress, 563 ("vm_object_terminate: pageout in progress")); 564 565 /* 566 * Clean and free the pages, as appropriate. All references to the 567 * object are gone, so we don't need to lock it. 568 */ 569 if (object->type == OBJT_VNODE) { 570 struct vnode *vp = (struct vnode *)object->handle; 571 572 /* 573 * Clean pages and flush buffers. 574 */ 575 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 576 VM_OBJECT_UNLOCK(object); 577 578 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0); 579 580 VM_OBJECT_LOCK(object); 581 } 582 583 KASSERT(object->ref_count == 0, 584 ("vm_object_terminate: object with references, ref_count=%d", 585 object->ref_count)); 586 587 /* 588 * Now free any remaining pages. For internal objects, this also 589 * removes them from paging queues. Don't free wired pages, just 590 * remove them from the object. 591 */ 592 s = splvm(); 593 vm_page_lock_queues(); 594 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 595 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0, 596 ("vm_object_terminate: freeing busy page %p " 597 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags)); 598 if (p->wire_count == 0) { 599 vm_page_busy(p); 600 vm_page_free(p); 601 cnt.v_pfree++; 602 } else { 603 vm_page_busy(p); 604 vm_page_remove(p); 605 } 606 } 607 vm_page_unlock_queues(); 608 splx(s); 609 610 /* 611 * Let the pager know object is dead. 612 */ 613 vm_pager_deallocate(object); 614 VM_OBJECT_UNLOCK(object); 615 616 /* 617 * Remove the object from the global object list. 618 */ 619 mtx_lock(&vm_object_list_mtx); 620 TAILQ_REMOVE(&vm_object_list, object, object_list); 621 mtx_unlock(&vm_object_list_mtx); 622 623 wakeup(object); 624 625 /* 626 * Free the space for the object. 627 */ 628 uma_zfree(obj_zone, object); 629 } 630 631 /* 632 * vm_object_page_clean 633 * 634 * Clean all dirty pages in the specified range of object. Leaves page 635 * on whatever queue it is currently on. If NOSYNC is set then do not 636 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC), 637 * leaving the object dirty. 638 * 639 * When stuffing pages asynchronously, allow clustering. XXX we need a 640 * synchronous clustering mode implementation. 641 * 642 * Odd semantics: if start == end, we clean everything. 643 * 644 * The object must be locked. 645 */ 646 void 647 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 648 { 649 vm_page_t p, np; 650 vm_pindex_t tstart, tend; 651 vm_pindex_t pi; 652 struct vnode *vp; 653 int clearobjflags; 654 int pagerflags; 655 int curgeneration; 656 657 GIANT_REQUIRED; 658 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 659 if (object->type != OBJT_VNODE || 660 (object->flags & OBJ_MIGHTBEDIRTY) == 0) 661 return; 662 663 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 664 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 665 666 vp = object->handle; 667 668 vm_object_set_flag(object, OBJ_CLEANING); 669 670 tstart = start; 671 if (end == 0) { 672 tend = object->size; 673 } else { 674 tend = end; 675 } 676 677 vm_page_lock_queues(); 678 /* 679 * If the caller is smart and only msync()s a range he knows is 680 * dirty, we may be able to avoid an object scan. This results in 681 * a phenominal improvement in performance. We cannot do this 682 * as a matter of course because the object may be huge - e.g. 683 * the size might be in the gigabytes or terrabytes. 684 */ 685 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 686 vm_pindex_t tscan; 687 int scanlimit; 688 int scanreset; 689 690 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 691 if (scanreset < 16) 692 scanreset = 16; 693 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 694 695 scanlimit = scanreset; 696 tscan = tstart; 697 while (tscan < tend) { 698 curgeneration = object->generation; 699 p = vm_page_lookup(object, tscan); 700 if (p == NULL || p->valid == 0 || 701 (p->queue - p->pc) == PQ_CACHE) { 702 if (--scanlimit == 0) 703 break; 704 ++tscan; 705 continue; 706 } 707 vm_page_test_dirty(p); 708 if ((p->dirty & p->valid) == 0) { 709 if (--scanlimit == 0) 710 break; 711 ++tscan; 712 continue; 713 } 714 /* 715 * If we have been asked to skip nosync pages and 716 * this is a nosync page, we can't continue. 717 */ 718 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 719 if (--scanlimit == 0) 720 break; 721 ++tscan; 722 continue; 723 } 724 scanlimit = scanreset; 725 726 /* 727 * This returns 0 if it was unable to busy the first 728 * page (i.e. had to sleep). 729 */ 730 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 731 } 732 733 /* 734 * If everything was dirty and we flushed it successfully, 735 * and the requested range is not the entire object, we 736 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 737 * return immediately. 738 */ 739 if (tscan >= tend && (tstart || tend < object->size)) { 740 vm_page_unlock_queues(); 741 vm_object_clear_flag(object, OBJ_CLEANING); 742 return; 743 } 744 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 745 } 746 747 /* 748 * Generally set CLEANCHK interlock and make the page read-only so 749 * we can then clear the object flags. 750 * 751 * However, if this is a nosync mmap then the object is likely to 752 * stay dirty so do not mess with the page and do not clear the 753 * object flags. 754 */ 755 clearobjflags = 1; 756 TAILQ_FOREACH(p, &object->memq, listq) { 757 vm_page_flag_set(p, PG_CLEANCHK); 758 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) 759 clearobjflags = 0; 760 else 761 pmap_page_protect(p, VM_PROT_READ); 762 } 763 764 if (clearobjflags && (tstart == 0) && (tend == object->size)) { 765 struct vnode *vp; 766 767 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 768 if (object->type == OBJT_VNODE && 769 (vp = (struct vnode *)object->handle) != NULL) { 770 VI_LOCK(vp); 771 if (vp->v_iflag & VI_OBJDIRTY) 772 vp->v_iflag &= ~VI_OBJDIRTY; 773 VI_UNLOCK(vp); 774 } 775 } 776 777 rescan: 778 curgeneration = object->generation; 779 780 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 781 int n; 782 783 np = TAILQ_NEXT(p, listq); 784 785 again: 786 pi = p->pindex; 787 if (((p->flags & PG_CLEANCHK) == 0) || 788 (pi < tstart) || (pi >= tend) || 789 (p->valid == 0) || 790 ((p->queue - p->pc) == PQ_CACHE)) { 791 vm_page_flag_clear(p, PG_CLEANCHK); 792 continue; 793 } 794 795 vm_page_test_dirty(p); 796 if ((p->dirty & p->valid) == 0) { 797 vm_page_flag_clear(p, PG_CLEANCHK); 798 continue; 799 } 800 801 /* 802 * If we have been asked to skip nosync pages and this is a 803 * nosync page, skip it. Note that the object flags were 804 * not cleared in this case so we do not have to set them. 805 */ 806 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) { 807 vm_page_flag_clear(p, PG_CLEANCHK); 808 continue; 809 } 810 811 n = vm_object_page_collect_flush(object, p, 812 curgeneration, pagerflags); 813 if (n == 0) 814 goto rescan; 815 816 if (object->generation != curgeneration) 817 goto rescan; 818 819 /* 820 * Try to optimize the next page. If we can't we pick up 821 * our (random) scan where we left off. 822 */ 823 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) { 824 if ((p = vm_page_lookup(object, pi + n)) != NULL) 825 goto again; 826 } 827 } 828 vm_page_unlock_queues(); 829 #if 0 830 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 831 #endif 832 833 vm_object_clear_flag(object, OBJ_CLEANING); 834 return; 835 } 836 837 static int 838 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 839 { 840 int runlen; 841 int s; 842 int maxf; 843 int chkb; 844 int maxb; 845 int i; 846 vm_pindex_t pi; 847 vm_page_t maf[vm_pageout_page_count]; 848 vm_page_t mab[vm_pageout_page_count]; 849 vm_page_t ma[vm_pageout_page_count]; 850 851 s = splvm(); 852 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 853 pi = p->pindex; 854 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 855 vm_page_lock_queues(); 856 if (object->generation != curgeneration) { 857 splx(s); 858 return(0); 859 } 860 } 861 maxf = 0; 862 for(i = 1; i < vm_pageout_page_count; i++) { 863 vm_page_t tp; 864 865 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 866 if ((tp->flags & PG_BUSY) || 867 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 868 (tp->flags & PG_CLEANCHK) == 0) || 869 (tp->busy != 0)) 870 break; 871 if((tp->queue - tp->pc) == PQ_CACHE) { 872 vm_page_flag_clear(tp, PG_CLEANCHK); 873 break; 874 } 875 vm_page_test_dirty(tp); 876 if ((tp->dirty & tp->valid) == 0) { 877 vm_page_flag_clear(tp, PG_CLEANCHK); 878 break; 879 } 880 maf[ i - 1 ] = tp; 881 maxf++; 882 continue; 883 } 884 break; 885 } 886 887 maxb = 0; 888 chkb = vm_pageout_page_count - maxf; 889 if (chkb) { 890 for(i = 1; i < chkb;i++) { 891 vm_page_t tp; 892 893 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 894 if ((tp->flags & PG_BUSY) || 895 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 896 (tp->flags & PG_CLEANCHK) == 0) || 897 (tp->busy != 0)) 898 break; 899 if ((tp->queue - tp->pc) == PQ_CACHE) { 900 vm_page_flag_clear(tp, PG_CLEANCHK); 901 break; 902 } 903 vm_page_test_dirty(tp); 904 if ((tp->dirty & tp->valid) == 0) { 905 vm_page_flag_clear(tp, PG_CLEANCHK); 906 break; 907 } 908 mab[ i - 1 ] = tp; 909 maxb++; 910 continue; 911 } 912 break; 913 } 914 } 915 916 for(i = 0; i < maxb; i++) { 917 int index = (maxb - i) - 1; 918 ma[index] = mab[i]; 919 vm_page_flag_clear(ma[index], PG_CLEANCHK); 920 } 921 vm_page_flag_clear(p, PG_CLEANCHK); 922 ma[maxb] = p; 923 for(i = 0; i < maxf; i++) { 924 int index = (maxb + i) + 1; 925 ma[index] = maf[i]; 926 vm_page_flag_clear(ma[index], PG_CLEANCHK); 927 } 928 runlen = maxb + maxf + 1; 929 930 splx(s); 931 vm_pageout_flush(ma, runlen, pagerflags, TRUE); 932 for (i = 0; i < runlen; i++) { 933 if (ma[i]->valid & ma[i]->dirty) { 934 pmap_page_protect(ma[i], VM_PROT_READ); 935 vm_page_flag_set(ma[i], PG_CLEANCHK); 936 937 /* 938 * maxf will end up being the actual number of pages 939 * we wrote out contiguously, non-inclusive of the 940 * first page. We do not count look-behind pages. 941 */ 942 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 943 maxf = i - maxb - 1; 944 } 945 } 946 return(maxf + 1); 947 } 948 949 /* 950 * vm_object_madvise: 951 * 952 * Implements the madvise function at the object/page level. 953 * 954 * MADV_WILLNEED (any object) 955 * 956 * Activate the specified pages if they are resident. 957 * 958 * MADV_DONTNEED (any object) 959 * 960 * Deactivate the specified pages if they are resident. 961 * 962 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 963 * OBJ_ONEMAPPING only) 964 * 965 * Deactivate and clean the specified pages if they are 966 * resident. This permits the process to reuse the pages 967 * without faulting or the kernel to reclaim the pages 968 * without I/O. 969 */ 970 void 971 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 972 { 973 vm_pindex_t end, tpindex; 974 vm_object_t backing_object, tobject; 975 vm_page_t m; 976 977 if (object == NULL) 978 return; 979 980 mtx_lock(&Giant); 981 982 end = pindex + count; 983 984 /* 985 * Locate and adjust resident pages 986 */ 987 for (; pindex < end; pindex += 1) { 988 relookup: 989 tobject = object; 990 tpindex = pindex; 991 VM_OBJECT_LOCK(tobject); 992 shadowlookup: 993 /* 994 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 995 * and those pages must be OBJ_ONEMAPPING. 996 */ 997 if (advise == MADV_FREE) { 998 if ((tobject->type != OBJT_DEFAULT && 999 tobject->type != OBJT_SWAP) || 1000 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1001 goto unlock_tobject; 1002 } 1003 } 1004 1005 m = vm_page_lookup(tobject, tpindex); 1006 1007 if (m == NULL) { 1008 /* 1009 * There may be swap even if there is no backing page 1010 */ 1011 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1012 swap_pager_freespace(tobject, tpindex, 1); 1013 1014 /* 1015 * next object 1016 */ 1017 backing_object = tobject->backing_object; 1018 if (backing_object == NULL) 1019 goto unlock_tobject; 1020 VM_OBJECT_LOCK(backing_object); 1021 VM_OBJECT_UNLOCK(tobject); 1022 tobject = backing_object; 1023 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1024 goto shadowlookup; 1025 } 1026 1027 /* 1028 * If the page is busy or not in a normal active state, 1029 * we skip it. If the page is not managed there are no 1030 * page queues to mess with. Things can break if we mess 1031 * with pages in any of the below states. 1032 */ 1033 vm_page_lock_queues(); 1034 if (m->hold_count || 1035 m->wire_count || 1036 (m->flags & PG_UNMANAGED) || 1037 m->valid != VM_PAGE_BITS_ALL) { 1038 vm_page_unlock_queues(); 1039 goto unlock_tobject; 1040 } 1041 if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) { 1042 VM_OBJECT_UNLOCK(tobject); 1043 goto relookup; 1044 } 1045 if (advise == MADV_WILLNEED) { 1046 vm_page_activate(m); 1047 } else if (advise == MADV_DONTNEED) { 1048 vm_page_dontneed(m); 1049 } else if (advise == MADV_FREE) { 1050 /* 1051 * Mark the page clean. This will allow the page 1052 * to be freed up by the system. However, such pages 1053 * are often reused quickly by malloc()/free() 1054 * so we do not do anything that would cause 1055 * a page fault if we can help it. 1056 * 1057 * Specifically, we do not try to actually free 1058 * the page now nor do we try to put it in the 1059 * cache (which would cause a page fault on reuse). 1060 * 1061 * But we do make the page is freeable as we 1062 * can without actually taking the step of unmapping 1063 * it. 1064 */ 1065 pmap_clear_modify(m); 1066 m->dirty = 0; 1067 m->act_count = 0; 1068 vm_page_dontneed(m); 1069 } 1070 vm_page_unlock_queues(); 1071 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1072 swap_pager_freespace(tobject, tpindex, 1); 1073 unlock_tobject: 1074 VM_OBJECT_UNLOCK(tobject); 1075 } 1076 mtx_unlock(&Giant); 1077 } 1078 1079 /* 1080 * vm_object_shadow: 1081 * 1082 * Create a new object which is backed by the 1083 * specified existing object range. The source 1084 * object reference is deallocated. 1085 * 1086 * The new object and offset into that object 1087 * are returned in the source parameters. 1088 */ 1089 void 1090 vm_object_shadow( 1091 vm_object_t *object, /* IN/OUT */ 1092 vm_ooffset_t *offset, /* IN/OUT */ 1093 vm_size_t length) 1094 { 1095 vm_object_t source; 1096 vm_object_t result; 1097 1098 GIANT_REQUIRED; 1099 1100 source = *object; 1101 1102 /* 1103 * Don't create the new object if the old object isn't shared. 1104 */ 1105 if (source != NULL) { 1106 VM_OBJECT_LOCK(source); 1107 if (source->ref_count == 1 && 1108 source->handle == NULL && 1109 (source->type == OBJT_DEFAULT || 1110 source->type == OBJT_SWAP)) { 1111 VM_OBJECT_UNLOCK(source); 1112 return; 1113 } 1114 VM_OBJECT_UNLOCK(source); 1115 } 1116 1117 /* 1118 * Allocate a new object with the given length. 1119 */ 1120 result = vm_object_allocate(OBJT_DEFAULT, length); 1121 1122 /* 1123 * The new object shadows the source object, adding a reference to it. 1124 * Our caller changes his reference to point to the new object, 1125 * removing a reference to the source object. Net result: no change 1126 * of reference count. 1127 * 1128 * Try to optimize the result object's page color when shadowing 1129 * in order to maintain page coloring consistency in the combined 1130 * shadowed object. 1131 */ 1132 result->backing_object = source; 1133 if (source != NULL) { 1134 VM_OBJECT_LOCK(source); 1135 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1136 source->shadow_count++; 1137 source->generation++; 1138 if (length < source->size) 1139 length = source->size; 1140 if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 || 1141 source->generation > 1) 1142 length = PQ_L2_SIZE / 3 + PQ_PRIME1; 1143 result->pg_color = (source->pg_color + 1144 length * source->generation) & PQ_L2_MASK; 1145 VM_OBJECT_UNLOCK(source); 1146 next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) & 1147 PQ_L2_MASK; 1148 } 1149 1150 /* 1151 * Store the offset into the source object, and fix up the offset into 1152 * the new object. 1153 */ 1154 result->backing_object_offset = *offset; 1155 1156 /* 1157 * Return the new things 1158 */ 1159 *offset = 0; 1160 *object = result; 1161 } 1162 1163 /* 1164 * vm_object_split: 1165 * 1166 * Split the pages in a map entry into a new object. This affords 1167 * easier removal of unused pages, and keeps object inheritance from 1168 * being a negative impact on memory usage. 1169 */ 1170 void 1171 vm_object_split(vm_map_entry_t entry) 1172 { 1173 vm_page_t m; 1174 vm_object_t orig_object, new_object, source; 1175 vm_offset_t s, e; 1176 vm_pindex_t offidxstart, offidxend; 1177 vm_size_t idx, size; 1178 vm_ooffset_t offset; 1179 1180 GIANT_REQUIRED; 1181 1182 orig_object = entry->object.vm_object; 1183 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1184 return; 1185 if (orig_object->ref_count <= 1) 1186 return; 1187 1188 offset = entry->offset; 1189 s = entry->start; 1190 e = entry->end; 1191 1192 offidxstart = OFF_TO_IDX(offset); 1193 offidxend = offidxstart + OFF_TO_IDX(e - s); 1194 size = offidxend - offidxstart; 1195 1196 new_object = vm_pager_allocate(orig_object->type, 1197 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 1198 if (new_object == NULL) 1199 return; 1200 1201 source = orig_object->backing_object; 1202 if (source != NULL) { 1203 vm_object_reference(source); /* Referenced by new_object */ 1204 VM_OBJECT_LOCK(source); 1205 LIST_INSERT_HEAD(&source->shadow_head, 1206 new_object, shadow_list); 1207 source->shadow_count++; 1208 source->generation++; 1209 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1210 VM_OBJECT_UNLOCK(source); 1211 new_object->backing_object_offset = 1212 orig_object->backing_object_offset + offset; 1213 new_object->backing_object = source; 1214 } 1215 VM_OBJECT_LOCK(orig_object); 1216 for (idx = 0; idx < size; idx++) { 1217 retry: 1218 m = vm_page_lookup(orig_object, offidxstart + idx); 1219 if (m == NULL) 1220 continue; 1221 1222 /* 1223 * We must wait for pending I/O to complete before we can 1224 * rename the page. 1225 * 1226 * We do not have to VM_PROT_NONE the page as mappings should 1227 * not be changed by this operation. 1228 */ 1229 vm_page_lock_queues(); 1230 if (vm_page_sleep_if_busy(m, TRUE, "spltwt")) 1231 goto retry; 1232 1233 vm_page_busy(m); 1234 vm_page_rename(m, new_object, idx); 1235 /* page automatically made dirty by rename and cache handled */ 1236 vm_page_busy(m); 1237 vm_page_unlock_queues(); 1238 } 1239 if (orig_object->type == OBJT_SWAP) { 1240 vm_object_pip_add(orig_object, 1); 1241 VM_OBJECT_UNLOCK(orig_object); 1242 /* 1243 * copy orig_object pages into new_object 1244 * and destroy unneeded pages in 1245 * shadow object. 1246 */ 1247 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1248 VM_OBJECT_LOCK(orig_object); 1249 vm_object_pip_wakeup(orig_object); 1250 } 1251 VM_OBJECT_UNLOCK(orig_object); 1252 vm_page_lock_queues(); 1253 TAILQ_FOREACH(m, &new_object->memq, listq) 1254 vm_page_wakeup(m); 1255 vm_page_unlock_queues(); 1256 entry->object.vm_object = new_object; 1257 entry->offset = 0LL; 1258 vm_object_deallocate(orig_object); 1259 } 1260 1261 #define OBSC_TEST_ALL_SHADOWED 0x0001 1262 #define OBSC_COLLAPSE_NOWAIT 0x0002 1263 #define OBSC_COLLAPSE_WAIT 0x0004 1264 1265 static __inline int 1266 vm_object_backing_scan(vm_object_t object, int op) 1267 { 1268 int s; 1269 int r = 1; 1270 vm_page_t p; 1271 vm_object_t backing_object; 1272 vm_pindex_t backing_offset_index; 1273 1274 s = splvm(); 1275 GIANT_REQUIRED; 1276 1277 backing_object = object->backing_object; 1278 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1279 1280 /* 1281 * Initial conditions 1282 */ 1283 if (op & OBSC_TEST_ALL_SHADOWED) { 1284 /* 1285 * We do not want to have to test for the existence of 1286 * swap pages in the backing object. XXX but with the 1287 * new swapper this would be pretty easy to do. 1288 * 1289 * XXX what about anonymous MAP_SHARED memory that hasn't 1290 * been ZFOD faulted yet? If we do not test for this, the 1291 * shadow test may succeed! XXX 1292 */ 1293 if (backing_object->type != OBJT_DEFAULT) { 1294 splx(s); 1295 return (0); 1296 } 1297 } 1298 if (op & OBSC_COLLAPSE_WAIT) { 1299 vm_object_set_flag(backing_object, OBJ_DEAD); 1300 } 1301 1302 /* 1303 * Our scan 1304 */ 1305 p = TAILQ_FIRST(&backing_object->memq); 1306 while (p) { 1307 vm_page_t next = TAILQ_NEXT(p, listq); 1308 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1309 1310 if (op & OBSC_TEST_ALL_SHADOWED) { 1311 vm_page_t pp; 1312 1313 /* 1314 * Ignore pages outside the parent object's range 1315 * and outside the parent object's mapping of the 1316 * backing object. 1317 * 1318 * note that we do not busy the backing object's 1319 * page. 1320 */ 1321 if ( 1322 p->pindex < backing_offset_index || 1323 new_pindex >= object->size 1324 ) { 1325 p = next; 1326 continue; 1327 } 1328 1329 /* 1330 * See if the parent has the page or if the parent's 1331 * object pager has the page. If the parent has the 1332 * page but the page is not valid, the parent's 1333 * object pager must have the page. 1334 * 1335 * If this fails, the parent does not completely shadow 1336 * the object and we might as well give up now. 1337 */ 1338 1339 pp = vm_page_lookup(object, new_pindex); 1340 if ( 1341 (pp == NULL || pp->valid == 0) && 1342 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1343 ) { 1344 r = 0; 1345 break; 1346 } 1347 } 1348 1349 /* 1350 * Check for busy page 1351 */ 1352 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1353 vm_page_t pp; 1354 1355 vm_page_lock_queues(); 1356 if (op & OBSC_COLLAPSE_NOWAIT) { 1357 if ((p->flags & PG_BUSY) || 1358 !p->valid || 1359 p->hold_count || 1360 p->wire_count || 1361 p->busy) { 1362 vm_page_unlock_queues(); 1363 p = next; 1364 continue; 1365 } 1366 } else if (op & OBSC_COLLAPSE_WAIT) { 1367 if (vm_page_sleep_if_busy(p, TRUE, "vmocol")) { 1368 /* 1369 * If we slept, anything could have 1370 * happened. Since the object is 1371 * marked dead, the backing offset 1372 * should not have changed so we 1373 * just restart our scan. 1374 */ 1375 p = TAILQ_FIRST(&backing_object->memq); 1376 continue; 1377 } 1378 } 1379 1380 /* 1381 * Busy the page 1382 */ 1383 vm_page_busy(p); 1384 vm_page_unlock_queues(); 1385 1386 KASSERT( 1387 p->object == backing_object, 1388 ("vm_object_qcollapse(): object mismatch") 1389 ); 1390 1391 /* 1392 * Destroy any associated swap 1393 */ 1394 if (backing_object->type == OBJT_SWAP) { 1395 swap_pager_freespace( 1396 backing_object, 1397 p->pindex, 1398 1 1399 ); 1400 } 1401 1402 if ( 1403 p->pindex < backing_offset_index || 1404 new_pindex >= object->size 1405 ) { 1406 /* 1407 * Page is out of the parent object's range, we 1408 * can simply destroy it. 1409 */ 1410 vm_page_lock_queues(); 1411 pmap_remove_all(p); 1412 vm_page_free(p); 1413 vm_page_unlock_queues(); 1414 p = next; 1415 continue; 1416 } 1417 1418 pp = vm_page_lookup(object, new_pindex); 1419 if ( 1420 pp != NULL || 1421 vm_pager_has_page(object, new_pindex, NULL, NULL) 1422 ) { 1423 /* 1424 * page already exists in parent OR swap exists 1425 * for this location in the parent. Destroy 1426 * the original page from the backing object. 1427 * 1428 * Leave the parent's page alone 1429 */ 1430 vm_page_lock_queues(); 1431 pmap_remove_all(p); 1432 vm_page_free(p); 1433 vm_page_unlock_queues(); 1434 p = next; 1435 continue; 1436 } 1437 1438 /* 1439 * Page does not exist in parent, rename the 1440 * page from the backing object to the main object. 1441 * 1442 * If the page was mapped to a process, it can remain 1443 * mapped through the rename. 1444 */ 1445 vm_page_lock_queues(); 1446 vm_page_rename(p, object, new_pindex); 1447 vm_page_unlock_queues(); 1448 /* page automatically made dirty by rename */ 1449 } 1450 p = next; 1451 } 1452 splx(s); 1453 return (r); 1454 } 1455 1456 1457 /* 1458 * this version of collapse allows the operation to occur earlier and 1459 * when paging_in_progress is true for an object... This is not a complete 1460 * operation, but should plug 99.9% of the rest of the leaks. 1461 */ 1462 static void 1463 vm_object_qcollapse(vm_object_t object) 1464 { 1465 vm_object_t backing_object = object->backing_object; 1466 1467 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1468 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1469 1470 if (backing_object->ref_count != 1) 1471 return; 1472 1473 backing_object->ref_count += 2; 1474 1475 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1476 1477 backing_object->ref_count -= 2; 1478 } 1479 1480 /* 1481 * vm_object_collapse: 1482 * 1483 * Collapse an object with the object backing it. 1484 * Pages in the backing object are moved into the 1485 * parent, and the backing object is deallocated. 1486 */ 1487 void 1488 vm_object_collapse(vm_object_t object) 1489 { 1490 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1491 1492 while (TRUE) { 1493 vm_object_t backing_object; 1494 1495 /* 1496 * Verify that the conditions are right for collapse: 1497 * 1498 * The object exists and the backing object exists. 1499 */ 1500 if ((backing_object = object->backing_object) == NULL) 1501 break; 1502 1503 /* 1504 * we check the backing object first, because it is most likely 1505 * not collapsable. 1506 */ 1507 VM_OBJECT_LOCK(backing_object); 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 VM_OBJECT_UNLOCK(backing_object); 1517 break; 1518 } 1519 1520 if ( 1521 object->paging_in_progress != 0 || 1522 backing_object->paging_in_progress != 0 1523 ) { 1524 vm_object_qcollapse(object); 1525 VM_OBJECT_UNLOCK(backing_object); 1526 break; 1527 } 1528 /* 1529 * We know that we can either collapse the backing object (if 1530 * the parent is the only reference to it) or (perhaps) have 1531 * the parent bypass the object if the parent happens to shadow 1532 * all the resident pages in the entire backing object. 1533 * 1534 * This is ignoring pager-backed pages such as swap pages. 1535 * vm_object_backing_scan fails the shadowing test in this 1536 * case. 1537 */ 1538 if (backing_object->ref_count == 1) { 1539 /* XXX */ VM_OBJECT_UNLOCK(object); 1540 /* 1541 * If there is exactly one reference to the backing 1542 * object, we can collapse it into the parent. 1543 */ 1544 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1545 1546 /* 1547 * Move the pager from backing_object to object. 1548 */ 1549 if (backing_object->type == OBJT_SWAP) { 1550 vm_object_pip_add(backing_object, 1); 1551 VM_OBJECT_UNLOCK(backing_object); 1552 /* 1553 * scrap the paging_offset junk and do a 1554 * discrete copy. This also removes major 1555 * assumptions about how the swap-pager 1556 * works from where it doesn't belong. The 1557 * new swapper is able to optimize the 1558 * destroy-source case. 1559 */ 1560 VM_OBJECT_LOCK(object); 1561 vm_object_pip_add(object, 1); 1562 VM_OBJECT_UNLOCK(object); 1563 swap_pager_copy( 1564 backing_object, 1565 object, 1566 OFF_TO_IDX(object->backing_object_offset), TRUE); 1567 VM_OBJECT_LOCK(object); 1568 vm_object_pip_wakeup(object); 1569 VM_OBJECT_UNLOCK(object); 1570 1571 VM_OBJECT_LOCK(backing_object); 1572 vm_object_pip_wakeup(backing_object); 1573 } 1574 /* 1575 * Object now shadows whatever backing_object did. 1576 * Note that the reference to 1577 * backing_object->backing_object moves from within 1578 * backing_object to within object. 1579 */ 1580 LIST_REMOVE(object, shadow_list); 1581 backing_object->shadow_count--; 1582 backing_object->generation++; 1583 if (backing_object->backing_object) { 1584 VM_OBJECT_LOCK(backing_object->backing_object); 1585 LIST_REMOVE(backing_object, shadow_list); 1586 backing_object->backing_object->shadow_count--; 1587 backing_object->backing_object->generation++; 1588 VM_OBJECT_UNLOCK(backing_object->backing_object); 1589 } 1590 object->backing_object = backing_object->backing_object; 1591 if (object->backing_object) { 1592 VM_OBJECT_LOCK(object->backing_object); 1593 LIST_INSERT_HEAD( 1594 &object->backing_object->shadow_head, 1595 object, 1596 shadow_list 1597 ); 1598 object->backing_object->shadow_count++; 1599 object->backing_object->generation++; 1600 VM_OBJECT_UNLOCK(object->backing_object); 1601 } 1602 1603 object->backing_object_offset += 1604 backing_object->backing_object_offset; 1605 1606 /* 1607 * Discard backing_object. 1608 * 1609 * Since the backing object has no pages, no pager left, 1610 * and no object references within it, all that is 1611 * necessary is to dispose of it. 1612 */ 1613 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object)); 1614 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object)); 1615 VM_OBJECT_UNLOCK(backing_object); 1616 1617 mtx_lock(&vm_object_list_mtx); 1618 TAILQ_REMOVE( 1619 &vm_object_list, 1620 backing_object, 1621 object_list 1622 ); 1623 mtx_unlock(&vm_object_list_mtx); 1624 1625 uma_zfree(obj_zone, backing_object); 1626 1627 object_collapses++; 1628 } else { 1629 vm_object_t new_backing_object; 1630 1631 /* 1632 * If we do not entirely shadow the backing object, 1633 * there is nothing we can do so we give up. 1634 */ 1635 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { 1636 VM_OBJECT_UNLOCK(backing_object); 1637 break; 1638 } 1639 1640 /* 1641 * Make the parent shadow the next object in the 1642 * chain. Deallocating backing_object will not remove 1643 * it, since its reference count is at least 2. 1644 */ 1645 LIST_REMOVE(object, shadow_list); 1646 backing_object->shadow_count--; 1647 backing_object->generation++; 1648 VM_OBJECT_UNLOCK(backing_object); 1649 /* XXX */ VM_OBJECT_UNLOCK(object); 1650 1651 new_backing_object = backing_object->backing_object; 1652 if ((object->backing_object = new_backing_object) != NULL) { 1653 vm_object_reference(new_backing_object); 1654 VM_OBJECT_LOCK(new_backing_object); 1655 LIST_INSERT_HEAD( 1656 &new_backing_object->shadow_head, 1657 object, 1658 shadow_list 1659 ); 1660 new_backing_object->shadow_count++; 1661 new_backing_object->generation++; 1662 VM_OBJECT_UNLOCK(new_backing_object); 1663 object->backing_object_offset += 1664 backing_object->backing_object_offset; 1665 } 1666 1667 /* 1668 * Drop the reference count on backing_object. Since 1669 * its ref_count was at least 2, it will not vanish; 1670 * so we don't need to call vm_object_deallocate, but 1671 * we do anyway. 1672 */ 1673 vm_object_deallocate(backing_object); 1674 object_bypasses++; 1675 } 1676 1677 /* 1678 * Try again with this object's new backing object. 1679 */ 1680 /* XXX */ VM_OBJECT_LOCK(object); 1681 } 1682 } 1683 1684 /* 1685 * vm_object_page_remove: 1686 * 1687 * Removes all physical pages in the given range from the 1688 * object's list of pages. If the range's end is zero, all 1689 * physical pages from the range's start to the end of the object 1690 * are deleted. 1691 * 1692 * The object must be locked. 1693 */ 1694 void 1695 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1696 boolean_t clean_only) 1697 { 1698 vm_page_t p, next; 1699 1700 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1701 if (object->resident_page_count == 0) 1702 return; 1703 1704 /* 1705 * Since physically-backed objects do not use managed pages, we can't 1706 * remove pages from the object (we must instead remove the page 1707 * references, and then destroy the object). 1708 */ 1709 KASSERT(object->type != OBJT_PHYS, 1710 ("attempt to remove pages from a physical object")); 1711 1712 vm_object_pip_add(object, 1); 1713 again: 1714 vm_page_lock_queues(); 1715 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1716 if (p->pindex < start) { 1717 p = vm_page_splay(start, object->root); 1718 if ((object->root = p)->pindex < start) 1719 p = TAILQ_NEXT(p, listq); 1720 } 1721 } 1722 /* 1723 * Assert: the variable p is either (1) the page with the 1724 * least pindex greater than or equal to the parameter pindex 1725 * or (2) NULL. 1726 */ 1727 for (; 1728 p != NULL && (p->pindex < end || end == 0); 1729 p = next) { 1730 next = TAILQ_NEXT(p, listq); 1731 1732 if (p->wire_count != 0) { 1733 pmap_remove_all(p); 1734 if (!clean_only) 1735 p->valid = 0; 1736 continue; 1737 } 1738 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1739 goto again; 1740 if (clean_only && p->valid) { 1741 vm_page_test_dirty(p); 1742 if (p->valid & p->dirty) 1743 continue; 1744 } 1745 vm_page_busy(p); 1746 pmap_remove_all(p); 1747 vm_page_free(p); 1748 } 1749 vm_page_unlock_queues(); 1750 vm_object_pip_wakeup(object); 1751 } 1752 1753 /* 1754 * Routine: vm_object_coalesce 1755 * Function: Coalesces two objects backing up adjoining 1756 * regions of memory into a single object. 1757 * 1758 * returns TRUE if objects were combined. 1759 * 1760 * NOTE: Only works at the moment if the second object is NULL - 1761 * if it's not, which object do we lock first? 1762 * 1763 * Parameters: 1764 * prev_object First object to coalesce 1765 * prev_offset Offset into prev_object 1766 * next_object Second object into coalesce 1767 * next_offset Offset into next_object 1768 * 1769 * prev_size Size of reference to prev_object 1770 * next_size Size of reference to next_object 1771 * 1772 * Conditions: 1773 * The object must *not* be locked. 1774 */ 1775 boolean_t 1776 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, 1777 vm_size_t prev_size, vm_size_t next_size) 1778 { 1779 vm_pindex_t next_pindex; 1780 1781 if (prev_object == NULL) 1782 return (TRUE); 1783 mtx_lock(&Giant); 1784 VM_OBJECT_LOCK(prev_object); 1785 if (prev_object->type != OBJT_DEFAULT && 1786 prev_object->type != OBJT_SWAP) { 1787 VM_OBJECT_UNLOCK(prev_object); 1788 mtx_unlock(&Giant); 1789 return (FALSE); 1790 } 1791 1792 /* 1793 * Try to collapse the object first 1794 */ 1795 vm_object_collapse(prev_object); 1796 1797 /* 1798 * Can't coalesce if: . more than one reference . paged out . shadows 1799 * another object . has a copy elsewhere (any of which mean that the 1800 * pages not mapped to prev_entry may be in use anyway) 1801 */ 1802 if (prev_object->backing_object != NULL) { 1803 VM_OBJECT_UNLOCK(prev_object); 1804 mtx_unlock(&Giant); 1805 return (FALSE); 1806 } 1807 1808 prev_size >>= PAGE_SHIFT; 1809 next_size >>= PAGE_SHIFT; 1810 next_pindex = prev_pindex + prev_size; 1811 1812 if ((prev_object->ref_count > 1) && 1813 (prev_object->size != next_pindex)) { 1814 VM_OBJECT_UNLOCK(prev_object); 1815 mtx_unlock(&Giant); 1816 return (FALSE); 1817 } 1818 1819 /* 1820 * Remove any pages that may still be in the object from a previous 1821 * deallocation. 1822 */ 1823 if (next_pindex < prev_object->size) { 1824 vm_object_page_remove(prev_object, 1825 next_pindex, 1826 next_pindex + next_size, FALSE); 1827 if (prev_object->type == OBJT_SWAP) 1828 swap_pager_freespace(prev_object, 1829 next_pindex, next_size); 1830 } 1831 1832 /* 1833 * Extend the object if necessary. 1834 */ 1835 if (next_pindex + next_size > prev_object->size) 1836 prev_object->size = next_pindex + next_size; 1837 1838 VM_OBJECT_UNLOCK(prev_object); 1839 mtx_unlock(&Giant); 1840 return (TRUE); 1841 } 1842 1843 void 1844 vm_object_set_writeable_dirty(vm_object_t object) 1845 { 1846 struct vnode *vp; 1847 1848 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 1849 if (object->type == OBJT_VNODE && 1850 (vp = (struct vnode *)object->handle) != NULL) { 1851 VI_LOCK(vp); 1852 if ((vp->v_iflag & VI_OBJDIRTY) == 0) 1853 vp->v_iflag |= VI_OBJDIRTY; 1854 VI_UNLOCK(vp); 1855 } 1856 } 1857 1858 #include "opt_ddb.h" 1859 #ifdef DDB 1860 #include <sys/kernel.h> 1861 1862 #include <sys/cons.h> 1863 1864 #include <ddb/ddb.h> 1865 1866 static int 1867 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 1868 { 1869 vm_map_t tmpm; 1870 vm_map_entry_t tmpe; 1871 vm_object_t obj; 1872 int entcount; 1873 1874 if (map == 0) 1875 return 0; 1876 1877 if (entry == 0) { 1878 tmpe = map->header.next; 1879 entcount = map->nentries; 1880 while (entcount-- && (tmpe != &map->header)) { 1881 if (_vm_object_in_map(map, object, tmpe)) { 1882 return 1; 1883 } 1884 tmpe = tmpe->next; 1885 } 1886 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 1887 tmpm = entry->object.sub_map; 1888 tmpe = tmpm->header.next; 1889 entcount = tmpm->nentries; 1890 while (entcount-- && tmpe != &tmpm->header) { 1891 if (_vm_object_in_map(tmpm, object, tmpe)) { 1892 return 1; 1893 } 1894 tmpe = tmpe->next; 1895 } 1896 } else if ((obj = entry->object.vm_object) != NULL) { 1897 for (; obj; obj = obj->backing_object) 1898 if (obj == object) { 1899 return 1; 1900 } 1901 } 1902 return 0; 1903 } 1904 1905 static int 1906 vm_object_in_map(vm_object_t object) 1907 { 1908 struct proc *p; 1909 1910 /* sx_slock(&allproc_lock); */ 1911 LIST_FOREACH(p, &allproc, p_list) { 1912 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 1913 continue; 1914 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 1915 /* sx_sunlock(&allproc_lock); */ 1916 return 1; 1917 } 1918 } 1919 /* sx_sunlock(&allproc_lock); */ 1920 if (_vm_object_in_map(kernel_map, object, 0)) 1921 return 1; 1922 if (_vm_object_in_map(kmem_map, object, 0)) 1923 return 1; 1924 if (_vm_object_in_map(pager_map, object, 0)) 1925 return 1; 1926 if (_vm_object_in_map(buffer_map, object, 0)) 1927 return 1; 1928 return 0; 1929 } 1930 1931 DB_SHOW_COMMAND(vmochk, vm_object_check) 1932 { 1933 vm_object_t object; 1934 1935 /* 1936 * make sure that internal objs are in a map somewhere 1937 * and none have zero ref counts. 1938 */ 1939 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1940 if (object->handle == NULL && 1941 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 1942 if (object->ref_count == 0) { 1943 db_printf("vmochk: internal obj has zero ref count: %ld\n", 1944 (long)object->size); 1945 } 1946 if (!vm_object_in_map(object)) { 1947 db_printf( 1948 "vmochk: internal obj is not in a map: " 1949 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 1950 object->ref_count, (u_long)object->size, 1951 (u_long)object->size, 1952 (void *)object->backing_object); 1953 } 1954 } 1955 } 1956 } 1957 1958 /* 1959 * vm_object_print: [ debug ] 1960 */ 1961 DB_SHOW_COMMAND(object, vm_object_print_static) 1962 { 1963 /* XXX convert args. */ 1964 vm_object_t object = (vm_object_t)addr; 1965 boolean_t full = have_addr; 1966 1967 vm_page_t p; 1968 1969 /* XXX count is an (unused) arg. Avoid shadowing it. */ 1970 #define count was_count 1971 1972 int count; 1973 1974 if (object == NULL) 1975 return; 1976 1977 db_iprintf( 1978 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n", 1979 object, (int)object->type, (uintmax_t)object->size, 1980 object->resident_page_count, object->ref_count, object->flags); 1981 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 1982 object->shadow_count, 1983 object->backing_object ? object->backing_object->ref_count : 0, 1984 object->backing_object, (uintmax_t)object->backing_object_offset); 1985 1986 if (!full) 1987 return; 1988 1989 db_indent += 2; 1990 count = 0; 1991 TAILQ_FOREACH(p, &object->memq, listq) { 1992 if (count == 0) 1993 db_iprintf("memory:="); 1994 else if (count == 6) { 1995 db_printf("\n"); 1996 db_iprintf(" ..."); 1997 count = 0; 1998 } else 1999 db_printf(","); 2000 count++; 2001 2002 db_printf("(off=0x%jx,page=0x%jx)", 2003 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2004 } 2005 if (count != 0) 2006 db_printf("\n"); 2007 db_indent -= 2; 2008 } 2009 2010 /* XXX. */ 2011 #undef count 2012 2013 /* XXX need this non-static entry for calling from vm_map_print. */ 2014 void 2015 vm_object_print( 2016 /* db_expr_t */ long addr, 2017 boolean_t have_addr, 2018 /* db_expr_t */ long count, 2019 char *modif) 2020 { 2021 vm_object_print_static(addr, have_addr, count, modif); 2022 } 2023 2024 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2025 { 2026 vm_object_t object; 2027 int nl = 0; 2028 int c; 2029 2030 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2031 vm_pindex_t idx, fidx; 2032 vm_pindex_t osize; 2033 vm_paddr_t pa = -1, padiff; 2034 int rcount; 2035 vm_page_t m; 2036 2037 db_printf("new object: %p\n", (void *)object); 2038 if (nl > 18) { 2039 c = cngetc(); 2040 if (c != ' ') 2041 return; 2042 nl = 0; 2043 } 2044 nl++; 2045 rcount = 0; 2046 fidx = 0; 2047 osize = object->size; 2048 if (osize > 128) 2049 osize = 128; 2050 for (idx = 0; idx < osize; idx++) { 2051 m = vm_page_lookup(object, idx); 2052 if (m == NULL) { 2053 if (rcount) { 2054 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2055 (long)fidx, rcount, (long)pa); 2056 if (nl > 18) { 2057 c = cngetc(); 2058 if (c != ' ') 2059 return; 2060 nl = 0; 2061 } 2062 nl++; 2063 rcount = 0; 2064 } 2065 continue; 2066 } 2067 2068 2069 if (rcount && 2070 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2071 ++rcount; 2072 continue; 2073 } 2074 if (rcount) { 2075 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m); 2076 padiff >>= PAGE_SHIFT; 2077 padiff &= PQ_L2_MASK; 2078 if (padiff == 0) { 2079 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE; 2080 ++rcount; 2081 continue; 2082 } 2083 db_printf(" index(%ld)run(%d)pa(0x%lx)", 2084 (long)fidx, rcount, (long)pa); 2085 db_printf("pd(%ld)\n", (long)padiff); 2086 if (nl > 18) { 2087 c = cngetc(); 2088 if (c != ' ') 2089 return; 2090 nl = 0; 2091 } 2092 nl++; 2093 } 2094 fidx = idx; 2095 pa = VM_PAGE_TO_PHYS(m); 2096 rcount = 1; 2097 } 2098 if (rcount) { 2099 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2100 (long)fidx, rcount, (long)pa); 2101 if (nl > 18) { 2102 c = cngetc(); 2103 if (c != ' ') 2104 return; 2105 nl = 0; 2106 } 2107 nl++; 2108 } 2109 } 2110 } 2111 #endif /* DDB */ 2112