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 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 33 * 34 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61 /* 62 * Virtual memory object module. 63 */ 64 65 #include <sys/cdefs.h> 66 __FBSDID("$FreeBSD$"); 67 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/lock.h> 73 #include <sys/mman.h> 74 #include <sys/mount.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/mutex.h> 78 #include <sys/proc.h> /* for curproc, pageproc */ 79 #include <sys/socket.h> 80 #include <sys/resourcevar.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/vm_reserv.h> 97 #include <vm/uma.h> 98 99 static int old_msync; 100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 101 "Use old (insecure) msync behavior"); 102 103 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 104 int pagerflags, int flags, boolean_t *clearobjflags, 105 boolean_t *eio); 106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 107 boolean_t *clearobjflags); 108 static void vm_object_qcollapse(vm_object_t object); 109 static void vm_object_vndeallocate(vm_object_t object); 110 111 /* 112 * Virtual memory objects maintain the actual data 113 * associated with allocated virtual memory. A given 114 * page of memory exists within exactly one object. 115 * 116 * An object is only deallocated when all "references" 117 * are given up. Only one "reference" to a given 118 * region of an object should be writeable. 119 * 120 * Associated with each object is a list of all resident 121 * memory pages belonging to that object; this list is 122 * maintained by the "vm_page" module, and locked by the object's 123 * lock. 124 * 125 * Each object also records a "pager" routine which is 126 * used to retrieve (and store) pages to the proper backing 127 * storage. In addition, objects may be backed by other 128 * objects from which they were virtual-copied. 129 * 130 * The only items within the object structure which are 131 * modified after time of creation are: 132 * reference count locked by object's lock 133 * pager routine locked by object's lock 134 * 135 */ 136 137 struct object_q vm_object_list; 138 struct mtx vm_object_list_mtx; /* lock for object list and count */ 139 140 struct vm_object kernel_object_store; 141 struct vm_object kmem_object_store; 142 143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 144 "VM object stats"); 145 146 static long object_collapses; 147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 148 &object_collapses, 0, "VM object collapses"); 149 150 static long object_bypasses; 151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 152 &object_bypasses, 0, "VM object bypasses"); 153 154 static uma_zone_t obj_zone; 155 156 static int vm_object_zinit(void *mem, int size, int flags); 157 158 #ifdef INVARIANTS 159 static void vm_object_zdtor(void *mem, int size, void *arg); 160 161 static void 162 vm_object_zdtor(void *mem, int size, void *arg) 163 { 164 vm_object_t object; 165 166 object = (vm_object_t)mem; 167 KASSERT(TAILQ_EMPTY(&object->memq), 168 ("object %p has resident pages", 169 object)); 170 #if VM_NRESERVLEVEL > 0 171 KASSERT(LIST_EMPTY(&object->rvq), 172 ("object %p has reservations", 173 object)); 174 #endif 175 KASSERT(object->cache == NULL, 176 ("object %p has cached pages", 177 object)); 178 KASSERT(object->paging_in_progress == 0, 179 ("object %p paging_in_progress = %d", 180 object, object->paging_in_progress)); 181 KASSERT(object->resident_page_count == 0, 182 ("object %p resident_page_count = %d", 183 object, object->resident_page_count)); 184 KASSERT(object->shadow_count == 0, 185 ("object %p shadow_count = %d", 186 object, object->shadow_count)); 187 } 188 #endif 189 190 static int 191 vm_object_zinit(void *mem, int size, int flags) 192 { 193 vm_object_t object; 194 195 object = (vm_object_t)mem; 196 bzero(&object->mtx, sizeof(object->mtx)); 197 VM_OBJECT_LOCK_INIT(object, "standard object"); 198 199 /* These are true for any object that has been freed */ 200 object->paging_in_progress = 0; 201 object->resident_page_count = 0; 202 object->shadow_count = 0; 203 return (0); 204 } 205 206 void 207 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 208 { 209 210 TAILQ_INIT(&object->memq); 211 LIST_INIT(&object->shadow_head); 212 213 object->root = NULL; 214 object->type = type; 215 object->size = size; 216 object->generation = 1; 217 object->ref_count = 1; 218 object->memattr = VM_MEMATTR_DEFAULT; 219 object->flags = 0; 220 object->cred = NULL; 221 object->charge = 0; 222 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 223 object->flags = OBJ_ONEMAPPING; 224 object->pg_color = 0; 225 object->handle = NULL; 226 object->backing_object = NULL; 227 object->backing_object_offset = (vm_ooffset_t) 0; 228 #if VM_NRESERVLEVEL > 0 229 LIST_INIT(&object->rvq); 230 #endif 231 object->cache = NULL; 232 233 mtx_lock(&vm_object_list_mtx); 234 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 235 mtx_unlock(&vm_object_list_mtx); 236 } 237 238 /* 239 * vm_object_init: 240 * 241 * Initialize the VM objects module. 242 */ 243 void 244 vm_object_init(void) 245 { 246 TAILQ_INIT(&vm_object_list); 247 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 248 249 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object"); 250 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 251 kernel_object); 252 #if VM_NRESERVLEVEL > 0 253 kernel_object->flags |= OBJ_COLORED; 254 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 255 #endif 256 257 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object"); 258 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 259 kmem_object); 260 #if VM_NRESERVLEVEL > 0 261 kmem_object->flags |= OBJ_COLORED; 262 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 263 #endif 264 265 /* 266 * The lock portion of struct vm_object must be type stable due 267 * to vm_pageout_fallback_object_lock locking a vm object 268 * without holding any references to it. 269 */ 270 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 271 #ifdef INVARIANTS 272 vm_object_zdtor, 273 #else 274 NULL, 275 #endif 276 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); 277 } 278 279 void 280 vm_object_clear_flag(vm_object_t object, u_short bits) 281 { 282 283 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 284 object->flags &= ~bits; 285 } 286 287 /* 288 * Sets the default memory attribute for the specified object. Pages 289 * that are allocated to this object are by default assigned this memory 290 * attribute. 291 * 292 * Presently, this function must be called before any pages are allocated 293 * to the object. In the future, this requirement may be relaxed for 294 * "default" and "swap" objects. 295 */ 296 int 297 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 298 { 299 300 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 301 switch (object->type) { 302 case OBJT_DEFAULT: 303 case OBJT_DEVICE: 304 case OBJT_PHYS: 305 case OBJT_SG: 306 case OBJT_SWAP: 307 case OBJT_VNODE: 308 if (!TAILQ_EMPTY(&object->memq)) 309 return (KERN_FAILURE); 310 break; 311 case OBJT_DEAD: 312 return (KERN_INVALID_ARGUMENT); 313 } 314 object->memattr = memattr; 315 return (KERN_SUCCESS); 316 } 317 318 void 319 vm_object_pip_add(vm_object_t object, short i) 320 { 321 322 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 323 object->paging_in_progress += i; 324 } 325 326 void 327 vm_object_pip_subtract(vm_object_t object, short i) 328 { 329 330 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 331 object->paging_in_progress -= i; 332 } 333 334 void 335 vm_object_pip_wakeup(vm_object_t object) 336 { 337 338 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 339 object->paging_in_progress--; 340 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 341 vm_object_clear_flag(object, OBJ_PIPWNT); 342 wakeup(object); 343 } 344 } 345 346 void 347 vm_object_pip_wakeupn(vm_object_t object, short i) 348 { 349 350 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 351 if (i) 352 object->paging_in_progress -= i; 353 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 354 vm_object_clear_flag(object, OBJ_PIPWNT); 355 wakeup(object); 356 } 357 } 358 359 void 360 vm_object_pip_wait(vm_object_t object, char *waitid) 361 { 362 363 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 364 while (object->paging_in_progress) { 365 object->flags |= OBJ_PIPWNT; 366 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 367 } 368 } 369 370 /* 371 * vm_object_allocate: 372 * 373 * Returns a new object with the given size. 374 */ 375 vm_object_t 376 vm_object_allocate(objtype_t type, vm_pindex_t size) 377 { 378 vm_object_t object; 379 380 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 381 _vm_object_allocate(type, size, object); 382 return (object); 383 } 384 385 386 /* 387 * vm_object_reference: 388 * 389 * Gets another reference to the given object. Note: OBJ_DEAD 390 * objects can be referenced during final cleaning. 391 */ 392 void 393 vm_object_reference(vm_object_t object) 394 { 395 if (object == NULL) 396 return; 397 VM_OBJECT_LOCK(object); 398 vm_object_reference_locked(object); 399 VM_OBJECT_UNLOCK(object); 400 } 401 402 /* 403 * vm_object_reference_locked: 404 * 405 * Gets another reference to the given object. 406 * 407 * The object must be locked. 408 */ 409 void 410 vm_object_reference_locked(vm_object_t object) 411 { 412 struct vnode *vp; 413 414 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 415 object->ref_count++; 416 if (object->type == OBJT_VNODE) { 417 vp = object->handle; 418 vref(vp); 419 } 420 } 421 422 /* 423 * Handle deallocating an object of type OBJT_VNODE. 424 */ 425 static void 426 vm_object_vndeallocate(vm_object_t object) 427 { 428 struct vnode *vp = (struct vnode *) object->handle; 429 430 VFS_ASSERT_GIANT(vp->v_mount); 431 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 432 KASSERT(object->type == OBJT_VNODE, 433 ("vm_object_vndeallocate: not a vnode object")); 434 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 435 #ifdef INVARIANTS 436 if (object->ref_count == 0) { 437 vprint("vm_object_vndeallocate", vp); 438 panic("vm_object_vndeallocate: bad object reference count"); 439 } 440 #endif 441 442 if (object->ref_count > 1) { 443 object->ref_count--; 444 VM_OBJECT_UNLOCK(object); 445 /* vrele may need the vnode lock. */ 446 vrele(vp); 447 } else { 448 vhold(vp); 449 VM_OBJECT_UNLOCK(object); 450 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 451 vdrop(vp); 452 VM_OBJECT_LOCK(object); 453 object->ref_count--; 454 if (object->type == OBJT_DEAD) { 455 VM_OBJECT_UNLOCK(object); 456 VOP_UNLOCK(vp, 0); 457 } else { 458 if (object->ref_count == 0) 459 VOP_UNSET_TEXT(vp); 460 VM_OBJECT_UNLOCK(object); 461 vput(vp); 462 } 463 } 464 } 465 466 /* 467 * vm_object_deallocate: 468 * 469 * Release a reference to the specified object, 470 * gained either through a vm_object_allocate 471 * or a vm_object_reference call. When all references 472 * are gone, storage associated with this object 473 * may be relinquished. 474 * 475 * No object may be locked. 476 */ 477 void 478 vm_object_deallocate(vm_object_t object) 479 { 480 vm_object_t temp; 481 482 while (object != NULL) { 483 int vfslocked; 484 485 vfslocked = 0; 486 restart: 487 VM_OBJECT_LOCK(object); 488 if (object->type == OBJT_VNODE) { 489 struct vnode *vp = (struct vnode *) object->handle; 490 491 /* 492 * Conditionally acquire Giant for a vnode-backed 493 * object. We have to be careful since the type of 494 * a vnode object can change while the object is 495 * unlocked. 496 */ 497 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) { 498 vfslocked = 1; 499 if (!mtx_trylock(&Giant)) { 500 VM_OBJECT_UNLOCK(object); 501 mtx_lock(&Giant); 502 goto restart; 503 } 504 } 505 vm_object_vndeallocate(object); 506 VFS_UNLOCK_GIANT(vfslocked); 507 return; 508 } else 509 /* 510 * This is to handle the case that the object 511 * changed type while we dropped its lock to 512 * obtain Giant. 513 */ 514 VFS_UNLOCK_GIANT(vfslocked); 515 516 KASSERT(object->ref_count != 0, 517 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 518 519 /* 520 * If the reference count goes to 0 we start calling 521 * vm_object_terminate() on the object chain. 522 * A ref count of 1 may be a special case depending on the 523 * shadow count being 0 or 1. 524 */ 525 object->ref_count--; 526 if (object->ref_count > 1) { 527 VM_OBJECT_UNLOCK(object); 528 return; 529 } else if (object->ref_count == 1) { 530 if (object->shadow_count == 0 && 531 object->handle == NULL && 532 (object->type == OBJT_DEFAULT || 533 object->type == OBJT_SWAP)) { 534 vm_object_set_flag(object, OBJ_ONEMAPPING); 535 } else if ((object->shadow_count == 1) && 536 (object->handle == NULL) && 537 (object->type == OBJT_DEFAULT || 538 object->type == OBJT_SWAP)) { 539 vm_object_t robject; 540 541 robject = LIST_FIRST(&object->shadow_head); 542 KASSERT(robject != NULL, 543 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 544 object->ref_count, 545 object->shadow_count)); 546 if (!VM_OBJECT_TRYLOCK(robject)) { 547 /* 548 * Avoid a potential deadlock. 549 */ 550 object->ref_count++; 551 VM_OBJECT_UNLOCK(object); 552 /* 553 * More likely than not the thread 554 * holding robject's lock has lower 555 * priority than the current thread. 556 * Let the lower priority thread run. 557 */ 558 pause("vmo_de", 1); 559 continue; 560 } 561 /* 562 * Collapse object into its shadow unless its 563 * shadow is dead. In that case, object will 564 * be deallocated by the thread that is 565 * deallocating its shadow. 566 */ 567 if ((robject->flags & OBJ_DEAD) == 0 && 568 (robject->handle == NULL) && 569 (robject->type == OBJT_DEFAULT || 570 robject->type == OBJT_SWAP)) { 571 572 robject->ref_count++; 573 retry: 574 if (robject->paging_in_progress) { 575 VM_OBJECT_UNLOCK(object); 576 vm_object_pip_wait(robject, 577 "objde1"); 578 temp = robject->backing_object; 579 if (object == temp) { 580 VM_OBJECT_LOCK(object); 581 goto retry; 582 } 583 } else if (object->paging_in_progress) { 584 VM_OBJECT_UNLOCK(robject); 585 object->flags |= OBJ_PIPWNT; 586 msleep(object, 587 VM_OBJECT_MTX(object), 588 PDROP | PVM, "objde2", 0); 589 VM_OBJECT_LOCK(robject); 590 temp = robject->backing_object; 591 if (object == temp) { 592 VM_OBJECT_LOCK(object); 593 goto retry; 594 } 595 } else 596 VM_OBJECT_UNLOCK(object); 597 598 if (robject->ref_count == 1) { 599 robject->ref_count--; 600 object = robject; 601 goto doterm; 602 } 603 object = robject; 604 vm_object_collapse(object); 605 VM_OBJECT_UNLOCK(object); 606 continue; 607 } 608 VM_OBJECT_UNLOCK(robject); 609 } 610 VM_OBJECT_UNLOCK(object); 611 return; 612 } 613 doterm: 614 temp = object->backing_object; 615 if (temp != NULL) { 616 VM_OBJECT_LOCK(temp); 617 LIST_REMOVE(object, shadow_list); 618 temp->shadow_count--; 619 VM_OBJECT_UNLOCK(temp); 620 object->backing_object = NULL; 621 } 622 /* 623 * Don't double-terminate, we could be in a termination 624 * recursion due to the terminate having to sync data 625 * to disk. 626 */ 627 if ((object->flags & OBJ_DEAD) == 0) 628 vm_object_terminate(object); 629 else 630 VM_OBJECT_UNLOCK(object); 631 object = temp; 632 } 633 } 634 635 /* 636 * vm_object_destroy removes the object from the global object list 637 * and frees the space for the object. 638 */ 639 void 640 vm_object_destroy(vm_object_t object) 641 { 642 643 /* 644 * Remove the object from the global object list. 645 */ 646 mtx_lock(&vm_object_list_mtx); 647 TAILQ_REMOVE(&vm_object_list, object, object_list); 648 mtx_unlock(&vm_object_list_mtx); 649 650 /* 651 * Release the allocation charge. 652 */ 653 if (object->cred != NULL) { 654 KASSERT(object->type == OBJT_DEFAULT || 655 object->type == OBJT_SWAP, 656 ("vm_object_terminate: non-swap obj %p has cred", 657 object)); 658 swap_release_by_cred(object->charge, object->cred); 659 object->charge = 0; 660 crfree(object->cred); 661 object->cred = NULL; 662 } 663 664 /* 665 * Free the space for the object. 666 */ 667 uma_zfree(obj_zone, object); 668 } 669 670 /* 671 * vm_object_terminate actually destroys the specified object, freeing 672 * up all previously used resources. 673 * 674 * The object must be locked. 675 * This routine may block. 676 */ 677 void 678 vm_object_terminate(vm_object_t object) 679 { 680 vm_page_t p, p_next; 681 682 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 683 684 /* 685 * Make sure no one uses us. 686 */ 687 vm_object_set_flag(object, OBJ_DEAD); 688 689 /* 690 * wait for the pageout daemon to be done with the object 691 */ 692 vm_object_pip_wait(object, "objtrm"); 693 694 KASSERT(!object->paging_in_progress, 695 ("vm_object_terminate: pageout in progress")); 696 697 /* 698 * Clean and free the pages, as appropriate. All references to the 699 * object are gone, so we don't need to lock it. 700 */ 701 if (object->type == OBJT_VNODE) { 702 struct vnode *vp = (struct vnode *)object->handle; 703 704 /* 705 * Clean pages and flush buffers. 706 */ 707 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 708 VM_OBJECT_UNLOCK(object); 709 710 vinvalbuf(vp, V_SAVE, 0, 0); 711 712 VM_OBJECT_LOCK(object); 713 } 714 715 KASSERT(object->ref_count == 0, 716 ("vm_object_terminate: object with references, ref_count=%d", 717 object->ref_count)); 718 719 /* 720 * Free any remaining pageable pages. This also removes them from the 721 * paging queues. However, don't free wired pages, just remove them 722 * from the object. Rather than incrementally removing each page from 723 * the object, the page and object are reset to any empty state. 724 */ 725 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 726 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0, 727 ("vm_object_terminate: freeing busy page %p", p)); 728 vm_page_lock(p); 729 /* 730 * Optimize the page's removal from the object by resetting 731 * its "object" field. Specifically, if the page is not 732 * wired, then the effect of this assignment is that 733 * vm_page_free()'s call to vm_page_remove() will return 734 * immediately without modifying the page or the object. 735 */ 736 p->object = NULL; 737 if (p->wire_count == 0) { 738 vm_page_free(p); 739 PCPU_INC(cnt.v_pfree); 740 } 741 vm_page_unlock(p); 742 } 743 /* 744 * If the object contained any pages, then reset it to an empty state. 745 * None of the object's fields, including "resident_page_count", were 746 * modified by the preceding loop. 747 */ 748 if (object->resident_page_count != 0) { 749 object->root = NULL; 750 TAILQ_INIT(&object->memq); 751 object->resident_page_count = 0; 752 if (object->type == OBJT_VNODE) 753 vdrop(object->handle); 754 } 755 756 #if VM_NRESERVLEVEL > 0 757 if (__predict_false(!LIST_EMPTY(&object->rvq))) 758 vm_reserv_break_all(object); 759 #endif 760 if (__predict_false(object->cache != NULL)) 761 vm_page_cache_free(object, 0, 0); 762 763 /* 764 * Let the pager know object is dead. 765 */ 766 vm_pager_deallocate(object); 767 VM_OBJECT_UNLOCK(object); 768 769 vm_object_destroy(object); 770 } 771 772 /* 773 * Make the page read-only so that we can clear the object flags. However, if 774 * this is a nosync mmap then the object is likely to stay dirty so do not 775 * mess with the page and do not clear the object flags. Returns TRUE if the 776 * page should be flushed, and FALSE otherwise. 777 */ 778 static boolean_t 779 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 780 { 781 782 /* 783 * If we have been asked to skip nosync pages and this is a 784 * nosync page, skip it. Note that the object flags were not 785 * cleared in this case so we do not have to set them. 786 */ 787 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 788 *clearobjflags = FALSE; 789 return (FALSE); 790 } else { 791 pmap_remove_write(p); 792 return (p->dirty != 0); 793 } 794 } 795 796 /* 797 * vm_object_page_clean 798 * 799 * Clean all dirty pages in the specified range of object. Leaves page 800 * on whatever queue it is currently on. If NOSYNC is set then do not 801 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 802 * leaving the object dirty. 803 * 804 * When stuffing pages asynchronously, allow clustering. XXX we need a 805 * synchronous clustering mode implementation. 806 * 807 * Odd semantics: if start == end, we clean everything. 808 * 809 * The object must be locked. 810 * 811 * Returns FALSE if some page from the range was not written, as 812 * reported by the pager, and TRUE otherwise. 813 */ 814 boolean_t 815 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 816 int flags) 817 { 818 vm_page_t np, p; 819 vm_pindex_t pi, tend, tstart; 820 int curgeneration, n, pagerflags; 821 boolean_t clearobjflags, eio, res; 822 823 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 824 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object")); 825 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 826 object->resident_page_count == 0) 827 return (TRUE); 828 829 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 830 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 831 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 832 833 tstart = OFF_TO_IDX(start); 834 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 835 clearobjflags = tstart == 0 && tend >= object->size; 836 res = TRUE; 837 838 rescan: 839 curgeneration = object->generation; 840 841 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 842 pi = p->pindex; 843 if (pi >= tend) 844 break; 845 np = TAILQ_NEXT(p, listq); 846 if (p->valid == 0) 847 continue; 848 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 849 if (object->generation != curgeneration) { 850 if ((flags & OBJPC_SYNC) != 0) 851 goto rescan; 852 else 853 clearobjflags = FALSE; 854 } 855 np = vm_page_find_least(object, pi); 856 continue; 857 } 858 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 859 continue; 860 861 n = vm_object_page_collect_flush(object, p, pagerflags, 862 flags, &clearobjflags, &eio); 863 if (eio) { 864 res = FALSE; 865 clearobjflags = FALSE; 866 } 867 if (object->generation != curgeneration) { 868 if ((flags & OBJPC_SYNC) != 0) 869 goto rescan; 870 else 871 clearobjflags = FALSE; 872 } 873 874 /* 875 * If the VOP_PUTPAGES() did a truncated write, so 876 * that even the first page of the run is not fully 877 * written, vm_pageout_flush() returns 0 as the run 878 * length. Since the condition that caused truncated 879 * write may be permanent, e.g. exhausted free space, 880 * accepting n == 0 would cause an infinite loop. 881 * 882 * Forwarding the iterator leaves the unwritten page 883 * behind, but there is not much we can do there if 884 * filesystem refuses to write it. 885 */ 886 if (n == 0) { 887 n = 1; 888 clearobjflags = FALSE; 889 } 890 np = vm_page_find_least(object, pi + n); 891 } 892 #if 0 893 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 894 #endif 895 896 if (clearobjflags) 897 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 898 return (res); 899 } 900 901 static int 902 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 903 int flags, boolean_t *clearobjflags, boolean_t *eio) 904 { 905 vm_page_t ma[vm_pageout_page_count], p_first, tp; 906 int count, i, mreq, runlen; 907 908 vm_page_lock_assert(p, MA_NOTOWNED); 909 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 910 911 count = 1; 912 mreq = 0; 913 914 for (tp = p; count < vm_pageout_page_count; count++) { 915 tp = vm_page_next(tp); 916 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0) 917 break; 918 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 919 break; 920 } 921 922 for (p_first = p; count < vm_pageout_page_count; count++) { 923 tp = vm_page_prev(p_first); 924 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0) 925 break; 926 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 927 break; 928 p_first = tp; 929 mreq++; 930 } 931 932 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 933 ma[i] = tp; 934 935 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 936 return (runlen); 937 } 938 939 /* 940 * Note that there is absolutely no sense in writing out 941 * anonymous objects, so we track down the vnode object 942 * to write out. 943 * We invalidate (remove) all pages from the address space 944 * for semantic correctness. 945 * 946 * If the backing object is a device object with unmanaged pages, then any 947 * mappings to the specified range of pages must be removed before this 948 * function is called. 949 * 950 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 951 * may start out with a NULL object. 952 */ 953 boolean_t 954 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 955 boolean_t syncio, boolean_t invalidate) 956 { 957 vm_object_t backing_object; 958 struct vnode *vp; 959 struct mount *mp; 960 int error, flags, fsync_after; 961 boolean_t res; 962 963 if (object == NULL) 964 return (TRUE); 965 res = TRUE; 966 error = 0; 967 VM_OBJECT_LOCK(object); 968 while ((backing_object = object->backing_object) != NULL) { 969 VM_OBJECT_LOCK(backing_object); 970 offset += object->backing_object_offset; 971 VM_OBJECT_UNLOCK(object); 972 object = backing_object; 973 if (object->size < OFF_TO_IDX(offset + size)) 974 size = IDX_TO_OFF(object->size) - offset; 975 } 976 /* 977 * Flush pages if writing is allowed, invalidate them 978 * if invalidation requested. Pages undergoing I/O 979 * will be ignored by vm_object_page_remove(). 980 * 981 * We cannot lock the vnode and then wait for paging 982 * to complete without deadlocking against vm_fault. 983 * Instead we simply call vm_object_page_remove() and 984 * allow it to block internally on a page-by-page 985 * basis when it encounters pages undergoing async 986 * I/O. 987 */ 988 if (object->type == OBJT_VNODE && 989 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 990 int vfslocked; 991 vp = object->handle; 992 VM_OBJECT_UNLOCK(object); 993 (void) vn_start_write(vp, &mp, V_WAIT); 994 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 995 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 996 if (syncio && !invalidate && offset == 0 && 997 OFF_TO_IDX(size) == object->size) { 998 /* 999 * If syncing the whole mapping of the file, 1000 * it is faster to schedule all the writes in 1001 * async mode, also allowing the clustering, 1002 * and then wait for i/o to complete. 1003 */ 1004 flags = 0; 1005 fsync_after = TRUE; 1006 } else { 1007 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1008 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1009 fsync_after = FALSE; 1010 } 1011 VM_OBJECT_LOCK(object); 1012 res = vm_object_page_clean(object, offset, offset + size, 1013 flags); 1014 VM_OBJECT_UNLOCK(object); 1015 if (fsync_after) 1016 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1017 VOP_UNLOCK(vp, 0); 1018 VFS_UNLOCK_GIANT(vfslocked); 1019 vn_finished_write(mp); 1020 if (error != 0) 1021 res = FALSE; 1022 VM_OBJECT_LOCK(object); 1023 } 1024 if ((object->type == OBJT_VNODE || 1025 object->type == OBJT_DEVICE) && invalidate) { 1026 if (object->type == OBJT_DEVICE) 1027 /* 1028 * The option OBJPR_NOTMAPPED must be passed here 1029 * because vm_object_page_remove() cannot remove 1030 * unmanaged mappings. 1031 */ 1032 flags = OBJPR_NOTMAPPED; 1033 else if (old_msync) 1034 flags = 0; 1035 else 1036 flags = OBJPR_CLEANONLY; 1037 vm_object_page_remove(object, OFF_TO_IDX(offset), 1038 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1039 } 1040 VM_OBJECT_UNLOCK(object); 1041 return (res); 1042 } 1043 1044 /* 1045 * vm_object_madvise: 1046 * 1047 * Implements the madvise function at the object/page level. 1048 * 1049 * MADV_WILLNEED (any object) 1050 * 1051 * Activate the specified pages if they are resident. 1052 * 1053 * MADV_DONTNEED (any object) 1054 * 1055 * Deactivate the specified pages if they are resident. 1056 * 1057 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1058 * OBJ_ONEMAPPING only) 1059 * 1060 * Deactivate and clean the specified pages if they are 1061 * resident. This permits the process to reuse the pages 1062 * without faulting or the kernel to reclaim the pages 1063 * without I/O. 1064 */ 1065 void 1066 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1067 int advise) 1068 { 1069 vm_pindex_t tpindex; 1070 vm_object_t backing_object, tobject; 1071 vm_page_t m; 1072 1073 if (object == NULL) 1074 return; 1075 VM_OBJECT_LOCK(object); 1076 /* 1077 * Locate and adjust resident pages 1078 */ 1079 for (; pindex < end; pindex += 1) { 1080 relookup: 1081 tobject = object; 1082 tpindex = pindex; 1083 shadowlookup: 1084 /* 1085 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1086 * and those pages must be OBJ_ONEMAPPING. 1087 */ 1088 if (advise == MADV_FREE) { 1089 if ((tobject->type != OBJT_DEFAULT && 1090 tobject->type != OBJT_SWAP) || 1091 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1092 goto unlock_tobject; 1093 } 1094 } else if (tobject->type == OBJT_PHYS) 1095 goto unlock_tobject; 1096 m = vm_page_lookup(tobject, tpindex); 1097 if (m == NULL && advise == MADV_WILLNEED) { 1098 /* 1099 * If the page is cached, reactivate it. 1100 */ 1101 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1102 VM_ALLOC_NOBUSY); 1103 } 1104 if (m == NULL) { 1105 /* 1106 * There may be swap even if there is no backing page 1107 */ 1108 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1109 swap_pager_freespace(tobject, tpindex, 1); 1110 /* 1111 * next object 1112 */ 1113 backing_object = tobject->backing_object; 1114 if (backing_object == NULL) 1115 goto unlock_tobject; 1116 VM_OBJECT_LOCK(backing_object); 1117 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1118 if (tobject != object) 1119 VM_OBJECT_UNLOCK(tobject); 1120 tobject = backing_object; 1121 goto shadowlookup; 1122 } else if (m->valid != VM_PAGE_BITS_ALL) 1123 goto unlock_tobject; 1124 /* 1125 * If the page is not in a normal state, skip it. 1126 */ 1127 vm_page_lock(m); 1128 if (m->hold_count != 0 || m->wire_count != 0) { 1129 vm_page_unlock(m); 1130 goto unlock_tobject; 1131 } 1132 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1133 ("vm_object_madvise: page %p is fictitious", m)); 1134 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1135 ("vm_object_madvise: page %p is not managed", m)); 1136 if ((m->oflags & VPO_BUSY) || m->busy) { 1137 if (advise == MADV_WILLNEED) { 1138 /* 1139 * Reference the page before unlocking and 1140 * sleeping so that the page daemon is less 1141 * likely to reclaim it. 1142 */ 1143 vm_page_aflag_set(m, PGA_REFERENCED); 1144 } 1145 vm_page_unlock(m); 1146 if (object != tobject) 1147 VM_OBJECT_UNLOCK(object); 1148 m->oflags |= VPO_WANTED; 1149 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 1150 0); 1151 VM_OBJECT_LOCK(object); 1152 goto relookup; 1153 } 1154 if (advise == MADV_WILLNEED) { 1155 vm_page_activate(m); 1156 } else if (advise == MADV_DONTNEED) { 1157 vm_page_dontneed(m); 1158 } else if (advise == MADV_FREE) { 1159 /* 1160 * Mark the page clean. This will allow the page 1161 * to be freed up by the system. However, such pages 1162 * are often reused quickly by malloc()/free() 1163 * so we do not do anything that would cause 1164 * a page fault if we can help it. 1165 * 1166 * Specifically, we do not try to actually free 1167 * the page now nor do we try to put it in the 1168 * cache (which would cause a page fault on reuse). 1169 * 1170 * But we do make the page is freeable as we 1171 * can without actually taking the step of unmapping 1172 * it. 1173 */ 1174 pmap_clear_modify(m); 1175 m->dirty = 0; 1176 m->act_count = 0; 1177 vm_page_dontneed(m); 1178 } 1179 vm_page_unlock(m); 1180 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1181 swap_pager_freespace(tobject, tpindex, 1); 1182 unlock_tobject: 1183 if (tobject != object) 1184 VM_OBJECT_UNLOCK(tobject); 1185 } 1186 VM_OBJECT_UNLOCK(object); 1187 } 1188 1189 /* 1190 * vm_object_shadow: 1191 * 1192 * Create a new object which is backed by the 1193 * specified existing object range. The source 1194 * object reference is deallocated. 1195 * 1196 * The new object and offset into that object 1197 * are returned in the source parameters. 1198 */ 1199 void 1200 vm_object_shadow( 1201 vm_object_t *object, /* IN/OUT */ 1202 vm_ooffset_t *offset, /* IN/OUT */ 1203 vm_size_t length) 1204 { 1205 vm_object_t source; 1206 vm_object_t result; 1207 1208 source = *object; 1209 1210 /* 1211 * Don't create the new object if the old object isn't shared. 1212 */ 1213 if (source != NULL) { 1214 VM_OBJECT_LOCK(source); 1215 if (source->ref_count == 1 && 1216 source->handle == NULL && 1217 (source->type == OBJT_DEFAULT || 1218 source->type == OBJT_SWAP)) { 1219 VM_OBJECT_UNLOCK(source); 1220 return; 1221 } 1222 VM_OBJECT_UNLOCK(source); 1223 } 1224 1225 /* 1226 * Allocate a new object with the given length. 1227 */ 1228 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1229 1230 /* 1231 * The new object shadows the source object, adding a reference to it. 1232 * Our caller changes his reference to point to the new object, 1233 * removing a reference to the source object. Net result: no change 1234 * of reference count. 1235 * 1236 * Try to optimize the result object's page color when shadowing 1237 * in order to maintain page coloring consistency in the combined 1238 * shadowed object. 1239 */ 1240 result->backing_object = source; 1241 /* 1242 * Store the offset into the source object, and fix up the offset into 1243 * the new object. 1244 */ 1245 result->backing_object_offset = *offset; 1246 if (source != NULL) { 1247 VM_OBJECT_LOCK(source); 1248 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1249 source->shadow_count++; 1250 #if VM_NRESERVLEVEL > 0 1251 result->flags |= source->flags & OBJ_COLORED; 1252 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1253 ((1 << (VM_NFREEORDER - 1)) - 1); 1254 #endif 1255 VM_OBJECT_UNLOCK(source); 1256 } 1257 1258 1259 /* 1260 * Return the new things 1261 */ 1262 *offset = 0; 1263 *object = result; 1264 } 1265 1266 /* 1267 * vm_object_split: 1268 * 1269 * Split the pages in a map entry into a new object. This affords 1270 * easier removal of unused pages, and keeps object inheritance from 1271 * being a negative impact on memory usage. 1272 */ 1273 void 1274 vm_object_split(vm_map_entry_t entry) 1275 { 1276 vm_page_t m, m_next; 1277 vm_object_t orig_object, new_object, source; 1278 vm_pindex_t idx, offidxstart; 1279 vm_size_t size; 1280 1281 orig_object = entry->object.vm_object; 1282 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1283 return; 1284 if (orig_object->ref_count <= 1) 1285 return; 1286 VM_OBJECT_UNLOCK(orig_object); 1287 1288 offidxstart = OFF_TO_IDX(entry->offset); 1289 size = atop(entry->end - entry->start); 1290 1291 /* 1292 * If swap_pager_copy() is later called, it will convert new_object 1293 * into a swap object. 1294 */ 1295 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1296 1297 /* 1298 * At this point, the new object is still private, so the order in 1299 * which the original and new objects are locked does not matter. 1300 */ 1301 VM_OBJECT_LOCK(new_object); 1302 VM_OBJECT_LOCK(orig_object); 1303 source = orig_object->backing_object; 1304 if (source != NULL) { 1305 VM_OBJECT_LOCK(source); 1306 if ((source->flags & OBJ_DEAD) != 0) { 1307 VM_OBJECT_UNLOCK(source); 1308 VM_OBJECT_UNLOCK(orig_object); 1309 VM_OBJECT_UNLOCK(new_object); 1310 vm_object_deallocate(new_object); 1311 VM_OBJECT_LOCK(orig_object); 1312 return; 1313 } 1314 LIST_INSERT_HEAD(&source->shadow_head, 1315 new_object, shadow_list); 1316 source->shadow_count++; 1317 vm_object_reference_locked(source); /* for new_object */ 1318 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1319 VM_OBJECT_UNLOCK(source); 1320 new_object->backing_object_offset = 1321 orig_object->backing_object_offset + entry->offset; 1322 new_object->backing_object = source; 1323 } 1324 if (orig_object->cred != NULL) { 1325 new_object->cred = orig_object->cred; 1326 crhold(orig_object->cred); 1327 new_object->charge = ptoa(size); 1328 KASSERT(orig_object->charge >= ptoa(size), 1329 ("orig_object->charge < 0")); 1330 orig_object->charge -= ptoa(size); 1331 } 1332 retry: 1333 m = vm_page_find_least(orig_object, offidxstart); 1334 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1335 m = m_next) { 1336 m_next = TAILQ_NEXT(m, listq); 1337 1338 /* 1339 * We must wait for pending I/O to complete before we can 1340 * rename the page. 1341 * 1342 * We do not have to VM_PROT_NONE the page as mappings should 1343 * not be changed by this operation. 1344 */ 1345 if ((m->oflags & VPO_BUSY) || m->busy) { 1346 VM_OBJECT_UNLOCK(new_object); 1347 m->oflags |= VPO_WANTED; 1348 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0); 1349 VM_OBJECT_LOCK(new_object); 1350 goto retry; 1351 } 1352 #if VM_NRESERVLEVEL > 0 1353 /* 1354 * If some of the reservation's allocated pages remain with 1355 * the original object, then transferring the reservation to 1356 * the new object is neither particularly beneficial nor 1357 * particularly harmful as compared to leaving the reservation 1358 * with the original object. If, however, all of the 1359 * reservation's allocated pages are transferred to the new 1360 * object, then transferring the reservation is typically 1361 * beneficial. Determining which of these two cases applies 1362 * would be more costly than unconditionally renaming the 1363 * reservation. 1364 */ 1365 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1366 #endif 1367 vm_page_lock(m); 1368 vm_page_rename(m, new_object, idx); 1369 vm_page_unlock(m); 1370 /* page automatically made dirty by rename and cache handled */ 1371 vm_page_busy(m); 1372 } 1373 if (orig_object->type == OBJT_SWAP) { 1374 /* 1375 * swap_pager_copy() can sleep, in which case the orig_object's 1376 * and new_object's locks are released and reacquired. 1377 */ 1378 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1379 1380 /* 1381 * Transfer any cached pages from orig_object to new_object. 1382 * If swap_pager_copy() found swapped out pages within the 1383 * specified range of orig_object, then it changed 1384 * new_object's type to OBJT_SWAP when it transferred those 1385 * pages to new_object. Otherwise, new_object's type 1386 * should still be OBJT_DEFAULT and orig_object should not 1387 * contain any cached pages within the specified range. 1388 */ 1389 if (__predict_false(orig_object->cache != NULL)) 1390 vm_page_cache_transfer(orig_object, offidxstart, 1391 new_object); 1392 } 1393 VM_OBJECT_UNLOCK(orig_object); 1394 TAILQ_FOREACH(m, &new_object->memq, listq) 1395 vm_page_wakeup(m); 1396 VM_OBJECT_UNLOCK(new_object); 1397 entry->object.vm_object = new_object; 1398 entry->offset = 0LL; 1399 vm_object_deallocate(orig_object); 1400 VM_OBJECT_LOCK(new_object); 1401 } 1402 1403 #define OBSC_TEST_ALL_SHADOWED 0x0001 1404 #define OBSC_COLLAPSE_NOWAIT 0x0002 1405 #define OBSC_COLLAPSE_WAIT 0x0004 1406 1407 static int 1408 vm_object_backing_scan(vm_object_t object, int op) 1409 { 1410 int r = 1; 1411 vm_page_t p; 1412 vm_object_t backing_object; 1413 vm_pindex_t backing_offset_index; 1414 1415 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1416 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); 1417 1418 backing_object = object->backing_object; 1419 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1420 1421 /* 1422 * Initial conditions 1423 */ 1424 if (op & OBSC_TEST_ALL_SHADOWED) { 1425 /* 1426 * We do not want to have to test for the existence of cache 1427 * or swap pages in the backing object. XXX but with the 1428 * new swapper this would be pretty easy to do. 1429 * 1430 * XXX what about anonymous MAP_SHARED memory that hasn't 1431 * been ZFOD faulted yet? If we do not test for this, the 1432 * shadow test may succeed! XXX 1433 */ 1434 if (backing_object->type != OBJT_DEFAULT) { 1435 return (0); 1436 } 1437 } 1438 if (op & OBSC_COLLAPSE_WAIT) { 1439 vm_object_set_flag(backing_object, OBJ_DEAD); 1440 } 1441 1442 /* 1443 * Our scan 1444 */ 1445 p = TAILQ_FIRST(&backing_object->memq); 1446 while (p) { 1447 vm_page_t next = TAILQ_NEXT(p, listq); 1448 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1449 1450 if (op & OBSC_TEST_ALL_SHADOWED) { 1451 vm_page_t pp; 1452 1453 /* 1454 * Ignore pages outside the parent object's range 1455 * and outside the parent object's mapping of the 1456 * backing object. 1457 * 1458 * note that we do not busy the backing object's 1459 * page. 1460 */ 1461 if ( 1462 p->pindex < backing_offset_index || 1463 new_pindex >= object->size 1464 ) { 1465 p = next; 1466 continue; 1467 } 1468 1469 /* 1470 * See if the parent has the page or if the parent's 1471 * object pager has the page. If the parent has the 1472 * page but the page is not valid, the parent's 1473 * object pager must have the page. 1474 * 1475 * If this fails, the parent does not completely shadow 1476 * the object and we might as well give up now. 1477 */ 1478 1479 pp = vm_page_lookup(object, new_pindex); 1480 if ( 1481 (pp == NULL || pp->valid == 0) && 1482 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1483 ) { 1484 r = 0; 1485 break; 1486 } 1487 } 1488 1489 /* 1490 * Check for busy page 1491 */ 1492 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1493 vm_page_t pp; 1494 1495 if (op & OBSC_COLLAPSE_NOWAIT) { 1496 if ((p->oflags & VPO_BUSY) || 1497 !p->valid || 1498 p->busy) { 1499 p = next; 1500 continue; 1501 } 1502 } else if (op & OBSC_COLLAPSE_WAIT) { 1503 if ((p->oflags & VPO_BUSY) || p->busy) { 1504 VM_OBJECT_UNLOCK(object); 1505 p->oflags |= VPO_WANTED; 1506 msleep(p, VM_OBJECT_MTX(backing_object), 1507 PDROP | PVM, "vmocol", 0); 1508 VM_OBJECT_LOCK(object); 1509 VM_OBJECT_LOCK(backing_object); 1510 /* 1511 * If we slept, anything could have 1512 * happened. Since the object is 1513 * marked dead, the backing offset 1514 * should not have changed so we 1515 * just restart our scan. 1516 */ 1517 p = TAILQ_FIRST(&backing_object->memq); 1518 continue; 1519 } 1520 } 1521 1522 KASSERT( 1523 p->object == backing_object, 1524 ("vm_object_backing_scan: object mismatch") 1525 ); 1526 1527 /* 1528 * Destroy any associated swap 1529 */ 1530 if (backing_object->type == OBJT_SWAP) { 1531 swap_pager_freespace( 1532 backing_object, 1533 p->pindex, 1534 1 1535 ); 1536 } 1537 1538 if ( 1539 p->pindex < backing_offset_index || 1540 new_pindex >= object->size 1541 ) { 1542 /* 1543 * Page is out of the parent object's range, we 1544 * can simply destroy it. 1545 */ 1546 vm_page_lock(p); 1547 KASSERT(!pmap_page_is_mapped(p), 1548 ("freeing mapped page %p", p)); 1549 if (p->wire_count == 0) 1550 vm_page_free(p); 1551 else 1552 vm_page_remove(p); 1553 vm_page_unlock(p); 1554 p = next; 1555 continue; 1556 } 1557 1558 pp = vm_page_lookup(object, new_pindex); 1559 if ( 1560 (op & OBSC_COLLAPSE_NOWAIT) != 0 && 1561 (pp != NULL && pp->valid == 0) 1562 ) { 1563 /* 1564 * The page in the parent is not (yet) valid. 1565 * We don't know anything about the state of 1566 * the original page. It might be mapped, 1567 * so we must avoid the next if here. 1568 * 1569 * This is due to a race in vm_fault() where 1570 * we must unbusy the original (backing_obj) 1571 * page before we can (re)lock the parent. 1572 * Hence we can get here. 1573 */ 1574 p = next; 1575 continue; 1576 } 1577 if ( 1578 pp != NULL || 1579 vm_pager_has_page(object, new_pindex, NULL, NULL) 1580 ) { 1581 /* 1582 * page already exists in parent OR swap exists 1583 * for this location in the parent. Destroy 1584 * the original page from the backing object. 1585 * 1586 * Leave the parent's page alone 1587 */ 1588 vm_page_lock(p); 1589 KASSERT(!pmap_page_is_mapped(p), 1590 ("freeing mapped page %p", p)); 1591 if (p->wire_count == 0) 1592 vm_page_free(p); 1593 else 1594 vm_page_remove(p); 1595 vm_page_unlock(p); 1596 p = next; 1597 continue; 1598 } 1599 1600 #if VM_NRESERVLEVEL > 0 1601 /* 1602 * Rename the reservation. 1603 */ 1604 vm_reserv_rename(p, object, backing_object, 1605 backing_offset_index); 1606 #endif 1607 1608 /* 1609 * Page does not exist in parent, rename the 1610 * page from the backing object to the main object. 1611 * 1612 * If the page was mapped to a process, it can remain 1613 * mapped through the rename. 1614 */ 1615 vm_page_lock(p); 1616 vm_page_rename(p, object, new_pindex); 1617 vm_page_unlock(p); 1618 /* page automatically made dirty by rename */ 1619 } 1620 p = next; 1621 } 1622 return (r); 1623 } 1624 1625 1626 /* 1627 * this version of collapse allows the operation to occur earlier and 1628 * when paging_in_progress is true for an object... This is not a complete 1629 * operation, but should plug 99.9% of the rest of the leaks. 1630 */ 1631 static void 1632 vm_object_qcollapse(vm_object_t object) 1633 { 1634 vm_object_t backing_object = object->backing_object; 1635 1636 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1637 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1638 1639 if (backing_object->ref_count != 1) 1640 return; 1641 1642 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1643 } 1644 1645 /* 1646 * vm_object_collapse: 1647 * 1648 * Collapse an object with the object backing it. 1649 * Pages in the backing object are moved into the 1650 * parent, and the backing object is deallocated. 1651 */ 1652 void 1653 vm_object_collapse(vm_object_t object) 1654 { 1655 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1656 1657 while (TRUE) { 1658 vm_object_t backing_object; 1659 1660 /* 1661 * Verify that the conditions are right for collapse: 1662 * 1663 * The object exists and the backing object exists. 1664 */ 1665 if ((backing_object = object->backing_object) == NULL) 1666 break; 1667 1668 /* 1669 * we check the backing object first, because it is most likely 1670 * not collapsable. 1671 */ 1672 VM_OBJECT_LOCK(backing_object); 1673 if (backing_object->handle != NULL || 1674 (backing_object->type != OBJT_DEFAULT && 1675 backing_object->type != OBJT_SWAP) || 1676 (backing_object->flags & OBJ_DEAD) || 1677 object->handle != NULL || 1678 (object->type != OBJT_DEFAULT && 1679 object->type != OBJT_SWAP) || 1680 (object->flags & OBJ_DEAD)) { 1681 VM_OBJECT_UNLOCK(backing_object); 1682 break; 1683 } 1684 1685 if ( 1686 object->paging_in_progress != 0 || 1687 backing_object->paging_in_progress != 0 1688 ) { 1689 vm_object_qcollapse(object); 1690 VM_OBJECT_UNLOCK(backing_object); 1691 break; 1692 } 1693 /* 1694 * We know that we can either collapse the backing object (if 1695 * the parent is the only reference to it) or (perhaps) have 1696 * the parent bypass the object if the parent happens to shadow 1697 * all the resident pages in the entire backing object. 1698 * 1699 * This is ignoring pager-backed pages such as swap pages. 1700 * vm_object_backing_scan fails the shadowing test in this 1701 * case. 1702 */ 1703 if (backing_object->ref_count == 1) { 1704 /* 1705 * If there is exactly one reference to the backing 1706 * object, we can collapse it into the parent. 1707 */ 1708 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1709 1710 #if VM_NRESERVLEVEL > 0 1711 /* 1712 * Break any reservations from backing_object. 1713 */ 1714 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1715 vm_reserv_break_all(backing_object); 1716 #endif 1717 1718 /* 1719 * Move the pager from backing_object to object. 1720 */ 1721 if (backing_object->type == OBJT_SWAP) { 1722 /* 1723 * swap_pager_copy() can sleep, in which case 1724 * the backing_object's and object's locks are 1725 * released and reacquired. 1726 * Since swap_pager_copy() is being asked to 1727 * destroy the source, it will change the 1728 * backing_object's type to OBJT_DEFAULT. 1729 */ 1730 swap_pager_copy( 1731 backing_object, 1732 object, 1733 OFF_TO_IDX(object->backing_object_offset), TRUE); 1734 1735 /* 1736 * Free any cached pages from backing_object. 1737 */ 1738 if (__predict_false(backing_object->cache != NULL)) 1739 vm_page_cache_free(backing_object, 0, 0); 1740 } 1741 /* 1742 * Object now shadows whatever backing_object did. 1743 * Note that the reference to 1744 * backing_object->backing_object moves from within 1745 * backing_object to within object. 1746 */ 1747 LIST_REMOVE(object, shadow_list); 1748 backing_object->shadow_count--; 1749 if (backing_object->backing_object) { 1750 VM_OBJECT_LOCK(backing_object->backing_object); 1751 LIST_REMOVE(backing_object, shadow_list); 1752 LIST_INSERT_HEAD( 1753 &backing_object->backing_object->shadow_head, 1754 object, shadow_list); 1755 /* 1756 * The shadow_count has not changed. 1757 */ 1758 VM_OBJECT_UNLOCK(backing_object->backing_object); 1759 } 1760 object->backing_object = backing_object->backing_object; 1761 object->backing_object_offset += 1762 backing_object->backing_object_offset; 1763 1764 /* 1765 * Discard backing_object. 1766 * 1767 * Since the backing object has no pages, no pager left, 1768 * and no object references within it, all that is 1769 * necessary is to dispose of it. 1770 */ 1771 KASSERT(backing_object->ref_count == 1, ( 1772 "backing_object %p was somehow re-referenced during collapse!", 1773 backing_object)); 1774 VM_OBJECT_UNLOCK(backing_object); 1775 vm_object_destroy(backing_object); 1776 1777 object_collapses++; 1778 } else { 1779 vm_object_t new_backing_object; 1780 1781 /* 1782 * If we do not entirely shadow the backing object, 1783 * there is nothing we can do so we give up. 1784 */ 1785 if (object->resident_page_count != object->size && 1786 vm_object_backing_scan(object, 1787 OBSC_TEST_ALL_SHADOWED) == 0) { 1788 VM_OBJECT_UNLOCK(backing_object); 1789 break; 1790 } 1791 1792 /* 1793 * Make the parent shadow the next object in the 1794 * chain. Deallocating backing_object will not remove 1795 * it, since its reference count is at least 2. 1796 */ 1797 LIST_REMOVE(object, shadow_list); 1798 backing_object->shadow_count--; 1799 1800 new_backing_object = backing_object->backing_object; 1801 if ((object->backing_object = new_backing_object) != NULL) { 1802 VM_OBJECT_LOCK(new_backing_object); 1803 LIST_INSERT_HEAD( 1804 &new_backing_object->shadow_head, 1805 object, 1806 shadow_list 1807 ); 1808 new_backing_object->shadow_count++; 1809 vm_object_reference_locked(new_backing_object); 1810 VM_OBJECT_UNLOCK(new_backing_object); 1811 object->backing_object_offset += 1812 backing_object->backing_object_offset; 1813 } 1814 1815 /* 1816 * Drop the reference count on backing_object. Since 1817 * its ref_count was at least 2, it will not vanish. 1818 */ 1819 backing_object->ref_count--; 1820 VM_OBJECT_UNLOCK(backing_object); 1821 object_bypasses++; 1822 } 1823 1824 /* 1825 * Try again with this object's new backing object. 1826 */ 1827 } 1828 } 1829 1830 /* 1831 * vm_object_page_remove: 1832 * 1833 * For the given object, either frees or invalidates each of the 1834 * specified pages. In general, a page is freed. However, if a page is 1835 * wired for any reason other than the existence of a managed, wired 1836 * mapping, then it may be invalidated but not removed from the object. 1837 * Pages are specified by the given range ["start", "end") and the option 1838 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1839 * extends from "start" to the end of the object. If the option 1840 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1841 * specified range are affected. If the option OBJPR_NOTMAPPED is 1842 * specified, then the pages within the specified range must have no 1843 * mappings. Otherwise, if this option is not specified, any mappings to 1844 * the specified pages are removed before the pages are freed or 1845 * invalidated. 1846 * 1847 * In general, this operation should only be performed on objects that 1848 * contain managed pages. There are, however, two exceptions. First, it 1849 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1850 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1851 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1852 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1853 * 1854 * The object must be locked. 1855 */ 1856 void 1857 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1858 int options) 1859 { 1860 vm_page_t p, next; 1861 int wirings; 1862 1863 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1864 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) || 1865 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1866 ("vm_object_page_remove: illegal options for object %p", object)); 1867 if (object->resident_page_count == 0) 1868 goto skipmemq; 1869 vm_object_pip_add(object, 1); 1870 again: 1871 p = vm_page_find_least(object, start); 1872 1873 /* 1874 * Here, the variable "p" is either (1) the page with the least pindex 1875 * greater than or equal to the parameter "start" or (2) NULL. 1876 */ 1877 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1878 next = TAILQ_NEXT(p, listq); 1879 1880 /* 1881 * If the page is wired for any reason besides the existence 1882 * of managed, wired mappings, then it cannot be freed. For 1883 * example, fictitious pages, which represent device memory, 1884 * are inherently wired and cannot be freed. They can, 1885 * however, be invalidated if the option OBJPR_CLEANONLY is 1886 * not specified. 1887 */ 1888 vm_page_lock(p); 1889 if ((wirings = p->wire_count) != 0 && 1890 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) { 1891 if ((options & OBJPR_NOTMAPPED) == 0) { 1892 pmap_remove_all(p); 1893 /* Account for removal of wired mappings. */ 1894 if (wirings != 0) 1895 p->wire_count -= wirings; 1896 } 1897 if ((options & OBJPR_CLEANONLY) == 0) { 1898 p->valid = 0; 1899 vm_page_undirty(p); 1900 } 1901 vm_page_unlock(p); 1902 continue; 1903 } 1904 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1905 goto again; 1906 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1907 ("vm_object_page_remove: page %p is fictitious", p)); 1908 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1909 if ((options & OBJPR_NOTMAPPED) == 0) 1910 pmap_remove_write(p); 1911 if (p->dirty) { 1912 vm_page_unlock(p); 1913 continue; 1914 } 1915 } 1916 if ((options & OBJPR_NOTMAPPED) == 0) { 1917 pmap_remove_all(p); 1918 /* Account for removal of wired mappings. */ 1919 if (wirings != 0) { 1920 KASSERT(p->wire_count == wirings, 1921 ("inconsistent wire count %d %d %p", 1922 p->wire_count, wirings, p)); 1923 p->wire_count = 0; 1924 atomic_subtract_int(&cnt.v_wire_count, 1); 1925 } 1926 } 1927 vm_page_free(p); 1928 vm_page_unlock(p); 1929 } 1930 vm_object_pip_wakeup(object); 1931 skipmemq: 1932 if (__predict_false(object->cache != NULL)) 1933 vm_page_cache_free(object, start, end); 1934 } 1935 1936 /* 1937 * vm_object_page_cache: 1938 * 1939 * For the given object, attempt to move the specified clean 1940 * pages to the cache queue. If a page is wired for any reason, 1941 * then it will not be changed. Pages are specified by the given 1942 * range ["start", "end"). As a special case, if "end" is zero, 1943 * then the range extends from "start" to the end of the object. 1944 * Any mappings to the specified pages are removed before the 1945 * pages are moved to the cache queue. 1946 * 1947 * This operation should only be performed on objects that 1948 * contain managed pages. 1949 * 1950 * The object must be locked. 1951 */ 1952 void 1953 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1954 { 1955 struct mtx *mtx, *new_mtx; 1956 vm_page_t p, next; 1957 1958 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1959 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG && 1960 object->type != OBJT_PHYS), 1961 ("vm_object_page_cache: illegal object %p", object)); 1962 if (object->resident_page_count == 0) 1963 return; 1964 p = vm_page_find_least(object, start); 1965 1966 /* 1967 * Here, the variable "p" is either (1) the page with the least pindex 1968 * greater than or equal to the parameter "start" or (2) NULL. 1969 */ 1970 mtx = NULL; 1971 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1972 next = TAILQ_NEXT(p, listq); 1973 1974 /* 1975 * Avoid releasing and reacquiring the same page lock. 1976 */ 1977 new_mtx = vm_page_lockptr(p); 1978 if (mtx != new_mtx) { 1979 if (mtx != NULL) 1980 mtx_unlock(mtx); 1981 mtx = new_mtx; 1982 mtx_lock(mtx); 1983 } 1984 vm_page_try_to_cache(p); 1985 } 1986 if (mtx != NULL) 1987 mtx_unlock(mtx); 1988 } 1989 1990 /* 1991 * Populate the specified range of the object with valid pages. Returns 1992 * TRUE if the range is successfully populated and FALSE otherwise. 1993 * 1994 * Note: This function should be optimized to pass a larger array of 1995 * pages to vm_pager_get_pages() before it is applied to a non- 1996 * OBJT_DEVICE object. 1997 * 1998 * The object must be locked. 1999 */ 2000 boolean_t 2001 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2002 { 2003 vm_page_t m, ma[1]; 2004 vm_pindex_t pindex; 2005 int rv; 2006 2007 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2008 for (pindex = start; pindex < end; pindex++) { 2009 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | 2010 VM_ALLOC_RETRY); 2011 if (m->valid != VM_PAGE_BITS_ALL) { 2012 ma[0] = m; 2013 rv = vm_pager_get_pages(object, ma, 1, 0); 2014 m = vm_page_lookup(object, pindex); 2015 if (m == NULL) 2016 break; 2017 if (rv != VM_PAGER_OK) { 2018 vm_page_lock(m); 2019 vm_page_free(m); 2020 vm_page_unlock(m); 2021 break; 2022 } 2023 } 2024 /* 2025 * Keep "m" busy because a subsequent iteration may unlock 2026 * the object. 2027 */ 2028 } 2029 if (pindex > start) { 2030 m = vm_page_lookup(object, start); 2031 while (m != NULL && m->pindex < pindex) { 2032 vm_page_wakeup(m); 2033 m = TAILQ_NEXT(m, listq); 2034 } 2035 } 2036 return (pindex == end); 2037 } 2038 2039 /* 2040 * Routine: vm_object_coalesce 2041 * Function: Coalesces two objects backing up adjoining 2042 * regions of memory into a single object. 2043 * 2044 * returns TRUE if objects were combined. 2045 * 2046 * NOTE: Only works at the moment if the second object is NULL - 2047 * if it's not, which object do we lock first? 2048 * 2049 * Parameters: 2050 * prev_object First object to coalesce 2051 * prev_offset Offset into prev_object 2052 * prev_size Size of reference to prev_object 2053 * next_size Size of reference to the second object 2054 * reserved Indicator that extension region has 2055 * swap accounted for 2056 * 2057 * Conditions: 2058 * The object must *not* be locked. 2059 */ 2060 boolean_t 2061 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2062 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2063 { 2064 vm_pindex_t next_pindex; 2065 2066 if (prev_object == NULL) 2067 return (TRUE); 2068 VM_OBJECT_LOCK(prev_object); 2069 if (prev_object->type != OBJT_DEFAULT && 2070 prev_object->type != OBJT_SWAP) { 2071 VM_OBJECT_UNLOCK(prev_object); 2072 return (FALSE); 2073 } 2074 2075 /* 2076 * Try to collapse the object first 2077 */ 2078 vm_object_collapse(prev_object); 2079 2080 /* 2081 * Can't coalesce if: . more than one reference . paged out . shadows 2082 * another object . has a copy elsewhere (any of which mean that the 2083 * pages not mapped to prev_entry may be in use anyway) 2084 */ 2085 if (prev_object->backing_object != NULL) { 2086 VM_OBJECT_UNLOCK(prev_object); 2087 return (FALSE); 2088 } 2089 2090 prev_size >>= PAGE_SHIFT; 2091 next_size >>= PAGE_SHIFT; 2092 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2093 2094 if ((prev_object->ref_count > 1) && 2095 (prev_object->size != next_pindex)) { 2096 VM_OBJECT_UNLOCK(prev_object); 2097 return (FALSE); 2098 } 2099 2100 /* 2101 * Account for the charge. 2102 */ 2103 if (prev_object->cred != NULL) { 2104 2105 /* 2106 * If prev_object was charged, then this mapping, 2107 * althought not charged now, may become writable 2108 * later. Non-NULL cred in the object would prevent 2109 * swap reservation during enabling of the write 2110 * access, so reserve swap now. Failed reservation 2111 * cause allocation of the separate object for the map 2112 * entry, and swap reservation for this entry is 2113 * managed in appropriate time. 2114 */ 2115 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2116 prev_object->cred)) { 2117 return (FALSE); 2118 } 2119 prev_object->charge += ptoa(next_size); 2120 } 2121 2122 /* 2123 * Remove any pages that may still be in the object from a previous 2124 * deallocation. 2125 */ 2126 if (next_pindex < prev_object->size) { 2127 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2128 next_size, 0); 2129 if (prev_object->type == OBJT_SWAP) 2130 swap_pager_freespace(prev_object, 2131 next_pindex, next_size); 2132 #if 0 2133 if (prev_object->cred != NULL) { 2134 KASSERT(prev_object->charge >= 2135 ptoa(prev_object->size - next_pindex), 2136 ("object %p overcharged 1 %jx %jx", prev_object, 2137 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2138 prev_object->charge -= ptoa(prev_object->size - 2139 next_pindex); 2140 } 2141 #endif 2142 } 2143 2144 /* 2145 * Extend the object if necessary. 2146 */ 2147 if (next_pindex + next_size > prev_object->size) 2148 prev_object->size = next_pindex + next_size; 2149 2150 VM_OBJECT_UNLOCK(prev_object); 2151 return (TRUE); 2152 } 2153 2154 void 2155 vm_object_set_writeable_dirty(vm_object_t object) 2156 { 2157 2158 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2159 if (object->type != OBJT_VNODE) 2160 return; 2161 object->generation++; 2162 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2163 return; 2164 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2165 } 2166 2167 #include "opt_ddb.h" 2168 #ifdef DDB 2169 #include <sys/kernel.h> 2170 2171 #include <sys/cons.h> 2172 2173 #include <ddb/ddb.h> 2174 2175 static int 2176 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2177 { 2178 vm_map_t tmpm; 2179 vm_map_entry_t tmpe; 2180 vm_object_t obj; 2181 int entcount; 2182 2183 if (map == 0) 2184 return 0; 2185 2186 if (entry == 0) { 2187 tmpe = map->header.next; 2188 entcount = map->nentries; 2189 while (entcount-- && (tmpe != &map->header)) { 2190 if (_vm_object_in_map(map, object, tmpe)) { 2191 return 1; 2192 } 2193 tmpe = tmpe->next; 2194 } 2195 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2196 tmpm = entry->object.sub_map; 2197 tmpe = tmpm->header.next; 2198 entcount = tmpm->nentries; 2199 while (entcount-- && tmpe != &tmpm->header) { 2200 if (_vm_object_in_map(tmpm, object, tmpe)) { 2201 return 1; 2202 } 2203 tmpe = tmpe->next; 2204 } 2205 } else if ((obj = entry->object.vm_object) != NULL) { 2206 for (; obj; obj = obj->backing_object) 2207 if (obj == object) { 2208 return 1; 2209 } 2210 } 2211 return 0; 2212 } 2213 2214 static int 2215 vm_object_in_map(vm_object_t object) 2216 { 2217 struct proc *p; 2218 2219 /* sx_slock(&allproc_lock); */ 2220 FOREACH_PROC_IN_SYSTEM(p) { 2221 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2222 continue; 2223 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2224 /* sx_sunlock(&allproc_lock); */ 2225 return 1; 2226 } 2227 } 2228 /* sx_sunlock(&allproc_lock); */ 2229 if (_vm_object_in_map(kernel_map, object, 0)) 2230 return 1; 2231 if (_vm_object_in_map(kmem_map, object, 0)) 2232 return 1; 2233 if (_vm_object_in_map(pager_map, object, 0)) 2234 return 1; 2235 if (_vm_object_in_map(buffer_map, object, 0)) 2236 return 1; 2237 return 0; 2238 } 2239 2240 DB_SHOW_COMMAND(vmochk, vm_object_check) 2241 { 2242 vm_object_t object; 2243 2244 /* 2245 * make sure that internal objs are in a map somewhere 2246 * and none have zero ref counts. 2247 */ 2248 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2249 if (object->handle == NULL && 2250 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2251 if (object->ref_count == 0) { 2252 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2253 (long)object->size); 2254 } 2255 if (!vm_object_in_map(object)) { 2256 db_printf( 2257 "vmochk: internal obj is not in a map: " 2258 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2259 object->ref_count, (u_long)object->size, 2260 (u_long)object->size, 2261 (void *)object->backing_object); 2262 } 2263 } 2264 } 2265 } 2266 2267 /* 2268 * vm_object_print: [ debug ] 2269 */ 2270 DB_SHOW_COMMAND(object, vm_object_print_static) 2271 { 2272 /* XXX convert args. */ 2273 vm_object_t object = (vm_object_t)addr; 2274 boolean_t full = have_addr; 2275 2276 vm_page_t p; 2277 2278 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2279 #define count was_count 2280 2281 int count; 2282 2283 if (object == NULL) 2284 return; 2285 2286 db_iprintf( 2287 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2288 object, (int)object->type, (uintmax_t)object->size, 2289 object->resident_page_count, object->ref_count, object->flags, 2290 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2291 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2292 object->shadow_count, 2293 object->backing_object ? object->backing_object->ref_count : 0, 2294 object->backing_object, (uintmax_t)object->backing_object_offset); 2295 2296 if (!full) 2297 return; 2298 2299 db_indent += 2; 2300 count = 0; 2301 TAILQ_FOREACH(p, &object->memq, listq) { 2302 if (count == 0) 2303 db_iprintf("memory:="); 2304 else if (count == 6) { 2305 db_printf("\n"); 2306 db_iprintf(" ..."); 2307 count = 0; 2308 } else 2309 db_printf(","); 2310 count++; 2311 2312 db_printf("(off=0x%jx,page=0x%jx)", 2313 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2314 } 2315 if (count != 0) 2316 db_printf("\n"); 2317 db_indent -= 2; 2318 } 2319 2320 /* XXX. */ 2321 #undef count 2322 2323 /* XXX need this non-static entry for calling from vm_map_print. */ 2324 void 2325 vm_object_print( 2326 /* db_expr_t */ long addr, 2327 boolean_t have_addr, 2328 /* db_expr_t */ long count, 2329 char *modif) 2330 { 2331 vm_object_print_static(addr, have_addr, count, modif); 2332 } 2333 2334 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2335 { 2336 vm_object_t object; 2337 vm_pindex_t fidx; 2338 vm_paddr_t pa; 2339 vm_page_t m, prev_m; 2340 int rcount, nl, c; 2341 2342 nl = 0; 2343 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2344 db_printf("new object: %p\n", (void *)object); 2345 if (nl > 18) { 2346 c = cngetc(); 2347 if (c != ' ') 2348 return; 2349 nl = 0; 2350 } 2351 nl++; 2352 rcount = 0; 2353 fidx = 0; 2354 pa = -1; 2355 TAILQ_FOREACH(m, &object->memq, listq) { 2356 if (m->pindex > 128) 2357 break; 2358 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2359 prev_m->pindex + 1 != m->pindex) { 2360 if (rcount) { 2361 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2362 (long)fidx, rcount, (long)pa); 2363 if (nl > 18) { 2364 c = cngetc(); 2365 if (c != ' ') 2366 return; 2367 nl = 0; 2368 } 2369 nl++; 2370 rcount = 0; 2371 } 2372 } 2373 if (rcount && 2374 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2375 ++rcount; 2376 continue; 2377 } 2378 if (rcount) { 2379 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2380 (long)fidx, rcount, (long)pa); 2381 if (nl > 18) { 2382 c = cngetc(); 2383 if (c != ' ') 2384 return; 2385 nl = 0; 2386 } 2387 nl++; 2388 } 2389 fidx = m->pindex; 2390 pa = VM_PAGE_TO_PHYS(m); 2391 rcount = 1; 2392 } 2393 if (rcount) { 2394 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2395 (long)fidx, rcount, (long)pa); 2396 if (nl > 18) { 2397 c = cngetc(); 2398 if (c != ' ') 2399 return; 2400 nl = 0; 2401 } 2402 nl++; 2403 } 2404 } 2405 } 2406 #endif /* DDB */ 2407