1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Virtual memory object module. 65 */ 66 67 #include <sys/cdefs.h> 68 __FBSDID("$FreeBSD$"); 69 70 #include "opt_vm.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/cpuset.h> 75 #include <sys/lock.h> 76 #include <sys/mman.h> 77 #include <sys/mount.h> 78 #include <sys/kernel.h> 79 #include <sys/pctrie.h> 80 #include <sys/sysctl.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> /* for curproc, pageproc */ 83 #include <sys/refcount.h> 84 #include <sys/socket.h> 85 #include <sys/resourcevar.h> 86 #include <sys/rwlock.h> 87 #include <sys/user.h> 88 #include <sys/vnode.h> 89 #include <sys/vmmeter.h> 90 #include <sys/sx.h> 91 92 #include <vm/vm.h> 93 #include <vm/vm_param.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_page.h> 98 #include <vm/vm_pageout.h> 99 #include <vm/vm_pager.h> 100 #include <vm/vm_phys.h> 101 #include <vm/vm_pagequeue.h> 102 #include <vm/swap_pager.h> 103 #include <vm/vm_kern.h> 104 #include <vm/vm_extern.h> 105 #include <vm/vm_radix.h> 106 #include <vm/vm_reserv.h> 107 #include <vm/uma.h> 108 109 static int old_msync; 110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 111 "Use old (insecure) msync behavior"); 112 113 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 114 int pagerflags, int flags, boolean_t *clearobjflags, 115 boolean_t *eio); 116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 117 boolean_t *clearobjflags); 118 static void vm_object_qcollapse(vm_object_t object); 119 static void vm_object_vndeallocate(vm_object_t object); 120 121 /* 122 * Virtual memory objects maintain the actual data 123 * associated with allocated virtual memory. A given 124 * page of memory exists within exactly one object. 125 * 126 * An object is only deallocated when all "references" 127 * are given up. Only one "reference" to a given 128 * region of an object should be writeable. 129 * 130 * Associated with each object is a list of all resident 131 * memory pages belonging to that object; this list is 132 * maintained by the "vm_page" module, and locked by the object's 133 * lock. 134 * 135 * Each object also records a "pager" routine which is 136 * used to retrieve (and store) pages to the proper backing 137 * storage. In addition, objects may be backed by other 138 * objects from which they were virtual-copied. 139 * 140 * The only items within the object structure which are 141 * modified after time of creation are: 142 * reference count locked by object's lock 143 * pager routine locked by object's lock 144 * 145 */ 146 147 struct object_q vm_object_list; 148 struct mtx vm_object_list_mtx; /* lock for object list and count */ 149 150 struct vm_object kernel_object_store; 151 152 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 153 "VM object stats"); 154 155 static counter_u64_t object_collapses = EARLY_COUNTER; 156 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 157 &object_collapses, 158 "VM object collapses"); 159 160 static counter_u64_t object_bypasses = EARLY_COUNTER; 161 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 162 &object_bypasses, 163 "VM object bypasses"); 164 165 static void 166 counter_startup(void) 167 { 168 169 object_collapses = counter_u64_alloc(M_WAITOK); 170 object_bypasses = counter_u64_alloc(M_WAITOK); 171 } 172 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); 173 174 static uma_zone_t obj_zone; 175 176 static int vm_object_zinit(void *mem, int size, int flags); 177 178 #ifdef INVARIANTS 179 static void vm_object_zdtor(void *mem, int size, void *arg); 180 181 static void 182 vm_object_zdtor(void *mem, int size, void *arg) 183 { 184 vm_object_t object; 185 186 object = (vm_object_t)mem; 187 KASSERT(object->ref_count == 0, 188 ("object %p ref_count = %d", object, object->ref_count)); 189 KASSERT(TAILQ_EMPTY(&object->memq), 190 ("object %p has resident pages in its memq", object)); 191 KASSERT(vm_radix_is_empty(&object->rtree), 192 ("object %p has resident pages in its trie", object)); 193 #if VM_NRESERVLEVEL > 0 194 KASSERT(LIST_EMPTY(&object->rvq), 195 ("object %p has reservations", 196 object)); 197 #endif 198 KASSERT(object->paging_in_progress == 0, 199 ("object %p paging_in_progress = %d", 200 object, object->paging_in_progress)); 201 KASSERT(object->resident_page_count == 0, 202 ("object %p resident_page_count = %d", 203 object, object->resident_page_count)); 204 KASSERT(object->shadow_count == 0, 205 ("object %p shadow_count = %d", 206 object, object->shadow_count)); 207 KASSERT(object->type == OBJT_DEAD, 208 ("object %p has non-dead type %d", 209 object, object->type)); 210 } 211 #endif 212 213 static int 214 vm_object_zinit(void *mem, int size, int flags) 215 { 216 vm_object_t object; 217 218 object = (vm_object_t)mem; 219 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 220 221 /* These are true for any object that has been freed */ 222 object->type = OBJT_DEAD; 223 object->ref_count = 0; 224 vm_radix_init(&object->rtree); 225 refcount_init(&object->paging_in_progress, 0); 226 object->resident_page_count = 0; 227 object->shadow_count = 0; 228 object->flags = OBJ_DEAD; 229 230 mtx_lock(&vm_object_list_mtx); 231 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 232 mtx_unlock(&vm_object_list_mtx); 233 return (0); 234 } 235 236 static void 237 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 238 { 239 240 TAILQ_INIT(&object->memq); 241 LIST_INIT(&object->shadow_head); 242 243 object->type = type; 244 if (type == OBJT_SWAP) 245 pctrie_init(&object->un_pager.swp.swp_blks); 246 247 /* 248 * Ensure that swap_pager_swapoff() iteration over object_list 249 * sees up to date type and pctrie head if it observed 250 * non-dead object. 251 */ 252 atomic_thread_fence_rel(); 253 254 switch (type) { 255 case OBJT_DEAD: 256 panic("_vm_object_allocate: can't create OBJT_DEAD"); 257 case OBJT_DEFAULT: 258 case OBJT_SWAP: 259 object->flags = OBJ_ONEMAPPING; 260 break; 261 case OBJT_DEVICE: 262 case OBJT_SG: 263 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 264 break; 265 case OBJT_MGTDEVICE: 266 object->flags = OBJ_FICTITIOUS; 267 break; 268 case OBJT_PHYS: 269 object->flags = OBJ_UNMANAGED; 270 break; 271 case OBJT_VNODE: 272 object->flags = 0; 273 break; 274 default: 275 panic("_vm_object_allocate: type %d is undefined", type); 276 } 277 object->size = size; 278 object->domain.dr_policy = NULL; 279 object->generation = 1; 280 object->ref_count = 1; 281 object->memattr = VM_MEMATTR_DEFAULT; 282 object->cred = NULL; 283 object->charge = 0; 284 object->handle = NULL; 285 object->backing_object = NULL; 286 object->backing_object_offset = (vm_ooffset_t) 0; 287 #if VM_NRESERVLEVEL > 0 288 LIST_INIT(&object->rvq); 289 #endif 290 umtx_shm_object_init(object); 291 } 292 293 /* 294 * vm_object_init: 295 * 296 * Initialize the VM objects module. 297 */ 298 void 299 vm_object_init(void) 300 { 301 TAILQ_INIT(&vm_object_list); 302 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 303 304 rw_init(&kernel_object->lock, "kernel vm object"); 305 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 306 VM_MIN_KERNEL_ADDRESS), kernel_object); 307 #if VM_NRESERVLEVEL > 0 308 kernel_object->flags |= OBJ_COLORED; 309 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 310 #endif 311 312 /* 313 * The lock portion of struct vm_object must be type stable due 314 * to vm_pageout_fallback_object_lock locking a vm object 315 * without holding any references to it. 316 */ 317 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 318 #ifdef INVARIANTS 319 vm_object_zdtor, 320 #else 321 NULL, 322 #endif 323 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 324 325 vm_radix_zinit(); 326 } 327 328 void 329 vm_object_clear_flag(vm_object_t object, u_short bits) 330 { 331 332 VM_OBJECT_ASSERT_WLOCKED(object); 333 object->flags &= ~bits; 334 } 335 336 /* 337 * Sets the default memory attribute for the specified object. Pages 338 * that are allocated to this object are by default assigned this memory 339 * attribute. 340 * 341 * Presently, this function must be called before any pages are allocated 342 * to the object. In the future, this requirement may be relaxed for 343 * "default" and "swap" objects. 344 */ 345 int 346 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 347 { 348 349 VM_OBJECT_ASSERT_WLOCKED(object); 350 switch (object->type) { 351 case OBJT_DEFAULT: 352 case OBJT_DEVICE: 353 case OBJT_MGTDEVICE: 354 case OBJT_PHYS: 355 case OBJT_SG: 356 case OBJT_SWAP: 357 case OBJT_VNODE: 358 if (!TAILQ_EMPTY(&object->memq)) 359 return (KERN_FAILURE); 360 break; 361 case OBJT_DEAD: 362 return (KERN_INVALID_ARGUMENT); 363 default: 364 panic("vm_object_set_memattr: object %p is of undefined type", 365 object); 366 } 367 object->memattr = memattr; 368 return (KERN_SUCCESS); 369 } 370 371 void 372 vm_object_pip_add(vm_object_t object, short i) 373 { 374 375 refcount_acquiren(&object->paging_in_progress, i); 376 } 377 378 void 379 vm_object_pip_wakeup(vm_object_t object) 380 { 381 382 refcount_release(&object->paging_in_progress); 383 } 384 385 void 386 vm_object_pip_wakeupn(vm_object_t object, short i) 387 { 388 389 refcount_releasen(&object->paging_in_progress, i); 390 } 391 392 void 393 vm_object_pip_wait(vm_object_t object, char *waitid) 394 { 395 396 VM_OBJECT_ASSERT_WLOCKED(object); 397 398 while (object->paging_in_progress) { 399 VM_OBJECT_WUNLOCK(object); 400 refcount_wait(&object->paging_in_progress, waitid, PVM); 401 VM_OBJECT_WLOCK(object); 402 } 403 } 404 405 void 406 vm_object_pip_wait_unlocked(vm_object_t object, char *waitid) 407 { 408 409 VM_OBJECT_ASSERT_UNLOCKED(object); 410 411 while (object->paging_in_progress) 412 refcount_wait(&object->paging_in_progress, waitid, PVM); 413 } 414 415 /* 416 * vm_object_allocate: 417 * 418 * Returns a new object with the given size. 419 */ 420 vm_object_t 421 vm_object_allocate(objtype_t type, vm_pindex_t size) 422 { 423 vm_object_t object; 424 425 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 426 _vm_object_allocate(type, size, object); 427 return (object); 428 } 429 430 431 /* 432 * vm_object_reference: 433 * 434 * Gets another reference to the given object. Note: OBJ_DEAD 435 * objects can be referenced during final cleaning. 436 */ 437 void 438 vm_object_reference(vm_object_t object) 439 { 440 if (object == NULL) 441 return; 442 VM_OBJECT_WLOCK(object); 443 vm_object_reference_locked(object); 444 VM_OBJECT_WUNLOCK(object); 445 } 446 447 /* 448 * vm_object_reference_locked: 449 * 450 * Gets another reference to the given object. 451 * 452 * The object must be locked. 453 */ 454 void 455 vm_object_reference_locked(vm_object_t object) 456 { 457 struct vnode *vp; 458 459 VM_OBJECT_ASSERT_WLOCKED(object); 460 object->ref_count++; 461 if (object->type == OBJT_VNODE) { 462 vp = object->handle; 463 vref(vp); 464 } 465 } 466 467 /* 468 * Handle deallocating an object of type OBJT_VNODE. 469 */ 470 static void 471 vm_object_vndeallocate(vm_object_t object) 472 { 473 struct vnode *vp = (struct vnode *) object->handle; 474 475 VM_OBJECT_ASSERT_WLOCKED(object); 476 KASSERT(object->type == OBJT_VNODE, 477 ("vm_object_vndeallocate: not a vnode object")); 478 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 479 #ifdef INVARIANTS 480 if (object->ref_count == 0) { 481 vn_printf(vp, "vm_object_vndeallocate "); 482 panic("vm_object_vndeallocate: bad object reference count"); 483 } 484 #endif 485 486 if (!umtx_shm_vnobj_persistent && object->ref_count == 1) 487 umtx_shm_object_terminated(object); 488 489 object->ref_count--; 490 491 /* vrele may need the vnode lock. */ 492 VM_OBJECT_WUNLOCK(object); 493 vrele(vp); 494 } 495 496 /* 497 * vm_object_deallocate: 498 * 499 * Release a reference to the specified object, 500 * gained either through a vm_object_allocate 501 * or a vm_object_reference call. When all references 502 * are gone, storage associated with this object 503 * may be relinquished. 504 * 505 * No object may be locked. 506 */ 507 void 508 vm_object_deallocate(vm_object_t object) 509 { 510 vm_object_t temp; 511 struct vnode *vp; 512 513 while (object != NULL) { 514 VM_OBJECT_WLOCK(object); 515 if (object->type == OBJT_VNODE) { 516 vm_object_vndeallocate(object); 517 return; 518 } 519 520 KASSERT(object->ref_count != 0, 521 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 522 523 /* 524 * If the reference count goes to 0 we start calling 525 * vm_object_terminate() on the object chain. 526 * A ref count of 1 may be a special case depending on the 527 * shadow count being 0 or 1. 528 */ 529 object->ref_count--; 530 if (object->ref_count > 1) { 531 VM_OBJECT_WUNLOCK(object); 532 return; 533 } else if (object->ref_count == 1) { 534 if (object->type == OBJT_SWAP && 535 (object->flags & OBJ_TMPFS) != 0) { 536 vp = object->un_pager.swp.swp_tmpfs; 537 vhold(vp); 538 VM_OBJECT_WUNLOCK(object); 539 vn_lock(vp, LK_SHARED | LK_RETRY); 540 VM_OBJECT_WLOCK(object); 541 if (object->type == OBJT_DEAD || 542 object->ref_count != 1) { 543 VM_OBJECT_WUNLOCK(object); 544 VOP_UNLOCK(vp, 0); 545 vdrop(vp); 546 return; 547 } 548 if ((object->flags & OBJ_TMPFS) != 0) 549 VOP_UNSET_TEXT(vp); 550 VOP_UNLOCK(vp, 0); 551 vdrop(vp); 552 } 553 if (object->shadow_count == 0 && 554 object->handle == NULL && 555 (object->type == OBJT_DEFAULT || 556 (object->type == OBJT_SWAP && 557 (object->flags & OBJ_TMPFS_NODE) == 0))) { 558 vm_object_set_flag(object, OBJ_ONEMAPPING); 559 } else if ((object->shadow_count == 1) && 560 (object->handle == NULL) && 561 (object->type == OBJT_DEFAULT || 562 object->type == OBJT_SWAP)) { 563 vm_object_t robject; 564 565 robject = LIST_FIRST(&object->shadow_head); 566 KASSERT(robject != NULL, 567 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 568 object->ref_count, 569 object->shadow_count)); 570 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 571 ("shadowed tmpfs v_object %p", object)); 572 if (!VM_OBJECT_TRYWLOCK(robject)) { 573 /* 574 * Avoid a potential deadlock. 575 */ 576 object->ref_count++; 577 VM_OBJECT_WUNLOCK(object); 578 /* 579 * More likely than not the thread 580 * holding robject's lock has lower 581 * priority than the current thread. 582 * Let the lower priority thread run. 583 */ 584 pause("vmo_de", 1); 585 continue; 586 } 587 /* 588 * Collapse object into its shadow unless its 589 * shadow is dead. In that case, object will 590 * be deallocated by the thread that is 591 * deallocating its shadow. 592 */ 593 if ((robject->flags & OBJ_DEAD) == 0 && 594 (robject->handle == NULL) && 595 (robject->type == OBJT_DEFAULT || 596 robject->type == OBJT_SWAP)) { 597 598 robject->ref_count++; 599 retry: 600 if (robject->paging_in_progress) { 601 VM_OBJECT_WUNLOCK(object); 602 vm_object_pip_wait(robject, 603 "objde1"); 604 temp = robject->backing_object; 605 if (object == temp) { 606 VM_OBJECT_WLOCK(object); 607 goto retry; 608 } 609 } else if (object->paging_in_progress) { 610 VM_OBJECT_WUNLOCK(robject); 611 VM_OBJECT_WUNLOCK(object); 612 refcount_wait( 613 &object->paging_in_progress, 614 "objde2", PVM); 615 VM_OBJECT_WLOCK(robject); 616 temp = robject->backing_object; 617 if (object == temp) { 618 VM_OBJECT_WLOCK(object); 619 goto retry; 620 } 621 } else 622 VM_OBJECT_WUNLOCK(object); 623 624 if (robject->ref_count == 1) { 625 robject->ref_count--; 626 object = robject; 627 goto doterm; 628 } 629 object = robject; 630 vm_object_collapse(object); 631 VM_OBJECT_WUNLOCK(object); 632 continue; 633 } 634 VM_OBJECT_WUNLOCK(robject); 635 } 636 VM_OBJECT_WUNLOCK(object); 637 return; 638 } 639 doterm: 640 umtx_shm_object_terminated(object); 641 temp = object->backing_object; 642 if (temp != NULL) { 643 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 644 ("shadowed tmpfs v_object 2 %p", object)); 645 VM_OBJECT_WLOCK(temp); 646 LIST_REMOVE(object, shadow_list); 647 temp->shadow_count--; 648 VM_OBJECT_WUNLOCK(temp); 649 object->backing_object = NULL; 650 } 651 /* 652 * Don't double-terminate, we could be in a termination 653 * recursion due to the terminate having to sync data 654 * to disk. 655 */ 656 if ((object->flags & OBJ_DEAD) == 0) { 657 vm_object_set_flag(object, OBJ_DEAD); 658 vm_object_terminate(object); 659 } else 660 VM_OBJECT_WUNLOCK(object); 661 object = temp; 662 } 663 } 664 665 /* 666 * vm_object_destroy removes the object from the global object list 667 * and frees the space for the object. 668 */ 669 void 670 vm_object_destroy(vm_object_t object) 671 { 672 673 /* 674 * Release the allocation charge. 675 */ 676 if (object->cred != NULL) { 677 swap_release_by_cred(object->charge, object->cred); 678 object->charge = 0; 679 crfree(object->cred); 680 object->cred = NULL; 681 } 682 683 /* 684 * Free the space for the object. 685 */ 686 uma_zfree(obj_zone, object); 687 } 688 689 /* 690 * vm_object_terminate_pages removes any remaining pageable pages 691 * from the object and resets the object to an empty state. 692 */ 693 static void 694 vm_object_terminate_pages(vm_object_t object) 695 { 696 vm_page_t p, p_next; 697 struct mtx *mtx; 698 699 VM_OBJECT_ASSERT_WLOCKED(object); 700 701 mtx = NULL; 702 703 /* 704 * Free any remaining pageable pages. This also removes them from the 705 * paging queues. However, don't free wired pages, just remove them 706 * from the object. Rather than incrementally removing each page from 707 * the object, the page and object are reset to any empty state. 708 */ 709 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 710 vm_page_assert_unbusied(p); 711 if ((object->flags & OBJ_UNMANAGED) == 0) 712 /* 713 * vm_page_free_prep() only needs the page 714 * lock for managed pages. 715 */ 716 vm_page_change_lock(p, &mtx); 717 p->object = NULL; 718 if (vm_page_wired(p)) 719 continue; 720 VM_CNT_INC(v_pfree); 721 vm_page_free(p); 722 } 723 if (mtx != NULL) 724 mtx_unlock(mtx); 725 726 /* 727 * If the object contained any pages, then reset it to an empty state. 728 * None of the object's fields, including "resident_page_count", were 729 * modified by the preceding loop. 730 */ 731 if (object->resident_page_count != 0) { 732 vm_radix_reclaim_allnodes(&object->rtree); 733 TAILQ_INIT(&object->memq); 734 object->resident_page_count = 0; 735 if (object->type == OBJT_VNODE) 736 vdrop(object->handle); 737 } 738 } 739 740 /* 741 * vm_object_terminate actually destroys the specified object, freeing 742 * up all previously used resources. 743 * 744 * The object must be locked. 745 * This routine may block. 746 */ 747 void 748 vm_object_terminate(vm_object_t object) 749 { 750 VM_OBJECT_ASSERT_WLOCKED(object); 751 KASSERT((object->flags & OBJ_DEAD) != 0, 752 ("terminating non-dead obj %p", object)); 753 754 /* 755 * wait for the pageout daemon to be done with the object 756 */ 757 vm_object_pip_wait(object, "objtrm"); 758 759 KASSERT(!object->paging_in_progress, 760 ("vm_object_terminate: pageout in progress")); 761 762 KASSERT(object->ref_count == 0, 763 ("vm_object_terminate: object with references, ref_count=%d", 764 object->ref_count)); 765 766 if ((object->flags & OBJ_PG_DTOR) == 0) 767 vm_object_terminate_pages(object); 768 769 #if VM_NRESERVLEVEL > 0 770 if (__predict_false(!LIST_EMPTY(&object->rvq))) 771 vm_reserv_break_all(object); 772 #endif 773 774 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 775 object->type == OBJT_SWAP, 776 ("%s: non-swap obj %p has cred", __func__, object)); 777 778 /* 779 * Let the pager know object is dead. 780 */ 781 vm_pager_deallocate(object); 782 VM_OBJECT_WUNLOCK(object); 783 784 vm_object_destroy(object); 785 } 786 787 /* 788 * Make the page read-only so that we can clear the object flags. However, if 789 * this is a nosync mmap then the object is likely to stay dirty so do not 790 * mess with the page and do not clear the object flags. Returns TRUE if the 791 * page should be flushed, and FALSE otherwise. 792 */ 793 static boolean_t 794 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 795 { 796 797 /* 798 * If we have been asked to skip nosync pages and this is a 799 * nosync page, skip it. Note that the object flags were not 800 * cleared in this case so we do not have to set them. 801 */ 802 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 803 *clearobjflags = FALSE; 804 return (FALSE); 805 } else { 806 pmap_remove_write(p); 807 return (p->dirty != 0); 808 } 809 } 810 811 /* 812 * vm_object_page_clean 813 * 814 * Clean all dirty pages in the specified range of object. Leaves page 815 * on whatever queue it is currently on. If NOSYNC is set then do not 816 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 817 * leaving the object dirty. 818 * 819 * When stuffing pages asynchronously, allow clustering. XXX we need a 820 * synchronous clustering mode implementation. 821 * 822 * Odd semantics: if start == end, we clean everything. 823 * 824 * The object must be locked. 825 * 826 * Returns FALSE if some page from the range was not written, as 827 * reported by the pager, and TRUE otherwise. 828 */ 829 boolean_t 830 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 831 int flags) 832 { 833 vm_page_t np, p; 834 vm_pindex_t pi, tend, tstart; 835 int curgeneration, n, pagerflags; 836 boolean_t clearobjflags, eio, res; 837 838 VM_OBJECT_ASSERT_WLOCKED(object); 839 840 /* 841 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 842 * objects. The check below prevents the function from 843 * operating on non-vnode objects. 844 */ 845 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 846 object->resident_page_count == 0) 847 return (TRUE); 848 849 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 850 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 851 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 852 853 tstart = OFF_TO_IDX(start); 854 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 855 clearobjflags = tstart == 0 && tend >= object->size; 856 res = TRUE; 857 858 rescan: 859 curgeneration = object->generation; 860 861 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 862 pi = p->pindex; 863 if (pi >= tend) 864 break; 865 np = TAILQ_NEXT(p, listq); 866 if (p->valid == 0) 867 continue; 868 if (vm_page_sleep_if_busy(p, "vpcwai")) { 869 if (object->generation != curgeneration) { 870 if ((flags & OBJPC_SYNC) != 0) 871 goto rescan; 872 else 873 clearobjflags = FALSE; 874 } 875 np = vm_page_find_least(object, pi); 876 continue; 877 } 878 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 879 continue; 880 881 n = vm_object_page_collect_flush(object, p, pagerflags, 882 flags, &clearobjflags, &eio); 883 if (eio) { 884 res = FALSE; 885 clearobjflags = FALSE; 886 } 887 if (object->generation != curgeneration) { 888 if ((flags & OBJPC_SYNC) != 0) 889 goto rescan; 890 else 891 clearobjflags = FALSE; 892 } 893 894 /* 895 * If the VOP_PUTPAGES() did a truncated write, so 896 * that even the first page of the run is not fully 897 * written, vm_pageout_flush() returns 0 as the run 898 * length. Since the condition that caused truncated 899 * write may be permanent, e.g. exhausted free space, 900 * accepting n == 0 would cause an infinite loop. 901 * 902 * Forwarding the iterator leaves the unwritten page 903 * behind, but there is not much we can do there if 904 * filesystem refuses to write it. 905 */ 906 if (n == 0) { 907 n = 1; 908 clearobjflags = FALSE; 909 } 910 np = vm_page_find_least(object, pi + n); 911 } 912 #if 0 913 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 914 #endif 915 916 if (clearobjflags) 917 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 918 return (res); 919 } 920 921 static int 922 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 923 int flags, boolean_t *clearobjflags, boolean_t *eio) 924 { 925 vm_page_t ma[vm_pageout_page_count], p_first, tp; 926 int count, i, mreq, runlen; 927 928 vm_page_lock_assert(p, MA_NOTOWNED); 929 VM_OBJECT_ASSERT_WLOCKED(object); 930 931 count = 1; 932 mreq = 0; 933 934 for (tp = p; count < vm_pageout_page_count; count++) { 935 tp = vm_page_next(tp); 936 if (tp == NULL || vm_page_busied(tp)) 937 break; 938 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 939 break; 940 } 941 942 for (p_first = p; count < vm_pageout_page_count; count++) { 943 tp = vm_page_prev(p_first); 944 if (tp == NULL || vm_page_busied(tp)) 945 break; 946 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 947 break; 948 p_first = tp; 949 mreq++; 950 } 951 952 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 953 ma[i] = tp; 954 955 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 956 return (runlen); 957 } 958 959 /* 960 * Note that there is absolutely no sense in writing out 961 * anonymous objects, so we track down the vnode object 962 * to write out. 963 * We invalidate (remove) all pages from the address space 964 * for semantic correctness. 965 * 966 * If the backing object is a device object with unmanaged pages, then any 967 * mappings to the specified range of pages must be removed before this 968 * function is called. 969 * 970 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 971 * may start out with a NULL object. 972 */ 973 boolean_t 974 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 975 boolean_t syncio, boolean_t invalidate) 976 { 977 vm_object_t backing_object; 978 struct vnode *vp; 979 struct mount *mp; 980 int error, flags, fsync_after; 981 boolean_t res; 982 983 if (object == NULL) 984 return (TRUE); 985 res = TRUE; 986 error = 0; 987 VM_OBJECT_WLOCK(object); 988 while ((backing_object = object->backing_object) != NULL) { 989 VM_OBJECT_WLOCK(backing_object); 990 offset += object->backing_object_offset; 991 VM_OBJECT_WUNLOCK(object); 992 object = backing_object; 993 if (object->size < OFF_TO_IDX(offset + size)) 994 size = IDX_TO_OFF(object->size) - offset; 995 } 996 /* 997 * Flush pages if writing is allowed, invalidate them 998 * if invalidation requested. Pages undergoing I/O 999 * will be ignored by vm_object_page_remove(). 1000 * 1001 * We cannot lock the vnode and then wait for paging 1002 * to complete without deadlocking against vm_fault. 1003 * Instead we simply call vm_object_page_remove() and 1004 * allow it to block internally on a page-by-page 1005 * basis when it encounters pages undergoing async 1006 * I/O. 1007 */ 1008 if (object->type == OBJT_VNODE && 1009 (object->flags & OBJ_MIGHTBEDIRTY) != 0 && 1010 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1011 VM_OBJECT_WUNLOCK(object); 1012 (void) vn_start_write(vp, &mp, V_WAIT); 1013 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1014 if (syncio && !invalidate && offset == 0 && 1015 atop(size) == object->size) { 1016 /* 1017 * If syncing the whole mapping of the file, 1018 * it is faster to schedule all the writes in 1019 * async mode, also allowing the clustering, 1020 * and then wait for i/o to complete. 1021 */ 1022 flags = 0; 1023 fsync_after = TRUE; 1024 } else { 1025 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1026 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1027 fsync_after = FALSE; 1028 } 1029 VM_OBJECT_WLOCK(object); 1030 res = vm_object_page_clean(object, offset, offset + size, 1031 flags); 1032 VM_OBJECT_WUNLOCK(object); 1033 if (fsync_after) 1034 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1035 VOP_UNLOCK(vp, 0); 1036 vn_finished_write(mp); 1037 if (error != 0) 1038 res = FALSE; 1039 VM_OBJECT_WLOCK(object); 1040 } 1041 if ((object->type == OBJT_VNODE || 1042 object->type == OBJT_DEVICE) && invalidate) { 1043 if (object->type == OBJT_DEVICE) 1044 /* 1045 * The option OBJPR_NOTMAPPED must be passed here 1046 * because vm_object_page_remove() cannot remove 1047 * unmanaged mappings. 1048 */ 1049 flags = OBJPR_NOTMAPPED; 1050 else if (old_msync) 1051 flags = 0; 1052 else 1053 flags = OBJPR_CLEANONLY; 1054 vm_object_page_remove(object, OFF_TO_IDX(offset), 1055 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1056 } 1057 VM_OBJECT_WUNLOCK(object); 1058 return (res); 1059 } 1060 1061 /* 1062 * Determine whether the given advice can be applied to the object. Advice is 1063 * not applied to unmanaged pages since they never belong to page queues, and 1064 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1065 * have been mapped at most once. 1066 */ 1067 static bool 1068 vm_object_advice_applies(vm_object_t object, int advice) 1069 { 1070 1071 if ((object->flags & OBJ_UNMANAGED) != 0) 1072 return (false); 1073 if (advice != MADV_FREE) 1074 return (true); 1075 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && 1076 (object->flags & OBJ_ONEMAPPING) != 0); 1077 } 1078 1079 static void 1080 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1081 vm_size_t size) 1082 { 1083 1084 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1085 swap_pager_freespace(object, pindex, size); 1086 } 1087 1088 /* 1089 * vm_object_madvise: 1090 * 1091 * Implements the madvise function at the object/page level. 1092 * 1093 * MADV_WILLNEED (any object) 1094 * 1095 * Activate the specified pages if they are resident. 1096 * 1097 * MADV_DONTNEED (any object) 1098 * 1099 * Deactivate the specified pages if they are resident. 1100 * 1101 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1102 * OBJ_ONEMAPPING only) 1103 * 1104 * Deactivate and clean the specified pages if they are 1105 * resident. This permits the process to reuse the pages 1106 * without faulting or the kernel to reclaim the pages 1107 * without I/O. 1108 */ 1109 void 1110 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1111 int advice) 1112 { 1113 vm_pindex_t tpindex; 1114 vm_object_t backing_object, tobject; 1115 vm_page_t m, tm; 1116 1117 if (object == NULL) 1118 return; 1119 1120 relookup: 1121 VM_OBJECT_WLOCK(object); 1122 if (!vm_object_advice_applies(object, advice)) { 1123 VM_OBJECT_WUNLOCK(object); 1124 return; 1125 } 1126 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1127 tobject = object; 1128 1129 /* 1130 * If the next page isn't resident in the top-level object, we 1131 * need to search the shadow chain. When applying MADV_FREE, we 1132 * take care to release any swap space used to store 1133 * non-resident pages. 1134 */ 1135 if (m == NULL || pindex < m->pindex) { 1136 /* 1137 * Optimize a common case: if the top-level object has 1138 * no backing object, we can skip over the non-resident 1139 * range in constant time. 1140 */ 1141 if (object->backing_object == NULL) { 1142 tpindex = (m != NULL && m->pindex < end) ? 1143 m->pindex : end; 1144 vm_object_madvise_freespace(object, advice, 1145 pindex, tpindex - pindex); 1146 if ((pindex = tpindex) == end) 1147 break; 1148 goto next_page; 1149 } 1150 1151 tpindex = pindex; 1152 do { 1153 vm_object_madvise_freespace(tobject, advice, 1154 tpindex, 1); 1155 /* 1156 * Prepare to search the next object in the 1157 * chain. 1158 */ 1159 backing_object = tobject->backing_object; 1160 if (backing_object == NULL) 1161 goto next_pindex; 1162 VM_OBJECT_WLOCK(backing_object); 1163 tpindex += 1164 OFF_TO_IDX(tobject->backing_object_offset); 1165 if (tobject != object) 1166 VM_OBJECT_WUNLOCK(tobject); 1167 tobject = backing_object; 1168 if (!vm_object_advice_applies(tobject, advice)) 1169 goto next_pindex; 1170 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1171 NULL); 1172 } else { 1173 next_page: 1174 tm = m; 1175 m = TAILQ_NEXT(m, listq); 1176 } 1177 1178 /* 1179 * If the page is not in a normal state, skip it. 1180 */ 1181 if (tm->valid != VM_PAGE_BITS_ALL) 1182 goto next_pindex; 1183 vm_page_lock(tm); 1184 if (vm_page_wired(tm)) { 1185 vm_page_unlock(tm); 1186 goto next_pindex; 1187 } 1188 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1189 ("vm_object_madvise: page %p is fictitious", tm)); 1190 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1191 ("vm_object_madvise: page %p is not managed", tm)); 1192 if (vm_page_busied(tm)) { 1193 if (object != tobject) 1194 VM_OBJECT_WUNLOCK(tobject); 1195 VM_OBJECT_WUNLOCK(object); 1196 if (advice == MADV_WILLNEED) { 1197 /* 1198 * Reference the page before unlocking and 1199 * sleeping so that the page daemon is less 1200 * likely to reclaim it. 1201 */ 1202 vm_page_aflag_set(tm, PGA_REFERENCED); 1203 } 1204 vm_page_busy_sleep(tm, "madvpo", false); 1205 goto relookup; 1206 } 1207 vm_page_advise(tm, advice); 1208 vm_page_unlock(tm); 1209 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1210 next_pindex: 1211 if (tobject != object) 1212 VM_OBJECT_WUNLOCK(tobject); 1213 } 1214 VM_OBJECT_WUNLOCK(object); 1215 } 1216 1217 /* 1218 * vm_object_shadow: 1219 * 1220 * Create a new object which is backed by the 1221 * specified existing object range. The source 1222 * object reference is deallocated. 1223 * 1224 * The new object and offset into that object 1225 * are returned in the source parameters. 1226 */ 1227 void 1228 vm_object_shadow( 1229 vm_object_t *object, /* IN/OUT */ 1230 vm_ooffset_t *offset, /* IN/OUT */ 1231 vm_size_t length) 1232 { 1233 vm_object_t source; 1234 vm_object_t result; 1235 1236 source = *object; 1237 1238 /* 1239 * Don't create the new object if the old object isn't shared. 1240 */ 1241 if (source != NULL) { 1242 VM_OBJECT_WLOCK(source); 1243 if (source->ref_count == 1 && 1244 source->handle == NULL && 1245 (source->type == OBJT_DEFAULT || 1246 source->type == OBJT_SWAP)) { 1247 VM_OBJECT_WUNLOCK(source); 1248 return; 1249 } 1250 VM_OBJECT_WUNLOCK(source); 1251 } 1252 1253 /* 1254 * Allocate a new object with the given length. 1255 */ 1256 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1257 1258 /* 1259 * The new object shadows the source object, adding a reference to it. 1260 * Our caller changes his reference to point to the new object, 1261 * removing a reference to the source object. Net result: no change 1262 * of reference count. 1263 * 1264 * Try to optimize the result object's page color when shadowing 1265 * in order to maintain page coloring consistency in the combined 1266 * shadowed object. 1267 */ 1268 result->backing_object = source; 1269 /* 1270 * Store the offset into the source object, and fix up the offset into 1271 * the new object. 1272 */ 1273 result->backing_object_offset = *offset; 1274 if (source != NULL) { 1275 VM_OBJECT_WLOCK(source); 1276 result->domain = source->domain; 1277 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1278 source->shadow_count++; 1279 #if VM_NRESERVLEVEL > 0 1280 result->flags |= source->flags & OBJ_COLORED; 1281 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1282 ((1 << (VM_NFREEORDER - 1)) - 1); 1283 #endif 1284 VM_OBJECT_WUNLOCK(source); 1285 } 1286 1287 1288 /* 1289 * Return the new things 1290 */ 1291 *offset = 0; 1292 *object = result; 1293 } 1294 1295 /* 1296 * vm_object_split: 1297 * 1298 * Split the pages in a map entry into a new object. This affords 1299 * easier removal of unused pages, and keeps object inheritance from 1300 * being a negative impact on memory usage. 1301 */ 1302 void 1303 vm_object_split(vm_map_entry_t entry) 1304 { 1305 vm_page_t m, m_next; 1306 vm_object_t orig_object, new_object, source; 1307 vm_pindex_t idx, offidxstart; 1308 vm_size_t size; 1309 1310 orig_object = entry->object.vm_object; 1311 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1312 return; 1313 if (orig_object->ref_count <= 1) 1314 return; 1315 VM_OBJECT_WUNLOCK(orig_object); 1316 1317 offidxstart = OFF_TO_IDX(entry->offset); 1318 size = atop(entry->end - entry->start); 1319 1320 /* 1321 * If swap_pager_copy() is later called, it will convert new_object 1322 * into a swap object. 1323 */ 1324 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1325 1326 /* 1327 * At this point, the new object is still private, so the order in 1328 * which the original and new objects are locked does not matter. 1329 */ 1330 VM_OBJECT_WLOCK(new_object); 1331 VM_OBJECT_WLOCK(orig_object); 1332 new_object->domain = orig_object->domain; 1333 source = orig_object->backing_object; 1334 if (source != NULL) { 1335 VM_OBJECT_WLOCK(source); 1336 if ((source->flags & OBJ_DEAD) != 0) { 1337 VM_OBJECT_WUNLOCK(source); 1338 VM_OBJECT_WUNLOCK(orig_object); 1339 VM_OBJECT_WUNLOCK(new_object); 1340 vm_object_deallocate(new_object); 1341 VM_OBJECT_WLOCK(orig_object); 1342 return; 1343 } 1344 LIST_INSERT_HEAD(&source->shadow_head, 1345 new_object, shadow_list); 1346 source->shadow_count++; 1347 vm_object_reference_locked(source); /* for new_object */ 1348 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1349 VM_OBJECT_WUNLOCK(source); 1350 new_object->backing_object_offset = 1351 orig_object->backing_object_offset + entry->offset; 1352 new_object->backing_object = source; 1353 } 1354 if (orig_object->cred != NULL) { 1355 new_object->cred = orig_object->cred; 1356 crhold(orig_object->cred); 1357 new_object->charge = ptoa(size); 1358 KASSERT(orig_object->charge >= ptoa(size), 1359 ("orig_object->charge < 0")); 1360 orig_object->charge -= ptoa(size); 1361 } 1362 retry: 1363 m = vm_page_find_least(orig_object, offidxstart); 1364 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1365 m = m_next) { 1366 m_next = TAILQ_NEXT(m, listq); 1367 1368 /* 1369 * We must wait for pending I/O to complete before we can 1370 * rename the page. 1371 * 1372 * We do not have to VM_PROT_NONE the page as mappings should 1373 * not be changed by this operation. 1374 */ 1375 if (vm_page_busied(m)) { 1376 VM_OBJECT_WUNLOCK(new_object); 1377 vm_page_lock(m); 1378 VM_OBJECT_WUNLOCK(orig_object); 1379 vm_page_busy_sleep(m, "spltwt", false); 1380 VM_OBJECT_WLOCK(orig_object); 1381 VM_OBJECT_WLOCK(new_object); 1382 goto retry; 1383 } 1384 1385 /* vm_page_rename() will dirty the page. */ 1386 if (vm_page_rename(m, new_object, idx)) { 1387 VM_OBJECT_WUNLOCK(new_object); 1388 VM_OBJECT_WUNLOCK(orig_object); 1389 vm_radix_wait(); 1390 VM_OBJECT_WLOCK(orig_object); 1391 VM_OBJECT_WLOCK(new_object); 1392 goto retry; 1393 } 1394 #if VM_NRESERVLEVEL > 0 1395 /* 1396 * If some of the reservation's allocated pages remain with 1397 * the original object, then transferring the reservation to 1398 * the new object is neither particularly beneficial nor 1399 * particularly harmful as compared to leaving the reservation 1400 * with the original object. If, however, all of the 1401 * reservation's allocated pages are transferred to the new 1402 * object, then transferring the reservation is typically 1403 * beneficial. Determining which of these two cases applies 1404 * would be more costly than unconditionally renaming the 1405 * reservation. 1406 */ 1407 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1408 #endif 1409 if (orig_object->type == OBJT_SWAP) 1410 vm_page_xbusy(m); 1411 } 1412 if (orig_object->type == OBJT_SWAP) { 1413 /* 1414 * swap_pager_copy() can sleep, in which case the orig_object's 1415 * and new_object's locks are released and reacquired. 1416 */ 1417 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1418 TAILQ_FOREACH(m, &new_object->memq, listq) 1419 vm_page_xunbusy(m); 1420 } 1421 VM_OBJECT_WUNLOCK(orig_object); 1422 VM_OBJECT_WUNLOCK(new_object); 1423 entry->object.vm_object = new_object; 1424 entry->offset = 0LL; 1425 vm_object_deallocate(orig_object); 1426 VM_OBJECT_WLOCK(new_object); 1427 } 1428 1429 #define OBSC_COLLAPSE_NOWAIT 0x0002 1430 #define OBSC_COLLAPSE_WAIT 0x0004 1431 1432 static vm_page_t 1433 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1434 int op) 1435 { 1436 vm_object_t backing_object; 1437 1438 VM_OBJECT_ASSERT_WLOCKED(object); 1439 backing_object = object->backing_object; 1440 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1441 1442 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1443 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1444 ("invalid ownership %p %p %p", p, object, backing_object)); 1445 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1446 return (next); 1447 if (p != NULL) 1448 vm_page_lock(p); 1449 VM_OBJECT_WUNLOCK(object); 1450 VM_OBJECT_WUNLOCK(backing_object); 1451 /* The page is only NULL when rename fails. */ 1452 if (p == NULL) 1453 vm_radix_wait(); 1454 else 1455 vm_page_busy_sleep(p, "vmocol", false); 1456 VM_OBJECT_WLOCK(object); 1457 VM_OBJECT_WLOCK(backing_object); 1458 return (TAILQ_FIRST(&backing_object->memq)); 1459 } 1460 1461 static bool 1462 vm_object_scan_all_shadowed(vm_object_t object) 1463 { 1464 vm_object_t backing_object; 1465 vm_page_t p, pp; 1466 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1467 1468 VM_OBJECT_ASSERT_WLOCKED(object); 1469 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1470 1471 backing_object = object->backing_object; 1472 1473 if (backing_object->type != OBJT_DEFAULT && 1474 backing_object->type != OBJT_SWAP) 1475 return (false); 1476 1477 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1478 p = vm_page_find_least(backing_object, pi); 1479 ps = swap_pager_find_least(backing_object, pi); 1480 1481 /* 1482 * Only check pages inside the parent object's range and 1483 * inside the parent object's mapping of the backing object. 1484 */ 1485 for (;; pi++) { 1486 if (p != NULL && p->pindex < pi) 1487 p = TAILQ_NEXT(p, listq); 1488 if (ps < pi) 1489 ps = swap_pager_find_least(backing_object, pi); 1490 if (p == NULL && ps >= backing_object->size) 1491 break; 1492 else if (p == NULL) 1493 pi = ps; 1494 else 1495 pi = MIN(p->pindex, ps); 1496 1497 new_pindex = pi - backing_offset_index; 1498 if (new_pindex >= object->size) 1499 break; 1500 1501 /* 1502 * See if the parent has the page or if the parent's object 1503 * pager has the page. If the parent has the page but the page 1504 * is not valid, the parent's object pager must have the page. 1505 * 1506 * If this fails, the parent does not completely shadow the 1507 * object and we might as well give up now. 1508 */ 1509 pp = vm_page_lookup(object, new_pindex); 1510 if ((pp == NULL || pp->valid == 0) && 1511 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1512 return (false); 1513 } 1514 return (true); 1515 } 1516 1517 static bool 1518 vm_object_collapse_scan(vm_object_t object, int op) 1519 { 1520 vm_object_t backing_object; 1521 vm_page_t next, p, pp; 1522 vm_pindex_t backing_offset_index, new_pindex; 1523 1524 VM_OBJECT_ASSERT_WLOCKED(object); 1525 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1526 1527 backing_object = object->backing_object; 1528 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1529 1530 /* 1531 * Initial conditions 1532 */ 1533 if ((op & OBSC_COLLAPSE_WAIT) != 0) 1534 vm_object_set_flag(backing_object, OBJ_DEAD); 1535 1536 /* 1537 * Our scan 1538 */ 1539 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1540 next = TAILQ_NEXT(p, listq); 1541 new_pindex = p->pindex - backing_offset_index; 1542 1543 /* 1544 * Check for busy page 1545 */ 1546 if (vm_page_busied(p)) { 1547 next = vm_object_collapse_scan_wait(object, p, next, op); 1548 continue; 1549 } 1550 1551 KASSERT(p->object == backing_object, 1552 ("vm_object_collapse_scan: object mismatch")); 1553 1554 if (p->pindex < backing_offset_index || 1555 new_pindex >= object->size) { 1556 if (backing_object->type == OBJT_SWAP) 1557 swap_pager_freespace(backing_object, p->pindex, 1558 1); 1559 1560 /* 1561 * Page is out of the parent object's range, we can 1562 * simply destroy it. 1563 */ 1564 vm_page_lock(p); 1565 KASSERT(!pmap_page_is_mapped(p), 1566 ("freeing mapped page %p", p)); 1567 if (vm_page_remove(p)) 1568 vm_page_free(p); 1569 vm_page_unlock(p); 1570 continue; 1571 } 1572 1573 pp = vm_page_lookup(object, new_pindex); 1574 if (pp != NULL && vm_page_busied(pp)) { 1575 /* 1576 * The page in the parent is busy and possibly not 1577 * (yet) valid. Until its state is finalized by the 1578 * busy bit owner, we can't tell whether it shadows the 1579 * original page. Therefore, we must either skip it 1580 * and the original (backing_object) page or wait for 1581 * its state to be finalized. 1582 * 1583 * This is due to a race with vm_fault() where we must 1584 * unbusy the original (backing_obj) page before we can 1585 * (re)lock the parent. Hence we can get here. 1586 */ 1587 next = vm_object_collapse_scan_wait(object, pp, next, 1588 op); 1589 continue; 1590 } 1591 1592 KASSERT(pp == NULL || pp->valid != 0, 1593 ("unbusy invalid page %p", pp)); 1594 1595 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1596 NULL)) { 1597 /* 1598 * The page already exists in the parent OR swap exists 1599 * for this location in the parent. Leave the parent's 1600 * page alone. Destroy the original page from the 1601 * backing object. 1602 */ 1603 if (backing_object->type == OBJT_SWAP) 1604 swap_pager_freespace(backing_object, p->pindex, 1605 1); 1606 vm_page_lock(p); 1607 KASSERT(!pmap_page_is_mapped(p), 1608 ("freeing mapped page %p", p)); 1609 if (vm_page_remove(p)) 1610 vm_page_free(p); 1611 vm_page_unlock(p); 1612 continue; 1613 } 1614 1615 /* 1616 * Page does not exist in parent, rename the page from the 1617 * backing object to the main object. 1618 * 1619 * If the page was mapped to a process, it can remain mapped 1620 * through the rename. vm_page_rename() will dirty the page. 1621 */ 1622 if (vm_page_rename(p, object, new_pindex)) { 1623 next = vm_object_collapse_scan_wait(object, NULL, next, 1624 op); 1625 continue; 1626 } 1627 1628 /* Use the old pindex to free the right page. */ 1629 if (backing_object->type == OBJT_SWAP) 1630 swap_pager_freespace(backing_object, 1631 new_pindex + backing_offset_index, 1); 1632 1633 #if VM_NRESERVLEVEL > 0 1634 /* 1635 * Rename the reservation. 1636 */ 1637 vm_reserv_rename(p, object, backing_object, 1638 backing_offset_index); 1639 #endif 1640 } 1641 return (true); 1642 } 1643 1644 1645 /* 1646 * this version of collapse allows the operation to occur earlier and 1647 * when paging_in_progress is true for an object... This is not a complete 1648 * operation, but should plug 99.9% of the rest of the leaks. 1649 */ 1650 static void 1651 vm_object_qcollapse(vm_object_t object) 1652 { 1653 vm_object_t backing_object = object->backing_object; 1654 1655 VM_OBJECT_ASSERT_WLOCKED(object); 1656 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1657 1658 if (backing_object->ref_count != 1) 1659 return; 1660 1661 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); 1662 } 1663 1664 /* 1665 * vm_object_collapse: 1666 * 1667 * Collapse an object with the object backing it. 1668 * Pages in the backing object are moved into the 1669 * parent, and the backing object is deallocated. 1670 */ 1671 void 1672 vm_object_collapse(vm_object_t object) 1673 { 1674 vm_object_t backing_object, new_backing_object; 1675 1676 VM_OBJECT_ASSERT_WLOCKED(object); 1677 1678 while (TRUE) { 1679 /* 1680 * Verify that the conditions are right for collapse: 1681 * 1682 * The object exists and the backing object exists. 1683 */ 1684 if ((backing_object = object->backing_object) == NULL) 1685 break; 1686 1687 /* 1688 * we check the backing object first, because it is most likely 1689 * not collapsable. 1690 */ 1691 VM_OBJECT_WLOCK(backing_object); 1692 if (backing_object->handle != NULL || 1693 (backing_object->type != OBJT_DEFAULT && 1694 backing_object->type != OBJT_SWAP) || 1695 (backing_object->flags & (OBJ_DEAD | OBJ_NOSPLIT)) != 0 || 1696 object->handle != NULL || 1697 (object->type != OBJT_DEFAULT && 1698 object->type != OBJT_SWAP) || 1699 (object->flags & OBJ_DEAD)) { 1700 VM_OBJECT_WUNLOCK(backing_object); 1701 break; 1702 } 1703 1704 if (object->paging_in_progress != 0 || 1705 backing_object->paging_in_progress != 0) { 1706 vm_object_qcollapse(object); 1707 VM_OBJECT_WUNLOCK(backing_object); 1708 break; 1709 } 1710 1711 /* 1712 * We know that we can either collapse the backing object (if 1713 * the parent is the only reference to it) or (perhaps) have 1714 * the parent bypass the object if the parent happens to shadow 1715 * all the resident pages in the entire backing object. 1716 * 1717 * This is ignoring pager-backed pages such as swap pages. 1718 * vm_object_collapse_scan fails the shadowing test in this 1719 * case. 1720 */ 1721 if (backing_object->ref_count == 1) { 1722 vm_object_pip_add(object, 1); 1723 vm_object_pip_add(backing_object, 1); 1724 1725 /* 1726 * If there is exactly one reference to the backing 1727 * object, we can collapse it into the parent. 1728 */ 1729 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); 1730 1731 #if VM_NRESERVLEVEL > 0 1732 /* 1733 * Break any reservations from backing_object. 1734 */ 1735 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1736 vm_reserv_break_all(backing_object); 1737 #endif 1738 1739 /* 1740 * Move the pager from backing_object to object. 1741 */ 1742 if (backing_object->type == OBJT_SWAP) { 1743 /* 1744 * swap_pager_copy() can sleep, in which case 1745 * the backing_object's and object's locks are 1746 * released and reacquired. 1747 * Since swap_pager_copy() is being asked to 1748 * destroy the source, it will change the 1749 * backing_object's type to OBJT_DEFAULT. 1750 */ 1751 swap_pager_copy( 1752 backing_object, 1753 object, 1754 OFF_TO_IDX(object->backing_object_offset), TRUE); 1755 } 1756 /* 1757 * Object now shadows whatever backing_object did. 1758 * Note that the reference to 1759 * backing_object->backing_object moves from within 1760 * backing_object to within object. 1761 */ 1762 LIST_REMOVE(object, shadow_list); 1763 backing_object->shadow_count--; 1764 if (backing_object->backing_object) { 1765 VM_OBJECT_WLOCK(backing_object->backing_object); 1766 LIST_REMOVE(backing_object, shadow_list); 1767 LIST_INSERT_HEAD( 1768 &backing_object->backing_object->shadow_head, 1769 object, shadow_list); 1770 /* 1771 * The shadow_count has not changed. 1772 */ 1773 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1774 } 1775 object->backing_object = backing_object->backing_object; 1776 object->backing_object_offset += 1777 backing_object->backing_object_offset; 1778 1779 /* 1780 * Discard backing_object. 1781 * 1782 * Since the backing object has no pages, no pager left, 1783 * and no object references within it, all that is 1784 * necessary is to dispose of it. 1785 */ 1786 KASSERT(backing_object->ref_count == 1, ( 1787 "backing_object %p was somehow re-referenced during collapse!", 1788 backing_object)); 1789 vm_object_pip_wakeup(backing_object); 1790 backing_object->type = OBJT_DEAD; 1791 backing_object->ref_count = 0; 1792 VM_OBJECT_WUNLOCK(backing_object); 1793 vm_object_destroy(backing_object); 1794 1795 vm_object_pip_wakeup(object); 1796 counter_u64_add(object_collapses, 1); 1797 } else { 1798 /* 1799 * If we do not entirely shadow the backing object, 1800 * there is nothing we can do so we give up. 1801 */ 1802 if (object->resident_page_count != object->size && 1803 !vm_object_scan_all_shadowed(object)) { 1804 VM_OBJECT_WUNLOCK(backing_object); 1805 break; 1806 } 1807 1808 /* 1809 * Make the parent shadow the next object in the 1810 * chain. Deallocating backing_object will not remove 1811 * it, since its reference count is at least 2. 1812 */ 1813 LIST_REMOVE(object, shadow_list); 1814 backing_object->shadow_count--; 1815 1816 new_backing_object = backing_object->backing_object; 1817 if ((object->backing_object = new_backing_object) != NULL) { 1818 VM_OBJECT_WLOCK(new_backing_object); 1819 LIST_INSERT_HEAD( 1820 &new_backing_object->shadow_head, 1821 object, 1822 shadow_list 1823 ); 1824 new_backing_object->shadow_count++; 1825 vm_object_reference_locked(new_backing_object); 1826 VM_OBJECT_WUNLOCK(new_backing_object); 1827 object->backing_object_offset += 1828 backing_object->backing_object_offset; 1829 } 1830 1831 /* 1832 * Drop the reference count on backing_object. Since 1833 * its ref_count was at least 2, it will not vanish. 1834 */ 1835 backing_object->ref_count--; 1836 VM_OBJECT_WUNLOCK(backing_object); 1837 counter_u64_add(object_bypasses, 1); 1838 } 1839 1840 /* 1841 * Try again with this object's new backing object. 1842 */ 1843 } 1844 } 1845 1846 /* 1847 * vm_object_page_remove: 1848 * 1849 * For the given object, either frees or invalidates each of the 1850 * specified pages. In general, a page is freed. However, if a page is 1851 * wired for any reason other than the existence of a managed, wired 1852 * mapping, then it may be invalidated but not removed from the object. 1853 * Pages are specified by the given range ["start", "end") and the option 1854 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1855 * extends from "start" to the end of the object. If the option 1856 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1857 * specified range are affected. If the option OBJPR_NOTMAPPED is 1858 * specified, then the pages within the specified range must have no 1859 * mappings. Otherwise, if this option is not specified, any mappings to 1860 * the specified pages are removed before the pages are freed or 1861 * invalidated. 1862 * 1863 * In general, this operation should only be performed on objects that 1864 * contain managed pages. There are, however, two exceptions. First, it 1865 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1866 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1867 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1868 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1869 * 1870 * The object must be locked. 1871 */ 1872 void 1873 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1874 int options) 1875 { 1876 vm_page_t p, next; 1877 struct mtx *mtx; 1878 1879 VM_OBJECT_ASSERT_WLOCKED(object); 1880 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1881 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1882 ("vm_object_page_remove: illegal options for object %p", object)); 1883 if (object->resident_page_count == 0) 1884 return; 1885 vm_object_pip_add(object, 1); 1886 again: 1887 p = vm_page_find_least(object, start); 1888 mtx = NULL; 1889 1890 /* 1891 * Here, the variable "p" is either (1) the page with the least pindex 1892 * greater than or equal to the parameter "start" or (2) NULL. 1893 */ 1894 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1895 next = TAILQ_NEXT(p, listq); 1896 1897 /* 1898 * If the page is wired for any reason besides the existence 1899 * of managed, wired mappings, then it cannot be freed. For 1900 * example, fictitious pages, which represent device memory, 1901 * are inherently wired and cannot be freed. They can, 1902 * however, be invalidated if the option OBJPR_CLEANONLY is 1903 * not specified. 1904 */ 1905 vm_page_change_lock(p, &mtx); 1906 if (vm_page_xbusied(p)) { 1907 VM_OBJECT_WUNLOCK(object); 1908 vm_page_busy_sleep(p, "vmopax", true); 1909 VM_OBJECT_WLOCK(object); 1910 goto again; 1911 } 1912 if (vm_page_wired(p)) { 1913 if ((options & OBJPR_NOTMAPPED) == 0 && 1914 object->ref_count != 0) 1915 pmap_remove_all(p); 1916 if ((options & OBJPR_CLEANONLY) == 0) { 1917 p->valid = 0; 1918 vm_page_undirty(p); 1919 } 1920 continue; 1921 } 1922 if (vm_page_busied(p)) { 1923 VM_OBJECT_WUNLOCK(object); 1924 vm_page_busy_sleep(p, "vmopar", false); 1925 VM_OBJECT_WLOCK(object); 1926 goto again; 1927 } 1928 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1929 ("vm_object_page_remove: page %p is fictitious", p)); 1930 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1931 if ((options & OBJPR_NOTMAPPED) == 0 && 1932 object->ref_count != 0) 1933 pmap_remove_write(p); 1934 if (p->dirty != 0) 1935 continue; 1936 } 1937 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) 1938 pmap_remove_all(p); 1939 vm_page_free(p); 1940 } 1941 if (mtx != NULL) 1942 mtx_unlock(mtx); 1943 vm_object_pip_wakeup(object); 1944 } 1945 1946 /* 1947 * vm_object_page_noreuse: 1948 * 1949 * For the given object, attempt to move the specified pages to 1950 * the head of the inactive queue. This bypasses regular LRU 1951 * operation and allows the pages to be reused quickly under memory 1952 * pressure. If a page is wired for any reason, then it will not 1953 * be queued. Pages are specified by the range ["start", "end"). 1954 * As a special case, if "end" is zero, then the range extends from 1955 * "start" to the end of the object. 1956 * 1957 * This operation should only be performed on objects that 1958 * contain non-fictitious, managed pages. 1959 * 1960 * The object must be locked. 1961 */ 1962 void 1963 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1964 { 1965 struct mtx *mtx; 1966 vm_page_t p, next; 1967 1968 VM_OBJECT_ASSERT_LOCKED(object); 1969 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 1970 ("vm_object_page_noreuse: illegal object %p", object)); 1971 if (object->resident_page_count == 0) 1972 return; 1973 p = vm_page_find_least(object, start); 1974 1975 /* 1976 * Here, the variable "p" is either (1) the page with the least pindex 1977 * greater than or equal to the parameter "start" or (2) NULL. 1978 */ 1979 mtx = NULL; 1980 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1981 next = TAILQ_NEXT(p, listq); 1982 vm_page_change_lock(p, &mtx); 1983 vm_page_deactivate_noreuse(p); 1984 } 1985 if (mtx != NULL) 1986 mtx_unlock(mtx); 1987 } 1988 1989 /* 1990 * Populate the specified range of the object with valid pages. Returns 1991 * TRUE if the range is successfully populated and FALSE otherwise. 1992 * 1993 * Note: This function should be optimized to pass a larger array of 1994 * pages to vm_pager_get_pages() before it is applied to a non- 1995 * OBJT_DEVICE object. 1996 * 1997 * The object must be locked. 1998 */ 1999 boolean_t 2000 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2001 { 2002 vm_page_t m; 2003 vm_pindex_t pindex; 2004 int rv; 2005 2006 VM_OBJECT_ASSERT_WLOCKED(object); 2007 for (pindex = start; pindex < end; pindex++) { 2008 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2009 if (m->valid != VM_PAGE_BITS_ALL) { 2010 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 2011 if (rv != VM_PAGER_OK) { 2012 vm_page_lock(m); 2013 vm_page_free(m); 2014 vm_page_unlock(m); 2015 break; 2016 } 2017 } 2018 /* 2019 * Keep "m" busy because a subsequent iteration may unlock 2020 * the object. 2021 */ 2022 } 2023 if (pindex > start) { 2024 m = vm_page_lookup(object, start); 2025 while (m != NULL && m->pindex < pindex) { 2026 vm_page_xunbusy(m); 2027 m = TAILQ_NEXT(m, listq); 2028 } 2029 } 2030 return (pindex == end); 2031 } 2032 2033 /* 2034 * Routine: vm_object_coalesce 2035 * Function: Coalesces two objects backing up adjoining 2036 * regions of memory into a single object. 2037 * 2038 * returns TRUE if objects were combined. 2039 * 2040 * NOTE: Only works at the moment if the second object is NULL - 2041 * if it's not, which object do we lock first? 2042 * 2043 * Parameters: 2044 * prev_object First object to coalesce 2045 * prev_offset Offset into prev_object 2046 * prev_size Size of reference to prev_object 2047 * next_size Size of reference to the second object 2048 * reserved Indicator that extension region has 2049 * swap accounted for 2050 * 2051 * Conditions: 2052 * The object must *not* be locked. 2053 */ 2054 boolean_t 2055 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2056 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2057 { 2058 vm_pindex_t next_pindex; 2059 2060 if (prev_object == NULL) 2061 return (TRUE); 2062 VM_OBJECT_WLOCK(prev_object); 2063 if ((prev_object->type != OBJT_DEFAULT && 2064 prev_object->type != OBJT_SWAP) || 2065 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2066 VM_OBJECT_WUNLOCK(prev_object); 2067 return (FALSE); 2068 } 2069 2070 /* 2071 * Try to collapse the object first 2072 */ 2073 vm_object_collapse(prev_object); 2074 2075 /* 2076 * Can't coalesce if: . more than one reference . paged out . shadows 2077 * another object . has a copy elsewhere (any of which mean that the 2078 * pages not mapped to prev_entry may be in use anyway) 2079 */ 2080 if (prev_object->backing_object != NULL) { 2081 VM_OBJECT_WUNLOCK(prev_object); 2082 return (FALSE); 2083 } 2084 2085 prev_size >>= PAGE_SHIFT; 2086 next_size >>= PAGE_SHIFT; 2087 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2088 2089 if (prev_object->ref_count > 1 && 2090 prev_object->size != next_pindex && 2091 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2092 VM_OBJECT_WUNLOCK(prev_object); 2093 return (FALSE); 2094 } 2095 2096 /* 2097 * Account for the charge. 2098 */ 2099 if (prev_object->cred != NULL) { 2100 2101 /* 2102 * If prev_object was charged, then this mapping, 2103 * although not charged now, may become writable 2104 * later. Non-NULL cred in the object would prevent 2105 * swap reservation during enabling of the write 2106 * access, so reserve swap now. Failed reservation 2107 * cause allocation of the separate object for the map 2108 * entry, and swap reservation for this entry is 2109 * managed in appropriate time. 2110 */ 2111 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2112 prev_object->cred)) { 2113 VM_OBJECT_WUNLOCK(prev_object); 2114 return (FALSE); 2115 } 2116 prev_object->charge += ptoa(next_size); 2117 } 2118 2119 /* 2120 * Remove any pages that may still be in the object from a previous 2121 * deallocation. 2122 */ 2123 if (next_pindex < prev_object->size) { 2124 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2125 next_size, 0); 2126 if (prev_object->type == OBJT_SWAP) 2127 swap_pager_freespace(prev_object, 2128 next_pindex, next_size); 2129 #if 0 2130 if (prev_object->cred != NULL) { 2131 KASSERT(prev_object->charge >= 2132 ptoa(prev_object->size - next_pindex), 2133 ("object %p overcharged 1 %jx %jx", prev_object, 2134 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2135 prev_object->charge -= ptoa(prev_object->size - 2136 next_pindex); 2137 } 2138 #endif 2139 } 2140 2141 /* 2142 * Extend the object if necessary. 2143 */ 2144 if (next_pindex + next_size > prev_object->size) 2145 prev_object->size = next_pindex + next_size; 2146 2147 VM_OBJECT_WUNLOCK(prev_object); 2148 return (TRUE); 2149 } 2150 2151 void 2152 vm_object_set_writeable_dirty(vm_object_t object) 2153 { 2154 2155 VM_OBJECT_ASSERT_WLOCKED(object); 2156 if (object->type != OBJT_VNODE) { 2157 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2158 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2159 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2160 } 2161 return; 2162 } 2163 object->generation++; 2164 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2165 return; 2166 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2167 } 2168 2169 /* 2170 * vm_object_unwire: 2171 * 2172 * For each page offset within the specified range of the given object, 2173 * find the highest-level page in the shadow chain and unwire it. A page 2174 * must exist at every page offset, and the highest-level page must be 2175 * wired. 2176 */ 2177 void 2178 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2179 uint8_t queue) 2180 { 2181 vm_object_t tobject, t1object; 2182 vm_page_t m, tm; 2183 vm_pindex_t end_pindex, pindex, tpindex; 2184 int depth, locked_depth; 2185 2186 KASSERT((offset & PAGE_MASK) == 0, 2187 ("vm_object_unwire: offset is not page aligned")); 2188 KASSERT((length & PAGE_MASK) == 0, 2189 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2190 /* The wired count of a fictitious page never changes. */ 2191 if ((object->flags & OBJ_FICTITIOUS) != 0) 2192 return; 2193 pindex = OFF_TO_IDX(offset); 2194 end_pindex = pindex + atop(length); 2195 again: 2196 locked_depth = 1; 2197 VM_OBJECT_RLOCK(object); 2198 m = vm_page_find_least(object, pindex); 2199 while (pindex < end_pindex) { 2200 if (m == NULL || pindex < m->pindex) { 2201 /* 2202 * The first object in the shadow chain doesn't 2203 * contain a page at the current index. Therefore, 2204 * the page must exist in a backing object. 2205 */ 2206 tobject = object; 2207 tpindex = pindex; 2208 depth = 0; 2209 do { 2210 tpindex += 2211 OFF_TO_IDX(tobject->backing_object_offset); 2212 tobject = tobject->backing_object; 2213 KASSERT(tobject != NULL, 2214 ("vm_object_unwire: missing page")); 2215 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2216 goto next_page; 2217 depth++; 2218 if (depth == locked_depth) { 2219 locked_depth++; 2220 VM_OBJECT_RLOCK(tobject); 2221 } 2222 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2223 NULL); 2224 } else { 2225 tm = m; 2226 m = TAILQ_NEXT(m, listq); 2227 } 2228 vm_page_lock(tm); 2229 if (vm_page_xbusied(tm)) { 2230 for (tobject = object; locked_depth >= 1; 2231 locked_depth--) { 2232 t1object = tobject->backing_object; 2233 VM_OBJECT_RUNLOCK(tobject); 2234 tobject = t1object; 2235 } 2236 vm_page_busy_sleep(tm, "unwbo", true); 2237 goto again; 2238 } 2239 vm_page_unwire(tm, queue); 2240 vm_page_unlock(tm); 2241 next_page: 2242 pindex++; 2243 } 2244 /* Release the accumulated object locks. */ 2245 for (tobject = object; locked_depth >= 1; locked_depth--) { 2246 t1object = tobject->backing_object; 2247 VM_OBJECT_RUNLOCK(tobject); 2248 tobject = t1object; 2249 } 2250 } 2251 2252 /* 2253 * Return the vnode for the given object, or NULL if none exists. 2254 * For tmpfs objects, the function may return NULL if there is 2255 * no vnode allocated at the time of the call. 2256 */ 2257 struct vnode * 2258 vm_object_vnode(vm_object_t object) 2259 { 2260 struct vnode *vp; 2261 2262 VM_OBJECT_ASSERT_LOCKED(object); 2263 if (object->type == OBJT_VNODE) { 2264 vp = object->handle; 2265 KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); 2266 } else if (object->type == OBJT_SWAP && 2267 (object->flags & OBJ_TMPFS) != 0) { 2268 vp = object->un_pager.swp.swp_tmpfs; 2269 KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); 2270 } else { 2271 vp = NULL; 2272 } 2273 return (vp); 2274 } 2275 2276 /* 2277 * Return the kvme type of the given object. 2278 * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. 2279 */ 2280 int 2281 vm_object_kvme_type(vm_object_t object, struct vnode **vpp) 2282 { 2283 2284 VM_OBJECT_ASSERT_LOCKED(object); 2285 if (vpp != NULL) 2286 *vpp = vm_object_vnode(object); 2287 switch (object->type) { 2288 case OBJT_DEFAULT: 2289 return (KVME_TYPE_DEFAULT); 2290 case OBJT_VNODE: 2291 return (KVME_TYPE_VNODE); 2292 case OBJT_SWAP: 2293 if ((object->flags & OBJ_TMPFS_NODE) != 0) 2294 return (KVME_TYPE_VNODE); 2295 return (KVME_TYPE_SWAP); 2296 case OBJT_DEVICE: 2297 return (KVME_TYPE_DEVICE); 2298 case OBJT_PHYS: 2299 return (KVME_TYPE_PHYS); 2300 case OBJT_DEAD: 2301 return (KVME_TYPE_DEAD); 2302 case OBJT_SG: 2303 return (KVME_TYPE_SG); 2304 case OBJT_MGTDEVICE: 2305 return (KVME_TYPE_MGTDEVICE); 2306 default: 2307 return (KVME_TYPE_UNKNOWN); 2308 } 2309 } 2310 2311 static int 2312 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2313 { 2314 struct kinfo_vmobject *kvo; 2315 char *fullpath, *freepath; 2316 struct vnode *vp; 2317 struct vattr va; 2318 vm_object_t obj; 2319 vm_page_t m; 2320 int count, error; 2321 2322 if (req->oldptr == NULL) { 2323 /* 2324 * If an old buffer has not been provided, generate an 2325 * estimate of the space needed for a subsequent call. 2326 */ 2327 mtx_lock(&vm_object_list_mtx); 2328 count = 0; 2329 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2330 if (obj->type == OBJT_DEAD) 2331 continue; 2332 count++; 2333 } 2334 mtx_unlock(&vm_object_list_mtx); 2335 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2336 count * 11 / 10)); 2337 } 2338 2339 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2340 error = 0; 2341 2342 /* 2343 * VM objects are type stable and are never removed from the 2344 * list once added. This allows us to safely read obj->object_list 2345 * after reacquiring the VM object lock. 2346 */ 2347 mtx_lock(&vm_object_list_mtx); 2348 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2349 if (obj->type == OBJT_DEAD) 2350 continue; 2351 VM_OBJECT_RLOCK(obj); 2352 if (obj->type == OBJT_DEAD) { 2353 VM_OBJECT_RUNLOCK(obj); 2354 continue; 2355 } 2356 mtx_unlock(&vm_object_list_mtx); 2357 kvo->kvo_size = ptoa(obj->size); 2358 kvo->kvo_resident = obj->resident_page_count; 2359 kvo->kvo_ref_count = obj->ref_count; 2360 kvo->kvo_shadow_count = obj->shadow_count; 2361 kvo->kvo_memattr = obj->memattr; 2362 kvo->kvo_active = 0; 2363 kvo->kvo_inactive = 0; 2364 TAILQ_FOREACH(m, &obj->memq, listq) { 2365 /* 2366 * A page may belong to the object but be 2367 * dequeued and set to PQ_NONE while the 2368 * object lock is not held. This makes the 2369 * reads of m->queue below racy, and we do not 2370 * count pages set to PQ_NONE. However, this 2371 * sysctl is only meant to give an 2372 * approximation of the system anyway. 2373 */ 2374 if (m->queue == PQ_ACTIVE) 2375 kvo->kvo_active++; 2376 else if (m->queue == PQ_INACTIVE) 2377 kvo->kvo_inactive++; 2378 } 2379 2380 kvo->kvo_vn_fileid = 0; 2381 kvo->kvo_vn_fsid = 0; 2382 kvo->kvo_vn_fsid_freebsd11 = 0; 2383 freepath = NULL; 2384 fullpath = ""; 2385 kvo->kvo_type = vm_object_kvme_type(obj, &vp); 2386 if (vp != NULL) 2387 vref(vp); 2388 VM_OBJECT_RUNLOCK(obj); 2389 if (vp != NULL) { 2390 vn_fullpath(curthread, vp, &fullpath, &freepath); 2391 vn_lock(vp, LK_SHARED | LK_RETRY); 2392 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2393 kvo->kvo_vn_fileid = va.va_fileid; 2394 kvo->kvo_vn_fsid = va.va_fsid; 2395 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2396 /* truncate */ 2397 } 2398 vput(vp); 2399 } 2400 2401 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2402 if (freepath != NULL) 2403 free(freepath, M_TEMP); 2404 2405 /* Pack record size down */ 2406 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2407 + strlen(kvo->kvo_path) + 1; 2408 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2409 sizeof(uint64_t)); 2410 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2411 mtx_lock(&vm_object_list_mtx); 2412 if (error) 2413 break; 2414 } 2415 mtx_unlock(&vm_object_list_mtx); 2416 free(kvo, M_TEMP); 2417 return (error); 2418 } 2419 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2420 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2421 "List of VM objects"); 2422 2423 #include "opt_ddb.h" 2424 #ifdef DDB 2425 #include <sys/kernel.h> 2426 2427 #include <sys/cons.h> 2428 2429 #include <ddb/ddb.h> 2430 2431 static int 2432 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2433 { 2434 vm_map_t tmpm; 2435 vm_map_entry_t tmpe; 2436 vm_object_t obj; 2437 int entcount; 2438 2439 if (map == 0) 2440 return 0; 2441 2442 if (entry == 0) { 2443 tmpe = map->header.next; 2444 entcount = map->nentries; 2445 while (entcount-- && (tmpe != &map->header)) { 2446 if (_vm_object_in_map(map, object, tmpe)) { 2447 return 1; 2448 } 2449 tmpe = tmpe->next; 2450 } 2451 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2452 tmpm = entry->object.sub_map; 2453 tmpe = tmpm->header.next; 2454 entcount = tmpm->nentries; 2455 while (entcount-- && tmpe != &tmpm->header) { 2456 if (_vm_object_in_map(tmpm, object, tmpe)) { 2457 return 1; 2458 } 2459 tmpe = tmpe->next; 2460 } 2461 } else if ((obj = entry->object.vm_object) != NULL) { 2462 for (; obj; obj = obj->backing_object) 2463 if (obj == object) { 2464 return 1; 2465 } 2466 } 2467 return 0; 2468 } 2469 2470 static int 2471 vm_object_in_map(vm_object_t object) 2472 { 2473 struct proc *p; 2474 2475 /* sx_slock(&allproc_lock); */ 2476 FOREACH_PROC_IN_SYSTEM(p) { 2477 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2478 continue; 2479 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2480 /* sx_sunlock(&allproc_lock); */ 2481 return 1; 2482 } 2483 } 2484 /* sx_sunlock(&allproc_lock); */ 2485 if (_vm_object_in_map(kernel_map, object, 0)) 2486 return 1; 2487 return 0; 2488 } 2489 2490 DB_SHOW_COMMAND(vmochk, vm_object_check) 2491 { 2492 vm_object_t object; 2493 2494 /* 2495 * make sure that internal objs are in a map somewhere 2496 * and none have zero ref counts. 2497 */ 2498 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2499 if (object->handle == NULL && 2500 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2501 if (object->ref_count == 0) { 2502 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2503 (long)object->size); 2504 } 2505 if (!vm_object_in_map(object)) { 2506 db_printf( 2507 "vmochk: internal obj is not in a map: " 2508 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2509 object->ref_count, (u_long)object->size, 2510 (u_long)object->size, 2511 (void *)object->backing_object); 2512 } 2513 } 2514 } 2515 } 2516 2517 /* 2518 * vm_object_print: [ debug ] 2519 */ 2520 DB_SHOW_COMMAND(object, vm_object_print_static) 2521 { 2522 /* XXX convert args. */ 2523 vm_object_t object = (vm_object_t)addr; 2524 boolean_t full = have_addr; 2525 2526 vm_page_t p; 2527 2528 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2529 #define count was_count 2530 2531 int count; 2532 2533 if (object == NULL) 2534 return; 2535 2536 db_iprintf( 2537 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2538 object, (int)object->type, (uintmax_t)object->size, 2539 object->resident_page_count, object->ref_count, object->flags, 2540 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2541 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2542 object->shadow_count, 2543 object->backing_object ? object->backing_object->ref_count : 0, 2544 object->backing_object, (uintmax_t)object->backing_object_offset); 2545 2546 if (!full) 2547 return; 2548 2549 db_indent += 2; 2550 count = 0; 2551 TAILQ_FOREACH(p, &object->memq, listq) { 2552 if (count == 0) 2553 db_iprintf("memory:="); 2554 else if (count == 6) { 2555 db_printf("\n"); 2556 db_iprintf(" ..."); 2557 count = 0; 2558 } else 2559 db_printf(","); 2560 count++; 2561 2562 db_printf("(off=0x%jx,page=0x%jx)", 2563 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2564 } 2565 if (count != 0) 2566 db_printf("\n"); 2567 db_indent -= 2; 2568 } 2569 2570 /* XXX. */ 2571 #undef count 2572 2573 /* XXX need this non-static entry for calling from vm_map_print. */ 2574 void 2575 vm_object_print( 2576 /* db_expr_t */ long addr, 2577 boolean_t have_addr, 2578 /* db_expr_t */ long count, 2579 char *modif) 2580 { 2581 vm_object_print_static(addr, have_addr, count, modif); 2582 } 2583 2584 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2585 { 2586 vm_object_t object; 2587 vm_pindex_t fidx; 2588 vm_paddr_t pa; 2589 vm_page_t m, prev_m; 2590 int rcount, nl, c; 2591 2592 nl = 0; 2593 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2594 db_printf("new object: %p\n", (void *)object); 2595 if (nl > 18) { 2596 c = cngetc(); 2597 if (c != ' ') 2598 return; 2599 nl = 0; 2600 } 2601 nl++; 2602 rcount = 0; 2603 fidx = 0; 2604 pa = -1; 2605 TAILQ_FOREACH(m, &object->memq, listq) { 2606 if (m->pindex > 128) 2607 break; 2608 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2609 prev_m->pindex + 1 != m->pindex) { 2610 if (rcount) { 2611 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2612 (long)fidx, rcount, (long)pa); 2613 if (nl > 18) { 2614 c = cngetc(); 2615 if (c != ' ') 2616 return; 2617 nl = 0; 2618 } 2619 nl++; 2620 rcount = 0; 2621 } 2622 } 2623 if (rcount && 2624 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2625 ++rcount; 2626 continue; 2627 } 2628 if (rcount) { 2629 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2630 (long)fidx, rcount, (long)pa); 2631 if (nl > 18) { 2632 c = cngetc(); 2633 if (c != ' ') 2634 return; 2635 nl = 0; 2636 } 2637 nl++; 2638 } 2639 fidx = m->pindex; 2640 pa = VM_PAGE_TO_PHYS(m); 2641 rcount = 1; 2642 } 2643 if (rcount) { 2644 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2645 (long)fidx, rcount, (long)pa); 2646 if (nl > 18) { 2647 c = cngetc(); 2648 if (c != ' ') 2649 return; 2650 nl = 0; 2651 } 2652 nl++; 2653 } 2654 } 2655 } 2656 #endif /* DDB */ 2657