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/socket.h> 84 #include <sys/resourcevar.h> 85 #include <sys/rwlock.h> 86 #include <sys/user.h> 87 #include <sys/vnode.h> 88 #include <sys/vmmeter.h> 89 #include <sys/sx.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_param.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_object.h> 96 #include <vm/vm_page.h> 97 #include <vm/vm_pageout.h> 98 #include <vm/vm_pager.h> 99 #include <vm/vm_phys.h> 100 #include <vm/vm_pagequeue.h> 101 #include <vm/swap_pager.h> 102 #include <vm/vm_kern.h> 103 #include <vm/vm_extern.h> 104 #include <vm/vm_radix.h> 105 #include <vm/vm_reserv.h> 106 #include <vm/uma.h> 107 108 static int old_msync; 109 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 110 "Use old (insecure) msync behavior"); 111 112 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 113 int pagerflags, int flags, boolean_t *clearobjflags, 114 boolean_t *eio); 115 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 116 boolean_t *clearobjflags); 117 static void vm_object_qcollapse(vm_object_t object); 118 static void vm_object_vndeallocate(vm_object_t object); 119 120 /* 121 * Virtual memory objects maintain the actual data 122 * associated with allocated virtual memory. A given 123 * page of memory exists within exactly one object. 124 * 125 * An object is only deallocated when all "references" 126 * are given up. Only one "reference" to a given 127 * region of an object should be writeable. 128 * 129 * Associated with each object is a list of all resident 130 * memory pages belonging to that object; this list is 131 * maintained by the "vm_page" module, and locked by the object's 132 * lock. 133 * 134 * Each object also records a "pager" routine which is 135 * used to retrieve (and store) pages to the proper backing 136 * storage. In addition, objects may be backed by other 137 * objects from which they were virtual-copied. 138 * 139 * The only items within the object structure which are 140 * modified after time of creation are: 141 * reference count locked by object's lock 142 * pager routine locked by object's lock 143 * 144 */ 145 146 struct object_q vm_object_list; 147 struct mtx vm_object_list_mtx; /* lock for object list and count */ 148 149 struct vm_object kernel_object_store; 150 151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 152 "VM object stats"); 153 154 static counter_u64_t object_collapses = EARLY_COUNTER; 155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 156 &object_collapses, 157 "VM object collapses"); 158 159 static counter_u64_t object_bypasses = EARLY_COUNTER; 160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 161 &object_bypasses, 162 "VM object bypasses"); 163 164 static void 165 counter_startup(void) 166 { 167 168 object_collapses = counter_u64_alloc(M_WAITOK); 169 object_bypasses = counter_u64_alloc(M_WAITOK); 170 } 171 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); 172 173 static uma_zone_t obj_zone; 174 175 static int vm_object_zinit(void *mem, int size, int flags); 176 177 #ifdef INVARIANTS 178 static void vm_object_zdtor(void *mem, int size, void *arg); 179 180 static void 181 vm_object_zdtor(void *mem, int size, void *arg) 182 { 183 vm_object_t object; 184 185 object = (vm_object_t)mem; 186 KASSERT(object->ref_count == 0, 187 ("object %p ref_count = %d", object, object->ref_count)); 188 KASSERT(TAILQ_EMPTY(&object->memq), 189 ("object %p has resident pages in its memq", object)); 190 KASSERT(vm_radix_is_empty(&object->rtree), 191 ("object %p has resident pages in its trie", object)); 192 #if VM_NRESERVLEVEL > 0 193 KASSERT(LIST_EMPTY(&object->rvq), 194 ("object %p has reservations", 195 object)); 196 #endif 197 KASSERT(object->paging_in_progress == 0, 198 ("object %p paging_in_progress = %d", 199 object, object->paging_in_progress)); 200 KASSERT(object->resident_page_count == 0, 201 ("object %p resident_page_count = %d", 202 object, object->resident_page_count)); 203 KASSERT(object->shadow_count == 0, 204 ("object %p shadow_count = %d", 205 object, object->shadow_count)); 206 KASSERT(object->type == OBJT_DEAD, 207 ("object %p has non-dead type %d", 208 object, object->type)); 209 } 210 #endif 211 212 static int 213 vm_object_zinit(void *mem, int size, int flags) 214 { 215 vm_object_t object; 216 217 object = (vm_object_t)mem; 218 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 219 220 /* These are true for any object that has been freed */ 221 object->type = OBJT_DEAD; 222 object->ref_count = 0; 223 vm_radix_init(&object->rtree); 224 object->paging_in_progress = 0; 225 object->resident_page_count = 0; 226 object->shadow_count = 0; 227 object->flags = OBJ_DEAD; 228 229 mtx_lock(&vm_object_list_mtx); 230 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 231 mtx_unlock(&vm_object_list_mtx); 232 return (0); 233 } 234 235 static void 236 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 237 { 238 239 TAILQ_INIT(&object->memq); 240 LIST_INIT(&object->shadow_head); 241 242 object->type = type; 243 if (type == OBJT_SWAP) 244 pctrie_init(&object->un_pager.swp.swp_blks); 245 246 /* 247 * Ensure that swap_pager_swapoff() iteration over object_list 248 * sees up to date type and pctrie head if it observed 249 * non-dead object. 250 */ 251 atomic_thread_fence_rel(); 252 253 switch (type) { 254 case OBJT_DEAD: 255 panic("_vm_object_allocate: can't create OBJT_DEAD"); 256 case OBJT_DEFAULT: 257 case OBJT_SWAP: 258 object->flags = OBJ_ONEMAPPING; 259 break; 260 case OBJT_DEVICE: 261 case OBJT_SG: 262 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 263 break; 264 case OBJT_MGTDEVICE: 265 object->flags = OBJ_FICTITIOUS; 266 break; 267 case OBJT_PHYS: 268 object->flags = OBJ_UNMANAGED; 269 break; 270 case OBJT_VNODE: 271 object->flags = 0; 272 break; 273 default: 274 panic("_vm_object_allocate: type %d is undefined", type); 275 } 276 object->size = size; 277 object->domain.dr_policy = NULL; 278 object->generation = 1; 279 object->ref_count = 1; 280 object->memattr = VM_MEMATTR_DEFAULT; 281 object->cred = NULL; 282 object->charge = 0; 283 object->handle = NULL; 284 object->backing_object = NULL; 285 object->backing_object_offset = (vm_ooffset_t) 0; 286 #if VM_NRESERVLEVEL > 0 287 LIST_INIT(&object->rvq); 288 #endif 289 umtx_shm_object_init(object); 290 } 291 292 /* 293 * vm_object_init: 294 * 295 * Initialize the VM objects module. 296 */ 297 void 298 vm_object_init(void) 299 { 300 TAILQ_INIT(&vm_object_list); 301 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 302 303 rw_init(&kernel_object->lock, "kernel vm object"); 304 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 305 VM_MIN_KERNEL_ADDRESS), kernel_object); 306 #if VM_NRESERVLEVEL > 0 307 kernel_object->flags |= OBJ_COLORED; 308 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 309 #endif 310 311 /* 312 * The lock portion of struct vm_object must be type stable due 313 * to vm_pageout_fallback_object_lock locking a vm object 314 * without holding any references to it. 315 */ 316 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 317 #ifdef INVARIANTS 318 vm_object_zdtor, 319 #else 320 NULL, 321 #endif 322 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 323 324 vm_radix_zinit(); 325 } 326 327 void 328 vm_object_clear_flag(vm_object_t object, u_short bits) 329 { 330 331 VM_OBJECT_ASSERT_WLOCKED(object); 332 object->flags &= ~bits; 333 } 334 335 /* 336 * Sets the default memory attribute for the specified object. Pages 337 * that are allocated to this object are by default assigned this memory 338 * attribute. 339 * 340 * Presently, this function must be called before any pages are allocated 341 * to the object. In the future, this requirement may be relaxed for 342 * "default" and "swap" objects. 343 */ 344 int 345 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 346 { 347 348 VM_OBJECT_ASSERT_WLOCKED(object); 349 switch (object->type) { 350 case OBJT_DEFAULT: 351 case OBJT_DEVICE: 352 case OBJT_MGTDEVICE: 353 case OBJT_PHYS: 354 case OBJT_SG: 355 case OBJT_SWAP: 356 case OBJT_VNODE: 357 if (!TAILQ_EMPTY(&object->memq)) 358 return (KERN_FAILURE); 359 break; 360 case OBJT_DEAD: 361 return (KERN_INVALID_ARGUMENT); 362 default: 363 panic("vm_object_set_memattr: object %p is of undefined type", 364 object); 365 } 366 object->memattr = memattr; 367 return (KERN_SUCCESS); 368 } 369 370 void 371 vm_object_pip_add(vm_object_t object, short i) 372 { 373 374 VM_OBJECT_ASSERT_WLOCKED(object); 375 object->paging_in_progress += i; 376 } 377 378 void 379 vm_object_pip_subtract(vm_object_t object, short i) 380 { 381 382 VM_OBJECT_ASSERT_WLOCKED(object); 383 object->paging_in_progress -= i; 384 } 385 386 void 387 vm_object_pip_wakeup(vm_object_t object) 388 { 389 390 VM_OBJECT_ASSERT_WLOCKED(object); 391 object->paging_in_progress--; 392 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 393 vm_object_clear_flag(object, OBJ_PIPWNT); 394 wakeup(object); 395 } 396 } 397 398 void 399 vm_object_pip_wakeupn(vm_object_t object, short i) 400 { 401 402 VM_OBJECT_ASSERT_WLOCKED(object); 403 if (i) 404 object->paging_in_progress -= i; 405 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 406 vm_object_clear_flag(object, OBJ_PIPWNT); 407 wakeup(object); 408 } 409 } 410 411 void 412 vm_object_pip_wait(vm_object_t object, char *waitid) 413 { 414 415 VM_OBJECT_ASSERT_WLOCKED(object); 416 while (object->paging_in_progress) { 417 object->flags |= OBJ_PIPWNT; 418 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 419 } 420 } 421 422 /* 423 * vm_object_allocate: 424 * 425 * Returns a new object with the given size. 426 */ 427 vm_object_t 428 vm_object_allocate(objtype_t type, vm_pindex_t size) 429 { 430 vm_object_t object; 431 432 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 433 _vm_object_allocate(type, size, object); 434 return (object); 435 } 436 437 438 /* 439 * vm_object_reference: 440 * 441 * Gets another reference to the given object. Note: OBJ_DEAD 442 * objects can be referenced during final cleaning. 443 */ 444 void 445 vm_object_reference(vm_object_t object) 446 { 447 if (object == NULL) 448 return; 449 VM_OBJECT_WLOCK(object); 450 vm_object_reference_locked(object); 451 VM_OBJECT_WUNLOCK(object); 452 } 453 454 /* 455 * vm_object_reference_locked: 456 * 457 * Gets another reference to the given object. 458 * 459 * The object must be locked. 460 */ 461 void 462 vm_object_reference_locked(vm_object_t object) 463 { 464 struct vnode *vp; 465 466 VM_OBJECT_ASSERT_WLOCKED(object); 467 object->ref_count++; 468 if (object->type == OBJT_VNODE) { 469 vp = object->handle; 470 vref(vp); 471 } 472 } 473 474 /* 475 * Handle deallocating an object of type OBJT_VNODE. 476 */ 477 static void 478 vm_object_vndeallocate(vm_object_t object) 479 { 480 struct vnode *vp = (struct vnode *) object->handle; 481 482 VM_OBJECT_ASSERT_WLOCKED(object); 483 KASSERT(object->type == OBJT_VNODE, 484 ("vm_object_vndeallocate: not a vnode object")); 485 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 486 #ifdef INVARIANTS 487 if (object->ref_count == 0) { 488 vn_printf(vp, "vm_object_vndeallocate "); 489 panic("vm_object_vndeallocate: bad object reference count"); 490 } 491 #endif 492 493 if (!umtx_shm_vnobj_persistent && object->ref_count == 1) 494 umtx_shm_object_terminated(object); 495 496 object->ref_count--; 497 498 /* vrele may need the vnode lock. */ 499 VM_OBJECT_WUNLOCK(object); 500 vrele(vp); 501 } 502 503 /* 504 * vm_object_deallocate: 505 * 506 * Release a reference to the specified object, 507 * gained either through a vm_object_allocate 508 * or a vm_object_reference call. When all references 509 * are gone, storage associated with this object 510 * may be relinquished. 511 * 512 * No object may be locked. 513 */ 514 void 515 vm_object_deallocate(vm_object_t object) 516 { 517 vm_object_t temp; 518 struct vnode *vp; 519 520 while (object != NULL) { 521 VM_OBJECT_WLOCK(object); 522 if (object->type == OBJT_VNODE) { 523 vm_object_vndeallocate(object); 524 return; 525 } 526 527 KASSERT(object->ref_count != 0, 528 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 529 530 /* 531 * If the reference count goes to 0 we start calling 532 * vm_object_terminate() on the object chain. 533 * A ref count of 1 may be a special case depending on the 534 * shadow count being 0 or 1. 535 */ 536 object->ref_count--; 537 if (object->ref_count > 1) { 538 VM_OBJECT_WUNLOCK(object); 539 return; 540 } else if (object->ref_count == 1) { 541 if (object->type == OBJT_SWAP && 542 (object->flags & OBJ_TMPFS) != 0) { 543 vp = object->un_pager.swp.swp_tmpfs; 544 vhold(vp); 545 VM_OBJECT_WUNLOCK(object); 546 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 547 VM_OBJECT_WLOCK(object); 548 if (object->type == OBJT_DEAD || 549 object->ref_count != 1) { 550 VM_OBJECT_WUNLOCK(object); 551 VOP_UNLOCK(vp, 0); 552 vdrop(vp); 553 return; 554 } 555 if ((object->flags & OBJ_TMPFS) != 0) 556 VOP_UNSET_TEXT(vp); 557 VOP_UNLOCK(vp, 0); 558 vdrop(vp); 559 } 560 if (object->shadow_count == 0 && 561 object->handle == NULL && 562 (object->type == OBJT_DEFAULT || 563 (object->type == OBJT_SWAP && 564 (object->flags & OBJ_TMPFS_NODE) == 0))) { 565 vm_object_set_flag(object, OBJ_ONEMAPPING); 566 } else if ((object->shadow_count == 1) && 567 (object->handle == NULL) && 568 (object->type == OBJT_DEFAULT || 569 object->type == OBJT_SWAP)) { 570 vm_object_t robject; 571 572 robject = LIST_FIRST(&object->shadow_head); 573 KASSERT(robject != NULL, 574 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 575 object->ref_count, 576 object->shadow_count)); 577 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 578 ("shadowed tmpfs v_object %p", object)); 579 if (!VM_OBJECT_TRYWLOCK(robject)) { 580 /* 581 * Avoid a potential deadlock. 582 */ 583 object->ref_count++; 584 VM_OBJECT_WUNLOCK(object); 585 /* 586 * More likely than not the thread 587 * holding robject's lock has lower 588 * priority than the current thread. 589 * Let the lower priority thread run. 590 */ 591 pause("vmo_de", 1); 592 continue; 593 } 594 /* 595 * Collapse object into its shadow unless its 596 * shadow is dead. In that case, object will 597 * be deallocated by the thread that is 598 * deallocating its shadow. 599 */ 600 if ((robject->flags & OBJ_DEAD) == 0 && 601 (robject->handle == NULL) && 602 (robject->type == OBJT_DEFAULT || 603 robject->type == OBJT_SWAP)) { 604 605 robject->ref_count++; 606 retry: 607 if (robject->paging_in_progress) { 608 VM_OBJECT_WUNLOCK(object); 609 vm_object_pip_wait(robject, 610 "objde1"); 611 temp = robject->backing_object; 612 if (object == temp) { 613 VM_OBJECT_WLOCK(object); 614 goto retry; 615 } 616 } else if (object->paging_in_progress) { 617 VM_OBJECT_WUNLOCK(robject); 618 object->flags |= OBJ_PIPWNT; 619 VM_OBJECT_SLEEP(object, object, 620 PDROP | PVM, "objde2", 0); 621 VM_OBJECT_WLOCK(robject); 622 temp = robject->backing_object; 623 if (object == temp) { 624 VM_OBJECT_WLOCK(object); 625 goto retry; 626 } 627 } else 628 VM_OBJECT_WUNLOCK(object); 629 630 if (robject->ref_count == 1) { 631 robject->ref_count--; 632 object = robject; 633 goto doterm; 634 } 635 object = robject; 636 vm_object_collapse(object); 637 VM_OBJECT_WUNLOCK(object); 638 continue; 639 } 640 VM_OBJECT_WUNLOCK(robject); 641 } 642 VM_OBJECT_WUNLOCK(object); 643 return; 644 } 645 doterm: 646 umtx_shm_object_terminated(object); 647 temp = object->backing_object; 648 if (temp != NULL) { 649 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 650 ("shadowed tmpfs v_object 2 %p", object)); 651 VM_OBJECT_WLOCK(temp); 652 LIST_REMOVE(object, shadow_list); 653 temp->shadow_count--; 654 VM_OBJECT_WUNLOCK(temp); 655 object->backing_object = NULL; 656 } 657 /* 658 * Don't double-terminate, we could be in a termination 659 * recursion due to the terminate having to sync data 660 * to disk. 661 */ 662 if ((object->flags & OBJ_DEAD) == 0) 663 vm_object_terminate(object); 664 else 665 VM_OBJECT_WUNLOCK(object); 666 object = temp; 667 } 668 } 669 670 /* 671 * vm_object_destroy removes the object from the global object list 672 * and frees the space for the object. 673 */ 674 void 675 vm_object_destroy(vm_object_t object) 676 { 677 678 /* 679 * Release the allocation charge. 680 */ 681 if (object->cred != NULL) { 682 swap_release_by_cred(object->charge, object->cred); 683 object->charge = 0; 684 crfree(object->cred); 685 object->cred = NULL; 686 } 687 688 /* 689 * Free the space for the object. 690 */ 691 uma_zfree(obj_zone, object); 692 } 693 694 /* 695 * vm_object_terminate_pages removes any remaining pageable pages 696 * from the object and resets the object to an empty state. 697 */ 698 static void 699 vm_object_terminate_pages(vm_object_t object) 700 { 701 vm_page_t p, p_next; 702 struct mtx *mtx; 703 704 VM_OBJECT_ASSERT_WLOCKED(object); 705 706 mtx = NULL; 707 708 /* 709 * Free any remaining pageable pages. This also removes them from the 710 * paging queues. However, don't free wired pages, just remove them 711 * from the object. Rather than incrementally removing each page from 712 * the object, the page and object are reset to any empty state. 713 */ 714 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 715 vm_page_assert_unbusied(p); 716 if ((object->flags & OBJ_UNMANAGED) == 0) 717 /* 718 * vm_page_free_prep() only needs the page 719 * lock for managed pages. 720 */ 721 vm_page_change_lock(p, &mtx); 722 p->object = NULL; 723 if (vm_page_wired(p)) 724 continue; 725 VM_CNT_INC(v_pfree); 726 vm_page_free(p); 727 } 728 if (mtx != NULL) 729 mtx_unlock(mtx); 730 731 /* 732 * If the object contained any pages, then reset it to an empty state. 733 * None of the object's fields, including "resident_page_count", were 734 * modified by the preceding loop. 735 */ 736 if (object->resident_page_count != 0) { 737 vm_radix_reclaim_allnodes(&object->rtree); 738 TAILQ_INIT(&object->memq); 739 object->resident_page_count = 0; 740 if (object->type == OBJT_VNODE) 741 vdrop(object->handle); 742 } 743 } 744 745 /* 746 * vm_object_terminate actually destroys the specified object, freeing 747 * up all previously used resources. 748 * 749 * The object must be locked. 750 * This routine may block. 751 */ 752 void 753 vm_object_terminate(vm_object_t object) 754 { 755 756 VM_OBJECT_ASSERT_WLOCKED(object); 757 758 /* 759 * Make sure no one uses us. 760 */ 761 vm_object_set_flag(object, OBJ_DEAD); 762 763 /* 764 * wait for the pageout daemon to be done with the object 765 */ 766 vm_object_pip_wait(object, "objtrm"); 767 768 KASSERT(!object->paging_in_progress, 769 ("vm_object_terminate: pageout in progress")); 770 771 /* 772 * Clean and free the pages, as appropriate. All references to the 773 * object are gone, so we don't need to lock it. 774 */ 775 if (object->type == OBJT_VNODE) { 776 struct vnode *vp = (struct vnode *)object->handle; 777 778 /* 779 * Clean pages and flush buffers. 780 */ 781 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 782 VM_OBJECT_WUNLOCK(object); 783 784 vinvalbuf(vp, V_SAVE, 0, 0); 785 786 BO_LOCK(&vp->v_bufobj); 787 vp->v_bufobj.bo_flag |= BO_DEAD; 788 BO_UNLOCK(&vp->v_bufobj); 789 790 VM_OBJECT_WLOCK(object); 791 } 792 793 KASSERT(object->ref_count == 0, 794 ("vm_object_terminate: object with references, ref_count=%d", 795 object->ref_count)); 796 797 if ((object->flags & OBJ_PG_DTOR) == 0) 798 vm_object_terminate_pages(object); 799 800 #if VM_NRESERVLEVEL > 0 801 if (__predict_false(!LIST_EMPTY(&object->rvq))) 802 vm_reserv_break_all(object); 803 #endif 804 805 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 806 object->type == OBJT_SWAP, 807 ("%s: non-swap obj %p has cred", __func__, object)); 808 809 /* 810 * Let the pager know object is dead. 811 */ 812 vm_pager_deallocate(object); 813 VM_OBJECT_WUNLOCK(object); 814 815 vm_object_destroy(object); 816 } 817 818 /* 819 * Make the page read-only so that we can clear the object flags. However, if 820 * this is a nosync mmap then the object is likely to stay dirty so do not 821 * mess with the page and do not clear the object flags. Returns TRUE if the 822 * page should be flushed, and FALSE otherwise. 823 */ 824 static boolean_t 825 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 826 { 827 828 /* 829 * If we have been asked to skip nosync pages and this is a 830 * nosync page, skip it. Note that the object flags were not 831 * cleared in this case so we do not have to set them. 832 */ 833 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 834 *clearobjflags = FALSE; 835 return (FALSE); 836 } else { 837 pmap_remove_write(p); 838 return (p->dirty != 0); 839 } 840 } 841 842 /* 843 * vm_object_page_clean 844 * 845 * Clean all dirty pages in the specified range of object. Leaves page 846 * on whatever queue it is currently on. If NOSYNC is set then do not 847 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 848 * leaving the object dirty. 849 * 850 * When stuffing pages asynchronously, allow clustering. XXX we need a 851 * synchronous clustering mode implementation. 852 * 853 * Odd semantics: if start == end, we clean everything. 854 * 855 * The object must be locked. 856 * 857 * Returns FALSE if some page from the range was not written, as 858 * reported by the pager, and TRUE otherwise. 859 */ 860 boolean_t 861 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 862 int flags) 863 { 864 vm_page_t np, p; 865 vm_pindex_t pi, tend, tstart; 866 int curgeneration, n, pagerflags; 867 boolean_t clearobjflags, eio, res; 868 869 VM_OBJECT_ASSERT_WLOCKED(object); 870 871 /* 872 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 873 * objects. The check below prevents the function from 874 * operating on non-vnode objects. 875 */ 876 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 877 object->resident_page_count == 0) 878 return (TRUE); 879 880 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 881 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 882 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 883 884 tstart = OFF_TO_IDX(start); 885 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 886 clearobjflags = tstart == 0 && tend >= object->size; 887 res = TRUE; 888 889 rescan: 890 curgeneration = object->generation; 891 892 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 893 pi = p->pindex; 894 if (pi >= tend) 895 break; 896 np = TAILQ_NEXT(p, listq); 897 if (p->valid == 0) 898 continue; 899 if (vm_page_sleep_if_busy(p, "vpcwai")) { 900 if (object->generation != curgeneration) { 901 if ((flags & OBJPC_SYNC) != 0) 902 goto rescan; 903 else 904 clearobjflags = FALSE; 905 } 906 np = vm_page_find_least(object, pi); 907 continue; 908 } 909 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 910 continue; 911 912 n = vm_object_page_collect_flush(object, p, pagerflags, 913 flags, &clearobjflags, &eio); 914 if (eio) { 915 res = FALSE; 916 clearobjflags = FALSE; 917 } 918 if (object->generation != curgeneration) { 919 if ((flags & OBJPC_SYNC) != 0) 920 goto rescan; 921 else 922 clearobjflags = FALSE; 923 } 924 925 /* 926 * If the VOP_PUTPAGES() did a truncated write, so 927 * that even the first page of the run is not fully 928 * written, vm_pageout_flush() returns 0 as the run 929 * length. Since the condition that caused truncated 930 * write may be permanent, e.g. exhausted free space, 931 * accepting n == 0 would cause an infinite loop. 932 * 933 * Forwarding the iterator leaves the unwritten page 934 * behind, but there is not much we can do there if 935 * filesystem refuses to write it. 936 */ 937 if (n == 0) { 938 n = 1; 939 clearobjflags = FALSE; 940 } 941 np = vm_page_find_least(object, pi + n); 942 } 943 #if 0 944 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 945 #endif 946 947 if (clearobjflags) 948 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 949 return (res); 950 } 951 952 static int 953 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 954 int flags, boolean_t *clearobjflags, boolean_t *eio) 955 { 956 vm_page_t ma[vm_pageout_page_count], p_first, tp; 957 int count, i, mreq, runlen; 958 959 vm_page_lock_assert(p, MA_NOTOWNED); 960 VM_OBJECT_ASSERT_WLOCKED(object); 961 962 count = 1; 963 mreq = 0; 964 965 for (tp = p; count < vm_pageout_page_count; count++) { 966 tp = vm_page_next(tp); 967 if (tp == NULL || vm_page_busied(tp)) 968 break; 969 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 970 break; 971 } 972 973 for (p_first = p; count < vm_pageout_page_count; count++) { 974 tp = vm_page_prev(p_first); 975 if (tp == NULL || vm_page_busied(tp)) 976 break; 977 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 978 break; 979 p_first = tp; 980 mreq++; 981 } 982 983 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 984 ma[i] = tp; 985 986 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 987 return (runlen); 988 } 989 990 /* 991 * Note that there is absolutely no sense in writing out 992 * anonymous objects, so we track down the vnode object 993 * to write out. 994 * We invalidate (remove) all pages from the address space 995 * for semantic correctness. 996 * 997 * If the backing object is a device object with unmanaged pages, then any 998 * mappings to the specified range of pages must be removed before this 999 * function is called. 1000 * 1001 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1002 * may start out with a NULL object. 1003 */ 1004 boolean_t 1005 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1006 boolean_t syncio, boolean_t invalidate) 1007 { 1008 vm_object_t backing_object; 1009 struct vnode *vp; 1010 struct mount *mp; 1011 int error, flags, fsync_after; 1012 boolean_t res; 1013 1014 if (object == NULL) 1015 return (TRUE); 1016 res = TRUE; 1017 error = 0; 1018 VM_OBJECT_WLOCK(object); 1019 while ((backing_object = object->backing_object) != NULL) { 1020 VM_OBJECT_WLOCK(backing_object); 1021 offset += object->backing_object_offset; 1022 VM_OBJECT_WUNLOCK(object); 1023 object = backing_object; 1024 if (object->size < OFF_TO_IDX(offset + size)) 1025 size = IDX_TO_OFF(object->size) - offset; 1026 } 1027 /* 1028 * Flush pages if writing is allowed, invalidate them 1029 * if invalidation requested. Pages undergoing I/O 1030 * will be ignored by vm_object_page_remove(). 1031 * 1032 * We cannot lock the vnode and then wait for paging 1033 * to complete without deadlocking against vm_fault. 1034 * Instead we simply call vm_object_page_remove() and 1035 * allow it to block internally on a page-by-page 1036 * basis when it encounters pages undergoing async 1037 * I/O. 1038 */ 1039 if (object->type == OBJT_VNODE && 1040 (object->flags & OBJ_MIGHTBEDIRTY) != 0 && 1041 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1042 VM_OBJECT_WUNLOCK(object); 1043 (void) vn_start_write(vp, &mp, V_WAIT); 1044 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1045 if (syncio && !invalidate && offset == 0 && 1046 atop(size) == object->size) { 1047 /* 1048 * If syncing the whole mapping of the file, 1049 * it is faster to schedule all the writes in 1050 * async mode, also allowing the clustering, 1051 * and then wait for i/o to complete. 1052 */ 1053 flags = 0; 1054 fsync_after = TRUE; 1055 } else { 1056 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1057 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1058 fsync_after = FALSE; 1059 } 1060 VM_OBJECT_WLOCK(object); 1061 res = vm_object_page_clean(object, offset, offset + size, 1062 flags); 1063 VM_OBJECT_WUNLOCK(object); 1064 if (fsync_after) 1065 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1066 VOP_UNLOCK(vp, 0); 1067 vn_finished_write(mp); 1068 if (error != 0) 1069 res = FALSE; 1070 VM_OBJECT_WLOCK(object); 1071 } 1072 if ((object->type == OBJT_VNODE || 1073 object->type == OBJT_DEVICE) && invalidate) { 1074 if (object->type == OBJT_DEVICE) 1075 /* 1076 * The option OBJPR_NOTMAPPED must be passed here 1077 * because vm_object_page_remove() cannot remove 1078 * unmanaged mappings. 1079 */ 1080 flags = OBJPR_NOTMAPPED; 1081 else if (old_msync) 1082 flags = 0; 1083 else 1084 flags = OBJPR_CLEANONLY; 1085 vm_object_page_remove(object, OFF_TO_IDX(offset), 1086 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1087 } 1088 VM_OBJECT_WUNLOCK(object); 1089 return (res); 1090 } 1091 1092 /* 1093 * Determine whether the given advice can be applied to the object. Advice is 1094 * not applied to unmanaged pages since they never belong to page queues, and 1095 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1096 * have been mapped at most once. 1097 */ 1098 static bool 1099 vm_object_advice_applies(vm_object_t object, int advice) 1100 { 1101 1102 if ((object->flags & OBJ_UNMANAGED) != 0) 1103 return (false); 1104 if (advice != MADV_FREE) 1105 return (true); 1106 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && 1107 (object->flags & OBJ_ONEMAPPING) != 0); 1108 } 1109 1110 static void 1111 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1112 vm_size_t size) 1113 { 1114 1115 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1116 swap_pager_freespace(object, pindex, size); 1117 } 1118 1119 /* 1120 * vm_object_madvise: 1121 * 1122 * Implements the madvise function at the object/page level. 1123 * 1124 * MADV_WILLNEED (any object) 1125 * 1126 * Activate the specified pages if they are resident. 1127 * 1128 * MADV_DONTNEED (any object) 1129 * 1130 * Deactivate the specified pages if they are resident. 1131 * 1132 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1133 * OBJ_ONEMAPPING only) 1134 * 1135 * Deactivate and clean the specified pages if they are 1136 * resident. This permits the process to reuse the pages 1137 * without faulting or the kernel to reclaim the pages 1138 * without I/O. 1139 */ 1140 void 1141 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1142 int advice) 1143 { 1144 vm_pindex_t tpindex; 1145 vm_object_t backing_object, tobject; 1146 vm_page_t m, tm; 1147 1148 if (object == NULL) 1149 return; 1150 1151 relookup: 1152 VM_OBJECT_WLOCK(object); 1153 if (!vm_object_advice_applies(object, advice)) { 1154 VM_OBJECT_WUNLOCK(object); 1155 return; 1156 } 1157 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1158 tobject = object; 1159 1160 /* 1161 * If the next page isn't resident in the top-level object, we 1162 * need to search the shadow chain. When applying MADV_FREE, we 1163 * take care to release any swap space used to store 1164 * non-resident pages. 1165 */ 1166 if (m == NULL || pindex < m->pindex) { 1167 /* 1168 * Optimize a common case: if the top-level object has 1169 * no backing object, we can skip over the non-resident 1170 * range in constant time. 1171 */ 1172 if (object->backing_object == NULL) { 1173 tpindex = (m != NULL && m->pindex < end) ? 1174 m->pindex : end; 1175 vm_object_madvise_freespace(object, advice, 1176 pindex, tpindex - pindex); 1177 if ((pindex = tpindex) == end) 1178 break; 1179 goto next_page; 1180 } 1181 1182 tpindex = pindex; 1183 do { 1184 vm_object_madvise_freespace(tobject, advice, 1185 tpindex, 1); 1186 /* 1187 * Prepare to search the next object in the 1188 * chain. 1189 */ 1190 backing_object = tobject->backing_object; 1191 if (backing_object == NULL) 1192 goto next_pindex; 1193 VM_OBJECT_WLOCK(backing_object); 1194 tpindex += 1195 OFF_TO_IDX(tobject->backing_object_offset); 1196 if (tobject != object) 1197 VM_OBJECT_WUNLOCK(tobject); 1198 tobject = backing_object; 1199 if (!vm_object_advice_applies(tobject, advice)) 1200 goto next_pindex; 1201 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1202 NULL); 1203 } else { 1204 next_page: 1205 tm = m; 1206 m = TAILQ_NEXT(m, listq); 1207 } 1208 1209 /* 1210 * If the page is not in a normal state, skip it. 1211 */ 1212 if (tm->valid != VM_PAGE_BITS_ALL) 1213 goto next_pindex; 1214 vm_page_lock(tm); 1215 if (vm_page_wired(tm)) { 1216 vm_page_unlock(tm); 1217 goto next_pindex; 1218 } 1219 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1220 ("vm_object_madvise: page %p is fictitious", tm)); 1221 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1222 ("vm_object_madvise: page %p is not managed", tm)); 1223 if (vm_page_busied(tm)) { 1224 if (object != tobject) 1225 VM_OBJECT_WUNLOCK(tobject); 1226 VM_OBJECT_WUNLOCK(object); 1227 if (advice == MADV_WILLNEED) { 1228 /* 1229 * Reference the page before unlocking and 1230 * sleeping so that the page daemon is less 1231 * likely to reclaim it. 1232 */ 1233 vm_page_aflag_set(tm, PGA_REFERENCED); 1234 } 1235 vm_page_busy_sleep(tm, "madvpo", false); 1236 goto relookup; 1237 } 1238 vm_page_advise(tm, advice); 1239 vm_page_unlock(tm); 1240 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1241 next_pindex: 1242 if (tobject != object) 1243 VM_OBJECT_WUNLOCK(tobject); 1244 } 1245 VM_OBJECT_WUNLOCK(object); 1246 } 1247 1248 /* 1249 * vm_object_shadow: 1250 * 1251 * Create a new object which is backed by the 1252 * specified existing object range. The source 1253 * object reference is deallocated. 1254 * 1255 * The new object and offset into that object 1256 * are returned in the source parameters. 1257 */ 1258 void 1259 vm_object_shadow( 1260 vm_object_t *object, /* IN/OUT */ 1261 vm_ooffset_t *offset, /* IN/OUT */ 1262 vm_size_t length) 1263 { 1264 vm_object_t source; 1265 vm_object_t result; 1266 1267 source = *object; 1268 1269 /* 1270 * Don't create the new object if the old object isn't shared. 1271 */ 1272 if (source != NULL) { 1273 VM_OBJECT_WLOCK(source); 1274 if (source->ref_count == 1 && 1275 source->handle == NULL && 1276 (source->type == OBJT_DEFAULT || 1277 source->type == OBJT_SWAP)) { 1278 VM_OBJECT_WUNLOCK(source); 1279 return; 1280 } 1281 VM_OBJECT_WUNLOCK(source); 1282 } 1283 1284 /* 1285 * Allocate a new object with the given length. 1286 */ 1287 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1288 1289 /* 1290 * The new object shadows the source object, adding a reference to it. 1291 * Our caller changes his reference to point to the new object, 1292 * removing a reference to the source object. Net result: no change 1293 * of reference count. 1294 * 1295 * Try to optimize the result object's page color when shadowing 1296 * in order to maintain page coloring consistency in the combined 1297 * shadowed object. 1298 */ 1299 result->backing_object = source; 1300 /* 1301 * Store the offset into the source object, and fix up the offset into 1302 * the new object. 1303 */ 1304 result->backing_object_offset = *offset; 1305 if (source != NULL) { 1306 VM_OBJECT_WLOCK(source); 1307 result->domain = source->domain; 1308 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1309 source->shadow_count++; 1310 #if VM_NRESERVLEVEL > 0 1311 result->flags |= source->flags & OBJ_COLORED; 1312 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1313 ((1 << (VM_NFREEORDER - 1)) - 1); 1314 #endif 1315 VM_OBJECT_WUNLOCK(source); 1316 } 1317 1318 1319 /* 1320 * Return the new things 1321 */ 1322 *offset = 0; 1323 *object = result; 1324 } 1325 1326 /* 1327 * vm_object_split: 1328 * 1329 * Split the pages in a map entry into a new object. This affords 1330 * easier removal of unused pages, and keeps object inheritance from 1331 * being a negative impact on memory usage. 1332 */ 1333 void 1334 vm_object_split(vm_map_entry_t entry) 1335 { 1336 vm_page_t m, m_next; 1337 vm_object_t orig_object, new_object, source; 1338 vm_pindex_t idx, offidxstart; 1339 vm_size_t size; 1340 1341 orig_object = entry->object.vm_object; 1342 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1343 return; 1344 if (orig_object->ref_count <= 1) 1345 return; 1346 VM_OBJECT_WUNLOCK(orig_object); 1347 1348 offidxstart = OFF_TO_IDX(entry->offset); 1349 size = atop(entry->end - entry->start); 1350 1351 /* 1352 * If swap_pager_copy() is later called, it will convert new_object 1353 * into a swap object. 1354 */ 1355 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1356 1357 /* 1358 * At this point, the new object is still private, so the order in 1359 * which the original and new objects are locked does not matter. 1360 */ 1361 VM_OBJECT_WLOCK(new_object); 1362 VM_OBJECT_WLOCK(orig_object); 1363 new_object->domain = orig_object->domain; 1364 source = orig_object->backing_object; 1365 if (source != NULL) { 1366 VM_OBJECT_WLOCK(source); 1367 if ((source->flags & OBJ_DEAD) != 0) { 1368 VM_OBJECT_WUNLOCK(source); 1369 VM_OBJECT_WUNLOCK(orig_object); 1370 VM_OBJECT_WUNLOCK(new_object); 1371 vm_object_deallocate(new_object); 1372 VM_OBJECT_WLOCK(orig_object); 1373 return; 1374 } 1375 LIST_INSERT_HEAD(&source->shadow_head, 1376 new_object, shadow_list); 1377 source->shadow_count++; 1378 vm_object_reference_locked(source); /* for new_object */ 1379 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1380 VM_OBJECT_WUNLOCK(source); 1381 new_object->backing_object_offset = 1382 orig_object->backing_object_offset + entry->offset; 1383 new_object->backing_object = source; 1384 } 1385 if (orig_object->cred != NULL) { 1386 new_object->cred = orig_object->cred; 1387 crhold(orig_object->cred); 1388 new_object->charge = ptoa(size); 1389 KASSERT(orig_object->charge >= ptoa(size), 1390 ("orig_object->charge < 0")); 1391 orig_object->charge -= ptoa(size); 1392 } 1393 retry: 1394 m = vm_page_find_least(orig_object, offidxstart); 1395 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1396 m = m_next) { 1397 m_next = TAILQ_NEXT(m, listq); 1398 1399 /* 1400 * We must wait for pending I/O to complete before we can 1401 * rename the page. 1402 * 1403 * We do not have to VM_PROT_NONE the page as mappings should 1404 * not be changed by this operation. 1405 */ 1406 if (vm_page_busied(m)) { 1407 VM_OBJECT_WUNLOCK(new_object); 1408 vm_page_lock(m); 1409 VM_OBJECT_WUNLOCK(orig_object); 1410 vm_page_busy_sleep(m, "spltwt", false); 1411 VM_OBJECT_WLOCK(orig_object); 1412 VM_OBJECT_WLOCK(new_object); 1413 goto retry; 1414 } 1415 1416 /* vm_page_rename() will dirty the page. */ 1417 if (vm_page_rename(m, new_object, idx)) { 1418 VM_OBJECT_WUNLOCK(new_object); 1419 VM_OBJECT_WUNLOCK(orig_object); 1420 vm_radix_wait(); 1421 VM_OBJECT_WLOCK(orig_object); 1422 VM_OBJECT_WLOCK(new_object); 1423 goto retry; 1424 } 1425 #if VM_NRESERVLEVEL > 0 1426 /* 1427 * If some of the reservation's allocated pages remain with 1428 * the original object, then transferring the reservation to 1429 * the new object is neither particularly beneficial nor 1430 * particularly harmful as compared to leaving the reservation 1431 * with the original object. If, however, all of the 1432 * reservation's allocated pages are transferred to the new 1433 * object, then transferring the reservation is typically 1434 * beneficial. Determining which of these two cases applies 1435 * would be more costly than unconditionally renaming the 1436 * reservation. 1437 */ 1438 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1439 #endif 1440 if (orig_object->type == OBJT_SWAP) 1441 vm_page_xbusy(m); 1442 } 1443 if (orig_object->type == OBJT_SWAP) { 1444 /* 1445 * swap_pager_copy() can sleep, in which case the orig_object's 1446 * and new_object's locks are released and reacquired. 1447 */ 1448 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1449 TAILQ_FOREACH(m, &new_object->memq, listq) 1450 vm_page_xunbusy(m); 1451 } 1452 VM_OBJECT_WUNLOCK(orig_object); 1453 VM_OBJECT_WUNLOCK(new_object); 1454 entry->object.vm_object = new_object; 1455 entry->offset = 0LL; 1456 vm_object_deallocate(orig_object); 1457 VM_OBJECT_WLOCK(new_object); 1458 } 1459 1460 #define OBSC_COLLAPSE_NOWAIT 0x0002 1461 #define OBSC_COLLAPSE_WAIT 0x0004 1462 1463 static vm_page_t 1464 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1465 int op) 1466 { 1467 vm_object_t backing_object; 1468 1469 VM_OBJECT_ASSERT_WLOCKED(object); 1470 backing_object = object->backing_object; 1471 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1472 1473 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1474 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1475 ("invalid ownership %p %p %p", p, object, backing_object)); 1476 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1477 return (next); 1478 if (p != NULL) 1479 vm_page_lock(p); 1480 VM_OBJECT_WUNLOCK(object); 1481 VM_OBJECT_WUNLOCK(backing_object); 1482 /* The page is only NULL when rename fails. */ 1483 if (p == NULL) 1484 vm_radix_wait(); 1485 else 1486 vm_page_busy_sleep(p, "vmocol", false); 1487 VM_OBJECT_WLOCK(object); 1488 VM_OBJECT_WLOCK(backing_object); 1489 return (TAILQ_FIRST(&backing_object->memq)); 1490 } 1491 1492 static bool 1493 vm_object_scan_all_shadowed(vm_object_t object) 1494 { 1495 vm_object_t backing_object; 1496 vm_page_t p, pp; 1497 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1498 1499 VM_OBJECT_ASSERT_WLOCKED(object); 1500 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1501 1502 backing_object = object->backing_object; 1503 1504 if (backing_object->type != OBJT_DEFAULT && 1505 backing_object->type != OBJT_SWAP) 1506 return (false); 1507 1508 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1509 p = vm_page_find_least(backing_object, pi); 1510 ps = swap_pager_find_least(backing_object, pi); 1511 1512 /* 1513 * Only check pages inside the parent object's range and 1514 * inside the parent object's mapping of the backing object. 1515 */ 1516 for (;; pi++) { 1517 if (p != NULL && p->pindex < pi) 1518 p = TAILQ_NEXT(p, listq); 1519 if (ps < pi) 1520 ps = swap_pager_find_least(backing_object, pi); 1521 if (p == NULL && ps >= backing_object->size) 1522 break; 1523 else if (p == NULL) 1524 pi = ps; 1525 else 1526 pi = MIN(p->pindex, ps); 1527 1528 new_pindex = pi - backing_offset_index; 1529 if (new_pindex >= object->size) 1530 break; 1531 1532 /* 1533 * See if the parent has the page or if the parent's object 1534 * pager has the page. If the parent has the page but the page 1535 * is not valid, the parent's object pager must have the page. 1536 * 1537 * If this fails, the parent does not completely shadow the 1538 * object and we might as well give up now. 1539 */ 1540 pp = vm_page_lookup(object, new_pindex); 1541 if ((pp == NULL || pp->valid == 0) && 1542 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1543 return (false); 1544 } 1545 return (true); 1546 } 1547 1548 static bool 1549 vm_object_collapse_scan(vm_object_t object, int op) 1550 { 1551 vm_object_t backing_object; 1552 vm_page_t next, p, pp; 1553 vm_pindex_t backing_offset_index, new_pindex; 1554 1555 VM_OBJECT_ASSERT_WLOCKED(object); 1556 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1557 1558 backing_object = object->backing_object; 1559 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1560 1561 /* 1562 * Initial conditions 1563 */ 1564 if ((op & OBSC_COLLAPSE_WAIT) != 0) 1565 vm_object_set_flag(backing_object, OBJ_DEAD); 1566 1567 /* 1568 * Our scan 1569 */ 1570 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1571 next = TAILQ_NEXT(p, listq); 1572 new_pindex = p->pindex - backing_offset_index; 1573 1574 /* 1575 * Check for busy page 1576 */ 1577 if (vm_page_busied(p)) { 1578 next = vm_object_collapse_scan_wait(object, p, next, op); 1579 continue; 1580 } 1581 1582 KASSERT(p->object == backing_object, 1583 ("vm_object_collapse_scan: object mismatch")); 1584 1585 if (p->pindex < backing_offset_index || 1586 new_pindex >= object->size) { 1587 if (backing_object->type == OBJT_SWAP) 1588 swap_pager_freespace(backing_object, p->pindex, 1589 1); 1590 1591 /* 1592 * Page is out of the parent object's range, we can 1593 * simply destroy it. 1594 */ 1595 vm_page_lock(p); 1596 KASSERT(!pmap_page_is_mapped(p), 1597 ("freeing mapped page %p", p)); 1598 if (vm_page_remove(p)) 1599 vm_page_free(p); 1600 vm_page_unlock(p); 1601 continue; 1602 } 1603 1604 pp = vm_page_lookup(object, new_pindex); 1605 if (pp != NULL && vm_page_busied(pp)) { 1606 /* 1607 * The page in the parent is busy and possibly not 1608 * (yet) valid. Until its state is finalized by the 1609 * busy bit owner, we can't tell whether it shadows the 1610 * original page. Therefore, we must either skip it 1611 * and the original (backing_object) page or wait for 1612 * its state to be finalized. 1613 * 1614 * This is due to a race with vm_fault() where we must 1615 * unbusy the original (backing_obj) page before we can 1616 * (re)lock the parent. Hence we can get here. 1617 */ 1618 next = vm_object_collapse_scan_wait(object, pp, next, 1619 op); 1620 continue; 1621 } 1622 1623 KASSERT(pp == NULL || pp->valid != 0, 1624 ("unbusy invalid page %p", pp)); 1625 1626 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1627 NULL)) { 1628 /* 1629 * The page already exists in the parent OR swap exists 1630 * for this location in the parent. Leave the parent's 1631 * page alone. Destroy the original page from the 1632 * backing object. 1633 */ 1634 if (backing_object->type == OBJT_SWAP) 1635 swap_pager_freespace(backing_object, p->pindex, 1636 1); 1637 vm_page_lock(p); 1638 KASSERT(!pmap_page_is_mapped(p), 1639 ("freeing mapped page %p", p)); 1640 if (vm_page_remove(p)) 1641 vm_page_free(p); 1642 vm_page_unlock(p); 1643 continue; 1644 } 1645 1646 /* 1647 * Page does not exist in parent, rename the page from the 1648 * backing object to the main object. 1649 * 1650 * If the page was mapped to a process, it can remain mapped 1651 * through the rename. vm_page_rename() will dirty the page. 1652 */ 1653 if (vm_page_rename(p, object, new_pindex)) { 1654 next = vm_object_collapse_scan_wait(object, NULL, next, 1655 op); 1656 continue; 1657 } 1658 1659 /* Use the old pindex to free the right page. */ 1660 if (backing_object->type == OBJT_SWAP) 1661 swap_pager_freespace(backing_object, 1662 new_pindex + backing_offset_index, 1); 1663 1664 #if VM_NRESERVLEVEL > 0 1665 /* 1666 * Rename the reservation. 1667 */ 1668 vm_reserv_rename(p, object, backing_object, 1669 backing_offset_index); 1670 #endif 1671 } 1672 return (true); 1673 } 1674 1675 1676 /* 1677 * this version of collapse allows the operation to occur earlier and 1678 * when paging_in_progress is true for an object... This is not a complete 1679 * operation, but should plug 99.9% of the rest of the leaks. 1680 */ 1681 static void 1682 vm_object_qcollapse(vm_object_t object) 1683 { 1684 vm_object_t backing_object = object->backing_object; 1685 1686 VM_OBJECT_ASSERT_WLOCKED(object); 1687 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1688 1689 if (backing_object->ref_count != 1) 1690 return; 1691 1692 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); 1693 } 1694 1695 /* 1696 * vm_object_collapse: 1697 * 1698 * Collapse an object with the object backing it. 1699 * Pages in the backing object are moved into the 1700 * parent, and the backing object is deallocated. 1701 */ 1702 void 1703 vm_object_collapse(vm_object_t object) 1704 { 1705 vm_object_t backing_object, new_backing_object; 1706 1707 VM_OBJECT_ASSERT_WLOCKED(object); 1708 1709 while (TRUE) { 1710 /* 1711 * Verify that the conditions are right for collapse: 1712 * 1713 * The object exists and the backing object exists. 1714 */ 1715 if ((backing_object = object->backing_object) == NULL) 1716 break; 1717 1718 /* 1719 * we check the backing object first, because it is most likely 1720 * not collapsable. 1721 */ 1722 VM_OBJECT_WLOCK(backing_object); 1723 if (backing_object->handle != NULL || 1724 (backing_object->type != OBJT_DEFAULT && 1725 backing_object->type != OBJT_SWAP) || 1726 (backing_object->flags & (OBJ_DEAD | OBJ_NOSPLIT)) != 0 || 1727 object->handle != NULL || 1728 (object->type != OBJT_DEFAULT && 1729 object->type != OBJT_SWAP) || 1730 (object->flags & OBJ_DEAD)) { 1731 VM_OBJECT_WUNLOCK(backing_object); 1732 break; 1733 } 1734 1735 if (object->paging_in_progress != 0 || 1736 backing_object->paging_in_progress != 0) { 1737 vm_object_qcollapse(object); 1738 VM_OBJECT_WUNLOCK(backing_object); 1739 break; 1740 } 1741 1742 /* 1743 * We know that we can either collapse the backing object (if 1744 * the parent is the only reference to it) or (perhaps) have 1745 * the parent bypass the object if the parent happens to shadow 1746 * all the resident pages in the entire backing object. 1747 * 1748 * This is ignoring pager-backed pages such as swap pages. 1749 * vm_object_collapse_scan fails the shadowing test in this 1750 * case. 1751 */ 1752 if (backing_object->ref_count == 1) { 1753 vm_object_pip_add(object, 1); 1754 vm_object_pip_add(backing_object, 1); 1755 1756 /* 1757 * If there is exactly one reference to the backing 1758 * object, we can collapse it into the parent. 1759 */ 1760 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); 1761 1762 #if VM_NRESERVLEVEL > 0 1763 /* 1764 * Break any reservations from backing_object. 1765 */ 1766 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1767 vm_reserv_break_all(backing_object); 1768 #endif 1769 1770 /* 1771 * Move the pager from backing_object to object. 1772 */ 1773 if (backing_object->type == OBJT_SWAP) { 1774 /* 1775 * swap_pager_copy() can sleep, in which case 1776 * the backing_object's and object's locks are 1777 * released and reacquired. 1778 * Since swap_pager_copy() is being asked to 1779 * destroy the source, it will change the 1780 * backing_object's type to OBJT_DEFAULT. 1781 */ 1782 swap_pager_copy( 1783 backing_object, 1784 object, 1785 OFF_TO_IDX(object->backing_object_offset), TRUE); 1786 } 1787 /* 1788 * Object now shadows whatever backing_object did. 1789 * Note that the reference to 1790 * backing_object->backing_object moves from within 1791 * backing_object to within object. 1792 */ 1793 LIST_REMOVE(object, shadow_list); 1794 backing_object->shadow_count--; 1795 if (backing_object->backing_object) { 1796 VM_OBJECT_WLOCK(backing_object->backing_object); 1797 LIST_REMOVE(backing_object, shadow_list); 1798 LIST_INSERT_HEAD( 1799 &backing_object->backing_object->shadow_head, 1800 object, shadow_list); 1801 /* 1802 * The shadow_count has not changed. 1803 */ 1804 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1805 } 1806 object->backing_object = backing_object->backing_object; 1807 object->backing_object_offset += 1808 backing_object->backing_object_offset; 1809 1810 /* 1811 * Discard backing_object. 1812 * 1813 * Since the backing object has no pages, no pager left, 1814 * and no object references within it, all that is 1815 * necessary is to dispose of it. 1816 */ 1817 KASSERT(backing_object->ref_count == 1, ( 1818 "backing_object %p was somehow re-referenced during collapse!", 1819 backing_object)); 1820 vm_object_pip_wakeup(backing_object); 1821 backing_object->type = OBJT_DEAD; 1822 backing_object->ref_count = 0; 1823 VM_OBJECT_WUNLOCK(backing_object); 1824 vm_object_destroy(backing_object); 1825 1826 vm_object_pip_wakeup(object); 1827 counter_u64_add(object_collapses, 1); 1828 } else { 1829 /* 1830 * If we do not entirely shadow the backing object, 1831 * there is nothing we can do so we give up. 1832 */ 1833 if (object->resident_page_count != object->size && 1834 !vm_object_scan_all_shadowed(object)) { 1835 VM_OBJECT_WUNLOCK(backing_object); 1836 break; 1837 } 1838 1839 /* 1840 * Make the parent shadow the next object in the 1841 * chain. Deallocating backing_object will not remove 1842 * it, since its reference count is at least 2. 1843 */ 1844 LIST_REMOVE(object, shadow_list); 1845 backing_object->shadow_count--; 1846 1847 new_backing_object = backing_object->backing_object; 1848 if ((object->backing_object = new_backing_object) != NULL) { 1849 VM_OBJECT_WLOCK(new_backing_object); 1850 LIST_INSERT_HEAD( 1851 &new_backing_object->shadow_head, 1852 object, 1853 shadow_list 1854 ); 1855 new_backing_object->shadow_count++; 1856 vm_object_reference_locked(new_backing_object); 1857 VM_OBJECT_WUNLOCK(new_backing_object); 1858 object->backing_object_offset += 1859 backing_object->backing_object_offset; 1860 } 1861 1862 /* 1863 * Drop the reference count on backing_object. Since 1864 * its ref_count was at least 2, it will not vanish. 1865 */ 1866 backing_object->ref_count--; 1867 VM_OBJECT_WUNLOCK(backing_object); 1868 counter_u64_add(object_bypasses, 1); 1869 } 1870 1871 /* 1872 * Try again with this object's new backing object. 1873 */ 1874 } 1875 } 1876 1877 /* 1878 * vm_object_page_remove: 1879 * 1880 * For the given object, either frees or invalidates each of the 1881 * specified pages. In general, a page is freed. However, if a page is 1882 * wired for any reason other than the existence of a managed, wired 1883 * mapping, then it may be invalidated but not removed from the object. 1884 * Pages are specified by the given range ["start", "end") and the option 1885 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1886 * extends from "start" to the end of the object. If the option 1887 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1888 * specified range are affected. If the option OBJPR_NOTMAPPED is 1889 * specified, then the pages within the specified range must have no 1890 * mappings. Otherwise, if this option is not specified, any mappings to 1891 * the specified pages are removed before the pages are freed or 1892 * invalidated. 1893 * 1894 * In general, this operation should only be performed on objects that 1895 * contain managed pages. There are, however, two exceptions. First, it 1896 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1897 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1898 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1899 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1900 * 1901 * The object must be locked. 1902 */ 1903 void 1904 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1905 int options) 1906 { 1907 vm_page_t p, next; 1908 struct mtx *mtx; 1909 1910 VM_OBJECT_ASSERT_WLOCKED(object); 1911 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1912 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1913 ("vm_object_page_remove: illegal options for object %p", object)); 1914 if (object->resident_page_count == 0) 1915 return; 1916 vm_object_pip_add(object, 1); 1917 again: 1918 p = vm_page_find_least(object, start); 1919 mtx = NULL; 1920 1921 /* 1922 * Here, the variable "p" is either (1) the page with the least pindex 1923 * greater than or equal to the parameter "start" or (2) NULL. 1924 */ 1925 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1926 next = TAILQ_NEXT(p, listq); 1927 1928 /* 1929 * If the page is wired for any reason besides the existence 1930 * of managed, wired mappings, then it cannot be freed. For 1931 * example, fictitious pages, which represent device memory, 1932 * are inherently wired and cannot be freed. They can, 1933 * however, be invalidated if the option OBJPR_CLEANONLY is 1934 * not specified. 1935 */ 1936 vm_page_change_lock(p, &mtx); 1937 if (vm_page_xbusied(p)) { 1938 VM_OBJECT_WUNLOCK(object); 1939 vm_page_busy_sleep(p, "vmopax", true); 1940 VM_OBJECT_WLOCK(object); 1941 goto again; 1942 } 1943 if (vm_page_wired(p)) { 1944 if ((options & OBJPR_NOTMAPPED) == 0 && 1945 object->ref_count != 0) 1946 pmap_remove_all(p); 1947 if ((options & OBJPR_CLEANONLY) == 0) { 1948 p->valid = 0; 1949 vm_page_undirty(p); 1950 } 1951 continue; 1952 } 1953 if (vm_page_busied(p)) { 1954 VM_OBJECT_WUNLOCK(object); 1955 vm_page_busy_sleep(p, "vmopar", false); 1956 VM_OBJECT_WLOCK(object); 1957 goto again; 1958 } 1959 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1960 ("vm_object_page_remove: page %p is fictitious", p)); 1961 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 1962 if ((options & OBJPR_NOTMAPPED) == 0 && 1963 object->ref_count != 0) 1964 pmap_remove_write(p); 1965 if (p->dirty != 0) 1966 continue; 1967 } 1968 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) 1969 pmap_remove_all(p); 1970 vm_page_free(p); 1971 } 1972 if (mtx != NULL) 1973 mtx_unlock(mtx); 1974 vm_object_pip_wakeup(object); 1975 } 1976 1977 /* 1978 * vm_object_page_noreuse: 1979 * 1980 * For the given object, attempt to move the specified pages to 1981 * the head of the inactive queue. This bypasses regular LRU 1982 * operation and allows the pages to be reused quickly under memory 1983 * pressure. If a page is wired for any reason, then it will not 1984 * be queued. Pages are specified by the range ["start", "end"). 1985 * As a special case, if "end" is zero, then the range extends from 1986 * "start" to the end of the object. 1987 * 1988 * This operation should only be performed on objects that 1989 * contain non-fictitious, managed pages. 1990 * 1991 * The object must be locked. 1992 */ 1993 void 1994 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1995 { 1996 struct mtx *mtx; 1997 vm_page_t p, next; 1998 1999 VM_OBJECT_ASSERT_LOCKED(object); 2000 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2001 ("vm_object_page_noreuse: illegal object %p", object)); 2002 if (object->resident_page_count == 0) 2003 return; 2004 p = vm_page_find_least(object, start); 2005 2006 /* 2007 * Here, the variable "p" is either (1) the page with the least pindex 2008 * greater than or equal to the parameter "start" or (2) NULL. 2009 */ 2010 mtx = NULL; 2011 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2012 next = TAILQ_NEXT(p, listq); 2013 vm_page_change_lock(p, &mtx); 2014 vm_page_deactivate_noreuse(p); 2015 } 2016 if (mtx != NULL) 2017 mtx_unlock(mtx); 2018 } 2019 2020 /* 2021 * Populate the specified range of the object with valid pages. Returns 2022 * TRUE if the range is successfully populated and FALSE otherwise. 2023 * 2024 * Note: This function should be optimized to pass a larger array of 2025 * pages to vm_pager_get_pages() before it is applied to a non- 2026 * OBJT_DEVICE object. 2027 * 2028 * The object must be locked. 2029 */ 2030 boolean_t 2031 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2032 { 2033 vm_page_t m; 2034 vm_pindex_t pindex; 2035 int rv; 2036 2037 VM_OBJECT_ASSERT_WLOCKED(object); 2038 for (pindex = start; pindex < end; pindex++) { 2039 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2040 if (m->valid != VM_PAGE_BITS_ALL) { 2041 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 2042 if (rv != VM_PAGER_OK) { 2043 vm_page_lock(m); 2044 vm_page_free(m); 2045 vm_page_unlock(m); 2046 break; 2047 } 2048 } 2049 /* 2050 * Keep "m" busy because a subsequent iteration may unlock 2051 * the object. 2052 */ 2053 } 2054 if (pindex > start) { 2055 m = vm_page_lookup(object, start); 2056 while (m != NULL && m->pindex < pindex) { 2057 vm_page_xunbusy(m); 2058 m = TAILQ_NEXT(m, listq); 2059 } 2060 } 2061 return (pindex == end); 2062 } 2063 2064 /* 2065 * Routine: vm_object_coalesce 2066 * Function: Coalesces two objects backing up adjoining 2067 * regions of memory into a single object. 2068 * 2069 * returns TRUE if objects were combined. 2070 * 2071 * NOTE: Only works at the moment if the second object is NULL - 2072 * if it's not, which object do we lock first? 2073 * 2074 * Parameters: 2075 * prev_object First object to coalesce 2076 * prev_offset Offset into prev_object 2077 * prev_size Size of reference to prev_object 2078 * next_size Size of reference to the second object 2079 * reserved Indicator that extension region has 2080 * swap accounted for 2081 * 2082 * Conditions: 2083 * The object must *not* be locked. 2084 */ 2085 boolean_t 2086 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2087 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2088 { 2089 vm_pindex_t next_pindex; 2090 2091 if (prev_object == NULL) 2092 return (TRUE); 2093 VM_OBJECT_WLOCK(prev_object); 2094 if ((prev_object->type != OBJT_DEFAULT && 2095 prev_object->type != OBJT_SWAP) || 2096 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2097 VM_OBJECT_WUNLOCK(prev_object); 2098 return (FALSE); 2099 } 2100 2101 /* 2102 * Try to collapse the object first 2103 */ 2104 vm_object_collapse(prev_object); 2105 2106 /* 2107 * Can't coalesce if: . more than one reference . paged out . shadows 2108 * another object . has a copy elsewhere (any of which mean that the 2109 * pages not mapped to prev_entry may be in use anyway) 2110 */ 2111 if (prev_object->backing_object != NULL) { 2112 VM_OBJECT_WUNLOCK(prev_object); 2113 return (FALSE); 2114 } 2115 2116 prev_size >>= PAGE_SHIFT; 2117 next_size >>= PAGE_SHIFT; 2118 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2119 2120 if (prev_object->ref_count > 1 && 2121 prev_object->size != next_pindex && 2122 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2123 VM_OBJECT_WUNLOCK(prev_object); 2124 return (FALSE); 2125 } 2126 2127 /* 2128 * Account for the charge. 2129 */ 2130 if (prev_object->cred != NULL) { 2131 2132 /* 2133 * If prev_object was charged, then this mapping, 2134 * although not charged now, may become writable 2135 * later. Non-NULL cred in the object would prevent 2136 * swap reservation during enabling of the write 2137 * access, so reserve swap now. Failed reservation 2138 * cause allocation of the separate object for the map 2139 * entry, and swap reservation for this entry is 2140 * managed in appropriate time. 2141 */ 2142 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2143 prev_object->cred)) { 2144 VM_OBJECT_WUNLOCK(prev_object); 2145 return (FALSE); 2146 } 2147 prev_object->charge += ptoa(next_size); 2148 } 2149 2150 /* 2151 * Remove any pages that may still be in the object from a previous 2152 * deallocation. 2153 */ 2154 if (next_pindex < prev_object->size) { 2155 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2156 next_size, 0); 2157 if (prev_object->type == OBJT_SWAP) 2158 swap_pager_freespace(prev_object, 2159 next_pindex, next_size); 2160 #if 0 2161 if (prev_object->cred != NULL) { 2162 KASSERT(prev_object->charge >= 2163 ptoa(prev_object->size - next_pindex), 2164 ("object %p overcharged 1 %jx %jx", prev_object, 2165 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2166 prev_object->charge -= ptoa(prev_object->size - 2167 next_pindex); 2168 } 2169 #endif 2170 } 2171 2172 /* 2173 * Extend the object if necessary. 2174 */ 2175 if (next_pindex + next_size > prev_object->size) 2176 prev_object->size = next_pindex + next_size; 2177 2178 VM_OBJECT_WUNLOCK(prev_object); 2179 return (TRUE); 2180 } 2181 2182 void 2183 vm_object_set_writeable_dirty(vm_object_t object) 2184 { 2185 2186 VM_OBJECT_ASSERT_WLOCKED(object); 2187 if (object->type != OBJT_VNODE) { 2188 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2189 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2190 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2191 } 2192 return; 2193 } 2194 object->generation++; 2195 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2196 return; 2197 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2198 } 2199 2200 /* 2201 * vm_object_unwire: 2202 * 2203 * For each page offset within the specified range of the given object, 2204 * find the highest-level page in the shadow chain and unwire it. A page 2205 * must exist at every page offset, and the highest-level page must be 2206 * wired. 2207 */ 2208 void 2209 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2210 uint8_t queue) 2211 { 2212 vm_object_t tobject, t1object; 2213 vm_page_t m, tm; 2214 vm_pindex_t end_pindex, pindex, tpindex; 2215 int depth, locked_depth; 2216 2217 KASSERT((offset & PAGE_MASK) == 0, 2218 ("vm_object_unwire: offset is not page aligned")); 2219 KASSERT((length & PAGE_MASK) == 0, 2220 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2221 /* The wired count of a fictitious page never changes. */ 2222 if ((object->flags & OBJ_FICTITIOUS) != 0) 2223 return; 2224 pindex = OFF_TO_IDX(offset); 2225 end_pindex = pindex + atop(length); 2226 again: 2227 locked_depth = 1; 2228 VM_OBJECT_RLOCK(object); 2229 m = vm_page_find_least(object, pindex); 2230 while (pindex < end_pindex) { 2231 if (m == NULL || pindex < m->pindex) { 2232 /* 2233 * The first object in the shadow chain doesn't 2234 * contain a page at the current index. Therefore, 2235 * the page must exist in a backing object. 2236 */ 2237 tobject = object; 2238 tpindex = pindex; 2239 depth = 0; 2240 do { 2241 tpindex += 2242 OFF_TO_IDX(tobject->backing_object_offset); 2243 tobject = tobject->backing_object; 2244 KASSERT(tobject != NULL, 2245 ("vm_object_unwire: missing page")); 2246 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2247 goto next_page; 2248 depth++; 2249 if (depth == locked_depth) { 2250 locked_depth++; 2251 VM_OBJECT_RLOCK(tobject); 2252 } 2253 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2254 NULL); 2255 } else { 2256 tm = m; 2257 m = TAILQ_NEXT(m, listq); 2258 } 2259 vm_page_lock(tm); 2260 if (vm_page_xbusied(tm)) { 2261 for (tobject = object; locked_depth >= 1; 2262 locked_depth--) { 2263 t1object = tobject->backing_object; 2264 VM_OBJECT_RUNLOCK(tobject); 2265 tobject = t1object; 2266 } 2267 vm_page_busy_sleep(tm, "unwbo", true); 2268 goto again; 2269 } 2270 vm_page_unwire(tm, queue); 2271 vm_page_unlock(tm); 2272 next_page: 2273 pindex++; 2274 } 2275 /* Release the accumulated object locks. */ 2276 for (tobject = object; locked_depth >= 1; locked_depth--) { 2277 t1object = tobject->backing_object; 2278 VM_OBJECT_RUNLOCK(tobject); 2279 tobject = t1object; 2280 } 2281 } 2282 2283 /* 2284 * Return the vnode for the given object, or NULL if none exists. 2285 * For tmpfs objects, the function may return NULL if there is 2286 * no vnode allocated at the time of the call. 2287 */ 2288 struct vnode * 2289 vm_object_vnode(vm_object_t object) 2290 { 2291 struct vnode *vp; 2292 2293 VM_OBJECT_ASSERT_LOCKED(object); 2294 if (object->type == OBJT_VNODE) { 2295 vp = object->handle; 2296 KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); 2297 } else if (object->type == OBJT_SWAP && 2298 (object->flags & OBJ_TMPFS) != 0) { 2299 vp = object->un_pager.swp.swp_tmpfs; 2300 KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); 2301 } else { 2302 vp = NULL; 2303 } 2304 return (vp); 2305 } 2306 2307 /* 2308 * Return the kvme type of the given object. 2309 * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. 2310 */ 2311 int 2312 vm_object_kvme_type(vm_object_t object, struct vnode **vpp) 2313 { 2314 2315 VM_OBJECT_ASSERT_LOCKED(object); 2316 if (vpp != NULL) 2317 *vpp = vm_object_vnode(object); 2318 switch (object->type) { 2319 case OBJT_DEFAULT: 2320 return (KVME_TYPE_DEFAULT); 2321 case OBJT_VNODE: 2322 return (KVME_TYPE_VNODE); 2323 case OBJT_SWAP: 2324 if ((object->flags & OBJ_TMPFS_NODE) != 0) 2325 return (KVME_TYPE_VNODE); 2326 return (KVME_TYPE_SWAP); 2327 case OBJT_DEVICE: 2328 return (KVME_TYPE_DEVICE); 2329 case OBJT_PHYS: 2330 return (KVME_TYPE_PHYS); 2331 case OBJT_DEAD: 2332 return (KVME_TYPE_DEAD); 2333 case OBJT_SG: 2334 return (KVME_TYPE_SG); 2335 case OBJT_MGTDEVICE: 2336 return (KVME_TYPE_MGTDEVICE); 2337 default: 2338 return (KVME_TYPE_UNKNOWN); 2339 } 2340 } 2341 2342 static int 2343 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2344 { 2345 struct kinfo_vmobject *kvo; 2346 char *fullpath, *freepath; 2347 struct vnode *vp; 2348 struct vattr va; 2349 vm_object_t obj; 2350 vm_page_t m; 2351 int count, error; 2352 2353 if (req->oldptr == NULL) { 2354 /* 2355 * If an old buffer has not been provided, generate an 2356 * estimate of the space needed for a subsequent call. 2357 */ 2358 mtx_lock(&vm_object_list_mtx); 2359 count = 0; 2360 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2361 if (obj->type == OBJT_DEAD) 2362 continue; 2363 count++; 2364 } 2365 mtx_unlock(&vm_object_list_mtx); 2366 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2367 count * 11 / 10)); 2368 } 2369 2370 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2371 error = 0; 2372 2373 /* 2374 * VM objects are type stable and are never removed from the 2375 * list once added. This allows us to safely read obj->object_list 2376 * after reacquiring the VM object lock. 2377 */ 2378 mtx_lock(&vm_object_list_mtx); 2379 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2380 if (obj->type == OBJT_DEAD) 2381 continue; 2382 VM_OBJECT_RLOCK(obj); 2383 if (obj->type == OBJT_DEAD) { 2384 VM_OBJECT_RUNLOCK(obj); 2385 continue; 2386 } 2387 mtx_unlock(&vm_object_list_mtx); 2388 kvo->kvo_size = ptoa(obj->size); 2389 kvo->kvo_resident = obj->resident_page_count; 2390 kvo->kvo_ref_count = obj->ref_count; 2391 kvo->kvo_shadow_count = obj->shadow_count; 2392 kvo->kvo_memattr = obj->memattr; 2393 kvo->kvo_active = 0; 2394 kvo->kvo_inactive = 0; 2395 TAILQ_FOREACH(m, &obj->memq, listq) { 2396 /* 2397 * A page may belong to the object but be 2398 * dequeued and set to PQ_NONE while the 2399 * object lock is not held. This makes the 2400 * reads of m->queue below racy, and we do not 2401 * count pages set to PQ_NONE. However, this 2402 * sysctl is only meant to give an 2403 * approximation of the system anyway. 2404 */ 2405 if (m->queue == PQ_ACTIVE) 2406 kvo->kvo_active++; 2407 else if (m->queue == PQ_INACTIVE) 2408 kvo->kvo_inactive++; 2409 } 2410 2411 kvo->kvo_vn_fileid = 0; 2412 kvo->kvo_vn_fsid = 0; 2413 kvo->kvo_vn_fsid_freebsd11 = 0; 2414 freepath = NULL; 2415 fullpath = ""; 2416 kvo->kvo_type = vm_object_kvme_type(obj, &vp); 2417 if (vp != NULL) 2418 vref(vp); 2419 VM_OBJECT_RUNLOCK(obj); 2420 if (vp != NULL) { 2421 vn_fullpath(curthread, vp, &fullpath, &freepath); 2422 vn_lock(vp, LK_SHARED | LK_RETRY); 2423 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2424 kvo->kvo_vn_fileid = va.va_fileid; 2425 kvo->kvo_vn_fsid = va.va_fsid; 2426 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2427 /* truncate */ 2428 } 2429 vput(vp); 2430 } 2431 2432 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2433 if (freepath != NULL) 2434 free(freepath, M_TEMP); 2435 2436 /* Pack record size down */ 2437 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2438 + strlen(kvo->kvo_path) + 1; 2439 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2440 sizeof(uint64_t)); 2441 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2442 mtx_lock(&vm_object_list_mtx); 2443 if (error) 2444 break; 2445 } 2446 mtx_unlock(&vm_object_list_mtx); 2447 free(kvo, M_TEMP); 2448 return (error); 2449 } 2450 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2451 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2452 "List of VM objects"); 2453 2454 #include "opt_ddb.h" 2455 #ifdef DDB 2456 #include <sys/kernel.h> 2457 2458 #include <sys/cons.h> 2459 2460 #include <ddb/ddb.h> 2461 2462 static int 2463 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2464 { 2465 vm_map_t tmpm; 2466 vm_map_entry_t tmpe; 2467 vm_object_t obj; 2468 int entcount; 2469 2470 if (map == 0) 2471 return 0; 2472 2473 if (entry == 0) { 2474 tmpe = map->header.next; 2475 entcount = map->nentries; 2476 while (entcount-- && (tmpe != &map->header)) { 2477 if (_vm_object_in_map(map, object, tmpe)) { 2478 return 1; 2479 } 2480 tmpe = tmpe->next; 2481 } 2482 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2483 tmpm = entry->object.sub_map; 2484 tmpe = tmpm->header.next; 2485 entcount = tmpm->nentries; 2486 while (entcount-- && tmpe != &tmpm->header) { 2487 if (_vm_object_in_map(tmpm, object, tmpe)) { 2488 return 1; 2489 } 2490 tmpe = tmpe->next; 2491 } 2492 } else if ((obj = entry->object.vm_object) != NULL) { 2493 for (; obj; obj = obj->backing_object) 2494 if (obj == object) { 2495 return 1; 2496 } 2497 } 2498 return 0; 2499 } 2500 2501 static int 2502 vm_object_in_map(vm_object_t object) 2503 { 2504 struct proc *p; 2505 2506 /* sx_slock(&allproc_lock); */ 2507 FOREACH_PROC_IN_SYSTEM(p) { 2508 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2509 continue; 2510 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2511 /* sx_sunlock(&allproc_lock); */ 2512 return 1; 2513 } 2514 } 2515 /* sx_sunlock(&allproc_lock); */ 2516 if (_vm_object_in_map(kernel_map, object, 0)) 2517 return 1; 2518 return 0; 2519 } 2520 2521 DB_SHOW_COMMAND(vmochk, vm_object_check) 2522 { 2523 vm_object_t object; 2524 2525 /* 2526 * make sure that internal objs are in a map somewhere 2527 * and none have zero ref counts. 2528 */ 2529 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2530 if (object->handle == NULL && 2531 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2532 if (object->ref_count == 0) { 2533 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2534 (long)object->size); 2535 } 2536 if (!vm_object_in_map(object)) { 2537 db_printf( 2538 "vmochk: internal obj is not in a map: " 2539 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2540 object->ref_count, (u_long)object->size, 2541 (u_long)object->size, 2542 (void *)object->backing_object); 2543 } 2544 } 2545 } 2546 } 2547 2548 /* 2549 * vm_object_print: [ debug ] 2550 */ 2551 DB_SHOW_COMMAND(object, vm_object_print_static) 2552 { 2553 /* XXX convert args. */ 2554 vm_object_t object = (vm_object_t)addr; 2555 boolean_t full = have_addr; 2556 2557 vm_page_t p; 2558 2559 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2560 #define count was_count 2561 2562 int count; 2563 2564 if (object == NULL) 2565 return; 2566 2567 db_iprintf( 2568 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2569 object, (int)object->type, (uintmax_t)object->size, 2570 object->resident_page_count, object->ref_count, object->flags, 2571 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2572 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2573 object->shadow_count, 2574 object->backing_object ? object->backing_object->ref_count : 0, 2575 object->backing_object, (uintmax_t)object->backing_object_offset); 2576 2577 if (!full) 2578 return; 2579 2580 db_indent += 2; 2581 count = 0; 2582 TAILQ_FOREACH(p, &object->memq, listq) { 2583 if (count == 0) 2584 db_iprintf("memory:="); 2585 else if (count == 6) { 2586 db_printf("\n"); 2587 db_iprintf(" ..."); 2588 count = 0; 2589 } else 2590 db_printf(","); 2591 count++; 2592 2593 db_printf("(off=0x%jx,page=0x%jx)", 2594 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2595 } 2596 if (count != 0) 2597 db_printf("\n"); 2598 db_indent -= 2; 2599 } 2600 2601 /* XXX. */ 2602 #undef count 2603 2604 /* XXX need this non-static entry for calling from vm_map_print. */ 2605 void 2606 vm_object_print( 2607 /* db_expr_t */ long addr, 2608 boolean_t have_addr, 2609 /* db_expr_t */ long count, 2610 char *modif) 2611 { 2612 vm_object_print_static(addr, have_addr, count, modif); 2613 } 2614 2615 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2616 { 2617 vm_object_t object; 2618 vm_pindex_t fidx; 2619 vm_paddr_t pa; 2620 vm_page_t m, prev_m; 2621 int rcount, nl, c; 2622 2623 nl = 0; 2624 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2625 db_printf("new object: %p\n", (void *)object); 2626 if (nl > 18) { 2627 c = cngetc(); 2628 if (c != ' ') 2629 return; 2630 nl = 0; 2631 } 2632 nl++; 2633 rcount = 0; 2634 fidx = 0; 2635 pa = -1; 2636 TAILQ_FOREACH(m, &object->memq, listq) { 2637 if (m->pindex > 128) 2638 break; 2639 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2640 prev_m->pindex + 1 != m->pindex) { 2641 if (rcount) { 2642 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2643 (long)fidx, rcount, (long)pa); 2644 if (nl > 18) { 2645 c = cngetc(); 2646 if (c != ' ') 2647 return; 2648 nl = 0; 2649 } 2650 nl++; 2651 rcount = 0; 2652 } 2653 } 2654 if (rcount && 2655 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2656 ++rcount; 2657 continue; 2658 } 2659 if (rcount) { 2660 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2661 (long)fidx, rcount, (long)pa); 2662 if (nl > 18) { 2663 c = cngetc(); 2664 if (c != ' ') 2665 return; 2666 nl = 0; 2667 } 2668 nl++; 2669 } 2670 fidx = m->pindex; 2671 pa = VM_PAGE_TO_PHYS(m); 2672 rcount = 1; 2673 } 2674 if (rcount) { 2675 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2676 (long)fidx, rcount, (long)pa); 2677 if (nl > 18) { 2678 c = cngetc(); 2679 if (c != ' ') 2680 return; 2681 nl = 0; 2682 } 2683 nl++; 2684 } 2685 } 2686 } 2687 #endif /* DDB */ 2688