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/lock.h> 75 #include <sys/mman.h> 76 #include <sys/mount.h> 77 #include <sys/kernel.h> 78 #include <sys/pctrie.h> 79 #include <sys/sysctl.h> 80 #include <sys/mutex.h> 81 #include <sys/proc.h> /* for curproc, pageproc */ 82 #include <sys/socket.h> 83 #include <sys/resourcevar.h> 84 #include <sys/rwlock.h> 85 #include <sys/user.h> 86 #include <sys/vnode.h> 87 #include <sys/vmmeter.h> 88 #include <sys/sx.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_param.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_object.h> 95 #include <vm/vm_page.h> 96 #include <vm/vm_pageout.h> 97 #include <vm/vm_pager.h> 98 #include <vm/swap_pager.h> 99 #include <vm/vm_kern.h> 100 #include <vm/vm_extern.h> 101 #include <vm/vm_radix.h> 102 #include <vm/vm_reserv.h> 103 #include <vm/uma.h> 104 105 static int old_msync; 106 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 107 "Use old (insecure) msync behavior"); 108 109 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 110 int pagerflags, int flags, boolean_t *clearobjflags, 111 boolean_t *eio); 112 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 113 boolean_t *clearobjflags); 114 static void vm_object_qcollapse(vm_object_t object); 115 static void vm_object_vndeallocate(vm_object_t object); 116 117 /* 118 * Virtual memory objects maintain the actual data 119 * associated with allocated virtual memory. A given 120 * page of memory exists within exactly one object. 121 * 122 * An object is only deallocated when all "references" 123 * are given up. Only one "reference" to a given 124 * region of an object should be writeable. 125 * 126 * Associated with each object is a list of all resident 127 * memory pages belonging to that object; this list is 128 * maintained by the "vm_page" module, and locked by the object's 129 * lock. 130 * 131 * Each object also records a "pager" routine which is 132 * used to retrieve (and store) pages to the proper backing 133 * storage. In addition, objects may be backed by other 134 * objects from which they were virtual-copied. 135 * 136 * The only items within the object structure which are 137 * modified after time of creation are: 138 * reference count locked by object's lock 139 * pager routine locked by object's lock 140 * 141 */ 142 143 struct object_q vm_object_list; 144 struct mtx vm_object_list_mtx; /* lock for object list and count */ 145 146 struct vm_object kernel_object_store; 147 148 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, 149 "VM object stats"); 150 151 static long object_collapses; 152 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 153 &object_collapses, 0, "VM object collapses"); 154 155 static long object_bypasses; 156 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 157 &object_bypasses, 0, "VM object bypasses"); 158 159 static uma_zone_t obj_zone; 160 161 static int vm_object_zinit(void *mem, int size, int flags); 162 163 #ifdef INVARIANTS 164 static void vm_object_zdtor(void *mem, int size, void *arg); 165 166 static void 167 vm_object_zdtor(void *mem, int size, void *arg) 168 { 169 vm_object_t object; 170 171 object = (vm_object_t)mem; 172 KASSERT(object->ref_count == 0, 173 ("object %p ref_count = %d", object, object->ref_count)); 174 KASSERT(TAILQ_EMPTY(&object->memq), 175 ("object %p has resident pages in its memq", object)); 176 KASSERT(vm_radix_is_empty(&object->rtree), 177 ("object %p has resident pages in its trie", object)); 178 #if VM_NRESERVLEVEL > 0 179 KASSERT(LIST_EMPTY(&object->rvq), 180 ("object %p has reservations", 181 object)); 182 #endif 183 KASSERT(object->paging_in_progress == 0, 184 ("object %p paging_in_progress = %d", 185 object, object->paging_in_progress)); 186 KASSERT(object->resident_page_count == 0, 187 ("object %p resident_page_count = %d", 188 object, object->resident_page_count)); 189 KASSERT(object->shadow_count == 0, 190 ("object %p shadow_count = %d", 191 object, object->shadow_count)); 192 KASSERT(object->type == OBJT_DEAD, 193 ("object %p has non-dead type %d", 194 object, object->type)); 195 } 196 #endif 197 198 static int 199 vm_object_zinit(void *mem, int size, int flags) 200 { 201 vm_object_t object; 202 203 object = (vm_object_t)mem; 204 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 205 206 /* These are true for any object that has been freed */ 207 object->type = OBJT_DEAD; 208 object->ref_count = 0; 209 vm_radix_init(&object->rtree); 210 object->paging_in_progress = 0; 211 object->resident_page_count = 0; 212 object->shadow_count = 0; 213 object->flags = OBJ_DEAD; 214 215 mtx_lock(&vm_object_list_mtx); 216 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 217 mtx_unlock(&vm_object_list_mtx); 218 return (0); 219 } 220 221 static void 222 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 223 { 224 225 TAILQ_INIT(&object->memq); 226 LIST_INIT(&object->shadow_head); 227 228 object->type = type; 229 if (type == OBJT_SWAP) 230 pctrie_init(&object->un_pager.swp.swp_blks); 231 232 /* 233 * Ensure that swap_pager_swapoff() iteration over object_list 234 * sees up to date type and pctrie head if it observed 235 * non-dead object. 236 */ 237 atomic_thread_fence_rel(); 238 239 switch (type) { 240 case OBJT_DEAD: 241 panic("_vm_object_allocate: can't create OBJT_DEAD"); 242 case OBJT_DEFAULT: 243 case OBJT_SWAP: 244 object->flags = OBJ_ONEMAPPING; 245 break; 246 case OBJT_DEVICE: 247 case OBJT_SG: 248 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 249 break; 250 case OBJT_MGTDEVICE: 251 object->flags = OBJ_FICTITIOUS; 252 break; 253 case OBJT_PHYS: 254 object->flags = OBJ_UNMANAGED; 255 break; 256 case OBJT_VNODE: 257 object->flags = 0; 258 break; 259 default: 260 panic("_vm_object_allocate: type %d is undefined", type); 261 } 262 object->size = size; 263 object->generation = 1; 264 object->ref_count = 1; 265 object->memattr = VM_MEMATTR_DEFAULT; 266 object->cred = NULL; 267 object->charge = 0; 268 object->handle = NULL; 269 object->backing_object = NULL; 270 object->backing_object_offset = (vm_ooffset_t) 0; 271 #if VM_NRESERVLEVEL > 0 272 LIST_INIT(&object->rvq); 273 #endif 274 umtx_shm_object_init(object); 275 } 276 277 /* 278 * vm_object_init: 279 * 280 * Initialize the VM objects module. 281 */ 282 void 283 vm_object_init(void) 284 { 285 TAILQ_INIT(&vm_object_list); 286 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 287 288 rw_init(&kernel_object->lock, "kernel vm object"); 289 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 290 VM_MIN_KERNEL_ADDRESS), kernel_object); 291 #if VM_NRESERVLEVEL > 0 292 kernel_object->flags |= OBJ_COLORED; 293 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 294 #endif 295 296 /* 297 * The lock portion of struct vm_object must be type stable due 298 * to vm_pageout_fallback_object_lock locking a vm object 299 * without holding any references to it. 300 */ 301 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 302 #ifdef INVARIANTS 303 vm_object_zdtor, 304 #else 305 NULL, 306 #endif 307 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 308 309 vm_radix_zinit(); 310 } 311 312 void 313 vm_object_clear_flag(vm_object_t object, u_short bits) 314 { 315 316 VM_OBJECT_ASSERT_WLOCKED(object); 317 object->flags &= ~bits; 318 } 319 320 /* 321 * Sets the default memory attribute for the specified object. Pages 322 * that are allocated to this object are by default assigned this memory 323 * attribute. 324 * 325 * Presently, this function must be called before any pages are allocated 326 * to the object. In the future, this requirement may be relaxed for 327 * "default" and "swap" objects. 328 */ 329 int 330 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 331 { 332 333 VM_OBJECT_ASSERT_WLOCKED(object); 334 switch (object->type) { 335 case OBJT_DEFAULT: 336 case OBJT_DEVICE: 337 case OBJT_MGTDEVICE: 338 case OBJT_PHYS: 339 case OBJT_SG: 340 case OBJT_SWAP: 341 case OBJT_VNODE: 342 if (!TAILQ_EMPTY(&object->memq)) 343 return (KERN_FAILURE); 344 break; 345 case OBJT_DEAD: 346 return (KERN_INVALID_ARGUMENT); 347 default: 348 panic("vm_object_set_memattr: object %p is of undefined type", 349 object); 350 } 351 object->memattr = memattr; 352 return (KERN_SUCCESS); 353 } 354 355 void 356 vm_object_pip_add(vm_object_t object, short i) 357 { 358 359 VM_OBJECT_ASSERT_WLOCKED(object); 360 object->paging_in_progress += i; 361 } 362 363 void 364 vm_object_pip_subtract(vm_object_t object, short i) 365 { 366 367 VM_OBJECT_ASSERT_WLOCKED(object); 368 object->paging_in_progress -= i; 369 } 370 371 void 372 vm_object_pip_wakeup(vm_object_t object) 373 { 374 375 VM_OBJECT_ASSERT_WLOCKED(object); 376 object->paging_in_progress--; 377 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 378 vm_object_clear_flag(object, OBJ_PIPWNT); 379 wakeup(object); 380 } 381 } 382 383 void 384 vm_object_pip_wakeupn(vm_object_t object, short i) 385 { 386 387 VM_OBJECT_ASSERT_WLOCKED(object); 388 if (i) 389 object->paging_in_progress -= i; 390 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 391 vm_object_clear_flag(object, OBJ_PIPWNT); 392 wakeup(object); 393 } 394 } 395 396 void 397 vm_object_pip_wait(vm_object_t object, char *waitid) 398 { 399 400 VM_OBJECT_ASSERT_WLOCKED(object); 401 while (object->paging_in_progress) { 402 object->flags |= OBJ_PIPWNT; 403 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0); 404 } 405 } 406 407 /* 408 * vm_object_allocate: 409 * 410 * Returns a new object with the given size. 411 */ 412 vm_object_t 413 vm_object_allocate(objtype_t type, vm_pindex_t size) 414 { 415 vm_object_t object; 416 417 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 418 _vm_object_allocate(type, size, object); 419 return (object); 420 } 421 422 423 /* 424 * vm_object_reference: 425 * 426 * Gets another reference to the given object. Note: OBJ_DEAD 427 * objects can be referenced during final cleaning. 428 */ 429 void 430 vm_object_reference(vm_object_t object) 431 { 432 if (object == NULL) 433 return; 434 VM_OBJECT_WLOCK(object); 435 vm_object_reference_locked(object); 436 VM_OBJECT_WUNLOCK(object); 437 } 438 439 /* 440 * vm_object_reference_locked: 441 * 442 * Gets another reference to the given object. 443 * 444 * The object must be locked. 445 */ 446 void 447 vm_object_reference_locked(vm_object_t object) 448 { 449 struct vnode *vp; 450 451 VM_OBJECT_ASSERT_WLOCKED(object); 452 object->ref_count++; 453 if (object->type == OBJT_VNODE) { 454 vp = object->handle; 455 vref(vp); 456 } 457 } 458 459 /* 460 * Handle deallocating an object of type OBJT_VNODE. 461 */ 462 static void 463 vm_object_vndeallocate(vm_object_t object) 464 { 465 struct vnode *vp = (struct vnode *) object->handle; 466 467 VM_OBJECT_ASSERT_WLOCKED(object); 468 KASSERT(object->type == OBJT_VNODE, 469 ("vm_object_vndeallocate: not a vnode object")); 470 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 471 #ifdef INVARIANTS 472 if (object->ref_count == 0) { 473 vn_printf(vp, "vm_object_vndeallocate "); 474 panic("vm_object_vndeallocate: bad object reference count"); 475 } 476 #endif 477 478 if (!umtx_shm_vnobj_persistent && object->ref_count == 1) 479 umtx_shm_object_terminated(object); 480 481 /* 482 * The test for text of vp vnode does not need a bypass to 483 * reach right VV_TEXT there, since it is obtained from 484 * object->handle. 485 */ 486 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) { 487 object->ref_count--; 488 VM_OBJECT_WUNLOCK(object); 489 /* vrele may need the vnode lock. */ 490 vrele(vp); 491 } else { 492 vhold(vp); 493 VM_OBJECT_WUNLOCK(object); 494 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 495 vdrop(vp); 496 VM_OBJECT_WLOCK(object); 497 object->ref_count--; 498 if (object->type == OBJT_DEAD) { 499 VM_OBJECT_WUNLOCK(object); 500 VOP_UNLOCK(vp, 0); 501 } else { 502 if (object->ref_count == 0) 503 VOP_UNSET_TEXT(vp); 504 VM_OBJECT_WUNLOCK(object); 505 vput(vp); 506 } 507 } 508 } 509 510 /* 511 * vm_object_deallocate: 512 * 513 * Release a reference to the specified object, 514 * gained either through a vm_object_allocate 515 * or a vm_object_reference call. When all references 516 * are gone, storage associated with this object 517 * may be relinquished. 518 * 519 * No object may be locked. 520 */ 521 void 522 vm_object_deallocate(vm_object_t object) 523 { 524 vm_object_t temp; 525 struct vnode *vp; 526 527 while (object != NULL) { 528 VM_OBJECT_WLOCK(object); 529 if (object->type == OBJT_VNODE) { 530 vm_object_vndeallocate(object); 531 return; 532 } 533 534 KASSERT(object->ref_count != 0, 535 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 536 537 /* 538 * If the reference count goes to 0 we start calling 539 * vm_object_terminate() on the object chain. 540 * A ref count of 1 may be a special case depending on the 541 * shadow count being 0 or 1. 542 */ 543 object->ref_count--; 544 if (object->ref_count > 1) { 545 VM_OBJECT_WUNLOCK(object); 546 return; 547 } else if (object->ref_count == 1) { 548 if (object->type == OBJT_SWAP && 549 (object->flags & OBJ_TMPFS) != 0) { 550 vp = object->un_pager.swp.swp_tmpfs; 551 vhold(vp); 552 VM_OBJECT_WUNLOCK(object); 553 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 554 VM_OBJECT_WLOCK(object); 555 if (object->type == OBJT_DEAD || 556 object->ref_count != 1) { 557 VM_OBJECT_WUNLOCK(object); 558 VOP_UNLOCK(vp, 0); 559 vdrop(vp); 560 return; 561 } 562 if ((object->flags & OBJ_TMPFS) != 0) 563 VOP_UNSET_TEXT(vp); 564 VOP_UNLOCK(vp, 0); 565 vdrop(vp); 566 } 567 if (object->shadow_count == 0 && 568 object->handle == NULL && 569 (object->type == OBJT_DEFAULT || 570 (object->type == OBJT_SWAP && 571 (object->flags & OBJ_TMPFS_NODE) == 0))) { 572 vm_object_set_flag(object, OBJ_ONEMAPPING); 573 } else if ((object->shadow_count == 1) && 574 (object->handle == NULL) && 575 (object->type == OBJT_DEFAULT || 576 object->type == OBJT_SWAP)) { 577 vm_object_t robject; 578 579 robject = LIST_FIRST(&object->shadow_head); 580 KASSERT(robject != NULL, 581 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 582 object->ref_count, 583 object->shadow_count)); 584 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, 585 ("shadowed tmpfs v_object %p", object)); 586 if (!VM_OBJECT_TRYWLOCK(robject)) { 587 /* 588 * Avoid a potential deadlock. 589 */ 590 object->ref_count++; 591 VM_OBJECT_WUNLOCK(object); 592 /* 593 * More likely than not the thread 594 * holding robject's lock has lower 595 * priority than the current thread. 596 * Let the lower priority thread run. 597 */ 598 pause("vmo_de", 1); 599 continue; 600 } 601 /* 602 * Collapse object into its shadow unless its 603 * shadow is dead. In that case, object will 604 * be deallocated by the thread that is 605 * deallocating its shadow. 606 */ 607 if ((robject->flags & OBJ_DEAD) == 0 && 608 (robject->handle == NULL) && 609 (robject->type == OBJT_DEFAULT || 610 robject->type == OBJT_SWAP)) { 611 612 robject->ref_count++; 613 retry: 614 if (robject->paging_in_progress) { 615 VM_OBJECT_WUNLOCK(object); 616 vm_object_pip_wait(robject, 617 "objde1"); 618 temp = robject->backing_object; 619 if (object == temp) { 620 VM_OBJECT_WLOCK(object); 621 goto retry; 622 } 623 } else if (object->paging_in_progress) { 624 VM_OBJECT_WUNLOCK(robject); 625 object->flags |= OBJ_PIPWNT; 626 VM_OBJECT_SLEEP(object, object, 627 PDROP | PVM, "objde2", 0); 628 VM_OBJECT_WLOCK(robject); 629 temp = robject->backing_object; 630 if (object == temp) { 631 VM_OBJECT_WLOCK(object); 632 goto retry; 633 } 634 } else 635 VM_OBJECT_WUNLOCK(object); 636 637 if (robject->ref_count == 1) { 638 robject->ref_count--; 639 object = robject; 640 goto doterm; 641 } 642 object = robject; 643 vm_object_collapse(object); 644 VM_OBJECT_WUNLOCK(object); 645 continue; 646 } 647 VM_OBJECT_WUNLOCK(robject); 648 } 649 VM_OBJECT_WUNLOCK(object); 650 return; 651 } 652 doterm: 653 umtx_shm_object_terminated(object); 654 temp = object->backing_object; 655 if (temp != NULL) { 656 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 657 ("shadowed tmpfs v_object 2 %p", object)); 658 VM_OBJECT_WLOCK(temp); 659 LIST_REMOVE(object, shadow_list); 660 temp->shadow_count--; 661 VM_OBJECT_WUNLOCK(temp); 662 object->backing_object = NULL; 663 } 664 /* 665 * Don't double-terminate, we could be in a termination 666 * recursion due to the terminate having to sync data 667 * to disk. 668 */ 669 if ((object->flags & OBJ_DEAD) == 0) 670 vm_object_terminate(object); 671 else 672 VM_OBJECT_WUNLOCK(object); 673 object = temp; 674 } 675 } 676 677 /* 678 * vm_object_destroy removes the object from the global object list 679 * and frees the space for the object. 680 */ 681 void 682 vm_object_destroy(vm_object_t object) 683 { 684 685 /* 686 * Release the allocation charge. 687 */ 688 if (object->cred != NULL) { 689 swap_release_by_cred(object->charge, object->cred); 690 object->charge = 0; 691 crfree(object->cred); 692 object->cred = NULL; 693 } 694 695 /* 696 * Free the space for the object. 697 */ 698 uma_zfree(obj_zone, object); 699 } 700 701 /* 702 * vm_object_terminate_pages removes any remaining pageable pages 703 * from the object and resets the object to an empty state. 704 */ 705 static void 706 vm_object_terminate_pages(vm_object_t object) 707 { 708 vm_page_t p, p_next; 709 struct mtx *mtx, *mtx1; 710 struct vm_pagequeue *pq, *pq1; 711 int dequeued; 712 713 VM_OBJECT_ASSERT_WLOCKED(object); 714 715 mtx = NULL; 716 pq = NULL; 717 718 /* 719 * Free any remaining pageable pages. This also removes them from the 720 * paging queues. However, don't free wired pages, just remove them 721 * from the object. Rather than incrementally removing each page from 722 * the object, the page and object are reset to any empty state. 723 */ 724 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 725 vm_page_assert_unbusied(p); 726 if ((object->flags & OBJ_UNMANAGED) == 0) { 727 /* 728 * vm_page_free_prep() only needs the page 729 * lock for managed pages. 730 */ 731 mtx1 = vm_page_lockptr(p); 732 if (mtx1 != mtx) { 733 if (mtx != NULL) 734 mtx_unlock(mtx); 735 if (pq != NULL) { 736 vm_pagequeue_cnt_add(pq, dequeued); 737 vm_pagequeue_unlock(pq); 738 pq = NULL; 739 } 740 mtx = mtx1; 741 mtx_lock(mtx); 742 } 743 } 744 p->object = NULL; 745 if (p->wire_count != 0) 746 goto unlist; 747 VM_CNT_INC(v_pfree); 748 p->flags &= ~PG_ZERO; 749 if (p->queue != PQ_NONE) { 750 KASSERT(p->queue < PQ_COUNT, ("vm_object_terminate: " 751 "page %p is not queued", p)); 752 pq1 = vm_page_pagequeue(p); 753 if (pq != pq1) { 754 if (pq != NULL) { 755 vm_pagequeue_cnt_add(pq, dequeued); 756 vm_pagequeue_unlock(pq); 757 } 758 pq = pq1; 759 vm_pagequeue_lock(pq); 760 dequeued = 0; 761 } 762 p->queue = PQ_NONE; 763 TAILQ_REMOVE(&pq->pq_pl, p, plinks.q); 764 dequeued--; 765 } 766 if (vm_page_free_prep(p, true)) 767 continue; 768 unlist: 769 TAILQ_REMOVE(&object->memq, p, listq); 770 } 771 if (pq != NULL) { 772 vm_pagequeue_cnt_add(pq, dequeued); 773 vm_pagequeue_unlock(pq); 774 } 775 if (mtx != NULL) 776 mtx_unlock(mtx); 777 778 vm_page_free_phys_pglist(&object->memq); 779 780 /* 781 * If the object contained any pages, then reset it to an empty state. 782 * None of the object's fields, including "resident_page_count", were 783 * modified by the preceding loop. 784 */ 785 if (object->resident_page_count != 0) { 786 vm_radix_reclaim_allnodes(&object->rtree); 787 TAILQ_INIT(&object->memq); 788 object->resident_page_count = 0; 789 if (object->type == OBJT_VNODE) 790 vdrop(object->handle); 791 } 792 } 793 794 /* 795 * vm_object_terminate actually destroys the specified object, freeing 796 * up all previously used resources. 797 * 798 * The object must be locked. 799 * This routine may block. 800 */ 801 void 802 vm_object_terminate(vm_object_t object) 803 { 804 805 VM_OBJECT_ASSERT_WLOCKED(object); 806 807 /* 808 * Make sure no one uses us. 809 */ 810 vm_object_set_flag(object, OBJ_DEAD); 811 812 /* 813 * wait for the pageout daemon to be done with the object 814 */ 815 vm_object_pip_wait(object, "objtrm"); 816 817 KASSERT(!object->paging_in_progress, 818 ("vm_object_terminate: pageout in progress")); 819 820 /* 821 * Clean and free the pages, as appropriate. All references to the 822 * object are gone, so we don't need to lock it. 823 */ 824 if (object->type == OBJT_VNODE) { 825 struct vnode *vp = (struct vnode *)object->handle; 826 827 /* 828 * Clean pages and flush buffers. 829 */ 830 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 831 VM_OBJECT_WUNLOCK(object); 832 833 vinvalbuf(vp, V_SAVE, 0, 0); 834 835 BO_LOCK(&vp->v_bufobj); 836 vp->v_bufobj.bo_flag |= BO_DEAD; 837 BO_UNLOCK(&vp->v_bufobj); 838 839 VM_OBJECT_WLOCK(object); 840 } 841 842 KASSERT(object->ref_count == 0, 843 ("vm_object_terminate: object with references, ref_count=%d", 844 object->ref_count)); 845 846 if ((object->flags & OBJ_PG_DTOR) == 0) 847 vm_object_terminate_pages(object); 848 849 #if VM_NRESERVLEVEL > 0 850 if (__predict_false(!LIST_EMPTY(&object->rvq))) 851 vm_reserv_break_all(object); 852 #endif 853 854 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 855 object->type == OBJT_SWAP, 856 ("%s: non-swap obj %p has cred", __func__, object)); 857 858 /* 859 * Let the pager know object is dead. 860 */ 861 vm_pager_deallocate(object); 862 VM_OBJECT_WUNLOCK(object); 863 864 vm_object_destroy(object); 865 } 866 867 /* 868 * Make the page read-only so that we can clear the object flags. However, if 869 * this is a nosync mmap then the object is likely to stay dirty so do not 870 * mess with the page and do not clear the object flags. Returns TRUE if the 871 * page should be flushed, and FALSE otherwise. 872 */ 873 static boolean_t 874 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags) 875 { 876 877 /* 878 * If we have been asked to skip nosync pages and this is a 879 * nosync page, skip it. Note that the object flags were not 880 * cleared in this case so we do not have to set them. 881 */ 882 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { 883 *clearobjflags = FALSE; 884 return (FALSE); 885 } else { 886 pmap_remove_write(p); 887 return (p->dirty != 0); 888 } 889 } 890 891 /* 892 * vm_object_page_clean 893 * 894 * Clean all dirty pages in the specified range of object. Leaves page 895 * on whatever queue it is currently on. If NOSYNC is set then do not 896 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 897 * leaving the object dirty. 898 * 899 * When stuffing pages asynchronously, allow clustering. XXX we need a 900 * synchronous clustering mode implementation. 901 * 902 * Odd semantics: if start == end, we clean everything. 903 * 904 * The object must be locked. 905 * 906 * Returns FALSE if some page from the range was not written, as 907 * reported by the pager, and TRUE otherwise. 908 */ 909 boolean_t 910 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 911 int flags) 912 { 913 vm_page_t np, p; 914 vm_pindex_t pi, tend, tstart; 915 int curgeneration, n, pagerflags; 916 boolean_t clearobjflags, eio, res; 917 918 VM_OBJECT_ASSERT_WLOCKED(object); 919 920 /* 921 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE 922 * objects. The check below prevents the function from 923 * operating on non-vnode objects. 924 */ 925 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || 926 object->resident_page_count == 0) 927 return (TRUE); 928 929 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 930 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 931 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 932 933 tstart = OFF_TO_IDX(start); 934 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 935 clearobjflags = tstart == 0 && tend >= object->size; 936 res = TRUE; 937 938 rescan: 939 curgeneration = object->generation; 940 941 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 942 pi = p->pindex; 943 if (pi >= tend) 944 break; 945 np = TAILQ_NEXT(p, listq); 946 if (p->valid == 0) 947 continue; 948 if (vm_page_sleep_if_busy(p, "vpcwai")) { 949 if (object->generation != curgeneration) { 950 if ((flags & OBJPC_SYNC) != 0) 951 goto rescan; 952 else 953 clearobjflags = FALSE; 954 } 955 np = vm_page_find_least(object, pi); 956 continue; 957 } 958 if (!vm_object_page_remove_write(p, flags, &clearobjflags)) 959 continue; 960 961 n = vm_object_page_collect_flush(object, p, pagerflags, 962 flags, &clearobjflags, &eio); 963 if (eio) { 964 res = FALSE; 965 clearobjflags = FALSE; 966 } 967 if (object->generation != curgeneration) { 968 if ((flags & OBJPC_SYNC) != 0) 969 goto rescan; 970 else 971 clearobjflags = FALSE; 972 } 973 974 /* 975 * If the VOP_PUTPAGES() did a truncated write, so 976 * that even the first page of the run is not fully 977 * written, vm_pageout_flush() returns 0 as the run 978 * length. Since the condition that caused truncated 979 * write may be permanent, e.g. exhausted free space, 980 * accepting n == 0 would cause an infinite loop. 981 * 982 * Forwarding the iterator leaves the unwritten page 983 * behind, but there is not much we can do there if 984 * filesystem refuses to write it. 985 */ 986 if (n == 0) { 987 n = 1; 988 clearobjflags = FALSE; 989 } 990 np = vm_page_find_least(object, pi + n); 991 } 992 #if 0 993 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 994 #endif 995 996 if (clearobjflags) 997 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 998 return (res); 999 } 1000 1001 static int 1002 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1003 int flags, boolean_t *clearobjflags, boolean_t *eio) 1004 { 1005 vm_page_t ma[vm_pageout_page_count], p_first, tp; 1006 int count, i, mreq, runlen; 1007 1008 vm_page_lock_assert(p, MA_NOTOWNED); 1009 VM_OBJECT_ASSERT_WLOCKED(object); 1010 1011 count = 1; 1012 mreq = 0; 1013 1014 for (tp = p; count < vm_pageout_page_count; count++) { 1015 tp = vm_page_next(tp); 1016 if (tp == NULL || vm_page_busied(tp)) 1017 break; 1018 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1019 break; 1020 } 1021 1022 for (p_first = p; count < vm_pageout_page_count; count++) { 1023 tp = vm_page_prev(p_first); 1024 if (tp == NULL || vm_page_busied(tp)) 1025 break; 1026 if (!vm_object_page_remove_write(tp, flags, clearobjflags)) 1027 break; 1028 p_first = tp; 1029 mreq++; 1030 } 1031 1032 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 1033 ma[i] = tp; 1034 1035 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 1036 return (runlen); 1037 } 1038 1039 /* 1040 * Note that there is absolutely no sense in writing out 1041 * anonymous objects, so we track down the vnode object 1042 * to write out. 1043 * We invalidate (remove) all pages from the address space 1044 * for semantic correctness. 1045 * 1046 * If the backing object is a device object with unmanaged pages, then any 1047 * mappings to the specified range of pages must be removed before this 1048 * function is called. 1049 * 1050 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1051 * may start out with a NULL object. 1052 */ 1053 boolean_t 1054 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1055 boolean_t syncio, boolean_t invalidate) 1056 { 1057 vm_object_t backing_object; 1058 struct vnode *vp; 1059 struct mount *mp; 1060 int error, flags, fsync_after; 1061 boolean_t res; 1062 1063 if (object == NULL) 1064 return (TRUE); 1065 res = TRUE; 1066 error = 0; 1067 VM_OBJECT_WLOCK(object); 1068 while ((backing_object = object->backing_object) != NULL) { 1069 VM_OBJECT_WLOCK(backing_object); 1070 offset += object->backing_object_offset; 1071 VM_OBJECT_WUNLOCK(object); 1072 object = backing_object; 1073 if (object->size < OFF_TO_IDX(offset + size)) 1074 size = IDX_TO_OFF(object->size) - offset; 1075 } 1076 /* 1077 * Flush pages if writing is allowed, invalidate them 1078 * if invalidation requested. Pages undergoing I/O 1079 * will be ignored by vm_object_page_remove(). 1080 * 1081 * We cannot lock the vnode and then wait for paging 1082 * to complete without deadlocking against vm_fault. 1083 * Instead we simply call vm_object_page_remove() and 1084 * allow it to block internally on a page-by-page 1085 * basis when it encounters pages undergoing async 1086 * I/O. 1087 */ 1088 if (object->type == OBJT_VNODE && 1089 (object->flags & OBJ_MIGHTBEDIRTY) != 0 && 1090 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1091 VM_OBJECT_WUNLOCK(object); 1092 (void) vn_start_write(vp, &mp, V_WAIT); 1093 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1094 if (syncio && !invalidate && offset == 0 && 1095 atop(size) == object->size) { 1096 /* 1097 * If syncing the whole mapping of the file, 1098 * it is faster to schedule all the writes in 1099 * async mode, also allowing the clustering, 1100 * and then wait for i/o to complete. 1101 */ 1102 flags = 0; 1103 fsync_after = TRUE; 1104 } else { 1105 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1106 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1107 fsync_after = FALSE; 1108 } 1109 VM_OBJECT_WLOCK(object); 1110 res = vm_object_page_clean(object, offset, offset + size, 1111 flags); 1112 VM_OBJECT_WUNLOCK(object); 1113 if (fsync_after) 1114 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1115 VOP_UNLOCK(vp, 0); 1116 vn_finished_write(mp); 1117 if (error != 0) 1118 res = FALSE; 1119 VM_OBJECT_WLOCK(object); 1120 } 1121 if ((object->type == OBJT_VNODE || 1122 object->type == OBJT_DEVICE) && invalidate) { 1123 if (object->type == OBJT_DEVICE) 1124 /* 1125 * The option OBJPR_NOTMAPPED must be passed here 1126 * because vm_object_page_remove() cannot remove 1127 * unmanaged mappings. 1128 */ 1129 flags = OBJPR_NOTMAPPED; 1130 else if (old_msync) 1131 flags = 0; 1132 else 1133 flags = OBJPR_CLEANONLY; 1134 vm_object_page_remove(object, OFF_TO_IDX(offset), 1135 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1136 } 1137 VM_OBJECT_WUNLOCK(object); 1138 return (res); 1139 } 1140 1141 /* 1142 * Determine whether the given advice can be applied to the object. Advice is 1143 * not applied to unmanaged pages since they never belong to page queues, and 1144 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1145 * have been mapped at most once. 1146 */ 1147 static bool 1148 vm_object_advice_applies(vm_object_t object, int advice) 1149 { 1150 1151 if ((object->flags & OBJ_UNMANAGED) != 0) 1152 return (false); 1153 if (advice != MADV_FREE) 1154 return (true); 1155 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) && 1156 (object->flags & OBJ_ONEMAPPING) != 0); 1157 } 1158 1159 static void 1160 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1161 vm_size_t size) 1162 { 1163 1164 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1165 swap_pager_freespace(object, pindex, size); 1166 } 1167 1168 /* 1169 * vm_object_madvise: 1170 * 1171 * Implements the madvise function at the object/page level. 1172 * 1173 * MADV_WILLNEED (any object) 1174 * 1175 * Activate the specified pages if they are resident. 1176 * 1177 * MADV_DONTNEED (any object) 1178 * 1179 * Deactivate the specified pages if they are resident. 1180 * 1181 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1182 * OBJ_ONEMAPPING only) 1183 * 1184 * Deactivate and clean the specified pages if they are 1185 * resident. This permits the process to reuse the pages 1186 * without faulting or the kernel to reclaim the pages 1187 * without I/O. 1188 */ 1189 void 1190 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1191 int advice) 1192 { 1193 vm_pindex_t tpindex; 1194 vm_object_t backing_object, tobject; 1195 vm_page_t m, tm; 1196 1197 if (object == NULL) 1198 return; 1199 1200 relookup: 1201 VM_OBJECT_WLOCK(object); 1202 if (!vm_object_advice_applies(object, advice)) { 1203 VM_OBJECT_WUNLOCK(object); 1204 return; 1205 } 1206 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1207 tobject = object; 1208 1209 /* 1210 * If the next page isn't resident in the top-level object, we 1211 * need to search the shadow chain. When applying MADV_FREE, we 1212 * take care to release any swap space used to store 1213 * non-resident pages. 1214 */ 1215 if (m == NULL || pindex < m->pindex) { 1216 /* 1217 * Optimize a common case: if the top-level object has 1218 * no backing object, we can skip over the non-resident 1219 * range in constant time. 1220 */ 1221 if (object->backing_object == NULL) { 1222 tpindex = (m != NULL && m->pindex < end) ? 1223 m->pindex : end; 1224 vm_object_madvise_freespace(object, advice, 1225 pindex, tpindex - pindex); 1226 if ((pindex = tpindex) == end) 1227 break; 1228 goto next_page; 1229 } 1230 1231 tpindex = pindex; 1232 do { 1233 vm_object_madvise_freespace(tobject, advice, 1234 tpindex, 1); 1235 /* 1236 * Prepare to search the next object in the 1237 * chain. 1238 */ 1239 backing_object = tobject->backing_object; 1240 if (backing_object == NULL) 1241 goto next_pindex; 1242 VM_OBJECT_WLOCK(backing_object); 1243 tpindex += 1244 OFF_TO_IDX(tobject->backing_object_offset); 1245 if (tobject != object) 1246 VM_OBJECT_WUNLOCK(tobject); 1247 tobject = backing_object; 1248 if (!vm_object_advice_applies(tobject, advice)) 1249 goto next_pindex; 1250 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1251 NULL); 1252 } else { 1253 next_page: 1254 tm = m; 1255 m = TAILQ_NEXT(m, listq); 1256 } 1257 1258 /* 1259 * If the page is not in a normal state, skip it. 1260 */ 1261 if (tm->valid != VM_PAGE_BITS_ALL) 1262 goto next_pindex; 1263 vm_page_lock(tm); 1264 if (tm->hold_count != 0 || tm->wire_count != 0) { 1265 vm_page_unlock(tm); 1266 goto next_pindex; 1267 } 1268 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1269 ("vm_object_madvise: page %p is fictitious", tm)); 1270 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1271 ("vm_object_madvise: page %p is not managed", tm)); 1272 if (vm_page_busied(tm)) { 1273 if (object != tobject) 1274 VM_OBJECT_WUNLOCK(tobject); 1275 VM_OBJECT_WUNLOCK(object); 1276 if (advice == MADV_WILLNEED) { 1277 /* 1278 * Reference the page before unlocking and 1279 * sleeping so that the page daemon is less 1280 * likely to reclaim it. 1281 */ 1282 vm_page_aflag_set(tm, PGA_REFERENCED); 1283 } 1284 vm_page_busy_sleep(tm, "madvpo", false); 1285 goto relookup; 1286 } 1287 vm_page_advise(tm, advice); 1288 vm_page_unlock(tm); 1289 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1290 next_pindex: 1291 if (tobject != object) 1292 VM_OBJECT_WUNLOCK(tobject); 1293 } 1294 VM_OBJECT_WUNLOCK(object); 1295 } 1296 1297 /* 1298 * vm_object_shadow: 1299 * 1300 * Create a new object which is backed by the 1301 * specified existing object range. The source 1302 * object reference is deallocated. 1303 * 1304 * The new object and offset into that object 1305 * are returned in the source parameters. 1306 */ 1307 void 1308 vm_object_shadow( 1309 vm_object_t *object, /* IN/OUT */ 1310 vm_ooffset_t *offset, /* IN/OUT */ 1311 vm_size_t length) 1312 { 1313 vm_object_t source; 1314 vm_object_t result; 1315 1316 source = *object; 1317 1318 /* 1319 * Don't create the new object if the old object isn't shared. 1320 */ 1321 if (source != NULL) { 1322 VM_OBJECT_WLOCK(source); 1323 if (source->ref_count == 1 && 1324 source->handle == NULL && 1325 (source->type == OBJT_DEFAULT || 1326 source->type == OBJT_SWAP)) { 1327 VM_OBJECT_WUNLOCK(source); 1328 return; 1329 } 1330 VM_OBJECT_WUNLOCK(source); 1331 } 1332 1333 /* 1334 * Allocate a new object with the given length. 1335 */ 1336 result = vm_object_allocate(OBJT_DEFAULT, atop(length)); 1337 1338 /* 1339 * The new object shadows the source object, adding a reference to it. 1340 * Our caller changes his reference to point to the new object, 1341 * removing a reference to the source object. Net result: no change 1342 * of reference count. 1343 * 1344 * Try to optimize the result object's page color when shadowing 1345 * in order to maintain page coloring consistency in the combined 1346 * shadowed object. 1347 */ 1348 result->backing_object = source; 1349 /* 1350 * Store the offset into the source object, and fix up the offset into 1351 * the new object. 1352 */ 1353 result->backing_object_offset = *offset; 1354 if (source != NULL) { 1355 VM_OBJECT_WLOCK(source); 1356 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1357 source->shadow_count++; 1358 #if VM_NRESERVLEVEL > 0 1359 result->flags |= source->flags & OBJ_COLORED; 1360 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1361 ((1 << (VM_NFREEORDER - 1)) - 1); 1362 #endif 1363 VM_OBJECT_WUNLOCK(source); 1364 } 1365 1366 1367 /* 1368 * Return the new things 1369 */ 1370 *offset = 0; 1371 *object = result; 1372 } 1373 1374 /* 1375 * vm_object_split: 1376 * 1377 * Split the pages in a map entry into a new object. This affords 1378 * easier removal of unused pages, and keeps object inheritance from 1379 * being a negative impact on memory usage. 1380 */ 1381 void 1382 vm_object_split(vm_map_entry_t entry) 1383 { 1384 vm_page_t m, m_next; 1385 vm_object_t orig_object, new_object, source; 1386 vm_pindex_t idx, offidxstart; 1387 vm_size_t size; 1388 1389 orig_object = entry->object.vm_object; 1390 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1391 return; 1392 if (orig_object->ref_count <= 1) 1393 return; 1394 VM_OBJECT_WUNLOCK(orig_object); 1395 1396 offidxstart = OFF_TO_IDX(entry->offset); 1397 size = atop(entry->end - entry->start); 1398 1399 /* 1400 * If swap_pager_copy() is later called, it will convert new_object 1401 * into a swap object. 1402 */ 1403 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1404 1405 /* 1406 * At this point, the new object is still private, so the order in 1407 * which the original and new objects are locked does not matter. 1408 */ 1409 VM_OBJECT_WLOCK(new_object); 1410 VM_OBJECT_WLOCK(orig_object); 1411 source = orig_object->backing_object; 1412 if (source != NULL) { 1413 VM_OBJECT_WLOCK(source); 1414 if ((source->flags & OBJ_DEAD) != 0) { 1415 VM_OBJECT_WUNLOCK(source); 1416 VM_OBJECT_WUNLOCK(orig_object); 1417 VM_OBJECT_WUNLOCK(new_object); 1418 vm_object_deallocate(new_object); 1419 VM_OBJECT_WLOCK(orig_object); 1420 return; 1421 } 1422 LIST_INSERT_HEAD(&source->shadow_head, 1423 new_object, shadow_list); 1424 source->shadow_count++; 1425 vm_object_reference_locked(source); /* for new_object */ 1426 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1427 VM_OBJECT_WUNLOCK(source); 1428 new_object->backing_object_offset = 1429 orig_object->backing_object_offset + entry->offset; 1430 new_object->backing_object = source; 1431 } 1432 if (orig_object->cred != NULL) { 1433 new_object->cred = orig_object->cred; 1434 crhold(orig_object->cred); 1435 new_object->charge = ptoa(size); 1436 KASSERT(orig_object->charge >= ptoa(size), 1437 ("orig_object->charge < 0")); 1438 orig_object->charge -= ptoa(size); 1439 } 1440 retry: 1441 m = vm_page_find_least(orig_object, offidxstart); 1442 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1443 m = m_next) { 1444 m_next = TAILQ_NEXT(m, listq); 1445 1446 /* 1447 * We must wait for pending I/O to complete before we can 1448 * rename the page. 1449 * 1450 * We do not have to VM_PROT_NONE the page as mappings should 1451 * not be changed by this operation. 1452 */ 1453 if (vm_page_busied(m)) { 1454 VM_OBJECT_WUNLOCK(new_object); 1455 vm_page_lock(m); 1456 VM_OBJECT_WUNLOCK(orig_object); 1457 vm_page_busy_sleep(m, "spltwt", false); 1458 VM_OBJECT_WLOCK(orig_object); 1459 VM_OBJECT_WLOCK(new_object); 1460 goto retry; 1461 } 1462 1463 /* vm_page_rename() will dirty the page. */ 1464 if (vm_page_rename(m, new_object, idx)) { 1465 VM_OBJECT_WUNLOCK(new_object); 1466 VM_OBJECT_WUNLOCK(orig_object); 1467 vm_radix_wait(); 1468 VM_OBJECT_WLOCK(orig_object); 1469 VM_OBJECT_WLOCK(new_object); 1470 goto retry; 1471 } 1472 #if VM_NRESERVLEVEL > 0 1473 /* 1474 * If some of the reservation's allocated pages remain with 1475 * the original object, then transferring the reservation to 1476 * the new object is neither particularly beneficial nor 1477 * particularly harmful as compared to leaving the reservation 1478 * with the original object. If, however, all of the 1479 * reservation's allocated pages are transferred to the new 1480 * object, then transferring the reservation is typically 1481 * beneficial. Determining which of these two cases applies 1482 * would be more costly than unconditionally renaming the 1483 * reservation. 1484 */ 1485 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1486 #endif 1487 if (orig_object->type == OBJT_SWAP) 1488 vm_page_xbusy(m); 1489 } 1490 if (orig_object->type == OBJT_SWAP) { 1491 /* 1492 * swap_pager_copy() can sleep, in which case the orig_object's 1493 * and new_object's locks are released and reacquired. 1494 */ 1495 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1496 TAILQ_FOREACH(m, &new_object->memq, listq) 1497 vm_page_xunbusy(m); 1498 } 1499 VM_OBJECT_WUNLOCK(orig_object); 1500 VM_OBJECT_WUNLOCK(new_object); 1501 entry->object.vm_object = new_object; 1502 entry->offset = 0LL; 1503 vm_object_deallocate(orig_object); 1504 VM_OBJECT_WLOCK(new_object); 1505 } 1506 1507 #define OBSC_COLLAPSE_NOWAIT 0x0002 1508 #define OBSC_COLLAPSE_WAIT 0x0004 1509 1510 static vm_page_t 1511 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, 1512 int op) 1513 { 1514 vm_object_t backing_object; 1515 1516 VM_OBJECT_ASSERT_WLOCKED(object); 1517 backing_object = object->backing_object; 1518 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1519 1520 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p)); 1521 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1522 ("invalid ownership %p %p %p", p, object, backing_object)); 1523 if ((op & OBSC_COLLAPSE_NOWAIT) != 0) 1524 return (next); 1525 if (p != NULL) 1526 vm_page_lock(p); 1527 VM_OBJECT_WUNLOCK(object); 1528 VM_OBJECT_WUNLOCK(backing_object); 1529 /* The page is only NULL when rename fails. */ 1530 if (p == NULL) 1531 vm_radix_wait(); 1532 else 1533 vm_page_busy_sleep(p, "vmocol", false); 1534 VM_OBJECT_WLOCK(object); 1535 VM_OBJECT_WLOCK(backing_object); 1536 return (TAILQ_FIRST(&backing_object->memq)); 1537 } 1538 1539 static bool 1540 vm_object_scan_all_shadowed(vm_object_t object) 1541 { 1542 vm_object_t backing_object; 1543 vm_page_t p, pp; 1544 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1545 1546 VM_OBJECT_ASSERT_WLOCKED(object); 1547 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1548 1549 backing_object = object->backing_object; 1550 1551 if (backing_object->type != OBJT_DEFAULT && 1552 backing_object->type != OBJT_SWAP) 1553 return (false); 1554 1555 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1556 p = vm_page_find_least(backing_object, pi); 1557 ps = swap_pager_find_least(backing_object, pi); 1558 1559 /* 1560 * Only check pages inside the parent object's range and 1561 * inside the parent object's mapping of the backing object. 1562 */ 1563 for (;; pi++) { 1564 if (p != NULL && p->pindex < pi) 1565 p = TAILQ_NEXT(p, listq); 1566 if (ps < pi) 1567 ps = swap_pager_find_least(backing_object, pi); 1568 if (p == NULL && ps >= backing_object->size) 1569 break; 1570 else if (p == NULL) 1571 pi = ps; 1572 else 1573 pi = MIN(p->pindex, ps); 1574 1575 new_pindex = pi - backing_offset_index; 1576 if (new_pindex >= object->size) 1577 break; 1578 1579 /* 1580 * See if the parent has the page or if the parent's object 1581 * pager has the page. If the parent has the page but the page 1582 * is not valid, the parent's object pager must have the page. 1583 * 1584 * If this fails, the parent does not completely shadow the 1585 * object and we might as well give up now. 1586 */ 1587 pp = vm_page_lookup(object, new_pindex); 1588 if ((pp == NULL || pp->valid == 0) && 1589 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1590 return (false); 1591 } 1592 return (true); 1593 } 1594 1595 static bool 1596 vm_object_collapse_scan(vm_object_t object, int op) 1597 { 1598 vm_object_t backing_object; 1599 vm_page_t next, p, pp; 1600 vm_pindex_t backing_offset_index, new_pindex; 1601 1602 VM_OBJECT_ASSERT_WLOCKED(object); 1603 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1604 1605 backing_object = object->backing_object; 1606 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1607 1608 /* 1609 * Initial conditions 1610 */ 1611 if ((op & OBSC_COLLAPSE_WAIT) != 0) 1612 vm_object_set_flag(backing_object, OBJ_DEAD); 1613 1614 /* 1615 * Our scan 1616 */ 1617 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1618 next = TAILQ_NEXT(p, listq); 1619 new_pindex = p->pindex - backing_offset_index; 1620 1621 /* 1622 * Check for busy page 1623 */ 1624 if (vm_page_busied(p)) { 1625 next = vm_object_collapse_scan_wait(object, p, next, op); 1626 continue; 1627 } 1628 1629 KASSERT(p->object == backing_object, 1630 ("vm_object_collapse_scan: object mismatch")); 1631 1632 if (p->pindex < backing_offset_index || 1633 new_pindex >= object->size) { 1634 if (backing_object->type == OBJT_SWAP) 1635 swap_pager_freespace(backing_object, p->pindex, 1636 1); 1637 1638 /* 1639 * Page is out of the parent object's range, we can 1640 * simply destroy it. 1641 */ 1642 vm_page_lock(p); 1643 KASSERT(!pmap_page_is_mapped(p), 1644 ("freeing mapped page %p", p)); 1645 if (p->wire_count == 0) 1646 vm_page_free(p); 1647 else 1648 vm_page_remove(p); 1649 vm_page_unlock(p); 1650 continue; 1651 } 1652 1653 pp = vm_page_lookup(object, new_pindex); 1654 if (pp != NULL && vm_page_busied(pp)) { 1655 /* 1656 * The page in the parent is busy and possibly not 1657 * (yet) valid. Until its state is finalized by the 1658 * busy bit owner, we can't tell whether it shadows the 1659 * original page. Therefore, we must either skip it 1660 * and the original (backing_object) page or wait for 1661 * its state to be finalized. 1662 * 1663 * This is due to a race with vm_fault() where we must 1664 * unbusy the original (backing_obj) page before we can 1665 * (re)lock the parent. Hence we can get here. 1666 */ 1667 next = vm_object_collapse_scan_wait(object, pp, next, 1668 op); 1669 continue; 1670 } 1671 1672 KASSERT(pp == NULL || pp->valid != 0, 1673 ("unbusy invalid page %p", pp)); 1674 1675 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1676 NULL)) { 1677 /* 1678 * The page already exists in the parent OR swap exists 1679 * for this location in the parent. Leave the parent's 1680 * page alone. Destroy the original page from the 1681 * backing object. 1682 */ 1683 if (backing_object->type == OBJT_SWAP) 1684 swap_pager_freespace(backing_object, p->pindex, 1685 1); 1686 vm_page_lock(p); 1687 KASSERT(!pmap_page_is_mapped(p), 1688 ("freeing mapped page %p", p)); 1689 if (p->wire_count == 0) 1690 vm_page_free(p); 1691 else 1692 vm_page_remove(p); 1693 vm_page_unlock(p); 1694 continue; 1695 } 1696 1697 /* 1698 * Page does not exist in parent, rename the page from the 1699 * backing object to the main object. 1700 * 1701 * If the page was mapped to a process, it can remain mapped 1702 * through the rename. vm_page_rename() will dirty the page. 1703 */ 1704 if (vm_page_rename(p, object, new_pindex)) { 1705 next = vm_object_collapse_scan_wait(object, NULL, next, 1706 op); 1707 continue; 1708 } 1709 1710 /* Use the old pindex to free the right page. */ 1711 if (backing_object->type == OBJT_SWAP) 1712 swap_pager_freespace(backing_object, 1713 new_pindex + backing_offset_index, 1); 1714 1715 #if VM_NRESERVLEVEL > 0 1716 /* 1717 * Rename the reservation. 1718 */ 1719 vm_reserv_rename(p, object, backing_object, 1720 backing_offset_index); 1721 #endif 1722 } 1723 return (true); 1724 } 1725 1726 1727 /* 1728 * this version of collapse allows the operation to occur earlier and 1729 * when paging_in_progress is true for an object... This is not a complete 1730 * operation, but should plug 99.9% of the rest of the leaks. 1731 */ 1732 static void 1733 vm_object_qcollapse(vm_object_t object) 1734 { 1735 vm_object_t backing_object = object->backing_object; 1736 1737 VM_OBJECT_ASSERT_WLOCKED(object); 1738 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1739 1740 if (backing_object->ref_count != 1) 1741 return; 1742 1743 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); 1744 } 1745 1746 /* 1747 * vm_object_collapse: 1748 * 1749 * Collapse an object with the object backing it. 1750 * Pages in the backing object are moved into the 1751 * parent, and the backing object is deallocated. 1752 */ 1753 void 1754 vm_object_collapse(vm_object_t object) 1755 { 1756 vm_object_t backing_object, new_backing_object; 1757 1758 VM_OBJECT_ASSERT_WLOCKED(object); 1759 1760 while (TRUE) { 1761 /* 1762 * Verify that the conditions are right for collapse: 1763 * 1764 * The object exists and the backing object exists. 1765 */ 1766 if ((backing_object = object->backing_object) == NULL) 1767 break; 1768 1769 /* 1770 * we check the backing object first, because it is most likely 1771 * not collapsable. 1772 */ 1773 VM_OBJECT_WLOCK(backing_object); 1774 if (backing_object->handle != NULL || 1775 (backing_object->type != OBJT_DEFAULT && 1776 backing_object->type != OBJT_SWAP) || 1777 (backing_object->flags & OBJ_DEAD) || 1778 object->handle != NULL || 1779 (object->type != OBJT_DEFAULT && 1780 object->type != OBJT_SWAP) || 1781 (object->flags & OBJ_DEAD)) { 1782 VM_OBJECT_WUNLOCK(backing_object); 1783 break; 1784 } 1785 1786 if (object->paging_in_progress != 0 || 1787 backing_object->paging_in_progress != 0) { 1788 vm_object_qcollapse(object); 1789 VM_OBJECT_WUNLOCK(backing_object); 1790 break; 1791 } 1792 1793 /* 1794 * We know that we can either collapse the backing object (if 1795 * the parent is the only reference to it) or (perhaps) have 1796 * the parent bypass the object if the parent happens to shadow 1797 * all the resident pages in the entire backing object. 1798 * 1799 * This is ignoring pager-backed pages such as swap pages. 1800 * vm_object_collapse_scan fails the shadowing test in this 1801 * case. 1802 */ 1803 if (backing_object->ref_count == 1) { 1804 vm_object_pip_add(object, 1); 1805 vm_object_pip_add(backing_object, 1); 1806 1807 /* 1808 * If there is exactly one reference to the backing 1809 * object, we can collapse it into the parent. 1810 */ 1811 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); 1812 1813 #if VM_NRESERVLEVEL > 0 1814 /* 1815 * Break any reservations from backing_object. 1816 */ 1817 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1818 vm_reserv_break_all(backing_object); 1819 #endif 1820 1821 /* 1822 * Move the pager from backing_object to object. 1823 */ 1824 if (backing_object->type == OBJT_SWAP) { 1825 /* 1826 * swap_pager_copy() can sleep, in which case 1827 * the backing_object's and object's locks are 1828 * released and reacquired. 1829 * Since swap_pager_copy() is being asked to 1830 * destroy the source, it will change the 1831 * backing_object's type to OBJT_DEFAULT. 1832 */ 1833 swap_pager_copy( 1834 backing_object, 1835 object, 1836 OFF_TO_IDX(object->backing_object_offset), TRUE); 1837 } 1838 /* 1839 * Object now shadows whatever backing_object did. 1840 * Note that the reference to 1841 * backing_object->backing_object moves from within 1842 * backing_object to within object. 1843 */ 1844 LIST_REMOVE(object, shadow_list); 1845 backing_object->shadow_count--; 1846 if (backing_object->backing_object) { 1847 VM_OBJECT_WLOCK(backing_object->backing_object); 1848 LIST_REMOVE(backing_object, shadow_list); 1849 LIST_INSERT_HEAD( 1850 &backing_object->backing_object->shadow_head, 1851 object, shadow_list); 1852 /* 1853 * The shadow_count has not changed. 1854 */ 1855 VM_OBJECT_WUNLOCK(backing_object->backing_object); 1856 } 1857 object->backing_object = backing_object->backing_object; 1858 object->backing_object_offset += 1859 backing_object->backing_object_offset; 1860 1861 /* 1862 * Discard backing_object. 1863 * 1864 * Since the backing object has no pages, no pager left, 1865 * and no object references within it, all that is 1866 * necessary is to dispose of it. 1867 */ 1868 KASSERT(backing_object->ref_count == 1, ( 1869 "backing_object %p was somehow re-referenced during collapse!", 1870 backing_object)); 1871 vm_object_pip_wakeup(backing_object); 1872 backing_object->type = OBJT_DEAD; 1873 backing_object->ref_count = 0; 1874 VM_OBJECT_WUNLOCK(backing_object); 1875 vm_object_destroy(backing_object); 1876 1877 vm_object_pip_wakeup(object); 1878 object_collapses++; 1879 } else { 1880 /* 1881 * If we do not entirely shadow the backing object, 1882 * there is nothing we can do so we give up. 1883 */ 1884 if (object->resident_page_count != object->size && 1885 !vm_object_scan_all_shadowed(object)) { 1886 VM_OBJECT_WUNLOCK(backing_object); 1887 break; 1888 } 1889 1890 /* 1891 * Make the parent shadow the next object in the 1892 * chain. Deallocating backing_object will not remove 1893 * it, since its reference count is at least 2. 1894 */ 1895 LIST_REMOVE(object, shadow_list); 1896 backing_object->shadow_count--; 1897 1898 new_backing_object = backing_object->backing_object; 1899 if ((object->backing_object = new_backing_object) != NULL) { 1900 VM_OBJECT_WLOCK(new_backing_object); 1901 LIST_INSERT_HEAD( 1902 &new_backing_object->shadow_head, 1903 object, 1904 shadow_list 1905 ); 1906 new_backing_object->shadow_count++; 1907 vm_object_reference_locked(new_backing_object); 1908 VM_OBJECT_WUNLOCK(new_backing_object); 1909 object->backing_object_offset += 1910 backing_object->backing_object_offset; 1911 } 1912 1913 /* 1914 * Drop the reference count on backing_object. Since 1915 * its ref_count was at least 2, it will not vanish. 1916 */ 1917 backing_object->ref_count--; 1918 VM_OBJECT_WUNLOCK(backing_object); 1919 object_bypasses++; 1920 } 1921 1922 /* 1923 * Try again with this object's new backing object. 1924 */ 1925 } 1926 } 1927 1928 /* 1929 * vm_object_page_remove: 1930 * 1931 * For the given object, either frees or invalidates each of the 1932 * specified pages. In general, a page is freed. However, if a page is 1933 * wired for any reason other than the existence of a managed, wired 1934 * mapping, then it may be invalidated but not removed from the object. 1935 * Pages are specified by the given range ["start", "end") and the option 1936 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1937 * extends from "start" to the end of the object. If the option 1938 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1939 * specified range are affected. If the option OBJPR_NOTMAPPED is 1940 * specified, then the pages within the specified range must have no 1941 * mappings. Otherwise, if this option is not specified, any mappings to 1942 * the specified pages are removed before the pages are freed or 1943 * invalidated. 1944 * 1945 * In general, this operation should only be performed on objects that 1946 * contain managed pages. There are, however, two exceptions. First, it 1947 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1948 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1949 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1950 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1951 * 1952 * The object must be locked. 1953 */ 1954 void 1955 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1956 int options) 1957 { 1958 vm_page_t p, next; 1959 struct mtx *mtx; 1960 struct pglist pgl; 1961 1962 VM_OBJECT_ASSERT_WLOCKED(object); 1963 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1964 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1965 ("vm_object_page_remove: illegal options for object %p", object)); 1966 if (object->resident_page_count == 0) 1967 return; 1968 vm_object_pip_add(object, 1); 1969 TAILQ_INIT(&pgl); 1970 again: 1971 p = vm_page_find_least(object, start); 1972 mtx = NULL; 1973 1974 /* 1975 * Here, the variable "p" is either (1) the page with the least pindex 1976 * greater than or equal to the parameter "start" or (2) NULL. 1977 */ 1978 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 1979 next = TAILQ_NEXT(p, listq); 1980 1981 /* 1982 * If the page is wired for any reason besides the existence 1983 * of managed, wired mappings, then it cannot be freed. For 1984 * example, fictitious pages, which represent device memory, 1985 * are inherently wired and cannot be freed. They can, 1986 * however, be invalidated if the option OBJPR_CLEANONLY is 1987 * not specified. 1988 */ 1989 vm_page_change_lock(p, &mtx); 1990 if (vm_page_xbusied(p)) { 1991 VM_OBJECT_WUNLOCK(object); 1992 vm_page_busy_sleep(p, "vmopax", true); 1993 VM_OBJECT_WLOCK(object); 1994 goto again; 1995 } 1996 if (p->wire_count != 0) { 1997 if ((options & OBJPR_NOTMAPPED) == 0 && 1998 object->ref_count != 0) 1999 pmap_remove_all(p); 2000 if ((options & OBJPR_CLEANONLY) == 0) { 2001 p->valid = 0; 2002 vm_page_undirty(p); 2003 } 2004 continue; 2005 } 2006 if (vm_page_busied(p)) { 2007 VM_OBJECT_WUNLOCK(object); 2008 vm_page_busy_sleep(p, "vmopar", false); 2009 VM_OBJECT_WLOCK(object); 2010 goto again; 2011 } 2012 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2013 ("vm_object_page_remove: page %p is fictitious", p)); 2014 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) { 2015 if ((options & OBJPR_NOTMAPPED) == 0 && 2016 object->ref_count != 0) 2017 pmap_remove_write(p); 2018 if (p->dirty != 0) 2019 continue; 2020 } 2021 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) 2022 pmap_remove_all(p); 2023 p->flags &= ~PG_ZERO; 2024 if (vm_page_free_prep(p, false)) 2025 TAILQ_INSERT_TAIL(&pgl, p, listq); 2026 } 2027 if (mtx != NULL) 2028 mtx_unlock(mtx); 2029 vm_page_free_phys_pglist(&pgl); 2030 vm_object_pip_wakeup(object); 2031 } 2032 2033 /* 2034 * vm_object_page_noreuse: 2035 * 2036 * For the given object, attempt to move the specified pages to 2037 * the head of the inactive queue. This bypasses regular LRU 2038 * operation and allows the pages to be reused quickly under memory 2039 * pressure. If a page is wired for any reason, then it will not 2040 * be queued. Pages are specified by the range ["start", "end"). 2041 * As a special case, if "end" is zero, then the range extends from 2042 * "start" to the end of the object. 2043 * 2044 * This operation should only be performed on objects that 2045 * contain non-fictitious, managed pages. 2046 * 2047 * The object must be locked. 2048 */ 2049 void 2050 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2051 { 2052 struct mtx *mtx; 2053 vm_page_t p, next; 2054 2055 VM_OBJECT_ASSERT_LOCKED(object); 2056 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2057 ("vm_object_page_noreuse: illegal object %p", object)); 2058 if (object->resident_page_count == 0) 2059 return; 2060 p = vm_page_find_least(object, start); 2061 2062 /* 2063 * Here, the variable "p" is either (1) the page with the least pindex 2064 * greater than or equal to the parameter "start" or (2) NULL. 2065 */ 2066 mtx = NULL; 2067 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2068 next = TAILQ_NEXT(p, listq); 2069 vm_page_change_lock(p, &mtx); 2070 vm_page_deactivate_noreuse(p); 2071 } 2072 if (mtx != NULL) 2073 mtx_unlock(mtx); 2074 } 2075 2076 /* 2077 * Populate the specified range of the object with valid pages. Returns 2078 * TRUE if the range is successfully populated and FALSE otherwise. 2079 * 2080 * Note: This function should be optimized to pass a larger array of 2081 * pages to vm_pager_get_pages() before it is applied to a non- 2082 * OBJT_DEVICE object. 2083 * 2084 * The object must be locked. 2085 */ 2086 boolean_t 2087 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2088 { 2089 vm_page_t m; 2090 vm_pindex_t pindex; 2091 int rv; 2092 2093 VM_OBJECT_ASSERT_WLOCKED(object); 2094 for (pindex = start; pindex < end; pindex++) { 2095 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 2096 if (m->valid != VM_PAGE_BITS_ALL) { 2097 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 2098 if (rv != VM_PAGER_OK) { 2099 vm_page_lock(m); 2100 vm_page_free(m); 2101 vm_page_unlock(m); 2102 break; 2103 } 2104 } 2105 /* 2106 * Keep "m" busy because a subsequent iteration may unlock 2107 * the object. 2108 */ 2109 } 2110 if (pindex > start) { 2111 m = vm_page_lookup(object, start); 2112 while (m != NULL && m->pindex < pindex) { 2113 vm_page_xunbusy(m); 2114 m = TAILQ_NEXT(m, listq); 2115 } 2116 } 2117 return (pindex == end); 2118 } 2119 2120 /* 2121 * Routine: vm_object_coalesce 2122 * Function: Coalesces two objects backing up adjoining 2123 * regions of memory into a single object. 2124 * 2125 * returns TRUE if objects were combined. 2126 * 2127 * NOTE: Only works at the moment if the second object is NULL - 2128 * if it's not, which object do we lock first? 2129 * 2130 * Parameters: 2131 * prev_object First object to coalesce 2132 * prev_offset Offset into prev_object 2133 * prev_size Size of reference to prev_object 2134 * next_size Size of reference to the second object 2135 * reserved Indicator that extension region has 2136 * swap accounted for 2137 * 2138 * Conditions: 2139 * The object must *not* be locked. 2140 */ 2141 boolean_t 2142 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2143 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2144 { 2145 vm_pindex_t next_pindex; 2146 2147 if (prev_object == NULL) 2148 return (TRUE); 2149 VM_OBJECT_WLOCK(prev_object); 2150 if ((prev_object->type != OBJT_DEFAULT && 2151 prev_object->type != OBJT_SWAP) || 2152 (prev_object->flags & OBJ_TMPFS_NODE) != 0) { 2153 VM_OBJECT_WUNLOCK(prev_object); 2154 return (FALSE); 2155 } 2156 2157 /* 2158 * Try to collapse the object first 2159 */ 2160 vm_object_collapse(prev_object); 2161 2162 /* 2163 * Can't coalesce if: . more than one reference . paged out . shadows 2164 * another object . has a copy elsewhere (any of which mean that the 2165 * pages not mapped to prev_entry may be in use anyway) 2166 */ 2167 if (prev_object->backing_object != NULL) { 2168 VM_OBJECT_WUNLOCK(prev_object); 2169 return (FALSE); 2170 } 2171 2172 prev_size >>= PAGE_SHIFT; 2173 next_size >>= PAGE_SHIFT; 2174 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2175 2176 if ((prev_object->ref_count > 1) && 2177 (prev_object->size != next_pindex)) { 2178 VM_OBJECT_WUNLOCK(prev_object); 2179 return (FALSE); 2180 } 2181 2182 /* 2183 * Account for the charge. 2184 */ 2185 if (prev_object->cred != NULL) { 2186 2187 /* 2188 * If prev_object was charged, then this mapping, 2189 * although not charged now, may become writable 2190 * later. Non-NULL cred in the object would prevent 2191 * swap reservation during enabling of the write 2192 * access, so reserve swap now. Failed reservation 2193 * cause allocation of the separate object for the map 2194 * entry, and swap reservation for this entry is 2195 * managed in appropriate time. 2196 */ 2197 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2198 prev_object->cred)) { 2199 VM_OBJECT_WUNLOCK(prev_object); 2200 return (FALSE); 2201 } 2202 prev_object->charge += ptoa(next_size); 2203 } 2204 2205 /* 2206 * Remove any pages that may still be in the object from a previous 2207 * deallocation. 2208 */ 2209 if (next_pindex < prev_object->size) { 2210 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2211 next_size, 0); 2212 if (prev_object->type == OBJT_SWAP) 2213 swap_pager_freespace(prev_object, 2214 next_pindex, next_size); 2215 #if 0 2216 if (prev_object->cred != NULL) { 2217 KASSERT(prev_object->charge >= 2218 ptoa(prev_object->size - next_pindex), 2219 ("object %p overcharged 1 %jx %jx", prev_object, 2220 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2221 prev_object->charge -= ptoa(prev_object->size - 2222 next_pindex); 2223 } 2224 #endif 2225 } 2226 2227 /* 2228 * Extend the object if necessary. 2229 */ 2230 if (next_pindex + next_size > prev_object->size) 2231 prev_object->size = next_pindex + next_size; 2232 2233 VM_OBJECT_WUNLOCK(prev_object); 2234 return (TRUE); 2235 } 2236 2237 void 2238 vm_object_set_writeable_dirty(vm_object_t object) 2239 { 2240 2241 VM_OBJECT_ASSERT_WLOCKED(object); 2242 if (object->type != OBJT_VNODE) { 2243 if ((object->flags & OBJ_TMPFS_NODE) != 0) { 2244 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs")); 2245 vm_object_set_flag(object, OBJ_TMPFS_DIRTY); 2246 } 2247 return; 2248 } 2249 object->generation++; 2250 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) 2251 return; 2252 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2253 } 2254 2255 /* 2256 * vm_object_unwire: 2257 * 2258 * For each page offset within the specified range of the given object, 2259 * find the highest-level page in the shadow chain and unwire it. A page 2260 * must exist at every page offset, and the highest-level page must be 2261 * wired. 2262 */ 2263 void 2264 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2265 uint8_t queue) 2266 { 2267 vm_object_t tobject; 2268 vm_page_t m, tm; 2269 vm_pindex_t end_pindex, pindex, tpindex; 2270 int depth, locked_depth; 2271 2272 KASSERT((offset & PAGE_MASK) == 0, 2273 ("vm_object_unwire: offset is not page aligned")); 2274 KASSERT((length & PAGE_MASK) == 0, 2275 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2276 /* The wired count of a fictitious page never changes. */ 2277 if ((object->flags & OBJ_FICTITIOUS) != 0) 2278 return; 2279 pindex = OFF_TO_IDX(offset); 2280 end_pindex = pindex + atop(length); 2281 locked_depth = 1; 2282 VM_OBJECT_RLOCK(object); 2283 m = vm_page_find_least(object, pindex); 2284 while (pindex < end_pindex) { 2285 if (m == NULL || pindex < m->pindex) { 2286 /* 2287 * The first object in the shadow chain doesn't 2288 * contain a page at the current index. Therefore, 2289 * the page must exist in a backing object. 2290 */ 2291 tobject = object; 2292 tpindex = pindex; 2293 depth = 0; 2294 do { 2295 tpindex += 2296 OFF_TO_IDX(tobject->backing_object_offset); 2297 tobject = tobject->backing_object; 2298 KASSERT(tobject != NULL, 2299 ("vm_object_unwire: missing page")); 2300 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2301 goto next_page; 2302 depth++; 2303 if (depth == locked_depth) { 2304 locked_depth++; 2305 VM_OBJECT_RLOCK(tobject); 2306 } 2307 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2308 NULL); 2309 } else { 2310 tm = m; 2311 m = TAILQ_NEXT(m, listq); 2312 } 2313 vm_page_lock(tm); 2314 vm_page_unwire(tm, queue); 2315 vm_page_unlock(tm); 2316 next_page: 2317 pindex++; 2318 } 2319 /* Release the accumulated object locks. */ 2320 for (depth = 0; depth < locked_depth; depth++) { 2321 tobject = object->backing_object; 2322 VM_OBJECT_RUNLOCK(object); 2323 object = tobject; 2324 } 2325 } 2326 2327 struct vnode * 2328 vm_object_vnode(vm_object_t object) 2329 { 2330 2331 VM_OBJECT_ASSERT_LOCKED(object); 2332 if (object->type == OBJT_VNODE) 2333 return (object->handle); 2334 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) 2335 return (object->un_pager.swp.swp_tmpfs); 2336 return (NULL); 2337 } 2338 2339 static int 2340 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2341 { 2342 struct kinfo_vmobject *kvo; 2343 char *fullpath, *freepath; 2344 struct vnode *vp; 2345 struct vattr va; 2346 vm_object_t obj; 2347 vm_page_t m; 2348 int count, error; 2349 2350 if (req->oldptr == NULL) { 2351 /* 2352 * If an old buffer has not been provided, generate an 2353 * estimate of the space needed for a subsequent call. 2354 */ 2355 mtx_lock(&vm_object_list_mtx); 2356 count = 0; 2357 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2358 if (obj->type == OBJT_DEAD) 2359 continue; 2360 count++; 2361 } 2362 mtx_unlock(&vm_object_list_mtx); 2363 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2364 count * 11 / 10)); 2365 } 2366 2367 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2368 error = 0; 2369 2370 /* 2371 * VM objects are type stable and are never removed from the 2372 * list once added. This allows us to safely read obj->object_list 2373 * after reacquiring the VM object lock. 2374 */ 2375 mtx_lock(&vm_object_list_mtx); 2376 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2377 if (obj->type == OBJT_DEAD) 2378 continue; 2379 VM_OBJECT_RLOCK(obj); 2380 if (obj->type == OBJT_DEAD) { 2381 VM_OBJECT_RUNLOCK(obj); 2382 continue; 2383 } 2384 mtx_unlock(&vm_object_list_mtx); 2385 kvo->kvo_size = ptoa(obj->size); 2386 kvo->kvo_resident = obj->resident_page_count; 2387 kvo->kvo_ref_count = obj->ref_count; 2388 kvo->kvo_shadow_count = obj->shadow_count; 2389 kvo->kvo_memattr = obj->memattr; 2390 kvo->kvo_active = 0; 2391 kvo->kvo_inactive = 0; 2392 TAILQ_FOREACH(m, &obj->memq, listq) { 2393 /* 2394 * A page may belong to the object but be 2395 * dequeued and set to PQ_NONE while the 2396 * object lock is not held. This makes the 2397 * reads of m->queue below racy, and we do not 2398 * count pages set to PQ_NONE. However, this 2399 * sysctl is only meant to give an 2400 * approximation of the system anyway. 2401 */ 2402 if (vm_page_active(m)) 2403 kvo->kvo_active++; 2404 else if (vm_page_inactive(m)) 2405 kvo->kvo_inactive++; 2406 } 2407 2408 kvo->kvo_vn_fileid = 0; 2409 kvo->kvo_vn_fsid = 0; 2410 kvo->kvo_vn_fsid_freebsd11 = 0; 2411 freepath = NULL; 2412 fullpath = ""; 2413 vp = NULL; 2414 switch (obj->type) { 2415 case OBJT_DEFAULT: 2416 kvo->kvo_type = KVME_TYPE_DEFAULT; 2417 break; 2418 case OBJT_VNODE: 2419 kvo->kvo_type = KVME_TYPE_VNODE; 2420 vp = obj->handle; 2421 vref(vp); 2422 break; 2423 case OBJT_SWAP: 2424 kvo->kvo_type = KVME_TYPE_SWAP; 2425 break; 2426 case OBJT_DEVICE: 2427 kvo->kvo_type = KVME_TYPE_DEVICE; 2428 break; 2429 case OBJT_PHYS: 2430 kvo->kvo_type = KVME_TYPE_PHYS; 2431 break; 2432 case OBJT_DEAD: 2433 kvo->kvo_type = KVME_TYPE_DEAD; 2434 break; 2435 case OBJT_SG: 2436 kvo->kvo_type = KVME_TYPE_SG; 2437 break; 2438 case OBJT_MGTDEVICE: 2439 kvo->kvo_type = KVME_TYPE_MGTDEVICE; 2440 break; 2441 default: 2442 kvo->kvo_type = KVME_TYPE_UNKNOWN; 2443 break; 2444 } 2445 VM_OBJECT_RUNLOCK(obj); 2446 if (vp != NULL) { 2447 vn_fullpath(curthread, vp, &fullpath, &freepath); 2448 vn_lock(vp, LK_SHARED | LK_RETRY); 2449 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2450 kvo->kvo_vn_fileid = va.va_fileid; 2451 kvo->kvo_vn_fsid = va.va_fsid; 2452 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2453 /* truncate */ 2454 } 2455 vput(vp); 2456 } 2457 2458 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2459 if (freepath != NULL) 2460 free(freepath, M_TEMP); 2461 2462 /* Pack record size down */ 2463 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2464 + strlen(kvo->kvo_path) + 1; 2465 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2466 sizeof(uint64_t)); 2467 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2468 mtx_lock(&vm_object_list_mtx); 2469 if (error) 2470 break; 2471 } 2472 mtx_unlock(&vm_object_list_mtx); 2473 free(kvo, M_TEMP); 2474 return (error); 2475 } 2476 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2477 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2478 "List of VM objects"); 2479 2480 #include "opt_ddb.h" 2481 #ifdef DDB 2482 #include <sys/kernel.h> 2483 2484 #include <sys/cons.h> 2485 2486 #include <ddb/ddb.h> 2487 2488 static int 2489 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2490 { 2491 vm_map_t tmpm; 2492 vm_map_entry_t tmpe; 2493 vm_object_t obj; 2494 int entcount; 2495 2496 if (map == 0) 2497 return 0; 2498 2499 if (entry == 0) { 2500 tmpe = map->header.next; 2501 entcount = map->nentries; 2502 while (entcount-- && (tmpe != &map->header)) { 2503 if (_vm_object_in_map(map, object, tmpe)) { 2504 return 1; 2505 } 2506 tmpe = tmpe->next; 2507 } 2508 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2509 tmpm = entry->object.sub_map; 2510 tmpe = tmpm->header.next; 2511 entcount = tmpm->nentries; 2512 while (entcount-- && tmpe != &tmpm->header) { 2513 if (_vm_object_in_map(tmpm, object, tmpe)) { 2514 return 1; 2515 } 2516 tmpe = tmpe->next; 2517 } 2518 } else if ((obj = entry->object.vm_object) != NULL) { 2519 for (; obj; obj = obj->backing_object) 2520 if (obj == object) { 2521 return 1; 2522 } 2523 } 2524 return 0; 2525 } 2526 2527 static int 2528 vm_object_in_map(vm_object_t object) 2529 { 2530 struct proc *p; 2531 2532 /* sx_slock(&allproc_lock); */ 2533 FOREACH_PROC_IN_SYSTEM(p) { 2534 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2535 continue; 2536 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2537 /* sx_sunlock(&allproc_lock); */ 2538 return 1; 2539 } 2540 } 2541 /* sx_sunlock(&allproc_lock); */ 2542 if (_vm_object_in_map(kernel_map, object, 0)) 2543 return 1; 2544 return 0; 2545 } 2546 2547 DB_SHOW_COMMAND(vmochk, vm_object_check) 2548 { 2549 vm_object_t object; 2550 2551 /* 2552 * make sure that internal objs are in a map somewhere 2553 * and none have zero ref counts. 2554 */ 2555 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2556 if (object->handle == NULL && 2557 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2558 if (object->ref_count == 0) { 2559 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2560 (long)object->size); 2561 } 2562 if (!vm_object_in_map(object)) { 2563 db_printf( 2564 "vmochk: internal obj is not in a map: " 2565 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2566 object->ref_count, (u_long)object->size, 2567 (u_long)object->size, 2568 (void *)object->backing_object); 2569 } 2570 } 2571 } 2572 } 2573 2574 /* 2575 * vm_object_print: [ debug ] 2576 */ 2577 DB_SHOW_COMMAND(object, vm_object_print_static) 2578 { 2579 /* XXX convert args. */ 2580 vm_object_t object = (vm_object_t)addr; 2581 boolean_t full = have_addr; 2582 2583 vm_page_t p; 2584 2585 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2586 #define count was_count 2587 2588 int count; 2589 2590 if (object == NULL) 2591 return; 2592 2593 db_iprintf( 2594 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2595 object, (int)object->type, (uintmax_t)object->size, 2596 object->resident_page_count, object->ref_count, object->flags, 2597 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2598 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2599 object->shadow_count, 2600 object->backing_object ? object->backing_object->ref_count : 0, 2601 object->backing_object, (uintmax_t)object->backing_object_offset); 2602 2603 if (!full) 2604 return; 2605 2606 db_indent += 2; 2607 count = 0; 2608 TAILQ_FOREACH(p, &object->memq, listq) { 2609 if (count == 0) 2610 db_iprintf("memory:="); 2611 else if (count == 6) { 2612 db_printf("\n"); 2613 db_iprintf(" ..."); 2614 count = 0; 2615 } else 2616 db_printf(","); 2617 count++; 2618 2619 db_printf("(off=0x%jx,page=0x%jx)", 2620 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2621 } 2622 if (count != 0) 2623 db_printf("\n"); 2624 db_indent -= 2; 2625 } 2626 2627 /* XXX. */ 2628 #undef count 2629 2630 /* XXX need this non-static entry for calling from vm_map_print. */ 2631 void 2632 vm_object_print( 2633 /* db_expr_t */ long addr, 2634 boolean_t have_addr, 2635 /* db_expr_t */ long count, 2636 char *modif) 2637 { 2638 vm_object_print_static(addr, have_addr, count, modif); 2639 } 2640 2641 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2642 { 2643 vm_object_t object; 2644 vm_pindex_t fidx; 2645 vm_paddr_t pa; 2646 vm_page_t m, prev_m; 2647 int rcount, nl, c; 2648 2649 nl = 0; 2650 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2651 db_printf("new object: %p\n", (void *)object); 2652 if (nl > 18) { 2653 c = cngetc(); 2654 if (c != ' ') 2655 return; 2656 nl = 0; 2657 } 2658 nl++; 2659 rcount = 0; 2660 fidx = 0; 2661 pa = -1; 2662 TAILQ_FOREACH(m, &object->memq, listq) { 2663 if (m->pindex > 128) 2664 break; 2665 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2666 prev_m->pindex + 1 != m->pindex) { 2667 if (rcount) { 2668 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2669 (long)fidx, rcount, (long)pa); 2670 if (nl > 18) { 2671 c = cngetc(); 2672 if (c != ' ') 2673 return; 2674 nl = 0; 2675 } 2676 nl++; 2677 rcount = 0; 2678 } 2679 } 2680 if (rcount && 2681 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2682 ++rcount; 2683 continue; 2684 } 2685 if (rcount) { 2686 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2687 (long)fidx, rcount, (long)pa); 2688 if (nl > 18) { 2689 c = cngetc(); 2690 if (c != ' ') 2691 return; 2692 nl = 0; 2693 } 2694 nl++; 2695 } 2696 fidx = m->pindex; 2697 pa = VM_PAGE_TO_PHYS(m); 2698 rcount = 1; 2699 } 2700 if (rcount) { 2701 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2702 (long)fidx, rcount, (long)pa); 2703 if (nl > 18) { 2704 c = cngetc(); 2705 if (c != ' ') 2706 return; 2707 nl = 0; 2708 } 2709 nl++; 2710 } 2711 } 2712 } 2713 #endif /* DDB */ 2714