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/blockcount.h> 75 #include <sys/cpuset.h> 76 #include <sys/limits.h> 77 #include <sys/lock.h> 78 #include <sys/mman.h> 79 #include <sys/mount.h> 80 #include <sys/kernel.h> 81 #include <sys/pctrie.h> 82 #include <sys/sysctl.h> 83 #include <sys/mutex.h> 84 #include <sys/proc.h> /* for curproc, pageproc */ 85 #include <sys/refcount.h> 86 #include <sys/socket.h> 87 #include <sys/resourcevar.h> 88 #include <sys/refcount.h> 89 #include <sys/rwlock.h> 90 #include <sys/user.h> 91 #include <sys/vnode.h> 92 #include <sys/vmmeter.h> 93 #include <sys/sx.h> 94 95 #include <vm/vm.h> 96 #include <vm/vm_param.h> 97 #include <vm/pmap.h> 98 #include <vm/vm_map.h> 99 #include <vm/vm_object.h> 100 #include <vm/vm_page.h> 101 #include <vm/vm_pageout.h> 102 #include <vm/vm_pager.h> 103 #include <vm/vm_phys.h> 104 #include <vm/vm_pagequeue.h> 105 #include <vm/swap_pager.h> 106 #include <vm/vm_kern.h> 107 #include <vm/vm_extern.h> 108 #include <vm/vm_radix.h> 109 #include <vm/vm_reserv.h> 110 #include <vm/uma.h> 111 112 static int old_msync; 113 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 114 "Use old (insecure) msync behavior"); 115 116 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 117 int pagerflags, int flags, boolean_t *allclean, 118 boolean_t *eio); 119 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 120 boolean_t *allclean); 121 static void vm_object_backing_remove(vm_object_t object); 122 123 /* 124 * Virtual memory objects maintain the actual data 125 * associated with allocated virtual memory. A given 126 * page of memory exists within exactly one object. 127 * 128 * An object is only deallocated when all "references" 129 * are given up. Only one "reference" to a given 130 * region of an object should be writeable. 131 * 132 * Associated with each object is a list of all resident 133 * memory pages belonging to that object; this list is 134 * maintained by the "vm_page" module, and locked by the object's 135 * lock. 136 * 137 * Each object also records a "pager" routine which is 138 * used to retrieve (and store) pages to the proper backing 139 * storage. In addition, objects may be backed by other 140 * objects from which they were virtual-copied. 141 * 142 * The only items within the object structure which are 143 * modified after time of creation are: 144 * reference count locked by object's lock 145 * pager routine locked by object's lock 146 * 147 */ 148 149 struct object_q vm_object_list; 150 struct mtx vm_object_list_mtx; /* lock for object list and count */ 151 152 struct vm_object kernel_object_store; 153 154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 155 "VM object stats"); 156 157 static COUNTER_U64_DEFINE_EARLY(object_collapses); 158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 159 &object_collapses, 160 "VM object collapses"); 161 162 static COUNTER_U64_DEFINE_EARLY(object_bypasses); 163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 164 &object_bypasses, 165 "VM object bypasses"); 166 167 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits); 168 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD, 169 &object_collapse_waits, 170 "Number of sleeps for collapse"); 171 172 static uma_zone_t obj_zone; 173 174 static int vm_object_zinit(void *mem, int size, int flags); 175 176 #ifdef INVARIANTS 177 static void vm_object_zdtor(void *mem, int size, void *arg); 178 179 static void 180 vm_object_zdtor(void *mem, int size, void *arg) 181 { 182 vm_object_t object; 183 184 object = (vm_object_t)mem; 185 KASSERT(object->ref_count == 0, 186 ("object %p ref_count = %d", object, object->ref_count)); 187 KASSERT(TAILQ_EMPTY(&object->memq), 188 ("object %p has resident pages in its memq", object)); 189 KASSERT(vm_radix_is_empty(&object->rtree), 190 ("object %p has resident pages in its trie", object)); 191 #if VM_NRESERVLEVEL > 0 192 KASSERT(LIST_EMPTY(&object->rvq), 193 ("object %p has reservations", 194 object)); 195 #endif 196 KASSERT(!vm_object_busied(object), 197 ("object %p busy = %d", object, blockcount_read(&object->busy))); 198 KASSERT(object->resident_page_count == 0, 199 ("object %p resident_page_count = %d", 200 object, object->resident_page_count)); 201 KASSERT(object->shadow_count == 0, 202 ("object %p shadow_count = %d", 203 object, object->shadow_count)); 204 KASSERT(object->type == OBJT_DEAD, 205 ("object %p has non-dead type %d", 206 object, object->type)); 207 } 208 #endif 209 210 static int 211 vm_object_zinit(void *mem, int size, int flags) 212 { 213 vm_object_t object; 214 215 object = (vm_object_t)mem; 216 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 217 218 /* These are true for any object that has been freed */ 219 object->type = OBJT_DEAD; 220 vm_radix_init(&object->rtree); 221 refcount_init(&object->ref_count, 0); 222 blockcount_init(&object->paging_in_progress); 223 blockcount_init(&object->busy); 224 object->resident_page_count = 0; 225 object->shadow_count = 0; 226 object->flags = OBJ_DEAD; 227 228 mtx_lock(&vm_object_list_mtx); 229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 230 mtx_unlock(&vm_object_list_mtx); 231 return (0); 232 } 233 234 static void 235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, 236 vm_object_t object, void *handle) 237 { 238 239 TAILQ_INIT(&object->memq); 240 LIST_INIT(&object->shadow_head); 241 242 object->type = type; 243 object->flags = flags; 244 if ((flags & OBJ_SWAP) != 0) 245 pctrie_init(&object->un_pager.swp.swp_blks); 246 247 /* 248 * Ensure that swap_pager_swapoff() iteration over object_list 249 * sees up to date type and pctrie head if it observed 250 * non-dead object. 251 */ 252 atomic_thread_fence_rel(); 253 254 object->pg_color = 0; 255 object->size = size; 256 object->domain.dr_policy = NULL; 257 object->generation = 1; 258 object->cleangeneration = 1; 259 refcount_init(&object->ref_count, 1); 260 object->memattr = VM_MEMATTR_DEFAULT; 261 object->cred = NULL; 262 object->charge = 0; 263 object->handle = handle; 264 object->backing_object = NULL; 265 object->backing_object_offset = (vm_ooffset_t) 0; 266 #if VM_NRESERVLEVEL > 0 267 LIST_INIT(&object->rvq); 268 #endif 269 umtx_shm_object_init(object); 270 } 271 272 /* 273 * vm_object_init: 274 * 275 * Initialize the VM objects module. 276 */ 277 void 278 vm_object_init(void) 279 { 280 TAILQ_INIT(&vm_object_list); 281 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 282 283 rw_init(&kernel_object->lock, "kernel vm object"); 284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 285 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); 286 #if VM_NRESERVLEVEL > 0 287 kernel_object->flags |= OBJ_COLORED; 288 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 289 #endif 290 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 291 292 /* 293 * The lock portion of struct vm_object must be type stable due 294 * to vm_pageout_fallback_object_lock locking a vm object 295 * without holding any references to it. 296 * 297 * paging_in_progress is valid always. Lockless references to 298 * the objects may acquire pip and then check OBJ_DEAD. 299 */ 300 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 301 #ifdef INVARIANTS 302 vm_object_zdtor, 303 #else 304 NULL, 305 #endif 306 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 307 308 vm_radix_zinit(); 309 } 310 311 void 312 vm_object_clear_flag(vm_object_t object, u_short bits) 313 { 314 315 VM_OBJECT_ASSERT_WLOCKED(object); 316 object->flags &= ~bits; 317 } 318 319 /* 320 * Sets the default memory attribute for the specified object. Pages 321 * that are allocated to this object are by default assigned this memory 322 * attribute. 323 * 324 * Presently, this function must be called before any pages are allocated 325 * to the object. In the future, this requirement may be relaxed for 326 * "default" and "swap" objects. 327 */ 328 int 329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 330 { 331 332 VM_OBJECT_ASSERT_WLOCKED(object); 333 334 if (object->type == OBJT_DEAD) 335 return (KERN_INVALID_ARGUMENT); 336 if (!TAILQ_EMPTY(&object->memq)) 337 return (KERN_FAILURE); 338 339 object->memattr = memattr; 340 return (KERN_SUCCESS); 341 } 342 343 void 344 vm_object_pip_add(vm_object_t object, short i) 345 { 346 347 if (i > 0) 348 blockcount_acquire(&object->paging_in_progress, i); 349 } 350 351 void 352 vm_object_pip_wakeup(vm_object_t object) 353 { 354 355 vm_object_pip_wakeupn(object, 1); 356 } 357 358 void 359 vm_object_pip_wakeupn(vm_object_t object, short i) 360 { 361 362 if (i > 0) 363 blockcount_release(&object->paging_in_progress, i); 364 } 365 366 /* 367 * Atomically drop the object lock and wait for pip to drain. This protects 368 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 369 * on return. 370 */ 371 static void 372 vm_object_pip_sleep(vm_object_t object, const char *waitid) 373 { 374 375 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 376 waitid, PVM | PDROP); 377 } 378 379 void 380 vm_object_pip_wait(vm_object_t object, const char *waitid) 381 { 382 383 VM_OBJECT_ASSERT_WLOCKED(object); 384 385 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 386 PVM); 387 } 388 389 void 390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 391 { 392 393 VM_OBJECT_ASSERT_UNLOCKED(object); 394 395 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 396 } 397 398 /* 399 * vm_object_allocate: 400 * 401 * Returns a new object with the given size. 402 */ 403 vm_object_t 404 vm_object_allocate(objtype_t type, vm_pindex_t size) 405 { 406 vm_object_t object; 407 u_short flags; 408 409 switch (type) { 410 case OBJT_DEAD: 411 panic("vm_object_allocate: can't create OBJT_DEAD"); 412 case OBJT_DEFAULT: 413 flags = OBJ_COLORED; 414 break; 415 case OBJT_SWAP: 416 flags = OBJ_COLORED | OBJ_SWAP; 417 break; 418 case OBJT_DEVICE: 419 case OBJT_SG: 420 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 421 break; 422 case OBJT_MGTDEVICE: 423 flags = OBJ_FICTITIOUS; 424 break; 425 case OBJT_PHYS: 426 flags = OBJ_UNMANAGED; 427 break; 428 case OBJT_VNODE: 429 flags = 0; 430 break; 431 default: 432 panic("vm_object_allocate: type %d is undefined or dynamic", 433 type); 434 } 435 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 436 _vm_object_allocate(type, size, flags, object, NULL); 437 438 return (object); 439 } 440 441 vm_object_t 442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags) 443 { 444 vm_object_t object; 445 446 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */); 447 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 448 _vm_object_allocate(dyntype, size, flags, object, NULL); 449 450 return (object); 451 } 452 453 /* 454 * vm_object_allocate_anon: 455 * 456 * Returns a new default object of the given size and marked as 457 * anonymous memory for special split/collapse handling. Color 458 * to be initialized by the caller. 459 */ 460 vm_object_t 461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 462 struct ucred *cred, vm_size_t charge) 463 { 464 vm_object_t handle, object; 465 466 if (backing_object == NULL) 467 handle = NULL; 468 else if ((backing_object->flags & OBJ_ANON) != 0) 469 handle = backing_object->handle; 470 else 471 handle = backing_object; 472 object = uma_zalloc(obj_zone, M_WAITOK); 473 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING, 474 object, handle); 475 object->cred = cred; 476 object->charge = cred != NULL ? charge : 0; 477 return (object); 478 } 479 480 static void 481 vm_object_reference_vnode(vm_object_t object) 482 { 483 u_int old; 484 485 /* 486 * vnode objects need the lock for the first reference 487 * to serialize with vnode_object_deallocate(). 488 */ 489 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 490 VM_OBJECT_RLOCK(object); 491 old = refcount_acquire(&object->ref_count); 492 if (object->type == OBJT_VNODE && old == 0) 493 vref(object->handle); 494 VM_OBJECT_RUNLOCK(object); 495 } 496 } 497 498 /* 499 * vm_object_reference: 500 * 501 * Acquires a reference to the given object. 502 */ 503 void 504 vm_object_reference(vm_object_t object) 505 { 506 507 if (object == NULL) 508 return; 509 510 if (object->type == OBJT_VNODE) 511 vm_object_reference_vnode(object); 512 else 513 refcount_acquire(&object->ref_count); 514 KASSERT((object->flags & OBJ_DEAD) == 0, 515 ("vm_object_reference: Referenced dead object.")); 516 } 517 518 /* 519 * vm_object_reference_locked: 520 * 521 * Gets another reference to the given object. 522 * 523 * The object must be locked. 524 */ 525 void 526 vm_object_reference_locked(vm_object_t object) 527 { 528 u_int old; 529 530 VM_OBJECT_ASSERT_LOCKED(object); 531 old = refcount_acquire(&object->ref_count); 532 if (object->type == OBJT_VNODE && old == 0) 533 vref(object->handle); 534 KASSERT((object->flags & OBJ_DEAD) == 0, 535 ("vm_object_reference: Referenced dead object.")); 536 } 537 538 /* 539 * Handle deallocating an object of type OBJT_VNODE. 540 */ 541 static void 542 vm_object_deallocate_vnode(vm_object_t object) 543 { 544 struct vnode *vp = (struct vnode *) object->handle; 545 bool last; 546 547 KASSERT(object->type == OBJT_VNODE, 548 ("vm_object_deallocate_vnode: not a vnode object")); 549 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 550 551 /* Object lock to protect handle lookup. */ 552 last = refcount_release(&object->ref_count); 553 VM_OBJECT_RUNLOCK(object); 554 555 if (!last) 556 return; 557 558 if (!umtx_shm_vnobj_persistent) 559 umtx_shm_object_terminated(object); 560 561 /* vrele may need the vnode lock. */ 562 vrele(vp); 563 } 564 565 /* 566 * We dropped a reference on an object and discovered that it had a 567 * single remaining shadow. This is a sibling of the reference we 568 * dropped. Attempt to collapse the sibling and backing object. 569 */ 570 static vm_object_t 571 vm_object_deallocate_anon(vm_object_t backing_object) 572 { 573 vm_object_t object; 574 575 /* Fetch the final shadow. */ 576 object = LIST_FIRST(&backing_object->shadow_head); 577 KASSERT(object != NULL && backing_object->shadow_count == 1, 578 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 579 backing_object->ref_count, backing_object->shadow_count)); 580 KASSERT((object->flags & OBJ_ANON) != 0, 581 ("invalid shadow object %p", object)); 582 583 if (!VM_OBJECT_TRYWLOCK(object)) { 584 /* 585 * Prevent object from disappearing since we do not have a 586 * reference. 587 */ 588 vm_object_pip_add(object, 1); 589 VM_OBJECT_WUNLOCK(backing_object); 590 VM_OBJECT_WLOCK(object); 591 vm_object_pip_wakeup(object); 592 } else 593 VM_OBJECT_WUNLOCK(backing_object); 594 595 /* 596 * Check for a collapse/terminate race with the last reference holder. 597 */ 598 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 599 !refcount_acquire_if_not_zero(&object->ref_count)) { 600 VM_OBJECT_WUNLOCK(object); 601 return (NULL); 602 } 603 backing_object = object->backing_object; 604 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 605 vm_object_collapse(object); 606 VM_OBJECT_WUNLOCK(object); 607 608 return (object); 609 } 610 611 /* 612 * vm_object_deallocate: 613 * 614 * Release a reference to the specified object, 615 * gained either through a vm_object_allocate 616 * or a vm_object_reference call. When all references 617 * are gone, storage associated with this object 618 * may be relinquished. 619 * 620 * No object may be locked. 621 */ 622 void 623 vm_object_deallocate(vm_object_t object) 624 { 625 vm_object_t temp; 626 bool released; 627 628 while (object != NULL) { 629 /* 630 * If the reference count goes to 0 we start calling 631 * vm_object_terminate() on the object chain. A ref count 632 * of 1 may be a special case depending on the shadow count 633 * being 0 or 1. These cases require a write lock on the 634 * object. 635 */ 636 if ((object->flags & OBJ_ANON) == 0) 637 released = refcount_release_if_gt(&object->ref_count, 1); 638 else 639 released = refcount_release_if_gt(&object->ref_count, 2); 640 if (released) 641 return; 642 643 if (object->type == OBJT_VNODE) { 644 VM_OBJECT_RLOCK(object); 645 if (object->type == OBJT_VNODE) { 646 vm_object_deallocate_vnode(object); 647 return; 648 } 649 VM_OBJECT_RUNLOCK(object); 650 } 651 652 VM_OBJECT_WLOCK(object); 653 KASSERT(object->ref_count > 0, 654 ("vm_object_deallocate: object deallocated too many times: %d", 655 object->type)); 656 657 /* 658 * If this is not the final reference to an anonymous 659 * object we may need to collapse the shadow chain. 660 */ 661 if (!refcount_release(&object->ref_count)) { 662 if (object->ref_count > 1 || 663 object->shadow_count == 0) { 664 if ((object->flags & OBJ_ANON) != 0 && 665 object->ref_count == 1) 666 vm_object_set_flag(object, 667 OBJ_ONEMAPPING); 668 VM_OBJECT_WUNLOCK(object); 669 return; 670 } 671 672 /* Handle collapsing last ref on anonymous objects. */ 673 object = vm_object_deallocate_anon(object); 674 continue; 675 } 676 677 /* 678 * Handle the final reference to an object. We restart 679 * the loop with the backing object to avoid recursion. 680 */ 681 umtx_shm_object_terminated(object); 682 temp = object->backing_object; 683 if (temp != NULL) { 684 KASSERT(object->type == OBJT_DEFAULT || 685 object->type == OBJT_SWAP, 686 ("shadowed tmpfs v_object 2 %p", object)); 687 vm_object_backing_remove(object); 688 } 689 690 KASSERT((object->flags & OBJ_DEAD) == 0, 691 ("vm_object_deallocate: Terminating dead object.")); 692 vm_object_set_flag(object, OBJ_DEAD); 693 vm_object_terminate(object); 694 object = temp; 695 } 696 } 697 698 /* 699 * vm_object_destroy removes the object from the global object list 700 * and frees the space for the object. 701 */ 702 void 703 vm_object_destroy(vm_object_t object) 704 { 705 706 /* 707 * Release the allocation charge. 708 */ 709 if (object->cred != NULL) { 710 swap_release_by_cred(object->charge, object->cred); 711 object->charge = 0; 712 crfree(object->cred); 713 object->cred = NULL; 714 } 715 716 /* 717 * Free the space for the object. 718 */ 719 uma_zfree(obj_zone, object); 720 } 721 722 static void 723 vm_object_backing_remove_locked(vm_object_t object) 724 { 725 vm_object_t backing_object; 726 727 backing_object = object->backing_object; 728 VM_OBJECT_ASSERT_WLOCKED(object); 729 VM_OBJECT_ASSERT_WLOCKED(backing_object); 730 731 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 732 ("vm_object_backing_remove: Removing collapsing object.")); 733 734 if ((object->flags & OBJ_SHADOWLIST) != 0) { 735 LIST_REMOVE(object, shadow_list); 736 backing_object->shadow_count--; 737 object->flags &= ~OBJ_SHADOWLIST; 738 } 739 object->backing_object = NULL; 740 } 741 742 static void 743 vm_object_backing_remove(vm_object_t object) 744 { 745 vm_object_t backing_object; 746 747 VM_OBJECT_ASSERT_WLOCKED(object); 748 749 if ((object->flags & OBJ_SHADOWLIST) != 0) { 750 backing_object = object->backing_object; 751 VM_OBJECT_WLOCK(backing_object); 752 vm_object_backing_remove_locked(object); 753 VM_OBJECT_WUNLOCK(backing_object); 754 } else 755 object->backing_object = NULL; 756 } 757 758 static void 759 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 760 { 761 762 VM_OBJECT_ASSERT_WLOCKED(object); 763 764 if ((backing_object->flags & OBJ_ANON) != 0) { 765 VM_OBJECT_ASSERT_WLOCKED(backing_object); 766 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 767 shadow_list); 768 backing_object->shadow_count++; 769 object->flags |= OBJ_SHADOWLIST; 770 } 771 object->backing_object = backing_object; 772 } 773 774 static void 775 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 776 { 777 778 VM_OBJECT_ASSERT_WLOCKED(object); 779 780 if ((backing_object->flags & OBJ_ANON) != 0) { 781 VM_OBJECT_WLOCK(backing_object); 782 vm_object_backing_insert_locked(object, backing_object); 783 VM_OBJECT_WUNLOCK(backing_object); 784 } else 785 object->backing_object = backing_object; 786 } 787 788 /* 789 * Insert an object into a backing_object's shadow list with an additional 790 * reference to the backing_object added. 791 */ 792 static void 793 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object) 794 { 795 796 VM_OBJECT_ASSERT_WLOCKED(object); 797 798 if ((backing_object->flags & OBJ_ANON) != 0) { 799 VM_OBJECT_WLOCK(backing_object); 800 KASSERT((backing_object->flags & OBJ_DEAD) == 0, 801 ("shadowing dead anonymous object")); 802 vm_object_reference_locked(backing_object); 803 vm_object_backing_insert_locked(object, backing_object); 804 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING); 805 VM_OBJECT_WUNLOCK(backing_object); 806 } else { 807 vm_object_reference(backing_object); 808 object->backing_object = backing_object; 809 } 810 } 811 812 /* 813 * Transfer a backing reference from backing_object to object. 814 */ 815 static void 816 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 817 { 818 vm_object_t new_backing_object; 819 820 /* 821 * Note that the reference to backing_object->backing_object 822 * moves from within backing_object to within object. 823 */ 824 vm_object_backing_remove_locked(object); 825 new_backing_object = backing_object->backing_object; 826 if (new_backing_object == NULL) 827 return; 828 if ((new_backing_object->flags & OBJ_ANON) != 0) { 829 VM_OBJECT_WLOCK(new_backing_object); 830 vm_object_backing_remove_locked(backing_object); 831 vm_object_backing_insert_locked(object, new_backing_object); 832 VM_OBJECT_WUNLOCK(new_backing_object); 833 } else { 834 object->backing_object = new_backing_object; 835 backing_object->backing_object = NULL; 836 } 837 } 838 839 /* 840 * Wait for a concurrent collapse to settle. 841 */ 842 static void 843 vm_object_collapse_wait(vm_object_t object) 844 { 845 846 VM_OBJECT_ASSERT_WLOCKED(object); 847 848 while ((object->flags & OBJ_COLLAPSING) != 0) { 849 vm_object_pip_wait(object, "vmcolwait"); 850 counter_u64_add(object_collapse_waits, 1); 851 } 852 } 853 854 /* 855 * Waits for a backing object to clear a pending collapse and returns 856 * it locked if it is an ANON object. 857 */ 858 static vm_object_t 859 vm_object_backing_collapse_wait(vm_object_t object) 860 { 861 vm_object_t backing_object; 862 863 VM_OBJECT_ASSERT_WLOCKED(object); 864 865 for (;;) { 866 backing_object = object->backing_object; 867 if (backing_object == NULL || 868 (backing_object->flags & OBJ_ANON) == 0) 869 return (NULL); 870 VM_OBJECT_WLOCK(backing_object); 871 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 872 break; 873 VM_OBJECT_WUNLOCK(object); 874 vm_object_pip_sleep(backing_object, "vmbckwait"); 875 counter_u64_add(object_collapse_waits, 1); 876 VM_OBJECT_WLOCK(object); 877 } 878 return (backing_object); 879 } 880 881 /* 882 * vm_object_terminate_pages removes any remaining pageable pages 883 * from the object and resets the object to an empty state. 884 */ 885 static void 886 vm_object_terminate_pages(vm_object_t object) 887 { 888 vm_page_t p, p_next; 889 890 VM_OBJECT_ASSERT_WLOCKED(object); 891 892 /* 893 * Free any remaining pageable pages. This also removes them from the 894 * paging queues. However, don't free wired pages, just remove them 895 * from the object. Rather than incrementally removing each page from 896 * the object, the page and object are reset to any empty state. 897 */ 898 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 899 vm_page_assert_unbusied(p); 900 KASSERT(p->object == object && 901 (p->ref_count & VPRC_OBJREF) != 0, 902 ("vm_object_terminate_pages: page %p is inconsistent", p)); 903 904 p->object = NULL; 905 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 906 VM_CNT_INC(v_pfree); 907 vm_page_free(p); 908 } 909 } 910 911 /* 912 * If the object contained any pages, then reset it to an empty state. 913 * None of the object's fields, including "resident_page_count", were 914 * modified by the preceding loop. 915 */ 916 if (object->resident_page_count != 0) { 917 vm_radix_reclaim_allnodes(&object->rtree); 918 TAILQ_INIT(&object->memq); 919 object->resident_page_count = 0; 920 if (object->type == OBJT_VNODE) 921 vdrop(object->handle); 922 } 923 } 924 925 /* 926 * vm_object_terminate actually destroys the specified object, freeing 927 * up all previously used resources. 928 * 929 * The object must be locked. 930 * This routine may block. 931 */ 932 void 933 vm_object_terminate(vm_object_t object) 934 { 935 936 VM_OBJECT_ASSERT_WLOCKED(object); 937 KASSERT((object->flags & OBJ_DEAD) != 0, 938 ("terminating non-dead obj %p", object)); 939 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 940 ("terminating collapsing obj %p", object)); 941 KASSERT(object->backing_object == NULL, 942 ("terminating shadow obj %p", object)); 943 944 /* 945 * Wait for the pageout daemon and other current users to be 946 * done with the object. Note that new paging_in_progress 947 * users can come after this wait, but they must check 948 * OBJ_DEAD flag set (without unlocking the object), and avoid 949 * the object being terminated. 950 */ 951 vm_object_pip_wait(object, "objtrm"); 952 953 KASSERT(object->ref_count == 0, 954 ("vm_object_terminate: object with references, ref_count=%d", 955 object->ref_count)); 956 957 if ((object->flags & OBJ_PG_DTOR) == 0) 958 vm_object_terminate_pages(object); 959 960 #if VM_NRESERVLEVEL > 0 961 if (__predict_false(!LIST_EMPTY(&object->rvq))) 962 vm_reserv_break_all(object); 963 #endif 964 965 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 966 (object->flags & OBJ_SWAP) != 0, 967 ("%s: non-swap obj %p has cred", __func__, object)); 968 969 /* 970 * Let the pager know object is dead. 971 */ 972 vm_pager_deallocate(object); 973 VM_OBJECT_WUNLOCK(object); 974 975 vm_object_destroy(object); 976 } 977 978 /* 979 * Make the page read-only so that we can clear the object flags. However, if 980 * this is a nosync mmap then the object is likely to stay dirty so do not 981 * mess with the page and do not clear the object flags. Returns TRUE if the 982 * page should be flushed, and FALSE otherwise. 983 */ 984 static boolean_t 985 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) 986 { 987 988 vm_page_assert_busied(p); 989 990 /* 991 * If we have been asked to skip nosync pages and this is a 992 * nosync page, skip it. Note that the object flags were not 993 * cleared in this case so we do not have to set them. 994 */ 995 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { 996 *allclean = FALSE; 997 return (FALSE); 998 } else { 999 pmap_remove_write(p); 1000 return (p->dirty != 0); 1001 } 1002 } 1003 1004 /* 1005 * vm_object_page_clean 1006 * 1007 * Clean all dirty pages in the specified range of object. Leaves page 1008 * on whatever queue it is currently on. If NOSYNC is set then do not 1009 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1010 * leaving the object dirty. 1011 * 1012 * For swap objects backing tmpfs regular files, do not flush anything, 1013 * but remove write protection on the mapped pages to update mtime through 1014 * mmaped writes. 1015 * 1016 * When stuffing pages asynchronously, allow clustering. XXX we need a 1017 * synchronous clustering mode implementation. 1018 * 1019 * Odd semantics: if start == end, we clean everything. 1020 * 1021 * The object must be locked. 1022 * 1023 * Returns FALSE if some page from the range was not written, as 1024 * reported by the pager, and TRUE otherwise. 1025 */ 1026 boolean_t 1027 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1028 int flags) 1029 { 1030 vm_page_t np, p; 1031 vm_pindex_t pi, tend, tstart; 1032 int curgeneration, n, pagerflags; 1033 boolean_t eio, res, allclean; 1034 1035 VM_OBJECT_ASSERT_WLOCKED(object); 1036 1037 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1038 return (TRUE); 1039 1040 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1041 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1042 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1043 1044 tstart = OFF_TO_IDX(start); 1045 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1046 allclean = tstart == 0 && tend >= object->size; 1047 res = TRUE; 1048 1049 rescan: 1050 curgeneration = object->generation; 1051 1052 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 1053 pi = p->pindex; 1054 if (pi >= tend) 1055 break; 1056 np = TAILQ_NEXT(p, listq); 1057 if (vm_page_none_valid(p)) 1058 continue; 1059 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { 1060 if (object->generation != curgeneration && 1061 (flags & OBJPC_SYNC) != 0) 1062 goto rescan; 1063 np = vm_page_find_least(object, pi); 1064 continue; 1065 } 1066 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1067 vm_page_xunbusy(p); 1068 continue; 1069 } 1070 if (object->type == OBJT_VNODE) { 1071 n = vm_object_page_collect_flush(object, p, pagerflags, 1072 flags, &allclean, &eio); 1073 if (eio) { 1074 res = FALSE; 1075 allclean = FALSE; 1076 } 1077 if (object->generation != curgeneration && 1078 (flags & OBJPC_SYNC) != 0) 1079 goto rescan; 1080 1081 /* 1082 * If the VOP_PUTPAGES() did a truncated write, so 1083 * that even the first page of the run is not fully 1084 * written, vm_pageout_flush() returns 0 as the run 1085 * length. Since the condition that caused truncated 1086 * write may be permanent, e.g. exhausted free space, 1087 * accepting n == 0 would cause an infinite loop. 1088 * 1089 * Forwarding the iterator leaves the unwritten page 1090 * behind, but there is not much we can do there if 1091 * filesystem refuses to write it. 1092 */ 1093 if (n == 0) { 1094 n = 1; 1095 allclean = FALSE; 1096 } 1097 } else { 1098 n = 1; 1099 vm_page_xunbusy(p); 1100 } 1101 np = vm_page_find_least(object, pi + n); 1102 } 1103 #if 0 1104 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1105 #endif 1106 1107 /* 1108 * Leave updating cleangeneration for tmpfs objects to tmpfs 1109 * scan. It needs to update mtime, which happens for other 1110 * filesystems during page writeouts. 1111 */ 1112 if (allclean && object->type == OBJT_VNODE) 1113 object->cleangeneration = curgeneration; 1114 return (res); 1115 } 1116 1117 static int 1118 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1119 int flags, boolean_t *allclean, boolean_t *eio) 1120 { 1121 vm_page_t ma[vm_pageout_page_count], p_first, tp; 1122 int count, i, mreq, runlen; 1123 1124 vm_page_lock_assert(p, MA_NOTOWNED); 1125 vm_page_assert_xbusied(p); 1126 VM_OBJECT_ASSERT_WLOCKED(object); 1127 1128 count = 1; 1129 mreq = 0; 1130 1131 for (tp = p; count < vm_pageout_page_count; count++) { 1132 tp = vm_page_next(tp); 1133 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1134 break; 1135 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1136 vm_page_xunbusy(tp); 1137 break; 1138 } 1139 } 1140 1141 for (p_first = p; count < vm_pageout_page_count; count++) { 1142 tp = vm_page_prev(p_first); 1143 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1144 break; 1145 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1146 vm_page_xunbusy(tp); 1147 break; 1148 } 1149 p_first = tp; 1150 mreq++; 1151 } 1152 1153 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 1154 ma[i] = tp; 1155 1156 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 1157 return (runlen); 1158 } 1159 1160 /* 1161 * Note that there is absolutely no sense in writing out 1162 * anonymous objects, so we track down the vnode object 1163 * to write out. 1164 * We invalidate (remove) all pages from the address space 1165 * for semantic correctness. 1166 * 1167 * If the backing object is a device object with unmanaged pages, then any 1168 * mappings to the specified range of pages must be removed before this 1169 * function is called. 1170 * 1171 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1172 * may start out with a NULL object. 1173 */ 1174 boolean_t 1175 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1176 boolean_t syncio, boolean_t invalidate) 1177 { 1178 vm_object_t backing_object; 1179 struct vnode *vp; 1180 struct mount *mp; 1181 int error, flags, fsync_after; 1182 boolean_t res; 1183 1184 if (object == NULL) 1185 return (TRUE); 1186 res = TRUE; 1187 error = 0; 1188 VM_OBJECT_WLOCK(object); 1189 while ((backing_object = object->backing_object) != NULL) { 1190 VM_OBJECT_WLOCK(backing_object); 1191 offset += object->backing_object_offset; 1192 VM_OBJECT_WUNLOCK(object); 1193 object = backing_object; 1194 if (object->size < OFF_TO_IDX(offset + size)) 1195 size = IDX_TO_OFF(object->size) - offset; 1196 } 1197 /* 1198 * Flush pages if writing is allowed, invalidate them 1199 * if invalidation requested. Pages undergoing I/O 1200 * will be ignored by vm_object_page_remove(). 1201 * 1202 * We cannot lock the vnode and then wait for paging 1203 * to complete without deadlocking against vm_fault. 1204 * Instead we simply call vm_object_page_remove() and 1205 * allow it to block internally on a page-by-page 1206 * basis when it encounters pages undergoing async 1207 * I/O. 1208 */ 1209 if (object->type == OBJT_VNODE && 1210 vm_object_mightbedirty(object) != 0 && 1211 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1212 VM_OBJECT_WUNLOCK(object); 1213 (void) vn_start_write(vp, &mp, V_WAIT); 1214 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1215 if (syncio && !invalidate && offset == 0 && 1216 atop(size) == object->size) { 1217 /* 1218 * If syncing the whole mapping of the file, 1219 * it is faster to schedule all the writes in 1220 * async mode, also allowing the clustering, 1221 * and then wait for i/o to complete. 1222 */ 1223 flags = 0; 1224 fsync_after = TRUE; 1225 } else { 1226 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1227 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1228 fsync_after = FALSE; 1229 } 1230 VM_OBJECT_WLOCK(object); 1231 res = vm_object_page_clean(object, offset, offset + size, 1232 flags); 1233 VM_OBJECT_WUNLOCK(object); 1234 if (fsync_after) 1235 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1236 VOP_UNLOCK(vp); 1237 vn_finished_write(mp); 1238 if (error != 0) 1239 res = FALSE; 1240 VM_OBJECT_WLOCK(object); 1241 } 1242 if ((object->type == OBJT_VNODE || 1243 object->type == OBJT_DEVICE) && invalidate) { 1244 if (object->type == OBJT_DEVICE) 1245 /* 1246 * The option OBJPR_NOTMAPPED must be passed here 1247 * because vm_object_page_remove() cannot remove 1248 * unmanaged mappings. 1249 */ 1250 flags = OBJPR_NOTMAPPED; 1251 else if (old_msync) 1252 flags = 0; 1253 else 1254 flags = OBJPR_CLEANONLY; 1255 vm_object_page_remove(object, OFF_TO_IDX(offset), 1256 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1257 } 1258 VM_OBJECT_WUNLOCK(object); 1259 return (res); 1260 } 1261 1262 /* 1263 * Determine whether the given advice can be applied to the object. Advice is 1264 * not applied to unmanaged pages since they never belong to page queues, and 1265 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1266 * have been mapped at most once. 1267 */ 1268 static bool 1269 vm_object_advice_applies(vm_object_t object, int advice) 1270 { 1271 1272 if ((object->flags & OBJ_UNMANAGED) != 0) 1273 return (false); 1274 if (advice != MADV_FREE) 1275 return (true); 1276 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1277 (OBJ_ONEMAPPING | OBJ_ANON)); 1278 } 1279 1280 static void 1281 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1282 vm_size_t size) 1283 { 1284 1285 if (advice == MADV_FREE) 1286 vm_pager_freespace(object, pindex, size); 1287 } 1288 1289 /* 1290 * vm_object_madvise: 1291 * 1292 * Implements the madvise function at the object/page level. 1293 * 1294 * MADV_WILLNEED (any object) 1295 * 1296 * Activate the specified pages if they are resident. 1297 * 1298 * MADV_DONTNEED (any object) 1299 * 1300 * Deactivate the specified pages if they are resident. 1301 * 1302 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1303 * OBJ_ONEMAPPING only) 1304 * 1305 * Deactivate and clean the specified pages if they are 1306 * resident. This permits the process to reuse the pages 1307 * without faulting or the kernel to reclaim the pages 1308 * without I/O. 1309 */ 1310 void 1311 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1312 int advice) 1313 { 1314 vm_pindex_t tpindex; 1315 vm_object_t backing_object, tobject; 1316 vm_page_t m, tm; 1317 1318 if (object == NULL) 1319 return; 1320 1321 relookup: 1322 VM_OBJECT_WLOCK(object); 1323 if (!vm_object_advice_applies(object, advice)) { 1324 VM_OBJECT_WUNLOCK(object); 1325 return; 1326 } 1327 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1328 tobject = object; 1329 1330 /* 1331 * If the next page isn't resident in the top-level object, we 1332 * need to search the shadow chain. When applying MADV_FREE, we 1333 * take care to release any swap space used to store 1334 * non-resident pages. 1335 */ 1336 if (m == NULL || pindex < m->pindex) { 1337 /* 1338 * Optimize a common case: if the top-level object has 1339 * no backing object, we can skip over the non-resident 1340 * range in constant time. 1341 */ 1342 if (object->backing_object == NULL) { 1343 tpindex = (m != NULL && m->pindex < end) ? 1344 m->pindex : end; 1345 vm_object_madvise_freespace(object, advice, 1346 pindex, tpindex - pindex); 1347 if ((pindex = tpindex) == end) 1348 break; 1349 goto next_page; 1350 } 1351 1352 tpindex = pindex; 1353 do { 1354 vm_object_madvise_freespace(tobject, advice, 1355 tpindex, 1); 1356 /* 1357 * Prepare to search the next object in the 1358 * chain. 1359 */ 1360 backing_object = tobject->backing_object; 1361 if (backing_object == NULL) 1362 goto next_pindex; 1363 VM_OBJECT_WLOCK(backing_object); 1364 tpindex += 1365 OFF_TO_IDX(tobject->backing_object_offset); 1366 if (tobject != object) 1367 VM_OBJECT_WUNLOCK(tobject); 1368 tobject = backing_object; 1369 if (!vm_object_advice_applies(tobject, advice)) 1370 goto next_pindex; 1371 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1372 NULL); 1373 } else { 1374 next_page: 1375 tm = m; 1376 m = TAILQ_NEXT(m, listq); 1377 } 1378 1379 /* 1380 * If the page is not in a normal state, skip it. The page 1381 * can not be invalidated while the object lock is held. 1382 */ 1383 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1384 goto next_pindex; 1385 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1386 ("vm_object_madvise: page %p is fictitious", tm)); 1387 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1388 ("vm_object_madvise: page %p is not managed", tm)); 1389 if (vm_page_tryxbusy(tm) == 0) { 1390 if (object != tobject) 1391 VM_OBJECT_WUNLOCK(object); 1392 if (advice == MADV_WILLNEED) { 1393 /* 1394 * Reference the page before unlocking and 1395 * sleeping so that the page daemon is less 1396 * likely to reclaim it. 1397 */ 1398 vm_page_aflag_set(tm, PGA_REFERENCED); 1399 } 1400 vm_page_busy_sleep(tm, "madvpo", false); 1401 goto relookup; 1402 } 1403 vm_page_advise(tm, advice); 1404 vm_page_xunbusy(tm); 1405 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1406 next_pindex: 1407 if (tobject != object) 1408 VM_OBJECT_WUNLOCK(tobject); 1409 } 1410 VM_OBJECT_WUNLOCK(object); 1411 } 1412 1413 /* 1414 * vm_object_shadow: 1415 * 1416 * Create a new object which is backed by the 1417 * specified existing object range. The source 1418 * object reference is deallocated. 1419 * 1420 * The new object and offset into that object 1421 * are returned in the source parameters. 1422 */ 1423 void 1424 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1425 struct ucred *cred, bool shared) 1426 { 1427 vm_object_t source; 1428 vm_object_t result; 1429 1430 source = *object; 1431 1432 /* 1433 * Don't create the new object if the old object isn't shared. 1434 * 1435 * If we hold the only reference we can guarantee that it won't 1436 * increase while we have the map locked. Otherwise the race is 1437 * harmless and we will end up with an extra shadow object that 1438 * will be collapsed later. 1439 */ 1440 if (source != NULL && source->ref_count == 1 && 1441 (source->flags & OBJ_ANON) != 0) 1442 return; 1443 1444 /* 1445 * Allocate a new object with the given length. 1446 */ 1447 result = vm_object_allocate_anon(atop(length), source, cred, length); 1448 1449 /* 1450 * Store the offset into the source object, and fix up the offset into 1451 * the new object. 1452 */ 1453 result->backing_object_offset = *offset; 1454 1455 if (shared || source != NULL) { 1456 VM_OBJECT_WLOCK(result); 1457 1458 /* 1459 * The new object shadows the source object, adding a 1460 * reference to it. Our caller changes his reference 1461 * to point to the new object, removing a reference to 1462 * the source object. Net result: no change of 1463 * reference count, unless the caller needs to add one 1464 * more reference due to forking a shared map entry. 1465 */ 1466 if (shared) { 1467 vm_object_reference_locked(result); 1468 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1469 } 1470 1471 /* 1472 * Try to optimize the result object's page color when 1473 * shadowing in order to maintain page coloring 1474 * consistency in the combined shadowed object. 1475 */ 1476 if (source != NULL) { 1477 vm_object_backing_insert(result, source); 1478 result->domain = source->domain; 1479 #if VM_NRESERVLEVEL > 0 1480 result->flags |= source->flags & OBJ_COLORED; 1481 result->pg_color = (source->pg_color + 1482 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1483 1)) - 1); 1484 #endif 1485 } 1486 VM_OBJECT_WUNLOCK(result); 1487 } 1488 1489 /* 1490 * Return the new things 1491 */ 1492 *offset = 0; 1493 *object = result; 1494 } 1495 1496 /* 1497 * vm_object_split: 1498 * 1499 * Split the pages in a map entry into a new object. This affords 1500 * easier removal of unused pages, and keeps object inheritance from 1501 * being a negative impact on memory usage. 1502 */ 1503 void 1504 vm_object_split(vm_map_entry_t entry) 1505 { 1506 vm_page_t m, m_busy, m_next; 1507 vm_object_t orig_object, new_object, backing_object; 1508 vm_pindex_t idx, offidxstart; 1509 vm_size_t size; 1510 1511 orig_object = entry->object.vm_object; 1512 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1513 ("vm_object_split: Splitting object with multiple mappings.")); 1514 if ((orig_object->flags & OBJ_ANON) == 0) 1515 return; 1516 if (orig_object->ref_count <= 1) 1517 return; 1518 VM_OBJECT_WUNLOCK(orig_object); 1519 1520 offidxstart = OFF_TO_IDX(entry->offset); 1521 size = atop(entry->end - entry->start); 1522 1523 /* 1524 * If swap_pager_copy() is later called, it will convert new_object 1525 * into a swap object. 1526 */ 1527 new_object = vm_object_allocate_anon(size, orig_object, 1528 orig_object->cred, ptoa(size)); 1529 1530 /* 1531 * We must wait for the orig_object to complete any in-progress 1532 * collapse so that the swap blocks are stable below. The 1533 * additional reference on backing_object by new object will 1534 * prevent further collapse operations until split completes. 1535 */ 1536 VM_OBJECT_WLOCK(orig_object); 1537 vm_object_collapse_wait(orig_object); 1538 1539 /* 1540 * At this point, the new object is still private, so the order in 1541 * which the original and new objects are locked does not matter. 1542 */ 1543 VM_OBJECT_WLOCK(new_object); 1544 new_object->domain = orig_object->domain; 1545 backing_object = orig_object->backing_object; 1546 if (backing_object != NULL) { 1547 vm_object_backing_insert_ref(new_object, backing_object); 1548 new_object->backing_object_offset = 1549 orig_object->backing_object_offset + entry->offset; 1550 } 1551 if (orig_object->cred != NULL) { 1552 crhold(orig_object->cred); 1553 KASSERT(orig_object->charge >= ptoa(size), 1554 ("orig_object->charge < 0")); 1555 orig_object->charge -= ptoa(size); 1556 } 1557 1558 /* 1559 * Mark the split operation so that swap_pager_getpages() knows 1560 * that the object is in transition. 1561 */ 1562 vm_object_set_flag(orig_object, OBJ_SPLIT); 1563 m_busy = NULL; 1564 #ifdef INVARIANTS 1565 idx = 0; 1566 #endif 1567 retry: 1568 m = vm_page_find_least(orig_object, offidxstart); 1569 KASSERT(m == NULL || idx <= m->pindex - offidxstart, 1570 ("%s: object %p was repopulated", __func__, orig_object)); 1571 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1572 m = m_next) { 1573 m_next = TAILQ_NEXT(m, listq); 1574 1575 /* 1576 * We must wait for pending I/O to complete before we can 1577 * rename the page. 1578 * 1579 * We do not have to VM_PROT_NONE the page as mappings should 1580 * not be changed by this operation. 1581 */ 1582 if (vm_page_tryxbusy(m) == 0) { 1583 VM_OBJECT_WUNLOCK(new_object); 1584 vm_page_sleep_if_busy(m, "spltwt"); 1585 VM_OBJECT_WLOCK(new_object); 1586 goto retry; 1587 } 1588 1589 /* 1590 * The page was left invalid. Likely placed there by 1591 * an incomplete fault. Just remove and ignore. 1592 */ 1593 if (vm_page_none_valid(m)) { 1594 if (vm_page_remove(m)) 1595 vm_page_free(m); 1596 continue; 1597 } 1598 1599 /* vm_page_rename() will dirty the page. */ 1600 if (vm_page_rename(m, new_object, idx)) { 1601 vm_page_xunbusy(m); 1602 VM_OBJECT_WUNLOCK(new_object); 1603 VM_OBJECT_WUNLOCK(orig_object); 1604 vm_radix_wait(); 1605 VM_OBJECT_WLOCK(orig_object); 1606 VM_OBJECT_WLOCK(new_object); 1607 goto retry; 1608 } 1609 1610 #if VM_NRESERVLEVEL > 0 1611 /* 1612 * If some of the reservation's allocated pages remain with 1613 * the original object, then transferring the reservation to 1614 * the new object is neither particularly beneficial nor 1615 * particularly harmful as compared to leaving the reservation 1616 * with the original object. If, however, all of the 1617 * reservation's allocated pages are transferred to the new 1618 * object, then transferring the reservation is typically 1619 * beneficial. Determining which of these two cases applies 1620 * would be more costly than unconditionally renaming the 1621 * reservation. 1622 */ 1623 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1624 #endif 1625 1626 /* 1627 * orig_object's type may change while sleeping, so keep track 1628 * of the beginning of the busied range. 1629 */ 1630 if (orig_object->type != OBJT_SWAP) 1631 vm_page_xunbusy(m); 1632 else if (m_busy == NULL) 1633 m_busy = m; 1634 } 1635 if ((orig_object->flags & OBJ_SWAP) != 0) { 1636 /* 1637 * swap_pager_copy() can sleep, in which case the orig_object's 1638 * and new_object's locks are released and reacquired. 1639 */ 1640 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1641 if (m_busy != NULL) 1642 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq) 1643 vm_page_xunbusy(m_busy); 1644 } 1645 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1646 VM_OBJECT_WUNLOCK(orig_object); 1647 VM_OBJECT_WUNLOCK(new_object); 1648 entry->object.vm_object = new_object; 1649 entry->offset = 0LL; 1650 vm_object_deallocate(orig_object); 1651 VM_OBJECT_WLOCK(new_object); 1652 } 1653 1654 static vm_page_t 1655 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p) 1656 { 1657 vm_object_t backing_object; 1658 1659 VM_OBJECT_ASSERT_WLOCKED(object); 1660 backing_object = object->backing_object; 1661 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1662 1663 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1664 ("invalid ownership %p %p %p", p, object, backing_object)); 1665 /* The page is only NULL when rename fails. */ 1666 if (p == NULL) { 1667 VM_OBJECT_WUNLOCK(object); 1668 VM_OBJECT_WUNLOCK(backing_object); 1669 vm_radix_wait(); 1670 } else { 1671 if (p->object == object) 1672 VM_OBJECT_WUNLOCK(backing_object); 1673 else 1674 VM_OBJECT_WUNLOCK(object); 1675 vm_page_busy_sleep(p, "vmocol", false); 1676 } 1677 VM_OBJECT_WLOCK(object); 1678 VM_OBJECT_WLOCK(backing_object); 1679 return (TAILQ_FIRST(&backing_object->memq)); 1680 } 1681 1682 static bool 1683 vm_object_scan_all_shadowed(vm_object_t object) 1684 { 1685 vm_object_t backing_object; 1686 vm_page_t p, pp; 1687 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1688 1689 VM_OBJECT_ASSERT_WLOCKED(object); 1690 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1691 1692 backing_object = object->backing_object; 1693 1694 if ((backing_object->flags & OBJ_ANON) == 0) 1695 return (false); 1696 1697 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1698 p = vm_page_find_least(backing_object, pi); 1699 ps = swap_pager_find_least(backing_object, pi); 1700 1701 /* 1702 * Only check pages inside the parent object's range and 1703 * inside the parent object's mapping of the backing object. 1704 */ 1705 for (;; pi++) { 1706 if (p != NULL && p->pindex < pi) 1707 p = TAILQ_NEXT(p, listq); 1708 if (ps < pi) 1709 ps = swap_pager_find_least(backing_object, pi); 1710 if (p == NULL && ps >= backing_object->size) 1711 break; 1712 else if (p == NULL) 1713 pi = ps; 1714 else 1715 pi = MIN(p->pindex, ps); 1716 1717 new_pindex = pi - backing_offset_index; 1718 if (new_pindex >= object->size) 1719 break; 1720 1721 if (p != NULL) { 1722 /* 1723 * If the backing object page is busy a 1724 * grandparent or older page may still be 1725 * undergoing CoW. It is not safe to collapse 1726 * the backing object until it is quiesced. 1727 */ 1728 if (vm_page_tryxbusy(p) == 0) 1729 return (false); 1730 1731 /* 1732 * We raced with the fault handler that left 1733 * newly allocated invalid page on the object 1734 * queue and retried. 1735 */ 1736 if (!vm_page_all_valid(p)) 1737 goto unbusy_ret; 1738 } 1739 1740 /* 1741 * See if the parent has the page or if the parent's object 1742 * pager has the page. If the parent has the page but the page 1743 * is not valid, the parent's object pager must have the page. 1744 * 1745 * If this fails, the parent does not completely shadow the 1746 * object and we might as well give up now. 1747 */ 1748 pp = vm_page_lookup(object, new_pindex); 1749 1750 /* 1751 * The valid check here is stable due to object lock 1752 * being required to clear valid and initiate paging. 1753 * Busy of p disallows fault handler to validate pp. 1754 */ 1755 if ((pp == NULL || vm_page_none_valid(pp)) && 1756 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1757 goto unbusy_ret; 1758 if (p != NULL) 1759 vm_page_xunbusy(p); 1760 } 1761 return (true); 1762 1763 unbusy_ret: 1764 if (p != NULL) 1765 vm_page_xunbusy(p); 1766 return (false); 1767 } 1768 1769 static void 1770 vm_object_collapse_scan(vm_object_t object) 1771 { 1772 vm_object_t backing_object; 1773 vm_page_t next, p, pp; 1774 vm_pindex_t backing_offset_index, new_pindex; 1775 1776 VM_OBJECT_ASSERT_WLOCKED(object); 1777 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1778 1779 backing_object = object->backing_object; 1780 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1781 1782 /* 1783 * Our scan 1784 */ 1785 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1786 next = TAILQ_NEXT(p, listq); 1787 new_pindex = p->pindex - backing_offset_index; 1788 1789 /* 1790 * Check for busy page 1791 */ 1792 if (vm_page_tryxbusy(p) == 0) { 1793 next = vm_object_collapse_scan_wait(object, p); 1794 continue; 1795 } 1796 1797 KASSERT(object->backing_object == backing_object, 1798 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1799 object->backing_object, backing_object)); 1800 KASSERT(p->object == backing_object, 1801 ("vm_object_collapse_scan: object mismatch %p != %p", 1802 p->object, backing_object)); 1803 1804 if (p->pindex < backing_offset_index || 1805 new_pindex >= object->size) { 1806 vm_pager_freespace(backing_object, p->pindex, 1); 1807 1808 KASSERT(!pmap_page_is_mapped(p), 1809 ("freeing mapped page %p", p)); 1810 if (vm_page_remove(p)) 1811 vm_page_free(p); 1812 continue; 1813 } 1814 1815 if (!vm_page_all_valid(p)) { 1816 KASSERT(!pmap_page_is_mapped(p), 1817 ("freeing mapped page %p", p)); 1818 if (vm_page_remove(p)) 1819 vm_page_free(p); 1820 continue; 1821 } 1822 1823 pp = vm_page_lookup(object, new_pindex); 1824 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1825 vm_page_xunbusy(p); 1826 /* 1827 * The page in the parent is busy and possibly not 1828 * (yet) valid. Until its state is finalized by the 1829 * busy bit owner, we can't tell whether it shadows the 1830 * original page. 1831 */ 1832 next = vm_object_collapse_scan_wait(object, pp); 1833 continue; 1834 } 1835 1836 if (pp != NULL && vm_page_none_valid(pp)) { 1837 /* 1838 * The page was invalid in the parent. Likely placed 1839 * there by an incomplete fault. Just remove and 1840 * ignore. p can replace it. 1841 */ 1842 if (vm_page_remove(pp)) 1843 vm_page_free(pp); 1844 pp = NULL; 1845 } 1846 1847 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1848 NULL)) { 1849 /* 1850 * The page already exists in the parent OR swap exists 1851 * for this location in the parent. Leave the parent's 1852 * page alone. Destroy the original page from the 1853 * backing object. 1854 */ 1855 vm_pager_freespace(backing_object, p->pindex, 1); 1856 KASSERT(!pmap_page_is_mapped(p), 1857 ("freeing mapped page %p", p)); 1858 if (vm_page_remove(p)) 1859 vm_page_free(p); 1860 if (pp != NULL) 1861 vm_page_xunbusy(pp); 1862 continue; 1863 } 1864 1865 /* 1866 * Page does not exist in parent, rename the page from the 1867 * backing object to the main object. 1868 * 1869 * If the page was mapped to a process, it can remain mapped 1870 * through the rename. vm_page_rename() will dirty the page. 1871 */ 1872 if (vm_page_rename(p, object, new_pindex)) { 1873 vm_page_xunbusy(p); 1874 next = vm_object_collapse_scan_wait(object, NULL); 1875 continue; 1876 } 1877 1878 /* Use the old pindex to free the right page. */ 1879 vm_pager_freespace(backing_object, new_pindex + 1880 backing_offset_index, 1); 1881 1882 #if VM_NRESERVLEVEL > 0 1883 /* 1884 * Rename the reservation. 1885 */ 1886 vm_reserv_rename(p, object, backing_object, 1887 backing_offset_index); 1888 #endif 1889 vm_page_xunbusy(p); 1890 } 1891 return; 1892 } 1893 1894 /* 1895 * vm_object_collapse: 1896 * 1897 * Collapse an object with the object backing it. 1898 * Pages in the backing object are moved into the 1899 * parent, and the backing object is deallocated. 1900 */ 1901 void 1902 vm_object_collapse(vm_object_t object) 1903 { 1904 vm_object_t backing_object, new_backing_object; 1905 1906 VM_OBJECT_ASSERT_WLOCKED(object); 1907 1908 while (TRUE) { 1909 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1910 ("collapsing invalid object")); 1911 1912 /* 1913 * Wait for the backing_object to finish any pending 1914 * collapse so that the caller sees the shortest possible 1915 * shadow chain. 1916 */ 1917 backing_object = vm_object_backing_collapse_wait(object); 1918 if (backing_object == NULL) 1919 return; 1920 1921 KASSERT(object->ref_count > 0 && 1922 object->ref_count > object->shadow_count, 1923 ("collapse with invalid ref %d or shadow %d count.", 1924 object->ref_count, object->shadow_count)); 1925 KASSERT((backing_object->flags & 1926 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1927 ("vm_object_collapse: Backing object already collapsing.")); 1928 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1929 ("vm_object_collapse: object is already collapsing.")); 1930 1931 /* 1932 * We know that we can either collapse the backing object if 1933 * the parent is the only reference to it, or (perhaps) have 1934 * the parent bypass the object if the parent happens to shadow 1935 * all the resident pages in the entire backing object. 1936 */ 1937 if (backing_object->ref_count == 1) { 1938 KASSERT(backing_object->shadow_count == 1, 1939 ("vm_object_collapse: shadow_count: %d", 1940 backing_object->shadow_count)); 1941 vm_object_pip_add(object, 1); 1942 vm_object_set_flag(object, OBJ_COLLAPSING); 1943 vm_object_pip_add(backing_object, 1); 1944 vm_object_set_flag(backing_object, OBJ_DEAD); 1945 1946 /* 1947 * If there is exactly one reference to the backing 1948 * object, we can collapse it into the parent. 1949 */ 1950 vm_object_collapse_scan(object); 1951 1952 #if VM_NRESERVLEVEL > 0 1953 /* 1954 * Break any reservations from backing_object. 1955 */ 1956 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1957 vm_reserv_break_all(backing_object); 1958 #endif 1959 1960 /* 1961 * Move the pager from backing_object to object. 1962 */ 1963 if ((backing_object->flags & OBJ_SWAP) != 0) { 1964 /* 1965 * swap_pager_copy() can sleep, in which case 1966 * the backing_object's and object's locks are 1967 * released and reacquired. 1968 * Since swap_pager_copy() is being asked to 1969 * destroy backing_object, it will change the 1970 * type to OBJT_DEFAULT. 1971 */ 1972 swap_pager_copy( 1973 backing_object, 1974 object, 1975 OFF_TO_IDX(object->backing_object_offset), TRUE); 1976 } 1977 1978 /* 1979 * Object now shadows whatever backing_object did. 1980 */ 1981 vm_object_clear_flag(object, OBJ_COLLAPSING); 1982 vm_object_backing_transfer(object, backing_object); 1983 object->backing_object_offset += 1984 backing_object->backing_object_offset; 1985 VM_OBJECT_WUNLOCK(object); 1986 vm_object_pip_wakeup(object); 1987 1988 /* 1989 * Discard backing_object. 1990 * 1991 * Since the backing object has no pages, no pager left, 1992 * and no object references within it, all that is 1993 * necessary is to dispose of it. 1994 */ 1995 KASSERT(backing_object->ref_count == 1, ( 1996 "backing_object %p was somehow re-referenced during collapse!", 1997 backing_object)); 1998 vm_object_pip_wakeup(backing_object); 1999 (void)refcount_release(&backing_object->ref_count); 2000 vm_object_terminate(backing_object); 2001 counter_u64_add(object_collapses, 1); 2002 VM_OBJECT_WLOCK(object); 2003 } else { 2004 /* 2005 * If we do not entirely shadow the backing object, 2006 * there is nothing we can do so we give up. 2007 * 2008 * The object lock and backing_object lock must not 2009 * be dropped during this sequence. 2010 */ 2011 if (!vm_object_scan_all_shadowed(object)) { 2012 VM_OBJECT_WUNLOCK(backing_object); 2013 break; 2014 } 2015 2016 /* 2017 * Make the parent shadow the next object in the 2018 * chain. Deallocating backing_object will not remove 2019 * it, since its reference count is at least 2. 2020 */ 2021 vm_object_backing_remove_locked(object); 2022 new_backing_object = backing_object->backing_object; 2023 if (new_backing_object != NULL) { 2024 vm_object_backing_insert_ref(object, 2025 new_backing_object); 2026 object->backing_object_offset += 2027 backing_object->backing_object_offset; 2028 } 2029 2030 /* 2031 * Drop the reference count on backing_object. Since 2032 * its ref_count was at least 2, it will not vanish. 2033 */ 2034 (void)refcount_release(&backing_object->ref_count); 2035 KASSERT(backing_object->ref_count >= 1, ( 2036 "backing_object %p was somehow dereferenced during collapse!", 2037 backing_object)); 2038 VM_OBJECT_WUNLOCK(backing_object); 2039 counter_u64_add(object_bypasses, 1); 2040 } 2041 2042 /* 2043 * Try again with this object's new backing object. 2044 */ 2045 } 2046 } 2047 2048 /* 2049 * vm_object_page_remove: 2050 * 2051 * For the given object, either frees or invalidates each of the 2052 * specified pages. In general, a page is freed. However, if a page is 2053 * wired for any reason other than the existence of a managed, wired 2054 * mapping, then it may be invalidated but not removed from the object. 2055 * Pages are specified by the given range ["start", "end") and the option 2056 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 2057 * extends from "start" to the end of the object. If the option 2058 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 2059 * specified range are affected. If the option OBJPR_NOTMAPPED is 2060 * specified, then the pages within the specified range must have no 2061 * mappings. Otherwise, if this option is not specified, any mappings to 2062 * the specified pages are removed before the pages are freed or 2063 * invalidated. 2064 * 2065 * In general, this operation should only be performed on objects that 2066 * contain managed pages. There are, however, two exceptions. First, it 2067 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 2068 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 2069 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 2070 * not be specified and the option OBJPR_NOTMAPPED must be specified. 2071 * 2072 * The object must be locked. 2073 */ 2074 void 2075 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2076 int options) 2077 { 2078 vm_page_t p, next; 2079 2080 VM_OBJECT_ASSERT_WLOCKED(object); 2081 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 2082 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 2083 ("vm_object_page_remove: illegal options for object %p", object)); 2084 if (object->resident_page_count == 0) 2085 return; 2086 vm_object_pip_add(object, 1); 2087 again: 2088 p = vm_page_find_least(object, start); 2089 2090 /* 2091 * Here, the variable "p" is either (1) the page with the least pindex 2092 * greater than or equal to the parameter "start" or (2) NULL. 2093 */ 2094 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2095 next = TAILQ_NEXT(p, listq); 2096 2097 /* 2098 * If the page is wired for any reason besides the existence 2099 * of managed, wired mappings, then it cannot be freed. For 2100 * example, fictitious pages, which represent device memory, 2101 * are inherently wired and cannot be freed. They can, 2102 * however, be invalidated if the option OBJPR_CLEANONLY is 2103 * not specified. 2104 */ 2105 if (vm_page_tryxbusy(p) == 0) { 2106 vm_page_sleep_if_busy(p, "vmopar"); 2107 goto again; 2108 } 2109 if (vm_page_wired(p)) { 2110 wired: 2111 if ((options & OBJPR_NOTMAPPED) == 0 && 2112 object->ref_count != 0) 2113 pmap_remove_all(p); 2114 if ((options & OBJPR_CLEANONLY) == 0) { 2115 vm_page_invalid(p); 2116 vm_page_undirty(p); 2117 } 2118 vm_page_xunbusy(p); 2119 continue; 2120 } 2121 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2122 ("vm_object_page_remove: page %p is fictitious", p)); 2123 if ((options & OBJPR_CLEANONLY) != 0 && 2124 !vm_page_none_valid(p)) { 2125 if ((options & OBJPR_NOTMAPPED) == 0 && 2126 object->ref_count != 0 && 2127 !vm_page_try_remove_write(p)) 2128 goto wired; 2129 if (p->dirty != 0) { 2130 vm_page_xunbusy(p); 2131 continue; 2132 } 2133 } 2134 if ((options & OBJPR_NOTMAPPED) == 0 && 2135 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2136 goto wired; 2137 vm_page_free(p); 2138 } 2139 vm_object_pip_wakeup(object); 2140 2141 vm_pager_freespace(object, start, (end == 0 ? object->size : end) - 2142 start); 2143 } 2144 2145 /* 2146 * vm_object_page_noreuse: 2147 * 2148 * For the given object, attempt to move the specified pages to 2149 * the head of the inactive queue. This bypasses regular LRU 2150 * operation and allows the pages to be reused quickly under memory 2151 * pressure. If a page is wired for any reason, then it will not 2152 * be queued. Pages are specified by the range ["start", "end"). 2153 * As a special case, if "end" is zero, then the range extends from 2154 * "start" to the end of the object. 2155 * 2156 * This operation should only be performed on objects that 2157 * contain non-fictitious, managed pages. 2158 * 2159 * The object must be locked. 2160 */ 2161 void 2162 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2163 { 2164 vm_page_t p, next; 2165 2166 VM_OBJECT_ASSERT_LOCKED(object); 2167 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2168 ("vm_object_page_noreuse: illegal object %p", object)); 2169 if (object->resident_page_count == 0) 2170 return; 2171 p = vm_page_find_least(object, start); 2172 2173 /* 2174 * Here, the variable "p" is either (1) the page with the least pindex 2175 * greater than or equal to the parameter "start" or (2) NULL. 2176 */ 2177 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2178 next = TAILQ_NEXT(p, listq); 2179 vm_page_deactivate_noreuse(p); 2180 } 2181 } 2182 2183 /* 2184 * Populate the specified range of the object with valid pages. Returns 2185 * TRUE if the range is successfully populated and FALSE otherwise. 2186 * 2187 * Note: This function should be optimized to pass a larger array of 2188 * pages to vm_pager_get_pages() before it is applied to a non- 2189 * OBJT_DEVICE object. 2190 * 2191 * The object must be locked. 2192 */ 2193 boolean_t 2194 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2195 { 2196 vm_page_t m; 2197 vm_pindex_t pindex; 2198 int rv; 2199 2200 VM_OBJECT_ASSERT_WLOCKED(object); 2201 for (pindex = start; pindex < end; pindex++) { 2202 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); 2203 if (rv != VM_PAGER_OK) 2204 break; 2205 2206 /* 2207 * Keep "m" busy because a subsequent iteration may unlock 2208 * the object. 2209 */ 2210 } 2211 if (pindex > start) { 2212 m = vm_page_lookup(object, start); 2213 while (m != NULL && m->pindex < pindex) { 2214 vm_page_xunbusy(m); 2215 m = TAILQ_NEXT(m, listq); 2216 } 2217 } 2218 return (pindex == end); 2219 } 2220 2221 /* 2222 * Routine: vm_object_coalesce 2223 * Function: Coalesces two objects backing up adjoining 2224 * regions of memory into a single object. 2225 * 2226 * returns TRUE if objects were combined. 2227 * 2228 * NOTE: Only works at the moment if the second object is NULL - 2229 * if it's not, which object do we lock first? 2230 * 2231 * Parameters: 2232 * prev_object First object to coalesce 2233 * prev_offset Offset into prev_object 2234 * prev_size Size of reference to prev_object 2235 * next_size Size of reference to the second object 2236 * reserved Indicator that extension region has 2237 * swap accounted for 2238 * 2239 * Conditions: 2240 * The object must *not* be locked. 2241 */ 2242 boolean_t 2243 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2244 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2245 { 2246 vm_pindex_t next_pindex; 2247 2248 if (prev_object == NULL) 2249 return (TRUE); 2250 if ((prev_object->flags & OBJ_ANON) == 0) 2251 return (FALSE); 2252 2253 VM_OBJECT_WLOCK(prev_object); 2254 /* 2255 * Try to collapse the object first. 2256 */ 2257 vm_object_collapse(prev_object); 2258 2259 /* 2260 * Can't coalesce if: . more than one reference . paged out . shadows 2261 * another object . has a copy elsewhere (any of which mean that the 2262 * pages not mapped to prev_entry may be in use anyway) 2263 */ 2264 if (prev_object->backing_object != NULL) { 2265 VM_OBJECT_WUNLOCK(prev_object); 2266 return (FALSE); 2267 } 2268 2269 prev_size >>= PAGE_SHIFT; 2270 next_size >>= PAGE_SHIFT; 2271 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2272 2273 if (prev_object->ref_count > 1 && 2274 prev_object->size != next_pindex && 2275 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2276 VM_OBJECT_WUNLOCK(prev_object); 2277 return (FALSE); 2278 } 2279 2280 /* 2281 * Account for the charge. 2282 */ 2283 if (prev_object->cred != NULL) { 2284 /* 2285 * If prev_object was charged, then this mapping, 2286 * although not charged now, may become writable 2287 * later. Non-NULL cred in the object would prevent 2288 * swap reservation during enabling of the write 2289 * access, so reserve swap now. Failed reservation 2290 * cause allocation of the separate object for the map 2291 * entry, and swap reservation for this entry is 2292 * managed in appropriate time. 2293 */ 2294 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2295 prev_object->cred)) { 2296 VM_OBJECT_WUNLOCK(prev_object); 2297 return (FALSE); 2298 } 2299 prev_object->charge += ptoa(next_size); 2300 } 2301 2302 /* 2303 * Remove any pages that may still be in the object from a previous 2304 * deallocation. 2305 */ 2306 if (next_pindex < prev_object->size) { 2307 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2308 next_size, 0); 2309 #if 0 2310 if (prev_object->cred != NULL) { 2311 KASSERT(prev_object->charge >= 2312 ptoa(prev_object->size - next_pindex), 2313 ("object %p overcharged 1 %jx %jx", prev_object, 2314 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2315 prev_object->charge -= ptoa(prev_object->size - 2316 next_pindex); 2317 } 2318 #endif 2319 } 2320 2321 /* 2322 * Extend the object if necessary. 2323 */ 2324 if (next_pindex + next_size > prev_object->size) 2325 prev_object->size = next_pindex + next_size; 2326 2327 VM_OBJECT_WUNLOCK(prev_object); 2328 return (TRUE); 2329 } 2330 2331 void 2332 vm_object_set_writeable_dirty_(vm_object_t object) 2333 { 2334 atomic_add_int(&object->generation, 1); 2335 } 2336 2337 bool 2338 vm_object_mightbedirty_(vm_object_t object) 2339 { 2340 return (object->generation != object->cleangeneration); 2341 } 2342 2343 /* 2344 * vm_object_unwire: 2345 * 2346 * For each page offset within the specified range of the given object, 2347 * find the highest-level page in the shadow chain and unwire it. A page 2348 * must exist at every page offset, and the highest-level page must be 2349 * wired. 2350 */ 2351 void 2352 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2353 uint8_t queue) 2354 { 2355 vm_object_t tobject, t1object; 2356 vm_page_t m, tm; 2357 vm_pindex_t end_pindex, pindex, tpindex; 2358 int depth, locked_depth; 2359 2360 KASSERT((offset & PAGE_MASK) == 0, 2361 ("vm_object_unwire: offset is not page aligned")); 2362 KASSERT((length & PAGE_MASK) == 0, 2363 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2364 /* The wired count of a fictitious page never changes. */ 2365 if ((object->flags & OBJ_FICTITIOUS) != 0) 2366 return; 2367 pindex = OFF_TO_IDX(offset); 2368 end_pindex = pindex + atop(length); 2369 again: 2370 locked_depth = 1; 2371 VM_OBJECT_RLOCK(object); 2372 m = vm_page_find_least(object, pindex); 2373 while (pindex < end_pindex) { 2374 if (m == NULL || pindex < m->pindex) { 2375 /* 2376 * The first object in the shadow chain doesn't 2377 * contain a page at the current index. Therefore, 2378 * the page must exist in a backing object. 2379 */ 2380 tobject = object; 2381 tpindex = pindex; 2382 depth = 0; 2383 do { 2384 tpindex += 2385 OFF_TO_IDX(tobject->backing_object_offset); 2386 tobject = tobject->backing_object; 2387 KASSERT(tobject != NULL, 2388 ("vm_object_unwire: missing page")); 2389 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2390 goto next_page; 2391 depth++; 2392 if (depth == locked_depth) { 2393 locked_depth++; 2394 VM_OBJECT_RLOCK(tobject); 2395 } 2396 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2397 NULL); 2398 } else { 2399 tm = m; 2400 m = TAILQ_NEXT(m, listq); 2401 } 2402 if (vm_page_trysbusy(tm) == 0) { 2403 for (tobject = object; locked_depth >= 1; 2404 locked_depth--) { 2405 t1object = tobject->backing_object; 2406 if (tm->object != tobject) 2407 VM_OBJECT_RUNLOCK(tobject); 2408 tobject = t1object; 2409 } 2410 vm_page_busy_sleep(tm, "unwbo", true); 2411 goto again; 2412 } 2413 vm_page_unwire(tm, queue); 2414 vm_page_sunbusy(tm); 2415 next_page: 2416 pindex++; 2417 } 2418 /* Release the accumulated object locks. */ 2419 for (tobject = object; locked_depth >= 1; locked_depth--) { 2420 t1object = tobject->backing_object; 2421 VM_OBJECT_RUNLOCK(tobject); 2422 tobject = t1object; 2423 } 2424 } 2425 2426 /* 2427 * Return the vnode for the given object, or NULL if none exists. 2428 * For tmpfs objects, the function may return NULL if there is 2429 * no vnode allocated at the time of the call. 2430 */ 2431 struct vnode * 2432 vm_object_vnode(vm_object_t object) 2433 { 2434 struct vnode *vp; 2435 2436 VM_OBJECT_ASSERT_LOCKED(object); 2437 vm_pager_getvp(object, &vp, NULL); 2438 return (vp); 2439 } 2440 2441 /* 2442 * Busy the vm object. This prevents new pages belonging to the object from 2443 * becoming busy. Existing pages persist as busy. Callers are responsible 2444 * for checking page state before proceeding. 2445 */ 2446 void 2447 vm_object_busy(vm_object_t obj) 2448 { 2449 2450 VM_OBJECT_ASSERT_LOCKED(obj); 2451 2452 blockcount_acquire(&obj->busy, 1); 2453 /* The fence is required to order loads of page busy. */ 2454 atomic_thread_fence_acq_rel(); 2455 } 2456 2457 void 2458 vm_object_unbusy(vm_object_t obj) 2459 { 2460 2461 blockcount_release(&obj->busy, 1); 2462 } 2463 2464 void 2465 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2466 { 2467 2468 VM_OBJECT_ASSERT_UNLOCKED(obj); 2469 2470 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2471 } 2472 2473 static int 2474 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2475 { 2476 struct kinfo_vmobject *kvo; 2477 char *fullpath, *freepath; 2478 struct vnode *vp; 2479 struct vattr va; 2480 vm_object_t obj; 2481 vm_page_t m; 2482 u_long sp; 2483 int count, error; 2484 2485 if (req->oldptr == NULL) { 2486 /* 2487 * If an old buffer has not been provided, generate an 2488 * estimate of the space needed for a subsequent call. 2489 */ 2490 mtx_lock(&vm_object_list_mtx); 2491 count = 0; 2492 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2493 if (obj->type == OBJT_DEAD) 2494 continue; 2495 count++; 2496 } 2497 mtx_unlock(&vm_object_list_mtx); 2498 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2499 count * 11 / 10)); 2500 } 2501 2502 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2503 error = 0; 2504 2505 /* 2506 * VM objects are type stable and are never removed from the 2507 * list once added. This allows us to safely read obj->object_list 2508 * after reacquiring the VM object lock. 2509 */ 2510 mtx_lock(&vm_object_list_mtx); 2511 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2512 if (obj->type == OBJT_DEAD) 2513 continue; 2514 VM_OBJECT_RLOCK(obj); 2515 if (obj->type == OBJT_DEAD) { 2516 VM_OBJECT_RUNLOCK(obj); 2517 continue; 2518 } 2519 mtx_unlock(&vm_object_list_mtx); 2520 kvo->kvo_size = ptoa(obj->size); 2521 kvo->kvo_resident = obj->resident_page_count; 2522 kvo->kvo_ref_count = obj->ref_count; 2523 kvo->kvo_shadow_count = obj->shadow_count; 2524 kvo->kvo_memattr = obj->memattr; 2525 kvo->kvo_active = 0; 2526 kvo->kvo_inactive = 0; 2527 TAILQ_FOREACH(m, &obj->memq, listq) { 2528 /* 2529 * A page may belong to the object but be 2530 * dequeued and set to PQ_NONE while the 2531 * object lock is not held. This makes the 2532 * reads of m->queue below racy, and we do not 2533 * count pages set to PQ_NONE. However, this 2534 * sysctl is only meant to give an 2535 * approximation of the system anyway. 2536 */ 2537 if (m->a.queue == PQ_ACTIVE) 2538 kvo->kvo_active++; 2539 else if (m->a.queue == PQ_INACTIVE) 2540 kvo->kvo_inactive++; 2541 } 2542 2543 kvo->kvo_vn_fileid = 0; 2544 kvo->kvo_vn_fsid = 0; 2545 kvo->kvo_vn_fsid_freebsd11 = 0; 2546 freepath = NULL; 2547 fullpath = ""; 2548 kvo->kvo_type = vm_object_kvme_type(obj, &vp); 2549 if (vp != NULL) { 2550 vref(vp); 2551 } else if ((obj->flags & OBJ_ANON) != 0) { 2552 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT || 2553 kvo->kvo_type == KVME_TYPE_SWAP); 2554 kvo->kvo_me = (uintptr_t)obj; 2555 /* tmpfs objs are reported as vnodes */ 2556 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object; 2557 sp = swap_pager_swapped_pages(obj); 2558 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp; 2559 } 2560 VM_OBJECT_RUNLOCK(obj); 2561 if (vp != NULL) { 2562 vn_fullpath(vp, &fullpath, &freepath); 2563 vn_lock(vp, LK_SHARED | LK_RETRY); 2564 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2565 kvo->kvo_vn_fileid = va.va_fileid; 2566 kvo->kvo_vn_fsid = va.va_fsid; 2567 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2568 /* truncate */ 2569 } 2570 vput(vp); 2571 } 2572 2573 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2574 if (freepath != NULL) 2575 free(freepath, M_TEMP); 2576 2577 /* Pack record size down */ 2578 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2579 + strlen(kvo->kvo_path) + 1; 2580 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2581 sizeof(uint64_t)); 2582 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2583 mtx_lock(&vm_object_list_mtx); 2584 if (error) 2585 break; 2586 } 2587 mtx_unlock(&vm_object_list_mtx); 2588 free(kvo, M_TEMP); 2589 return (error); 2590 } 2591 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2592 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2593 "List of VM objects"); 2594 2595 #include "opt_ddb.h" 2596 #ifdef DDB 2597 #include <sys/kernel.h> 2598 2599 #include <sys/cons.h> 2600 2601 #include <ddb/ddb.h> 2602 2603 static int 2604 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2605 { 2606 vm_map_t tmpm; 2607 vm_map_entry_t tmpe; 2608 vm_object_t obj; 2609 2610 if (map == 0) 2611 return 0; 2612 2613 if (entry == 0) { 2614 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2615 if (_vm_object_in_map(map, object, tmpe)) { 2616 return 1; 2617 } 2618 } 2619 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2620 tmpm = entry->object.sub_map; 2621 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2622 if (_vm_object_in_map(tmpm, object, tmpe)) { 2623 return 1; 2624 } 2625 } 2626 } else if ((obj = entry->object.vm_object) != NULL) { 2627 for (; obj; obj = obj->backing_object) 2628 if (obj == object) { 2629 return 1; 2630 } 2631 } 2632 return 0; 2633 } 2634 2635 static int 2636 vm_object_in_map(vm_object_t object) 2637 { 2638 struct proc *p; 2639 2640 /* sx_slock(&allproc_lock); */ 2641 FOREACH_PROC_IN_SYSTEM(p) { 2642 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2643 continue; 2644 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2645 /* sx_sunlock(&allproc_lock); */ 2646 return 1; 2647 } 2648 } 2649 /* sx_sunlock(&allproc_lock); */ 2650 if (_vm_object_in_map(kernel_map, object, 0)) 2651 return 1; 2652 return 0; 2653 } 2654 2655 DB_SHOW_COMMAND(vmochk, vm_object_check) 2656 { 2657 vm_object_t object; 2658 2659 /* 2660 * make sure that internal objs are in a map somewhere 2661 * and none have zero ref counts. 2662 */ 2663 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2664 if ((object->flags & OBJ_ANON) != 0) { 2665 if (object->ref_count == 0) { 2666 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2667 (long)object->size); 2668 } 2669 if (!vm_object_in_map(object)) { 2670 db_printf( 2671 "vmochk: internal obj is not in a map: " 2672 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2673 object->ref_count, (u_long)object->size, 2674 (u_long)object->size, 2675 (void *)object->backing_object); 2676 } 2677 } 2678 if (db_pager_quit) 2679 return; 2680 } 2681 } 2682 2683 /* 2684 * vm_object_print: [ debug ] 2685 */ 2686 DB_SHOW_COMMAND(object, vm_object_print_static) 2687 { 2688 /* XXX convert args. */ 2689 vm_object_t object = (vm_object_t)addr; 2690 boolean_t full = have_addr; 2691 2692 vm_page_t p; 2693 2694 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2695 #define count was_count 2696 2697 int count; 2698 2699 if (object == NULL) 2700 return; 2701 2702 db_iprintf( 2703 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2704 object, (int)object->type, (uintmax_t)object->size, 2705 object->resident_page_count, object->ref_count, object->flags, 2706 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2707 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2708 object->shadow_count, 2709 object->backing_object ? object->backing_object->ref_count : 0, 2710 object->backing_object, (uintmax_t)object->backing_object_offset); 2711 2712 if (!full) 2713 return; 2714 2715 db_indent += 2; 2716 count = 0; 2717 TAILQ_FOREACH(p, &object->memq, listq) { 2718 if (count == 0) 2719 db_iprintf("memory:="); 2720 else if (count == 6) { 2721 db_printf("\n"); 2722 db_iprintf(" ..."); 2723 count = 0; 2724 } else 2725 db_printf(","); 2726 count++; 2727 2728 db_printf("(off=0x%jx,page=0x%jx)", 2729 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2730 2731 if (db_pager_quit) 2732 break; 2733 } 2734 if (count != 0) 2735 db_printf("\n"); 2736 db_indent -= 2; 2737 } 2738 2739 /* XXX. */ 2740 #undef count 2741 2742 /* XXX need this non-static entry for calling from vm_map_print. */ 2743 void 2744 vm_object_print( 2745 /* db_expr_t */ long addr, 2746 boolean_t have_addr, 2747 /* db_expr_t */ long count, 2748 char *modif) 2749 { 2750 vm_object_print_static(addr, have_addr, count, modif); 2751 } 2752 2753 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2754 { 2755 vm_object_t object; 2756 vm_pindex_t fidx; 2757 vm_paddr_t pa; 2758 vm_page_t m, prev_m; 2759 int rcount, nl, c; 2760 2761 nl = 0; 2762 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2763 db_printf("new object: %p\n", (void *)object); 2764 if (nl > 18) { 2765 c = cngetc(); 2766 if (c != ' ') 2767 return; 2768 nl = 0; 2769 } 2770 nl++; 2771 rcount = 0; 2772 fidx = 0; 2773 pa = -1; 2774 TAILQ_FOREACH(m, &object->memq, listq) { 2775 if (m->pindex > 128) 2776 break; 2777 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2778 prev_m->pindex + 1 != m->pindex) { 2779 if (rcount) { 2780 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2781 (long)fidx, rcount, (long)pa); 2782 if (nl > 18) { 2783 c = cngetc(); 2784 if (c != ' ') 2785 return; 2786 nl = 0; 2787 } 2788 nl++; 2789 rcount = 0; 2790 } 2791 } 2792 if (rcount && 2793 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2794 ++rcount; 2795 continue; 2796 } 2797 if (rcount) { 2798 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2799 (long)fidx, rcount, (long)pa); 2800 if (nl > 18) { 2801 c = cngetc(); 2802 if (c != ' ') 2803 return; 2804 nl = 0; 2805 } 2806 nl++; 2807 } 2808 fidx = m->pindex; 2809 pa = VM_PAGE_TO_PHYS(m); 2810 rcount = 1; 2811 } 2812 if (rcount) { 2813 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2814 (long)fidx, rcount, (long)pa); 2815 if (nl > 18) { 2816 c = cngetc(); 2817 if (c != ' ') 2818 return; 2819 nl = 0; 2820 } 2821 nl++; 2822 } 2823 } 2824 } 2825 #endif /* DDB */ 2826