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/lock.h> 77 #include <sys/mman.h> 78 #include <sys/mount.h> 79 #include <sys/kernel.h> 80 #include <sys/pctrie.h> 81 #include <sys/sysctl.h> 82 #include <sys/mutex.h> 83 #include <sys/proc.h> /* for curproc, pageproc */ 84 #include <sys/refcount.h> 85 #include <sys/socket.h> 86 #include <sys/resourcevar.h> 87 #include <sys/refcount.h> 88 #include <sys/rwlock.h> 89 #include <sys/user.h> 90 #include <sys/vnode.h> 91 #include <sys/vmmeter.h> 92 #include <sys/sx.h> 93 94 #include <vm/vm.h> 95 #include <vm/vm_param.h> 96 #include <vm/pmap.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_pageout.h> 101 #include <vm/vm_pager.h> 102 #include <vm/vm_phys.h> 103 #include <vm/vm_pagequeue.h> 104 #include <vm/swap_pager.h> 105 #include <vm/vm_kern.h> 106 #include <vm/vm_extern.h> 107 #include <vm/vm_radix.h> 108 #include <vm/vm_reserv.h> 109 #include <vm/uma.h> 110 111 static int old_msync; 112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 113 "Use old (insecure) msync behavior"); 114 115 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 116 int pagerflags, int flags, boolean_t *allclean, 117 boolean_t *eio); 118 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 119 boolean_t *allclean); 120 static void vm_object_backing_remove(vm_object_t object); 121 122 /* 123 * Virtual memory objects maintain the actual data 124 * associated with allocated virtual memory. A given 125 * page of memory exists within exactly one object. 126 * 127 * An object is only deallocated when all "references" 128 * are given up. Only one "reference" to a given 129 * region of an object should be writeable. 130 * 131 * Associated with each object is a list of all resident 132 * memory pages belonging to that object; this list is 133 * maintained by the "vm_page" module, and locked by the object's 134 * lock. 135 * 136 * Each object also records a "pager" routine which is 137 * used to retrieve (and store) pages to the proper backing 138 * storage. In addition, objects may be backed by other 139 * objects from which they were virtual-copied. 140 * 141 * The only items within the object structure which are 142 * modified after time of creation are: 143 * reference count locked by object's lock 144 * pager routine locked by object's lock 145 * 146 */ 147 148 struct object_q vm_object_list; 149 struct mtx vm_object_list_mtx; /* lock for object list and count */ 150 151 struct vm_object kernel_object_store; 152 153 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 154 "VM object stats"); 155 156 static COUNTER_U64_DEFINE_EARLY(object_collapses); 157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 158 &object_collapses, 159 "VM object collapses"); 160 161 static COUNTER_U64_DEFINE_EARLY(object_bypasses); 162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 163 &object_bypasses, 164 "VM object bypasses"); 165 166 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits); 167 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD, 168 &object_collapse_waits, 169 "Number of sleeps for collapse"); 170 171 static uma_zone_t obj_zone; 172 173 static int vm_object_zinit(void *mem, int size, int flags); 174 175 #ifdef INVARIANTS 176 static void vm_object_zdtor(void *mem, int size, void *arg); 177 178 static void 179 vm_object_zdtor(void *mem, int size, void *arg) 180 { 181 vm_object_t object; 182 183 object = (vm_object_t)mem; 184 KASSERT(object->ref_count == 0, 185 ("object %p ref_count = %d", object, object->ref_count)); 186 KASSERT(TAILQ_EMPTY(&object->memq), 187 ("object %p has resident pages in its memq", object)); 188 KASSERT(vm_radix_is_empty(&object->rtree), 189 ("object %p has resident pages in its trie", object)); 190 #if VM_NRESERVLEVEL > 0 191 KASSERT(LIST_EMPTY(&object->rvq), 192 ("object %p has reservations", 193 object)); 194 #endif 195 KASSERT(!vm_object_busied(object), 196 ("object %p busy = %d", object, blockcount_read(&object->busy))); 197 KASSERT(object->resident_page_count == 0, 198 ("object %p resident_page_count = %d", 199 object, object->resident_page_count)); 200 KASSERT(object->shadow_count == 0, 201 ("object %p shadow_count = %d", 202 object, object->shadow_count)); 203 KASSERT(object->type == OBJT_DEAD, 204 ("object %p has non-dead type %d", 205 object, object->type)); 206 } 207 #endif 208 209 static int 210 vm_object_zinit(void *mem, int size, int flags) 211 { 212 vm_object_t object; 213 214 object = (vm_object_t)mem; 215 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); 216 217 /* These are true for any object that has been freed */ 218 object->type = OBJT_DEAD; 219 vm_radix_init(&object->rtree); 220 refcount_init(&object->ref_count, 0); 221 blockcount_init(&object->paging_in_progress); 222 blockcount_init(&object->busy); 223 object->resident_page_count = 0; 224 object->shadow_count = 0; 225 object->flags = OBJ_DEAD; 226 227 mtx_lock(&vm_object_list_mtx); 228 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 229 mtx_unlock(&vm_object_list_mtx); 230 return (0); 231 } 232 233 static void 234 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, 235 vm_object_t object, void *handle) 236 { 237 238 TAILQ_INIT(&object->memq); 239 LIST_INIT(&object->shadow_head); 240 241 object->type = type; 242 if (type == OBJT_SWAP) 243 pctrie_init(&object->un_pager.swp.swp_blks); 244 245 /* 246 * Ensure that swap_pager_swapoff() iteration over object_list 247 * sees up to date type and pctrie head if it observed 248 * non-dead object. 249 */ 250 atomic_thread_fence_rel(); 251 252 object->pg_color = 0; 253 object->flags = flags; 254 object->size = size; 255 object->domain.dr_policy = NULL; 256 object->generation = 1; 257 object->cleangeneration = 1; 258 refcount_init(&object->ref_count, 1); 259 object->memattr = VM_MEMATTR_DEFAULT; 260 object->cred = NULL; 261 object->charge = 0; 262 object->handle = handle; 263 object->backing_object = NULL; 264 object->backing_object_offset = (vm_ooffset_t) 0; 265 #if VM_NRESERVLEVEL > 0 266 LIST_INIT(&object->rvq); 267 #endif 268 umtx_shm_object_init(object); 269 } 270 271 /* 272 * vm_object_init: 273 * 274 * Initialize the VM objects module. 275 */ 276 void 277 vm_object_init(void) 278 { 279 TAILQ_INIT(&vm_object_list); 280 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 281 282 rw_init(&kernel_object->lock, "kernel vm object"); 283 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 284 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); 285 #if VM_NRESERVLEVEL > 0 286 kernel_object->flags |= OBJ_COLORED; 287 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 288 #endif 289 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 290 291 /* 292 * The lock portion of struct vm_object must be type stable due 293 * to vm_pageout_fallback_object_lock locking a vm object 294 * without holding any references to it. 295 * 296 * paging_in_progress is valid always. Lockless references to 297 * the objects may acquire pip and then check OBJ_DEAD. 298 */ 299 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 300 #ifdef INVARIANTS 301 vm_object_zdtor, 302 #else 303 NULL, 304 #endif 305 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 306 307 vm_radix_zinit(); 308 } 309 310 void 311 vm_object_clear_flag(vm_object_t object, u_short bits) 312 { 313 314 VM_OBJECT_ASSERT_WLOCKED(object); 315 object->flags &= ~bits; 316 } 317 318 /* 319 * Sets the default memory attribute for the specified object. Pages 320 * that are allocated to this object are by default assigned this memory 321 * attribute. 322 * 323 * Presently, this function must be called before any pages are allocated 324 * to the object. In the future, this requirement may be relaxed for 325 * "default" and "swap" objects. 326 */ 327 int 328 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 329 { 330 331 VM_OBJECT_ASSERT_WLOCKED(object); 332 switch (object->type) { 333 case OBJT_DEFAULT: 334 case OBJT_DEVICE: 335 case OBJT_MGTDEVICE: 336 case OBJT_PHYS: 337 case OBJT_SG: 338 case OBJT_SWAP: 339 case OBJT_VNODE: 340 if (!TAILQ_EMPTY(&object->memq)) 341 return (KERN_FAILURE); 342 break; 343 case OBJT_DEAD: 344 return (KERN_INVALID_ARGUMENT); 345 default: 346 panic("vm_object_set_memattr: object %p is of undefined type", 347 object); 348 } 349 object->memattr = memattr; 350 return (KERN_SUCCESS); 351 } 352 353 void 354 vm_object_pip_add(vm_object_t object, short i) 355 { 356 357 if (i > 0) 358 blockcount_acquire(&object->paging_in_progress, i); 359 } 360 361 void 362 vm_object_pip_wakeup(vm_object_t object) 363 { 364 365 vm_object_pip_wakeupn(object, 1); 366 } 367 368 void 369 vm_object_pip_wakeupn(vm_object_t object, short i) 370 { 371 372 if (i > 0) 373 blockcount_release(&object->paging_in_progress, i); 374 } 375 376 /* 377 * Atomically drop the object lock and wait for pip to drain. This protects 378 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 379 * on return. 380 */ 381 static void 382 vm_object_pip_sleep(vm_object_t object, const char *waitid) 383 { 384 385 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 386 waitid, PVM | PDROP); 387 } 388 389 void 390 vm_object_pip_wait(vm_object_t object, const char *waitid) 391 { 392 393 VM_OBJECT_ASSERT_WLOCKED(object); 394 395 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 396 PVM); 397 } 398 399 void 400 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 401 { 402 403 VM_OBJECT_ASSERT_UNLOCKED(object); 404 405 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 406 } 407 408 /* 409 * vm_object_allocate: 410 * 411 * Returns a new object with the given size. 412 */ 413 vm_object_t 414 vm_object_allocate(objtype_t type, vm_pindex_t size) 415 { 416 vm_object_t object; 417 u_short flags; 418 419 switch (type) { 420 case OBJT_DEAD: 421 panic("vm_object_allocate: can't create OBJT_DEAD"); 422 case OBJT_DEFAULT: 423 case OBJT_SWAP: 424 flags = OBJ_COLORED; 425 break; 426 case OBJT_DEVICE: 427 case OBJT_SG: 428 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 429 break; 430 case OBJT_MGTDEVICE: 431 flags = OBJ_FICTITIOUS; 432 break; 433 case OBJT_PHYS: 434 flags = OBJ_UNMANAGED; 435 break; 436 case OBJT_VNODE: 437 flags = 0; 438 break; 439 default: 440 panic("vm_object_allocate: type %d is undefined", type); 441 } 442 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 443 _vm_object_allocate(type, size, flags, object, NULL); 444 445 return (object); 446 } 447 448 /* 449 * vm_object_allocate_anon: 450 * 451 * Returns a new default object of the given size and marked as 452 * anonymous memory for special split/collapse handling. Color 453 * to be initialized by the caller. 454 */ 455 vm_object_t 456 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 457 struct ucred *cred, vm_size_t charge) 458 { 459 vm_object_t handle, object; 460 461 if (backing_object == NULL) 462 handle = NULL; 463 else if ((backing_object->flags & OBJ_ANON) != 0) 464 handle = backing_object->handle; 465 else 466 handle = backing_object; 467 object = uma_zalloc(obj_zone, M_WAITOK); 468 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING, 469 object, handle); 470 object->cred = cred; 471 object->charge = cred != NULL ? charge : 0; 472 return (object); 473 } 474 475 static void 476 vm_object_reference_vnode(vm_object_t object) 477 { 478 u_int old; 479 480 /* 481 * vnode objects need the lock for the first reference 482 * to serialize with vnode_object_deallocate(). 483 */ 484 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 485 VM_OBJECT_RLOCK(object); 486 old = refcount_acquire(&object->ref_count); 487 if (object->type == OBJT_VNODE && old == 0) 488 vref(object->handle); 489 VM_OBJECT_RUNLOCK(object); 490 } 491 } 492 493 /* 494 * vm_object_reference: 495 * 496 * Acquires a reference to the given object. 497 */ 498 void 499 vm_object_reference(vm_object_t object) 500 { 501 502 if (object == NULL) 503 return; 504 505 if (object->type == OBJT_VNODE) 506 vm_object_reference_vnode(object); 507 else 508 refcount_acquire(&object->ref_count); 509 KASSERT((object->flags & OBJ_DEAD) == 0, 510 ("vm_object_reference: Referenced dead object.")); 511 } 512 513 /* 514 * vm_object_reference_locked: 515 * 516 * Gets another reference to the given object. 517 * 518 * The object must be locked. 519 */ 520 void 521 vm_object_reference_locked(vm_object_t object) 522 { 523 u_int old; 524 525 VM_OBJECT_ASSERT_LOCKED(object); 526 old = refcount_acquire(&object->ref_count); 527 if (object->type == OBJT_VNODE && old == 0) 528 vref(object->handle); 529 KASSERT((object->flags & OBJ_DEAD) == 0, 530 ("vm_object_reference: Referenced dead object.")); 531 } 532 533 /* 534 * Handle deallocating an object of type OBJT_VNODE. 535 */ 536 static void 537 vm_object_deallocate_vnode(vm_object_t object) 538 { 539 struct vnode *vp = (struct vnode *) object->handle; 540 bool last; 541 542 KASSERT(object->type == OBJT_VNODE, 543 ("vm_object_deallocate_vnode: not a vnode object")); 544 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 545 546 /* Object lock to protect handle lookup. */ 547 last = refcount_release(&object->ref_count); 548 VM_OBJECT_RUNLOCK(object); 549 550 if (!last) 551 return; 552 553 if (!umtx_shm_vnobj_persistent) 554 umtx_shm_object_terminated(object); 555 556 /* vrele may need the vnode lock. */ 557 vrele(vp); 558 } 559 560 /* 561 * We dropped a reference on an object and discovered that it had a 562 * single remaining shadow. This is a sibling of the reference we 563 * dropped. Attempt to collapse the sibling and backing object. 564 */ 565 static vm_object_t 566 vm_object_deallocate_anon(vm_object_t backing_object) 567 { 568 vm_object_t object; 569 570 /* Fetch the final shadow. */ 571 object = LIST_FIRST(&backing_object->shadow_head); 572 KASSERT(object != NULL && backing_object->shadow_count == 1, 573 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 574 backing_object->ref_count, backing_object->shadow_count)); 575 KASSERT((object->flags & (OBJ_TMPFS_NODE | OBJ_ANON)) == OBJ_ANON, 576 ("invalid shadow object %p", object)); 577 578 if (!VM_OBJECT_TRYWLOCK(object)) { 579 /* 580 * Prevent object from disappearing since we do not have a 581 * reference. 582 */ 583 vm_object_pip_add(object, 1); 584 VM_OBJECT_WUNLOCK(backing_object); 585 VM_OBJECT_WLOCK(object); 586 vm_object_pip_wakeup(object); 587 } else 588 VM_OBJECT_WUNLOCK(backing_object); 589 590 /* 591 * Check for a collapse/terminate race with the last reference holder. 592 */ 593 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 594 !refcount_acquire_if_not_zero(&object->ref_count)) { 595 VM_OBJECT_WUNLOCK(object); 596 return (NULL); 597 } 598 backing_object = object->backing_object; 599 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 600 vm_object_collapse(object); 601 VM_OBJECT_WUNLOCK(object); 602 603 return (object); 604 } 605 606 /* 607 * vm_object_deallocate: 608 * 609 * Release a reference to the specified object, 610 * gained either through a vm_object_allocate 611 * or a vm_object_reference call. When all references 612 * are gone, storage associated with this object 613 * may be relinquished. 614 * 615 * No object may be locked. 616 */ 617 void 618 vm_object_deallocate(vm_object_t object) 619 { 620 vm_object_t temp; 621 bool released; 622 623 while (object != NULL) { 624 /* 625 * If the reference count goes to 0 we start calling 626 * vm_object_terminate() on the object chain. A ref count 627 * of 1 may be a special case depending on the shadow count 628 * being 0 or 1. These cases require a write lock on the 629 * object. 630 */ 631 if ((object->flags & OBJ_ANON) == 0) 632 released = refcount_release_if_gt(&object->ref_count, 1); 633 else 634 released = refcount_release_if_gt(&object->ref_count, 2); 635 if (released) 636 return; 637 638 if (object->type == OBJT_VNODE) { 639 VM_OBJECT_RLOCK(object); 640 if (object->type == OBJT_VNODE) { 641 vm_object_deallocate_vnode(object); 642 return; 643 } 644 VM_OBJECT_RUNLOCK(object); 645 } 646 647 VM_OBJECT_WLOCK(object); 648 KASSERT(object->ref_count > 0, 649 ("vm_object_deallocate: object deallocated too many times: %d", 650 object->type)); 651 652 /* 653 * If this is not the final reference to an anonymous 654 * object we may need to collapse the shadow chain. 655 */ 656 if (!refcount_release(&object->ref_count)) { 657 if (object->ref_count > 1 || 658 object->shadow_count == 0) { 659 if ((object->flags & OBJ_ANON) != 0 && 660 object->ref_count == 1) 661 vm_object_set_flag(object, 662 OBJ_ONEMAPPING); 663 VM_OBJECT_WUNLOCK(object); 664 return; 665 } 666 667 /* Handle collapsing last ref on anonymous objects. */ 668 object = vm_object_deallocate_anon(object); 669 continue; 670 } 671 672 /* 673 * Handle the final reference to an object. We restart 674 * the loop with the backing object to avoid recursion. 675 */ 676 umtx_shm_object_terminated(object); 677 temp = object->backing_object; 678 if (temp != NULL) { 679 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, 680 ("shadowed tmpfs v_object 2 %p", object)); 681 vm_object_backing_remove(object); 682 } 683 684 KASSERT((object->flags & OBJ_DEAD) == 0, 685 ("vm_object_deallocate: Terminating dead object.")); 686 vm_object_set_flag(object, OBJ_DEAD); 687 vm_object_terminate(object); 688 object = temp; 689 } 690 } 691 692 /* 693 * vm_object_destroy removes the object from the global object list 694 * and frees the space for the object. 695 */ 696 void 697 vm_object_destroy(vm_object_t object) 698 { 699 700 /* 701 * Release the allocation charge. 702 */ 703 if (object->cred != NULL) { 704 swap_release_by_cred(object->charge, object->cred); 705 object->charge = 0; 706 crfree(object->cred); 707 object->cred = NULL; 708 } 709 710 /* 711 * Free the space for the object. 712 */ 713 uma_zfree(obj_zone, object); 714 } 715 716 static void 717 vm_object_backing_remove_locked(vm_object_t object) 718 { 719 vm_object_t backing_object; 720 721 backing_object = object->backing_object; 722 VM_OBJECT_ASSERT_WLOCKED(object); 723 VM_OBJECT_ASSERT_WLOCKED(backing_object); 724 725 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 726 ("vm_object_backing_remove: Removing collapsing object.")); 727 728 if ((object->flags & OBJ_SHADOWLIST) != 0) { 729 LIST_REMOVE(object, shadow_list); 730 backing_object->shadow_count--; 731 object->flags &= ~OBJ_SHADOWLIST; 732 } 733 object->backing_object = NULL; 734 } 735 736 static void 737 vm_object_backing_remove(vm_object_t object) 738 { 739 vm_object_t backing_object; 740 741 VM_OBJECT_ASSERT_WLOCKED(object); 742 743 if ((object->flags & OBJ_SHADOWLIST) != 0) { 744 backing_object = object->backing_object; 745 VM_OBJECT_WLOCK(backing_object); 746 vm_object_backing_remove_locked(object); 747 VM_OBJECT_WUNLOCK(backing_object); 748 } else 749 object->backing_object = NULL; 750 } 751 752 static void 753 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 754 { 755 756 VM_OBJECT_ASSERT_WLOCKED(object); 757 758 if ((backing_object->flags & OBJ_ANON) != 0) { 759 VM_OBJECT_ASSERT_WLOCKED(backing_object); 760 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 761 shadow_list); 762 backing_object->shadow_count++; 763 object->flags |= OBJ_SHADOWLIST; 764 } 765 object->backing_object = backing_object; 766 } 767 768 static void 769 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 770 { 771 772 VM_OBJECT_ASSERT_WLOCKED(object); 773 774 if ((backing_object->flags & OBJ_ANON) != 0) { 775 VM_OBJECT_WLOCK(backing_object); 776 vm_object_backing_insert_locked(object, backing_object); 777 VM_OBJECT_WUNLOCK(backing_object); 778 } else 779 object->backing_object = backing_object; 780 } 781 782 /* 783 * Insert an object into a backing_object's shadow list with an additional 784 * reference to the backing_object added. 785 */ 786 static void 787 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object) 788 { 789 790 VM_OBJECT_ASSERT_WLOCKED(object); 791 792 if ((backing_object->flags & OBJ_ANON) != 0) { 793 VM_OBJECT_WLOCK(backing_object); 794 KASSERT((backing_object->flags & OBJ_DEAD) == 0, 795 ("shadowing dead anonymous object")); 796 vm_object_reference_locked(backing_object); 797 vm_object_backing_insert_locked(object, backing_object); 798 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING); 799 VM_OBJECT_WUNLOCK(backing_object); 800 } else { 801 vm_object_reference(backing_object); 802 object->backing_object = backing_object; 803 } 804 } 805 806 /* 807 * Transfer a backing reference from backing_object to object. 808 */ 809 static void 810 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 811 { 812 vm_object_t new_backing_object; 813 814 /* 815 * Note that the reference to backing_object->backing_object 816 * moves from within backing_object to within object. 817 */ 818 vm_object_backing_remove_locked(object); 819 new_backing_object = backing_object->backing_object; 820 if (new_backing_object == NULL) 821 return; 822 if ((new_backing_object->flags & OBJ_ANON) != 0) { 823 VM_OBJECT_WLOCK(new_backing_object); 824 vm_object_backing_remove_locked(backing_object); 825 vm_object_backing_insert_locked(object, new_backing_object); 826 VM_OBJECT_WUNLOCK(new_backing_object); 827 } else { 828 object->backing_object = new_backing_object; 829 backing_object->backing_object = NULL; 830 } 831 } 832 833 /* 834 * Wait for a concurrent collapse to settle. 835 */ 836 static void 837 vm_object_collapse_wait(vm_object_t object) 838 { 839 840 VM_OBJECT_ASSERT_WLOCKED(object); 841 842 while ((object->flags & OBJ_COLLAPSING) != 0) { 843 vm_object_pip_wait(object, "vmcolwait"); 844 counter_u64_add(object_collapse_waits, 1); 845 } 846 } 847 848 /* 849 * Waits for a backing object to clear a pending collapse and returns 850 * it locked if it is an ANON object. 851 */ 852 static vm_object_t 853 vm_object_backing_collapse_wait(vm_object_t object) 854 { 855 vm_object_t backing_object; 856 857 VM_OBJECT_ASSERT_WLOCKED(object); 858 859 for (;;) { 860 backing_object = object->backing_object; 861 if (backing_object == NULL || 862 (backing_object->flags & OBJ_ANON) == 0) 863 return (NULL); 864 VM_OBJECT_WLOCK(backing_object); 865 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 866 break; 867 VM_OBJECT_WUNLOCK(object); 868 vm_object_pip_sleep(backing_object, "vmbckwait"); 869 counter_u64_add(object_collapse_waits, 1); 870 VM_OBJECT_WLOCK(object); 871 } 872 return (backing_object); 873 } 874 875 /* 876 * vm_object_terminate_pages removes any remaining pageable pages 877 * from the object and resets the object to an empty state. 878 */ 879 static void 880 vm_object_terminate_pages(vm_object_t object) 881 { 882 vm_page_t p, p_next; 883 884 VM_OBJECT_ASSERT_WLOCKED(object); 885 886 /* 887 * Free any remaining pageable pages. This also removes them from the 888 * paging queues. However, don't free wired pages, just remove them 889 * from the object. Rather than incrementally removing each page from 890 * the object, the page and object are reset to any empty state. 891 */ 892 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 893 vm_page_assert_unbusied(p); 894 KASSERT(p->object == object && 895 (p->ref_count & VPRC_OBJREF) != 0, 896 ("vm_object_terminate_pages: page %p is inconsistent", p)); 897 898 p->object = NULL; 899 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 900 VM_CNT_INC(v_pfree); 901 vm_page_free(p); 902 } 903 } 904 905 /* 906 * If the object contained any pages, then reset it to an empty state. 907 * None of the object's fields, including "resident_page_count", were 908 * modified by the preceding loop. 909 */ 910 if (object->resident_page_count != 0) { 911 vm_radix_reclaim_allnodes(&object->rtree); 912 TAILQ_INIT(&object->memq); 913 object->resident_page_count = 0; 914 if (object->type == OBJT_VNODE) 915 vdrop(object->handle); 916 } 917 } 918 919 /* 920 * vm_object_terminate actually destroys the specified object, freeing 921 * up all previously used resources. 922 * 923 * The object must be locked. 924 * This routine may block. 925 */ 926 void 927 vm_object_terminate(vm_object_t object) 928 { 929 930 VM_OBJECT_ASSERT_WLOCKED(object); 931 KASSERT((object->flags & OBJ_DEAD) != 0, 932 ("terminating non-dead obj %p", object)); 933 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 934 ("terminating collapsing obj %p", object)); 935 KASSERT(object->backing_object == NULL, 936 ("terminating shadow obj %p", object)); 937 938 /* 939 * Wait for the pageout daemon and other current users to be 940 * done with the object. Note that new paging_in_progress 941 * users can come after this wait, but they must check 942 * OBJ_DEAD flag set (without unlocking the object), and avoid 943 * the object being terminated. 944 */ 945 vm_object_pip_wait(object, "objtrm"); 946 947 KASSERT(object->ref_count == 0, 948 ("vm_object_terminate: object with references, ref_count=%d", 949 object->ref_count)); 950 951 if ((object->flags & OBJ_PG_DTOR) == 0) 952 vm_object_terminate_pages(object); 953 954 #if VM_NRESERVLEVEL > 0 955 if (__predict_false(!LIST_EMPTY(&object->rvq))) 956 vm_reserv_break_all(object); 957 #endif 958 959 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || 960 object->type == OBJT_SWAP, 961 ("%s: non-swap obj %p has cred", __func__, object)); 962 963 /* 964 * Let the pager know object is dead. 965 */ 966 vm_pager_deallocate(object); 967 VM_OBJECT_WUNLOCK(object); 968 969 vm_object_destroy(object); 970 } 971 972 /* 973 * Make the page read-only so that we can clear the object flags. However, if 974 * this is a nosync mmap then the object is likely to stay dirty so do not 975 * mess with the page and do not clear the object flags. Returns TRUE if the 976 * page should be flushed, and FALSE otherwise. 977 */ 978 static boolean_t 979 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) 980 { 981 982 vm_page_assert_busied(p); 983 984 /* 985 * If we have been asked to skip nosync pages and this is a 986 * nosync page, skip it. Note that the object flags were not 987 * cleared in this case so we do not have to set them. 988 */ 989 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { 990 *allclean = FALSE; 991 return (FALSE); 992 } else { 993 pmap_remove_write(p); 994 return (p->dirty != 0); 995 } 996 } 997 998 /* 999 * vm_object_page_clean 1000 * 1001 * Clean all dirty pages in the specified range of object. Leaves page 1002 * on whatever queue it is currently on. If NOSYNC is set then do not 1003 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1004 * leaving the object dirty. 1005 * 1006 * For swap objects backing tmpfs regular files, do not flush anything, 1007 * but remove write protection on the mapped pages to update mtime through 1008 * mmaped writes. 1009 * 1010 * When stuffing pages asynchronously, allow clustering. XXX we need a 1011 * synchronous clustering mode implementation. 1012 * 1013 * Odd semantics: if start == end, we clean everything. 1014 * 1015 * The object must be locked. 1016 * 1017 * Returns FALSE if some page from the range was not written, as 1018 * reported by the pager, and TRUE otherwise. 1019 */ 1020 boolean_t 1021 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1022 int flags) 1023 { 1024 vm_page_t np, p; 1025 vm_pindex_t pi, tend, tstart; 1026 int curgeneration, n, pagerflags; 1027 boolean_t eio, res, allclean; 1028 1029 VM_OBJECT_ASSERT_WLOCKED(object); 1030 1031 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1032 return (TRUE); 1033 1034 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1035 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1036 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1037 1038 tstart = OFF_TO_IDX(start); 1039 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1040 allclean = tstart == 0 && tend >= object->size; 1041 res = TRUE; 1042 1043 rescan: 1044 curgeneration = object->generation; 1045 1046 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 1047 pi = p->pindex; 1048 if (pi >= tend) 1049 break; 1050 np = TAILQ_NEXT(p, listq); 1051 if (vm_page_none_valid(p)) 1052 continue; 1053 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { 1054 if (object->generation != curgeneration && 1055 (flags & OBJPC_SYNC) != 0) 1056 goto rescan; 1057 np = vm_page_find_least(object, pi); 1058 continue; 1059 } 1060 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1061 vm_page_xunbusy(p); 1062 continue; 1063 } 1064 if (object->type == OBJT_VNODE) { 1065 n = vm_object_page_collect_flush(object, p, pagerflags, 1066 flags, &allclean, &eio); 1067 if (eio) { 1068 res = FALSE; 1069 allclean = FALSE; 1070 } 1071 if (object->generation != curgeneration && 1072 (flags & OBJPC_SYNC) != 0) 1073 goto rescan; 1074 1075 /* 1076 * If the VOP_PUTPAGES() did a truncated write, so 1077 * that even the first page of the run is not fully 1078 * written, vm_pageout_flush() returns 0 as the run 1079 * length. Since the condition that caused truncated 1080 * write may be permanent, e.g. exhausted free space, 1081 * accepting n == 0 would cause an infinite loop. 1082 * 1083 * Forwarding the iterator leaves the unwritten page 1084 * behind, but there is not much we can do there if 1085 * filesystem refuses to write it. 1086 */ 1087 if (n == 0) { 1088 n = 1; 1089 allclean = FALSE; 1090 } 1091 } else { 1092 n = 1; 1093 vm_page_xunbusy(p); 1094 } 1095 np = vm_page_find_least(object, pi + n); 1096 } 1097 #if 0 1098 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1099 #endif 1100 1101 /* 1102 * Leave updating cleangeneration for tmpfs objects to tmpfs 1103 * scan. It needs to update mtime, which happens for other 1104 * filesystems during page writeouts. 1105 */ 1106 if (allclean && object->type == OBJT_VNODE) 1107 object->cleangeneration = curgeneration; 1108 return (res); 1109 } 1110 1111 static int 1112 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1113 int flags, boolean_t *allclean, boolean_t *eio) 1114 { 1115 vm_page_t ma[vm_pageout_page_count], p_first, tp; 1116 int count, i, mreq, runlen; 1117 1118 vm_page_lock_assert(p, MA_NOTOWNED); 1119 vm_page_assert_xbusied(p); 1120 VM_OBJECT_ASSERT_WLOCKED(object); 1121 1122 count = 1; 1123 mreq = 0; 1124 1125 for (tp = p; count < vm_pageout_page_count; count++) { 1126 tp = vm_page_next(tp); 1127 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1128 break; 1129 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1130 vm_page_xunbusy(tp); 1131 break; 1132 } 1133 } 1134 1135 for (p_first = p; count < vm_pageout_page_count; count++) { 1136 tp = vm_page_prev(p_first); 1137 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1138 break; 1139 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1140 vm_page_xunbusy(tp); 1141 break; 1142 } 1143 p_first = tp; 1144 mreq++; 1145 } 1146 1147 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) 1148 ma[i] = tp; 1149 1150 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); 1151 return (runlen); 1152 } 1153 1154 /* 1155 * Note that there is absolutely no sense in writing out 1156 * anonymous objects, so we track down the vnode object 1157 * to write out. 1158 * We invalidate (remove) all pages from the address space 1159 * for semantic correctness. 1160 * 1161 * If the backing object is a device object with unmanaged pages, then any 1162 * mappings to the specified range of pages must be removed before this 1163 * function is called. 1164 * 1165 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1166 * may start out with a NULL object. 1167 */ 1168 boolean_t 1169 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1170 boolean_t syncio, boolean_t invalidate) 1171 { 1172 vm_object_t backing_object; 1173 struct vnode *vp; 1174 struct mount *mp; 1175 int error, flags, fsync_after; 1176 boolean_t res; 1177 1178 if (object == NULL) 1179 return (TRUE); 1180 res = TRUE; 1181 error = 0; 1182 VM_OBJECT_WLOCK(object); 1183 while ((backing_object = object->backing_object) != NULL) { 1184 VM_OBJECT_WLOCK(backing_object); 1185 offset += object->backing_object_offset; 1186 VM_OBJECT_WUNLOCK(object); 1187 object = backing_object; 1188 if (object->size < OFF_TO_IDX(offset + size)) 1189 size = IDX_TO_OFF(object->size) - offset; 1190 } 1191 /* 1192 * Flush pages if writing is allowed, invalidate them 1193 * if invalidation requested. Pages undergoing I/O 1194 * will be ignored by vm_object_page_remove(). 1195 * 1196 * We cannot lock the vnode and then wait for paging 1197 * to complete without deadlocking against vm_fault. 1198 * Instead we simply call vm_object_page_remove() and 1199 * allow it to block internally on a page-by-page 1200 * basis when it encounters pages undergoing async 1201 * I/O. 1202 */ 1203 if (object->type == OBJT_VNODE && 1204 vm_object_mightbedirty(object) != 0 && 1205 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1206 VM_OBJECT_WUNLOCK(object); 1207 (void) vn_start_write(vp, &mp, V_WAIT); 1208 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1209 if (syncio && !invalidate && offset == 0 && 1210 atop(size) == object->size) { 1211 /* 1212 * If syncing the whole mapping of the file, 1213 * it is faster to schedule all the writes in 1214 * async mode, also allowing the clustering, 1215 * and then wait for i/o to complete. 1216 */ 1217 flags = 0; 1218 fsync_after = TRUE; 1219 } else { 1220 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1221 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1222 fsync_after = FALSE; 1223 } 1224 VM_OBJECT_WLOCK(object); 1225 res = vm_object_page_clean(object, offset, offset + size, 1226 flags); 1227 VM_OBJECT_WUNLOCK(object); 1228 if (fsync_after) 1229 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1230 VOP_UNLOCK(vp); 1231 vn_finished_write(mp); 1232 if (error != 0) 1233 res = FALSE; 1234 VM_OBJECT_WLOCK(object); 1235 } 1236 if ((object->type == OBJT_VNODE || 1237 object->type == OBJT_DEVICE) && invalidate) { 1238 if (object->type == OBJT_DEVICE) 1239 /* 1240 * The option OBJPR_NOTMAPPED must be passed here 1241 * because vm_object_page_remove() cannot remove 1242 * unmanaged mappings. 1243 */ 1244 flags = OBJPR_NOTMAPPED; 1245 else if (old_msync) 1246 flags = 0; 1247 else 1248 flags = OBJPR_CLEANONLY; 1249 vm_object_page_remove(object, OFF_TO_IDX(offset), 1250 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1251 } 1252 VM_OBJECT_WUNLOCK(object); 1253 return (res); 1254 } 1255 1256 /* 1257 * Determine whether the given advice can be applied to the object. Advice is 1258 * not applied to unmanaged pages since they never belong to page queues, and 1259 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1260 * have been mapped at most once. 1261 */ 1262 static bool 1263 vm_object_advice_applies(vm_object_t object, int advice) 1264 { 1265 1266 if ((object->flags & OBJ_UNMANAGED) != 0) 1267 return (false); 1268 if (advice != MADV_FREE) 1269 return (true); 1270 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1271 (OBJ_ONEMAPPING | OBJ_ANON)); 1272 } 1273 1274 static void 1275 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1276 vm_size_t size) 1277 { 1278 1279 if (advice == MADV_FREE && object->type == OBJT_SWAP) 1280 swap_pager_freespace(object, pindex, size); 1281 } 1282 1283 /* 1284 * vm_object_madvise: 1285 * 1286 * Implements the madvise function at the object/page level. 1287 * 1288 * MADV_WILLNEED (any object) 1289 * 1290 * Activate the specified pages if they are resident. 1291 * 1292 * MADV_DONTNEED (any object) 1293 * 1294 * Deactivate the specified pages if they are resident. 1295 * 1296 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1297 * OBJ_ONEMAPPING only) 1298 * 1299 * Deactivate and clean the specified pages if they are 1300 * resident. This permits the process to reuse the pages 1301 * without faulting or the kernel to reclaim the pages 1302 * without I/O. 1303 */ 1304 void 1305 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1306 int advice) 1307 { 1308 vm_pindex_t tpindex; 1309 vm_object_t backing_object, tobject; 1310 vm_page_t m, tm; 1311 1312 if (object == NULL) 1313 return; 1314 1315 relookup: 1316 VM_OBJECT_WLOCK(object); 1317 if (!vm_object_advice_applies(object, advice)) { 1318 VM_OBJECT_WUNLOCK(object); 1319 return; 1320 } 1321 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1322 tobject = object; 1323 1324 /* 1325 * If the next page isn't resident in the top-level object, we 1326 * need to search the shadow chain. When applying MADV_FREE, we 1327 * take care to release any swap space used to store 1328 * non-resident pages. 1329 */ 1330 if (m == NULL || pindex < m->pindex) { 1331 /* 1332 * Optimize a common case: if the top-level object has 1333 * no backing object, we can skip over the non-resident 1334 * range in constant time. 1335 */ 1336 if (object->backing_object == NULL) { 1337 tpindex = (m != NULL && m->pindex < end) ? 1338 m->pindex : end; 1339 vm_object_madvise_freespace(object, advice, 1340 pindex, tpindex - pindex); 1341 if ((pindex = tpindex) == end) 1342 break; 1343 goto next_page; 1344 } 1345 1346 tpindex = pindex; 1347 do { 1348 vm_object_madvise_freespace(tobject, advice, 1349 tpindex, 1); 1350 /* 1351 * Prepare to search the next object in the 1352 * chain. 1353 */ 1354 backing_object = tobject->backing_object; 1355 if (backing_object == NULL) 1356 goto next_pindex; 1357 VM_OBJECT_WLOCK(backing_object); 1358 tpindex += 1359 OFF_TO_IDX(tobject->backing_object_offset); 1360 if (tobject != object) 1361 VM_OBJECT_WUNLOCK(tobject); 1362 tobject = backing_object; 1363 if (!vm_object_advice_applies(tobject, advice)) 1364 goto next_pindex; 1365 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1366 NULL); 1367 } else { 1368 next_page: 1369 tm = m; 1370 m = TAILQ_NEXT(m, listq); 1371 } 1372 1373 /* 1374 * If the page is not in a normal state, skip it. The page 1375 * can not be invalidated while the object lock is held. 1376 */ 1377 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1378 goto next_pindex; 1379 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1380 ("vm_object_madvise: page %p is fictitious", tm)); 1381 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1382 ("vm_object_madvise: page %p is not managed", tm)); 1383 if (vm_page_tryxbusy(tm) == 0) { 1384 if (object != tobject) 1385 VM_OBJECT_WUNLOCK(object); 1386 if (advice == MADV_WILLNEED) { 1387 /* 1388 * Reference the page before unlocking and 1389 * sleeping so that the page daemon is less 1390 * likely to reclaim it. 1391 */ 1392 vm_page_aflag_set(tm, PGA_REFERENCED); 1393 } 1394 vm_page_busy_sleep(tm, "madvpo", false); 1395 goto relookup; 1396 } 1397 vm_page_advise(tm, advice); 1398 vm_page_xunbusy(tm); 1399 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1400 next_pindex: 1401 if (tobject != object) 1402 VM_OBJECT_WUNLOCK(tobject); 1403 } 1404 VM_OBJECT_WUNLOCK(object); 1405 } 1406 1407 /* 1408 * vm_object_shadow: 1409 * 1410 * Create a new object which is backed by the 1411 * specified existing object range. The source 1412 * object reference is deallocated. 1413 * 1414 * The new object and offset into that object 1415 * are returned in the source parameters. 1416 */ 1417 void 1418 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1419 struct ucred *cred, bool shared) 1420 { 1421 vm_object_t source; 1422 vm_object_t result; 1423 1424 source = *object; 1425 1426 /* 1427 * Don't create the new object if the old object isn't shared. 1428 * 1429 * If we hold the only reference we can guarantee that it won't 1430 * increase while we have the map locked. Otherwise the race is 1431 * harmless and we will end up with an extra shadow object that 1432 * will be collapsed later. 1433 */ 1434 if (source != NULL && source->ref_count == 1 && 1435 (source->flags & OBJ_ANON) != 0) 1436 return; 1437 1438 /* 1439 * Allocate a new object with the given length. 1440 */ 1441 result = vm_object_allocate_anon(atop(length), source, cred, length); 1442 1443 /* 1444 * Store the offset into the source object, and fix up the offset into 1445 * the new object. 1446 */ 1447 result->backing_object_offset = *offset; 1448 1449 if (shared || source != NULL) { 1450 VM_OBJECT_WLOCK(result); 1451 1452 /* 1453 * The new object shadows the source object, adding a 1454 * reference to it. Our caller changes his reference 1455 * to point to the new object, removing a reference to 1456 * the source object. Net result: no change of 1457 * reference count, unless the caller needs to add one 1458 * more reference due to forking a shared map entry. 1459 */ 1460 if (shared) { 1461 vm_object_reference_locked(result); 1462 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1463 } 1464 1465 /* 1466 * Try to optimize the result object's page color when 1467 * shadowing in order to maintain page coloring 1468 * consistency in the combined shadowed object. 1469 */ 1470 if (source != NULL) { 1471 vm_object_backing_insert(result, source); 1472 result->domain = source->domain; 1473 #if VM_NRESERVLEVEL > 0 1474 result->flags |= source->flags & OBJ_COLORED; 1475 result->pg_color = (source->pg_color + 1476 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1477 1)) - 1); 1478 #endif 1479 } 1480 VM_OBJECT_WUNLOCK(result); 1481 } 1482 1483 /* 1484 * Return the new things 1485 */ 1486 *offset = 0; 1487 *object = result; 1488 } 1489 1490 /* 1491 * vm_object_split: 1492 * 1493 * Split the pages in a map entry into a new object. This affords 1494 * easier removal of unused pages, and keeps object inheritance from 1495 * being a negative impact on memory usage. 1496 */ 1497 void 1498 vm_object_split(vm_map_entry_t entry) 1499 { 1500 vm_page_t m, m_busy, m_next; 1501 vm_object_t orig_object, new_object, backing_object; 1502 vm_pindex_t idx, offidxstart; 1503 vm_size_t size; 1504 1505 orig_object = entry->object.vm_object; 1506 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1507 ("vm_object_split: Splitting object with multiple mappings.")); 1508 if ((orig_object->flags & OBJ_ANON) == 0) 1509 return; 1510 if (orig_object->ref_count <= 1) 1511 return; 1512 VM_OBJECT_WUNLOCK(orig_object); 1513 1514 offidxstart = OFF_TO_IDX(entry->offset); 1515 size = atop(entry->end - entry->start); 1516 1517 /* 1518 * If swap_pager_copy() is later called, it will convert new_object 1519 * into a swap object. 1520 */ 1521 new_object = vm_object_allocate_anon(size, orig_object, 1522 orig_object->cred, ptoa(size)); 1523 1524 /* 1525 * We must wait for the orig_object to complete any in-progress 1526 * collapse so that the swap blocks are stable below. The 1527 * additional reference on backing_object by new object will 1528 * prevent further collapse operations until split completes. 1529 */ 1530 VM_OBJECT_WLOCK(orig_object); 1531 vm_object_collapse_wait(orig_object); 1532 1533 /* 1534 * At this point, the new object is still private, so the order in 1535 * which the original and new objects are locked does not matter. 1536 */ 1537 VM_OBJECT_WLOCK(new_object); 1538 new_object->domain = orig_object->domain; 1539 backing_object = orig_object->backing_object; 1540 if (backing_object != NULL) { 1541 vm_object_backing_insert_ref(new_object, backing_object); 1542 new_object->backing_object_offset = 1543 orig_object->backing_object_offset + entry->offset; 1544 } 1545 if (orig_object->cred != NULL) { 1546 crhold(orig_object->cred); 1547 KASSERT(orig_object->charge >= ptoa(size), 1548 ("orig_object->charge < 0")); 1549 orig_object->charge -= ptoa(size); 1550 } 1551 1552 /* 1553 * Mark the split operation so that swap_pager_getpages() knows 1554 * that the object is in transition. 1555 */ 1556 vm_object_set_flag(orig_object, OBJ_SPLIT); 1557 m_busy = NULL; 1558 #ifdef INVARIANTS 1559 idx = 0; 1560 #endif 1561 retry: 1562 m = vm_page_find_least(orig_object, offidxstart); 1563 KASSERT(m == NULL || idx <= m->pindex - offidxstart, 1564 ("%s: object %p was repopulated", __func__, orig_object)); 1565 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1566 m = m_next) { 1567 m_next = TAILQ_NEXT(m, listq); 1568 1569 /* 1570 * We must wait for pending I/O to complete before we can 1571 * rename the page. 1572 * 1573 * We do not have to VM_PROT_NONE the page as mappings should 1574 * not be changed by this operation. 1575 */ 1576 if (vm_page_tryxbusy(m) == 0) { 1577 VM_OBJECT_WUNLOCK(new_object); 1578 vm_page_sleep_if_busy(m, "spltwt"); 1579 VM_OBJECT_WLOCK(new_object); 1580 goto retry; 1581 } 1582 1583 /* 1584 * The page was left invalid. Likely placed there by 1585 * an incomplete fault. Just remove and ignore. 1586 */ 1587 if (vm_page_none_valid(m)) { 1588 if (vm_page_remove(m)) 1589 vm_page_free(m); 1590 continue; 1591 } 1592 1593 /* vm_page_rename() will dirty the page. */ 1594 if (vm_page_rename(m, new_object, idx)) { 1595 vm_page_xunbusy(m); 1596 VM_OBJECT_WUNLOCK(new_object); 1597 VM_OBJECT_WUNLOCK(orig_object); 1598 vm_radix_wait(); 1599 VM_OBJECT_WLOCK(orig_object); 1600 VM_OBJECT_WLOCK(new_object); 1601 goto retry; 1602 } 1603 1604 #if VM_NRESERVLEVEL > 0 1605 /* 1606 * If some of the reservation's allocated pages remain with 1607 * the original object, then transferring the reservation to 1608 * the new object is neither particularly beneficial nor 1609 * particularly harmful as compared to leaving the reservation 1610 * with the original object. If, however, all of the 1611 * reservation's allocated pages are transferred to the new 1612 * object, then transferring the reservation is typically 1613 * beneficial. Determining which of these two cases applies 1614 * would be more costly than unconditionally renaming the 1615 * reservation. 1616 */ 1617 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1618 #endif 1619 1620 /* 1621 * orig_object's type may change while sleeping, so keep track 1622 * of the beginning of the busied range. 1623 */ 1624 if (orig_object->type != OBJT_SWAP) 1625 vm_page_xunbusy(m); 1626 else if (m_busy == NULL) 1627 m_busy = m; 1628 } 1629 if (orig_object->type == OBJT_SWAP) { 1630 /* 1631 * swap_pager_copy() can sleep, in which case the orig_object's 1632 * and new_object's locks are released and reacquired. 1633 */ 1634 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1635 if (m_busy != NULL) 1636 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq) 1637 vm_page_xunbusy(m_busy); 1638 } 1639 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1640 VM_OBJECT_WUNLOCK(orig_object); 1641 VM_OBJECT_WUNLOCK(new_object); 1642 entry->object.vm_object = new_object; 1643 entry->offset = 0LL; 1644 vm_object_deallocate(orig_object); 1645 VM_OBJECT_WLOCK(new_object); 1646 } 1647 1648 static vm_page_t 1649 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p) 1650 { 1651 vm_object_t backing_object; 1652 1653 VM_OBJECT_ASSERT_WLOCKED(object); 1654 backing_object = object->backing_object; 1655 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1656 1657 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1658 ("invalid ownership %p %p %p", p, object, backing_object)); 1659 /* The page is only NULL when rename fails. */ 1660 if (p == NULL) { 1661 VM_OBJECT_WUNLOCK(object); 1662 VM_OBJECT_WUNLOCK(backing_object); 1663 vm_radix_wait(); 1664 } else { 1665 if (p->object == object) 1666 VM_OBJECT_WUNLOCK(backing_object); 1667 else 1668 VM_OBJECT_WUNLOCK(object); 1669 vm_page_busy_sleep(p, "vmocol", false); 1670 } 1671 VM_OBJECT_WLOCK(object); 1672 VM_OBJECT_WLOCK(backing_object); 1673 return (TAILQ_FIRST(&backing_object->memq)); 1674 } 1675 1676 static bool 1677 vm_object_scan_all_shadowed(vm_object_t object) 1678 { 1679 vm_object_t backing_object; 1680 vm_page_t p, pp; 1681 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1682 1683 VM_OBJECT_ASSERT_WLOCKED(object); 1684 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1685 1686 backing_object = object->backing_object; 1687 1688 if ((backing_object->flags & OBJ_ANON) == 0) 1689 return (false); 1690 1691 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1692 p = vm_page_find_least(backing_object, pi); 1693 ps = swap_pager_find_least(backing_object, pi); 1694 1695 /* 1696 * Only check pages inside the parent object's range and 1697 * inside the parent object's mapping of the backing object. 1698 */ 1699 for (;; pi++) { 1700 if (p != NULL && p->pindex < pi) 1701 p = TAILQ_NEXT(p, listq); 1702 if (ps < pi) 1703 ps = swap_pager_find_least(backing_object, pi); 1704 if (p == NULL && ps >= backing_object->size) 1705 break; 1706 else if (p == NULL) 1707 pi = ps; 1708 else 1709 pi = MIN(p->pindex, ps); 1710 1711 new_pindex = pi - backing_offset_index; 1712 if (new_pindex >= object->size) 1713 break; 1714 1715 if (p != NULL) { 1716 /* 1717 * If the backing object page is busy a 1718 * grandparent or older page may still be 1719 * undergoing CoW. It is not safe to collapse 1720 * the backing object until it is quiesced. 1721 */ 1722 if (vm_page_tryxbusy(p) == 0) 1723 return (false); 1724 1725 /* 1726 * We raced with the fault handler that left 1727 * newly allocated invalid page on the object 1728 * queue and retried. 1729 */ 1730 if (!vm_page_all_valid(p)) 1731 goto unbusy_ret; 1732 } 1733 1734 /* 1735 * See if the parent has the page or if the parent's object 1736 * pager has the page. If the parent has the page but the page 1737 * is not valid, the parent's object pager must have the page. 1738 * 1739 * If this fails, the parent does not completely shadow the 1740 * object and we might as well give up now. 1741 */ 1742 pp = vm_page_lookup(object, new_pindex); 1743 1744 /* 1745 * The valid check here is stable due to object lock 1746 * being required to clear valid and initiate paging. 1747 * Busy of p disallows fault handler to validate pp. 1748 */ 1749 if ((pp == NULL || vm_page_none_valid(pp)) && 1750 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1751 goto unbusy_ret; 1752 if (p != NULL) 1753 vm_page_xunbusy(p); 1754 } 1755 return (true); 1756 1757 unbusy_ret: 1758 if (p != NULL) 1759 vm_page_xunbusy(p); 1760 return (false); 1761 } 1762 1763 static void 1764 vm_object_collapse_scan(vm_object_t object) 1765 { 1766 vm_object_t backing_object; 1767 vm_page_t next, p, pp; 1768 vm_pindex_t backing_offset_index, new_pindex; 1769 1770 VM_OBJECT_ASSERT_WLOCKED(object); 1771 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1772 1773 backing_object = object->backing_object; 1774 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1775 1776 /* 1777 * Our scan 1778 */ 1779 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1780 next = TAILQ_NEXT(p, listq); 1781 new_pindex = p->pindex - backing_offset_index; 1782 1783 /* 1784 * Check for busy page 1785 */ 1786 if (vm_page_tryxbusy(p) == 0) { 1787 next = vm_object_collapse_scan_wait(object, p); 1788 continue; 1789 } 1790 1791 KASSERT(object->backing_object == backing_object, 1792 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1793 object->backing_object, backing_object)); 1794 KASSERT(p->object == backing_object, 1795 ("vm_object_collapse_scan: object mismatch %p != %p", 1796 p->object, backing_object)); 1797 1798 if (p->pindex < backing_offset_index || 1799 new_pindex >= object->size) { 1800 if (backing_object->type == OBJT_SWAP) 1801 swap_pager_freespace(backing_object, p->pindex, 1802 1); 1803 1804 KASSERT(!pmap_page_is_mapped(p), 1805 ("freeing mapped page %p", p)); 1806 if (vm_page_remove(p)) 1807 vm_page_free(p); 1808 continue; 1809 } 1810 1811 if (!vm_page_all_valid(p)) { 1812 KASSERT(!pmap_page_is_mapped(p), 1813 ("freeing mapped page %p", p)); 1814 if (vm_page_remove(p)) 1815 vm_page_free(p); 1816 continue; 1817 } 1818 1819 pp = vm_page_lookup(object, new_pindex); 1820 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1821 vm_page_xunbusy(p); 1822 /* 1823 * The page in the parent is busy and possibly not 1824 * (yet) valid. Until its state is finalized by the 1825 * busy bit owner, we can't tell whether it shadows the 1826 * original page. 1827 */ 1828 next = vm_object_collapse_scan_wait(object, pp); 1829 continue; 1830 } 1831 1832 if (pp != NULL && vm_page_none_valid(pp)) { 1833 /* 1834 * The page was invalid in the parent. Likely placed 1835 * there by an incomplete fault. Just remove and 1836 * ignore. p can replace it. 1837 */ 1838 if (vm_page_remove(pp)) 1839 vm_page_free(pp); 1840 pp = NULL; 1841 } 1842 1843 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1844 NULL)) { 1845 /* 1846 * The page already exists in the parent OR swap exists 1847 * for this location in the parent. Leave the parent's 1848 * page alone. Destroy the original page from the 1849 * backing object. 1850 */ 1851 if (backing_object->type == OBJT_SWAP) 1852 swap_pager_freespace(backing_object, p->pindex, 1853 1); 1854 KASSERT(!pmap_page_is_mapped(p), 1855 ("freeing mapped page %p", p)); 1856 if (vm_page_remove(p)) 1857 vm_page_free(p); 1858 if (pp != NULL) 1859 vm_page_xunbusy(pp); 1860 continue; 1861 } 1862 1863 /* 1864 * Page does not exist in parent, rename the page from the 1865 * backing object to the main object. 1866 * 1867 * If the page was mapped to a process, it can remain mapped 1868 * through the rename. vm_page_rename() will dirty the page. 1869 */ 1870 if (vm_page_rename(p, object, new_pindex)) { 1871 vm_page_xunbusy(p); 1872 next = vm_object_collapse_scan_wait(object, NULL); 1873 continue; 1874 } 1875 1876 /* Use the old pindex to free the right page. */ 1877 if (backing_object->type == OBJT_SWAP) 1878 swap_pager_freespace(backing_object, 1879 new_pindex + backing_offset_index, 1); 1880 1881 #if VM_NRESERVLEVEL > 0 1882 /* 1883 * Rename the reservation. 1884 */ 1885 vm_reserv_rename(p, object, backing_object, 1886 backing_offset_index); 1887 #endif 1888 vm_page_xunbusy(p); 1889 } 1890 return; 1891 } 1892 1893 /* 1894 * vm_object_collapse: 1895 * 1896 * Collapse an object with the object backing it. 1897 * Pages in the backing object are moved into the 1898 * parent, and the backing object is deallocated. 1899 */ 1900 void 1901 vm_object_collapse(vm_object_t object) 1902 { 1903 vm_object_t backing_object, new_backing_object; 1904 1905 VM_OBJECT_ASSERT_WLOCKED(object); 1906 1907 while (TRUE) { 1908 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1909 ("collapsing invalid object")); 1910 1911 /* 1912 * Wait for the backing_object to finish any pending 1913 * collapse so that the caller sees the shortest possible 1914 * shadow chain. 1915 */ 1916 backing_object = vm_object_backing_collapse_wait(object); 1917 if (backing_object == NULL) 1918 return; 1919 1920 KASSERT(object->ref_count > 0 && 1921 object->ref_count > object->shadow_count, 1922 ("collapse with invalid ref %d or shadow %d count.", 1923 object->ref_count, object->shadow_count)); 1924 KASSERT((backing_object->flags & 1925 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1926 ("vm_object_collapse: Backing object already collapsing.")); 1927 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1928 ("vm_object_collapse: object is already collapsing.")); 1929 1930 /* 1931 * We know that we can either collapse the backing object if 1932 * the parent is the only reference to it, or (perhaps) have 1933 * the parent bypass the object if the parent happens to shadow 1934 * all the resident pages in the entire backing object. 1935 */ 1936 if (backing_object->ref_count == 1) { 1937 KASSERT(backing_object->shadow_count == 1, 1938 ("vm_object_collapse: shadow_count: %d", 1939 backing_object->shadow_count)); 1940 vm_object_pip_add(object, 1); 1941 vm_object_set_flag(object, OBJ_COLLAPSING); 1942 vm_object_pip_add(backing_object, 1); 1943 vm_object_set_flag(backing_object, OBJ_DEAD); 1944 1945 /* 1946 * If there is exactly one reference to the backing 1947 * object, we can collapse it into the parent. 1948 */ 1949 vm_object_collapse_scan(object); 1950 1951 #if VM_NRESERVLEVEL > 0 1952 /* 1953 * Break any reservations from backing_object. 1954 */ 1955 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1956 vm_reserv_break_all(backing_object); 1957 #endif 1958 1959 /* 1960 * Move the pager from backing_object to object. 1961 */ 1962 if (backing_object->type == OBJT_SWAP) { 1963 /* 1964 * swap_pager_copy() can sleep, in which case 1965 * the backing_object's and object's locks are 1966 * released and reacquired. 1967 * Since swap_pager_copy() is being asked to 1968 * destroy backing_object, it will change the 1969 * type to OBJT_DEFAULT. 1970 */ 1971 swap_pager_copy( 1972 backing_object, 1973 object, 1974 OFF_TO_IDX(object->backing_object_offset), TRUE); 1975 } 1976 1977 /* 1978 * Object now shadows whatever backing_object did. 1979 */ 1980 vm_object_clear_flag(object, OBJ_COLLAPSING); 1981 vm_object_backing_transfer(object, backing_object); 1982 object->backing_object_offset += 1983 backing_object->backing_object_offset; 1984 VM_OBJECT_WUNLOCK(object); 1985 vm_object_pip_wakeup(object); 1986 1987 /* 1988 * Discard backing_object. 1989 * 1990 * Since the backing object has no pages, no pager left, 1991 * and no object references within it, all that is 1992 * necessary is to dispose of it. 1993 */ 1994 KASSERT(backing_object->ref_count == 1, ( 1995 "backing_object %p was somehow re-referenced during collapse!", 1996 backing_object)); 1997 vm_object_pip_wakeup(backing_object); 1998 (void)refcount_release(&backing_object->ref_count); 1999 vm_object_terminate(backing_object); 2000 counter_u64_add(object_collapses, 1); 2001 VM_OBJECT_WLOCK(object); 2002 } else { 2003 /* 2004 * If we do not entirely shadow the backing object, 2005 * there is nothing we can do so we give up. 2006 * 2007 * The object lock and backing_object lock must not 2008 * be dropped during this sequence. 2009 */ 2010 if (!vm_object_scan_all_shadowed(object)) { 2011 VM_OBJECT_WUNLOCK(backing_object); 2012 break; 2013 } 2014 2015 /* 2016 * Make the parent shadow the next object in the 2017 * chain. Deallocating backing_object will not remove 2018 * it, since its reference count is at least 2. 2019 */ 2020 vm_object_backing_remove_locked(object); 2021 new_backing_object = backing_object->backing_object; 2022 if (new_backing_object != NULL) { 2023 vm_object_backing_insert_ref(object, 2024 new_backing_object); 2025 object->backing_object_offset += 2026 backing_object->backing_object_offset; 2027 } 2028 2029 /* 2030 * Drop the reference count on backing_object. Since 2031 * its ref_count was at least 2, it will not vanish. 2032 */ 2033 (void)refcount_release(&backing_object->ref_count); 2034 KASSERT(backing_object->ref_count >= 1, ( 2035 "backing_object %p was somehow dereferenced during collapse!", 2036 backing_object)); 2037 VM_OBJECT_WUNLOCK(backing_object); 2038 counter_u64_add(object_bypasses, 1); 2039 } 2040 2041 /* 2042 * Try again with this object's new backing object. 2043 */ 2044 } 2045 } 2046 2047 /* 2048 * vm_object_page_remove: 2049 * 2050 * For the given object, either frees or invalidates each of the 2051 * specified pages. In general, a page is freed. However, if a page is 2052 * wired for any reason other than the existence of a managed, wired 2053 * mapping, then it may be invalidated but not removed from the object. 2054 * Pages are specified by the given range ["start", "end") and the option 2055 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 2056 * extends from "start" to the end of the object. If the option 2057 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 2058 * specified range are affected. If the option OBJPR_NOTMAPPED is 2059 * specified, then the pages within the specified range must have no 2060 * mappings. Otherwise, if this option is not specified, any mappings to 2061 * the specified pages are removed before the pages are freed or 2062 * invalidated. 2063 * 2064 * In general, this operation should only be performed on objects that 2065 * contain managed pages. There are, however, two exceptions. First, it 2066 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 2067 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 2068 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 2069 * not be specified and the option OBJPR_NOTMAPPED must be specified. 2070 * 2071 * The object must be locked. 2072 */ 2073 void 2074 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2075 int options) 2076 { 2077 vm_page_t p, next; 2078 2079 VM_OBJECT_ASSERT_WLOCKED(object); 2080 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 2081 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 2082 ("vm_object_page_remove: illegal options for object %p", object)); 2083 if (object->resident_page_count == 0) 2084 return; 2085 vm_object_pip_add(object, 1); 2086 again: 2087 p = vm_page_find_least(object, start); 2088 2089 /* 2090 * Here, the variable "p" is either (1) the page with the least pindex 2091 * greater than or equal to the parameter "start" or (2) NULL. 2092 */ 2093 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2094 next = TAILQ_NEXT(p, listq); 2095 2096 /* 2097 * If the page is wired for any reason besides the existence 2098 * of managed, wired mappings, then it cannot be freed. For 2099 * example, fictitious pages, which represent device memory, 2100 * are inherently wired and cannot be freed. They can, 2101 * however, be invalidated if the option OBJPR_CLEANONLY is 2102 * not specified. 2103 */ 2104 if (vm_page_tryxbusy(p) == 0) { 2105 vm_page_sleep_if_busy(p, "vmopar"); 2106 goto again; 2107 } 2108 if (vm_page_wired(p)) { 2109 wired: 2110 if ((options & OBJPR_NOTMAPPED) == 0 && 2111 object->ref_count != 0) 2112 pmap_remove_all(p); 2113 if ((options & OBJPR_CLEANONLY) == 0) { 2114 vm_page_invalid(p); 2115 vm_page_undirty(p); 2116 } 2117 vm_page_xunbusy(p); 2118 continue; 2119 } 2120 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2121 ("vm_object_page_remove: page %p is fictitious", p)); 2122 if ((options & OBJPR_CLEANONLY) != 0 && 2123 !vm_page_none_valid(p)) { 2124 if ((options & OBJPR_NOTMAPPED) == 0 && 2125 object->ref_count != 0 && 2126 !vm_page_try_remove_write(p)) 2127 goto wired; 2128 if (p->dirty != 0) { 2129 vm_page_xunbusy(p); 2130 continue; 2131 } 2132 } 2133 if ((options & OBJPR_NOTMAPPED) == 0 && 2134 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2135 goto wired; 2136 vm_page_free(p); 2137 } 2138 vm_object_pip_wakeup(object); 2139 2140 if (object->type == OBJT_SWAP) { 2141 if (end == 0) 2142 end = object->size; 2143 swap_pager_freespace(object, start, end - start); 2144 } 2145 } 2146 2147 /* 2148 * vm_object_page_noreuse: 2149 * 2150 * For the given object, attempt to move the specified pages to 2151 * the head of the inactive queue. This bypasses regular LRU 2152 * operation and allows the pages to be reused quickly under memory 2153 * pressure. If a page is wired for any reason, then it will not 2154 * be queued. Pages are specified by the range ["start", "end"). 2155 * As a special case, if "end" is zero, then the range extends from 2156 * "start" to the end of the object. 2157 * 2158 * This operation should only be performed on objects that 2159 * contain non-fictitious, managed pages. 2160 * 2161 * The object must be locked. 2162 */ 2163 void 2164 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2165 { 2166 vm_page_t p, next; 2167 2168 VM_OBJECT_ASSERT_LOCKED(object); 2169 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2170 ("vm_object_page_noreuse: illegal object %p", object)); 2171 if (object->resident_page_count == 0) 2172 return; 2173 p = vm_page_find_least(object, start); 2174 2175 /* 2176 * Here, the variable "p" is either (1) the page with the least pindex 2177 * greater than or equal to the parameter "start" or (2) NULL. 2178 */ 2179 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2180 next = TAILQ_NEXT(p, listq); 2181 vm_page_deactivate_noreuse(p); 2182 } 2183 } 2184 2185 /* 2186 * Populate the specified range of the object with valid pages. Returns 2187 * TRUE if the range is successfully populated and FALSE otherwise. 2188 * 2189 * Note: This function should be optimized to pass a larger array of 2190 * pages to vm_pager_get_pages() before it is applied to a non- 2191 * OBJT_DEVICE object. 2192 * 2193 * The object must be locked. 2194 */ 2195 boolean_t 2196 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2197 { 2198 vm_page_t m; 2199 vm_pindex_t pindex; 2200 int rv; 2201 2202 VM_OBJECT_ASSERT_WLOCKED(object); 2203 for (pindex = start; pindex < end; pindex++) { 2204 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); 2205 if (rv != VM_PAGER_OK) 2206 break; 2207 2208 /* 2209 * Keep "m" busy because a subsequent iteration may unlock 2210 * the object. 2211 */ 2212 } 2213 if (pindex > start) { 2214 m = vm_page_lookup(object, start); 2215 while (m != NULL && m->pindex < pindex) { 2216 vm_page_xunbusy(m); 2217 m = TAILQ_NEXT(m, listq); 2218 } 2219 } 2220 return (pindex == end); 2221 } 2222 2223 /* 2224 * Routine: vm_object_coalesce 2225 * Function: Coalesces two objects backing up adjoining 2226 * regions of memory into a single object. 2227 * 2228 * returns TRUE if objects were combined. 2229 * 2230 * NOTE: Only works at the moment if the second object is NULL - 2231 * if it's not, which object do we lock first? 2232 * 2233 * Parameters: 2234 * prev_object First object to coalesce 2235 * prev_offset Offset into prev_object 2236 * prev_size Size of reference to prev_object 2237 * next_size Size of reference to the second object 2238 * reserved Indicator that extension region has 2239 * swap accounted for 2240 * 2241 * Conditions: 2242 * The object must *not* be locked. 2243 */ 2244 boolean_t 2245 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2246 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2247 { 2248 vm_pindex_t next_pindex; 2249 2250 if (prev_object == NULL) 2251 return (TRUE); 2252 if ((prev_object->flags & OBJ_ANON) == 0) 2253 return (FALSE); 2254 2255 VM_OBJECT_WLOCK(prev_object); 2256 /* 2257 * Try to collapse the object first. 2258 */ 2259 vm_object_collapse(prev_object); 2260 2261 /* 2262 * Can't coalesce if: . more than one reference . paged out . shadows 2263 * another object . has a copy elsewhere (any of which mean that the 2264 * pages not mapped to prev_entry may be in use anyway) 2265 */ 2266 if (prev_object->backing_object != NULL) { 2267 VM_OBJECT_WUNLOCK(prev_object); 2268 return (FALSE); 2269 } 2270 2271 prev_size >>= PAGE_SHIFT; 2272 next_size >>= PAGE_SHIFT; 2273 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2274 2275 if (prev_object->ref_count > 1 && 2276 prev_object->size != next_pindex && 2277 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2278 VM_OBJECT_WUNLOCK(prev_object); 2279 return (FALSE); 2280 } 2281 2282 /* 2283 * Account for the charge. 2284 */ 2285 if (prev_object->cred != NULL) { 2286 /* 2287 * If prev_object was charged, then this mapping, 2288 * although not charged now, may become writable 2289 * later. Non-NULL cred in the object would prevent 2290 * swap reservation during enabling of the write 2291 * access, so reserve swap now. Failed reservation 2292 * cause allocation of the separate object for the map 2293 * entry, and swap reservation for this entry is 2294 * managed in appropriate time. 2295 */ 2296 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2297 prev_object->cred)) { 2298 VM_OBJECT_WUNLOCK(prev_object); 2299 return (FALSE); 2300 } 2301 prev_object->charge += ptoa(next_size); 2302 } 2303 2304 /* 2305 * Remove any pages that may still be in the object from a previous 2306 * deallocation. 2307 */ 2308 if (next_pindex < prev_object->size) { 2309 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2310 next_size, 0); 2311 #if 0 2312 if (prev_object->cred != NULL) { 2313 KASSERT(prev_object->charge >= 2314 ptoa(prev_object->size - next_pindex), 2315 ("object %p overcharged 1 %jx %jx", prev_object, 2316 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2317 prev_object->charge -= ptoa(prev_object->size - 2318 next_pindex); 2319 } 2320 #endif 2321 } 2322 2323 /* 2324 * Extend the object if necessary. 2325 */ 2326 if (next_pindex + next_size > prev_object->size) 2327 prev_object->size = next_pindex + next_size; 2328 2329 VM_OBJECT_WUNLOCK(prev_object); 2330 return (TRUE); 2331 } 2332 2333 void 2334 vm_object_set_writeable_dirty(vm_object_t object) 2335 { 2336 2337 /* Only set for vnodes & tmpfs */ 2338 if (object->type != OBJT_VNODE && 2339 (object->flags & OBJ_TMPFS_NODE) == 0) 2340 return; 2341 atomic_add_int(&object->generation, 1); 2342 } 2343 2344 /* 2345 * vm_object_unwire: 2346 * 2347 * For each page offset within the specified range of the given object, 2348 * find the highest-level page in the shadow chain and unwire it. A page 2349 * must exist at every page offset, and the highest-level page must be 2350 * wired. 2351 */ 2352 void 2353 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2354 uint8_t queue) 2355 { 2356 vm_object_t tobject, t1object; 2357 vm_page_t m, tm; 2358 vm_pindex_t end_pindex, pindex, tpindex; 2359 int depth, locked_depth; 2360 2361 KASSERT((offset & PAGE_MASK) == 0, 2362 ("vm_object_unwire: offset is not page aligned")); 2363 KASSERT((length & PAGE_MASK) == 0, 2364 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2365 /* The wired count of a fictitious page never changes. */ 2366 if ((object->flags & OBJ_FICTITIOUS) != 0) 2367 return; 2368 pindex = OFF_TO_IDX(offset); 2369 end_pindex = pindex + atop(length); 2370 again: 2371 locked_depth = 1; 2372 VM_OBJECT_RLOCK(object); 2373 m = vm_page_find_least(object, pindex); 2374 while (pindex < end_pindex) { 2375 if (m == NULL || pindex < m->pindex) { 2376 /* 2377 * The first object in the shadow chain doesn't 2378 * contain a page at the current index. Therefore, 2379 * the page must exist in a backing object. 2380 */ 2381 tobject = object; 2382 tpindex = pindex; 2383 depth = 0; 2384 do { 2385 tpindex += 2386 OFF_TO_IDX(tobject->backing_object_offset); 2387 tobject = tobject->backing_object; 2388 KASSERT(tobject != NULL, 2389 ("vm_object_unwire: missing page")); 2390 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2391 goto next_page; 2392 depth++; 2393 if (depth == locked_depth) { 2394 locked_depth++; 2395 VM_OBJECT_RLOCK(tobject); 2396 } 2397 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2398 NULL); 2399 } else { 2400 tm = m; 2401 m = TAILQ_NEXT(m, listq); 2402 } 2403 if (vm_page_trysbusy(tm) == 0) { 2404 for (tobject = object; locked_depth >= 1; 2405 locked_depth--) { 2406 t1object = tobject->backing_object; 2407 if (tm->object != tobject) 2408 VM_OBJECT_RUNLOCK(tobject); 2409 tobject = t1object; 2410 } 2411 vm_page_busy_sleep(tm, "unwbo", true); 2412 goto again; 2413 } 2414 vm_page_unwire(tm, queue); 2415 vm_page_sunbusy(tm); 2416 next_page: 2417 pindex++; 2418 } 2419 /* Release the accumulated object locks. */ 2420 for (tobject = object; locked_depth >= 1; locked_depth--) { 2421 t1object = tobject->backing_object; 2422 VM_OBJECT_RUNLOCK(tobject); 2423 tobject = t1object; 2424 } 2425 } 2426 2427 /* 2428 * Return the vnode for the given object, or NULL if none exists. 2429 * For tmpfs objects, the function may return NULL if there is 2430 * no vnode allocated at the time of the call. 2431 */ 2432 struct vnode * 2433 vm_object_vnode(vm_object_t object) 2434 { 2435 struct vnode *vp; 2436 2437 VM_OBJECT_ASSERT_LOCKED(object); 2438 if (object->type == OBJT_VNODE) { 2439 vp = object->handle; 2440 KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); 2441 } else if (object->type == OBJT_SWAP && 2442 (object->flags & OBJ_TMPFS) != 0) { 2443 vp = object->un_pager.swp.swp_tmpfs; 2444 KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); 2445 } else { 2446 vp = NULL; 2447 } 2448 return (vp); 2449 } 2450 2451 /* 2452 * Busy the vm object. This prevents new pages belonging to the object from 2453 * becoming busy. Existing pages persist as busy. Callers are responsible 2454 * for checking page state before proceeding. 2455 */ 2456 void 2457 vm_object_busy(vm_object_t obj) 2458 { 2459 2460 VM_OBJECT_ASSERT_LOCKED(obj); 2461 2462 blockcount_acquire(&obj->busy, 1); 2463 /* The fence is required to order loads of page busy. */ 2464 atomic_thread_fence_acq_rel(); 2465 } 2466 2467 void 2468 vm_object_unbusy(vm_object_t obj) 2469 { 2470 2471 blockcount_release(&obj->busy, 1); 2472 } 2473 2474 void 2475 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2476 { 2477 2478 VM_OBJECT_ASSERT_UNLOCKED(obj); 2479 2480 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2481 } 2482 2483 /* 2484 * Return the kvme type of the given object. 2485 * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. 2486 */ 2487 int 2488 vm_object_kvme_type(vm_object_t object, struct vnode **vpp) 2489 { 2490 2491 VM_OBJECT_ASSERT_LOCKED(object); 2492 if (vpp != NULL) 2493 *vpp = vm_object_vnode(object); 2494 switch (object->type) { 2495 case OBJT_DEFAULT: 2496 return (KVME_TYPE_DEFAULT); 2497 case OBJT_VNODE: 2498 return (KVME_TYPE_VNODE); 2499 case OBJT_SWAP: 2500 if ((object->flags & OBJ_TMPFS_NODE) != 0) 2501 return (KVME_TYPE_VNODE); 2502 return (KVME_TYPE_SWAP); 2503 case OBJT_DEVICE: 2504 return (KVME_TYPE_DEVICE); 2505 case OBJT_PHYS: 2506 return (KVME_TYPE_PHYS); 2507 case OBJT_DEAD: 2508 return (KVME_TYPE_DEAD); 2509 case OBJT_SG: 2510 return (KVME_TYPE_SG); 2511 case OBJT_MGTDEVICE: 2512 return (KVME_TYPE_MGTDEVICE); 2513 default: 2514 return (KVME_TYPE_UNKNOWN); 2515 } 2516 } 2517 2518 static int 2519 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2520 { 2521 struct kinfo_vmobject *kvo; 2522 char *fullpath, *freepath; 2523 struct vnode *vp; 2524 struct vattr va; 2525 vm_object_t obj; 2526 vm_page_t m; 2527 int count, error; 2528 2529 if (req->oldptr == NULL) { 2530 /* 2531 * If an old buffer has not been provided, generate an 2532 * estimate of the space needed for a subsequent call. 2533 */ 2534 mtx_lock(&vm_object_list_mtx); 2535 count = 0; 2536 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2537 if (obj->type == OBJT_DEAD) 2538 continue; 2539 count++; 2540 } 2541 mtx_unlock(&vm_object_list_mtx); 2542 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2543 count * 11 / 10)); 2544 } 2545 2546 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); 2547 error = 0; 2548 2549 /* 2550 * VM objects are type stable and are never removed from the 2551 * list once added. This allows us to safely read obj->object_list 2552 * after reacquiring the VM object lock. 2553 */ 2554 mtx_lock(&vm_object_list_mtx); 2555 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2556 if (obj->type == OBJT_DEAD) 2557 continue; 2558 VM_OBJECT_RLOCK(obj); 2559 if (obj->type == OBJT_DEAD) { 2560 VM_OBJECT_RUNLOCK(obj); 2561 continue; 2562 } 2563 mtx_unlock(&vm_object_list_mtx); 2564 kvo->kvo_size = ptoa(obj->size); 2565 kvo->kvo_resident = obj->resident_page_count; 2566 kvo->kvo_ref_count = obj->ref_count; 2567 kvo->kvo_shadow_count = obj->shadow_count; 2568 kvo->kvo_memattr = obj->memattr; 2569 kvo->kvo_active = 0; 2570 kvo->kvo_inactive = 0; 2571 TAILQ_FOREACH(m, &obj->memq, listq) { 2572 /* 2573 * A page may belong to the object but be 2574 * dequeued and set to PQ_NONE while the 2575 * object lock is not held. This makes the 2576 * reads of m->queue below racy, and we do not 2577 * count pages set to PQ_NONE. However, this 2578 * sysctl is only meant to give an 2579 * approximation of the system anyway. 2580 */ 2581 if (m->a.queue == PQ_ACTIVE) 2582 kvo->kvo_active++; 2583 else if (m->a.queue == PQ_INACTIVE) 2584 kvo->kvo_inactive++; 2585 } 2586 2587 kvo->kvo_vn_fileid = 0; 2588 kvo->kvo_vn_fsid = 0; 2589 kvo->kvo_vn_fsid_freebsd11 = 0; 2590 freepath = NULL; 2591 fullpath = ""; 2592 kvo->kvo_type = vm_object_kvme_type(obj, &vp); 2593 if (vp != NULL) 2594 vref(vp); 2595 VM_OBJECT_RUNLOCK(obj); 2596 if (vp != NULL) { 2597 vn_fullpath(vp, &fullpath, &freepath); 2598 vn_lock(vp, LK_SHARED | LK_RETRY); 2599 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2600 kvo->kvo_vn_fileid = va.va_fileid; 2601 kvo->kvo_vn_fsid = va.va_fsid; 2602 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2603 /* truncate */ 2604 } 2605 vput(vp); 2606 } 2607 2608 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2609 if (freepath != NULL) 2610 free(freepath, M_TEMP); 2611 2612 /* Pack record size down */ 2613 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2614 + strlen(kvo->kvo_path) + 1; 2615 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2616 sizeof(uint64_t)); 2617 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2618 mtx_lock(&vm_object_list_mtx); 2619 if (error) 2620 break; 2621 } 2622 mtx_unlock(&vm_object_list_mtx); 2623 free(kvo, M_TEMP); 2624 return (error); 2625 } 2626 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2627 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2628 "List of VM objects"); 2629 2630 #include "opt_ddb.h" 2631 #ifdef DDB 2632 #include <sys/kernel.h> 2633 2634 #include <sys/cons.h> 2635 2636 #include <ddb/ddb.h> 2637 2638 static int 2639 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2640 { 2641 vm_map_t tmpm; 2642 vm_map_entry_t tmpe; 2643 vm_object_t obj; 2644 2645 if (map == 0) 2646 return 0; 2647 2648 if (entry == 0) { 2649 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2650 if (_vm_object_in_map(map, object, tmpe)) { 2651 return 1; 2652 } 2653 } 2654 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2655 tmpm = entry->object.sub_map; 2656 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2657 if (_vm_object_in_map(tmpm, object, tmpe)) { 2658 return 1; 2659 } 2660 } 2661 } else if ((obj = entry->object.vm_object) != NULL) { 2662 for (; obj; obj = obj->backing_object) 2663 if (obj == object) { 2664 return 1; 2665 } 2666 } 2667 return 0; 2668 } 2669 2670 static int 2671 vm_object_in_map(vm_object_t object) 2672 { 2673 struct proc *p; 2674 2675 /* sx_slock(&allproc_lock); */ 2676 FOREACH_PROC_IN_SYSTEM(p) { 2677 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2678 continue; 2679 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2680 /* sx_sunlock(&allproc_lock); */ 2681 return 1; 2682 } 2683 } 2684 /* sx_sunlock(&allproc_lock); */ 2685 if (_vm_object_in_map(kernel_map, object, 0)) 2686 return 1; 2687 return 0; 2688 } 2689 2690 DB_SHOW_COMMAND(vmochk, vm_object_check) 2691 { 2692 vm_object_t object; 2693 2694 /* 2695 * make sure that internal objs are in a map somewhere 2696 * and none have zero ref counts. 2697 */ 2698 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2699 if ((object->flags & OBJ_ANON) != 0) { 2700 if (object->ref_count == 0) { 2701 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2702 (long)object->size); 2703 } 2704 if (!vm_object_in_map(object)) { 2705 db_printf( 2706 "vmochk: internal obj is not in a map: " 2707 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2708 object->ref_count, (u_long)object->size, 2709 (u_long)object->size, 2710 (void *)object->backing_object); 2711 } 2712 } 2713 if (db_pager_quit) 2714 return; 2715 } 2716 } 2717 2718 /* 2719 * vm_object_print: [ debug ] 2720 */ 2721 DB_SHOW_COMMAND(object, vm_object_print_static) 2722 { 2723 /* XXX convert args. */ 2724 vm_object_t object = (vm_object_t)addr; 2725 boolean_t full = have_addr; 2726 2727 vm_page_t p; 2728 2729 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2730 #define count was_count 2731 2732 int count; 2733 2734 if (object == NULL) 2735 return; 2736 2737 db_iprintf( 2738 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2739 object, (int)object->type, (uintmax_t)object->size, 2740 object->resident_page_count, object->ref_count, object->flags, 2741 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2742 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2743 object->shadow_count, 2744 object->backing_object ? object->backing_object->ref_count : 0, 2745 object->backing_object, (uintmax_t)object->backing_object_offset); 2746 2747 if (!full) 2748 return; 2749 2750 db_indent += 2; 2751 count = 0; 2752 TAILQ_FOREACH(p, &object->memq, listq) { 2753 if (count == 0) 2754 db_iprintf("memory:="); 2755 else if (count == 6) { 2756 db_printf("\n"); 2757 db_iprintf(" ..."); 2758 count = 0; 2759 } else 2760 db_printf(","); 2761 count++; 2762 2763 db_printf("(off=0x%jx,page=0x%jx)", 2764 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2765 2766 if (db_pager_quit) 2767 break; 2768 } 2769 if (count != 0) 2770 db_printf("\n"); 2771 db_indent -= 2; 2772 } 2773 2774 /* XXX. */ 2775 #undef count 2776 2777 /* XXX need this non-static entry for calling from vm_map_print. */ 2778 void 2779 vm_object_print( 2780 /* db_expr_t */ long addr, 2781 boolean_t have_addr, 2782 /* db_expr_t */ long count, 2783 char *modif) 2784 { 2785 vm_object_print_static(addr, have_addr, count, modif); 2786 } 2787 2788 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2789 { 2790 vm_object_t object; 2791 vm_pindex_t fidx; 2792 vm_paddr_t pa; 2793 vm_page_t m, prev_m; 2794 int rcount, nl, c; 2795 2796 nl = 0; 2797 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2798 db_printf("new object: %p\n", (void *)object); 2799 if (nl > 18) { 2800 c = cngetc(); 2801 if (c != ' ') 2802 return; 2803 nl = 0; 2804 } 2805 nl++; 2806 rcount = 0; 2807 fidx = 0; 2808 pa = -1; 2809 TAILQ_FOREACH(m, &object->memq, listq) { 2810 if (m->pindex > 128) 2811 break; 2812 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2813 prev_m->pindex + 1 != m->pindex) { 2814 if (rcount) { 2815 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2816 (long)fidx, rcount, (long)pa); 2817 if (nl > 18) { 2818 c = cngetc(); 2819 if (c != ' ') 2820 return; 2821 nl = 0; 2822 } 2823 nl++; 2824 rcount = 0; 2825 } 2826 } 2827 if (rcount && 2828 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2829 ++rcount; 2830 continue; 2831 } 2832 if (rcount) { 2833 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2834 (long)fidx, rcount, (long)pa); 2835 if (nl > 18) { 2836 c = cngetc(); 2837 if (c != ' ') 2838 return; 2839 nl = 0; 2840 } 2841 nl++; 2842 } 2843 fidx = m->pindex; 2844 pa = VM_PAGE_TO_PHYS(m); 2845 rcount = 1; 2846 } 2847 if (rcount) { 2848 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2849 (long)fidx, rcount, (long)pa); 2850 if (nl > 18) { 2851 c = cngetc(); 2852 if (c != ' ') 2853 return; 2854 nl = 0; 2855 } 2856 nl++; 2857 } 2858 } 2859 } 2860 #endif /* DDB */ 2861