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