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 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61 /* 62 * Virtual memory object module. 63 */ 64 65 #include "opt_vm.h" 66 67 #include <sys/systm.h> 68 #include <sys/blockcount.h> 69 #include <sys/cpuset.h> 70 #include <sys/ipc.h> 71 #include <sys/jail.h> 72 #include <sys/limits.h> 73 #include <sys/lock.h> 74 #include <sys/mman.h> 75 #include <sys/mount.h> 76 #include <sys/kernel.h> 77 #include <sys/mutex.h> 78 #include <sys/pctrie.h> 79 #include <sys/proc.h> 80 #include <sys/refcount.h> 81 #include <sys/shm.h> 82 #include <sys/sx.h> 83 #include <sys/sysctl.h> 84 #include <sys/resourcevar.h> 85 #include <sys/refcount.h> 86 #include <sys/rwlock.h> 87 #include <sys/user.h> 88 #include <sys/vnode.h> 89 #include <sys/vmmeter.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_param.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_map.h> 95 #include <vm/vm_object.h> 96 #include <vm/vm_page.h> 97 #include <vm/vm_pageout.h> 98 #include <vm/vm_pager.h> 99 #include <vm/vm_phys.h> 100 #include <vm/vm_pagequeue.h> 101 #include <vm/swap_pager.h> 102 #include <vm/vm_kern.h> 103 #include <vm/vm_extern.h> 104 #include <vm/vm_radix.h> 105 #include <vm/vm_reserv.h> 106 #include <vm/uma.h> 107 108 static int old_msync; 109 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 110 "Use old (insecure) msync behavior"); 111 112 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 113 int pagerflags, int flags, boolean_t *allclean, 114 boolean_t *eio); 115 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 116 boolean_t *allclean); 117 static void vm_object_backing_remove(vm_object_t object); 118 119 /* 120 * Virtual memory objects maintain the actual data 121 * associated with allocated virtual memory. A given 122 * page of memory exists within exactly one object. 123 * 124 * An object is only deallocated when all "references" 125 * are given up. Only one "reference" to a given 126 * region of an object should be writeable. 127 * 128 * Associated with each object is a list of all resident 129 * memory pages belonging to that object; this list is 130 * maintained by the "vm_page" module, and locked by the object's 131 * lock. 132 * 133 * Each object also records a "pager" routine which is 134 * used to retrieve (and store) pages to the proper backing 135 * storage. In addition, objects may be backed by other 136 * objects from which they were virtual-copied. 137 * 138 * The only items within the object structure which are 139 * modified after time of creation are: 140 * reference count locked by object's lock 141 * pager routine locked by object's lock 142 * 143 */ 144 145 struct object_q vm_object_list; 146 struct mtx vm_object_list_mtx; /* lock for object list and count */ 147 148 struct vm_object kernel_object_store; 149 150 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 151 "VM object stats"); 152 153 static COUNTER_U64_DEFINE_EARLY(object_collapses); 154 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 155 &object_collapses, 156 "VM object collapses"); 157 158 static COUNTER_U64_DEFINE_EARLY(object_bypasses); 159 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 160 &object_bypasses, 161 "VM object bypasses"); 162 163 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits); 164 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD, 165 &object_collapse_waits, 166 "Number of sleeps for collapse"); 167 168 static uma_zone_t obj_zone; 169 170 static int vm_object_zinit(void *mem, int size, int flags); 171 172 #ifdef INVARIANTS 173 static void vm_object_zdtor(void *mem, int size, void *arg); 174 175 static void 176 vm_object_zdtor(void *mem, int size, void *arg) 177 { 178 vm_object_t object; 179 180 object = (vm_object_t)mem; 181 KASSERT(object->ref_count == 0, 182 ("object %p ref_count = %d", object, object->ref_count)); 183 KASSERT(TAILQ_EMPTY(&object->memq), 184 ("object %p has resident pages in its memq", object)); 185 KASSERT(vm_radix_is_empty(&object->rtree), 186 ("object %p has resident pages in its trie", object)); 187 #if VM_NRESERVLEVEL > 0 188 KASSERT(LIST_EMPTY(&object->rvq), 189 ("object %p has reservations", 190 object)); 191 #endif 192 KASSERT(!vm_object_busied(object), 193 ("object %p busy = %d", object, blockcount_read(&object->busy))); 194 KASSERT(object->resident_page_count == 0, 195 ("object %p resident_page_count = %d", 196 object, object->resident_page_count)); 197 KASSERT(atomic_load_int(&object->shadow_count) == 0, 198 ("object %p shadow_count = %d", 199 object, atomic_load_int(&object->shadow_count))); 200 KASSERT(object->type == OBJT_DEAD, 201 ("object %p has non-dead type %d", 202 object, object->type)); 203 KASSERT(object->charge == 0 && object->cred == NULL, 204 ("object %p has non-zero charge %ju (%p)", 205 object, (uintmax_t)object->charge, object->cred)); 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, "vmobject", 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 atomic_store_int(&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 object->flags = flags; 243 if ((flags & OBJ_SWAP) != 0) { 244 pctrie_init(&object->un_pager.swp.swp_blks); 245 object->un_pager.swp.writemappings = 0; 246 } 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->size = size; 257 object->domain.dr_policy = NULL; 258 object->generation = 1; 259 object->cleangeneration = 1; 260 refcount_init(&object->ref_count, 1); 261 object->memattr = VM_MEMATTR_DEFAULT; 262 object->cred = NULL; 263 object->charge = 0; 264 object->handle = handle; 265 object->backing_object = NULL; 266 object->backing_object_offset = (vm_ooffset_t) 0; 267 #if VM_NRESERVLEVEL > 0 268 LIST_INIT(&object->rvq); 269 #endif 270 umtx_shm_object_init(object); 271 } 272 273 /* 274 * vm_object_init: 275 * 276 * Initialize the VM objects module. 277 */ 278 void 279 vm_object_init(void) 280 { 281 TAILQ_INIT(&vm_object_list); 282 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 283 284 rw_init(&kernel_object->lock, "kernel vm object"); 285 vm_radix_init(&kernel_object->rtree); 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 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 293 294 /* 295 * The lock portion of struct vm_object must be type stable due 296 * to vm_pageout_fallback_object_lock locking a vm object 297 * without holding any references to it. 298 * 299 * paging_in_progress is valid always. Lockless references to 300 * the objects may acquire pip and then check OBJ_DEAD. 301 */ 302 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 303 #ifdef INVARIANTS 304 vm_object_zdtor, 305 #else 306 NULL, 307 #endif 308 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 309 310 vm_radix_zinit(); 311 } 312 313 void 314 vm_object_clear_flag(vm_object_t object, u_short bits) 315 { 316 317 VM_OBJECT_ASSERT_WLOCKED(object); 318 object->flags &= ~bits; 319 } 320 321 /* 322 * Sets the default memory attribute for the specified object. Pages 323 * that are allocated to this object are by default assigned this memory 324 * attribute. 325 * 326 * Presently, this function must be called before any pages are allocated 327 * to the object. In the future, this requirement may be relaxed for 328 * "default" and "swap" objects. 329 */ 330 int 331 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 332 { 333 334 VM_OBJECT_ASSERT_WLOCKED(object); 335 336 if (object->type == OBJT_DEAD) 337 return (KERN_INVALID_ARGUMENT); 338 if (!TAILQ_EMPTY(&object->memq)) 339 return (KERN_FAILURE); 340 341 object->memattr = memattr; 342 return (KERN_SUCCESS); 343 } 344 345 void 346 vm_object_pip_add(vm_object_t object, short i) 347 { 348 349 if (i > 0) 350 blockcount_acquire(&object->paging_in_progress, i); 351 } 352 353 void 354 vm_object_pip_wakeup(vm_object_t object) 355 { 356 357 vm_object_pip_wakeupn(object, 1); 358 } 359 360 void 361 vm_object_pip_wakeupn(vm_object_t object, short i) 362 { 363 364 if (i > 0) 365 blockcount_release(&object->paging_in_progress, i); 366 } 367 368 /* 369 * Atomically drop the object lock and wait for pip to drain. This protects 370 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 371 * on return. 372 */ 373 static void 374 vm_object_pip_sleep(vm_object_t object, const char *waitid) 375 { 376 377 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 378 waitid, PVM | PDROP); 379 } 380 381 void 382 vm_object_pip_wait(vm_object_t object, const char *waitid) 383 { 384 385 VM_OBJECT_ASSERT_WLOCKED(object); 386 387 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 388 PVM); 389 } 390 391 void 392 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 393 { 394 395 VM_OBJECT_ASSERT_UNLOCKED(object); 396 397 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 398 } 399 400 /* 401 * vm_object_allocate: 402 * 403 * Returns a new object with the given size. 404 */ 405 vm_object_t 406 vm_object_allocate(objtype_t type, vm_pindex_t size) 407 { 408 vm_object_t object; 409 u_short flags; 410 411 switch (type) { 412 case OBJT_DEAD: 413 panic("vm_object_allocate: can't create OBJT_DEAD"); 414 case OBJT_SWAP: 415 flags = OBJ_COLORED | OBJ_SWAP; 416 break; 417 case OBJT_DEVICE: 418 case OBJT_SG: 419 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 420 break; 421 case OBJT_MGTDEVICE: 422 flags = OBJ_FICTITIOUS; 423 break; 424 case OBJT_PHYS: 425 flags = OBJ_UNMANAGED; 426 break; 427 case OBJT_VNODE: 428 flags = 0; 429 break; 430 default: 431 panic("vm_object_allocate: type %d is undefined or dynamic", 432 type); 433 } 434 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 435 _vm_object_allocate(type, size, flags, object, NULL); 436 437 return (object); 438 } 439 440 vm_object_t 441 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags) 442 { 443 vm_object_t object; 444 445 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */); 446 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 447 _vm_object_allocate(dyntype, size, flags, object, NULL); 448 449 return (object); 450 } 451 452 /* 453 * vm_object_allocate_anon: 454 * 455 * Returns a new default object of the given size and marked as 456 * anonymous memory for special split/collapse handling. Color 457 * to be initialized by the caller. 458 */ 459 vm_object_t 460 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 461 struct ucred *cred, vm_size_t charge) 462 { 463 vm_object_t handle, object; 464 465 if (backing_object == NULL) 466 handle = NULL; 467 else if ((backing_object->flags & OBJ_ANON) != 0) 468 handle = backing_object->handle; 469 else 470 handle = backing_object; 471 object = uma_zalloc(obj_zone, M_WAITOK); 472 _vm_object_allocate(OBJT_SWAP, size, 473 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle); 474 object->cred = cred; 475 object->charge = cred != NULL ? charge : 0; 476 return (object); 477 } 478 479 static void 480 vm_object_reference_vnode(vm_object_t object) 481 { 482 u_int old; 483 484 /* 485 * vnode objects need the lock for the first reference 486 * to serialize with vnode_object_deallocate(). 487 */ 488 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 489 VM_OBJECT_RLOCK(object); 490 old = refcount_acquire(&object->ref_count); 491 if (object->type == OBJT_VNODE && old == 0) 492 vref(object->handle); 493 VM_OBJECT_RUNLOCK(object); 494 } 495 } 496 497 /* 498 * vm_object_reference: 499 * 500 * Acquires a reference to the given object. 501 */ 502 void 503 vm_object_reference(vm_object_t object) 504 { 505 506 if (object == NULL) 507 return; 508 509 if (object->type == OBJT_VNODE) 510 vm_object_reference_vnode(object); 511 else 512 refcount_acquire(&object->ref_count); 513 KASSERT((object->flags & OBJ_DEAD) == 0, 514 ("vm_object_reference: Referenced dead object.")); 515 } 516 517 /* 518 * vm_object_reference_locked: 519 * 520 * Gets another reference to the given object. 521 * 522 * The object must be locked. 523 */ 524 void 525 vm_object_reference_locked(vm_object_t object) 526 { 527 u_int old; 528 529 VM_OBJECT_ASSERT_LOCKED(object); 530 old = refcount_acquire(&object->ref_count); 531 if (object->type == OBJT_VNODE && old == 0) 532 vref(object->handle); 533 KASSERT((object->flags & OBJ_DEAD) == 0, 534 ("vm_object_reference: Referenced dead object.")); 535 } 536 537 /* 538 * Handle deallocating an object of type OBJT_VNODE. 539 */ 540 static void 541 vm_object_deallocate_vnode(vm_object_t object) 542 { 543 struct vnode *vp = (struct vnode *) object->handle; 544 bool last; 545 546 KASSERT(object->type == OBJT_VNODE, 547 ("vm_object_deallocate_vnode: not a vnode object")); 548 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 549 550 /* Object lock to protect handle lookup. */ 551 last = refcount_release(&object->ref_count); 552 VM_OBJECT_RUNLOCK(object); 553 554 if (!last) 555 return; 556 557 if (!umtx_shm_vnobj_persistent) 558 umtx_shm_object_terminated(object); 559 560 /* vrele may need the vnode lock. */ 561 vrele(vp); 562 } 563 564 /* 565 * We dropped a reference on an object and discovered that it had a 566 * single remaining shadow. This is a sibling of the reference we 567 * dropped. Attempt to collapse the sibling and backing object. 568 */ 569 static vm_object_t 570 vm_object_deallocate_anon(vm_object_t backing_object) 571 { 572 vm_object_t object; 573 574 /* Fetch the final shadow. */ 575 object = LIST_FIRST(&backing_object->shadow_head); 576 KASSERT(object != NULL && 577 atomic_load_int(&backing_object->shadow_count) == 1, 578 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 579 backing_object->ref_count, 580 atomic_load_int(&backing_object->shadow_count))); 581 KASSERT((object->flags & OBJ_ANON) != 0, 582 ("invalid shadow object %p", object)); 583 584 if (!VM_OBJECT_TRYWLOCK(object)) { 585 /* 586 * Prevent object from disappearing since we do not have a 587 * reference. 588 */ 589 vm_object_pip_add(object, 1); 590 VM_OBJECT_WUNLOCK(backing_object); 591 VM_OBJECT_WLOCK(object); 592 vm_object_pip_wakeup(object); 593 } else 594 VM_OBJECT_WUNLOCK(backing_object); 595 596 /* 597 * Check for a collapse/terminate race with the last reference holder. 598 */ 599 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 600 !refcount_acquire_if_not_zero(&object->ref_count)) { 601 VM_OBJECT_WUNLOCK(object); 602 return (NULL); 603 } 604 backing_object = object->backing_object; 605 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 606 vm_object_collapse(object); 607 VM_OBJECT_WUNLOCK(object); 608 609 return (object); 610 } 611 612 /* 613 * vm_object_deallocate: 614 * 615 * Release a reference to the specified object, 616 * gained either through a vm_object_allocate 617 * or a vm_object_reference call. When all references 618 * are gone, storage associated with this object 619 * may be relinquished. 620 * 621 * No object may be locked. 622 */ 623 void 624 vm_object_deallocate(vm_object_t object) 625 { 626 vm_object_t temp; 627 bool released; 628 629 while (object != NULL) { 630 /* 631 * If the reference count goes to 0 we start calling 632 * vm_object_terminate() on the object chain. A ref count 633 * of 1 may be a special case depending on the shadow count 634 * being 0 or 1. These cases require a write lock on the 635 * object. 636 */ 637 if ((object->flags & OBJ_ANON) == 0) 638 released = refcount_release_if_gt(&object->ref_count, 1); 639 else 640 released = refcount_release_if_gt(&object->ref_count, 2); 641 if (released) 642 return; 643 644 if (object->type == OBJT_VNODE) { 645 VM_OBJECT_RLOCK(object); 646 if (object->type == OBJT_VNODE) { 647 vm_object_deallocate_vnode(object); 648 return; 649 } 650 VM_OBJECT_RUNLOCK(object); 651 } 652 653 VM_OBJECT_WLOCK(object); 654 KASSERT(object->ref_count > 0, 655 ("vm_object_deallocate: object deallocated too many times: %d", 656 object->type)); 657 658 /* 659 * If this is not the final reference to an anonymous 660 * object we may need to collapse the shadow chain. 661 */ 662 if (!refcount_release(&object->ref_count)) { 663 if (object->ref_count > 1 || 664 atomic_load_int(&object->shadow_count) == 0) { 665 if ((object->flags & OBJ_ANON) != 0 && 666 object->ref_count == 1) 667 vm_object_set_flag(object, 668 OBJ_ONEMAPPING); 669 VM_OBJECT_WUNLOCK(object); 670 return; 671 } 672 673 /* Handle collapsing last ref on anonymous objects. */ 674 object = vm_object_deallocate_anon(object); 675 continue; 676 } 677 678 /* 679 * Handle the final reference to an object. We restart 680 * the loop with the backing object to avoid recursion. 681 */ 682 umtx_shm_object_terminated(object); 683 temp = object->backing_object; 684 if (temp != NULL) { 685 KASSERT(object->type == OBJT_SWAP, 686 ("shadowed tmpfs v_object 2 %p", object)); 687 vm_object_backing_remove(object); 688 } 689 690 KASSERT((object->flags & OBJ_DEAD) == 0, 691 ("vm_object_deallocate: Terminating dead object.")); 692 vm_object_set_flag(object, OBJ_DEAD); 693 vm_object_terminate(object); 694 object = temp; 695 } 696 } 697 698 void 699 vm_object_destroy(vm_object_t object) 700 { 701 uma_zfree(obj_zone, object); 702 } 703 704 static void 705 vm_object_sub_shadow(vm_object_t object) 706 { 707 KASSERT(object->shadow_count >= 1, 708 ("object %p sub_shadow count zero", object)); 709 atomic_subtract_int(&object->shadow_count, 1); 710 } 711 712 static void 713 vm_object_backing_remove_locked(vm_object_t object) 714 { 715 vm_object_t backing_object; 716 717 backing_object = object->backing_object; 718 VM_OBJECT_ASSERT_WLOCKED(object); 719 VM_OBJECT_ASSERT_WLOCKED(backing_object); 720 721 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 722 ("vm_object_backing_remove: Removing collapsing object.")); 723 724 vm_object_sub_shadow(backing_object); 725 if ((object->flags & OBJ_SHADOWLIST) != 0) { 726 LIST_REMOVE(object, shadow_list); 727 vm_object_clear_flag(object, OBJ_SHADOWLIST); 728 } 729 object->backing_object = NULL; 730 } 731 732 static void 733 vm_object_backing_remove(vm_object_t object) 734 { 735 vm_object_t backing_object; 736 737 VM_OBJECT_ASSERT_WLOCKED(object); 738 739 backing_object = object->backing_object; 740 if ((object->flags & OBJ_SHADOWLIST) != 0) { 741 VM_OBJECT_WLOCK(backing_object); 742 vm_object_backing_remove_locked(object); 743 VM_OBJECT_WUNLOCK(backing_object); 744 } else { 745 object->backing_object = NULL; 746 vm_object_sub_shadow(backing_object); 747 } 748 } 749 750 static void 751 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 752 { 753 754 VM_OBJECT_ASSERT_WLOCKED(object); 755 756 atomic_add_int(&backing_object->shadow_count, 1); 757 if ((backing_object->flags & OBJ_ANON) != 0) { 758 VM_OBJECT_ASSERT_WLOCKED(backing_object); 759 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 760 shadow_list); 761 vm_object_set_flag(object, OBJ_SHADOWLIST); 762 } 763 object->backing_object = backing_object; 764 } 765 766 static void 767 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 768 { 769 770 VM_OBJECT_ASSERT_WLOCKED(object); 771 772 if ((backing_object->flags & OBJ_ANON) != 0) { 773 VM_OBJECT_WLOCK(backing_object); 774 vm_object_backing_insert_locked(object, backing_object); 775 VM_OBJECT_WUNLOCK(backing_object); 776 } else { 777 object->backing_object = backing_object; 778 atomic_add_int(&backing_object->shadow_count, 1); 779 } 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 atomic_add_int(&backing_object->shadow_count, 1); 803 object->backing_object = backing_object; 804 } 805 } 806 807 /* 808 * Transfer a backing reference from backing_object to object. 809 */ 810 static void 811 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 812 { 813 vm_object_t new_backing_object; 814 815 /* 816 * Note that the reference to backing_object->backing_object 817 * moves from within backing_object to within object. 818 */ 819 vm_object_backing_remove_locked(object); 820 new_backing_object = backing_object->backing_object; 821 if (new_backing_object == NULL) 822 return; 823 if ((new_backing_object->flags & OBJ_ANON) != 0) { 824 VM_OBJECT_WLOCK(new_backing_object); 825 vm_object_backing_remove_locked(backing_object); 826 vm_object_backing_insert_locked(object, new_backing_object); 827 VM_OBJECT_WUNLOCK(new_backing_object); 828 } else { 829 /* 830 * shadow_count for new_backing_object is left 831 * unchanged, its reference provided by backing_object 832 * is replaced by object. 833 */ 834 object->backing_object = new_backing_object; 835 backing_object->backing_object = NULL; 836 } 837 } 838 839 /* 840 * Wait for a concurrent collapse to settle. 841 */ 842 static void 843 vm_object_collapse_wait(vm_object_t object) 844 { 845 846 VM_OBJECT_ASSERT_WLOCKED(object); 847 848 while ((object->flags & OBJ_COLLAPSING) != 0) { 849 vm_object_pip_wait(object, "vmcolwait"); 850 counter_u64_add(object_collapse_waits, 1); 851 } 852 } 853 854 /* 855 * Waits for a backing object to clear a pending collapse and returns 856 * it locked if it is an ANON object. 857 */ 858 static vm_object_t 859 vm_object_backing_collapse_wait(vm_object_t object) 860 { 861 vm_object_t backing_object; 862 863 VM_OBJECT_ASSERT_WLOCKED(object); 864 865 for (;;) { 866 backing_object = object->backing_object; 867 if (backing_object == NULL || 868 (backing_object->flags & OBJ_ANON) == 0) 869 return (NULL); 870 VM_OBJECT_WLOCK(backing_object); 871 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 872 break; 873 VM_OBJECT_WUNLOCK(object); 874 vm_object_pip_sleep(backing_object, "vmbckwait"); 875 counter_u64_add(object_collapse_waits, 1); 876 VM_OBJECT_WLOCK(object); 877 } 878 return (backing_object); 879 } 880 881 /* 882 * vm_object_terminate_single_page removes a pageable page from the object, 883 * and removes it from the paging queues and frees it, if it is not wired. 884 * It is invoked via callback from vm_object_terminate_pages. 885 */ 886 static void 887 vm_object_terminate_single_page(vm_page_t p, void *objectv) 888 { 889 vm_object_t object __diagused = objectv; 890 891 vm_page_assert_unbusied(p); 892 KASSERT(p->object == object && 893 (p->ref_count & VPRC_OBJREF) != 0, 894 ("%s: page %p is inconsistent", __func__, p)); 895 p->object = NULL; 896 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 897 VM_CNT_INC(v_pfree); 898 vm_page_free(p); 899 } 900 } 901 902 /* 903 * vm_object_terminate_pages removes any remaining pageable pages 904 * from the object and resets the object to an empty state. 905 */ 906 static void 907 vm_object_terminate_pages(vm_object_t object) 908 { 909 VM_OBJECT_ASSERT_WLOCKED(object); 910 911 /* 912 * If the object contained any pages, then reset it to an empty state. 913 * Rather than incrementally removing each page from the object, the 914 * page and object are reset to any empty state. 915 */ 916 if (object->resident_page_count == 0) 917 return; 918 919 vm_radix_reclaim_callback(&object->rtree, 920 vm_object_terminate_single_page, object); 921 TAILQ_INIT(&object->memq); 922 object->resident_page_count = 0; 923 if (object->type == OBJT_VNODE) 924 vdrop(object->handle); 925 } 926 927 /* 928 * vm_object_terminate actually destroys the specified object, freeing 929 * up all previously used resources. 930 * 931 * The object must be locked. 932 * This routine may block. 933 */ 934 void 935 vm_object_terminate(vm_object_t object) 936 { 937 938 VM_OBJECT_ASSERT_WLOCKED(object); 939 KASSERT((object->flags & OBJ_DEAD) != 0, 940 ("terminating non-dead obj %p", object)); 941 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 942 ("terminating collapsing obj %p", object)); 943 KASSERT(object->backing_object == NULL, 944 ("terminating shadow obj %p", object)); 945 946 /* 947 * Wait for the pageout daemon and other current users to be 948 * done with the object. Note that new paging_in_progress 949 * users can come after this wait, but they must check 950 * OBJ_DEAD flag set (without unlocking the object), and avoid 951 * the object being terminated. 952 */ 953 vm_object_pip_wait(object, "objtrm"); 954 955 KASSERT(object->ref_count == 0, 956 ("vm_object_terminate: object with references, ref_count=%d", 957 object->ref_count)); 958 959 if ((object->flags & OBJ_PG_DTOR) == 0) 960 vm_object_terminate_pages(object); 961 962 #if VM_NRESERVLEVEL > 0 963 if (__predict_false(!LIST_EMPTY(&object->rvq))) 964 vm_reserv_break_all(object); 965 #endif 966 967 KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0, 968 ("%s: non-swap obj %p has cred", __func__, object)); 969 970 /* 971 * Let the pager know object is dead. 972 */ 973 vm_pager_deallocate(object); 974 VM_OBJECT_WUNLOCK(object); 975 976 vm_object_destroy(object); 977 } 978 979 /* 980 * Make the page read-only so that we can clear the object flags. However, if 981 * this is a nosync mmap then the object is likely to stay dirty so do not 982 * mess with the page and do not clear the object flags. Returns TRUE if the 983 * page should be flushed, and FALSE otherwise. 984 */ 985 static boolean_t 986 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) 987 { 988 989 vm_page_assert_busied(p); 990 991 /* 992 * If we have been asked to skip nosync pages and this is a 993 * nosync page, skip it. Note that the object flags were not 994 * cleared in this case so we do not have to set them. 995 */ 996 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { 997 *allclean = FALSE; 998 return (FALSE); 999 } else { 1000 pmap_remove_write(p); 1001 return (p->dirty != 0); 1002 } 1003 } 1004 1005 /* 1006 * vm_object_page_clean 1007 * 1008 * Clean all dirty pages in the specified range of object. Leaves page 1009 * on whatever queue it is currently on. If NOSYNC is set then do not 1010 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1011 * leaving the object dirty. 1012 * 1013 * For swap objects backing tmpfs regular files, do not flush anything, 1014 * but remove write protection on the mapped pages to update mtime through 1015 * mmaped writes. 1016 * 1017 * When stuffing pages asynchronously, allow clustering. XXX we need a 1018 * synchronous clustering mode implementation. 1019 * 1020 * Odd semantics: if start == end, we clean everything. 1021 * 1022 * The object must be locked. 1023 * 1024 * Returns FALSE if some page from the range was not written, as 1025 * reported by the pager, and TRUE otherwise. 1026 */ 1027 boolean_t 1028 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1029 int flags) 1030 { 1031 vm_page_t np, p; 1032 vm_pindex_t pi, tend, tstart; 1033 int curgeneration, n, pagerflags; 1034 boolean_t eio, res, allclean; 1035 1036 VM_OBJECT_ASSERT_WLOCKED(object); 1037 1038 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1039 return (TRUE); 1040 1041 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1042 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1043 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1044 1045 tstart = OFF_TO_IDX(start); 1046 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1047 allclean = tstart == 0 && tend >= object->size; 1048 res = TRUE; 1049 1050 rescan: 1051 curgeneration = object->generation; 1052 1053 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 1054 pi = p->pindex; 1055 if (pi >= tend) 1056 break; 1057 np = TAILQ_NEXT(p, listq); 1058 if (vm_page_none_valid(p)) 1059 continue; 1060 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { 1061 if (object->generation != curgeneration && 1062 (flags & OBJPC_SYNC) != 0) 1063 goto rescan; 1064 np = vm_page_find_least(object, pi); 1065 continue; 1066 } 1067 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1068 vm_page_xunbusy(p); 1069 continue; 1070 } 1071 if (object->type == OBJT_VNODE) { 1072 n = vm_object_page_collect_flush(object, p, pagerflags, 1073 flags, &allclean, &eio); 1074 if (eio) { 1075 res = FALSE; 1076 allclean = FALSE; 1077 } 1078 if (object->generation != curgeneration && 1079 (flags & OBJPC_SYNC) != 0) 1080 goto rescan; 1081 1082 /* 1083 * If the VOP_PUTPAGES() did a truncated write, so 1084 * that even the first page of the run is not fully 1085 * written, vm_pageout_flush() returns 0 as the run 1086 * length. Since the condition that caused truncated 1087 * write may be permanent, e.g. exhausted free space, 1088 * accepting n == 0 would cause an infinite loop. 1089 * 1090 * Forwarding the iterator leaves the unwritten page 1091 * behind, but there is not much we can do there if 1092 * filesystem refuses to write it. 1093 */ 1094 if (n == 0) { 1095 n = 1; 1096 allclean = FALSE; 1097 } 1098 } else { 1099 n = 1; 1100 vm_page_xunbusy(p); 1101 } 1102 np = vm_page_find_least(object, pi + n); 1103 } 1104 #if 0 1105 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1106 #endif 1107 1108 /* 1109 * Leave updating cleangeneration for tmpfs objects to tmpfs 1110 * scan. It needs to update mtime, which happens for other 1111 * filesystems during page writeouts. 1112 */ 1113 if (allclean && object->type == OBJT_VNODE) 1114 object->cleangeneration = curgeneration; 1115 return (res); 1116 } 1117 1118 static int 1119 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1120 int flags, boolean_t *allclean, boolean_t *eio) 1121 { 1122 vm_page_t ma[2 * vm_pageout_page_count - 1], tp; 1123 int base, count, runlen; 1124 1125 vm_page_lock_assert(p, MA_NOTOWNED); 1126 vm_page_assert_xbusied(p); 1127 VM_OBJECT_ASSERT_WLOCKED(object); 1128 base = nitems(ma) / 2; 1129 ma[base] = p; 1130 for (count = 1, tp = p; count < vm_pageout_page_count; count++) { 1131 tp = vm_page_next(tp); 1132 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1133 break; 1134 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1135 vm_page_xunbusy(tp); 1136 break; 1137 } 1138 ma[base + count] = tp; 1139 } 1140 1141 for (tp = p; count < vm_pageout_page_count; count++) { 1142 tp = vm_page_prev(tp); 1143 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1144 break; 1145 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1146 vm_page_xunbusy(tp); 1147 break; 1148 } 1149 ma[--base] = tp; 1150 } 1151 1152 vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base, 1153 &runlen, eio); 1154 return (runlen); 1155 } 1156 1157 /* 1158 * Note that there is absolutely no sense in writing out 1159 * anonymous objects, so we track down the vnode object 1160 * to write out. 1161 * We invalidate (remove) all pages from the address space 1162 * for semantic correctness. 1163 * 1164 * If the backing object is a device object with unmanaged pages, then any 1165 * mappings to the specified range of pages must be removed before this 1166 * function is called. 1167 * 1168 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1169 * may start out with a NULL object. 1170 */ 1171 boolean_t 1172 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1173 boolean_t syncio, boolean_t invalidate) 1174 { 1175 vm_object_t backing_object; 1176 struct vnode *vp; 1177 struct mount *mp; 1178 int error, flags, fsync_after; 1179 boolean_t res; 1180 1181 if (object == NULL) 1182 return (TRUE); 1183 res = TRUE; 1184 error = 0; 1185 VM_OBJECT_WLOCK(object); 1186 while ((backing_object = object->backing_object) != NULL) { 1187 VM_OBJECT_WLOCK(backing_object); 1188 offset += object->backing_object_offset; 1189 VM_OBJECT_WUNLOCK(object); 1190 object = backing_object; 1191 if (object->size < OFF_TO_IDX(offset + size)) 1192 size = IDX_TO_OFF(object->size) - offset; 1193 } 1194 /* 1195 * Flush pages if writing is allowed, invalidate them 1196 * if invalidation requested. Pages undergoing I/O 1197 * will be ignored by vm_object_page_remove(). 1198 * 1199 * We cannot lock the vnode and then wait for paging 1200 * to complete without deadlocking against vm_fault. 1201 * Instead we simply call vm_object_page_remove() and 1202 * allow it to block internally on a page-by-page 1203 * basis when it encounters pages undergoing async 1204 * I/O. 1205 */ 1206 if (object->type == OBJT_VNODE && 1207 vm_object_mightbedirty(object) != 0 && 1208 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1209 VM_OBJECT_WUNLOCK(object); 1210 (void)vn_start_write(vp, &mp, V_WAIT); 1211 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1212 if (syncio && !invalidate && offset == 0 && 1213 atop(size) == object->size) { 1214 /* 1215 * If syncing the whole mapping of the file, 1216 * it is faster to schedule all the writes in 1217 * async mode, also allowing the clustering, 1218 * and then wait for i/o to complete. 1219 */ 1220 flags = 0; 1221 fsync_after = TRUE; 1222 } else { 1223 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1224 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1225 fsync_after = FALSE; 1226 } 1227 VM_OBJECT_WLOCK(object); 1228 res = vm_object_page_clean(object, offset, offset + size, 1229 flags); 1230 VM_OBJECT_WUNLOCK(object); 1231 if (fsync_after) { 1232 for (;;) { 1233 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1234 if (error != ERELOOKUP) 1235 break; 1236 1237 /* 1238 * Allow SU/bufdaemon to handle more 1239 * dependencies in the meantime. 1240 */ 1241 VOP_UNLOCK(vp); 1242 vn_finished_write(mp); 1243 1244 (void)vn_start_write(vp, &mp, V_WAIT); 1245 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1246 } 1247 } 1248 VOP_UNLOCK(vp); 1249 vn_finished_write(mp); 1250 if (error != 0) 1251 res = FALSE; 1252 VM_OBJECT_WLOCK(object); 1253 } 1254 if ((object->type == OBJT_VNODE || 1255 object->type == OBJT_DEVICE) && invalidate) { 1256 if (object->type == OBJT_DEVICE) 1257 /* 1258 * The option OBJPR_NOTMAPPED must be passed here 1259 * because vm_object_page_remove() cannot remove 1260 * unmanaged mappings. 1261 */ 1262 flags = OBJPR_NOTMAPPED; 1263 else if (old_msync) 1264 flags = 0; 1265 else 1266 flags = OBJPR_CLEANONLY; 1267 vm_object_page_remove(object, OFF_TO_IDX(offset), 1268 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1269 } 1270 VM_OBJECT_WUNLOCK(object); 1271 return (res); 1272 } 1273 1274 /* 1275 * Determine whether the given advice can be applied to the object. Advice is 1276 * not applied to unmanaged pages since they never belong to page queues, and 1277 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1278 * have been mapped at most once. 1279 */ 1280 static bool 1281 vm_object_advice_applies(vm_object_t object, int advice) 1282 { 1283 1284 if ((object->flags & OBJ_UNMANAGED) != 0) 1285 return (false); 1286 if (advice != MADV_FREE) 1287 return (true); 1288 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1289 (OBJ_ONEMAPPING | OBJ_ANON)); 1290 } 1291 1292 static void 1293 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1294 vm_size_t size) 1295 { 1296 1297 if (advice == MADV_FREE) 1298 vm_pager_freespace(object, pindex, size); 1299 } 1300 1301 /* 1302 * vm_object_madvise: 1303 * 1304 * Implements the madvise function at the object/page level. 1305 * 1306 * MADV_WILLNEED (any object) 1307 * 1308 * Activate the specified pages if they are resident. 1309 * 1310 * MADV_DONTNEED (any object) 1311 * 1312 * Deactivate the specified pages if they are resident. 1313 * 1314 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only) 1315 * 1316 * Deactivate and clean the specified pages if they are 1317 * resident. This permits the process to reuse the pages 1318 * without faulting or the kernel to reclaim the pages 1319 * without I/O. 1320 */ 1321 void 1322 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1323 int advice) 1324 { 1325 vm_pindex_t tpindex; 1326 vm_object_t backing_object, tobject; 1327 vm_page_t m, tm; 1328 1329 if (object == NULL) 1330 return; 1331 1332 relookup: 1333 VM_OBJECT_WLOCK(object); 1334 if (!vm_object_advice_applies(object, advice)) { 1335 VM_OBJECT_WUNLOCK(object); 1336 return; 1337 } 1338 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1339 tobject = object; 1340 1341 /* 1342 * If the next page isn't resident in the top-level object, we 1343 * need to search the shadow chain. When applying MADV_FREE, we 1344 * take care to release any swap space used to store 1345 * non-resident pages. 1346 */ 1347 if (m == NULL || pindex < m->pindex) { 1348 /* 1349 * Optimize a common case: if the top-level object has 1350 * no backing object, we can skip over the non-resident 1351 * range in constant time. 1352 */ 1353 if (object->backing_object == NULL) { 1354 tpindex = (m != NULL && m->pindex < end) ? 1355 m->pindex : end; 1356 vm_object_madvise_freespace(object, advice, 1357 pindex, tpindex - pindex); 1358 if ((pindex = tpindex) == end) 1359 break; 1360 goto next_page; 1361 } 1362 1363 tpindex = pindex; 1364 do { 1365 vm_object_madvise_freespace(tobject, advice, 1366 tpindex, 1); 1367 /* 1368 * Prepare to search the next object in the 1369 * chain. 1370 */ 1371 backing_object = tobject->backing_object; 1372 if (backing_object == NULL) 1373 goto next_pindex; 1374 VM_OBJECT_WLOCK(backing_object); 1375 tpindex += 1376 OFF_TO_IDX(tobject->backing_object_offset); 1377 if (tobject != object) 1378 VM_OBJECT_WUNLOCK(tobject); 1379 tobject = backing_object; 1380 if (!vm_object_advice_applies(tobject, advice)) 1381 goto next_pindex; 1382 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1383 NULL); 1384 } else { 1385 next_page: 1386 tm = m; 1387 m = TAILQ_NEXT(m, listq); 1388 } 1389 1390 /* 1391 * If the page is not in a normal state, skip it. The page 1392 * can not be invalidated while the object lock is held. 1393 */ 1394 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1395 goto next_pindex; 1396 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1397 ("vm_object_madvise: page %p is fictitious", tm)); 1398 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1399 ("vm_object_madvise: page %p is not managed", tm)); 1400 if (vm_page_tryxbusy(tm) == 0) { 1401 if (object != tobject) 1402 VM_OBJECT_WUNLOCK(object); 1403 if (advice == MADV_WILLNEED) { 1404 /* 1405 * Reference the page before unlocking and 1406 * sleeping so that the page daemon is less 1407 * likely to reclaim it. 1408 */ 1409 vm_page_aflag_set(tm, PGA_REFERENCED); 1410 } 1411 if (!vm_page_busy_sleep(tm, "madvpo", 0)) 1412 VM_OBJECT_WUNLOCK(tobject); 1413 goto relookup; 1414 } 1415 vm_page_advise(tm, advice); 1416 vm_page_xunbusy(tm); 1417 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1418 next_pindex: 1419 if (tobject != object) 1420 VM_OBJECT_WUNLOCK(tobject); 1421 } 1422 VM_OBJECT_WUNLOCK(object); 1423 } 1424 1425 /* 1426 * vm_object_shadow: 1427 * 1428 * Create a new object which is backed by the 1429 * specified existing object range. The source 1430 * object reference is deallocated. 1431 * 1432 * The new object and offset into that object 1433 * are returned in the source parameters. 1434 */ 1435 void 1436 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1437 struct ucred *cred, bool shared) 1438 { 1439 vm_object_t source; 1440 vm_object_t result; 1441 1442 source = *object; 1443 1444 /* 1445 * Don't create the new object if the old object isn't shared. 1446 * 1447 * If we hold the only reference we can guarantee that it won't 1448 * increase while we have the map locked. Otherwise the race is 1449 * harmless and we will end up with an extra shadow object that 1450 * will be collapsed later. 1451 */ 1452 if (source != NULL && source->ref_count == 1 && 1453 (source->flags & OBJ_ANON) != 0) 1454 return; 1455 1456 /* 1457 * Allocate a new object with the given length. 1458 */ 1459 result = vm_object_allocate_anon(atop(length), source, cred, length); 1460 1461 /* 1462 * Store the offset into the source object, and fix up the offset into 1463 * the new object. 1464 */ 1465 result->backing_object_offset = *offset; 1466 1467 if (shared || source != NULL) { 1468 VM_OBJECT_WLOCK(result); 1469 1470 /* 1471 * The new object shadows the source object, adding a 1472 * reference to it. Our caller changes his reference 1473 * to point to the new object, removing a reference to 1474 * the source object. Net result: no change of 1475 * reference count, unless the caller needs to add one 1476 * more reference due to forking a shared map entry. 1477 */ 1478 if (shared) { 1479 vm_object_reference_locked(result); 1480 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1481 } 1482 1483 /* 1484 * Try to optimize the result object's page color when 1485 * shadowing in order to maintain page coloring 1486 * consistency in the combined shadowed object. 1487 */ 1488 if (source != NULL) { 1489 vm_object_backing_insert(result, source); 1490 result->domain = source->domain; 1491 #if VM_NRESERVLEVEL > 0 1492 vm_object_set_flag(result, 1493 (source->flags & OBJ_COLORED)); 1494 result->pg_color = (source->pg_color + 1495 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1496 1)) - 1); 1497 #endif 1498 } 1499 VM_OBJECT_WUNLOCK(result); 1500 } 1501 1502 /* 1503 * Return the new things 1504 */ 1505 *offset = 0; 1506 *object = result; 1507 } 1508 1509 /* 1510 * vm_object_split: 1511 * 1512 * Split the pages in a map entry into a new object. This affords 1513 * easier removal of unused pages, and keeps object inheritance from 1514 * being a negative impact on memory usage. 1515 */ 1516 void 1517 vm_object_split(vm_map_entry_t entry) 1518 { 1519 vm_page_t m, m_next; 1520 vm_object_t orig_object, new_object, backing_object; 1521 vm_pindex_t idx, offidxstart; 1522 vm_size_t size; 1523 1524 orig_object = entry->object.vm_object; 1525 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1526 ("vm_object_split: Splitting object with multiple mappings.")); 1527 if ((orig_object->flags & OBJ_ANON) == 0) 1528 return; 1529 if (orig_object->ref_count <= 1) 1530 return; 1531 VM_OBJECT_WUNLOCK(orig_object); 1532 1533 offidxstart = OFF_TO_IDX(entry->offset); 1534 size = atop(entry->end - entry->start); 1535 1536 new_object = vm_object_allocate_anon(size, orig_object, 1537 orig_object->cred, ptoa(size)); 1538 1539 /* 1540 * We must wait for the orig_object to complete any in-progress 1541 * collapse so that the swap blocks are stable below. The 1542 * additional reference on backing_object by new object will 1543 * prevent further collapse operations until split completes. 1544 */ 1545 VM_OBJECT_WLOCK(orig_object); 1546 vm_object_collapse_wait(orig_object); 1547 1548 /* 1549 * At this point, the new object is still private, so the order in 1550 * which the original and new objects are locked does not matter. 1551 */ 1552 VM_OBJECT_WLOCK(new_object); 1553 new_object->domain = orig_object->domain; 1554 backing_object = orig_object->backing_object; 1555 if (backing_object != NULL) { 1556 vm_object_backing_insert_ref(new_object, backing_object); 1557 new_object->backing_object_offset = 1558 orig_object->backing_object_offset + entry->offset; 1559 } 1560 if (orig_object->cred != NULL) { 1561 crhold(orig_object->cred); 1562 KASSERT(orig_object->charge >= ptoa(size), 1563 ("orig_object->charge < 0")); 1564 orig_object->charge -= ptoa(size); 1565 } 1566 1567 /* 1568 * Mark the split operation so that swap_pager_getpages() knows 1569 * that the object is in transition. 1570 */ 1571 vm_object_set_flag(orig_object, OBJ_SPLIT); 1572 #ifdef INVARIANTS 1573 idx = 0; 1574 #endif 1575 retry: 1576 m = vm_page_find_least(orig_object, offidxstart); 1577 KASSERT(m == NULL || idx <= m->pindex - offidxstart, 1578 ("%s: object %p was repopulated", __func__, orig_object)); 1579 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1580 m = m_next) { 1581 m_next = TAILQ_NEXT(m, listq); 1582 1583 /* 1584 * We must wait for pending I/O to complete before we can 1585 * rename the page. 1586 * 1587 * We do not have to VM_PROT_NONE the page as mappings should 1588 * not be changed by this operation. 1589 */ 1590 if (vm_page_tryxbusy(m) == 0) { 1591 VM_OBJECT_WUNLOCK(new_object); 1592 if (vm_page_busy_sleep(m, "spltwt", 0)) 1593 VM_OBJECT_WLOCK(orig_object); 1594 VM_OBJECT_WLOCK(new_object); 1595 goto retry; 1596 } 1597 1598 /* 1599 * The page was left invalid. Likely placed there by 1600 * an incomplete fault. Just remove and ignore. 1601 */ 1602 if (vm_page_none_valid(m)) { 1603 if (vm_page_remove(m)) 1604 vm_page_free(m); 1605 continue; 1606 } 1607 1608 /* vm_page_rename() will dirty the page. */ 1609 if (vm_page_rename(m, new_object, idx)) { 1610 vm_page_xunbusy(m); 1611 VM_OBJECT_WUNLOCK(new_object); 1612 VM_OBJECT_WUNLOCK(orig_object); 1613 vm_radix_wait(); 1614 VM_OBJECT_WLOCK(orig_object); 1615 VM_OBJECT_WLOCK(new_object); 1616 goto retry; 1617 } 1618 1619 #if VM_NRESERVLEVEL > 0 1620 /* 1621 * If some of the reservation's allocated pages remain with 1622 * the original object, then transferring the reservation to 1623 * the new object is neither particularly beneficial nor 1624 * particularly harmful as compared to leaving the reservation 1625 * with the original object. If, however, all of the 1626 * reservation's allocated pages are transferred to the new 1627 * object, then transferring the reservation is typically 1628 * beneficial. Determining which of these two cases applies 1629 * would be more costly than unconditionally renaming the 1630 * reservation. 1631 */ 1632 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1633 #endif 1634 } 1635 1636 /* 1637 * swap_pager_copy() can sleep, in which case the orig_object's 1638 * and new_object's locks are released and reacquired. 1639 */ 1640 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1641 1642 TAILQ_FOREACH(m, &new_object->memq, listq) 1643 vm_page_xunbusy(m); 1644 1645 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1646 VM_OBJECT_WUNLOCK(orig_object); 1647 VM_OBJECT_WUNLOCK(new_object); 1648 entry->object.vm_object = new_object; 1649 entry->offset = 0LL; 1650 vm_object_deallocate(orig_object); 1651 VM_OBJECT_WLOCK(new_object); 1652 } 1653 1654 static vm_page_t 1655 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p) 1656 { 1657 vm_object_t backing_object; 1658 1659 VM_OBJECT_ASSERT_WLOCKED(object); 1660 backing_object = object->backing_object; 1661 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1662 1663 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1664 ("invalid ownership %p %p %p", p, object, backing_object)); 1665 /* The page is only NULL when rename fails. */ 1666 if (p == NULL) { 1667 VM_OBJECT_WUNLOCK(object); 1668 VM_OBJECT_WUNLOCK(backing_object); 1669 vm_radix_wait(); 1670 VM_OBJECT_WLOCK(object); 1671 } else if (p->object == object) { 1672 VM_OBJECT_WUNLOCK(backing_object); 1673 if (vm_page_busy_sleep(p, "vmocol", 0)) 1674 VM_OBJECT_WLOCK(object); 1675 } else { 1676 VM_OBJECT_WUNLOCK(object); 1677 if (!vm_page_busy_sleep(p, "vmocol", 0)) 1678 VM_OBJECT_WUNLOCK(backing_object); 1679 VM_OBJECT_WLOCK(object); 1680 } 1681 VM_OBJECT_WLOCK(backing_object); 1682 return (TAILQ_FIRST(&backing_object->memq)); 1683 } 1684 1685 static bool 1686 vm_object_scan_all_shadowed(vm_object_t object) 1687 { 1688 vm_object_t backing_object; 1689 vm_page_t p, pp; 1690 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1691 1692 VM_OBJECT_ASSERT_WLOCKED(object); 1693 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1694 1695 backing_object = object->backing_object; 1696 1697 if ((backing_object->flags & OBJ_ANON) == 0) 1698 return (false); 1699 1700 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1701 p = vm_page_find_least(backing_object, pi); 1702 ps = swap_pager_find_least(backing_object, pi); 1703 1704 /* 1705 * Only check pages inside the parent object's range and 1706 * inside the parent object's mapping of the backing object. 1707 */ 1708 for (;; pi++) { 1709 if (p != NULL && p->pindex < pi) 1710 p = TAILQ_NEXT(p, listq); 1711 if (ps < pi) 1712 ps = swap_pager_find_least(backing_object, pi); 1713 if (p == NULL && ps >= backing_object->size) 1714 break; 1715 else if (p == NULL) 1716 pi = ps; 1717 else 1718 pi = MIN(p->pindex, ps); 1719 1720 new_pindex = pi - backing_offset_index; 1721 if (new_pindex >= object->size) 1722 break; 1723 1724 if (p != NULL) { 1725 /* 1726 * If the backing object page is busy a 1727 * grandparent or older page may still be 1728 * undergoing CoW. It is not safe to collapse 1729 * the backing object until it is quiesced. 1730 */ 1731 if (vm_page_tryxbusy(p) == 0) 1732 return (false); 1733 1734 /* 1735 * We raced with the fault handler that left 1736 * newly allocated invalid page on the object 1737 * queue and retried. 1738 */ 1739 if (!vm_page_all_valid(p)) 1740 goto unbusy_ret; 1741 } 1742 1743 /* 1744 * See if the parent has the page or if the parent's object 1745 * pager has the page. If the parent has the page but the page 1746 * is not valid, the parent's object pager must have the page. 1747 * 1748 * If this fails, the parent does not completely shadow the 1749 * object and we might as well give up now. 1750 */ 1751 pp = vm_page_lookup(object, new_pindex); 1752 1753 /* 1754 * The valid check here is stable due to object lock 1755 * being required to clear valid and initiate paging. 1756 * Busy of p disallows fault handler to validate pp. 1757 */ 1758 if ((pp == NULL || vm_page_none_valid(pp)) && 1759 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1760 goto unbusy_ret; 1761 if (p != NULL) 1762 vm_page_xunbusy(p); 1763 } 1764 return (true); 1765 1766 unbusy_ret: 1767 if (p != NULL) 1768 vm_page_xunbusy(p); 1769 return (false); 1770 } 1771 1772 static void 1773 vm_object_collapse_scan(vm_object_t object) 1774 { 1775 vm_object_t backing_object; 1776 vm_page_t next, p, pp; 1777 vm_pindex_t backing_offset_index, new_pindex; 1778 1779 VM_OBJECT_ASSERT_WLOCKED(object); 1780 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1781 1782 backing_object = object->backing_object; 1783 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1784 1785 /* 1786 * Our scan 1787 */ 1788 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1789 next = TAILQ_NEXT(p, listq); 1790 new_pindex = p->pindex - backing_offset_index; 1791 1792 /* 1793 * Check for busy page 1794 */ 1795 if (vm_page_tryxbusy(p) == 0) { 1796 next = vm_object_collapse_scan_wait(object, p); 1797 continue; 1798 } 1799 1800 KASSERT(object->backing_object == backing_object, 1801 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1802 object->backing_object, backing_object)); 1803 KASSERT(p->object == backing_object, 1804 ("vm_object_collapse_scan: object mismatch %p != %p", 1805 p->object, backing_object)); 1806 1807 if (p->pindex < backing_offset_index || 1808 new_pindex >= object->size) { 1809 vm_pager_freespace(backing_object, p->pindex, 1); 1810 1811 KASSERT(!pmap_page_is_mapped(p), 1812 ("freeing mapped page %p", p)); 1813 if (vm_page_remove(p)) 1814 vm_page_free(p); 1815 continue; 1816 } 1817 1818 if (!vm_page_all_valid(p)) { 1819 KASSERT(!pmap_page_is_mapped(p), 1820 ("freeing mapped page %p", p)); 1821 if (vm_page_remove(p)) 1822 vm_page_free(p); 1823 continue; 1824 } 1825 1826 pp = vm_page_lookup(object, new_pindex); 1827 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1828 vm_page_xunbusy(p); 1829 /* 1830 * The page in the parent is busy and possibly not 1831 * (yet) valid. Until its state is finalized by the 1832 * busy bit owner, we can't tell whether it shadows the 1833 * original page. 1834 */ 1835 next = vm_object_collapse_scan_wait(object, pp); 1836 continue; 1837 } 1838 1839 if (pp != NULL && vm_page_none_valid(pp)) { 1840 /* 1841 * The page was invalid in the parent. Likely placed 1842 * there by an incomplete fault. Just remove and 1843 * ignore. p can replace it. 1844 */ 1845 if (vm_page_remove(pp)) 1846 vm_page_free(pp); 1847 pp = NULL; 1848 } 1849 1850 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1851 NULL)) { 1852 /* 1853 * The page already exists in the parent OR swap exists 1854 * for this location in the parent. Leave the parent's 1855 * page alone. Destroy the original page from the 1856 * backing object. 1857 */ 1858 vm_pager_freespace(backing_object, p->pindex, 1); 1859 KASSERT(!pmap_page_is_mapped(p), 1860 ("freeing mapped page %p", p)); 1861 if (vm_page_remove(p)) 1862 vm_page_free(p); 1863 if (pp != NULL) 1864 vm_page_xunbusy(pp); 1865 continue; 1866 } 1867 1868 /* 1869 * Page does not exist in parent, rename the page from the 1870 * backing object to the main object. 1871 * 1872 * If the page was mapped to a process, it can remain mapped 1873 * through the rename. vm_page_rename() will dirty the page. 1874 */ 1875 if (vm_page_rename(p, object, new_pindex)) { 1876 vm_page_xunbusy(p); 1877 next = vm_object_collapse_scan_wait(object, NULL); 1878 continue; 1879 } 1880 1881 /* Use the old pindex to free the right page. */ 1882 vm_pager_freespace(backing_object, new_pindex + 1883 backing_offset_index, 1); 1884 1885 #if VM_NRESERVLEVEL > 0 1886 /* 1887 * Rename the reservation. 1888 */ 1889 vm_reserv_rename(p, object, backing_object, 1890 backing_offset_index); 1891 #endif 1892 vm_page_xunbusy(p); 1893 } 1894 return; 1895 } 1896 1897 /* 1898 * vm_object_collapse: 1899 * 1900 * Collapse an object with the object backing it. 1901 * Pages in the backing object are moved into the 1902 * parent, and the backing object is deallocated. 1903 */ 1904 void 1905 vm_object_collapse(vm_object_t object) 1906 { 1907 vm_object_t backing_object, new_backing_object; 1908 1909 VM_OBJECT_ASSERT_WLOCKED(object); 1910 1911 while (TRUE) { 1912 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1913 ("collapsing invalid object")); 1914 1915 /* 1916 * Wait for the backing_object to finish any pending 1917 * collapse so that the caller sees the shortest possible 1918 * shadow chain. 1919 */ 1920 backing_object = vm_object_backing_collapse_wait(object); 1921 if (backing_object == NULL) 1922 return; 1923 1924 KASSERT(object->ref_count > 0 && 1925 object->ref_count > atomic_load_int(&object->shadow_count), 1926 ("collapse with invalid ref %d or shadow %d count.", 1927 object->ref_count, atomic_load_int(&object->shadow_count))); 1928 KASSERT((backing_object->flags & 1929 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1930 ("vm_object_collapse: Backing object already collapsing.")); 1931 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1932 ("vm_object_collapse: object is already collapsing.")); 1933 1934 /* 1935 * We know that we can either collapse the backing object if 1936 * the parent is the only reference to it, or (perhaps) have 1937 * the parent bypass the object if the parent happens to shadow 1938 * all the resident pages in the entire backing object. 1939 */ 1940 if (backing_object->ref_count == 1) { 1941 KASSERT(atomic_load_int(&backing_object->shadow_count) 1942 == 1, 1943 ("vm_object_collapse: shadow_count: %d", 1944 atomic_load_int(&backing_object->shadow_count))); 1945 vm_object_pip_add(object, 1); 1946 vm_object_set_flag(object, OBJ_COLLAPSING); 1947 vm_object_pip_add(backing_object, 1); 1948 vm_object_set_flag(backing_object, OBJ_DEAD); 1949 1950 /* 1951 * If there is exactly one reference to the backing 1952 * object, we can collapse it into the parent. 1953 */ 1954 vm_object_collapse_scan(object); 1955 1956 /* 1957 * Move the pager from backing_object to object. 1958 * 1959 * swap_pager_copy() can sleep, in which case the 1960 * backing_object's and object's locks are released and 1961 * reacquired. 1962 */ 1963 swap_pager_copy(backing_object, object, 1964 OFF_TO_IDX(object->backing_object_offset), TRUE); 1965 1966 /* 1967 * Object now shadows whatever backing_object did. 1968 */ 1969 vm_object_clear_flag(object, OBJ_COLLAPSING); 1970 vm_object_backing_transfer(object, backing_object); 1971 object->backing_object_offset += 1972 backing_object->backing_object_offset; 1973 VM_OBJECT_WUNLOCK(object); 1974 vm_object_pip_wakeup(object); 1975 1976 /* 1977 * Discard backing_object. 1978 * 1979 * Since the backing object has no pages, no pager left, 1980 * and no object references within it, all that is 1981 * necessary is to dispose of it. 1982 */ 1983 KASSERT(backing_object->ref_count == 1, ( 1984 "backing_object %p was somehow re-referenced during collapse!", 1985 backing_object)); 1986 vm_object_pip_wakeup(backing_object); 1987 (void)refcount_release(&backing_object->ref_count); 1988 umtx_shm_object_terminated(backing_object); 1989 vm_object_terminate(backing_object); 1990 counter_u64_add(object_collapses, 1); 1991 VM_OBJECT_WLOCK(object); 1992 } else { 1993 /* 1994 * If we do not entirely shadow the backing object, 1995 * there is nothing we can do so we give up. 1996 * 1997 * The object lock and backing_object lock must not 1998 * be dropped during this sequence. 1999 */ 2000 if (!vm_object_scan_all_shadowed(object)) { 2001 VM_OBJECT_WUNLOCK(backing_object); 2002 break; 2003 } 2004 2005 /* 2006 * Make the parent shadow the next object in the 2007 * chain. Deallocating backing_object will not remove 2008 * it, since its reference count is at least 2. 2009 */ 2010 vm_object_backing_remove_locked(object); 2011 new_backing_object = backing_object->backing_object; 2012 if (new_backing_object != NULL) { 2013 vm_object_backing_insert_ref(object, 2014 new_backing_object); 2015 object->backing_object_offset += 2016 backing_object->backing_object_offset; 2017 } 2018 2019 /* 2020 * Drop the reference count on backing_object. Since 2021 * its ref_count was at least 2, it will not vanish. 2022 */ 2023 (void)refcount_release(&backing_object->ref_count); 2024 KASSERT(backing_object->ref_count >= 1, ( 2025 "backing_object %p was somehow dereferenced during collapse!", 2026 backing_object)); 2027 VM_OBJECT_WUNLOCK(backing_object); 2028 counter_u64_add(object_bypasses, 1); 2029 } 2030 2031 /* 2032 * Try again with this object's new backing object. 2033 */ 2034 } 2035 } 2036 2037 /* 2038 * vm_object_page_remove: 2039 * 2040 * For the given object, either frees or invalidates each of the 2041 * specified pages. In general, a page is freed. However, if a page is 2042 * wired for any reason other than the existence of a managed, wired 2043 * mapping, then it may be invalidated but not removed from the object. 2044 * Pages are specified by the given range ["start", "end") and the option 2045 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 2046 * extends from "start" to the end of the object. If the option 2047 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 2048 * specified range are affected. If the option OBJPR_NOTMAPPED is 2049 * specified, then the pages within the specified range must have no 2050 * mappings. Otherwise, if this option is not specified, any mappings to 2051 * the specified pages are removed before the pages are freed or 2052 * invalidated. 2053 * 2054 * In general, this operation should only be performed on objects that 2055 * contain managed pages. There are, however, two exceptions. First, it 2056 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 2057 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 2058 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 2059 * not be specified and the option OBJPR_NOTMAPPED must be specified. 2060 * 2061 * The object must be locked. 2062 */ 2063 void 2064 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2065 int options) 2066 { 2067 vm_page_t p, next; 2068 2069 VM_OBJECT_ASSERT_WLOCKED(object); 2070 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 2071 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 2072 ("vm_object_page_remove: illegal options for object %p", object)); 2073 if (object->resident_page_count == 0) 2074 return; 2075 vm_object_pip_add(object, 1); 2076 again: 2077 p = vm_page_find_least(object, start); 2078 2079 /* 2080 * Here, the variable "p" is either (1) the page with the least pindex 2081 * greater than or equal to the parameter "start" or (2) NULL. 2082 */ 2083 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2084 next = TAILQ_NEXT(p, listq); 2085 2086 /* 2087 * Skip invalid pages if asked to do so. Try to avoid acquiring 2088 * the busy lock, as some consumers rely on this to avoid 2089 * deadlocks. 2090 * 2091 * A thread may concurrently transition the page from invalid to 2092 * valid using only the busy lock, so the result of this check 2093 * is immediately stale. It is up to consumers to handle this, 2094 * for instance by ensuring that all invalid->valid transitions 2095 * happen with a mutex held, as may be possible for a 2096 * filesystem. 2097 */ 2098 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) 2099 continue; 2100 2101 /* 2102 * If the page is wired for any reason besides the existence 2103 * of managed, wired mappings, then it cannot be freed. For 2104 * example, fictitious pages, which represent device memory, 2105 * are inherently wired and cannot be freed. They can, 2106 * however, be invalidated if the option OBJPR_CLEANONLY is 2107 * not specified. 2108 */ 2109 if (vm_page_tryxbusy(p) == 0) { 2110 if (vm_page_busy_sleep(p, "vmopar", 0)) 2111 VM_OBJECT_WLOCK(object); 2112 goto again; 2113 } 2114 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) { 2115 vm_page_xunbusy(p); 2116 continue; 2117 } 2118 if (vm_page_wired(p)) { 2119 wired: 2120 if ((options & OBJPR_NOTMAPPED) == 0 && 2121 object->ref_count != 0) 2122 pmap_remove_all(p); 2123 if ((options & OBJPR_CLEANONLY) == 0) { 2124 vm_page_invalid(p); 2125 vm_page_undirty(p); 2126 } 2127 vm_page_xunbusy(p); 2128 continue; 2129 } 2130 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2131 ("vm_object_page_remove: page %p is fictitious", p)); 2132 if ((options & OBJPR_CLEANONLY) != 0 && 2133 !vm_page_none_valid(p)) { 2134 if ((options & OBJPR_NOTMAPPED) == 0 && 2135 object->ref_count != 0 && 2136 !vm_page_try_remove_write(p)) 2137 goto wired; 2138 if (p->dirty != 0) { 2139 vm_page_xunbusy(p); 2140 continue; 2141 } 2142 } 2143 if ((options & OBJPR_NOTMAPPED) == 0 && 2144 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2145 goto wired; 2146 vm_page_free(p); 2147 } 2148 vm_object_pip_wakeup(object); 2149 2150 vm_pager_freespace(object, start, (end == 0 ? object->size : end) - 2151 start); 2152 } 2153 2154 /* 2155 * vm_object_page_noreuse: 2156 * 2157 * For the given object, attempt to move the specified pages to 2158 * the head of the inactive queue. This bypasses regular LRU 2159 * operation and allows the pages to be reused quickly under memory 2160 * pressure. If a page is wired for any reason, then it will not 2161 * be queued. Pages are specified by the range ["start", "end"). 2162 * As a special case, if "end" is zero, then the range extends from 2163 * "start" to the end of the object. 2164 * 2165 * This operation should only be performed on objects that 2166 * contain non-fictitious, managed pages. 2167 * 2168 * The object must be locked. 2169 */ 2170 void 2171 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2172 { 2173 vm_page_t p, next; 2174 2175 VM_OBJECT_ASSERT_LOCKED(object); 2176 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2177 ("vm_object_page_noreuse: illegal object %p", object)); 2178 if (object->resident_page_count == 0) 2179 return; 2180 p = vm_page_find_least(object, start); 2181 2182 /* 2183 * Here, the variable "p" is either (1) the page with the least pindex 2184 * greater than or equal to the parameter "start" or (2) NULL. 2185 */ 2186 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2187 next = TAILQ_NEXT(p, listq); 2188 vm_page_deactivate_noreuse(p); 2189 } 2190 } 2191 2192 /* 2193 * Populate the specified range of the object with valid pages. Returns 2194 * TRUE if the range is successfully populated and FALSE otherwise. 2195 * 2196 * Note: This function should be optimized to pass a larger array of 2197 * pages to vm_pager_get_pages() before it is applied to a non- 2198 * OBJT_DEVICE object. 2199 * 2200 * The object must be locked. 2201 */ 2202 boolean_t 2203 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2204 { 2205 vm_page_t m; 2206 vm_pindex_t pindex; 2207 int rv; 2208 2209 VM_OBJECT_ASSERT_WLOCKED(object); 2210 for (pindex = start; pindex < end; pindex++) { 2211 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); 2212 if (rv != VM_PAGER_OK) 2213 break; 2214 2215 /* 2216 * Keep "m" busy because a subsequent iteration may unlock 2217 * the object. 2218 */ 2219 } 2220 if (pindex > start) { 2221 m = vm_page_lookup(object, start); 2222 while (m != NULL && m->pindex < pindex) { 2223 vm_page_xunbusy(m); 2224 m = TAILQ_NEXT(m, listq); 2225 } 2226 } 2227 return (pindex == end); 2228 } 2229 2230 /* 2231 * Routine: vm_object_coalesce 2232 * Function: Coalesces two objects backing up adjoining 2233 * regions of memory into a single object. 2234 * 2235 * returns TRUE if objects were combined. 2236 * 2237 * NOTE: Only works at the moment if the second object is NULL - 2238 * if it's not, which object do we lock first? 2239 * 2240 * Parameters: 2241 * prev_object First object to coalesce 2242 * prev_offset Offset into prev_object 2243 * prev_size Size of reference to prev_object 2244 * next_size Size of reference to the second object 2245 * reserved Indicator that extension region has 2246 * swap accounted for 2247 * 2248 * Conditions: 2249 * The object must *not* be locked. 2250 */ 2251 boolean_t 2252 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2253 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2254 { 2255 vm_pindex_t next_pindex; 2256 2257 if (prev_object == NULL) 2258 return (TRUE); 2259 if ((prev_object->flags & OBJ_ANON) == 0) 2260 return (FALSE); 2261 2262 VM_OBJECT_WLOCK(prev_object); 2263 /* 2264 * Try to collapse the object first. 2265 */ 2266 vm_object_collapse(prev_object); 2267 2268 /* 2269 * Can't coalesce if: . more than one reference . paged out . shadows 2270 * another object . has a copy elsewhere (any of which mean that the 2271 * pages not mapped to prev_entry may be in use anyway) 2272 */ 2273 if (prev_object->backing_object != NULL) { 2274 VM_OBJECT_WUNLOCK(prev_object); 2275 return (FALSE); 2276 } 2277 2278 prev_size >>= PAGE_SHIFT; 2279 next_size >>= PAGE_SHIFT; 2280 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2281 2282 if (prev_object->ref_count > 1 && 2283 prev_object->size != next_pindex && 2284 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2285 VM_OBJECT_WUNLOCK(prev_object); 2286 return (FALSE); 2287 } 2288 2289 /* 2290 * Account for the charge. 2291 */ 2292 if (prev_object->cred != NULL) { 2293 /* 2294 * If prev_object was charged, then this mapping, 2295 * although not charged now, may become writable 2296 * later. Non-NULL cred in the object would prevent 2297 * swap reservation during enabling of the write 2298 * access, so reserve swap now. Failed reservation 2299 * cause allocation of the separate object for the map 2300 * entry, and swap reservation for this entry is 2301 * managed in appropriate time. 2302 */ 2303 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2304 prev_object->cred)) { 2305 VM_OBJECT_WUNLOCK(prev_object); 2306 return (FALSE); 2307 } 2308 prev_object->charge += ptoa(next_size); 2309 } 2310 2311 /* 2312 * Remove any pages that may still be in the object from a previous 2313 * deallocation. 2314 */ 2315 if (next_pindex < prev_object->size) { 2316 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2317 next_size, 0); 2318 #if 0 2319 if (prev_object->cred != NULL) { 2320 KASSERT(prev_object->charge >= 2321 ptoa(prev_object->size - next_pindex), 2322 ("object %p overcharged 1 %jx %jx", prev_object, 2323 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2324 prev_object->charge -= ptoa(prev_object->size - 2325 next_pindex); 2326 } 2327 #endif 2328 } 2329 2330 /* 2331 * Extend the object if necessary. 2332 */ 2333 if (next_pindex + next_size > prev_object->size) 2334 prev_object->size = next_pindex + next_size; 2335 2336 VM_OBJECT_WUNLOCK(prev_object); 2337 return (TRUE); 2338 } 2339 2340 void 2341 vm_object_set_writeable_dirty_(vm_object_t object) 2342 { 2343 atomic_add_int(&object->generation, 1); 2344 } 2345 2346 bool 2347 vm_object_mightbedirty_(vm_object_t object) 2348 { 2349 return (object->generation != object->cleangeneration); 2350 } 2351 2352 /* 2353 * vm_object_unwire: 2354 * 2355 * For each page offset within the specified range of the given object, 2356 * find the highest-level page in the shadow chain and unwire it. A page 2357 * must exist at every page offset, and the highest-level page must be 2358 * wired. 2359 */ 2360 void 2361 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2362 uint8_t queue) 2363 { 2364 vm_object_t tobject, t1object; 2365 vm_page_t m, tm; 2366 vm_pindex_t end_pindex, pindex, tpindex; 2367 int depth, locked_depth; 2368 2369 KASSERT((offset & PAGE_MASK) == 0, 2370 ("vm_object_unwire: offset is not page aligned")); 2371 KASSERT((length & PAGE_MASK) == 0, 2372 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2373 /* The wired count of a fictitious page never changes. */ 2374 if ((object->flags & OBJ_FICTITIOUS) != 0) 2375 return; 2376 pindex = OFF_TO_IDX(offset); 2377 end_pindex = pindex + atop(length); 2378 again: 2379 locked_depth = 1; 2380 VM_OBJECT_RLOCK(object); 2381 m = vm_page_find_least(object, pindex); 2382 while (pindex < end_pindex) { 2383 if (m == NULL || pindex < m->pindex) { 2384 /* 2385 * The first object in the shadow chain doesn't 2386 * contain a page at the current index. Therefore, 2387 * the page must exist in a backing object. 2388 */ 2389 tobject = object; 2390 tpindex = pindex; 2391 depth = 0; 2392 do { 2393 tpindex += 2394 OFF_TO_IDX(tobject->backing_object_offset); 2395 tobject = tobject->backing_object; 2396 KASSERT(tobject != NULL, 2397 ("vm_object_unwire: missing page")); 2398 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2399 goto next_page; 2400 depth++; 2401 if (depth == locked_depth) { 2402 locked_depth++; 2403 VM_OBJECT_RLOCK(tobject); 2404 } 2405 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2406 NULL); 2407 } else { 2408 tm = m; 2409 m = TAILQ_NEXT(m, listq); 2410 } 2411 if (vm_page_trysbusy(tm) == 0) { 2412 for (tobject = object; locked_depth >= 1; 2413 locked_depth--) { 2414 t1object = tobject->backing_object; 2415 if (tm->object != tobject) 2416 VM_OBJECT_RUNLOCK(tobject); 2417 tobject = t1object; 2418 } 2419 tobject = tm->object; 2420 if (!vm_page_busy_sleep(tm, "unwbo", 2421 VM_ALLOC_IGN_SBUSY)) 2422 VM_OBJECT_RUNLOCK(tobject); 2423 goto again; 2424 } 2425 vm_page_unwire(tm, queue); 2426 vm_page_sunbusy(tm); 2427 next_page: 2428 pindex++; 2429 } 2430 /* Release the accumulated object locks. */ 2431 for (tobject = object; locked_depth >= 1; locked_depth--) { 2432 t1object = tobject->backing_object; 2433 VM_OBJECT_RUNLOCK(tobject); 2434 tobject = t1object; 2435 } 2436 } 2437 2438 /* 2439 * Return the vnode for the given object, or NULL if none exists. 2440 * For tmpfs objects, the function may return NULL if there is 2441 * no vnode allocated at the time of the call. 2442 */ 2443 struct vnode * 2444 vm_object_vnode(vm_object_t object) 2445 { 2446 struct vnode *vp; 2447 2448 VM_OBJECT_ASSERT_LOCKED(object); 2449 vm_pager_getvp(object, &vp, NULL); 2450 return (vp); 2451 } 2452 2453 /* 2454 * Busy the vm object. This prevents new pages belonging to the object from 2455 * becoming busy. Existing pages persist as busy. Callers are responsible 2456 * for checking page state before proceeding. 2457 */ 2458 void 2459 vm_object_busy(vm_object_t obj) 2460 { 2461 2462 VM_OBJECT_ASSERT_LOCKED(obj); 2463 2464 blockcount_acquire(&obj->busy, 1); 2465 /* The fence is required to order loads of page busy. */ 2466 atomic_thread_fence_acq_rel(); 2467 } 2468 2469 void 2470 vm_object_unbusy(vm_object_t obj) 2471 { 2472 2473 blockcount_release(&obj->busy, 1); 2474 } 2475 2476 void 2477 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2478 { 2479 2480 VM_OBJECT_ASSERT_UNLOCKED(obj); 2481 2482 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2483 } 2484 2485 /* 2486 * This function aims to determine if the object is mapped, 2487 * specifically, if it is referenced by a vm_map_entry. Because 2488 * objects occasionally acquire transient references that do not 2489 * represent a mapping, the method used here is inexact. However, it 2490 * has very low overhead and is good enough for the advisory 2491 * vm.vmtotal sysctl. 2492 */ 2493 bool 2494 vm_object_is_active(vm_object_t obj) 2495 { 2496 2497 return (obj->ref_count > atomic_load_int(&obj->shadow_count)); 2498 } 2499 2500 static int 2501 vm_object_list_handler(struct sysctl_req *req, bool swap_only) 2502 { 2503 struct kinfo_vmobject *kvo; 2504 char *fullpath, *freepath; 2505 struct vnode *vp; 2506 struct vattr va; 2507 vm_object_t obj; 2508 vm_page_t m; 2509 u_long sp; 2510 int count, error; 2511 key_t key; 2512 unsigned short seq; 2513 bool want_path; 2514 2515 if (req->oldptr == NULL) { 2516 /* 2517 * If an old buffer has not been provided, generate an 2518 * estimate of the space needed for a subsequent call. 2519 */ 2520 mtx_lock(&vm_object_list_mtx); 2521 count = 0; 2522 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2523 if (obj->type == OBJT_DEAD) 2524 continue; 2525 count++; 2526 } 2527 mtx_unlock(&vm_object_list_mtx); 2528 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2529 count * 11 / 10)); 2530 } 2531 2532 want_path = !(swap_only || jailed(curthread->td_ucred)); 2533 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO); 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 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) 2545 continue; 2546 VM_OBJECT_RLOCK(obj); 2547 if (obj->type == OBJT_DEAD || 2548 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) { 2549 VM_OBJECT_RUNLOCK(obj); 2550 continue; 2551 } 2552 mtx_unlock(&vm_object_list_mtx); 2553 kvo->kvo_size = ptoa(obj->size); 2554 kvo->kvo_resident = obj->resident_page_count; 2555 kvo->kvo_ref_count = obj->ref_count; 2556 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count); 2557 kvo->kvo_memattr = obj->memattr; 2558 kvo->kvo_active = 0; 2559 kvo->kvo_inactive = 0; 2560 kvo->kvo_flags = 0; 2561 if (!swap_only) { 2562 TAILQ_FOREACH(m, &obj->memq, listq) { 2563 /* 2564 * A page may belong to the object but be 2565 * dequeued and set to PQ_NONE while the 2566 * object lock is not held. This makes the 2567 * reads of m->queue below racy, and we do not 2568 * count pages set to PQ_NONE. However, this 2569 * sysctl is only meant to give an 2570 * approximation of the system anyway. 2571 */ 2572 if (m->a.queue == PQ_ACTIVE) 2573 kvo->kvo_active++; 2574 else if (m->a.queue == PQ_INACTIVE) 2575 kvo->kvo_inactive++; 2576 } 2577 } 2578 2579 kvo->kvo_vn_fileid = 0; 2580 kvo->kvo_vn_fsid = 0; 2581 kvo->kvo_vn_fsid_freebsd11 = 0; 2582 freepath = NULL; 2583 fullpath = ""; 2584 vp = NULL; 2585 kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp : 2586 NULL); 2587 if (vp != NULL) { 2588 vref(vp); 2589 } else if ((obj->flags & OBJ_ANON) != 0) { 2590 MPASS(kvo->kvo_type == KVME_TYPE_SWAP); 2591 kvo->kvo_me = (uintptr_t)obj; 2592 /* tmpfs objs are reported as vnodes */ 2593 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object; 2594 sp = swap_pager_swapped_pages(obj); 2595 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp; 2596 } 2597 VM_OBJECT_RUNLOCK(obj); 2598 if ((obj->flags & OBJ_SYSVSHM) != 0) { 2599 kvo->kvo_flags |= KVMO_FLAG_SYSVSHM; 2600 shmobjinfo(obj, &key, &seq); 2601 kvo->kvo_vn_fileid = key; 2602 kvo->kvo_vn_fsid_freebsd11 = seq; 2603 } 2604 if (vp != NULL) { 2605 vn_fullpath(vp, &fullpath, &freepath); 2606 vn_lock(vp, LK_SHARED | LK_RETRY); 2607 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2608 kvo->kvo_vn_fileid = va.va_fileid; 2609 kvo->kvo_vn_fsid = va.va_fsid; 2610 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2611 /* truncate */ 2612 } 2613 vput(vp); 2614 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2615 free(freepath, M_TEMP); 2616 } 2617 2618 /* Pack record size down */ 2619 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2620 + strlen(kvo->kvo_path) + 1; 2621 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2622 sizeof(uint64_t)); 2623 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2624 maybe_yield(); 2625 mtx_lock(&vm_object_list_mtx); 2626 if (error) 2627 break; 2628 } 2629 mtx_unlock(&vm_object_list_mtx); 2630 free(kvo, M_TEMP); 2631 return (error); 2632 } 2633 2634 static int 2635 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2636 { 2637 return (vm_object_list_handler(req, false)); 2638 } 2639 2640 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2641 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2642 "List of VM objects"); 2643 2644 static int 2645 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS) 2646 { 2647 return (vm_object_list_handler(req, true)); 2648 } 2649 2650 /* 2651 * This sysctl returns list of the anonymous or swap objects. Intent 2652 * is to provide stripped optimized list useful to analyze swap use. 2653 * Since technically non-swap (default) objects participate in the 2654 * shadow chains, and are converted to swap type as needed by swap 2655 * pager, we must report them. 2656 */ 2657 SYSCTL_PROC(_vm, OID_AUTO, swap_objects, 2658 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, 2659 sysctl_vm_object_list_swap, "S,kinfo_vmobject", 2660 "List of swap VM objects"); 2661 2662 #include "opt_ddb.h" 2663 #ifdef DDB 2664 #include <sys/kernel.h> 2665 2666 #include <sys/cons.h> 2667 2668 #include <ddb/ddb.h> 2669 2670 static int 2671 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2672 { 2673 vm_map_t tmpm; 2674 vm_map_entry_t tmpe; 2675 vm_object_t obj; 2676 2677 if (map == 0) 2678 return 0; 2679 2680 if (entry == 0) { 2681 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2682 if (_vm_object_in_map(map, object, tmpe)) { 2683 return 1; 2684 } 2685 } 2686 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2687 tmpm = entry->object.sub_map; 2688 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2689 if (_vm_object_in_map(tmpm, object, tmpe)) { 2690 return 1; 2691 } 2692 } 2693 } else if ((obj = entry->object.vm_object) != NULL) { 2694 for (; obj; obj = obj->backing_object) 2695 if (obj == object) { 2696 return 1; 2697 } 2698 } 2699 return 0; 2700 } 2701 2702 static int 2703 vm_object_in_map(vm_object_t object) 2704 { 2705 struct proc *p; 2706 2707 /* sx_slock(&allproc_lock); */ 2708 FOREACH_PROC_IN_SYSTEM(p) { 2709 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2710 continue; 2711 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2712 /* sx_sunlock(&allproc_lock); */ 2713 return 1; 2714 } 2715 } 2716 /* sx_sunlock(&allproc_lock); */ 2717 if (_vm_object_in_map(kernel_map, object, 0)) 2718 return 1; 2719 return 0; 2720 } 2721 2722 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE) 2723 { 2724 vm_object_t object; 2725 2726 /* 2727 * make sure that internal objs are in a map somewhere 2728 * and none have zero ref counts. 2729 */ 2730 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2731 if ((object->flags & OBJ_ANON) != 0) { 2732 if (object->ref_count == 0) { 2733 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2734 (long)object->size); 2735 } 2736 if (!vm_object_in_map(object)) { 2737 db_printf( 2738 "vmochk: internal obj is not in a map: " 2739 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2740 object->ref_count, (u_long)object->size, 2741 (u_long)object->size, 2742 (void *)object->backing_object); 2743 } 2744 } 2745 if (db_pager_quit) 2746 return; 2747 } 2748 } 2749 2750 /* 2751 * vm_object_print: [ debug ] 2752 */ 2753 DB_SHOW_COMMAND(object, vm_object_print_static) 2754 { 2755 /* XXX convert args. */ 2756 vm_object_t object = (vm_object_t)addr; 2757 boolean_t full = have_addr; 2758 2759 vm_page_t p; 2760 2761 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2762 #define count was_count 2763 2764 int count; 2765 2766 if (object == NULL) 2767 return; 2768 2769 db_iprintf( 2770 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2771 object, (int)object->type, (uintmax_t)object->size, 2772 object->resident_page_count, object->ref_count, object->flags, 2773 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2774 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2775 atomic_load_int(&object->shadow_count), 2776 object->backing_object ? object->backing_object->ref_count : 0, 2777 object->backing_object, (uintmax_t)object->backing_object_offset); 2778 2779 if (!full) 2780 return; 2781 2782 db_indent += 2; 2783 count = 0; 2784 TAILQ_FOREACH(p, &object->memq, listq) { 2785 if (count == 0) 2786 db_iprintf("memory:="); 2787 else if (count == 6) { 2788 db_printf("\n"); 2789 db_iprintf(" ..."); 2790 count = 0; 2791 } else 2792 db_printf(","); 2793 count++; 2794 2795 db_printf("(off=0x%jx,page=0x%jx)", 2796 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2797 2798 if (db_pager_quit) 2799 break; 2800 } 2801 if (count != 0) 2802 db_printf("\n"); 2803 db_indent -= 2; 2804 } 2805 2806 /* XXX. */ 2807 #undef count 2808 2809 /* XXX need this non-static entry for calling from vm_map_print. */ 2810 void 2811 vm_object_print( 2812 /* db_expr_t */ long addr, 2813 boolean_t have_addr, 2814 /* db_expr_t */ long count, 2815 char *modif) 2816 { 2817 vm_object_print_static(addr, have_addr, count, modif); 2818 } 2819 2820 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE) 2821 { 2822 vm_object_t object; 2823 vm_pindex_t fidx; 2824 vm_paddr_t pa; 2825 vm_page_t m, prev_m; 2826 int rcount; 2827 2828 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2829 db_printf("new object: %p\n", (void *)object); 2830 if (db_pager_quit) 2831 return; 2832 2833 rcount = 0; 2834 fidx = 0; 2835 pa = -1; 2836 TAILQ_FOREACH(m, &object->memq, listq) { 2837 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2838 prev_m->pindex + 1 != m->pindex) { 2839 if (rcount) { 2840 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2841 (long)fidx, rcount, (long)pa); 2842 if (db_pager_quit) 2843 return; 2844 rcount = 0; 2845 } 2846 } 2847 if (rcount && 2848 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2849 ++rcount; 2850 continue; 2851 } 2852 if (rcount) { 2853 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2854 (long)fidx, rcount, (long)pa); 2855 if (db_pager_quit) 2856 return; 2857 } 2858 fidx = m->pindex; 2859 pa = VM_PAGE_TO_PHYS(m); 2860 rcount = 1; 2861 } 2862 if (rcount) { 2863 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2864 (long)fidx, rcount, (long)pa); 2865 if (db_pager_quit) 2866 return; 2867 } 2868 } 2869 } 2870 #endif /* DDB */ 2871