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/jail.h> 71 #include <sys/limits.h> 72 #include <sys/lock.h> 73 #include <sys/mman.h> 74 #include <sys/mount.h> 75 #include <sys/kernel.h> 76 #include <sys/mutex.h> 77 #include <sys/pctrie.h> 78 #include <sys/proc.h> 79 #include <sys/refcount.h> 80 #include <sys/sx.h> 81 #include <sys/sysctl.h> 82 #include <sys/resourcevar.h> 83 #include <sys/refcount.h> 84 #include <sys/rwlock.h> 85 #include <sys/user.h> 86 #include <sys/vnode.h> 87 #include <sys/vmmeter.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_param.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_pageout.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vm_phys.h> 98 #include <vm/vm_pagequeue.h> 99 #include <vm/swap_pager.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_extern.h> 102 #include <vm/vm_radix.h> 103 #include <vm/vm_reserv.h> 104 #include <vm/uma.h> 105 106 static int old_msync; 107 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 108 "Use old (insecure) msync behavior"); 109 110 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, 111 int pagerflags, int flags, boolean_t *allclean, 112 boolean_t *eio); 113 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 114 boolean_t *allclean); 115 static void vm_object_backing_remove(vm_object_t object); 116 117 /* 118 * Virtual memory objects maintain the actual data 119 * associated with allocated virtual memory. A given 120 * page of memory exists within exactly one object. 121 * 122 * An object is only deallocated when all "references" 123 * are given up. Only one "reference" to a given 124 * region of an object should be writeable. 125 * 126 * Associated with each object is a list of all resident 127 * memory pages belonging to that object; this list is 128 * maintained by the "vm_page" module, and locked by the object's 129 * lock. 130 * 131 * Each object also records a "pager" routine which is 132 * used to retrieve (and store) pages to the proper backing 133 * storage. In addition, objects may be backed by other 134 * objects from which they were virtual-copied. 135 * 136 * The only items within the object structure which are 137 * modified after time of creation are: 138 * reference count locked by object's lock 139 * pager routine locked by object's lock 140 * 141 */ 142 143 struct object_q vm_object_list; 144 struct mtx vm_object_list_mtx; /* lock for object list and count */ 145 146 struct vm_object kernel_object_store; 147 148 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 149 "VM object stats"); 150 151 static COUNTER_U64_DEFINE_EARLY(object_collapses); 152 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 153 &object_collapses, 154 "VM object collapses"); 155 156 static COUNTER_U64_DEFINE_EARLY(object_bypasses); 157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 158 &object_bypasses, 159 "VM object bypasses"); 160 161 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits); 162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD, 163 &object_collapse_waits, 164 "Number of sleeps for collapse"); 165 166 static uma_zone_t obj_zone; 167 168 static int vm_object_zinit(void *mem, int size, int flags); 169 170 #ifdef INVARIANTS 171 static void vm_object_zdtor(void *mem, int size, void *arg); 172 173 static void 174 vm_object_zdtor(void *mem, int size, void *arg) 175 { 176 vm_object_t object; 177 178 object = (vm_object_t)mem; 179 KASSERT(object->ref_count == 0, 180 ("object %p ref_count = %d", object, object->ref_count)); 181 KASSERT(TAILQ_EMPTY(&object->memq), 182 ("object %p has resident pages in its memq", object)); 183 KASSERT(vm_radix_is_empty(&object->rtree), 184 ("object %p has resident pages in its trie", object)); 185 #if VM_NRESERVLEVEL > 0 186 KASSERT(LIST_EMPTY(&object->rvq), 187 ("object %p has reservations", 188 object)); 189 #endif 190 KASSERT(!vm_object_busied(object), 191 ("object %p busy = %d", object, blockcount_read(&object->busy))); 192 KASSERT(object->resident_page_count == 0, 193 ("object %p resident_page_count = %d", 194 object, object->resident_page_count)); 195 KASSERT(atomic_load_int(&object->shadow_count) == 0, 196 ("object %p shadow_count = %d", 197 object, atomic_load_int(&object->shadow_count))); 198 KASSERT(object->type == OBJT_DEAD, 199 ("object %p has non-dead type %d", 200 object, object->type)); 201 KASSERT(object->charge == 0 && object->cred == NULL, 202 ("object %p has non-zero charge %ju (%p)", 203 object, (uintmax_t)object->charge, object->cred)); 204 } 205 #endif 206 207 static int 208 vm_object_zinit(void *mem, int size, int flags) 209 { 210 vm_object_t object; 211 212 object = (vm_object_t)mem; 213 rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW); 214 215 /* These are true for any object that has been freed */ 216 object->type = OBJT_DEAD; 217 vm_radix_init(&object->rtree); 218 refcount_init(&object->ref_count, 0); 219 blockcount_init(&object->paging_in_progress); 220 blockcount_init(&object->busy); 221 object->resident_page_count = 0; 222 atomic_store_int(&object->shadow_count, 0); 223 object->flags = OBJ_DEAD; 224 225 mtx_lock(&vm_object_list_mtx); 226 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 227 mtx_unlock(&vm_object_list_mtx); 228 return (0); 229 } 230 231 static void 232 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, 233 vm_object_t object, void *handle) 234 { 235 236 TAILQ_INIT(&object->memq); 237 LIST_INIT(&object->shadow_head); 238 239 object->type = type; 240 object->flags = flags; 241 if ((flags & OBJ_SWAP) != 0) { 242 pctrie_init(&object->un_pager.swp.swp_blks); 243 object->un_pager.swp.writemappings = 0; 244 } 245 246 /* 247 * Ensure that swap_pager_swapoff() iteration over object_list 248 * sees up to date type and pctrie head if it observed 249 * non-dead object. 250 */ 251 atomic_thread_fence_rel(); 252 253 object->pg_color = 0; 254 object->size = size; 255 object->domain.dr_policy = NULL; 256 object->generation = 1; 257 object->cleangeneration = 1; 258 refcount_init(&object->ref_count, 1); 259 object->memattr = VM_MEMATTR_DEFAULT; 260 object->cred = NULL; 261 object->charge = 0; 262 object->handle = handle; 263 object->backing_object = NULL; 264 object->backing_object_offset = (vm_ooffset_t) 0; 265 #if VM_NRESERVLEVEL > 0 266 LIST_INIT(&object->rvq); 267 #endif 268 umtx_shm_object_init(object); 269 } 270 271 /* 272 * vm_object_init: 273 * 274 * Initialize the VM objects module. 275 */ 276 void 277 vm_object_init(void) 278 { 279 TAILQ_INIT(&vm_object_list); 280 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 281 282 rw_init(&kernel_object->lock, "kernel vm object"); 283 vm_radix_init(&kernel_object->rtree); 284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 285 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); 286 #if VM_NRESERVLEVEL > 0 287 kernel_object->flags |= OBJ_COLORED; 288 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 289 #endif 290 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 291 292 /* 293 * The lock portion of struct vm_object must be type stable due 294 * to vm_pageout_fallback_object_lock locking a vm object 295 * without holding any references to it. 296 * 297 * paging_in_progress is valid always. Lockless references to 298 * the objects may acquire pip and then check OBJ_DEAD. 299 */ 300 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 301 #ifdef INVARIANTS 302 vm_object_zdtor, 303 #else 304 NULL, 305 #endif 306 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 307 308 vm_radix_zinit(); 309 } 310 311 void 312 vm_object_clear_flag(vm_object_t object, u_short bits) 313 { 314 315 VM_OBJECT_ASSERT_WLOCKED(object); 316 object->flags &= ~bits; 317 } 318 319 /* 320 * Sets the default memory attribute for the specified object. Pages 321 * that are allocated to this object are by default assigned this memory 322 * attribute. 323 * 324 * Presently, this function must be called before any pages are allocated 325 * to the object. In the future, this requirement may be relaxed for 326 * "default" and "swap" objects. 327 */ 328 int 329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 330 { 331 332 VM_OBJECT_ASSERT_WLOCKED(object); 333 334 if (object->type == OBJT_DEAD) 335 return (KERN_INVALID_ARGUMENT); 336 if (!TAILQ_EMPTY(&object->memq)) 337 return (KERN_FAILURE); 338 339 object->memattr = memattr; 340 return (KERN_SUCCESS); 341 } 342 343 void 344 vm_object_pip_add(vm_object_t object, short i) 345 { 346 347 if (i > 0) 348 blockcount_acquire(&object->paging_in_progress, i); 349 } 350 351 void 352 vm_object_pip_wakeup(vm_object_t object) 353 { 354 355 vm_object_pip_wakeupn(object, 1); 356 } 357 358 void 359 vm_object_pip_wakeupn(vm_object_t object, short i) 360 { 361 362 if (i > 0) 363 blockcount_release(&object->paging_in_progress, i); 364 } 365 366 /* 367 * Atomically drop the object lock and wait for pip to drain. This protects 368 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 369 * on return. 370 */ 371 static void 372 vm_object_pip_sleep(vm_object_t object, const char *waitid) 373 { 374 375 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 376 waitid, PVM | PDROP); 377 } 378 379 void 380 vm_object_pip_wait(vm_object_t object, const char *waitid) 381 { 382 383 VM_OBJECT_ASSERT_WLOCKED(object); 384 385 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 386 PVM); 387 } 388 389 void 390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 391 { 392 393 VM_OBJECT_ASSERT_UNLOCKED(object); 394 395 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 396 } 397 398 /* 399 * vm_object_allocate: 400 * 401 * Returns a new object with the given size. 402 */ 403 vm_object_t 404 vm_object_allocate(objtype_t type, vm_pindex_t size) 405 { 406 vm_object_t object; 407 u_short flags; 408 409 switch (type) { 410 case OBJT_DEAD: 411 panic("vm_object_allocate: can't create OBJT_DEAD"); 412 case OBJT_SWAP: 413 flags = OBJ_COLORED | OBJ_SWAP; 414 break; 415 case OBJT_DEVICE: 416 case OBJT_SG: 417 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 418 break; 419 case OBJT_MGTDEVICE: 420 flags = OBJ_FICTITIOUS; 421 break; 422 case OBJT_PHYS: 423 flags = OBJ_UNMANAGED; 424 break; 425 case OBJT_VNODE: 426 flags = 0; 427 break; 428 default: 429 panic("vm_object_allocate: type %d is undefined or dynamic", 430 type); 431 } 432 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 433 _vm_object_allocate(type, size, flags, object, NULL); 434 435 return (object); 436 } 437 438 vm_object_t 439 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags) 440 { 441 vm_object_t object; 442 443 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */); 444 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 445 _vm_object_allocate(dyntype, size, flags, object, NULL); 446 447 return (object); 448 } 449 450 /* 451 * vm_object_allocate_anon: 452 * 453 * Returns a new default object of the given size and marked as 454 * anonymous memory for special split/collapse handling. Color 455 * to be initialized by the caller. 456 */ 457 vm_object_t 458 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 459 struct ucred *cred, vm_size_t charge) 460 { 461 vm_object_t handle, object; 462 463 if (backing_object == NULL) 464 handle = NULL; 465 else if ((backing_object->flags & OBJ_ANON) != 0) 466 handle = backing_object->handle; 467 else 468 handle = backing_object; 469 object = uma_zalloc(obj_zone, M_WAITOK); 470 _vm_object_allocate(OBJT_SWAP, size, 471 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle); 472 object->cred = cred; 473 object->charge = cred != NULL ? charge : 0; 474 return (object); 475 } 476 477 static void 478 vm_object_reference_vnode(vm_object_t object) 479 { 480 u_int old; 481 482 /* 483 * vnode objects need the lock for the first reference 484 * to serialize with vnode_object_deallocate(). 485 */ 486 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 487 VM_OBJECT_RLOCK(object); 488 old = refcount_acquire(&object->ref_count); 489 if (object->type == OBJT_VNODE && old == 0) 490 vref(object->handle); 491 VM_OBJECT_RUNLOCK(object); 492 } 493 } 494 495 /* 496 * vm_object_reference: 497 * 498 * Acquires a reference to the given object. 499 */ 500 void 501 vm_object_reference(vm_object_t object) 502 { 503 504 if (object == NULL) 505 return; 506 507 if (object->type == OBJT_VNODE) 508 vm_object_reference_vnode(object); 509 else 510 refcount_acquire(&object->ref_count); 511 KASSERT((object->flags & OBJ_DEAD) == 0, 512 ("vm_object_reference: Referenced dead object.")); 513 } 514 515 /* 516 * vm_object_reference_locked: 517 * 518 * Gets another reference to the given object. 519 * 520 * The object must be locked. 521 */ 522 void 523 vm_object_reference_locked(vm_object_t object) 524 { 525 u_int old; 526 527 VM_OBJECT_ASSERT_LOCKED(object); 528 old = refcount_acquire(&object->ref_count); 529 if (object->type == OBJT_VNODE && old == 0) 530 vref(object->handle); 531 KASSERT((object->flags & OBJ_DEAD) == 0, 532 ("vm_object_reference: Referenced dead object.")); 533 } 534 535 /* 536 * Handle deallocating an object of type OBJT_VNODE. 537 */ 538 static void 539 vm_object_deallocate_vnode(vm_object_t object) 540 { 541 struct vnode *vp = (struct vnode *) object->handle; 542 bool last; 543 544 KASSERT(object->type == OBJT_VNODE, 545 ("vm_object_deallocate_vnode: not a vnode object")); 546 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 547 548 /* Object lock to protect handle lookup. */ 549 last = refcount_release(&object->ref_count); 550 VM_OBJECT_RUNLOCK(object); 551 552 if (!last) 553 return; 554 555 if (!umtx_shm_vnobj_persistent) 556 umtx_shm_object_terminated(object); 557 558 /* vrele may need the vnode lock. */ 559 vrele(vp); 560 } 561 562 /* 563 * We dropped a reference on an object and discovered that it had a 564 * single remaining shadow. This is a sibling of the reference we 565 * dropped. Attempt to collapse the sibling and backing object. 566 */ 567 static vm_object_t 568 vm_object_deallocate_anon(vm_object_t backing_object) 569 { 570 vm_object_t object; 571 572 /* Fetch the final shadow. */ 573 object = LIST_FIRST(&backing_object->shadow_head); 574 KASSERT(object != NULL && 575 atomic_load_int(&backing_object->shadow_count) == 1, 576 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 577 backing_object->ref_count, 578 atomic_load_int(&backing_object->shadow_count))); 579 KASSERT((object->flags & OBJ_ANON) != 0, 580 ("invalid shadow object %p", object)); 581 582 if (!VM_OBJECT_TRYWLOCK(object)) { 583 /* 584 * Prevent object from disappearing since we do not have a 585 * reference. 586 */ 587 vm_object_pip_add(object, 1); 588 VM_OBJECT_WUNLOCK(backing_object); 589 VM_OBJECT_WLOCK(object); 590 vm_object_pip_wakeup(object); 591 } else 592 VM_OBJECT_WUNLOCK(backing_object); 593 594 /* 595 * Check for a collapse/terminate race with the last reference holder. 596 */ 597 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 598 !refcount_acquire_if_not_zero(&object->ref_count)) { 599 VM_OBJECT_WUNLOCK(object); 600 return (NULL); 601 } 602 backing_object = object->backing_object; 603 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 604 vm_object_collapse(object); 605 VM_OBJECT_WUNLOCK(object); 606 607 return (object); 608 } 609 610 /* 611 * vm_object_deallocate: 612 * 613 * Release a reference to the specified object, 614 * gained either through a vm_object_allocate 615 * or a vm_object_reference call. When all references 616 * are gone, storage associated with this object 617 * may be relinquished. 618 * 619 * No object may be locked. 620 */ 621 void 622 vm_object_deallocate(vm_object_t object) 623 { 624 vm_object_t temp; 625 bool released; 626 627 while (object != NULL) { 628 /* 629 * If the reference count goes to 0 we start calling 630 * vm_object_terminate() on the object chain. A ref count 631 * of 1 may be a special case depending on the shadow count 632 * being 0 or 1. These cases require a write lock on the 633 * object. 634 */ 635 if ((object->flags & OBJ_ANON) == 0) 636 released = refcount_release_if_gt(&object->ref_count, 1); 637 else 638 released = refcount_release_if_gt(&object->ref_count, 2); 639 if (released) 640 return; 641 642 if (object->type == OBJT_VNODE) { 643 VM_OBJECT_RLOCK(object); 644 if (object->type == OBJT_VNODE) { 645 vm_object_deallocate_vnode(object); 646 return; 647 } 648 VM_OBJECT_RUNLOCK(object); 649 } 650 651 VM_OBJECT_WLOCK(object); 652 KASSERT(object->ref_count > 0, 653 ("vm_object_deallocate: object deallocated too many times: %d", 654 object->type)); 655 656 /* 657 * If this is not the final reference to an anonymous 658 * object we may need to collapse the shadow chain. 659 */ 660 if (!refcount_release(&object->ref_count)) { 661 if (object->ref_count > 1 || 662 atomic_load_int(&object->shadow_count) == 0) { 663 if ((object->flags & OBJ_ANON) != 0 && 664 object->ref_count == 1) 665 vm_object_set_flag(object, 666 OBJ_ONEMAPPING); 667 VM_OBJECT_WUNLOCK(object); 668 return; 669 } 670 671 /* Handle collapsing last ref on anonymous objects. */ 672 object = vm_object_deallocate_anon(object); 673 continue; 674 } 675 676 /* 677 * Handle the final reference to an object. We restart 678 * the loop with the backing object to avoid recursion. 679 */ 680 umtx_shm_object_terminated(object); 681 temp = object->backing_object; 682 if (temp != NULL) { 683 KASSERT(object->type == OBJT_SWAP, 684 ("shadowed tmpfs v_object 2 %p", object)); 685 vm_object_backing_remove(object); 686 } 687 688 KASSERT((object->flags & OBJ_DEAD) == 0, 689 ("vm_object_deallocate: Terminating dead object.")); 690 vm_object_set_flag(object, OBJ_DEAD); 691 vm_object_terminate(object); 692 object = temp; 693 } 694 } 695 696 void 697 vm_object_destroy(vm_object_t object) 698 { 699 uma_zfree(obj_zone, object); 700 } 701 702 static void 703 vm_object_sub_shadow(vm_object_t object) 704 { 705 KASSERT(object->shadow_count >= 1, 706 ("object %p sub_shadow count zero", object)); 707 atomic_subtract_int(&object->shadow_count, 1); 708 } 709 710 static void 711 vm_object_backing_remove_locked(vm_object_t object) 712 { 713 vm_object_t backing_object; 714 715 backing_object = object->backing_object; 716 VM_OBJECT_ASSERT_WLOCKED(object); 717 VM_OBJECT_ASSERT_WLOCKED(backing_object); 718 719 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 720 ("vm_object_backing_remove: Removing collapsing object.")); 721 722 vm_object_sub_shadow(backing_object); 723 if ((object->flags & OBJ_SHADOWLIST) != 0) { 724 LIST_REMOVE(object, shadow_list); 725 vm_object_clear_flag(object, OBJ_SHADOWLIST); 726 } 727 object->backing_object = NULL; 728 } 729 730 static void 731 vm_object_backing_remove(vm_object_t object) 732 { 733 vm_object_t backing_object; 734 735 VM_OBJECT_ASSERT_WLOCKED(object); 736 737 backing_object = object->backing_object; 738 if ((object->flags & OBJ_SHADOWLIST) != 0) { 739 VM_OBJECT_WLOCK(backing_object); 740 vm_object_backing_remove_locked(object); 741 VM_OBJECT_WUNLOCK(backing_object); 742 } else { 743 object->backing_object = NULL; 744 vm_object_sub_shadow(backing_object); 745 } 746 } 747 748 static void 749 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 750 { 751 752 VM_OBJECT_ASSERT_WLOCKED(object); 753 754 atomic_add_int(&backing_object->shadow_count, 1); 755 if ((backing_object->flags & OBJ_ANON) != 0) { 756 VM_OBJECT_ASSERT_WLOCKED(backing_object); 757 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 758 shadow_list); 759 vm_object_set_flag(object, OBJ_SHADOWLIST); 760 } 761 object->backing_object = backing_object; 762 } 763 764 static void 765 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 766 { 767 768 VM_OBJECT_ASSERT_WLOCKED(object); 769 770 if ((backing_object->flags & OBJ_ANON) != 0) { 771 VM_OBJECT_WLOCK(backing_object); 772 vm_object_backing_insert_locked(object, backing_object); 773 VM_OBJECT_WUNLOCK(backing_object); 774 } else { 775 object->backing_object = backing_object; 776 atomic_add_int(&backing_object->shadow_count, 1); 777 } 778 } 779 780 /* 781 * Insert an object into a backing_object's shadow list with an additional 782 * reference to the backing_object added. 783 */ 784 static void 785 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object) 786 { 787 788 VM_OBJECT_ASSERT_WLOCKED(object); 789 790 if ((backing_object->flags & OBJ_ANON) != 0) { 791 VM_OBJECT_WLOCK(backing_object); 792 KASSERT((backing_object->flags & OBJ_DEAD) == 0, 793 ("shadowing dead anonymous object")); 794 vm_object_reference_locked(backing_object); 795 vm_object_backing_insert_locked(object, backing_object); 796 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING); 797 VM_OBJECT_WUNLOCK(backing_object); 798 } else { 799 vm_object_reference(backing_object); 800 atomic_add_int(&backing_object->shadow_count, 1); 801 object->backing_object = backing_object; 802 } 803 } 804 805 /* 806 * Transfer a backing reference from backing_object to object. 807 */ 808 static void 809 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 810 { 811 vm_object_t new_backing_object; 812 813 /* 814 * Note that the reference to backing_object->backing_object 815 * moves from within backing_object to within object. 816 */ 817 vm_object_backing_remove_locked(object); 818 new_backing_object = backing_object->backing_object; 819 if (new_backing_object == NULL) 820 return; 821 if ((new_backing_object->flags & OBJ_ANON) != 0) { 822 VM_OBJECT_WLOCK(new_backing_object); 823 vm_object_backing_remove_locked(backing_object); 824 vm_object_backing_insert_locked(object, new_backing_object); 825 VM_OBJECT_WUNLOCK(new_backing_object); 826 } else { 827 /* 828 * shadow_count for new_backing_object is left 829 * unchanged, its reference provided by backing_object 830 * is replaced by object. 831 */ 832 object->backing_object = new_backing_object; 833 backing_object->backing_object = NULL; 834 } 835 } 836 837 /* 838 * Wait for a concurrent collapse to settle. 839 */ 840 static void 841 vm_object_collapse_wait(vm_object_t object) 842 { 843 844 VM_OBJECT_ASSERT_WLOCKED(object); 845 846 while ((object->flags & OBJ_COLLAPSING) != 0) { 847 vm_object_pip_wait(object, "vmcolwait"); 848 counter_u64_add(object_collapse_waits, 1); 849 } 850 } 851 852 /* 853 * Waits for a backing object to clear a pending collapse and returns 854 * it locked if it is an ANON object. 855 */ 856 static vm_object_t 857 vm_object_backing_collapse_wait(vm_object_t object) 858 { 859 vm_object_t backing_object; 860 861 VM_OBJECT_ASSERT_WLOCKED(object); 862 863 for (;;) { 864 backing_object = object->backing_object; 865 if (backing_object == NULL || 866 (backing_object->flags & OBJ_ANON) == 0) 867 return (NULL); 868 VM_OBJECT_WLOCK(backing_object); 869 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 870 break; 871 VM_OBJECT_WUNLOCK(object); 872 vm_object_pip_sleep(backing_object, "vmbckwait"); 873 counter_u64_add(object_collapse_waits, 1); 874 VM_OBJECT_WLOCK(object); 875 } 876 return (backing_object); 877 } 878 879 /* 880 * vm_object_terminate_pages removes any remaining pageable pages 881 * from the object and resets the object to an empty state. 882 */ 883 static void 884 vm_object_terminate_pages(vm_object_t object) 885 { 886 vm_page_t p, p_next; 887 888 VM_OBJECT_ASSERT_WLOCKED(object); 889 890 /* 891 * Free any remaining pageable pages. This also removes them from the 892 * paging queues. However, don't free wired pages, just remove them 893 * from the object. Rather than incrementally removing each page from 894 * the object, the page and object are reset to any empty state. 895 */ 896 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { 897 vm_page_assert_unbusied(p); 898 KASSERT(p->object == object && 899 (p->ref_count & VPRC_OBJREF) != 0, 900 ("vm_object_terminate_pages: page %p is inconsistent", p)); 901 902 p->object = NULL; 903 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 904 VM_CNT_INC(v_pfree); 905 vm_page_free(p); 906 } 907 } 908 909 /* 910 * If the object contained any pages, then reset it to an empty state. 911 * None of the object's fields, including "resident_page_count", were 912 * modified by the preceding loop. 913 */ 914 if (object->resident_page_count != 0) { 915 vm_radix_reclaim_allnodes(&object->rtree); 916 TAILQ_INIT(&object->memq); 917 object->resident_page_count = 0; 918 if (object->type == OBJT_VNODE) 919 vdrop(object->handle); 920 } 921 } 922 923 /* 924 * vm_object_terminate actually destroys the specified object, freeing 925 * up all previously used resources. 926 * 927 * The object must be locked. 928 * This routine may block. 929 */ 930 void 931 vm_object_terminate(vm_object_t object) 932 { 933 934 VM_OBJECT_ASSERT_WLOCKED(object); 935 KASSERT((object->flags & OBJ_DEAD) != 0, 936 ("terminating non-dead obj %p", object)); 937 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 938 ("terminating collapsing obj %p", object)); 939 KASSERT(object->backing_object == NULL, 940 ("terminating shadow obj %p", object)); 941 942 /* 943 * Wait for the pageout daemon and other current users to be 944 * done with the object. Note that new paging_in_progress 945 * users can come after this wait, but they must check 946 * OBJ_DEAD flag set (without unlocking the object), and avoid 947 * the object being terminated. 948 */ 949 vm_object_pip_wait(object, "objtrm"); 950 951 KASSERT(object->ref_count == 0, 952 ("vm_object_terminate: object with references, ref_count=%d", 953 object->ref_count)); 954 955 if ((object->flags & OBJ_PG_DTOR) == 0) 956 vm_object_terminate_pages(object); 957 958 #if VM_NRESERVLEVEL > 0 959 if (__predict_false(!LIST_EMPTY(&object->rvq))) 960 vm_reserv_break_all(object); 961 #endif 962 963 KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0, 964 ("%s: non-swap obj %p has cred", __func__, object)); 965 966 /* 967 * Let the pager know object is dead. 968 */ 969 vm_pager_deallocate(object); 970 VM_OBJECT_WUNLOCK(object); 971 972 vm_object_destroy(object); 973 } 974 975 /* 976 * Make the page read-only so that we can clear the object flags. However, if 977 * this is a nosync mmap then the object is likely to stay dirty so do not 978 * mess with the page and do not clear the object flags. Returns TRUE if the 979 * page should be flushed, and FALSE otherwise. 980 */ 981 static boolean_t 982 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) 983 { 984 985 vm_page_assert_busied(p); 986 987 /* 988 * If we have been asked to skip nosync pages and this is a 989 * nosync page, skip it. Note that the object flags were not 990 * cleared in this case so we do not have to set them. 991 */ 992 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { 993 *allclean = FALSE; 994 return (FALSE); 995 } else { 996 pmap_remove_write(p); 997 return (p->dirty != 0); 998 } 999 } 1000 1001 /* 1002 * vm_object_page_clean 1003 * 1004 * Clean all dirty pages in the specified range of object. Leaves page 1005 * on whatever queue it is currently on. If NOSYNC is set then do not 1006 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1007 * leaving the object dirty. 1008 * 1009 * For swap objects backing tmpfs regular files, do not flush anything, 1010 * but remove write protection on the mapped pages to update mtime through 1011 * mmaped writes. 1012 * 1013 * When stuffing pages asynchronously, allow clustering. XXX we need a 1014 * synchronous clustering mode implementation. 1015 * 1016 * Odd semantics: if start == end, we clean everything. 1017 * 1018 * The object must be locked. 1019 * 1020 * Returns FALSE if some page from the range was not written, as 1021 * reported by the pager, and TRUE otherwise. 1022 */ 1023 boolean_t 1024 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1025 int flags) 1026 { 1027 vm_page_t np, p; 1028 vm_pindex_t pi, tend, tstart; 1029 int curgeneration, n, pagerflags; 1030 boolean_t eio, res, allclean; 1031 1032 VM_OBJECT_ASSERT_WLOCKED(object); 1033 1034 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1035 return (TRUE); 1036 1037 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1038 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1039 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1040 1041 tstart = OFF_TO_IDX(start); 1042 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1043 allclean = tstart == 0 && tend >= object->size; 1044 res = TRUE; 1045 1046 rescan: 1047 curgeneration = object->generation; 1048 1049 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { 1050 pi = p->pindex; 1051 if (pi >= tend) 1052 break; 1053 np = TAILQ_NEXT(p, listq); 1054 if (vm_page_none_valid(p)) 1055 continue; 1056 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { 1057 if (object->generation != curgeneration && 1058 (flags & OBJPC_SYNC) != 0) 1059 goto rescan; 1060 np = vm_page_find_least(object, pi); 1061 continue; 1062 } 1063 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1064 vm_page_xunbusy(p); 1065 continue; 1066 } 1067 if (object->type == OBJT_VNODE) { 1068 n = vm_object_page_collect_flush(object, p, pagerflags, 1069 flags, &allclean, &eio); 1070 if (eio) { 1071 res = FALSE; 1072 allclean = FALSE; 1073 } 1074 if (object->generation != curgeneration && 1075 (flags & OBJPC_SYNC) != 0) 1076 goto rescan; 1077 1078 /* 1079 * If the VOP_PUTPAGES() did a truncated write, so 1080 * that even the first page of the run is not fully 1081 * written, vm_pageout_flush() returns 0 as the run 1082 * length. Since the condition that caused truncated 1083 * write may be permanent, e.g. exhausted free space, 1084 * accepting n == 0 would cause an infinite loop. 1085 * 1086 * Forwarding the iterator leaves the unwritten page 1087 * behind, but there is not much we can do there if 1088 * filesystem refuses to write it. 1089 */ 1090 if (n == 0) { 1091 n = 1; 1092 allclean = FALSE; 1093 } 1094 } else { 1095 n = 1; 1096 vm_page_xunbusy(p); 1097 } 1098 np = vm_page_find_least(object, pi + n); 1099 } 1100 #if 0 1101 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1102 #endif 1103 1104 /* 1105 * Leave updating cleangeneration for tmpfs objects to tmpfs 1106 * scan. It needs to update mtime, which happens for other 1107 * filesystems during page writeouts. 1108 */ 1109 if (allclean && object->type == OBJT_VNODE) 1110 object->cleangeneration = curgeneration; 1111 return (res); 1112 } 1113 1114 static int 1115 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, 1116 int flags, boolean_t *allclean, boolean_t *eio) 1117 { 1118 vm_page_t ma[2 * vm_pageout_page_count], tp; 1119 int count, mreq, runlen; 1120 1121 vm_page_lock_assert(p, MA_NOTOWNED); 1122 vm_page_assert_xbusied(p); 1123 VM_OBJECT_ASSERT_WLOCKED(object); 1124 ma[vm_pageout_page_count] = p; 1125 for (count = 1, tp = p; count < vm_pageout_page_count; count++) { 1126 tp = vm_page_next(tp); 1127 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1128 break; 1129 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1130 vm_page_xunbusy(tp); 1131 break; 1132 } 1133 ma[vm_pageout_page_count + count] = tp; 1134 } 1135 1136 for (mreq = 0, tp = p; count < vm_pageout_page_count; count++, mreq++) { 1137 tp = vm_page_prev(tp); 1138 if (tp == NULL || vm_page_tryxbusy(tp) == 0) 1139 break; 1140 if (!vm_object_page_remove_write(tp, flags, allclean)) { 1141 vm_page_xunbusy(tp); 1142 break; 1143 } 1144 ma[vm_pageout_page_count - 1 - mreq] = tp; 1145 } 1146 1147 vm_pageout_flush(&ma[vm_pageout_page_count - mreq], count, pagerflags, 1148 mreq, &runlen, eio); 1149 return (runlen); 1150 } 1151 1152 /* 1153 * Note that there is absolutely no sense in writing out 1154 * anonymous objects, so we track down the vnode object 1155 * to write out. 1156 * We invalidate (remove) all pages from the address space 1157 * for semantic correctness. 1158 * 1159 * If the backing object is a device object with unmanaged pages, then any 1160 * mappings to the specified range of pages must be removed before this 1161 * function is called. 1162 * 1163 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1164 * may start out with a NULL object. 1165 */ 1166 boolean_t 1167 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1168 boolean_t syncio, boolean_t invalidate) 1169 { 1170 vm_object_t backing_object; 1171 struct vnode *vp; 1172 struct mount *mp; 1173 int error, flags, fsync_after; 1174 boolean_t res; 1175 1176 if (object == NULL) 1177 return (TRUE); 1178 res = TRUE; 1179 error = 0; 1180 VM_OBJECT_WLOCK(object); 1181 while ((backing_object = object->backing_object) != NULL) { 1182 VM_OBJECT_WLOCK(backing_object); 1183 offset += object->backing_object_offset; 1184 VM_OBJECT_WUNLOCK(object); 1185 object = backing_object; 1186 if (object->size < OFF_TO_IDX(offset + size)) 1187 size = IDX_TO_OFF(object->size) - offset; 1188 } 1189 /* 1190 * Flush pages if writing is allowed, invalidate them 1191 * if invalidation requested. Pages undergoing I/O 1192 * will be ignored by vm_object_page_remove(). 1193 * 1194 * We cannot lock the vnode and then wait for paging 1195 * to complete without deadlocking against vm_fault. 1196 * Instead we simply call vm_object_page_remove() and 1197 * allow it to block internally on a page-by-page 1198 * basis when it encounters pages undergoing async 1199 * I/O. 1200 */ 1201 if (object->type == OBJT_VNODE && 1202 vm_object_mightbedirty(object) != 0 && 1203 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1204 VM_OBJECT_WUNLOCK(object); 1205 (void)vn_start_write(vp, &mp, V_WAIT); 1206 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1207 if (syncio && !invalidate && offset == 0 && 1208 atop(size) == object->size) { 1209 /* 1210 * If syncing the whole mapping of the file, 1211 * it is faster to schedule all the writes in 1212 * async mode, also allowing the clustering, 1213 * and then wait for i/o to complete. 1214 */ 1215 flags = 0; 1216 fsync_after = TRUE; 1217 } else { 1218 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1219 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1220 fsync_after = FALSE; 1221 } 1222 VM_OBJECT_WLOCK(object); 1223 res = vm_object_page_clean(object, offset, offset + size, 1224 flags); 1225 VM_OBJECT_WUNLOCK(object); 1226 if (fsync_after) { 1227 for (;;) { 1228 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1229 if (error != ERELOOKUP) 1230 break; 1231 1232 /* 1233 * Allow SU/bufdaemon to handle more 1234 * dependencies in the meantime. 1235 */ 1236 VOP_UNLOCK(vp); 1237 vn_finished_write(mp); 1238 1239 (void)vn_start_write(vp, &mp, V_WAIT); 1240 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1241 } 1242 } 1243 VOP_UNLOCK(vp); 1244 vn_finished_write(mp); 1245 if (error != 0) 1246 res = FALSE; 1247 VM_OBJECT_WLOCK(object); 1248 } 1249 if ((object->type == OBJT_VNODE || 1250 object->type == OBJT_DEVICE) && invalidate) { 1251 if (object->type == OBJT_DEVICE) 1252 /* 1253 * The option OBJPR_NOTMAPPED must be passed here 1254 * because vm_object_page_remove() cannot remove 1255 * unmanaged mappings. 1256 */ 1257 flags = OBJPR_NOTMAPPED; 1258 else if (old_msync) 1259 flags = 0; 1260 else 1261 flags = OBJPR_CLEANONLY; 1262 vm_object_page_remove(object, OFF_TO_IDX(offset), 1263 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1264 } 1265 VM_OBJECT_WUNLOCK(object); 1266 return (res); 1267 } 1268 1269 /* 1270 * Determine whether the given advice can be applied to the object. Advice is 1271 * not applied to unmanaged pages since they never belong to page queues, and 1272 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1273 * have been mapped at most once. 1274 */ 1275 static bool 1276 vm_object_advice_applies(vm_object_t object, int advice) 1277 { 1278 1279 if ((object->flags & OBJ_UNMANAGED) != 0) 1280 return (false); 1281 if (advice != MADV_FREE) 1282 return (true); 1283 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1284 (OBJ_ONEMAPPING | OBJ_ANON)); 1285 } 1286 1287 static void 1288 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1289 vm_size_t size) 1290 { 1291 1292 if (advice == MADV_FREE) 1293 vm_pager_freespace(object, pindex, size); 1294 } 1295 1296 /* 1297 * vm_object_madvise: 1298 * 1299 * Implements the madvise function at the object/page level. 1300 * 1301 * MADV_WILLNEED (any object) 1302 * 1303 * Activate the specified pages if they are resident. 1304 * 1305 * MADV_DONTNEED (any object) 1306 * 1307 * Deactivate the specified pages if they are resident. 1308 * 1309 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only) 1310 * 1311 * Deactivate and clean the specified pages if they are 1312 * resident. This permits the process to reuse the pages 1313 * without faulting or the kernel to reclaim the pages 1314 * without I/O. 1315 */ 1316 void 1317 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1318 int advice) 1319 { 1320 vm_pindex_t tpindex; 1321 vm_object_t backing_object, tobject; 1322 vm_page_t m, tm; 1323 1324 if (object == NULL) 1325 return; 1326 1327 relookup: 1328 VM_OBJECT_WLOCK(object); 1329 if (!vm_object_advice_applies(object, advice)) { 1330 VM_OBJECT_WUNLOCK(object); 1331 return; 1332 } 1333 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1334 tobject = object; 1335 1336 /* 1337 * If the next page isn't resident in the top-level object, we 1338 * need to search the shadow chain. When applying MADV_FREE, we 1339 * take care to release any swap space used to store 1340 * non-resident pages. 1341 */ 1342 if (m == NULL || pindex < m->pindex) { 1343 /* 1344 * Optimize a common case: if the top-level object has 1345 * no backing object, we can skip over the non-resident 1346 * range in constant time. 1347 */ 1348 if (object->backing_object == NULL) { 1349 tpindex = (m != NULL && m->pindex < end) ? 1350 m->pindex : end; 1351 vm_object_madvise_freespace(object, advice, 1352 pindex, tpindex - pindex); 1353 if ((pindex = tpindex) == end) 1354 break; 1355 goto next_page; 1356 } 1357 1358 tpindex = pindex; 1359 do { 1360 vm_object_madvise_freespace(tobject, advice, 1361 tpindex, 1); 1362 /* 1363 * Prepare to search the next object in the 1364 * chain. 1365 */ 1366 backing_object = tobject->backing_object; 1367 if (backing_object == NULL) 1368 goto next_pindex; 1369 VM_OBJECT_WLOCK(backing_object); 1370 tpindex += 1371 OFF_TO_IDX(tobject->backing_object_offset); 1372 if (tobject != object) 1373 VM_OBJECT_WUNLOCK(tobject); 1374 tobject = backing_object; 1375 if (!vm_object_advice_applies(tobject, advice)) 1376 goto next_pindex; 1377 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1378 NULL); 1379 } else { 1380 next_page: 1381 tm = m; 1382 m = TAILQ_NEXT(m, listq); 1383 } 1384 1385 /* 1386 * If the page is not in a normal state, skip it. The page 1387 * can not be invalidated while the object lock is held. 1388 */ 1389 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1390 goto next_pindex; 1391 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1392 ("vm_object_madvise: page %p is fictitious", tm)); 1393 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1394 ("vm_object_madvise: page %p is not managed", tm)); 1395 if (vm_page_tryxbusy(tm) == 0) { 1396 if (object != tobject) 1397 VM_OBJECT_WUNLOCK(object); 1398 if (advice == MADV_WILLNEED) { 1399 /* 1400 * Reference the page before unlocking and 1401 * sleeping so that the page daemon is less 1402 * likely to reclaim it. 1403 */ 1404 vm_page_aflag_set(tm, PGA_REFERENCED); 1405 } 1406 if (!vm_page_busy_sleep(tm, "madvpo", 0)) 1407 VM_OBJECT_WUNLOCK(tobject); 1408 goto relookup; 1409 } 1410 vm_page_advise(tm, advice); 1411 vm_page_xunbusy(tm); 1412 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1413 next_pindex: 1414 if (tobject != object) 1415 VM_OBJECT_WUNLOCK(tobject); 1416 } 1417 VM_OBJECT_WUNLOCK(object); 1418 } 1419 1420 /* 1421 * vm_object_shadow: 1422 * 1423 * Create a new object which is backed by the 1424 * specified existing object range. The source 1425 * object reference is deallocated. 1426 * 1427 * The new object and offset into that object 1428 * are returned in the source parameters. 1429 */ 1430 void 1431 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1432 struct ucred *cred, bool shared) 1433 { 1434 vm_object_t source; 1435 vm_object_t result; 1436 1437 source = *object; 1438 1439 /* 1440 * Don't create the new object if the old object isn't shared. 1441 * 1442 * If we hold the only reference we can guarantee that it won't 1443 * increase while we have the map locked. Otherwise the race is 1444 * harmless and we will end up with an extra shadow object that 1445 * will be collapsed later. 1446 */ 1447 if (source != NULL && source->ref_count == 1 && 1448 (source->flags & OBJ_ANON) != 0) 1449 return; 1450 1451 /* 1452 * Allocate a new object with the given length. 1453 */ 1454 result = vm_object_allocate_anon(atop(length), source, cred, length); 1455 1456 /* 1457 * Store the offset into the source object, and fix up the offset into 1458 * the new object. 1459 */ 1460 result->backing_object_offset = *offset; 1461 1462 if (shared || source != NULL) { 1463 VM_OBJECT_WLOCK(result); 1464 1465 /* 1466 * The new object shadows the source object, adding a 1467 * reference to it. Our caller changes his reference 1468 * to point to the new object, removing a reference to 1469 * the source object. Net result: no change of 1470 * reference count, unless the caller needs to add one 1471 * more reference due to forking a shared map entry. 1472 */ 1473 if (shared) { 1474 vm_object_reference_locked(result); 1475 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1476 } 1477 1478 /* 1479 * Try to optimize the result object's page color when 1480 * shadowing in order to maintain page coloring 1481 * consistency in the combined shadowed object. 1482 */ 1483 if (source != NULL) { 1484 vm_object_backing_insert(result, source); 1485 result->domain = source->domain; 1486 #if VM_NRESERVLEVEL > 0 1487 vm_object_set_flag(result, 1488 (source->flags & OBJ_COLORED)); 1489 result->pg_color = (source->pg_color + 1490 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1491 1)) - 1); 1492 #endif 1493 } 1494 VM_OBJECT_WUNLOCK(result); 1495 } 1496 1497 /* 1498 * Return the new things 1499 */ 1500 *offset = 0; 1501 *object = result; 1502 } 1503 1504 /* 1505 * vm_object_split: 1506 * 1507 * Split the pages in a map entry into a new object. This affords 1508 * easier removal of unused pages, and keeps object inheritance from 1509 * being a negative impact on memory usage. 1510 */ 1511 void 1512 vm_object_split(vm_map_entry_t entry) 1513 { 1514 vm_page_t m, m_next; 1515 vm_object_t orig_object, new_object, backing_object; 1516 vm_pindex_t idx, offidxstart; 1517 vm_size_t size; 1518 1519 orig_object = entry->object.vm_object; 1520 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1521 ("vm_object_split: Splitting object with multiple mappings.")); 1522 if ((orig_object->flags & OBJ_ANON) == 0) 1523 return; 1524 if (orig_object->ref_count <= 1) 1525 return; 1526 VM_OBJECT_WUNLOCK(orig_object); 1527 1528 offidxstart = OFF_TO_IDX(entry->offset); 1529 size = atop(entry->end - entry->start); 1530 1531 new_object = vm_object_allocate_anon(size, orig_object, 1532 orig_object->cred, ptoa(size)); 1533 1534 /* 1535 * We must wait for the orig_object to complete any in-progress 1536 * collapse so that the swap blocks are stable below. The 1537 * additional reference on backing_object by new object will 1538 * prevent further collapse operations until split completes. 1539 */ 1540 VM_OBJECT_WLOCK(orig_object); 1541 vm_object_collapse_wait(orig_object); 1542 1543 /* 1544 * At this point, the new object is still private, so the order in 1545 * which the original and new objects are locked does not matter. 1546 */ 1547 VM_OBJECT_WLOCK(new_object); 1548 new_object->domain = orig_object->domain; 1549 backing_object = orig_object->backing_object; 1550 if (backing_object != NULL) { 1551 vm_object_backing_insert_ref(new_object, backing_object); 1552 new_object->backing_object_offset = 1553 orig_object->backing_object_offset + entry->offset; 1554 } 1555 if (orig_object->cred != NULL) { 1556 crhold(orig_object->cred); 1557 KASSERT(orig_object->charge >= ptoa(size), 1558 ("orig_object->charge < 0")); 1559 orig_object->charge -= ptoa(size); 1560 } 1561 1562 /* 1563 * Mark the split operation so that swap_pager_getpages() knows 1564 * that the object is in transition. 1565 */ 1566 vm_object_set_flag(orig_object, OBJ_SPLIT); 1567 #ifdef INVARIANTS 1568 idx = 0; 1569 #endif 1570 retry: 1571 m = vm_page_find_least(orig_object, offidxstart); 1572 KASSERT(m == NULL || idx <= m->pindex - offidxstart, 1573 ("%s: object %p was repopulated", __func__, orig_object)); 1574 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1575 m = m_next) { 1576 m_next = TAILQ_NEXT(m, listq); 1577 1578 /* 1579 * We must wait for pending I/O to complete before we can 1580 * rename the page. 1581 * 1582 * We do not have to VM_PROT_NONE the page as mappings should 1583 * not be changed by this operation. 1584 */ 1585 if (vm_page_tryxbusy(m) == 0) { 1586 VM_OBJECT_WUNLOCK(new_object); 1587 if (vm_page_busy_sleep(m, "spltwt", 0)) 1588 VM_OBJECT_WLOCK(orig_object); 1589 VM_OBJECT_WLOCK(new_object); 1590 goto retry; 1591 } 1592 1593 /* 1594 * The page was left invalid. Likely placed there by 1595 * an incomplete fault. Just remove and ignore. 1596 */ 1597 if (vm_page_none_valid(m)) { 1598 if (vm_page_remove(m)) 1599 vm_page_free(m); 1600 continue; 1601 } 1602 1603 /* vm_page_rename() will dirty the page. */ 1604 if (vm_page_rename(m, new_object, idx)) { 1605 vm_page_xunbusy(m); 1606 VM_OBJECT_WUNLOCK(new_object); 1607 VM_OBJECT_WUNLOCK(orig_object); 1608 vm_radix_wait(); 1609 VM_OBJECT_WLOCK(orig_object); 1610 VM_OBJECT_WLOCK(new_object); 1611 goto retry; 1612 } 1613 1614 #if VM_NRESERVLEVEL > 0 1615 /* 1616 * If some of the reservation's allocated pages remain with 1617 * the original object, then transferring the reservation to 1618 * the new object is neither particularly beneficial nor 1619 * particularly harmful as compared to leaving the reservation 1620 * with the original object. If, however, all of the 1621 * reservation's allocated pages are transferred to the new 1622 * object, then transferring the reservation is typically 1623 * beneficial. Determining which of these two cases applies 1624 * would be more costly than unconditionally renaming the 1625 * reservation. 1626 */ 1627 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1628 #endif 1629 } 1630 1631 /* 1632 * swap_pager_copy() can sleep, in which case the orig_object's 1633 * and new_object's locks are released and reacquired. 1634 */ 1635 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1636 1637 TAILQ_FOREACH(m, &new_object->memq, listq) 1638 vm_page_xunbusy(m); 1639 1640 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1641 VM_OBJECT_WUNLOCK(orig_object); 1642 VM_OBJECT_WUNLOCK(new_object); 1643 entry->object.vm_object = new_object; 1644 entry->offset = 0LL; 1645 vm_object_deallocate(orig_object); 1646 VM_OBJECT_WLOCK(new_object); 1647 } 1648 1649 static vm_page_t 1650 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p) 1651 { 1652 vm_object_t backing_object; 1653 1654 VM_OBJECT_ASSERT_WLOCKED(object); 1655 backing_object = object->backing_object; 1656 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1657 1658 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1659 ("invalid ownership %p %p %p", p, object, backing_object)); 1660 /* The page is only NULL when rename fails. */ 1661 if (p == NULL) { 1662 VM_OBJECT_WUNLOCK(object); 1663 VM_OBJECT_WUNLOCK(backing_object); 1664 vm_radix_wait(); 1665 VM_OBJECT_WLOCK(object); 1666 } else if (p->object == object) { 1667 VM_OBJECT_WUNLOCK(backing_object); 1668 if (vm_page_busy_sleep(p, "vmocol", 0)) 1669 VM_OBJECT_WLOCK(object); 1670 } else { 1671 VM_OBJECT_WUNLOCK(object); 1672 if (!vm_page_busy_sleep(p, "vmocol", 0)) 1673 VM_OBJECT_WUNLOCK(backing_object); 1674 VM_OBJECT_WLOCK(object); 1675 } 1676 VM_OBJECT_WLOCK(backing_object); 1677 return (TAILQ_FIRST(&backing_object->memq)); 1678 } 1679 1680 static bool 1681 vm_object_scan_all_shadowed(vm_object_t object) 1682 { 1683 vm_object_t backing_object; 1684 vm_page_t p, pp; 1685 vm_pindex_t backing_offset_index, new_pindex, pi, ps; 1686 1687 VM_OBJECT_ASSERT_WLOCKED(object); 1688 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1689 1690 backing_object = object->backing_object; 1691 1692 if ((backing_object->flags & OBJ_ANON) == 0) 1693 return (false); 1694 1695 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1696 p = vm_page_find_least(backing_object, pi); 1697 ps = swap_pager_find_least(backing_object, pi); 1698 1699 /* 1700 * Only check pages inside the parent object's range and 1701 * inside the parent object's mapping of the backing object. 1702 */ 1703 for (;; pi++) { 1704 if (p != NULL && p->pindex < pi) 1705 p = TAILQ_NEXT(p, listq); 1706 if (ps < pi) 1707 ps = swap_pager_find_least(backing_object, pi); 1708 if (p == NULL && ps >= backing_object->size) 1709 break; 1710 else if (p == NULL) 1711 pi = ps; 1712 else 1713 pi = MIN(p->pindex, ps); 1714 1715 new_pindex = pi - backing_offset_index; 1716 if (new_pindex >= object->size) 1717 break; 1718 1719 if (p != NULL) { 1720 /* 1721 * If the backing object page is busy a 1722 * grandparent or older page may still be 1723 * undergoing CoW. It is not safe to collapse 1724 * the backing object until it is quiesced. 1725 */ 1726 if (vm_page_tryxbusy(p) == 0) 1727 return (false); 1728 1729 /* 1730 * We raced with the fault handler that left 1731 * newly allocated invalid page on the object 1732 * queue and retried. 1733 */ 1734 if (!vm_page_all_valid(p)) 1735 goto unbusy_ret; 1736 } 1737 1738 /* 1739 * See if the parent has the page or if the parent's object 1740 * pager has the page. If the parent has the page but the page 1741 * is not valid, the parent's object pager must have the page. 1742 * 1743 * If this fails, the parent does not completely shadow the 1744 * object and we might as well give up now. 1745 */ 1746 pp = vm_page_lookup(object, new_pindex); 1747 1748 /* 1749 * The valid check here is stable due to object lock 1750 * being required to clear valid and initiate paging. 1751 * Busy of p disallows fault handler to validate pp. 1752 */ 1753 if ((pp == NULL || vm_page_none_valid(pp)) && 1754 !vm_pager_has_page(object, new_pindex, NULL, NULL)) 1755 goto unbusy_ret; 1756 if (p != NULL) 1757 vm_page_xunbusy(p); 1758 } 1759 return (true); 1760 1761 unbusy_ret: 1762 if (p != NULL) 1763 vm_page_xunbusy(p); 1764 return (false); 1765 } 1766 1767 static void 1768 vm_object_collapse_scan(vm_object_t object) 1769 { 1770 vm_object_t backing_object; 1771 vm_page_t next, p, pp; 1772 vm_pindex_t backing_offset_index, new_pindex; 1773 1774 VM_OBJECT_ASSERT_WLOCKED(object); 1775 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1776 1777 backing_object = object->backing_object; 1778 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1779 1780 /* 1781 * Our scan 1782 */ 1783 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { 1784 next = TAILQ_NEXT(p, listq); 1785 new_pindex = p->pindex - backing_offset_index; 1786 1787 /* 1788 * Check for busy page 1789 */ 1790 if (vm_page_tryxbusy(p) == 0) { 1791 next = vm_object_collapse_scan_wait(object, p); 1792 continue; 1793 } 1794 1795 KASSERT(object->backing_object == backing_object, 1796 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1797 object->backing_object, backing_object)); 1798 KASSERT(p->object == backing_object, 1799 ("vm_object_collapse_scan: object mismatch %p != %p", 1800 p->object, backing_object)); 1801 1802 if (p->pindex < backing_offset_index || 1803 new_pindex >= object->size) { 1804 vm_pager_freespace(backing_object, p->pindex, 1); 1805 1806 KASSERT(!pmap_page_is_mapped(p), 1807 ("freeing mapped page %p", p)); 1808 if (vm_page_remove(p)) 1809 vm_page_free(p); 1810 continue; 1811 } 1812 1813 if (!vm_page_all_valid(p)) { 1814 KASSERT(!pmap_page_is_mapped(p), 1815 ("freeing mapped page %p", p)); 1816 if (vm_page_remove(p)) 1817 vm_page_free(p); 1818 continue; 1819 } 1820 1821 pp = vm_page_lookup(object, new_pindex); 1822 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1823 vm_page_xunbusy(p); 1824 /* 1825 * The page in the parent is busy and possibly not 1826 * (yet) valid. Until its state is finalized by the 1827 * busy bit owner, we can't tell whether it shadows the 1828 * original page. 1829 */ 1830 next = vm_object_collapse_scan_wait(object, pp); 1831 continue; 1832 } 1833 1834 if (pp != NULL && vm_page_none_valid(pp)) { 1835 /* 1836 * The page was invalid in the parent. Likely placed 1837 * there by an incomplete fault. Just remove and 1838 * ignore. p can replace it. 1839 */ 1840 if (vm_page_remove(pp)) 1841 vm_page_free(pp); 1842 pp = NULL; 1843 } 1844 1845 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1846 NULL)) { 1847 /* 1848 * The page already exists in the parent OR swap exists 1849 * for this location in the parent. Leave the parent's 1850 * page alone. Destroy the original page from the 1851 * backing object. 1852 */ 1853 vm_pager_freespace(backing_object, p->pindex, 1); 1854 KASSERT(!pmap_page_is_mapped(p), 1855 ("freeing mapped page %p", p)); 1856 if (vm_page_remove(p)) 1857 vm_page_free(p); 1858 if (pp != NULL) 1859 vm_page_xunbusy(pp); 1860 continue; 1861 } 1862 1863 /* 1864 * Page does not exist in parent, rename the page from the 1865 * backing object to the main object. 1866 * 1867 * If the page was mapped to a process, it can remain mapped 1868 * through the rename. vm_page_rename() will dirty the page. 1869 */ 1870 if (vm_page_rename(p, object, new_pindex)) { 1871 vm_page_xunbusy(p); 1872 next = vm_object_collapse_scan_wait(object, NULL); 1873 continue; 1874 } 1875 1876 /* Use the old pindex to free the right page. */ 1877 vm_pager_freespace(backing_object, new_pindex + 1878 backing_offset_index, 1); 1879 1880 #if VM_NRESERVLEVEL > 0 1881 /* 1882 * Rename the reservation. 1883 */ 1884 vm_reserv_rename(p, object, backing_object, 1885 backing_offset_index); 1886 #endif 1887 vm_page_xunbusy(p); 1888 } 1889 return; 1890 } 1891 1892 /* 1893 * vm_object_collapse: 1894 * 1895 * Collapse an object with the object backing it. 1896 * Pages in the backing object are moved into the 1897 * parent, and the backing object is deallocated. 1898 */ 1899 void 1900 vm_object_collapse(vm_object_t object) 1901 { 1902 vm_object_t backing_object, new_backing_object; 1903 1904 VM_OBJECT_ASSERT_WLOCKED(object); 1905 1906 while (TRUE) { 1907 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1908 ("collapsing invalid object")); 1909 1910 /* 1911 * Wait for the backing_object to finish any pending 1912 * collapse so that the caller sees the shortest possible 1913 * shadow chain. 1914 */ 1915 backing_object = vm_object_backing_collapse_wait(object); 1916 if (backing_object == NULL) 1917 return; 1918 1919 KASSERT(object->ref_count > 0 && 1920 object->ref_count > atomic_load_int(&object->shadow_count), 1921 ("collapse with invalid ref %d or shadow %d count.", 1922 object->ref_count, atomic_load_int(&object->shadow_count))); 1923 KASSERT((backing_object->flags & 1924 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1925 ("vm_object_collapse: Backing object already collapsing.")); 1926 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1927 ("vm_object_collapse: object is already collapsing.")); 1928 1929 /* 1930 * We know that we can either collapse the backing object if 1931 * the parent is the only reference to it, or (perhaps) have 1932 * the parent bypass the object if the parent happens to shadow 1933 * all the resident pages in the entire backing object. 1934 */ 1935 if (backing_object->ref_count == 1) { 1936 KASSERT(atomic_load_int(&backing_object->shadow_count) 1937 == 1, 1938 ("vm_object_collapse: shadow_count: %d", 1939 atomic_load_int(&backing_object->shadow_count))); 1940 vm_object_pip_add(object, 1); 1941 vm_object_set_flag(object, OBJ_COLLAPSING); 1942 vm_object_pip_add(backing_object, 1); 1943 vm_object_set_flag(backing_object, OBJ_DEAD); 1944 1945 /* 1946 * If there is exactly one reference to the backing 1947 * object, we can collapse it into the parent. 1948 */ 1949 vm_object_collapse_scan(object); 1950 1951 /* 1952 * Move the pager from backing_object to object. 1953 * 1954 * swap_pager_copy() can sleep, in which case the 1955 * backing_object's and object's locks are released and 1956 * reacquired. 1957 */ 1958 swap_pager_copy(backing_object, object, 1959 OFF_TO_IDX(object->backing_object_offset), TRUE); 1960 1961 /* 1962 * Object now shadows whatever backing_object did. 1963 */ 1964 vm_object_clear_flag(object, OBJ_COLLAPSING); 1965 vm_object_backing_transfer(object, backing_object); 1966 object->backing_object_offset += 1967 backing_object->backing_object_offset; 1968 VM_OBJECT_WUNLOCK(object); 1969 vm_object_pip_wakeup(object); 1970 1971 /* 1972 * Discard backing_object. 1973 * 1974 * Since the backing object has no pages, no pager left, 1975 * and no object references within it, all that is 1976 * necessary is to dispose of it. 1977 */ 1978 KASSERT(backing_object->ref_count == 1, ( 1979 "backing_object %p was somehow re-referenced during collapse!", 1980 backing_object)); 1981 vm_object_pip_wakeup(backing_object); 1982 (void)refcount_release(&backing_object->ref_count); 1983 umtx_shm_object_terminated(backing_object); 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 * Skip invalid pages if asked to do so. Try to avoid acquiring 2083 * the busy lock, as some consumers rely on this to avoid 2084 * deadlocks. 2085 * 2086 * A thread may concurrently transition the page from invalid to 2087 * valid using only the busy lock, so the result of this check 2088 * is immediately stale. It is up to consumers to handle this, 2089 * for instance by ensuring that all invalid->valid transitions 2090 * happen with a mutex held, as may be possible for a 2091 * filesystem. 2092 */ 2093 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) 2094 continue; 2095 2096 /* 2097 * If the page is wired for any reason besides the existence 2098 * of managed, wired mappings, then it cannot be freed. For 2099 * example, fictitious pages, which represent device memory, 2100 * are inherently wired and cannot be freed. They can, 2101 * however, be invalidated if the option OBJPR_CLEANONLY is 2102 * not specified. 2103 */ 2104 if (vm_page_tryxbusy(p) == 0) { 2105 if (vm_page_busy_sleep(p, "vmopar", 0)) 2106 VM_OBJECT_WLOCK(object); 2107 goto again; 2108 } 2109 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) { 2110 vm_page_xunbusy(p); 2111 continue; 2112 } 2113 if (vm_page_wired(p)) { 2114 wired: 2115 if ((options & OBJPR_NOTMAPPED) == 0 && 2116 object->ref_count != 0) 2117 pmap_remove_all(p); 2118 if ((options & OBJPR_CLEANONLY) == 0) { 2119 vm_page_invalid(p); 2120 vm_page_undirty(p); 2121 } 2122 vm_page_xunbusy(p); 2123 continue; 2124 } 2125 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2126 ("vm_object_page_remove: page %p is fictitious", p)); 2127 if ((options & OBJPR_CLEANONLY) != 0 && 2128 !vm_page_none_valid(p)) { 2129 if ((options & OBJPR_NOTMAPPED) == 0 && 2130 object->ref_count != 0 && 2131 !vm_page_try_remove_write(p)) 2132 goto wired; 2133 if (p->dirty != 0) { 2134 vm_page_xunbusy(p); 2135 continue; 2136 } 2137 } 2138 if ((options & OBJPR_NOTMAPPED) == 0 && 2139 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2140 goto wired; 2141 vm_page_free(p); 2142 } 2143 vm_object_pip_wakeup(object); 2144 2145 vm_pager_freespace(object, start, (end == 0 ? object->size : end) - 2146 start); 2147 } 2148 2149 /* 2150 * vm_object_page_noreuse: 2151 * 2152 * For the given object, attempt to move the specified pages to 2153 * the head of the inactive queue. This bypasses regular LRU 2154 * operation and allows the pages to be reused quickly under memory 2155 * pressure. If a page is wired for any reason, then it will not 2156 * be queued. Pages are specified by the range ["start", "end"). 2157 * As a special case, if "end" is zero, then the range extends from 2158 * "start" to the end of the object. 2159 * 2160 * This operation should only be performed on objects that 2161 * contain non-fictitious, managed pages. 2162 * 2163 * The object must be locked. 2164 */ 2165 void 2166 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2167 { 2168 vm_page_t p, next; 2169 2170 VM_OBJECT_ASSERT_LOCKED(object); 2171 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2172 ("vm_object_page_noreuse: illegal object %p", object)); 2173 if (object->resident_page_count == 0) 2174 return; 2175 p = vm_page_find_least(object, start); 2176 2177 /* 2178 * Here, the variable "p" is either (1) the page with the least pindex 2179 * greater than or equal to the parameter "start" or (2) NULL. 2180 */ 2181 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2182 next = TAILQ_NEXT(p, listq); 2183 vm_page_deactivate_noreuse(p); 2184 } 2185 } 2186 2187 /* 2188 * Populate the specified range of the object with valid pages. Returns 2189 * TRUE if the range is successfully populated and FALSE otherwise. 2190 * 2191 * Note: This function should be optimized to pass a larger array of 2192 * pages to vm_pager_get_pages() before it is applied to a non- 2193 * OBJT_DEVICE object. 2194 * 2195 * The object must be locked. 2196 */ 2197 boolean_t 2198 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2199 { 2200 vm_page_t m; 2201 vm_pindex_t pindex; 2202 int rv; 2203 2204 VM_OBJECT_ASSERT_WLOCKED(object); 2205 for (pindex = start; pindex < end; pindex++) { 2206 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); 2207 if (rv != VM_PAGER_OK) 2208 break; 2209 2210 /* 2211 * Keep "m" busy because a subsequent iteration may unlock 2212 * the object. 2213 */ 2214 } 2215 if (pindex > start) { 2216 m = vm_page_lookup(object, start); 2217 while (m != NULL && m->pindex < pindex) { 2218 vm_page_xunbusy(m); 2219 m = TAILQ_NEXT(m, listq); 2220 } 2221 } 2222 return (pindex == end); 2223 } 2224 2225 /* 2226 * Routine: vm_object_coalesce 2227 * Function: Coalesces two objects backing up adjoining 2228 * regions of memory into a single object. 2229 * 2230 * returns TRUE if objects were combined. 2231 * 2232 * NOTE: Only works at the moment if the second object is NULL - 2233 * if it's not, which object do we lock first? 2234 * 2235 * Parameters: 2236 * prev_object First object to coalesce 2237 * prev_offset Offset into prev_object 2238 * prev_size Size of reference to prev_object 2239 * next_size Size of reference to the second object 2240 * reserved Indicator that extension region has 2241 * swap accounted for 2242 * 2243 * Conditions: 2244 * The object must *not* be locked. 2245 */ 2246 boolean_t 2247 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2248 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2249 { 2250 vm_pindex_t next_pindex; 2251 2252 if (prev_object == NULL) 2253 return (TRUE); 2254 if ((prev_object->flags & OBJ_ANON) == 0) 2255 return (FALSE); 2256 2257 VM_OBJECT_WLOCK(prev_object); 2258 /* 2259 * Try to collapse the object first. 2260 */ 2261 vm_object_collapse(prev_object); 2262 2263 /* 2264 * Can't coalesce if: . more than one reference . paged out . shadows 2265 * another object . has a copy elsewhere (any of which mean that the 2266 * pages not mapped to prev_entry may be in use anyway) 2267 */ 2268 if (prev_object->backing_object != NULL) { 2269 VM_OBJECT_WUNLOCK(prev_object); 2270 return (FALSE); 2271 } 2272 2273 prev_size >>= PAGE_SHIFT; 2274 next_size >>= PAGE_SHIFT; 2275 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2276 2277 if (prev_object->ref_count > 1 && 2278 prev_object->size != next_pindex && 2279 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2280 VM_OBJECT_WUNLOCK(prev_object); 2281 return (FALSE); 2282 } 2283 2284 /* 2285 * Account for the charge. 2286 */ 2287 if (prev_object->cred != NULL) { 2288 /* 2289 * If prev_object was charged, then this mapping, 2290 * although not charged now, may become writable 2291 * later. Non-NULL cred in the object would prevent 2292 * swap reservation during enabling of the write 2293 * access, so reserve swap now. Failed reservation 2294 * cause allocation of the separate object for the map 2295 * entry, and swap reservation for this entry is 2296 * managed in appropriate time. 2297 */ 2298 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2299 prev_object->cred)) { 2300 VM_OBJECT_WUNLOCK(prev_object); 2301 return (FALSE); 2302 } 2303 prev_object->charge += ptoa(next_size); 2304 } 2305 2306 /* 2307 * Remove any pages that may still be in the object from a previous 2308 * deallocation. 2309 */ 2310 if (next_pindex < prev_object->size) { 2311 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2312 next_size, 0); 2313 #if 0 2314 if (prev_object->cred != NULL) { 2315 KASSERT(prev_object->charge >= 2316 ptoa(prev_object->size - next_pindex), 2317 ("object %p overcharged 1 %jx %jx", prev_object, 2318 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2319 prev_object->charge -= ptoa(prev_object->size - 2320 next_pindex); 2321 } 2322 #endif 2323 } 2324 2325 /* 2326 * Extend the object if necessary. 2327 */ 2328 if (next_pindex + next_size > prev_object->size) 2329 prev_object->size = next_pindex + next_size; 2330 2331 VM_OBJECT_WUNLOCK(prev_object); 2332 return (TRUE); 2333 } 2334 2335 void 2336 vm_object_set_writeable_dirty_(vm_object_t object) 2337 { 2338 atomic_add_int(&object->generation, 1); 2339 } 2340 2341 bool 2342 vm_object_mightbedirty_(vm_object_t object) 2343 { 2344 return (object->generation != object->cleangeneration); 2345 } 2346 2347 /* 2348 * vm_object_unwire: 2349 * 2350 * For each page offset within the specified range of the given object, 2351 * find the highest-level page in the shadow chain and unwire it. A page 2352 * must exist at every page offset, and the highest-level page must be 2353 * wired. 2354 */ 2355 void 2356 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2357 uint8_t queue) 2358 { 2359 vm_object_t tobject, t1object; 2360 vm_page_t m, tm; 2361 vm_pindex_t end_pindex, pindex, tpindex; 2362 int depth, locked_depth; 2363 2364 KASSERT((offset & PAGE_MASK) == 0, 2365 ("vm_object_unwire: offset is not page aligned")); 2366 KASSERT((length & PAGE_MASK) == 0, 2367 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2368 /* The wired count of a fictitious page never changes. */ 2369 if ((object->flags & OBJ_FICTITIOUS) != 0) 2370 return; 2371 pindex = OFF_TO_IDX(offset); 2372 end_pindex = pindex + atop(length); 2373 again: 2374 locked_depth = 1; 2375 VM_OBJECT_RLOCK(object); 2376 m = vm_page_find_least(object, pindex); 2377 while (pindex < end_pindex) { 2378 if (m == NULL || pindex < m->pindex) { 2379 /* 2380 * The first object in the shadow chain doesn't 2381 * contain a page at the current index. Therefore, 2382 * the page must exist in a backing object. 2383 */ 2384 tobject = object; 2385 tpindex = pindex; 2386 depth = 0; 2387 do { 2388 tpindex += 2389 OFF_TO_IDX(tobject->backing_object_offset); 2390 tobject = tobject->backing_object; 2391 KASSERT(tobject != NULL, 2392 ("vm_object_unwire: missing page")); 2393 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2394 goto next_page; 2395 depth++; 2396 if (depth == locked_depth) { 2397 locked_depth++; 2398 VM_OBJECT_RLOCK(tobject); 2399 } 2400 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2401 NULL); 2402 } else { 2403 tm = m; 2404 m = TAILQ_NEXT(m, listq); 2405 } 2406 if (vm_page_trysbusy(tm) == 0) { 2407 for (tobject = object; locked_depth >= 1; 2408 locked_depth--) { 2409 t1object = tobject->backing_object; 2410 if (tm->object != tobject) 2411 VM_OBJECT_RUNLOCK(tobject); 2412 tobject = t1object; 2413 } 2414 tobject = tm->object; 2415 if (!vm_page_busy_sleep(tm, "unwbo", 2416 VM_ALLOC_IGN_SBUSY)) 2417 VM_OBJECT_RUNLOCK(tobject); 2418 goto again; 2419 } 2420 vm_page_unwire(tm, queue); 2421 vm_page_sunbusy(tm); 2422 next_page: 2423 pindex++; 2424 } 2425 /* Release the accumulated object locks. */ 2426 for (tobject = object; locked_depth >= 1; locked_depth--) { 2427 t1object = tobject->backing_object; 2428 VM_OBJECT_RUNLOCK(tobject); 2429 tobject = t1object; 2430 } 2431 } 2432 2433 /* 2434 * Return the vnode for the given object, or NULL if none exists. 2435 * For tmpfs objects, the function may return NULL if there is 2436 * no vnode allocated at the time of the call. 2437 */ 2438 struct vnode * 2439 vm_object_vnode(vm_object_t object) 2440 { 2441 struct vnode *vp; 2442 2443 VM_OBJECT_ASSERT_LOCKED(object); 2444 vm_pager_getvp(object, &vp, NULL); 2445 return (vp); 2446 } 2447 2448 /* 2449 * Busy the vm object. This prevents new pages belonging to the object from 2450 * becoming busy. Existing pages persist as busy. Callers are responsible 2451 * for checking page state before proceeding. 2452 */ 2453 void 2454 vm_object_busy(vm_object_t obj) 2455 { 2456 2457 VM_OBJECT_ASSERT_LOCKED(obj); 2458 2459 blockcount_acquire(&obj->busy, 1); 2460 /* The fence is required to order loads of page busy. */ 2461 atomic_thread_fence_acq_rel(); 2462 } 2463 2464 void 2465 vm_object_unbusy(vm_object_t obj) 2466 { 2467 2468 blockcount_release(&obj->busy, 1); 2469 } 2470 2471 void 2472 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2473 { 2474 2475 VM_OBJECT_ASSERT_UNLOCKED(obj); 2476 2477 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2478 } 2479 2480 /* 2481 * This function aims to determine if the object is mapped, 2482 * specifically, if it is referenced by a vm_map_entry. Because 2483 * objects occasionally acquire transient references that do not 2484 * represent a mapping, the method used here is inexact. However, it 2485 * has very low overhead and is good enough for the advisory 2486 * vm.vmtotal sysctl. 2487 */ 2488 bool 2489 vm_object_is_active(vm_object_t obj) 2490 { 2491 2492 return (obj->ref_count > atomic_load_int(&obj->shadow_count)); 2493 } 2494 2495 static int 2496 vm_object_list_handler(struct sysctl_req *req, bool swap_only) 2497 { 2498 struct kinfo_vmobject *kvo; 2499 char *fullpath, *freepath; 2500 struct vnode *vp; 2501 struct vattr va; 2502 vm_object_t obj; 2503 vm_page_t m; 2504 u_long sp; 2505 int count, error; 2506 bool want_path; 2507 2508 if (req->oldptr == NULL) { 2509 /* 2510 * If an old buffer has not been provided, generate an 2511 * estimate of the space needed for a subsequent call. 2512 */ 2513 mtx_lock(&vm_object_list_mtx); 2514 count = 0; 2515 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2516 if (obj->type == OBJT_DEAD) 2517 continue; 2518 count++; 2519 } 2520 mtx_unlock(&vm_object_list_mtx); 2521 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2522 count * 11 / 10)); 2523 } 2524 2525 want_path = !(swap_only || jailed(curthread->td_ucred)); 2526 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO); 2527 error = 0; 2528 2529 /* 2530 * VM objects are type stable and are never removed from the 2531 * list once added. This allows us to safely read obj->object_list 2532 * after reacquiring the VM object lock. 2533 */ 2534 mtx_lock(&vm_object_list_mtx); 2535 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2536 if (obj->type == OBJT_DEAD || 2537 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) 2538 continue; 2539 VM_OBJECT_RLOCK(obj); 2540 if (obj->type == OBJT_DEAD || 2541 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) { 2542 VM_OBJECT_RUNLOCK(obj); 2543 continue; 2544 } 2545 mtx_unlock(&vm_object_list_mtx); 2546 kvo->kvo_size = ptoa(obj->size); 2547 kvo->kvo_resident = obj->resident_page_count; 2548 kvo->kvo_ref_count = obj->ref_count; 2549 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count); 2550 kvo->kvo_memattr = obj->memattr; 2551 kvo->kvo_active = 0; 2552 kvo->kvo_inactive = 0; 2553 if (!swap_only) { 2554 TAILQ_FOREACH(m, &obj->memq, listq) { 2555 /* 2556 * A page may belong to the object but be 2557 * dequeued and set to PQ_NONE while the 2558 * object lock is not held. This makes the 2559 * reads of m->queue below racy, and we do not 2560 * count pages set to PQ_NONE. However, this 2561 * sysctl is only meant to give an 2562 * approximation of the system anyway. 2563 */ 2564 if (m->a.queue == PQ_ACTIVE) 2565 kvo->kvo_active++; 2566 else if (m->a.queue == PQ_INACTIVE) 2567 kvo->kvo_inactive++; 2568 } 2569 } 2570 2571 kvo->kvo_vn_fileid = 0; 2572 kvo->kvo_vn_fsid = 0; 2573 kvo->kvo_vn_fsid_freebsd11 = 0; 2574 freepath = NULL; 2575 fullpath = ""; 2576 vp = NULL; 2577 kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp : 2578 NULL); 2579 if (vp != NULL) { 2580 vref(vp); 2581 } else if ((obj->flags & OBJ_ANON) != 0) { 2582 MPASS(kvo->kvo_type == KVME_TYPE_SWAP); 2583 kvo->kvo_me = (uintptr_t)obj; 2584 /* tmpfs objs are reported as vnodes */ 2585 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object; 2586 sp = swap_pager_swapped_pages(obj); 2587 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp; 2588 } 2589 VM_OBJECT_RUNLOCK(obj); 2590 if (vp != NULL) { 2591 vn_fullpath(vp, &fullpath, &freepath); 2592 vn_lock(vp, LK_SHARED | LK_RETRY); 2593 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2594 kvo->kvo_vn_fileid = va.va_fileid; 2595 kvo->kvo_vn_fsid = va.va_fsid; 2596 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2597 /* truncate */ 2598 } 2599 vput(vp); 2600 } 2601 2602 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2603 free(freepath, M_TEMP); 2604 2605 /* Pack record size down */ 2606 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2607 + strlen(kvo->kvo_path) + 1; 2608 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2609 sizeof(uint64_t)); 2610 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2611 maybe_yield(); 2612 mtx_lock(&vm_object_list_mtx); 2613 if (error) 2614 break; 2615 } 2616 mtx_unlock(&vm_object_list_mtx); 2617 free(kvo, M_TEMP); 2618 return (error); 2619 } 2620 2621 static int 2622 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2623 { 2624 return (vm_object_list_handler(req, false)); 2625 } 2626 2627 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2628 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2629 "List of VM objects"); 2630 2631 static int 2632 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS) 2633 { 2634 return (vm_object_list_handler(req, true)); 2635 } 2636 2637 /* 2638 * This sysctl returns list of the anonymous or swap objects. Intent 2639 * is to provide stripped optimized list useful to analyze swap use. 2640 * Since technically non-swap (default) objects participate in the 2641 * shadow chains, and are converted to swap type as needed by swap 2642 * pager, we must report them. 2643 */ 2644 SYSCTL_PROC(_vm, OID_AUTO, swap_objects, 2645 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, 2646 sysctl_vm_object_list_swap, "S,kinfo_vmobject", 2647 "List of swap VM objects"); 2648 2649 #include "opt_ddb.h" 2650 #ifdef DDB 2651 #include <sys/kernel.h> 2652 2653 #include <sys/cons.h> 2654 2655 #include <ddb/ddb.h> 2656 2657 static int 2658 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2659 { 2660 vm_map_t tmpm; 2661 vm_map_entry_t tmpe; 2662 vm_object_t obj; 2663 2664 if (map == 0) 2665 return 0; 2666 2667 if (entry == 0) { 2668 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2669 if (_vm_object_in_map(map, object, tmpe)) { 2670 return 1; 2671 } 2672 } 2673 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2674 tmpm = entry->object.sub_map; 2675 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2676 if (_vm_object_in_map(tmpm, object, tmpe)) { 2677 return 1; 2678 } 2679 } 2680 } else if ((obj = entry->object.vm_object) != NULL) { 2681 for (; obj; obj = obj->backing_object) 2682 if (obj == object) { 2683 return 1; 2684 } 2685 } 2686 return 0; 2687 } 2688 2689 static int 2690 vm_object_in_map(vm_object_t object) 2691 { 2692 struct proc *p; 2693 2694 /* sx_slock(&allproc_lock); */ 2695 FOREACH_PROC_IN_SYSTEM(p) { 2696 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2697 continue; 2698 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2699 /* sx_sunlock(&allproc_lock); */ 2700 return 1; 2701 } 2702 } 2703 /* sx_sunlock(&allproc_lock); */ 2704 if (_vm_object_in_map(kernel_map, object, 0)) 2705 return 1; 2706 return 0; 2707 } 2708 2709 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE) 2710 { 2711 vm_object_t object; 2712 2713 /* 2714 * make sure that internal objs are in a map somewhere 2715 * and none have zero ref counts. 2716 */ 2717 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2718 if ((object->flags & OBJ_ANON) != 0) { 2719 if (object->ref_count == 0) { 2720 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2721 (long)object->size); 2722 } 2723 if (!vm_object_in_map(object)) { 2724 db_printf( 2725 "vmochk: internal obj is not in a map: " 2726 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2727 object->ref_count, (u_long)object->size, 2728 (u_long)object->size, 2729 (void *)object->backing_object); 2730 } 2731 } 2732 if (db_pager_quit) 2733 return; 2734 } 2735 } 2736 2737 /* 2738 * vm_object_print: [ debug ] 2739 */ 2740 DB_SHOW_COMMAND(object, vm_object_print_static) 2741 { 2742 /* XXX convert args. */ 2743 vm_object_t object = (vm_object_t)addr; 2744 boolean_t full = have_addr; 2745 2746 vm_page_t p; 2747 2748 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2749 #define count was_count 2750 2751 int count; 2752 2753 if (object == NULL) 2754 return; 2755 2756 db_iprintf( 2757 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", 2758 object, (int)object->type, (uintmax_t)object->size, 2759 object->resident_page_count, object->ref_count, object->flags, 2760 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); 2761 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2762 atomic_load_int(&object->shadow_count), 2763 object->backing_object ? object->backing_object->ref_count : 0, 2764 object->backing_object, (uintmax_t)object->backing_object_offset); 2765 2766 if (!full) 2767 return; 2768 2769 db_indent += 2; 2770 count = 0; 2771 TAILQ_FOREACH(p, &object->memq, listq) { 2772 if (count == 0) 2773 db_iprintf("memory:="); 2774 else if (count == 6) { 2775 db_printf("\n"); 2776 db_iprintf(" ..."); 2777 count = 0; 2778 } else 2779 db_printf(","); 2780 count++; 2781 2782 db_printf("(off=0x%jx,page=0x%jx)", 2783 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2784 2785 if (db_pager_quit) 2786 break; 2787 } 2788 if (count != 0) 2789 db_printf("\n"); 2790 db_indent -= 2; 2791 } 2792 2793 /* XXX. */ 2794 #undef count 2795 2796 /* XXX need this non-static entry for calling from vm_map_print. */ 2797 void 2798 vm_object_print( 2799 /* db_expr_t */ long addr, 2800 boolean_t have_addr, 2801 /* db_expr_t */ long count, 2802 char *modif) 2803 { 2804 vm_object_print_static(addr, have_addr, count, modif); 2805 } 2806 2807 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE) 2808 { 2809 vm_object_t object; 2810 vm_pindex_t fidx; 2811 vm_paddr_t pa; 2812 vm_page_t m, prev_m; 2813 int rcount; 2814 2815 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2816 db_printf("new object: %p\n", (void *)object); 2817 if (db_pager_quit) 2818 return; 2819 2820 rcount = 0; 2821 fidx = 0; 2822 pa = -1; 2823 TAILQ_FOREACH(m, &object->memq, listq) { 2824 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2825 prev_m->pindex + 1 != m->pindex) { 2826 if (rcount) { 2827 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2828 (long)fidx, rcount, (long)pa); 2829 if (db_pager_quit) 2830 return; 2831 rcount = 0; 2832 } 2833 } 2834 if (rcount && 2835 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2836 ++rcount; 2837 continue; 2838 } 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 } 2845 fidx = m->pindex; 2846 pa = VM_PAGE_TO_PHYS(m); 2847 rcount = 1; 2848 } 2849 if (rcount) { 2850 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2851 (long)fidx, rcount, (long)pa); 2852 if (db_pager_quit) 2853 return; 2854 } 2855 } 2856 } 2857 #endif /* DDB */ 2858