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/conf.h> 70 #include <sys/cpuset.h> 71 #include <sys/ipc.h> 72 #include <sys/jail.h> 73 #include <sys/limits.h> 74 #include <sys/lock.h> 75 #include <sys/mman.h> 76 #include <sys/mount.h> 77 #include <sys/kernel.h> 78 #include <sys/mutex.h> 79 #include <sys/pctrie.h> 80 #include <sys/proc.h> 81 #include <sys/refcount.h> 82 #include <sys/shm.h> 83 #include <sys/sx.h> 84 #include <sys/sysctl.h> 85 #include <sys/resourcevar.h> 86 #include <sys/refcount.h> 87 #include <sys/rwlock.h> 88 #include <sys/user.h> 89 #include <sys/vnode.h> 90 #include <sys/vmmeter.h> 91 92 #include <vm/vm.h> 93 #include <vm/vm_param.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_page.h> 98 #include <vm/vm_pageout.h> 99 #include <vm/vm_pager.h> 100 #include <vm/vm_phys.h> 101 #include <vm/vm_pagequeue.h> 102 #include <vm/swap_pager.h> 103 #include <vm/vm_kern.h> 104 #include <vm/vm_extern.h> 105 #include <vm/vm_radix.h> 106 #include <vm/vm_reserv.h> 107 #include <vm/uma.h> 108 109 static int old_msync; 110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 111 "Use old (insecure) msync behavior"); 112 113 static int vm_object_page_collect_flush(struct pctrie_iter *pages, 114 vm_page_t p, int pagerflags, int flags, boolean_t *allclean, 115 boolean_t *eio); 116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, 117 boolean_t *allclean); 118 static void vm_object_backing_remove(vm_object_t object); 119 120 /* 121 * Virtual memory objects maintain the actual data 122 * associated with allocated virtual memory. A given 123 * page of memory exists within exactly one object. 124 * 125 * An object is only deallocated when all "references" 126 * are given up. Only one "reference" to a given 127 * region of an object should be writeable. 128 * 129 * Associated with each object is a list of all resident 130 * memory pages belonging to that object; this list is 131 * maintained by the "vm_page" module, and locked by the object's 132 * lock. 133 * 134 * Each object also records a "pager" routine which is 135 * used to retrieve (and store) pages to the proper backing 136 * storage. In addition, objects may be backed by other 137 * objects from which they were virtual-copied. 138 * 139 * The only items within the object structure which are 140 * modified after time of creation are: 141 * reference count locked by object's lock 142 * pager routine locked by object's lock 143 * 144 */ 145 146 struct object_q vm_object_list; 147 struct mtx vm_object_list_mtx; /* lock for object list and count */ 148 149 struct vm_object kernel_object_store; 150 151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 152 "VM object stats"); 153 154 static COUNTER_U64_DEFINE_EARLY(object_collapses); 155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 156 &object_collapses, 157 "VM object collapses"); 158 159 static COUNTER_U64_DEFINE_EARLY(object_bypasses); 160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 161 &object_bypasses, 162 "VM object bypasses"); 163 164 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits); 165 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD, 166 &object_collapse_waits, 167 "Number of sleeps for collapse"); 168 169 static uma_zone_t obj_zone; 170 171 static int vm_object_zinit(void *mem, int size, int flags); 172 173 #ifdef INVARIANTS 174 static void vm_object_zdtor(void *mem, int size, void *arg); 175 176 static void 177 vm_object_zdtor(void *mem, int size, void *arg) 178 { 179 vm_object_t object; 180 181 object = (vm_object_t)mem; 182 KASSERT(object->ref_count == 0, 183 ("object %p ref_count = %d", object, object->ref_count)); 184 KASSERT(TAILQ_EMPTY(&object->memq), 185 ("object %p has resident pages in its memq", object)); 186 KASSERT(vm_radix_is_empty(&object->rtree), 187 ("object %p has resident pages in its trie", object)); 188 #if VM_NRESERVLEVEL > 0 189 KASSERT(LIST_EMPTY(&object->rvq), 190 ("object %p has reservations", 191 object)); 192 #endif 193 KASSERT(!vm_object_busied(object), 194 ("object %p busy = %d", object, blockcount_read(&object->busy))); 195 KASSERT(object->resident_page_count == 0, 196 ("object %p resident_page_count = %d", 197 object, object->resident_page_count)); 198 KASSERT(atomic_load_int(&object->shadow_count) == 0, 199 ("object %p shadow_count = %d", 200 object, atomic_load_int(&object->shadow_count))); 201 KASSERT(object->type == OBJT_DEAD, 202 ("object %p has non-dead type %d", 203 object, object->type)); 204 KASSERT(object->charge == 0 && object->cred == NULL, 205 ("object %p has non-zero charge %ju (%p)", 206 object, (uintmax_t)object->charge, object->cred)); 207 } 208 #endif 209 210 static int 211 vm_object_zinit(void *mem, int size, int flags) 212 { 213 vm_object_t object; 214 215 object = (vm_object_t)mem; 216 rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW); 217 218 /* These are true for any object that has been freed */ 219 object->type = OBJT_DEAD; 220 vm_radix_init(&object->rtree); 221 refcount_init(&object->ref_count, 0); 222 blockcount_init(&object->paging_in_progress); 223 blockcount_init(&object->busy); 224 object->resident_page_count = 0; 225 atomic_store_int(&object->shadow_count, 0); 226 object->flags = OBJ_DEAD; 227 228 mtx_lock(&vm_object_list_mtx); 229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 230 mtx_unlock(&vm_object_list_mtx); 231 return (0); 232 } 233 234 static void 235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, 236 vm_object_t object, void *handle) 237 { 238 239 TAILQ_INIT(&object->memq); 240 LIST_INIT(&object->shadow_head); 241 242 object->type = type; 243 object->flags = flags; 244 if ((flags & OBJ_SWAP) != 0) { 245 pctrie_init(&object->un_pager.swp.swp_blks); 246 object->un_pager.swp.writemappings = 0; 247 } 248 249 /* 250 * Ensure that swap_pager_swapoff() iteration over object_list 251 * sees up to date type and pctrie head if it observed 252 * non-dead object. 253 */ 254 atomic_thread_fence_rel(); 255 256 object->pg_color = 0; 257 object->size = size; 258 object->domain.dr_policy = NULL; 259 object->generation = 1; 260 object->cleangeneration = 1; 261 refcount_init(&object->ref_count, 1); 262 object->memattr = VM_MEMATTR_DEFAULT; 263 object->cred = NULL; 264 object->charge = 0; 265 object->handle = handle; 266 object->backing_object = NULL; 267 object->backing_object_offset = (vm_ooffset_t) 0; 268 #if VM_NRESERVLEVEL > 0 269 LIST_INIT(&object->rvq); 270 #endif 271 umtx_shm_object_init(object); 272 } 273 274 /* 275 * vm_object_init: 276 * 277 * Initialize the VM objects module. 278 */ 279 void 280 vm_object_init(void) 281 { 282 TAILQ_INIT(&vm_object_list); 283 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 284 285 rw_init(&kernel_object->lock, "kernel vm object"); 286 vm_radix_init(&kernel_object->rtree); 287 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 288 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); 289 #if VM_NRESERVLEVEL > 0 290 kernel_object->flags |= OBJ_COLORED; 291 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 292 #endif 293 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 294 295 /* 296 * The lock portion of struct vm_object must be type stable due 297 * to vm_pageout_fallback_object_lock locking a vm object 298 * without holding any references to it. 299 * 300 * paging_in_progress is valid always. Lockless references to 301 * the objects may acquire pip and then check OBJ_DEAD. 302 */ 303 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 304 #ifdef INVARIANTS 305 vm_object_zdtor, 306 #else 307 NULL, 308 #endif 309 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 310 311 vm_radix_zinit(); 312 } 313 314 void 315 vm_object_clear_flag(vm_object_t object, u_short bits) 316 { 317 318 VM_OBJECT_ASSERT_WLOCKED(object); 319 object->flags &= ~bits; 320 } 321 322 /* 323 * Sets the default memory attribute for the specified object. Pages 324 * that are allocated to this object are by default assigned this memory 325 * attribute. 326 * 327 * Presently, this function must be called before any pages are allocated 328 * to the object. In the future, this requirement may be relaxed for 329 * "default" and "swap" objects. 330 */ 331 int 332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 333 { 334 335 VM_OBJECT_ASSERT_WLOCKED(object); 336 337 if (object->type == OBJT_DEAD) 338 return (KERN_INVALID_ARGUMENT); 339 if (!TAILQ_EMPTY(&object->memq)) 340 return (KERN_FAILURE); 341 342 object->memattr = memattr; 343 return (KERN_SUCCESS); 344 } 345 346 void 347 vm_object_pip_add(vm_object_t object, short i) 348 { 349 350 if (i > 0) 351 blockcount_acquire(&object->paging_in_progress, i); 352 } 353 354 void 355 vm_object_pip_wakeup(vm_object_t object) 356 { 357 358 vm_object_pip_wakeupn(object, 1); 359 } 360 361 void 362 vm_object_pip_wakeupn(vm_object_t object, short i) 363 { 364 365 if (i > 0) 366 blockcount_release(&object->paging_in_progress, i); 367 } 368 369 /* 370 * Atomically drop the object lock and wait for pip to drain. This protects 371 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 372 * on return. 373 */ 374 static void 375 vm_object_pip_sleep(vm_object_t object, const char *waitid) 376 { 377 378 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 379 waitid, PVM | PDROP); 380 } 381 382 void 383 vm_object_pip_wait(vm_object_t object, const char *waitid) 384 { 385 386 VM_OBJECT_ASSERT_WLOCKED(object); 387 388 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 389 PVM); 390 } 391 392 void 393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 394 { 395 396 VM_OBJECT_ASSERT_UNLOCKED(object); 397 398 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 399 } 400 401 /* 402 * vm_object_allocate: 403 * 404 * Returns a new object with the given size. 405 */ 406 vm_object_t 407 vm_object_allocate(objtype_t type, vm_pindex_t size) 408 { 409 vm_object_t object; 410 u_short flags; 411 412 switch (type) { 413 case OBJT_DEAD: 414 panic("vm_object_allocate: can't create OBJT_DEAD"); 415 case OBJT_SWAP: 416 flags = OBJ_COLORED | OBJ_SWAP; 417 break; 418 case OBJT_DEVICE: 419 case OBJT_SG: 420 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 421 break; 422 case OBJT_MGTDEVICE: 423 flags = OBJ_FICTITIOUS; 424 break; 425 case OBJT_PHYS: 426 flags = OBJ_UNMANAGED; 427 break; 428 case OBJT_VNODE: 429 flags = 0; 430 break; 431 default: 432 panic("vm_object_allocate: type %d is undefined or dynamic", 433 type); 434 } 435 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 436 _vm_object_allocate(type, size, flags, object, NULL); 437 438 return (object); 439 } 440 441 vm_object_t 442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags) 443 { 444 vm_object_t object; 445 446 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */); 447 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 448 _vm_object_allocate(dyntype, size, flags, object, NULL); 449 450 return (object); 451 } 452 453 /* 454 * vm_object_allocate_anon: 455 * 456 * Returns a new default object of the given size and marked as 457 * anonymous memory for special split/collapse handling. Color 458 * to be initialized by the caller. 459 */ 460 vm_object_t 461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 462 struct ucred *cred, vm_size_t charge) 463 { 464 vm_object_t handle, object; 465 466 if (backing_object == NULL) 467 handle = NULL; 468 else if ((backing_object->flags & OBJ_ANON) != 0) 469 handle = backing_object->handle; 470 else 471 handle = backing_object; 472 object = uma_zalloc(obj_zone, M_WAITOK); 473 _vm_object_allocate(OBJT_SWAP, size, 474 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle); 475 object->cred = cred; 476 object->charge = cred != NULL ? charge : 0; 477 return (object); 478 } 479 480 static void 481 vm_object_reference_vnode(vm_object_t object) 482 { 483 u_int old; 484 485 /* 486 * vnode objects need the lock for the first reference 487 * to serialize with vnode_object_deallocate(). 488 */ 489 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 490 VM_OBJECT_RLOCK(object); 491 old = refcount_acquire(&object->ref_count); 492 if (object->type == OBJT_VNODE && old == 0) 493 vref(object->handle); 494 VM_OBJECT_RUNLOCK(object); 495 } 496 } 497 498 /* 499 * vm_object_reference: 500 * 501 * Acquires a reference to the given object. 502 */ 503 void 504 vm_object_reference(vm_object_t object) 505 { 506 507 if (object == NULL) 508 return; 509 510 if (object->type == OBJT_VNODE) 511 vm_object_reference_vnode(object); 512 else 513 refcount_acquire(&object->ref_count); 514 KASSERT((object->flags & OBJ_DEAD) == 0, 515 ("vm_object_reference: Referenced dead object.")); 516 } 517 518 /* 519 * vm_object_reference_locked: 520 * 521 * Gets another reference to the given object. 522 * 523 * The object must be locked. 524 */ 525 void 526 vm_object_reference_locked(vm_object_t object) 527 { 528 u_int old; 529 530 VM_OBJECT_ASSERT_LOCKED(object); 531 old = refcount_acquire(&object->ref_count); 532 if (object->type == OBJT_VNODE && old == 0) 533 vref(object->handle); 534 KASSERT((object->flags & OBJ_DEAD) == 0, 535 ("vm_object_reference: Referenced dead object.")); 536 } 537 538 /* 539 * Handle deallocating an object of type OBJT_VNODE. 540 */ 541 static void 542 vm_object_deallocate_vnode(vm_object_t object) 543 { 544 struct vnode *vp = (struct vnode *) object->handle; 545 bool last; 546 547 KASSERT(object->type == OBJT_VNODE, 548 ("vm_object_deallocate_vnode: not a vnode object")); 549 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 550 551 /* Object lock to protect handle lookup. */ 552 last = refcount_release(&object->ref_count); 553 VM_OBJECT_RUNLOCK(object); 554 555 if (!last) 556 return; 557 558 if (!umtx_shm_vnobj_persistent) 559 umtx_shm_object_terminated(object); 560 561 /* vrele may need the vnode lock. */ 562 vrele(vp); 563 } 564 565 /* 566 * We dropped a reference on an object and discovered that it had a 567 * single remaining shadow. This is a sibling of the reference we 568 * dropped. Attempt to collapse the sibling and backing object. 569 */ 570 static vm_object_t 571 vm_object_deallocate_anon(vm_object_t backing_object) 572 { 573 vm_object_t object; 574 575 /* Fetch the final shadow. */ 576 object = LIST_FIRST(&backing_object->shadow_head); 577 KASSERT(object != NULL && 578 atomic_load_int(&backing_object->shadow_count) == 1, 579 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 580 backing_object->ref_count, 581 atomic_load_int(&backing_object->shadow_count))); 582 KASSERT((object->flags & OBJ_ANON) != 0, 583 ("invalid shadow object %p", object)); 584 585 if (!VM_OBJECT_TRYWLOCK(object)) { 586 /* 587 * Prevent object from disappearing since we do not have a 588 * reference. 589 */ 590 vm_object_pip_add(object, 1); 591 VM_OBJECT_WUNLOCK(backing_object); 592 VM_OBJECT_WLOCK(object); 593 vm_object_pip_wakeup(object); 594 } else 595 VM_OBJECT_WUNLOCK(backing_object); 596 597 /* 598 * Check for a collapse/terminate race with the last reference holder. 599 */ 600 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 601 !refcount_acquire_if_not_zero(&object->ref_count)) { 602 VM_OBJECT_WUNLOCK(object); 603 return (NULL); 604 } 605 backing_object = object->backing_object; 606 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 607 vm_object_collapse(object); 608 VM_OBJECT_WUNLOCK(object); 609 610 return (object); 611 } 612 613 /* 614 * vm_object_deallocate: 615 * 616 * Release a reference to the specified object, 617 * gained either through a vm_object_allocate 618 * or a vm_object_reference call. When all references 619 * are gone, storage associated with this object 620 * may be relinquished. 621 * 622 * No object may be locked. 623 */ 624 void 625 vm_object_deallocate(vm_object_t object) 626 { 627 vm_object_t temp; 628 bool released; 629 630 while (object != NULL) { 631 /* 632 * If the reference count goes to 0 we start calling 633 * vm_object_terminate() on the object chain. A ref count 634 * of 1 may be a special case depending on the shadow count 635 * being 0 or 1. These cases require a write lock on the 636 * object. 637 */ 638 if ((object->flags & OBJ_ANON) == 0) 639 released = refcount_release_if_gt(&object->ref_count, 1); 640 else 641 released = refcount_release_if_gt(&object->ref_count, 2); 642 if (released) 643 return; 644 645 if (object->type == OBJT_VNODE) { 646 VM_OBJECT_RLOCK(object); 647 if (object->type == OBJT_VNODE) { 648 vm_object_deallocate_vnode(object); 649 return; 650 } 651 VM_OBJECT_RUNLOCK(object); 652 } 653 654 VM_OBJECT_WLOCK(object); 655 KASSERT(object->ref_count > 0, 656 ("vm_object_deallocate: object deallocated too many times: %d", 657 object->type)); 658 659 /* 660 * If this is not the final reference to an anonymous 661 * object we may need to collapse the shadow chain. 662 */ 663 if (!refcount_release(&object->ref_count)) { 664 if (object->ref_count > 1 || 665 atomic_load_int(&object->shadow_count) == 0) { 666 if ((object->flags & OBJ_ANON) != 0 && 667 object->ref_count == 1) 668 vm_object_set_flag(object, 669 OBJ_ONEMAPPING); 670 VM_OBJECT_WUNLOCK(object); 671 return; 672 } 673 674 /* Handle collapsing last ref on anonymous objects. */ 675 object = vm_object_deallocate_anon(object); 676 continue; 677 } 678 679 /* 680 * Handle the final reference to an object. We restart 681 * the loop with the backing object to avoid recursion. 682 */ 683 umtx_shm_object_terminated(object); 684 temp = object->backing_object; 685 if (temp != NULL) { 686 KASSERT(object->type == OBJT_SWAP, 687 ("shadowed tmpfs v_object 2 %p", object)); 688 vm_object_backing_remove(object); 689 } 690 691 KASSERT((object->flags & OBJ_DEAD) == 0, 692 ("vm_object_deallocate: Terminating dead object.")); 693 vm_object_set_flag(object, OBJ_DEAD); 694 vm_object_terminate(object); 695 object = temp; 696 } 697 } 698 699 void 700 vm_object_destroy(vm_object_t object) 701 { 702 uma_zfree(obj_zone, object); 703 } 704 705 static void 706 vm_object_sub_shadow(vm_object_t object) 707 { 708 KASSERT(object->shadow_count >= 1, 709 ("object %p sub_shadow count zero", object)); 710 atomic_subtract_int(&object->shadow_count, 1); 711 } 712 713 static void 714 vm_object_backing_remove_locked(vm_object_t object) 715 { 716 vm_object_t backing_object; 717 718 backing_object = object->backing_object; 719 VM_OBJECT_ASSERT_WLOCKED(object); 720 VM_OBJECT_ASSERT_WLOCKED(backing_object); 721 722 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 723 ("vm_object_backing_remove: Removing collapsing object.")); 724 725 vm_object_sub_shadow(backing_object); 726 if ((object->flags & OBJ_SHADOWLIST) != 0) { 727 LIST_REMOVE(object, shadow_list); 728 vm_object_clear_flag(object, OBJ_SHADOWLIST); 729 } 730 object->backing_object = NULL; 731 } 732 733 static void 734 vm_object_backing_remove(vm_object_t object) 735 { 736 vm_object_t backing_object; 737 738 VM_OBJECT_ASSERT_WLOCKED(object); 739 740 backing_object = object->backing_object; 741 if ((object->flags & OBJ_SHADOWLIST) != 0) { 742 VM_OBJECT_WLOCK(backing_object); 743 vm_object_backing_remove_locked(object); 744 VM_OBJECT_WUNLOCK(backing_object); 745 } else { 746 object->backing_object = NULL; 747 vm_object_sub_shadow(backing_object); 748 } 749 } 750 751 static void 752 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 753 { 754 755 VM_OBJECT_ASSERT_WLOCKED(object); 756 757 atomic_add_int(&backing_object->shadow_count, 1); 758 if ((backing_object->flags & OBJ_ANON) != 0) { 759 VM_OBJECT_ASSERT_WLOCKED(backing_object); 760 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 761 shadow_list); 762 vm_object_set_flag(object, OBJ_SHADOWLIST); 763 } 764 object->backing_object = backing_object; 765 } 766 767 static void 768 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 769 { 770 771 VM_OBJECT_ASSERT_WLOCKED(object); 772 773 if ((backing_object->flags & OBJ_ANON) != 0) { 774 VM_OBJECT_WLOCK(backing_object); 775 vm_object_backing_insert_locked(object, backing_object); 776 VM_OBJECT_WUNLOCK(backing_object); 777 } else { 778 object->backing_object = backing_object; 779 atomic_add_int(&backing_object->shadow_count, 1); 780 } 781 } 782 783 /* 784 * Insert an object into a backing_object's shadow list with an additional 785 * reference to the backing_object added. 786 */ 787 static void 788 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object) 789 { 790 791 VM_OBJECT_ASSERT_WLOCKED(object); 792 793 if ((backing_object->flags & OBJ_ANON) != 0) { 794 VM_OBJECT_WLOCK(backing_object); 795 KASSERT((backing_object->flags & OBJ_DEAD) == 0, 796 ("shadowing dead anonymous object")); 797 vm_object_reference_locked(backing_object); 798 vm_object_backing_insert_locked(object, backing_object); 799 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING); 800 VM_OBJECT_WUNLOCK(backing_object); 801 } else { 802 vm_object_reference(backing_object); 803 atomic_add_int(&backing_object->shadow_count, 1); 804 object->backing_object = backing_object; 805 } 806 } 807 808 /* 809 * Transfer a backing reference from backing_object to object. 810 */ 811 static void 812 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 813 { 814 vm_object_t new_backing_object; 815 816 /* 817 * Note that the reference to backing_object->backing_object 818 * moves from within backing_object to within object. 819 */ 820 vm_object_backing_remove_locked(object); 821 new_backing_object = backing_object->backing_object; 822 if (new_backing_object == NULL) 823 return; 824 if ((new_backing_object->flags & OBJ_ANON) != 0) { 825 VM_OBJECT_WLOCK(new_backing_object); 826 vm_object_backing_remove_locked(backing_object); 827 vm_object_backing_insert_locked(object, new_backing_object); 828 VM_OBJECT_WUNLOCK(new_backing_object); 829 } else { 830 /* 831 * shadow_count for new_backing_object is left 832 * unchanged, its reference provided by backing_object 833 * is replaced by object. 834 */ 835 object->backing_object = new_backing_object; 836 backing_object->backing_object = NULL; 837 } 838 } 839 840 /* 841 * Wait for a concurrent collapse to settle. 842 */ 843 static void 844 vm_object_collapse_wait(vm_object_t object) 845 { 846 847 VM_OBJECT_ASSERT_WLOCKED(object); 848 849 while ((object->flags & OBJ_COLLAPSING) != 0) { 850 vm_object_pip_wait(object, "vmcolwait"); 851 counter_u64_add(object_collapse_waits, 1); 852 } 853 } 854 855 /* 856 * Waits for a backing object to clear a pending collapse and returns 857 * it locked if it is an ANON object. 858 */ 859 static vm_object_t 860 vm_object_backing_collapse_wait(vm_object_t object) 861 { 862 vm_object_t backing_object; 863 864 VM_OBJECT_ASSERT_WLOCKED(object); 865 866 for (;;) { 867 backing_object = object->backing_object; 868 if (backing_object == NULL || 869 (backing_object->flags & OBJ_ANON) == 0) 870 return (NULL); 871 VM_OBJECT_WLOCK(backing_object); 872 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 873 break; 874 VM_OBJECT_WUNLOCK(object); 875 vm_object_pip_sleep(backing_object, "vmbckwait"); 876 counter_u64_add(object_collapse_waits, 1); 877 VM_OBJECT_WLOCK(object); 878 } 879 return (backing_object); 880 } 881 882 /* 883 * vm_object_terminate_single_page removes a pageable page from the object, 884 * and removes it from the paging queues and frees it, if it is not wired. 885 * It is invoked via callback from vm_object_terminate_pages. 886 */ 887 static void 888 vm_object_terminate_single_page(vm_page_t p, void *objectv) 889 { 890 vm_object_t object __diagused = objectv; 891 892 vm_page_assert_unbusied(p); 893 KASSERT(p->object == object && 894 (p->ref_count & VPRC_OBJREF) != 0, 895 ("%s: page %p is inconsistent", __func__, p)); 896 p->object = NULL; 897 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 898 KASSERT((object->flags & OBJ_UNMANAGED) != 0 || 899 vm_page_astate_load(p).queue != PQ_NONE, 900 ("%s: page %p does not belong to a queue", __func__, p)); 901 VM_CNT_INC(v_pfree); 902 vm_page_free(p); 903 } 904 } 905 906 /* 907 * vm_object_terminate_pages removes any remaining pageable pages 908 * from the object and resets the object to an empty state. 909 */ 910 static void 911 vm_object_terminate_pages(vm_object_t object) 912 { 913 VM_OBJECT_ASSERT_WLOCKED(object); 914 915 /* 916 * If the object contained any pages, then reset it to an empty state. 917 * Rather than incrementally removing each page from the object, the 918 * page and object are reset to any empty state. 919 */ 920 if (object->resident_page_count == 0) 921 return; 922 923 vm_radix_reclaim_callback(&object->rtree, 924 vm_object_terminate_single_page, object); 925 TAILQ_INIT(&object->memq); 926 object->resident_page_count = 0; 927 if (object->type == OBJT_VNODE) 928 vdrop(object->handle); 929 } 930 931 /* 932 * vm_object_terminate actually destroys the specified object, freeing 933 * up all previously used resources. 934 * 935 * The object must be locked. 936 * This routine may block. 937 */ 938 void 939 vm_object_terminate(vm_object_t object) 940 { 941 942 VM_OBJECT_ASSERT_WLOCKED(object); 943 KASSERT((object->flags & OBJ_DEAD) != 0, 944 ("terminating non-dead obj %p", object)); 945 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 946 ("terminating collapsing obj %p", object)); 947 KASSERT(object->backing_object == NULL, 948 ("terminating shadow obj %p", object)); 949 950 /* 951 * Wait for the pageout daemon and other current users to be 952 * done with the object. Note that new paging_in_progress 953 * users can come after this wait, but they must check 954 * OBJ_DEAD flag set (without unlocking the object), and avoid 955 * the object being terminated. 956 */ 957 vm_object_pip_wait(object, "objtrm"); 958 959 KASSERT(object->ref_count == 0, 960 ("vm_object_terminate: object with references, ref_count=%d", 961 object->ref_count)); 962 963 if ((object->flags & OBJ_PG_DTOR) == 0) 964 vm_object_terminate_pages(object); 965 966 #if VM_NRESERVLEVEL > 0 967 if (__predict_false(!LIST_EMPTY(&object->rvq))) 968 vm_reserv_break_all(object); 969 #endif 970 971 KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0, 972 ("%s: non-swap obj %p has cred", __func__, object)); 973 974 /* 975 * Let the pager know object is dead. 976 */ 977 vm_pager_deallocate(object); 978 VM_OBJECT_WUNLOCK(object); 979 980 vm_object_destroy(object); 981 } 982 983 /* 984 * Make the page read-only so that we can clear the object flags. However, if 985 * this is a nosync mmap then the object is likely to stay dirty so do not 986 * mess with the page and do not clear the object flags. Returns TRUE if the 987 * page should be flushed, and FALSE otherwise. 988 */ 989 static boolean_t 990 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) 991 { 992 993 vm_page_assert_busied(p); 994 995 /* 996 * If we have been asked to skip nosync pages and this is a 997 * nosync page, skip it. Note that the object flags were not 998 * cleared in this case so we do not have to set them. 999 */ 1000 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { 1001 *allclean = FALSE; 1002 return (FALSE); 1003 } else { 1004 pmap_remove_write(p); 1005 return (p->dirty != 0); 1006 } 1007 } 1008 1009 /* 1010 * vm_object_page_clean 1011 * 1012 * Clean all dirty pages in the specified range of object. Leaves page 1013 * on whatever queue it is currently on. If NOSYNC is set then do not 1014 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1015 * leaving the object dirty. 1016 * 1017 * For swap objects backing tmpfs regular files, do not flush anything, 1018 * but remove write protection on the mapped pages to update mtime through 1019 * mmaped writes. 1020 * 1021 * When stuffing pages asynchronously, allow clustering. XXX we need a 1022 * synchronous clustering mode implementation. 1023 * 1024 * Odd semantics: if start == end, we clean everything. 1025 * 1026 * The object must be locked. 1027 * 1028 * Returns FALSE if some page from the range was not written, as 1029 * reported by the pager, and TRUE otherwise. 1030 */ 1031 boolean_t 1032 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1033 int flags) 1034 { 1035 struct pctrie_iter pages; 1036 vm_page_t np, p; 1037 vm_pindex_t pi, tend, tstart; 1038 int curgeneration, n, pagerflags; 1039 boolean_t eio, res, allclean; 1040 1041 VM_OBJECT_ASSERT_WLOCKED(object); 1042 1043 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1044 return (TRUE); 1045 1046 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1047 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1048 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1049 1050 tstart = OFF_TO_IDX(start); 1051 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1052 allclean = tstart == 0 && tend >= object->size; 1053 res = TRUE; 1054 vm_page_iter_init(&pages, object); 1055 1056 rescan: 1057 curgeneration = object->generation; 1058 1059 for (p = vm_radix_iter_lookup_ge(&pages, tstart); p != NULL; p = np) { 1060 pi = p->pindex; 1061 if (pi >= tend) 1062 break; 1063 if (vm_page_none_valid(p)) { 1064 np = vm_radix_iter_step(&pages); 1065 continue; 1066 } 1067 if (!vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL)) { 1068 if (object->generation != curgeneration && 1069 (flags & OBJPC_SYNC) != 0) { 1070 pctrie_iter_reset(&pages); 1071 goto rescan; 1072 } 1073 np = vm_radix_iter_lookup_ge(&pages, pi); 1074 continue; 1075 } 1076 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1077 np = vm_radix_iter_step(&pages); 1078 vm_page_xunbusy(p); 1079 continue; 1080 } 1081 if (object->type == OBJT_VNODE) { 1082 n = vm_object_page_collect_flush(&pages, p, pagerflags, 1083 flags, &allclean, &eio); 1084 pctrie_iter_reset(&pages); 1085 if (eio) { 1086 res = FALSE; 1087 allclean = FALSE; 1088 } 1089 if (object->generation != curgeneration && 1090 (flags & OBJPC_SYNC) != 0) 1091 goto rescan; 1092 1093 /* 1094 * If the VOP_PUTPAGES() did a truncated write, so 1095 * that even the first page of the run is not fully 1096 * written, vm_pageout_flush() returns 0 as the run 1097 * length. Since the condition that caused truncated 1098 * write may be permanent, e.g. exhausted free space, 1099 * accepting n == 0 would cause an infinite loop. 1100 * 1101 * Forwarding the iterator leaves the unwritten page 1102 * behind, but there is not much we can do there if 1103 * filesystem refuses to write it. 1104 */ 1105 if (n == 0) { 1106 n = 1; 1107 allclean = FALSE; 1108 } 1109 } else { 1110 n = 1; 1111 vm_page_xunbusy(p); 1112 } 1113 np = vm_radix_iter_lookup_ge(&pages, pi + n); 1114 } 1115 #if 0 1116 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1117 #endif 1118 1119 /* 1120 * Leave updating cleangeneration for tmpfs objects to tmpfs 1121 * scan. It needs to update mtime, which happens for other 1122 * filesystems during page writeouts. 1123 */ 1124 if (allclean && object->type == OBJT_VNODE) 1125 object->cleangeneration = curgeneration; 1126 return (res); 1127 } 1128 1129 static int 1130 vm_object_page_collect_flush(struct pctrie_iter *pages, vm_page_t p, 1131 int pagerflags, int flags, boolean_t *allclean, boolean_t *eio) 1132 { 1133 vm_page_t ma[2 * vm_pageout_page_count - 1]; 1134 int base, count, runlen; 1135 1136 vm_page_lock_assert(p, MA_NOTOWNED); 1137 vm_page_assert_xbusied(p); 1138 base = nitems(ma) / 2; 1139 ma[base] = p; 1140 for (count = 1; count < vm_pageout_page_count; count++) { 1141 p = vm_radix_iter_next(pages); 1142 if (p == NULL || vm_page_tryxbusy(p) == 0) 1143 break; 1144 if (!vm_object_page_remove_write(p, flags, allclean)) { 1145 vm_page_xunbusy(p); 1146 break; 1147 } 1148 ma[base + count] = p; 1149 } 1150 1151 pages->index = ma[base]->pindex; 1152 for (; count < vm_pageout_page_count; count++) { 1153 p = vm_radix_iter_prev(pages); 1154 if (p == NULL || vm_page_tryxbusy(p) == 0) 1155 break; 1156 if (!vm_object_page_remove_write(p, flags, allclean)) { 1157 vm_page_xunbusy(p); 1158 break; 1159 } 1160 ma[--base] = p; 1161 } 1162 1163 vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base, 1164 &runlen, eio); 1165 return (runlen); 1166 } 1167 1168 /* 1169 * Note that there is absolutely no sense in writing out 1170 * anonymous objects, so we track down the vnode object 1171 * to write out. 1172 * We invalidate (remove) all pages from the address space 1173 * for semantic correctness. 1174 * 1175 * If the backing object is a device object with unmanaged pages, then any 1176 * mappings to the specified range of pages must be removed before this 1177 * function is called. 1178 * 1179 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1180 * may start out with a NULL object. 1181 */ 1182 boolean_t 1183 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1184 boolean_t syncio, boolean_t invalidate) 1185 { 1186 vm_object_t backing_object; 1187 struct vnode *vp; 1188 struct mount *mp; 1189 int error, flags, fsync_after; 1190 boolean_t res; 1191 1192 if (object == NULL) 1193 return (TRUE); 1194 res = TRUE; 1195 error = 0; 1196 VM_OBJECT_WLOCK(object); 1197 while ((backing_object = object->backing_object) != NULL) { 1198 VM_OBJECT_WLOCK(backing_object); 1199 offset += object->backing_object_offset; 1200 VM_OBJECT_WUNLOCK(object); 1201 object = backing_object; 1202 if (object->size < OFF_TO_IDX(offset + size)) 1203 size = IDX_TO_OFF(object->size) - offset; 1204 } 1205 /* 1206 * Flush pages if writing is allowed, invalidate them 1207 * if invalidation requested. Pages undergoing I/O 1208 * will be ignored by vm_object_page_remove(). 1209 * 1210 * We cannot lock the vnode and then wait for paging 1211 * to complete without deadlocking against vm_fault. 1212 * Instead we simply call vm_object_page_remove() and 1213 * allow it to block internally on a page-by-page 1214 * basis when it encounters pages undergoing async 1215 * I/O. 1216 */ 1217 if (object->type == OBJT_VNODE && 1218 vm_object_mightbedirty(object) != 0 && 1219 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1220 VM_OBJECT_WUNLOCK(object); 1221 (void)vn_start_write(vp, &mp, V_WAIT); 1222 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1223 if (syncio && !invalidate && offset == 0 && 1224 atop(size) == object->size) { 1225 /* 1226 * If syncing the whole mapping of the file, 1227 * it is faster to schedule all the writes in 1228 * async mode, also allowing the clustering, 1229 * and then wait for i/o to complete. 1230 */ 1231 flags = 0; 1232 fsync_after = TRUE; 1233 } else { 1234 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1235 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1236 fsync_after = FALSE; 1237 } 1238 VM_OBJECT_WLOCK(object); 1239 res = vm_object_page_clean(object, offset, offset + size, 1240 flags); 1241 VM_OBJECT_WUNLOCK(object); 1242 if (fsync_after) { 1243 for (;;) { 1244 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1245 if (error != ERELOOKUP) 1246 break; 1247 1248 /* 1249 * Allow SU/bufdaemon to handle more 1250 * dependencies in the meantime. 1251 */ 1252 VOP_UNLOCK(vp); 1253 vn_finished_write(mp); 1254 1255 (void)vn_start_write(vp, &mp, V_WAIT); 1256 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1257 } 1258 } 1259 VOP_UNLOCK(vp); 1260 vn_finished_write(mp); 1261 if (error != 0) 1262 res = FALSE; 1263 VM_OBJECT_WLOCK(object); 1264 } 1265 if ((object->type == OBJT_VNODE || 1266 object->type == OBJT_DEVICE) && invalidate) { 1267 if (object->type == OBJT_DEVICE) 1268 /* 1269 * The option OBJPR_NOTMAPPED must be passed here 1270 * because vm_object_page_remove() cannot remove 1271 * unmanaged mappings. 1272 */ 1273 flags = OBJPR_NOTMAPPED; 1274 else if (old_msync) 1275 flags = 0; 1276 else 1277 flags = OBJPR_CLEANONLY; 1278 vm_object_page_remove(object, OFF_TO_IDX(offset), 1279 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1280 } 1281 VM_OBJECT_WUNLOCK(object); 1282 return (res); 1283 } 1284 1285 /* 1286 * Determine whether the given advice can be applied to the object. Advice is 1287 * not applied to unmanaged pages since they never belong to page queues, and 1288 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1289 * have been mapped at most once. 1290 */ 1291 static bool 1292 vm_object_advice_applies(vm_object_t object, int advice) 1293 { 1294 1295 if ((object->flags & OBJ_UNMANAGED) != 0) 1296 return (false); 1297 if (advice != MADV_FREE) 1298 return (true); 1299 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1300 (OBJ_ONEMAPPING | OBJ_ANON)); 1301 } 1302 1303 static void 1304 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1305 vm_size_t size) 1306 { 1307 1308 if (advice == MADV_FREE) 1309 vm_pager_freespace(object, pindex, size); 1310 } 1311 1312 /* 1313 * vm_object_madvise: 1314 * 1315 * Implements the madvise function at the object/page level. 1316 * 1317 * MADV_WILLNEED (any object) 1318 * 1319 * Activate the specified pages if they are resident. 1320 * 1321 * MADV_DONTNEED (any object) 1322 * 1323 * Deactivate the specified pages if they are resident. 1324 * 1325 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only) 1326 * 1327 * Deactivate and clean the specified pages if they are 1328 * resident. This permits the process to reuse the pages 1329 * without faulting or the kernel to reclaim the pages 1330 * without I/O. 1331 */ 1332 void 1333 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1334 int advice) 1335 { 1336 vm_pindex_t tpindex; 1337 vm_object_t backing_object, tobject; 1338 vm_page_t m, tm; 1339 1340 if (object == NULL) 1341 return; 1342 1343 relookup: 1344 VM_OBJECT_WLOCK(object); 1345 if (!vm_object_advice_applies(object, advice)) { 1346 VM_OBJECT_WUNLOCK(object); 1347 return; 1348 } 1349 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { 1350 tobject = object; 1351 1352 /* 1353 * If the next page isn't resident in the top-level object, we 1354 * need to search the shadow chain. When applying MADV_FREE, we 1355 * take care to release any swap space used to store 1356 * non-resident pages. 1357 */ 1358 if (m == NULL || pindex < m->pindex) { 1359 /* 1360 * Optimize a common case: if the top-level object has 1361 * no backing object, we can skip over the non-resident 1362 * range in constant time. 1363 */ 1364 if (object->backing_object == NULL) { 1365 tpindex = (m != NULL && m->pindex < end) ? 1366 m->pindex : end; 1367 vm_object_madvise_freespace(object, advice, 1368 pindex, tpindex - pindex); 1369 if ((pindex = tpindex) == end) 1370 break; 1371 goto next_page; 1372 } 1373 1374 tpindex = pindex; 1375 do { 1376 vm_object_madvise_freespace(tobject, advice, 1377 tpindex, 1); 1378 /* 1379 * Prepare to search the next object in the 1380 * chain. 1381 */ 1382 backing_object = tobject->backing_object; 1383 if (backing_object == NULL) 1384 goto next_pindex; 1385 VM_OBJECT_WLOCK(backing_object); 1386 tpindex += 1387 OFF_TO_IDX(tobject->backing_object_offset); 1388 if (tobject != object) 1389 VM_OBJECT_WUNLOCK(tobject); 1390 tobject = backing_object; 1391 if (!vm_object_advice_applies(tobject, advice)) 1392 goto next_pindex; 1393 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1394 NULL); 1395 } else { 1396 next_page: 1397 tm = m; 1398 m = TAILQ_NEXT(m, listq); 1399 } 1400 1401 /* 1402 * If the page is not in a normal state, skip it. The page 1403 * can not be invalidated while the object lock is held. 1404 */ 1405 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1406 goto next_pindex; 1407 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1408 ("vm_object_madvise: page %p is fictitious", tm)); 1409 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1410 ("vm_object_madvise: page %p is not managed", tm)); 1411 if (vm_page_tryxbusy(tm) == 0) { 1412 if (object != tobject) 1413 VM_OBJECT_WUNLOCK(object); 1414 if (advice == MADV_WILLNEED) { 1415 /* 1416 * Reference the page before unlocking and 1417 * sleeping so that the page daemon is less 1418 * likely to reclaim it. 1419 */ 1420 vm_page_aflag_set(tm, PGA_REFERENCED); 1421 } 1422 if (!vm_page_busy_sleep(tm, "madvpo", 0)) 1423 VM_OBJECT_WUNLOCK(tobject); 1424 goto relookup; 1425 } 1426 vm_page_advise(tm, advice); 1427 vm_page_xunbusy(tm); 1428 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1429 next_pindex: 1430 if (tobject != object) 1431 VM_OBJECT_WUNLOCK(tobject); 1432 } 1433 VM_OBJECT_WUNLOCK(object); 1434 } 1435 1436 /* 1437 * vm_object_shadow: 1438 * 1439 * Create a new object which is backed by the 1440 * specified existing object range. The source 1441 * object reference is deallocated. 1442 * 1443 * The new object and offset into that object 1444 * are returned in the source parameters. 1445 */ 1446 void 1447 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1448 struct ucred *cred, bool shared) 1449 { 1450 vm_object_t source; 1451 vm_object_t result; 1452 1453 source = *object; 1454 1455 /* 1456 * Don't create the new object if the old object isn't shared. 1457 * 1458 * If we hold the only reference we can guarantee that it won't 1459 * increase while we have the map locked. Otherwise the race is 1460 * harmless and we will end up with an extra shadow object that 1461 * will be collapsed later. 1462 */ 1463 if (source != NULL && source->ref_count == 1 && 1464 (source->flags & OBJ_ANON) != 0) 1465 return; 1466 1467 /* 1468 * Allocate a new object with the given length. 1469 */ 1470 result = vm_object_allocate_anon(atop(length), source, cred, length); 1471 1472 /* 1473 * Store the offset into the source object, and fix up the offset into 1474 * the new object. 1475 */ 1476 result->backing_object_offset = *offset; 1477 1478 if (shared || source != NULL) { 1479 VM_OBJECT_WLOCK(result); 1480 1481 /* 1482 * The new object shadows the source object, adding a 1483 * reference to it. Our caller changes his reference 1484 * to point to the new object, removing a reference to 1485 * the source object. Net result: no change of 1486 * reference count, unless the caller needs to add one 1487 * more reference due to forking a shared map entry. 1488 */ 1489 if (shared) { 1490 vm_object_reference_locked(result); 1491 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1492 } 1493 1494 /* 1495 * Try to optimize the result object's page color when 1496 * shadowing in order to maintain page coloring 1497 * consistency in the combined shadowed object. 1498 */ 1499 if (source != NULL) { 1500 vm_object_backing_insert(result, source); 1501 result->domain = source->domain; 1502 #if VM_NRESERVLEVEL > 0 1503 vm_object_set_flag(result, 1504 (source->flags & OBJ_COLORED)); 1505 result->pg_color = (source->pg_color + 1506 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1507 1)) - 1); 1508 #endif 1509 } 1510 VM_OBJECT_WUNLOCK(result); 1511 } 1512 1513 /* 1514 * Return the new things 1515 */ 1516 *offset = 0; 1517 *object = result; 1518 } 1519 1520 /* 1521 * vm_object_split: 1522 * 1523 * Split the pages in a map entry into a new object. This affords 1524 * easier removal of unused pages, and keeps object inheritance from 1525 * being a negative impact on memory usage. 1526 */ 1527 void 1528 vm_object_split(vm_map_entry_t entry) 1529 { 1530 struct pctrie_iter pages; 1531 vm_page_t m; 1532 vm_object_t orig_object, new_object, backing_object; 1533 vm_pindex_t offidxstart; 1534 vm_size_t size; 1535 1536 orig_object = entry->object.vm_object; 1537 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1538 ("vm_object_split: Splitting object with multiple mappings.")); 1539 if ((orig_object->flags & OBJ_ANON) == 0) 1540 return; 1541 if (orig_object->ref_count <= 1) 1542 return; 1543 VM_OBJECT_WUNLOCK(orig_object); 1544 1545 offidxstart = OFF_TO_IDX(entry->offset); 1546 size = atop(entry->end - entry->start); 1547 1548 new_object = vm_object_allocate_anon(size, orig_object, 1549 orig_object->cred, ptoa(size)); 1550 1551 /* 1552 * We must wait for the orig_object to complete any in-progress 1553 * collapse so that the swap blocks are stable below. The 1554 * additional reference on backing_object by new object will 1555 * prevent further collapse operations until split completes. 1556 */ 1557 VM_OBJECT_WLOCK(orig_object); 1558 vm_object_collapse_wait(orig_object); 1559 1560 /* 1561 * At this point, the new object is still private, so the order in 1562 * which the original and new objects are locked does not matter. 1563 */ 1564 VM_OBJECT_WLOCK(new_object); 1565 new_object->domain = orig_object->domain; 1566 backing_object = orig_object->backing_object; 1567 if (backing_object != NULL) { 1568 vm_object_backing_insert_ref(new_object, backing_object); 1569 new_object->backing_object_offset = 1570 orig_object->backing_object_offset + entry->offset; 1571 } 1572 if (orig_object->cred != NULL) { 1573 crhold(orig_object->cred); 1574 KASSERT(orig_object->charge >= ptoa(size), 1575 ("orig_object->charge < 0")); 1576 orig_object->charge -= ptoa(size); 1577 } 1578 1579 /* 1580 * Mark the split operation so that swap_pager_getpages() knows 1581 * that the object is in transition. 1582 */ 1583 vm_object_set_flag(orig_object, OBJ_SPLIT); 1584 vm_page_iter_limit_init(&pages, orig_object, offidxstart + size); 1585 retry: 1586 KASSERT(pctrie_iter_is_reset(&pages), 1587 ("%s: pctrie_iter not reset for retry", __func__)); 1588 for (m = vm_radix_iter_lookup_ge(&pages, offidxstart); m != NULL; 1589 m = vm_radix_iter_step(&pages)) { 1590 /* 1591 * We must wait for pending I/O to complete before we can 1592 * rename the page. 1593 * 1594 * We do not have to VM_PROT_NONE the page as mappings should 1595 * not be changed by this operation. 1596 */ 1597 if (vm_page_tryxbusy(m) == 0) { 1598 VM_OBJECT_WUNLOCK(new_object); 1599 if (vm_page_busy_sleep(m, "spltwt", 0)) 1600 VM_OBJECT_WLOCK(orig_object); 1601 pctrie_iter_reset(&pages); 1602 VM_OBJECT_WLOCK(new_object); 1603 goto retry; 1604 } 1605 1606 /* 1607 * If the page was left invalid, it was likely placed there by 1608 * an incomplete fault. Just remove and ignore. 1609 * 1610 * One other possibility is that the map entry is wired, in 1611 * which case we must hang on to the page to avoid leaking it, 1612 * as the map entry owns the wiring. This case can arise if the 1613 * backing object is truncated by the pager. 1614 */ 1615 if (vm_page_none_valid(m) && entry->wired_count == 0) { 1616 if (vm_page_iter_remove(&pages, m)) 1617 vm_page_free(m); 1618 continue; 1619 } 1620 1621 /* vm_page_iter_rename() will dirty the page if it is valid. */ 1622 if (!vm_page_iter_rename(&pages, m, new_object, m->pindex - 1623 offidxstart)) { 1624 vm_page_xunbusy(m); 1625 VM_OBJECT_WUNLOCK(new_object); 1626 VM_OBJECT_WUNLOCK(orig_object); 1627 vm_radix_wait(); 1628 pctrie_iter_reset(&pages); 1629 VM_OBJECT_WLOCK(orig_object); 1630 VM_OBJECT_WLOCK(new_object); 1631 goto retry; 1632 } 1633 1634 #if VM_NRESERVLEVEL > 0 1635 /* 1636 * If some of the reservation's allocated pages remain with 1637 * the original object, then transferring the reservation to 1638 * the new object is neither particularly beneficial nor 1639 * particularly harmful as compared to leaving the reservation 1640 * with the original object. If, however, all of the 1641 * reservation's allocated pages are transferred to the new 1642 * object, then transferring the reservation is typically 1643 * beneficial. Determining which of these two cases applies 1644 * would be more costly than unconditionally renaming the 1645 * reservation. 1646 */ 1647 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1648 #endif 1649 } 1650 1651 /* 1652 * swap_pager_copy() can sleep, in which case the orig_object's 1653 * and new_object's locks are released and reacquired. 1654 */ 1655 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1656 1657 TAILQ_FOREACH(m, &new_object->memq, listq) 1658 vm_page_xunbusy(m); 1659 1660 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1661 VM_OBJECT_WUNLOCK(orig_object); 1662 VM_OBJECT_WUNLOCK(new_object); 1663 entry->object.vm_object = new_object; 1664 entry->offset = 0LL; 1665 vm_object_deallocate(orig_object); 1666 VM_OBJECT_WLOCK(new_object); 1667 } 1668 1669 static vm_page_t 1670 vm_object_collapse_scan_wait(struct pctrie_iter *pages, vm_object_t object, 1671 vm_page_t p) 1672 { 1673 vm_object_t backing_object; 1674 1675 VM_OBJECT_ASSERT_WLOCKED(object); 1676 backing_object = object->backing_object; 1677 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1678 1679 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1680 ("invalid ownership %p %p %p", p, object, backing_object)); 1681 /* The page is only NULL when rename fails. */ 1682 if (p == NULL) { 1683 VM_OBJECT_WUNLOCK(object); 1684 VM_OBJECT_WUNLOCK(backing_object); 1685 vm_radix_wait(); 1686 VM_OBJECT_WLOCK(object); 1687 } else if (p->object == object) { 1688 VM_OBJECT_WUNLOCK(backing_object); 1689 if (vm_page_busy_sleep(p, "vmocol", 0)) 1690 VM_OBJECT_WLOCK(object); 1691 } else { 1692 VM_OBJECT_WUNLOCK(object); 1693 if (!vm_page_busy_sleep(p, "vmocol", 0)) 1694 VM_OBJECT_WUNLOCK(backing_object); 1695 VM_OBJECT_WLOCK(object); 1696 } 1697 VM_OBJECT_WLOCK(backing_object); 1698 vm_page_iter_init(pages, backing_object); 1699 return (vm_radix_iter_lookup_ge(pages, 0)); 1700 } 1701 1702 static void 1703 vm_object_collapse_scan(vm_object_t object) 1704 { 1705 struct pctrie_iter pages; 1706 vm_object_t backing_object; 1707 vm_page_t next, p, pp; 1708 vm_pindex_t backing_offset_index, new_pindex; 1709 1710 VM_OBJECT_ASSERT_WLOCKED(object); 1711 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1712 1713 backing_object = object->backing_object; 1714 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1715 1716 /* 1717 * Our scan 1718 */ 1719 vm_page_iter_init(&pages, backing_object); 1720 for (p = vm_radix_iter_lookup_ge(&pages, 0); p != NULL; p = next) { 1721 /* 1722 * Check for busy page 1723 */ 1724 if (vm_page_tryxbusy(p) == 0) { 1725 next = vm_object_collapse_scan_wait(&pages, object, p); 1726 continue; 1727 } 1728 1729 KASSERT(object->backing_object == backing_object, 1730 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1731 object->backing_object, backing_object)); 1732 KASSERT(p->object == backing_object, 1733 ("vm_object_collapse_scan: object mismatch %p != %p", 1734 p->object, backing_object)); 1735 1736 if (p->pindex < backing_offset_index || object->size <= 1737 (new_pindex = p->pindex - backing_offset_index)) { 1738 vm_pager_freespace(backing_object, p->pindex, 1); 1739 1740 KASSERT(!pmap_page_is_mapped(p), 1741 ("freeing mapped page %p", p)); 1742 if (vm_page_iter_remove(&pages, p)) 1743 vm_page_free(p); 1744 next = vm_radix_iter_step(&pages); 1745 continue; 1746 } 1747 1748 if (!vm_page_all_valid(p)) { 1749 KASSERT(!pmap_page_is_mapped(p), 1750 ("freeing mapped page %p", p)); 1751 if (vm_page_iter_remove(&pages, p)) 1752 vm_page_free(p); 1753 next = vm_radix_iter_step(&pages); 1754 continue; 1755 } 1756 1757 pp = vm_page_lookup(object, new_pindex); 1758 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1759 vm_page_xunbusy(p); 1760 /* 1761 * The page in the parent is busy and possibly not 1762 * (yet) valid. Until its state is finalized by the 1763 * busy bit owner, we can't tell whether it shadows the 1764 * original page. 1765 */ 1766 next = vm_object_collapse_scan_wait(&pages, object, pp); 1767 continue; 1768 } 1769 1770 if (pp != NULL && vm_page_none_valid(pp)) { 1771 /* 1772 * The page was invalid in the parent. Likely placed 1773 * there by an incomplete fault. Just remove and 1774 * ignore. p can replace it. 1775 */ 1776 if (vm_page_remove(pp)) 1777 vm_page_free(pp); 1778 pp = NULL; 1779 } 1780 1781 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1782 NULL)) { 1783 /* 1784 * The page already exists in the parent OR swap exists 1785 * for this location in the parent. Leave the parent's 1786 * page alone. Destroy the original page from the 1787 * backing object. 1788 */ 1789 vm_pager_freespace(backing_object, p->pindex, 1); 1790 KASSERT(!pmap_page_is_mapped(p), 1791 ("freeing mapped page %p", p)); 1792 if (pp != NULL) 1793 vm_page_xunbusy(pp); 1794 if (vm_page_iter_remove(&pages, p)) 1795 vm_page_free(p); 1796 next = vm_radix_iter_step(&pages); 1797 continue; 1798 } 1799 1800 /* 1801 * Page does not exist in parent, rename the page from the 1802 * backing object to the main object. 1803 * 1804 * If the page was mapped to a process, it can remain mapped 1805 * through the rename. vm_page_iter_rename() will dirty the 1806 * page. 1807 */ 1808 if (!vm_page_iter_rename(&pages, p, object, new_pindex)) { 1809 vm_page_xunbusy(p); 1810 next = vm_object_collapse_scan_wait(&pages, object, 1811 NULL); 1812 continue; 1813 } 1814 1815 /* Use the old pindex to free the right page. */ 1816 vm_pager_freespace(backing_object, new_pindex + 1817 backing_offset_index, 1); 1818 1819 #if VM_NRESERVLEVEL > 0 1820 /* 1821 * Rename the reservation. 1822 */ 1823 vm_reserv_rename(p, object, backing_object, 1824 backing_offset_index); 1825 #endif 1826 vm_page_xunbusy(p); 1827 next = vm_radix_iter_step(&pages); 1828 } 1829 return; 1830 } 1831 1832 /* 1833 * vm_object_collapse: 1834 * 1835 * Collapse an object with the object backing it. 1836 * Pages in the backing object are moved into the 1837 * parent, and the backing object is deallocated. 1838 */ 1839 void 1840 vm_object_collapse(vm_object_t object) 1841 { 1842 vm_object_t backing_object, new_backing_object; 1843 1844 VM_OBJECT_ASSERT_WLOCKED(object); 1845 1846 while (TRUE) { 1847 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1848 ("collapsing invalid object")); 1849 1850 /* 1851 * Wait for the backing_object to finish any pending 1852 * collapse so that the caller sees the shortest possible 1853 * shadow chain. 1854 */ 1855 backing_object = vm_object_backing_collapse_wait(object); 1856 if (backing_object == NULL) 1857 return; 1858 1859 KASSERT(object->ref_count > 0 && 1860 object->ref_count > atomic_load_int(&object->shadow_count), 1861 ("collapse with invalid ref %d or shadow %d count.", 1862 object->ref_count, atomic_load_int(&object->shadow_count))); 1863 KASSERT((backing_object->flags & 1864 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1865 ("vm_object_collapse: Backing object already collapsing.")); 1866 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1867 ("vm_object_collapse: object is already collapsing.")); 1868 1869 /* 1870 * We know that we can either collapse the backing object if 1871 * the parent is the only reference to it, or (perhaps) have 1872 * the parent bypass the object if the parent happens to shadow 1873 * all the resident pages in the entire backing object. 1874 */ 1875 if (backing_object->ref_count == 1) { 1876 KASSERT(atomic_load_int(&backing_object->shadow_count) 1877 == 1, 1878 ("vm_object_collapse: shadow_count: %d", 1879 atomic_load_int(&backing_object->shadow_count))); 1880 vm_object_pip_add(object, 1); 1881 vm_object_set_flag(object, OBJ_COLLAPSING); 1882 vm_object_pip_add(backing_object, 1); 1883 vm_object_set_flag(backing_object, OBJ_DEAD); 1884 1885 /* 1886 * If there is exactly one reference to the backing 1887 * object, we can collapse it into the parent. 1888 */ 1889 vm_object_collapse_scan(object); 1890 1891 /* 1892 * Move the pager from backing_object to object. 1893 * 1894 * swap_pager_copy() can sleep, in which case the 1895 * backing_object's and object's locks are released and 1896 * reacquired. 1897 */ 1898 swap_pager_copy(backing_object, object, 1899 OFF_TO_IDX(object->backing_object_offset), TRUE); 1900 1901 /* 1902 * Object now shadows whatever backing_object did. 1903 */ 1904 vm_object_clear_flag(object, OBJ_COLLAPSING); 1905 vm_object_backing_transfer(object, backing_object); 1906 object->backing_object_offset += 1907 backing_object->backing_object_offset; 1908 VM_OBJECT_WUNLOCK(object); 1909 vm_object_pip_wakeup(object); 1910 1911 /* 1912 * Discard backing_object. 1913 * 1914 * Since the backing object has no pages, no pager left, 1915 * and no object references within it, all that is 1916 * necessary is to dispose of it. 1917 */ 1918 KASSERT(backing_object->ref_count == 1, ( 1919 "backing_object %p was somehow re-referenced during collapse!", 1920 backing_object)); 1921 vm_object_pip_wakeup(backing_object); 1922 (void)refcount_release(&backing_object->ref_count); 1923 umtx_shm_object_terminated(backing_object); 1924 vm_object_terminate(backing_object); 1925 counter_u64_add(object_collapses, 1); 1926 VM_OBJECT_WLOCK(object); 1927 } else { 1928 /* 1929 * If we do not entirely shadow the backing object, 1930 * there is nothing we can do so we give up. 1931 * 1932 * The object lock and backing_object lock must not 1933 * be dropped during this sequence. 1934 */ 1935 if (!swap_pager_scan_all_shadowed(object)) { 1936 VM_OBJECT_WUNLOCK(backing_object); 1937 break; 1938 } 1939 1940 /* 1941 * Make the parent shadow the next object in the 1942 * chain. Deallocating backing_object will not remove 1943 * it, since its reference count is at least 2. 1944 */ 1945 vm_object_backing_remove_locked(object); 1946 new_backing_object = backing_object->backing_object; 1947 if (new_backing_object != NULL) { 1948 vm_object_backing_insert_ref(object, 1949 new_backing_object); 1950 object->backing_object_offset += 1951 backing_object->backing_object_offset; 1952 } 1953 1954 /* 1955 * Drop the reference count on backing_object. Since 1956 * its ref_count was at least 2, it will not vanish. 1957 */ 1958 (void)refcount_release(&backing_object->ref_count); 1959 KASSERT(backing_object->ref_count >= 1, ( 1960 "backing_object %p was somehow dereferenced during collapse!", 1961 backing_object)); 1962 VM_OBJECT_WUNLOCK(backing_object); 1963 counter_u64_add(object_bypasses, 1); 1964 } 1965 1966 /* 1967 * Try again with this object's new backing object. 1968 */ 1969 } 1970 } 1971 1972 /* 1973 * vm_object_page_remove: 1974 * 1975 * For the given object, either frees or invalidates each of the 1976 * specified pages. In general, a page is freed. However, if a page is 1977 * wired for any reason other than the existence of a managed, wired 1978 * mapping, then it may be invalidated but not removed from the object. 1979 * Pages are specified by the given range ["start", "end") and the option 1980 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1981 * extends from "start" to the end of the object. If the option 1982 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1983 * specified range are affected. If the option OBJPR_NOTMAPPED is 1984 * specified, then the pages within the specified range must have no 1985 * mappings. Otherwise, if this option is not specified, any mappings to 1986 * the specified pages are removed before the pages are freed or 1987 * invalidated. 1988 * 1989 * In general, this operation should only be performed on objects that 1990 * contain managed pages. There are, however, two exceptions. First, it 1991 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1992 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1993 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1994 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1995 * 1996 * The object must be locked. 1997 */ 1998 void 1999 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 2000 int options) 2001 { 2002 struct pctrie_iter pages; 2003 vm_page_t p; 2004 2005 VM_OBJECT_ASSERT_WLOCKED(object); 2006 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 2007 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 2008 ("vm_object_page_remove: illegal options for object %p", object)); 2009 if (object->resident_page_count == 0) 2010 return; 2011 vm_object_pip_add(object, 1); 2012 vm_page_iter_limit_init(&pages, object, end); 2013 again: 2014 KASSERT(pctrie_iter_is_reset(&pages), 2015 ("%s: pctrie_iter not reset for retry", __func__)); 2016 for (p = vm_radix_iter_lookup_ge(&pages, start); p != NULL; 2017 p = vm_radix_iter_step(&pages)) { 2018 /* 2019 * Skip invalid pages if asked to do so. Try to avoid acquiring 2020 * the busy lock, as some consumers rely on this to avoid 2021 * deadlocks. 2022 * 2023 * A thread may concurrently transition the page from invalid to 2024 * valid using only the busy lock, so the result of this check 2025 * is immediately stale. It is up to consumers to handle this, 2026 * for instance by ensuring that all invalid->valid transitions 2027 * happen with a mutex held, as may be possible for a 2028 * filesystem. 2029 */ 2030 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) 2031 continue; 2032 2033 /* 2034 * If the page is wired for any reason besides the existence 2035 * of managed, wired mappings, then it cannot be freed. For 2036 * example, fictitious pages, which represent device memory, 2037 * are inherently wired and cannot be freed. They can, 2038 * however, be invalidated if the option OBJPR_CLEANONLY is 2039 * not specified. 2040 */ 2041 if (vm_page_tryxbusy(p) == 0) { 2042 if (vm_page_busy_sleep(p, "vmopar", 0)) 2043 VM_OBJECT_WLOCK(object); 2044 pctrie_iter_reset(&pages); 2045 goto again; 2046 } 2047 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) { 2048 vm_page_xunbusy(p); 2049 continue; 2050 } 2051 if (vm_page_wired(p)) { 2052 wired: 2053 if ((options & OBJPR_NOTMAPPED) == 0 && 2054 object->ref_count != 0) 2055 pmap_remove_all(p); 2056 if ((options & OBJPR_CLEANONLY) == 0) { 2057 vm_page_invalid(p); 2058 vm_page_undirty(p); 2059 } 2060 vm_page_xunbusy(p); 2061 continue; 2062 } 2063 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2064 ("vm_object_page_remove: page %p is fictitious", p)); 2065 if ((options & OBJPR_CLEANONLY) != 0 && 2066 !vm_page_none_valid(p)) { 2067 if ((options & OBJPR_NOTMAPPED) == 0 && 2068 object->ref_count != 0 && 2069 !vm_page_try_remove_write(p)) 2070 goto wired; 2071 if (p->dirty != 0) { 2072 vm_page_xunbusy(p); 2073 continue; 2074 } 2075 } 2076 if ((options & OBJPR_NOTMAPPED) == 0 && 2077 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2078 goto wired; 2079 vm_page_iter_free(&pages, p); 2080 } 2081 vm_object_pip_wakeup(object); 2082 2083 vm_pager_freespace(object, start, (end == 0 ? object->size : end) - 2084 start); 2085 } 2086 2087 /* 2088 * vm_object_page_noreuse: 2089 * 2090 * For the given object, attempt to move the specified pages to 2091 * the head of the inactive queue. This bypasses regular LRU 2092 * operation and allows the pages to be reused quickly under memory 2093 * pressure. If a page is wired for any reason, then it will not 2094 * be queued. Pages are specified by the range ["start", "end"). 2095 * As a special case, if "end" is zero, then the range extends from 2096 * "start" to the end of the object. 2097 * 2098 * This operation should only be performed on objects that 2099 * contain non-fictitious, managed pages. 2100 * 2101 * The object must be locked. 2102 */ 2103 void 2104 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2105 { 2106 vm_page_t p, next; 2107 2108 VM_OBJECT_ASSERT_LOCKED(object); 2109 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2110 ("vm_object_page_noreuse: illegal object %p", object)); 2111 if (object->resident_page_count == 0) 2112 return; 2113 p = vm_page_find_least(object, start); 2114 2115 /* 2116 * Here, the variable "p" is either (1) the page with the least pindex 2117 * greater than or equal to the parameter "start" or (2) NULL. 2118 */ 2119 for (; p != NULL && (p->pindex < end || end == 0); p = next) { 2120 next = TAILQ_NEXT(p, listq); 2121 vm_page_deactivate_noreuse(p); 2122 } 2123 } 2124 2125 /* 2126 * Populate the specified range of the object with valid pages. Returns 2127 * TRUE if the range is successfully populated and FALSE otherwise. 2128 * 2129 * Note: This function should be optimized to pass a larger array of 2130 * pages to vm_pager_get_pages() before it is applied to a non- 2131 * OBJT_DEVICE object. 2132 * 2133 * The object must be locked. 2134 */ 2135 boolean_t 2136 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2137 { 2138 vm_page_t m; 2139 vm_pindex_t pindex; 2140 int rv; 2141 2142 VM_OBJECT_ASSERT_WLOCKED(object); 2143 for (pindex = start; pindex < end; pindex++) { 2144 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); 2145 if (rv != VM_PAGER_OK) 2146 break; 2147 2148 /* 2149 * Keep "m" busy because a subsequent iteration may unlock 2150 * the object. 2151 */ 2152 } 2153 if (pindex > start) { 2154 m = vm_page_lookup(object, start); 2155 while (m != NULL && m->pindex < pindex) { 2156 vm_page_xunbusy(m); 2157 m = TAILQ_NEXT(m, listq); 2158 } 2159 } 2160 return (pindex == end); 2161 } 2162 2163 /* 2164 * Routine: vm_object_coalesce 2165 * Function: Coalesces two objects backing up adjoining 2166 * regions of memory into a single object. 2167 * 2168 * returns TRUE if objects were combined. 2169 * 2170 * NOTE: Only works at the moment if the second object is NULL - 2171 * if it's not, which object do we lock first? 2172 * 2173 * Parameters: 2174 * prev_object First object to coalesce 2175 * prev_offset Offset into prev_object 2176 * prev_size Size of reference to prev_object 2177 * next_size Size of reference to the second object 2178 * reserved Indicator that extension region has 2179 * swap accounted for 2180 * 2181 * Conditions: 2182 * The object must *not* be locked. 2183 */ 2184 boolean_t 2185 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2186 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2187 { 2188 vm_pindex_t next_pindex; 2189 2190 if (prev_object == NULL) 2191 return (TRUE); 2192 if ((prev_object->flags & OBJ_ANON) == 0) 2193 return (FALSE); 2194 2195 VM_OBJECT_WLOCK(prev_object); 2196 /* 2197 * Try to collapse the object first. 2198 */ 2199 vm_object_collapse(prev_object); 2200 2201 /* 2202 * Can't coalesce if: . more than one reference . paged out . shadows 2203 * another object . has a copy elsewhere (any of which mean that the 2204 * pages not mapped to prev_entry may be in use anyway) 2205 */ 2206 if (prev_object->backing_object != NULL) { 2207 VM_OBJECT_WUNLOCK(prev_object); 2208 return (FALSE); 2209 } 2210 2211 prev_size >>= PAGE_SHIFT; 2212 next_size >>= PAGE_SHIFT; 2213 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2214 2215 if (prev_object->ref_count > 1 && 2216 prev_object->size != next_pindex && 2217 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2218 VM_OBJECT_WUNLOCK(prev_object); 2219 return (FALSE); 2220 } 2221 2222 /* 2223 * Account for the charge. 2224 */ 2225 if (prev_object->cred != NULL) { 2226 /* 2227 * If prev_object was charged, then this mapping, 2228 * although not charged now, may become writable 2229 * later. Non-NULL cred in the object would prevent 2230 * swap reservation during enabling of the write 2231 * access, so reserve swap now. Failed reservation 2232 * cause allocation of the separate object for the map 2233 * entry, and swap reservation for this entry is 2234 * managed in appropriate time. 2235 */ 2236 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2237 prev_object->cred)) { 2238 VM_OBJECT_WUNLOCK(prev_object); 2239 return (FALSE); 2240 } 2241 prev_object->charge += ptoa(next_size); 2242 } 2243 2244 /* 2245 * Remove any pages that may still be in the object from a previous 2246 * deallocation. 2247 */ 2248 if (next_pindex < prev_object->size) { 2249 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2250 next_size, 0); 2251 #if 0 2252 if (prev_object->cred != NULL) { 2253 KASSERT(prev_object->charge >= 2254 ptoa(prev_object->size - next_pindex), 2255 ("object %p overcharged 1 %jx %jx", prev_object, 2256 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2257 prev_object->charge -= ptoa(prev_object->size - 2258 next_pindex); 2259 } 2260 #endif 2261 } 2262 2263 /* 2264 * Extend the object if necessary. 2265 */ 2266 if (next_pindex + next_size > prev_object->size) 2267 prev_object->size = next_pindex + next_size; 2268 2269 VM_OBJECT_WUNLOCK(prev_object); 2270 return (TRUE); 2271 } 2272 2273 /* 2274 * Fill in the m_dst array with up to *rbehind optional pages before m_src[0] 2275 * and up to *rahead optional pages after m_src[count - 1]. In both cases, stop 2276 * the filling-in short on encountering a cached page, an object boundary limit, 2277 * or an allocation error. Update *rbehind and *rahead to indicate the number 2278 * of pages allocated. Copy elements of m_src into array elements from 2279 * m_dst[*rbehind] to m_dst[*rbehind + count -1]. 2280 */ 2281 void 2282 vm_object_prepare_buf_pages(vm_object_t object, vm_page_t *ma_dst, int count, 2283 int *rbehind, int *rahead, vm_page_t *ma_src) 2284 { 2285 struct pctrie_iter pages; 2286 vm_pindex_t pindex; 2287 vm_page_t m, mpred, msucc; 2288 2289 vm_page_iter_init(&pages, object); 2290 VM_OBJECT_ASSERT_LOCKED(object); 2291 if (*rbehind != 0) { 2292 m = ma_src[0]; 2293 pindex = m->pindex; 2294 mpred = vm_radix_iter_lookup_lt(&pages, pindex); 2295 *rbehind = MIN(*rbehind, 2296 pindex - (mpred != NULL ? mpred->pindex + 1 : 0)); 2297 /* Stepping backward from pindex, mpred doesn't change. */ 2298 for (int i = 0; i < *rbehind; i++) { 2299 m = vm_page_alloc_after(object, &pages, pindex - i - 1, 2300 VM_ALLOC_NORMAL, mpred); 2301 if (m == NULL) { 2302 /* Shift the array. */ 2303 for (int j = 0; j < i; j++) 2304 ma_dst[j] = ma_dst[j + *rbehind - i]; 2305 *rbehind = i; 2306 *rahead = 0; 2307 break; 2308 } 2309 ma_dst[*rbehind - i - 1] = m; 2310 } 2311 } 2312 for (int i = 0; i < count; i++) 2313 ma_dst[*rbehind + i] = ma_src[i]; 2314 if (*rahead != 0) { 2315 m = ma_src[count - 1]; 2316 pindex = m->pindex + 1; 2317 msucc = vm_radix_iter_lookup_ge(&pages, pindex); 2318 *rahead = MIN(*rahead, 2319 (msucc != NULL ? msucc->pindex : object->size) - pindex); 2320 mpred = m; 2321 for (int i = 0; i < *rahead; i++) { 2322 m = vm_page_alloc_after(object, &pages, pindex + i, 2323 VM_ALLOC_NORMAL, mpred); 2324 if (m == NULL) { 2325 *rahead = i; 2326 break; 2327 } 2328 ma_dst[*rbehind + count + i] = mpred = m; 2329 } 2330 } 2331 } 2332 2333 void 2334 vm_object_set_writeable_dirty_(vm_object_t object) 2335 { 2336 atomic_add_int(&object->generation, 1); 2337 } 2338 2339 bool 2340 vm_object_mightbedirty_(vm_object_t object) 2341 { 2342 return (object->generation != object->cleangeneration); 2343 } 2344 2345 /* 2346 * vm_object_unwire: 2347 * 2348 * For each page offset within the specified range of the given object, 2349 * find the highest-level page in the shadow chain and unwire it. A page 2350 * must exist at every page offset, and the highest-level page must be 2351 * wired. 2352 */ 2353 void 2354 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2355 uint8_t queue) 2356 { 2357 vm_object_t tobject, t1object; 2358 vm_page_t m, tm; 2359 vm_pindex_t end_pindex, pindex, tpindex; 2360 int depth, locked_depth; 2361 2362 KASSERT((offset & PAGE_MASK) == 0, 2363 ("vm_object_unwire: offset is not page aligned")); 2364 KASSERT((length & PAGE_MASK) == 0, 2365 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2366 /* The wired count of a fictitious page never changes. */ 2367 if ((object->flags & OBJ_FICTITIOUS) != 0) 2368 return; 2369 pindex = OFF_TO_IDX(offset); 2370 end_pindex = pindex + atop(length); 2371 again: 2372 locked_depth = 1; 2373 VM_OBJECT_RLOCK(object); 2374 m = vm_page_find_least(object, pindex); 2375 while (pindex < end_pindex) { 2376 if (m == NULL || pindex < m->pindex) { 2377 /* 2378 * The first object in the shadow chain doesn't 2379 * contain a page at the current index. Therefore, 2380 * the page must exist in a backing object. 2381 */ 2382 tobject = object; 2383 tpindex = pindex; 2384 depth = 0; 2385 do { 2386 tpindex += 2387 OFF_TO_IDX(tobject->backing_object_offset); 2388 tobject = tobject->backing_object; 2389 KASSERT(tobject != NULL, 2390 ("vm_object_unwire: missing page")); 2391 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2392 goto next_page; 2393 depth++; 2394 if (depth == locked_depth) { 2395 locked_depth++; 2396 VM_OBJECT_RLOCK(tobject); 2397 } 2398 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2399 NULL); 2400 } else { 2401 tm = m; 2402 m = TAILQ_NEXT(m, listq); 2403 } 2404 if (vm_page_trysbusy(tm) == 0) { 2405 for (tobject = object; locked_depth >= 1; 2406 locked_depth--) { 2407 t1object = tobject->backing_object; 2408 if (tm->object != tobject) 2409 VM_OBJECT_RUNLOCK(tobject); 2410 tobject = t1object; 2411 } 2412 tobject = tm->object; 2413 if (!vm_page_busy_sleep(tm, "unwbo", 2414 VM_ALLOC_IGN_SBUSY)) 2415 VM_OBJECT_RUNLOCK(tobject); 2416 goto again; 2417 } 2418 vm_page_unwire(tm, queue); 2419 vm_page_sunbusy(tm); 2420 next_page: 2421 pindex++; 2422 } 2423 /* Release the accumulated object locks. */ 2424 for (tobject = object; locked_depth >= 1; locked_depth--) { 2425 t1object = tobject->backing_object; 2426 VM_OBJECT_RUNLOCK(tobject); 2427 tobject = t1object; 2428 } 2429 } 2430 2431 /* 2432 * Return the vnode for the given object, or NULL if none exists. 2433 * For tmpfs objects, the function may return NULL if there is 2434 * no vnode allocated at the time of the call. 2435 */ 2436 struct vnode * 2437 vm_object_vnode(vm_object_t object) 2438 { 2439 struct vnode *vp; 2440 2441 VM_OBJECT_ASSERT_LOCKED(object); 2442 vm_pager_getvp(object, &vp, NULL); 2443 return (vp); 2444 } 2445 2446 /* 2447 * Busy the vm object. This prevents new pages belonging to the object from 2448 * becoming busy. Existing pages persist as busy. Callers are responsible 2449 * for checking page state before proceeding. 2450 */ 2451 void 2452 vm_object_busy(vm_object_t obj) 2453 { 2454 2455 VM_OBJECT_ASSERT_LOCKED(obj); 2456 2457 blockcount_acquire(&obj->busy, 1); 2458 /* The fence is required to order loads of page busy. */ 2459 atomic_thread_fence_acq_rel(); 2460 } 2461 2462 void 2463 vm_object_unbusy(vm_object_t obj) 2464 { 2465 2466 blockcount_release(&obj->busy, 1); 2467 } 2468 2469 void 2470 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2471 { 2472 2473 VM_OBJECT_ASSERT_UNLOCKED(obj); 2474 2475 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2476 } 2477 2478 /* 2479 * This function aims to determine if the object is mapped, 2480 * specifically, if it is referenced by a vm_map_entry. Because 2481 * objects occasionally acquire transient references that do not 2482 * represent a mapping, the method used here is inexact. However, it 2483 * has very low overhead and is good enough for the advisory 2484 * vm.vmtotal sysctl. 2485 */ 2486 bool 2487 vm_object_is_active(vm_object_t obj) 2488 { 2489 2490 return (obj->ref_count > atomic_load_int(&obj->shadow_count)); 2491 } 2492 2493 static int 2494 vm_object_list_handler(struct sysctl_req *req, bool swap_only) 2495 { 2496 struct kinfo_vmobject *kvo; 2497 char *fullpath, *freepath; 2498 struct vnode *vp; 2499 struct vattr va; 2500 vm_object_t obj; 2501 vm_page_t m; 2502 u_long sp; 2503 int count, error; 2504 key_t key; 2505 unsigned short seq; 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 kvo->kvo_flags = 0; 2554 if (!swap_only) { 2555 TAILQ_FOREACH(m, &obj->memq, listq) { 2556 /* 2557 * A page may belong to the object but be 2558 * dequeued and set to PQ_NONE while the 2559 * object lock is not held. This makes the 2560 * reads of m->queue below racy, and we do not 2561 * count pages set to PQ_NONE. However, this 2562 * sysctl is only meant to give an 2563 * approximation of the system anyway. 2564 */ 2565 if (vm_page_active(m)) 2566 kvo->kvo_active++; 2567 else if (vm_page_inactive(m)) 2568 kvo->kvo_inactive++; 2569 else if (vm_page_in_laundry(m)) 2570 kvo->kvo_laundry++; 2571 } 2572 } 2573 2574 kvo->kvo_vn_fileid = 0; 2575 kvo->kvo_vn_fsid = 0; 2576 kvo->kvo_vn_fsid_freebsd11 = 0; 2577 freepath = NULL; 2578 fullpath = ""; 2579 vp = NULL; 2580 kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp : 2581 NULL); 2582 if (vp != NULL) { 2583 vref(vp); 2584 } else if ((obj->flags & OBJ_ANON) != 0) { 2585 MPASS(kvo->kvo_type == KVME_TYPE_SWAP); 2586 kvo->kvo_me = (uintptr_t)obj; 2587 /* tmpfs objs are reported as vnodes */ 2588 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object; 2589 sp = swap_pager_swapped_pages(obj); 2590 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp; 2591 } 2592 if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) { 2593 cdev_pager_get_path(obj, kvo->kvo_path, 2594 sizeof(kvo->kvo_path)); 2595 } 2596 VM_OBJECT_RUNLOCK(obj); 2597 if ((obj->flags & OBJ_SYSVSHM) != 0) { 2598 kvo->kvo_flags |= KVMO_FLAG_SYSVSHM; 2599 shmobjinfo(obj, &key, &seq); 2600 kvo->kvo_vn_fileid = key; 2601 kvo->kvo_vn_fsid_freebsd11 = seq; 2602 } 2603 if ((obj->flags & OBJ_POSIXSHM) != 0) { 2604 kvo->kvo_flags |= KVMO_FLAG_POSIXSHM; 2605 shm_get_path(obj, kvo->kvo_path, 2606 sizeof(kvo->kvo_path)); 2607 } 2608 if (vp != NULL) { 2609 vn_fullpath(vp, &fullpath, &freepath); 2610 vn_lock(vp, LK_SHARED | LK_RETRY); 2611 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2612 kvo->kvo_vn_fileid = va.va_fileid; 2613 kvo->kvo_vn_fsid = va.va_fsid; 2614 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2615 /* truncate */ 2616 } 2617 vput(vp); 2618 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2619 free(freepath, M_TEMP); 2620 } 2621 2622 /* Pack record size down */ 2623 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2624 + strlen(kvo->kvo_path) + 1; 2625 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2626 sizeof(uint64_t)); 2627 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2628 maybe_yield(); 2629 mtx_lock(&vm_object_list_mtx); 2630 if (error) 2631 break; 2632 } 2633 mtx_unlock(&vm_object_list_mtx); 2634 free(kvo, M_TEMP); 2635 return (error); 2636 } 2637 2638 static int 2639 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2640 { 2641 return (vm_object_list_handler(req, false)); 2642 } 2643 2644 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2645 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2646 "List of VM objects"); 2647 2648 static int 2649 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS) 2650 { 2651 return (vm_object_list_handler(req, true)); 2652 } 2653 2654 /* 2655 * This sysctl returns list of the anonymous or swap objects. Intent 2656 * is to provide stripped optimized list useful to analyze swap use. 2657 * Since technically non-swap (default) objects participate in the 2658 * shadow chains, and are converted to swap type as needed by swap 2659 * pager, we must report them. 2660 */ 2661 SYSCTL_PROC(_vm, OID_AUTO, swap_objects, 2662 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, 2663 sysctl_vm_object_list_swap, "S,kinfo_vmobject", 2664 "List of swap VM objects"); 2665 2666 #include "opt_ddb.h" 2667 #ifdef DDB 2668 #include <sys/kernel.h> 2669 2670 #include <sys/cons.h> 2671 2672 #include <ddb/ddb.h> 2673 2674 static int 2675 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2676 { 2677 vm_map_t tmpm; 2678 vm_map_entry_t tmpe; 2679 vm_object_t obj; 2680 2681 if (map == 0) 2682 return 0; 2683 2684 if (entry == 0) { 2685 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2686 if (_vm_object_in_map(map, object, tmpe)) { 2687 return 1; 2688 } 2689 } 2690 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2691 tmpm = entry->object.sub_map; 2692 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2693 if (_vm_object_in_map(tmpm, object, tmpe)) { 2694 return 1; 2695 } 2696 } 2697 } else if ((obj = entry->object.vm_object) != NULL) { 2698 for (; obj; obj = obj->backing_object) 2699 if (obj == object) { 2700 return 1; 2701 } 2702 } 2703 return 0; 2704 } 2705 2706 static int 2707 vm_object_in_map(vm_object_t object) 2708 { 2709 struct proc *p; 2710 2711 /* sx_slock(&allproc_lock); */ 2712 FOREACH_PROC_IN_SYSTEM(p) { 2713 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2714 continue; 2715 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2716 /* sx_sunlock(&allproc_lock); */ 2717 return 1; 2718 } 2719 } 2720 /* sx_sunlock(&allproc_lock); */ 2721 if (_vm_object_in_map(kernel_map, object, 0)) 2722 return 1; 2723 return 0; 2724 } 2725 2726 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE) 2727 { 2728 vm_object_t object; 2729 2730 /* 2731 * make sure that internal objs are in a map somewhere 2732 * and none have zero ref counts. 2733 */ 2734 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2735 if ((object->flags & OBJ_ANON) != 0) { 2736 if (object->ref_count == 0) { 2737 db_printf( 2738 "vmochk: internal obj has zero ref count: %lu\n", 2739 (u_long)object->size); 2740 } 2741 if (!vm_object_in_map(object)) { 2742 db_printf( 2743 "vmochk: internal obj is not in a map: " 2744 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2745 object->ref_count, (u_long)object->size, 2746 (u_long)object->size, 2747 (void *)object->backing_object); 2748 } 2749 } 2750 if (db_pager_quit) 2751 return; 2752 } 2753 } 2754 2755 /* 2756 * vm_object_print: [ debug ] 2757 */ 2758 DB_SHOW_COMMAND(object, vm_object_print_static) 2759 { 2760 /* XXX convert args. */ 2761 vm_object_t object = (vm_object_t)addr; 2762 boolean_t full = have_addr; 2763 2764 vm_page_t p; 2765 2766 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2767 #define count was_count 2768 2769 int count; 2770 2771 if (object == NULL) 2772 return; 2773 2774 db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x", 2775 object, (int)object->type, (uintmax_t)object->size, 2776 object->resident_page_count, object->ref_count, object->flags); 2777 db_iprintf(" ruid %d charge %jx\n", 2778 object->cred ? object->cred->cr_ruid : -1, 2779 (uintmax_t)object->charge); 2780 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2781 atomic_load_int(&object->shadow_count), 2782 object->backing_object ? object->backing_object->ref_count : 0, 2783 object->backing_object, (uintmax_t)object->backing_object_offset); 2784 2785 if (!full) 2786 return; 2787 2788 db_indent += 2; 2789 count = 0; 2790 TAILQ_FOREACH(p, &object->memq, listq) { 2791 if (count == 0) 2792 db_iprintf("memory:="); 2793 else if (count == 6) { 2794 db_printf("\n"); 2795 db_iprintf(" ..."); 2796 count = 0; 2797 } else 2798 db_printf(","); 2799 count++; 2800 2801 db_printf("(off=0x%jx,page=0x%jx)", 2802 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2803 2804 if (db_pager_quit) 2805 break; 2806 } 2807 if (count != 0) 2808 db_printf("\n"); 2809 db_indent -= 2; 2810 } 2811 2812 /* XXX. */ 2813 #undef count 2814 2815 /* XXX need this non-static entry for calling from vm_map_print. */ 2816 void 2817 vm_object_print( 2818 /* db_expr_t */ long addr, 2819 boolean_t have_addr, 2820 /* db_expr_t */ long count, 2821 char *modif) 2822 { 2823 vm_object_print_static(addr, have_addr, count, modif); 2824 } 2825 2826 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE) 2827 { 2828 vm_object_t object; 2829 vm_pindex_t fidx; 2830 vm_paddr_t pa; 2831 vm_page_t m, prev_m; 2832 int rcount; 2833 2834 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2835 db_printf("new object: %p\n", (void *)object); 2836 if (db_pager_quit) 2837 return; 2838 2839 rcount = 0; 2840 fidx = 0; 2841 pa = -1; 2842 TAILQ_FOREACH(m, &object->memq, listq) { 2843 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2844 prev_m->pindex + 1 != m->pindex) { 2845 if (rcount) { 2846 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2847 (long)fidx, rcount, (long)pa); 2848 if (db_pager_quit) 2849 return; 2850 rcount = 0; 2851 } 2852 } 2853 if (rcount && 2854 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2855 ++rcount; 2856 continue; 2857 } 2858 if (rcount) { 2859 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2860 (long)fidx, rcount, (long)pa); 2861 if (db_pager_quit) 2862 return; 2863 } 2864 fidx = m->pindex; 2865 pa = VM_PAGE_TO_PHYS(m); 2866 rcount = 1; 2867 } 2868 if (rcount) { 2869 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2870 (long)fidx, rcount, (long)pa); 2871 if (db_pager_quit) 2872 return; 2873 } 2874 } 2875 } 2876 #endif /* DDB */ 2877