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 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(vm_radix_is_empty(&object->rtree), 182 ("object %p has resident pages in its trie", object)); 183 #if VM_NRESERVLEVEL > 0 184 KASSERT(LIST_EMPTY(&object->rvq), 185 ("object %p has reservations", 186 object)); 187 #endif 188 KASSERT(!vm_object_busied(object), 189 ("object %p busy = %d", object, blockcount_read(&object->busy))); 190 KASSERT(object->resident_page_count == 0, 191 ("object %p resident_page_count = %d", 192 object, object->resident_page_count)); 193 KASSERT(atomic_load_int(&object->shadow_count) == 0, 194 ("object %p shadow_count = %d", 195 object, atomic_load_int(&object->shadow_count))); 196 KASSERT(object->type == OBJT_DEAD, 197 ("object %p has non-dead type %d", 198 object, object->type)); 199 KASSERT(object->charge == 0 && object->cred == NULL, 200 ("object %p has non-zero charge %ju (%p)", 201 object, (uintmax_t)object->charge, object->cred)); 202 } 203 #endif 204 205 static int 206 vm_object_zinit(void *mem, int size, int flags) 207 { 208 vm_object_t object; 209 210 object = (vm_object_t)mem; 211 rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW); 212 213 /* These are true for any object that has been freed */ 214 object->type = OBJT_DEAD; 215 vm_radix_init(&object->rtree); 216 refcount_init(&object->ref_count, 0); 217 blockcount_init(&object->paging_in_progress); 218 blockcount_init(&object->busy); 219 object->resident_page_count = 0; 220 atomic_store_int(&object->shadow_count, 0); 221 object->flags = OBJ_DEAD; 222 223 mtx_lock(&vm_object_list_mtx); 224 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 225 mtx_unlock(&vm_object_list_mtx); 226 return (0); 227 } 228 229 static void 230 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, 231 vm_object_t object, void *handle) 232 { 233 LIST_INIT(&object->shadow_head); 234 235 object->type = type; 236 object->flags = flags; 237 if ((flags & OBJ_SWAP) != 0) { 238 pctrie_init(&object->un_pager.swp.swp_blks); 239 object->un_pager.swp.writemappings = 0; 240 } 241 242 /* 243 * Ensure that swap_pager_swapoff() iteration over object_list 244 * sees up to date type and pctrie head if it observed 245 * non-dead object. 246 */ 247 atomic_thread_fence_rel(); 248 249 object->pg_color = 0; 250 object->size = size; 251 object->domain.dr_policy = NULL; 252 object->generation = 1; 253 object->cleangeneration = 1; 254 refcount_init(&object->ref_count, 1); 255 object->memattr = VM_MEMATTR_DEFAULT; 256 object->cred = NULL; 257 object->charge = 0; 258 object->handle = handle; 259 object->backing_object = NULL; 260 object->backing_object_offset = (vm_ooffset_t) 0; 261 #if VM_NRESERVLEVEL > 0 262 LIST_INIT(&object->rvq); 263 #endif 264 umtx_shm_object_init(object); 265 } 266 267 /* 268 * vm_object_init: 269 * 270 * Initialize the VM objects module. 271 */ 272 void 273 vm_object_init(void) 274 { 275 TAILQ_INIT(&vm_object_list); 276 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 277 278 rw_init(&kernel_object->lock, "kernel vm object"); 279 vm_radix_init(&kernel_object->rtree); 280 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - 281 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); 282 #if VM_NRESERVLEVEL > 0 283 kernel_object->flags |= OBJ_COLORED; 284 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 285 #endif 286 kernel_object->un_pager.phys.ops = &default_phys_pg_ops; 287 288 /* 289 * The lock portion of struct vm_object must be type stable due 290 * to vm_pageout_fallback_object_lock locking a vm object 291 * without holding any references to it. 292 * 293 * paging_in_progress is valid always. Lockless references to 294 * the objects may acquire pip and then check OBJ_DEAD. 295 */ 296 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 297 #ifdef INVARIANTS 298 vm_object_zdtor, 299 #else 300 NULL, 301 #endif 302 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 303 304 vm_radix_zinit(); 305 } 306 307 void 308 vm_object_clear_flag(vm_object_t object, u_short bits) 309 { 310 311 VM_OBJECT_ASSERT_WLOCKED(object); 312 object->flags &= ~bits; 313 } 314 315 /* 316 * Sets the default memory attribute for the specified object. Pages 317 * that are allocated to this object are by default assigned this memory 318 * attribute. 319 * 320 * Presently, this function must be called before any pages are allocated 321 * to the object. In the future, this requirement may be relaxed for 322 * "default" and "swap" objects. 323 */ 324 int 325 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 326 { 327 328 VM_OBJECT_ASSERT_WLOCKED(object); 329 330 if (object->type == OBJT_DEAD) 331 return (KERN_INVALID_ARGUMENT); 332 if (!vm_radix_is_empty(&object->rtree)) 333 return (KERN_FAILURE); 334 335 object->memattr = memattr; 336 return (KERN_SUCCESS); 337 } 338 339 void 340 vm_object_pip_add(vm_object_t object, short i) 341 { 342 343 if (i > 0) 344 blockcount_acquire(&object->paging_in_progress, i); 345 } 346 347 void 348 vm_object_pip_wakeup(vm_object_t object) 349 { 350 351 vm_object_pip_wakeupn(object, 1); 352 } 353 354 void 355 vm_object_pip_wakeupn(vm_object_t object, short i) 356 { 357 358 if (i > 0) 359 blockcount_release(&object->paging_in_progress, i); 360 } 361 362 /* 363 * Atomically drop the object lock and wait for pip to drain. This protects 364 * from sleep/wakeup races due to identity changes. The lock is not re-acquired 365 * on return. 366 */ 367 static void 368 vm_object_pip_sleep(vm_object_t object, const char *waitid) 369 { 370 371 (void)blockcount_sleep(&object->paging_in_progress, &object->lock, 372 waitid, PVM | PDROP); 373 } 374 375 void 376 vm_object_pip_wait(vm_object_t object, const char *waitid) 377 { 378 379 VM_OBJECT_ASSERT_WLOCKED(object); 380 381 blockcount_wait(&object->paging_in_progress, &object->lock, waitid, 382 PVM); 383 } 384 385 void 386 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid) 387 { 388 389 VM_OBJECT_ASSERT_UNLOCKED(object); 390 391 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM); 392 } 393 394 /* 395 * vm_object_allocate: 396 * 397 * Returns a new object with the given size. 398 */ 399 vm_object_t 400 vm_object_allocate(objtype_t type, vm_pindex_t size) 401 { 402 vm_object_t object; 403 u_short flags; 404 405 switch (type) { 406 case OBJT_DEAD: 407 panic("vm_object_allocate: can't create OBJT_DEAD"); 408 case OBJT_SWAP: 409 flags = OBJ_COLORED | OBJ_SWAP; 410 break; 411 case OBJT_DEVICE: 412 case OBJT_SG: 413 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; 414 break; 415 case OBJT_MGTDEVICE: 416 flags = OBJ_FICTITIOUS; 417 break; 418 case OBJT_PHYS: 419 flags = OBJ_UNMANAGED; 420 break; 421 case OBJT_VNODE: 422 flags = 0; 423 break; 424 default: 425 panic("vm_object_allocate: type %d is undefined or dynamic", 426 type); 427 } 428 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 429 _vm_object_allocate(type, size, flags, object, NULL); 430 431 return (object); 432 } 433 434 vm_object_t 435 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags) 436 { 437 vm_object_t object; 438 439 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */); 440 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 441 _vm_object_allocate(dyntype, size, flags, object, NULL); 442 443 return (object); 444 } 445 446 /* 447 * vm_object_allocate_anon: 448 * 449 * Returns a new default object of the given size and marked as 450 * anonymous memory for special split/collapse handling. Color 451 * to be initialized by the caller. 452 */ 453 vm_object_t 454 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, 455 struct ucred *cred, vm_size_t charge) 456 { 457 vm_object_t handle, object; 458 459 if (backing_object == NULL) 460 handle = NULL; 461 else if ((backing_object->flags & OBJ_ANON) != 0) 462 handle = backing_object->handle; 463 else 464 handle = backing_object; 465 object = uma_zalloc(obj_zone, M_WAITOK); 466 _vm_object_allocate(OBJT_SWAP, size, 467 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle); 468 object->cred = cred; 469 object->charge = cred != NULL ? charge : 0; 470 return (object); 471 } 472 473 static void 474 vm_object_reference_vnode(vm_object_t object) 475 { 476 u_int old; 477 478 /* 479 * vnode objects need the lock for the first reference 480 * to serialize with vnode_object_deallocate(). 481 */ 482 if (!refcount_acquire_if_gt(&object->ref_count, 0)) { 483 VM_OBJECT_RLOCK(object); 484 old = refcount_acquire(&object->ref_count); 485 if (object->type == OBJT_VNODE && old == 0) 486 vref(object->handle); 487 VM_OBJECT_RUNLOCK(object); 488 } 489 } 490 491 /* 492 * vm_object_reference: 493 * 494 * Acquires a reference to the given object. 495 */ 496 void 497 vm_object_reference(vm_object_t object) 498 { 499 500 if (object == NULL) 501 return; 502 503 if (object->type == OBJT_VNODE) 504 vm_object_reference_vnode(object); 505 else 506 refcount_acquire(&object->ref_count); 507 KASSERT((object->flags & OBJ_DEAD) == 0, 508 ("vm_object_reference: Referenced dead object.")); 509 } 510 511 /* 512 * vm_object_reference_locked: 513 * 514 * Gets another reference to the given object. 515 * 516 * The object must be locked. 517 */ 518 void 519 vm_object_reference_locked(vm_object_t object) 520 { 521 u_int old; 522 523 VM_OBJECT_ASSERT_LOCKED(object); 524 old = refcount_acquire(&object->ref_count); 525 if (object->type == OBJT_VNODE && old == 0) 526 vref(object->handle); 527 KASSERT((object->flags & OBJ_DEAD) == 0, 528 ("vm_object_reference: Referenced dead object.")); 529 } 530 531 /* 532 * Handle deallocating an object of type OBJT_VNODE. 533 */ 534 static void 535 vm_object_deallocate_vnode(vm_object_t object) 536 { 537 struct vnode *vp = (struct vnode *) object->handle; 538 bool last; 539 540 KASSERT(object->type == OBJT_VNODE, 541 ("vm_object_deallocate_vnode: not a vnode object")); 542 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp")); 543 544 /* Object lock to protect handle lookup. */ 545 last = refcount_release(&object->ref_count); 546 VM_OBJECT_RUNLOCK(object); 547 548 if (!last) 549 return; 550 551 if (!umtx_shm_vnobj_persistent) 552 umtx_shm_object_terminated(object); 553 554 /* vrele may need the vnode lock. */ 555 vrele(vp); 556 } 557 558 /* 559 * We dropped a reference on an object and discovered that it had a 560 * single remaining shadow. This is a sibling of the reference we 561 * dropped. Attempt to collapse the sibling and backing object. 562 */ 563 static vm_object_t 564 vm_object_deallocate_anon(vm_object_t backing_object) 565 { 566 vm_object_t object; 567 568 /* Fetch the final shadow. */ 569 object = LIST_FIRST(&backing_object->shadow_head); 570 KASSERT(object != NULL && 571 atomic_load_int(&backing_object->shadow_count) == 1, 572 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d", 573 backing_object->ref_count, 574 atomic_load_int(&backing_object->shadow_count))); 575 KASSERT((object->flags & OBJ_ANON) != 0, 576 ("invalid shadow object %p", object)); 577 578 if (!VM_OBJECT_TRYWLOCK(object)) { 579 /* 580 * Prevent object from disappearing since we do not have a 581 * reference. 582 */ 583 vm_object_pip_add(object, 1); 584 VM_OBJECT_WUNLOCK(backing_object); 585 VM_OBJECT_WLOCK(object); 586 vm_object_pip_wakeup(object); 587 } else 588 VM_OBJECT_WUNLOCK(backing_object); 589 590 /* 591 * Check for a collapse/terminate race with the last reference holder. 592 */ 593 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 || 594 !refcount_acquire_if_not_zero(&object->ref_count)) { 595 VM_OBJECT_WUNLOCK(object); 596 return (NULL); 597 } 598 backing_object = object->backing_object; 599 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0) 600 vm_object_collapse(object); 601 VM_OBJECT_WUNLOCK(object); 602 603 return (object); 604 } 605 606 /* 607 * vm_object_deallocate: 608 * 609 * Release a reference to the specified object, 610 * gained either through a vm_object_allocate 611 * or a vm_object_reference call. When all references 612 * are gone, storage associated with this object 613 * may be relinquished. 614 * 615 * No object may be locked. 616 */ 617 void 618 vm_object_deallocate(vm_object_t object) 619 { 620 vm_object_t temp; 621 bool released; 622 623 while (object != NULL) { 624 /* 625 * If the reference count goes to 0 we start calling 626 * vm_object_terminate() on the object chain. A ref count 627 * of 1 may be a special case depending on the shadow count 628 * being 0 or 1. These cases require a write lock on the 629 * object. 630 */ 631 if ((object->flags & OBJ_ANON) == 0) 632 released = refcount_release_if_gt(&object->ref_count, 1); 633 else 634 released = refcount_release_if_gt(&object->ref_count, 2); 635 if (released) 636 return; 637 638 if (object->type == OBJT_VNODE) { 639 VM_OBJECT_RLOCK(object); 640 if (object->type == OBJT_VNODE) { 641 vm_object_deallocate_vnode(object); 642 return; 643 } 644 VM_OBJECT_RUNLOCK(object); 645 } 646 647 VM_OBJECT_WLOCK(object); 648 KASSERT(object->ref_count > 0, 649 ("vm_object_deallocate: object deallocated too many times: %d", 650 object->type)); 651 652 /* 653 * If this is not the final reference to an anonymous 654 * object we may need to collapse the shadow chain. 655 */ 656 if (!refcount_release(&object->ref_count)) { 657 if (object->ref_count > 1 || 658 atomic_load_int(&object->shadow_count) == 0) { 659 if ((object->flags & OBJ_ANON) != 0 && 660 object->ref_count == 1) 661 vm_object_set_flag(object, 662 OBJ_ONEMAPPING); 663 VM_OBJECT_WUNLOCK(object); 664 return; 665 } 666 667 /* Handle collapsing last ref on anonymous objects. */ 668 object = vm_object_deallocate_anon(object); 669 continue; 670 } 671 672 /* 673 * Handle the final reference to an object. We restart 674 * the loop with the backing object to avoid recursion. 675 */ 676 umtx_shm_object_terminated(object); 677 temp = object->backing_object; 678 if (temp != NULL) { 679 KASSERT(object->type == OBJT_SWAP, 680 ("shadowed tmpfs v_object 2 %p", object)); 681 vm_object_backing_remove(object); 682 } 683 684 KASSERT((object->flags & OBJ_DEAD) == 0, 685 ("vm_object_deallocate: Terminating dead object.")); 686 vm_object_set_flag(object, OBJ_DEAD); 687 vm_object_terminate(object); 688 object = temp; 689 } 690 } 691 692 void 693 vm_object_destroy(vm_object_t object) 694 { 695 uma_zfree(obj_zone, object); 696 } 697 698 static void 699 vm_object_sub_shadow(vm_object_t object) 700 { 701 KASSERT(object->shadow_count >= 1, 702 ("object %p sub_shadow count zero", object)); 703 atomic_subtract_int(&object->shadow_count, 1); 704 } 705 706 static void 707 vm_object_backing_remove_locked(vm_object_t object) 708 { 709 vm_object_t backing_object; 710 711 backing_object = object->backing_object; 712 VM_OBJECT_ASSERT_WLOCKED(object); 713 VM_OBJECT_ASSERT_WLOCKED(backing_object); 714 715 KASSERT((object->flags & OBJ_COLLAPSING) == 0, 716 ("vm_object_backing_remove: Removing collapsing object.")); 717 718 vm_object_sub_shadow(backing_object); 719 if ((object->flags & OBJ_SHADOWLIST) != 0) { 720 LIST_REMOVE(object, shadow_list); 721 vm_object_clear_flag(object, OBJ_SHADOWLIST); 722 } 723 object->backing_object = NULL; 724 } 725 726 static void 727 vm_object_backing_remove(vm_object_t object) 728 { 729 vm_object_t backing_object; 730 731 VM_OBJECT_ASSERT_WLOCKED(object); 732 733 backing_object = object->backing_object; 734 if ((object->flags & OBJ_SHADOWLIST) != 0) { 735 VM_OBJECT_WLOCK(backing_object); 736 vm_object_backing_remove_locked(object); 737 VM_OBJECT_WUNLOCK(backing_object); 738 } else { 739 object->backing_object = NULL; 740 vm_object_sub_shadow(backing_object); 741 } 742 } 743 744 static void 745 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) 746 { 747 748 VM_OBJECT_ASSERT_WLOCKED(object); 749 750 atomic_add_int(&backing_object->shadow_count, 1); 751 if ((backing_object->flags & OBJ_ANON) != 0) { 752 VM_OBJECT_ASSERT_WLOCKED(backing_object); 753 LIST_INSERT_HEAD(&backing_object->shadow_head, object, 754 shadow_list); 755 vm_object_set_flag(object, OBJ_SHADOWLIST); 756 } 757 object->backing_object = backing_object; 758 } 759 760 static void 761 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) 762 { 763 764 VM_OBJECT_ASSERT_WLOCKED(object); 765 766 if ((backing_object->flags & OBJ_ANON) != 0) { 767 VM_OBJECT_WLOCK(backing_object); 768 vm_object_backing_insert_locked(object, backing_object); 769 VM_OBJECT_WUNLOCK(backing_object); 770 } else { 771 object->backing_object = backing_object; 772 atomic_add_int(&backing_object->shadow_count, 1); 773 } 774 } 775 776 /* 777 * Insert an object into a backing_object's shadow list with an additional 778 * reference to the backing_object added. 779 */ 780 static void 781 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object) 782 { 783 784 VM_OBJECT_ASSERT_WLOCKED(object); 785 786 if ((backing_object->flags & OBJ_ANON) != 0) { 787 VM_OBJECT_WLOCK(backing_object); 788 KASSERT((backing_object->flags & OBJ_DEAD) == 0, 789 ("shadowing dead anonymous object")); 790 vm_object_reference_locked(backing_object); 791 vm_object_backing_insert_locked(object, backing_object); 792 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING); 793 VM_OBJECT_WUNLOCK(backing_object); 794 } else { 795 vm_object_reference(backing_object); 796 atomic_add_int(&backing_object->shadow_count, 1); 797 object->backing_object = backing_object; 798 } 799 } 800 801 /* 802 * Transfer a backing reference from backing_object to object. 803 */ 804 static void 805 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object) 806 { 807 vm_object_t new_backing_object; 808 809 /* 810 * Note that the reference to backing_object->backing_object 811 * moves from within backing_object to within object. 812 */ 813 vm_object_backing_remove_locked(object); 814 new_backing_object = backing_object->backing_object; 815 if (new_backing_object == NULL) 816 return; 817 if ((new_backing_object->flags & OBJ_ANON) != 0) { 818 VM_OBJECT_WLOCK(new_backing_object); 819 vm_object_backing_remove_locked(backing_object); 820 vm_object_backing_insert_locked(object, new_backing_object); 821 VM_OBJECT_WUNLOCK(new_backing_object); 822 } else { 823 /* 824 * shadow_count for new_backing_object is left 825 * unchanged, its reference provided by backing_object 826 * is replaced by object. 827 */ 828 object->backing_object = new_backing_object; 829 backing_object->backing_object = NULL; 830 } 831 } 832 833 /* 834 * Wait for a concurrent collapse to settle. 835 */ 836 static void 837 vm_object_collapse_wait(vm_object_t object) 838 { 839 840 VM_OBJECT_ASSERT_WLOCKED(object); 841 842 while ((object->flags & OBJ_COLLAPSING) != 0) { 843 vm_object_pip_wait(object, "vmcolwait"); 844 counter_u64_add(object_collapse_waits, 1); 845 } 846 } 847 848 /* 849 * Waits for a backing object to clear a pending collapse and returns 850 * it locked if it is an ANON object. 851 */ 852 static vm_object_t 853 vm_object_backing_collapse_wait(vm_object_t object) 854 { 855 vm_object_t backing_object; 856 857 VM_OBJECT_ASSERT_WLOCKED(object); 858 859 for (;;) { 860 backing_object = object->backing_object; 861 if (backing_object == NULL || 862 (backing_object->flags & OBJ_ANON) == 0) 863 return (NULL); 864 VM_OBJECT_WLOCK(backing_object); 865 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0) 866 break; 867 VM_OBJECT_WUNLOCK(object); 868 vm_object_pip_sleep(backing_object, "vmbckwait"); 869 counter_u64_add(object_collapse_waits, 1); 870 VM_OBJECT_WLOCK(object); 871 } 872 return (backing_object); 873 } 874 875 /* 876 * vm_object_terminate_single_page removes a pageable page from the object, 877 * and removes it from the paging queues and frees it, if it is not wired. 878 * It is invoked via callback from vm_object_terminate_pages. 879 */ 880 static void 881 vm_object_terminate_single_page(vm_page_t p, void *objectv) 882 { 883 vm_object_t object __diagused = objectv; 884 885 vm_page_assert_unbusied(p); 886 KASSERT(p->object == object && 887 (p->ref_count & VPRC_OBJREF) != 0, 888 ("%s: page %p is inconsistent", __func__, p)); 889 p->object = NULL; 890 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { 891 KASSERT((object->flags & OBJ_UNMANAGED) != 0 || 892 vm_page_astate_load(p).queue != PQ_NONE, 893 ("%s: page %p does not belong to a queue", __func__, p)); 894 VM_CNT_INC(v_pfree); 895 vm_page_free(p); 896 } 897 } 898 899 /* 900 * vm_object_terminate_pages removes any remaining pageable pages 901 * from the object and resets the object to an empty state. 902 */ 903 static void 904 vm_object_terminate_pages(vm_object_t object) 905 { 906 VM_OBJECT_ASSERT_WLOCKED(object); 907 908 /* 909 * If the object contained any pages, then reset it to an empty state. 910 * Rather than incrementally removing each page from the object, the 911 * page and object are reset to any empty state. 912 */ 913 if (object->resident_page_count == 0) 914 return; 915 916 vm_radix_reclaim_callback(&object->rtree, 917 vm_object_terminate_single_page, object); 918 object->resident_page_count = 0; 919 if (object->type == OBJT_VNODE) 920 vdrop(object->handle); 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 static int 1002 vm_object_page_clean_flush(struct pctrie_iter *pages, vm_page_t p, 1003 int pagerflags, int flags, boolean_t *allclean, bool *eio) 1004 { 1005 vm_page_t ma[vm_pageout_page_count]; 1006 int count, runlen; 1007 1008 vm_page_assert_xbusied(p); 1009 ma[0] = p; 1010 runlen = vm_radix_iter_lookup_range(pages, p->pindex + 1, 1011 &ma[1], vm_pageout_page_count - 1); 1012 for (count = 1; count <= runlen; count++) { 1013 p = ma[count]; 1014 if (vm_page_tryxbusy(p) == 0) 1015 break; 1016 if (!vm_object_page_remove_write(p, flags, allclean)) { 1017 vm_page_xunbusy(p); 1018 break; 1019 } 1020 } 1021 1022 return (vm_pageout_flush(ma, count, pagerflags, eio)); 1023 } 1024 1025 /* 1026 * vm_object_page_clean 1027 * 1028 * Clean all dirty pages in the specified range of object. Leaves page 1029 * on whatever queue it is currently on. If NOSYNC is set then do not 1030 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), 1031 * leaving the object dirty. 1032 * 1033 * For swap objects backing tmpfs regular files, do not flush anything, 1034 * but remove write protection on the mapped pages to update mtime through 1035 * mmaped writes. 1036 * 1037 * When stuffing pages asynchronously, allow clustering. XXX we need a 1038 * synchronous clustering mode implementation. 1039 * 1040 * Odd semantics: if start == end, we clean everything. 1041 * 1042 * The object must be locked. 1043 * 1044 * Returns FALSE if some page from the range was not written, as 1045 * reported by the pager, and TRUE otherwise. 1046 */ 1047 boolean_t 1048 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, 1049 int flags) 1050 { 1051 struct pctrie_iter pages; 1052 vm_page_t np, p; 1053 vm_pindex_t pi, tend, tstart; 1054 int curgeneration, n, pagerflags; 1055 boolean_t res, allclean; 1056 bool eio; 1057 1058 VM_OBJECT_ASSERT_WLOCKED(object); 1059 1060 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0) 1061 return (TRUE); 1062 1063 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? 1064 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 1065 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; 1066 1067 tstart = OFF_TO_IDX(start); 1068 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); 1069 allclean = tstart == 0 && tend >= object->size; 1070 res = TRUE; 1071 vm_page_iter_init(&pages, object); 1072 1073 rescan: 1074 curgeneration = object->generation; 1075 1076 for (p = vm_radix_iter_lookup_ge(&pages, tstart); p != NULL; p = np) { 1077 pi = p->pindex; 1078 if (pi >= tend) 1079 break; 1080 if (vm_page_none_valid(p)) { 1081 np = vm_radix_iter_step(&pages); 1082 continue; 1083 } 1084 if (!vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL)) { 1085 pctrie_iter_reset(&pages); 1086 if (object->generation != curgeneration && 1087 (flags & OBJPC_SYNC) != 0) 1088 goto rescan; 1089 np = vm_radix_iter_lookup_ge(&pages, pi); 1090 continue; 1091 } 1092 if (!vm_object_page_remove_write(p, flags, &allclean)) { 1093 np = vm_radix_iter_step(&pages); 1094 vm_page_xunbusy(p); 1095 continue; 1096 } 1097 if (object->type == OBJT_VNODE) { 1098 n = vm_object_page_clean_flush(&pages, p, pagerflags, 1099 flags, &allclean, &eio); 1100 pctrie_iter_reset(&pages); 1101 if (eio) { 1102 res = FALSE; 1103 allclean = FALSE; 1104 } 1105 if (object->generation != curgeneration && 1106 (flags & OBJPC_SYNC) != 0) 1107 goto rescan; 1108 1109 /* 1110 * If the VOP_PUTPAGES() did a truncated write, so 1111 * that even the first page of the run is not fully 1112 * written, vm_pageout_flush() returns 0 as the run 1113 * length. Since the condition that caused truncated 1114 * write may be permanent, e.g. exhausted free space, 1115 * accepting n == 0 would cause an infinite loop. 1116 * 1117 * Forwarding the iterator leaves the unwritten page 1118 * behind, but there is not much we can do there if 1119 * filesystem refuses to write it. 1120 */ 1121 if (n == 0) { 1122 n = 1; 1123 allclean = FALSE; 1124 } 1125 } else { 1126 n = 1; 1127 vm_page_xunbusy(p); 1128 } 1129 np = vm_radix_iter_lookup_ge(&pages, pi + n); 1130 } 1131 #if 0 1132 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); 1133 #endif 1134 1135 /* 1136 * Leave updating cleangeneration for tmpfs objects to tmpfs 1137 * scan. It needs to update mtime, which happens for other 1138 * filesystems during page writeouts. 1139 */ 1140 if (allclean && object->type == OBJT_VNODE) 1141 object->cleangeneration = curgeneration; 1142 return (res); 1143 } 1144 1145 /* 1146 * Note that there is absolutely no sense in writing out 1147 * anonymous objects, so we track down the vnode object 1148 * to write out. 1149 * We invalidate (remove) all pages from the address space 1150 * for semantic correctness. 1151 * 1152 * If the backing object is a device object with unmanaged pages, then any 1153 * mappings to the specified range of pages must be removed before this 1154 * function is called. 1155 * 1156 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1157 * may start out with a NULL object. 1158 */ 1159 boolean_t 1160 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1161 boolean_t syncio, boolean_t invalidate) 1162 { 1163 vm_object_t backing_object; 1164 struct vnode *vp; 1165 struct mount *mp; 1166 int error, flags, fsync_after; 1167 boolean_t res; 1168 1169 if (object == NULL) 1170 return (TRUE); 1171 res = TRUE; 1172 error = 0; 1173 VM_OBJECT_WLOCK(object); 1174 while ((backing_object = object->backing_object) != NULL) { 1175 VM_OBJECT_WLOCK(backing_object); 1176 offset += object->backing_object_offset; 1177 VM_OBJECT_WUNLOCK(object); 1178 object = backing_object; 1179 if (object->size < OFF_TO_IDX(offset + size)) 1180 size = IDX_TO_OFF(object->size) - offset; 1181 } 1182 /* 1183 * Flush pages if writing is allowed, invalidate them 1184 * if invalidation requested. Pages undergoing I/O 1185 * will be ignored by vm_object_page_remove(). 1186 * 1187 * We cannot lock the vnode and then wait for paging 1188 * to complete without deadlocking against vm_fault. 1189 * Instead we simply call vm_object_page_remove() and 1190 * allow it to block internally on a page-by-page 1191 * basis when it encounters pages undergoing async 1192 * I/O. 1193 */ 1194 if (object->type == OBJT_VNODE && 1195 vm_object_mightbedirty(object) != 0 && 1196 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { 1197 VM_OBJECT_WUNLOCK(object); 1198 (void)vn_start_write(vp, &mp, V_WAIT); 1199 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1200 if (syncio && !invalidate && offset == 0 && 1201 atop(size) == object->size) { 1202 /* 1203 * If syncing the whole mapping of the file, 1204 * it is faster to schedule all the writes in 1205 * async mode, also allowing the clustering, 1206 * and then wait for i/o to complete. 1207 */ 1208 flags = 0; 1209 fsync_after = TRUE; 1210 } else { 1211 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1212 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; 1213 fsync_after = FALSE; 1214 } 1215 VM_OBJECT_WLOCK(object); 1216 res = vm_object_page_clean(object, offset, offset + size, 1217 flags); 1218 VM_OBJECT_WUNLOCK(object); 1219 if (fsync_after) { 1220 for (;;) { 1221 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 1222 if (error != ERELOOKUP) 1223 break; 1224 1225 /* 1226 * Allow SU/bufdaemon to handle more 1227 * dependencies in the meantime. 1228 */ 1229 VOP_UNLOCK(vp); 1230 vn_finished_write(mp); 1231 1232 (void)vn_start_write(vp, &mp, V_WAIT); 1233 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1234 } 1235 } 1236 VOP_UNLOCK(vp); 1237 vn_finished_write(mp); 1238 if (error != 0) 1239 res = FALSE; 1240 VM_OBJECT_WLOCK(object); 1241 } 1242 if ((object->type == OBJT_VNODE || 1243 object->type == OBJT_DEVICE) && invalidate) { 1244 if (object->type == OBJT_DEVICE) 1245 /* 1246 * The option OBJPR_NOTMAPPED must be passed here 1247 * because vm_object_page_remove() cannot remove 1248 * unmanaged mappings. 1249 */ 1250 flags = OBJPR_NOTMAPPED; 1251 else if (old_msync) 1252 flags = 0; 1253 else 1254 flags = OBJPR_CLEANONLY; 1255 vm_object_page_remove(object, OFF_TO_IDX(offset), 1256 OFF_TO_IDX(offset + size + PAGE_MASK), flags); 1257 } 1258 VM_OBJECT_WUNLOCK(object); 1259 return (res); 1260 } 1261 1262 /* 1263 * Determine whether the given advice can be applied to the object. Advice is 1264 * not applied to unmanaged pages since they never belong to page queues, and 1265 * since MADV_FREE is destructive, it can apply only to anonymous pages that 1266 * have been mapped at most once. 1267 */ 1268 static bool 1269 vm_object_advice_applies(vm_object_t object, int advice) 1270 { 1271 1272 if ((object->flags & OBJ_UNMANAGED) != 0) 1273 return (false); 1274 if (advice != MADV_FREE) 1275 return (true); 1276 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == 1277 (OBJ_ONEMAPPING | OBJ_ANON)); 1278 } 1279 1280 static void 1281 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, 1282 vm_size_t size) 1283 { 1284 1285 if (advice == MADV_FREE) 1286 vm_pager_freespace(object, pindex, size); 1287 } 1288 1289 /* 1290 * vm_object_madvise: 1291 * 1292 * Implements the madvise function at the object/page level. 1293 * 1294 * MADV_WILLNEED (any object) 1295 * 1296 * Activate the specified pages if they are resident. 1297 * 1298 * MADV_DONTNEED (any object) 1299 * 1300 * Deactivate the specified pages if they are resident. 1301 * 1302 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only) 1303 * 1304 * Deactivate and clean the specified pages if they are 1305 * resident. This permits the process to reuse the pages 1306 * without faulting or the kernel to reclaim the pages 1307 * without I/O. 1308 */ 1309 void 1310 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, 1311 int advice) 1312 { 1313 struct pctrie_iter pages; 1314 vm_pindex_t tpindex; 1315 vm_object_t backing_object, tobject; 1316 vm_page_t m, tm; 1317 1318 if (object == NULL) 1319 return; 1320 1321 vm_page_iter_init(&pages, object); 1322 relookup: 1323 VM_OBJECT_WLOCK(object); 1324 if (!vm_object_advice_applies(object, advice)) { 1325 VM_OBJECT_WUNLOCK(object); 1326 return; 1327 } 1328 for (m = vm_radix_iter_lookup_ge(&pages, pindex); pindex < end; 1329 pindex++) { 1330 tobject = object; 1331 1332 /* 1333 * If the next page isn't resident in the top-level object, we 1334 * need to search the shadow chain. When applying MADV_FREE, we 1335 * take care to release any swap space used to store 1336 * non-resident pages. 1337 */ 1338 if (m == NULL || pindex < m->pindex) { 1339 /* 1340 * Optimize a common case: if the top-level object has 1341 * no backing object, we can skip over the non-resident 1342 * range in constant time. 1343 */ 1344 if (object->backing_object == NULL) { 1345 tpindex = (m != NULL && m->pindex < end) ? 1346 m->pindex : end; 1347 vm_object_madvise_freespace(object, advice, 1348 pindex, tpindex - pindex); 1349 if ((pindex = tpindex) == end) 1350 break; 1351 goto next_page; 1352 } 1353 1354 tpindex = pindex; 1355 do { 1356 vm_object_madvise_freespace(tobject, advice, 1357 tpindex, 1); 1358 /* 1359 * Prepare to search the next object in the 1360 * chain. 1361 */ 1362 backing_object = tobject->backing_object; 1363 if (backing_object == NULL) 1364 goto next_pindex; 1365 VM_OBJECT_WLOCK(backing_object); 1366 tpindex += 1367 OFF_TO_IDX(tobject->backing_object_offset); 1368 if (tobject != object) 1369 VM_OBJECT_WUNLOCK(tobject); 1370 tobject = backing_object; 1371 if (!vm_object_advice_applies(tobject, advice)) 1372 goto next_pindex; 1373 } while ((tm = vm_page_lookup(tobject, tpindex)) == 1374 NULL); 1375 } else { 1376 next_page: 1377 tm = m; 1378 m = vm_radix_iter_step(&pages); 1379 } 1380 1381 /* 1382 * If the page is not in a normal state, skip it. The page 1383 * can not be invalidated while the object lock is held. 1384 */ 1385 if (!vm_page_all_valid(tm) || vm_page_wired(tm)) 1386 goto next_pindex; 1387 KASSERT((tm->flags & PG_FICTITIOUS) == 0, 1388 ("vm_object_madvise: page %p is fictitious", tm)); 1389 KASSERT((tm->oflags & VPO_UNMANAGED) == 0, 1390 ("vm_object_madvise: page %p is not managed", tm)); 1391 if (vm_page_tryxbusy(tm) == 0) { 1392 if (object != tobject) 1393 VM_OBJECT_WUNLOCK(object); 1394 if (advice == MADV_WILLNEED) { 1395 /* 1396 * Reference the page before unlocking and 1397 * sleeping so that the page daemon is less 1398 * likely to reclaim it. 1399 */ 1400 vm_page_aflag_set(tm, PGA_REFERENCED); 1401 } 1402 if (!vm_page_busy_sleep(tm, "madvpo", 0)) 1403 VM_OBJECT_WUNLOCK(tobject); 1404 pctrie_iter_reset(&pages); 1405 goto relookup; 1406 } 1407 vm_page_advise(tm, advice); 1408 vm_page_xunbusy(tm); 1409 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); 1410 next_pindex: 1411 if (tobject != object) 1412 VM_OBJECT_WUNLOCK(tobject); 1413 } 1414 VM_OBJECT_WUNLOCK(object); 1415 } 1416 1417 /* 1418 * vm_object_shadow: 1419 * 1420 * Create a new object which is backed by the 1421 * specified existing object range. The source 1422 * object reference is deallocated. 1423 * 1424 * The new object and offset into that object 1425 * are returned in the source parameters. 1426 */ 1427 void 1428 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, 1429 struct ucred *cred, bool shared) 1430 { 1431 vm_object_t source; 1432 vm_object_t result; 1433 1434 source = *object; 1435 1436 /* 1437 * Don't create the new object if the old object isn't shared. 1438 * 1439 * If we hold the only reference we can guarantee that it won't 1440 * increase while we have the map locked. Otherwise the race is 1441 * harmless and we will end up with an extra shadow object that 1442 * will be collapsed later. 1443 */ 1444 if (source != NULL && source->ref_count == 1 && 1445 (source->flags & OBJ_ANON) != 0) 1446 return; 1447 1448 /* 1449 * Allocate a new object with the given length. 1450 */ 1451 result = vm_object_allocate_anon(atop(length), source, cred, length); 1452 1453 /* 1454 * Store the offset into the source object, and fix up the offset into 1455 * the new object. 1456 */ 1457 result->backing_object_offset = *offset; 1458 1459 if (shared || source != NULL) { 1460 VM_OBJECT_WLOCK(result); 1461 1462 /* 1463 * The new object shadows the source object, adding a 1464 * reference to it. Our caller changes his reference 1465 * to point to the new object, removing a reference to 1466 * the source object. Net result: no change of 1467 * reference count, unless the caller needs to add one 1468 * more reference due to forking a shared map entry. 1469 */ 1470 if (shared) { 1471 vm_object_reference_locked(result); 1472 vm_object_clear_flag(result, OBJ_ONEMAPPING); 1473 } 1474 1475 /* 1476 * Try to optimize the result object's page color when 1477 * shadowing in order to maintain page coloring 1478 * consistency in the combined shadowed object. 1479 */ 1480 if (source != NULL) { 1481 vm_object_backing_insert(result, source); 1482 result->domain = source->domain; 1483 #if VM_NRESERVLEVEL > 0 1484 vm_object_set_flag(result, 1485 (source->flags & OBJ_COLORED)); 1486 result->pg_color = (source->pg_color + 1487 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1488 1)) - 1); 1489 #endif 1490 } 1491 VM_OBJECT_WUNLOCK(result); 1492 } 1493 1494 /* 1495 * Return the new things 1496 */ 1497 *offset = 0; 1498 *object = result; 1499 } 1500 1501 /* 1502 * vm_object_split: 1503 * 1504 * Split the pages in a map entry into a new object. This affords 1505 * easier removal of unused pages, and keeps object inheritance from 1506 * being a negative impact on memory usage. 1507 */ 1508 void 1509 vm_object_split(vm_map_entry_t entry) 1510 { 1511 struct pctrie_iter pages; 1512 vm_page_t m; 1513 vm_object_t orig_object, new_object, backing_object; 1514 vm_pindex_t offidxstart; 1515 vm_size_t size; 1516 1517 orig_object = entry->object.vm_object; 1518 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0, 1519 ("vm_object_split: Splitting object with multiple mappings.")); 1520 if ((orig_object->flags & OBJ_ANON) == 0) 1521 return; 1522 if (orig_object->ref_count <= 1) 1523 return; 1524 VM_OBJECT_WUNLOCK(orig_object); 1525 1526 offidxstart = OFF_TO_IDX(entry->offset); 1527 size = atop(entry->end - entry->start); 1528 1529 new_object = vm_object_allocate_anon(size, orig_object, 1530 orig_object->cred, ptoa(size)); 1531 1532 /* 1533 * We must wait for the orig_object to complete any in-progress 1534 * collapse so that the swap blocks are stable below. The 1535 * additional reference on backing_object by new object will 1536 * prevent further collapse operations until split completes. 1537 */ 1538 VM_OBJECT_WLOCK(orig_object); 1539 vm_object_collapse_wait(orig_object); 1540 1541 /* 1542 * At this point, the new object is still private, so the order in 1543 * which the original and new objects are locked does not matter. 1544 */ 1545 VM_OBJECT_WLOCK(new_object); 1546 new_object->domain = orig_object->domain; 1547 backing_object = orig_object->backing_object; 1548 if (backing_object != NULL) { 1549 vm_object_backing_insert_ref(new_object, backing_object); 1550 new_object->backing_object_offset = 1551 orig_object->backing_object_offset + entry->offset; 1552 } 1553 if (orig_object->cred != NULL) { 1554 crhold(orig_object->cred); 1555 KASSERT(orig_object->charge >= ptoa(size), 1556 ("orig_object->charge < 0")); 1557 orig_object->charge -= ptoa(size); 1558 } 1559 1560 /* 1561 * Mark the split operation so that swap_pager_getpages() knows 1562 * that the object is in transition. 1563 */ 1564 vm_object_set_flag(orig_object, OBJ_SPLIT); 1565 vm_page_iter_limit_init(&pages, orig_object, offidxstart + size); 1566 retry: 1567 KASSERT(pctrie_iter_is_reset(&pages), 1568 ("%s: pctrie_iter not reset for retry", __func__)); 1569 for (m = vm_radix_iter_lookup_ge(&pages, offidxstart); m != NULL; 1570 m = vm_radix_iter_step(&pages)) { 1571 /* 1572 * We must wait for pending I/O to complete before we can 1573 * rename the page. 1574 * 1575 * We do not have to VM_PROT_NONE the page as mappings should 1576 * not be changed by this operation. 1577 */ 1578 if (vm_page_tryxbusy(m) == 0) { 1579 VM_OBJECT_WUNLOCK(new_object); 1580 if (vm_page_busy_sleep(m, "spltwt", 0)) 1581 VM_OBJECT_WLOCK(orig_object); 1582 pctrie_iter_reset(&pages); 1583 VM_OBJECT_WLOCK(new_object); 1584 goto retry; 1585 } 1586 1587 /* 1588 * If the page was left invalid, it was likely placed there by 1589 * an incomplete fault. Just remove and ignore. 1590 * 1591 * One other possibility is that the map entry is wired, in 1592 * which case we must hang on to the page to avoid leaking it, 1593 * as the map entry owns the wiring. This case can arise if the 1594 * backing object is truncated by the pager. 1595 */ 1596 if (vm_page_none_valid(m) && entry->wired_count == 0) { 1597 if (vm_page_iter_remove(&pages, m)) 1598 vm_page_free(m); 1599 continue; 1600 } 1601 1602 /* vm_page_iter_rename() will dirty the page if it is valid. */ 1603 if (!vm_page_iter_rename(&pages, m, new_object, m->pindex - 1604 offidxstart)) { 1605 vm_page_xunbusy(m); 1606 VM_OBJECT_WUNLOCK(new_object); 1607 VM_OBJECT_WUNLOCK(orig_object); 1608 vm_radix_wait(); 1609 pctrie_iter_reset(&pages); 1610 VM_OBJECT_WLOCK(orig_object); 1611 VM_OBJECT_WLOCK(new_object); 1612 goto retry; 1613 } 1614 1615 #if VM_NRESERVLEVEL > 0 1616 /* 1617 * If some of the reservation's allocated pages remain with 1618 * the original object, then transferring the reservation to 1619 * the new object is neither particularly beneficial nor 1620 * particularly harmful as compared to leaving the reservation 1621 * with the original object. If, however, all of the 1622 * reservation's allocated pages are transferred to the new 1623 * object, then transferring the reservation is typically 1624 * beneficial. Determining which of these two cases applies 1625 * would be more costly than unconditionally renaming the 1626 * reservation. 1627 */ 1628 vm_reserv_rename(m, new_object, orig_object, offidxstart); 1629 #endif 1630 } 1631 1632 /* 1633 * swap_pager_copy() can sleep, in which case the orig_object's 1634 * and new_object's locks are released and reacquired. 1635 */ 1636 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1637 vm_page_iter_init(&pages, new_object); 1638 VM_RADIX_FOREACH(m, &pages) 1639 vm_page_xunbusy(m); 1640 1641 vm_object_clear_flag(orig_object, OBJ_SPLIT); 1642 VM_OBJECT_WUNLOCK(orig_object); 1643 VM_OBJECT_WUNLOCK(new_object); 1644 entry->object.vm_object = new_object; 1645 entry->offset = 0LL; 1646 vm_object_deallocate(orig_object); 1647 VM_OBJECT_WLOCK(new_object); 1648 } 1649 1650 static vm_page_t 1651 vm_object_collapse_scan_wait(struct pctrie_iter *pages, vm_object_t object, 1652 vm_page_t p) 1653 { 1654 vm_object_t backing_object; 1655 1656 VM_OBJECT_ASSERT_WLOCKED(object); 1657 backing_object = object->backing_object; 1658 VM_OBJECT_ASSERT_WLOCKED(backing_object); 1659 1660 KASSERT(p == NULL || p->object == object || p->object == backing_object, 1661 ("invalid ownership %p %p %p", p, object, backing_object)); 1662 /* The page is only NULL when rename fails. */ 1663 if (p == NULL) { 1664 VM_OBJECT_WUNLOCK(object); 1665 VM_OBJECT_WUNLOCK(backing_object); 1666 vm_radix_wait(); 1667 VM_OBJECT_WLOCK(object); 1668 } else if (p->object == object) { 1669 VM_OBJECT_WUNLOCK(backing_object); 1670 if (vm_page_busy_sleep(p, "vmocol", 0)) 1671 VM_OBJECT_WLOCK(object); 1672 } else { 1673 VM_OBJECT_WUNLOCK(object); 1674 if (!vm_page_busy_sleep(p, "vmocol", 0)) 1675 VM_OBJECT_WUNLOCK(backing_object); 1676 VM_OBJECT_WLOCK(object); 1677 } 1678 VM_OBJECT_WLOCK(backing_object); 1679 vm_page_iter_init(pages, backing_object); 1680 return (vm_radix_iter_lookup_ge(pages, 0)); 1681 } 1682 1683 static void 1684 vm_object_collapse_scan(vm_object_t object) 1685 { 1686 struct pctrie_iter pages; 1687 vm_object_t backing_object; 1688 vm_page_t next, p, pp; 1689 vm_pindex_t backing_offset_index, new_pindex; 1690 1691 VM_OBJECT_ASSERT_WLOCKED(object); 1692 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 1693 1694 backing_object = object->backing_object; 1695 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1696 1697 /* 1698 * Our scan 1699 */ 1700 vm_page_iter_init(&pages, backing_object); 1701 for (p = vm_radix_iter_lookup_ge(&pages, 0); p != NULL; p = next) { 1702 /* 1703 * Check for busy page 1704 */ 1705 if (vm_page_tryxbusy(p) == 0) { 1706 next = vm_object_collapse_scan_wait(&pages, object, p); 1707 continue; 1708 } 1709 1710 KASSERT(object->backing_object == backing_object, 1711 ("vm_object_collapse_scan: backing object mismatch %p != %p", 1712 object->backing_object, backing_object)); 1713 KASSERT(p->object == backing_object, 1714 ("vm_object_collapse_scan: object mismatch %p != %p", 1715 p->object, backing_object)); 1716 1717 if (p->pindex < backing_offset_index || object->size <= 1718 (new_pindex = p->pindex - backing_offset_index)) { 1719 vm_pager_freespace(backing_object, p->pindex, 1); 1720 1721 KASSERT(!pmap_page_is_mapped(p), 1722 ("freeing mapped page %p", p)); 1723 if (vm_page_iter_remove(&pages, p)) 1724 vm_page_free(p); 1725 next = vm_radix_iter_step(&pages); 1726 continue; 1727 } 1728 1729 if (!vm_page_all_valid(p)) { 1730 KASSERT(!pmap_page_is_mapped(p), 1731 ("freeing mapped page %p", p)); 1732 if (vm_page_iter_remove(&pages, p)) 1733 vm_page_free(p); 1734 next = vm_radix_iter_step(&pages); 1735 continue; 1736 } 1737 1738 pp = vm_page_lookup(object, new_pindex); 1739 if (pp != NULL && vm_page_tryxbusy(pp) == 0) { 1740 vm_page_xunbusy(p); 1741 /* 1742 * The page in the parent is busy and possibly not 1743 * (yet) valid. Until its state is finalized by the 1744 * busy bit owner, we can't tell whether it shadows the 1745 * original page. 1746 */ 1747 next = vm_object_collapse_scan_wait(&pages, object, pp); 1748 continue; 1749 } 1750 1751 if (pp != NULL && vm_page_none_valid(pp)) { 1752 /* 1753 * The page was invalid in the parent. Likely placed 1754 * there by an incomplete fault. Just remove and 1755 * ignore. p can replace it. 1756 */ 1757 if (vm_page_remove(pp)) 1758 vm_page_free(pp); 1759 pp = NULL; 1760 } 1761 1762 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, 1763 NULL)) { 1764 /* 1765 * The page already exists in the parent OR swap exists 1766 * for this location in the parent. Leave the parent's 1767 * page alone. Destroy the original page from the 1768 * backing object. 1769 */ 1770 vm_pager_freespace(backing_object, p->pindex, 1); 1771 KASSERT(!pmap_page_is_mapped(p), 1772 ("freeing mapped page %p", p)); 1773 if (pp != NULL) 1774 vm_page_xunbusy(pp); 1775 if (vm_page_iter_remove(&pages, p)) 1776 vm_page_free(p); 1777 next = vm_radix_iter_step(&pages); 1778 continue; 1779 } 1780 1781 /* 1782 * Page does not exist in parent, rename the page from the 1783 * backing object to the main object. 1784 * 1785 * If the page was mapped to a process, it can remain mapped 1786 * through the rename. vm_page_iter_rename() will dirty the 1787 * page. 1788 */ 1789 if (!vm_page_iter_rename(&pages, p, object, new_pindex)) { 1790 vm_page_xunbusy(p); 1791 next = vm_object_collapse_scan_wait(&pages, object, 1792 NULL); 1793 continue; 1794 } 1795 1796 /* Use the old pindex to free the right page. */ 1797 vm_pager_freespace(backing_object, new_pindex + 1798 backing_offset_index, 1); 1799 1800 #if VM_NRESERVLEVEL > 0 1801 /* 1802 * Rename the reservation. 1803 */ 1804 vm_reserv_rename(p, object, backing_object, 1805 backing_offset_index); 1806 #endif 1807 vm_page_xunbusy(p); 1808 next = vm_radix_iter_step(&pages); 1809 } 1810 return; 1811 } 1812 1813 /* 1814 * vm_object_collapse: 1815 * 1816 * Collapse an object with the object backing it. 1817 * Pages in the backing object are moved into the 1818 * parent, and the backing object is deallocated. 1819 */ 1820 void 1821 vm_object_collapse(vm_object_t object) 1822 { 1823 vm_object_t backing_object, new_backing_object; 1824 1825 VM_OBJECT_ASSERT_WLOCKED(object); 1826 1827 while (TRUE) { 1828 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON, 1829 ("collapsing invalid object")); 1830 1831 /* 1832 * Wait for the backing_object to finish any pending 1833 * collapse so that the caller sees the shortest possible 1834 * shadow chain. 1835 */ 1836 backing_object = vm_object_backing_collapse_wait(object); 1837 if (backing_object == NULL) 1838 return; 1839 1840 KASSERT(object->ref_count > 0 && 1841 object->ref_count > atomic_load_int(&object->shadow_count), 1842 ("collapse with invalid ref %d or shadow %d count.", 1843 object->ref_count, atomic_load_int(&object->shadow_count))); 1844 KASSERT((backing_object->flags & 1845 (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1846 ("vm_object_collapse: Backing object already collapsing.")); 1847 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0, 1848 ("vm_object_collapse: object is already collapsing.")); 1849 1850 /* 1851 * We know that we can either collapse the backing object if 1852 * the parent is the only reference to it, or (perhaps) have 1853 * the parent bypass the object if the parent happens to shadow 1854 * all the resident pages in the entire backing object. 1855 */ 1856 if (backing_object->ref_count == 1) { 1857 KASSERT(atomic_load_int(&backing_object->shadow_count) 1858 == 1, 1859 ("vm_object_collapse: shadow_count: %d", 1860 atomic_load_int(&backing_object->shadow_count))); 1861 vm_object_pip_add(object, 1); 1862 vm_object_set_flag(object, OBJ_COLLAPSING); 1863 vm_object_pip_add(backing_object, 1); 1864 vm_object_set_flag(backing_object, OBJ_DEAD); 1865 1866 /* 1867 * If there is exactly one reference to the backing 1868 * object, we can collapse it into the parent. 1869 */ 1870 vm_object_collapse_scan(object); 1871 1872 /* 1873 * Move the pager from backing_object to object. 1874 * 1875 * swap_pager_copy() can sleep, in which case the 1876 * backing_object's and object's locks are released and 1877 * reacquired. 1878 */ 1879 swap_pager_copy(backing_object, object, 1880 OFF_TO_IDX(object->backing_object_offset), TRUE); 1881 1882 /* 1883 * Object now shadows whatever backing_object did. 1884 */ 1885 vm_object_clear_flag(object, OBJ_COLLAPSING); 1886 vm_object_backing_transfer(object, backing_object); 1887 object->backing_object_offset += 1888 backing_object->backing_object_offset; 1889 VM_OBJECT_WUNLOCK(object); 1890 vm_object_pip_wakeup(object); 1891 1892 /* 1893 * Discard backing_object. 1894 * 1895 * Since the backing object has no pages, no pager left, 1896 * and no object references within it, all that is 1897 * necessary is to dispose of it. 1898 */ 1899 KASSERT(backing_object->ref_count == 1, ( 1900 "backing_object %p was somehow re-referenced during collapse!", 1901 backing_object)); 1902 vm_object_pip_wakeup(backing_object); 1903 (void)refcount_release(&backing_object->ref_count); 1904 umtx_shm_object_terminated(backing_object); 1905 vm_object_terminate(backing_object); 1906 counter_u64_add(object_collapses, 1); 1907 VM_OBJECT_WLOCK(object); 1908 } else { 1909 /* 1910 * If we do not entirely shadow the backing object, 1911 * there is nothing we can do so we give up. 1912 * 1913 * The object lock and backing_object lock must not 1914 * be dropped during this sequence. 1915 */ 1916 if (!swap_pager_scan_all_shadowed(object)) { 1917 VM_OBJECT_WUNLOCK(backing_object); 1918 break; 1919 } 1920 1921 /* 1922 * Make the parent shadow the next object in the 1923 * chain. Deallocating backing_object will not remove 1924 * it, since its reference count is at least 2. 1925 */ 1926 vm_object_backing_remove_locked(object); 1927 new_backing_object = backing_object->backing_object; 1928 if (new_backing_object != NULL) { 1929 vm_object_backing_insert_ref(object, 1930 new_backing_object); 1931 object->backing_object_offset += 1932 backing_object->backing_object_offset; 1933 } 1934 1935 /* 1936 * Drop the reference count on backing_object. Since 1937 * its ref_count was at least 2, it will not vanish. 1938 */ 1939 (void)refcount_release(&backing_object->ref_count); 1940 KASSERT(backing_object->ref_count >= 1, ( 1941 "backing_object %p was somehow dereferenced during collapse!", 1942 backing_object)); 1943 VM_OBJECT_WUNLOCK(backing_object); 1944 counter_u64_add(object_bypasses, 1); 1945 } 1946 1947 /* 1948 * Try again with this object's new backing object. 1949 */ 1950 } 1951 } 1952 1953 /* 1954 * vm_object_page_remove: 1955 * 1956 * For the given object, either frees or invalidates each of the 1957 * specified pages. In general, a page is freed. However, if a page is 1958 * wired for any reason other than the existence of a managed, wired 1959 * mapping, then it may be invalidated but not removed from the object. 1960 * Pages are specified by the given range ["start", "end") and the option 1961 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range 1962 * extends from "start" to the end of the object. If the option 1963 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the 1964 * specified range are affected. If the option OBJPR_NOTMAPPED is 1965 * specified, then the pages within the specified range must have no 1966 * mappings. Otherwise, if this option is not specified, any mappings to 1967 * the specified pages are removed before the pages are freed or 1968 * invalidated. 1969 * 1970 * In general, this operation should only be performed on objects that 1971 * contain managed pages. There are, however, two exceptions. First, it 1972 * is performed on the kernel and kmem objects by vm_map_entry_delete(). 1973 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- 1974 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must 1975 * not be specified and the option OBJPR_NOTMAPPED must be specified. 1976 * 1977 * The object must be locked. 1978 */ 1979 void 1980 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1981 int options) 1982 { 1983 struct pctrie_iter pages; 1984 vm_page_t p; 1985 1986 VM_OBJECT_ASSERT_WLOCKED(object); 1987 KASSERT((object->flags & OBJ_UNMANAGED) == 0 || 1988 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, 1989 ("vm_object_page_remove: illegal options for object %p", object)); 1990 if (object->resident_page_count == 0) 1991 return; 1992 vm_object_pip_add(object, 1); 1993 vm_page_iter_limit_init(&pages, object, end); 1994 again: 1995 KASSERT(pctrie_iter_is_reset(&pages), 1996 ("%s: pctrie_iter not reset for retry", __func__)); 1997 for (p = vm_radix_iter_lookup_ge(&pages, start); p != NULL; 1998 p = vm_radix_iter_step(&pages)) { 1999 /* 2000 * Skip invalid pages if asked to do so. Try to avoid acquiring 2001 * the busy lock, as some consumers rely on this to avoid 2002 * deadlocks. 2003 * 2004 * A thread may concurrently transition the page from invalid to 2005 * valid using only the busy lock, so the result of this check 2006 * is immediately stale. It is up to consumers to handle this, 2007 * for instance by ensuring that all invalid->valid transitions 2008 * happen with a mutex held, as may be possible for a 2009 * filesystem. 2010 */ 2011 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) 2012 continue; 2013 2014 /* 2015 * If the page is wired for any reason besides the existence 2016 * of managed, wired mappings, then it cannot be freed. For 2017 * example, fictitious pages, which represent device memory, 2018 * are inherently wired and cannot be freed. They can, 2019 * however, be invalidated if the option OBJPR_CLEANONLY is 2020 * not specified. 2021 */ 2022 if (vm_page_tryxbusy(p) == 0) { 2023 if (vm_page_busy_sleep(p, "vmopar", 0)) 2024 VM_OBJECT_WLOCK(object); 2025 pctrie_iter_reset(&pages); 2026 goto again; 2027 } 2028 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) { 2029 vm_page_xunbusy(p); 2030 continue; 2031 } 2032 if (vm_page_wired(p)) { 2033 wired: 2034 if ((options & OBJPR_NOTMAPPED) == 0 && 2035 object->ref_count != 0) 2036 pmap_remove_all(p); 2037 if ((options & OBJPR_CLEANONLY) == 0) { 2038 vm_page_invalid(p); 2039 vm_page_undirty(p); 2040 } 2041 vm_page_xunbusy(p); 2042 continue; 2043 } 2044 KASSERT((p->flags & PG_FICTITIOUS) == 0, 2045 ("vm_object_page_remove: page %p is fictitious", p)); 2046 if ((options & OBJPR_CLEANONLY) != 0 && 2047 !vm_page_none_valid(p)) { 2048 if ((options & OBJPR_NOTMAPPED) == 0 && 2049 object->ref_count != 0 && 2050 !vm_page_try_remove_write(p)) 2051 goto wired; 2052 if (p->dirty != 0) { 2053 vm_page_xunbusy(p); 2054 continue; 2055 } 2056 } 2057 if ((options & OBJPR_NOTMAPPED) == 0 && 2058 object->ref_count != 0 && !vm_page_try_remove_all(p)) 2059 goto wired; 2060 vm_page_iter_free(&pages, p); 2061 } 2062 vm_object_pip_wakeup(object); 2063 2064 vm_pager_freespace(object, start, (end == 0 ? object->size : end) - 2065 start); 2066 } 2067 2068 /* 2069 * vm_object_page_noreuse: 2070 * 2071 * For the given object, attempt to move the specified pages to 2072 * the head of the inactive queue. This bypasses regular LRU 2073 * operation and allows the pages to be reused quickly under memory 2074 * pressure. If a page is wired for any reason, then it will not 2075 * be queued. Pages are specified by the range ["start", "end"). 2076 * As a special case, if "end" is zero, then the range extends from 2077 * "start" to the end of the object. 2078 * 2079 * This operation should only be performed on objects that 2080 * contain non-fictitious, managed pages. 2081 * 2082 * The object must be locked. 2083 */ 2084 void 2085 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2086 { 2087 struct pctrie_iter pages; 2088 vm_page_t p; 2089 2090 VM_OBJECT_ASSERT_LOCKED(object); 2091 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, 2092 ("vm_object_page_noreuse: illegal object %p", object)); 2093 if (object->resident_page_count == 0) 2094 return; 2095 2096 vm_page_iter_limit_init(&pages, object, end); 2097 VM_RADIX_FOREACH_FROM(p, &pages, start) 2098 vm_page_deactivate_noreuse(p); 2099 } 2100 2101 /* 2102 * Populate the specified range of the object with valid pages. Returns 2103 * TRUE if the range is successfully populated and FALSE otherwise. 2104 * 2105 * Note: This function should be optimized to pass a larger array of 2106 * pages to vm_pager_get_pages() before it is applied to a non- 2107 * OBJT_DEVICE object. 2108 * 2109 * The object must be locked. 2110 */ 2111 boolean_t 2112 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2113 { 2114 struct pctrie_iter pages; 2115 vm_page_t m; 2116 vm_pindex_t pindex; 2117 int rv; 2118 2119 vm_page_iter_init(&pages, object); 2120 VM_OBJECT_ASSERT_WLOCKED(object); 2121 for (pindex = start; pindex < end; pindex++) { 2122 rv = vm_page_grab_valid_iter(&m, object, pindex, 2123 VM_ALLOC_NORMAL, &pages); 2124 if (rv != VM_PAGER_OK) 2125 break; 2126 2127 /* 2128 * Keep "m" busy because a subsequent iteration may unlock 2129 * the object. 2130 */ 2131 } 2132 if (pindex > start) { 2133 pages.limit = pindex; 2134 VM_RADIX_FORALL_FROM(m, &pages, start) 2135 vm_page_xunbusy(m); 2136 } 2137 return (pindex == end); 2138 } 2139 2140 /* 2141 * Routine: vm_object_coalesce 2142 * Function: Coalesces two objects backing up adjoining 2143 * regions of memory into a single object. 2144 * 2145 * returns TRUE if objects were combined. 2146 * 2147 * NOTE: Only works at the moment if the second object is NULL - 2148 * if it's not, which object do we lock first? 2149 * 2150 * Parameters: 2151 * prev_object First object to coalesce 2152 * prev_offset Offset into prev_object 2153 * prev_size Size of reference to prev_object 2154 * next_size Size of reference to the second object 2155 * reserved Indicator that extension region has 2156 * swap accounted for 2157 * 2158 * Conditions: 2159 * The object must *not* be locked. 2160 */ 2161 boolean_t 2162 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2163 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2164 { 2165 vm_pindex_t next_pindex; 2166 2167 if (prev_object == NULL) 2168 return (TRUE); 2169 if ((prev_object->flags & OBJ_ANON) == 0) 2170 return (FALSE); 2171 2172 VM_OBJECT_WLOCK(prev_object); 2173 /* 2174 * Try to collapse the object first. 2175 */ 2176 vm_object_collapse(prev_object); 2177 2178 /* 2179 * Can't coalesce if: . more than one reference . paged out . shadows 2180 * another object . has a copy elsewhere (any of which mean that the 2181 * pages not mapped to prev_entry may be in use anyway) 2182 */ 2183 if (prev_object->backing_object != NULL) { 2184 VM_OBJECT_WUNLOCK(prev_object); 2185 return (FALSE); 2186 } 2187 2188 prev_size >>= PAGE_SHIFT; 2189 next_size >>= PAGE_SHIFT; 2190 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2191 2192 if (prev_object->ref_count > 1 && 2193 prev_object->size != next_pindex && 2194 (prev_object->flags & OBJ_ONEMAPPING) == 0) { 2195 VM_OBJECT_WUNLOCK(prev_object); 2196 return (FALSE); 2197 } 2198 2199 /* 2200 * Account for the charge. 2201 */ 2202 if (prev_object->cred != NULL) { 2203 /* 2204 * If prev_object was charged, then this mapping, 2205 * although not charged now, may become writable 2206 * later. Non-NULL cred in the object would prevent 2207 * swap reservation during enabling of the write 2208 * access, so reserve swap now. Failed reservation 2209 * cause allocation of the separate object for the map 2210 * entry, and swap reservation for this entry is 2211 * managed in appropriate time. 2212 */ 2213 if (!reserved && !swap_reserve_by_cred(ptoa(next_size), 2214 prev_object->cred)) { 2215 VM_OBJECT_WUNLOCK(prev_object); 2216 return (FALSE); 2217 } 2218 prev_object->charge += ptoa(next_size); 2219 } 2220 2221 /* 2222 * Remove any pages that may still be in the object from a previous 2223 * deallocation. 2224 */ 2225 if (next_pindex < prev_object->size) { 2226 vm_object_page_remove(prev_object, next_pindex, next_pindex + 2227 next_size, 0); 2228 #if 0 2229 if (prev_object->cred != NULL) { 2230 KASSERT(prev_object->charge >= 2231 ptoa(prev_object->size - next_pindex), 2232 ("object %p overcharged 1 %jx %jx", prev_object, 2233 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2234 prev_object->charge -= ptoa(prev_object->size - 2235 next_pindex); 2236 } 2237 #endif 2238 } 2239 2240 /* 2241 * Extend the object if necessary. 2242 */ 2243 if (next_pindex + next_size > prev_object->size) 2244 prev_object->size = next_pindex + next_size; 2245 2246 VM_OBJECT_WUNLOCK(prev_object); 2247 return (TRUE); 2248 } 2249 2250 /* 2251 * Fill in the m_dst array with up to *rbehind optional pages before m_src[0] 2252 * and up to *rahead optional pages after m_src[count - 1]. In both cases, stop 2253 * the filling-in short on encountering a cached page, an object boundary limit, 2254 * or an allocation error. Update *rbehind and *rahead to indicate the number 2255 * of pages allocated. Copy elements of m_src into array elements from 2256 * m_dst[*rbehind] to m_dst[*rbehind + count -1]. 2257 */ 2258 void 2259 vm_object_prepare_buf_pages(vm_object_t object, vm_page_t *ma_dst, int count, 2260 int *rbehind, int *rahead, vm_page_t *ma_src) 2261 { 2262 struct pctrie_iter pages; 2263 vm_pindex_t pindex; 2264 vm_page_t m, mpred, msucc; 2265 2266 vm_page_iter_init(&pages, object); 2267 VM_OBJECT_ASSERT_LOCKED(object); 2268 if (*rbehind != 0) { 2269 m = ma_src[0]; 2270 pindex = m->pindex; 2271 mpred = vm_radix_iter_lookup_lt(&pages, pindex); 2272 *rbehind = MIN(*rbehind, 2273 pindex - (mpred != NULL ? mpred->pindex + 1 : 0)); 2274 for (int i = 0; i < *rbehind; i++) { 2275 m = vm_page_alloc_iter(object, pindex - i - 1, 2276 VM_ALLOC_NORMAL, &pages); 2277 if (m == NULL) { 2278 /* Shift the array. */ 2279 for (int j = 0; j < i; j++) 2280 ma_dst[j] = ma_dst[j + *rbehind - i]; 2281 *rbehind = i; 2282 *rahead = 0; 2283 break; 2284 } 2285 ma_dst[*rbehind - i - 1] = m; 2286 } 2287 } 2288 for (int i = 0; i < count; i++) 2289 ma_dst[*rbehind + i] = ma_src[i]; 2290 if (*rahead != 0) { 2291 m = ma_src[count - 1]; 2292 pindex = m->pindex + 1; 2293 msucc = vm_radix_iter_lookup_ge(&pages, pindex); 2294 *rahead = MIN(*rahead, 2295 (msucc != NULL ? msucc->pindex : object->size) - pindex); 2296 for (int i = 0; i < *rahead; i++) { 2297 m = vm_page_alloc_iter(object, pindex + i, 2298 VM_ALLOC_NORMAL, &pages); 2299 if (m == NULL) { 2300 *rahead = i; 2301 break; 2302 } 2303 ma_dst[*rbehind + count + i] = m; 2304 } 2305 } 2306 } 2307 2308 void 2309 vm_object_set_writeable_dirty_(vm_object_t object) 2310 { 2311 atomic_add_int(&object->generation, 1); 2312 } 2313 2314 bool 2315 vm_object_mightbedirty_(vm_object_t object) 2316 { 2317 return (object->generation != object->cleangeneration); 2318 } 2319 2320 /* 2321 * vm_object_unwire: 2322 * 2323 * For each page offset within the specified range of the given object, 2324 * find the highest-level page in the shadow chain and unwire it. A page 2325 * must exist at every page offset, and the highest-level page must be 2326 * wired. 2327 */ 2328 void 2329 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, 2330 uint8_t queue) 2331 { 2332 struct pctrie_iter pages; 2333 vm_object_t tobject, t1object; 2334 vm_page_t m, tm; 2335 vm_pindex_t end_pindex, pindex, tpindex; 2336 int depth, locked_depth; 2337 2338 KASSERT((offset & PAGE_MASK) == 0, 2339 ("vm_object_unwire: offset is not page aligned")); 2340 KASSERT((length & PAGE_MASK) == 0, 2341 ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); 2342 /* The wired count of a fictitious page never changes. */ 2343 if ((object->flags & OBJ_FICTITIOUS) != 0) 2344 return; 2345 pindex = OFF_TO_IDX(offset); 2346 end_pindex = pindex + atop(length); 2347 vm_page_iter_init(&pages, object); 2348 again: 2349 locked_depth = 1; 2350 VM_OBJECT_RLOCK(object); 2351 m = vm_radix_iter_lookup_ge(&pages, pindex); 2352 while (pindex < end_pindex) { 2353 if (m == NULL || pindex < m->pindex) { 2354 /* 2355 * The first object in the shadow chain doesn't 2356 * contain a page at the current index. Therefore, 2357 * the page must exist in a backing object. 2358 */ 2359 tobject = object; 2360 tpindex = pindex; 2361 depth = 0; 2362 do { 2363 tpindex += 2364 OFF_TO_IDX(tobject->backing_object_offset); 2365 tobject = tobject->backing_object; 2366 KASSERT(tobject != NULL, 2367 ("vm_object_unwire: missing page")); 2368 if ((tobject->flags & OBJ_FICTITIOUS) != 0) 2369 goto next_page; 2370 depth++; 2371 if (depth == locked_depth) { 2372 locked_depth++; 2373 VM_OBJECT_RLOCK(tobject); 2374 } 2375 } while ((tm = vm_page_lookup(tobject, tpindex)) == 2376 NULL); 2377 } else { 2378 tm = m; 2379 m = vm_radix_iter_step(&pages); 2380 } 2381 if (vm_page_trysbusy(tm) == 0) { 2382 for (tobject = object; locked_depth >= 1; 2383 locked_depth--) { 2384 t1object = tobject->backing_object; 2385 if (tm->object != tobject) 2386 VM_OBJECT_RUNLOCK(tobject); 2387 tobject = t1object; 2388 } 2389 tobject = tm->object; 2390 if (!vm_page_busy_sleep(tm, "unwbo", 2391 VM_ALLOC_IGN_SBUSY)) 2392 VM_OBJECT_RUNLOCK(tobject); 2393 pctrie_iter_reset(&pages); 2394 goto again; 2395 } 2396 vm_page_unwire(tm, queue); 2397 vm_page_sunbusy(tm); 2398 next_page: 2399 pindex++; 2400 } 2401 /* Release the accumulated object locks. */ 2402 for (tobject = object; locked_depth >= 1; locked_depth--) { 2403 t1object = tobject->backing_object; 2404 VM_OBJECT_RUNLOCK(tobject); 2405 tobject = t1object; 2406 } 2407 } 2408 2409 /* 2410 * Return the vnode for the given object, or NULL if none exists. 2411 * For tmpfs objects, the function may return NULL if there is 2412 * no vnode allocated at the time of the call. 2413 */ 2414 struct vnode * 2415 vm_object_vnode(vm_object_t object) 2416 { 2417 struct vnode *vp; 2418 2419 VM_OBJECT_ASSERT_LOCKED(object); 2420 vm_pager_getvp(object, &vp, NULL); 2421 return (vp); 2422 } 2423 2424 /* 2425 * Busy the vm object. This prevents new pages belonging to the object from 2426 * becoming busy. Existing pages persist as busy. Callers are responsible 2427 * for checking page state before proceeding. 2428 */ 2429 void 2430 vm_object_busy(vm_object_t obj) 2431 { 2432 2433 VM_OBJECT_ASSERT_LOCKED(obj); 2434 2435 blockcount_acquire(&obj->busy, 1); 2436 /* The fence is required to order loads of page busy. */ 2437 atomic_thread_fence_acq_rel(); 2438 } 2439 2440 void 2441 vm_object_unbusy(vm_object_t obj) 2442 { 2443 2444 blockcount_release(&obj->busy, 1); 2445 } 2446 2447 void 2448 vm_object_busy_wait(vm_object_t obj, const char *wmesg) 2449 { 2450 2451 VM_OBJECT_ASSERT_UNLOCKED(obj); 2452 2453 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM); 2454 } 2455 2456 /* 2457 * This function aims to determine if the object is mapped, 2458 * specifically, if it is referenced by a vm_map_entry. Because 2459 * objects occasionally acquire transient references that do not 2460 * represent a mapping, the method used here is inexact. However, it 2461 * has very low overhead and is good enough for the advisory 2462 * vm.vmtotal sysctl. 2463 */ 2464 bool 2465 vm_object_is_active(vm_object_t obj) 2466 { 2467 2468 return (obj->ref_count > atomic_load_int(&obj->shadow_count)); 2469 } 2470 2471 static int 2472 vm_object_list_handler(struct sysctl_req *req, bool swap_only) 2473 { 2474 struct pctrie_iter pages; 2475 struct kinfo_vmobject *kvo; 2476 char *fullpath, *freepath; 2477 struct vnode *vp; 2478 struct vattr va; 2479 vm_object_t obj; 2480 vm_page_t m; 2481 u_long sp; 2482 int count, error; 2483 key_t key; 2484 unsigned short seq; 2485 bool want_path; 2486 2487 if (req->oldptr == NULL) { 2488 /* 2489 * If an old buffer has not been provided, generate an 2490 * estimate of the space needed for a subsequent call. 2491 */ 2492 mtx_lock(&vm_object_list_mtx); 2493 count = 0; 2494 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2495 if (obj->type == OBJT_DEAD) 2496 continue; 2497 count++; 2498 } 2499 mtx_unlock(&vm_object_list_mtx); 2500 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * 2501 count * 11 / 10)); 2502 } 2503 2504 want_path = !(swap_only || jailed(curthread->td_ucred)); 2505 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO); 2506 error = 0; 2507 2508 /* 2509 * VM objects are type stable and are never removed from the 2510 * list once added. This allows us to safely read obj->object_list 2511 * after reacquiring the VM object lock. 2512 */ 2513 mtx_lock(&vm_object_list_mtx); 2514 TAILQ_FOREACH(obj, &vm_object_list, object_list) { 2515 if (obj->type == OBJT_DEAD || 2516 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) 2517 continue; 2518 VM_OBJECT_RLOCK(obj); 2519 if (obj->type == OBJT_DEAD || 2520 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) { 2521 VM_OBJECT_RUNLOCK(obj); 2522 continue; 2523 } 2524 mtx_unlock(&vm_object_list_mtx); 2525 kvo->kvo_size = ptoa(obj->size); 2526 kvo->kvo_resident = obj->resident_page_count; 2527 kvo->kvo_ref_count = obj->ref_count; 2528 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count); 2529 kvo->kvo_memattr = obj->memattr; 2530 kvo->kvo_active = 0; 2531 kvo->kvo_inactive = 0; 2532 kvo->kvo_flags = 0; 2533 if (!swap_only) { 2534 vm_page_iter_init(&pages, obj); 2535 VM_RADIX_FOREACH(m, &pages) { 2536 /* 2537 * A page may belong to the object but be 2538 * dequeued and set to PQ_NONE while the 2539 * object lock is not held. This makes the 2540 * reads of m->queue below racy, and we do not 2541 * count pages set to PQ_NONE. However, this 2542 * sysctl is only meant to give an 2543 * approximation of the system anyway. 2544 */ 2545 if (vm_page_active(m)) 2546 kvo->kvo_active++; 2547 else if (vm_page_inactive(m)) 2548 kvo->kvo_inactive++; 2549 else if (vm_page_in_laundry(m)) 2550 kvo->kvo_laundry++; 2551 } 2552 } 2553 2554 kvo->kvo_vn_fileid = 0; 2555 kvo->kvo_vn_fsid = 0; 2556 kvo->kvo_vn_fsid_freebsd11 = 0; 2557 freepath = NULL; 2558 fullpath = ""; 2559 vp = NULL; 2560 kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp : 2561 NULL); 2562 if (vp != NULL) { 2563 vref(vp); 2564 } else if ((obj->flags & OBJ_ANON) != 0) { 2565 MPASS(kvo->kvo_type == KVME_TYPE_SWAP); 2566 kvo->kvo_me = (uintptr_t)obj; 2567 /* tmpfs objs are reported as vnodes */ 2568 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object; 2569 sp = swap_pager_swapped_pages(obj); 2570 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp; 2571 } 2572 if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) { 2573 cdev_pager_get_path(obj, kvo->kvo_path, 2574 sizeof(kvo->kvo_path)); 2575 } 2576 VM_OBJECT_RUNLOCK(obj); 2577 if ((obj->flags & OBJ_SYSVSHM) != 0) { 2578 kvo->kvo_flags |= KVMO_FLAG_SYSVSHM; 2579 shmobjinfo(obj, &key, &seq); 2580 kvo->kvo_vn_fileid = key; 2581 kvo->kvo_vn_fsid_freebsd11 = seq; 2582 } 2583 if ((obj->flags & OBJ_POSIXSHM) != 0) { 2584 kvo->kvo_flags |= KVMO_FLAG_POSIXSHM; 2585 shm_get_path(obj, kvo->kvo_path, 2586 sizeof(kvo->kvo_path)); 2587 } 2588 if (vp != NULL) { 2589 vn_fullpath(vp, &fullpath, &freepath); 2590 vn_lock(vp, LK_SHARED | LK_RETRY); 2591 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { 2592 kvo->kvo_vn_fileid = va.va_fileid; 2593 kvo->kvo_vn_fsid = va.va_fsid; 2594 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; 2595 /* truncate */ 2596 } 2597 vput(vp); 2598 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); 2599 free(freepath, M_TEMP); 2600 } 2601 2602 /* Pack record size down */ 2603 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) 2604 + strlen(kvo->kvo_path) + 1; 2605 kvo->kvo_structsize = roundup(kvo->kvo_structsize, 2606 sizeof(uint64_t)); 2607 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); 2608 maybe_yield(); 2609 mtx_lock(&vm_object_list_mtx); 2610 if (error) 2611 break; 2612 } 2613 mtx_unlock(&vm_object_list_mtx); 2614 free(kvo, M_TEMP); 2615 return (error); 2616 } 2617 2618 static int 2619 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) 2620 { 2621 return (vm_object_list_handler(req, false)); 2622 } 2623 2624 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | 2625 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", 2626 "List of VM objects"); 2627 2628 static int 2629 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS) 2630 { 2631 return (vm_object_list_handler(req, true)); 2632 } 2633 2634 /* 2635 * This sysctl returns list of the anonymous or swap objects. Intent 2636 * is to provide stripped optimized list useful to analyze swap use. 2637 * Since technically non-swap (default) objects participate in the 2638 * shadow chains, and are converted to swap type as needed by swap 2639 * pager, we must report them. 2640 */ 2641 SYSCTL_PROC(_vm, OID_AUTO, swap_objects, 2642 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, 2643 sysctl_vm_object_list_swap, "S,kinfo_vmobject", 2644 "List of swap VM objects"); 2645 2646 #include "opt_ddb.h" 2647 #ifdef DDB 2648 #include <sys/kernel.h> 2649 2650 #include <sys/cons.h> 2651 2652 #include <ddb/ddb.h> 2653 2654 static int 2655 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2656 { 2657 vm_map_t tmpm; 2658 vm_map_entry_t tmpe; 2659 vm_object_t obj; 2660 2661 if (map == 0) 2662 return 0; 2663 2664 if (entry == 0) { 2665 VM_MAP_ENTRY_FOREACH(tmpe, map) { 2666 if (_vm_object_in_map(map, object, tmpe)) { 2667 return 1; 2668 } 2669 } 2670 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2671 tmpm = entry->object.sub_map; 2672 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { 2673 if (_vm_object_in_map(tmpm, object, tmpe)) { 2674 return 1; 2675 } 2676 } 2677 } else if ((obj = entry->object.vm_object) != NULL) { 2678 for (; obj; obj = obj->backing_object) 2679 if (obj == object) { 2680 return 1; 2681 } 2682 } 2683 return 0; 2684 } 2685 2686 static int 2687 vm_object_in_map(vm_object_t object) 2688 { 2689 struct proc *p; 2690 2691 /* sx_slock(&allproc_lock); */ 2692 FOREACH_PROC_IN_SYSTEM(p) { 2693 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2694 continue; 2695 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2696 /* sx_sunlock(&allproc_lock); */ 2697 return 1; 2698 } 2699 } 2700 /* sx_sunlock(&allproc_lock); */ 2701 if (_vm_object_in_map(kernel_map, object, 0)) 2702 return 1; 2703 return 0; 2704 } 2705 2706 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE) 2707 { 2708 vm_object_t object; 2709 2710 /* 2711 * make sure that internal objs are in a map somewhere 2712 * and none have zero ref counts. 2713 */ 2714 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2715 if ((object->flags & OBJ_ANON) != 0) { 2716 if (object->ref_count == 0) { 2717 db_printf( 2718 "vmochk: internal obj has zero ref count: %lu\n", 2719 (u_long)object->size); 2720 } 2721 if (!vm_object_in_map(object)) { 2722 db_printf( 2723 "vmochk: internal obj is not in a map: " 2724 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2725 object->ref_count, (u_long)object->size, 2726 (u_long)object->size, 2727 (void *)object->backing_object); 2728 } 2729 } 2730 if (db_pager_quit) 2731 return; 2732 } 2733 } 2734 2735 /* 2736 * vm_object_print: [ debug ] 2737 */ 2738 DB_SHOW_COMMAND(object, vm_object_print_static) 2739 { 2740 struct pctrie_iter pages; 2741 /* XXX convert args. */ 2742 vm_object_t object = (vm_object_t)addr; 2743 boolean_t full = have_addr; 2744 2745 vm_page_t p; 2746 2747 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2748 #define count was_count 2749 2750 int count; 2751 2752 if (object == NULL) 2753 return; 2754 2755 db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x", 2756 object, (int)object->type, (uintmax_t)object->size, 2757 object->resident_page_count, object->ref_count, object->flags); 2758 db_iprintf(" ruid %d charge %jx\n", 2759 object->cred ? object->cred->cr_ruid : -1, 2760 (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 vm_page_iter_init(&pages, object); 2772 VM_RADIX_FOREACH(p, &pages) { 2773 if (count == 0) 2774 db_iprintf("memory:="); 2775 else if (count == 6) { 2776 db_printf("\n"); 2777 db_iprintf(" ..."); 2778 count = 0; 2779 } else 2780 db_printf(","); 2781 count++; 2782 2783 db_printf("(off=0x%jx,page=0x%jx)", 2784 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2785 2786 if (db_pager_quit) 2787 break; 2788 } 2789 if (count != 0) 2790 db_printf("\n"); 2791 db_indent -= 2; 2792 } 2793 2794 /* XXX. */ 2795 #undef count 2796 2797 /* XXX need this non-static entry for calling from vm_map_print. */ 2798 void 2799 vm_object_print( 2800 /* db_expr_t */ long addr, 2801 boolean_t have_addr, 2802 /* db_expr_t */ long count, 2803 char *modif) 2804 { 2805 vm_object_print_static(addr, have_addr, count, modif); 2806 } 2807 2808 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE) 2809 { 2810 struct pctrie_iter pages; 2811 vm_object_t object; 2812 vm_page_t m, start_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 start_m = NULL; 2820 vm_page_iter_init(&pages, object); 2821 VM_RADIX_FOREACH(m, &pages) { 2822 if (start_m == NULL) { 2823 start_m = m; 2824 rcount = 0; 2825 } else if (start_m->pindex + rcount != m->pindex || 2826 VM_PAGE_TO_PHYS(start_m) + ptoa(rcount) != 2827 VM_PAGE_TO_PHYS(m)) { 2828 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2829 (long)start_m->pindex, rcount, 2830 (long)VM_PAGE_TO_PHYS(start_m)); 2831 if (db_pager_quit) 2832 return; 2833 start_m = m; 2834 rcount = 0; 2835 } 2836 rcount++; 2837 } 2838 if (start_m != NULL) { 2839 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2840 (long)start_m->pindex, rcount, 2841 (long)VM_PAGE_TO_PHYS(start_m)); 2842 if (db_pager_quit) 2843 return; 2844 } 2845 } 2846 } 2847 #endif /* DDB */ 2848