1 /*- 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 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 <sys/cdefs.h> 66 __FBSDID("$FreeBSD$"); 67 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/lock.h> 73 #include <sys/mman.h> 74 #include <sys/mount.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/mutex.h> 78 #include <sys/proc.h> /* for curproc, pageproc */ 79 #include <sys/socket.h> 80 #include <sys/resourcevar.h> 81 #include <sys/vnode.h> 82 #include <sys/vmmeter.h> 83 #include <sys/sx.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pageout.h> 92 #include <vm/vm_pager.h> 93 #include <vm/swap_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/vm_reserv.h> 97 #include <vm/uma.h> 98 99 #define EASY_SCAN_FACTOR 8 100 101 #define MSYNC_FLUSH_HARDSEQ 0x01 102 #define MSYNC_FLUSH_SOFTSEQ 0x02 103 104 /* 105 * msync / VM object flushing optimizations 106 */ 107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ; 108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0, 109 "Enable sequential iteration optimization"); 110 111 static int old_msync; 112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, 113 "Use old (insecure) msync behavior"); 114 115 static void vm_object_qcollapse(vm_object_t object); 116 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags); 117 static void vm_object_vndeallocate(vm_object_t object); 118 119 /* 120 * Virtual memory objects maintain the actual data 121 * associated with allocated virtual memory. A given 122 * page of memory exists within exactly one object. 123 * 124 * An object is only deallocated when all "references" 125 * are given up. Only one "reference" to a given 126 * region of an object should be writeable. 127 * 128 * Associated with each object is a list of all resident 129 * memory pages belonging to that object; this list is 130 * maintained by the "vm_page" module, and locked by the object's 131 * lock. 132 * 133 * Each object also records a "pager" routine which is 134 * used to retrieve (and store) pages to the proper backing 135 * storage. In addition, objects may be backed by other 136 * objects from which they were virtual-copied. 137 * 138 * The only items within the object structure which are 139 * modified after time of creation are: 140 * reference count locked by object's lock 141 * pager routine locked by object's lock 142 * 143 */ 144 145 struct object_q vm_object_list; 146 struct mtx vm_object_list_mtx; /* lock for object list and count */ 147 148 struct vm_object kernel_object_store; 149 struct vm_object kmem_object_store; 150 151 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); 152 153 static long object_collapses; 154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, 155 &object_collapses, 0, "VM object collapses"); 156 157 static long object_bypasses; 158 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, 159 &object_bypasses, 0, "VM object bypasses"); 160 161 static uma_zone_t obj_zone; 162 163 static int vm_object_zinit(void *mem, int size, int flags); 164 165 #ifdef INVARIANTS 166 static void vm_object_zdtor(void *mem, int size, void *arg); 167 168 static void 169 vm_object_zdtor(void *mem, int size, void *arg) 170 { 171 vm_object_t object; 172 173 object = (vm_object_t)mem; 174 KASSERT(TAILQ_EMPTY(&object->memq), 175 ("object %p has resident pages", 176 object)); 177 #if VM_NRESERVLEVEL > 0 178 KASSERT(LIST_EMPTY(&object->rvq), 179 ("object %p has reservations", 180 object)); 181 #endif 182 KASSERT(object->cache == NULL, 183 ("object %p has cached pages", 184 object)); 185 KASSERT(object->paging_in_progress == 0, 186 ("object %p paging_in_progress = %d", 187 object, object->paging_in_progress)); 188 KASSERT(object->resident_page_count == 0, 189 ("object %p resident_page_count = %d", 190 object, object->resident_page_count)); 191 KASSERT(object->shadow_count == 0, 192 ("object %p shadow_count = %d", 193 object, object->shadow_count)); 194 } 195 #endif 196 197 static int 198 vm_object_zinit(void *mem, int size, int flags) 199 { 200 vm_object_t object; 201 202 object = (vm_object_t)mem; 203 bzero(&object->mtx, sizeof(object->mtx)); 204 VM_OBJECT_LOCK_INIT(object, "standard object"); 205 206 /* These are true for any object that has been freed */ 207 object->paging_in_progress = 0; 208 object->resident_page_count = 0; 209 object->shadow_count = 0; 210 return (0); 211 } 212 213 void 214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) 215 { 216 217 TAILQ_INIT(&object->memq); 218 LIST_INIT(&object->shadow_head); 219 220 object->root = NULL; 221 object->type = type; 222 object->size = size; 223 object->generation = 1; 224 object->ref_count = 1; 225 object->memattr = VM_MEMATTR_DEFAULT; 226 object->flags = 0; 227 object->uip = NULL; 228 object->charge = 0; 229 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) 230 object->flags = OBJ_ONEMAPPING; 231 object->pg_color = 0; 232 object->handle = NULL; 233 object->backing_object = NULL; 234 object->backing_object_offset = (vm_ooffset_t) 0; 235 #if VM_NRESERVLEVEL > 0 236 LIST_INIT(&object->rvq); 237 #endif 238 object->cache = NULL; 239 240 mtx_lock(&vm_object_list_mtx); 241 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); 242 mtx_unlock(&vm_object_list_mtx); 243 } 244 245 /* 246 * vm_object_init: 247 * 248 * Initialize the VM objects module. 249 */ 250 void 251 vm_object_init(void) 252 { 253 TAILQ_INIT(&vm_object_list); 254 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); 255 256 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object"); 257 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 258 kernel_object); 259 #if VM_NRESERVLEVEL > 0 260 kernel_object->flags |= OBJ_COLORED; 261 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 262 #endif 263 264 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object"); 265 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), 266 kmem_object); 267 #if VM_NRESERVLEVEL > 0 268 kmem_object->flags |= OBJ_COLORED; 269 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); 270 #endif 271 272 /* 273 * The lock portion of struct vm_object must be type stable due 274 * to vm_pageout_fallback_object_lock locking a vm object 275 * without holding any references to it. 276 */ 277 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, 278 #ifdef INVARIANTS 279 vm_object_zdtor, 280 #else 281 NULL, 282 #endif 283 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); 284 } 285 286 void 287 vm_object_clear_flag(vm_object_t object, u_short bits) 288 { 289 290 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 291 object->flags &= ~bits; 292 } 293 294 /* 295 * Sets the default memory attribute for the specified object. Pages 296 * that are allocated to this object are by default assigned this memory 297 * attribute. 298 * 299 * Presently, this function must be called before any pages are allocated 300 * to the object. In the future, this requirement may be relaxed for 301 * "default" and "swap" objects. 302 */ 303 int 304 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) 305 { 306 307 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 308 switch (object->type) { 309 case OBJT_DEFAULT: 310 case OBJT_DEVICE: 311 case OBJT_PHYS: 312 case OBJT_SG: 313 case OBJT_SWAP: 314 case OBJT_VNODE: 315 if (!TAILQ_EMPTY(&object->memq)) 316 return (KERN_FAILURE); 317 break; 318 case OBJT_DEAD: 319 return (KERN_INVALID_ARGUMENT); 320 } 321 object->memattr = memattr; 322 return (KERN_SUCCESS); 323 } 324 325 void 326 vm_object_pip_add(vm_object_t object, short i) 327 { 328 329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 330 object->paging_in_progress += i; 331 } 332 333 void 334 vm_object_pip_subtract(vm_object_t object, short i) 335 { 336 337 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 338 object->paging_in_progress -= i; 339 } 340 341 void 342 vm_object_pip_wakeup(vm_object_t object) 343 { 344 345 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 346 object->paging_in_progress--; 347 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 348 vm_object_clear_flag(object, OBJ_PIPWNT); 349 wakeup(object); 350 } 351 } 352 353 void 354 vm_object_pip_wakeupn(vm_object_t object, short i) 355 { 356 357 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 358 if (i) 359 object->paging_in_progress -= i; 360 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { 361 vm_object_clear_flag(object, OBJ_PIPWNT); 362 wakeup(object); 363 } 364 } 365 366 void 367 vm_object_pip_wait(vm_object_t object, char *waitid) 368 { 369 370 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 371 while (object->paging_in_progress) { 372 object->flags |= OBJ_PIPWNT; 373 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); 374 } 375 } 376 377 /* 378 * vm_object_allocate: 379 * 380 * Returns a new object with the given size. 381 */ 382 vm_object_t 383 vm_object_allocate(objtype_t type, vm_pindex_t size) 384 { 385 vm_object_t object; 386 387 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); 388 _vm_object_allocate(type, size, object); 389 return (object); 390 } 391 392 393 /* 394 * vm_object_reference: 395 * 396 * Gets another reference to the given object. Note: OBJ_DEAD 397 * objects can be referenced during final cleaning. 398 */ 399 void 400 vm_object_reference(vm_object_t object) 401 { 402 if (object == NULL) 403 return; 404 VM_OBJECT_LOCK(object); 405 vm_object_reference_locked(object); 406 VM_OBJECT_UNLOCK(object); 407 } 408 409 /* 410 * vm_object_reference_locked: 411 * 412 * Gets another reference to the given object. 413 * 414 * The object must be locked. 415 */ 416 void 417 vm_object_reference_locked(vm_object_t object) 418 { 419 struct vnode *vp; 420 421 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 422 object->ref_count++; 423 if (object->type == OBJT_VNODE) { 424 vp = object->handle; 425 vref(vp); 426 } 427 } 428 429 /* 430 * Handle deallocating an object of type OBJT_VNODE. 431 */ 432 static void 433 vm_object_vndeallocate(vm_object_t object) 434 { 435 struct vnode *vp = (struct vnode *) object->handle; 436 437 VFS_ASSERT_GIANT(vp->v_mount); 438 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 439 KASSERT(object->type == OBJT_VNODE, 440 ("vm_object_vndeallocate: not a vnode object")); 441 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); 442 #ifdef INVARIANTS 443 if (object->ref_count == 0) { 444 vprint("vm_object_vndeallocate", vp); 445 panic("vm_object_vndeallocate: bad object reference count"); 446 } 447 #endif 448 449 object->ref_count--; 450 if (object->ref_count == 0) { 451 mp_fixme("Unlocked vflag access."); 452 vp->v_vflag &= ~VV_TEXT; 453 } 454 VM_OBJECT_UNLOCK(object); 455 /* 456 * vrele may need a vop lock 457 */ 458 vrele(vp); 459 } 460 461 /* 462 * vm_object_deallocate: 463 * 464 * Release a reference to the specified object, 465 * gained either through a vm_object_allocate 466 * or a vm_object_reference call. When all references 467 * are gone, storage associated with this object 468 * may be relinquished. 469 * 470 * No object may be locked. 471 */ 472 void 473 vm_object_deallocate(vm_object_t object) 474 { 475 vm_object_t temp; 476 477 while (object != NULL) { 478 int vfslocked; 479 480 vfslocked = 0; 481 restart: 482 VM_OBJECT_LOCK(object); 483 if (object->type == OBJT_VNODE) { 484 struct vnode *vp = (struct vnode *) object->handle; 485 486 /* 487 * Conditionally acquire Giant for a vnode-backed 488 * object. We have to be careful since the type of 489 * a vnode object can change while the object is 490 * unlocked. 491 */ 492 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) { 493 vfslocked = 1; 494 if (!mtx_trylock(&Giant)) { 495 VM_OBJECT_UNLOCK(object); 496 mtx_lock(&Giant); 497 goto restart; 498 } 499 } 500 vm_object_vndeallocate(object); 501 VFS_UNLOCK_GIANT(vfslocked); 502 return; 503 } else 504 /* 505 * This is to handle the case that the object 506 * changed type while we dropped its lock to 507 * obtain Giant. 508 */ 509 VFS_UNLOCK_GIANT(vfslocked); 510 511 KASSERT(object->ref_count != 0, 512 ("vm_object_deallocate: object deallocated too many times: %d", object->type)); 513 514 /* 515 * If the reference count goes to 0 we start calling 516 * vm_object_terminate() on the object chain. 517 * A ref count of 1 may be a special case depending on the 518 * shadow count being 0 or 1. 519 */ 520 object->ref_count--; 521 if (object->ref_count > 1) { 522 VM_OBJECT_UNLOCK(object); 523 return; 524 } else if (object->ref_count == 1) { 525 if (object->shadow_count == 0 && 526 object->handle == NULL && 527 (object->type == OBJT_DEFAULT || 528 object->type == OBJT_SWAP)) { 529 vm_object_set_flag(object, OBJ_ONEMAPPING); 530 } else if ((object->shadow_count == 1) && 531 (object->handle == NULL) && 532 (object->type == OBJT_DEFAULT || 533 object->type == OBJT_SWAP)) { 534 vm_object_t robject; 535 536 robject = LIST_FIRST(&object->shadow_head); 537 KASSERT(robject != NULL, 538 ("vm_object_deallocate: ref_count: %d, shadow_count: %d", 539 object->ref_count, 540 object->shadow_count)); 541 if (!VM_OBJECT_TRYLOCK(robject)) { 542 /* 543 * Avoid a potential deadlock. 544 */ 545 object->ref_count++; 546 VM_OBJECT_UNLOCK(object); 547 /* 548 * More likely than not the thread 549 * holding robject's lock has lower 550 * priority than the current thread. 551 * Let the lower priority thread run. 552 */ 553 pause("vmo_de", 1); 554 continue; 555 } 556 /* 557 * Collapse object into its shadow unless its 558 * shadow is dead. In that case, object will 559 * be deallocated by the thread that is 560 * deallocating its shadow. 561 */ 562 if ((robject->flags & OBJ_DEAD) == 0 && 563 (robject->handle == NULL) && 564 (robject->type == OBJT_DEFAULT || 565 robject->type == OBJT_SWAP)) { 566 567 robject->ref_count++; 568 retry: 569 if (robject->paging_in_progress) { 570 VM_OBJECT_UNLOCK(object); 571 vm_object_pip_wait(robject, 572 "objde1"); 573 temp = robject->backing_object; 574 if (object == temp) { 575 VM_OBJECT_LOCK(object); 576 goto retry; 577 } 578 } else if (object->paging_in_progress) { 579 VM_OBJECT_UNLOCK(robject); 580 object->flags |= OBJ_PIPWNT; 581 msleep(object, 582 VM_OBJECT_MTX(object), 583 PDROP | PVM, "objde2", 0); 584 VM_OBJECT_LOCK(robject); 585 temp = robject->backing_object; 586 if (object == temp) { 587 VM_OBJECT_LOCK(object); 588 goto retry; 589 } 590 } else 591 VM_OBJECT_UNLOCK(object); 592 593 if (robject->ref_count == 1) { 594 robject->ref_count--; 595 object = robject; 596 goto doterm; 597 } 598 object = robject; 599 vm_object_collapse(object); 600 VM_OBJECT_UNLOCK(object); 601 continue; 602 } 603 VM_OBJECT_UNLOCK(robject); 604 } 605 VM_OBJECT_UNLOCK(object); 606 return; 607 } 608 doterm: 609 temp = object->backing_object; 610 if (temp != NULL) { 611 VM_OBJECT_LOCK(temp); 612 LIST_REMOVE(object, shadow_list); 613 temp->shadow_count--; 614 temp->generation++; 615 VM_OBJECT_UNLOCK(temp); 616 object->backing_object = NULL; 617 } 618 /* 619 * Don't double-terminate, we could be in a termination 620 * recursion due to the terminate having to sync data 621 * to disk. 622 */ 623 if ((object->flags & OBJ_DEAD) == 0) 624 vm_object_terminate(object); 625 else 626 VM_OBJECT_UNLOCK(object); 627 object = temp; 628 } 629 } 630 631 /* 632 * vm_object_destroy removes the object from the global object list 633 * and frees the space for the object. 634 */ 635 void 636 vm_object_destroy(vm_object_t object) 637 { 638 639 /* 640 * Remove the object from the global object list. 641 */ 642 mtx_lock(&vm_object_list_mtx); 643 TAILQ_REMOVE(&vm_object_list, object, object_list); 644 mtx_unlock(&vm_object_list_mtx); 645 646 /* 647 * Release the allocation charge. 648 */ 649 if (object->uip != NULL) { 650 KASSERT(object->type == OBJT_DEFAULT || 651 object->type == OBJT_SWAP, 652 ("vm_object_terminate: non-swap obj %p has uip", 653 object)); 654 swap_release_by_uid(object->charge, object->uip); 655 object->charge = 0; 656 uifree(object->uip); 657 object->uip = NULL; 658 } 659 660 /* 661 * Free the space for the object. 662 */ 663 uma_zfree(obj_zone, object); 664 } 665 666 /* 667 * vm_object_terminate actually destroys the specified object, freeing 668 * up all previously used resources. 669 * 670 * The object must be locked. 671 * This routine may block. 672 */ 673 void 674 vm_object_terminate(vm_object_t object) 675 { 676 vm_page_t p; 677 678 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 679 680 /* 681 * Make sure no one uses us. 682 */ 683 vm_object_set_flag(object, OBJ_DEAD); 684 685 /* 686 * wait for the pageout daemon to be done with the object 687 */ 688 vm_object_pip_wait(object, "objtrm"); 689 690 KASSERT(!object->paging_in_progress, 691 ("vm_object_terminate: pageout in progress")); 692 693 /* 694 * Clean and free the pages, as appropriate. All references to the 695 * object are gone, so we don't need to lock it. 696 */ 697 if (object->type == OBJT_VNODE) { 698 struct vnode *vp = (struct vnode *)object->handle; 699 700 /* 701 * Clean pages and flush buffers. 702 */ 703 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 704 VM_OBJECT_UNLOCK(object); 705 706 vinvalbuf(vp, V_SAVE, 0, 0); 707 708 VM_OBJECT_LOCK(object); 709 } 710 711 KASSERT(object->ref_count == 0, 712 ("vm_object_terminate: object with references, ref_count=%d", 713 object->ref_count)); 714 715 /* 716 * Now free any remaining pages. For internal objects, this also 717 * removes them from paging queues. Don't free wired pages, just 718 * remove them from the object. 719 */ 720 while ((p = TAILQ_FIRST(&object->memq)) != NULL) { 721 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0, 722 ("vm_object_terminate: freeing busy page %p " 723 "p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags)); 724 vm_page_lock(p); 725 if (p->wire_count == 0) { 726 vm_page_free(p); 727 PCPU_INC(cnt.v_pfree); 728 } else 729 vm_page_remove(p); 730 vm_page_unlock(p); 731 } 732 733 #if VM_NRESERVLEVEL > 0 734 if (__predict_false(!LIST_EMPTY(&object->rvq))) 735 vm_reserv_break_all(object); 736 #endif 737 if (__predict_false(object->cache != NULL)) 738 vm_page_cache_free(object, 0, 0); 739 740 /* 741 * Let the pager know object is dead. 742 */ 743 vm_pager_deallocate(object); 744 VM_OBJECT_UNLOCK(object); 745 746 vm_object_destroy(object); 747 } 748 749 /* 750 * vm_object_page_clean 751 * 752 * Clean all dirty pages in the specified range of object. Leaves page 753 * on whatever queue it is currently on. If NOSYNC is set then do not 754 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), 755 * leaving the object dirty. 756 * 757 * When stuffing pages asynchronously, allow clustering. XXX we need a 758 * synchronous clustering mode implementation. 759 * 760 * Odd semantics: if start == end, we clean everything. 761 * 762 * The object must be locked. 763 */ 764 void 765 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags) 766 { 767 vm_page_t p, np; 768 vm_pindex_t tstart, tend; 769 vm_pindex_t pi; 770 int clearobjflags; 771 int pagerflags; 772 int curgeneration; 773 774 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 775 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 776 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0) 777 return; 778 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object")); 779 780 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; 781 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0; 782 783 vm_object_set_flag(object, OBJ_CLEANING); 784 785 tstart = start; 786 if (end == 0) { 787 tend = object->size; 788 } else { 789 tend = end; 790 } 791 792 /* 793 * If the caller is smart and only msync()s a range he knows is 794 * dirty, we may be able to avoid an object scan. This results in 795 * a phenominal improvement in performance. We cannot do this 796 * as a matter of course because the object may be huge - e.g. 797 * the size might be in the gigabytes or terrabytes. 798 */ 799 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) { 800 vm_pindex_t tscan; 801 int scanlimit; 802 int scanreset; 803 804 scanreset = object->resident_page_count / EASY_SCAN_FACTOR; 805 if (scanreset < 16) 806 scanreset = 16; 807 pagerflags |= VM_PAGER_IGNORE_CLEANCHK; 808 809 scanlimit = scanreset; 810 tscan = tstart; 811 while (tscan < tend) { 812 curgeneration = object->generation; 813 p = vm_page_lookup(object, tscan); 814 if (p == NULL || p->valid == 0) { 815 if (--scanlimit == 0) 816 break; 817 ++tscan; 818 continue; 819 } 820 vm_page_lock(p); 821 vm_page_lock_queues(); 822 vm_page_test_dirty(p); 823 if (p->dirty == 0) { 824 vm_page_unlock_queues(); 825 vm_page_unlock(p); 826 if (--scanlimit == 0) 827 break; 828 ++tscan; 829 continue; 830 } 831 vm_page_unlock_queues(); 832 vm_page_unlock(p); 833 /* 834 * If we have been asked to skip nosync pages and 835 * this is a nosync page, we can't continue. 836 */ 837 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) { 838 if (--scanlimit == 0) 839 break; 840 ++tscan; 841 continue; 842 } 843 scanlimit = scanreset; 844 845 /* 846 * This returns 0 if it was unable to busy the first 847 * page (i.e. had to sleep). 848 */ 849 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags); 850 851 } 852 853 /* 854 * If everything was dirty and we flushed it successfully, 855 * and the requested range is not the entire object, we 856 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can 857 * return immediately. 858 */ 859 if (tscan >= tend && (tstart || tend < object->size)) { 860 vm_object_clear_flag(object, OBJ_CLEANING); 861 return; 862 } 863 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK; 864 } 865 866 /* 867 * Generally set CLEANCHK interlock and make the page read-only so 868 * we can then clear the object flags. 869 * 870 * However, if this is a nosync mmap then the object is likely to 871 * stay dirty so do not mess with the page and do not clear the 872 * object flags. 873 */ 874 clearobjflags = 1; 875 TAILQ_FOREACH(p, &object->memq, listq) { 876 p->oflags |= VPO_CLEANCHK; 877 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) 878 clearobjflags = 0; 879 else 880 pmap_remove_write(p); 881 } 882 883 if (clearobjflags && (tstart == 0) && (tend == object->size)) 884 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); 885 886 rescan: 887 curgeneration = object->generation; 888 889 for (p = TAILQ_FIRST(&object->memq); p; p = np) { 890 int n; 891 892 np = TAILQ_NEXT(p, listq); 893 894 again: 895 pi = p->pindex; 896 if ((p->oflags & VPO_CLEANCHK) == 0 || 897 (pi < tstart) || (pi >= tend) || 898 p->valid == 0) { 899 p->oflags &= ~VPO_CLEANCHK; 900 continue; 901 } 902 903 vm_page_lock(p); 904 vm_page_lock_queues(); 905 vm_page_test_dirty(p); 906 if (p->dirty == 0) { 907 vm_page_unlock_queues(); 908 vm_page_unlock(p); 909 p->oflags &= ~VPO_CLEANCHK; 910 continue; 911 } 912 vm_page_unlock_queues(); 913 vm_page_unlock(p); 914 /* 915 * If we have been asked to skip nosync pages and this is a 916 * nosync page, skip it. Note that the object flags were 917 * not cleared in this case so we do not have to set them. 918 */ 919 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) { 920 p->oflags &= ~VPO_CLEANCHK; 921 continue; 922 } 923 924 n = vm_object_page_collect_flush(object, p, 925 curgeneration, pagerflags); 926 if (n == 0) 927 goto rescan; 928 929 if (object->generation != curgeneration) 930 goto rescan; 931 932 /* 933 * Try to optimize the next page. If we can't we pick up 934 * our (random) scan where we left off. 935 */ 936 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) 937 if ((p = vm_page_lookup(object, pi + n)) != NULL) 938 goto again; 939 } 940 #if 0 941 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc); 942 #endif 943 944 vm_object_clear_flag(object, OBJ_CLEANING); 945 return; 946 } 947 948 static int 949 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags) 950 { 951 int runlen; 952 int maxf; 953 int chkb; 954 int maxb; 955 int i; 956 vm_pindex_t pi; 957 vm_page_t maf[vm_pageout_page_count]; 958 vm_page_t mab[vm_pageout_page_count]; 959 vm_page_t ma[vm_pageout_page_count]; 960 961 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 962 vm_page_lock_assert(p, MA_NOTOWNED); 963 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 964 pi = p->pindex; 965 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { 966 if (object->generation != curgeneration) { 967 return(0); 968 } 969 } 970 maxf = 0; 971 for(i = 1; i < vm_pageout_page_count; i++) { 972 vm_page_t tp; 973 974 if ((tp = vm_page_lookup(object, pi + i)) != NULL) { 975 if ((tp->oflags & VPO_BUSY) || 976 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 977 (tp->oflags & VPO_CLEANCHK) == 0) || 978 (tp->busy != 0)) 979 break; 980 vm_page_lock(tp); 981 vm_page_lock_queues(); 982 vm_page_test_dirty(tp); 983 if (tp->dirty == 0) { 984 vm_page_unlock(tp); 985 vm_page_unlock_queues(); 986 tp->oflags &= ~VPO_CLEANCHK; 987 break; 988 } 989 vm_page_unlock(tp); 990 vm_page_unlock_queues(); 991 maf[ i - 1 ] = tp; 992 maxf++; 993 continue; 994 } 995 break; 996 } 997 998 maxb = 0; 999 chkb = vm_pageout_page_count - maxf; 1000 if (chkb) { 1001 for(i = 1; i < chkb;i++) { 1002 vm_page_t tp; 1003 1004 if ((tp = vm_page_lookup(object, pi - i)) != NULL) { 1005 if ((tp->oflags & VPO_BUSY) || 1006 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 1007 (tp->oflags & VPO_CLEANCHK) == 0) || 1008 (tp->busy != 0)) 1009 break; 1010 vm_page_lock(tp); 1011 vm_page_lock_queues(); 1012 vm_page_test_dirty(tp); 1013 if (tp->dirty == 0) { 1014 vm_page_unlock_queues(); 1015 vm_page_unlock(tp); 1016 tp->oflags &= ~VPO_CLEANCHK; 1017 break; 1018 } 1019 vm_page_unlock_queues(); 1020 vm_page_unlock(tp); 1021 mab[ i - 1 ] = tp; 1022 maxb++; 1023 continue; 1024 } 1025 break; 1026 } 1027 } 1028 1029 for(i = 0; i < maxb; i++) { 1030 int index = (maxb - i) - 1; 1031 ma[index] = mab[i]; 1032 ma[index]->oflags &= ~VPO_CLEANCHK; 1033 } 1034 p->oflags &= ~VPO_CLEANCHK; 1035 ma[maxb] = p; 1036 for(i = 0; i < maxf; i++) { 1037 int index = (maxb + i) + 1; 1038 ma[index] = maf[i]; 1039 ma[index]->oflags &= ~VPO_CLEANCHK; 1040 } 1041 runlen = maxb + maxf + 1; 1042 1043 vm_pageout_flush(ma, runlen, pagerflags); 1044 for (i = 0; i < runlen; i++) { 1045 if (ma[i]->dirty) { 1046 pmap_remove_write(ma[i]); 1047 ma[i]->oflags |= VPO_CLEANCHK; 1048 1049 /* 1050 * maxf will end up being the actual number of pages 1051 * we wrote out contiguously, non-inclusive of the 1052 * first page. We do not count look-behind pages. 1053 */ 1054 if (i >= maxb + 1 && (maxf > i - maxb - 1)) 1055 maxf = i - maxb - 1; 1056 } 1057 } 1058 return(maxf + 1); 1059 } 1060 1061 /* 1062 * Note that there is absolutely no sense in writing out 1063 * anonymous objects, so we track down the vnode object 1064 * to write out. 1065 * We invalidate (remove) all pages from the address space 1066 * for semantic correctness. 1067 * 1068 * Note: certain anonymous maps, such as MAP_NOSYNC maps, 1069 * may start out with a NULL object. 1070 */ 1071 void 1072 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, 1073 boolean_t syncio, boolean_t invalidate) 1074 { 1075 vm_object_t backing_object; 1076 struct vnode *vp; 1077 struct mount *mp; 1078 int flags; 1079 1080 if (object == NULL) 1081 return; 1082 VM_OBJECT_LOCK(object); 1083 while ((backing_object = object->backing_object) != NULL) { 1084 VM_OBJECT_LOCK(backing_object); 1085 offset += object->backing_object_offset; 1086 VM_OBJECT_UNLOCK(object); 1087 object = backing_object; 1088 if (object->size < OFF_TO_IDX(offset + size)) 1089 size = IDX_TO_OFF(object->size) - offset; 1090 } 1091 /* 1092 * Flush pages if writing is allowed, invalidate them 1093 * if invalidation requested. Pages undergoing I/O 1094 * will be ignored by vm_object_page_remove(). 1095 * 1096 * We cannot lock the vnode and then wait for paging 1097 * to complete without deadlocking against vm_fault. 1098 * Instead we simply call vm_object_page_remove() and 1099 * allow it to block internally on a page-by-page 1100 * basis when it encounters pages undergoing async 1101 * I/O. 1102 */ 1103 if (object->type == OBJT_VNODE && 1104 (object->flags & OBJ_MIGHTBEDIRTY) != 0) { 1105 int vfslocked; 1106 vp = object->handle; 1107 VM_OBJECT_UNLOCK(object); 1108 (void) vn_start_write(vp, &mp, V_WAIT); 1109 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1110 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1111 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 1112 flags |= invalidate ? OBJPC_INVAL : 0; 1113 VM_OBJECT_LOCK(object); 1114 vm_object_page_clean(object, 1115 OFF_TO_IDX(offset), 1116 OFF_TO_IDX(offset + size + PAGE_MASK), 1117 flags); 1118 VM_OBJECT_UNLOCK(object); 1119 VOP_UNLOCK(vp, 0); 1120 VFS_UNLOCK_GIANT(vfslocked); 1121 vn_finished_write(mp); 1122 VM_OBJECT_LOCK(object); 1123 } 1124 if ((object->type == OBJT_VNODE || 1125 object->type == OBJT_DEVICE) && invalidate) { 1126 boolean_t purge; 1127 purge = old_msync || (object->type == OBJT_DEVICE); 1128 vm_object_page_remove(object, 1129 OFF_TO_IDX(offset), 1130 OFF_TO_IDX(offset + size + PAGE_MASK), 1131 purge ? FALSE : TRUE); 1132 } 1133 VM_OBJECT_UNLOCK(object); 1134 } 1135 1136 /* 1137 * vm_object_madvise: 1138 * 1139 * Implements the madvise function at the object/page level. 1140 * 1141 * MADV_WILLNEED (any object) 1142 * 1143 * Activate the specified pages if they are resident. 1144 * 1145 * MADV_DONTNEED (any object) 1146 * 1147 * Deactivate the specified pages if they are resident. 1148 * 1149 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, 1150 * OBJ_ONEMAPPING only) 1151 * 1152 * Deactivate and clean the specified pages if they are 1153 * resident. This permits the process to reuse the pages 1154 * without faulting or the kernel to reclaim the pages 1155 * without I/O. 1156 */ 1157 void 1158 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) 1159 { 1160 vm_pindex_t end, tpindex; 1161 vm_object_t backing_object, tobject; 1162 vm_page_t m; 1163 1164 if (object == NULL) 1165 return; 1166 VM_OBJECT_LOCK(object); 1167 end = pindex + count; 1168 /* 1169 * Locate and adjust resident pages 1170 */ 1171 for (; pindex < end; pindex += 1) { 1172 relookup: 1173 tobject = object; 1174 tpindex = pindex; 1175 shadowlookup: 1176 /* 1177 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages 1178 * and those pages must be OBJ_ONEMAPPING. 1179 */ 1180 if (advise == MADV_FREE) { 1181 if ((tobject->type != OBJT_DEFAULT && 1182 tobject->type != OBJT_SWAP) || 1183 (tobject->flags & OBJ_ONEMAPPING) == 0) { 1184 goto unlock_tobject; 1185 } 1186 } else if (tobject->type == OBJT_PHYS) 1187 goto unlock_tobject; 1188 m = vm_page_lookup(tobject, tpindex); 1189 if (m == NULL && advise == MADV_WILLNEED) { 1190 /* 1191 * If the page is cached, reactivate it. 1192 */ 1193 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | 1194 VM_ALLOC_NOBUSY); 1195 } 1196 if (m == NULL) { 1197 /* 1198 * There may be swap even if there is no backing page 1199 */ 1200 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1201 swap_pager_freespace(tobject, tpindex, 1); 1202 /* 1203 * next object 1204 */ 1205 backing_object = tobject->backing_object; 1206 if (backing_object == NULL) 1207 goto unlock_tobject; 1208 VM_OBJECT_LOCK(backing_object); 1209 tpindex += OFF_TO_IDX(tobject->backing_object_offset); 1210 if (tobject != object) 1211 VM_OBJECT_UNLOCK(tobject); 1212 tobject = backing_object; 1213 goto shadowlookup; 1214 } else if (m->valid != VM_PAGE_BITS_ALL) 1215 goto unlock_tobject; 1216 /* 1217 * If the page is not in a normal state, skip it. 1218 */ 1219 vm_page_lock(m); 1220 vm_page_lock_queues(); 1221 if (m->hold_count != 0 || m->wire_count != 0) { 1222 vm_page_unlock_queues(); 1223 vm_page_unlock(m); 1224 goto unlock_tobject; 1225 } 1226 if ((m->oflags & VPO_BUSY) || m->busy) { 1227 if (advise == MADV_WILLNEED) 1228 /* 1229 * Reference the page before unlocking and 1230 * sleeping so that the page daemon is less 1231 * likely to reclaim it. 1232 */ 1233 vm_page_flag_set(m, PG_REFERENCED); 1234 vm_page_unlock_queues(); 1235 vm_page_unlock(m); 1236 if (object != tobject) 1237 VM_OBJECT_UNLOCK(object); 1238 m->oflags |= VPO_WANTED; 1239 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 1240 0); 1241 VM_OBJECT_LOCK(object); 1242 goto relookup; 1243 } 1244 if (advise == MADV_WILLNEED) { 1245 vm_page_activate(m); 1246 } else if (advise == MADV_DONTNEED) { 1247 vm_page_dontneed(m); 1248 } else if (advise == MADV_FREE) { 1249 /* 1250 * Mark the page clean. This will allow the page 1251 * to be freed up by the system. However, such pages 1252 * are often reused quickly by malloc()/free() 1253 * so we do not do anything that would cause 1254 * a page fault if we can help it. 1255 * 1256 * Specifically, we do not try to actually free 1257 * the page now nor do we try to put it in the 1258 * cache (which would cause a page fault on reuse). 1259 * 1260 * But we do make the page is freeable as we 1261 * can without actually taking the step of unmapping 1262 * it. 1263 */ 1264 pmap_clear_modify(m); 1265 m->dirty = 0; 1266 m->act_count = 0; 1267 vm_page_dontneed(m); 1268 } 1269 vm_page_unlock_queues(); 1270 vm_page_unlock(m); 1271 if (advise == MADV_FREE && tobject->type == OBJT_SWAP) 1272 swap_pager_freespace(tobject, tpindex, 1); 1273 unlock_tobject: 1274 if (tobject != object) 1275 VM_OBJECT_UNLOCK(tobject); 1276 } 1277 VM_OBJECT_UNLOCK(object); 1278 } 1279 1280 /* 1281 * vm_object_shadow: 1282 * 1283 * Create a new object which is backed by the 1284 * specified existing object range. The source 1285 * object reference is deallocated. 1286 * 1287 * The new object and offset into that object 1288 * are returned in the source parameters. 1289 */ 1290 void 1291 vm_object_shadow( 1292 vm_object_t *object, /* IN/OUT */ 1293 vm_ooffset_t *offset, /* IN/OUT */ 1294 vm_size_t length) 1295 { 1296 vm_object_t source; 1297 vm_object_t result; 1298 1299 source = *object; 1300 1301 /* 1302 * Don't create the new object if the old object isn't shared. 1303 */ 1304 if (source != NULL) { 1305 VM_OBJECT_LOCK(source); 1306 if (source->ref_count == 1 && 1307 source->handle == NULL && 1308 (source->type == OBJT_DEFAULT || 1309 source->type == OBJT_SWAP)) { 1310 VM_OBJECT_UNLOCK(source); 1311 return; 1312 } 1313 VM_OBJECT_UNLOCK(source); 1314 } 1315 1316 /* 1317 * Allocate a new object with the given length. 1318 */ 1319 result = vm_object_allocate(OBJT_DEFAULT, length); 1320 1321 /* 1322 * The new object shadows the source object, adding a reference to it. 1323 * Our caller changes his reference to point to the new object, 1324 * removing a reference to the source object. Net result: no change 1325 * of reference count. 1326 * 1327 * Try to optimize the result object's page color when shadowing 1328 * in order to maintain page coloring consistency in the combined 1329 * shadowed object. 1330 */ 1331 result->backing_object = source; 1332 /* 1333 * Store the offset into the source object, and fix up the offset into 1334 * the new object. 1335 */ 1336 result->backing_object_offset = *offset; 1337 if (source != NULL) { 1338 VM_OBJECT_LOCK(source); 1339 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); 1340 source->shadow_count++; 1341 source->generation++; 1342 #if VM_NRESERVLEVEL > 0 1343 result->flags |= source->flags & OBJ_COLORED; 1344 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & 1345 ((1 << (VM_NFREEORDER - 1)) - 1); 1346 #endif 1347 VM_OBJECT_UNLOCK(source); 1348 } 1349 1350 1351 /* 1352 * Return the new things 1353 */ 1354 *offset = 0; 1355 *object = result; 1356 } 1357 1358 /* 1359 * vm_object_split: 1360 * 1361 * Split the pages in a map entry into a new object. This affords 1362 * easier removal of unused pages, and keeps object inheritance from 1363 * being a negative impact on memory usage. 1364 */ 1365 void 1366 vm_object_split(vm_map_entry_t entry) 1367 { 1368 vm_page_t m, m_next; 1369 vm_object_t orig_object, new_object, source; 1370 vm_pindex_t idx, offidxstart; 1371 vm_size_t size; 1372 1373 orig_object = entry->object.vm_object; 1374 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 1375 return; 1376 if (orig_object->ref_count <= 1) 1377 return; 1378 VM_OBJECT_UNLOCK(orig_object); 1379 1380 offidxstart = OFF_TO_IDX(entry->offset); 1381 size = atop(entry->end - entry->start); 1382 1383 /* 1384 * If swap_pager_copy() is later called, it will convert new_object 1385 * into a swap object. 1386 */ 1387 new_object = vm_object_allocate(OBJT_DEFAULT, size); 1388 1389 /* 1390 * At this point, the new object is still private, so the order in 1391 * which the original and new objects are locked does not matter. 1392 */ 1393 VM_OBJECT_LOCK(new_object); 1394 VM_OBJECT_LOCK(orig_object); 1395 source = orig_object->backing_object; 1396 if (source != NULL) { 1397 VM_OBJECT_LOCK(source); 1398 if ((source->flags & OBJ_DEAD) != 0) { 1399 VM_OBJECT_UNLOCK(source); 1400 VM_OBJECT_UNLOCK(orig_object); 1401 VM_OBJECT_UNLOCK(new_object); 1402 vm_object_deallocate(new_object); 1403 VM_OBJECT_LOCK(orig_object); 1404 return; 1405 } 1406 LIST_INSERT_HEAD(&source->shadow_head, 1407 new_object, shadow_list); 1408 source->shadow_count++; 1409 source->generation++; 1410 vm_object_reference_locked(source); /* for new_object */ 1411 vm_object_clear_flag(source, OBJ_ONEMAPPING); 1412 VM_OBJECT_UNLOCK(source); 1413 new_object->backing_object_offset = 1414 orig_object->backing_object_offset + entry->offset; 1415 new_object->backing_object = source; 1416 } 1417 if (orig_object->uip != NULL) { 1418 new_object->uip = orig_object->uip; 1419 uihold(orig_object->uip); 1420 new_object->charge = ptoa(size); 1421 KASSERT(orig_object->charge >= ptoa(size), 1422 ("orig_object->charge < 0")); 1423 orig_object->charge -= ptoa(size); 1424 } 1425 retry: 1426 if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) { 1427 if (m->pindex < offidxstart) { 1428 m = vm_page_splay(offidxstart, orig_object->root); 1429 if ((orig_object->root = m)->pindex < offidxstart) 1430 m = TAILQ_NEXT(m, listq); 1431 } 1432 } 1433 for (; m != NULL && (idx = m->pindex - offidxstart) < size; 1434 m = m_next) { 1435 m_next = TAILQ_NEXT(m, listq); 1436 1437 /* 1438 * We must wait for pending I/O to complete before we can 1439 * rename the page. 1440 * 1441 * We do not have to VM_PROT_NONE the page as mappings should 1442 * not be changed by this operation. 1443 */ 1444 if ((m->oflags & VPO_BUSY) || m->busy) { 1445 VM_OBJECT_UNLOCK(new_object); 1446 m->oflags |= VPO_WANTED; 1447 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0); 1448 VM_OBJECT_LOCK(new_object); 1449 goto retry; 1450 } 1451 vm_page_lock(m); 1452 vm_page_rename(m, new_object, idx); 1453 vm_page_unlock(m); 1454 /* page automatically made dirty by rename and cache handled */ 1455 vm_page_busy(m); 1456 } 1457 if (orig_object->type == OBJT_SWAP) { 1458 /* 1459 * swap_pager_copy() can sleep, in which case the orig_object's 1460 * and new_object's locks are released and reacquired. 1461 */ 1462 swap_pager_copy(orig_object, new_object, offidxstart, 0); 1463 1464 /* 1465 * Transfer any cached pages from orig_object to new_object. 1466 */ 1467 if (__predict_false(orig_object->cache != NULL)) 1468 vm_page_cache_transfer(orig_object, offidxstart, 1469 new_object); 1470 } 1471 VM_OBJECT_UNLOCK(orig_object); 1472 TAILQ_FOREACH(m, &new_object->memq, listq) 1473 vm_page_wakeup(m); 1474 VM_OBJECT_UNLOCK(new_object); 1475 entry->object.vm_object = new_object; 1476 entry->offset = 0LL; 1477 vm_object_deallocate(orig_object); 1478 VM_OBJECT_LOCK(new_object); 1479 } 1480 1481 #define OBSC_TEST_ALL_SHADOWED 0x0001 1482 #define OBSC_COLLAPSE_NOWAIT 0x0002 1483 #define OBSC_COLLAPSE_WAIT 0x0004 1484 1485 static int 1486 vm_object_backing_scan(vm_object_t object, int op) 1487 { 1488 int r = 1; 1489 vm_page_t p; 1490 vm_object_t backing_object; 1491 vm_pindex_t backing_offset_index; 1492 1493 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1494 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); 1495 1496 backing_object = object->backing_object; 1497 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 1498 1499 /* 1500 * Initial conditions 1501 */ 1502 if (op & OBSC_TEST_ALL_SHADOWED) { 1503 /* 1504 * We do not want to have to test for the existence of cache 1505 * or swap pages in the backing object. XXX but with the 1506 * new swapper this would be pretty easy to do. 1507 * 1508 * XXX what about anonymous MAP_SHARED memory that hasn't 1509 * been ZFOD faulted yet? If we do not test for this, the 1510 * shadow test may succeed! XXX 1511 */ 1512 if (backing_object->type != OBJT_DEFAULT) { 1513 return (0); 1514 } 1515 } 1516 if (op & OBSC_COLLAPSE_WAIT) { 1517 vm_object_set_flag(backing_object, OBJ_DEAD); 1518 } 1519 1520 /* 1521 * Our scan 1522 */ 1523 p = TAILQ_FIRST(&backing_object->memq); 1524 while (p) { 1525 vm_page_t next = TAILQ_NEXT(p, listq); 1526 vm_pindex_t new_pindex = p->pindex - backing_offset_index; 1527 1528 if (op & OBSC_TEST_ALL_SHADOWED) { 1529 vm_page_t pp; 1530 1531 /* 1532 * Ignore pages outside the parent object's range 1533 * and outside the parent object's mapping of the 1534 * backing object. 1535 * 1536 * note that we do not busy the backing object's 1537 * page. 1538 */ 1539 if ( 1540 p->pindex < backing_offset_index || 1541 new_pindex >= object->size 1542 ) { 1543 p = next; 1544 continue; 1545 } 1546 1547 /* 1548 * See if the parent has the page or if the parent's 1549 * object pager has the page. If the parent has the 1550 * page but the page is not valid, the parent's 1551 * object pager must have the page. 1552 * 1553 * If this fails, the parent does not completely shadow 1554 * the object and we might as well give up now. 1555 */ 1556 1557 pp = vm_page_lookup(object, new_pindex); 1558 if ( 1559 (pp == NULL || pp->valid == 0) && 1560 !vm_pager_has_page(object, new_pindex, NULL, NULL) 1561 ) { 1562 r = 0; 1563 break; 1564 } 1565 } 1566 1567 /* 1568 * Check for busy page 1569 */ 1570 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { 1571 vm_page_t pp; 1572 1573 if (op & OBSC_COLLAPSE_NOWAIT) { 1574 if ((p->oflags & VPO_BUSY) || 1575 !p->valid || 1576 p->busy) { 1577 p = next; 1578 continue; 1579 } 1580 } else if (op & OBSC_COLLAPSE_WAIT) { 1581 if ((p->oflags & VPO_BUSY) || p->busy) { 1582 VM_OBJECT_UNLOCK(object); 1583 p->oflags |= VPO_WANTED; 1584 msleep(p, VM_OBJECT_MTX(backing_object), 1585 PDROP | PVM, "vmocol", 0); 1586 VM_OBJECT_LOCK(object); 1587 VM_OBJECT_LOCK(backing_object); 1588 /* 1589 * If we slept, anything could have 1590 * happened. Since the object is 1591 * marked dead, the backing offset 1592 * should not have changed so we 1593 * just restart our scan. 1594 */ 1595 p = TAILQ_FIRST(&backing_object->memq); 1596 continue; 1597 } 1598 } 1599 1600 KASSERT( 1601 p->object == backing_object, 1602 ("vm_object_backing_scan: object mismatch") 1603 ); 1604 1605 /* 1606 * Destroy any associated swap 1607 */ 1608 if (backing_object->type == OBJT_SWAP) { 1609 swap_pager_freespace( 1610 backing_object, 1611 p->pindex, 1612 1 1613 ); 1614 } 1615 1616 if ( 1617 p->pindex < backing_offset_index || 1618 new_pindex >= object->size 1619 ) { 1620 /* 1621 * Page is out of the parent object's range, we 1622 * can simply destroy it. 1623 */ 1624 vm_page_lock(p); 1625 KASSERT(!pmap_page_is_mapped(p), 1626 ("freeing mapped page %p", p)); 1627 if (p->wire_count == 0) 1628 vm_page_free(p); 1629 else 1630 vm_page_remove(p); 1631 vm_page_unlock(p); 1632 p = next; 1633 continue; 1634 } 1635 1636 pp = vm_page_lookup(object, new_pindex); 1637 if ( 1638 pp != NULL || 1639 vm_pager_has_page(object, new_pindex, NULL, NULL) 1640 ) { 1641 /* 1642 * page already exists in parent OR swap exists 1643 * for this location in the parent. Destroy 1644 * the original page from the backing object. 1645 * 1646 * Leave the parent's page alone 1647 */ 1648 vm_page_lock(p); 1649 KASSERT(!pmap_page_is_mapped(p), 1650 ("freeing mapped page %p", p)); 1651 if (p->wire_count == 0) 1652 vm_page_free(p); 1653 else 1654 vm_page_remove(p); 1655 vm_page_unlock(p); 1656 p = next; 1657 continue; 1658 } 1659 1660 #if VM_NRESERVLEVEL > 0 1661 /* 1662 * Rename the reservation. 1663 */ 1664 vm_reserv_rename(p, object, backing_object, 1665 backing_offset_index); 1666 #endif 1667 1668 /* 1669 * Page does not exist in parent, rename the 1670 * page from the backing object to the main object. 1671 * 1672 * If the page was mapped to a process, it can remain 1673 * mapped through the rename. 1674 */ 1675 vm_page_lock(p); 1676 vm_page_rename(p, object, new_pindex); 1677 vm_page_unlock(p); 1678 /* page automatically made dirty by rename */ 1679 } 1680 p = next; 1681 } 1682 return (r); 1683 } 1684 1685 1686 /* 1687 * this version of collapse allows the operation to occur earlier and 1688 * when paging_in_progress is true for an object... This is not a complete 1689 * operation, but should plug 99.9% of the rest of the leaks. 1690 */ 1691 static void 1692 vm_object_qcollapse(vm_object_t object) 1693 { 1694 vm_object_t backing_object = object->backing_object; 1695 1696 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1697 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); 1698 1699 if (backing_object->ref_count != 1) 1700 return; 1701 1702 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); 1703 } 1704 1705 /* 1706 * vm_object_collapse: 1707 * 1708 * Collapse an object with the object backing it. 1709 * Pages in the backing object are moved into the 1710 * parent, and the backing object is deallocated. 1711 */ 1712 void 1713 vm_object_collapse(vm_object_t object) 1714 { 1715 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1716 1717 while (TRUE) { 1718 vm_object_t backing_object; 1719 1720 /* 1721 * Verify that the conditions are right for collapse: 1722 * 1723 * The object exists and the backing object exists. 1724 */ 1725 if ((backing_object = object->backing_object) == NULL) 1726 break; 1727 1728 /* 1729 * we check the backing object first, because it is most likely 1730 * not collapsable. 1731 */ 1732 VM_OBJECT_LOCK(backing_object); 1733 if (backing_object->handle != NULL || 1734 (backing_object->type != OBJT_DEFAULT && 1735 backing_object->type != OBJT_SWAP) || 1736 (backing_object->flags & OBJ_DEAD) || 1737 object->handle != NULL || 1738 (object->type != OBJT_DEFAULT && 1739 object->type != OBJT_SWAP) || 1740 (object->flags & OBJ_DEAD)) { 1741 VM_OBJECT_UNLOCK(backing_object); 1742 break; 1743 } 1744 1745 if ( 1746 object->paging_in_progress != 0 || 1747 backing_object->paging_in_progress != 0 1748 ) { 1749 vm_object_qcollapse(object); 1750 VM_OBJECT_UNLOCK(backing_object); 1751 break; 1752 } 1753 /* 1754 * We know that we can either collapse the backing object (if 1755 * the parent is the only reference to it) or (perhaps) have 1756 * the parent bypass the object if the parent happens to shadow 1757 * all the resident pages in the entire backing object. 1758 * 1759 * This is ignoring pager-backed pages such as swap pages. 1760 * vm_object_backing_scan fails the shadowing test in this 1761 * case. 1762 */ 1763 if (backing_object->ref_count == 1) { 1764 /* 1765 * If there is exactly one reference to the backing 1766 * object, we can collapse it into the parent. 1767 */ 1768 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); 1769 1770 #if VM_NRESERVLEVEL > 0 1771 /* 1772 * Break any reservations from backing_object. 1773 */ 1774 if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) 1775 vm_reserv_break_all(backing_object); 1776 #endif 1777 1778 /* 1779 * Move the pager from backing_object to object. 1780 */ 1781 if (backing_object->type == OBJT_SWAP) { 1782 /* 1783 * swap_pager_copy() can sleep, in which case 1784 * the backing_object's and object's locks are 1785 * released and reacquired. 1786 */ 1787 swap_pager_copy( 1788 backing_object, 1789 object, 1790 OFF_TO_IDX(object->backing_object_offset), TRUE); 1791 1792 /* 1793 * Free any cached pages from backing_object. 1794 */ 1795 if (__predict_false(backing_object->cache != NULL)) 1796 vm_page_cache_free(backing_object, 0, 0); 1797 } 1798 /* 1799 * Object now shadows whatever backing_object did. 1800 * Note that the reference to 1801 * backing_object->backing_object moves from within 1802 * backing_object to within object. 1803 */ 1804 LIST_REMOVE(object, shadow_list); 1805 backing_object->shadow_count--; 1806 backing_object->generation++; 1807 if (backing_object->backing_object) { 1808 VM_OBJECT_LOCK(backing_object->backing_object); 1809 LIST_REMOVE(backing_object, shadow_list); 1810 LIST_INSERT_HEAD( 1811 &backing_object->backing_object->shadow_head, 1812 object, shadow_list); 1813 /* 1814 * The shadow_count has not changed. 1815 */ 1816 backing_object->backing_object->generation++; 1817 VM_OBJECT_UNLOCK(backing_object->backing_object); 1818 } 1819 object->backing_object = backing_object->backing_object; 1820 object->backing_object_offset += 1821 backing_object->backing_object_offset; 1822 1823 /* 1824 * Discard backing_object. 1825 * 1826 * Since the backing object has no pages, no pager left, 1827 * and no object references within it, all that is 1828 * necessary is to dispose of it. 1829 */ 1830 KASSERT(backing_object->ref_count == 1, ( 1831 "backing_object %p was somehow re-referenced during collapse!", 1832 backing_object)); 1833 VM_OBJECT_UNLOCK(backing_object); 1834 vm_object_destroy(backing_object); 1835 1836 object_collapses++; 1837 } else { 1838 vm_object_t new_backing_object; 1839 1840 /* 1841 * If we do not entirely shadow the backing object, 1842 * there is nothing we can do so we give up. 1843 */ 1844 if (object->resident_page_count != object->size && 1845 vm_object_backing_scan(object, 1846 OBSC_TEST_ALL_SHADOWED) == 0) { 1847 VM_OBJECT_UNLOCK(backing_object); 1848 break; 1849 } 1850 1851 /* 1852 * Make the parent shadow the next object in the 1853 * chain. Deallocating backing_object will not remove 1854 * it, since its reference count is at least 2. 1855 */ 1856 LIST_REMOVE(object, shadow_list); 1857 backing_object->shadow_count--; 1858 backing_object->generation++; 1859 1860 new_backing_object = backing_object->backing_object; 1861 if ((object->backing_object = new_backing_object) != NULL) { 1862 VM_OBJECT_LOCK(new_backing_object); 1863 LIST_INSERT_HEAD( 1864 &new_backing_object->shadow_head, 1865 object, 1866 shadow_list 1867 ); 1868 new_backing_object->shadow_count++; 1869 new_backing_object->generation++; 1870 vm_object_reference_locked(new_backing_object); 1871 VM_OBJECT_UNLOCK(new_backing_object); 1872 object->backing_object_offset += 1873 backing_object->backing_object_offset; 1874 } 1875 1876 /* 1877 * Drop the reference count on backing_object. Since 1878 * its ref_count was at least 2, it will not vanish. 1879 */ 1880 backing_object->ref_count--; 1881 VM_OBJECT_UNLOCK(backing_object); 1882 object_bypasses++; 1883 } 1884 1885 /* 1886 * Try again with this object's new backing object. 1887 */ 1888 } 1889 } 1890 1891 /* 1892 * vm_object_page_remove: 1893 * 1894 * For the given object, either frees or invalidates each of the 1895 * specified pages. In general, a page is freed. However, if a 1896 * page is wired for any reason other than the existence of a 1897 * managed, wired mapping, then it may be invalidated but not 1898 * removed from the object. Pages are specified by the given 1899 * range ["start", "end") and Boolean "clean_only". As a 1900 * special case, if "end" is zero, then the range extends from 1901 * "start" to the end of the object. If "clean_only" is TRUE, 1902 * then only the non-dirty pages within the specified range are 1903 * affected. 1904 * 1905 * In general, this operation should only be performed on objects 1906 * that contain managed pages. There are two exceptions. First, 1907 * it may be performed on the kernel and kmem objects. Second, 1908 * it may be used by msync(..., MS_INVALIDATE) to invalidate 1909 * device-backed pages. In both of these cases, "clean_only" 1910 * must be FALSE. 1911 * 1912 * The object must be locked. 1913 */ 1914 void 1915 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, 1916 boolean_t clean_only) 1917 { 1918 vm_page_t p, next; 1919 int wirings; 1920 1921 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1922 if (object->resident_page_count == 0) 1923 goto skipmemq; 1924 1925 /* 1926 * Since physically-backed objects do not use managed pages, we can't 1927 * remove pages from the object (we must instead remove the page 1928 * references, and then destroy the object). 1929 */ 1930 KASSERT(object->type != OBJT_PHYS || object == kernel_object || 1931 object == kmem_object, 1932 ("attempt to remove pages from a physical object")); 1933 1934 vm_object_pip_add(object, 1); 1935 again: 1936 if ((p = TAILQ_FIRST(&object->memq)) != NULL) { 1937 if (p->pindex < start) { 1938 p = vm_page_splay(start, object->root); 1939 if ((object->root = p)->pindex < start) 1940 p = TAILQ_NEXT(p, listq); 1941 } 1942 } 1943 1944 /* 1945 * Assert: the variable p is either (1) the page with the 1946 * least pindex greater than or equal to the parameter pindex 1947 * or (2) NULL. 1948 */ 1949 for (; 1950 p != NULL && (p->pindex < end || end == 0); 1951 p = next) { 1952 next = TAILQ_NEXT(p, listq); 1953 1954 /* 1955 * If the page is wired for any reason besides the 1956 * existence of managed, wired mappings, then it cannot 1957 * be freed. For example, fictitious pages, which 1958 * represent device memory, are inherently wired and 1959 * cannot be freed. They can, however, be invalidated 1960 * if "clean_only" is FALSE. 1961 */ 1962 vm_page_lock(p); 1963 if ((wirings = p->wire_count) != 0 && 1964 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) { 1965 /* Fictitious pages do not have managed mappings. */ 1966 if ((p->flags & PG_FICTITIOUS) == 0) 1967 pmap_remove_all(p); 1968 /* Account for removal of managed, wired mappings. */ 1969 p->wire_count -= wirings; 1970 if (!clean_only) { 1971 p->valid = 0; 1972 vm_page_undirty(p); 1973 } 1974 vm_page_unlock(p); 1975 continue; 1976 } 1977 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) 1978 goto again; 1979 KASSERT((p->flags & PG_FICTITIOUS) == 0, 1980 ("vm_object_page_remove: page %p is fictitious", p)); 1981 if (clean_only && p->valid) { 1982 pmap_remove_write(p); 1983 if (p->dirty) { 1984 vm_page_unlock(p); 1985 continue; 1986 } 1987 } 1988 pmap_remove_all(p); 1989 /* Account for removal of managed, wired mappings. */ 1990 if (wirings != 0) 1991 p->wire_count -= wirings; 1992 vm_page_free(p); 1993 vm_page_unlock(p); 1994 } 1995 vm_object_pip_wakeup(object); 1996 skipmemq: 1997 if (__predict_false(object->cache != NULL)) 1998 vm_page_cache_free(object, start, end); 1999 } 2000 2001 /* 2002 * Populate the specified range of the object with valid pages. Returns 2003 * TRUE if the range is successfully populated and FALSE otherwise. 2004 * 2005 * Note: This function should be optimized to pass a larger array of 2006 * pages to vm_pager_get_pages() before it is applied to a non- 2007 * OBJT_DEVICE object. 2008 * 2009 * The object must be locked. 2010 */ 2011 boolean_t 2012 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 2013 { 2014 vm_page_t m, ma[1]; 2015 vm_pindex_t pindex; 2016 int rv; 2017 2018 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2019 for (pindex = start; pindex < end; pindex++) { 2020 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | 2021 VM_ALLOC_RETRY); 2022 if (m->valid != VM_PAGE_BITS_ALL) { 2023 ma[0] = m; 2024 rv = vm_pager_get_pages(object, ma, 1, 0); 2025 m = vm_page_lookup(object, pindex); 2026 if (m == NULL) 2027 break; 2028 if (rv != VM_PAGER_OK) { 2029 vm_page_lock(m); 2030 vm_page_free(m); 2031 vm_page_unlock(m); 2032 break; 2033 } 2034 } 2035 /* 2036 * Keep "m" busy because a subsequent iteration may unlock 2037 * the object. 2038 */ 2039 } 2040 if (pindex > start) { 2041 m = vm_page_lookup(object, start); 2042 while (m != NULL && m->pindex < pindex) { 2043 vm_page_wakeup(m); 2044 m = TAILQ_NEXT(m, listq); 2045 } 2046 } 2047 return (pindex == end); 2048 } 2049 2050 /* 2051 * Routine: vm_object_coalesce 2052 * Function: Coalesces two objects backing up adjoining 2053 * regions of memory into a single object. 2054 * 2055 * returns TRUE if objects were combined. 2056 * 2057 * NOTE: Only works at the moment if the second object is NULL - 2058 * if it's not, which object do we lock first? 2059 * 2060 * Parameters: 2061 * prev_object First object to coalesce 2062 * prev_offset Offset into prev_object 2063 * prev_size Size of reference to prev_object 2064 * next_size Size of reference to the second object 2065 * reserved Indicator that extension region has 2066 * swap accounted for 2067 * 2068 * Conditions: 2069 * The object must *not* be locked. 2070 */ 2071 boolean_t 2072 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, 2073 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) 2074 { 2075 vm_pindex_t next_pindex; 2076 2077 if (prev_object == NULL) 2078 return (TRUE); 2079 VM_OBJECT_LOCK(prev_object); 2080 if (prev_object->type != OBJT_DEFAULT && 2081 prev_object->type != OBJT_SWAP) { 2082 VM_OBJECT_UNLOCK(prev_object); 2083 return (FALSE); 2084 } 2085 2086 /* 2087 * Try to collapse the object first 2088 */ 2089 vm_object_collapse(prev_object); 2090 2091 /* 2092 * Can't coalesce if: . more than one reference . paged out . shadows 2093 * another object . has a copy elsewhere (any of which mean that the 2094 * pages not mapped to prev_entry may be in use anyway) 2095 */ 2096 if (prev_object->backing_object != NULL) { 2097 VM_OBJECT_UNLOCK(prev_object); 2098 return (FALSE); 2099 } 2100 2101 prev_size >>= PAGE_SHIFT; 2102 next_size >>= PAGE_SHIFT; 2103 next_pindex = OFF_TO_IDX(prev_offset) + prev_size; 2104 2105 if ((prev_object->ref_count > 1) && 2106 (prev_object->size != next_pindex)) { 2107 VM_OBJECT_UNLOCK(prev_object); 2108 return (FALSE); 2109 } 2110 2111 /* 2112 * Account for the charge. 2113 */ 2114 if (prev_object->uip != NULL) { 2115 2116 /* 2117 * If prev_object was charged, then this mapping, 2118 * althought not charged now, may become writable 2119 * later. Non-NULL uip in the object would prevent 2120 * swap reservation during enabling of the write 2121 * access, so reserve swap now. Failed reservation 2122 * cause allocation of the separate object for the map 2123 * entry, and swap reservation for this entry is 2124 * managed in appropriate time. 2125 */ 2126 if (!reserved && !swap_reserve_by_uid(ptoa(next_size), 2127 prev_object->uip)) { 2128 return (FALSE); 2129 } 2130 prev_object->charge += ptoa(next_size); 2131 } 2132 2133 /* 2134 * Remove any pages that may still be in the object from a previous 2135 * deallocation. 2136 */ 2137 if (next_pindex < prev_object->size) { 2138 vm_object_page_remove(prev_object, 2139 next_pindex, 2140 next_pindex + next_size, FALSE); 2141 if (prev_object->type == OBJT_SWAP) 2142 swap_pager_freespace(prev_object, 2143 next_pindex, next_size); 2144 #if 0 2145 if (prev_object->uip != NULL) { 2146 KASSERT(prev_object->charge >= 2147 ptoa(prev_object->size - next_pindex), 2148 ("object %p overcharged 1 %jx %jx", prev_object, 2149 (uintmax_t)next_pindex, (uintmax_t)next_size)); 2150 prev_object->charge -= ptoa(prev_object->size - 2151 next_pindex); 2152 } 2153 #endif 2154 } 2155 2156 /* 2157 * Extend the object if necessary. 2158 */ 2159 if (next_pindex + next_size > prev_object->size) 2160 prev_object->size = next_pindex + next_size; 2161 2162 VM_OBJECT_UNLOCK(prev_object); 2163 return (TRUE); 2164 } 2165 2166 void 2167 vm_object_set_writeable_dirty(vm_object_t object) 2168 { 2169 2170 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2171 if (object->type != OBJT_VNODE || 2172 (object->flags & OBJ_MIGHTBEDIRTY) != 0) 2173 return; 2174 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); 2175 } 2176 2177 #include "opt_ddb.h" 2178 #ifdef DDB 2179 #include <sys/kernel.h> 2180 2181 #include <sys/cons.h> 2182 2183 #include <ddb/ddb.h> 2184 2185 static int 2186 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) 2187 { 2188 vm_map_t tmpm; 2189 vm_map_entry_t tmpe; 2190 vm_object_t obj; 2191 int entcount; 2192 2193 if (map == 0) 2194 return 0; 2195 2196 if (entry == 0) { 2197 tmpe = map->header.next; 2198 entcount = map->nentries; 2199 while (entcount-- && (tmpe != &map->header)) { 2200 if (_vm_object_in_map(map, object, tmpe)) { 2201 return 1; 2202 } 2203 tmpe = tmpe->next; 2204 } 2205 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 2206 tmpm = entry->object.sub_map; 2207 tmpe = tmpm->header.next; 2208 entcount = tmpm->nentries; 2209 while (entcount-- && tmpe != &tmpm->header) { 2210 if (_vm_object_in_map(tmpm, object, tmpe)) { 2211 return 1; 2212 } 2213 tmpe = tmpe->next; 2214 } 2215 } else if ((obj = entry->object.vm_object) != NULL) { 2216 for (; obj; obj = obj->backing_object) 2217 if (obj == object) { 2218 return 1; 2219 } 2220 } 2221 return 0; 2222 } 2223 2224 static int 2225 vm_object_in_map(vm_object_t object) 2226 { 2227 struct proc *p; 2228 2229 /* sx_slock(&allproc_lock); */ 2230 FOREACH_PROC_IN_SYSTEM(p) { 2231 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) 2232 continue; 2233 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { 2234 /* sx_sunlock(&allproc_lock); */ 2235 return 1; 2236 } 2237 } 2238 /* sx_sunlock(&allproc_lock); */ 2239 if (_vm_object_in_map(kernel_map, object, 0)) 2240 return 1; 2241 if (_vm_object_in_map(kmem_map, object, 0)) 2242 return 1; 2243 if (_vm_object_in_map(pager_map, object, 0)) 2244 return 1; 2245 if (_vm_object_in_map(buffer_map, object, 0)) 2246 return 1; 2247 return 0; 2248 } 2249 2250 DB_SHOW_COMMAND(vmochk, vm_object_check) 2251 { 2252 vm_object_t object; 2253 2254 /* 2255 * make sure that internal objs are in a map somewhere 2256 * and none have zero ref counts. 2257 */ 2258 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2259 if (object->handle == NULL && 2260 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2261 if (object->ref_count == 0) { 2262 db_printf("vmochk: internal obj has zero ref count: %ld\n", 2263 (long)object->size); 2264 } 2265 if (!vm_object_in_map(object)) { 2266 db_printf( 2267 "vmochk: internal obj is not in a map: " 2268 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", 2269 object->ref_count, (u_long)object->size, 2270 (u_long)object->size, 2271 (void *)object->backing_object); 2272 } 2273 } 2274 } 2275 } 2276 2277 /* 2278 * vm_object_print: [ debug ] 2279 */ 2280 DB_SHOW_COMMAND(object, vm_object_print_static) 2281 { 2282 /* XXX convert args. */ 2283 vm_object_t object = (vm_object_t)addr; 2284 boolean_t full = have_addr; 2285 2286 vm_page_t p; 2287 2288 /* XXX count is an (unused) arg. Avoid shadowing it. */ 2289 #define count was_count 2290 2291 int count; 2292 2293 if (object == NULL) 2294 return; 2295 2296 db_iprintf( 2297 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n", 2298 object, (int)object->type, (uintmax_t)object->size, 2299 object->resident_page_count, object->ref_count, object->flags, 2300 object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge); 2301 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", 2302 object->shadow_count, 2303 object->backing_object ? object->backing_object->ref_count : 0, 2304 object->backing_object, (uintmax_t)object->backing_object_offset); 2305 2306 if (!full) 2307 return; 2308 2309 db_indent += 2; 2310 count = 0; 2311 TAILQ_FOREACH(p, &object->memq, listq) { 2312 if (count == 0) 2313 db_iprintf("memory:="); 2314 else if (count == 6) { 2315 db_printf("\n"); 2316 db_iprintf(" ..."); 2317 count = 0; 2318 } else 2319 db_printf(","); 2320 count++; 2321 2322 db_printf("(off=0x%jx,page=0x%jx)", 2323 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); 2324 } 2325 if (count != 0) 2326 db_printf("\n"); 2327 db_indent -= 2; 2328 } 2329 2330 /* XXX. */ 2331 #undef count 2332 2333 /* XXX need this non-static entry for calling from vm_map_print. */ 2334 void 2335 vm_object_print( 2336 /* db_expr_t */ long addr, 2337 boolean_t have_addr, 2338 /* db_expr_t */ long count, 2339 char *modif) 2340 { 2341 vm_object_print_static(addr, have_addr, count, modif); 2342 } 2343 2344 DB_SHOW_COMMAND(vmopag, vm_object_print_pages) 2345 { 2346 vm_object_t object; 2347 vm_pindex_t fidx; 2348 vm_paddr_t pa; 2349 vm_page_t m, prev_m; 2350 int rcount, nl, c; 2351 2352 nl = 0; 2353 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2354 db_printf("new object: %p\n", (void *)object); 2355 if (nl > 18) { 2356 c = cngetc(); 2357 if (c != ' ') 2358 return; 2359 nl = 0; 2360 } 2361 nl++; 2362 rcount = 0; 2363 fidx = 0; 2364 pa = -1; 2365 TAILQ_FOREACH(m, &object->memq, listq) { 2366 if (m->pindex > 128) 2367 break; 2368 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && 2369 prev_m->pindex + 1 != m->pindex) { 2370 if (rcount) { 2371 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2372 (long)fidx, rcount, (long)pa); 2373 if (nl > 18) { 2374 c = cngetc(); 2375 if (c != ' ') 2376 return; 2377 nl = 0; 2378 } 2379 nl++; 2380 rcount = 0; 2381 } 2382 } 2383 if (rcount && 2384 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { 2385 ++rcount; 2386 continue; 2387 } 2388 if (rcount) { 2389 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2390 (long)fidx, rcount, (long)pa); 2391 if (nl > 18) { 2392 c = cngetc(); 2393 if (c != ' ') 2394 return; 2395 nl = 0; 2396 } 2397 nl++; 2398 } 2399 fidx = m->pindex; 2400 pa = VM_PAGE_TO_PHYS(m); 2401 rcount = 1; 2402 } 2403 if (rcount) { 2404 db_printf(" index(%ld)run(%d)pa(0x%lx)\n", 2405 (long)fidx, rcount, (long)pa); 2406 if (nl > 18) { 2407 c = cngetc(); 2408 if (c != ' ') 2409 return; 2410 nl = 0; 2411 } 2412 nl++; 2413 } 2414 } 2415 } 2416 #endif /* DDB */ 2417