| 1 | /* |
| 2 | * Copyright (c) 1991 Regents of the University of California. |
| 3 | * 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 | * 3. All advertising materials mentioning features or use of this software |
| 17 | * must display the following acknowledgement: |
| 18 | * This product includes software developed by the University of |
| 19 | * California, Berkeley and its contributors. |
| 20 | * 4. Neither the name of the University nor the names of its contributors |
| 21 | * may be used to endorse or promote products derived from this software |
| 22 | * without specific prior written permission. |
| 23 | * |
| 24 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 25 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 26 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 27 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 28 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 29 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 30 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 31 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 32 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 33 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 34 | * SUCH DAMAGE. |
| 35 | * |
| 36 | * from: @(#)vm_object.c 7.4 (Berkeley) 5/7/91 |
| 37 | * $Id: vm_object.c,v 1.21 1994/02/10 00:15:51 davidg Exp $ |
| 38 | * |
| 39 | * |
| 40 | * Copyright (c) 1987, 1990 Carnegie-Mellon University. |
| 41 | * All rights reserved. |
| 42 | * |
| 43 | * Authors: Avadis Tevanian, Jr., Michael Wayne Young |
| 44 | * |
| 45 | * Permission to use, copy, modify and distribute this software and |
| 46 | * its documentation is hereby granted, provided that both the copyright |
| 47 | * notice and this permission notice appear in all copies of the |
| 48 | * software, derivative works or modified versions, and any portions |
| 49 | * thereof, and that both notices appear in supporting documentation. |
| 50 | * |
| 51 | * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" |
| 52 | * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND |
| 53 | * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. |
| 54 | * |
| 55 | * Carnegie Mellon requests users of this software to return to |
| 56 | * |
| 57 | * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU |
| 58 | * School of Computer Science |
| 59 | * Carnegie Mellon University |
| 60 | * Pittsburgh PA 15213-3890 |
| 61 | * |
| 62 | * any improvements or extensions that they make and grant Carnegie the |
| 63 | * rights to redistribute these changes. |
| 64 | */ |
| 65 | |
| 66 | /* |
| 67 | * Virtual memory object module. |
| 68 | */ |
| 69 | |
| 70 | #include "ddb.h" |
| 71 | #include "param.h" |
| 72 | #include "malloc.h" |
| 73 | #include "systm.h" |
| 74 | |
| 75 | #include "vm.h" |
| 76 | #include "vm_page.h" |
| 77 | #include "proc.h" |
| 78 | |
| 79 | |
| 80 | static void _vm_object_allocate(vm_size_t, vm_object_t); |
| 81 | void vm_object_deactivate_pages(vm_object_t); |
| 82 | static void vm_object_cache_trim(void); |
| 83 | static void vm_object_remove(vm_pager_t); |
| 84 | |
| 85 | /* |
| 86 | * Virtual memory objects maintain the actual data |
| 87 | * associated with allocated virtual memory. A given |
| 88 | * page of memory exists within exactly one object. |
| 89 | * |
| 90 | * An object is only deallocated when all "references" |
| 91 | * are given up. Only one "reference" to a given |
| 92 | * region of an object should be writeable. |
| 93 | * |
| 94 | * Associated with each object is a list of all resident |
| 95 | * memory pages belonging to that object; this list is |
| 96 | * maintained by the "vm_page" module, and locked by the object's |
| 97 | * lock. |
| 98 | * |
| 99 | * Each object also records a "pager" routine which is |
| 100 | * used to retrieve (and store) pages to the proper backing |
| 101 | * storage. In addition, objects may be backed by other |
| 102 | * objects from which they were virtual-copied. |
| 103 | * |
| 104 | * The only items within the object structure which are |
| 105 | * modified after time of creation are: |
| 106 | * reference count locked by object's lock |
| 107 | * pager routine locked by object's lock |
| 108 | * |
| 109 | */ |
| 110 | |
| 111 | |
| 112 | queue_head_t vm_object_cached_list; /* list of objects persisting */ |
| 113 | int vm_object_cached; /* size of cached list */ |
| 114 | simple_lock_data_t vm_cache_lock; /* lock for object cache */ |
| 115 | |
| 116 | queue_head_t vm_object_list; /* list of allocated objects */ |
| 117 | long vm_object_count; /* count of all objects */ |
| 118 | simple_lock_data_t vm_object_list_lock; |
| 119 | /* lock for object list and count */ |
| 120 | |
| 121 | vm_object_t kernel_object; /* the single kernel object */ |
| 122 | vm_object_t kmem_object; /* the kernel malloc object */ |
| 123 | struct vm_object kernel_object_store; |
| 124 | struct vm_object kmem_object_store; |
| 125 | |
| 126 | extern int vm_cache_max; |
| 127 | #define VM_OBJECT_HASH_COUNT 127 |
| 128 | |
| 129 | queue_head_t vm_object_hashtable[VM_OBJECT_HASH_COUNT]; |
| 130 | |
| 131 | long object_collapses = 0; |
| 132 | long object_bypasses = 0; |
| 133 | |
| 134 | /* |
| 135 | * internal version of vm_object_allocate |
| 136 | */ |
| 137 | static inline void |
| 138 | _vm_object_allocate(size, object) |
| 139 | vm_size_t size; |
| 140 | register vm_object_t object; |
| 141 | { |
| 142 | queue_init(&object->memq); |
| 143 | vm_object_lock_init(object); |
| 144 | object->ref_count = 1; |
| 145 | object->resident_page_count = 0; |
| 146 | object->size = size; |
| 147 | object->can_persist = FALSE; |
| 148 | object->paging_in_progress = 0; |
| 149 | object->copy = NULL; |
| 150 | |
| 151 | /* |
| 152 | * Object starts out read-write, with no pager. |
| 153 | */ |
| 154 | |
| 155 | object->pager = NULL; |
| 156 | object->internal = TRUE; /* vm_allocate_with_pager will reset */ |
| 157 | object->paging_offset = 0; |
| 158 | object->shadow = NULL; |
| 159 | object->shadow_offset = (vm_offset_t) 0; |
| 160 | |
| 161 | simple_lock(&vm_object_list_lock); |
| 162 | queue_enter(&vm_object_list, object, vm_object_t, object_list); |
| 163 | vm_object_count++; |
| 164 | simple_unlock(&vm_object_list_lock); |
| 165 | } |
| 166 | |
| 167 | /* |
| 168 | * vm_object_init: |
| 169 | * |
| 170 | * Initialize the VM objects module. |
| 171 | */ |
| 172 | void |
| 173 | vm_object_init() |
| 174 | { |
| 175 | register int i; |
| 176 | |
| 177 | queue_init(&vm_object_cached_list); |
| 178 | queue_init(&vm_object_list); |
| 179 | vm_object_count = 0; |
| 180 | simple_lock_init(&vm_cache_lock); |
| 181 | simple_lock_init(&vm_object_list_lock); |
| 182 | |
| 183 | for (i = 0; i < VM_OBJECT_HASH_COUNT; i++) |
| 184 | queue_init(&vm_object_hashtable[i]); |
| 185 | |
| 186 | kernel_object = &kernel_object_store; |
| 187 | _vm_object_allocate(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS, |
| 188 | kernel_object); |
| 189 | |
| 190 | kmem_object = &kmem_object_store; |
| 191 | _vm_object_allocate(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS, |
| 192 | kmem_object); |
| 193 | } |
| 194 | |
| 195 | /* |
| 196 | * vm_object_allocate: |
| 197 | * |
| 198 | * Returns a new object with the given size. |
| 199 | */ |
| 200 | |
| 201 | vm_object_t |
| 202 | vm_object_allocate(size) |
| 203 | vm_size_t size; |
| 204 | { |
| 205 | register vm_object_t result; |
| 206 | int s; |
| 207 | |
| 208 | result = (vm_object_t) |
| 209 | malloc((u_long)sizeof *result, M_VMOBJ, M_WAITOK); |
| 210 | |
| 211 | |
| 212 | _vm_object_allocate(size, result); |
| 213 | |
| 214 | return(result); |
| 215 | } |
| 216 | |
| 217 | |
| 218 | /* |
| 219 | * vm_object_reference: |
| 220 | * |
| 221 | * Gets another reference to the given object. |
| 222 | */ |
| 223 | inline void |
| 224 | vm_object_reference(object) |
| 225 | register vm_object_t object; |
| 226 | { |
| 227 | if (object == NULL) |
| 228 | return; |
| 229 | |
| 230 | vm_object_lock(object); |
| 231 | object->ref_count++; |
| 232 | vm_object_unlock(object); |
| 233 | } |
| 234 | |
| 235 | /* |
| 236 | * vm_object_deallocate: |
| 237 | * |
| 238 | * Release a reference to the specified object, |
| 239 | * gained either through a vm_object_allocate |
| 240 | * or a vm_object_reference call. When all references |
| 241 | * are gone, storage associated with this object |
| 242 | * may be relinquished. |
| 243 | * |
| 244 | * No object may be locked. |
| 245 | */ |
| 246 | void |
| 247 | vm_object_deallocate(object) |
| 248 | vm_object_t object; |
| 249 | { |
| 250 | vm_object_t temp; |
| 251 | |
| 252 | while (object != NULL) { |
| 253 | |
| 254 | /* |
| 255 | * The cache holds a reference (uncounted) to |
| 256 | * the object; we must lock it before removing |
| 257 | * the object. |
| 258 | */ |
| 259 | |
| 260 | vm_object_cache_lock(); |
| 261 | |
| 262 | /* |
| 263 | * Lose the reference |
| 264 | */ |
| 265 | vm_object_lock(object); |
| 266 | if (--(object->ref_count) != 0) { |
| 267 | |
| 268 | vm_object_unlock(object); |
| 269 | /* |
| 270 | * If there are still references, then |
| 271 | * we are done. |
| 272 | */ |
| 273 | vm_object_cache_unlock(); |
| 274 | return; |
| 275 | } |
| 276 | |
| 277 | /* |
| 278 | * See if this object can persist. If so, enter |
| 279 | * it in the cache, then deactivate all of its |
| 280 | * pages. |
| 281 | */ |
| 282 | |
| 283 | if (object->can_persist) { |
| 284 | |
| 285 | queue_enter(&vm_object_cached_list, object, |
| 286 | vm_object_t, cached_list); |
| 287 | vm_object_cached++; |
| 288 | vm_object_cache_unlock(); |
| 289 | |
| 290 | /* vm_object_deactivate_pages(object); */ |
| 291 | vm_object_unlock(object); |
| 292 | |
| 293 | vm_object_cache_trim(); |
| 294 | return; |
| 295 | } |
| 296 | |
| 297 | /* |
| 298 | * Make sure no one can look us up now. |
| 299 | */ |
| 300 | vm_object_remove(object->pager); |
| 301 | vm_object_cache_unlock(); |
| 302 | |
| 303 | temp = object->shadow; |
| 304 | vm_object_terminate(object); |
| 305 | /* unlocks and deallocates object */ |
| 306 | object = temp; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | /* |
| 311 | * vm_object_terminate actually destroys the specified object, freeing |
| 312 | * up all previously used resources. |
| 313 | * |
| 314 | * The object must be locked. |
| 315 | */ |
| 316 | void |
| 317 | vm_object_terminate(object) |
| 318 | register vm_object_t object; |
| 319 | { |
| 320 | register vm_page_t p; |
| 321 | vm_object_t shadow_object; |
| 322 | int s; |
| 323 | |
| 324 | /* |
| 325 | * Detach the object from its shadow if we are the shadow's |
| 326 | * copy. |
| 327 | */ |
| 328 | if ((shadow_object = object->shadow) != NULL) { |
| 329 | vm_object_lock(shadow_object); |
| 330 | if (shadow_object->copy == object) |
| 331 | shadow_object->copy = NULL; |
| 332 | /* |
| 333 | else if (shadow_object->copy != NULL) |
| 334 | panic("vm_object_terminate: copy/shadow inconsistency"); |
| 335 | */ |
| 336 | vm_object_unlock(shadow_object); |
| 337 | } |
| 338 | |
| 339 | /* |
| 340 | * optim: get rid of any pages that we can right now |
| 341 | * so the pageout daemon can't get any more to page |
| 342 | * out at rundown. |
| 343 | */ |
| 344 | #if 0 |
| 345 | p = (vm_page_t) queue_first(&object->memq); |
| 346 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 347 | vm_page_t next = (vm_page_t) queue_next(&p->listq); |
| 348 | VM_PAGE_CHECK(p); |
| 349 | vm_page_lock_queues(); |
| 350 | |
| 351 | if (p->flags & PG_BUSY) { |
| 352 | p = next; |
| 353 | vm_page_unlock_queues(); |
| 354 | continue; |
| 355 | } |
| 356 | if (!object->internal) { |
| 357 | if ((p->flags & PG_CLEAN) == 0) { |
| 358 | p = next; |
| 359 | vm_page_unlock_queues(); |
| 360 | continue; |
| 361 | } |
| 362 | |
| 363 | if (pmap_is_modified(VM_PAGE_TO_PHYS(p))) { |
| 364 | p->flags &= ~PG_CLEAN; |
| 365 | p = next; |
| 366 | vm_page_unlock_queues(); |
| 367 | continue; |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | vm_page_free(p); |
| 372 | vm_page_unlock_queues(); |
| 373 | p = next; |
| 374 | } |
| 375 | #endif |
| 376 | |
| 377 | /* |
| 378 | * Wait until the pageout daemon is through |
| 379 | * with the object. |
| 380 | */ |
| 381 | |
| 382 | while (object->paging_in_progress != 0) { |
| 383 | vm_object_sleep(object, object, FALSE); |
| 384 | vm_object_lock(object); |
| 385 | } |
| 386 | |
| 387 | /* |
| 388 | * While the paging system is locked, |
| 389 | * pull the object's pages off the active |
| 390 | * and inactive queues. This keeps the |
| 391 | * pageout daemon from playing with them |
| 392 | * during vm_pager_deallocate. |
| 393 | * |
| 394 | * We can't free the pages yet, because the |
| 395 | * object's pager may have to write them out |
| 396 | * before deallocating the paging space. |
| 397 | */ |
| 398 | |
| 399 | p = (vm_page_t) queue_first(&object->memq); |
| 400 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 401 | VM_PAGE_CHECK(p); |
| 402 | |
| 403 | vm_page_lock_queues(); |
| 404 | s = vm_disable_intr(); |
| 405 | if (p->flags & PG_ACTIVE) { |
| 406 | queue_remove(&vm_page_queue_active, p, vm_page_t, |
| 407 | pageq); |
| 408 | p->flags &= ~PG_ACTIVE; |
| 409 | vm_page_active_count--; |
| 410 | } |
| 411 | |
| 412 | if (p->flags & PG_INACTIVE) { |
| 413 | queue_remove(&vm_page_queue_inactive, p, vm_page_t, |
| 414 | pageq); |
| 415 | p->flags &= ~PG_INACTIVE; |
| 416 | vm_page_inactive_count--; |
| 417 | } |
| 418 | vm_set_intr(s); |
| 419 | vm_page_unlock_queues(); |
| 420 | p = (vm_page_t) queue_next(&p->listq); |
| 421 | } |
| 422 | |
| 423 | vm_object_unlock(object); |
| 424 | |
| 425 | if (object->paging_in_progress != 0) |
| 426 | panic("vm_object_deallocate: pageout in progress"); |
| 427 | |
| 428 | /* |
| 429 | * Clean and free the pages, as appropriate. |
| 430 | * All references to the object are gone, |
| 431 | * so we don't need to lock it. |
| 432 | */ |
| 433 | |
| 434 | if (!object->internal) { |
| 435 | vm_object_lock(object); |
| 436 | vm_object_page_clean(object, 0, 0); |
| 437 | vm_object_unlock(object); |
| 438 | } |
| 439 | |
| 440 | while (!queue_empty(&object->memq)) { |
| 441 | p = (vm_page_t) queue_first(&object->memq); |
| 442 | |
| 443 | VM_PAGE_CHECK(p); |
| 444 | |
| 445 | vm_page_lock_queues(); |
| 446 | vm_page_free(p); |
| 447 | vm_page_unlock_queues(); |
| 448 | } |
| 449 | |
| 450 | /* |
| 451 | * Let the pager know object is dead. |
| 452 | */ |
| 453 | |
| 454 | if (object->pager != NULL) { |
| 455 | vm_pager_deallocate(object->pager); |
| 456 | } |
| 457 | |
| 458 | |
| 459 | simple_lock(&vm_object_list_lock); |
| 460 | queue_remove(&vm_object_list, object, vm_object_t, object_list); |
| 461 | vm_object_count--; |
| 462 | simple_unlock(&vm_object_list_lock); |
| 463 | |
| 464 | /* |
| 465 | * Free the space for the object. |
| 466 | */ |
| 467 | |
| 468 | free((caddr_t)object, M_VMOBJ); |
| 469 | } |
| 470 | |
| 471 | /* |
| 472 | * vm_object_page_clean |
| 473 | * |
| 474 | * Clean all dirty pages in the specified range of object. |
| 475 | * Leaves page on whatever queue it is currently on. |
| 476 | * |
| 477 | * Odd semantics: if start == end, we clean everything. |
| 478 | * |
| 479 | * The object must be locked. |
| 480 | */ |
| 481 | void |
| 482 | vm_object_page_clean(object, start, end) |
| 483 | register vm_object_t object; |
| 484 | register vm_offset_t start; |
| 485 | register vm_offset_t end; |
| 486 | { |
| 487 | register vm_page_t p; |
| 488 | int s; |
| 489 | int size; |
| 490 | |
| 491 | if (object->pager == NULL) |
| 492 | return; |
| 493 | |
| 494 | if (start != end) { |
| 495 | start = trunc_page(start); |
| 496 | end = round_page(end); |
| 497 | } |
| 498 | size = end - start; |
| 499 | |
| 500 | again: |
| 501 | p = (vm_page_t) queue_first(&object->memq); |
| 502 | while (!queue_end(&object->memq, (queue_entry_t) p) && ((start == end) || (size != 0) ) ) { |
| 503 | if (start == end || (p->offset >= start && p->offset < end)) { |
| 504 | if (p->flags & PG_BUSY) |
| 505 | goto next; |
| 506 | |
| 507 | size -= PAGE_SIZE; |
| 508 | |
| 509 | if ((p->flags & PG_CLEAN) |
| 510 | && pmap_is_modified(VM_PAGE_TO_PHYS(p))) |
| 511 | p->flags &= ~PG_CLEAN; |
| 512 | |
| 513 | if (p->flags & PG_ACTIVE) |
| 514 | vm_page_deactivate(p); |
| 515 | |
| 516 | if ((p->flags & PG_CLEAN) == 0) { |
| 517 | p->flags |= PG_BUSY; |
| 518 | object->paging_in_progress++; |
| 519 | vm_object_unlock(object); |
| 520 | (void) vm_pager_put(object->pager, p, TRUE); |
| 521 | vm_object_lock(object); |
| 522 | object->paging_in_progress--; |
| 523 | if (object->paging_in_progress == 0) |
| 524 | wakeup((caddr_t) object); |
| 525 | PAGE_WAKEUP(p); |
| 526 | goto again; |
| 527 | } |
| 528 | } |
| 529 | next: |
| 530 | p = (vm_page_t) queue_next(&p->listq); |
| 531 | } |
| 532 | wakeup((caddr_t)object); |
| 533 | } |
| 534 | |
| 535 | /* |
| 536 | * vm_object_deactivate_pages |
| 537 | * |
| 538 | * Deactivate all pages in the specified object. (Keep its pages |
| 539 | * in memory even though it is no longer referenced.) |
| 540 | * |
| 541 | * The object must be locked. |
| 542 | */ |
| 543 | void |
| 544 | vm_object_deactivate_pages(object) |
| 545 | register vm_object_t object; |
| 546 | { |
| 547 | register vm_page_t p, next; |
| 548 | |
| 549 | p = (vm_page_t) queue_first(&object->memq); |
| 550 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 551 | next = (vm_page_t) queue_next(&p->listq); |
| 552 | vm_page_lock_queues(); |
| 553 | if ((p->flags & (PG_INACTIVE|PG_BUSY)) == 0 && |
| 554 | p->wire_count == 0) |
| 555 | vm_page_deactivate(p); /* optimisation from mach 3.0 - |
| 556 | * andrew@werple.apana.org.au, |
| 557 | * Feb '93 |
| 558 | */ |
| 559 | vm_page_unlock_queues(); |
| 560 | p = next; |
| 561 | } |
| 562 | } |
| 563 | |
| 564 | /* |
| 565 | * Trim the object cache to size. |
| 566 | */ |
| 567 | void |
| 568 | vm_object_cache_trim() |
| 569 | { |
| 570 | register vm_object_t object; |
| 571 | |
| 572 | vm_object_cache_lock(); |
| 573 | while (vm_object_cached > vm_cache_max) { |
| 574 | object = (vm_object_t) queue_first(&vm_object_cached_list); |
| 575 | vm_object_cache_unlock(); |
| 576 | |
| 577 | if (object != vm_object_lookup(object->pager)) |
| 578 | panic("vm_object_deactivate: I'm sooo confused."); |
| 579 | |
| 580 | pager_cache(object, FALSE); |
| 581 | |
| 582 | vm_object_cache_lock(); |
| 583 | } |
| 584 | vm_object_cache_unlock(); |
| 585 | } |
| 586 | |
| 587 | /* |
| 588 | * vm_object_shutdown() |
| 589 | * |
| 590 | * Shut down the object system. Unfortunately, while we |
| 591 | * may be trying to do this, init is happily waiting for |
| 592 | * processes to exit, and therefore will be causing some objects |
| 593 | * to be deallocated. To handle this, we gain a fake reference |
| 594 | * to all objects we release paging areas for. This will prevent |
| 595 | * a duplicate deallocation. This routine is probably full of |
| 596 | * race conditions! |
| 597 | */ |
| 598 | |
| 599 | #if 0 |
| 600 | void |
| 601 | vm_object_shutdown() |
| 602 | { |
| 603 | register vm_object_t object; |
| 604 | |
| 605 | /* |
| 606 | * Clean up the object cache *before* we screw up the reference |
| 607 | * counts on all of the objects. |
| 608 | */ |
| 609 | |
| 610 | vm_object_cache_clear(); |
| 611 | |
| 612 | |
| 613 | /* |
| 614 | * First we gain a reference to each object so that |
| 615 | * no one else will deallocate them. |
| 616 | */ |
| 617 | |
| 618 | simple_lock(&vm_object_list_lock); |
| 619 | object = (vm_object_t) queue_first(&vm_object_list); |
| 620 | while (!queue_end(&vm_object_list, (queue_entry_t) object)) { |
| 621 | vm_object_reference(object); |
| 622 | object = (vm_object_t) queue_next(&object->object_list); |
| 623 | } |
| 624 | simple_unlock(&vm_object_list_lock); |
| 625 | |
| 626 | /* |
| 627 | * Now we deallocate all the paging areas. We don't need |
| 628 | * to lock anything because we've reduced to a single |
| 629 | * processor while shutting down. This also assumes that |
| 630 | * no new objects are being created. |
| 631 | */ |
| 632 | |
| 633 | object = (vm_object_t) queue_first(&vm_object_list); |
| 634 | while (!queue_end(&vm_object_list, (queue_entry_t) object)) { |
| 635 | if (object->pager != NULL) |
| 636 | vm_pager_deallocate(object->pager); |
| 637 | object = (vm_object_t) queue_next(&object->object_list); |
| 638 | printf("."); |
| 639 | } |
| 640 | printf("done.\n"); |
| 641 | } |
| 642 | #endif |
| 643 | /* |
| 644 | * vm_object_pmap_copy: |
| 645 | * |
| 646 | * Makes all physical pages in the specified |
| 647 | * object range copy-on-write. No writeable |
| 648 | * references to these pages should remain. |
| 649 | * |
| 650 | * The object must *not* be locked. |
| 651 | */ |
| 652 | void |
| 653 | vm_object_pmap_copy(object, start, end) |
| 654 | register vm_object_t object; |
| 655 | register vm_offset_t start; |
| 656 | register vm_offset_t end; |
| 657 | { |
| 658 | register vm_page_t p; |
| 659 | vm_offset_t amount; |
| 660 | |
| 661 | start = trunc_page(start); |
| 662 | end = round_page(end); |
| 663 | |
| 664 | amount = ((end - start) + PAGE_SIZE - 1) / PAGE_SIZE; |
| 665 | |
| 666 | if (object == NULL) |
| 667 | return; |
| 668 | |
| 669 | vm_object_lock(object); |
| 670 | p = (vm_page_t) queue_first(&object->memq); |
| 671 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 672 | if ((start <= p->offset) && (p->offset < end)) { |
| 673 | pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_READ); |
| 674 | p->flags |= PG_COPY_ON_WRITE; |
| 675 | amount -= 1; |
| 676 | if (amount <= 0) |
| 677 | break; |
| 678 | } |
| 679 | p = (vm_page_t) queue_next(&p->listq); |
| 680 | } |
| 681 | vm_object_unlock(object); |
| 682 | } |
| 683 | |
| 684 | /* |
| 685 | * vm_object_pmap_remove: |
| 686 | * |
| 687 | * Removes all physical pages in the specified |
| 688 | * object range from all physical maps. |
| 689 | * |
| 690 | * The object must *not* be locked. |
| 691 | */ |
| 692 | void |
| 693 | vm_object_pmap_remove(object, start, end) |
| 694 | register vm_object_t object; |
| 695 | register vm_offset_t start; |
| 696 | register vm_offset_t end; |
| 697 | { |
| 698 | register vm_page_t p; |
| 699 | vm_offset_t size; |
| 700 | |
| 701 | if (object == NULL) |
| 702 | return; |
| 703 | |
| 704 | vm_object_lock(object); |
| 705 | again: |
| 706 | size = ((end - start) + PAGE_SIZE - 1) / PAGE_SIZE; |
| 707 | p = (vm_page_t) queue_first(&object->memq); |
| 708 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 709 | if ((start <= p->offset) && (p->offset < end)) { |
| 710 | if (p->flags & PG_BUSY) { |
| 711 | p->flags |= PG_WANTED; |
| 712 | tsleep((caddr_t) p, PVM, "vmopmr", 0); |
| 713 | goto again; |
| 714 | } |
| 715 | pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE); |
| 716 | if ((p->flags & PG_CLEAN) == 0) |
| 717 | p->flags |= PG_LAUNDRY; |
| 718 | if (--size <= 0) break; |
| 719 | } |
| 720 | p = (vm_page_t) queue_next(&p->listq); |
| 721 | } |
| 722 | vm_object_unlock(object); |
| 723 | } |
| 724 | |
| 725 | /* |
| 726 | * vm_object_copy: |
| 727 | * |
| 728 | * Create a new object which is a copy of an existing |
| 729 | * object, and mark all of the pages in the existing |
| 730 | * object 'copy-on-write'. The new object has one reference. |
| 731 | * Returns the new object. |
| 732 | * |
| 733 | * May defer the copy until later if the object is not backed |
| 734 | * up by a non-default pager. |
| 735 | */ |
| 736 | void |
| 737 | vm_object_copy(src_object, src_offset, size, dst_object, dst_offset, src_needs_copy) |
| 738 | register vm_object_t src_object; |
| 739 | vm_offset_t src_offset; |
| 740 | vm_size_t size; |
| 741 | vm_object_t *dst_object; /* OUT */ |
| 742 | vm_offset_t *dst_offset; /* OUT */ |
| 743 | boolean_t *src_needs_copy; /* OUT */ |
| 744 | { |
| 745 | register vm_object_t new_copy; |
| 746 | register vm_object_t old_copy; |
| 747 | vm_offset_t new_start, new_end; |
| 748 | vm_offset_t src_offset_end; |
| 749 | vm_offset_t tmpsize; |
| 750 | |
| 751 | register vm_page_t p; |
| 752 | |
| 753 | if (src_object == NULL) { |
| 754 | /* |
| 755 | * Nothing to copy |
| 756 | */ |
| 757 | *dst_object = NULL; |
| 758 | *dst_offset = 0; |
| 759 | *src_needs_copy = FALSE; |
| 760 | return; |
| 761 | } |
| 762 | |
| 763 | /* |
| 764 | * If the object's pager is null_pager or the |
| 765 | * default pager, we don't have to make a copy |
| 766 | * of it. Instead, we set the needs copy flag and |
| 767 | * make a shadow later. |
| 768 | * DYSON: check for swap(default) pager too.... |
| 769 | */ |
| 770 | |
| 771 | vm_object_lock(src_object); |
| 772 | |
| 773 | /* |
| 774 | * Try to collapse the object before copying it. |
| 775 | */ |
| 776 | |
| 777 | vm_object_collapse(src_object); |
| 778 | |
| 779 | if (src_object->pager == NULL || |
| 780 | src_object->pager->pg_type == PG_SWAP || |
| 781 | src_object->internal) { |
| 782 | |
| 783 | /* |
| 784 | * Make another reference to the object |
| 785 | */ |
| 786 | src_object->ref_count++; |
| 787 | |
| 788 | /* |
| 789 | * Mark all of the pages copy-on-write. |
| 790 | */ |
| 791 | tmpsize = size; |
| 792 | src_offset_end = src_offset + size; |
| 793 | for (p = (vm_page_t) queue_first(&src_object->memq); |
| 794 | !queue_end(&src_object->memq, (queue_entry_t)p) && tmpsize > 0; |
| 795 | p = (vm_page_t) queue_next(&p->listq)) { |
| 796 | if (src_offset <= p->offset && |
| 797 | p->offset < src_offset_end) { |
| 798 | p->flags |= PG_COPY_ON_WRITE; |
| 799 | tmpsize -= PAGE_SIZE; |
| 800 | } |
| 801 | } |
| 802 | vm_object_unlock(src_object); |
| 803 | |
| 804 | *dst_object = src_object; |
| 805 | *dst_offset = src_offset; |
| 806 | |
| 807 | /* |
| 808 | * Must make a shadow when write is desired |
| 809 | */ |
| 810 | *src_needs_copy = TRUE; |
| 811 | return; |
| 812 | } |
| 813 | |
| 814 | /* |
| 815 | * If the object has a pager, the pager wants to |
| 816 | * see all of the changes. We need a copy-object |
| 817 | * for the changed pages. |
| 818 | * |
| 819 | * If there is a copy-object, and it is empty, |
| 820 | * no changes have been made to the object since the |
| 821 | * copy-object was made. We can use the same copy- |
| 822 | * object. |
| 823 | */ |
| 824 | |
| 825 | Retry1: |
| 826 | old_copy = src_object->copy; |
| 827 | if (old_copy != NULL) { |
| 828 | /* |
| 829 | * Try to get the locks (out of order) |
| 830 | */ |
| 831 | if (!vm_object_lock_try(old_copy)) { |
| 832 | vm_object_unlock(src_object); |
| 833 | |
| 834 | /* should spin a bit here... */ |
| 835 | vm_object_lock(src_object); |
| 836 | goto Retry1; |
| 837 | } |
| 838 | |
| 839 | if (old_copy->resident_page_count == 0 && |
| 840 | old_copy->pager == NULL) { |
| 841 | /* |
| 842 | * Return another reference to |
| 843 | * the existing copy-object. |
| 844 | */ |
| 845 | old_copy->ref_count++; |
| 846 | vm_object_unlock(old_copy); |
| 847 | vm_object_unlock(src_object); |
| 848 | *dst_object = old_copy; |
| 849 | *dst_offset = src_offset; |
| 850 | *src_needs_copy = FALSE; |
| 851 | return; |
| 852 | } |
| 853 | vm_object_unlock(old_copy); |
| 854 | } |
| 855 | vm_object_unlock(src_object); |
| 856 | |
| 857 | /* |
| 858 | * If the object has a pager, the pager wants |
| 859 | * to see all of the changes. We must make |
| 860 | * a copy-object and put the changed pages there. |
| 861 | * |
| 862 | * The copy-object is always made large enough to |
| 863 | * completely shadow the original object, since |
| 864 | * it may have several users who want to shadow |
| 865 | * the original object at different points. |
| 866 | */ |
| 867 | |
| 868 | new_copy = vm_object_allocate(src_object->size); |
| 869 | |
| 870 | Retry2: |
| 871 | vm_object_lock(src_object); |
| 872 | /* |
| 873 | * Copy object may have changed while we were unlocked |
| 874 | */ |
| 875 | old_copy = src_object->copy; |
| 876 | if (old_copy != NULL) { |
| 877 | /* |
| 878 | * Try to get the locks (out of order) |
| 879 | */ |
| 880 | if (!vm_object_lock_try(old_copy)) { |
| 881 | vm_object_unlock(src_object); |
| 882 | goto Retry2; |
| 883 | } |
| 884 | |
| 885 | /* |
| 886 | * Consistency check |
| 887 | */ |
| 888 | if (old_copy->shadow != src_object || |
| 889 | old_copy->shadow_offset != (vm_offset_t) 0) |
| 890 | panic("vm_object_copy: copy/shadow inconsistency"); |
| 891 | |
| 892 | /* |
| 893 | * Make the old copy-object shadow the new one. |
| 894 | * It will receive no more pages from the original |
| 895 | * object. |
| 896 | */ |
| 897 | |
| 898 | src_object->ref_count--; /* remove ref. from old_copy */ |
| 899 | old_copy->shadow = new_copy; |
| 900 | new_copy->ref_count++; /* locking not needed - we |
| 901 | have the only pointer */ |
| 902 | vm_object_unlock(old_copy); /* done with old_copy */ |
| 903 | } |
| 904 | |
| 905 | new_start = (vm_offset_t) 0; /* always shadow original at 0 */ |
| 906 | new_end = (vm_offset_t) new_copy->size; /* for the whole object */ |
| 907 | |
| 908 | /* |
| 909 | * Point the new copy at the existing object. |
| 910 | */ |
| 911 | |
| 912 | new_copy->shadow = src_object; |
| 913 | new_copy->shadow_offset = new_start; |
| 914 | src_object->ref_count++; |
| 915 | src_object->copy = new_copy; |
| 916 | |
| 917 | /* |
| 918 | * Mark all the affected pages of the existing object |
| 919 | * copy-on-write. |
| 920 | */ |
| 921 | tmpsize = size; |
| 922 | p = (vm_page_t) queue_first(&src_object->memq); |
| 923 | while (!queue_end(&src_object->memq, (queue_entry_t) p) && tmpsize > 0) { |
| 924 | if ((new_start <= p->offset) && (p->offset < new_end)) { |
| 925 | p->flags |= PG_COPY_ON_WRITE; |
| 926 | tmpsize -= PAGE_SIZE; |
| 927 | } |
| 928 | p = (vm_page_t) queue_next(&p->listq); |
| 929 | } |
| 930 | |
| 931 | vm_object_unlock(src_object); |
| 932 | |
| 933 | *dst_object = new_copy; |
| 934 | *dst_offset = src_offset - new_start; |
| 935 | *src_needs_copy = FALSE; |
| 936 | } |
| 937 | |
| 938 | /* |
| 939 | * vm_object_shadow: |
| 940 | * |
| 941 | * Create a new object which is backed by the |
| 942 | * specified existing object range. The source |
| 943 | * object reference is deallocated. |
| 944 | * |
| 945 | * The new object and offset into that object |
| 946 | * are returned in the source parameters. |
| 947 | */ |
| 948 | |
| 949 | void |
| 950 | vm_object_shadow(object, offset, length) |
| 951 | vm_object_t *object; /* IN/OUT */ |
| 952 | vm_offset_t *offset; /* IN/OUT */ |
| 953 | vm_size_t length; |
| 954 | { |
| 955 | register vm_object_t source; |
| 956 | register vm_object_t result; |
| 957 | |
| 958 | source = *object; |
| 959 | |
| 960 | /* |
| 961 | * Allocate a new object with the given length |
| 962 | */ |
| 963 | |
| 964 | if ((result = vm_object_allocate(length)) == NULL) |
| 965 | panic("vm_object_shadow: no object for shadowing"); |
| 966 | |
| 967 | /* |
| 968 | * The new object shadows the source object, adding |
| 969 | * a reference to it. Our caller changes his reference |
| 970 | * to point to the new object, removing a reference to |
| 971 | * the source object. Net result: no change of reference |
| 972 | * count. |
| 973 | */ |
| 974 | result->shadow = source; |
| 975 | |
| 976 | /* |
| 977 | * Store the offset into the source object, |
| 978 | * and fix up the offset into the new object. |
| 979 | */ |
| 980 | |
| 981 | result->shadow_offset = *offset; |
| 982 | |
| 983 | /* |
| 984 | * Return the new things |
| 985 | */ |
| 986 | |
| 987 | *offset = 0; |
| 988 | *object = result; |
| 989 | } |
| 990 | |
| 991 | /* |
| 992 | * Set the specified object's pager to the specified pager. |
| 993 | */ |
| 994 | |
| 995 | void |
| 996 | vm_object_setpager(object, pager, paging_offset, |
| 997 | read_only) |
| 998 | vm_object_t object; |
| 999 | vm_pager_t pager; |
| 1000 | vm_offset_t paging_offset; |
| 1001 | boolean_t read_only; |
| 1002 | { |
| 1003 | #ifdef lint |
| 1004 | read_only++; /* No longer used */ |
| 1005 | #endif lint |
| 1006 | |
| 1007 | vm_object_lock(object); /* XXX ? */ |
| 1008 | if (object->pager && object->pager != pager) { |
| 1009 | panic("!!!pager already allocated!!!\n"); |
| 1010 | } |
| 1011 | object->pager = pager; |
| 1012 | object->paging_offset = paging_offset; |
| 1013 | vm_object_unlock(object); /* XXX ? */ |
| 1014 | } |
| 1015 | |
| 1016 | /* |
| 1017 | * vm_object_hash hashes the pager/id pair. |
| 1018 | */ |
| 1019 | |
| 1020 | #define vm_object_hash(pager) \ |
| 1021 | ((((unsigned)pager) >> 5)%VM_OBJECT_HASH_COUNT) |
| 1022 | |
| 1023 | /* |
| 1024 | * vm_object_lookup looks in the object cache for an object with the |
| 1025 | * specified pager and paging id. |
| 1026 | */ |
| 1027 | |
| 1028 | vm_object_t |
| 1029 | vm_object_lookup(pager) |
| 1030 | vm_pager_t pager; |
| 1031 | { |
| 1032 | register queue_t bucket; |
| 1033 | register vm_object_hash_entry_t entry; |
| 1034 | vm_object_t object; |
| 1035 | |
| 1036 | bucket = &vm_object_hashtable[vm_object_hash(pager)]; |
| 1037 | |
| 1038 | vm_object_cache_lock(); |
| 1039 | |
| 1040 | entry = (vm_object_hash_entry_t) queue_first(bucket); |
| 1041 | while (!queue_end(bucket, (queue_entry_t) entry)) { |
| 1042 | object = entry->object; |
| 1043 | if (object->pager == pager) { |
| 1044 | vm_object_lock(object); |
| 1045 | if (object->ref_count == 0) { |
| 1046 | queue_remove(&vm_object_cached_list, object, |
| 1047 | vm_object_t, cached_list); |
| 1048 | vm_object_cached--; |
| 1049 | } |
| 1050 | object->ref_count++; |
| 1051 | vm_object_unlock(object); |
| 1052 | vm_object_cache_unlock(); |
| 1053 | return(object); |
| 1054 | } |
| 1055 | entry = (vm_object_hash_entry_t) queue_next(&entry->hash_links); |
| 1056 | } |
| 1057 | |
| 1058 | vm_object_cache_unlock(); |
| 1059 | return(NULL); |
| 1060 | } |
| 1061 | |
| 1062 | /* |
| 1063 | * vm_object_enter enters the specified object/pager/id into |
| 1064 | * the hash table. |
| 1065 | */ |
| 1066 | |
| 1067 | void |
| 1068 | vm_object_enter(object, pager) |
| 1069 | vm_object_t object; |
| 1070 | vm_pager_t pager; |
| 1071 | { |
| 1072 | register queue_t bucket; |
| 1073 | register vm_object_hash_entry_t entry; |
| 1074 | |
| 1075 | /* |
| 1076 | * We don't cache null objects, and we can't cache |
| 1077 | * objects with the null pager. |
| 1078 | */ |
| 1079 | |
| 1080 | if (object == NULL) |
| 1081 | return; |
| 1082 | if (pager == NULL) |
| 1083 | return; |
| 1084 | |
| 1085 | bucket = &vm_object_hashtable[vm_object_hash(pager)]; |
| 1086 | entry = (vm_object_hash_entry_t) |
| 1087 | malloc((u_long)sizeof *entry, M_VMOBJHASH, M_WAITOK); |
| 1088 | entry->object = object; |
| 1089 | object->can_persist = TRUE; |
| 1090 | |
| 1091 | vm_object_cache_lock(); |
| 1092 | queue_enter(bucket, entry, vm_object_hash_entry_t, hash_links); |
| 1093 | vm_object_cache_unlock(); |
| 1094 | } |
| 1095 | |
| 1096 | /* |
| 1097 | * vm_object_remove: |
| 1098 | * |
| 1099 | * Remove the pager from the hash table. |
| 1100 | * Note: This assumes that the object cache |
| 1101 | * is locked. XXX this should be fixed |
| 1102 | * by reorganizing vm_object_deallocate. |
| 1103 | */ |
| 1104 | void |
| 1105 | vm_object_remove(pager) |
| 1106 | register vm_pager_t pager; |
| 1107 | { |
| 1108 | register queue_t bucket; |
| 1109 | register vm_object_hash_entry_t entry; |
| 1110 | register vm_object_t object; |
| 1111 | |
| 1112 | bucket = &vm_object_hashtable[vm_object_hash(pager)]; |
| 1113 | |
| 1114 | entry = (vm_object_hash_entry_t) queue_first(bucket); |
| 1115 | while (!queue_end(bucket, (queue_entry_t) entry)) { |
| 1116 | object = entry->object; |
| 1117 | if (object->pager == pager) { |
| 1118 | queue_remove(bucket, entry, vm_object_hash_entry_t, |
| 1119 | hash_links); |
| 1120 | free((caddr_t)entry, M_VMOBJHASH); |
| 1121 | break; |
| 1122 | } |
| 1123 | entry = (vm_object_hash_entry_t) queue_next(&entry->hash_links); |
| 1124 | } |
| 1125 | } |
| 1126 | |
| 1127 | /* |
| 1128 | * vm_object_cache_clear removes all objects from the cache. |
| 1129 | * |
| 1130 | */ |
| 1131 | void |
| 1132 | vm_object_cache_clear() |
| 1133 | { |
| 1134 | register vm_object_t object; |
| 1135 | |
| 1136 | /* |
| 1137 | * Remove each object in the cache by scanning down the |
| 1138 | * list of cached objects. |
| 1139 | */ |
| 1140 | vm_object_cache_lock(); |
| 1141 | while (!queue_empty(&vm_object_cached_list)) { |
| 1142 | object = (vm_object_t) queue_first(&vm_object_cached_list); |
| 1143 | vm_object_cache_unlock(); |
| 1144 | |
| 1145 | /* |
| 1146 | * Note: it is important that we use vm_object_lookup |
| 1147 | * to gain a reference, and not vm_object_reference, because |
| 1148 | * the logic for removing an object from the cache lies in |
| 1149 | * lookup. |
| 1150 | */ |
| 1151 | if (object != vm_object_lookup(object->pager)) |
| 1152 | panic("vm_object_cache_clear: I'm sooo confused."); |
| 1153 | pager_cache(object, FALSE); |
| 1154 | |
| 1155 | vm_object_cache_lock(); |
| 1156 | } |
| 1157 | vm_object_cache_unlock(); |
| 1158 | } |
| 1159 | |
| 1160 | boolean_t vm_object_collapse_allowed = TRUE; |
| 1161 | /* |
| 1162 | * vm_object_collapse: |
| 1163 | * |
| 1164 | * Collapse an object with the object backing it. |
| 1165 | * Pages in the backing object are moved into the |
| 1166 | * parent, and the backing object is deallocated. |
| 1167 | * |
| 1168 | * Requires that the object be locked and the page |
| 1169 | * queues be unlocked. |
| 1170 | * |
| 1171 | * This routine has significant changes by John S. Dyson |
| 1172 | * to fix some swap memory leaks. 18 Dec 93 |
| 1173 | * |
| 1174 | */ |
| 1175 | void |
| 1176 | vm_object_collapse(object) |
| 1177 | register vm_object_t object; |
| 1178 | |
| 1179 | { |
| 1180 | register vm_object_t backing_object; |
| 1181 | register vm_offset_t backing_offset; |
| 1182 | register vm_size_t size; |
| 1183 | register vm_offset_t new_offset; |
| 1184 | register vm_page_t p, pp; |
| 1185 | |
| 1186 | if (!vm_object_collapse_allowed) |
| 1187 | return; |
| 1188 | |
| 1189 | while (TRUE) { |
| 1190 | /* |
| 1191 | * Verify that the conditions are right for collapse: |
| 1192 | * |
| 1193 | * The object exists and no pages in it are currently |
| 1194 | * being paged out. |
| 1195 | */ |
| 1196 | if (object == NULL || |
| 1197 | object->paging_in_progress != 0) |
| 1198 | return; |
| 1199 | |
| 1200 | /* |
| 1201 | * There is a backing object, and |
| 1202 | */ |
| 1203 | |
| 1204 | if ((backing_object = object->shadow) == NULL) |
| 1205 | return; |
| 1206 | |
| 1207 | vm_object_lock(backing_object); |
| 1208 | /* |
| 1209 | * ... |
| 1210 | * The backing object is not read_only, |
| 1211 | * and no pages in the backing object are |
| 1212 | * currently being paged out. |
| 1213 | * The backing object is internal. |
| 1214 | */ |
| 1215 | |
| 1216 | if (!backing_object->internal || |
| 1217 | backing_object->paging_in_progress != 0 ) { |
| 1218 | vm_object_unlock(backing_object); |
| 1219 | return; |
| 1220 | } |
| 1221 | |
| 1222 | /* |
| 1223 | * The backing object can't be a copy-object: |
| 1224 | * the shadow_offset for the copy-object must stay |
| 1225 | * as 0. Furthermore (for the 'we have all the |
| 1226 | * pages' case), if we bypass backing_object and |
| 1227 | * just shadow the next object in the chain, old |
| 1228 | * pages from that object would then have to be copied |
| 1229 | * BOTH into the (former) backing_object and into the |
| 1230 | * parent object. |
| 1231 | */ |
| 1232 | if (backing_object->shadow != NULL && |
| 1233 | backing_object->shadow->copy == backing_object) { |
| 1234 | vm_object_unlock(backing_object); |
| 1235 | return; |
| 1236 | } |
| 1237 | |
| 1238 | /* |
| 1239 | * we can deal only with the swap pager |
| 1240 | */ |
| 1241 | if ((object->pager && |
| 1242 | object->pager->pg_type != PG_SWAP) || |
| 1243 | (backing_object->pager && |
| 1244 | backing_object->pager->pg_type != PG_SWAP)) { |
| 1245 | vm_object_unlock(backing_object); |
| 1246 | return; |
| 1247 | } |
| 1248 | |
| 1249 | |
| 1250 | /* |
| 1251 | * We know that we can either collapse the backing |
| 1252 | * object (if the parent is the only reference to |
| 1253 | * it) or (perhaps) remove the parent's reference |
| 1254 | * to it. |
| 1255 | */ |
| 1256 | |
| 1257 | backing_offset = object->shadow_offset; |
| 1258 | size = object->size; |
| 1259 | |
| 1260 | /* |
| 1261 | * If there is exactly one reference to the backing |
| 1262 | * object, we can collapse it into the parent. |
| 1263 | */ |
| 1264 | |
| 1265 | if (backing_object->ref_count == 1) { |
| 1266 | |
| 1267 | /* |
| 1268 | * We can collapse the backing object. |
| 1269 | * |
| 1270 | * Move all in-memory pages from backing_object |
| 1271 | * to the parent. Pages that have been paged out |
| 1272 | * will be overwritten by any of the parent's |
| 1273 | * pages that shadow them. |
| 1274 | */ |
| 1275 | |
| 1276 | while (!queue_empty(&backing_object->memq)) { |
| 1277 | |
| 1278 | p = (vm_page_t) |
| 1279 | queue_first(&backing_object->memq); |
| 1280 | |
| 1281 | new_offset = (p->offset - backing_offset); |
| 1282 | |
| 1283 | /* |
| 1284 | * If the parent has a page here, or if |
| 1285 | * this page falls outside the parent, |
| 1286 | * dispose of it. |
| 1287 | * |
| 1288 | * Otherwise, move it as planned. |
| 1289 | */ |
| 1290 | |
| 1291 | if (p->offset < backing_offset || |
| 1292 | new_offset >= size) { |
| 1293 | vm_page_lock_queues(); |
| 1294 | vm_page_free(p); |
| 1295 | vm_page_unlock_queues(); |
| 1296 | } else { |
| 1297 | pp = vm_page_lookup(object, new_offset); |
| 1298 | if (pp != NULL || (object->pager && vm_pager_has_page(object->pager, |
| 1299 | object->paging_offset + new_offset))) { |
| 1300 | vm_page_lock_queues(); |
| 1301 | vm_page_free(p); |
| 1302 | vm_page_unlock_queues(); |
| 1303 | } else { |
| 1304 | vm_page_rename(p, object, new_offset); |
| 1305 | } |
| 1306 | } |
| 1307 | } |
| 1308 | |
| 1309 | /* |
| 1310 | * Move the pager from backing_object to object. |
| 1311 | */ |
| 1312 | |
| 1313 | if (backing_object->pager) { |
| 1314 | backing_object->paging_in_progress++; |
| 1315 | if (object->pager) { |
| 1316 | vm_pager_t bopager; |
| 1317 | object->paging_in_progress++; |
| 1318 | /* |
| 1319 | * copy shadow object pages into ours |
| 1320 | * and destroy unneeded pages in shadow object. |
| 1321 | */ |
| 1322 | bopager = backing_object->pager; |
| 1323 | backing_object->pager = NULL; |
| 1324 | vm_object_remove(backing_object->pager); |
| 1325 | swap_pager_copy( |
| 1326 | bopager, backing_object->paging_offset, |
| 1327 | object->pager, object->paging_offset, |
| 1328 | object->shadow_offset); |
| 1329 | object->paging_in_progress--; |
| 1330 | if (object->paging_in_progress == 0) |
| 1331 | wakeup((caddr_t)object); |
| 1332 | } else { |
| 1333 | object->paging_in_progress++; |
| 1334 | /* |
| 1335 | * grab the shadow objects pager |
| 1336 | */ |
| 1337 | object->pager = backing_object->pager; |
| 1338 | object->paging_offset = backing_object->paging_offset + backing_offset; |
| 1339 | vm_object_remove(backing_object->pager); |
| 1340 | backing_object->pager = NULL; |
| 1341 | /* |
| 1342 | * free unnecessary blocks |
| 1343 | */ |
| 1344 | swap_pager_freespace(object->pager, 0, object->paging_offset); |
| 1345 | object->paging_in_progress--; |
| 1346 | if (object->paging_in_progress == 0) |
| 1347 | wakeup((caddr_t)object); |
| 1348 | } |
| 1349 | backing_object->paging_in_progress--; |
| 1350 | if (backing_object->paging_in_progress == 0) |
| 1351 | wakeup((caddr_t)backing_object); |
| 1352 | } |
| 1353 | |
| 1354 | |
| 1355 | /* |
| 1356 | * Object now shadows whatever backing_object did. |
| 1357 | * Note that the reference to backing_object->shadow |
| 1358 | * moves from within backing_object to within object. |
| 1359 | */ |
| 1360 | |
| 1361 | object->shadow = backing_object->shadow; |
| 1362 | object->shadow_offset += backing_object->shadow_offset; |
| 1363 | if (object->shadow != NULL && |
| 1364 | object->shadow->copy != NULL) { |
| 1365 | panic("vm_object_collapse: we collapsed a copy-object!"); |
| 1366 | } |
| 1367 | /* |
| 1368 | * Discard backing_object. |
| 1369 | * |
| 1370 | * Since the backing object has no pages, no |
| 1371 | * pager left, and no object references within it, |
| 1372 | * all that is necessary is to dispose of it. |
| 1373 | */ |
| 1374 | |
| 1375 | vm_object_unlock(backing_object); |
| 1376 | |
| 1377 | simple_lock(&vm_object_list_lock); |
| 1378 | queue_remove(&vm_object_list, backing_object, |
| 1379 | vm_object_t, object_list); |
| 1380 | vm_object_count--; |
| 1381 | simple_unlock(&vm_object_list_lock); |
| 1382 | |
| 1383 | free((caddr_t)backing_object, M_VMOBJ); |
| 1384 | |
| 1385 | object_collapses++; |
| 1386 | } |
| 1387 | else { |
| 1388 | /* |
| 1389 | * If all of the pages in the backing object are |
| 1390 | * shadowed by the parent object, the parent |
| 1391 | * object no longer has to shadow the backing |
| 1392 | * object; it can shadow the next one in the |
| 1393 | * chain. |
| 1394 | * |
| 1395 | * The backing object must not be paged out - we'd |
| 1396 | * have to check all of the paged-out pages, as |
| 1397 | * well. |
| 1398 | */ |
| 1399 | |
| 1400 | if (backing_object->pager != NULL) { |
| 1401 | vm_object_unlock(backing_object); |
| 1402 | return; |
| 1403 | } |
| 1404 | |
| 1405 | /* |
| 1406 | * Should have a check for a 'small' number |
| 1407 | * of pages here. |
| 1408 | */ |
| 1409 | |
| 1410 | p = (vm_page_t) queue_first(&backing_object->memq); |
| 1411 | while (!queue_end(&backing_object->memq, |
| 1412 | (queue_entry_t) p)) { |
| 1413 | |
| 1414 | new_offset = (p->offset - backing_offset); |
| 1415 | |
| 1416 | /* |
| 1417 | * If the parent has a page here, or if |
| 1418 | * this page falls outside the parent, |
| 1419 | * keep going. |
| 1420 | * |
| 1421 | * Otherwise, the backing_object must be |
| 1422 | * left in the chain. |
| 1423 | */ |
| 1424 | |
| 1425 | if (p->offset >= backing_offset && |
| 1426 | new_offset <= size && |
| 1427 | ((pp = vm_page_lookup(object, new_offset)) == NULL || (pp->flags & PG_FAKE)) && |
| 1428 | (!object->pager || !vm_pager_has_page(object->pager, object->paging_offset+new_offset))) { |
| 1429 | /* |
| 1430 | * Page still needed. |
| 1431 | * Can't go any further. |
| 1432 | */ |
| 1433 | vm_object_unlock(backing_object); |
| 1434 | return; |
| 1435 | } |
| 1436 | p = (vm_page_t) queue_next(&p->listq); |
| 1437 | } |
| 1438 | |
| 1439 | /* |
| 1440 | * Make the parent shadow the next object |
| 1441 | * in the chain. Deallocating backing_object |
| 1442 | * will not remove it, since its reference |
| 1443 | * count is at least 2. |
| 1444 | */ |
| 1445 | |
| 1446 | vm_object_reference(object->shadow = backing_object->shadow); |
| 1447 | object->shadow_offset += backing_object->shadow_offset; |
| 1448 | |
| 1449 | #if 1 |
| 1450 | /* Mach 3.0 code */ |
| 1451 | /* andrew@werple.apana.org.au, 12 Feb 1993 */ |
| 1452 | |
| 1453 | /* |
| 1454 | * Backing object might have had a copy pointer |
| 1455 | * to us. If it did, clear it. |
| 1456 | */ |
| 1457 | if (backing_object->copy == object) |
| 1458 | backing_object->copy = NULL; |
| 1459 | #endif |
| 1460 | |
| 1461 | /* Drop the reference count on backing_object. |
| 1462 | * Since its ref_count was at least 2, it |
| 1463 | * will not vanish; so we don't need to call |
| 1464 | * vm_object_deallocate. |
| 1465 | */ |
| 1466 | if (backing_object->ref_count == 1) |
| 1467 | printf("should have called obj deallocate\n"); |
| 1468 | backing_object->ref_count--; |
| 1469 | vm_object_unlock(backing_object); |
| 1470 | |
| 1471 | object_bypasses ++; |
| 1472 | |
| 1473 | } |
| 1474 | |
| 1475 | /* |
| 1476 | * Try again with this object's new backing object. |
| 1477 | */ |
| 1478 | } |
| 1479 | } |
| 1480 | |
| 1481 | /* |
| 1482 | * vm_object_page_remove: [internal] |
| 1483 | * |
| 1484 | * Removes all physical pages in the specified |
| 1485 | * object range from the object's list of pages. |
| 1486 | * |
| 1487 | * The object must be locked. |
| 1488 | */ |
| 1489 | void |
| 1490 | vm_object_page_remove(object, start, end) |
| 1491 | register vm_object_t object; |
| 1492 | register vm_offset_t start; |
| 1493 | register vm_offset_t end; |
| 1494 | { |
| 1495 | register vm_page_t p, next; |
| 1496 | vm_offset_t size; |
| 1497 | int cnt; |
| 1498 | int s; |
| 1499 | |
| 1500 | if (object == NULL) |
| 1501 | return; |
| 1502 | |
| 1503 | start = trunc_page(start); |
| 1504 | end = round_page(end); |
| 1505 | again: |
| 1506 | size = end-start; |
| 1507 | if (size > 4*PAGE_SIZE || size >= object->size/4) { |
| 1508 | p = (vm_page_t) queue_first(&object->memq); |
| 1509 | while (!queue_end(&object->memq, (queue_entry_t) p) && size > 0) { |
| 1510 | next = (vm_page_t) queue_next(&p->listq); |
| 1511 | if ((start <= p->offset) && (p->offset < end)) { |
| 1512 | if (p->flags & PG_BUSY) { |
| 1513 | p->flags |= PG_WANTED; |
| 1514 | tsleep((caddr_t) p, PVM, "vmopar", 0); |
| 1515 | goto again; |
| 1516 | } |
| 1517 | pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE); |
| 1518 | vm_page_lock_queues(); |
| 1519 | vm_page_free(p); |
| 1520 | vm_page_unlock_queues(); |
| 1521 | size -= PAGE_SIZE; |
| 1522 | } |
| 1523 | p = next; |
| 1524 | } |
| 1525 | } else { |
| 1526 | while (size > 0) { |
| 1527 | while (p = vm_page_lookup(object, start)) { |
| 1528 | if (p->flags & PG_BUSY) { |
| 1529 | p->flags |= PG_WANTED; |
| 1530 | tsleep((caddr_t) p, PVM, "vmopar", 0); |
| 1531 | goto again; |
| 1532 | } |
| 1533 | pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE); |
| 1534 | vm_page_lock_queues(); |
| 1535 | vm_page_free(p); |
| 1536 | vm_page_unlock_queues(); |
| 1537 | } |
| 1538 | start += PAGE_SIZE; |
| 1539 | size -= PAGE_SIZE; |
| 1540 | } |
| 1541 | } |
| 1542 | } |
| 1543 | |
| 1544 | /* |
| 1545 | * Routine: vm_object_coalesce |
| 1546 | * Function: Coalesces two objects backing up adjoining |
| 1547 | * regions of memory into a single object. |
| 1548 | * |
| 1549 | * returns TRUE if objects were combined. |
| 1550 | * |
| 1551 | * NOTE: Only works at the moment if the second object is NULL - |
| 1552 | * if it's not, which object do we lock first? |
| 1553 | * |
| 1554 | * Parameters: |
| 1555 | * prev_object First object to coalesce |
| 1556 | * prev_offset Offset into prev_object |
| 1557 | * next_object Second object into coalesce |
| 1558 | * next_offset Offset into next_object |
| 1559 | * |
| 1560 | * prev_size Size of reference to prev_object |
| 1561 | * next_size Size of reference to next_object |
| 1562 | * |
| 1563 | * Conditions: |
| 1564 | * The object must *not* be locked. |
| 1565 | */ |
| 1566 | boolean_t |
| 1567 | vm_object_coalesce(prev_object, next_object, prev_offset, next_offset, prev_size, next_size) |
| 1568 | register vm_object_t prev_object; |
| 1569 | vm_object_t next_object; |
| 1570 | vm_offset_t prev_offset, next_offset; |
| 1571 | vm_size_t prev_size, next_size; |
| 1572 | { |
| 1573 | vm_size_t newsize; |
| 1574 | |
| 1575 | #ifdef lint |
| 1576 | next_offset++; |
| 1577 | #endif lint |
| 1578 | |
| 1579 | if (next_object != NULL) { |
| 1580 | return(FALSE); |
| 1581 | } |
| 1582 | |
| 1583 | if (prev_object == NULL) { |
| 1584 | return(TRUE); |
| 1585 | } |
| 1586 | |
| 1587 | vm_object_lock(prev_object); |
| 1588 | |
| 1589 | /* |
| 1590 | * Try to collapse the object first |
| 1591 | */ |
| 1592 | vm_object_collapse(prev_object); |
| 1593 | |
| 1594 | /* |
| 1595 | * Can't coalesce if: |
| 1596 | * . more than one reference |
| 1597 | * . paged out |
| 1598 | * . shadows another object |
| 1599 | * . has a copy elsewhere |
| 1600 | * (any of which mean that the pages not mapped to |
| 1601 | * prev_entry may be in use anyway) |
| 1602 | */ |
| 1603 | if (prev_object->ref_count > 1 || |
| 1604 | prev_object->pager != NULL || |
| 1605 | prev_object->shadow != NULL || |
| 1606 | prev_object->copy != NULL) { |
| 1607 | vm_object_unlock(prev_object); |
| 1608 | return(FALSE); |
| 1609 | } |
| 1610 | |
| 1611 | /* |
| 1612 | * Remove any pages that may still be in the object from |
| 1613 | * a previous deallocation. |
| 1614 | */ |
| 1615 | vm_object_page_remove(prev_object, |
| 1616 | prev_offset + prev_size, |
| 1617 | prev_offset + prev_size + next_size); |
| 1618 | |
| 1619 | /* |
| 1620 | * Extend the object if necessary. |
| 1621 | */ |
| 1622 | newsize = prev_offset + prev_size + next_size; |
| 1623 | if (newsize > prev_object->size) |
| 1624 | prev_object->size = newsize; |
| 1625 | |
| 1626 | vm_object_unlock(prev_object); |
| 1627 | return(TRUE); |
| 1628 | } |
| 1629 | |
| 1630 | /* |
| 1631 | * returns page after looking up in shadow chain |
| 1632 | */ |
| 1633 | |
| 1634 | vm_page_t |
| 1635 | vm_object_page_lookup(object, offset) |
| 1636 | vm_object_t object; |
| 1637 | vm_offset_t offset; |
| 1638 | { |
| 1639 | vm_page_t m; |
| 1640 | if (!(m=vm_page_lookup(object, offset))) { |
| 1641 | if (!object->shadow) |
| 1642 | return 0; |
| 1643 | else |
| 1644 | return vm_object_page_lookup(object->shadow, offset + object->shadow_offset); |
| 1645 | } |
| 1646 | return m; |
| 1647 | } |
| 1648 | |
| 1649 | #if defined(DEBUG) || (NDDB > 0) |
| 1650 | /* |
| 1651 | * vm_object_print: [ debug ] |
| 1652 | */ |
| 1653 | void |
| 1654 | vm_object_print(object, full) |
| 1655 | vm_object_t object; |
| 1656 | boolean_t full; |
| 1657 | { |
| 1658 | register vm_page_t p; |
| 1659 | extern indent; |
| 1660 | |
| 1661 | register int count; |
| 1662 | |
| 1663 | if (object == NULL) |
| 1664 | return; |
| 1665 | |
| 1666 | iprintf("Object 0x%x: size=0x%x, res=%d, ref=%d, ", |
| 1667 | (int) object, (int) object->size, |
| 1668 | object->resident_page_count, object->ref_count); |
| 1669 | printf("pager=0x%x+0x%x, shadow=(0x%x)+0x%x\n", |
| 1670 | (int) object->pager, (int) object->paging_offset, |
| 1671 | (int) object->shadow, (int) object->shadow_offset); |
| 1672 | printf("cache: next=0x%x, prev=0x%x\n", |
| 1673 | object->cached_list.next, object->cached_list.prev); |
| 1674 | |
| 1675 | if (!full) |
| 1676 | return; |
| 1677 | |
| 1678 | indent += 2; |
| 1679 | count = 0; |
| 1680 | p = (vm_page_t) queue_first(&object->memq); |
| 1681 | while (!queue_end(&object->memq, (queue_entry_t) p)) { |
| 1682 | if (count == 0) |
| 1683 | iprintf("memory:="); |
| 1684 | else if (count == 6) { |
| 1685 | printf("\n"); |
| 1686 | iprintf(" ..."); |
| 1687 | count = 0; |
| 1688 | } else |
| 1689 | printf(","); |
| 1690 | count++; |
| 1691 | |
| 1692 | printf("(off=0x%x,page=0x%x)", p->offset, VM_PAGE_TO_PHYS(p)); |
| 1693 | p = (vm_page_t) queue_next(&p->listq); |
| 1694 | } |
| 1695 | if (count != 0) |
| 1696 | printf("\n"); |
| 1697 | indent -= 2; |
| 1698 | } |
| 1699 | #endif /* defined(DEBUG) || (NDDB > 0) */ |