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_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 38 * $DragonFly: src/sys/vm/vm_page.c,v 1.30 2005/06/02 20:57:21 swildner Exp $ 39 */ 40 41 /* 42 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 43 * All rights reserved. 44 * 45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 46 * 47 * Permission to use, copy, modify and distribute this software and 48 * its documentation is hereby granted, provided that both the copyright 49 * notice and this permission notice appear in all copies of the 50 * software, derivative works or modified versions, and any portions 51 * thereof, and that both notices appear in supporting documentation. 52 * 53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 56 * 57 * Carnegie Mellon requests users of this software to return to 58 * 59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 60 * School of Computer Science 61 * Carnegie Mellon University 62 * Pittsburgh PA 15213-3890 63 * 64 * any improvements or extensions that they make and grant Carnegie the 65 * rights to redistribute these changes. 66 */ 67 /* 68 * Resident memory management module. The module manipulates 'VM pages'. 69 * A VM page is the core building block for memory management. 70 */ 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/malloc.h> 75 #include <sys/proc.h> 76 #include <sys/vmmeter.h> 77 #include <sys/vnode.h> 78 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <sys/lock.h> 82 #include <vm/vm_kern.h> 83 #include <vm/pmap.h> 84 #include <vm/vm_map.h> 85 #include <vm/vm_object.h> 86 #include <vm/vm_page.h> 87 #include <vm/vm_pageout.h> 88 #include <vm/vm_pager.h> 89 #include <vm/vm_extern.h> 90 #include <vm/vm_page2.h> 91 92 static void vm_page_queue_init(void); 93 static void vm_page_free_wakeup(void); 94 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 95 static vm_page_t _vm_page_list_find2(int basequeue, int index); 96 97 static int vm_page_bucket_count; /* How big is array? */ 98 static int vm_page_hash_mask; /* Mask for hash function */ 99 static struct vm_page **vm_page_buckets; /* Array of buckets */ 100 static volatile int vm_page_bucket_generation; 101 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 102 103 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread)); 104 105 static void 106 vm_page_queue_init(void) 107 { 108 int i; 109 110 for (i = 0; i < PQ_L2_SIZE; i++) 111 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 112 for (i = 0; i < PQ_L2_SIZE; i++) 113 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 114 115 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 116 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 117 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 118 /* PQ_NONE has no queue */ 119 120 for (i = 0; i < PQ_COUNT; i++) 121 TAILQ_INIT(&vm_page_queues[i].pl); 122 } 123 124 /* 125 * note: place in initialized data section? Is this necessary? 126 */ 127 long first_page = 0; 128 int vm_page_array_size = 0; 129 int vm_page_zero_count = 0; 130 vm_page_t vm_page_array = 0; 131 132 /* 133 * (low level boot) 134 * 135 * Sets the page size, perhaps based upon the memory size. 136 * Must be called before any use of page-size dependent functions. 137 */ 138 void 139 vm_set_page_size(void) 140 { 141 if (vmstats.v_page_size == 0) 142 vmstats.v_page_size = PAGE_SIZE; 143 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 144 panic("vm_set_page_size: page size not a power of two"); 145 } 146 147 /* 148 * (low level boot) 149 * 150 * Add a new page to the freelist for use by the system. New pages 151 * are added to both the head and tail of the associated free page 152 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 153 * requests pull 'recent' adds (higher physical addresses) first. 154 * 155 * Must be called in a critical section. 156 */ 157 vm_page_t 158 vm_add_new_page(vm_paddr_t pa) 159 { 160 struct vpgqueues *vpq; 161 vm_page_t m; 162 163 ++vmstats.v_page_count; 164 ++vmstats.v_free_count; 165 m = PHYS_TO_VM_PAGE(pa); 166 m->phys_addr = pa; 167 m->flags = 0; 168 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 169 m->queue = m->pc + PQ_FREE; 170 171 vpq = &vm_page_queues[m->queue]; 172 if (vpq->flipflop) 173 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 174 else 175 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 176 vpq->flipflop = 1 - vpq->flipflop; 177 178 vm_page_queues[m->queue].lcnt++; 179 return (m); 180 } 181 182 /* 183 * (low level boot) 184 * 185 * Initializes the resident memory module. 186 * 187 * Allocates memory for the page cells, and for the object/offset-to-page 188 * hash table headers. Each page cell is initialized and placed on the 189 * free list. 190 */ 191 vm_offset_t 192 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 193 { 194 vm_offset_t mapped; 195 struct vm_page **bucket; 196 vm_size_t npages; 197 vm_paddr_t page_range; 198 vm_paddr_t new_end; 199 int i; 200 vm_paddr_t pa; 201 int nblocks; 202 vm_paddr_t last_pa; 203 204 /* the biggest memory array is the second group of pages */ 205 vm_paddr_t end; 206 vm_paddr_t biggestone, biggestsize; 207 208 vm_paddr_t total; 209 210 total = 0; 211 biggestsize = 0; 212 biggestone = 0; 213 nblocks = 0; 214 vaddr = round_page(vaddr); 215 216 for (i = 0; phys_avail[i + 1]; i += 2) { 217 phys_avail[i] = round_page(phys_avail[i]); 218 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 219 } 220 221 for (i = 0; phys_avail[i + 1]; i += 2) { 222 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 223 224 if (size > biggestsize) { 225 biggestone = i; 226 biggestsize = size; 227 } 228 ++nblocks; 229 total += size; 230 } 231 232 end = phys_avail[biggestone+1]; 233 234 /* 235 * Initialize the queue headers for the free queue, the active queue 236 * and the inactive queue. 237 */ 238 239 vm_page_queue_init(); 240 241 /* 242 * Allocate (and initialize) the hash table buckets. 243 * 244 * The number of buckets MUST BE a power of 2, and the actual value is 245 * the next power of 2 greater than the number of physical pages in 246 * the system. 247 * 248 * We make the hash table approximately 2x the number of pages to 249 * reduce the chain length. This is about the same size using the 250 * singly-linked list as the 1x hash table we were using before 251 * using TAILQ but the chain length will be smaller. 252 * 253 * Note: This computation can be tweaked if desired. 254 */ 255 vm_page_buckets = (struct vm_page **)vaddr; 256 bucket = vm_page_buckets; 257 if (vm_page_bucket_count == 0) { 258 vm_page_bucket_count = 1; 259 while (vm_page_bucket_count < atop(total)) 260 vm_page_bucket_count <<= 1; 261 } 262 vm_page_bucket_count <<= 1; 263 vm_page_hash_mask = vm_page_bucket_count - 1; 264 265 /* 266 * Validate these addresses. 267 */ 268 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 269 new_end = trunc_page(new_end); 270 mapped = round_page(vaddr); 271 vaddr = pmap_map(mapped, new_end, end, 272 VM_PROT_READ | VM_PROT_WRITE); 273 vaddr = round_page(vaddr); 274 bzero((caddr_t) mapped, vaddr - mapped); 275 276 for (i = 0; i < vm_page_bucket_count; i++) { 277 *bucket = NULL; 278 bucket++; 279 } 280 281 /* 282 * Compute the number of pages of memory that will be available for 283 * use (taking into account the overhead of a page structure per 284 * page). 285 */ 286 first_page = phys_avail[0] / PAGE_SIZE; 287 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 288 npages = (total - (page_range * sizeof(struct vm_page)) - 289 (end - new_end)) / PAGE_SIZE; 290 291 end = new_end; 292 293 /* 294 * Initialize the mem entry structures now, and put them in the free 295 * queue. 296 */ 297 vm_page_array = (vm_page_t) vaddr; 298 mapped = vaddr; 299 300 /* 301 * Validate these addresses. 302 */ 303 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 304 mapped = pmap_map(mapped, new_end, end, 305 VM_PROT_READ | VM_PROT_WRITE); 306 307 /* 308 * Clear all of the page structures 309 */ 310 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 311 vm_page_array_size = page_range; 312 313 /* 314 * Construct the free queue(s) in ascending order (by physical 315 * address) so that the first 16MB of physical memory is allocated 316 * last rather than first. On large-memory machines, this avoids 317 * the exhaustion of low physical memory before isa_dmainit has run. 318 */ 319 vmstats.v_page_count = 0; 320 vmstats.v_free_count = 0; 321 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 322 pa = phys_avail[i]; 323 if (i == biggestone) 324 last_pa = new_end; 325 else 326 last_pa = phys_avail[i + 1]; 327 while (pa < last_pa && npages-- > 0) { 328 vm_add_new_page(pa); 329 pa += PAGE_SIZE; 330 } 331 } 332 return (mapped); 333 } 334 335 /* 336 * Distributes the object/offset key pair among hash buckets. 337 * 338 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 339 * This routine may not block. 340 * 341 * We try to randomize the hash based on the object to spread the pages 342 * out in the hash table without it costing us too much. 343 */ 344 static __inline int 345 vm_page_hash(vm_object_t object, vm_pindex_t pindex) 346 { 347 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 348 349 return(i & vm_page_hash_mask); 350 } 351 352 /* 353 * The opposite of vm_page_hold(). A page can be freed while being held, 354 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 355 * in this case to actually free it once the hold count drops to 0. 356 * 357 * This routine must be called at splvm(). 358 */ 359 void 360 vm_page_unhold(vm_page_t mem) 361 { 362 --mem->hold_count; 363 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 364 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) { 365 vm_page_busy(mem); 366 vm_page_free_toq(mem); 367 } 368 } 369 370 /* 371 * Inserts the given mem entry into the object and object list. 372 * 373 * The pagetables are not updated but will presumably fault the page 374 * in if necessary, or if a kernel page the caller will at some point 375 * enter the page into the kernel's pmap. We are not allowed to block 376 * here so we *can't* do this anyway. 377 * 378 * This routine may not block. 379 * This routine must be called with a critical section held. 380 */ 381 void 382 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 383 { 384 struct vm_page **bucket; 385 386 ASSERT_IN_CRIT_SECTION(); 387 if (m->object != NULL) 388 panic("vm_page_insert: already inserted"); 389 390 /* 391 * Record the object/offset pair in this page 392 */ 393 m->object = object; 394 m->pindex = pindex; 395 396 /* 397 * Insert it into the object_object/offset hash table 398 */ 399 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 400 m->hnext = *bucket; 401 *bucket = m; 402 vm_page_bucket_generation++; 403 404 /* 405 * Now link into the object's list of backed pages. 406 */ 407 TAILQ_INSERT_TAIL(&object->memq, m, listq); 408 object->generation++; 409 410 /* 411 * show that the object has one more resident page. 412 */ 413 object->resident_page_count++; 414 415 /* 416 * Since we are inserting a new and possibly dirty page, 417 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 418 */ 419 if (m->flags & PG_WRITEABLE) 420 vm_object_set_writeable_dirty(object); 421 } 422 423 /* 424 * Removes the given vm_page_t from the global (object,index) hash table 425 * and from the object's memq. 426 * 427 * The underlying pmap entry (if any) is NOT removed here. 428 * This routine may not block. 429 * 430 * The page must be BUSY and will remain BUSY on return. No spl needs to be 431 * held on call to this routine. 432 * 433 * note: FreeBSD side effect was to unbusy the page on return. We leave 434 * it busy. 435 */ 436 void 437 vm_page_remove(vm_page_t m) 438 { 439 vm_object_t object; 440 struct vm_page **bucket; 441 442 crit_enter(); 443 if (m->object == NULL) { 444 crit_exit(); 445 return; 446 } 447 448 if ((m->flags & PG_BUSY) == 0) 449 panic("vm_page_remove: page not busy"); 450 451 object = m->object; 452 453 /* 454 * Remove from the object_object/offset hash table. The object 455 * must be on the hash queue, we will panic if it isn't 456 * 457 * Note: we must NULL-out m->hnext to prevent loops in detached 458 * buffers with vm_page_lookup(). 459 */ 460 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 461 while (*bucket != m) { 462 if (*bucket == NULL) 463 panic("vm_page_remove(): page not found in hash"); 464 bucket = &(*bucket)->hnext; 465 } 466 *bucket = m->hnext; 467 m->hnext = NULL; 468 vm_page_bucket_generation++; 469 470 /* 471 * Now remove from the object's list of backed pages. 472 */ 473 TAILQ_REMOVE(&object->memq, m, listq); 474 475 /* 476 * And show that the object has one fewer resident page. 477 */ 478 object->resident_page_count--; 479 object->generation++; 480 481 m->object = NULL; 482 crit_exit(); 483 } 484 485 /* 486 * Locate and return the page at (object, pindex), or NULL if the 487 * page could not be found. 488 * 489 * This routine will operate properly without spl protection, but 490 * the returned page could be in flux if it is busy. Because an 491 * interrupt can race a caller's busy check (unbusying and freeing the 492 * page we return before the caller is able to check the busy bit), 493 * the caller should generally call this routine with a critical 494 * section held. 495 * 496 * Callers may call this routine without spl protection if they know 497 * 'for sure' that the page will not be ripped out from under them 498 * by an interrupt. 499 */ 500 vm_page_t 501 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 502 { 503 vm_page_t m; 504 struct vm_page **bucket; 505 int generation; 506 507 /* 508 * Search the hash table for this object/offset pair 509 */ 510 retry: 511 generation = vm_page_bucket_generation; 512 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 513 for (m = *bucket; m != NULL; m = m->hnext) { 514 if ((m->object == object) && (m->pindex == pindex)) { 515 if (vm_page_bucket_generation != generation) 516 goto retry; 517 return (m); 518 } 519 } 520 if (vm_page_bucket_generation != generation) 521 goto retry; 522 return (NULL); 523 } 524 525 /* 526 * vm_page_rename() 527 * 528 * Move the given memory entry from its current object to the specified 529 * target object/offset. 530 * 531 * The object must be locked. 532 * This routine may not block. 533 * 534 * Note: This routine will raise itself to splvm(), the caller need not. 535 * 536 * Note: Swap associated with the page must be invalidated by the move. We 537 * have to do this for several reasons: (1) we aren't freeing the 538 * page, (2) we are dirtying the page, (3) the VM system is probably 539 * moving the page from object A to B, and will then later move 540 * the backing store from A to B and we can't have a conflict. 541 * 542 * Note: We *always* dirty the page. It is necessary both for the 543 * fact that we moved it, and because we may be invalidating 544 * swap. If the page is on the cache, we have to deactivate it 545 * or vm_page_dirty() will panic. Dirty pages are not allowed 546 * on the cache. 547 */ 548 void 549 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 550 { 551 crit_enter(); 552 vm_page_remove(m); 553 vm_page_insert(m, new_object, new_pindex); 554 if (m->queue - m->pc == PQ_CACHE) 555 vm_page_deactivate(m); 556 vm_page_dirty(m); 557 vm_page_wakeup(m); 558 crit_exit(); 559 } 560 561 /* 562 * vm_page_unqueue() without any wakeup. This routine is used when a page 563 * is being moved between queues or otherwise is to remain BUSYied by the 564 * caller. 565 * 566 * This routine must be called at splhigh(). 567 * This routine may not block. 568 */ 569 void 570 vm_page_unqueue_nowakeup(vm_page_t m) 571 { 572 int queue = m->queue; 573 struct vpgqueues *pq; 574 575 if (queue != PQ_NONE) { 576 pq = &vm_page_queues[queue]; 577 m->queue = PQ_NONE; 578 TAILQ_REMOVE(&pq->pl, m, pageq); 579 (*pq->cnt)--; 580 pq->lcnt--; 581 } 582 } 583 584 /* 585 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 586 * if necessary. 587 * 588 * This routine must be called at splhigh(). 589 * This routine may not block. 590 */ 591 void 592 vm_page_unqueue(vm_page_t m) 593 { 594 int queue = m->queue; 595 struct vpgqueues *pq; 596 597 if (queue != PQ_NONE) { 598 m->queue = PQ_NONE; 599 pq = &vm_page_queues[queue]; 600 TAILQ_REMOVE(&pq->pl, m, pageq); 601 (*pq->cnt)--; 602 pq->lcnt--; 603 if ((queue - m->pc) == PQ_CACHE) { 604 if (vm_paging_needed()) 605 pagedaemon_wakeup(); 606 } 607 } 608 } 609 610 /* 611 * vm_page_list_find() 612 * 613 * Find a page on the specified queue with color optimization. 614 * 615 * The page coloring optimization attempts to locate a page that does 616 * not overload other nearby pages in the object in the cpu's L1 or L2 617 * caches. We need this optimization because cpu caches tend to be 618 * physical caches, while object spaces tend to be virtual. 619 * 620 * This routine must be called at splvm(). 621 * This routine may not block. 622 * 623 * Note that this routine is carefully inlined. A non-inlined version 624 * is available for outside callers but the only critical path is 625 * from within this source file. 626 */ 627 static __inline 628 vm_page_t 629 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 630 { 631 vm_page_t m; 632 633 if (prefer_zero) 634 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 635 else 636 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 637 if (m == NULL) 638 m = _vm_page_list_find2(basequeue, index); 639 return(m); 640 } 641 642 static vm_page_t 643 _vm_page_list_find2(int basequeue, int index) 644 { 645 int i; 646 vm_page_t m = NULL; 647 struct vpgqueues *pq; 648 649 pq = &vm_page_queues[basequeue]; 650 651 /* 652 * Note that for the first loop, index+i and index-i wind up at the 653 * same place. Even though this is not totally optimal, we've already 654 * blown it by missing the cache case so we do not care. 655 */ 656 657 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 658 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 659 break; 660 661 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 662 break; 663 } 664 return(m); 665 } 666 667 vm_page_t 668 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 669 { 670 return(_vm_page_list_find(basequeue, index, prefer_zero)); 671 } 672 673 /* 674 * Find a page on the cache queue with color optimization. As pages 675 * might be found, but not applicable, they are deactivated. This 676 * keeps us from using potentially busy cached pages. 677 * 678 * This routine must be called with a critical section held. 679 * This routine may not block. 680 */ 681 vm_page_t 682 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 683 { 684 vm_page_t m; 685 686 while (TRUE) { 687 m = _vm_page_list_find( 688 PQ_CACHE, 689 (pindex + object->pg_color) & PQ_L2_MASK, 690 FALSE 691 ); 692 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 693 m->hold_count || m->wire_count)) { 694 vm_page_deactivate(m); 695 continue; 696 } 697 return m; 698 } 699 /* not reached */ 700 } 701 702 /* 703 * Find a free or zero page, with specified preference. We attempt to 704 * inline the nominal case and fall back to _vm_page_select_free() 705 * otherwise. 706 * 707 * This routine must be called with a critical section held. 708 * This routine may not block. 709 */ 710 static __inline vm_page_t 711 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 712 { 713 vm_page_t m; 714 715 m = _vm_page_list_find( 716 PQ_FREE, 717 (pindex + object->pg_color) & PQ_L2_MASK, 718 prefer_zero 719 ); 720 return(m); 721 } 722 723 /* 724 * vm_page_alloc() 725 * 726 * Allocate and return a memory cell associated with this VM object/offset 727 * pair. 728 * 729 * page_req classes: 730 * 731 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 732 * VM_ALLOC_SYSTEM greater free drain 733 * VM_ALLOC_INTERRUPT allow free list to be completely drained 734 * VM_ALLOC_ZERO advisory request for pre-zero'd page 735 * 736 * The object must be locked. 737 * This routine may not block. 738 * The returned page will be marked PG_BUSY 739 * 740 * Additional special handling is required when called from an interrupt 741 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 742 * in this case. 743 */ 744 vm_page_t 745 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 746 { 747 vm_page_t m = NULL; 748 749 KASSERT(!vm_page_lookup(object, pindex), 750 ("vm_page_alloc: page already allocated")); 751 KKASSERT(page_req & 752 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 753 754 /* 755 * The pager is allowed to eat deeper into the free page list. 756 */ 757 if (curthread == pagethread) 758 page_req |= VM_ALLOC_SYSTEM; 759 760 crit_enter(); 761 loop: 762 if (vmstats.v_free_count > vmstats.v_free_reserved || 763 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 764 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 765 vmstats.v_free_count > vmstats.v_interrupt_free_min) 766 ) { 767 /* 768 * The free queue has sufficient free pages to take one out. 769 */ 770 if (page_req & VM_ALLOC_ZERO) 771 m = vm_page_select_free(object, pindex, TRUE); 772 else 773 m = vm_page_select_free(object, pindex, FALSE); 774 } else if (page_req & VM_ALLOC_NORMAL) { 775 /* 776 * Allocatable from the cache (non-interrupt only). On 777 * success, we must free the page and try again, thus 778 * ensuring that vmstats.v_*_free_min counters are replenished. 779 */ 780 #ifdef INVARIANTS 781 if (curthread->td_preempted) { 782 printf("vm_page_alloc(): warning, attempt to allocate" 783 " cache page from preempting interrupt\n"); 784 m = NULL; 785 } else { 786 m = vm_page_select_cache(object, pindex); 787 } 788 #else 789 m = vm_page_select_cache(object, pindex); 790 #endif 791 /* 792 * On success move the page into the free queue and loop. 793 */ 794 if (m != NULL) { 795 KASSERT(m->dirty == 0, 796 ("Found dirty cache page %p", m)); 797 vm_page_busy(m); 798 vm_page_protect(m, VM_PROT_NONE); 799 vm_page_free(m); 800 goto loop; 801 } 802 803 /* 804 * On failure return NULL 805 */ 806 crit_exit(); 807 #if defined(DIAGNOSTIC) 808 if (vmstats.v_cache_count > 0) 809 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 810 #endif 811 vm_pageout_deficit++; 812 pagedaemon_wakeup(); 813 return (NULL); 814 } else { 815 /* 816 * No pages available, wakeup the pageout daemon and give up. 817 */ 818 crit_exit(); 819 vm_pageout_deficit++; 820 pagedaemon_wakeup(); 821 return (NULL); 822 } 823 824 /* 825 * Good page found. The page has not yet been busied. We are in 826 * a critical section. 827 */ 828 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 829 830 /* 831 * Remove from free queue 832 */ 833 vm_page_unqueue_nowakeup(m); 834 835 /* 836 * Initialize structure. Only the PG_ZERO flag is inherited. Set 837 * the page PG_BUSY 838 */ 839 if (m->flags & PG_ZERO) { 840 vm_page_zero_count--; 841 m->flags = PG_ZERO | PG_BUSY; 842 } else { 843 m->flags = PG_BUSY; 844 } 845 m->wire_count = 0; 846 m->hold_count = 0; 847 m->act_count = 0; 848 m->busy = 0; 849 m->valid = 0; 850 KASSERT(m->dirty == 0, 851 ("vm_page_alloc: free/cache page %p was dirty", m)); 852 853 /* 854 * vm_page_insert() is safe prior to the crit_exit(). Note also that 855 * inserting a page here does not insert it into the pmap (which 856 * could cause us to block allocating memory). We cannot block 857 * anywhere. 858 */ 859 vm_page_insert(m, object, pindex); 860 861 /* 862 * Don't wakeup too often - wakeup the pageout daemon when 863 * we would be nearly out of memory. 864 */ 865 if (vm_paging_needed()) 866 pagedaemon_wakeup(); 867 868 crit_exit(); 869 870 /* 871 * A PG_BUSY page is returned. 872 */ 873 return (m); 874 } 875 876 /* 877 * Block until free pages are available for allocation, called in various 878 * places before memory allocations. 879 */ 880 void 881 vm_wait(void) 882 { 883 crit_enter(); 884 if (curthread == pagethread) { 885 vm_pageout_pages_needed = 1; 886 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0); 887 } else { 888 if (!vm_pages_needed) { 889 vm_pages_needed = 1; 890 wakeup(&vm_pages_needed); 891 } 892 tsleep(&vmstats.v_free_count, 0, "vmwait", 0); 893 } 894 crit_exit(); 895 } 896 897 /* 898 * Block until free pages are available for allocation 899 * 900 * Called only in vm_fault so that processes page faulting can be 901 * easily tracked. 902 * 903 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing 904 * processes will be able to grab memory first. Do not change 905 * this balance without careful testing first. 906 */ 907 void 908 vm_waitpfault(void) 909 { 910 crit_enter(); 911 if (!vm_pages_needed) { 912 vm_pages_needed = 1; 913 wakeup(&vm_pages_needed); 914 } 915 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 916 crit_exit(); 917 } 918 919 /* 920 * Put the specified page on the active list (if appropriate). Ensure 921 * that act_count is at least ACT_INIT but do not otherwise mess with it. 922 * 923 * The page queues must be locked. 924 * This routine may not block. 925 */ 926 void 927 vm_page_activate(vm_page_t m) 928 { 929 crit_enter(); 930 if (m->queue != PQ_ACTIVE) { 931 if ((m->queue - m->pc) == PQ_CACHE) 932 mycpu->gd_cnt.v_reactivated++; 933 934 vm_page_unqueue(m); 935 936 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 937 m->queue = PQ_ACTIVE; 938 vm_page_queues[PQ_ACTIVE].lcnt++; 939 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 940 m, pageq); 941 if (m->act_count < ACT_INIT) 942 m->act_count = ACT_INIT; 943 vmstats.v_active_count++; 944 } 945 } else { 946 if (m->act_count < ACT_INIT) 947 m->act_count = ACT_INIT; 948 } 949 crit_exit(); 950 } 951 952 /* 953 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 954 * routine is called when a page has been added to the cache or free 955 * queues. 956 * 957 * This routine may not block. 958 * This routine must be called at splvm() 959 */ 960 static __inline void 961 vm_page_free_wakeup(void) 962 { 963 /* 964 * if pageout daemon needs pages, then tell it that there are 965 * some free. 966 */ 967 if (vm_pageout_pages_needed && 968 vmstats.v_cache_count + vmstats.v_free_count >= 969 vmstats.v_pageout_free_min 970 ) { 971 wakeup(&vm_pageout_pages_needed); 972 vm_pageout_pages_needed = 0; 973 } 974 975 /* 976 * wakeup processes that are waiting on memory if we hit a 977 * high water mark. And wakeup scheduler process if we have 978 * lots of memory. this process will swapin processes. 979 */ 980 if (vm_pages_needed && !vm_page_count_min()) { 981 vm_pages_needed = 0; 982 wakeup(&vmstats.v_free_count); 983 } 984 } 985 986 /* 987 * vm_page_free_toq: 988 * 989 * Returns the given page to the PQ_FREE list, disassociating it with 990 * any VM object. 991 * 992 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 993 * return (the page will have been freed). No particular spl is required 994 * on entry. 995 * 996 * This routine may not block. 997 */ 998 void 999 vm_page_free_toq(vm_page_t m) 1000 { 1001 struct vpgqueues *pq; 1002 1003 crit_enter(); 1004 mycpu->gd_cnt.v_tfree++; 1005 1006 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1007 printf( 1008 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1009 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1010 m->hold_count); 1011 if ((m->queue - m->pc) == PQ_FREE) 1012 panic("vm_page_free: freeing free page"); 1013 else 1014 panic("vm_page_free: freeing busy page"); 1015 } 1016 1017 /* 1018 * unqueue, then remove page. Note that we cannot destroy 1019 * the page here because we do not want to call the pager's 1020 * callback routine until after we've put the page on the 1021 * appropriate free queue. 1022 */ 1023 vm_page_unqueue_nowakeup(m); 1024 vm_page_remove(m); 1025 1026 /* 1027 * No further management of fictitious pages occurs beyond object 1028 * and queue removal. 1029 */ 1030 if ((m->flags & PG_FICTITIOUS) != 0) { 1031 vm_page_wakeup(m); 1032 crit_exit(); 1033 return; 1034 } 1035 1036 m->valid = 0; 1037 vm_page_undirty(m); 1038 1039 if (m->wire_count != 0) { 1040 if (m->wire_count > 1) { 1041 panic( 1042 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1043 m->wire_count, (long)m->pindex); 1044 } 1045 panic("vm_page_free: freeing wired page"); 1046 } 1047 1048 /* 1049 * Clear the UNMANAGED flag when freeing an unmanaged page. 1050 */ 1051 if (m->flags & PG_UNMANAGED) { 1052 m->flags &= ~PG_UNMANAGED; 1053 } 1054 1055 if (m->hold_count != 0) { 1056 m->flags &= ~PG_ZERO; 1057 m->queue = PQ_HOLD; 1058 } else { 1059 m->queue = PQ_FREE + m->pc; 1060 } 1061 pq = &vm_page_queues[m->queue]; 1062 pq->lcnt++; 1063 ++(*pq->cnt); 1064 1065 /* 1066 * Put zero'd pages on the end ( where we look for zero'd pages 1067 * first ) and non-zerod pages at the head. 1068 */ 1069 if (m->flags & PG_ZERO) { 1070 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1071 ++vm_page_zero_count; 1072 } else { 1073 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1074 } 1075 vm_page_wakeup(m); 1076 vm_page_free_wakeup(); 1077 crit_exit(); 1078 } 1079 1080 /* 1081 * vm_page_unmanage() 1082 * 1083 * Prevent PV management from being done on the page. The page is 1084 * removed from the paging queues as if it were wired, and as a 1085 * consequence of no longer being managed the pageout daemon will not 1086 * touch it (since there is no way to locate the pte mappings for the 1087 * page). madvise() calls that mess with the pmap will also no longer 1088 * operate on the page. 1089 * 1090 * Beyond that the page is still reasonably 'normal'. Freeing the page 1091 * will clear the flag. 1092 * 1093 * This routine is used by OBJT_PHYS objects - objects using unswappable 1094 * physical memory as backing store rather then swap-backed memory and 1095 * will eventually be extended to support 4MB unmanaged physical 1096 * mappings. 1097 * 1098 * Must be called with a critical section held. 1099 */ 1100 void 1101 vm_page_unmanage(vm_page_t m) 1102 { 1103 ASSERT_IN_CRIT_SECTION(); 1104 if ((m->flags & PG_UNMANAGED) == 0) { 1105 if (m->wire_count == 0) 1106 vm_page_unqueue(m); 1107 } 1108 vm_page_flag_set(m, PG_UNMANAGED); 1109 } 1110 1111 /* 1112 * Mark this page as wired down by yet another map, removing it from 1113 * paging queues as necessary. 1114 * 1115 * The page queues must be locked. 1116 * This routine may not block. 1117 */ 1118 void 1119 vm_page_wire(vm_page_t m) 1120 { 1121 /* 1122 * Only bump the wire statistics if the page is not already wired, 1123 * and only unqueue the page if it is on some queue (if it is unmanaged 1124 * it is already off the queues). Don't do anything with fictitious 1125 * pages because they are always wired. 1126 */ 1127 crit_enter(); 1128 if ((m->flags & PG_FICTITIOUS) == 0) { 1129 if (m->wire_count == 0) { 1130 if ((m->flags & PG_UNMANAGED) == 0) 1131 vm_page_unqueue(m); 1132 vmstats.v_wire_count++; 1133 } 1134 m->wire_count++; 1135 KASSERT(m->wire_count != 0, 1136 ("vm_page_wire: wire_count overflow m=%p", m)); 1137 } 1138 vm_page_flag_set(m, PG_MAPPED); 1139 crit_exit(); 1140 } 1141 1142 /* 1143 * Release one wiring of this page, potentially enabling it to be paged again. 1144 * 1145 * Many pages placed on the inactive queue should actually go 1146 * into the cache, but it is difficult to figure out which. What 1147 * we do instead, if the inactive target is well met, is to put 1148 * clean pages at the head of the inactive queue instead of the tail. 1149 * This will cause them to be moved to the cache more quickly and 1150 * if not actively re-referenced, freed more quickly. If we just 1151 * stick these pages at the end of the inactive queue, heavy filesystem 1152 * meta-data accesses can cause an unnecessary paging load on memory bound 1153 * processes. This optimization causes one-time-use metadata to be 1154 * reused more quickly. 1155 * 1156 * BUT, if we are in a low-memory situation we have no choice but to 1157 * put clean pages on the cache queue. 1158 * 1159 * A number of routines use vm_page_unwire() to guarantee that the page 1160 * will go into either the inactive or active queues, and will NEVER 1161 * be placed in the cache - for example, just after dirtying a page. 1162 * dirty pages in the cache are not allowed. 1163 * 1164 * The page queues must be locked. 1165 * This routine may not block. 1166 */ 1167 void 1168 vm_page_unwire(vm_page_t m, int activate) 1169 { 1170 crit_enter(); 1171 if (m->flags & PG_FICTITIOUS) { 1172 /* do nothing */ 1173 } else if (m->wire_count <= 0) { 1174 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1175 } else { 1176 if (--m->wire_count == 0) { 1177 --vmstats.v_wire_count; 1178 if (m->flags & PG_UNMANAGED) { 1179 ; 1180 } else if (activate) { 1181 TAILQ_INSERT_TAIL( 1182 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1183 m->queue = PQ_ACTIVE; 1184 vm_page_queues[PQ_ACTIVE].lcnt++; 1185 vmstats.v_active_count++; 1186 } else { 1187 vm_page_flag_clear(m, PG_WINATCFLS); 1188 TAILQ_INSERT_TAIL( 1189 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1190 m->queue = PQ_INACTIVE; 1191 vm_page_queues[PQ_INACTIVE].lcnt++; 1192 vmstats.v_inactive_count++; 1193 } 1194 } 1195 } 1196 crit_exit(); 1197 } 1198 1199 1200 /* 1201 * Move the specified page to the inactive queue. If the page has 1202 * any associated swap, the swap is deallocated. 1203 * 1204 * Normally athead is 0 resulting in LRU operation. athead is set 1205 * to 1 if we want this page to be 'as if it were placed in the cache', 1206 * except without unmapping it from the process address space. 1207 * 1208 * This routine may not block. 1209 */ 1210 static __inline void 1211 _vm_page_deactivate(vm_page_t m, int athead) 1212 { 1213 /* 1214 * Ignore if already inactive. 1215 */ 1216 if (m->queue == PQ_INACTIVE) 1217 return; 1218 1219 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1220 if ((m->queue - m->pc) == PQ_CACHE) 1221 mycpu->gd_cnt.v_reactivated++; 1222 vm_page_flag_clear(m, PG_WINATCFLS); 1223 vm_page_unqueue(m); 1224 if (athead) 1225 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1226 else 1227 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1228 m->queue = PQ_INACTIVE; 1229 vm_page_queues[PQ_INACTIVE].lcnt++; 1230 vmstats.v_inactive_count++; 1231 } 1232 } 1233 1234 void 1235 vm_page_deactivate(vm_page_t m) 1236 { 1237 crit_enter(); 1238 _vm_page_deactivate(m, 0); 1239 crit_exit(); 1240 } 1241 1242 /* 1243 * vm_page_try_to_cache: 1244 * 1245 * Returns 0 on failure, 1 on success 1246 */ 1247 int 1248 vm_page_try_to_cache(vm_page_t m) 1249 { 1250 crit_enter(); 1251 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1252 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1253 return(0); 1254 } 1255 vm_page_test_dirty(m); 1256 if (m->dirty) { 1257 crit_exit(); 1258 return(0); 1259 } 1260 vm_page_cache(m); 1261 crit_exit(); 1262 return(1); 1263 } 1264 1265 /* 1266 * Attempt to free the page. If we cannot free it, we do nothing. 1267 * 1 is returned on success, 0 on failure. 1268 */ 1269 int 1270 vm_page_try_to_free(vm_page_t m) 1271 { 1272 crit_enter(); 1273 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1274 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1275 crit_exit(); 1276 return(0); 1277 } 1278 vm_page_test_dirty(m); 1279 if (m->dirty) { 1280 crit_exit(); 1281 return(0); 1282 } 1283 vm_page_busy(m); 1284 vm_page_protect(m, VM_PROT_NONE); 1285 vm_page_free(m); 1286 crit_exit(); 1287 return(1); 1288 } 1289 1290 /* 1291 * vm_page_cache 1292 * 1293 * Put the specified page onto the page cache queue (if appropriate). 1294 * 1295 * This routine may not block. 1296 */ 1297 void 1298 vm_page_cache(vm_page_t m) 1299 { 1300 ASSERT_IN_CRIT_SECTION(); 1301 1302 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1303 m->wire_count || m->hold_count) { 1304 printf("vm_page_cache: attempting to cache busy/held page\n"); 1305 return; 1306 } 1307 if ((m->queue - m->pc) == PQ_CACHE) 1308 return; 1309 1310 /* 1311 * Remove all pmaps and indicate that the page is not 1312 * writeable or mapped. 1313 */ 1314 1315 vm_page_protect(m, VM_PROT_NONE); 1316 if (m->dirty != 0) { 1317 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1318 (long)m->pindex); 1319 } 1320 vm_page_unqueue_nowakeup(m); 1321 m->queue = PQ_CACHE + m->pc; 1322 vm_page_queues[m->queue].lcnt++; 1323 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1324 vmstats.v_cache_count++; 1325 vm_page_free_wakeup(); 1326 } 1327 1328 /* 1329 * vm_page_dontneed() 1330 * 1331 * Cache, deactivate, or do nothing as appropriate. This routine 1332 * is typically used by madvise() MADV_DONTNEED. 1333 * 1334 * Generally speaking we want to move the page into the cache so 1335 * it gets reused quickly. However, this can result in a silly syndrome 1336 * due to the page recycling too quickly. Small objects will not be 1337 * fully cached. On the otherhand, if we move the page to the inactive 1338 * queue we wind up with a problem whereby very large objects 1339 * unnecessarily blow away our inactive and cache queues. 1340 * 1341 * The solution is to move the pages based on a fixed weighting. We 1342 * either leave them alone, deactivate them, or move them to the cache, 1343 * where moving them to the cache has the highest weighting. 1344 * By forcing some pages into other queues we eventually force the 1345 * system to balance the queues, potentially recovering other unrelated 1346 * space from active. The idea is to not force this to happen too 1347 * often. 1348 */ 1349 void 1350 vm_page_dontneed(vm_page_t m) 1351 { 1352 static int dnweight; 1353 int dnw; 1354 int head; 1355 1356 dnw = ++dnweight; 1357 1358 /* 1359 * occassionally leave the page alone 1360 */ 1361 crit_enter(); 1362 if ((dnw & 0x01F0) == 0 || 1363 m->queue == PQ_INACTIVE || 1364 m->queue - m->pc == PQ_CACHE 1365 ) { 1366 if (m->act_count >= ACT_INIT) 1367 --m->act_count; 1368 crit_exit(); 1369 return; 1370 } 1371 1372 if (m->dirty == 0) 1373 vm_page_test_dirty(m); 1374 1375 if (m->dirty || (dnw & 0x0070) == 0) { 1376 /* 1377 * Deactivate the page 3 times out of 32. 1378 */ 1379 head = 0; 1380 } else { 1381 /* 1382 * Cache the page 28 times out of every 32. Note that 1383 * the page is deactivated instead of cached, but placed 1384 * at the head of the queue instead of the tail. 1385 */ 1386 head = 1; 1387 } 1388 _vm_page_deactivate(m, head); 1389 crit_exit(); 1390 } 1391 1392 /* 1393 * Grab a page, blocking if it is busy and allocating a page if necessary. 1394 * A busy page is returned or NULL. 1395 * 1396 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1397 * If VM_ALLOC_RETRY is not specified 1398 * 1399 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1400 * always returned if we had blocked. 1401 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1402 * This routine may not be called from an interrupt. 1403 * The returned page may not be entirely valid. 1404 * 1405 * This routine may be called from mainline code without spl protection and 1406 * be guarenteed a busied page associated with the object at the specified 1407 * index. 1408 */ 1409 vm_page_t 1410 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1411 { 1412 vm_page_t m; 1413 int generation; 1414 1415 KKASSERT(allocflags & 1416 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1417 crit_enter(); 1418 retrylookup: 1419 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1420 if (m->busy || (m->flags & PG_BUSY)) { 1421 generation = object->generation; 1422 1423 while ((object->generation == generation) && 1424 (m->busy || (m->flags & PG_BUSY))) { 1425 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1426 tsleep(m, 0, "pgrbwt", 0); 1427 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1428 m = NULL; 1429 goto done; 1430 } 1431 } 1432 goto retrylookup; 1433 } else { 1434 vm_page_busy(m); 1435 goto done; 1436 } 1437 } 1438 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1439 if (m == NULL) { 1440 vm_wait(); 1441 if ((allocflags & VM_ALLOC_RETRY) == 0) 1442 goto done; 1443 goto retrylookup; 1444 } 1445 done: 1446 crit_exit(); 1447 return(m); 1448 } 1449 1450 /* 1451 * Mapping function for valid bits or for dirty bits in 1452 * a page. May not block. 1453 * 1454 * Inputs are required to range within a page. 1455 */ 1456 __inline int 1457 vm_page_bits(int base, int size) 1458 { 1459 int first_bit; 1460 int last_bit; 1461 1462 KASSERT( 1463 base + size <= PAGE_SIZE, 1464 ("vm_page_bits: illegal base/size %d/%d", base, size) 1465 ); 1466 1467 if (size == 0) /* handle degenerate case */ 1468 return(0); 1469 1470 first_bit = base >> DEV_BSHIFT; 1471 last_bit = (base + size - 1) >> DEV_BSHIFT; 1472 1473 return ((2 << last_bit) - (1 << first_bit)); 1474 } 1475 1476 /* 1477 * Sets portions of a page valid and clean. The arguments are expected 1478 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1479 * of any partial chunks touched by the range. The invalid portion of 1480 * such chunks will be zero'd. 1481 * 1482 * This routine may not block. 1483 * 1484 * (base + size) must be less then or equal to PAGE_SIZE. 1485 */ 1486 void 1487 vm_page_set_validclean(vm_page_t m, int base, int size) 1488 { 1489 int pagebits; 1490 int frag; 1491 int endoff; 1492 1493 if (size == 0) /* handle degenerate case */ 1494 return; 1495 1496 /* 1497 * If the base is not DEV_BSIZE aligned and the valid 1498 * bit is clear, we have to zero out a portion of the 1499 * first block. 1500 */ 1501 1502 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1503 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1504 ) { 1505 pmap_zero_page_area( 1506 VM_PAGE_TO_PHYS(m), 1507 frag, 1508 base - frag 1509 ); 1510 } 1511 1512 /* 1513 * If the ending offset is not DEV_BSIZE aligned and the 1514 * valid bit is clear, we have to zero out a portion of 1515 * the last block. 1516 */ 1517 1518 endoff = base + size; 1519 1520 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1521 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1522 ) { 1523 pmap_zero_page_area( 1524 VM_PAGE_TO_PHYS(m), 1525 endoff, 1526 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1527 ); 1528 } 1529 1530 /* 1531 * Set valid, clear dirty bits. If validating the entire 1532 * page we can safely clear the pmap modify bit. We also 1533 * use this opportunity to clear the PG_NOSYNC flag. If a process 1534 * takes a write fault on a MAP_NOSYNC memory area the flag will 1535 * be set again. 1536 * 1537 * We set valid bits inclusive of any overlap, but we can only 1538 * clear dirty bits for DEV_BSIZE chunks that are fully within 1539 * the range. 1540 */ 1541 1542 pagebits = vm_page_bits(base, size); 1543 m->valid |= pagebits; 1544 #if 0 /* NOT YET */ 1545 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1546 frag = DEV_BSIZE - frag; 1547 base += frag; 1548 size -= frag; 1549 if (size < 0) 1550 size = 0; 1551 } 1552 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1553 #endif 1554 m->dirty &= ~pagebits; 1555 if (base == 0 && size == PAGE_SIZE) { 1556 pmap_clear_modify(m); 1557 vm_page_flag_clear(m, PG_NOSYNC); 1558 } 1559 } 1560 1561 void 1562 vm_page_clear_dirty(vm_page_t m, int base, int size) 1563 { 1564 m->dirty &= ~vm_page_bits(base, size); 1565 } 1566 1567 /* 1568 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1569 * valid and dirty bits for the effected areas are cleared. 1570 * 1571 * May not block. 1572 */ 1573 void 1574 vm_page_set_invalid(vm_page_t m, int base, int size) 1575 { 1576 int bits; 1577 1578 bits = vm_page_bits(base, size); 1579 m->valid &= ~bits; 1580 m->dirty &= ~bits; 1581 m->object->generation++; 1582 } 1583 1584 /* 1585 * The kernel assumes that the invalid portions of a page contain 1586 * garbage, but such pages can be mapped into memory by user code. 1587 * When this occurs, we must zero out the non-valid portions of the 1588 * page so user code sees what it expects. 1589 * 1590 * Pages are most often semi-valid when the end of a file is mapped 1591 * into memory and the file's size is not page aligned. 1592 */ 1593 void 1594 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1595 { 1596 int b; 1597 int i; 1598 1599 /* 1600 * Scan the valid bits looking for invalid sections that 1601 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1602 * valid bit may be set ) have already been zerod by 1603 * vm_page_set_validclean(). 1604 */ 1605 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1606 if (i == (PAGE_SIZE / DEV_BSIZE) || 1607 (m->valid & (1 << i)) 1608 ) { 1609 if (i > b) { 1610 pmap_zero_page_area( 1611 VM_PAGE_TO_PHYS(m), 1612 b << DEV_BSHIFT, 1613 (i - b) << DEV_BSHIFT 1614 ); 1615 } 1616 b = i + 1; 1617 } 1618 } 1619 1620 /* 1621 * setvalid is TRUE when we can safely set the zero'd areas 1622 * as being valid. We can do this if there are no cache consistency 1623 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1624 */ 1625 if (setvalid) 1626 m->valid = VM_PAGE_BITS_ALL; 1627 } 1628 1629 /* 1630 * Is a (partial) page valid? Note that the case where size == 0 1631 * will return FALSE in the degenerate case where the page is entirely 1632 * invalid, and TRUE otherwise. 1633 * 1634 * May not block. 1635 */ 1636 int 1637 vm_page_is_valid(vm_page_t m, int base, int size) 1638 { 1639 int bits = vm_page_bits(base, size); 1640 1641 if (m->valid && ((m->valid & bits) == bits)) 1642 return 1; 1643 else 1644 return 0; 1645 } 1646 1647 /* 1648 * update dirty bits from pmap/mmu. May not block. 1649 */ 1650 void 1651 vm_page_test_dirty(vm_page_t m) 1652 { 1653 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1654 vm_page_dirty(m); 1655 } 1656 } 1657 1658 #include "opt_ddb.h" 1659 #ifdef DDB 1660 #include <sys/kernel.h> 1661 1662 #include <ddb/ddb.h> 1663 1664 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1665 { 1666 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1667 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1668 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1669 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1670 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1671 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1672 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1673 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1674 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1675 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1676 } 1677 1678 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1679 { 1680 int i; 1681 db_printf("PQ_FREE:"); 1682 for(i=0;i<PQ_L2_SIZE;i++) { 1683 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1684 } 1685 db_printf("\n"); 1686 1687 db_printf("PQ_CACHE:"); 1688 for(i=0;i<PQ_L2_SIZE;i++) { 1689 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1690 } 1691 db_printf("\n"); 1692 1693 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1694 vm_page_queues[PQ_ACTIVE].lcnt, 1695 vm_page_queues[PQ_INACTIVE].lcnt); 1696 } 1697 #endif /* DDB */ 1698