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 1.28 2004/12/10 23:07:10 dillon 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 #include <sys/thread2.h> 93 94 static void vm_page_queue_init(void); 95 static void vm_page_free_wakeup(void); 96 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 97 static vm_page_t _vm_page_list_find2(int basequeue, int index); 98 99 static int vm_page_bucket_count; /* How big is array? */ 100 static int vm_page_hash_mask; /* Mask for hash function */ 101 static struct vm_page **vm_page_buckets; /* Array of buckets */ 102 static volatile int vm_page_bucket_generation; 103 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 104 105 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread)); 106 107 static void 108 vm_page_queue_init(void) 109 { 110 int i; 111 112 for (i = 0; i < PQ_L2_SIZE; i++) 113 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 114 for (i = 0; i < PQ_L2_SIZE; i++) 115 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 116 117 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 118 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 119 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 120 /* PQ_NONE has no queue */ 121 122 for (i = 0; i < PQ_COUNT; i++) 123 TAILQ_INIT(&vm_page_queues[i].pl); 124 } 125 126 /* 127 * note: place in initialized data section? Is this necessary? 128 */ 129 long first_page = 0; 130 int vm_page_array_size = 0; 131 int vm_page_zero_count = 0; 132 vm_page_t vm_page_array = 0; 133 134 /* 135 * (low level boot) 136 * 137 * Sets the page size, perhaps based upon the memory size. 138 * Must be called before any use of page-size dependent functions. 139 */ 140 void 141 vm_set_page_size(void) 142 { 143 if (vmstats.v_page_size == 0) 144 vmstats.v_page_size = PAGE_SIZE; 145 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 146 panic("vm_set_page_size: page size not a power of two"); 147 } 148 149 /* 150 * (low level boot) 151 * 152 * Add a new page to the freelist for use by the system. New pages 153 * are added to both the head and tail of the associated free page 154 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 155 * requests pull 'recent' adds (higher physical addresses) first. 156 * 157 * Must be called in a critical section. 158 */ 159 vm_page_t 160 vm_add_new_page(vm_paddr_t pa) 161 { 162 struct vpgqueues *vpq; 163 vm_page_t m; 164 165 ++vmstats.v_page_count; 166 ++vmstats.v_free_count; 167 m = PHYS_TO_VM_PAGE(pa); 168 m->phys_addr = pa; 169 m->flags = 0; 170 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 171 m->queue = m->pc + PQ_FREE; 172 173 vpq = &vm_page_queues[m->queue]; 174 if (vpq->flipflop) 175 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 176 else 177 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 178 vpq->flipflop = 1 - vpq->flipflop; 179 180 vm_page_queues[m->queue].lcnt++; 181 return (m); 182 } 183 184 /* 185 * (low level boot) 186 * 187 * Initializes the resident memory module. 188 * 189 * Allocates memory for the page cells, and for the object/offset-to-page 190 * hash table headers. Each page cell is initialized and placed on the 191 * free list. 192 */ 193 vm_offset_t 194 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 195 { 196 vm_offset_t mapped; 197 struct vm_page **bucket; 198 vm_size_t npages; 199 vm_paddr_t page_range; 200 vm_paddr_t new_end; 201 int i; 202 vm_paddr_t pa; 203 int nblocks; 204 vm_paddr_t last_pa; 205 206 /* the biggest memory array is the second group of pages */ 207 vm_paddr_t end; 208 vm_paddr_t biggestone, biggestsize; 209 210 vm_paddr_t total; 211 212 total = 0; 213 biggestsize = 0; 214 biggestone = 0; 215 nblocks = 0; 216 vaddr = round_page(vaddr); 217 218 for (i = 0; phys_avail[i + 1]; i += 2) { 219 phys_avail[i] = round_page(phys_avail[i]); 220 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 221 } 222 223 for (i = 0; phys_avail[i + 1]; i += 2) { 224 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 225 226 if (size > biggestsize) { 227 biggestone = i; 228 biggestsize = size; 229 } 230 ++nblocks; 231 total += size; 232 } 233 234 end = phys_avail[biggestone+1]; 235 236 /* 237 * Initialize the queue headers for the free queue, the active queue 238 * and the inactive queue. 239 */ 240 241 vm_page_queue_init(); 242 243 /* 244 * Allocate (and initialize) the hash table buckets. 245 * 246 * The number of buckets MUST BE a power of 2, and the actual value is 247 * the next power of 2 greater than the number of physical pages in 248 * the system. 249 * 250 * We make the hash table approximately 2x the number of pages to 251 * reduce the chain length. This is about the same size using the 252 * singly-linked list as the 1x hash table we were using before 253 * using TAILQ but the chain length will be smaller. 254 * 255 * Note: This computation can be tweaked if desired. 256 */ 257 vm_page_buckets = (struct vm_page **)vaddr; 258 bucket = vm_page_buckets; 259 if (vm_page_bucket_count == 0) { 260 vm_page_bucket_count = 1; 261 while (vm_page_bucket_count < atop(total)) 262 vm_page_bucket_count <<= 1; 263 } 264 vm_page_bucket_count <<= 1; 265 vm_page_hash_mask = vm_page_bucket_count - 1; 266 267 /* 268 * Validate these addresses. 269 */ 270 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 271 new_end = trunc_page(new_end); 272 mapped = round_page(vaddr); 273 vaddr = pmap_map(mapped, new_end, end, 274 VM_PROT_READ | VM_PROT_WRITE); 275 vaddr = round_page(vaddr); 276 bzero((caddr_t) mapped, vaddr - mapped); 277 278 for (i = 0; i < vm_page_bucket_count; i++) { 279 *bucket = NULL; 280 bucket++; 281 } 282 283 /* 284 * Compute the number of pages of memory that will be available for 285 * use (taking into account the overhead of a page structure per 286 * page). 287 */ 288 first_page = phys_avail[0] / PAGE_SIZE; 289 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 290 npages = (total - (page_range * sizeof(struct vm_page)) - 291 (end - new_end)) / PAGE_SIZE; 292 293 end = new_end; 294 295 /* 296 * Initialize the mem entry structures now, and put them in the free 297 * queue. 298 */ 299 vm_page_array = (vm_page_t) vaddr; 300 mapped = vaddr; 301 302 /* 303 * Validate these addresses. 304 */ 305 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 306 mapped = pmap_map(mapped, new_end, end, 307 VM_PROT_READ | VM_PROT_WRITE); 308 309 /* 310 * Clear all of the page structures 311 */ 312 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 313 vm_page_array_size = page_range; 314 315 /* 316 * Construct the free queue(s) in ascending order (by physical 317 * address) so that the first 16MB of physical memory is allocated 318 * last rather than first. On large-memory machines, this avoids 319 * the exhaustion of low physical memory before isa_dmainit has run. 320 */ 321 vmstats.v_page_count = 0; 322 vmstats.v_free_count = 0; 323 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 324 pa = phys_avail[i]; 325 if (i == biggestone) 326 last_pa = new_end; 327 else 328 last_pa = phys_avail[i + 1]; 329 while (pa < last_pa && npages-- > 0) { 330 vm_add_new_page(pa); 331 pa += PAGE_SIZE; 332 } 333 } 334 return (mapped); 335 } 336 337 /* 338 * Distributes the object/offset key pair among hash buckets. 339 * 340 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 341 * This routine may not block. 342 * 343 * We try to randomize the hash based on the object to spread the pages 344 * out in the hash table without it costing us too much. 345 */ 346 static __inline int 347 vm_page_hash(vm_object_t object, vm_pindex_t pindex) 348 { 349 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 350 351 return(i & vm_page_hash_mask); 352 } 353 354 /* 355 * The opposite of vm_page_hold(). A page can be freed while being held, 356 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 357 * in this case to actually free it once the hold count drops to 0. 358 * 359 * This routine must be called at splvm(). 360 */ 361 void 362 vm_page_unhold(vm_page_t mem) 363 { 364 --mem->hold_count; 365 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 366 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) { 367 vm_page_busy(mem); 368 vm_page_free_toq(mem); 369 } 370 } 371 372 /* 373 * Inserts the given mem entry into the object and object list. 374 * 375 * The pagetables are not updated but will presumably fault the page 376 * in if necessary, or if a kernel page the caller will at some point 377 * enter the page into the kernel's pmap. We are not allowed to block 378 * here so we *can't* do this anyway. 379 * 380 * This routine may not block. 381 * This routine must be called with a critical section held. 382 */ 383 void 384 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 385 { 386 struct vm_page **bucket; 387 388 ASSERT_IN_CRIT_SECTION(); 389 if (m->object != NULL) 390 panic("vm_page_insert: already inserted"); 391 392 /* 393 * Record the object/offset pair in this page 394 */ 395 m->object = object; 396 m->pindex = pindex; 397 398 /* 399 * Insert it into the object_object/offset hash table 400 */ 401 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 402 m->hnext = *bucket; 403 *bucket = m; 404 vm_page_bucket_generation++; 405 406 /* 407 * Now link into the object's list of backed pages. 408 */ 409 TAILQ_INSERT_TAIL(&object->memq, m, listq); 410 object->generation++; 411 412 /* 413 * show that the object has one more resident page. 414 */ 415 object->resident_page_count++; 416 417 /* 418 * Since we are inserting a new and possibly dirty page, 419 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 420 */ 421 if (m->flags & PG_WRITEABLE) 422 vm_object_set_writeable_dirty(object); 423 } 424 425 /* 426 * Removes the given vm_page_t from the global (object,index) hash table 427 * and from the object's memq. 428 * 429 * The underlying pmap entry (if any) is NOT removed here. 430 * This routine may not block. 431 * 432 * The page must be BUSY and will remain BUSY on return. No spl needs to be 433 * held on call to this routine. 434 * 435 * note: FreeBSD side effect was to unbusy the page on return. We leave 436 * it busy. 437 */ 438 void 439 vm_page_remove(vm_page_t m) 440 { 441 vm_object_t object; 442 struct vm_page **bucket; 443 444 crit_enter(); 445 if (m->object == NULL) { 446 crit_exit(); 447 return; 448 } 449 450 if ((m->flags & PG_BUSY) == 0) 451 panic("vm_page_remove: page not busy"); 452 453 object = m->object; 454 455 /* 456 * Remove from the object_object/offset hash table. The object 457 * must be on the hash queue, we will panic if it isn't 458 * 459 * Note: we must NULL-out m->hnext to prevent loops in detached 460 * buffers with vm_page_lookup(). 461 */ 462 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 463 while (*bucket != m) { 464 if (*bucket == NULL) 465 panic("vm_page_remove(): page not found in hash"); 466 bucket = &(*bucket)->hnext; 467 } 468 *bucket = m->hnext; 469 m->hnext = NULL; 470 vm_page_bucket_generation++; 471 472 /* 473 * Now remove from the object's list of backed pages. 474 */ 475 TAILQ_REMOVE(&object->memq, m, listq); 476 477 /* 478 * And show that the object has one fewer resident page. 479 */ 480 object->resident_page_count--; 481 object->generation++; 482 483 m->object = NULL; 484 crit_exit(); 485 } 486 487 /* 488 * Locate and return the page at (object, pindex), or NULL if the 489 * page could not be found. 490 * 491 * This routine will operate properly without spl protection, but 492 * the returned page could be in flux if it is busy. Because an 493 * interrupt can race a caller's busy check (unbusying and freeing the 494 * page we return before the caller is able to check the busy bit), 495 * the caller should generally call this routine with a critical 496 * section held. 497 * 498 * Callers may call this routine without spl protection if they know 499 * 'for sure' that the page will not be ripped out from under them 500 * by an interrupt. 501 */ 502 vm_page_t 503 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 504 { 505 vm_page_t m; 506 struct vm_page **bucket; 507 int generation; 508 509 /* 510 * Search the hash table for this object/offset pair 511 */ 512 retry: 513 generation = vm_page_bucket_generation; 514 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 515 for (m = *bucket; m != NULL; m = m->hnext) { 516 if ((m->object == object) && (m->pindex == pindex)) { 517 if (vm_page_bucket_generation != generation) 518 goto retry; 519 return (m); 520 } 521 } 522 if (vm_page_bucket_generation != generation) 523 goto retry; 524 return (NULL); 525 } 526 527 /* 528 * vm_page_rename() 529 * 530 * Move the given memory entry from its current object to the specified 531 * target object/offset. 532 * 533 * The object must be locked. 534 * This routine may not block. 535 * 536 * Note: This routine will raise itself to splvm(), the caller need not. 537 * 538 * Note: Swap associated with the page must be invalidated by the move. We 539 * have to do this for several reasons: (1) we aren't freeing the 540 * page, (2) we are dirtying the page, (3) the VM system is probably 541 * moving the page from object A to B, and will then later move 542 * the backing store from A to B and we can't have a conflict. 543 * 544 * Note: We *always* dirty the page. It is necessary both for the 545 * fact that we moved it, and because we may be invalidating 546 * swap. If the page is on the cache, we have to deactivate it 547 * or vm_page_dirty() will panic. Dirty pages are not allowed 548 * on the cache. 549 */ 550 void 551 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 552 { 553 crit_enter(); 554 vm_page_remove(m); 555 vm_page_insert(m, new_object, new_pindex); 556 if (m->queue - m->pc == PQ_CACHE) 557 vm_page_deactivate(m); 558 vm_page_dirty(m); 559 vm_page_wakeup(m); 560 crit_exit(); 561 } 562 563 /* 564 * vm_page_unqueue() without any wakeup. This routine is used when a page 565 * is being moved between queues or otherwise is to remain BUSYied by the 566 * caller. 567 * 568 * This routine must be called at splhigh(). 569 * This routine may not block. 570 */ 571 void 572 vm_page_unqueue_nowakeup(vm_page_t m) 573 { 574 int queue = m->queue; 575 struct vpgqueues *pq; 576 577 if (queue != PQ_NONE) { 578 pq = &vm_page_queues[queue]; 579 m->queue = PQ_NONE; 580 TAILQ_REMOVE(&pq->pl, m, pageq); 581 (*pq->cnt)--; 582 pq->lcnt--; 583 } 584 } 585 586 /* 587 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 588 * if necessary. 589 * 590 * This routine must be called at splhigh(). 591 * This routine may not block. 592 */ 593 void 594 vm_page_unqueue(vm_page_t m) 595 { 596 int queue = m->queue; 597 struct vpgqueues *pq; 598 599 if (queue != PQ_NONE) { 600 m->queue = PQ_NONE; 601 pq = &vm_page_queues[queue]; 602 TAILQ_REMOVE(&pq->pl, m, pageq); 603 (*pq->cnt)--; 604 pq->lcnt--; 605 if ((queue - m->pc) == PQ_CACHE) { 606 if (vm_paging_needed()) 607 pagedaemon_wakeup(); 608 } 609 } 610 } 611 612 /* 613 * vm_page_list_find() 614 * 615 * Find a page on the specified queue with color optimization. 616 * 617 * The page coloring optimization attempts to locate a page that does 618 * not overload other nearby pages in the object in the cpu's L1 or L2 619 * caches. We need this optimization because cpu caches tend to be 620 * physical caches, while object spaces tend to be virtual. 621 * 622 * This routine must be called at splvm(). 623 * This routine may not block. 624 * 625 * Note that this routine is carefully inlined. A non-inlined version 626 * is available for outside callers but the only critical path is 627 * from within this source file. 628 */ 629 static __inline 630 vm_page_t 631 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 632 { 633 vm_page_t m; 634 635 if (prefer_zero) 636 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 637 else 638 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 639 if (m == NULL) 640 m = _vm_page_list_find2(basequeue, index); 641 return(m); 642 } 643 644 static vm_page_t 645 _vm_page_list_find2(int basequeue, int index) 646 { 647 int i; 648 vm_page_t m = NULL; 649 struct vpgqueues *pq; 650 651 pq = &vm_page_queues[basequeue]; 652 653 /* 654 * Note that for the first loop, index+i and index-i wind up at the 655 * same place. Even though this is not totally optimal, we've already 656 * blown it by missing the cache case so we do not care. 657 */ 658 659 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 660 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 661 break; 662 663 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 664 break; 665 } 666 return(m); 667 } 668 669 vm_page_t 670 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 671 { 672 return(_vm_page_list_find(basequeue, index, prefer_zero)); 673 } 674 675 /* 676 * Find a page on the cache queue with color optimization. As pages 677 * might be found, but not applicable, they are deactivated. This 678 * keeps us from using potentially busy cached pages. 679 * 680 * This routine must be called with a critical section held. 681 * This routine may not block. 682 */ 683 vm_page_t 684 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 685 { 686 vm_page_t m; 687 688 while (TRUE) { 689 m = _vm_page_list_find( 690 PQ_CACHE, 691 (pindex + object->pg_color) & PQ_L2_MASK, 692 FALSE 693 ); 694 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 695 m->hold_count || m->wire_count)) { 696 vm_page_deactivate(m); 697 continue; 698 } 699 return m; 700 } 701 /* not reached */ 702 } 703 704 /* 705 * Find a free or zero page, with specified preference. We attempt to 706 * inline the nominal case and fall back to _vm_page_select_free() 707 * otherwise. 708 * 709 * This routine must be called with a critical section held. 710 * This routine may not block. 711 */ 712 static __inline vm_page_t 713 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 714 { 715 vm_page_t m; 716 717 m = _vm_page_list_find( 718 PQ_FREE, 719 (pindex + object->pg_color) & PQ_L2_MASK, 720 prefer_zero 721 ); 722 return(m); 723 } 724 725 /* 726 * vm_page_alloc() 727 * 728 * Allocate and return a memory cell associated with this VM object/offset 729 * pair. 730 * 731 * page_req classes: 732 * 733 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 734 * VM_ALLOC_SYSTEM greater free drain 735 * VM_ALLOC_INTERRUPT allow free list to be completely drained 736 * VM_ALLOC_ZERO advisory request for pre-zero'd page 737 * 738 * The object must be locked. 739 * This routine may not block. 740 * The returned page will be marked PG_BUSY 741 * 742 * Additional special handling is required when called from an interrupt 743 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 744 * in this case. 745 */ 746 vm_page_t 747 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 748 { 749 vm_page_t m = NULL; 750 751 KASSERT(!vm_page_lookup(object, pindex), 752 ("vm_page_alloc: page already allocated")); 753 KKASSERT(page_req & 754 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 755 756 /* 757 * The pager is allowed to eat deeper into the free page list. 758 */ 759 if (curthread == pagethread) 760 page_req |= VM_ALLOC_SYSTEM; 761 762 crit_enter(); 763 loop: 764 if (vmstats.v_free_count > vmstats.v_free_reserved || 765 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 766 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 767 vmstats.v_free_count > vmstats.v_interrupt_free_min) 768 ) { 769 /* 770 * The free queue has sufficient free pages to take one out. 771 */ 772 if (page_req & VM_ALLOC_ZERO) 773 m = vm_page_select_free(object, pindex, TRUE); 774 else 775 m = vm_page_select_free(object, pindex, FALSE); 776 } else if (page_req & VM_ALLOC_NORMAL) { 777 /* 778 * Allocatable from the cache (non-interrupt only). On 779 * success, we must free the page and try again, thus 780 * ensuring that vmstats.v_*_free_min counters are replenished. 781 */ 782 #ifdef INVARIANTS 783 if (curthread->td_preempted) { 784 printf("vm_page_alloc(): warning, attempt to allocate" 785 " cache page from preempting interrupt\n"); 786 m = NULL; 787 } else { 788 m = vm_page_select_cache(object, pindex); 789 } 790 #else 791 m = vm_page_select_cache(object, pindex); 792 #endif 793 /* 794 * On success move the page into the free queue and loop. 795 */ 796 if (m != NULL) { 797 KASSERT(m->dirty == 0, 798 ("Found dirty cache page %p", m)); 799 vm_page_busy(m); 800 vm_page_protect(m, VM_PROT_NONE); 801 vm_page_free(m); 802 goto loop; 803 } 804 805 /* 806 * On failure return NULL 807 */ 808 crit_exit(); 809 #if defined(DIAGNOSTIC) 810 if (vmstats.v_cache_count > 0) 811 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 812 #endif 813 vm_pageout_deficit++; 814 pagedaemon_wakeup(); 815 return (NULL); 816 } else { 817 /* 818 * No pages available, wakeup the pageout daemon and give up. 819 */ 820 crit_exit(); 821 vm_pageout_deficit++; 822 pagedaemon_wakeup(); 823 return (NULL); 824 } 825 826 /* 827 * Good page found. The page has not yet been busied. We are in 828 * a critical section. 829 */ 830 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 831 832 /* 833 * Remove from free queue 834 */ 835 vm_page_unqueue_nowakeup(m); 836 837 /* 838 * Initialize structure. Only the PG_ZERO flag is inherited. Set 839 * the page PG_BUSY 840 */ 841 if (m->flags & PG_ZERO) { 842 vm_page_zero_count--; 843 m->flags = PG_ZERO | PG_BUSY; 844 } else { 845 m->flags = PG_BUSY; 846 } 847 m->wire_count = 0; 848 m->hold_count = 0; 849 m->act_count = 0; 850 m->busy = 0; 851 m->valid = 0; 852 KASSERT(m->dirty == 0, 853 ("vm_page_alloc: free/cache page %p was dirty", m)); 854 855 /* 856 * vm_page_insert() is safe prior to the crit_exit(). Note also that 857 * inserting a page here does not insert it into the pmap (which 858 * could cause us to block allocating memory). We cannot block 859 * anywhere. 860 */ 861 vm_page_insert(m, object, pindex); 862 863 /* 864 * Don't wakeup too often - wakeup the pageout daemon when 865 * we would be nearly out of memory. 866 */ 867 if (vm_paging_needed()) 868 pagedaemon_wakeup(); 869 870 crit_exit(); 871 872 /* 873 * A PG_BUSY page is returned. 874 */ 875 return (m); 876 } 877 878 /* 879 * Block until free pages are available for allocation, called in various 880 * places before memory allocations. 881 */ 882 void 883 vm_wait(void) 884 { 885 int s; 886 887 s = splvm(); 888 if (curthread == pagethread) { 889 vm_pageout_pages_needed = 1; 890 tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0); 891 } else { 892 if (!vm_pages_needed) { 893 vm_pages_needed = 1; 894 wakeup(&vm_pages_needed); 895 } 896 tsleep(&vmstats.v_free_count, 0, "vmwait", 0); 897 } 898 splx(s); 899 } 900 901 /* 902 * Block until free pages are available for allocation 903 * 904 * Called only in vm_fault so that processes page faulting can be 905 * easily tracked. 906 * 907 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing 908 * processes will be able to grab memory first. Do not change 909 * this balance without careful testing first. 910 */ 911 void 912 vm_waitpfault(void) 913 { 914 int s; 915 916 s = splvm(); 917 if (!vm_pages_needed) { 918 vm_pages_needed = 1; 919 wakeup(&vm_pages_needed); 920 } 921 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 922 splx(s); 923 } 924 925 /* 926 * Put the specified page on the active list (if appropriate). Ensure 927 * that act_count is at least ACT_INIT but do not otherwise mess with it. 928 * 929 * The page queues must be locked. 930 * This routine may not block. 931 */ 932 void 933 vm_page_activate(vm_page_t m) 934 { 935 crit_enter(); 936 if (m->queue != PQ_ACTIVE) { 937 if ((m->queue - m->pc) == PQ_CACHE) 938 mycpu->gd_cnt.v_reactivated++; 939 940 vm_page_unqueue(m); 941 942 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 943 m->queue = PQ_ACTIVE; 944 vm_page_queues[PQ_ACTIVE].lcnt++; 945 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 946 m, pageq); 947 if (m->act_count < ACT_INIT) 948 m->act_count = ACT_INIT; 949 vmstats.v_active_count++; 950 } 951 } else { 952 if (m->act_count < ACT_INIT) 953 m->act_count = ACT_INIT; 954 } 955 crit_exit(); 956 } 957 958 /* 959 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 960 * routine is called when a page has been added to the cache or free 961 * queues. 962 * 963 * This routine may not block. 964 * This routine must be called at splvm() 965 */ 966 static __inline void 967 vm_page_free_wakeup(void) 968 { 969 /* 970 * if pageout daemon needs pages, then tell it that there are 971 * some free. 972 */ 973 if (vm_pageout_pages_needed && 974 vmstats.v_cache_count + vmstats.v_free_count >= 975 vmstats.v_pageout_free_min 976 ) { 977 wakeup(&vm_pageout_pages_needed); 978 vm_pageout_pages_needed = 0; 979 } 980 981 /* 982 * wakeup processes that are waiting on memory if we hit a 983 * high water mark. And wakeup scheduler process if we have 984 * lots of memory. this process will swapin processes. 985 */ 986 if (vm_pages_needed && !vm_page_count_min()) { 987 vm_pages_needed = 0; 988 wakeup(&vmstats.v_free_count); 989 } 990 } 991 992 /* 993 * vm_page_free_toq: 994 * 995 * Returns the given page to the PQ_FREE list, disassociating it with 996 * any VM object. 997 * 998 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 999 * return (the page will have been freed). No particular spl is required 1000 * on entry. 1001 * 1002 * This routine may not block. 1003 */ 1004 void 1005 vm_page_free_toq(vm_page_t m) 1006 { 1007 struct vpgqueues *pq; 1008 1009 crit_enter(); 1010 mycpu->gd_cnt.v_tfree++; 1011 1012 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1013 printf( 1014 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1015 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1016 m->hold_count); 1017 if ((m->queue - m->pc) == PQ_FREE) 1018 panic("vm_page_free: freeing free page"); 1019 else 1020 panic("vm_page_free: freeing busy page"); 1021 } 1022 1023 /* 1024 * unqueue, then remove page. Note that we cannot destroy 1025 * the page here because we do not want to call the pager's 1026 * callback routine until after we've put the page on the 1027 * appropriate free queue. 1028 */ 1029 vm_page_unqueue_nowakeup(m); 1030 vm_page_remove(m); 1031 1032 /* 1033 * No further management of fictitious pages occurs beyond object 1034 * and queue removal. 1035 */ 1036 if ((m->flags & PG_FICTITIOUS) != 0) { 1037 vm_page_wakeup(m); 1038 crit_exit(); 1039 return; 1040 } 1041 1042 m->valid = 0; 1043 vm_page_undirty(m); 1044 1045 if (m->wire_count != 0) { 1046 if (m->wire_count > 1) { 1047 panic( 1048 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1049 m->wire_count, (long)m->pindex); 1050 } 1051 panic("vm_page_free: freeing wired page"); 1052 } 1053 1054 /* 1055 * Clear the UNMANAGED flag when freeing an unmanaged page. 1056 */ 1057 if (m->flags & PG_UNMANAGED) { 1058 m->flags &= ~PG_UNMANAGED; 1059 } else { 1060 #ifdef __alpha__ 1061 pmap_page_is_free(m); 1062 #endif 1063 } 1064 1065 if (m->hold_count != 0) { 1066 m->flags &= ~PG_ZERO; 1067 m->queue = PQ_HOLD; 1068 } else { 1069 m->queue = PQ_FREE + m->pc; 1070 } 1071 pq = &vm_page_queues[m->queue]; 1072 pq->lcnt++; 1073 ++(*pq->cnt); 1074 1075 /* 1076 * Put zero'd pages on the end ( where we look for zero'd pages 1077 * first ) and non-zerod pages at the head. 1078 */ 1079 if (m->flags & PG_ZERO) { 1080 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1081 ++vm_page_zero_count; 1082 } else { 1083 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1084 } 1085 vm_page_wakeup(m); 1086 vm_page_free_wakeup(); 1087 crit_exit(); 1088 } 1089 1090 /* 1091 * vm_page_unmanage() 1092 * 1093 * Prevent PV management from being done on the page. The page is 1094 * removed from the paging queues as if it were wired, and as a 1095 * consequence of no longer being managed the pageout daemon will not 1096 * touch it (since there is no way to locate the pte mappings for the 1097 * page). madvise() calls that mess with the pmap will also no longer 1098 * operate on the page. 1099 * 1100 * Beyond that the page is still reasonably 'normal'. Freeing the page 1101 * will clear the flag. 1102 * 1103 * This routine is used by OBJT_PHYS objects - objects using unswappable 1104 * physical memory as backing store rather then swap-backed memory and 1105 * will eventually be extended to support 4MB unmanaged physical 1106 * mappings. 1107 * 1108 * Must be called with a critical section held. 1109 */ 1110 void 1111 vm_page_unmanage(vm_page_t m) 1112 { 1113 ASSERT_IN_CRIT_SECTION(); 1114 if ((m->flags & PG_UNMANAGED) == 0) { 1115 if (m->wire_count == 0) 1116 vm_page_unqueue(m); 1117 } 1118 vm_page_flag_set(m, PG_UNMANAGED); 1119 } 1120 1121 /* 1122 * Mark this page as wired down by yet another map, removing it from 1123 * paging queues as necessary. 1124 * 1125 * The page queues must be locked. 1126 * This routine may not block. 1127 */ 1128 void 1129 vm_page_wire(vm_page_t m) 1130 { 1131 /* 1132 * Only bump the wire statistics if the page is not already wired, 1133 * and only unqueue the page if it is on some queue (if it is unmanaged 1134 * it is already off the queues). Don't do anything with fictitious 1135 * pages because they are always wired. 1136 */ 1137 crit_enter(); 1138 if ((m->flags & PG_FICTITIOUS) == 0) { 1139 if (m->wire_count == 0) { 1140 if ((m->flags & PG_UNMANAGED) == 0) 1141 vm_page_unqueue(m); 1142 vmstats.v_wire_count++; 1143 } 1144 m->wire_count++; 1145 KASSERT(m->wire_count != 0, 1146 ("vm_page_wire: wire_count overflow m=%p", m)); 1147 } 1148 vm_page_flag_set(m, PG_MAPPED); 1149 crit_exit(); 1150 } 1151 1152 /* 1153 * Release one wiring of this page, potentially enabling it to be paged again. 1154 * 1155 * Many pages placed on the inactive queue should actually go 1156 * into the cache, but it is difficult to figure out which. What 1157 * we do instead, if the inactive target is well met, is to put 1158 * clean pages at the head of the inactive queue instead of the tail. 1159 * This will cause them to be moved to the cache more quickly and 1160 * if not actively re-referenced, freed more quickly. If we just 1161 * stick these pages at the end of the inactive queue, heavy filesystem 1162 * meta-data accesses can cause an unnecessary paging load on memory bound 1163 * processes. This optimization causes one-time-use metadata to be 1164 * reused more quickly. 1165 * 1166 * BUT, if we are in a low-memory situation we have no choice but to 1167 * put clean pages on the cache queue. 1168 * 1169 * A number of routines use vm_page_unwire() to guarantee that the page 1170 * will go into either the inactive or active queues, and will NEVER 1171 * be placed in the cache - for example, just after dirtying a page. 1172 * dirty pages in the cache are not allowed. 1173 * 1174 * The page queues must be locked. 1175 * This routine may not block. 1176 */ 1177 void 1178 vm_page_unwire(vm_page_t m, int activate) 1179 { 1180 crit_enter(); 1181 if (m->flags & PG_FICTITIOUS) { 1182 /* do nothing */ 1183 } else if (m->wire_count <= 0) { 1184 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1185 } else { 1186 if (--m->wire_count == 0) { 1187 --vmstats.v_wire_count; 1188 if (m->flags & PG_UNMANAGED) { 1189 ; 1190 } else if (activate) { 1191 TAILQ_INSERT_TAIL( 1192 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1193 m->queue = PQ_ACTIVE; 1194 vm_page_queues[PQ_ACTIVE].lcnt++; 1195 vmstats.v_active_count++; 1196 } else { 1197 vm_page_flag_clear(m, PG_WINATCFLS); 1198 TAILQ_INSERT_TAIL( 1199 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1200 m->queue = PQ_INACTIVE; 1201 vm_page_queues[PQ_INACTIVE].lcnt++; 1202 vmstats.v_inactive_count++; 1203 } 1204 } 1205 } 1206 crit_exit(); 1207 } 1208 1209 1210 /* 1211 * Move the specified page to the inactive queue. If the page has 1212 * any associated swap, the swap is deallocated. 1213 * 1214 * Normally athead is 0 resulting in LRU operation. athead is set 1215 * to 1 if we want this page to be 'as if it were placed in the cache', 1216 * except without unmapping it from the process address space. 1217 * 1218 * This routine may not block. 1219 */ 1220 static __inline void 1221 _vm_page_deactivate(vm_page_t m, int athead) 1222 { 1223 /* 1224 * Ignore if already inactive. 1225 */ 1226 if (m->queue == PQ_INACTIVE) 1227 return; 1228 1229 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1230 if ((m->queue - m->pc) == PQ_CACHE) 1231 mycpu->gd_cnt.v_reactivated++; 1232 vm_page_flag_clear(m, PG_WINATCFLS); 1233 vm_page_unqueue(m); 1234 if (athead) 1235 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1236 else 1237 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1238 m->queue = PQ_INACTIVE; 1239 vm_page_queues[PQ_INACTIVE].lcnt++; 1240 vmstats.v_inactive_count++; 1241 } 1242 } 1243 1244 void 1245 vm_page_deactivate(vm_page_t m) 1246 { 1247 crit_enter(); 1248 _vm_page_deactivate(m, 0); 1249 crit_exit(); 1250 } 1251 1252 /* 1253 * vm_page_try_to_cache: 1254 * 1255 * Returns 0 on failure, 1 on success 1256 */ 1257 int 1258 vm_page_try_to_cache(vm_page_t m) 1259 { 1260 crit_enter(); 1261 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1262 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1263 return(0); 1264 } 1265 vm_page_test_dirty(m); 1266 if (m->dirty) { 1267 crit_exit(); 1268 return(0); 1269 } 1270 vm_page_cache(m); 1271 crit_exit(); 1272 return(1); 1273 } 1274 1275 /* 1276 * Attempt to free the page. If we cannot free it, we do nothing. 1277 * 1 is returned on success, 0 on failure. 1278 */ 1279 int 1280 vm_page_try_to_free(vm_page_t m) 1281 { 1282 crit_enter(); 1283 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1284 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1285 crit_exit(); 1286 return(0); 1287 } 1288 vm_page_test_dirty(m); 1289 if (m->dirty) { 1290 crit_exit(); 1291 return(0); 1292 } 1293 vm_page_busy(m); 1294 vm_page_protect(m, VM_PROT_NONE); 1295 vm_page_free(m); 1296 crit_exit(); 1297 return(1); 1298 } 1299 1300 /* 1301 * vm_page_cache 1302 * 1303 * Put the specified page onto the page cache queue (if appropriate). 1304 * 1305 * This routine may not block. 1306 */ 1307 void 1308 vm_page_cache(vm_page_t m) 1309 { 1310 ASSERT_IN_CRIT_SECTION(); 1311 1312 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1313 m->wire_count || m->hold_count) { 1314 printf("vm_page_cache: attempting to cache busy/held page\n"); 1315 return; 1316 } 1317 if ((m->queue - m->pc) == PQ_CACHE) 1318 return; 1319 1320 /* 1321 * Remove all pmaps and indicate that the page is not 1322 * writeable or mapped. 1323 */ 1324 1325 vm_page_protect(m, VM_PROT_NONE); 1326 if (m->dirty != 0) { 1327 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1328 (long)m->pindex); 1329 } 1330 vm_page_unqueue_nowakeup(m); 1331 m->queue = PQ_CACHE + m->pc; 1332 vm_page_queues[m->queue].lcnt++; 1333 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1334 vmstats.v_cache_count++; 1335 vm_page_free_wakeup(); 1336 } 1337 1338 /* 1339 * vm_page_dontneed() 1340 * 1341 * Cache, deactivate, or do nothing as appropriate. This routine 1342 * is typically used by madvise() MADV_DONTNEED. 1343 * 1344 * Generally speaking we want to move the page into the cache so 1345 * it gets reused quickly. However, this can result in a silly syndrome 1346 * due to the page recycling too quickly. Small objects will not be 1347 * fully cached. On the otherhand, if we move the page to the inactive 1348 * queue we wind up with a problem whereby very large objects 1349 * unnecessarily blow away our inactive and cache queues. 1350 * 1351 * The solution is to move the pages based on a fixed weighting. We 1352 * either leave them alone, deactivate them, or move them to the cache, 1353 * where moving them to the cache has the highest weighting. 1354 * By forcing some pages into other queues we eventually force the 1355 * system to balance the queues, potentially recovering other unrelated 1356 * space from active. The idea is to not force this to happen too 1357 * often. 1358 */ 1359 void 1360 vm_page_dontneed(vm_page_t m) 1361 { 1362 static int dnweight; 1363 int dnw; 1364 int head; 1365 1366 dnw = ++dnweight; 1367 1368 /* 1369 * occassionally leave the page alone 1370 */ 1371 crit_enter(); 1372 if ((dnw & 0x01F0) == 0 || 1373 m->queue == PQ_INACTIVE || 1374 m->queue - m->pc == PQ_CACHE 1375 ) { 1376 if (m->act_count >= ACT_INIT) 1377 --m->act_count; 1378 crit_exit(); 1379 return; 1380 } 1381 1382 if (m->dirty == 0) 1383 vm_page_test_dirty(m); 1384 1385 if (m->dirty || (dnw & 0x0070) == 0) { 1386 /* 1387 * Deactivate the page 3 times out of 32. 1388 */ 1389 head = 0; 1390 } else { 1391 /* 1392 * Cache the page 28 times out of every 32. Note that 1393 * the page is deactivated instead of cached, but placed 1394 * at the head of the queue instead of the tail. 1395 */ 1396 head = 1; 1397 } 1398 _vm_page_deactivate(m, head); 1399 crit_exit(); 1400 } 1401 1402 /* 1403 * Grab a page, blocking if it is busy and allocating a page if necessary. 1404 * A busy page is returned or NULL. 1405 * 1406 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1407 * If VM_ALLOC_RETRY is not specified 1408 * 1409 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1410 * always returned if we had blocked. 1411 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1412 * This routine may not be called from an interrupt. 1413 * The returned page may not be entirely valid. 1414 * 1415 * This routine may be called from mainline code without spl protection and 1416 * be guarenteed a busied page associated with the object at the specified 1417 * index. 1418 */ 1419 vm_page_t 1420 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1421 { 1422 vm_page_t m; 1423 int generation; 1424 1425 KKASSERT(allocflags & 1426 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1427 crit_enter(); 1428 retrylookup: 1429 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1430 if (m->busy || (m->flags & PG_BUSY)) { 1431 generation = object->generation; 1432 1433 while ((object->generation == generation) && 1434 (m->busy || (m->flags & PG_BUSY))) { 1435 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1436 tsleep(m, 0, "pgrbwt", 0); 1437 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1438 m = NULL; 1439 goto done; 1440 } 1441 } 1442 goto retrylookup; 1443 } else { 1444 vm_page_busy(m); 1445 goto done; 1446 } 1447 } 1448 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1449 if (m == NULL) { 1450 vm_wait(); 1451 if ((allocflags & VM_ALLOC_RETRY) == 0) 1452 goto done; 1453 goto retrylookup; 1454 } 1455 done: 1456 crit_exit(); 1457 return(m); 1458 } 1459 1460 /* 1461 * Mapping function for valid bits or for dirty bits in 1462 * a page. May not block. 1463 * 1464 * Inputs are required to range within a page. 1465 */ 1466 __inline int 1467 vm_page_bits(int base, int size) 1468 { 1469 int first_bit; 1470 int last_bit; 1471 1472 KASSERT( 1473 base + size <= PAGE_SIZE, 1474 ("vm_page_bits: illegal base/size %d/%d", base, size) 1475 ); 1476 1477 if (size == 0) /* handle degenerate case */ 1478 return(0); 1479 1480 first_bit = base >> DEV_BSHIFT; 1481 last_bit = (base + size - 1) >> DEV_BSHIFT; 1482 1483 return ((2 << last_bit) - (1 << first_bit)); 1484 } 1485 1486 /* 1487 * Sets portions of a page valid and clean. The arguments are expected 1488 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1489 * of any partial chunks touched by the range. The invalid portion of 1490 * such chunks will be zero'd. 1491 * 1492 * This routine may not block. 1493 * 1494 * (base + size) must be less then or equal to PAGE_SIZE. 1495 */ 1496 void 1497 vm_page_set_validclean(vm_page_t m, int base, int size) 1498 { 1499 int pagebits; 1500 int frag; 1501 int endoff; 1502 1503 if (size == 0) /* handle degenerate case */ 1504 return; 1505 1506 /* 1507 * If the base is not DEV_BSIZE aligned and the valid 1508 * bit is clear, we have to zero out a portion of the 1509 * first block. 1510 */ 1511 1512 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1513 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1514 ) { 1515 pmap_zero_page_area( 1516 VM_PAGE_TO_PHYS(m), 1517 frag, 1518 base - frag 1519 ); 1520 } 1521 1522 /* 1523 * If the ending offset is not DEV_BSIZE aligned and the 1524 * valid bit is clear, we have to zero out a portion of 1525 * the last block. 1526 */ 1527 1528 endoff = base + size; 1529 1530 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1531 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1532 ) { 1533 pmap_zero_page_area( 1534 VM_PAGE_TO_PHYS(m), 1535 endoff, 1536 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1537 ); 1538 } 1539 1540 /* 1541 * Set valid, clear dirty bits. If validating the entire 1542 * page we can safely clear the pmap modify bit. We also 1543 * use this opportunity to clear the PG_NOSYNC flag. If a process 1544 * takes a write fault on a MAP_NOSYNC memory area the flag will 1545 * be set again. 1546 * 1547 * We set valid bits inclusive of any overlap, but we can only 1548 * clear dirty bits for DEV_BSIZE chunks that are fully within 1549 * the range. 1550 */ 1551 1552 pagebits = vm_page_bits(base, size); 1553 m->valid |= pagebits; 1554 #if 0 /* NOT YET */ 1555 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1556 frag = DEV_BSIZE - frag; 1557 base += frag; 1558 size -= frag; 1559 if (size < 0) 1560 size = 0; 1561 } 1562 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1563 #endif 1564 m->dirty &= ~pagebits; 1565 if (base == 0 && size == PAGE_SIZE) { 1566 pmap_clear_modify(m); 1567 vm_page_flag_clear(m, PG_NOSYNC); 1568 } 1569 } 1570 1571 void 1572 vm_page_clear_dirty(vm_page_t m, int base, int size) 1573 { 1574 m->dirty &= ~vm_page_bits(base, size); 1575 } 1576 1577 /* 1578 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1579 * valid and dirty bits for the effected areas are cleared. 1580 * 1581 * May not block. 1582 */ 1583 void 1584 vm_page_set_invalid(vm_page_t m, int base, int size) 1585 { 1586 int bits; 1587 1588 bits = vm_page_bits(base, size); 1589 m->valid &= ~bits; 1590 m->dirty &= ~bits; 1591 m->object->generation++; 1592 } 1593 1594 /* 1595 * The kernel assumes that the invalid portions of a page contain 1596 * garbage, but such pages can be mapped into memory by user code. 1597 * When this occurs, we must zero out the non-valid portions of the 1598 * page so user code sees what it expects. 1599 * 1600 * Pages are most often semi-valid when the end of a file is mapped 1601 * into memory and the file's size is not page aligned. 1602 */ 1603 void 1604 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1605 { 1606 int b; 1607 int i; 1608 1609 /* 1610 * Scan the valid bits looking for invalid sections that 1611 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1612 * valid bit may be set ) have already been zerod by 1613 * vm_page_set_validclean(). 1614 */ 1615 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1616 if (i == (PAGE_SIZE / DEV_BSIZE) || 1617 (m->valid & (1 << i)) 1618 ) { 1619 if (i > b) { 1620 pmap_zero_page_area( 1621 VM_PAGE_TO_PHYS(m), 1622 b << DEV_BSHIFT, 1623 (i - b) << DEV_BSHIFT 1624 ); 1625 } 1626 b = i + 1; 1627 } 1628 } 1629 1630 /* 1631 * setvalid is TRUE when we can safely set the zero'd areas 1632 * as being valid. We can do this if there are no cache consistency 1633 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1634 */ 1635 if (setvalid) 1636 m->valid = VM_PAGE_BITS_ALL; 1637 } 1638 1639 /* 1640 * Is a (partial) page valid? Note that the case where size == 0 1641 * will return FALSE in the degenerate case where the page is entirely 1642 * invalid, and TRUE otherwise. 1643 * 1644 * May not block. 1645 */ 1646 int 1647 vm_page_is_valid(vm_page_t m, int base, int size) 1648 { 1649 int bits = vm_page_bits(base, size); 1650 1651 if (m->valid && ((m->valid & bits) == bits)) 1652 return 1; 1653 else 1654 return 0; 1655 } 1656 1657 /* 1658 * update dirty bits from pmap/mmu. May not block. 1659 */ 1660 void 1661 vm_page_test_dirty(vm_page_t m) 1662 { 1663 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1664 vm_page_dirty(m); 1665 } 1666 } 1667 1668 #include "opt_ddb.h" 1669 #ifdef DDB 1670 #include <sys/kernel.h> 1671 1672 #include <ddb/ddb.h> 1673 1674 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1675 { 1676 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1677 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1678 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1679 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1680 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1681 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1682 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1683 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1684 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1685 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1686 } 1687 1688 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1689 { 1690 int i; 1691 db_printf("PQ_FREE:"); 1692 for(i=0;i<PQ_L2_SIZE;i++) { 1693 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1694 } 1695 db_printf("\n"); 1696 1697 db_printf("PQ_CACHE:"); 1698 for(i=0;i<PQ_L2_SIZE;i++) { 1699 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1700 } 1701 db_printf("\n"); 1702 1703 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1704 vm_page_queues[PQ_ACTIVE].lcnt, 1705 vm_page_queues[PQ_INACTIVE].lcnt); 1706 } 1707 #endif /* DDB */ 1708