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 1.40 2008/08/25 17:01:42 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 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 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 98 99 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread)); 100 101 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 102 vm_pindex_t, pindex); 103 104 static void 105 vm_page_queue_init(void) 106 { 107 int i; 108 109 for (i = 0; i < PQ_L2_SIZE; i++) 110 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 111 for (i = 0; i < PQ_L2_SIZE; i++) 112 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 113 114 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 115 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 116 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 117 /* PQ_NONE has no queue */ 118 119 for (i = 0; i < PQ_COUNT; i++) 120 TAILQ_INIT(&vm_page_queues[i].pl); 121 } 122 123 /* 124 * note: place in initialized data section? Is this necessary? 125 */ 126 long first_page = 0; 127 int vm_page_array_size = 0; 128 int vm_page_zero_count = 0; 129 vm_page_t vm_page_array = 0; 130 131 /* 132 * (low level boot) 133 * 134 * Sets the page size, perhaps based upon the memory size. 135 * Must be called before any use of page-size dependent functions. 136 */ 137 void 138 vm_set_page_size(void) 139 { 140 if (vmstats.v_page_size == 0) 141 vmstats.v_page_size = PAGE_SIZE; 142 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 143 panic("vm_set_page_size: page size not a power of two"); 144 } 145 146 /* 147 * (low level boot) 148 * 149 * Add a new page to the freelist for use by the system. New pages 150 * are added to both the head and tail of the associated free page 151 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 152 * requests pull 'recent' adds (higher physical addresses) first. 153 * 154 * Must be called in a critical section. 155 */ 156 vm_page_t 157 vm_add_new_page(vm_paddr_t pa) 158 { 159 struct vpgqueues *vpq; 160 vm_page_t m; 161 162 ++vmstats.v_page_count; 163 ++vmstats.v_free_count; 164 m = PHYS_TO_VM_PAGE(pa); 165 m->phys_addr = pa; 166 m->flags = 0; 167 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 168 m->queue = m->pc + PQ_FREE; 169 KKASSERT(m->dirty == 0); 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 * starta/enda represents the range of physical memory addresses available 192 * for use (skipping memory already used by the kernel), subject to 193 * phys_avail[]. Note that phys_avail[] has already mapped out memory 194 * already in use by the kernel. 195 */ 196 vm_offset_t 197 vm_page_startup(vm_offset_t vaddr) 198 { 199 vm_offset_t mapped; 200 vm_size_t npages; 201 vm_paddr_t page_range; 202 vm_paddr_t new_end; 203 int i; 204 vm_paddr_t pa; 205 int nblocks; 206 vm_paddr_t last_pa; 207 vm_paddr_t end; 208 vm_paddr_t biggestone, biggestsize; 209 vm_paddr_t total; 210 211 total = 0; 212 biggestsize = 0; 213 biggestone = 0; 214 nblocks = 0; 215 vaddr = round_page(vaddr); 216 217 for (i = 0; phys_avail[i + 1]; i += 2) { 218 phys_avail[i] = round_page(phys_avail[i]); 219 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 220 } 221 222 for (i = 0; phys_avail[i + 1]; i += 2) { 223 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 224 225 if (size > biggestsize) { 226 biggestone = i; 227 biggestsize = size; 228 } 229 ++nblocks; 230 total += size; 231 } 232 233 end = phys_avail[biggestone+1]; 234 end = trunc_page(end); 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 * Compute the number of pages of memory that will be available for 245 * use (taking into account the overhead of a page structure per 246 * page). 247 */ 248 first_page = phys_avail[0] / PAGE_SIZE; 249 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 250 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 251 252 /* 253 * Initialize the mem entry structures now, and put them in the free 254 * queue. 255 */ 256 vm_page_array = (vm_page_t) vaddr; 257 mapped = vaddr; 258 259 /* 260 * Validate these addresses. 261 */ 262 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 263 mapped = pmap_map(mapped, new_end, end, 264 VM_PROT_READ | VM_PROT_WRITE); 265 266 /* 267 * Clear all of the page structures 268 */ 269 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 270 vm_page_array_size = page_range; 271 272 /* 273 * Construct the free queue(s) in ascending order (by physical 274 * address) so that the first 16MB of physical memory is allocated 275 * last rather than first. On large-memory machines, this avoids 276 * the exhaustion of low physical memory before isa_dmainit has run. 277 */ 278 vmstats.v_page_count = 0; 279 vmstats.v_free_count = 0; 280 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 281 pa = phys_avail[i]; 282 if (i == biggestone) 283 last_pa = new_end; 284 else 285 last_pa = phys_avail[i + 1]; 286 while (pa < last_pa && npages-- > 0) { 287 vm_add_new_page(pa); 288 pa += PAGE_SIZE; 289 } 290 } 291 return (mapped); 292 } 293 294 /* 295 * Scan comparison function for Red-Black tree scans. An inclusive 296 * (start,end) is expected. Other fields are not used. 297 */ 298 int 299 rb_vm_page_scancmp(struct vm_page *p, void *data) 300 { 301 struct rb_vm_page_scan_info *info = data; 302 303 if (p->pindex < info->start_pindex) 304 return(-1); 305 if (p->pindex > info->end_pindex) 306 return(1); 307 return(0); 308 } 309 310 int 311 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 312 { 313 if (p1->pindex < p2->pindex) 314 return(-1); 315 if (p1->pindex > p2->pindex) 316 return(1); 317 return(0); 318 } 319 320 /* 321 * The opposite of vm_page_hold(). A page can be freed while being held, 322 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 323 * in this case to actually free it once the hold count drops to 0. 324 * 325 * This routine must be called at splvm(). 326 */ 327 void 328 vm_page_unhold(vm_page_t mem) 329 { 330 --mem->hold_count; 331 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 332 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) { 333 vm_page_busy(mem); 334 vm_page_free_toq(mem); 335 } 336 } 337 338 /* 339 * Inserts the given mem entry into the object and object list. 340 * 341 * The pagetables are not updated but will presumably fault the page 342 * in if necessary, or if a kernel page the caller will at some point 343 * enter the page into the kernel's pmap. We are not allowed to block 344 * here so we *can't* do this anyway. 345 * 346 * This routine may not block. 347 * This routine must be called with a critical section held. 348 */ 349 void 350 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 351 { 352 ASSERT_IN_CRIT_SECTION(); 353 if (m->object != NULL) 354 panic("vm_page_insert: already inserted"); 355 356 /* 357 * Record the object/offset pair in this page 358 */ 359 m->object = object; 360 m->pindex = pindex; 361 362 /* 363 * Insert it into the object. 364 */ 365 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 366 object->generation++; 367 368 /* 369 * show that the object has one more resident page. 370 */ 371 object->resident_page_count++; 372 373 /* 374 * Since we are inserting a new and possibly dirty page, 375 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 376 */ 377 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 378 vm_object_set_writeable_dirty(object); 379 } 380 381 /* 382 * Removes the given vm_page_t from the global (object,index) hash table 383 * and from the object's memq. 384 * 385 * The underlying pmap entry (if any) is NOT removed here. 386 * This routine may not block. 387 * 388 * The page must be BUSY and will remain BUSY on return. No spl needs to be 389 * held on call to this routine. 390 * 391 * note: FreeBSD side effect was to unbusy the page on return. We leave 392 * it busy. 393 */ 394 void 395 vm_page_remove(vm_page_t m) 396 { 397 vm_object_t object; 398 399 crit_enter(); 400 if (m->object == NULL) { 401 crit_exit(); 402 return; 403 } 404 405 if ((m->flags & PG_BUSY) == 0) 406 panic("vm_page_remove: page not busy"); 407 408 object = m->object; 409 410 /* 411 * Remove the page from the object and update the object. 412 */ 413 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 414 object->resident_page_count--; 415 object->generation++; 416 m->object = NULL; 417 418 crit_exit(); 419 } 420 421 /* 422 * Locate and return the page at (object, pindex), or NULL if the 423 * page could not be found. 424 * 425 * This routine will operate properly without spl protection, but 426 * the returned page could be in flux if it is busy. Because an 427 * interrupt can race a caller's busy check (unbusying and freeing the 428 * page we return before the caller is able to check the busy bit), 429 * the caller should generally call this routine with a critical 430 * section held. 431 * 432 * Callers may call this routine without spl protection if they know 433 * 'for sure' that the page will not be ripped out from under them 434 * by an interrupt. 435 */ 436 vm_page_t 437 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 438 { 439 vm_page_t m; 440 441 /* 442 * Search the hash table for this object/offset pair 443 */ 444 crit_enter(); 445 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 446 crit_exit(); 447 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 448 return(m); 449 } 450 451 /* 452 * vm_page_rename() 453 * 454 * Move the given memory entry from its current object to the specified 455 * target object/offset. 456 * 457 * The object must be locked. 458 * This routine may not block. 459 * 460 * Note: This routine will raise itself to splvm(), the caller need not. 461 * 462 * Note: Swap associated with the page must be invalidated by the move. We 463 * have to do this for several reasons: (1) we aren't freeing the 464 * page, (2) we are dirtying the page, (3) the VM system is probably 465 * moving the page from object A to B, and will then later move 466 * the backing store from A to B and we can't have a conflict. 467 * 468 * Note: We *always* dirty the page. It is necessary both for the 469 * fact that we moved it, and because we may be invalidating 470 * swap. If the page is on the cache, we have to deactivate it 471 * or vm_page_dirty() will panic. Dirty pages are not allowed 472 * on the cache. 473 */ 474 void 475 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 476 { 477 crit_enter(); 478 vm_page_remove(m); 479 vm_page_insert(m, new_object, new_pindex); 480 if (m->queue - m->pc == PQ_CACHE) 481 vm_page_deactivate(m); 482 vm_page_dirty(m); 483 vm_page_wakeup(m); 484 crit_exit(); 485 } 486 487 /* 488 * vm_page_unqueue() without any wakeup. This routine is used when a page 489 * is being moved between queues or otherwise is to remain BUSYied by the 490 * caller. 491 * 492 * This routine must be called at splhigh(). 493 * This routine may not block. 494 */ 495 void 496 vm_page_unqueue_nowakeup(vm_page_t m) 497 { 498 int queue = m->queue; 499 struct vpgqueues *pq; 500 501 if (queue != PQ_NONE) { 502 pq = &vm_page_queues[queue]; 503 m->queue = PQ_NONE; 504 TAILQ_REMOVE(&pq->pl, m, pageq); 505 (*pq->cnt)--; 506 pq->lcnt--; 507 } 508 } 509 510 /* 511 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 512 * if necessary. 513 * 514 * This routine must be called at splhigh(). 515 * This routine may not block. 516 */ 517 void 518 vm_page_unqueue(vm_page_t m) 519 { 520 int queue = m->queue; 521 struct vpgqueues *pq; 522 523 if (queue != PQ_NONE) { 524 m->queue = PQ_NONE; 525 pq = &vm_page_queues[queue]; 526 TAILQ_REMOVE(&pq->pl, m, pageq); 527 (*pq->cnt)--; 528 pq->lcnt--; 529 if ((queue - m->pc) == PQ_CACHE) { 530 if (vm_paging_needed()) 531 pagedaemon_wakeup(); 532 } 533 } 534 } 535 536 /* 537 * vm_page_list_find() 538 * 539 * Find a page on the specified queue with color optimization. 540 * 541 * The page coloring optimization attempts to locate a page that does 542 * not overload other nearby pages in the object in the cpu's L1 or L2 543 * caches. We need this optimization because cpu caches tend to be 544 * physical caches, while object spaces tend to be virtual. 545 * 546 * This routine must be called at splvm(). 547 * This routine may not block. 548 * 549 * Note that this routine is carefully inlined. A non-inlined version 550 * is available for outside callers but the only critical path is 551 * from within this source file. 552 */ 553 static __inline 554 vm_page_t 555 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 556 { 557 vm_page_t m; 558 559 if (prefer_zero) 560 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 561 else 562 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 563 if (m == NULL) 564 m = _vm_page_list_find2(basequeue, index); 565 return(m); 566 } 567 568 static vm_page_t 569 _vm_page_list_find2(int basequeue, int index) 570 { 571 int i; 572 vm_page_t m = NULL; 573 struct vpgqueues *pq; 574 575 pq = &vm_page_queues[basequeue]; 576 577 /* 578 * Note that for the first loop, index+i and index-i wind up at the 579 * same place. Even though this is not totally optimal, we've already 580 * blown it by missing the cache case so we do not care. 581 */ 582 583 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 584 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 585 break; 586 587 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 588 break; 589 } 590 return(m); 591 } 592 593 vm_page_t 594 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 595 { 596 return(_vm_page_list_find(basequeue, index, prefer_zero)); 597 } 598 599 /* 600 * Find a page on the cache queue with color optimization. As pages 601 * might be found, but not applicable, they are deactivated. This 602 * keeps us from using potentially busy cached pages. 603 * 604 * This routine must be called with a critical section held. 605 * This routine may not block. 606 */ 607 vm_page_t 608 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 609 { 610 vm_page_t m; 611 612 while (TRUE) { 613 m = _vm_page_list_find( 614 PQ_CACHE, 615 (pindex + object->pg_color) & PQ_L2_MASK, 616 FALSE 617 ); 618 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 619 m->hold_count || m->wire_count)) { 620 vm_page_deactivate(m); 621 continue; 622 } 623 return m; 624 } 625 /* not reached */ 626 } 627 628 /* 629 * Find a free or zero page, with specified preference. We attempt to 630 * inline the nominal case and fall back to _vm_page_select_free() 631 * otherwise. 632 * 633 * This routine must be called with a critical section held. 634 * This routine may not block. 635 */ 636 static __inline vm_page_t 637 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 638 { 639 vm_page_t m; 640 641 m = _vm_page_list_find( 642 PQ_FREE, 643 (pindex + object->pg_color) & PQ_L2_MASK, 644 prefer_zero 645 ); 646 return(m); 647 } 648 649 /* 650 * vm_page_alloc() 651 * 652 * Allocate and return a memory cell associated with this VM object/offset 653 * pair. 654 * 655 * page_req classes: 656 * 657 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 658 * VM_ALLOC_SYSTEM greater free drain 659 * VM_ALLOC_INTERRUPT allow free list to be completely drained 660 * VM_ALLOC_ZERO advisory request for pre-zero'd page 661 * 662 * The object must be locked. 663 * This routine may not block. 664 * The returned page will be marked PG_BUSY 665 * 666 * Additional special handling is required when called from an interrupt 667 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 668 * in this case. 669 */ 670 vm_page_t 671 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 672 { 673 vm_page_t m = NULL; 674 675 KKASSERT(object != NULL); 676 KASSERT(!vm_page_lookup(object, pindex), 677 ("vm_page_alloc: page already allocated")); 678 KKASSERT(page_req & 679 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 680 681 /* 682 * Certain system threads (pageout daemon, buf_daemon's) are 683 * allowed to eat deeper into the free page list. 684 */ 685 if (curthread->td_flags & TDF_SYSTHREAD) 686 page_req |= VM_ALLOC_SYSTEM; 687 688 crit_enter(); 689 loop: 690 if (vmstats.v_free_count > vmstats.v_free_reserved || 691 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 692 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 693 vmstats.v_free_count > vmstats.v_interrupt_free_min) 694 ) { 695 /* 696 * The free queue has sufficient free pages to take one out. 697 */ 698 if (page_req & VM_ALLOC_ZERO) 699 m = vm_page_select_free(object, pindex, TRUE); 700 else 701 m = vm_page_select_free(object, pindex, FALSE); 702 } else if (page_req & VM_ALLOC_NORMAL) { 703 /* 704 * Allocatable from the cache (non-interrupt only). On 705 * success, we must free the page and try again, thus 706 * ensuring that vmstats.v_*_free_min counters are replenished. 707 */ 708 #ifdef INVARIANTS 709 if (curthread->td_preempted) { 710 kprintf("vm_page_alloc(): warning, attempt to allocate" 711 " cache page from preempting interrupt\n"); 712 m = NULL; 713 } else { 714 m = vm_page_select_cache(object, pindex); 715 } 716 #else 717 m = vm_page_select_cache(object, pindex); 718 #endif 719 /* 720 * On success move the page into the free queue and loop. 721 */ 722 if (m != NULL) { 723 KASSERT(m->dirty == 0, 724 ("Found dirty cache page %p", m)); 725 vm_page_busy(m); 726 vm_page_protect(m, VM_PROT_NONE); 727 vm_page_free(m); 728 goto loop; 729 } 730 731 /* 732 * On failure return NULL 733 */ 734 crit_exit(); 735 #if defined(DIAGNOSTIC) 736 if (vmstats.v_cache_count > 0) 737 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 738 #endif 739 vm_pageout_deficit++; 740 pagedaemon_wakeup(); 741 return (NULL); 742 } else { 743 /* 744 * No pages available, wakeup the pageout daemon and give up. 745 */ 746 crit_exit(); 747 vm_pageout_deficit++; 748 pagedaemon_wakeup(); 749 return (NULL); 750 } 751 752 /* 753 * Good page found. The page has not yet been busied. We are in 754 * a critical section. 755 */ 756 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 757 KASSERT(m->dirty == 0, 758 ("vm_page_alloc: free/cache page %p was dirty", m)); 759 760 /* 761 * Remove from free queue 762 */ 763 vm_page_unqueue_nowakeup(m); 764 765 /* 766 * Initialize structure. Only the PG_ZERO flag is inherited. Set 767 * the page PG_BUSY 768 */ 769 if (m->flags & PG_ZERO) { 770 vm_page_zero_count--; 771 m->flags = PG_ZERO | PG_BUSY; 772 } else { 773 m->flags = PG_BUSY; 774 } 775 m->wire_count = 0; 776 m->hold_count = 0; 777 m->act_count = 0; 778 m->busy = 0; 779 m->valid = 0; 780 781 /* 782 * vm_page_insert() is safe prior to the crit_exit(). Note also that 783 * inserting a page here does not insert it into the pmap (which 784 * could cause us to block allocating memory). We cannot block 785 * anywhere. 786 */ 787 vm_page_insert(m, object, pindex); 788 789 /* 790 * Don't wakeup too often - wakeup the pageout daemon when 791 * we would be nearly out of memory. 792 */ 793 if (vm_paging_needed()) 794 pagedaemon_wakeup(); 795 796 crit_exit(); 797 798 /* 799 * A PG_BUSY page is returned. 800 */ 801 return (m); 802 } 803 804 /* 805 * Block until free pages are available for allocation, called in various 806 * places before memory allocations. 807 */ 808 void 809 vm_wait(int timo) 810 { 811 crit_enter(); 812 if (curthread == pagethread) { 813 vm_pageout_pages_needed = 1; 814 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 815 } else { 816 if (!vm_pages_needed) { 817 vm_pages_needed = 1; 818 wakeup(&vm_pages_needed); 819 } 820 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 821 } 822 crit_exit(); 823 } 824 825 /* 826 * Block until free pages are available for allocation 827 * 828 * Called only in vm_fault so that processes page faulting can be 829 * easily tracked. 830 * 831 * Sleeps at a lower priority than vm_wait() so that vm_wait()ing 832 * processes will be able to grab memory first. Do not change 833 * this balance without careful testing first. 834 */ 835 void 836 vm_waitpfault(void) 837 { 838 crit_enter(); 839 if (!vm_pages_needed) { 840 vm_pages_needed = 1; 841 wakeup(&vm_pages_needed); 842 } 843 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 844 crit_exit(); 845 } 846 847 /* 848 * Put the specified page on the active list (if appropriate). Ensure 849 * that act_count is at least ACT_INIT but do not otherwise mess with it. 850 * 851 * The page queues must be locked. 852 * This routine may not block. 853 */ 854 void 855 vm_page_activate(vm_page_t m) 856 { 857 crit_enter(); 858 if (m->queue != PQ_ACTIVE) { 859 if ((m->queue - m->pc) == PQ_CACHE) 860 mycpu->gd_cnt.v_reactivated++; 861 862 vm_page_unqueue(m); 863 864 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 865 m->queue = PQ_ACTIVE; 866 vm_page_queues[PQ_ACTIVE].lcnt++; 867 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 868 m, pageq); 869 if (m->act_count < ACT_INIT) 870 m->act_count = ACT_INIT; 871 vmstats.v_active_count++; 872 } 873 } else { 874 if (m->act_count < ACT_INIT) 875 m->act_count = ACT_INIT; 876 } 877 crit_exit(); 878 } 879 880 /* 881 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 882 * routine is called when a page has been added to the cache or free 883 * queues. 884 * 885 * This routine may not block. 886 * This routine must be called at splvm() 887 */ 888 static __inline void 889 vm_page_free_wakeup(void) 890 { 891 /* 892 * if pageout daemon needs pages, then tell it that there are 893 * some free. 894 */ 895 if (vm_pageout_pages_needed && 896 vmstats.v_cache_count + vmstats.v_free_count >= 897 vmstats.v_pageout_free_min 898 ) { 899 wakeup(&vm_pageout_pages_needed); 900 vm_pageout_pages_needed = 0; 901 } 902 903 /* 904 * wakeup processes that are waiting on memory if we hit a 905 * high water mark. And wakeup scheduler process if we have 906 * lots of memory. this process will swapin processes. 907 */ 908 if (vm_pages_needed && !vm_page_count_min()) { 909 vm_pages_needed = 0; 910 wakeup(&vmstats.v_free_count); 911 } 912 } 913 914 /* 915 * vm_page_free_toq: 916 * 917 * Returns the given page to the PQ_FREE list, disassociating it with 918 * any VM object. 919 * 920 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 921 * return (the page will have been freed). No particular spl is required 922 * on entry. 923 * 924 * This routine may not block. 925 */ 926 void 927 vm_page_free_toq(vm_page_t m) 928 { 929 struct vpgqueues *pq; 930 931 crit_enter(); 932 mycpu->gd_cnt.v_tfree++; 933 934 KKASSERT((m->flags & PG_MAPPED) == 0); 935 936 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 937 kprintf( 938 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 939 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 940 m->hold_count); 941 if ((m->queue - m->pc) == PQ_FREE) 942 panic("vm_page_free: freeing free page"); 943 else 944 panic("vm_page_free: freeing busy page"); 945 } 946 947 /* 948 * unqueue, then remove page. Note that we cannot destroy 949 * the page here because we do not want to call the pager's 950 * callback routine until after we've put the page on the 951 * appropriate free queue. 952 */ 953 vm_page_unqueue_nowakeup(m); 954 vm_page_remove(m); 955 956 /* 957 * No further management of fictitious pages occurs beyond object 958 * and queue removal. 959 */ 960 if ((m->flags & PG_FICTITIOUS) != 0) { 961 vm_page_wakeup(m); 962 crit_exit(); 963 return; 964 } 965 966 m->valid = 0; 967 vm_page_undirty(m); 968 969 if (m->wire_count != 0) { 970 if (m->wire_count > 1) { 971 panic( 972 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 973 m->wire_count, (long)m->pindex); 974 } 975 panic("vm_page_free: freeing wired page"); 976 } 977 978 /* 979 * Clear the UNMANAGED flag when freeing an unmanaged page. 980 */ 981 if (m->flags & PG_UNMANAGED) { 982 m->flags &= ~PG_UNMANAGED; 983 } 984 985 if (m->hold_count != 0) { 986 m->flags &= ~PG_ZERO; 987 m->queue = PQ_HOLD; 988 } else { 989 m->queue = PQ_FREE + m->pc; 990 } 991 pq = &vm_page_queues[m->queue]; 992 pq->lcnt++; 993 ++(*pq->cnt); 994 995 /* 996 * Put zero'd pages on the end ( where we look for zero'd pages 997 * first ) and non-zerod pages at the head. 998 */ 999 if (m->flags & PG_ZERO) { 1000 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1001 ++vm_page_zero_count; 1002 } else { 1003 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1004 } 1005 vm_page_wakeup(m); 1006 vm_page_free_wakeup(); 1007 crit_exit(); 1008 } 1009 1010 /* 1011 * vm_page_unmanage() 1012 * 1013 * Prevent PV management from being done on the page. The page is 1014 * removed from the paging queues as if it were wired, and as a 1015 * consequence of no longer being managed the pageout daemon will not 1016 * touch it (since there is no way to locate the pte mappings for the 1017 * page). madvise() calls that mess with the pmap will also no longer 1018 * operate on the page. 1019 * 1020 * Beyond that the page is still reasonably 'normal'. Freeing the page 1021 * will clear the flag. 1022 * 1023 * This routine is used by OBJT_PHYS objects - objects using unswappable 1024 * physical memory as backing store rather then swap-backed memory and 1025 * will eventually be extended to support 4MB unmanaged physical 1026 * mappings. 1027 * 1028 * Must be called with a critical section held. 1029 */ 1030 void 1031 vm_page_unmanage(vm_page_t m) 1032 { 1033 ASSERT_IN_CRIT_SECTION(); 1034 if ((m->flags & PG_UNMANAGED) == 0) { 1035 if (m->wire_count == 0) 1036 vm_page_unqueue(m); 1037 } 1038 vm_page_flag_set(m, PG_UNMANAGED); 1039 } 1040 1041 /* 1042 * Mark this page as wired down by yet another map, removing it from 1043 * paging queues as necessary. 1044 * 1045 * The page queues must be locked. 1046 * This routine may not block. 1047 */ 1048 void 1049 vm_page_wire(vm_page_t m) 1050 { 1051 /* 1052 * Only bump the wire statistics if the page is not already wired, 1053 * and only unqueue the page if it is on some queue (if it is unmanaged 1054 * it is already off the queues). Don't do anything with fictitious 1055 * pages because they are always wired. 1056 */ 1057 crit_enter(); 1058 if ((m->flags & PG_FICTITIOUS) == 0) { 1059 if (m->wire_count == 0) { 1060 if ((m->flags & PG_UNMANAGED) == 0) 1061 vm_page_unqueue(m); 1062 vmstats.v_wire_count++; 1063 } 1064 m->wire_count++; 1065 KASSERT(m->wire_count != 0, 1066 ("vm_page_wire: wire_count overflow m=%p", m)); 1067 } 1068 crit_exit(); 1069 } 1070 1071 /* 1072 * Release one wiring of this page, potentially enabling it to be paged again. 1073 * 1074 * Many pages placed on the inactive queue should actually go 1075 * into the cache, but it is difficult to figure out which. What 1076 * we do instead, if the inactive target is well met, is to put 1077 * clean pages at the head of the inactive queue instead of the tail. 1078 * This will cause them to be moved to the cache more quickly and 1079 * if not actively re-referenced, freed more quickly. If we just 1080 * stick these pages at the end of the inactive queue, heavy filesystem 1081 * meta-data accesses can cause an unnecessary paging load on memory bound 1082 * processes. This optimization causes one-time-use metadata to be 1083 * reused more quickly. 1084 * 1085 * BUT, if we are in a low-memory situation we have no choice but to 1086 * put clean pages on the cache queue. 1087 * 1088 * A number of routines use vm_page_unwire() to guarantee that the page 1089 * will go into either the inactive or active queues, and will NEVER 1090 * be placed in the cache - for example, just after dirtying a page. 1091 * dirty pages in the cache are not allowed. 1092 * 1093 * The page queues must be locked. 1094 * This routine may not block. 1095 */ 1096 void 1097 vm_page_unwire(vm_page_t m, int activate) 1098 { 1099 crit_enter(); 1100 if (m->flags & PG_FICTITIOUS) { 1101 /* do nothing */ 1102 } else if (m->wire_count <= 0) { 1103 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1104 } else { 1105 if (--m->wire_count == 0) { 1106 --vmstats.v_wire_count; 1107 if (m->flags & PG_UNMANAGED) { 1108 ; 1109 } else if (activate) { 1110 TAILQ_INSERT_TAIL( 1111 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1112 m->queue = PQ_ACTIVE; 1113 vm_page_queues[PQ_ACTIVE].lcnt++; 1114 vmstats.v_active_count++; 1115 } else { 1116 vm_page_flag_clear(m, PG_WINATCFLS); 1117 TAILQ_INSERT_TAIL( 1118 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1119 m->queue = PQ_INACTIVE; 1120 vm_page_queues[PQ_INACTIVE].lcnt++; 1121 vmstats.v_inactive_count++; 1122 } 1123 } 1124 } 1125 crit_exit(); 1126 } 1127 1128 1129 /* 1130 * Move the specified page to the inactive queue. If the page has 1131 * any associated swap, the swap is deallocated. 1132 * 1133 * Normally athead is 0 resulting in LRU operation. athead is set 1134 * to 1 if we want this page to be 'as if it were placed in the cache', 1135 * except without unmapping it from the process address space. 1136 * 1137 * This routine may not block. 1138 */ 1139 static __inline void 1140 _vm_page_deactivate(vm_page_t m, int athead) 1141 { 1142 /* 1143 * Ignore if already inactive. 1144 */ 1145 if (m->queue == PQ_INACTIVE) 1146 return; 1147 1148 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1149 if ((m->queue - m->pc) == PQ_CACHE) 1150 mycpu->gd_cnt.v_reactivated++; 1151 vm_page_flag_clear(m, PG_WINATCFLS); 1152 vm_page_unqueue(m); 1153 if (athead) 1154 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1155 else 1156 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1157 m->queue = PQ_INACTIVE; 1158 vm_page_queues[PQ_INACTIVE].lcnt++; 1159 vmstats.v_inactive_count++; 1160 } 1161 } 1162 1163 void 1164 vm_page_deactivate(vm_page_t m) 1165 { 1166 crit_enter(); 1167 _vm_page_deactivate(m, 0); 1168 crit_exit(); 1169 } 1170 1171 /* 1172 * vm_page_try_to_cache: 1173 * 1174 * Returns 0 on failure, 1 on success 1175 */ 1176 int 1177 vm_page_try_to_cache(vm_page_t m) 1178 { 1179 crit_enter(); 1180 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1181 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1182 crit_exit(); 1183 return(0); 1184 } 1185 vm_page_test_dirty(m); 1186 if (m->dirty) { 1187 crit_exit(); 1188 return(0); 1189 } 1190 vm_page_cache(m); 1191 crit_exit(); 1192 return(1); 1193 } 1194 1195 /* 1196 * Attempt to free the page. If we cannot free it, we do nothing. 1197 * 1 is returned on success, 0 on failure. 1198 */ 1199 int 1200 vm_page_try_to_free(vm_page_t m) 1201 { 1202 crit_enter(); 1203 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1204 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1205 crit_exit(); 1206 return(0); 1207 } 1208 vm_page_test_dirty(m); 1209 if (m->dirty) { 1210 crit_exit(); 1211 return(0); 1212 } 1213 vm_page_busy(m); 1214 vm_page_protect(m, VM_PROT_NONE); 1215 vm_page_free(m); 1216 crit_exit(); 1217 return(1); 1218 } 1219 1220 /* 1221 * vm_page_cache 1222 * 1223 * Put the specified page onto the page cache queue (if appropriate). 1224 * 1225 * This routine may not block. 1226 */ 1227 void 1228 vm_page_cache(vm_page_t m) 1229 { 1230 ASSERT_IN_CRIT_SECTION(); 1231 1232 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1233 m->wire_count || m->hold_count) { 1234 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1235 return; 1236 } 1237 1238 /* 1239 * Already in the cache (and thus not mapped) 1240 */ 1241 if ((m->queue - m->pc) == PQ_CACHE) { 1242 KKASSERT((m->flags & PG_MAPPED) == 0); 1243 return; 1244 } 1245 1246 /* 1247 * Caller is required to test m->dirty, but note that the act of 1248 * removing the page from its maps can cause it to become dirty 1249 * on an SMP system due to another cpu running in usermode. 1250 */ 1251 if (m->dirty) { 1252 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1253 (long)m->pindex); 1254 } 1255 1256 /* 1257 * Remove all pmaps and indicate that the page is not 1258 * writeable or mapped. Our vm_page_protect() call may 1259 * have blocked (especially w/ VM_PROT_NONE), so recheck 1260 * everything. 1261 */ 1262 vm_page_busy(m); 1263 vm_page_protect(m, VM_PROT_NONE); 1264 vm_page_wakeup(m); 1265 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy || 1266 m->wire_count || m->hold_count) { 1267 /* do nothing */ 1268 } else if (m->dirty) { 1269 vm_page_deactivate(m); 1270 } else { 1271 vm_page_unqueue_nowakeup(m); 1272 m->queue = PQ_CACHE + m->pc; 1273 vm_page_queues[m->queue].lcnt++; 1274 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1275 vmstats.v_cache_count++; 1276 vm_page_free_wakeup(); 1277 } 1278 } 1279 1280 /* 1281 * vm_page_dontneed() 1282 * 1283 * Cache, deactivate, or do nothing as appropriate. This routine 1284 * is typically used by madvise() MADV_DONTNEED. 1285 * 1286 * Generally speaking we want to move the page into the cache so 1287 * it gets reused quickly. However, this can result in a silly syndrome 1288 * due to the page recycling too quickly. Small objects will not be 1289 * fully cached. On the otherhand, if we move the page to the inactive 1290 * queue we wind up with a problem whereby very large objects 1291 * unnecessarily blow away our inactive and cache queues. 1292 * 1293 * The solution is to move the pages based on a fixed weighting. We 1294 * either leave them alone, deactivate them, or move them to the cache, 1295 * where moving them to the cache has the highest weighting. 1296 * By forcing some pages into other queues we eventually force the 1297 * system to balance the queues, potentially recovering other unrelated 1298 * space from active. The idea is to not force this to happen too 1299 * often. 1300 */ 1301 void 1302 vm_page_dontneed(vm_page_t m) 1303 { 1304 static int dnweight; 1305 int dnw; 1306 int head; 1307 1308 dnw = ++dnweight; 1309 1310 /* 1311 * occassionally leave the page alone 1312 */ 1313 crit_enter(); 1314 if ((dnw & 0x01F0) == 0 || 1315 m->queue == PQ_INACTIVE || 1316 m->queue - m->pc == PQ_CACHE 1317 ) { 1318 if (m->act_count >= ACT_INIT) 1319 --m->act_count; 1320 crit_exit(); 1321 return; 1322 } 1323 1324 if (m->dirty == 0) 1325 vm_page_test_dirty(m); 1326 1327 if (m->dirty || (dnw & 0x0070) == 0) { 1328 /* 1329 * Deactivate the page 3 times out of 32. 1330 */ 1331 head = 0; 1332 } else { 1333 /* 1334 * Cache the page 28 times out of every 32. Note that 1335 * the page is deactivated instead of cached, but placed 1336 * at the head of the queue instead of the tail. 1337 */ 1338 head = 1; 1339 } 1340 _vm_page_deactivate(m, head); 1341 crit_exit(); 1342 } 1343 1344 /* 1345 * Grab a page, blocking if it is busy and allocating a page if necessary. 1346 * A busy page is returned or NULL. 1347 * 1348 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1349 * If VM_ALLOC_RETRY is not specified 1350 * 1351 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1352 * always returned if we had blocked. 1353 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1354 * This routine may not be called from an interrupt. 1355 * The returned page may not be entirely valid. 1356 * 1357 * This routine may be called from mainline code without spl protection and 1358 * be guarenteed a busied page associated with the object at the specified 1359 * index. 1360 */ 1361 vm_page_t 1362 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1363 { 1364 vm_page_t m; 1365 int generation; 1366 1367 KKASSERT(allocflags & 1368 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1369 crit_enter(); 1370 retrylookup: 1371 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1372 if (m->busy || (m->flags & PG_BUSY)) { 1373 generation = object->generation; 1374 1375 while ((object->generation == generation) && 1376 (m->busy || (m->flags & PG_BUSY))) { 1377 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1378 tsleep(m, 0, "pgrbwt", 0); 1379 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1380 m = NULL; 1381 goto done; 1382 } 1383 } 1384 goto retrylookup; 1385 } else { 1386 vm_page_busy(m); 1387 goto done; 1388 } 1389 } 1390 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1391 if (m == NULL) { 1392 vm_wait(0); 1393 if ((allocflags & VM_ALLOC_RETRY) == 0) 1394 goto done; 1395 goto retrylookup; 1396 } 1397 done: 1398 crit_exit(); 1399 return(m); 1400 } 1401 1402 /* 1403 * Mapping function for valid bits or for dirty bits in 1404 * a page. May not block. 1405 * 1406 * Inputs are required to range within a page. 1407 */ 1408 __inline int 1409 vm_page_bits(int base, int size) 1410 { 1411 int first_bit; 1412 int last_bit; 1413 1414 KASSERT( 1415 base + size <= PAGE_SIZE, 1416 ("vm_page_bits: illegal base/size %d/%d", base, size) 1417 ); 1418 1419 if (size == 0) /* handle degenerate case */ 1420 return(0); 1421 1422 first_bit = base >> DEV_BSHIFT; 1423 last_bit = (base + size - 1) >> DEV_BSHIFT; 1424 1425 return ((2 << last_bit) - (1 << first_bit)); 1426 } 1427 1428 /* 1429 * Sets portions of a page valid and clean. The arguments are expected 1430 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1431 * of any partial chunks touched by the range. The invalid portion of 1432 * such chunks will be zero'd. 1433 * 1434 * This routine may not block. 1435 * 1436 * (base + size) must be less then or equal to PAGE_SIZE. 1437 */ 1438 void 1439 vm_page_set_validclean(vm_page_t m, int base, int size) 1440 { 1441 int pagebits; 1442 int frag; 1443 int endoff; 1444 1445 if (size == 0) /* handle degenerate case */ 1446 return; 1447 1448 /* 1449 * If the base is not DEV_BSIZE aligned and the valid 1450 * bit is clear, we have to zero out a portion of the 1451 * first block. 1452 */ 1453 1454 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1455 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1456 ) { 1457 pmap_zero_page_area( 1458 VM_PAGE_TO_PHYS(m), 1459 frag, 1460 base - frag 1461 ); 1462 } 1463 1464 /* 1465 * If the ending offset is not DEV_BSIZE aligned and the 1466 * valid bit is clear, we have to zero out a portion of 1467 * the last block. 1468 */ 1469 1470 endoff = base + size; 1471 1472 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1473 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1474 ) { 1475 pmap_zero_page_area( 1476 VM_PAGE_TO_PHYS(m), 1477 endoff, 1478 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1479 ); 1480 } 1481 1482 /* 1483 * Set valid, clear dirty bits. If validating the entire 1484 * page we can safely clear the pmap modify bit. We also 1485 * use this opportunity to clear the PG_NOSYNC flag. If a process 1486 * takes a write fault on a MAP_NOSYNC memory area the flag will 1487 * be set again. 1488 * 1489 * We set valid bits inclusive of any overlap, but we can only 1490 * clear dirty bits for DEV_BSIZE chunks that are fully within 1491 * the range. 1492 */ 1493 1494 pagebits = vm_page_bits(base, size); 1495 m->valid |= pagebits; 1496 #if 0 /* NOT YET */ 1497 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1498 frag = DEV_BSIZE - frag; 1499 base += frag; 1500 size -= frag; 1501 if (size < 0) 1502 size = 0; 1503 } 1504 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1505 #endif 1506 m->dirty &= ~pagebits; 1507 if (base == 0 && size == PAGE_SIZE) { 1508 pmap_clear_modify(m); 1509 vm_page_flag_clear(m, PG_NOSYNC); 1510 } 1511 } 1512 1513 void 1514 vm_page_clear_dirty(vm_page_t m, int base, int size) 1515 { 1516 m->dirty &= ~vm_page_bits(base, size); 1517 } 1518 1519 /* 1520 * Make the page all-dirty. 1521 * 1522 * Also make sure the related object and vnode reflect the fact that the 1523 * object may now contain a dirty page. 1524 */ 1525 void 1526 vm_page_dirty(vm_page_t m) 1527 { 1528 #ifdef INVARIANTS 1529 int pqtype = m->queue - m->pc; 1530 #endif 1531 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1532 ("vm_page_dirty: page in free/cache queue!")); 1533 if (m->dirty != VM_PAGE_BITS_ALL) { 1534 m->dirty = VM_PAGE_BITS_ALL; 1535 if (m->object) 1536 vm_object_set_writeable_dirty(m->object); 1537 } 1538 } 1539 1540 /* 1541 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1542 * valid and dirty bits for the effected areas are cleared. 1543 * 1544 * May not block. 1545 */ 1546 void 1547 vm_page_set_invalid(vm_page_t m, int base, int size) 1548 { 1549 int bits; 1550 1551 bits = vm_page_bits(base, size); 1552 m->valid &= ~bits; 1553 m->dirty &= ~bits; 1554 m->object->generation++; 1555 } 1556 1557 /* 1558 * The kernel assumes that the invalid portions of a page contain 1559 * garbage, but such pages can be mapped into memory by user code. 1560 * When this occurs, we must zero out the non-valid portions of the 1561 * page so user code sees what it expects. 1562 * 1563 * Pages are most often semi-valid when the end of a file is mapped 1564 * into memory and the file's size is not page aligned. 1565 */ 1566 void 1567 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1568 { 1569 int b; 1570 int i; 1571 1572 /* 1573 * Scan the valid bits looking for invalid sections that 1574 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1575 * valid bit may be set ) have already been zerod by 1576 * vm_page_set_validclean(). 1577 */ 1578 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1579 if (i == (PAGE_SIZE / DEV_BSIZE) || 1580 (m->valid & (1 << i)) 1581 ) { 1582 if (i > b) { 1583 pmap_zero_page_area( 1584 VM_PAGE_TO_PHYS(m), 1585 b << DEV_BSHIFT, 1586 (i - b) << DEV_BSHIFT 1587 ); 1588 } 1589 b = i + 1; 1590 } 1591 } 1592 1593 /* 1594 * setvalid is TRUE when we can safely set the zero'd areas 1595 * as being valid. We can do this if there are no cache consistency 1596 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1597 */ 1598 if (setvalid) 1599 m->valid = VM_PAGE_BITS_ALL; 1600 } 1601 1602 /* 1603 * Is a (partial) page valid? Note that the case where size == 0 1604 * will return FALSE in the degenerate case where the page is entirely 1605 * invalid, and TRUE otherwise. 1606 * 1607 * May not block. 1608 */ 1609 int 1610 vm_page_is_valid(vm_page_t m, int base, int size) 1611 { 1612 int bits = vm_page_bits(base, size); 1613 1614 if (m->valid && ((m->valid & bits) == bits)) 1615 return 1; 1616 else 1617 return 0; 1618 } 1619 1620 /* 1621 * update dirty bits from pmap/mmu. May not block. 1622 */ 1623 void 1624 vm_page_test_dirty(vm_page_t m) 1625 { 1626 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1627 vm_page_dirty(m); 1628 } 1629 } 1630 1631 /* 1632 * Issue an event on a VM page. Corresponding action structures are 1633 * removed from the page's list and called. 1634 */ 1635 void 1636 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1637 { 1638 struct vm_page_action *scan, *next; 1639 1640 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) { 1641 if (scan->event == event) { 1642 scan->event = VMEVENT_NONE; 1643 LIST_REMOVE(scan, entry); 1644 scan->func(m, scan); 1645 } 1646 } 1647 } 1648 1649 #include "opt_ddb.h" 1650 #ifdef DDB 1651 #include <sys/kernel.h> 1652 1653 #include <ddb/ddb.h> 1654 1655 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1656 { 1657 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1658 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1659 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1660 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1661 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1662 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1663 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1664 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1665 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1666 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1667 } 1668 1669 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1670 { 1671 int i; 1672 db_printf("PQ_FREE:"); 1673 for(i=0;i<PQ_L2_SIZE;i++) { 1674 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1675 } 1676 db_printf("\n"); 1677 1678 db_printf("PQ_CACHE:"); 1679 for(i=0;i<PQ_L2_SIZE;i++) { 1680 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1681 } 1682 db_printf("\n"); 1683 1684 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1685 vm_page_queues[PQ_ACTIVE].lcnt, 1686 vm_page_queues[PQ_INACTIVE].lcnt); 1687 } 1688 #endif /* DDB */ 1689