1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 39 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 40 * $DragonFly: src/sys/vm/vm_page.c,v 1.40 2008/08/25 17:01:42 dillon Exp $ 41 */ 42 43 /* 44 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 45 * All rights reserved. 46 * 47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 /* 70 * Resident memory management module. The module manipulates 'VM pages'. 71 * A VM page is the core building block for memory management. 72 */ 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/malloc.h> 77 #include <sys/proc.h> 78 #include <sys/vmmeter.h> 79 #include <sys/vnode.h> 80 81 #include <vm/vm.h> 82 #include <vm/vm_param.h> 83 #include <sys/lock.h> 84 #include <vm/vm_kern.h> 85 #include <vm/pmap.h> 86 #include <vm/vm_map.h> 87 #include <vm/vm_object.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_pageout.h> 90 #include <vm/vm_pager.h> 91 #include <vm/vm_extern.h> 92 #include <vm/swap_pager.h> 93 94 #include <machine/md_var.h> 95 96 #include <vm/vm_page2.h> 97 #include <sys/mplock2.h> 98 99 static void vm_page_queue_init(void); 100 static void vm_page_free_wakeup(void); 101 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 102 static vm_page_t _vm_page_list_find2(int basequeue, int index); 103 104 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 105 106 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread)); 107 108 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 109 vm_pindex_t, pindex); 110 111 static void 112 vm_page_queue_init(void) 113 { 114 int i; 115 116 for (i = 0; i < PQ_L2_SIZE; i++) 117 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 118 for (i = 0; i < PQ_L2_SIZE; i++) 119 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 120 121 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 122 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 123 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 124 /* PQ_NONE has no queue */ 125 126 for (i = 0; i < PQ_COUNT; i++) 127 TAILQ_INIT(&vm_page_queues[i].pl); 128 } 129 130 /* 131 * note: place in initialized data section? Is this necessary? 132 */ 133 long first_page = 0; 134 int vm_page_array_size = 0; 135 int vm_page_zero_count = 0; 136 vm_page_t vm_page_array = 0; 137 138 /* 139 * (low level boot) 140 * 141 * Sets the page size, perhaps based upon the memory size. 142 * Must be called before any use of page-size dependent functions. 143 */ 144 void 145 vm_set_page_size(void) 146 { 147 if (vmstats.v_page_size == 0) 148 vmstats.v_page_size = PAGE_SIZE; 149 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 150 panic("vm_set_page_size: page size not a power of two"); 151 } 152 153 /* 154 * (low level boot) 155 * 156 * Add a new page to the freelist for use by the system. New pages 157 * are added to both the head and tail of the associated free page 158 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 159 * requests pull 'recent' adds (higher physical addresses) first. 160 * 161 * Must be called in a critical section. 162 */ 163 vm_page_t 164 vm_add_new_page(vm_paddr_t pa) 165 { 166 struct vpgqueues *vpq; 167 vm_page_t m; 168 169 ++vmstats.v_page_count; 170 ++vmstats.v_free_count; 171 m = PHYS_TO_VM_PAGE(pa); 172 m->phys_addr = pa; 173 m->flags = 0; 174 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 175 m->queue = m->pc + PQ_FREE; 176 KKASSERT(m->dirty == 0); 177 178 vpq = &vm_page_queues[m->queue]; 179 if (vpq->flipflop) 180 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 181 else 182 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 183 vpq->flipflop = 1 - vpq->flipflop; 184 185 vm_page_queues[m->queue].lcnt++; 186 return (m); 187 } 188 189 /* 190 * (low level boot) 191 * 192 * Initializes the resident memory module. 193 * 194 * Allocates memory for the page cells, and for the object/offset-to-page 195 * hash table headers. Each page cell is initialized and placed on the 196 * free list. 197 * 198 * starta/enda represents the range of physical memory addresses available 199 * for use (skipping memory already used by the kernel), subject to 200 * phys_avail[]. Note that phys_avail[] has already mapped out memory 201 * already in use by the kernel. 202 */ 203 vm_offset_t 204 vm_page_startup(vm_offset_t vaddr) 205 { 206 vm_offset_t mapped; 207 vm_size_t npages; 208 vm_paddr_t page_range; 209 vm_paddr_t new_end; 210 int i; 211 vm_paddr_t pa; 212 int nblocks; 213 vm_paddr_t last_pa; 214 vm_paddr_t end; 215 vm_paddr_t biggestone, biggestsize; 216 vm_paddr_t total; 217 218 total = 0; 219 biggestsize = 0; 220 biggestone = 0; 221 nblocks = 0; 222 vaddr = round_page(vaddr); 223 224 for (i = 0; phys_avail[i + 1]; i += 2) { 225 phys_avail[i] = round_page64(phys_avail[i]); 226 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 227 } 228 229 for (i = 0; phys_avail[i + 1]; i += 2) { 230 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 231 232 if (size > biggestsize) { 233 biggestone = i; 234 biggestsize = size; 235 } 236 ++nblocks; 237 total += size; 238 } 239 240 end = phys_avail[biggestone+1]; 241 end = trunc_page(end); 242 243 /* 244 * Initialize the queue headers for the free queue, the active queue 245 * and the inactive queue. 246 */ 247 248 vm_page_queue_init(); 249 250 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */ 251 #if !defined(_KERNEL_VIRTUAL) 252 /* 253 * Allocate a bitmap to indicate that a random physical page 254 * needs to be included in a minidump. 255 * 256 * The amd64 port needs this to indicate which direct map pages 257 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 258 * 259 * However, i386 still needs this workspace internally within the 260 * minidump code. In theory, they are not needed on i386, but are 261 * included should the sf_buf code decide to use them. 262 */ 263 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 264 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 265 end -= vm_page_dump_size; 266 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 267 VM_PROT_READ | VM_PROT_WRITE); 268 bzero((void *)vm_page_dump, vm_page_dump_size); 269 #endif 270 271 /* 272 * Compute the number of pages of memory that will be available for 273 * use (taking into account the overhead of a page structure per 274 * page). 275 */ 276 first_page = phys_avail[0] / PAGE_SIZE; 277 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 278 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 279 280 /* 281 * Initialize the mem entry structures now, and put them in the free 282 * queue. 283 */ 284 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 285 mapped = pmap_map(&vaddr, new_end, end, 286 VM_PROT_READ | VM_PROT_WRITE); 287 vm_page_array = (vm_page_t)mapped; 288 289 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 290 /* 291 * since pmap_map on amd64 returns stuff out of a direct-map region, 292 * we have to manually add these pages to the minidump tracking so 293 * that they can be dumped, including the vm_page_array. 294 */ 295 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 296 dump_add_page(pa); 297 #endif 298 299 /* 300 * Clear all of the page structures 301 */ 302 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 303 vm_page_array_size = page_range; 304 305 /* 306 * Construct the free queue(s) in ascending order (by physical 307 * address) so that the first 16MB of physical memory is allocated 308 * last rather than first. On large-memory machines, this avoids 309 * the exhaustion of low physical memory before isa_dmainit has run. 310 */ 311 vmstats.v_page_count = 0; 312 vmstats.v_free_count = 0; 313 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 314 pa = phys_avail[i]; 315 if (i == biggestone) 316 last_pa = new_end; 317 else 318 last_pa = phys_avail[i + 1]; 319 while (pa < last_pa && npages-- > 0) { 320 vm_add_new_page(pa); 321 pa += PAGE_SIZE; 322 } 323 } 324 return (vaddr); 325 } 326 327 /* 328 * Scan comparison function for Red-Black tree scans. An inclusive 329 * (start,end) is expected. Other fields are not used. 330 */ 331 int 332 rb_vm_page_scancmp(struct vm_page *p, void *data) 333 { 334 struct rb_vm_page_scan_info *info = data; 335 336 if (p->pindex < info->start_pindex) 337 return(-1); 338 if (p->pindex > info->end_pindex) 339 return(1); 340 return(0); 341 } 342 343 int 344 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 345 { 346 if (p1->pindex < p2->pindex) 347 return(-1); 348 if (p1->pindex > p2->pindex) 349 return(1); 350 return(0); 351 } 352 353 /* 354 * Holding a page keeps it from being reused. Other parts of the system 355 * can still disassociate the page from its current object and free it, or 356 * perform read or write I/O on it and/or otherwise manipulate the page, 357 * but if the page is held the VM system will leave the page and its data 358 * intact and not reuse the page for other purposes until the last hold 359 * reference is released. (see vm_page_wire() if you want to prevent the 360 * page from being disassociated from its object too). 361 * 362 * The caller must hold vm_token. 363 * 364 * The caller must still validate the contents of the page and, if necessary, 365 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 366 * before manipulating the page. 367 */ 368 void 369 vm_page_hold(vm_page_t m) 370 { 371 ASSERT_LWKT_TOKEN_HELD(&vm_token); 372 ++m->hold_count; 373 } 374 375 /* 376 * The opposite of vm_page_hold(). A page can be freed while being held, 377 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 378 * in this case to actually free it once the hold count drops to 0. 379 * 380 * The caller must hold vm_token if non-blocking operation is desired, 381 * but otherwise does not need to. 382 */ 383 void 384 vm_page_unhold(vm_page_t m) 385 { 386 lwkt_gettoken(&vm_token); 387 --m->hold_count; 388 KASSERT(m->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 389 if (m->hold_count == 0 && m->queue == PQ_HOLD) { 390 vm_page_busy(m); 391 vm_page_free_toq(m); 392 } 393 lwkt_reltoken(&vm_token); 394 } 395 396 /* 397 * Inserts the given vm_page into the object and object list. 398 * 399 * The pagetables are not updated but will presumably fault the page 400 * in if necessary, or if a kernel page the caller will at some point 401 * enter the page into the kernel's pmap. We are not allowed to block 402 * here so we *can't* do this anyway. 403 * 404 * This routine may not block. 405 * This routine must be called with the vm_token held. 406 * This routine must be called with a critical section held. 407 */ 408 void 409 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 410 { 411 ASSERT_IN_CRIT_SECTION(); 412 ASSERT_LWKT_TOKEN_HELD(&vm_token); 413 if (m->object != NULL) 414 panic("vm_page_insert: already inserted"); 415 416 /* 417 * Record the object/offset pair in this page 418 */ 419 m->object = object; 420 m->pindex = pindex; 421 422 /* 423 * Insert it into the object. 424 */ 425 ASSERT_MP_LOCK_HELD(curthread); 426 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 427 object->generation++; 428 429 /* 430 * show that the object has one more resident page. 431 */ 432 object->resident_page_count++; 433 434 /* 435 * Since we are inserting a new and possibly dirty page, 436 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 437 */ 438 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 439 vm_object_set_writeable_dirty(object); 440 441 /* 442 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 443 */ 444 swap_pager_page_inserted(m); 445 } 446 447 /* 448 * Removes the given vm_page_t from the global (object,index) hash table 449 * and from the object's memq. 450 * 451 * The underlying pmap entry (if any) is NOT removed here. 452 * This routine may not block. 453 * 454 * The page must be BUSY and will remain BUSY on return. 455 * No other requirements. 456 * 457 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 458 * it busy. 459 */ 460 void 461 vm_page_remove(vm_page_t m) 462 { 463 vm_object_t object; 464 465 crit_enter(); 466 lwkt_gettoken(&vm_token); 467 if (m->object == NULL) { 468 lwkt_reltoken(&vm_token); 469 crit_exit(); 470 return; 471 } 472 473 if ((m->flags & PG_BUSY) == 0) 474 panic("vm_page_remove: page not busy"); 475 476 object = m->object; 477 478 /* 479 * Remove the page from the object and update the object. 480 */ 481 ASSERT_MP_LOCK_HELD(curthread); 482 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 483 object->resident_page_count--; 484 object->generation++; 485 m->object = NULL; 486 487 lwkt_reltoken(&vm_token); 488 crit_exit(); 489 } 490 491 /* 492 * Locate and return the page at (object, pindex), or NULL if the 493 * page could not be found. 494 * 495 * The caller must hold vm_token if non-blocking operation is desired. 496 */ 497 vm_page_t 498 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 499 { 500 vm_page_t m; 501 502 /* 503 * Search the hash table for this object/offset pair 504 */ 505 ASSERT_MP_LOCK_HELD(curthread); 506 crit_enter(); 507 lwkt_gettoken(&vm_token); 508 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 509 lwkt_reltoken(&vm_token); 510 crit_exit(); 511 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 512 return(m); 513 } 514 515 /* 516 * vm_page_rename() 517 * 518 * Move the given memory entry from its current object to the specified 519 * target object/offset. 520 * 521 * The object must be locked. 522 * This routine may not block. 523 * 524 * Note: This routine will raise itself to splvm(), the caller need not. 525 * 526 * Note: Swap associated with the page must be invalidated by the move. We 527 * have to do this for several reasons: (1) we aren't freeing the 528 * page, (2) we are dirtying the page, (3) the VM system is probably 529 * moving the page from object A to B, and will then later move 530 * the backing store from A to B and we can't have a conflict. 531 * 532 * Note: We *always* dirty the page. It is necessary both for the 533 * fact that we moved it, and because we may be invalidating 534 * swap. If the page is on the cache, we have to deactivate it 535 * or vm_page_dirty() will panic. Dirty pages are not allowed 536 * on the cache. 537 */ 538 void 539 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 540 { 541 crit_enter(); 542 lwkt_gettoken(&vm_token); 543 vm_page_remove(m); 544 vm_page_insert(m, new_object, new_pindex); 545 if (m->queue - m->pc == PQ_CACHE) 546 vm_page_deactivate(m); 547 vm_page_dirty(m); 548 vm_page_wakeup(m); 549 lwkt_reltoken(&vm_token); 550 crit_exit(); 551 } 552 553 /* 554 * vm_page_unqueue() without any wakeup. This routine is used when a page 555 * is being moved between queues or otherwise is to remain BUSYied by the 556 * caller. 557 * 558 * The caller must hold vm_token 559 * This routine may not block. 560 */ 561 void 562 vm_page_unqueue_nowakeup(vm_page_t m) 563 { 564 int queue = m->queue; 565 struct vpgqueues *pq; 566 567 ASSERT_LWKT_TOKEN_HELD(&vm_token); 568 if (queue != PQ_NONE) { 569 pq = &vm_page_queues[queue]; 570 m->queue = PQ_NONE; 571 TAILQ_REMOVE(&pq->pl, m, pageq); 572 (*pq->cnt)--; 573 pq->lcnt--; 574 } 575 } 576 577 /* 578 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 579 * if necessary. 580 * 581 * The caller must hold vm_token 582 * This routine may not block. 583 */ 584 void 585 vm_page_unqueue(vm_page_t m) 586 { 587 int queue = m->queue; 588 struct vpgqueues *pq; 589 590 ASSERT_LWKT_TOKEN_HELD(&vm_token); 591 if (queue != PQ_NONE) { 592 m->queue = PQ_NONE; 593 pq = &vm_page_queues[queue]; 594 TAILQ_REMOVE(&pq->pl, m, pageq); 595 (*pq->cnt)--; 596 pq->lcnt--; 597 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 598 pagedaemon_wakeup(); 599 } 600 } 601 602 /* 603 * vm_page_list_find() 604 * 605 * Find a page on the specified queue with color optimization. 606 * 607 * The page coloring optimization attempts to locate a page that does 608 * not overload other nearby pages in the object in the cpu's L1 or L2 609 * caches. We need this optimization because cpu caches tend to be 610 * physical caches, while object spaces tend to be virtual. 611 * 612 * Must be called with vm_token held. 613 * This routine may not block. 614 * 615 * Note that this routine is carefully inlined. A non-inlined version 616 * is available for outside callers but the only critical path is 617 * from within this source file. 618 */ 619 static __inline 620 vm_page_t 621 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 622 { 623 vm_page_t m; 624 625 if (prefer_zero) 626 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 627 else 628 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 629 if (m == NULL) 630 m = _vm_page_list_find2(basequeue, index); 631 return(m); 632 } 633 634 static vm_page_t 635 _vm_page_list_find2(int basequeue, int index) 636 { 637 int i; 638 vm_page_t m = NULL; 639 struct vpgqueues *pq; 640 641 pq = &vm_page_queues[basequeue]; 642 643 /* 644 * Note that for the first loop, index+i and index-i wind up at the 645 * same place. Even though this is not totally optimal, we've already 646 * blown it by missing the cache case so we do not care. 647 */ 648 649 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 650 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 651 break; 652 653 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 654 break; 655 } 656 return(m); 657 } 658 659 /* 660 * Must be called with vm_token held if the caller desired non-blocking 661 * operation and a stable result. 662 */ 663 vm_page_t 664 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 665 { 666 return(_vm_page_list_find(basequeue, index, prefer_zero)); 667 } 668 669 /* 670 * Find a page on the cache queue with color optimization. As pages 671 * might be found, but not applicable, they are deactivated. This 672 * keeps us from using potentially busy cached pages. 673 * 674 * This routine may not block. 675 * Must be called with vm_token held. 676 */ 677 vm_page_t 678 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 679 { 680 vm_page_t m; 681 682 ASSERT_LWKT_TOKEN_HELD(&vm_token); 683 while (TRUE) { 684 m = _vm_page_list_find( 685 PQ_CACHE, 686 (pindex + object->pg_color) & PQ_L2_MASK, 687 FALSE 688 ); 689 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 690 m->hold_count || m->wire_count)) { 691 vm_page_deactivate(m); 692 continue; 693 } 694 return m; 695 } 696 /* not reached */ 697 } 698 699 /* 700 * Find a free or zero page, with specified preference. We attempt to 701 * inline the nominal case and fall back to _vm_page_select_free() 702 * otherwise. 703 * 704 * This routine must be called with a critical section held. 705 * This routine may not block. 706 */ 707 static __inline vm_page_t 708 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 709 { 710 vm_page_t m; 711 712 m = _vm_page_list_find( 713 PQ_FREE, 714 (pindex + object->pg_color) & PQ_L2_MASK, 715 prefer_zero 716 ); 717 return(m); 718 } 719 720 /* 721 * vm_page_alloc() 722 * 723 * Allocate and return a memory cell associated with this VM object/offset 724 * pair. 725 * 726 * page_req classes: 727 * 728 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 729 * VM_ALLOC_QUICK like normal but cannot use cache 730 * VM_ALLOC_SYSTEM greater free drain 731 * VM_ALLOC_INTERRUPT allow free list to be completely drained 732 * VM_ALLOC_ZERO advisory request for pre-zero'd page 733 * 734 * The object must be locked. 735 * This routine may not block. 736 * The returned page will be marked PG_BUSY 737 * 738 * Additional special handling is required when called from an interrupt 739 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 740 * in this case. 741 */ 742 vm_page_t 743 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 744 { 745 vm_page_t m = NULL; 746 747 crit_enter(); 748 lwkt_gettoken(&vm_token); 749 750 KKASSERT(object != NULL); 751 KASSERT(!vm_page_lookup(object, pindex), 752 ("vm_page_alloc: page already allocated")); 753 KKASSERT(page_req & 754 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 755 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 756 757 /* 758 * Certain system threads (pageout daemon, buf_daemon's) are 759 * allowed to eat deeper into the free page list. 760 */ 761 if (curthread->td_flags & TDF_SYSTHREAD) 762 page_req |= VM_ALLOC_SYSTEM; 763 764 loop: 765 if (vmstats.v_free_count > vmstats.v_free_reserved || 766 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 767 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 768 vmstats.v_free_count > vmstats.v_interrupt_free_min) 769 ) { 770 /* 771 * The free queue has sufficient free pages to take one out. 772 */ 773 if (page_req & VM_ALLOC_ZERO) 774 m = vm_page_select_free(object, pindex, TRUE); 775 else 776 m = vm_page_select_free(object, pindex, FALSE); 777 } else if (page_req & VM_ALLOC_NORMAL) { 778 /* 779 * Allocatable from the cache (non-interrupt only). On 780 * success, we must free the page and try again, thus 781 * ensuring that vmstats.v_*_free_min counters are replenished. 782 */ 783 #ifdef INVARIANTS 784 if (curthread->td_preempted) { 785 kprintf("vm_page_alloc(): warning, attempt to allocate" 786 " cache page from preempting interrupt\n"); 787 m = NULL; 788 } else { 789 m = vm_page_select_cache(object, pindex); 790 } 791 #else 792 m = vm_page_select_cache(object, pindex); 793 #endif 794 /* 795 * On success move the page into the free queue and loop. 796 */ 797 if (m != NULL) { 798 KASSERT(m->dirty == 0, 799 ("Found dirty cache page %p", m)); 800 vm_page_busy(m); 801 vm_page_protect(m, VM_PROT_NONE); 802 vm_page_free(m); 803 goto loop; 804 } 805 806 /* 807 * On failure return NULL 808 */ 809 lwkt_reltoken(&vm_token); 810 crit_exit(); 811 #if defined(DIAGNOSTIC) 812 if (vmstats.v_cache_count > 0) 813 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 814 #endif 815 vm_pageout_deficit++; 816 pagedaemon_wakeup(); 817 return (NULL); 818 } else { 819 /* 820 * No pages available, wakeup the pageout daemon and give up. 821 */ 822 lwkt_reltoken(&vm_token); 823 crit_exit(); 824 vm_pageout_deficit++; 825 pagedaemon_wakeup(); 826 return (NULL); 827 } 828 829 /* 830 * Good page found. The page has not yet been busied. We are in 831 * a critical section. 832 */ 833 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 834 KASSERT(m->dirty == 0, 835 ("vm_page_alloc: free/cache page %p was dirty", m)); 836 837 /* 838 * Remove from free queue 839 */ 840 vm_page_unqueue_nowakeup(m); 841 842 /* 843 * Initialize structure. Only the PG_ZERO flag is inherited. Set 844 * the page PG_BUSY 845 */ 846 if (m->flags & PG_ZERO) { 847 vm_page_zero_count--; 848 m->flags = PG_ZERO | PG_BUSY; 849 } else { 850 m->flags = PG_BUSY; 851 } 852 m->wire_count = 0; 853 m->hold_count = 0; 854 m->act_count = 0; 855 m->busy = 0; 856 m->valid = 0; 857 858 /* 859 * vm_page_insert() is safe prior to the crit_exit(). Note also that 860 * inserting a page here does not insert it into the pmap (which 861 * could cause us to block allocating memory). We cannot block 862 * anywhere. 863 */ 864 vm_page_insert(m, object, pindex); 865 866 /* 867 * Don't wakeup too often - wakeup the pageout daemon when 868 * we would be nearly out of memory. 869 */ 870 pagedaemon_wakeup(); 871 872 lwkt_reltoken(&vm_token); 873 crit_exit(); 874 875 /* 876 * A PG_BUSY page is returned. 877 */ 878 return (m); 879 } 880 881 /* 882 * Wait for sufficient free memory for nominal heavy memory use kernel 883 * operations. 884 */ 885 void 886 vm_wait_nominal(void) 887 { 888 while (vm_page_count_min(0)) 889 vm_wait(0); 890 } 891 892 /* 893 * Test if vm_wait_nominal() would block. 894 */ 895 int 896 vm_test_nominal(void) 897 { 898 if (vm_page_count_min(0)) 899 return(1); 900 return(0); 901 } 902 903 /* 904 * Block until free pages are available for allocation, called in various 905 * places before memory allocations. 906 */ 907 void 908 vm_wait(int timo) 909 { 910 crit_enter(); 911 lwkt_gettoken(&vm_token); 912 if (curthread == pagethread) { 913 vm_pageout_pages_needed = 1; 914 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 915 } else { 916 if (vm_pages_needed == 0) { 917 vm_pages_needed = 1; 918 wakeup(&vm_pages_needed); 919 } 920 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 921 } 922 lwkt_reltoken(&vm_token); 923 crit_exit(); 924 } 925 926 /* 927 * Block until free pages are available for allocation 928 * 929 * Called only in vm_fault so that processes page faulting can be 930 * easily tracked. 931 */ 932 void 933 vm_waitpfault(void) 934 { 935 crit_enter(); 936 lwkt_gettoken(&vm_token); 937 if (vm_pages_needed == 0) { 938 vm_pages_needed = 1; 939 wakeup(&vm_pages_needed); 940 } 941 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 942 lwkt_reltoken(&vm_token); 943 crit_exit(); 944 } 945 946 /* 947 * Put the specified page on the active list (if appropriate). Ensure 948 * that act_count is at least ACT_INIT but do not otherwise mess with it. 949 * 950 * The page queues must be locked. 951 * This routine may not block. 952 */ 953 void 954 vm_page_activate(vm_page_t m) 955 { 956 crit_enter(); 957 lwkt_gettoken(&vm_token); 958 if (m->queue != PQ_ACTIVE) { 959 if ((m->queue - m->pc) == PQ_CACHE) 960 mycpu->gd_cnt.v_reactivated++; 961 962 vm_page_unqueue(m); 963 964 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 965 m->queue = PQ_ACTIVE; 966 vm_page_queues[PQ_ACTIVE].lcnt++; 967 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 968 m, pageq); 969 if (m->act_count < ACT_INIT) 970 m->act_count = ACT_INIT; 971 vmstats.v_active_count++; 972 } 973 } else { 974 if (m->act_count < ACT_INIT) 975 m->act_count = ACT_INIT; 976 } 977 lwkt_reltoken(&vm_token); 978 crit_exit(); 979 } 980 981 /* 982 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 983 * routine is called when a page has been added to the cache or free 984 * queues. 985 * 986 * This routine may not block. 987 * This routine must be called at splvm() 988 */ 989 static __inline void 990 vm_page_free_wakeup(void) 991 { 992 /* 993 * if pageout daemon needs pages, then tell it that there are 994 * some free. 995 */ 996 if (vm_pageout_pages_needed && 997 vmstats.v_cache_count + vmstats.v_free_count >= 998 vmstats.v_pageout_free_min 999 ) { 1000 wakeup(&vm_pageout_pages_needed); 1001 vm_pageout_pages_needed = 0; 1002 } 1003 1004 /* 1005 * wakeup processes that are waiting on memory if we hit a 1006 * high water mark. And wakeup scheduler process if we have 1007 * lots of memory. this process will swapin processes. 1008 */ 1009 if (vm_pages_needed && !vm_page_count_min(0)) { 1010 vm_pages_needed = 0; 1011 wakeup(&vmstats.v_free_count); 1012 } 1013 } 1014 1015 /* 1016 * vm_page_free_toq: 1017 * 1018 * Returns the given page to the PQ_FREE list, disassociating it with 1019 * any VM object. 1020 * 1021 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1022 * return (the page will have been freed). No particular spl is required 1023 * on entry. 1024 * 1025 * This routine may not block. 1026 */ 1027 void 1028 vm_page_free_toq(vm_page_t m) 1029 { 1030 struct vpgqueues *pq; 1031 1032 crit_enter(); 1033 lwkt_gettoken(&vm_token); 1034 mycpu->gd_cnt.v_tfree++; 1035 1036 KKASSERT((m->flags & PG_MAPPED) == 0); 1037 1038 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1039 kprintf( 1040 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1041 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1042 m->hold_count); 1043 if ((m->queue - m->pc) == PQ_FREE) 1044 panic("vm_page_free: freeing free page"); 1045 else 1046 panic("vm_page_free: freeing busy page"); 1047 } 1048 1049 /* 1050 * unqueue, then remove page. Note that we cannot destroy 1051 * the page here because we do not want to call the pager's 1052 * callback routine until after we've put the page on the 1053 * appropriate free queue. 1054 */ 1055 vm_page_unqueue_nowakeup(m); 1056 vm_page_remove(m); 1057 1058 /* 1059 * No further management of fictitious pages occurs beyond object 1060 * and queue removal. 1061 */ 1062 if ((m->flags & PG_FICTITIOUS) != 0) { 1063 vm_page_wakeup(m); 1064 lwkt_reltoken(&vm_token); 1065 crit_exit(); 1066 return; 1067 } 1068 1069 m->valid = 0; 1070 vm_page_undirty(m); 1071 1072 if (m->wire_count != 0) { 1073 if (m->wire_count > 1) { 1074 panic( 1075 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1076 m->wire_count, (long)m->pindex); 1077 } 1078 panic("vm_page_free: freeing wired page"); 1079 } 1080 1081 /* 1082 * Clear the UNMANAGED flag when freeing an unmanaged page. 1083 */ 1084 if (m->flags & PG_UNMANAGED) { 1085 m->flags &= ~PG_UNMANAGED; 1086 } 1087 1088 if (m->hold_count != 0) { 1089 m->flags &= ~PG_ZERO; 1090 m->queue = PQ_HOLD; 1091 } else { 1092 m->queue = PQ_FREE + m->pc; 1093 } 1094 pq = &vm_page_queues[m->queue]; 1095 pq->lcnt++; 1096 ++(*pq->cnt); 1097 1098 /* 1099 * Put zero'd pages on the end ( where we look for zero'd pages 1100 * first ) and non-zerod pages at the head. 1101 */ 1102 if (m->flags & PG_ZERO) { 1103 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1104 ++vm_page_zero_count; 1105 } else { 1106 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1107 } 1108 vm_page_wakeup(m); 1109 vm_page_free_wakeup(); 1110 lwkt_reltoken(&vm_token); 1111 crit_exit(); 1112 } 1113 1114 /* 1115 * vm_page_free_fromq_fast() 1116 * 1117 * Remove a non-zero page from one of the free queues; the page is removed for 1118 * zeroing, so do not issue a wakeup. 1119 * 1120 * MPUNSAFE 1121 */ 1122 vm_page_t 1123 vm_page_free_fromq_fast(void) 1124 { 1125 static int qi; 1126 vm_page_t m; 1127 int i; 1128 1129 crit_enter(); 1130 lwkt_gettoken(&vm_token); 1131 for (i = 0; i < PQ_L2_SIZE; ++i) { 1132 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1133 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1134 if (m && (m->flags & PG_ZERO) == 0) { 1135 vm_page_unqueue_nowakeup(m); 1136 vm_page_busy(m); 1137 break; 1138 } 1139 m = NULL; 1140 } 1141 lwkt_reltoken(&vm_token); 1142 crit_exit(); 1143 return (m); 1144 } 1145 1146 /* 1147 * vm_page_unmanage() 1148 * 1149 * Prevent PV management from being done on the page. The page is 1150 * removed from the paging queues as if it were wired, and as a 1151 * consequence of no longer being managed the pageout daemon will not 1152 * touch it (since there is no way to locate the pte mappings for the 1153 * page). madvise() calls that mess with the pmap will also no longer 1154 * operate on the page. 1155 * 1156 * Beyond that the page is still reasonably 'normal'. Freeing the page 1157 * will clear the flag. 1158 * 1159 * This routine is used by OBJT_PHYS objects - objects using unswappable 1160 * physical memory as backing store rather then swap-backed memory and 1161 * will eventually be extended to support 4MB unmanaged physical 1162 * mappings. 1163 * 1164 * Must be called with a critical section held. 1165 * Must be called with vm_token held. 1166 */ 1167 void 1168 vm_page_unmanage(vm_page_t m) 1169 { 1170 ASSERT_IN_CRIT_SECTION(); 1171 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1172 if ((m->flags & PG_UNMANAGED) == 0) { 1173 if (m->wire_count == 0) 1174 vm_page_unqueue(m); 1175 } 1176 vm_page_flag_set(m, PG_UNMANAGED); 1177 } 1178 1179 /* 1180 * Mark this page as wired down by yet another map, removing it from 1181 * paging queues as necessary. 1182 * 1183 * The page queues must be locked. 1184 * This routine may not block. 1185 */ 1186 void 1187 vm_page_wire(vm_page_t m) 1188 { 1189 /* 1190 * Only bump the wire statistics if the page is not already wired, 1191 * and only unqueue the page if it is on some queue (if it is unmanaged 1192 * it is already off the queues). Don't do anything with fictitious 1193 * pages because they are always wired. 1194 */ 1195 crit_enter(); 1196 lwkt_gettoken(&vm_token); 1197 if ((m->flags & PG_FICTITIOUS) == 0) { 1198 if (m->wire_count == 0) { 1199 if ((m->flags & PG_UNMANAGED) == 0) 1200 vm_page_unqueue(m); 1201 vmstats.v_wire_count++; 1202 } 1203 m->wire_count++; 1204 KASSERT(m->wire_count != 0, 1205 ("vm_page_wire: wire_count overflow m=%p", m)); 1206 } 1207 lwkt_reltoken(&vm_token); 1208 crit_exit(); 1209 } 1210 1211 /* 1212 * Release one wiring of this page, potentially enabling it to be paged again. 1213 * 1214 * Many pages placed on the inactive queue should actually go 1215 * into the cache, but it is difficult to figure out which. What 1216 * we do instead, if the inactive target is well met, is to put 1217 * clean pages at the head of the inactive queue instead of the tail. 1218 * This will cause them to be moved to the cache more quickly and 1219 * if not actively re-referenced, freed more quickly. If we just 1220 * stick these pages at the end of the inactive queue, heavy filesystem 1221 * meta-data accesses can cause an unnecessary paging load on memory bound 1222 * processes. This optimization causes one-time-use metadata to be 1223 * reused more quickly. 1224 * 1225 * BUT, if we are in a low-memory situation we have no choice but to 1226 * put clean pages on the cache queue. 1227 * 1228 * A number of routines use vm_page_unwire() to guarantee that the page 1229 * will go into either the inactive or active queues, and will NEVER 1230 * be placed in the cache - for example, just after dirtying a page. 1231 * dirty pages in the cache are not allowed. 1232 * 1233 * The page queues must be locked. 1234 * This routine may not block. 1235 */ 1236 void 1237 vm_page_unwire(vm_page_t m, int activate) 1238 { 1239 crit_enter(); 1240 lwkt_gettoken(&vm_token); 1241 if (m->flags & PG_FICTITIOUS) { 1242 /* do nothing */ 1243 } else if (m->wire_count <= 0) { 1244 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1245 } else { 1246 if (--m->wire_count == 0) { 1247 --vmstats.v_wire_count; 1248 if (m->flags & PG_UNMANAGED) { 1249 ; 1250 } else if (activate) { 1251 TAILQ_INSERT_TAIL( 1252 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1253 m->queue = PQ_ACTIVE; 1254 vm_page_queues[PQ_ACTIVE].lcnt++; 1255 vmstats.v_active_count++; 1256 } else { 1257 vm_page_flag_clear(m, PG_WINATCFLS); 1258 TAILQ_INSERT_TAIL( 1259 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1260 m->queue = PQ_INACTIVE; 1261 vm_page_queues[PQ_INACTIVE].lcnt++; 1262 vmstats.v_inactive_count++; 1263 ++vm_swapcache_inactive_heuristic; 1264 } 1265 } 1266 } 1267 lwkt_reltoken(&vm_token); 1268 crit_exit(); 1269 } 1270 1271 1272 /* 1273 * Move the specified page to the inactive queue. If the page has 1274 * any associated swap, the swap is deallocated. 1275 * 1276 * Normally athead is 0 resulting in LRU operation. athead is set 1277 * to 1 if we want this page to be 'as if it were placed in the cache', 1278 * except without unmapping it from the process address space. 1279 * 1280 * This routine may not block. 1281 * The caller must hold vm_token. 1282 */ 1283 static __inline void 1284 _vm_page_deactivate(vm_page_t m, int athead) 1285 { 1286 /* 1287 * Ignore if already inactive. 1288 */ 1289 if (m->queue == PQ_INACTIVE) 1290 return; 1291 1292 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1293 if ((m->queue - m->pc) == PQ_CACHE) 1294 mycpu->gd_cnt.v_reactivated++; 1295 vm_page_flag_clear(m, PG_WINATCFLS); 1296 vm_page_unqueue(m); 1297 if (athead) { 1298 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1299 m, pageq); 1300 } else { 1301 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1302 m, pageq); 1303 ++vm_swapcache_inactive_heuristic; 1304 } 1305 m->queue = PQ_INACTIVE; 1306 vm_page_queues[PQ_INACTIVE].lcnt++; 1307 vmstats.v_inactive_count++; 1308 } 1309 } 1310 1311 /* 1312 * Attempt to deactivate a page. 1313 * 1314 * No requirements. 1315 */ 1316 void 1317 vm_page_deactivate(vm_page_t m) 1318 { 1319 crit_enter(); 1320 lwkt_gettoken(&vm_token); 1321 _vm_page_deactivate(m, 0); 1322 lwkt_reltoken(&vm_token); 1323 crit_exit(); 1324 } 1325 1326 /* 1327 * Attempt to move a page to PQ_CACHE. 1328 * Returns 0 on failure, 1 on success 1329 * 1330 * No requirements. 1331 */ 1332 int 1333 vm_page_try_to_cache(vm_page_t m) 1334 { 1335 crit_enter(); 1336 lwkt_gettoken(&vm_token); 1337 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1338 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1339 lwkt_reltoken(&vm_token); 1340 crit_exit(); 1341 return(0); 1342 } 1343 vm_page_test_dirty(m); 1344 if (m->dirty) { 1345 lwkt_reltoken(&vm_token); 1346 crit_exit(); 1347 return(0); 1348 } 1349 vm_page_cache(m); 1350 lwkt_reltoken(&vm_token); 1351 crit_exit(); 1352 return(1); 1353 } 1354 1355 /* 1356 * Attempt to free the page. If we cannot free it, we do nothing. 1357 * 1 is returned on success, 0 on failure. 1358 * 1359 * No requirements. 1360 */ 1361 int 1362 vm_page_try_to_free(vm_page_t m) 1363 { 1364 crit_enter(); 1365 lwkt_gettoken(&vm_token); 1366 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1367 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1368 lwkt_reltoken(&vm_token); 1369 crit_exit(); 1370 return(0); 1371 } 1372 vm_page_test_dirty(m); 1373 if (m->dirty) { 1374 lwkt_reltoken(&vm_token); 1375 crit_exit(); 1376 return(0); 1377 } 1378 vm_page_busy(m); 1379 vm_page_protect(m, VM_PROT_NONE); 1380 vm_page_free(m); 1381 lwkt_reltoken(&vm_token); 1382 crit_exit(); 1383 return(1); 1384 } 1385 1386 /* 1387 * vm_page_cache 1388 * 1389 * Put the specified page onto the page cache queue (if appropriate). 1390 * 1391 * The caller must hold vm_token. 1392 * This routine may not block. 1393 */ 1394 void 1395 vm_page_cache(vm_page_t m) 1396 { 1397 ASSERT_IN_CRIT_SECTION(); 1398 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1399 1400 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1401 m->wire_count || m->hold_count) { 1402 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1403 return; 1404 } 1405 1406 /* 1407 * Already in the cache (and thus not mapped) 1408 */ 1409 if ((m->queue - m->pc) == PQ_CACHE) { 1410 KKASSERT((m->flags & PG_MAPPED) == 0); 1411 return; 1412 } 1413 1414 /* 1415 * Caller is required to test m->dirty, but note that the act of 1416 * removing the page from its maps can cause it to become dirty 1417 * on an SMP system due to another cpu running in usermode. 1418 */ 1419 if (m->dirty) { 1420 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1421 (long)m->pindex); 1422 } 1423 1424 /* 1425 * Remove all pmaps and indicate that the page is not 1426 * writeable or mapped. Our vm_page_protect() call may 1427 * have blocked (especially w/ VM_PROT_NONE), so recheck 1428 * everything. 1429 */ 1430 vm_page_busy(m); 1431 vm_page_protect(m, VM_PROT_NONE); 1432 vm_page_wakeup(m); 1433 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy || 1434 m->wire_count || m->hold_count) { 1435 /* do nothing */ 1436 } else if (m->dirty) { 1437 vm_page_deactivate(m); 1438 } else { 1439 vm_page_unqueue_nowakeup(m); 1440 m->queue = PQ_CACHE + m->pc; 1441 vm_page_queues[m->queue].lcnt++; 1442 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1443 vmstats.v_cache_count++; 1444 vm_page_free_wakeup(); 1445 } 1446 } 1447 1448 /* 1449 * vm_page_dontneed() 1450 * 1451 * Cache, deactivate, or do nothing as appropriate. This routine 1452 * is typically used by madvise() MADV_DONTNEED. 1453 * 1454 * Generally speaking we want to move the page into the cache so 1455 * it gets reused quickly. However, this can result in a silly syndrome 1456 * due to the page recycling too quickly. Small objects will not be 1457 * fully cached. On the otherhand, if we move the page to the inactive 1458 * queue we wind up with a problem whereby very large objects 1459 * unnecessarily blow away our inactive and cache queues. 1460 * 1461 * The solution is to move the pages based on a fixed weighting. We 1462 * either leave them alone, deactivate them, or move them to the cache, 1463 * where moving them to the cache has the highest weighting. 1464 * By forcing some pages into other queues we eventually force the 1465 * system to balance the queues, potentially recovering other unrelated 1466 * space from active. The idea is to not force this to happen too 1467 * often. 1468 * 1469 * No requirements. 1470 */ 1471 void 1472 vm_page_dontneed(vm_page_t m) 1473 { 1474 static int dnweight; 1475 int dnw; 1476 int head; 1477 1478 dnw = ++dnweight; 1479 1480 /* 1481 * occassionally leave the page alone 1482 */ 1483 crit_enter(); 1484 lwkt_gettoken(&vm_token); 1485 if ((dnw & 0x01F0) == 0 || 1486 m->queue == PQ_INACTIVE || 1487 m->queue - m->pc == PQ_CACHE 1488 ) { 1489 if (m->act_count >= ACT_INIT) 1490 --m->act_count; 1491 lwkt_reltoken(&vm_token); 1492 crit_exit(); 1493 return; 1494 } 1495 1496 if (m->dirty == 0) 1497 vm_page_test_dirty(m); 1498 1499 if (m->dirty || (dnw & 0x0070) == 0) { 1500 /* 1501 * Deactivate the page 3 times out of 32. 1502 */ 1503 head = 0; 1504 } else { 1505 /* 1506 * Cache the page 28 times out of every 32. Note that 1507 * the page is deactivated instead of cached, but placed 1508 * at the head of the queue instead of the tail. 1509 */ 1510 head = 1; 1511 } 1512 _vm_page_deactivate(m, head); 1513 lwkt_reltoken(&vm_token); 1514 crit_exit(); 1515 } 1516 1517 /* 1518 * Grab a page, blocking if it is busy and allocating a page if necessary. 1519 * A busy page is returned or NULL. 1520 * 1521 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1522 * If VM_ALLOC_RETRY is not specified 1523 * 1524 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1525 * always returned if we had blocked. 1526 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1527 * This routine may not be called from an interrupt. 1528 * The returned page may not be entirely valid. 1529 * 1530 * This routine may be called from mainline code without spl protection and 1531 * be guarenteed a busied page associated with the object at the specified 1532 * index. 1533 * 1534 * No requirements. 1535 */ 1536 vm_page_t 1537 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1538 { 1539 vm_page_t m; 1540 int generation; 1541 1542 KKASSERT(allocflags & 1543 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1544 crit_enter(); 1545 lwkt_gettoken(&vm_token); 1546 retrylookup: 1547 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1548 if (m->busy || (m->flags & PG_BUSY)) { 1549 generation = object->generation; 1550 1551 while ((object->generation == generation) && 1552 (m->busy || (m->flags & PG_BUSY))) { 1553 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1554 tsleep(m, 0, "pgrbwt", 0); 1555 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1556 m = NULL; 1557 goto done; 1558 } 1559 } 1560 goto retrylookup; 1561 } else { 1562 vm_page_busy(m); 1563 goto done; 1564 } 1565 } 1566 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1567 if (m == NULL) { 1568 vm_wait(0); 1569 if ((allocflags & VM_ALLOC_RETRY) == 0) 1570 goto done; 1571 goto retrylookup; 1572 } 1573 done: 1574 lwkt_reltoken(&vm_token); 1575 crit_exit(); 1576 return(m); 1577 } 1578 1579 /* 1580 * Mapping function for valid bits or for dirty bits in 1581 * a page. May not block. 1582 * 1583 * Inputs are required to range within a page. 1584 * 1585 * No requirements. 1586 * Non blocking. 1587 */ 1588 int 1589 vm_page_bits(int base, int size) 1590 { 1591 int first_bit; 1592 int last_bit; 1593 1594 KASSERT( 1595 base + size <= PAGE_SIZE, 1596 ("vm_page_bits: illegal base/size %d/%d", base, size) 1597 ); 1598 1599 if (size == 0) /* handle degenerate case */ 1600 return(0); 1601 1602 first_bit = base >> DEV_BSHIFT; 1603 last_bit = (base + size - 1) >> DEV_BSHIFT; 1604 1605 return ((2 << last_bit) - (1 << first_bit)); 1606 } 1607 1608 /* 1609 * Sets portions of a page valid and clean. The arguments are expected 1610 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1611 * of any partial chunks touched by the range. The invalid portion of 1612 * such chunks will be zero'd. 1613 * 1614 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1615 * align base to DEV_BSIZE so as not to mark clean a partially 1616 * truncated device block. Otherwise the dirty page status might be 1617 * lost. 1618 * 1619 * This routine may not block. 1620 * 1621 * (base + size) must be less then or equal to PAGE_SIZE. 1622 */ 1623 static void 1624 _vm_page_zero_valid(vm_page_t m, int base, int size) 1625 { 1626 int frag; 1627 int endoff; 1628 1629 if (size == 0) /* handle degenerate case */ 1630 return; 1631 1632 /* 1633 * If the base is not DEV_BSIZE aligned and the valid 1634 * bit is clear, we have to zero out a portion of the 1635 * first block. 1636 */ 1637 1638 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1639 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1640 ) { 1641 pmap_zero_page_area( 1642 VM_PAGE_TO_PHYS(m), 1643 frag, 1644 base - frag 1645 ); 1646 } 1647 1648 /* 1649 * If the ending offset is not DEV_BSIZE aligned and the 1650 * valid bit is clear, we have to zero out a portion of 1651 * the last block. 1652 */ 1653 1654 endoff = base + size; 1655 1656 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1657 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1658 ) { 1659 pmap_zero_page_area( 1660 VM_PAGE_TO_PHYS(m), 1661 endoff, 1662 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1663 ); 1664 } 1665 } 1666 1667 /* 1668 * Set valid, clear dirty bits. If validating the entire 1669 * page we can safely clear the pmap modify bit. We also 1670 * use this opportunity to clear the PG_NOSYNC flag. If a process 1671 * takes a write fault on a MAP_NOSYNC memory area the flag will 1672 * be set again. 1673 * 1674 * We set valid bits inclusive of any overlap, but we can only 1675 * clear dirty bits for DEV_BSIZE chunks that are fully within 1676 * the range. 1677 * 1678 * Page must be busied? 1679 * No other requirements. 1680 */ 1681 void 1682 vm_page_set_valid(vm_page_t m, int base, int size) 1683 { 1684 _vm_page_zero_valid(m, base, size); 1685 m->valid |= vm_page_bits(base, size); 1686 } 1687 1688 1689 /* 1690 * Set valid bits and clear dirty bits. 1691 * 1692 * NOTE: This function does not clear the pmap modified bit. 1693 * Also note that e.g. NFS may use a byte-granular base 1694 * and size. 1695 * 1696 * Page must be busied? 1697 * No other requirements. 1698 */ 1699 void 1700 vm_page_set_validclean(vm_page_t m, int base, int size) 1701 { 1702 int pagebits; 1703 1704 _vm_page_zero_valid(m, base, size); 1705 pagebits = vm_page_bits(base, size); 1706 m->valid |= pagebits; 1707 m->dirty &= ~pagebits; 1708 if (base == 0 && size == PAGE_SIZE) { 1709 /*pmap_clear_modify(m);*/ 1710 vm_page_flag_clear(m, PG_NOSYNC); 1711 } 1712 } 1713 1714 /* 1715 * Set valid & dirty. Used by buwrite() 1716 * 1717 * Page must be busied? 1718 * No other requirements. 1719 */ 1720 void 1721 vm_page_set_validdirty(vm_page_t m, int base, int size) 1722 { 1723 int pagebits; 1724 1725 pagebits = vm_page_bits(base, size); 1726 m->valid |= pagebits; 1727 m->dirty |= pagebits; 1728 if (m->object) 1729 vm_object_set_writeable_dirty(m->object); 1730 } 1731 1732 /* 1733 * Clear dirty bits. 1734 * 1735 * NOTE: This function does not clear the pmap modified bit. 1736 * Also note that e.g. NFS may use a byte-granular base 1737 * and size. 1738 * 1739 * Page must be busied? 1740 * No other requirements. 1741 */ 1742 void 1743 vm_page_clear_dirty(vm_page_t m, int base, int size) 1744 { 1745 m->dirty &= ~vm_page_bits(base, size); 1746 if (base == 0 && size == PAGE_SIZE) { 1747 /*pmap_clear_modify(m);*/ 1748 vm_page_flag_clear(m, PG_NOSYNC); 1749 } 1750 } 1751 1752 /* 1753 * Make the page all-dirty. 1754 * 1755 * Also make sure the related object and vnode reflect the fact that the 1756 * object may now contain a dirty page. 1757 * 1758 * Page must be busied? 1759 * No other requirements. 1760 */ 1761 void 1762 vm_page_dirty(vm_page_t m) 1763 { 1764 #ifdef INVARIANTS 1765 int pqtype = m->queue - m->pc; 1766 #endif 1767 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1768 ("vm_page_dirty: page in free/cache queue!")); 1769 if (m->dirty != VM_PAGE_BITS_ALL) { 1770 m->dirty = VM_PAGE_BITS_ALL; 1771 if (m->object) 1772 vm_object_set_writeable_dirty(m->object); 1773 } 1774 } 1775 1776 /* 1777 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1778 * valid and dirty bits for the effected areas are cleared. 1779 * 1780 * Page must be busied? 1781 * Does not block. 1782 * No other requirements. 1783 */ 1784 void 1785 vm_page_set_invalid(vm_page_t m, int base, int size) 1786 { 1787 int bits; 1788 1789 bits = vm_page_bits(base, size); 1790 m->valid &= ~bits; 1791 m->dirty &= ~bits; 1792 m->object->generation++; 1793 } 1794 1795 /* 1796 * The kernel assumes that the invalid portions of a page contain 1797 * garbage, but such pages can be mapped into memory by user code. 1798 * When this occurs, we must zero out the non-valid portions of the 1799 * page so user code sees what it expects. 1800 * 1801 * Pages are most often semi-valid when the end of a file is mapped 1802 * into memory and the file's size is not page aligned. 1803 * 1804 * Page must be busied? 1805 * No other requirements. 1806 */ 1807 void 1808 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1809 { 1810 int b; 1811 int i; 1812 1813 /* 1814 * Scan the valid bits looking for invalid sections that 1815 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1816 * valid bit may be set ) have already been zerod by 1817 * vm_page_set_validclean(). 1818 */ 1819 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1820 if (i == (PAGE_SIZE / DEV_BSIZE) || 1821 (m->valid & (1 << i)) 1822 ) { 1823 if (i > b) { 1824 pmap_zero_page_area( 1825 VM_PAGE_TO_PHYS(m), 1826 b << DEV_BSHIFT, 1827 (i - b) << DEV_BSHIFT 1828 ); 1829 } 1830 b = i + 1; 1831 } 1832 } 1833 1834 /* 1835 * setvalid is TRUE when we can safely set the zero'd areas 1836 * as being valid. We can do this if there are no cache consistency 1837 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1838 */ 1839 if (setvalid) 1840 m->valid = VM_PAGE_BITS_ALL; 1841 } 1842 1843 /* 1844 * Is a (partial) page valid? Note that the case where size == 0 1845 * will return FALSE in the degenerate case where the page is entirely 1846 * invalid, and TRUE otherwise. 1847 * 1848 * Does not block. 1849 * No other requirements. 1850 */ 1851 int 1852 vm_page_is_valid(vm_page_t m, int base, int size) 1853 { 1854 int bits = vm_page_bits(base, size); 1855 1856 if (m->valid && ((m->valid & bits) == bits)) 1857 return 1; 1858 else 1859 return 0; 1860 } 1861 1862 /* 1863 * update dirty bits from pmap/mmu. May not block. 1864 * 1865 * Caller must hold vm_token if non-blocking operation desired. 1866 * No other requirements. 1867 */ 1868 void 1869 vm_page_test_dirty(vm_page_t m) 1870 { 1871 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1872 vm_page_dirty(m); 1873 } 1874 } 1875 1876 /* 1877 * Issue an event on a VM page. Corresponding action structures are 1878 * removed from the page's list and called. 1879 */ 1880 void 1881 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1882 { 1883 struct vm_page_action *scan, *next; 1884 1885 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) { 1886 if (scan->event == event) { 1887 scan->event = VMEVENT_NONE; 1888 LIST_REMOVE(scan, entry); 1889 scan->func(m, scan); 1890 } 1891 } 1892 } 1893 1894 1895 #include "opt_ddb.h" 1896 #ifdef DDB 1897 #include <sys/kernel.h> 1898 1899 #include <ddb/ddb.h> 1900 1901 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1902 { 1903 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1904 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1905 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1906 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1907 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1908 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1909 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1910 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1911 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1912 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1913 } 1914 1915 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1916 { 1917 int i; 1918 db_printf("PQ_FREE:"); 1919 for(i=0;i<PQ_L2_SIZE;i++) { 1920 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1921 } 1922 db_printf("\n"); 1923 1924 db_printf("PQ_CACHE:"); 1925 for(i=0;i<PQ_L2_SIZE;i++) { 1926 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1927 } 1928 db_printf("\n"); 1929 1930 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1931 vm_page_queues[PQ_ACTIVE].lcnt, 1932 vm_page_queues[PQ_INACTIVE].lcnt); 1933 } 1934 #endif /* DDB */ 1935