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 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 426 object->generation++; 427 428 /* 429 * show that the object has one more resident page. 430 */ 431 object->resident_page_count++; 432 433 /* 434 * Since we are inserting a new and possibly dirty page, 435 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 436 */ 437 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 438 vm_object_set_writeable_dirty(object); 439 440 /* 441 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 442 */ 443 swap_pager_page_inserted(m); 444 } 445 446 /* 447 * Removes the given vm_page_t from the global (object,index) hash table 448 * and from the object's memq. 449 * 450 * The underlying pmap entry (if any) is NOT removed here. 451 * This routine may not block. 452 * 453 * The page must be BUSY and will remain BUSY on return. 454 * No other requirements. 455 * 456 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 457 * it busy. 458 */ 459 void 460 vm_page_remove(vm_page_t m) 461 { 462 vm_object_t object; 463 464 crit_enter(); 465 lwkt_gettoken(&vm_token); 466 if (m->object == NULL) { 467 lwkt_reltoken(&vm_token); 468 crit_exit(); 469 return; 470 } 471 472 if ((m->flags & PG_BUSY) == 0) 473 panic("vm_page_remove: page not busy"); 474 475 object = m->object; 476 477 /* 478 * Remove the page from the object and update the object. 479 */ 480 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 481 object->resident_page_count--; 482 object->generation++; 483 m->object = NULL; 484 485 lwkt_reltoken(&vm_token); 486 crit_exit(); 487 } 488 489 /* 490 * Locate and return the page at (object, pindex), or NULL if the 491 * page could not be found. 492 * 493 * The caller must hold vm_token. 494 */ 495 vm_page_t 496 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 497 { 498 vm_page_t m; 499 500 /* 501 * Search the hash table for this object/offset pair 502 */ 503 ASSERT_LWKT_TOKEN_HELD(&vm_token); 504 crit_enter(); 505 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 506 crit_exit(); 507 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 508 return(m); 509 } 510 511 /* 512 * vm_page_rename() 513 * 514 * Move the given memory entry from its current object to the specified 515 * target object/offset. 516 * 517 * The object must be locked. 518 * This routine may not block. 519 * 520 * Note: This routine will raise itself to splvm(), the caller need not. 521 * 522 * Note: Swap associated with the page must be invalidated by the move. We 523 * have to do this for several reasons: (1) we aren't freeing the 524 * page, (2) we are dirtying the page, (3) the VM system is probably 525 * moving the page from object A to B, and will then later move 526 * the backing store from A to B and we can't have a conflict. 527 * 528 * Note: We *always* dirty the page. It is necessary both for the 529 * fact that we moved it, and because we may be invalidating 530 * swap. If the page is on the cache, we have to deactivate it 531 * or vm_page_dirty() will panic. Dirty pages are not allowed 532 * on the cache. 533 */ 534 void 535 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 536 { 537 crit_enter(); 538 lwkt_gettoken(&vm_token); 539 vm_page_remove(m); 540 vm_page_insert(m, new_object, new_pindex); 541 if (m->queue - m->pc == PQ_CACHE) 542 vm_page_deactivate(m); 543 vm_page_dirty(m); 544 vm_page_wakeup(m); 545 lwkt_reltoken(&vm_token); 546 crit_exit(); 547 } 548 549 /* 550 * vm_page_unqueue() without any wakeup. This routine is used when a page 551 * is being moved between queues or otherwise is to remain BUSYied by the 552 * caller. 553 * 554 * The caller must hold vm_token 555 * This routine may not block. 556 */ 557 void 558 vm_page_unqueue_nowakeup(vm_page_t m) 559 { 560 int queue = m->queue; 561 struct vpgqueues *pq; 562 563 ASSERT_LWKT_TOKEN_HELD(&vm_token); 564 if (queue != PQ_NONE) { 565 pq = &vm_page_queues[queue]; 566 m->queue = PQ_NONE; 567 TAILQ_REMOVE(&pq->pl, m, pageq); 568 (*pq->cnt)--; 569 pq->lcnt--; 570 } 571 } 572 573 /* 574 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 575 * if necessary. 576 * 577 * The caller must hold vm_token 578 * This routine may not block. 579 */ 580 void 581 vm_page_unqueue(vm_page_t m) 582 { 583 int queue = m->queue; 584 struct vpgqueues *pq; 585 586 ASSERT_LWKT_TOKEN_HELD(&vm_token); 587 if (queue != PQ_NONE) { 588 m->queue = PQ_NONE; 589 pq = &vm_page_queues[queue]; 590 TAILQ_REMOVE(&pq->pl, m, pageq); 591 (*pq->cnt)--; 592 pq->lcnt--; 593 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 594 pagedaemon_wakeup(); 595 } 596 } 597 598 /* 599 * vm_page_list_find() 600 * 601 * Find a page on the specified queue with color optimization. 602 * 603 * The page coloring optimization attempts to locate a page that does 604 * not overload other nearby pages in the object in the cpu's L1 or L2 605 * caches. We need this optimization because cpu caches tend to be 606 * physical caches, while object spaces tend to be virtual. 607 * 608 * Must be called with vm_token held. 609 * This routine may not block. 610 * 611 * Note that this routine is carefully inlined. A non-inlined version 612 * is available for outside callers but the only critical path is 613 * from within this source file. 614 */ 615 static __inline 616 vm_page_t 617 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 618 { 619 vm_page_t m; 620 621 if (prefer_zero) 622 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 623 else 624 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 625 if (m == NULL) 626 m = _vm_page_list_find2(basequeue, index); 627 return(m); 628 } 629 630 static vm_page_t 631 _vm_page_list_find2(int basequeue, int index) 632 { 633 int i; 634 vm_page_t m = NULL; 635 struct vpgqueues *pq; 636 637 pq = &vm_page_queues[basequeue]; 638 639 /* 640 * Note that for the first loop, index+i and index-i wind up at the 641 * same place. Even though this is not totally optimal, we've already 642 * blown it by missing the cache case so we do not care. 643 */ 644 645 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 646 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 647 break; 648 649 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 650 break; 651 } 652 return(m); 653 } 654 655 /* 656 * Must be called with vm_token held if the caller desired non-blocking 657 * operation and a stable result. 658 */ 659 vm_page_t 660 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 661 { 662 return(_vm_page_list_find(basequeue, index, prefer_zero)); 663 } 664 665 /* 666 * Find a page on the cache queue with color optimization. As pages 667 * might be found, but not applicable, they are deactivated. This 668 * keeps us from using potentially busy cached pages. 669 * 670 * This routine may not block. 671 * Must be called with vm_token held. 672 */ 673 vm_page_t 674 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 675 { 676 vm_page_t m; 677 678 ASSERT_LWKT_TOKEN_HELD(&vm_token); 679 while (TRUE) { 680 m = _vm_page_list_find( 681 PQ_CACHE, 682 (pindex + object->pg_color) & PQ_L2_MASK, 683 FALSE 684 ); 685 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 686 m->hold_count || m->wire_count)) { 687 vm_page_deactivate(m); 688 continue; 689 } 690 return m; 691 } 692 /* not reached */ 693 } 694 695 /* 696 * Find a free or zero page, with specified preference. We attempt to 697 * inline the nominal case and fall back to _vm_page_select_free() 698 * otherwise. 699 * 700 * This routine must be called with a critical section held. 701 * This routine may not block. 702 */ 703 static __inline vm_page_t 704 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 705 { 706 vm_page_t m; 707 708 m = _vm_page_list_find( 709 PQ_FREE, 710 (pindex + object->pg_color) & PQ_L2_MASK, 711 prefer_zero 712 ); 713 return(m); 714 } 715 716 /* 717 * vm_page_alloc() 718 * 719 * Allocate and return a memory cell associated with this VM object/offset 720 * pair. 721 * 722 * page_req classes: 723 * 724 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 725 * VM_ALLOC_QUICK like normal but cannot use cache 726 * VM_ALLOC_SYSTEM greater free drain 727 * VM_ALLOC_INTERRUPT allow free list to be completely drained 728 * VM_ALLOC_ZERO advisory request for pre-zero'd page 729 * 730 * The object must be locked. 731 * This routine may not block. 732 * The returned page will be marked PG_BUSY 733 * 734 * Additional special handling is required when called from an interrupt 735 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 736 * in this case. 737 */ 738 vm_page_t 739 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 740 { 741 vm_page_t m = NULL; 742 743 crit_enter(); 744 lwkt_gettoken(&vm_token); 745 746 KKASSERT(object != NULL); 747 KASSERT(!vm_page_lookup(object, pindex), 748 ("vm_page_alloc: page already allocated")); 749 KKASSERT(page_req & 750 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 751 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 752 753 /* 754 * Certain system threads (pageout daemon, buf_daemon's) are 755 * allowed to eat deeper into the free page list. 756 */ 757 if (curthread->td_flags & TDF_SYSTHREAD) 758 page_req |= VM_ALLOC_SYSTEM; 759 760 loop: 761 if (vmstats.v_free_count > vmstats.v_free_reserved || 762 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 763 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 764 vmstats.v_free_count > vmstats.v_interrupt_free_min) 765 ) { 766 /* 767 * The free queue has sufficient free pages to take one out. 768 */ 769 if (page_req & VM_ALLOC_ZERO) 770 m = vm_page_select_free(object, pindex, TRUE); 771 else 772 m = vm_page_select_free(object, pindex, FALSE); 773 } else if (page_req & VM_ALLOC_NORMAL) { 774 /* 775 * Allocatable from the cache (non-interrupt only). On 776 * success, we must free the page and try again, thus 777 * ensuring that vmstats.v_*_free_min counters are replenished. 778 */ 779 #ifdef INVARIANTS 780 if (curthread->td_preempted) { 781 kprintf("vm_page_alloc(): warning, attempt to allocate" 782 " cache page from preempting interrupt\n"); 783 m = NULL; 784 } else { 785 m = vm_page_select_cache(object, pindex); 786 } 787 #else 788 m = vm_page_select_cache(object, pindex); 789 #endif 790 /* 791 * On success move the page into the free queue and loop. 792 */ 793 if (m != NULL) { 794 KASSERT(m->dirty == 0, 795 ("Found dirty cache page %p", m)); 796 vm_page_busy(m); 797 vm_page_protect(m, VM_PROT_NONE); 798 vm_page_free(m); 799 goto loop; 800 } 801 802 /* 803 * On failure return NULL 804 */ 805 lwkt_reltoken(&vm_token); 806 crit_exit(); 807 #if defined(DIAGNOSTIC) 808 if (vmstats.v_cache_count > 0) 809 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 810 #endif 811 vm_pageout_deficit++; 812 pagedaemon_wakeup(); 813 return (NULL); 814 } else { 815 /* 816 * No pages available, wakeup the pageout daemon and give up. 817 */ 818 lwkt_reltoken(&vm_token); 819 crit_exit(); 820 vm_pageout_deficit++; 821 pagedaemon_wakeup(); 822 return (NULL); 823 } 824 825 /* 826 * Good page found. The page has not yet been busied. We are in 827 * a critical section. 828 */ 829 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 830 KASSERT(m->dirty == 0, 831 ("vm_page_alloc: free/cache page %p was dirty", m)); 832 833 /* 834 * Remove from free queue 835 */ 836 vm_page_unqueue_nowakeup(m); 837 838 /* 839 * Initialize structure. Only the PG_ZERO flag is inherited. Set 840 * the page PG_BUSY 841 */ 842 if (m->flags & PG_ZERO) { 843 vm_page_zero_count--; 844 m->flags = PG_ZERO | PG_BUSY; 845 } else { 846 m->flags = PG_BUSY; 847 } 848 m->wire_count = 0; 849 m->hold_count = 0; 850 m->act_count = 0; 851 m->busy = 0; 852 m->valid = 0; 853 854 /* 855 * vm_page_insert() is safe prior to the crit_exit(). Note also that 856 * inserting a page here does not insert it into the pmap (which 857 * could cause us to block allocating memory). We cannot block 858 * anywhere. 859 */ 860 vm_page_insert(m, object, pindex); 861 862 /* 863 * Don't wakeup too often - wakeup the pageout daemon when 864 * we would be nearly out of memory. 865 */ 866 pagedaemon_wakeup(); 867 868 lwkt_reltoken(&vm_token); 869 crit_exit(); 870 871 /* 872 * A PG_BUSY page is returned. 873 */ 874 return (m); 875 } 876 877 /* 878 * Wait for sufficient free memory for nominal heavy memory use kernel 879 * operations. 880 */ 881 void 882 vm_wait_nominal(void) 883 { 884 while (vm_page_count_min(0)) 885 vm_wait(0); 886 } 887 888 /* 889 * Test if vm_wait_nominal() would block. 890 */ 891 int 892 vm_test_nominal(void) 893 { 894 if (vm_page_count_min(0)) 895 return(1); 896 return(0); 897 } 898 899 /* 900 * Block until free pages are available for allocation, called in various 901 * places before memory allocations. 902 */ 903 void 904 vm_wait(int timo) 905 { 906 crit_enter(); 907 lwkt_gettoken(&vm_token); 908 if (curthread == pagethread) { 909 vm_pageout_pages_needed = 1; 910 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 911 } else { 912 if (vm_pages_needed == 0) { 913 vm_pages_needed = 1; 914 wakeup(&vm_pages_needed); 915 } 916 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 917 } 918 lwkt_reltoken(&vm_token); 919 crit_exit(); 920 } 921 922 /* 923 * Block until free pages are available for allocation 924 * 925 * Called only in vm_fault so that processes page faulting can be 926 * easily tracked. 927 */ 928 void 929 vm_waitpfault(void) 930 { 931 crit_enter(); 932 lwkt_gettoken(&vm_token); 933 if (vm_pages_needed == 0) { 934 vm_pages_needed = 1; 935 wakeup(&vm_pages_needed); 936 } 937 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 938 lwkt_reltoken(&vm_token); 939 crit_exit(); 940 } 941 942 /* 943 * Put the specified page on the active list (if appropriate). Ensure 944 * that act_count is at least ACT_INIT but do not otherwise mess with it. 945 * 946 * The page queues must be locked. 947 * This routine may not block. 948 */ 949 void 950 vm_page_activate(vm_page_t m) 951 { 952 crit_enter(); 953 lwkt_gettoken(&vm_token); 954 if (m->queue != PQ_ACTIVE) { 955 if ((m->queue - m->pc) == PQ_CACHE) 956 mycpu->gd_cnt.v_reactivated++; 957 958 vm_page_unqueue(m); 959 960 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 961 m->queue = PQ_ACTIVE; 962 vm_page_queues[PQ_ACTIVE].lcnt++; 963 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 964 m, pageq); 965 if (m->act_count < ACT_INIT) 966 m->act_count = ACT_INIT; 967 vmstats.v_active_count++; 968 } 969 } else { 970 if (m->act_count < ACT_INIT) 971 m->act_count = ACT_INIT; 972 } 973 lwkt_reltoken(&vm_token); 974 crit_exit(); 975 } 976 977 /* 978 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 979 * routine is called when a page has been added to the cache or free 980 * queues. 981 * 982 * This routine may not block. 983 * This routine must be called at splvm() 984 */ 985 static __inline void 986 vm_page_free_wakeup(void) 987 { 988 /* 989 * if pageout daemon needs pages, then tell it that there are 990 * some free. 991 */ 992 if (vm_pageout_pages_needed && 993 vmstats.v_cache_count + vmstats.v_free_count >= 994 vmstats.v_pageout_free_min 995 ) { 996 wakeup(&vm_pageout_pages_needed); 997 vm_pageout_pages_needed = 0; 998 } 999 1000 /* 1001 * wakeup processes that are waiting on memory if we hit a 1002 * high water mark. And wakeup scheduler process if we have 1003 * lots of memory. this process will swapin processes. 1004 */ 1005 if (vm_pages_needed && !vm_page_count_min(0)) { 1006 vm_pages_needed = 0; 1007 wakeup(&vmstats.v_free_count); 1008 } 1009 } 1010 1011 /* 1012 * vm_page_free_toq: 1013 * 1014 * Returns the given page to the PQ_FREE list, disassociating it with 1015 * any VM object. 1016 * 1017 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1018 * return (the page will have been freed). No particular spl is required 1019 * on entry. 1020 * 1021 * This routine may not block. 1022 */ 1023 void 1024 vm_page_free_toq(vm_page_t m) 1025 { 1026 struct vpgqueues *pq; 1027 1028 crit_enter(); 1029 lwkt_gettoken(&vm_token); 1030 mycpu->gd_cnt.v_tfree++; 1031 1032 KKASSERT((m->flags & PG_MAPPED) == 0); 1033 1034 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1035 kprintf( 1036 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1037 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1038 m->hold_count); 1039 if ((m->queue - m->pc) == PQ_FREE) 1040 panic("vm_page_free: freeing free page"); 1041 else 1042 panic("vm_page_free: freeing busy page"); 1043 } 1044 1045 /* 1046 * unqueue, then remove page. Note that we cannot destroy 1047 * the page here because we do not want to call the pager's 1048 * callback routine until after we've put the page on the 1049 * appropriate free queue. 1050 */ 1051 vm_page_unqueue_nowakeup(m); 1052 vm_page_remove(m); 1053 1054 /* 1055 * No further management of fictitious pages occurs beyond object 1056 * and queue removal. 1057 */ 1058 if ((m->flags & PG_FICTITIOUS) != 0) { 1059 vm_page_wakeup(m); 1060 lwkt_reltoken(&vm_token); 1061 crit_exit(); 1062 return; 1063 } 1064 1065 m->valid = 0; 1066 vm_page_undirty(m); 1067 1068 if (m->wire_count != 0) { 1069 if (m->wire_count > 1) { 1070 panic( 1071 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1072 m->wire_count, (long)m->pindex); 1073 } 1074 panic("vm_page_free: freeing wired page"); 1075 } 1076 1077 /* 1078 * Clear the UNMANAGED flag when freeing an unmanaged page. 1079 */ 1080 if (m->flags & PG_UNMANAGED) { 1081 m->flags &= ~PG_UNMANAGED; 1082 } 1083 1084 if (m->hold_count != 0) { 1085 m->flags &= ~PG_ZERO; 1086 m->queue = PQ_HOLD; 1087 } else { 1088 m->queue = PQ_FREE + m->pc; 1089 } 1090 pq = &vm_page_queues[m->queue]; 1091 pq->lcnt++; 1092 ++(*pq->cnt); 1093 1094 /* 1095 * Put zero'd pages on the end ( where we look for zero'd pages 1096 * first ) and non-zerod pages at the head. 1097 */ 1098 if (m->flags & PG_ZERO) { 1099 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1100 ++vm_page_zero_count; 1101 } else { 1102 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1103 } 1104 vm_page_wakeup(m); 1105 vm_page_free_wakeup(); 1106 lwkt_reltoken(&vm_token); 1107 crit_exit(); 1108 } 1109 1110 /* 1111 * vm_page_free_fromq_fast() 1112 * 1113 * Remove a non-zero page from one of the free queues; the page is removed for 1114 * zeroing, so do not issue a wakeup. 1115 * 1116 * MPUNSAFE 1117 */ 1118 vm_page_t 1119 vm_page_free_fromq_fast(void) 1120 { 1121 static int qi; 1122 vm_page_t m; 1123 int i; 1124 1125 crit_enter(); 1126 lwkt_gettoken(&vm_token); 1127 for (i = 0; i < PQ_L2_SIZE; ++i) { 1128 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1129 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1130 if (m && (m->flags & PG_ZERO) == 0) { 1131 vm_page_unqueue_nowakeup(m); 1132 vm_page_busy(m); 1133 break; 1134 } 1135 m = NULL; 1136 } 1137 lwkt_reltoken(&vm_token); 1138 crit_exit(); 1139 return (m); 1140 } 1141 1142 /* 1143 * vm_page_unmanage() 1144 * 1145 * Prevent PV management from being done on the page. The page is 1146 * removed from the paging queues as if it were wired, and as a 1147 * consequence of no longer being managed the pageout daemon will not 1148 * touch it (since there is no way to locate the pte mappings for the 1149 * page). madvise() calls that mess with the pmap will also no longer 1150 * operate on the page. 1151 * 1152 * Beyond that the page is still reasonably 'normal'. Freeing the page 1153 * will clear the flag. 1154 * 1155 * This routine is used by OBJT_PHYS objects - objects using unswappable 1156 * physical memory as backing store rather then swap-backed memory and 1157 * will eventually be extended to support 4MB unmanaged physical 1158 * mappings. 1159 * 1160 * Must be called with a critical section held. 1161 * Must be called with vm_token held. 1162 */ 1163 void 1164 vm_page_unmanage(vm_page_t m) 1165 { 1166 ASSERT_IN_CRIT_SECTION(); 1167 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1168 if ((m->flags & PG_UNMANAGED) == 0) { 1169 if (m->wire_count == 0) 1170 vm_page_unqueue(m); 1171 } 1172 vm_page_flag_set(m, PG_UNMANAGED); 1173 } 1174 1175 /* 1176 * Mark this page as wired down by yet another map, removing it from 1177 * paging queues as necessary. 1178 * 1179 * The page queues must be locked. 1180 * This routine may not block. 1181 */ 1182 void 1183 vm_page_wire(vm_page_t m) 1184 { 1185 /* 1186 * Only bump the wire statistics if the page is not already wired, 1187 * and only unqueue the page if it is on some queue (if it is unmanaged 1188 * it is already off the queues). Don't do anything with fictitious 1189 * pages because they are always wired. 1190 */ 1191 crit_enter(); 1192 lwkt_gettoken(&vm_token); 1193 if ((m->flags & PG_FICTITIOUS) == 0) { 1194 if (m->wire_count == 0) { 1195 if ((m->flags & PG_UNMANAGED) == 0) 1196 vm_page_unqueue(m); 1197 vmstats.v_wire_count++; 1198 } 1199 m->wire_count++; 1200 KASSERT(m->wire_count != 0, 1201 ("vm_page_wire: wire_count overflow m=%p", m)); 1202 } 1203 lwkt_reltoken(&vm_token); 1204 crit_exit(); 1205 } 1206 1207 /* 1208 * Release one wiring of this page, potentially enabling it to be paged again. 1209 * 1210 * Many pages placed on the inactive queue should actually go 1211 * into the cache, but it is difficult to figure out which. What 1212 * we do instead, if the inactive target is well met, is to put 1213 * clean pages at the head of the inactive queue instead of the tail. 1214 * This will cause them to be moved to the cache more quickly and 1215 * if not actively re-referenced, freed more quickly. If we just 1216 * stick these pages at the end of the inactive queue, heavy filesystem 1217 * meta-data accesses can cause an unnecessary paging load on memory bound 1218 * processes. This optimization causes one-time-use metadata to be 1219 * reused more quickly. 1220 * 1221 * BUT, if we are in a low-memory situation we have no choice but to 1222 * put clean pages on the cache queue. 1223 * 1224 * A number of routines use vm_page_unwire() to guarantee that the page 1225 * will go into either the inactive or active queues, and will NEVER 1226 * be placed in the cache - for example, just after dirtying a page. 1227 * dirty pages in the cache are not allowed. 1228 * 1229 * The page queues must be locked. 1230 * This routine may not block. 1231 */ 1232 void 1233 vm_page_unwire(vm_page_t m, int activate) 1234 { 1235 crit_enter(); 1236 lwkt_gettoken(&vm_token); 1237 if (m->flags & PG_FICTITIOUS) { 1238 /* do nothing */ 1239 } else if (m->wire_count <= 0) { 1240 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1241 } else { 1242 if (--m->wire_count == 0) { 1243 --vmstats.v_wire_count; 1244 if (m->flags & PG_UNMANAGED) { 1245 ; 1246 } else if (activate) { 1247 TAILQ_INSERT_TAIL( 1248 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1249 m->queue = PQ_ACTIVE; 1250 vm_page_queues[PQ_ACTIVE].lcnt++; 1251 vmstats.v_active_count++; 1252 } else { 1253 vm_page_flag_clear(m, PG_WINATCFLS); 1254 TAILQ_INSERT_TAIL( 1255 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1256 m->queue = PQ_INACTIVE; 1257 vm_page_queues[PQ_INACTIVE].lcnt++; 1258 vmstats.v_inactive_count++; 1259 ++vm_swapcache_inactive_heuristic; 1260 } 1261 } 1262 } 1263 lwkt_reltoken(&vm_token); 1264 crit_exit(); 1265 } 1266 1267 1268 /* 1269 * Move the specified page to the inactive queue. If the page has 1270 * any associated swap, the swap is deallocated. 1271 * 1272 * Normally athead is 0 resulting in LRU operation. athead is set 1273 * to 1 if we want this page to be 'as if it were placed in the cache', 1274 * except without unmapping it from the process address space. 1275 * 1276 * This routine may not block. 1277 * The caller must hold vm_token. 1278 */ 1279 static __inline void 1280 _vm_page_deactivate(vm_page_t m, int athead) 1281 { 1282 /* 1283 * Ignore if already inactive. 1284 */ 1285 if (m->queue == PQ_INACTIVE) 1286 return; 1287 1288 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1289 if ((m->queue - m->pc) == PQ_CACHE) 1290 mycpu->gd_cnt.v_reactivated++; 1291 vm_page_flag_clear(m, PG_WINATCFLS); 1292 vm_page_unqueue(m); 1293 if (athead) { 1294 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1295 m, pageq); 1296 } else { 1297 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1298 m, pageq); 1299 ++vm_swapcache_inactive_heuristic; 1300 } 1301 m->queue = PQ_INACTIVE; 1302 vm_page_queues[PQ_INACTIVE].lcnt++; 1303 vmstats.v_inactive_count++; 1304 } 1305 } 1306 1307 /* 1308 * Attempt to deactivate a page. 1309 * 1310 * No requirements. 1311 */ 1312 void 1313 vm_page_deactivate(vm_page_t m) 1314 { 1315 crit_enter(); 1316 lwkt_gettoken(&vm_token); 1317 _vm_page_deactivate(m, 0); 1318 lwkt_reltoken(&vm_token); 1319 crit_exit(); 1320 } 1321 1322 /* 1323 * Attempt to move a page to PQ_CACHE. 1324 * Returns 0 on failure, 1 on success 1325 * 1326 * No requirements. 1327 */ 1328 int 1329 vm_page_try_to_cache(vm_page_t m) 1330 { 1331 crit_enter(); 1332 lwkt_gettoken(&vm_token); 1333 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1334 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1335 lwkt_reltoken(&vm_token); 1336 crit_exit(); 1337 return(0); 1338 } 1339 vm_page_test_dirty(m); 1340 if (m->dirty) { 1341 lwkt_reltoken(&vm_token); 1342 crit_exit(); 1343 return(0); 1344 } 1345 vm_page_cache(m); 1346 lwkt_reltoken(&vm_token); 1347 crit_exit(); 1348 return(1); 1349 } 1350 1351 /* 1352 * Attempt to free the page. If we cannot free it, we do nothing. 1353 * 1 is returned on success, 0 on failure. 1354 * 1355 * No requirements. 1356 */ 1357 int 1358 vm_page_try_to_free(vm_page_t m) 1359 { 1360 crit_enter(); 1361 lwkt_gettoken(&vm_token); 1362 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1363 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1364 lwkt_reltoken(&vm_token); 1365 crit_exit(); 1366 return(0); 1367 } 1368 vm_page_test_dirty(m); 1369 if (m->dirty) { 1370 lwkt_reltoken(&vm_token); 1371 crit_exit(); 1372 return(0); 1373 } 1374 vm_page_busy(m); 1375 vm_page_protect(m, VM_PROT_NONE); 1376 vm_page_free(m); 1377 lwkt_reltoken(&vm_token); 1378 crit_exit(); 1379 return(1); 1380 } 1381 1382 /* 1383 * vm_page_cache 1384 * 1385 * Put the specified page onto the page cache queue (if appropriate). 1386 * 1387 * The caller must hold vm_token. 1388 * This routine may not block. 1389 */ 1390 void 1391 vm_page_cache(vm_page_t m) 1392 { 1393 ASSERT_IN_CRIT_SECTION(); 1394 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1395 1396 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1397 m->wire_count || m->hold_count) { 1398 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1399 return; 1400 } 1401 1402 /* 1403 * Already in the cache (and thus not mapped) 1404 */ 1405 if ((m->queue - m->pc) == PQ_CACHE) { 1406 KKASSERT((m->flags & PG_MAPPED) == 0); 1407 return; 1408 } 1409 1410 /* 1411 * Caller is required to test m->dirty, but note that the act of 1412 * removing the page from its maps can cause it to become dirty 1413 * on an SMP system due to another cpu running in usermode. 1414 */ 1415 if (m->dirty) { 1416 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1417 (long)m->pindex); 1418 } 1419 1420 /* 1421 * Remove all pmaps and indicate that the page is not 1422 * writeable or mapped. Our vm_page_protect() call may 1423 * have blocked (especially w/ VM_PROT_NONE), so recheck 1424 * everything. 1425 */ 1426 vm_page_busy(m); 1427 vm_page_protect(m, VM_PROT_NONE); 1428 vm_page_wakeup(m); 1429 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy || 1430 m->wire_count || m->hold_count) { 1431 /* do nothing */ 1432 } else if (m->dirty) { 1433 vm_page_deactivate(m); 1434 } else { 1435 vm_page_unqueue_nowakeup(m); 1436 m->queue = PQ_CACHE + m->pc; 1437 vm_page_queues[m->queue].lcnt++; 1438 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1439 vmstats.v_cache_count++; 1440 vm_page_free_wakeup(); 1441 } 1442 } 1443 1444 /* 1445 * vm_page_dontneed() 1446 * 1447 * Cache, deactivate, or do nothing as appropriate. This routine 1448 * is typically used by madvise() MADV_DONTNEED. 1449 * 1450 * Generally speaking we want to move the page into the cache so 1451 * it gets reused quickly. However, this can result in a silly syndrome 1452 * due to the page recycling too quickly. Small objects will not be 1453 * fully cached. On the otherhand, if we move the page to the inactive 1454 * queue we wind up with a problem whereby very large objects 1455 * unnecessarily blow away our inactive and cache queues. 1456 * 1457 * The solution is to move the pages based on a fixed weighting. We 1458 * either leave them alone, deactivate them, or move them to the cache, 1459 * where moving them to the cache has the highest weighting. 1460 * By forcing some pages into other queues we eventually force the 1461 * system to balance the queues, potentially recovering other unrelated 1462 * space from active. The idea is to not force this to happen too 1463 * often. 1464 * 1465 * No requirements. 1466 */ 1467 void 1468 vm_page_dontneed(vm_page_t m) 1469 { 1470 static int dnweight; 1471 int dnw; 1472 int head; 1473 1474 dnw = ++dnweight; 1475 1476 /* 1477 * occassionally leave the page alone 1478 */ 1479 crit_enter(); 1480 lwkt_gettoken(&vm_token); 1481 if ((dnw & 0x01F0) == 0 || 1482 m->queue == PQ_INACTIVE || 1483 m->queue - m->pc == PQ_CACHE 1484 ) { 1485 if (m->act_count >= ACT_INIT) 1486 --m->act_count; 1487 lwkt_reltoken(&vm_token); 1488 crit_exit(); 1489 return; 1490 } 1491 1492 if (m->dirty == 0) 1493 vm_page_test_dirty(m); 1494 1495 if (m->dirty || (dnw & 0x0070) == 0) { 1496 /* 1497 * Deactivate the page 3 times out of 32. 1498 */ 1499 head = 0; 1500 } else { 1501 /* 1502 * Cache the page 28 times out of every 32. Note that 1503 * the page is deactivated instead of cached, but placed 1504 * at the head of the queue instead of the tail. 1505 */ 1506 head = 1; 1507 } 1508 _vm_page_deactivate(m, head); 1509 lwkt_reltoken(&vm_token); 1510 crit_exit(); 1511 } 1512 1513 /* 1514 * Grab a page, blocking if it is busy and allocating a page if necessary. 1515 * A busy page is returned or NULL. 1516 * 1517 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1518 * If VM_ALLOC_RETRY is not specified 1519 * 1520 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1521 * always returned if we had blocked. 1522 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1523 * This routine may not be called from an interrupt. 1524 * The returned page may not be entirely valid. 1525 * 1526 * This routine may be called from mainline code without spl protection and 1527 * be guarenteed a busied page associated with the object at the specified 1528 * index. 1529 * 1530 * No requirements. 1531 */ 1532 vm_page_t 1533 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1534 { 1535 vm_page_t m; 1536 int generation; 1537 1538 KKASSERT(allocflags & 1539 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1540 crit_enter(); 1541 lwkt_gettoken(&vm_token); 1542 retrylookup: 1543 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1544 if (m->busy || (m->flags & PG_BUSY)) { 1545 generation = object->generation; 1546 1547 while ((object->generation == generation) && 1548 (m->busy || (m->flags & PG_BUSY))) { 1549 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1550 tsleep(m, 0, "pgrbwt", 0); 1551 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1552 m = NULL; 1553 goto done; 1554 } 1555 } 1556 goto retrylookup; 1557 } else { 1558 vm_page_busy(m); 1559 goto done; 1560 } 1561 } 1562 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1563 if (m == NULL) { 1564 vm_wait(0); 1565 if ((allocflags & VM_ALLOC_RETRY) == 0) 1566 goto done; 1567 goto retrylookup; 1568 } 1569 done: 1570 lwkt_reltoken(&vm_token); 1571 crit_exit(); 1572 return(m); 1573 } 1574 1575 /* 1576 * Mapping function for valid bits or for dirty bits in 1577 * a page. May not block. 1578 * 1579 * Inputs are required to range within a page. 1580 * 1581 * No requirements. 1582 * Non blocking. 1583 */ 1584 int 1585 vm_page_bits(int base, int size) 1586 { 1587 int first_bit; 1588 int last_bit; 1589 1590 KASSERT( 1591 base + size <= PAGE_SIZE, 1592 ("vm_page_bits: illegal base/size %d/%d", base, size) 1593 ); 1594 1595 if (size == 0) /* handle degenerate case */ 1596 return(0); 1597 1598 first_bit = base >> DEV_BSHIFT; 1599 last_bit = (base + size - 1) >> DEV_BSHIFT; 1600 1601 return ((2 << last_bit) - (1 << first_bit)); 1602 } 1603 1604 /* 1605 * Sets portions of a page valid and clean. The arguments are expected 1606 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1607 * of any partial chunks touched by the range. The invalid portion of 1608 * such chunks will be zero'd. 1609 * 1610 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1611 * align base to DEV_BSIZE so as not to mark clean a partially 1612 * truncated device block. Otherwise the dirty page status might be 1613 * lost. 1614 * 1615 * This routine may not block. 1616 * 1617 * (base + size) must be less then or equal to PAGE_SIZE. 1618 */ 1619 static void 1620 _vm_page_zero_valid(vm_page_t m, int base, int size) 1621 { 1622 int frag; 1623 int endoff; 1624 1625 if (size == 0) /* handle degenerate case */ 1626 return; 1627 1628 /* 1629 * If the base is not DEV_BSIZE aligned and the valid 1630 * bit is clear, we have to zero out a portion of the 1631 * first block. 1632 */ 1633 1634 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1635 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1636 ) { 1637 pmap_zero_page_area( 1638 VM_PAGE_TO_PHYS(m), 1639 frag, 1640 base - frag 1641 ); 1642 } 1643 1644 /* 1645 * If the ending offset is not DEV_BSIZE aligned and the 1646 * valid bit is clear, we have to zero out a portion of 1647 * the last block. 1648 */ 1649 1650 endoff = base + size; 1651 1652 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1653 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1654 ) { 1655 pmap_zero_page_area( 1656 VM_PAGE_TO_PHYS(m), 1657 endoff, 1658 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1659 ); 1660 } 1661 } 1662 1663 /* 1664 * Set valid, clear dirty bits. If validating the entire 1665 * page we can safely clear the pmap modify bit. We also 1666 * use this opportunity to clear the PG_NOSYNC flag. If a process 1667 * takes a write fault on a MAP_NOSYNC memory area the flag will 1668 * be set again. 1669 * 1670 * We set valid bits inclusive of any overlap, but we can only 1671 * clear dirty bits for DEV_BSIZE chunks that are fully within 1672 * the range. 1673 * 1674 * Page must be busied? 1675 * No other requirements. 1676 */ 1677 void 1678 vm_page_set_valid(vm_page_t m, int base, int size) 1679 { 1680 _vm_page_zero_valid(m, base, size); 1681 m->valid |= vm_page_bits(base, size); 1682 } 1683 1684 1685 /* 1686 * Set valid bits and clear dirty bits. 1687 * 1688 * NOTE: This function does not clear the pmap modified bit. 1689 * Also note that e.g. NFS may use a byte-granular base 1690 * and size. 1691 * 1692 * Page must be busied? 1693 * No other requirements. 1694 */ 1695 void 1696 vm_page_set_validclean(vm_page_t m, int base, int size) 1697 { 1698 int pagebits; 1699 1700 _vm_page_zero_valid(m, base, size); 1701 pagebits = vm_page_bits(base, size); 1702 m->valid |= pagebits; 1703 m->dirty &= ~pagebits; 1704 if (base == 0 && size == PAGE_SIZE) { 1705 /*pmap_clear_modify(m);*/ 1706 vm_page_flag_clear(m, PG_NOSYNC); 1707 } 1708 } 1709 1710 /* 1711 * Set valid & dirty. Used by buwrite() 1712 * 1713 * Page must be busied? 1714 * No other requirements. 1715 */ 1716 void 1717 vm_page_set_validdirty(vm_page_t m, int base, int size) 1718 { 1719 int pagebits; 1720 1721 pagebits = vm_page_bits(base, size); 1722 m->valid |= pagebits; 1723 m->dirty |= pagebits; 1724 if (m->object) 1725 vm_object_set_writeable_dirty(m->object); 1726 } 1727 1728 /* 1729 * Clear dirty bits. 1730 * 1731 * NOTE: This function does not clear the pmap modified bit. 1732 * Also note that e.g. NFS may use a byte-granular base 1733 * and size. 1734 * 1735 * Page must be busied? 1736 * No other requirements. 1737 */ 1738 void 1739 vm_page_clear_dirty(vm_page_t m, int base, int size) 1740 { 1741 m->dirty &= ~vm_page_bits(base, size); 1742 if (base == 0 && size == PAGE_SIZE) { 1743 /*pmap_clear_modify(m);*/ 1744 vm_page_flag_clear(m, PG_NOSYNC); 1745 } 1746 } 1747 1748 /* 1749 * Make the page all-dirty. 1750 * 1751 * Also make sure the related object and vnode reflect the fact that the 1752 * object may now contain a dirty page. 1753 * 1754 * Page must be busied? 1755 * No other requirements. 1756 */ 1757 void 1758 vm_page_dirty(vm_page_t m) 1759 { 1760 #ifdef INVARIANTS 1761 int pqtype = m->queue - m->pc; 1762 #endif 1763 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1764 ("vm_page_dirty: page in free/cache queue!")); 1765 if (m->dirty != VM_PAGE_BITS_ALL) { 1766 m->dirty = VM_PAGE_BITS_ALL; 1767 if (m->object) 1768 vm_object_set_writeable_dirty(m->object); 1769 } 1770 } 1771 1772 /* 1773 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1774 * valid and dirty bits for the effected areas are cleared. 1775 * 1776 * Page must be busied? 1777 * Does not block. 1778 * No other requirements. 1779 */ 1780 void 1781 vm_page_set_invalid(vm_page_t m, int base, int size) 1782 { 1783 int bits; 1784 1785 bits = vm_page_bits(base, size); 1786 m->valid &= ~bits; 1787 m->dirty &= ~bits; 1788 m->object->generation++; 1789 } 1790 1791 /* 1792 * The kernel assumes that the invalid portions of a page contain 1793 * garbage, but such pages can be mapped into memory by user code. 1794 * When this occurs, we must zero out the non-valid portions of the 1795 * page so user code sees what it expects. 1796 * 1797 * Pages are most often semi-valid when the end of a file is mapped 1798 * into memory and the file's size is not page aligned. 1799 * 1800 * Page must be busied? 1801 * No other requirements. 1802 */ 1803 void 1804 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1805 { 1806 int b; 1807 int i; 1808 1809 /* 1810 * Scan the valid bits looking for invalid sections that 1811 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1812 * valid bit may be set ) have already been zerod by 1813 * vm_page_set_validclean(). 1814 */ 1815 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1816 if (i == (PAGE_SIZE / DEV_BSIZE) || 1817 (m->valid & (1 << i)) 1818 ) { 1819 if (i > b) { 1820 pmap_zero_page_area( 1821 VM_PAGE_TO_PHYS(m), 1822 b << DEV_BSHIFT, 1823 (i - b) << DEV_BSHIFT 1824 ); 1825 } 1826 b = i + 1; 1827 } 1828 } 1829 1830 /* 1831 * setvalid is TRUE when we can safely set the zero'd areas 1832 * as being valid. We can do this if there are no cache consistency 1833 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1834 */ 1835 if (setvalid) 1836 m->valid = VM_PAGE_BITS_ALL; 1837 } 1838 1839 /* 1840 * Is a (partial) page valid? Note that the case where size == 0 1841 * will return FALSE in the degenerate case where the page is entirely 1842 * invalid, and TRUE otherwise. 1843 * 1844 * Does not block. 1845 * No other requirements. 1846 */ 1847 int 1848 vm_page_is_valid(vm_page_t m, int base, int size) 1849 { 1850 int bits = vm_page_bits(base, size); 1851 1852 if (m->valid && ((m->valid & bits) == bits)) 1853 return 1; 1854 else 1855 return 0; 1856 } 1857 1858 /* 1859 * update dirty bits from pmap/mmu. May not block. 1860 * 1861 * Caller must hold vm_token if non-blocking operation desired. 1862 * No other requirements. 1863 */ 1864 void 1865 vm_page_test_dirty(vm_page_t m) 1866 { 1867 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1868 vm_page_dirty(m); 1869 } 1870 } 1871 1872 /* 1873 * Issue an event on a VM page. Corresponding action structures are 1874 * removed from the page's list and called. 1875 */ 1876 void 1877 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1878 { 1879 struct vm_page_action *scan, *next; 1880 1881 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) { 1882 if (scan->event == event) { 1883 scan->event = VMEVENT_NONE; 1884 LIST_REMOVE(scan, entry); 1885 scan->func(m, scan); 1886 } 1887 } 1888 } 1889 1890 1891 #include "opt_ddb.h" 1892 #ifdef DDB 1893 #include <sys/kernel.h> 1894 1895 #include <ddb/ddb.h> 1896 1897 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1898 { 1899 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1900 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1901 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1902 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1903 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1904 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1905 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1906 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1907 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1908 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1909 } 1910 1911 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1912 { 1913 int i; 1914 db_printf("PQ_FREE:"); 1915 for(i=0;i<PQ_L2_SIZE;i++) { 1916 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1917 } 1918 db_printf("\n"); 1919 1920 db_printf("PQ_CACHE:"); 1921 for(i=0;i<PQ_L2_SIZE;i++) { 1922 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1923 } 1924 db_printf("\n"); 1925 1926 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1927 vm_page_queues[PQ_ACTIVE].lcnt, 1928 vm_page_queues[PQ_INACTIVE].lcnt); 1929 } 1930 #endif /* DDB */ 1931