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 #include <sys/kernel.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <sys/lock.h> 85 #include <vm/vm_kern.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pageout.h> 91 #include <vm/vm_pager.h> 92 #include <vm/vm_extern.h> 93 #include <vm/swap_pager.h> 94 95 #include <machine/md_var.h> 96 97 #include <vm/vm_page2.h> 98 99 #define VMACTION_HSIZE 256 100 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 101 102 static void vm_page_queue_init(void); 103 static void vm_page_free_wakeup(void); 104 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 105 static vm_page_t _vm_page_list_find2(int basequeue, int index); 106 107 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 108 109 LIST_HEAD(vm_page_action_list, vm_page_action); 110 struct vm_page_action_list action_list[VMACTION_HSIZE]; 111 static volatile int vm_pages_waiting; 112 113 114 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 115 vm_pindex_t, pindex); 116 117 static void 118 vm_page_queue_init(void) 119 { 120 int i; 121 122 for (i = 0; i < PQ_L2_SIZE; i++) 123 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 124 for (i = 0; i < PQ_L2_SIZE; i++) 125 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 126 127 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 128 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 129 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 130 /* PQ_NONE has no queue */ 131 132 for (i = 0; i < PQ_COUNT; i++) 133 TAILQ_INIT(&vm_page_queues[i].pl); 134 135 for (i = 0; i < VMACTION_HSIZE; i++) 136 LIST_INIT(&action_list[i]); 137 } 138 139 /* 140 * note: place in initialized data section? Is this necessary? 141 */ 142 long first_page = 0; 143 int vm_page_array_size = 0; 144 int vm_page_zero_count = 0; 145 vm_page_t vm_page_array = 0; 146 147 /* 148 * (low level boot) 149 * 150 * Sets the page size, perhaps based upon the memory size. 151 * Must be called before any use of page-size dependent functions. 152 */ 153 void 154 vm_set_page_size(void) 155 { 156 if (vmstats.v_page_size == 0) 157 vmstats.v_page_size = PAGE_SIZE; 158 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 159 panic("vm_set_page_size: page size not a power of two"); 160 } 161 162 /* 163 * (low level boot) 164 * 165 * Add a new page to the freelist for use by the system. New pages 166 * are added to both the head and tail of the associated free page 167 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 168 * requests pull 'recent' adds (higher physical addresses) first. 169 * 170 * Must be called in a critical section. 171 */ 172 vm_page_t 173 vm_add_new_page(vm_paddr_t pa) 174 { 175 struct vpgqueues *vpq; 176 vm_page_t m; 177 178 ++vmstats.v_page_count; 179 ++vmstats.v_free_count; 180 m = PHYS_TO_VM_PAGE(pa); 181 m->phys_addr = pa; 182 m->flags = 0; 183 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 184 m->queue = m->pc + PQ_FREE; 185 KKASSERT(m->dirty == 0); 186 187 vpq = &vm_page_queues[m->queue]; 188 if (vpq->flipflop) 189 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 190 else 191 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 192 vpq->flipflop = 1 - vpq->flipflop; 193 194 vm_page_queues[m->queue].lcnt++; 195 return (m); 196 } 197 198 /* 199 * (low level boot) 200 * 201 * Initializes the resident memory module. 202 * 203 * Preallocates memory for critical VM structures and arrays prior to 204 * kernel_map becoming available. 205 * 206 * Memory is allocated from (virtual2_start, virtual2_end) if available, 207 * otherwise memory is allocated from (virtual_start, virtual_end). 208 * 209 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 210 * large enough to hold vm_page_array & other structures for machines with 211 * large amounts of ram, so we want to use virtual2* when available. 212 */ 213 void 214 vm_page_startup(void) 215 { 216 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 217 vm_offset_t mapped; 218 vm_size_t npages; 219 vm_paddr_t page_range; 220 vm_paddr_t new_end; 221 int i; 222 vm_paddr_t pa; 223 int nblocks; 224 vm_paddr_t last_pa; 225 vm_paddr_t end; 226 vm_paddr_t biggestone, biggestsize; 227 vm_paddr_t total; 228 229 total = 0; 230 biggestsize = 0; 231 biggestone = 0; 232 nblocks = 0; 233 vaddr = round_page(vaddr); 234 235 for (i = 0; phys_avail[i + 1]; i += 2) { 236 phys_avail[i] = round_page64(phys_avail[i]); 237 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 238 } 239 240 for (i = 0; phys_avail[i + 1]; i += 2) { 241 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 242 243 if (size > biggestsize) { 244 biggestone = i; 245 biggestsize = size; 246 } 247 ++nblocks; 248 total += size; 249 } 250 251 end = phys_avail[biggestone+1]; 252 end = trunc_page(end); 253 254 /* 255 * Initialize the queue headers for the free queue, the active queue 256 * and the inactive queue. 257 */ 258 259 vm_page_queue_init(); 260 261 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */ 262 #if !defined(_KERNEL_VIRTUAL) 263 /* 264 * Allocate a bitmap to indicate that a random physical page 265 * needs to be included in a minidump. 266 * 267 * The amd64 port needs this to indicate which direct map pages 268 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 269 * 270 * However, i386 still needs this workspace internally within the 271 * minidump code. In theory, they are not needed on i386, but are 272 * included should the sf_buf code decide to use them. 273 */ 274 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 275 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 276 end -= vm_page_dump_size; 277 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 278 VM_PROT_READ | VM_PROT_WRITE); 279 bzero((void *)vm_page_dump, vm_page_dump_size); 280 #endif 281 282 /* 283 * Compute the number of pages of memory that will be available for 284 * use (taking into account the overhead of a page structure per 285 * page). 286 */ 287 first_page = phys_avail[0] / PAGE_SIZE; 288 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 289 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 290 291 /* 292 * Initialize the mem entry structures now, and put them in the free 293 * queue. 294 */ 295 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 296 mapped = pmap_map(&vaddr, new_end, end, 297 VM_PROT_READ | VM_PROT_WRITE); 298 vm_page_array = (vm_page_t)mapped; 299 300 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 301 /* 302 * since pmap_map on amd64 returns stuff out of a direct-map region, 303 * we have to manually add these pages to the minidump tracking so 304 * that they can be dumped, including the vm_page_array. 305 */ 306 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 307 dump_add_page(pa); 308 #endif 309 310 /* 311 * Clear all of the page structures 312 */ 313 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 314 vm_page_array_size = page_range; 315 316 /* 317 * Construct the free queue(s) in ascending order (by physical 318 * address) so that the first 16MB of physical memory is allocated 319 * last rather than first. On large-memory machines, this avoids 320 * the exhaustion of low physical memory before isa_dmainit has run. 321 */ 322 vmstats.v_page_count = 0; 323 vmstats.v_free_count = 0; 324 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 325 pa = phys_avail[i]; 326 if (i == biggestone) 327 last_pa = new_end; 328 else 329 last_pa = phys_avail[i + 1]; 330 while (pa < last_pa && npages-- > 0) { 331 vm_add_new_page(pa); 332 pa += PAGE_SIZE; 333 } 334 } 335 if (virtual2_start) 336 virtual2_start = vaddr; 337 else 338 virtual_start = vaddr; 339 } 340 341 /* 342 * Scan comparison function for Red-Black tree scans. An inclusive 343 * (start,end) is expected. Other fields are not used. 344 */ 345 int 346 rb_vm_page_scancmp(struct vm_page *p, void *data) 347 { 348 struct rb_vm_page_scan_info *info = data; 349 350 if (p->pindex < info->start_pindex) 351 return(-1); 352 if (p->pindex > info->end_pindex) 353 return(1); 354 return(0); 355 } 356 357 int 358 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 359 { 360 if (p1->pindex < p2->pindex) 361 return(-1); 362 if (p1->pindex > p2->pindex) 363 return(1); 364 return(0); 365 } 366 367 /* 368 * Holding a page keeps it from being reused. Other parts of the system 369 * can still disassociate the page from its current object and free it, or 370 * perform read or write I/O on it and/or otherwise manipulate the page, 371 * but if the page is held the VM system will leave the page and its data 372 * intact and not reuse the page for other purposes until the last hold 373 * reference is released. (see vm_page_wire() if you want to prevent the 374 * page from being disassociated from its object too). 375 * 376 * The caller must hold vm_token. 377 * 378 * The caller must still validate the contents of the page and, if necessary, 379 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 380 * before manipulating the page. 381 */ 382 void 383 vm_page_hold(vm_page_t m) 384 { 385 ASSERT_LWKT_TOKEN_HELD(&vm_token); 386 ++m->hold_count; 387 } 388 389 /* 390 * The opposite of vm_page_hold(). A page can be freed while being held, 391 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 392 * in this case to actually free it once the hold count drops to 0. 393 * 394 * The caller must hold vm_token if non-blocking operation is desired, 395 * but otherwise does not need to. 396 */ 397 void 398 vm_page_unhold(vm_page_t m) 399 { 400 lwkt_gettoken(&vm_token); 401 --m->hold_count; 402 KASSERT(m->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 403 if (m->hold_count == 0 && m->queue == PQ_HOLD) { 404 vm_page_busy(m); 405 vm_page_free_toq(m); 406 } 407 lwkt_reltoken(&vm_token); 408 } 409 410 /* 411 * Inserts the given vm_page into the object and object list. 412 * 413 * The pagetables are not updated but will presumably fault the page 414 * in if necessary, or if a kernel page the caller will at some point 415 * enter the page into the kernel's pmap. We are not allowed to block 416 * here so we *can't* do this anyway. 417 * 418 * This routine may not block. 419 * This routine must be called with the vm_token held. 420 * This routine must be called with a critical section held. 421 */ 422 void 423 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 424 { 425 ASSERT_LWKT_TOKEN_HELD(&vm_token); 426 if (m->object != NULL) 427 panic("vm_page_insert: already inserted"); 428 429 /* 430 * Record the object/offset pair in this page 431 */ 432 m->object = object; 433 m->pindex = pindex; 434 435 /* 436 * Insert it into the object. 437 */ 438 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 439 object->generation++; 440 441 /* 442 * show that the object has one more resident page. 443 */ 444 object->resident_page_count++; 445 446 /* 447 * Add the pv_list_cout of the page when its inserted in 448 * the object 449 */ 450 object->agg_pv_list_count = object->agg_pv_list_count + m->md.pv_list_count; 451 452 /* 453 * Since we are inserting a new and possibly dirty page, 454 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 455 */ 456 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 457 vm_object_set_writeable_dirty(object); 458 459 /* 460 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 461 */ 462 swap_pager_page_inserted(m); 463 } 464 465 /* 466 * Removes the given vm_page_t from the global (object,index) hash table 467 * and from the object's memq. 468 * 469 * The underlying pmap entry (if any) is NOT removed here. 470 * This routine may not block. 471 * 472 * The page must be BUSY and will remain BUSY on return. 473 * No other requirements. 474 * 475 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 476 * it busy. 477 */ 478 void 479 vm_page_remove(vm_page_t m) 480 { 481 vm_object_t object; 482 483 lwkt_gettoken(&vm_token); 484 if (m->object == NULL) { 485 lwkt_reltoken(&vm_token); 486 return; 487 } 488 489 if ((m->flags & PG_BUSY) == 0) 490 panic("vm_page_remove: page not busy"); 491 492 object = m->object; 493 494 /* 495 * Remove the page from the object and update the object. 496 */ 497 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 498 object->resident_page_count--; 499 object->agg_pv_list_count = object->agg_pv_list_count - m->md.pv_list_count; 500 object->generation++; 501 m->object = NULL; 502 503 lwkt_reltoken(&vm_token); 504 } 505 506 /* 507 * Locate and return the page at (object, pindex), or NULL if the 508 * page could not be found. 509 * 510 * The caller must hold vm_token. 511 */ 512 vm_page_t 513 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 514 { 515 vm_page_t m; 516 517 /* 518 * Search the hash table for this object/offset pair 519 */ 520 ASSERT_LWKT_TOKEN_HELD(&vm_token); 521 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 522 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 523 return(m); 524 } 525 526 /* 527 * vm_page_rename() 528 * 529 * Move the given memory entry from its current object to the specified 530 * target object/offset. 531 * 532 * The object must be locked. 533 * This routine may not block. 534 * 535 * Note: This routine will raise itself to splvm(), the caller need not. 536 * 537 * Note: Swap associated with the page must be invalidated by the move. We 538 * have to do this for several reasons: (1) we aren't freeing the 539 * page, (2) we are dirtying the page, (3) the VM system is probably 540 * moving the page from object A to B, and will then later move 541 * the backing store from A to B and we can't have a conflict. 542 * 543 * Note: We *always* dirty the page. It is necessary both for the 544 * fact that we moved it, and because we may be invalidating 545 * swap. If the page is on the cache, we have to deactivate it 546 * or vm_page_dirty() will panic. Dirty pages are not allowed 547 * on the cache. 548 */ 549 void 550 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 551 { 552 lwkt_gettoken(&vm_token); 553 vm_page_remove(m); 554 vm_page_insert(m, new_object, new_pindex); 555 if (m->queue - m->pc == PQ_CACHE) 556 vm_page_deactivate(m); 557 vm_page_dirty(m); 558 vm_page_wakeup(m); 559 lwkt_reltoken(&vm_token); 560 } 561 562 /* 563 * vm_page_unqueue() without any wakeup. This routine is used when a page 564 * is being moved between queues or otherwise is to remain BUSYied by the 565 * caller. 566 * 567 * The caller must hold vm_token 568 * This routine may not block. 569 */ 570 void 571 vm_page_unqueue_nowakeup(vm_page_t m) 572 { 573 int queue = m->queue; 574 struct vpgqueues *pq; 575 576 ASSERT_LWKT_TOKEN_HELD(&vm_token); 577 if (queue != PQ_NONE) { 578 pq = &vm_page_queues[queue]; 579 m->queue = PQ_NONE; 580 TAILQ_REMOVE(&pq->pl, m, pageq); 581 (*pq->cnt)--; 582 pq->lcnt--; 583 } 584 } 585 586 /* 587 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 588 * if necessary. 589 * 590 * The caller must hold vm_token 591 * This routine may not block. 592 */ 593 void 594 vm_page_unqueue(vm_page_t m) 595 { 596 int queue = m->queue; 597 struct vpgqueues *pq; 598 599 ASSERT_LWKT_TOKEN_HELD(&vm_token); 600 if (queue != PQ_NONE) { 601 m->queue = PQ_NONE; 602 pq = &vm_page_queues[queue]; 603 TAILQ_REMOVE(&pq->pl, m, pageq); 604 (*pq->cnt)--; 605 pq->lcnt--; 606 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 607 pagedaemon_wakeup(); 608 } 609 } 610 611 /* 612 * vm_page_list_find() 613 * 614 * Find a page on the specified queue with color optimization. 615 * 616 * The page coloring optimization attempts to locate a page that does 617 * not overload other nearby pages in the object in the cpu's L1 or L2 618 * caches. We need this optimization because cpu caches tend to be 619 * physical caches, while object spaces tend to be virtual. 620 * 621 * Must be called with vm_token held. 622 * This routine may not block. 623 * 624 * Note that this routine is carefully inlined. A non-inlined version 625 * is available for outside callers but the only critical path is 626 * from within this source file. 627 */ 628 static __inline 629 vm_page_t 630 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 631 { 632 vm_page_t m; 633 634 if (prefer_zero) 635 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 636 else 637 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 638 if (m == NULL) 639 m = _vm_page_list_find2(basequeue, index); 640 return(m); 641 } 642 643 static vm_page_t 644 _vm_page_list_find2(int basequeue, int index) 645 { 646 int i; 647 vm_page_t m = NULL; 648 struct vpgqueues *pq; 649 650 pq = &vm_page_queues[basequeue]; 651 652 /* 653 * Note that for the first loop, index+i and index-i wind up at the 654 * same place. Even though this is not totally optimal, we've already 655 * blown it by missing the cache case so we do not care. 656 */ 657 658 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 659 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 660 break; 661 662 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 663 break; 664 } 665 return(m); 666 } 667 668 /* 669 * Must be called with vm_token held if the caller desired non-blocking 670 * operation and a stable result. 671 */ 672 vm_page_t 673 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 674 { 675 return(_vm_page_list_find(basequeue, index, prefer_zero)); 676 } 677 678 /* 679 * Find a page on the cache queue with color optimization. As pages 680 * might be found, but not applicable, they are deactivated. This 681 * keeps us from using potentially busy cached pages. 682 * 683 * This routine may not block. 684 * Must be called with vm_token held. 685 */ 686 vm_page_t 687 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 688 { 689 vm_page_t m; 690 691 ASSERT_LWKT_TOKEN_HELD(&vm_token); 692 while (TRUE) { 693 m = _vm_page_list_find( 694 PQ_CACHE, 695 (pindex + object->pg_color) & PQ_L2_MASK, 696 FALSE 697 ); 698 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 699 m->hold_count || m->wire_count)) { 700 /* cache page found busy */ 701 vm_page_deactivate(m); 702 #ifdef INVARIANTS 703 kprintf("Warning: busy page %p found in cache\n", m); 704 #endif 705 continue; 706 } 707 return m; 708 } 709 /* not reached */ 710 } 711 712 /* 713 * Find a free or zero page, with specified preference. We attempt to 714 * inline the nominal case and fall back to _vm_page_select_free() 715 * otherwise. 716 * 717 * This routine must be called with a critical section held. 718 * This routine may not block. 719 */ 720 static __inline vm_page_t 721 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 722 { 723 vm_page_t m; 724 725 m = _vm_page_list_find( 726 PQ_FREE, 727 (pindex + object->pg_color) & PQ_L2_MASK, 728 prefer_zero 729 ); 730 return(m); 731 } 732 733 /* 734 * vm_page_alloc() 735 * 736 * Allocate and return a memory cell associated with this VM object/offset 737 * pair. 738 * 739 * page_req classes: 740 * 741 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 742 * VM_ALLOC_QUICK like normal but cannot use cache 743 * VM_ALLOC_SYSTEM greater free drain 744 * VM_ALLOC_INTERRUPT allow free list to be completely drained 745 * VM_ALLOC_ZERO advisory request for pre-zero'd page 746 * 747 * The object must be locked. 748 * This routine may not block. 749 * The returned page will be marked PG_BUSY 750 * 751 * Additional special handling is required when called from an interrupt 752 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 753 * in this case. 754 */ 755 vm_page_t 756 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 757 { 758 vm_page_t m = NULL; 759 760 lwkt_gettoken(&vm_token); 761 762 KKASSERT(object != NULL); 763 KASSERT(!vm_page_lookup(object, pindex), 764 ("vm_page_alloc: page already allocated")); 765 KKASSERT(page_req & 766 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 767 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 768 769 /* 770 * Certain system threads (pageout daemon, buf_daemon's) are 771 * allowed to eat deeper into the free page list. 772 */ 773 if (curthread->td_flags & TDF_SYSTHREAD) 774 page_req |= VM_ALLOC_SYSTEM; 775 776 loop: 777 if (vmstats.v_free_count > vmstats.v_free_reserved || 778 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 779 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 780 vmstats.v_free_count > vmstats.v_interrupt_free_min) 781 ) { 782 /* 783 * The free queue has sufficient free pages to take one out. 784 */ 785 if (page_req & VM_ALLOC_ZERO) 786 m = vm_page_select_free(object, pindex, TRUE); 787 else 788 m = vm_page_select_free(object, pindex, FALSE); 789 } else if (page_req & VM_ALLOC_NORMAL) { 790 /* 791 * Allocatable from the cache (non-interrupt only). On 792 * success, we must free the page and try again, thus 793 * ensuring that vmstats.v_*_free_min counters are replenished. 794 */ 795 #ifdef INVARIANTS 796 if (curthread->td_preempted) { 797 kprintf("vm_page_alloc(): warning, attempt to allocate" 798 " cache page from preempting interrupt\n"); 799 m = NULL; 800 } else { 801 m = vm_page_select_cache(object, pindex); 802 } 803 #else 804 m = vm_page_select_cache(object, pindex); 805 #endif 806 /* 807 * On success move the page into the free queue and loop. 808 */ 809 if (m != NULL) { 810 KASSERT(m->dirty == 0, 811 ("Found dirty cache page %p", m)); 812 vm_page_busy(m); 813 vm_page_protect(m, VM_PROT_NONE); 814 vm_page_free(m); 815 goto loop; 816 } 817 818 /* 819 * On failure return NULL 820 */ 821 lwkt_reltoken(&vm_token); 822 #if defined(DIAGNOSTIC) 823 if (vmstats.v_cache_count > 0) 824 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 825 #endif 826 vm_pageout_deficit++; 827 pagedaemon_wakeup(); 828 return (NULL); 829 } else { 830 /* 831 * No pages available, wakeup the pageout daemon and give up. 832 */ 833 lwkt_reltoken(&vm_token); 834 vm_pageout_deficit++; 835 pagedaemon_wakeup(); 836 return (NULL); 837 } 838 839 /* 840 * Good page found. The page has not yet been busied. We are in 841 * a critical section. 842 */ 843 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 844 KASSERT(m->dirty == 0, 845 ("vm_page_alloc: free/cache page %p was dirty", m)); 846 847 /* 848 * Remove from free queue 849 */ 850 vm_page_unqueue_nowakeup(m); 851 852 /* 853 * Initialize structure. Only the PG_ZERO flag is inherited. Set 854 * the page PG_BUSY 855 */ 856 if (m->flags & PG_ZERO) { 857 vm_page_zero_count--; 858 m->flags = PG_ZERO | PG_BUSY; 859 } else { 860 m->flags = PG_BUSY; 861 } 862 m->wire_count = 0; 863 m->hold_count = 0; 864 m->act_count = 0; 865 m->busy = 0; 866 m->valid = 0; 867 868 /* 869 * vm_page_insert() is safe while holding vm_token. Note also that 870 * inserting a page here does not insert it into the pmap (which 871 * could cause us to block allocating memory). We cannot block 872 * anywhere. 873 */ 874 vm_page_insert(m, object, pindex); 875 876 /* 877 * Don't wakeup too often - wakeup the pageout daemon when 878 * we would be nearly out of memory. 879 */ 880 pagedaemon_wakeup(); 881 882 lwkt_reltoken(&vm_token); 883 884 /* 885 * A PG_BUSY page is returned. 886 */ 887 return (m); 888 } 889 890 /* 891 * Wait for sufficient free memory for nominal heavy memory use kernel 892 * operations. 893 */ 894 void 895 vm_wait_nominal(void) 896 { 897 while (vm_page_count_min(0)) 898 vm_wait(0); 899 } 900 901 /* 902 * Test if vm_wait_nominal() would block. 903 */ 904 int 905 vm_test_nominal(void) 906 { 907 if (vm_page_count_min(0)) 908 return(1); 909 return(0); 910 } 911 912 /* 913 * Block until free pages are available for allocation, called in various 914 * places before memory allocations. 915 * 916 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 917 * more generous then that. 918 */ 919 void 920 vm_wait(int timo) 921 { 922 /* 923 * never wait forever 924 */ 925 if (timo == 0) 926 timo = hz; 927 lwkt_gettoken(&vm_token); 928 929 if (curthread == pagethread) { 930 /* 931 * The pageout daemon itself needs pages, this is bad. 932 */ 933 if (vm_page_count_min(0)) { 934 vm_pageout_pages_needed = 1; 935 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 936 } 937 } else { 938 /* 939 * Wakeup the pageout daemon if necessary and wait. 940 */ 941 if (vm_page_count_target()) { 942 if (vm_pages_needed == 0) { 943 vm_pages_needed = 1; 944 wakeup(&vm_pages_needed); 945 } 946 ++vm_pages_waiting; /* SMP race ok */ 947 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 948 } 949 } 950 lwkt_reltoken(&vm_token); 951 } 952 953 /* 954 * Block until free pages are available for allocation 955 * 956 * Called only from vm_fault so that processes page faulting can be 957 * easily tracked. 958 */ 959 void 960 vm_waitpfault(void) 961 { 962 /* 963 * Wakeup the pageout daemon if necessary and wait. 964 */ 965 if (vm_page_count_target()) { 966 lwkt_gettoken(&vm_token); 967 if (vm_page_count_target()) { 968 if (vm_pages_needed == 0) { 969 vm_pages_needed = 1; 970 wakeup(&vm_pages_needed); 971 } 972 ++vm_pages_waiting; /* SMP race ok */ 973 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 974 } 975 lwkt_reltoken(&vm_token); 976 } 977 } 978 979 /* 980 * Put the specified page on the active list (if appropriate). Ensure 981 * that act_count is at least ACT_INIT but do not otherwise mess with it. 982 * 983 * The page queues must be locked. 984 * This routine may not block. 985 */ 986 void 987 vm_page_activate(vm_page_t m) 988 { 989 lwkt_gettoken(&vm_token); 990 if (m->queue != PQ_ACTIVE) { 991 if ((m->queue - m->pc) == PQ_CACHE) 992 mycpu->gd_cnt.v_reactivated++; 993 994 vm_page_unqueue(m); 995 996 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 997 m->queue = PQ_ACTIVE; 998 vm_page_queues[PQ_ACTIVE].lcnt++; 999 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 1000 m, pageq); 1001 if (m->act_count < ACT_INIT) 1002 m->act_count = ACT_INIT; 1003 vmstats.v_active_count++; 1004 } 1005 } else { 1006 if (m->act_count < ACT_INIT) 1007 m->act_count = ACT_INIT; 1008 } 1009 lwkt_reltoken(&vm_token); 1010 } 1011 1012 /* 1013 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1014 * routine is called when a page has been added to the cache or free 1015 * queues. 1016 * 1017 * This routine may not block. 1018 * This routine must be called at splvm() 1019 */ 1020 static __inline void 1021 vm_page_free_wakeup(void) 1022 { 1023 /* 1024 * If the pageout daemon itself needs pages, then tell it that 1025 * there are some free. 1026 */ 1027 if (vm_pageout_pages_needed && 1028 vmstats.v_cache_count + vmstats.v_free_count >= 1029 vmstats.v_pageout_free_min 1030 ) { 1031 wakeup(&vm_pageout_pages_needed); 1032 vm_pageout_pages_needed = 0; 1033 } 1034 1035 /* 1036 * Wakeup processes that are waiting on memory. 1037 * 1038 * NOTE: vm_paging_target() is the pageout daemon's target, while 1039 * vm_page_count_target() is somewhere inbetween. We want 1040 * to wake processes up prior to the pageout daemon reaching 1041 * its target to provide some hysteresis. 1042 */ 1043 if (vm_pages_waiting) { 1044 if (!vm_page_count_target()) { 1045 /* 1046 * Plenty of pages are free, wakeup everyone. 1047 */ 1048 vm_pages_waiting = 0; 1049 wakeup(&vmstats.v_free_count); 1050 ++mycpu->gd_cnt.v_ppwakeups; 1051 } else if (!vm_page_count_min(0)) { 1052 /* 1053 * Some pages are free, wakeup someone. 1054 */ 1055 int wcount = vm_pages_waiting; 1056 if (wcount > 0) 1057 --wcount; 1058 vm_pages_waiting = wcount; 1059 wakeup_one(&vmstats.v_free_count); 1060 ++mycpu->gd_cnt.v_ppwakeups; 1061 } 1062 } 1063 } 1064 1065 /* 1066 * vm_page_free_toq: 1067 * 1068 * Returns the given page to the PQ_FREE list, disassociating it with 1069 * any VM object. 1070 * 1071 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1072 * return (the page will have been freed). No particular spl is required 1073 * on entry. 1074 * 1075 * This routine may not block. 1076 */ 1077 void 1078 vm_page_free_toq(vm_page_t m) 1079 { 1080 struct vpgqueues *pq; 1081 1082 lwkt_gettoken(&vm_token); 1083 mycpu->gd_cnt.v_tfree++; 1084 1085 KKASSERT((m->flags & PG_MAPPED) == 0); 1086 1087 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1088 kprintf( 1089 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1090 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1091 m->hold_count); 1092 if ((m->queue - m->pc) == PQ_FREE) 1093 panic("vm_page_free: freeing free page"); 1094 else 1095 panic("vm_page_free: freeing busy page"); 1096 } 1097 1098 /* 1099 * unqueue, then remove page. Note that we cannot destroy 1100 * the page here because we do not want to call the pager's 1101 * callback routine until after we've put the page on the 1102 * appropriate free queue. 1103 */ 1104 vm_page_unqueue_nowakeup(m); 1105 vm_page_remove(m); 1106 1107 /* 1108 * No further management of fictitious pages occurs beyond object 1109 * and queue removal. 1110 */ 1111 if ((m->flags & PG_FICTITIOUS) != 0) { 1112 vm_page_wakeup(m); 1113 lwkt_reltoken(&vm_token); 1114 return; 1115 } 1116 1117 m->valid = 0; 1118 vm_page_undirty(m); 1119 1120 if (m->wire_count != 0) { 1121 if (m->wire_count > 1) { 1122 panic( 1123 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1124 m->wire_count, (long)m->pindex); 1125 } 1126 panic("vm_page_free: freeing wired page"); 1127 } 1128 1129 /* 1130 * Clear the UNMANAGED flag when freeing an unmanaged page. 1131 */ 1132 if (m->flags & PG_UNMANAGED) { 1133 m->flags &= ~PG_UNMANAGED; 1134 } 1135 1136 if (m->hold_count != 0) { 1137 m->flags &= ~PG_ZERO; 1138 m->queue = PQ_HOLD; 1139 } else { 1140 m->queue = PQ_FREE + m->pc; 1141 } 1142 pq = &vm_page_queues[m->queue]; 1143 pq->lcnt++; 1144 ++(*pq->cnt); 1145 1146 /* 1147 * Put zero'd pages on the end ( where we look for zero'd pages 1148 * first ) and non-zerod pages at the head. 1149 */ 1150 if (m->flags & PG_ZERO) { 1151 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1152 ++vm_page_zero_count; 1153 } else { 1154 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1155 } 1156 vm_page_wakeup(m); 1157 vm_page_free_wakeup(); 1158 lwkt_reltoken(&vm_token); 1159 } 1160 1161 /* 1162 * vm_page_free_fromq_fast() 1163 * 1164 * Remove a non-zero page from one of the free queues; the page is removed for 1165 * zeroing, so do not issue a wakeup. 1166 * 1167 * MPUNSAFE 1168 */ 1169 vm_page_t 1170 vm_page_free_fromq_fast(void) 1171 { 1172 static int qi; 1173 vm_page_t m; 1174 int i; 1175 1176 lwkt_gettoken(&vm_token); 1177 for (i = 0; i < PQ_L2_SIZE; ++i) { 1178 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1179 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1180 if (m && (m->flags & PG_ZERO) == 0) { 1181 KKASSERT(m->busy == 0 && (m->flags & PG_BUSY) == 0); 1182 vm_page_unqueue_nowakeup(m); 1183 vm_page_busy(m); 1184 break; 1185 } 1186 m = NULL; 1187 } 1188 lwkt_reltoken(&vm_token); 1189 return (m); 1190 } 1191 1192 /* 1193 * vm_page_unmanage() 1194 * 1195 * Prevent PV management from being done on the page. The page is 1196 * removed from the paging queues as if it were wired, and as a 1197 * consequence of no longer being managed the pageout daemon will not 1198 * touch it (since there is no way to locate the pte mappings for the 1199 * page). madvise() calls that mess with the pmap will also no longer 1200 * operate on the page. 1201 * 1202 * Beyond that the page is still reasonably 'normal'. Freeing the page 1203 * will clear the flag. 1204 * 1205 * This routine is used by OBJT_PHYS objects - objects using unswappable 1206 * physical memory as backing store rather then swap-backed memory and 1207 * will eventually be extended to support 4MB unmanaged physical 1208 * mappings. 1209 * 1210 * Must be called with a critical section held. 1211 * Must be called with vm_token held. 1212 */ 1213 void 1214 vm_page_unmanage(vm_page_t m) 1215 { 1216 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1217 if ((m->flags & PG_UNMANAGED) == 0) { 1218 if (m->wire_count == 0) 1219 vm_page_unqueue(m); 1220 } 1221 vm_page_flag_set(m, PG_UNMANAGED); 1222 } 1223 1224 /* 1225 * Mark this page as wired down by yet another map, removing it from 1226 * paging queues as necessary. 1227 * 1228 * The page queues must be locked. 1229 * This routine may not block. 1230 */ 1231 void 1232 vm_page_wire(vm_page_t m) 1233 { 1234 /* 1235 * Only bump the wire statistics if the page is not already wired, 1236 * and only unqueue the page if it is on some queue (if it is unmanaged 1237 * it is already off the queues). Don't do anything with fictitious 1238 * pages because they are always wired. 1239 */ 1240 lwkt_gettoken(&vm_token); 1241 if ((m->flags & PG_FICTITIOUS) == 0) { 1242 if (m->wire_count == 0) { 1243 if ((m->flags & PG_UNMANAGED) == 0) 1244 vm_page_unqueue(m); 1245 vmstats.v_wire_count++; 1246 } 1247 m->wire_count++; 1248 KASSERT(m->wire_count != 0, 1249 ("vm_page_wire: wire_count overflow m=%p", m)); 1250 } 1251 lwkt_reltoken(&vm_token); 1252 } 1253 1254 /* 1255 * Release one wiring of this page, potentially enabling it to be paged again. 1256 * 1257 * Many pages placed on the inactive queue should actually go 1258 * into the cache, but it is difficult to figure out which. What 1259 * we do instead, if the inactive target is well met, is to put 1260 * clean pages at the head of the inactive queue instead of the tail. 1261 * This will cause them to be moved to the cache more quickly and 1262 * if not actively re-referenced, freed more quickly. If we just 1263 * stick these pages at the end of the inactive queue, heavy filesystem 1264 * meta-data accesses can cause an unnecessary paging load on memory bound 1265 * processes. This optimization causes one-time-use metadata to be 1266 * reused more quickly. 1267 * 1268 * BUT, if we are in a low-memory situation we have no choice but to 1269 * put clean pages on the cache queue. 1270 * 1271 * A number of routines use vm_page_unwire() to guarantee that the page 1272 * will go into either the inactive or active queues, and will NEVER 1273 * be placed in the cache - for example, just after dirtying a page. 1274 * dirty pages in the cache are not allowed. 1275 * 1276 * The page queues must be locked. 1277 * This routine may not block. 1278 */ 1279 void 1280 vm_page_unwire(vm_page_t m, int activate) 1281 { 1282 lwkt_gettoken(&vm_token); 1283 if (m->flags & PG_FICTITIOUS) { 1284 /* do nothing */ 1285 } else if (m->wire_count <= 0) { 1286 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1287 } else { 1288 if (--m->wire_count == 0) { 1289 --vmstats.v_wire_count; 1290 if (m->flags & PG_UNMANAGED) { 1291 ; 1292 } else if (activate) { 1293 TAILQ_INSERT_TAIL( 1294 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1295 m->queue = PQ_ACTIVE; 1296 vm_page_queues[PQ_ACTIVE].lcnt++; 1297 vmstats.v_active_count++; 1298 } else { 1299 vm_page_flag_clear(m, PG_WINATCFLS); 1300 TAILQ_INSERT_TAIL( 1301 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1302 m->queue = PQ_INACTIVE; 1303 vm_page_queues[PQ_INACTIVE].lcnt++; 1304 vmstats.v_inactive_count++; 1305 ++vm_swapcache_inactive_heuristic; 1306 } 1307 } 1308 } 1309 lwkt_reltoken(&vm_token); 1310 } 1311 1312 1313 /* 1314 * Move the specified page to the inactive queue. If the page has 1315 * any associated swap, the swap is deallocated. 1316 * 1317 * Normally athead is 0 resulting in LRU operation. athead is set 1318 * to 1 if we want this page to be 'as if it were placed in the cache', 1319 * except without unmapping it from the process address space. 1320 * 1321 * This routine may not block. 1322 * The caller must hold vm_token. 1323 */ 1324 static __inline void 1325 _vm_page_deactivate(vm_page_t m, int athead) 1326 { 1327 /* 1328 * Ignore if already inactive. 1329 */ 1330 if (m->queue == PQ_INACTIVE) 1331 return; 1332 1333 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1334 if ((m->queue - m->pc) == PQ_CACHE) 1335 mycpu->gd_cnt.v_reactivated++; 1336 vm_page_flag_clear(m, PG_WINATCFLS); 1337 vm_page_unqueue(m); 1338 if (athead) { 1339 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1340 m, pageq); 1341 } else { 1342 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1343 m, pageq); 1344 ++vm_swapcache_inactive_heuristic; 1345 } 1346 m->queue = PQ_INACTIVE; 1347 vm_page_queues[PQ_INACTIVE].lcnt++; 1348 vmstats.v_inactive_count++; 1349 } 1350 } 1351 1352 /* 1353 * Attempt to deactivate a page. 1354 * 1355 * No requirements. 1356 */ 1357 void 1358 vm_page_deactivate(vm_page_t m) 1359 { 1360 lwkt_gettoken(&vm_token); 1361 _vm_page_deactivate(m, 0); 1362 lwkt_reltoken(&vm_token); 1363 } 1364 1365 /* 1366 * Attempt to move a page to PQ_CACHE. 1367 * Returns 0 on failure, 1 on success 1368 * 1369 * No requirements. 1370 */ 1371 int 1372 vm_page_try_to_cache(vm_page_t m) 1373 { 1374 lwkt_gettoken(&vm_token); 1375 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1376 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1377 lwkt_reltoken(&vm_token); 1378 return(0); 1379 } 1380 vm_page_busy(m); 1381 vm_page_test_dirty(m); 1382 if (m->dirty) { 1383 lwkt_reltoken(&vm_token); 1384 return(0); 1385 } 1386 vm_page_cache(m); 1387 lwkt_reltoken(&vm_token); 1388 return(1); 1389 } 1390 1391 /* 1392 * Attempt to free the page. If we cannot free it, we do nothing. 1393 * 1 is returned on success, 0 on failure. 1394 * 1395 * No requirements. 1396 */ 1397 int 1398 vm_page_try_to_free(vm_page_t m) 1399 { 1400 lwkt_gettoken(&vm_token); 1401 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1402 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1403 lwkt_reltoken(&vm_token); 1404 return(0); 1405 } 1406 vm_page_test_dirty(m); 1407 if (m->dirty) { 1408 lwkt_reltoken(&vm_token); 1409 return(0); 1410 } 1411 vm_page_busy(m); 1412 vm_page_protect(m, VM_PROT_NONE); 1413 vm_page_free(m); 1414 lwkt_reltoken(&vm_token); 1415 return(1); 1416 } 1417 1418 /* 1419 * vm_page_cache 1420 * 1421 * Put the specified page onto the page cache queue (if appropriate). 1422 * 1423 * The caller must hold vm_token. 1424 * This routine may not block. 1425 * The page must be busy, and this routine will release the busy and 1426 * possibly even free the page. 1427 */ 1428 void 1429 vm_page_cache(vm_page_t m) 1430 { 1431 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1432 1433 if ((m->flags & PG_UNMANAGED) || m->busy || 1434 m->wire_count || m->hold_count) { 1435 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1436 vm_page_wakeup(m); 1437 return; 1438 } 1439 1440 /* 1441 * Already in the cache (and thus not mapped) 1442 */ 1443 if ((m->queue - m->pc) == PQ_CACHE) { 1444 KKASSERT((m->flags & PG_MAPPED) == 0); 1445 vm_page_wakeup(m); 1446 return; 1447 } 1448 1449 /* 1450 * Caller is required to test m->dirty, but note that the act of 1451 * removing the page from its maps can cause it to become dirty 1452 * on an SMP system due to another cpu running in usermode. 1453 */ 1454 if (m->dirty) { 1455 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1456 (long)m->pindex); 1457 } 1458 1459 /* 1460 * Remove all pmaps and indicate that the page is not 1461 * writeable or mapped. Our vm_page_protect() call may 1462 * have blocked (especially w/ VM_PROT_NONE), so recheck 1463 * everything. 1464 */ 1465 vm_page_protect(m, VM_PROT_NONE); 1466 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy || 1467 m->wire_count || m->hold_count) { 1468 vm_page_wakeup(m); 1469 } else if (m->dirty) { 1470 vm_page_deactivate(m); 1471 vm_page_wakeup(m); 1472 } else { 1473 vm_page_unqueue_nowakeup(m); 1474 m->queue = PQ_CACHE + m->pc; 1475 vm_page_queues[m->queue].lcnt++; 1476 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1477 vmstats.v_cache_count++; 1478 vm_page_wakeup(m); 1479 vm_page_free_wakeup(); 1480 } 1481 } 1482 1483 /* 1484 * vm_page_dontneed() 1485 * 1486 * Cache, deactivate, or do nothing as appropriate. This routine 1487 * is typically used by madvise() MADV_DONTNEED. 1488 * 1489 * Generally speaking we want to move the page into the cache so 1490 * it gets reused quickly. However, this can result in a silly syndrome 1491 * due to the page recycling too quickly. Small objects will not be 1492 * fully cached. On the otherhand, if we move the page to the inactive 1493 * queue we wind up with a problem whereby very large objects 1494 * unnecessarily blow away our inactive and cache queues. 1495 * 1496 * The solution is to move the pages based on a fixed weighting. We 1497 * either leave them alone, deactivate them, or move them to the cache, 1498 * where moving them to the cache has the highest weighting. 1499 * By forcing some pages into other queues we eventually force the 1500 * system to balance the queues, potentially recovering other unrelated 1501 * space from active. The idea is to not force this to happen too 1502 * often. 1503 * 1504 * No requirements. 1505 */ 1506 void 1507 vm_page_dontneed(vm_page_t m) 1508 { 1509 static int dnweight; 1510 int dnw; 1511 int head; 1512 1513 dnw = ++dnweight; 1514 1515 /* 1516 * occassionally leave the page alone 1517 */ 1518 lwkt_gettoken(&vm_token); 1519 if ((dnw & 0x01F0) == 0 || 1520 m->queue == PQ_INACTIVE || 1521 m->queue - m->pc == PQ_CACHE 1522 ) { 1523 if (m->act_count >= ACT_INIT) 1524 --m->act_count; 1525 lwkt_reltoken(&vm_token); 1526 return; 1527 } 1528 1529 if (m->dirty == 0) 1530 vm_page_test_dirty(m); 1531 1532 if (m->dirty || (dnw & 0x0070) == 0) { 1533 /* 1534 * Deactivate the page 3 times out of 32. 1535 */ 1536 head = 0; 1537 } else { 1538 /* 1539 * Cache the page 28 times out of every 32. Note that 1540 * the page is deactivated instead of cached, but placed 1541 * at the head of the queue instead of the tail. 1542 */ 1543 head = 1; 1544 } 1545 _vm_page_deactivate(m, head); 1546 lwkt_reltoken(&vm_token); 1547 } 1548 1549 /* 1550 * Grab a page, blocking if it is busy and allocating a page if necessary. 1551 * A busy page is returned or NULL. 1552 * 1553 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1554 * If VM_ALLOC_RETRY is not specified 1555 * 1556 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1557 * always returned if we had blocked. 1558 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1559 * This routine may not be called from an interrupt. 1560 * The returned page may not be entirely valid. 1561 * 1562 * This routine may be called from mainline code without spl protection and 1563 * be guarenteed a busied page associated with the object at the specified 1564 * index. 1565 * 1566 * No requirements. 1567 */ 1568 vm_page_t 1569 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1570 { 1571 vm_page_t m; 1572 int generation; 1573 1574 KKASSERT(allocflags & 1575 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1576 lwkt_gettoken(&vm_token); 1577 retrylookup: 1578 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1579 if (m->busy || (m->flags & PG_BUSY)) { 1580 generation = object->generation; 1581 1582 while ((object->generation == generation) && 1583 (m->busy || (m->flags & PG_BUSY))) { 1584 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1585 tsleep(m, 0, "pgrbwt", 0); 1586 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1587 m = NULL; 1588 goto done; 1589 } 1590 } 1591 goto retrylookup; 1592 } else { 1593 vm_page_busy(m); 1594 goto done; 1595 } 1596 } 1597 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1598 if (m == NULL) { 1599 vm_wait(0); 1600 if ((allocflags & VM_ALLOC_RETRY) == 0) 1601 goto done; 1602 goto retrylookup; 1603 } 1604 done: 1605 lwkt_reltoken(&vm_token); 1606 return(m); 1607 } 1608 1609 /* 1610 * Mapping function for valid bits or for dirty bits in 1611 * a page. May not block. 1612 * 1613 * Inputs are required to range within a page. 1614 * 1615 * No requirements. 1616 * Non blocking. 1617 */ 1618 int 1619 vm_page_bits(int base, int size) 1620 { 1621 int first_bit; 1622 int last_bit; 1623 1624 KASSERT( 1625 base + size <= PAGE_SIZE, 1626 ("vm_page_bits: illegal base/size %d/%d", base, size) 1627 ); 1628 1629 if (size == 0) /* handle degenerate case */ 1630 return(0); 1631 1632 first_bit = base >> DEV_BSHIFT; 1633 last_bit = (base + size - 1) >> DEV_BSHIFT; 1634 1635 return ((2 << last_bit) - (1 << first_bit)); 1636 } 1637 1638 /* 1639 * Sets portions of a page valid and clean. The arguments are expected 1640 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1641 * of any partial chunks touched by the range. The invalid portion of 1642 * such chunks will be zero'd. 1643 * 1644 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1645 * align base to DEV_BSIZE so as not to mark clean a partially 1646 * truncated device block. Otherwise the dirty page status might be 1647 * lost. 1648 * 1649 * This routine may not block. 1650 * 1651 * (base + size) must be less then or equal to PAGE_SIZE. 1652 */ 1653 static void 1654 _vm_page_zero_valid(vm_page_t m, int base, int size) 1655 { 1656 int frag; 1657 int endoff; 1658 1659 if (size == 0) /* handle degenerate case */ 1660 return; 1661 1662 /* 1663 * If the base is not DEV_BSIZE aligned and the valid 1664 * bit is clear, we have to zero out a portion of the 1665 * first block. 1666 */ 1667 1668 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1669 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1670 ) { 1671 pmap_zero_page_area( 1672 VM_PAGE_TO_PHYS(m), 1673 frag, 1674 base - frag 1675 ); 1676 } 1677 1678 /* 1679 * If the ending offset is not DEV_BSIZE aligned and the 1680 * valid bit is clear, we have to zero out a portion of 1681 * the last block. 1682 */ 1683 1684 endoff = base + size; 1685 1686 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1687 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1688 ) { 1689 pmap_zero_page_area( 1690 VM_PAGE_TO_PHYS(m), 1691 endoff, 1692 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1693 ); 1694 } 1695 } 1696 1697 /* 1698 * Set valid, clear dirty bits. If validating the entire 1699 * page we can safely clear the pmap modify bit. We also 1700 * use this opportunity to clear the PG_NOSYNC flag. If a process 1701 * takes a write fault on a MAP_NOSYNC memory area the flag will 1702 * be set again. 1703 * 1704 * We set valid bits inclusive of any overlap, but we can only 1705 * clear dirty bits for DEV_BSIZE chunks that are fully within 1706 * the range. 1707 * 1708 * Page must be busied? 1709 * No other requirements. 1710 */ 1711 void 1712 vm_page_set_valid(vm_page_t m, int base, int size) 1713 { 1714 _vm_page_zero_valid(m, base, size); 1715 m->valid |= vm_page_bits(base, size); 1716 } 1717 1718 1719 /* 1720 * Set valid bits and clear dirty bits. 1721 * 1722 * NOTE: This function does not clear the pmap modified bit. 1723 * Also note that e.g. NFS may use a byte-granular base 1724 * and size. 1725 * 1726 * WARNING: Page must be busied? But vfs_clean_one_page() will call 1727 * this without necessarily busying the page (via bdwrite()). 1728 * So for now vm_token must also be held. 1729 * 1730 * No other requirements. 1731 */ 1732 void 1733 vm_page_set_validclean(vm_page_t m, int base, int size) 1734 { 1735 int pagebits; 1736 1737 _vm_page_zero_valid(m, base, size); 1738 pagebits = vm_page_bits(base, size); 1739 m->valid |= pagebits; 1740 m->dirty &= ~pagebits; 1741 if (base == 0 && size == PAGE_SIZE) { 1742 /*pmap_clear_modify(m);*/ 1743 vm_page_flag_clear(m, PG_NOSYNC); 1744 } 1745 } 1746 1747 /* 1748 * Set valid & dirty. Used by buwrite() 1749 * 1750 * WARNING: Page must be busied? But vfs_dirty_one_page() will 1751 * call this function in buwrite() so for now vm_token must 1752 * be held. 1753 * 1754 * No other requirements. 1755 */ 1756 void 1757 vm_page_set_validdirty(vm_page_t m, int base, int size) 1758 { 1759 int pagebits; 1760 1761 pagebits = vm_page_bits(base, size); 1762 m->valid |= pagebits; 1763 m->dirty |= pagebits; 1764 if (m->object) 1765 vm_object_set_writeable_dirty(m->object); 1766 } 1767 1768 /* 1769 * Clear dirty bits. 1770 * 1771 * NOTE: This function does not clear the pmap modified bit. 1772 * Also note that e.g. NFS may use a byte-granular base 1773 * and size. 1774 * 1775 * Page must be busied? 1776 * No other requirements. 1777 */ 1778 void 1779 vm_page_clear_dirty(vm_page_t m, int base, int size) 1780 { 1781 m->dirty &= ~vm_page_bits(base, size); 1782 if (base == 0 && size == PAGE_SIZE) { 1783 /*pmap_clear_modify(m);*/ 1784 vm_page_flag_clear(m, PG_NOSYNC); 1785 } 1786 } 1787 1788 /* 1789 * Make the page all-dirty. 1790 * 1791 * Also make sure the related object and vnode reflect the fact that the 1792 * object may now contain a dirty page. 1793 * 1794 * Page must be busied? 1795 * No other requirements. 1796 */ 1797 void 1798 vm_page_dirty(vm_page_t m) 1799 { 1800 #ifdef INVARIANTS 1801 int pqtype = m->queue - m->pc; 1802 #endif 1803 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1804 ("vm_page_dirty: page in free/cache queue!")); 1805 if (m->dirty != VM_PAGE_BITS_ALL) { 1806 m->dirty = VM_PAGE_BITS_ALL; 1807 if (m->object) 1808 vm_object_set_writeable_dirty(m->object); 1809 } 1810 } 1811 1812 /* 1813 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1814 * valid and dirty bits for the effected areas are cleared. 1815 * 1816 * Page must be busied? 1817 * Does not block. 1818 * No other requirements. 1819 */ 1820 void 1821 vm_page_set_invalid(vm_page_t m, int base, int size) 1822 { 1823 int bits; 1824 1825 bits = vm_page_bits(base, size); 1826 m->valid &= ~bits; 1827 m->dirty &= ~bits; 1828 m->object->generation++; 1829 } 1830 1831 /* 1832 * The kernel assumes that the invalid portions of a page contain 1833 * garbage, but such pages can be mapped into memory by user code. 1834 * When this occurs, we must zero out the non-valid portions of the 1835 * page so user code sees what it expects. 1836 * 1837 * Pages are most often semi-valid when the end of a file is mapped 1838 * into memory and the file's size is not page aligned. 1839 * 1840 * Page must be busied? 1841 * No other requirements. 1842 */ 1843 void 1844 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1845 { 1846 int b; 1847 int i; 1848 1849 /* 1850 * Scan the valid bits looking for invalid sections that 1851 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1852 * valid bit may be set ) have already been zerod by 1853 * vm_page_set_validclean(). 1854 */ 1855 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1856 if (i == (PAGE_SIZE / DEV_BSIZE) || 1857 (m->valid & (1 << i)) 1858 ) { 1859 if (i > b) { 1860 pmap_zero_page_area( 1861 VM_PAGE_TO_PHYS(m), 1862 b << DEV_BSHIFT, 1863 (i - b) << DEV_BSHIFT 1864 ); 1865 } 1866 b = i + 1; 1867 } 1868 } 1869 1870 /* 1871 * setvalid is TRUE when we can safely set the zero'd areas 1872 * as being valid. We can do this if there are no cache consistency 1873 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1874 */ 1875 if (setvalid) 1876 m->valid = VM_PAGE_BITS_ALL; 1877 } 1878 1879 /* 1880 * Is a (partial) page valid? Note that the case where size == 0 1881 * will return FALSE in the degenerate case where the page is entirely 1882 * invalid, and TRUE otherwise. 1883 * 1884 * Does not block. 1885 * No other requirements. 1886 */ 1887 int 1888 vm_page_is_valid(vm_page_t m, int base, int size) 1889 { 1890 int bits = vm_page_bits(base, size); 1891 1892 if (m->valid && ((m->valid & bits) == bits)) 1893 return 1; 1894 else 1895 return 0; 1896 } 1897 1898 /* 1899 * update dirty bits from pmap/mmu. May not block. 1900 * 1901 * Caller must hold vm_token if non-blocking operation desired. 1902 * No other requirements. 1903 */ 1904 void 1905 vm_page_test_dirty(vm_page_t m) 1906 { 1907 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1908 vm_page_dirty(m); 1909 } 1910 } 1911 1912 /* 1913 * Register an action, associating it with its vm_page 1914 */ 1915 void 1916 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 1917 { 1918 struct vm_page_action_list *list; 1919 int hv; 1920 1921 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1922 list = &action_list[hv]; 1923 1924 lwkt_gettoken(&vm_token); 1925 vm_page_flag_set(action->m, PG_ACTIONLIST); 1926 action->event = event; 1927 LIST_INSERT_HEAD(list, action, entry); 1928 lwkt_reltoken(&vm_token); 1929 } 1930 1931 /* 1932 * Unregister an action, disassociating it from its related vm_page 1933 */ 1934 void 1935 vm_page_unregister_action(vm_page_action_t action) 1936 { 1937 struct vm_page_action_list *list; 1938 int hv; 1939 1940 lwkt_gettoken(&vm_token); 1941 if (action->event != VMEVENT_NONE) { 1942 action->event = VMEVENT_NONE; 1943 LIST_REMOVE(action, entry); 1944 1945 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1946 list = &action_list[hv]; 1947 if (LIST_EMPTY(list)) 1948 vm_page_flag_clear(action->m, PG_ACTIONLIST); 1949 } 1950 lwkt_reltoken(&vm_token); 1951 } 1952 1953 /* 1954 * Issue an event on a VM page. Corresponding action structures are 1955 * removed from the page's list and called. 1956 * 1957 * If the vm_page has no more pending action events we clear its 1958 * PG_ACTIONLIST flag. 1959 */ 1960 void 1961 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1962 { 1963 struct vm_page_action_list *list; 1964 struct vm_page_action *scan; 1965 struct vm_page_action *next; 1966 int hv; 1967 int all; 1968 1969 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 1970 list = &action_list[hv]; 1971 all = 1; 1972 1973 lwkt_gettoken(&vm_token); 1974 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 1975 if (scan->m == m) { 1976 if (scan->event == event) { 1977 scan->event = VMEVENT_NONE; 1978 LIST_REMOVE(scan, entry); 1979 scan->func(m, scan); 1980 /* XXX */ 1981 } else { 1982 all = 0; 1983 } 1984 } 1985 } 1986 if (all) 1987 vm_page_flag_clear(m, PG_ACTIONLIST); 1988 lwkt_reltoken(&vm_token); 1989 } 1990 1991 1992 #include "opt_ddb.h" 1993 #ifdef DDB 1994 #include <sys/kernel.h> 1995 1996 #include <ddb/ddb.h> 1997 1998 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1999 { 2000 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 2001 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 2002 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 2003 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 2004 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 2005 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 2006 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 2007 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 2008 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 2009 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 2010 } 2011 2012 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2013 { 2014 int i; 2015 db_printf("PQ_FREE:"); 2016 for(i=0;i<PQ_L2_SIZE;i++) { 2017 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 2018 } 2019 db_printf("\n"); 2020 2021 db_printf("PQ_CACHE:"); 2022 for(i=0;i<PQ_L2_SIZE;i++) { 2023 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 2024 } 2025 db_printf("\n"); 2026 2027 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2028 vm_page_queues[PQ_ACTIVE].lcnt, 2029 vm_page_queues[PQ_INACTIVE].lcnt); 2030 } 2031 #endif /* DDB */ 2032