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 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 36 */ 37 38 /* 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 /* 65 * Resident memory management module. The module manipulates 'VM pages'. 66 * A VM page is the core building block for memory management. 67 */ 68 69 #include <sys/param.h> 70 #include <sys/systm.h> 71 #include <sys/malloc.h> 72 #include <sys/proc.h> 73 #include <sys/vmmeter.h> 74 #include <sys/vnode.h> 75 #include <sys/kernel.h> 76 77 #include <vm/vm.h> 78 #include <vm/vm_param.h> 79 #include <sys/lock.h> 80 #include <vm/vm_kern.h> 81 #include <vm/pmap.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_object.h> 84 #include <vm/vm_page.h> 85 #include <vm/vm_pageout.h> 86 #include <vm/vm_pager.h> 87 #include <vm/vm_extern.h> 88 #include <vm/swap_pager.h> 89 90 #include <machine/md_var.h> 91 92 #include <vm/vm_page2.h> 93 94 #define VMACTION_HSIZE 256 95 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 96 97 static void vm_page_queue_init(void); 98 static void vm_page_free_wakeup(void); 99 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 100 static vm_page_t _vm_page_list_find2(int basequeue, int index); 101 102 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 103 104 LIST_HEAD(vm_page_action_list, vm_page_action); 105 struct vm_page_action_list action_list[VMACTION_HSIZE]; 106 static volatile int vm_pages_waiting; 107 108 109 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 110 vm_pindex_t, pindex); 111 112 static void 113 vm_page_queue_init(void) 114 { 115 int i; 116 117 for (i = 0; i < PQ_L2_SIZE; i++) 118 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 119 for (i = 0; i < PQ_L2_SIZE; i++) 120 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 121 122 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 123 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 124 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 125 /* PQ_NONE has no queue */ 126 127 for (i = 0; i < PQ_COUNT; i++) 128 TAILQ_INIT(&vm_page_queues[i].pl); 129 130 for (i = 0; i < VMACTION_HSIZE; i++) 131 LIST_INIT(&action_list[i]); 132 } 133 134 /* 135 * note: place in initialized data section? Is this necessary? 136 */ 137 long first_page = 0; 138 int vm_page_array_size = 0; 139 int vm_page_zero_count = 0; 140 vm_page_t vm_page_array = 0; 141 142 /* 143 * (low level boot) 144 * 145 * Sets the page size, perhaps based upon the memory size. 146 * Must be called before any use of page-size dependent functions. 147 */ 148 void 149 vm_set_page_size(void) 150 { 151 if (vmstats.v_page_size == 0) 152 vmstats.v_page_size = PAGE_SIZE; 153 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 154 panic("vm_set_page_size: page size not a power of two"); 155 } 156 157 /* 158 * (low level boot) 159 * 160 * Add a new page to the freelist for use by the system. New pages 161 * are added to both the head and tail of the associated free page 162 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 163 * requests pull 'recent' adds (higher physical addresses) first. 164 * 165 * Must be called in a critical section. 166 */ 167 vm_page_t 168 vm_add_new_page(vm_paddr_t pa) 169 { 170 struct vpgqueues *vpq; 171 vm_page_t m; 172 173 ++vmstats.v_page_count; 174 ++vmstats.v_free_count; 175 m = PHYS_TO_VM_PAGE(pa); 176 m->phys_addr = pa; 177 m->flags = 0; 178 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 179 m->queue = m->pc + PQ_FREE; 180 KKASSERT(m->dirty == 0); 181 182 vpq = &vm_page_queues[m->queue]; 183 if (vpq->flipflop) 184 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 185 else 186 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 187 vpq->flipflop = 1 - vpq->flipflop; 188 189 vm_page_queues[m->queue].lcnt++; 190 return (m); 191 } 192 193 /* 194 * (low level boot) 195 * 196 * Initializes the resident memory module. 197 * 198 * Preallocates memory for critical VM structures and arrays prior to 199 * kernel_map becoming available. 200 * 201 * Memory is allocated from (virtual2_start, virtual2_end) if available, 202 * otherwise memory is allocated from (virtual_start, virtual_end). 203 * 204 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 205 * large enough to hold vm_page_array & other structures for machines with 206 * large amounts of ram, so we want to use virtual2* when available. 207 */ 208 void 209 vm_page_startup(void) 210 { 211 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 212 vm_offset_t mapped; 213 vm_size_t npages; 214 vm_paddr_t page_range; 215 vm_paddr_t new_end; 216 int i; 217 vm_paddr_t pa; 218 int nblocks; 219 vm_paddr_t last_pa; 220 vm_paddr_t end; 221 vm_paddr_t biggestone, biggestsize; 222 vm_paddr_t total; 223 224 total = 0; 225 biggestsize = 0; 226 biggestone = 0; 227 nblocks = 0; 228 vaddr = round_page(vaddr); 229 230 for (i = 0; phys_avail[i + 1]; i += 2) { 231 phys_avail[i] = round_page64(phys_avail[i]); 232 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 233 } 234 235 for (i = 0; phys_avail[i + 1]; i += 2) { 236 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 237 238 if (size > biggestsize) { 239 biggestone = i; 240 biggestsize = size; 241 } 242 ++nblocks; 243 total += size; 244 } 245 246 end = phys_avail[biggestone+1]; 247 end = trunc_page(end); 248 249 /* 250 * Initialize the queue headers for the free queue, the active queue 251 * and the inactive queue. 252 */ 253 254 vm_page_queue_init(); 255 256 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */ 257 #if !defined(_KERNEL_VIRTUAL) 258 /* 259 * Allocate a bitmap to indicate that a random physical page 260 * needs to be included in a minidump. 261 * 262 * The amd64 port needs this to indicate which direct map pages 263 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 264 * 265 * However, i386 still needs this workspace internally within the 266 * minidump code. In theory, they are not needed on i386, but are 267 * included should the sf_buf code decide to use them. 268 */ 269 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 270 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 271 end -= vm_page_dump_size; 272 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 273 VM_PROT_READ | VM_PROT_WRITE); 274 bzero((void *)vm_page_dump, vm_page_dump_size); 275 #endif 276 277 /* 278 * Compute the number of pages of memory that will be available for 279 * use (taking into account the overhead of a page structure per 280 * page). 281 */ 282 first_page = phys_avail[0] / PAGE_SIZE; 283 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 284 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 285 286 /* 287 * Initialize the mem entry structures now, and put them in the free 288 * queue. 289 */ 290 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 291 mapped = pmap_map(&vaddr, new_end, end, 292 VM_PROT_READ | VM_PROT_WRITE); 293 vm_page_array = (vm_page_t)mapped; 294 295 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 296 /* 297 * since pmap_map on amd64 returns stuff out of a direct-map region, 298 * we have to manually add these pages to the minidump tracking so 299 * that they can be dumped, including the vm_page_array. 300 */ 301 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 302 dump_add_page(pa); 303 #endif 304 305 /* 306 * Clear all of the page structures 307 */ 308 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 309 vm_page_array_size = page_range; 310 311 /* 312 * Construct the free queue(s) in ascending order (by physical 313 * address) so that the first 16MB of physical memory is allocated 314 * last rather than first. On large-memory machines, this avoids 315 * the exhaustion of low physical memory before isa_dmainit has run. 316 */ 317 vmstats.v_page_count = 0; 318 vmstats.v_free_count = 0; 319 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 320 pa = phys_avail[i]; 321 if (i == biggestone) 322 last_pa = new_end; 323 else 324 last_pa = phys_avail[i + 1]; 325 while (pa < last_pa && npages-- > 0) { 326 vm_add_new_page(pa); 327 pa += PAGE_SIZE; 328 } 329 } 330 if (virtual2_start) 331 virtual2_start = vaddr; 332 else 333 virtual_start = vaddr; 334 } 335 336 /* 337 * Scan comparison function for Red-Black tree scans. An inclusive 338 * (start,end) is expected. Other fields are not used. 339 */ 340 int 341 rb_vm_page_scancmp(struct vm_page *p, void *data) 342 { 343 struct rb_vm_page_scan_info *info = data; 344 345 if (p->pindex < info->start_pindex) 346 return(-1); 347 if (p->pindex > info->end_pindex) 348 return(1); 349 return(0); 350 } 351 352 int 353 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 354 { 355 if (p1->pindex < p2->pindex) 356 return(-1); 357 if (p1->pindex > p2->pindex) 358 return(1); 359 return(0); 360 } 361 362 /* 363 * Holding a page keeps it from being reused. Other parts of the system 364 * can still disassociate the page from its current object and free it, or 365 * perform read or write I/O on it and/or otherwise manipulate the page, 366 * but if the page is held the VM system will leave the page and its data 367 * intact and not reuse the page for other purposes until the last hold 368 * reference is released. (see vm_page_wire() if you want to prevent the 369 * page from being disassociated from its object too). 370 * 371 * The caller must hold vm_token. 372 * 373 * The caller must still validate the contents of the page and, if necessary, 374 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 375 * before manipulating the page. 376 */ 377 void 378 vm_page_hold(vm_page_t m) 379 { 380 ASSERT_LWKT_TOKEN_HELD(&vm_token); 381 ++m->hold_count; 382 } 383 384 /* 385 * The opposite of vm_page_hold(). A page can be freed while being held, 386 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 387 * in this case to actually free it once the hold count drops to 0. 388 * 389 * The caller must hold vm_token if non-blocking operation is desired, 390 * but otherwise does not need to. 391 */ 392 void 393 vm_page_unhold(vm_page_t m) 394 { 395 lwkt_gettoken(&vm_token); 396 --m->hold_count; 397 KASSERT(m->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 398 if (m->hold_count == 0 && m->queue == PQ_HOLD) { 399 vm_page_busy(m); 400 vm_page_free_toq(m); 401 } 402 lwkt_reltoken(&vm_token); 403 } 404 405 /* 406 * Inserts the given vm_page into the object and object list. 407 * 408 * The pagetables are not updated but will presumably fault the page 409 * in if necessary, or if a kernel page the caller will at some point 410 * enter the page into the kernel's pmap. We are not allowed to block 411 * here so we *can't* do this anyway. 412 * 413 * This routine may not block. 414 * This routine must be called with the vm_token held. 415 * This routine must be called with the vm_object held. 416 * This routine must be called with a critical section held. 417 */ 418 void 419 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 420 { 421 ASSERT_LWKT_TOKEN_HELD(&vm_token); 422 if (m->object != NULL) 423 panic("vm_page_insert: already inserted"); 424 425 /* 426 * Record the object/offset pair in this page 427 */ 428 m->object = object; 429 m->pindex = pindex; 430 431 /* 432 * Insert it into the object. 433 */ 434 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 435 object->generation++; 436 437 /* 438 * show that the object has one more resident page. 439 */ 440 object->resident_page_count++; 441 442 /* 443 * Add the pv_list_cout of the page when its inserted in 444 * the object 445 */ 446 object->agg_pv_list_count = object->agg_pv_list_count + m->md.pv_list_count; 447 448 /* 449 * Since we are inserting a new and possibly dirty page, 450 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 451 */ 452 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 453 vm_object_set_writeable_dirty(object); 454 455 /* 456 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 457 */ 458 swap_pager_page_inserted(m); 459 } 460 461 /* 462 * Removes the given vm_page_t from the global (object,index) hash table 463 * and from the object's memq. 464 * 465 * The underlying pmap entry (if any) is NOT removed here. 466 * This routine may not block. 467 * 468 * The page must be BUSY and will remain BUSY on return. 469 * No other requirements. 470 * 471 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 472 * it busy. 473 */ 474 void 475 vm_page_remove(vm_page_t m) 476 { 477 vm_object_t object; 478 479 lwkt_gettoken(&vm_token); 480 if (m->object == NULL) { 481 lwkt_reltoken(&vm_token); 482 return; 483 } 484 485 if ((m->flags & PG_BUSY) == 0) 486 panic("vm_page_remove: page not busy"); 487 488 object = m->object; 489 490 vm_object_hold(object); 491 492 /* 493 * Remove the page from the object and update the object. 494 */ 495 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 496 object->resident_page_count--; 497 object->agg_pv_list_count = object->agg_pv_list_count - m->md.pv_list_count; 498 object->generation++; 499 m->object = NULL; 500 501 vm_object_drop(object); 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_object_hold(new_object); 554 vm_page_remove(m); 555 vm_page_insert(m, new_object, new_pindex); 556 if (m->queue - m->pc == PQ_CACHE) 557 vm_page_deactivate(m); 558 vm_page_dirty(m); 559 vm_page_wakeup(m); 560 vm_object_drop(new_object); 561 lwkt_reltoken(&vm_token); 562 } 563 564 /* 565 * vm_page_unqueue() without any wakeup. This routine is used when a page 566 * is being moved between queues or otherwise is to remain BUSYied by the 567 * caller. 568 * 569 * The caller must hold vm_token 570 * This routine may not block. 571 */ 572 void 573 vm_page_unqueue_nowakeup(vm_page_t m) 574 { 575 int queue = m->queue; 576 struct vpgqueues *pq; 577 578 ASSERT_LWKT_TOKEN_HELD(&vm_token); 579 if (queue != PQ_NONE) { 580 pq = &vm_page_queues[queue]; 581 m->queue = PQ_NONE; 582 TAILQ_REMOVE(&pq->pl, m, pageq); 583 (*pq->cnt)--; 584 pq->lcnt--; 585 } 586 } 587 588 /* 589 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 590 * if necessary. 591 * 592 * The caller must hold vm_token 593 * This routine may not block. 594 */ 595 void 596 vm_page_unqueue(vm_page_t m) 597 { 598 int queue = m->queue; 599 struct vpgqueues *pq; 600 601 ASSERT_LWKT_TOKEN_HELD(&vm_token); 602 if (queue != PQ_NONE) { 603 m->queue = PQ_NONE; 604 pq = &vm_page_queues[queue]; 605 TAILQ_REMOVE(&pq->pl, m, pageq); 606 (*pq->cnt)--; 607 pq->lcnt--; 608 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 609 pagedaemon_wakeup(); 610 } 611 } 612 613 /* 614 * vm_page_list_find() 615 * 616 * Find a page on the specified queue with color optimization. 617 * 618 * The page coloring optimization attempts to locate a page that does 619 * not overload other nearby pages in the object in the cpu's L1 or L2 620 * caches. We need this optimization because cpu caches tend to be 621 * physical caches, while object spaces tend to be virtual. 622 * 623 * Must be called with vm_token held. 624 * This routine may not block. 625 * 626 * Note that this routine is carefully inlined. A non-inlined version 627 * is available for outside callers but the only critical path is 628 * from within this source file. 629 */ 630 static __inline 631 vm_page_t 632 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 633 { 634 vm_page_t m; 635 636 if (prefer_zero) 637 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 638 else 639 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 640 if (m == NULL) 641 m = _vm_page_list_find2(basequeue, index); 642 return(m); 643 } 644 645 static vm_page_t 646 _vm_page_list_find2(int basequeue, int index) 647 { 648 int i; 649 vm_page_t m = NULL; 650 struct vpgqueues *pq; 651 652 pq = &vm_page_queues[basequeue]; 653 654 /* 655 * Note that for the first loop, index+i and index-i wind up at the 656 * same place. Even though this is not totally optimal, we've already 657 * blown it by missing the cache case so we do not care. 658 */ 659 660 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 661 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 662 break; 663 664 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 665 break; 666 } 667 return(m); 668 } 669 670 /* 671 * Must be called with vm_token held if the caller desired non-blocking 672 * operation and a stable result. 673 */ 674 vm_page_t 675 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 676 { 677 return(_vm_page_list_find(basequeue, index, prefer_zero)); 678 } 679 680 /* 681 * Find a page on the cache queue with color optimization. As pages 682 * might be found, but not applicable, they are deactivated. This 683 * keeps us from using potentially busy cached pages. 684 * 685 * This routine may not block. 686 * Must be called with vm_token held. 687 */ 688 vm_page_t 689 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 690 { 691 vm_page_t m; 692 693 ASSERT_LWKT_TOKEN_HELD(&vm_token); 694 while (TRUE) { 695 m = _vm_page_list_find( 696 PQ_CACHE, 697 (pindex + object->pg_color) & PQ_L2_MASK, 698 FALSE 699 ); 700 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 701 m->hold_count || m->wire_count)) { 702 /* cache page found busy */ 703 vm_page_deactivate(m); 704 #ifdef INVARIANTS 705 kprintf("Warning: busy page %p found in cache\n", m); 706 #endif 707 continue; 708 } 709 return m; 710 } 711 /* not reached */ 712 } 713 714 /* 715 * Find a free or zero page, with specified preference. We attempt to 716 * inline the nominal case and fall back to _vm_page_select_free() 717 * otherwise. 718 * 719 * This routine must be called with a critical section held. 720 * This routine may not block. 721 */ 722 static __inline vm_page_t 723 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 724 { 725 vm_page_t m; 726 727 m = _vm_page_list_find( 728 PQ_FREE, 729 (pindex + object->pg_color) & PQ_L2_MASK, 730 prefer_zero 731 ); 732 return(m); 733 } 734 735 /* 736 * vm_page_alloc() 737 * 738 * Allocate and return a memory cell associated with this VM object/offset 739 * pair. 740 * 741 * page_req classes: 742 * 743 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 744 * VM_ALLOC_QUICK like normal but cannot use cache 745 * VM_ALLOC_SYSTEM greater free drain 746 * VM_ALLOC_INTERRUPT allow free list to be completely drained 747 * VM_ALLOC_ZERO advisory request for pre-zero'd page 748 * 749 * The object must be locked. 750 * This routine may not block. 751 * The returned page will be marked PG_BUSY 752 * 753 * Additional special handling is required when called from an interrupt 754 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 755 * in this case. 756 */ 757 vm_page_t 758 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 759 { 760 vm_page_t m = NULL; 761 762 lwkt_gettoken(&vm_token); 763 764 KKASSERT(object != NULL); 765 KASSERT(!vm_page_lookup(object, pindex), 766 ("vm_page_alloc: page already allocated")); 767 KKASSERT(page_req & 768 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 769 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 770 771 /* 772 * Certain system threads (pageout daemon, buf_daemon's) are 773 * allowed to eat deeper into the free page list. 774 */ 775 if (curthread->td_flags & TDF_SYSTHREAD) 776 page_req |= VM_ALLOC_SYSTEM; 777 778 loop: 779 if (vmstats.v_free_count > vmstats.v_free_reserved || 780 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 781 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 782 vmstats.v_free_count > vmstats.v_interrupt_free_min) 783 ) { 784 /* 785 * The free queue has sufficient free pages to take one out. 786 */ 787 if (page_req & VM_ALLOC_ZERO) 788 m = vm_page_select_free(object, pindex, TRUE); 789 else 790 m = vm_page_select_free(object, pindex, FALSE); 791 } else if (page_req & VM_ALLOC_NORMAL) { 792 /* 793 * Allocatable from the cache (non-interrupt only). On 794 * success, we must free the page and try again, thus 795 * ensuring that vmstats.v_*_free_min counters are replenished. 796 */ 797 #ifdef INVARIANTS 798 if (curthread->td_preempted) { 799 kprintf("vm_page_alloc(): warning, attempt to allocate" 800 " cache page from preempting interrupt\n"); 801 m = NULL; 802 } else { 803 m = vm_page_select_cache(object, pindex); 804 } 805 #else 806 m = vm_page_select_cache(object, pindex); 807 #endif 808 /* 809 * On success move the page into the free queue and loop. 810 */ 811 if (m != NULL) { 812 KASSERT(m->dirty == 0, 813 ("Found dirty cache page %p", m)); 814 vm_page_busy(m); 815 vm_page_protect(m, VM_PROT_NONE); 816 vm_page_free(m); 817 goto loop; 818 } 819 820 /* 821 * On failure return NULL 822 */ 823 lwkt_reltoken(&vm_token); 824 #if defined(DIAGNOSTIC) 825 if (vmstats.v_cache_count > 0) 826 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 827 #endif 828 vm_pageout_deficit++; 829 pagedaemon_wakeup(); 830 return (NULL); 831 } else { 832 /* 833 * No pages available, wakeup the pageout daemon and give up. 834 */ 835 lwkt_reltoken(&vm_token); 836 vm_pageout_deficit++; 837 pagedaemon_wakeup(); 838 return (NULL); 839 } 840 841 /* 842 * Good page found. The page has not yet been busied. We are in 843 * a critical section. 844 */ 845 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 846 KASSERT(m->dirty == 0, 847 ("vm_page_alloc: free/cache page %p was dirty", m)); 848 849 /* 850 * Remove from free queue 851 */ 852 vm_page_unqueue_nowakeup(m); 853 854 /* 855 * Initialize structure. Only the PG_ZERO flag is inherited. Set 856 * the page PG_BUSY 857 */ 858 if (m->flags & PG_ZERO) { 859 vm_page_zero_count--; 860 m->flags = PG_ZERO | PG_BUSY; 861 } else { 862 m->flags = PG_BUSY; 863 } 864 m->wire_count = 0; 865 m->hold_count = 0; 866 m->act_count = 0; 867 m->busy = 0; 868 m->valid = 0; 869 870 /* 871 * vm_page_insert() is safe while holding vm_token. Note also that 872 * inserting a page here does not insert it into the pmap (which 873 * could cause us to block allocating memory). We cannot block 874 * anywhere. 875 */ 876 vm_page_insert(m, object, pindex); 877 878 /* 879 * Don't wakeup too often - wakeup the pageout daemon when 880 * we would be nearly out of memory. 881 */ 882 pagedaemon_wakeup(); 883 884 lwkt_reltoken(&vm_token); 885 886 /* 887 * A PG_BUSY page is returned. 888 */ 889 return (m); 890 } 891 892 /* 893 * Wait for sufficient free memory for nominal heavy memory use kernel 894 * operations. 895 */ 896 void 897 vm_wait_nominal(void) 898 { 899 while (vm_page_count_min(0)) 900 vm_wait(0); 901 } 902 903 /* 904 * Test if vm_wait_nominal() would block. 905 */ 906 int 907 vm_test_nominal(void) 908 { 909 if (vm_page_count_min(0)) 910 return(1); 911 return(0); 912 } 913 914 /* 915 * Block until free pages are available for allocation, called in various 916 * places before memory allocations. 917 * 918 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 919 * more generous then that. 920 */ 921 void 922 vm_wait(int timo) 923 { 924 /* 925 * never wait forever 926 */ 927 if (timo == 0) 928 timo = hz; 929 lwkt_gettoken(&vm_token); 930 931 if (curthread == pagethread) { 932 /* 933 * The pageout daemon itself needs pages, this is bad. 934 */ 935 if (vm_page_count_min(0)) { 936 vm_pageout_pages_needed = 1; 937 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 938 } 939 } else { 940 /* 941 * Wakeup the pageout daemon if necessary and wait. 942 */ 943 if (vm_page_count_target()) { 944 if (vm_pages_needed == 0) { 945 vm_pages_needed = 1; 946 wakeup(&vm_pages_needed); 947 } 948 ++vm_pages_waiting; /* SMP race ok */ 949 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 950 } 951 } 952 lwkt_reltoken(&vm_token); 953 } 954 955 /* 956 * Block until free pages are available for allocation 957 * 958 * Called only from vm_fault so that processes page faulting can be 959 * easily tracked. 960 */ 961 void 962 vm_waitpfault(void) 963 { 964 /* 965 * Wakeup the pageout daemon if necessary and wait. 966 */ 967 if (vm_page_count_target()) { 968 lwkt_gettoken(&vm_token); 969 if (vm_page_count_target()) { 970 if (vm_pages_needed == 0) { 971 vm_pages_needed = 1; 972 wakeup(&vm_pages_needed); 973 } 974 ++vm_pages_waiting; /* SMP race ok */ 975 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 976 } 977 lwkt_reltoken(&vm_token); 978 } 979 } 980 981 /* 982 * Put the specified page on the active list (if appropriate). Ensure 983 * that act_count is at least ACT_INIT but do not otherwise mess with it. 984 * 985 * The page queues must be locked. 986 * This routine may not block. 987 */ 988 void 989 vm_page_activate(vm_page_t m) 990 { 991 lwkt_gettoken(&vm_token); 992 if (m->queue != PQ_ACTIVE) { 993 if ((m->queue - m->pc) == PQ_CACHE) 994 mycpu->gd_cnt.v_reactivated++; 995 996 vm_page_unqueue(m); 997 998 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 999 m->queue = PQ_ACTIVE; 1000 vm_page_queues[PQ_ACTIVE].lcnt++; 1001 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 1002 m, pageq); 1003 if (m->act_count < ACT_INIT) 1004 m->act_count = ACT_INIT; 1005 vmstats.v_active_count++; 1006 } 1007 } else { 1008 if (m->act_count < ACT_INIT) 1009 m->act_count = ACT_INIT; 1010 } 1011 lwkt_reltoken(&vm_token); 1012 } 1013 1014 /* 1015 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1016 * routine is called when a page has been added to the cache or free 1017 * queues. 1018 * 1019 * This routine may not block. 1020 * This routine must be called at splvm() 1021 */ 1022 static __inline void 1023 vm_page_free_wakeup(void) 1024 { 1025 /* 1026 * If the pageout daemon itself needs pages, then tell it that 1027 * there are some free. 1028 */ 1029 if (vm_pageout_pages_needed && 1030 vmstats.v_cache_count + vmstats.v_free_count >= 1031 vmstats.v_pageout_free_min 1032 ) { 1033 wakeup(&vm_pageout_pages_needed); 1034 vm_pageout_pages_needed = 0; 1035 } 1036 1037 /* 1038 * Wakeup processes that are waiting on memory. 1039 * 1040 * NOTE: vm_paging_target() is the pageout daemon's target, while 1041 * vm_page_count_target() is somewhere inbetween. We want 1042 * to wake processes up prior to the pageout daemon reaching 1043 * its target to provide some hysteresis. 1044 */ 1045 if (vm_pages_waiting) { 1046 if (!vm_page_count_target()) { 1047 /* 1048 * Plenty of pages are free, wakeup everyone. 1049 */ 1050 vm_pages_waiting = 0; 1051 wakeup(&vmstats.v_free_count); 1052 ++mycpu->gd_cnt.v_ppwakeups; 1053 } else if (!vm_page_count_min(0)) { 1054 /* 1055 * Some pages are free, wakeup someone. 1056 */ 1057 int wcount = vm_pages_waiting; 1058 if (wcount > 0) 1059 --wcount; 1060 vm_pages_waiting = wcount; 1061 wakeup_one(&vmstats.v_free_count); 1062 ++mycpu->gd_cnt.v_ppwakeups; 1063 } 1064 } 1065 } 1066 1067 /* 1068 * vm_page_free_toq: 1069 * 1070 * Returns the given page to the PQ_FREE list, disassociating it with 1071 * any VM object. 1072 * 1073 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1074 * return (the page will have been freed). No particular spl is required 1075 * on entry. 1076 * 1077 * This routine may not block. 1078 */ 1079 void 1080 vm_page_free_toq(vm_page_t m) 1081 { 1082 struct vpgqueues *pq; 1083 1084 lwkt_gettoken(&vm_token); 1085 mycpu->gd_cnt.v_tfree++; 1086 1087 KKASSERT((m->flags & PG_MAPPED) == 0); 1088 1089 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1090 kprintf( 1091 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1092 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1093 m->hold_count); 1094 if ((m->queue - m->pc) == PQ_FREE) 1095 panic("vm_page_free: freeing free page"); 1096 else 1097 panic("vm_page_free: freeing busy page"); 1098 } 1099 1100 /* 1101 * unqueue, then remove page. Note that we cannot destroy 1102 * the page here because we do not want to call the pager's 1103 * callback routine until after we've put the page on the 1104 * appropriate free queue. 1105 */ 1106 vm_page_unqueue_nowakeup(m); 1107 vm_page_remove(m); 1108 1109 /* 1110 * No further management of fictitious pages occurs beyond object 1111 * and queue removal. 1112 */ 1113 if ((m->flags & PG_FICTITIOUS) != 0) { 1114 vm_page_wakeup(m); 1115 lwkt_reltoken(&vm_token); 1116 return; 1117 } 1118 1119 m->valid = 0; 1120 vm_page_undirty(m); 1121 1122 if (m->wire_count != 0) { 1123 if (m->wire_count > 1) { 1124 panic( 1125 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1126 m->wire_count, (long)m->pindex); 1127 } 1128 panic("vm_page_free: freeing wired page"); 1129 } 1130 1131 /* 1132 * Clear the UNMANAGED flag when freeing an unmanaged page. 1133 */ 1134 if (m->flags & PG_UNMANAGED) { 1135 vm_page_flag_clear(m, PG_UNMANAGED); 1136 } 1137 1138 if (m->hold_count != 0) { 1139 vm_page_flag_clear(m, PG_ZERO); 1140 m->queue = PQ_HOLD; 1141 } else { 1142 m->queue = PQ_FREE + m->pc; 1143 } 1144 pq = &vm_page_queues[m->queue]; 1145 pq->lcnt++; 1146 ++(*pq->cnt); 1147 1148 /* 1149 * Put zero'd pages on the end ( where we look for zero'd pages 1150 * first ) and non-zerod pages at the head. 1151 */ 1152 if (m->flags & PG_ZERO) { 1153 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1154 ++vm_page_zero_count; 1155 } else { 1156 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1157 } 1158 vm_page_wakeup(m); 1159 vm_page_free_wakeup(); 1160 lwkt_reltoken(&vm_token); 1161 } 1162 1163 /* 1164 * vm_page_free_fromq_fast() 1165 * 1166 * Remove a non-zero page from one of the free queues; the page is removed for 1167 * zeroing, so do not issue a wakeup. 1168 * 1169 * MPUNSAFE 1170 */ 1171 vm_page_t 1172 vm_page_free_fromq_fast(void) 1173 { 1174 static int qi; 1175 vm_page_t m; 1176 int i; 1177 1178 lwkt_gettoken(&vm_token); 1179 for (i = 0; i < PQ_L2_SIZE; ++i) { 1180 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1181 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1182 if (m && (m->flags & PG_ZERO) == 0) { 1183 KKASSERT(m->busy == 0 && (m->flags & PG_BUSY) == 0); 1184 vm_page_unqueue_nowakeup(m); 1185 vm_page_busy(m); 1186 break; 1187 } 1188 m = NULL; 1189 } 1190 lwkt_reltoken(&vm_token); 1191 return (m); 1192 } 1193 1194 /* 1195 * vm_page_unmanage() 1196 * 1197 * Prevent PV management from being done on the page. The page is 1198 * removed from the paging queues as if it were wired, and as a 1199 * consequence of no longer being managed the pageout daemon will not 1200 * touch it (since there is no way to locate the pte mappings for the 1201 * page). madvise() calls that mess with the pmap will also no longer 1202 * operate on the page. 1203 * 1204 * Beyond that the page is still reasonably 'normal'. Freeing the page 1205 * will clear the flag. 1206 * 1207 * This routine is used by OBJT_PHYS objects - objects using unswappable 1208 * physical memory as backing store rather then swap-backed memory and 1209 * will eventually be extended to support 4MB unmanaged physical 1210 * mappings. 1211 * 1212 * Must be called with a critical section held. 1213 * Must be called with vm_token held. 1214 */ 1215 void 1216 vm_page_unmanage(vm_page_t m) 1217 { 1218 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1219 if ((m->flags & PG_UNMANAGED) == 0) { 1220 if (m->wire_count == 0) 1221 vm_page_unqueue(m); 1222 } 1223 vm_page_flag_set(m, PG_UNMANAGED); 1224 } 1225 1226 /* 1227 * Mark this page as wired down by yet another map, removing it from 1228 * paging queues as necessary. 1229 * 1230 * The page queues must be locked. 1231 * This routine may not block. 1232 */ 1233 void 1234 vm_page_wire(vm_page_t m) 1235 { 1236 /* 1237 * Only bump the wire statistics if the page is not already wired, 1238 * and only unqueue the page if it is on some queue (if it is unmanaged 1239 * it is already off the queues). Don't do anything with fictitious 1240 * pages because they are always wired. 1241 */ 1242 lwkt_gettoken(&vm_token); 1243 if ((m->flags & PG_FICTITIOUS) == 0) { 1244 if (m->wire_count == 0) { 1245 if ((m->flags & PG_UNMANAGED) == 0) 1246 vm_page_unqueue(m); 1247 vmstats.v_wire_count++; 1248 } 1249 m->wire_count++; 1250 KASSERT(m->wire_count != 0, 1251 ("vm_page_wire: wire_count overflow m=%p", m)); 1252 } 1253 lwkt_reltoken(&vm_token); 1254 } 1255 1256 /* 1257 * Release one wiring of this page, potentially enabling it to be paged again. 1258 * 1259 * Many pages placed on the inactive queue should actually go 1260 * into the cache, but it is difficult to figure out which. What 1261 * we do instead, if the inactive target is well met, is to put 1262 * clean pages at the head of the inactive queue instead of the tail. 1263 * This will cause them to be moved to the cache more quickly and 1264 * if not actively re-referenced, freed more quickly. If we just 1265 * stick these pages at the end of the inactive queue, heavy filesystem 1266 * meta-data accesses can cause an unnecessary paging load on memory bound 1267 * processes. This optimization causes one-time-use metadata to be 1268 * reused more quickly. 1269 * 1270 * BUT, if we are in a low-memory situation we have no choice but to 1271 * put clean pages on the cache queue. 1272 * 1273 * A number of routines use vm_page_unwire() to guarantee that the page 1274 * will go into either the inactive or active queues, and will NEVER 1275 * be placed in the cache - for example, just after dirtying a page. 1276 * dirty pages in the cache are not allowed. 1277 * 1278 * The page queues must be locked. 1279 * This routine may not block. 1280 */ 1281 void 1282 vm_page_unwire(vm_page_t m, int activate) 1283 { 1284 lwkt_gettoken(&vm_token); 1285 if (m->flags & PG_FICTITIOUS) { 1286 /* do nothing */ 1287 } else if (m->wire_count <= 0) { 1288 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1289 } else { 1290 if (--m->wire_count == 0) { 1291 --vmstats.v_wire_count; 1292 if (m->flags & PG_UNMANAGED) { 1293 ; 1294 } else if (activate) { 1295 TAILQ_INSERT_TAIL( 1296 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1297 m->queue = PQ_ACTIVE; 1298 vm_page_queues[PQ_ACTIVE].lcnt++; 1299 vmstats.v_active_count++; 1300 } else { 1301 vm_page_flag_clear(m, PG_WINATCFLS); 1302 TAILQ_INSERT_TAIL( 1303 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1304 m->queue = PQ_INACTIVE; 1305 vm_page_queues[PQ_INACTIVE].lcnt++; 1306 vmstats.v_inactive_count++; 1307 ++vm_swapcache_inactive_heuristic; 1308 } 1309 } 1310 } 1311 lwkt_reltoken(&vm_token); 1312 } 1313 1314 1315 /* 1316 * Move the specified page to the inactive queue. If the page has 1317 * any associated swap, the swap is deallocated. 1318 * 1319 * Normally athead is 0 resulting in LRU operation. athead is set 1320 * to 1 if we want this page to be 'as if it were placed in the cache', 1321 * except without unmapping it from the process address space. 1322 * 1323 * This routine may not block. 1324 * The caller must hold vm_token. 1325 */ 1326 static __inline void 1327 _vm_page_deactivate(vm_page_t m, int athead) 1328 { 1329 /* 1330 * Ignore if already inactive. 1331 */ 1332 if (m->queue == PQ_INACTIVE) 1333 return; 1334 1335 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1336 if ((m->queue - m->pc) == PQ_CACHE) 1337 mycpu->gd_cnt.v_reactivated++; 1338 vm_page_flag_clear(m, PG_WINATCFLS); 1339 vm_page_unqueue(m); 1340 if (athead) { 1341 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1342 m, pageq); 1343 } else { 1344 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1345 m, pageq); 1346 ++vm_swapcache_inactive_heuristic; 1347 } 1348 m->queue = PQ_INACTIVE; 1349 vm_page_queues[PQ_INACTIVE].lcnt++; 1350 vmstats.v_inactive_count++; 1351 } 1352 } 1353 1354 /* 1355 * Attempt to deactivate a page. 1356 * 1357 * No requirements. 1358 */ 1359 void 1360 vm_page_deactivate(vm_page_t m) 1361 { 1362 lwkt_gettoken(&vm_token); 1363 _vm_page_deactivate(m, 0); 1364 lwkt_reltoken(&vm_token); 1365 } 1366 1367 /* 1368 * Attempt to move a page to PQ_CACHE. 1369 * Returns 0 on failure, 1 on success 1370 * 1371 * No requirements. 1372 */ 1373 int 1374 vm_page_try_to_cache(vm_page_t m) 1375 { 1376 lwkt_gettoken(&vm_token); 1377 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1378 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1379 lwkt_reltoken(&vm_token); 1380 return(0); 1381 } 1382 vm_page_busy(m); 1383 vm_page_test_dirty(m); 1384 if (m->dirty) { 1385 vm_page_wakeup(m); 1386 lwkt_reltoken(&vm_token); 1387 return(0); 1388 } 1389 vm_page_cache(m); 1390 lwkt_reltoken(&vm_token); 1391 return(1); 1392 } 1393 1394 /* 1395 * Attempt to free the page. If we cannot free it, we do nothing. 1396 * 1 is returned on success, 0 on failure. 1397 * 1398 * No requirements. 1399 */ 1400 int 1401 vm_page_try_to_free(vm_page_t m) 1402 { 1403 lwkt_gettoken(&vm_token); 1404 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1405 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1406 lwkt_reltoken(&vm_token); 1407 return(0); 1408 } 1409 vm_page_test_dirty(m); 1410 if (m->dirty) { 1411 lwkt_reltoken(&vm_token); 1412 return(0); 1413 } 1414 vm_page_busy(m); 1415 vm_page_protect(m, VM_PROT_NONE); 1416 vm_page_free(m); 1417 lwkt_reltoken(&vm_token); 1418 return(1); 1419 } 1420 1421 /* 1422 * vm_page_cache 1423 * 1424 * Put the specified page onto the page cache queue (if appropriate). 1425 * 1426 * The caller must hold vm_token. 1427 * This routine may not block. 1428 * The page must be busy, and this routine will release the busy and 1429 * possibly even free the page. 1430 */ 1431 void 1432 vm_page_cache(vm_page_t m) 1433 { 1434 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1435 1436 if ((m->flags & PG_UNMANAGED) || m->busy || 1437 m->wire_count || m->hold_count) { 1438 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1439 vm_page_wakeup(m); 1440 return; 1441 } 1442 1443 /* 1444 * Already in the cache (and thus not mapped) 1445 */ 1446 if ((m->queue - m->pc) == PQ_CACHE) { 1447 KKASSERT((m->flags & PG_MAPPED) == 0); 1448 vm_page_wakeup(m); 1449 return; 1450 } 1451 1452 /* 1453 * Caller is required to test m->dirty, but note that the act of 1454 * removing the page from its maps can cause it to become dirty 1455 * on an SMP system due to another cpu running in usermode. 1456 */ 1457 if (m->dirty) { 1458 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1459 (long)m->pindex); 1460 } 1461 1462 /* 1463 * Remove all pmaps and indicate that the page is not 1464 * writeable or mapped. Our vm_page_protect() call may 1465 * have blocked (especially w/ VM_PROT_NONE), so recheck 1466 * everything. 1467 */ 1468 vm_page_protect(m, VM_PROT_NONE); 1469 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy || 1470 m->wire_count || m->hold_count) { 1471 vm_page_wakeup(m); 1472 } else if (m->dirty) { 1473 vm_page_deactivate(m); 1474 vm_page_wakeup(m); 1475 } else { 1476 vm_page_unqueue_nowakeup(m); 1477 m->queue = PQ_CACHE + m->pc; 1478 vm_page_queues[m->queue].lcnt++; 1479 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1480 vmstats.v_cache_count++; 1481 vm_page_wakeup(m); 1482 vm_page_free_wakeup(); 1483 } 1484 } 1485 1486 /* 1487 * vm_page_dontneed() 1488 * 1489 * Cache, deactivate, or do nothing as appropriate. This routine 1490 * is typically used by madvise() MADV_DONTNEED. 1491 * 1492 * Generally speaking we want to move the page into the cache so 1493 * it gets reused quickly. However, this can result in a silly syndrome 1494 * due to the page recycling too quickly. Small objects will not be 1495 * fully cached. On the otherhand, if we move the page to the inactive 1496 * queue we wind up with a problem whereby very large objects 1497 * unnecessarily blow away our inactive and cache queues. 1498 * 1499 * The solution is to move the pages based on a fixed weighting. We 1500 * either leave them alone, deactivate them, or move them to the cache, 1501 * where moving them to the cache has the highest weighting. 1502 * By forcing some pages into other queues we eventually force the 1503 * system to balance the queues, potentially recovering other unrelated 1504 * space from active. The idea is to not force this to happen too 1505 * often. 1506 * 1507 * No requirements. 1508 */ 1509 void 1510 vm_page_dontneed(vm_page_t m) 1511 { 1512 static int dnweight; 1513 int dnw; 1514 int head; 1515 1516 dnw = ++dnweight; 1517 1518 /* 1519 * occassionally leave the page alone 1520 */ 1521 lwkt_gettoken(&vm_token); 1522 if ((dnw & 0x01F0) == 0 || 1523 m->queue == PQ_INACTIVE || 1524 m->queue - m->pc == PQ_CACHE 1525 ) { 1526 if (m->act_count >= ACT_INIT) 1527 --m->act_count; 1528 lwkt_reltoken(&vm_token); 1529 return; 1530 } 1531 1532 if (m->dirty == 0) 1533 vm_page_test_dirty(m); 1534 1535 if (m->dirty || (dnw & 0x0070) == 0) { 1536 /* 1537 * Deactivate the page 3 times out of 32. 1538 */ 1539 head = 0; 1540 } else { 1541 /* 1542 * Cache the page 28 times out of every 32. Note that 1543 * the page is deactivated instead of cached, but placed 1544 * at the head of the queue instead of the tail. 1545 */ 1546 head = 1; 1547 } 1548 _vm_page_deactivate(m, head); 1549 lwkt_reltoken(&vm_token); 1550 } 1551 1552 /* 1553 * Grab a page, blocking if it is busy and allocating a page if necessary. 1554 * A busy page is returned or NULL. 1555 * 1556 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1557 * If VM_ALLOC_RETRY is not specified 1558 * 1559 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1560 * always returned if we had blocked. 1561 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1562 * This routine may not be called from an interrupt. 1563 * The returned page may not be entirely valid. 1564 * 1565 * This routine may be called from mainline code without spl protection and 1566 * be guarenteed a busied page associated with the object at the specified 1567 * index. 1568 * 1569 * No requirements. 1570 */ 1571 vm_page_t 1572 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1573 { 1574 vm_page_t m; 1575 int generation; 1576 1577 KKASSERT(allocflags & 1578 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1579 lwkt_gettoken(&vm_token); 1580 vm_object_hold(object); 1581 retrylookup: 1582 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1583 if (m->busy || (m->flags & PG_BUSY)) { 1584 generation = object->generation; 1585 1586 while ((object->generation == generation) && 1587 (m->busy || (m->flags & PG_BUSY))) { 1588 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1589 tsleep(m, 0, "pgrbwt", 0); 1590 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1591 m = NULL; 1592 goto done; 1593 } 1594 } 1595 goto retrylookup; 1596 } else { 1597 vm_page_busy(m); 1598 goto done; 1599 } 1600 } 1601 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1602 if (m == NULL) { 1603 vm_wait(0); 1604 if ((allocflags & VM_ALLOC_RETRY) == 0) 1605 goto done; 1606 goto retrylookup; 1607 } 1608 done: 1609 vm_object_drop(object); 1610 lwkt_reltoken(&vm_token); 1611 return(m); 1612 } 1613 1614 /* 1615 * Mapping function for valid bits or for dirty bits in 1616 * a page. May not block. 1617 * 1618 * Inputs are required to range within a page. 1619 * 1620 * No requirements. 1621 * Non blocking. 1622 */ 1623 int 1624 vm_page_bits(int base, int size) 1625 { 1626 int first_bit; 1627 int last_bit; 1628 1629 KASSERT( 1630 base + size <= PAGE_SIZE, 1631 ("vm_page_bits: illegal base/size %d/%d", base, size) 1632 ); 1633 1634 if (size == 0) /* handle degenerate case */ 1635 return(0); 1636 1637 first_bit = base >> DEV_BSHIFT; 1638 last_bit = (base + size - 1) >> DEV_BSHIFT; 1639 1640 return ((2 << last_bit) - (1 << first_bit)); 1641 } 1642 1643 /* 1644 * Sets portions of a page valid and clean. The arguments are expected 1645 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1646 * of any partial chunks touched by the range. The invalid portion of 1647 * such chunks will be zero'd. 1648 * 1649 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1650 * align base to DEV_BSIZE so as not to mark clean a partially 1651 * truncated device block. Otherwise the dirty page status might be 1652 * lost. 1653 * 1654 * This routine may not block. 1655 * 1656 * (base + size) must be less then or equal to PAGE_SIZE. 1657 */ 1658 static void 1659 _vm_page_zero_valid(vm_page_t m, int base, int size) 1660 { 1661 int frag; 1662 int endoff; 1663 1664 if (size == 0) /* handle degenerate case */ 1665 return; 1666 1667 /* 1668 * If the base is not DEV_BSIZE aligned and the valid 1669 * bit is clear, we have to zero out a portion of the 1670 * first block. 1671 */ 1672 1673 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1674 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1675 ) { 1676 pmap_zero_page_area( 1677 VM_PAGE_TO_PHYS(m), 1678 frag, 1679 base - frag 1680 ); 1681 } 1682 1683 /* 1684 * If the ending offset is not DEV_BSIZE aligned and the 1685 * valid bit is clear, we have to zero out a portion of 1686 * the last block. 1687 */ 1688 1689 endoff = base + size; 1690 1691 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1692 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1693 ) { 1694 pmap_zero_page_area( 1695 VM_PAGE_TO_PHYS(m), 1696 endoff, 1697 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1698 ); 1699 } 1700 } 1701 1702 /* 1703 * Set valid, clear dirty bits. If validating the entire 1704 * page we can safely clear the pmap modify bit. We also 1705 * use this opportunity to clear the PG_NOSYNC flag. If a process 1706 * takes a write fault on a MAP_NOSYNC memory area the flag will 1707 * be set again. 1708 * 1709 * We set valid bits inclusive of any overlap, but we can only 1710 * clear dirty bits for DEV_BSIZE chunks that are fully within 1711 * the range. 1712 * 1713 * Page must be busied? 1714 * No other requirements. 1715 */ 1716 void 1717 vm_page_set_valid(vm_page_t m, int base, int size) 1718 { 1719 _vm_page_zero_valid(m, base, size); 1720 m->valid |= vm_page_bits(base, size); 1721 } 1722 1723 1724 /* 1725 * Set valid bits and clear dirty bits. 1726 * 1727 * NOTE: This function does not clear the pmap modified bit. 1728 * Also note that e.g. NFS may use a byte-granular base 1729 * and size. 1730 * 1731 * WARNING: Page must be busied? But vfs_clean_one_page() will call 1732 * this without necessarily busying the page (via bdwrite()). 1733 * So for now vm_token must also be held. 1734 * 1735 * No other requirements. 1736 */ 1737 void 1738 vm_page_set_validclean(vm_page_t m, int base, int size) 1739 { 1740 int pagebits; 1741 1742 _vm_page_zero_valid(m, base, size); 1743 pagebits = vm_page_bits(base, size); 1744 m->valid |= pagebits; 1745 m->dirty &= ~pagebits; 1746 if (base == 0 && size == PAGE_SIZE) { 1747 /*pmap_clear_modify(m);*/ 1748 vm_page_flag_clear(m, PG_NOSYNC); 1749 } 1750 } 1751 1752 /* 1753 * Set valid & dirty. Used by buwrite() 1754 * 1755 * WARNING: Page must be busied? But vfs_dirty_one_page() will 1756 * call this function in buwrite() so for now vm_token must 1757 * be held. 1758 * 1759 * No other requirements. 1760 */ 1761 void 1762 vm_page_set_validdirty(vm_page_t m, int base, int size) 1763 { 1764 int pagebits; 1765 1766 pagebits = vm_page_bits(base, size); 1767 m->valid |= pagebits; 1768 m->dirty |= pagebits; 1769 if (m->object) 1770 vm_object_set_writeable_dirty(m->object); 1771 } 1772 1773 /* 1774 * Clear dirty bits. 1775 * 1776 * NOTE: This function does not clear the pmap modified bit. 1777 * Also note that e.g. NFS may use a byte-granular base 1778 * and size. 1779 * 1780 * Page must be busied? 1781 * No other requirements. 1782 */ 1783 void 1784 vm_page_clear_dirty(vm_page_t m, int base, int size) 1785 { 1786 m->dirty &= ~vm_page_bits(base, size); 1787 if (base == 0 && size == PAGE_SIZE) { 1788 /*pmap_clear_modify(m);*/ 1789 vm_page_flag_clear(m, PG_NOSYNC); 1790 } 1791 } 1792 1793 /* 1794 * Make the page all-dirty. 1795 * 1796 * Also make sure the related object and vnode reflect the fact that the 1797 * object may now contain a dirty page. 1798 * 1799 * Page must be busied? 1800 * No other requirements. 1801 */ 1802 void 1803 vm_page_dirty(vm_page_t m) 1804 { 1805 #ifdef INVARIANTS 1806 int pqtype = m->queue - m->pc; 1807 #endif 1808 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1809 ("vm_page_dirty: page in free/cache queue!")); 1810 if (m->dirty != VM_PAGE_BITS_ALL) { 1811 m->dirty = VM_PAGE_BITS_ALL; 1812 if (m->object) 1813 vm_object_set_writeable_dirty(m->object); 1814 } 1815 } 1816 1817 /* 1818 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1819 * valid and dirty bits for the effected areas are cleared. 1820 * 1821 * Page must be busied? 1822 * Does not block. 1823 * No other requirements. 1824 */ 1825 void 1826 vm_page_set_invalid(vm_page_t m, int base, int size) 1827 { 1828 int bits; 1829 1830 bits = vm_page_bits(base, size); 1831 m->valid &= ~bits; 1832 m->dirty &= ~bits; 1833 m->object->generation++; 1834 } 1835 1836 /* 1837 * The kernel assumes that the invalid portions of a page contain 1838 * garbage, but such pages can be mapped into memory by user code. 1839 * When this occurs, we must zero out the non-valid portions of the 1840 * page so user code sees what it expects. 1841 * 1842 * Pages are most often semi-valid when the end of a file is mapped 1843 * into memory and the file's size is not page aligned. 1844 * 1845 * Page must be busied? 1846 * No other requirements. 1847 */ 1848 void 1849 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1850 { 1851 int b; 1852 int i; 1853 1854 /* 1855 * Scan the valid bits looking for invalid sections that 1856 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1857 * valid bit may be set ) have already been zerod by 1858 * vm_page_set_validclean(). 1859 */ 1860 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1861 if (i == (PAGE_SIZE / DEV_BSIZE) || 1862 (m->valid & (1 << i)) 1863 ) { 1864 if (i > b) { 1865 pmap_zero_page_area( 1866 VM_PAGE_TO_PHYS(m), 1867 b << DEV_BSHIFT, 1868 (i - b) << DEV_BSHIFT 1869 ); 1870 } 1871 b = i + 1; 1872 } 1873 } 1874 1875 /* 1876 * setvalid is TRUE when we can safely set the zero'd areas 1877 * as being valid. We can do this if there are no cache consistency 1878 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1879 */ 1880 if (setvalid) 1881 m->valid = VM_PAGE_BITS_ALL; 1882 } 1883 1884 /* 1885 * Is a (partial) page valid? Note that the case where size == 0 1886 * will return FALSE in the degenerate case where the page is entirely 1887 * invalid, and TRUE otherwise. 1888 * 1889 * Does not block. 1890 * No other requirements. 1891 */ 1892 int 1893 vm_page_is_valid(vm_page_t m, int base, int size) 1894 { 1895 int bits = vm_page_bits(base, size); 1896 1897 if (m->valid && ((m->valid & bits) == bits)) 1898 return 1; 1899 else 1900 return 0; 1901 } 1902 1903 /* 1904 * update dirty bits from pmap/mmu. May not block. 1905 * 1906 * Caller must hold vm_token if non-blocking operation desired. 1907 * No other requirements. 1908 */ 1909 void 1910 vm_page_test_dirty(vm_page_t m) 1911 { 1912 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1913 vm_page_dirty(m); 1914 } 1915 } 1916 1917 /* 1918 * Register an action, associating it with its vm_page 1919 */ 1920 void 1921 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 1922 { 1923 struct vm_page_action_list *list; 1924 int hv; 1925 1926 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1927 list = &action_list[hv]; 1928 1929 lwkt_gettoken(&vm_token); 1930 vm_page_flag_set(action->m, PG_ACTIONLIST); 1931 action->event = event; 1932 LIST_INSERT_HEAD(list, action, entry); 1933 lwkt_reltoken(&vm_token); 1934 } 1935 1936 /* 1937 * Unregister an action, disassociating it from its related vm_page 1938 */ 1939 void 1940 vm_page_unregister_action(vm_page_action_t action) 1941 { 1942 struct vm_page_action_list *list; 1943 int hv; 1944 1945 lwkt_gettoken(&vm_token); 1946 if (action->event != VMEVENT_NONE) { 1947 action->event = VMEVENT_NONE; 1948 LIST_REMOVE(action, entry); 1949 1950 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1951 list = &action_list[hv]; 1952 if (LIST_EMPTY(list)) 1953 vm_page_flag_clear(action->m, PG_ACTIONLIST); 1954 } 1955 lwkt_reltoken(&vm_token); 1956 } 1957 1958 /* 1959 * Issue an event on a VM page. Corresponding action structures are 1960 * removed from the page's list and called. 1961 * 1962 * If the vm_page has no more pending action events we clear its 1963 * PG_ACTIONLIST flag. 1964 */ 1965 void 1966 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1967 { 1968 struct vm_page_action_list *list; 1969 struct vm_page_action *scan; 1970 struct vm_page_action *next; 1971 int hv; 1972 int all; 1973 1974 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 1975 list = &action_list[hv]; 1976 all = 1; 1977 1978 lwkt_gettoken(&vm_token); 1979 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 1980 if (scan->m == m) { 1981 if (scan->event == event) { 1982 scan->event = VMEVENT_NONE; 1983 LIST_REMOVE(scan, entry); 1984 scan->func(m, scan); 1985 /* XXX */ 1986 } else { 1987 all = 0; 1988 } 1989 } 1990 } 1991 if (all) 1992 vm_page_flag_clear(m, PG_ACTIONLIST); 1993 lwkt_reltoken(&vm_token); 1994 } 1995 1996 #include "opt_ddb.h" 1997 #ifdef DDB 1998 #include <sys/kernel.h> 1999 2000 #include <ddb/ddb.h> 2001 2002 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2003 { 2004 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 2005 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 2006 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 2007 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 2008 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 2009 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 2010 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 2011 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 2012 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 2013 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 2014 } 2015 2016 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2017 { 2018 int i; 2019 db_printf("PQ_FREE:"); 2020 for(i=0;i<PQ_L2_SIZE;i++) { 2021 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 2022 } 2023 db_printf("\n"); 2024 2025 db_printf("PQ_CACHE:"); 2026 for(i=0;i<PQ_L2_SIZE;i++) { 2027 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 2028 } 2029 db_printf("\n"); 2030 2031 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2032 vm_page_queues[PQ_ACTIVE].lcnt, 2033 vm_page_queues[PQ_INACTIVE].lcnt); 2034 } 2035 #endif /* DDB */ 2036