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 #include <sys/mplock2.h> 99 100 #define VMACTION_HSIZE 256 101 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 102 103 static void vm_page_queue_init(void); 104 static void vm_page_free_wakeup(void); 105 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 106 static vm_page_t _vm_page_list_find2(int basequeue, int index); 107 108 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 109 110 LIST_HEAD(vm_page_action_list, vm_page_action); 111 struct vm_page_action_list action_list[VMACTION_HSIZE]; 112 static volatile int vm_pages_waiting; 113 114 115 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 116 vm_pindex_t, pindex); 117 118 static void 119 vm_page_queue_init(void) 120 { 121 int i; 122 123 for (i = 0; i < PQ_L2_SIZE; i++) 124 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 125 for (i = 0; i < PQ_L2_SIZE; i++) 126 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 127 128 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 129 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 130 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 131 /* PQ_NONE has no queue */ 132 133 for (i = 0; i < PQ_COUNT; i++) 134 TAILQ_INIT(&vm_page_queues[i].pl); 135 136 for (i = 0; i < VMACTION_HSIZE; i++) 137 LIST_INIT(&action_list[i]); 138 } 139 140 /* 141 * note: place in initialized data section? Is this necessary? 142 */ 143 long first_page = 0; 144 int vm_page_array_size = 0; 145 int vm_page_zero_count = 0; 146 vm_page_t vm_page_array = 0; 147 148 /* 149 * (low level boot) 150 * 151 * Sets the page size, perhaps based upon the memory size. 152 * Must be called before any use of page-size dependent functions. 153 */ 154 void 155 vm_set_page_size(void) 156 { 157 if (vmstats.v_page_size == 0) 158 vmstats.v_page_size = PAGE_SIZE; 159 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 160 panic("vm_set_page_size: page size not a power of two"); 161 } 162 163 /* 164 * (low level boot) 165 * 166 * Add a new page to the freelist for use by the system. New pages 167 * are added to both the head and tail of the associated free page 168 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 169 * requests pull 'recent' adds (higher physical addresses) first. 170 * 171 * Must be called in a critical section. 172 */ 173 vm_page_t 174 vm_add_new_page(vm_paddr_t pa) 175 { 176 struct vpgqueues *vpq; 177 vm_page_t m; 178 179 ++vmstats.v_page_count; 180 ++vmstats.v_free_count; 181 m = PHYS_TO_VM_PAGE(pa); 182 m->phys_addr = pa; 183 m->flags = 0; 184 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 185 m->queue = m->pc + PQ_FREE; 186 KKASSERT(m->dirty == 0); 187 188 vpq = &vm_page_queues[m->queue]; 189 if (vpq->flipflop) 190 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 191 else 192 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 193 vpq->flipflop = 1 - vpq->flipflop; 194 195 vm_page_queues[m->queue].lcnt++; 196 return (m); 197 } 198 199 /* 200 * (low level boot) 201 * 202 * Initializes the resident memory module. 203 * 204 * Preallocates memory for critical VM structures and arrays prior to 205 * kernel_map becoming available. 206 * 207 * Memory is allocated from (virtual2_start, virtual2_end) if available, 208 * otherwise memory is allocated from (virtual_start, virtual_end). 209 * 210 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 211 * large enough to hold vm_page_array & other structures for machines with 212 * large amounts of ram, so we want to use virtual2* when available. 213 */ 214 void 215 vm_page_startup(void) 216 { 217 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 218 vm_offset_t mapped; 219 vm_size_t npages; 220 vm_paddr_t page_range; 221 vm_paddr_t new_end; 222 int i; 223 vm_paddr_t pa; 224 int nblocks; 225 vm_paddr_t last_pa; 226 vm_paddr_t end; 227 vm_paddr_t biggestone, biggestsize; 228 vm_paddr_t total; 229 230 total = 0; 231 biggestsize = 0; 232 biggestone = 0; 233 nblocks = 0; 234 vaddr = round_page(vaddr); 235 236 for (i = 0; phys_avail[i + 1]; i += 2) { 237 phys_avail[i] = round_page64(phys_avail[i]); 238 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 239 } 240 241 for (i = 0; phys_avail[i + 1]; i += 2) { 242 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 243 244 if (size > biggestsize) { 245 biggestone = i; 246 biggestsize = size; 247 } 248 ++nblocks; 249 total += size; 250 } 251 252 end = phys_avail[biggestone+1]; 253 end = trunc_page(end); 254 255 /* 256 * Initialize the queue headers for the free queue, the active queue 257 * and the inactive queue. 258 */ 259 260 vm_page_queue_init(); 261 262 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */ 263 #if !defined(_KERNEL_VIRTUAL) 264 /* 265 * Allocate a bitmap to indicate that a random physical page 266 * needs to be included in a minidump. 267 * 268 * The amd64 port needs this to indicate which direct map pages 269 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 270 * 271 * However, i386 still needs this workspace internally within the 272 * minidump code. In theory, they are not needed on i386, but are 273 * included should the sf_buf code decide to use them. 274 */ 275 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 276 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 277 end -= vm_page_dump_size; 278 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 279 VM_PROT_READ | VM_PROT_WRITE); 280 bzero((void *)vm_page_dump, vm_page_dump_size); 281 #endif 282 283 /* 284 * Compute the number of pages of memory that will be available for 285 * use (taking into account the overhead of a page structure per 286 * page). 287 */ 288 first_page = phys_avail[0] / PAGE_SIZE; 289 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 290 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 291 292 /* 293 * Initialize the mem entry structures now, and put them in the free 294 * queue. 295 */ 296 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 297 mapped = pmap_map(&vaddr, new_end, end, 298 VM_PROT_READ | VM_PROT_WRITE); 299 vm_page_array = (vm_page_t)mapped; 300 301 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 302 /* 303 * since pmap_map on amd64 returns stuff out of a direct-map region, 304 * we have to manually add these pages to the minidump tracking so 305 * that they can be dumped, including the vm_page_array. 306 */ 307 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 308 dump_add_page(pa); 309 #endif 310 311 /* 312 * Clear all of the page structures 313 */ 314 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 315 vm_page_array_size = page_range; 316 317 /* 318 * Construct the free queue(s) in ascending order (by physical 319 * address) so that the first 16MB of physical memory is allocated 320 * last rather than first. On large-memory machines, this avoids 321 * the exhaustion of low physical memory before isa_dmainit has run. 322 */ 323 vmstats.v_page_count = 0; 324 vmstats.v_free_count = 0; 325 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 326 pa = phys_avail[i]; 327 if (i == biggestone) 328 last_pa = new_end; 329 else 330 last_pa = phys_avail[i + 1]; 331 while (pa < last_pa && npages-- > 0) { 332 vm_add_new_page(pa); 333 pa += PAGE_SIZE; 334 } 335 } 336 if (virtual2_start) 337 virtual2_start = vaddr; 338 else 339 virtual_start = vaddr; 340 } 341 342 /* 343 * Scan comparison function for Red-Black tree scans. An inclusive 344 * (start,end) is expected. Other fields are not used. 345 */ 346 int 347 rb_vm_page_scancmp(struct vm_page *p, void *data) 348 { 349 struct rb_vm_page_scan_info *info = data; 350 351 if (p->pindex < info->start_pindex) 352 return(-1); 353 if (p->pindex > info->end_pindex) 354 return(1); 355 return(0); 356 } 357 358 int 359 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 360 { 361 if (p1->pindex < p2->pindex) 362 return(-1); 363 if (p1->pindex > p2->pindex) 364 return(1); 365 return(0); 366 } 367 368 /* 369 * Holding a page keeps it from being reused. Other parts of the system 370 * can still disassociate the page from its current object and free it, or 371 * perform read or write I/O on it and/or otherwise manipulate the page, 372 * but if the page is held the VM system will leave the page and its data 373 * intact and not reuse the page for other purposes until the last hold 374 * reference is released. (see vm_page_wire() if you want to prevent the 375 * page from being disassociated from its object too). 376 * 377 * The caller must hold vm_token. 378 * 379 * The caller must still validate the contents of the page and, if necessary, 380 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 381 * before manipulating the page. 382 */ 383 void 384 vm_page_hold(vm_page_t m) 385 { 386 ASSERT_LWKT_TOKEN_HELD(&vm_token); 387 ++m->hold_count; 388 } 389 390 /* 391 * The opposite of vm_page_hold(). A page can be freed while being held, 392 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 393 * in this case to actually free it once the hold count drops to 0. 394 * 395 * The caller must hold vm_token if non-blocking operation is desired, 396 * but otherwise does not need to. 397 */ 398 void 399 vm_page_unhold(vm_page_t m) 400 { 401 lwkt_gettoken(&vm_token); 402 --m->hold_count; 403 KASSERT(m->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 404 if (m->hold_count == 0 && m->queue == PQ_HOLD) { 405 vm_page_busy(m); 406 vm_page_free_toq(m); 407 } 408 lwkt_reltoken(&vm_token); 409 } 410 411 /* 412 * Inserts the given vm_page into the object and object list. 413 * 414 * The pagetables are not updated but will presumably fault the page 415 * in if necessary, or if a kernel page the caller will at some point 416 * enter the page into the kernel's pmap. We are not allowed to block 417 * here so we *can't* do this anyway. 418 * 419 * This routine may not block. 420 * This routine must be called with the vm_token held. 421 * This routine must be called with a critical section held. 422 */ 423 void 424 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 425 { 426 ASSERT_LWKT_TOKEN_HELD(&vm_token); 427 if (m->object != NULL) 428 panic("vm_page_insert: already inserted"); 429 430 /* 431 * Record the object/offset pair in this page 432 */ 433 m->object = object; 434 m->pindex = pindex; 435 436 /* 437 * Insert it into the object. 438 */ 439 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 440 object->generation++; 441 442 /* 443 * show that the object has one more resident page. 444 */ 445 object->resident_page_count++; 446 447 /* 448 * Add the pv_list_cout of the page when its inserted in 449 * the object 450 */ 451 object->agg_pv_list_count = object->agg_pv_list_count + m->md.pv_list_count; 452 453 /* 454 * Since we are inserting a new and possibly dirty page, 455 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 456 */ 457 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 458 vm_object_set_writeable_dirty(object); 459 460 /* 461 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 462 */ 463 swap_pager_page_inserted(m); 464 } 465 466 /* 467 * Removes the given vm_page_t from the global (object,index) hash table 468 * and from the object's memq. 469 * 470 * The underlying pmap entry (if any) is NOT removed here. 471 * This routine may not block. 472 * 473 * The page must be BUSY and will remain BUSY on return. 474 * No other requirements. 475 * 476 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 477 * it busy. 478 */ 479 void 480 vm_page_remove(vm_page_t m) 481 { 482 vm_object_t object; 483 484 lwkt_gettoken(&vm_token); 485 if (m->object == NULL) { 486 lwkt_reltoken(&vm_token); 487 return; 488 } 489 490 if ((m->flags & PG_BUSY) == 0) 491 panic("vm_page_remove: page not busy"); 492 493 object = m->object; 494 495 /* 496 * Remove the page from the object and update the object. 497 */ 498 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 499 object->resident_page_count--; 500 object->agg_pv_list_count = object->agg_pv_list_count - m->md.pv_list_count; 501 object->generation++; 502 m->object = NULL; 503 504 lwkt_reltoken(&vm_token); 505 } 506 507 /* 508 * Locate and return the page at (object, pindex), or NULL if the 509 * page could not be found. 510 * 511 * The caller must hold vm_token. 512 */ 513 vm_page_t 514 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 515 { 516 vm_page_t m; 517 518 /* 519 * Search the hash table for this object/offset pair 520 */ 521 ASSERT_LWKT_TOKEN_HELD(&vm_token); 522 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 523 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 524 return(m); 525 } 526 527 /* 528 * vm_page_rename() 529 * 530 * Move the given memory entry from its current object to the specified 531 * target object/offset. 532 * 533 * The object must be locked. 534 * This routine may not block. 535 * 536 * Note: This routine will raise itself to splvm(), the caller need not. 537 * 538 * Note: Swap associated with the page must be invalidated by the move. We 539 * have to do this for several reasons: (1) we aren't freeing the 540 * page, (2) we are dirtying the page, (3) the VM system is probably 541 * moving the page from object A to B, and will then later move 542 * the backing store from A to B and we can't have a conflict. 543 * 544 * Note: We *always* dirty the page. It is necessary both for the 545 * fact that we moved it, and because we may be invalidating 546 * swap. If the page is on the cache, we have to deactivate it 547 * or vm_page_dirty() will panic. Dirty pages are not allowed 548 * on the cache. 549 */ 550 void 551 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 552 { 553 lwkt_gettoken(&vm_token); 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 lwkt_reltoken(&vm_token); 561 } 562 563 /* 564 * vm_page_unqueue() without any wakeup. This routine is used when a page 565 * is being moved between queues or otherwise is to remain BUSYied by the 566 * caller. 567 * 568 * The caller must hold vm_token 569 * This routine may not block. 570 */ 571 void 572 vm_page_unqueue_nowakeup(vm_page_t m) 573 { 574 int queue = m->queue; 575 struct vpgqueues *pq; 576 577 ASSERT_LWKT_TOKEN_HELD(&vm_token); 578 if (queue != PQ_NONE) { 579 pq = &vm_page_queues[queue]; 580 m->queue = PQ_NONE; 581 TAILQ_REMOVE(&pq->pl, m, pageq); 582 (*pq->cnt)--; 583 pq->lcnt--; 584 } 585 } 586 587 /* 588 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 589 * if necessary. 590 * 591 * The caller must hold vm_token 592 * This routine may not block. 593 */ 594 void 595 vm_page_unqueue(vm_page_t m) 596 { 597 int queue = m->queue; 598 struct vpgqueues *pq; 599 600 ASSERT_LWKT_TOKEN_HELD(&vm_token); 601 if (queue != PQ_NONE) { 602 m->queue = PQ_NONE; 603 pq = &vm_page_queues[queue]; 604 TAILQ_REMOVE(&pq->pl, m, pageq); 605 (*pq->cnt)--; 606 pq->lcnt--; 607 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 608 pagedaemon_wakeup(); 609 } 610 } 611 612 /* 613 * vm_page_list_find() 614 * 615 * Find a page on the specified queue with color optimization. 616 * 617 * The page coloring optimization attempts to locate a page that does 618 * not overload other nearby pages in the object in the cpu's L1 or L2 619 * caches. We need this optimization because cpu caches tend to be 620 * physical caches, while object spaces tend to be virtual. 621 * 622 * Must be called with vm_token held. 623 * This routine may not block. 624 * 625 * Note that this routine is carefully inlined. A non-inlined version 626 * is available for outside callers but the only critical path is 627 * from within this source file. 628 */ 629 static __inline 630 vm_page_t 631 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 632 { 633 vm_page_t m; 634 635 if (prefer_zero) 636 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 637 else 638 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 639 if (m == NULL) 640 m = _vm_page_list_find2(basequeue, index); 641 return(m); 642 } 643 644 static vm_page_t 645 _vm_page_list_find2(int basequeue, int index) 646 { 647 int i; 648 vm_page_t m = NULL; 649 struct vpgqueues *pq; 650 651 pq = &vm_page_queues[basequeue]; 652 653 /* 654 * Note that for the first loop, index+i and index-i wind up at the 655 * same place. Even though this is not totally optimal, we've already 656 * blown it by missing the cache case so we do not care. 657 */ 658 659 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 660 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 661 break; 662 663 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 664 break; 665 } 666 return(m); 667 } 668 669 /* 670 * Must be called with vm_token held if the caller desired non-blocking 671 * operation and a stable result. 672 */ 673 vm_page_t 674 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 675 { 676 return(_vm_page_list_find(basequeue, index, prefer_zero)); 677 } 678 679 /* 680 * Find a page on the cache queue with color optimization. As pages 681 * might be found, but not applicable, they are deactivated. This 682 * keeps us from using potentially busy cached pages. 683 * 684 * This routine may not block. 685 * Must be called with vm_token held. 686 */ 687 vm_page_t 688 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 689 { 690 vm_page_t m; 691 692 ASSERT_LWKT_TOKEN_HELD(&vm_token); 693 while (TRUE) { 694 m = _vm_page_list_find( 695 PQ_CACHE, 696 (pindex + object->pg_color) & PQ_L2_MASK, 697 FALSE 698 ); 699 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 700 m->hold_count || m->wire_count)) { 701 /* cache page found busy */ 702 vm_page_deactivate(m); 703 #ifdef INVARIANTS 704 kprintf("Warning: busy page %p found in cache\n", m); 705 #endif 706 continue; 707 } 708 return m; 709 } 710 /* not reached */ 711 } 712 713 /* 714 * Find a free or zero page, with specified preference. We attempt to 715 * inline the nominal case and fall back to _vm_page_select_free() 716 * otherwise. 717 * 718 * This routine must be called with a critical section held. 719 * This routine may not block. 720 */ 721 static __inline vm_page_t 722 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 723 { 724 vm_page_t m; 725 726 m = _vm_page_list_find( 727 PQ_FREE, 728 (pindex + object->pg_color) & PQ_L2_MASK, 729 prefer_zero 730 ); 731 return(m); 732 } 733 734 /* 735 * vm_page_alloc() 736 * 737 * Allocate and return a memory cell associated with this VM object/offset 738 * pair. 739 * 740 * page_req classes: 741 * 742 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 743 * VM_ALLOC_QUICK like normal but cannot use cache 744 * VM_ALLOC_SYSTEM greater free drain 745 * VM_ALLOC_INTERRUPT allow free list to be completely drained 746 * VM_ALLOC_ZERO advisory request for pre-zero'd page 747 * 748 * The object must be locked. 749 * This routine may not block. 750 * The returned page will be marked PG_BUSY 751 * 752 * Additional special handling is required when called from an interrupt 753 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 754 * in this case. 755 */ 756 vm_page_t 757 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 758 { 759 vm_page_t m = NULL; 760 761 lwkt_gettoken(&vm_token); 762 763 KKASSERT(object != NULL); 764 KASSERT(!vm_page_lookup(object, pindex), 765 ("vm_page_alloc: page already allocated")); 766 KKASSERT(page_req & 767 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 768 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 769 770 /* 771 * Certain system threads (pageout daemon, buf_daemon's) are 772 * allowed to eat deeper into the free page list. 773 */ 774 if (curthread->td_flags & TDF_SYSTHREAD) 775 page_req |= VM_ALLOC_SYSTEM; 776 777 loop: 778 if (vmstats.v_free_count > vmstats.v_free_reserved || 779 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 780 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 781 vmstats.v_free_count > vmstats.v_interrupt_free_min) 782 ) { 783 /* 784 * The free queue has sufficient free pages to take one out. 785 */ 786 if (page_req & VM_ALLOC_ZERO) 787 m = vm_page_select_free(object, pindex, TRUE); 788 else 789 m = vm_page_select_free(object, pindex, FALSE); 790 } else if (page_req & VM_ALLOC_NORMAL) { 791 /* 792 * Allocatable from the cache (non-interrupt only). On 793 * success, we must free the page and try again, thus 794 * ensuring that vmstats.v_*_free_min counters are replenished. 795 */ 796 #ifdef INVARIANTS 797 if (curthread->td_preempted) { 798 kprintf("vm_page_alloc(): warning, attempt to allocate" 799 " cache page from preempting interrupt\n"); 800 m = NULL; 801 } else { 802 m = vm_page_select_cache(object, pindex); 803 } 804 #else 805 m = vm_page_select_cache(object, pindex); 806 #endif 807 /* 808 * On success move the page into the free queue and loop. 809 */ 810 if (m != NULL) { 811 KASSERT(m->dirty == 0, 812 ("Found dirty cache page %p", m)); 813 vm_page_busy(m); 814 vm_page_protect(m, VM_PROT_NONE); 815 vm_page_free(m); 816 goto loop; 817 } 818 819 /* 820 * On failure return NULL 821 */ 822 lwkt_reltoken(&vm_token); 823 #if defined(DIAGNOSTIC) 824 if (vmstats.v_cache_count > 0) 825 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 826 #endif 827 vm_pageout_deficit++; 828 pagedaemon_wakeup(); 829 return (NULL); 830 } else { 831 /* 832 * No pages available, wakeup the pageout daemon and give up. 833 */ 834 lwkt_reltoken(&vm_token); 835 vm_pageout_deficit++; 836 pagedaemon_wakeup(); 837 return (NULL); 838 } 839 840 /* 841 * Good page found. The page has not yet been busied. We are in 842 * a critical section. 843 */ 844 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 845 KASSERT(m->dirty == 0, 846 ("vm_page_alloc: free/cache page %p was dirty", m)); 847 848 /* 849 * Remove from free queue 850 */ 851 vm_page_unqueue_nowakeup(m); 852 853 /* 854 * Initialize structure. Only the PG_ZERO flag is inherited. Set 855 * the page PG_BUSY 856 */ 857 if (m->flags & PG_ZERO) { 858 vm_page_zero_count--; 859 m->flags = PG_ZERO | PG_BUSY; 860 } else { 861 m->flags = PG_BUSY; 862 } 863 m->wire_count = 0; 864 m->hold_count = 0; 865 m->act_count = 0; 866 m->busy = 0; 867 m->valid = 0; 868 869 /* 870 * vm_page_insert() is safe while holding vm_token. Note also that 871 * inserting a page here does not insert it into the pmap (which 872 * could cause us to block allocating memory). We cannot block 873 * anywhere. 874 */ 875 vm_page_insert(m, object, pindex); 876 877 /* 878 * Don't wakeup too often - wakeup the pageout daemon when 879 * we would be nearly out of memory. 880 */ 881 pagedaemon_wakeup(); 882 883 lwkt_reltoken(&vm_token); 884 885 /* 886 * A PG_BUSY page is returned. 887 */ 888 return (m); 889 } 890 891 /* 892 * Wait for sufficient free memory for nominal heavy memory use kernel 893 * operations. 894 */ 895 void 896 vm_wait_nominal(void) 897 { 898 while (vm_page_count_min(0)) 899 vm_wait(0); 900 } 901 902 /* 903 * Test if vm_wait_nominal() would block. 904 */ 905 int 906 vm_test_nominal(void) 907 { 908 if (vm_page_count_min(0)) 909 return(1); 910 return(0); 911 } 912 913 /* 914 * Block until free pages are available for allocation, called in various 915 * places before memory allocations. 916 * 917 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 918 * more generous then that. 919 */ 920 void 921 vm_wait(int timo) 922 { 923 /* 924 * never wait forever 925 */ 926 if (timo == 0) 927 timo = hz; 928 lwkt_gettoken(&vm_token); 929 930 if (curthread == pagethread) { 931 /* 932 * The pageout daemon itself needs pages, this is bad. 933 */ 934 if (vm_page_count_min(0)) { 935 vm_pageout_pages_needed = 1; 936 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 937 } 938 } else { 939 /* 940 * Wakeup the pageout daemon if necessary and wait. 941 */ 942 if (vm_page_count_target()) { 943 if (vm_pages_needed == 0) { 944 vm_pages_needed = 1; 945 wakeup(&vm_pages_needed); 946 } 947 ++vm_pages_waiting; /* SMP race ok */ 948 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 949 } 950 } 951 lwkt_reltoken(&vm_token); 952 } 953 954 /* 955 * Block until free pages are available for allocation 956 * 957 * Called only from vm_fault so that processes page faulting can be 958 * easily tracked. 959 */ 960 void 961 vm_waitpfault(void) 962 { 963 /* 964 * Wakeup the pageout daemon if necessary and wait. 965 */ 966 if (vm_page_count_target()) { 967 lwkt_gettoken(&vm_token); 968 if (vm_page_count_target()) { 969 if (vm_pages_needed == 0) { 970 vm_pages_needed = 1; 971 wakeup(&vm_pages_needed); 972 } 973 ++vm_pages_waiting; /* SMP race ok */ 974 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 975 } 976 lwkt_reltoken(&vm_token); 977 } 978 } 979 980 /* 981 * Put the specified page on the active list (if appropriate). Ensure 982 * that act_count is at least ACT_INIT but do not otherwise mess with it. 983 * 984 * The page queues must be locked. 985 * This routine may not block. 986 */ 987 void 988 vm_page_activate(vm_page_t m) 989 { 990 lwkt_gettoken(&vm_token); 991 if (m->queue != PQ_ACTIVE) { 992 if ((m->queue - m->pc) == PQ_CACHE) 993 mycpu->gd_cnt.v_reactivated++; 994 995 vm_page_unqueue(m); 996 997 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 998 m->queue = PQ_ACTIVE; 999 vm_page_queues[PQ_ACTIVE].lcnt++; 1000 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 1001 m, pageq); 1002 if (m->act_count < ACT_INIT) 1003 m->act_count = ACT_INIT; 1004 vmstats.v_active_count++; 1005 } 1006 } else { 1007 if (m->act_count < ACT_INIT) 1008 m->act_count = ACT_INIT; 1009 } 1010 lwkt_reltoken(&vm_token); 1011 } 1012 1013 /* 1014 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1015 * routine is called when a page has been added to the cache or free 1016 * queues. 1017 * 1018 * This routine may not block. 1019 * This routine must be called at splvm() 1020 */ 1021 static __inline void 1022 vm_page_free_wakeup(void) 1023 { 1024 /* 1025 * If the pageout daemon itself needs pages, then tell it that 1026 * there are some free. 1027 */ 1028 if (vm_pageout_pages_needed && 1029 vmstats.v_cache_count + vmstats.v_free_count >= 1030 vmstats.v_pageout_free_min 1031 ) { 1032 wakeup(&vm_pageout_pages_needed); 1033 vm_pageout_pages_needed = 0; 1034 } 1035 1036 /* 1037 * Wakeup processes that are waiting on memory. 1038 * 1039 * NOTE: vm_paging_target() is the pageout daemon's target, while 1040 * vm_page_count_target() is somewhere inbetween. We want 1041 * to wake processes up prior to the pageout daemon reaching 1042 * its target to provide some hysteresis. 1043 */ 1044 if (vm_pages_waiting) { 1045 if (!vm_page_count_target()) { 1046 /* 1047 * Plenty of pages are free, wakeup everyone. 1048 */ 1049 vm_pages_waiting = 0; 1050 wakeup(&vmstats.v_free_count); 1051 ++mycpu->gd_cnt.v_ppwakeups; 1052 } else if (!vm_page_count_min(0)) { 1053 /* 1054 * Some pages are free, wakeup someone. 1055 */ 1056 int wcount = vm_pages_waiting; 1057 if (wcount > 0) 1058 --wcount; 1059 vm_pages_waiting = wcount; 1060 wakeup_one(&vmstats.v_free_count); 1061 ++mycpu->gd_cnt.v_ppwakeups; 1062 } 1063 } 1064 } 1065 1066 /* 1067 * vm_page_free_toq: 1068 * 1069 * Returns the given page to the PQ_FREE list, disassociating it with 1070 * any VM object. 1071 * 1072 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1073 * return (the page will have been freed). No particular spl is required 1074 * on entry. 1075 * 1076 * This routine may not block. 1077 */ 1078 void 1079 vm_page_free_toq(vm_page_t m) 1080 { 1081 struct vpgqueues *pq; 1082 1083 lwkt_gettoken(&vm_token); 1084 mycpu->gd_cnt.v_tfree++; 1085 1086 KKASSERT((m->flags & PG_MAPPED) == 0); 1087 1088 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1089 kprintf( 1090 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1091 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1092 m->hold_count); 1093 if ((m->queue - m->pc) == PQ_FREE) 1094 panic("vm_page_free: freeing free page"); 1095 else 1096 panic("vm_page_free: freeing busy page"); 1097 } 1098 1099 /* 1100 * unqueue, then remove page. Note that we cannot destroy 1101 * the page here because we do not want to call the pager's 1102 * callback routine until after we've put the page on the 1103 * appropriate free queue. 1104 */ 1105 vm_page_unqueue_nowakeup(m); 1106 vm_page_remove(m); 1107 1108 /* 1109 * No further management of fictitious pages occurs beyond object 1110 * and queue removal. 1111 */ 1112 if ((m->flags & PG_FICTITIOUS) != 0) { 1113 vm_page_wakeup(m); 1114 lwkt_reltoken(&vm_token); 1115 return; 1116 } 1117 1118 m->valid = 0; 1119 vm_page_undirty(m); 1120 1121 if (m->wire_count != 0) { 1122 if (m->wire_count > 1) { 1123 panic( 1124 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1125 m->wire_count, (long)m->pindex); 1126 } 1127 panic("vm_page_free: freeing wired page"); 1128 } 1129 1130 /* 1131 * Clear the UNMANAGED flag when freeing an unmanaged page. 1132 */ 1133 if (m->flags & PG_UNMANAGED) { 1134 m->flags &= ~PG_UNMANAGED; 1135 } 1136 1137 if (m->hold_count != 0) { 1138 m->flags &= ~PG_ZERO; 1139 m->queue = PQ_HOLD; 1140 } else { 1141 m->queue = PQ_FREE + m->pc; 1142 } 1143 pq = &vm_page_queues[m->queue]; 1144 pq->lcnt++; 1145 ++(*pq->cnt); 1146 1147 /* 1148 * Put zero'd pages on the end ( where we look for zero'd pages 1149 * first ) and non-zerod pages at the head. 1150 */ 1151 if (m->flags & PG_ZERO) { 1152 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1153 ++vm_page_zero_count; 1154 } else { 1155 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1156 } 1157 vm_page_wakeup(m); 1158 vm_page_free_wakeup(); 1159 lwkt_reltoken(&vm_token); 1160 } 1161 1162 /* 1163 * vm_page_free_fromq_fast() 1164 * 1165 * Remove a non-zero page from one of the free queues; the page is removed for 1166 * zeroing, so do not issue a wakeup. 1167 * 1168 * MPUNSAFE 1169 */ 1170 vm_page_t 1171 vm_page_free_fromq_fast(void) 1172 { 1173 static int qi; 1174 vm_page_t m; 1175 int i; 1176 1177 lwkt_gettoken(&vm_token); 1178 for (i = 0; i < PQ_L2_SIZE; ++i) { 1179 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1180 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1181 if (m && (m->flags & PG_ZERO) == 0) { 1182 KKASSERT(m->busy == 0 && (m->flags & PG_BUSY) == 0); 1183 vm_page_unqueue_nowakeup(m); 1184 vm_page_busy(m); 1185 break; 1186 } 1187 m = NULL; 1188 } 1189 lwkt_reltoken(&vm_token); 1190 return (m); 1191 } 1192 1193 /* 1194 * vm_page_unmanage() 1195 * 1196 * Prevent PV management from being done on the page. The page is 1197 * removed from the paging queues as if it were wired, and as a 1198 * consequence of no longer being managed the pageout daemon will not 1199 * touch it (since there is no way to locate the pte mappings for the 1200 * page). madvise() calls that mess with the pmap will also no longer 1201 * operate on the page. 1202 * 1203 * Beyond that the page is still reasonably 'normal'. Freeing the page 1204 * will clear the flag. 1205 * 1206 * This routine is used by OBJT_PHYS objects - objects using unswappable 1207 * physical memory as backing store rather then swap-backed memory and 1208 * will eventually be extended to support 4MB unmanaged physical 1209 * mappings. 1210 * 1211 * Must be called with a critical section held. 1212 * Must be called with vm_token held. 1213 */ 1214 void 1215 vm_page_unmanage(vm_page_t m) 1216 { 1217 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1218 if ((m->flags & PG_UNMANAGED) == 0) { 1219 if (m->wire_count == 0) 1220 vm_page_unqueue(m); 1221 } 1222 vm_page_flag_set(m, PG_UNMANAGED); 1223 } 1224 1225 /* 1226 * Mark this page as wired down by yet another map, removing it from 1227 * paging queues as necessary. 1228 * 1229 * The page queues must be locked. 1230 * This routine may not block. 1231 */ 1232 void 1233 vm_page_wire(vm_page_t m) 1234 { 1235 /* 1236 * Only bump the wire statistics if the page is not already wired, 1237 * and only unqueue the page if it is on some queue (if it is unmanaged 1238 * it is already off the queues). Don't do anything with fictitious 1239 * pages because they are always wired. 1240 */ 1241 lwkt_gettoken(&vm_token); 1242 if ((m->flags & PG_FICTITIOUS) == 0) { 1243 if (m->wire_count == 0) { 1244 if ((m->flags & PG_UNMANAGED) == 0) 1245 vm_page_unqueue(m); 1246 vmstats.v_wire_count++; 1247 } 1248 m->wire_count++; 1249 KASSERT(m->wire_count != 0, 1250 ("vm_page_wire: wire_count overflow m=%p", m)); 1251 } 1252 lwkt_reltoken(&vm_token); 1253 } 1254 1255 /* 1256 * Release one wiring of this page, potentially enabling it to be paged again. 1257 * 1258 * Many pages placed on the inactive queue should actually go 1259 * into the cache, but it is difficult to figure out which. What 1260 * we do instead, if the inactive target is well met, is to put 1261 * clean pages at the head of the inactive queue instead of the tail. 1262 * This will cause them to be moved to the cache more quickly and 1263 * if not actively re-referenced, freed more quickly. If we just 1264 * stick these pages at the end of the inactive queue, heavy filesystem 1265 * meta-data accesses can cause an unnecessary paging load on memory bound 1266 * processes. This optimization causes one-time-use metadata to be 1267 * reused more quickly. 1268 * 1269 * BUT, if we are in a low-memory situation we have no choice but to 1270 * put clean pages on the cache queue. 1271 * 1272 * A number of routines use vm_page_unwire() to guarantee that the page 1273 * will go into either the inactive or active queues, and will NEVER 1274 * be placed in the cache - for example, just after dirtying a page. 1275 * dirty pages in the cache are not allowed. 1276 * 1277 * The page queues must be locked. 1278 * This routine may not block. 1279 */ 1280 void 1281 vm_page_unwire(vm_page_t m, int activate) 1282 { 1283 lwkt_gettoken(&vm_token); 1284 if (m->flags & PG_FICTITIOUS) { 1285 /* do nothing */ 1286 } else if (m->wire_count <= 0) { 1287 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1288 } else { 1289 if (--m->wire_count == 0) { 1290 --vmstats.v_wire_count; 1291 if (m->flags & PG_UNMANAGED) { 1292 ; 1293 } else if (activate) { 1294 TAILQ_INSERT_TAIL( 1295 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1296 m->queue = PQ_ACTIVE; 1297 vm_page_queues[PQ_ACTIVE].lcnt++; 1298 vmstats.v_active_count++; 1299 } else { 1300 vm_page_flag_clear(m, PG_WINATCFLS); 1301 TAILQ_INSERT_TAIL( 1302 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1303 m->queue = PQ_INACTIVE; 1304 vm_page_queues[PQ_INACTIVE].lcnt++; 1305 vmstats.v_inactive_count++; 1306 ++vm_swapcache_inactive_heuristic; 1307 } 1308 } 1309 } 1310 lwkt_reltoken(&vm_token); 1311 } 1312 1313 1314 /* 1315 * Move the specified page to the inactive queue. If the page has 1316 * any associated swap, the swap is deallocated. 1317 * 1318 * Normally athead is 0 resulting in LRU operation. athead is set 1319 * to 1 if we want this page to be 'as if it were placed in the cache', 1320 * except without unmapping it from the process address space. 1321 * 1322 * This routine may not block. 1323 * The caller must hold vm_token. 1324 */ 1325 static __inline void 1326 _vm_page_deactivate(vm_page_t m, int athead) 1327 { 1328 /* 1329 * Ignore if already inactive. 1330 */ 1331 if (m->queue == PQ_INACTIVE) 1332 return; 1333 1334 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1335 if ((m->queue - m->pc) == PQ_CACHE) 1336 mycpu->gd_cnt.v_reactivated++; 1337 vm_page_flag_clear(m, PG_WINATCFLS); 1338 vm_page_unqueue(m); 1339 if (athead) { 1340 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1341 m, pageq); 1342 } else { 1343 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1344 m, pageq); 1345 ++vm_swapcache_inactive_heuristic; 1346 } 1347 m->queue = PQ_INACTIVE; 1348 vm_page_queues[PQ_INACTIVE].lcnt++; 1349 vmstats.v_inactive_count++; 1350 } 1351 } 1352 1353 /* 1354 * Attempt to deactivate a page. 1355 * 1356 * No requirements. 1357 */ 1358 void 1359 vm_page_deactivate(vm_page_t m) 1360 { 1361 lwkt_gettoken(&vm_token); 1362 _vm_page_deactivate(m, 0); 1363 lwkt_reltoken(&vm_token); 1364 } 1365 1366 /* 1367 * Attempt to move a page to PQ_CACHE. 1368 * Returns 0 on failure, 1 on success 1369 * 1370 * No requirements. 1371 */ 1372 int 1373 vm_page_try_to_cache(vm_page_t m) 1374 { 1375 lwkt_gettoken(&vm_token); 1376 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1377 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1378 lwkt_reltoken(&vm_token); 1379 return(0); 1380 } 1381 vm_page_busy(m); 1382 vm_page_test_dirty(m); 1383 if (m->dirty) { 1384 lwkt_reltoken(&vm_token); 1385 return(0); 1386 } 1387 vm_page_cache(m); 1388 lwkt_reltoken(&vm_token); 1389 return(1); 1390 } 1391 1392 /* 1393 * Attempt to free the page. If we cannot free it, we do nothing. 1394 * 1 is returned on success, 0 on failure. 1395 * 1396 * No requirements. 1397 */ 1398 int 1399 vm_page_try_to_free(vm_page_t m) 1400 { 1401 lwkt_gettoken(&vm_token); 1402 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1403 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1404 lwkt_reltoken(&vm_token); 1405 return(0); 1406 } 1407 vm_page_test_dirty(m); 1408 if (m->dirty) { 1409 lwkt_reltoken(&vm_token); 1410 return(0); 1411 } 1412 vm_page_busy(m); 1413 vm_page_protect(m, VM_PROT_NONE); 1414 vm_page_free(m); 1415 lwkt_reltoken(&vm_token); 1416 return(1); 1417 } 1418 1419 /* 1420 * vm_page_cache 1421 * 1422 * Put the specified page onto the page cache queue (if appropriate). 1423 * 1424 * The caller must hold vm_token. 1425 * This routine may not block. 1426 * The page must be busy, and this routine will release the busy and 1427 * possibly even free the page. 1428 */ 1429 void 1430 vm_page_cache(vm_page_t m) 1431 { 1432 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1433 1434 if ((m->flags & PG_UNMANAGED) || m->busy || 1435 m->wire_count || m->hold_count) { 1436 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1437 vm_page_wakeup(m); 1438 return; 1439 } 1440 1441 /* 1442 * Already in the cache (and thus not mapped) 1443 */ 1444 if ((m->queue - m->pc) == PQ_CACHE) { 1445 KKASSERT((m->flags & PG_MAPPED) == 0); 1446 vm_page_wakeup(m); 1447 return; 1448 } 1449 1450 /* 1451 * Caller is required to test m->dirty, but note that the act of 1452 * removing the page from its maps can cause it to become dirty 1453 * on an SMP system due to another cpu running in usermode. 1454 */ 1455 if (m->dirty) { 1456 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1457 (long)m->pindex); 1458 } 1459 1460 /* 1461 * Remove all pmaps and indicate that the page is not 1462 * writeable or mapped. Our vm_page_protect() call may 1463 * have blocked (especially w/ VM_PROT_NONE), so recheck 1464 * everything. 1465 */ 1466 vm_page_protect(m, VM_PROT_NONE); 1467 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy || 1468 m->wire_count || m->hold_count) { 1469 vm_page_wakeup(m); 1470 } else if (m->dirty) { 1471 vm_page_deactivate(m); 1472 vm_page_wakeup(m); 1473 } else { 1474 vm_page_unqueue_nowakeup(m); 1475 m->queue = PQ_CACHE + m->pc; 1476 vm_page_queues[m->queue].lcnt++; 1477 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1478 vmstats.v_cache_count++; 1479 vm_page_wakeup(m); 1480 vm_page_free_wakeup(); 1481 } 1482 } 1483 1484 /* 1485 * vm_page_dontneed() 1486 * 1487 * Cache, deactivate, or do nothing as appropriate. This routine 1488 * is typically used by madvise() MADV_DONTNEED. 1489 * 1490 * Generally speaking we want to move the page into the cache so 1491 * it gets reused quickly. However, this can result in a silly syndrome 1492 * due to the page recycling too quickly. Small objects will not be 1493 * fully cached. On the otherhand, if we move the page to the inactive 1494 * queue we wind up with a problem whereby very large objects 1495 * unnecessarily blow away our inactive and cache queues. 1496 * 1497 * The solution is to move the pages based on a fixed weighting. We 1498 * either leave them alone, deactivate them, or move them to the cache, 1499 * where moving them to the cache has the highest weighting. 1500 * By forcing some pages into other queues we eventually force the 1501 * system to balance the queues, potentially recovering other unrelated 1502 * space from active. The idea is to not force this to happen too 1503 * often. 1504 * 1505 * No requirements. 1506 */ 1507 void 1508 vm_page_dontneed(vm_page_t m) 1509 { 1510 static int dnweight; 1511 int dnw; 1512 int head; 1513 1514 dnw = ++dnweight; 1515 1516 /* 1517 * occassionally leave the page alone 1518 */ 1519 lwkt_gettoken(&vm_token); 1520 if ((dnw & 0x01F0) == 0 || 1521 m->queue == PQ_INACTIVE || 1522 m->queue - m->pc == PQ_CACHE 1523 ) { 1524 if (m->act_count >= ACT_INIT) 1525 --m->act_count; 1526 lwkt_reltoken(&vm_token); 1527 return; 1528 } 1529 1530 if (m->dirty == 0) 1531 vm_page_test_dirty(m); 1532 1533 if (m->dirty || (dnw & 0x0070) == 0) { 1534 /* 1535 * Deactivate the page 3 times out of 32. 1536 */ 1537 head = 0; 1538 } else { 1539 /* 1540 * Cache the page 28 times out of every 32. Note that 1541 * the page is deactivated instead of cached, but placed 1542 * at the head of the queue instead of the tail. 1543 */ 1544 head = 1; 1545 } 1546 _vm_page_deactivate(m, head); 1547 lwkt_reltoken(&vm_token); 1548 } 1549 1550 /* 1551 * Grab a page, blocking if it is busy and allocating a page if necessary. 1552 * A busy page is returned or NULL. 1553 * 1554 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1555 * If VM_ALLOC_RETRY is not specified 1556 * 1557 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1558 * always returned if we had blocked. 1559 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1560 * This routine may not be called from an interrupt. 1561 * The returned page may not be entirely valid. 1562 * 1563 * This routine may be called from mainline code without spl protection and 1564 * be guarenteed a busied page associated with the object at the specified 1565 * index. 1566 * 1567 * No requirements. 1568 */ 1569 vm_page_t 1570 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1571 { 1572 vm_page_t m; 1573 int generation; 1574 1575 KKASSERT(allocflags & 1576 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1577 lwkt_gettoken(&vm_token); 1578 retrylookup: 1579 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1580 if (m->busy || (m->flags & PG_BUSY)) { 1581 generation = object->generation; 1582 1583 while ((object->generation == generation) && 1584 (m->busy || (m->flags & PG_BUSY))) { 1585 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1586 tsleep(m, 0, "pgrbwt", 0); 1587 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1588 m = NULL; 1589 goto done; 1590 } 1591 } 1592 goto retrylookup; 1593 } else { 1594 vm_page_busy(m); 1595 goto done; 1596 } 1597 } 1598 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1599 if (m == NULL) { 1600 vm_wait(0); 1601 if ((allocflags & VM_ALLOC_RETRY) == 0) 1602 goto done; 1603 goto retrylookup; 1604 } 1605 done: 1606 lwkt_reltoken(&vm_token); 1607 return(m); 1608 } 1609 1610 /* 1611 * Mapping function for valid bits or for dirty bits in 1612 * a page. May not block. 1613 * 1614 * Inputs are required to range within a page. 1615 * 1616 * No requirements. 1617 * Non blocking. 1618 */ 1619 int 1620 vm_page_bits(int base, int size) 1621 { 1622 int first_bit; 1623 int last_bit; 1624 1625 KASSERT( 1626 base + size <= PAGE_SIZE, 1627 ("vm_page_bits: illegal base/size %d/%d", base, size) 1628 ); 1629 1630 if (size == 0) /* handle degenerate case */ 1631 return(0); 1632 1633 first_bit = base >> DEV_BSHIFT; 1634 last_bit = (base + size - 1) >> DEV_BSHIFT; 1635 1636 return ((2 << last_bit) - (1 << first_bit)); 1637 } 1638 1639 /* 1640 * Sets portions of a page valid and clean. The arguments are expected 1641 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1642 * of any partial chunks touched by the range. The invalid portion of 1643 * such chunks will be zero'd. 1644 * 1645 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1646 * align base to DEV_BSIZE so as not to mark clean a partially 1647 * truncated device block. Otherwise the dirty page status might be 1648 * lost. 1649 * 1650 * This routine may not block. 1651 * 1652 * (base + size) must be less then or equal to PAGE_SIZE. 1653 */ 1654 static void 1655 _vm_page_zero_valid(vm_page_t m, int base, int size) 1656 { 1657 int frag; 1658 int endoff; 1659 1660 if (size == 0) /* handle degenerate case */ 1661 return; 1662 1663 /* 1664 * If the base is not DEV_BSIZE aligned and the valid 1665 * bit is clear, we have to zero out a portion of the 1666 * first block. 1667 */ 1668 1669 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1670 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1671 ) { 1672 pmap_zero_page_area( 1673 VM_PAGE_TO_PHYS(m), 1674 frag, 1675 base - frag 1676 ); 1677 } 1678 1679 /* 1680 * If the ending offset is not DEV_BSIZE aligned and the 1681 * valid bit is clear, we have to zero out a portion of 1682 * the last block. 1683 */ 1684 1685 endoff = base + size; 1686 1687 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1688 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1689 ) { 1690 pmap_zero_page_area( 1691 VM_PAGE_TO_PHYS(m), 1692 endoff, 1693 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1694 ); 1695 } 1696 } 1697 1698 /* 1699 * Set valid, clear dirty bits. If validating the entire 1700 * page we can safely clear the pmap modify bit. We also 1701 * use this opportunity to clear the PG_NOSYNC flag. If a process 1702 * takes a write fault on a MAP_NOSYNC memory area the flag will 1703 * be set again. 1704 * 1705 * We set valid bits inclusive of any overlap, but we can only 1706 * clear dirty bits for DEV_BSIZE chunks that are fully within 1707 * the range. 1708 * 1709 * Page must be busied? 1710 * No other requirements. 1711 */ 1712 void 1713 vm_page_set_valid(vm_page_t m, int base, int size) 1714 { 1715 _vm_page_zero_valid(m, base, size); 1716 m->valid |= vm_page_bits(base, size); 1717 } 1718 1719 1720 /* 1721 * Set valid bits and clear dirty bits. 1722 * 1723 * NOTE: This function does not clear the pmap modified bit. 1724 * Also note that e.g. NFS may use a byte-granular base 1725 * and size. 1726 * 1727 * WARNING: Page must be busied? But vfs_clean_one_page() will call 1728 * this without necessarily busying the page (via bdwrite()). 1729 * So for now vm_token must also be held. 1730 * 1731 * No other requirements. 1732 */ 1733 void 1734 vm_page_set_validclean(vm_page_t m, int base, int size) 1735 { 1736 int pagebits; 1737 1738 _vm_page_zero_valid(m, base, size); 1739 pagebits = vm_page_bits(base, size); 1740 m->valid |= pagebits; 1741 m->dirty &= ~pagebits; 1742 if (base == 0 && size == PAGE_SIZE) { 1743 /*pmap_clear_modify(m);*/ 1744 vm_page_flag_clear(m, PG_NOSYNC); 1745 } 1746 } 1747 1748 /* 1749 * Set valid & dirty. Used by buwrite() 1750 * 1751 * WARNING: Page must be busied? But vfs_dirty_one_page() will 1752 * call this function in buwrite() so for now vm_token must 1753 * be held. 1754 * 1755 * No other requirements. 1756 */ 1757 void 1758 vm_page_set_validdirty(vm_page_t m, int base, int size) 1759 { 1760 int pagebits; 1761 1762 pagebits = vm_page_bits(base, size); 1763 m->valid |= pagebits; 1764 m->dirty |= pagebits; 1765 if (m->object) 1766 vm_object_set_writeable_dirty(m->object); 1767 } 1768 1769 /* 1770 * Clear dirty bits. 1771 * 1772 * NOTE: This function does not clear the pmap modified bit. 1773 * Also note that e.g. NFS may use a byte-granular base 1774 * and size. 1775 * 1776 * Page must be busied? 1777 * No other requirements. 1778 */ 1779 void 1780 vm_page_clear_dirty(vm_page_t m, int base, int size) 1781 { 1782 m->dirty &= ~vm_page_bits(base, size); 1783 if (base == 0 && size == PAGE_SIZE) { 1784 /*pmap_clear_modify(m);*/ 1785 vm_page_flag_clear(m, PG_NOSYNC); 1786 } 1787 } 1788 1789 /* 1790 * Make the page all-dirty. 1791 * 1792 * Also make sure the related object and vnode reflect the fact that the 1793 * object may now contain a dirty page. 1794 * 1795 * Page must be busied? 1796 * No other requirements. 1797 */ 1798 void 1799 vm_page_dirty(vm_page_t m) 1800 { 1801 #ifdef INVARIANTS 1802 int pqtype = m->queue - m->pc; 1803 #endif 1804 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1805 ("vm_page_dirty: page in free/cache queue!")); 1806 if (m->dirty != VM_PAGE_BITS_ALL) { 1807 m->dirty = VM_PAGE_BITS_ALL; 1808 if (m->object) 1809 vm_object_set_writeable_dirty(m->object); 1810 } 1811 } 1812 1813 /* 1814 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1815 * valid and dirty bits for the effected areas are cleared. 1816 * 1817 * Page must be busied? 1818 * Does not block. 1819 * No other requirements. 1820 */ 1821 void 1822 vm_page_set_invalid(vm_page_t m, int base, int size) 1823 { 1824 int bits; 1825 1826 bits = vm_page_bits(base, size); 1827 m->valid &= ~bits; 1828 m->dirty &= ~bits; 1829 m->object->generation++; 1830 } 1831 1832 /* 1833 * The kernel assumes that the invalid portions of a page contain 1834 * garbage, but such pages can be mapped into memory by user code. 1835 * When this occurs, we must zero out the non-valid portions of the 1836 * page so user code sees what it expects. 1837 * 1838 * Pages are most often semi-valid when the end of a file is mapped 1839 * into memory and the file's size is not page aligned. 1840 * 1841 * Page must be busied? 1842 * No other requirements. 1843 */ 1844 void 1845 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1846 { 1847 int b; 1848 int i; 1849 1850 /* 1851 * Scan the valid bits looking for invalid sections that 1852 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1853 * valid bit may be set ) have already been zerod by 1854 * vm_page_set_validclean(). 1855 */ 1856 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1857 if (i == (PAGE_SIZE / DEV_BSIZE) || 1858 (m->valid & (1 << i)) 1859 ) { 1860 if (i > b) { 1861 pmap_zero_page_area( 1862 VM_PAGE_TO_PHYS(m), 1863 b << DEV_BSHIFT, 1864 (i - b) << DEV_BSHIFT 1865 ); 1866 } 1867 b = i + 1; 1868 } 1869 } 1870 1871 /* 1872 * setvalid is TRUE when we can safely set the zero'd areas 1873 * as being valid. We can do this if there are no cache consistency 1874 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1875 */ 1876 if (setvalid) 1877 m->valid = VM_PAGE_BITS_ALL; 1878 } 1879 1880 /* 1881 * Is a (partial) page valid? Note that the case where size == 0 1882 * will return FALSE in the degenerate case where the page is entirely 1883 * invalid, and TRUE otherwise. 1884 * 1885 * Does not block. 1886 * No other requirements. 1887 */ 1888 int 1889 vm_page_is_valid(vm_page_t m, int base, int size) 1890 { 1891 int bits = vm_page_bits(base, size); 1892 1893 if (m->valid && ((m->valid & bits) == bits)) 1894 return 1; 1895 else 1896 return 0; 1897 } 1898 1899 /* 1900 * update dirty bits from pmap/mmu. May not block. 1901 * 1902 * Caller must hold vm_token if non-blocking operation desired. 1903 * No other requirements. 1904 */ 1905 void 1906 vm_page_test_dirty(vm_page_t m) 1907 { 1908 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1909 vm_page_dirty(m); 1910 } 1911 } 1912 1913 /* 1914 * Register an action, associating it with its vm_page 1915 */ 1916 void 1917 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 1918 { 1919 struct vm_page_action_list *list; 1920 int hv; 1921 1922 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1923 list = &action_list[hv]; 1924 1925 lwkt_gettoken(&vm_token); 1926 vm_page_flag_set(action->m, PG_ACTIONLIST); 1927 action->event = event; 1928 LIST_INSERT_HEAD(list, action, entry); 1929 lwkt_reltoken(&vm_token); 1930 } 1931 1932 /* 1933 * Unregister an action, disassociating it from its related vm_page 1934 */ 1935 void 1936 vm_page_unregister_action(vm_page_action_t action) 1937 { 1938 struct vm_page_action_list *list; 1939 int hv; 1940 1941 lwkt_gettoken(&vm_token); 1942 if (action->event != VMEVENT_NONE) { 1943 action->event = VMEVENT_NONE; 1944 LIST_REMOVE(action, entry); 1945 1946 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1947 list = &action_list[hv]; 1948 if (LIST_EMPTY(list)) 1949 vm_page_flag_clear(action->m, PG_ACTIONLIST); 1950 } 1951 lwkt_reltoken(&vm_token); 1952 } 1953 1954 /* 1955 * Issue an event on a VM page. Corresponding action structures are 1956 * removed from the page's list and called. 1957 * 1958 * If the vm_page has no more pending action events we clear its 1959 * PG_ACTIONLIST flag. 1960 */ 1961 void 1962 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1963 { 1964 struct vm_page_action_list *list; 1965 struct vm_page_action *scan; 1966 struct vm_page_action *next; 1967 int hv; 1968 int all; 1969 1970 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 1971 list = &action_list[hv]; 1972 all = 1; 1973 1974 lwkt_gettoken(&vm_token); 1975 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 1976 if (scan->m == m) { 1977 if (scan->event == event) { 1978 scan->event = VMEVENT_NONE; 1979 LIST_REMOVE(scan, entry); 1980 scan->func(m, scan); 1981 /* XXX */ 1982 } else { 1983 all = 0; 1984 } 1985 } 1986 } 1987 if (all) 1988 vm_page_flag_clear(m, PG_ACTIONLIST); 1989 lwkt_reltoken(&vm_token); 1990 } 1991 1992 1993 #include "opt_ddb.h" 1994 #ifdef DDB 1995 #include <sys/kernel.h> 1996 1997 #include <ddb/ddb.h> 1998 1999 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2000 { 2001 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 2002 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 2003 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 2004 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 2005 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 2006 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 2007 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 2008 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 2009 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 2010 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 2011 } 2012 2013 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2014 { 2015 int i; 2016 db_printf("PQ_FREE:"); 2017 for(i=0;i<PQ_L2_SIZE;i++) { 2018 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 2019 } 2020 db_printf("\n"); 2021 2022 db_printf("PQ_CACHE:"); 2023 for(i=0;i<PQ_L2_SIZE;i++) { 2024 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 2025 } 2026 db_printf("\n"); 2027 2028 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2029 vm_page_queues[PQ_ACTIVE].lcnt, 2030 vm_page_queues[PQ_INACTIVE].lcnt); 2031 } 2032 #endif /* DDB */ 2033