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 vm_page_deactivate(m); 702 continue; 703 } 704 return m; 705 } 706 /* not reached */ 707 } 708 709 /* 710 * Find a free or zero page, with specified preference. We attempt to 711 * inline the nominal case and fall back to _vm_page_select_free() 712 * otherwise. 713 * 714 * This routine must be called with a critical section held. 715 * This routine may not block. 716 */ 717 static __inline vm_page_t 718 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 719 { 720 vm_page_t m; 721 722 m = _vm_page_list_find( 723 PQ_FREE, 724 (pindex + object->pg_color) & PQ_L2_MASK, 725 prefer_zero 726 ); 727 return(m); 728 } 729 730 /* 731 * vm_page_alloc() 732 * 733 * Allocate and return a memory cell associated with this VM object/offset 734 * pair. 735 * 736 * page_req classes: 737 * 738 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 739 * VM_ALLOC_QUICK like normal but cannot use cache 740 * VM_ALLOC_SYSTEM greater free drain 741 * VM_ALLOC_INTERRUPT allow free list to be completely drained 742 * VM_ALLOC_ZERO advisory request for pre-zero'd page 743 * 744 * The object must be locked. 745 * This routine may not block. 746 * The returned page will be marked PG_BUSY 747 * 748 * Additional special handling is required when called from an interrupt 749 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 750 * in this case. 751 */ 752 vm_page_t 753 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 754 { 755 vm_page_t m = NULL; 756 757 lwkt_gettoken(&vm_token); 758 759 KKASSERT(object != NULL); 760 KASSERT(!vm_page_lookup(object, pindex), 761 ("vm_page_alloc: page already allocated")); 762 KKASSERT(page_req & 763 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 764 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 765 766 /* 767 * Certain system threads (pageout daemon, buf_daemon's) are 768 * allowed to eat deeper into the free page list. 769 */ 770 if (curthread->td_flags & TDF_SYSTHREAD) 771 page_req |= VM_ALLOC_SYSTEM; 772 773 loop: 774 if (vmstats.v_free_count > vmstats.v_free_reserved || 775 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 776 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 777 vmstats.v_free_count > vmstats.v_interrupt_free_min) 778 ) { 779 /* 780 * The free queue has sufficient free pages to take one out. 781 */ 782 if (page_req & VM_ALLOC_ZERO) 783 m = vm_page_select_free(object, pindex, TRUE); 784 else 785 m = vm_page_select_free(object, pindex, FALSE); 786 } else if (page_req & VM_ALLOC_NORMAL) { 787 /* 788 * Allocatable from the cache (non-interrupt only). On 789 * success, we must free the page and try again, thus 790 * ensuring that vmstats.v_*_free_min counters are replenished. 791 */ 792 #ifdef INVARIANTS 793 if (curthread->td_preempted) { 794 kprintf("vm_page_alloc(): warning, attempt to allocate" 795 " cache page from preempting interrupt\n"); 796 m = NULL; 797 } else { 798 m = vm_page_select_cache(object, pindex); 799 } 800 #else 801 m = vm_page_select_cache(object, pindex); 802 #endif 803 /* 804 * On success move the page into the free queue and loop. 805 */ 806 if (m != NULL) { 807 KASSERT(m->dirty == 0, 808 ("Found dirty cache page %p", m)); 809 vm_page_busy(m); 810 vm_page_protect(m, VM_PROT_NONE); 811 vm_page_free(m); 812 goto loop; 813 } 814 815 /* 816 * On failure return NULL 817 */ 818 lwkt_reltoken(&vm_token); 819 #if defined(DIAGNOSTIC) 820 if (vmstats.v_cache_count > 0) 821 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 822 #endif 823 vm_pageout_deficit++; 824 pagedaemon_wakeup(); 825 return (NULL); 826 } else { 827 /* 828 * No pages available, wakeup the pageout daemon and give up. 829 */ 830 lwkt_reltoken(&vm_token); 831 vm_pageout_deficit++; 832 pagedaemon_wakeup(); 833 return (NULL); 834 } 835 836 /* 837 * Good page found. The page has not yet been busied. We are in 838 * a critical section. 839 */ 840 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 841 KASSERT(m->dirty == 0, 842 ("vm_page_alloc: free/cache page %p was dirty", m)); 843 844 /* 845 * Remove from free queue 846 */ 847 vm_page_unqueue_nowakeup(m); 848 849 /* 850 * Initialize structure. Only the PG_ZERO flag is inherited. Set 851 * the page PG_BUSY 852 */ 853 if (m->flags & PG_ZERO) { 854 vm_page_zero_count--; 855 m->flags = PG_ZERO | PG_BUSY; 856 } else { 857 m->flags = PG_BUSY; 858 } 859 m->wire_count = 0; 860 m->hold_count = 0; 861 m->act_count = 0; 862 m->busy = 0; 863 m->valid = 0; 864 865 /* 866 * vm_page_insert() is safe while holding vm_token. Note also that 867 * inserting a page here does not insert it into the pmap (which 868 * could cause us to block allocating memory). We cannot block 869 * anywhere. 870 */ 871 vm_page_insert(m, object, pindex); 872 873 /* 874 * Don't wakeup too often - wakeup the pageout daemon when 875 * we would be nearly out of memory. 876 */ 877 pagedaemon_wakeup(); 878 879 lwkt_reltoken(&vm_token); 880 881 /* 882 * A PG_BUSY page is returned. 883 */ 884 return (m); 885 } 886 887 /* 888 * Wait for sufficient free memory for nominal heavy memory use kernel 889 * operations. 890 */ 891 void 892 vm_wait_nominal(void) 893 { 894 while (vm_page_count_min(0)) 895 vm_wait(0); 896 } 897 898 /* 899 * Test if vm_wait_nominal() would block. 900 */ 901 int 902 vm_test_nominal(void) 903 { 904 if (vm_page_count_min(0)) 905 return(1); 906 return(0); 907 } 908 909 /* 910 * Block until free pages are available for allocation, called in various 911 * places before memory allocations. 912 * 913 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 914 * more generous then that. 915 */ 916 void 917 vm_wait(int timo) 918 { 919 /* 920 * never wait forever 921 */ 922 if (timo == 0) 923 timo = hz; 924 lwkt_gettoken(&vm_token); 925 926 if (curthread == pagethread) { 927 /* 928 * The pageout daemon itself needs pages, this is bad. 929 */ 930 if (vm_page_count_min(0)) { 931 vm_pageout_pages_needed = 1; 932 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 933 } 934 } else { 935 /* 936 * Wakeup the pageout daemon if necessary and wait. 937 */ 938 if (vm_page_count_target()) { 939 if (vm_pages_needed == 0) { 940 vm_pages_needed = 1; 941 wakeup(&vm_pages_needed); 942 } 943 ++vm_pages_waiting; /* SMP race ok */ 944 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 945 } 946 } 947 lwkt_reltoken(&vm_token); 948 } 949 950 /* 951 * Block until free pages are available for allocation 952 * 953 * Called only from vm_fault so that processes page faulting can be 954 * easily tracked. 955 */ 956 void 957 vm_waitpfault(void) 958 { 959 /* 960 * Wakeup the pageout daemon if necessary and wait. 961 */ 962 if (vm_page_count_target()) { 963 lwkt_gettoken(&vm_token); 964 if (vm_page_count_target()) { 965 if (vm_pages_needed == 0) { 966 vm_pages_needed = 1; 967 wakeup(&vm_pages_needed); 968 } 969 ++vm_pages_waiting; /* SMP race ok */ 970 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 971 } 972 lwkt_reltoken(&vm_token); 973 } 974 } 975 976 /* 977 * Put the specified page on the active list (if appropriate). Ensure 978 * that act_count is at least ACT_INIT but do not otherwise mess with it. 979 * 980 * The page queues must be locked. 981 * This routine may not block. 982 */ 983 void 984 vm_page_activate(vm_page_t m) 985 { 986 lwkt_gettoken(&vm_token); 987 if (m->queue != PQ_ACTIVE) { 988 if ((m->queue - m->pc) == PQ_CACHE) 989 mycpu->gd_cnt.v_reactivated++; 990 991 vm_page_unqueue(m); 992 993 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 994 m->queue = PQ_ACTIVE; 995 vm_page_queues[PQ_ACTIVE].lcnt++; 996 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 997 m, pageq); 998 if (m->act_count < ACT_INIT) 999 m->act_count = ACT_INIT; 1000 vmstats.v_active_count++; 1001 } 1002 } else { 1003 if (m->act_count < ACT_INIT) 1004 m->act_count = ACT_INIT; 1005 } 1006 lwkt_reltoken(&vm_token); 1007 } 1008 1009 /* 1010 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1011 * routine is called when a page has been added to the cache or free 1012 * queues. 1013 * 1014 * This routine may not block. 1015 * This routine must be called at splvm() 1016 */ 1017 static __inline void 1018 vm_page_free_wakeup(void) 1019 { 1020 /* 1021 * If the pageout daemon itself needs pages, then tell it that 1022 * there are some free. 1023 */ 1024 if (vm_pageout_pages_needed && 1025 vmstats.v_cache_count + vmstats.v_free_count >= 1026 vmstats.v_pageout_free_min 1027 ) { 1028 wakeup(&vm_pageout_pages_needed); 1029 vm_pageout_pages_needed = 0; 1030 } 1031 1032 /* 1033 * Wakeup processes that are waiting on memory. 1034 * 1035 * NOTE: vm_paging_target() is the pageout daemon's target, while 1036 * vm_page_count_target() is somewhere inbetween. We want 1037 * to wake processes up prior to the pageout daemon reaching 1038 * its target to provide some hysteresis. 1039 */ 1040 if (vm_pages_waiting) { 1041 if (!vm_page_count_target()) { 1042 /* 1043 * Plenty of pages are free, wakeup everyone. 1044 */ 1045 vm_pages_waiting = 0; 1046 wakeup(&vmstats.v_free_count); 1047 ++mycpu->gd_cnt.v_ppwakeups; 1048 } else if (!vm_page_count_min(0)) { 1049 /* 1050 * Some pages are free, wakeup someone. 1051 */ 1052 int wcount = vm_pages_waiting; 1053 if (wcount > 0) 1054 --wcount; 1055 vm_pages_waiting = wcount; 1056 wakeup_one(&vmstats.v_free_count); 1057 ++mycpu->gd_cnt.v_ppwakeups; 1058 } 1059 } 1060 } 1061 1062 /* 1063 * vm_page_free_toq: 1064 * 1065 * Returns the given page to the PQ_FREE list, disassociating it with 1066 * any VM object. 1067 * 1068 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1069 * return (the page will have been freed). No particular spl is required 1070 * on entry. 1071 * 1072 * This routine may not block. 1073 */ 1074 void 1075 vm_page_free_toq(vm_page_t m) 1076 { 1077 struct vpgqueues *pq; 1078 1079 lwkt_gettoken(&vm_token); 1080 mycpu->gd_cnt.v_tfree++; 1081 1082 KKASSERT((m->flags & PG_MAPPED) == 0); 1083 1084 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1085 kprintf( 1086 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1087 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1088 m->hold_count); 1089 if ((m->queue - m->pc) == PQ_FREE) 1090 panic("vm_page_free: freeing free page"); 1091 else 1092 panic("vm_page_free: freeing busy page"); 1093 } 1094 1095 /* 1096 * unqueue, then remove page. Note that we cannot destroy 1097 * the page here because we do not want to call the pager's 1098 * callback routine until after we've put the page on the 1099 * appropriate free queue. 1100 */ 1101 vm_page_unqueue_nowakeup(m); 1102 vm_page_remove(m); 1103 1104 /* 1105 * No further management of fictitious pages occurs beyond object 1106 * and queue removal. 1107 */ 1108 if ((m->flags & PG_FICTITIOUS) != 0) { 1109 vm_page_wakeup(m); 1110 lwkt_reltoken(&vm_token); 1111 return; 1112 } 1113 1114 m->valid = 0; 1115 vm_page_undirty(m); 1116 1117 if (m->wire_count != 0) { 1118 if (m->wire_count > 1) { 1119 panic( 1120 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1121 m->wire_count, (long)m->pindex); 1122 } 1123 panic("vm_page_free: freeing wired page"); 1124 } 1125 1126 /* 1127 * Clear the UNMANAGED flag when freeing an unmanaged page. 1128 */ 1129 if (m->flags & PG_UNMANAGED) { 1130 m->flags &= ~PG_UNMANAGED; 1131 } 1132 1133 if (m->hold_count != 0) { 1134 m->flags &= ~PG_ZERO; 1135 m->queue = PQ_HOLD; 1136 } else { 1137 m->queue = PQ_FREE + m->pc; 1138 } 1139 pq = &vm_page_queues[m->queue]; 1140 pq->lcnt++; 1141 ++(*pq->cnt); 1142 1143 /* 1144 * Put zero'd pages on the end ( where we look for zero'd pages 1145 * first ) and non-zerod pages at the head. 1146 */ 1147 if (m->flags & PG_ZERO) { 1148 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1149 ++vm_page_zero_count; 1150 } else { 1151 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1152 } 1153 vm_page_wakeup(m); 1154 vm_page_free_wakeup(); 1155 lwkt_reltoken(&vm_token); 1156 } 1157 1158 /* 1159 * vm_page_free_fromq_fast() 1160 * 1161 * Remove a non-zero page from one of the free queues; the page is removed for 1162 * zeroing, so do not issue a wakeup. 1163 * 1164 * MPUNSAFE 1165 */ 1166 vm_page_t 1167 vm_page_free_fromq_fast(void) 1168 { 1169 static int qi; 1170 vm_page_t m; 1171 int i; 1172 1173 lwkt_gettoken(&vm_token); 1174 for (i = 0; i < PQ_L2_SIZE; ++i) { 1175 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1176 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1177 if (m && (m->flags & PG_ZERO) == 0) { 1178 KKASSERT(m->busy == 0 && (m->flags & PG_BUSY) == 0); 1179 vm_page_unqueue_nowakeup(m); 1180 vm_page_busy(m); 1181 break; 1182 } 1183 m = NULL; 1184 } 1185 lwkt_reltoken(&vm_token); 1186 return (m); 1187 } 1188 1189 /* 1190 * vm_page_unmanage() 1191 * 1192 * Prevent PV management from being done on the page. The page is 1193 * removed from the paging queues as if it were wired, and as a 1194 * consequence of no longer being managed the pageout daemon will not 1195 * touch it (since there is no way to locate the pte mappings for the 1196 * page). madvise() calls that mess with the pmap will also no longer 1197 * operate on the page. 1198 * 1199 * Beyond that the page is still reasonably 'normal'. Freeing the page 1200 * will clear the flag. 1201 * 1202 * This routine is used by OBJT_PHYS objects - objects using unswappable 1203 * physical memory as backing store rather then swap-backed memory and 1204 * will eventually be extended to support 4MB unmanaged physical 1205 * mappings. 1206 * 1207 * Must be called with a critical section held. 1208 * Must be called with vm_token held. 1209 */ 1210 void 1211 vm_page_unmanage(vm_page_t m) 1212 { 1213 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1214 if ((m->flags & PG_UNMANAGED) == 0) { 1215 if (m->wire_count == 0) 1216 vm_page_unqueue(m); 1217 } 1218 vm_page_flag_set(m, PG_UNMANAGED); 1219 } 1220 1221 /* 1222 * Mark this page as wired down by yet another map, removing it from 1223 * paging queues as necessary. 1224 * 1225 * The page queues must be locked. 1226 * This routine may not block. 1227 */ 1228 void 1229 vm_page_wire(vm_page_t m) 1230 { 1231 /* 1232 * Only bump the wire statistics if the page is not already wired, 1233 * and only unqueue the page if it is on some queue (if it is unmanaged 1234 * it is already off the queues). Don't do anything with fictitious 1235 * pages because they are always wired. 1236 */ 1237 lwkt_gettoken(&vm_token); 1238 if ((m->flags & PG_FICTITIOUS) == 0) { 1239 if (m->wire_count == 0) { 1240 if ((m->flags & PG_UNMANAGED) == 0) 1241 vm_page_unqueue(m); 1242 vmstats.v_wire_count++; 1243 } 1244 m->wire_count++; 1245 KASSERT(m->wire_count != 0, 1246 ("vm_page_wire: wire_count overflow m=%p", m)); 1247 } 1248 lwkt_reltoken(&vm_token); 1249 } 1250 1251 /* 1252 * Release one wiring of this page, potentially enabling it to be paged again. 1253 * 1254 * Many pages placed on the inactive queue should actually go 1255 * into the cache, but it is difficult to figure out which. What 1256 * we do instead, if the inactive target is well met, is to put 1257 * clean pages at the head of the inactive queue instead of the tail. 1258 * This will cause them to be moved to the cache more quickly and 1259 * if not actively re-referenced, freed more quickly. If we just 1260 * stick these pages at the end of the inactive queue, heavy filesystem 1261 * meta-data accesses can cause an unnecessary paging load on memory bound 1262 * processes. This optimization causes one-time-use metadata to be 1263 * reused more quickly. 1264 * 1265 * BUT, if we are in a low-memory situation we have no choice but to 1266 * put clean pages on the cache queue. 1267 * 1268 * A number of routines use vm_page_unwire() to guarantee that the page 1269 * will go into either the inactive or active queues, and will NEVER 1270 * be placed in the cache - for example, just after dirtying a page. 1271 * dirty pages in the cache are not allowed. 1272 * 1273 * The page queues must be locked. 1274 * This routine may not block. 1275 */ 1276 void 1277 vm_page_unwire(vm_page_t m, int activate) 1278 { 1279 lwkt_gettoken(&vm_token); 1280 if (m->flags & PG_FICTITIOUS) { 1281 /* do nothing */ 1282 } else if (m->wire_count <= 0) { 1283 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1284 } else { 1285 if (--m->wire_count == 0) { 1286 --vmstats.v_wire_count; 1287 if (m->flags & PG_UNMANAGED) { 1288 ; 1289 } else if (activate) { 1290 TAILQ_INSERT_TAIL( 1291 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1292 m->queue = PQ_ACTIVE; 1293 vm_page_queues[PQ_ACTIVE].lcnt++; 1294 vmstats.v_active_count++; 1295 } else { 1296 vm_page_flag_clear(m, PG_WINATCFLS); 1297 TAILQ_INSERT_TAIL( 1298 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1299 m->queue = PQ_INACTIVE; 1300 vm_page_queues[PQ_INACTIVE].lcnt++; 1301 vmstats.v_inactive_count++; 1302 ++vm_swapcache_inactive_heuristic; 1303 } 1304 } 1305 } 1306 lwkt_reltoken(&vm_token); 1307 } 1308 1309 1310 /* 1311 * Move the specified page to the inactive queue. If the page has 1312 * any associated swap, the swap is deallocated. 1313 * 1314 * Normally athead is 0 resulting in LRU operation. athead is set 1315 * to 1 if we want this page to be 'as if it were placed in the cache', 1316 * except without unmapping it from the process address space. 1317 * 1318 * This routine may not block. 1319 * The caller must hold vm_token. 1320 */ 1321 static __inline void 1322 _vm_page_deactivate(vm_page_t m, int athead) 1323 { 1324 /* 1325 * Ignore if already inactive. 1326 */ 1327 if (m->queue == PQ_INACTIVE) 1328 return; 1329 1330 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1331 if ((m->queue - m->pc) == PQ_CACHE) 1332 mycpu->gd_cnt.v_reactivated++; 1333 vm_page_flag_clear(m, PG_WINATCFLS); 1334 vm_page_unqueue(m); 1335 if (athead) { 1336 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1337 m, pageq); 1338 } else { 1339 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1340 m, pageq); 1341 ++vm_swapcache_inactive_heuristic; 1342 } 1343 m->queue = PQ_INACTIVE; 1344 vm_page_queues[PQ_INACTIVE].lcnt++; 1345 vmstats.v_inactive_count++; 1346 } 1347 } 1348 1349 /* 1350 * Attempt to deactivate a page. 1351 * 1352 * No requirements. 1353 */ 1354 void 1355 vm_page_deactivate(vm_page_t m) 1356 { 1357 lwkt_gettoken(&vm_token); 1358 _vm_page_deactivate(m, 0); 1359 lwkt_reltoken(&vm_token); 1360 } 1361 1362 /* 1363 * Attempt to move a page to PQ_CACHE. 1364 * Returns 0 on failure, 1 on success 1365 * 1366 * No requirements. 1367 */ 1368 int 1369 vm_page_try_to_cache(vm_page_t m) 1370 { 1371 lwkt_gettoken(&vm_token); 1372 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1373 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1374 lwkt_reltoken(&vm_token); 1375 return(0); 1376 } 1377 vm_page_test_dirty(m); 1378 if (m->dirty) { 1379 lwkt_reltoken(&vm_token); 1380 return(0); 1381 } 1382 vm_page_cache(m); 1383 lwkt_reltoken(&vm_token); 1384 return(1); 1385 } 1386 1387 /* 1388 * Attempt to free the page. If we cannot free it, we do nothing. 1389 * 1 is returned on success, 0 on failure. 1390 * 1391 * No requirements. 1392 */ 1393 int 1394 vm_page_try_to_free(vm_page_t m) 1395 { 1396 lwkt_gettoken(&vm_token); 1397 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1398 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1399 lwkt_reltoken(&vm_token); 1400 return(0); 1401 } 1402 vm_page_test_dirty(m); 1403 if (m->dirty) { 1404 lwkt_reltoken(&vm_token); 1405 return(0); 1406 } 1407 vm_page_busy(m); 1408 vm_page_protect(m, VM_PROT_NONE); 1409 vm_page_free(m); 1410 lwkt_reltoken(&vm_token); 1411 return(1); 1412 } 1413 1414 /* 1415 * vm_page_cache 1416 * 1417 * Put the specified page onto the page cache queue (if appropriate). 1418 * 1419 * The caller must hold vm_token. 1420 * This routine may not block. 1421 */ 1422 void 1423 vm_page_cache(vm_page_t m) 1424 { 1425 ASSERT_LWKT_TOKEN_HELD(&vm_token); 1426 1427 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1428 m->wire_count || m->hold_count) { 1429 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1430 return; 1431 } 1432 1433 /* 1434 * Already in the cache (and thus not mapped) 1435 */ 1436 if ((m->queue - m->pc) == PQ_CACHE) { 1437 KKASSERT((m->flags & PG_MAPPED) == 0); 1438 return; 1439 } 1440 1441 /* 1442 * Caller is required to test m->dirty, but note that the act of 1443 * removing the page from its maps can cause it to become dirty 1444 * on an SMP system due to another cpu running in usermode. 1445 */ 1446 if (m->dirty) { 1447 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1448 (long)m->pindex); 1449 } 1450 1451 /* 1452 * Remove all pmaps and indicate that the page is not 1453 * writeable or mapped. Our vm_page_protect() call may 1454 * have blocked (especially w/ VM_PROT_NONE), so recheck 1455 * everything. 1456 */ 1457 vm_page_busy(m); 1458 vm_page_protect(m, VM_PROT_NONE); 1459 vm_page_wakeup(m); 1460 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy || 1461 m->wire_count || m->hold_count) { 1462 /* do nothing */ 1463 } else if (m->dirty) { 1464 vm_page_deactivate(m); 1465 } else { 1466 vm_page_unqueue_nowakeup(m); 1467 m->queue = PQ_CACHE + m->pc; 1468 vm_page_queues[m->queue].lcnt++; 1469 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1470 vmstats.v_cache_count++; 1471 vm_page_free_wakeup(); 1472 } 1473 } 1474 1475 /* 1476 * vm_page_dontneed() 1477 * 1478 * Cache, deactivate, or do nothing as appropriate. This routine 1479 * is typically used by madvise() MADV_DONTNEED. 1480 * 1481 * Generally speaking we want to move the page into the cache so 1482 * it gets reused quickly. However, this can result in a silly syndrome 1483 * due to the page recycling too quickly. Small objects will not be 1484 * fully cached. On the otherhand, if we move the page to the inactive 1485 * queue we wind up with a problem whereby very large objects 1486 * unnecessarily blow away our inactive and cache queues. 1487 * 1488 * The solution is to move the pages based on a fixed weighting. We 1489 * either leave them alone, deactivate them, or move them to the cache, 1490 * where moving them to the cache has the highest weighting. 1491 * By forcing some pages into other queues we eventually force the 1492 * system to balance the queues, potentially recovering other unrelated 1493 * space from active. The idea is to not force this to happen too 1494 * often. 1495 * 1496 * No requirements. 1497 */ 1498 void 1499 vm_page_dontneed(vm_page_t m) 1500 { 1501 static int dnweight; 1502 int dnw; 1503 int head; 1504 1505 dnw = ++dnweight; 1506 1507 /* 1508 * occassionally leave the page alone 1509 */ 1510 lwkt_gettoken(&vm_token); 1511 if ((dnw & 0x01F0) == 0 || 1512 m->queue == PQ_INACTIVE || 1513 m->queue - m->pc == PQ_CACHE 1514 ) { 1515 if (m->act_count >= ACT_INIT) 1516 --m->act_count; 1517 lwkt_reltoken(&vm_token); 1518 return; 1519 } 1520 1521 if (m->dirty == 0) 1522 vm_page_test_dirty(m); 1523 1524 if (m->dirty || (dnw & 0x0070) == 0) { 1525 /* 1526 * Deactivate the page 3 times out of 32. 1527 */ 1528 head = 0; 1529 } else { 1530 /* 1531 * Cache the page 28 times out of every 32. Note that 1532 * the page is deactivated instead of cached, but placed 1533 * at the head of the queue instead of the tail. 1534 */ 1535 head = 1; 1536 } 1537 _vm_page_deactivate(m, head); 1538 lwkt_reltoken(&vm_token); 1539 } 1540 1541 /* 1542 * Grab a page, blocking if it is busy and allocating a page if necessary. 1543 * A busy page is returned or NULL. 1544 * 1545 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1546 * If VM_ALLOC_RETRY is not specified 1547 * 1548 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1549 * always returned if we had blocked. 1550 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1551 * This routine may not be called from an interrupt. 1552 * The returned page may not be entirely valid. 1553 * 1554 * This routine may be called from mainline code without spl protection and 1555 * be guarenteed a busied page associated with the object at the specified 1556 * index. 1557 * 1558 * No requirements. 1559 */ 1560 vm_page_t 1561 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1562 { 1563 vm_page_t m; 1564 int generation; 1565 1566 KKASSERT(allocflags & 1567 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1568 lwkt_gettoken(&vm_token); 1569 retrylookup: 1570 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1571 if (m->busy || (m->flags & PG_BUSY)) { 1572 generation = object->generation; 1573 1574 while ((object->generation == generation) && 1575 (m->busy || (m->flags & PG_BUSY))) { 1576 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1577 tsleep(m, 0, "pgrbwt", 0); 1578 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1579 m = NULL; 1580 goto done; 1581 } 1582 } 1583 goto retrylookup; 1584 } else { 1585 vm_page_busy(m); 1586 goto done; 1587 } 1588 } 1589 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1590 if (m == NULL) { 1591 vm_wait(0); 1592 if ((allocflags & VM_ALLOC_RETRY) == 0) 1593 goto done; 1594 goto retrylookup; 1595 } 1596 done: 1597 lwkt_reltoken(&vm_token); 1598 return(m); 1599 } 1600 1601 /* 1602 * Mapping function for valid bits or for dirty bits in 1603 * a page. May not block. 1604 * 1605 * Inputs are required to range within a page. 1606 * 1607 * No requirements. 1608 * Non blocking. 1609 */ 1610 int 1611 vm_page_bits(int base, int size) 1612 { 1613 int first_bit; 1614 int last_bit; 1615 1616 KASSERT( 1617 base + size <= PAGE_SIZE, 1618 ("vm_page_bits: illegal base/size %d/%d", base, size) 1619 ); 1620 1621 if (size == 0) /* handle degenerate case */ 1622 return(0); 1623 1624 first_bit = base >> DEV_BSHIFT; 1625 last_bit = (base + size - 1) >> DEV_BSHIFT; 1626 1627 return ((2 << last_bit) - (1 << first_bit)); 1628 } 1629 1630 /* 1631 * Sets portions of a page valid and clean. The arguments are expected 1632 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1633 * of any partial chunks touched by the range. The invalid portion of 1634 * such chunks will be zero'd. 1635 * 1636 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1637 * align base to DEV_BSIZE so as not to mark clean a partially 1638 * truncated device block. Otherwise the dirty page status might be 1639 * lost. 1640 * 1641 * This routine may not block. 1642 * 1643 * (base + size) must be less then or equal to PAGE_SIZE. 1644 */ 1645 static void 1646 _vm_page_zero_valid(vm_page_t m, int base, int size) 1647 { 1648 int frag; 1649 int endoff; 1650 1651 if (size == 0) /* handle degenerate case */ 1652 return; 1653 1654 /* 1655 * If the base is not DEV_BSIZE aligned and the valid 1656 * bit is clear, we have to zero out a portion of the 1657 * first block. 1658 */ 1659 1660 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1661 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1662 ) { 1663 pmap_zero_page_area( 1664 VM_PAGE_TO_PHYS(m), 1665 frag, 1666 base - frag 1667 ); 1668 } 1669 1670 /* 1671 * If the ending offset is not DEV_BSIZE aligned and the 1672 * valid bit is clear, we have to zero out a portion of 1673 * the last block. 1674 */ 1675 1676 endoff = base + size; 1677 1678 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1679 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1680 ) { 1681 pmap_zero_page_area( 1682 VM_PAGE_TO_PHYS(m), 1683 endoff, 1684 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1685 ); 1686 } 1687 } 1688 1689 /* 1690 * Set valid, clear dirty bits. If validating the entire 1691 * page we can safely clear the pmap modify bit. We also 1692 * use this opportunity to clear the PG_NOSYNC flag. If a process 1693 * takes a write fault on a MAP_NOSYNC memory area the flag will 1694 * be set again. 1695 * 1696 * We set valid bits inclusive of any overlap, but we can only 1697 * clear dirty bits for DEV_BSIZE chunks that are fully within 1698 * the range. 1699 * 1700 * Page must be busied? 1701 * No other requirements. 1702 */ 1703 void 1704 vm_page_set_valid(vm_page_t m, int base, int size) 1705 { 1706 _vm_page_zero_valid(m, base, size); 1707 m->valid |= vm_page_bits(base, size); 1708 } 1709 1710 1711 /* 1712 * Set valid bits and clear dirty bits. 1713 * 1714 * NOTE: This function does not clear the pmap modified bit. 1715 * Also note that e.g. NFS may use a byte-granular base 1716 * and size. 1717 * 1718 * WARNING: Page must be busied? But vfs_clean_one_page() will call 1719 * this without necessarily busying the page (via bdwrite()). 1720 * So for now vm_token must also be held. 1721 * 1722 * No other requirements. 1723 */ 1724 void 1725 vm_page_set_validclean(vm_page_t m, int base, int size) 1726 { 1727 int pagebits; 1728 1729 _vm_page_zero_valid(m, base, size); 1730 pagebits = vm_page_bits(base, size); 1731 m->valid |= pagebits; 1732 m->dirty &= ~pagebits; 1733 if (base == 0 && size == PAGE_SIZE) { 1734 /*pmap_clear_modify(m);*/ 1735 vm_page_flag_clear(m, PG_NOSYNC); 1736 } 1737 } 1738 1739 /* 1740 * Set valid & dirty. Used by buwrite() 1741 * 1742 * WARNING: Page must be busied? But vfs_dirty_one_page() will 1743 * call this function in buwrite() so for now vm_token must 1744 * be held. 1745 * 1746 * No other requirements. 1747 */ 1748 void 1749 vm_page_set_validdirty(vm_page_t m, int base, int size) 1750 { 1751 int pagebits; 1752 1753 pagebits = vm_page_bits(base, size); 1754 m->valid |= pagebits; 1755 m->dirty |= pagebits; 1756 if (m->object) 1757 vm_object_set_writeable_dirty(m->object); 1758 } 1759 1760 /* 1761 * Clear dirty bits. 1762 * 1763 * NOTE: This function does not clear the pmap modified bit. 1764 * Also note that e.g. NFS may use a byte-granular base 1765 * and size. 1766 * 1767 * Page must be busied? 1768 * No other requirements. 1769 */ 1770 void 1771 vm_page_clear_dirty(vm_page_t m, int base, int size) 1772 { 1773 m->dirty &= ~vm_page_bits(base, size); 1774 if (base == 0 && size == PAGE_SIZE) { 1775 /*pmap_clear_modify(m);*/ 1776 vm_page_flag_clear(m, PG_NOSYNC); 1777 } 1778 } 1779 1780 /* 1781 * Make the page all-dirty. 1782 * 1783 * Also make sure the related object and vnode reflect the fact that the 1784 * object may now contain a dirty page. 1785 * 1786 * Page must be busied? 1787 * No other requirements. 1788 */ 1789 void 1790 vm_page_dirty(vm_page_t m) 1791 { 1792 #ifdef INVARIANTS 1793 int pqtype = m->queue - m->pc; 1794 #endif 1795 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1796 ("vm_page_dirty: page in free/cache queue!")); 1797 if (m->dirty != VM_PAGE_BITS_ALL) { 1798 m->dirty = VM_PAGE_BITS_ALL; 1799 if (m->object) 1800 vm_object_set_writeable_dirty(m->object); 1801 } 1802 } 1803 1804 /* 1805 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1806 * valid and dirty bits for the effected areas are cleared. 1807 * 1808 * Page must be busied? 1809 * Does not block. 1810 * No other requirements. 1811 */ 1812 void 1813 vm_page_set_invalid(vm_page_t m, int base, int size) 1814 { 1815 int bits; 1816 1817 bits = vm_page_bits(base, size); 1818 m->valid &= ~bits; 1819 m->dirty &= ~bits; 1820 m->object->generation++; 1821 } 1822 1823 /* 1824 * The kernel assumes that the invalid portions of a page contain 1825 * garbage, but such pages can be mapped into memory by user code. 1826 * When this occurs, we must zero out the non-valid portions of the 1827 * page so user code sees what it expects. 1828 * 1829 * Pages are most often semi-valid when the end of a file is mapped 1830 * into memory and the file's size is not page aligned. 1831 * 1832 * Page must be busied? 1833 * No other requirements. 1834 */ 1835 void 1836 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1837 { 1838 int b; 1839 int i; 1840 1841 /* 1842 * Scan the valid bits looking for invalid sections that 1843 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1844 * valid bit may be set ) have already been zerod by 1845 * vm_page_set_validclean(). 1846 */ 1847 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1848 if (i == (PAGE_SIZE / DEV_BSIZE) || 1849 (m->valid & (1 << i)) 1850 ) { 1851 if (i > b) { 1852 pmap_zero_page_area( 1853 VM_PAGE_TO_PHYS(m), 1854 b << DEV_BSHIFT, 1855 (i - b) << DEV_BSHIFT 1856 ); 1857 } 1858 b = i + 1; 1859 } 1860 } 1861 1862 /* 1863 * setvalid is TRUE when we can safely set the zero'd areas 1864 * as being valid. We can do this if there are no cache consistency 1865 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1866 */ 1867 if (setvalid) 1868 m->valid = VM_PAGE_BITS_ALL; 1869 } 1870 1871 /* 1872 * Is a (partial) page valid? Note that the case where size == 0 1873 * will return FALSE in the degenerate case where the page is entirely 1874 * invalid, and TRUE otherwise. 1875 * 1876 * Does not block. 1877 * No other requirements. 1878 */ 1879 int 1880 vm_page_is_valid(vm_page_t m, int base, int size) 1881 { 1882 int bits = vm_page_bits(base, size); 1883 1884 if (m->valid && ((m->valid & bits) == bits)) 1885 return 1; 1886 else 1887 return 0; 1888 } 1889 1890 /* 1891 * update dirty bits from pmap/mmu. May not block. 1892 * 1893 * Caller must hold vm_token if non-blocking operation desired. 1894 * No other requirements. 1895 */ 1896 void 1897 vm_page_test_dirty(vm_page_t m) 1898 { 1899 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1900 vm_page_dirty(m); 1901 } 1902 } 1903 1904 /* 1905 * Register an action, associating it with its vm_page 1906 */ 1907 void 1908 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 1909 { 1910 struct vm_page_action_list *list; 1911 int hv; 1912 1913 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1914 list = &action_list[hv]; 1915 1916 lwkt_gettoken(&vm_token); 1917 vm_page_flag_set(action->m, PG_ACTIONLIST); 1918 action->event = event; 1919 LIST_INSERT_HEAD(list, action, entry); 1920 lwkt_reltoken(&vm_token); 1921 } 1922 1923 /* 1924 * Unregister an action, disassociating it from its related vm_page 1925 */ 1926 void 1927 vm_page_unregister_action(vm_page_action_t action) 1928 { 1929 struct vm_page_action_list *list; 1930 int hv; 1931 1932 lwkt_gettoken(&vm_token); 1933 if (action->event != VMEVENT_NONE) { 1934 action->event = VMEVENT_NONE; 1935 LIST_REMOVE(action, entry); 1936 1937 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 1938 list = &action_list[hv]; 1939 if (LIST_EMPTY(list)) 1940 vm_page_flag_clear(action->m, PG_ACTIONLIST); 1941 } 1942 lwkt_reltoken(&vm_token); 1943 } 1944 1945 /* 1946 * Issue an event on a VM page. Corresponding action structures are 1947 * removed from the page's list and called. 1948 * 1949 * If the vm_page has no more pending action events we clear its 1950 * PG_ACTIONLIST flag. 1951 */ 1952 void 1953 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1954 { 1955 struct vm_page_action_list *list; 1956 struct vm_page_action *scan; 1957 struct vm_page_action *next; 1958 int hv; 1959 int all; 1960 1961 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 1962 list = &action_list[hv]; 1963 all = 1; 1964 1965 lwkt_gettoken(&vm_token); 1966 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 1967 if (scan->m == m) { 1968 if (scan->event == event) { 1969 scan->event = VMEVENT_NONE; 1970 LIST_REMOVE(scan, entry); 1971 scan->func(m, scan); 1972 /* XXX */ 1973 } else { 1974 all = 0; 1975 } 1976 } 1977 } 1978 if (all) 1979 vm_page_flag_clear(m, PG_ACTIONLIST); 1980 lwkt_reltoken(&vm_token); 1981 } 1982 1983 1984 #include "opt_ddb.h" 1985 #ifdef DDB 1986 #include <sys/kernel.h> 1987 1988 #include <ddb/ddb.h> 1989 1990 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1991 { 1992 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1993 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1994 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1995 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1996 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1997 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1998 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1999 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 2000 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 2001 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 2002 } 2003 2004 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2005 { 2006 int i; 2007 db_printf("PQ_FREE:"); 2008 for(i=0;i<PQ_L2_SIZE;i++) { 2009 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 2010 } 2011 db_printf("\n"); 2012 2013 db_printf("PQ_CACHE:"); 2014 for(i=0;i<PQ_L2_SIZE;i++) { 2015 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 2016 } 2017 db_printf("\n"); 2018 2019 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2020 vm_page_queues[PQ_ACTIVE].lcnt, 2021 vm_page_queues[PQ_INACTIVE].lcnt); 2022 } 2023 #endif /* DDB */ 2024