1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 39 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 40 * $DragonFly: src/sys/vm/vm_page.c,v 1.40 2008/08/25 17:01:42 dillon Exp $ 41 */ 42 43 /* 44 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 45 * All rights reserved. 46 * 47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 /* 70 * Resident memory management module. The module manipulates 'VM pages'. 71 * A VM page is the core building block for memory management. 72 */ 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/malloc.h> 77 #include <sys/proc.h> 78 #include <sys/vmmeter.h> 79 #include <sys/vnode.h> 80 81 #include <vm/vm.h> 82 #include <vm/vm_param.h> 83 #include <sys/lock.h> 84 #include <vm/vm_kern.h> 85 #include <vm/pmap.h> 86 #include <vm/vm_map.h> 87 #include <vm/vm_object.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_pageout.h> 90 #include <vm/vm_pager.h> 91 #include <vm/vm_extern.h> 92 #include <vm/swap_pager.h> 93 94 #include <machine/md_var.h> 95 96 #include <vm/vm_page2.h> 97 #include <sys/mplock2.h> 98 99 static void vm_page_queue_init(void); 100 static void vm_page_free_wakeup(void); 101 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t); 102 static vm_page_t _vm_page_list_find2(int basequeue, int index); 103 104 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */ 105 106 #define ASSERT_IN_CRIT_SECTION() KKASSERT(crit_test(curthread)); 107 108 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 109 vm_pindex_t, pindex); 110 111 static void 112 vm_page_queue_init(void) 113 { 114 int i; 115 116 for (i = 0; i < PQ_L2_SIZE; i++) 117 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 118 for (i = 0; i < PQ_L2_SIZE; i++) 119 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 120 121 vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count; 122 vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count; 123 vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count; 124 /* PQ_NONE has no queue */ 125 126 for (i = 0; i < PQ_COUNT; i++) 127 TAILQ_INIT(&vm_page_queues[i].pl); 128 } 129 130 /* 131 * note: place in initialized data section? Is this necessary? 132 */ 133 long first_page = 0; 134 int vm_page_array_size = 0; 135 int vm_page_zero_count = 0; 136 vm_page_t vm_page_array = 0; 137 138 /* 139 * (low level boot) 140 * 141 * Sets the page size, perhaps based upon the memory size. 142 * Must be called before any use of page-size dependent functions. 143 */ 144 void 145 vm_set_page_size(void) 146 { 147 if (vmstats.v_page_size == 0) 148 vmstats.v_page_size = PAGE_SIZE; 149 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 150 panic("vm_set_page_size: page size not a power of two"); 151 } 152 153 /* 154 * (low level boot) 155 * 156 * Add a new page to the freelist for use by the system. New pages 157 * are added to both the head and tail of the associated free page 158 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 159 * requests pull 'recent' adds (higher physical addresses) first. 160 * 161 * Must be called in a critical section. 162 */ 163 vm_page_t 164 vm_add_new_page(vm_paddr_t pa) 165 { 166 struct vpgqueues *vpq; 167 vm_page_t m; 168 169 ++vmstats.v_page_count; 170 ++vmstats.v_free_count; 171 m = PHYS_TO_VM_PAGE(pa); 172 m->phys_addr = pa; 173 m->flags = 0; 174 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 175 m->queue = m->pc + PQ_FREE; 176 KKASSERT(m->dirty == 0); 177 178 vpq = &vm_page_queues[m->queue]; 179 if (vpq->flipflop) 180 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 181 else 182 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 183 vpq->flipflop = 1 - vpq->flipflop; 184 185 vm_page_queues[m->queue].lcnt++; 186 return (m); 187 } 188 189 /* 190 * (low level boot) 191 * 192 * Initializes the resident memory module. 193 * 194 * Allocates memory for the page cells, and for the object/offset-to-page 195 * hash table headers. Each page cell is initialized and placed on the 196 * free list. 197 * 198 * starta/enda represents the range of physical memory addresses available 199 * for use (skipping memory already used by the kernel), subject to 200 * phys_avail[]. Note that phys_avail[] has already mapped out memory 201 * already in use by the kernel. 202 */ 203 vm_offset_t 204 vm_page_startup(vm_offset_t vaddr) 205 { 206 vm_offset_t mapped; 207 vm_size_t npages; 208 vm_paddr_t page_range; 209 vm_paddr_t new_end; 210 int i; 211 vm_paddr_t pa; 212 int nblocks; 213 vm_paddr_t last_pa; 214 vm_paddr_t end; 215 vm_paddr_t biggestone, biggestsize; 216 vm_paddr_t total; 217 218 total = 0; 219 biggestsize = 0; 220 biggestone = 0; 221 nblocks = 0; 222 vaddr = round_page(vaddr); 223 224 for (i = 0; phys_avail[i + 1]; i += 2) { 225 phys_avail[i] = round_page64(phys_avail[i]); 226 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 227 } 228 229 for (i = 0; phys_avail[i + 1]; i += 2) { 230 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 231 232 if (size > biggestsize) { 233 biggestone = i; 234 biggestsize = size; 235 } 236 ++nblocks; 237 total += size; 238 } 239 240 end = phys_avail[biggestone+1]; 241 end = trunc_page(end); 242 243 /* 244 * Initialize the queue headers for the free queue, the active queue 245 * and the inactive queue. 246 */ 247 248 vm_page_queue_init(); 249 250 /* VKERNELs don't support minidumps and as such don't need vm_page_dump */ 251 #if !defined(_KERNEL_VIRTUAL) 252 /* 253 * Allocate a bitmap to indicate that a random physical page 254 * needs to be included in a minidump. 255 * 256 * The amd64 port needs this to indicate which direct map pages 257 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 258 * 259 * However, i386 still needs this workspace internally within the 260 * minidump code. In theory, they are not needed on i386, but are 261 * included should the sf_buf code decide to use them. 262 */ 263 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 264 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 265 end -= vm_page_dump_size; 266 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 267 VM_PROT_READ | VM_PROT_WRITE); 268 bzero((void *)vm_page_dump, vm_page_dump_size); 269 #endif 270 271 /* 272 * Compute the number of pages of memory that will be available for 273 * use (taking into account the overhead of a page structure per 274 * page). 275 */ 276 first_page = phys_avail[0] / PAGE_SIZE; 277 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 278 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 279 280 /* 281 * Initialize the mem entry structures now, and put them in the free 282 * queue. 283 */ 284 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 285 mapped = pmap_map(&vaddr, new_end, end, 286 VM_PROT_READ | VM_PROT_WRITE); 287 vm_page_array = (vm_page_t)mapped; 288 289 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 290 /* 291 * since pmap_map on amd64 returns stuff out of a direct-map region, 292 * we have to manually add these pages to the minidump tracking so 293 * that they can be dumped, including the vm_page_array. 294 */ 295 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 296 dump_add_page(pa); 297 #endif 298 299 /* 300 * Clear all of the page structures 301 */ 302 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 303 vm_page_array_size = page_range; 304 305 /* 306 * Construct the free queue(s) in ascending order (by physical 307 * address) so that the first 16MB of physical memory is allocated 308 * last rather than first. On large-memory machines, this avoids 309 * the exhaustion of low physical memory before isa_dmainit has run. 310 */ 311 vmstats.v_page_count = 0; 312 vmstats.v_free_count = 0; 313 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 314 pa = phys_avail[i]; 315 if (i == biggestone) 316 last_pa = new_end; 317 else 318 last_pa = phys_avail[i + 1]; 319 while (pa < last_pa && npages-- > 0) { 320 vm_add_new_page(pa); 321 pa += PAGE_SIZE; 322 } 323 } 324 return (vaddr); 325 } 326 327 /* 328 * Scan comparison function for Red-Black tree scans. An inclusive 329 * (start,end) is expected. Other fields are not used. 330 */ 331 int 332 rb_vm_page_scancmp(struct vm_page *p, void *data) 333 { 334 struct rb_vm_page_scan_info *info = data; 335 336 if (p->pindex < info->start_pindex) 337 return(-1); 338 if (p->pindex > info->end_pindex) 339 return(1); 340 return(0); 341 } 342 343 int 344 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 345 { 346 if (p1->pindex < p2->pindex) 347 return(-1); 348 if (p1->pindex > p2->pindex) 349 return(1); 350 return(0); 351 } 352 353 /* 354 * The opposite of vm_page_hold(). A page can be freed while being held, 355 * which places it on the PQ_HOLD queue. We must call vm_page_free_toq() 356 * in this case to actually free it once the hold count drops to 0. 357 * 358 * This routine must be called at splvm(). 359 */ 360 void 361 vm_page_unhold(vm_page_t mem) 362 { 363 --mem->hold_count; 364 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 365 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) { 366 vm_page_busy(mem); 367 vm_page_free_toq(mem); 368 } 369 } 370 371 /* 372 * Inserts the given mem entry into the object and object list. 373 * 374 * The pagetables are not updated but will presumably fault the page 375 * in if necessary, or if a kernel page the caller will at some point 376 * enter the page into the kernel's pmap. We are not allowed to block 377 * here so we *can't* do this anyway. 378 * 379 * This routine may not block. 380 * This routine must be called with a critical section held. 381 */ 382 void 383 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 384 { 385 ASSERT_IN_CRIT_SECTION(); 386 if (m->object != NULL) 387 panic("vm_page_insert: already inserted"); 388 389 /* 390 * Record the object/offset pair in this page 391 */ 392 m->object = object; 393 m->pindex = pindex; 394 395 /* 396 * Insert it into the object. 397 */ 398 ASSERT_MP_LOCK_HELD(curthread); 399 vm_page_rb_tree_RB_INSERT(&object->rb_memq, m); 400 object->generation++; 401 402 /* 403 * show that the object has one more resident page. 404 */ 405 object->resident_page_count++; 406 407 /* 408 * Since we are inserting a new and possibly dirty page, 409 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 410 */ 411 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 412 vm_object_set_writeable_dirty(object); 413 414 /* 415 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 416 */ 417 swap_pager_page_inserted(m); 418 } 419 420 /* 421 * Removes the given vm_page_t from the global (object,index) hash table 422 * and from the object's memq. 423 * 424 * The underlying pmap entry (if any) is NOT removed here. 425 * This routine may not block. 426 * 427 * The page must be BUSY and will remain BUSY on return. 428 * No other requirements. 429 * 430 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 431 * it busy. 432 */ 433 void 434 vm_page_remove(vm_page_t m) 435 { 436 vm_object_t object; 437 438 crit_enter(); 439 lwkt_gettoken(&vm_token); 440 if (m->object == NULL) { 441 lwkt_reltoken(&vm_token); 442 crit_exit(); 443 return; 444 } 445 446 if ((m->flags & PG_BUSY) == 0) 447 panic("vm_page_remove: page not busy"); 448 449 object = m->object; 450 451 /* 452 * Remove the page from the object and update the object. 453 */ 454 ASSERT_MP_LOCK_HELD(curthread); 455 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 456 object->resident_page_count--; 457 object->generation++; 458 m->object = NULL; 459 460 lwkt_reltoken(&vm_token); 461 crit_exit(); 462 } 463 464 /* 465 * Locate and return the page at (object, pindex), or NULL if the 466 * page could not be found. 467 * 468 * This routine will operate properly without spl protection, but 469 * the returned page could be in flux if it is busy. Because an 470 * interrupt can race a caller's busy check (unbusying and freeing the 471 * page we return before the caller is able to check the busy bit), 472 * the caller should generally call this routine with a critical 473 * section held. 474 * 475 * Callers may call this routine without spl protection if they know 476 * 'for sure' that the page will not be ripped out from under them 477 * by an interrupt. 478 */ 479 vm_page_t 480 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 481 { 482 vm_page_t m; 483 484 /* 485 * Search the hash table for this object/offset pair 486 */ 487 ASSERT_MP_LOCK_HELD(curthread); 488 crit_enter(); 489 lwkt_gettoken(&vm_token); 490 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 491 lwkt_reltoken(&vm_token); 492 crit_exit(); 493 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 494 return(m); 495 } 496 497 /* 498 * vm_page_rename() 499 * 500 * Move the given memory entry from its current object to the specified 501 * target object/offset. 502 * 503 * The object must be locked. 504 * This routine may not block. 505 * 506 * Note: This routine will raise itself to splvm(), the caller need not. 507 * 508 * Note: Swap associated with the page must be invalidated by the move. We 509 * have to do this for several reasons: (1) we aren't freeing the 510 * page, (2) we are dirtying the page, (3) the VM system is probably 511 * moving the page from object A to B, and will then later move 512 * the backing store from A to B and we can't have a conflict. 513 * 514 * Note: We *always* dirty the page. It is necessary both for the 515 * fact that we moved it, and because we may be invalidating 516 * swap. If the page is on the cache, we have to deactivate it 517 * or vm_page_dirty() will panic. Dirty pages are not allowed 518 * on the cache. 519 */ 520 void 521 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 522 { 523 crit_enter(); 524 lwkt_gettoken(&vm_token); 525 vm_page_remove(m); 526 vm_page_insert(m, new_object, new_pindex); 527 if (m->queue - m->pc == PQ_CACHE) 528 vm_page_deactivate(m); 529 vm_page_dirty(m); 530 vm_page_wakeup(m); 531 lwkt_reltoken(&vm_token); 532 crit_exit(); 533 } 534 535 /* 536 * vm_page_unqueue() without any wakeup. This routine is used when a page 537 * is being moved between queues or otherwise is to remain BUSYied by the 538 * caller. 539 * 540 * This routine must be called at splhigh(). 541 * This routine may not block. 542 */ 543 void 544 vm_page_unqueue_nowakeup(vm_page_t m) 545 { 546 int queue = m->queue; 547 struct vpgqueues *pq; 548 549 if (queue != PQ_NONE) { 550 pq = &vm_page_queues[queue]; 551 m->queue = PQ_NONE; 552 TAILQ_REMOVE(&pq->pl, m, pageq); 553 (*pq->cnt)--; 554 pq->lcnt--; 555 } 556 } 557 558 /* 559 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 560 * if necessary. 561 * 562 * This routine must be called at splhigh(). 563 * This routine may not block. 564 */ 565 void 566 vm_page_unqueue(vm_page_t m) 567 { 568 int queue = m->queue; 569 struct vpgqueues *pq; 570 571 if (queue != PQ_NONE) { 572 m->queue = PQ_NONE; 573 pq = &vm_page_queues[queue]; 574 TAILQ_REMOVE(&pq->pl, m, pageq); 575 (*pq->cnt)--; 576 pq->lcnt--; 577 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 578 pagedaemon_wakeup(); 579 } 580 } 581 582 /* 583 * vm_page_list_find() 584 * 585 * Find a page on the specified queue with color optimization. 586 * 587 * The page coloring optimization attempts to locate a page that does 588 * not overload other nearby pages in the object in the cpu's L1 or L2 589 * caches. We need this optimization because cpu caches tend to be 590 * physical caches, while object spaces tend to be virtual. 591 * 592 * This routine must be called at splvm(). 593 * This routine may not block. 594 * 595 * Note that this routine is carefully inlined. A non-inlined version 596 * is available for outside callers but the only critical path is 597 * from within this source file. 598 */ 599 static __inline 600 vm_page_t 601 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 602 { 603 vm_page_t m; 604 605 if (prefer_zero) 606 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 607 else 608 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 609 if (m == NULL) 610 m = _vm_page_list_find2(basequeue, index); 611 return(m); 612 } 613 614 static vm_page_t 615 _vm_page_list_find2(int basequeue, int index) 616 { 617 int i; 618 vm_page_t m = NULL; 619 struct vpgqueues *pq; 620 621 pq = &vm_page_queues[basequeue]; 622 623 /* 624 * Note that for the first loop, index+i and index-i wind up at the 625 * same place. Even though this is not totally optimal, we've already 626 * blown it by missing the cache case so we do not care. 627 */ 628 629 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 630 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 631 break; 632 633 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 634 break; 635 } 636 return(m); 637 } 638 639 vm_page_t 640 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 641 { 642 return(_vm_page_list_find(basequeue, index, prefer_zero)); 643 } 644 645 /* 646 * Find a page on the cache queue with color optimization. As pages 647 * might be found, but not applicable, they are deactivated. This 648 * keeps us from using potentially busy cached pages. 649 * 650 * This routine must be called with a critical section held. 651 * This routine may not block. 652 */ 653 vm_page_t 654 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 655 { 656 vm_page_t m; 657 658 while (TRUE) { 659 m = _vm_page_list_find( 660 PQ_CACHE, 661 (pindex + object->pg_color) & PQ_L2_MASK, 662 FALSE 663 ); 664 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 665 m->hold_count || m->wire_count)) { 666 vm_page_deactivate(m); 667 continue; 668 } 669 return m; 670 } 671 /* not reached */ 672 } 673 674 /* 675 * Find a free or zero page, with specified preference. We attempt to 676 * inline the nominal case and fall back to _vm_page_select_free() 677 * otherwise. 678 * 679 * This routine must be called with a critical section held. 680 * This routine may not block. 681 */ 682 static __inline vm_page_t 683 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 684 { 685 vm_page_t m; 686 687 m = _vm_page_list_find( 688 PQ_FREE, 689 (pindex + object->pg_color) & PQ_L2_MASK, 690 prefer_zero 691 ); 692 return(m); 693 } 694 695 /* 696 * vm_page_alloc() 697 * 698 * Allocate and return a memory cell associated with this VM object/offset 699 * pair. 700 * 701 * page_req classes: 702 * 703 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 704 * VM_ALLOC_QUICK like normal but cannot use cache 705 * VM_ALLOC_SYSTEM greater free drain 706 * VM_ALLOC_INTERRUPT allow free list to be completely drained 707 * VM_ALLOC_ZERO advisory request for pre-zero'd page 708 * 709 * The object must be locked. 710 * This routine may not block. 711 * The returned page will be marked PG_BUSY 712 * 713 * Additional special handling is required when called from an interrupt 714 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 715 * in this case. 716 */ 717 vm_page_t 718 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 719 { 720 vm_page_t m = NULL; 721 722 crit_enter(); 723 lwkt_gettoken(&vm_token); 724 725 KKASSERT(object != NULL); 726 KASSERT(!vm_page_lookup(object, pindex), 727 ("vm_page_alloc: page already allocated")); 728 KKASSERT(page_req & 729 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 730 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 731 732 /* 733 * Certain system threads (pageout daemon, buf_daemon's) are 734 * allowed to eat deeper into the free page list. 735 */ 736 if (curthread->td_flags & TDF_SYSTHREAD) 737 page_req |= VM_ALLOC_SYSTEM; 738 739 loop: 740 if (vmstats.v_free_count > vmstats.v_free_reserved || 741 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 742 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 743 vmstats.v_free_count > vmstats.v_interrupt_free_min) 744 ) { 745 /* 746 * The free queue has sufficient free pages to take one out. 747 */ 748 if (page_req & VM_ALLOC_ZERO) 749 m = vm_page_select_free(object, pindex, TRUE); 750 else 751 m = vm_page_select_free(object, pindex, FALSE); 752 } else if (page_req & VM_ALLOC_NORMAL) { 753 /* 754 * Allocatable from the cache (non-interrupt only). On 755 * success, we must free the page and try again, thus 756 * ensuring that vmstats.v_*_free_min counters are replenished. 757 */ 758 #ifdef INVARIANTS 759 if (curthread->td_preempted) { 760 kprintf("vm_page_alloc(): warning, attempt to allocate" 761 " cache page from preempting interrupt\n"); 762 m = NULL; 763 } else { 764 m = vm_page_select_cache(object, pindex); 765 } 766 #else 767 m = vm_page_select_cache(object, pindex); 768 #endif 769 /* 770 * On success move the page into the free queue and loop. 771 */ 772 if (m != NULL) { 773 KASSERT(m->dirty == 0, 774 ("Found dirty cache page %p", m)); 775 vm_page_busy(m); 776 vm_page_protect(m, VM_PROT_NONE); 777 vm_page_free(m); 778 goto loop; 779 } 780 781 /* 782 * On failure return NULL 783 */ 784 lwkt_reltoken(&vm_token); 785 crit_exit(); 786 #if defined(DIAGNOSTIC) 787 if (vmstats.v_cache_count > 0) 788 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 789 #endif 790 vm_pageout_deficit++; 791 pagedaemon_wakeup(); 792 return (NULL); 793 } else { 794 /* 795 * No pages available, wakeup the pageout daemon and give up. 796 */ 797 lwkt_reltoken(&vm_token); 798 crit_exit(); 799 vm_pageout_deficit++; 800 pagedaemon_wakeup(); 801 return (NULL); 802 } 803 804 /* 805 * Good page found. The page has not yet been busied. We are in 806 * a critical section. 807 */ 808 KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n")); 809 KASSERT(m->dirty == 0, 810 ("vm_page_alloc: free/cache page %p was dirty", m)); 811 812 /* 813 * Remove from free queue 814 */ 815 vm_page_unqueue_nowakeup(m); 816 817 /* 818 * Initialize structure. Only the PG_ZERO flag is inherited. Set 819 * the page PG_BUSY 820 */ 821 if (m->flags & PG_ZERO) { 822 vm_page_zero_count--; 823 m->flags = PG_ZERO | PG_BUSY; 824 } else { 825 m->flags = PG_BUSY; 826 } 827 m->wire_count = 0; 828 m->hold_count = 0; 829 m->act_count = 0; 830 m->busy = 0; 831 m->valid = 0; 832 833 /* 834 * vm_page_insert() is safe prior to the crit_exit(). Note also that 835 * inserting a page here does not insert it into the pmap (which 836 * could cause us to block allocating memory). We cannot block 837 * anywhere. 838 */ 839 vm_page_insert(m, object, pindex); 840 841 /* 842 * Don't wakeup too often - wakeup the pageout daemon when 843 * we would be nearly out of memory. 844 */ 845 pagedaemon_wakeup(); 846 847 lwkt_reltoken(&vm_token); 848 crit_exit(); 849 850 /* 851 * A PG_BUSY page is returned. 852 */ 853 return (m); 854 } 855 856 /* 857 * Wait for sufficient free memory for nominal heavy memory use kernel 858 * operations. 859 */ 860 void 861 vm_wait_nominal(void) 862 { 863 while (vm_page_count_min(0)) 864 vm_wait(0); 865 } 866 867 /* 868 * Test if vm_wait_nominal() would block. 869 */ 870 int 871 vm_test_nominal(void) 872 { 873 if (vm_page_count_min(0)) 874 return(1); 875 return(0); 876 } 877 878 /* 879 * Block until free pages are available for allocation, called in various 880 * places before memory allocations. 881 */ 882 void 883 vm_wait(int timo) 884 { 885 crit_enter(); 886 lwkt_gettoken(&vm_token); 887 if (curthread == pagethread) { 888 vm_pageout_pages_needed = 1; 889 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 890 } else { 891 if (vm_pages_needed == 0) { 892 vm_pages_needed = 1; 893 wakeup(&vm_pages_needed); 894 } 895 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 896 } 897 lwkt_reltoken(&vm_token); 898 crit_exit(); 899 } 900 901 /* 902 * Block until free pages are available for allocation 903 * 904 * Called only in vm_fault so that processes page faulting can be 905 * easily tracked. 906 */ 907 void 908 vm_waitpfault(void) 909 { 910 crit_enter(); 911 lwkt_gettoken(&vm_token); 912 if (vm_pages_needed == 0) { 913 vm_pages_needed = 1; 914 wakeup(&vm_pages_needed); 915 } 916 tsleep(&vmstats.v_free_count, 0, "pfault", 0); 917 lwkt_reltoken(&vm_token); 918 crit_exit(); 919 } 920 921 /* 922 * Put the specified page on the active list (if appropriate). Ensure 923 * that act_count is at least ACT_INIT but do not otherwise mess with it. 924 * 925 * The page queues must be locked. 926 * This routine may not block. 927 */ 928 void 929 vm_page_activate(vm_page_t m) 930 { 931 crit_enter(); 932 lwkt_gettoken(&vm_token); 933 if (m->queue != PQ_ACTIVE) { 934 if ((m->queue - m->pc) == PQ_CACHE) 935 mycpu->gd_cnt.v_reactivated++; 936 937 vm_page_unqueue(m); 938 939 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 940 m->queue = PQ_ACTIVE; 941 vm_page_queues[PQ_ACTIVE].lcnt++; 942 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, 943 m, pageq); 944 if (m->act_count < ACT_INIT) 945 m->act_count = ACT_INIT; 946 vmstats.v_active_count++; 947 } 948 } else { 949 if (m->act_count < ACT_INIT) 950 m->act_count = ACT_INIT; 951 } 952 lwkt_reltoken(&vm_token); 953 crit_exit(); 954 } 955 956 /* 957 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 958 * routine is called when a page has been added to the cache or free 959 * queues. 960 * 961 * This routine may not block. 962 * This routine must be called at splvm() 963 */ 964 static __inline void 965 vm_page_free_wakeup(void) 966 { 967 /* 968 * if pageout daemon needs pages, then tell it that there are 969 * some free. 970 */ 971 if (vm_pageout_pages_needed && 972 vmstats.v_cache_count + vmstats.v_free_count >= 973 vmstats.v_pageout_free_min 974 ) { 975 wakeup(&vm_pageout_pages_needed); 976 vm_pageout_pages_needed = 0; 977 } 978 979 /* 980 * wakeup processes that are waiting on memory if we hit a 981 * high water mark. And wakeup scheduler process if we have 982 * lots of memory. this process will swapin processes. 983 */ 984 if (vm_pages_needed && !vm_page_count_min(0)) { 985 vm_pages_needed = 0; 986 wakeup(&vmstats.v_free_count); 987 } 988 } 989 990 /* 991 * vm_page_free_toq: 992 * 993 * Returns the given page to the PQ_FREE list, disassociating it with 994 * any VM object. 995 * 996 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 997 * return (the page will have been freed). No particular spl is required 998 * on entry. 999 * 1000 * This routine may not block. 1001 */ 1002 void 1003 vm_page_free_toq(vm_page_t m) 1004 { 1005 struct vpgqueues *pq; 1006 1007 crit_enter(); 1008 lwkt_gettoken(&vm_token); 1009 mycpu->gd_cnt.v_tfree++; 1010 1011 KKASSERT((m->flags & PG_MAPPED) == 0); 1012 1013 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1014 kprintf( 1015 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1016 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1017 m->hold_count); 1018 if ((m->queue - m->pc) == PQ_FREE) 1019 panic("vm_page_free: freeing free page"); 1020 else 1021 panic("vm_page_free: freeing busy page"); 1022 } 1023 1024 /* 1025 * unqueue, then remove page. Note that we cannot destroy 1026 * the page here because we do not want to call the pager's 1027 * callback routine until after we've put the page on the 1028 * appropriate free queue. 1029 */ 1030 vm_page_unqueue_nowakeup(m); 1031 vm_page_remove(m); 1032 1033 /* 1034 * No further management of fictitious pages occurs beyond object 1035 * and queue removal. 1036 */ 1037 if ((m->flags & PG_FICTITIOUS) != 0) { 1038 vm_page_wakeup(m); 1039 lwkt_reltoken(&vm_token); 1040 crit_exit(); 1041 return; 1042 } 1043 1044 m->valid = 0; 1045 vm_page_undirty(m); 1046 1047 if (m->wire_count != 0) { 1048 if (m->wire_count > 1) { 1049 panic( 1050 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1051 m->wire_count, (long)m->pindex); 1052 } 1053 panic("vm_page_free: freeing wired page"); 1054 } 1055 1056 /* 1057 * Clear the UNMANAGED flag when freeing an unmanaged page. 1058 */ 1059 if (m->flags & PG_UNMANAGED) { 1060 m->flags &= ~PG_UNMANAGED; 1061 } 1062 1063 if (m->hold_count != 0) { 1064 m->flags &= ~PG_ZERO; 1065 m->queue = PQ_HOLD; 1066 } else { 1067 m->queue = PQ_FREE + m->pc; 1068 } 1069 pq = &vm_page_queues[m->queue]; 1070 pq->lcnt++; 1071 ++(*pq->cnt); 1072 1073 /* 1074 * Put zero'd pages on the end ( where we look for zero'd pages 1075 * first ) and non-zerod pages at the head. 1076 */ 1077 if (m->flags & PG_ZERO) { 1078 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1079 ++vm_page_zero_count; 1080 } else { 1081 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1082 } 1083 vm_page_wakeup(m); 1084 vm_page_free_wakeup(); 1085 lwkt_reltoken(&vm_token); 1086 crit_exit(); 1087 } 1088 1089 /* 1090 * vm_page_free_fromq_fast() 1091 * 1092 * Remove a non-zero page from one of the free queues; the page is removed for 1093 * zeroing, so do not issue a wakeup. 1094 * 1095 * MPUNSAFE 1096 */ 1097 vm_page_t 1098 vm_page_free_fromq_fast(void) 1099 { 1100 static int qi; 1101 vm_page_t m; 1102 int i; 1103 1104 crit_enter(); 1105 lwkt_gettoken(&vm_token); 1106 for (i = 0; i < PQ_L2_SIZE; ++i) { 1107 m = vm_page_list_find(PQ_FREE, qi, FALSE); 1108 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 1109 if (m && (m->flags & PG_ZERO) == 0) { 1110 vm_page_unqueue_nowakeup(m); 1111 vm_page_busy(m); 1112 break; 1113 } 1114 m = NULL; 1115 } 1116 lwkt_reltoken(&vm_token); 1117 crit_exit(); 1118 return (m); 1119 } 1120 1121 /* 1122 * vm_page_unmanage() 1123 * 1124 * Prevent PV management from being done on the page. The page is 1125 * removed from the paging queues as if it were wired, and as a 1126 * consequence of no longer being managed the pageout daemon will not 1127 * touch it (since there is no way to locate the pte mappings for the 1128 * page). madvise() calls that mess with the pmap will also no longer 1129 * operate on the page. 1130 * 1131 * Beyond that the page is still reasonably 'normal'. Freeing the page 1132 * will clear the flag. 1133 * 1134 * This routine is used by OBJT_PHYS objects - objects using unswappable 1135 * physical memory as backing store rather then swap-backed memory and 1136 * will eventually be extended to support 4MB unmanaged physical 1137 * mappings. 1138 * 1139 * Must be called with a critical section held. 1140 */ 1141 void 1142 vm_page_unmanage(vm_page_t m) 1143 { 1144 ASSERT_IN_CRIT_SECTION(); 1145 if ((m->flags & PG_UNMANAGED) == 0) { 1146 if (m->wire_count == 0) 1147 vm_page_unqueue(m); 1148 } 1149 vm_page_flag_set(m, PG_UNMANAGED); 1150 } 1151 1152 /* 1153 * Mark this page as wired down by yet another map, removing it from 1154 * paging queues as necessary. 1155 * 1156 * The page queues must be locked. 1157 * This routine may not block. 1158 */ 1159 void 1160 vm_page_wire(vm_page_t m) 1161 { 1162 /* 1163 * Only bump the wire statistics if the page is not already wired, 1164 * and only unqueue the page if it is on some queue (if it is unmanaged 1165 * it is already off the queues). Don't do anything with fictitious 1166 * pages because they are always wired. 1167 */ 1168 crit_enter(); 1169 lwkt_gettoken(&vm_token); 1170 if ((m->flags & PG_FICTITIOUS) == 0) { 1171 if (m->wire_count == 0) { 1172 if ((m->flags & PG_UNMANAGED) == 0) 1173 vm_page_unqueue(m); 1174 vmstats.v_wire_count++; 1175 } 1176 m->wire_count++; 1177 KASSERT(m->wire_count != 0, 1178 ("vm_page_wire: wire_count overflow m=%p", m)); 1179 } 1180 lwkt_reltoken(&vm_token); 1181 crit_exit(); 1182 } 1183 1184 /* 1185 * Release one wiring of this page, potentially enabling it to be paged again. 1186 * 1187 * Many pages placed on the inactive queue should actually go 1188 * into the cache, but it is difficult to figure out which. What 1189 * we do instead, if the inactive target is well met, is to put 1190 * clean pages at the head of the inactive queue instead of the tail. 1191 * This will cause them to be moved to the cache more quickly and 1192 * if not actively re-referenced, freed more quickly. If we just 1193 * stick these pages at the end of the inactive queue, heavy filesystem 1194 * meta-data accesses can cause an unnecessary paging load on memory bound 1195 * processes. This optimization causes one-time-use metadata to be 1196 * reused more quickly. 1197 * 1198 * BUT, if we are in a low-memory situation we have no choice but to 1199 * put clean pages on the cache queue. 1200 * 1201 * A number of routines use vm_page_unwire() to guarantee that the page 1202 * will go into either the inactive or active queues, and will NEVER 1203 * be placed in the cache - for example, just after dirtying a page. 1204 * dirty pages in the cache are not allowed. 1205 * 1206 * The page queues must be locked. 1207 * This routine may not block. 1208 */ 1209 void 1210 vm_page_unwire(vm_page_t m, int activate) 1211 { 1212 crit_enter(); 1213 lwkt_gettoken(&vm_token); 1214 if (m->flags & PG_FICTITIOUS) { 1215 /* do nothing */ 1216 } else if (m->wire_count <= 0) { 1217 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1218 } else { 1219 if (--m->wire_count == 0) { 1220 --vmstats.v_wire_count; 1221 if (m->flags & PG_UNMANAGED) { 1222 ; 1223 } else if (activate) { 1224 TAILQ_INSERT_TAIL( 1225 &vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1226 m->queue = PQ_ACTIVE; 1227 vm_page_queues[PQ_ACTIVE].lcnt++; 1228 vmstats.v_active_count++; 1229 } else { 1230 vm_page_flag_clear(m, PG_WINATCFLS); 1231 TAILQ_INSERT_TAIL( 1232 &vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1233 m->queue = PQ_INACTIVE; 1234 vm_page_queues[PQ_INACTIVE].lcnt++; 1235 vmstats.v_inactive_count++; 1236 ++vm_swapcache_inactive_heuristic; 1237 } 1238 } 1239 } 1240 lwkt_reltoken(&vm_token); 1241 crit_exit(); 1242 } 1243 1244 1245 /* 1246 * Move the specified page to the inactive queue. If the page has 1247 * any associated swap, the swap is deallocated. 1248 * 1249 * Normally athead is 0 resulting in LRU operation. athead is set 1250 * to 1 if we want this page to be 'as if it were placed in the cache', 1251 * except without unmapping it from the process address space. 1252 * 1253 * This routine may not block. 1254 */ 1255 static __inline void 1256 _vm_page_deactivate(vm_page_t m, int athead) 1257 { 1258 /* 1259 * Ignore if already inactive. 1260 */ 1261 if (m->queue == PQ_INACTIVE) 1262 return; 1263 1264 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1265 if ((m->queue - m->pc) == PQ_CACHE) 1266 mycpu->gd_cnt.v_reactivated++; 1267 vm_page_flag_clear(m, PG_WINATCFLS); 1268 vm_page_unqueue(m); 1269 if (athead) { 1270 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, 1271 m, pageq); 1272 } else { 1273 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, 1274 m, pageq); 1275 ++vm_swapcache_inactive_heuristic; 1276 } 1277 m->queue = PQ_INACTIVE; 1278 vm_page_queues[PQ_INACTIVE].lcnt++; 1279 vmstats.v_inactive_count++; 1280 } 1281 } 1282 1283 void 1284 vm_page_deactivate(vm_page_t m) 1285 { 1286 crit_enter(); 1287 lwkt_gettoken(&vm_token); 1288 _vm_page_deactivate(m, 0); 1289 lwkt_reltoken(&vm_token); 1290 crit_exit(); 1291 } 1292 1293 /* 1294 * vm_page_try_to_cache: 1295 * 1296 * Returns 0 on failure, 1 on success 1297 */ 1298 int 1299 vm_page_try_to_cache(vm_page_t m) 1300 { 1301 crit_enter(); 1302 lwkt_gettoken(&vm_token); 1303 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1304 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1305 lwkt_reltoken(&vm_token); 1306 crit_exit(); 1307 return(0); 1308 } 1309 vm_page_test_dirty(m); 1310 if (m->dirty) { 1311 lwkt_reltoken(&vm_token); 1312 crit_exit(); 1313 return(0); 1314 } 1315 vm_page_cache(m); 1316 lwkt_reltoken(&vm_token); 1317 crit_exit(); 1318 return(1); 1319 } 1320 1321 /* 1322 * Attempt to free the page. If we cannot free it, we do nothing. 1323 * 1 is returned on success, 0 on failure. 1324 */ 1325 int 1326 vm_page_try_to_free(vm_page_t m) 1327 { 1328 crit_enter(); 1329 lwkt_gettoken(&vm_token); 1330 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1331 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1332 lwkt_reltoken(&vm_token); 1333 crit_exit(); 1334 return(0); 1335 } 1336 vm_page_test_dirty(m); 1337 if (m->dirty) { 1338 lwkt_reltoken(&vm_token); 1339 crit_exit(); 1340 return(0); 1341 } 1342 vm_page_busy(m); 1343 vm_page_protect(m, VM_PROT_NONE); 1344 vm_page_free(m); 1345 lwkt_reltoken(&vm_token); 1346 crit_exit(); 1347 return(1); 1348 } 1349 1350 /* 1351 * vm_page_cache 1352 * 1353 * Put the specified page onto the page cache queue (if appropriate). 1354 * 1355 * This routine may not block. 1356 */ 1357 void 1358 vm_page_cache(vm_page_t m) 1359 { 1360 ASSERT_IN_CRIT_SECTION(); 1361 1362 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1363 m->wire_count || m->hold_count) { 1364 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 1365 return; 1366 } 1367 1368 /* 1369 * Already in the cache (and thus not mapped) 1370 */ 1371 if ((m->queue - m->pc) == PQ_CACHE) { 1372 KKASSERT((m->flags & PG_MAPPED) == 0); 1373 return; 1374 } 1375 1376 /* 1377 * Caller is required to test m->dirty, but note that the act of 1378 * removing the page from its maps can cause it to become dirty 1379 * on an SMP system due to another cpu running in usermode. 1380 */ 1381 if (m->dirty) { 1382 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1383 (long)m->pindex); 1384 } 1385 1386 /* 1387 * Remove all pmaps and indicate that the page is not 1388 * writeable or mapped. Our vm_page_protect() call may 1389 * have blocked (especially w/ VM_PROT_NONE), so recheck 1390 * everything. 1391 */ 1392 vm_page_busy(m); 1393 vm_page_protect(m, VM_PROT_NONE); 1394 vm_page_wakeup(m); 1395 if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy || 1396 m->wire_count || m->hold_count) { 1397 /* do nothing */ 1398 } else if (m->dirty) { 1399 vm_page_deactivate(m); 1400 } else { 1401 vm_page_unqueue_nowakeup(m); 1402 m->queue = PQ_CACHE + m->pc; 1403 vm_page_queues[m->queue].lcnt++; 1404 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1405 vmstats.v_cache_count++; 1406 vm_page_free_wakeup(); 1407 } 1408 } 1409 1410 /* 1411 * vm_page_dontneed() 1412 * 1413 * Cache, deactivate, or do nothing as appropriate. This routine 1414 * is typically used by madvise() MADV_DONTNEED. 1415 * 1416 * Generally speaking we want to move the page into the cache so 1417 * it gets reused quickly. However, this can result in a silly syndrome 1418 * due to the page recycling too quickly. Small objects will not be 1419 * fully cached. On the otherhand, if we move the page to the inactive 1420 * queue we wind up with a problem whereby very large objects 1421 * unnecessarily blow away our inactive and cache queues. 1422 * 1423 * The solution is to move the pages based on a fixed weighting. We 1424 * either leave them alone, deactivate them, or move them to the cache, 1425 * where moving them to the cache has the highest weighting. 1426 * By forcing some pages into other queues we eventually force the 1427 * system to balance the queues, potentially recovering other unrelated 1428 * space from active. The idea is to not force this to happen too 1429 * often. 1430 */ 1431 void 1432 vm_page_dontneed(vm_page_t m) 1433 { 1434 static int dnweight; 1435 int dnw; 1436 int head; 1437 1438 dnw = ++dnweight; 1439 1440 /* 1441 * occassionally leave the page alone 1442 */ 1443 crit_enter(); 1444 lwkt_gettoken(&vm_token); 1445 if ((dnw & 0x01F0) == 0 || 1446 m->queue == PQ_INACTIVE || 1447 m->queue - m->pc == PQ_CACHE 1448 ) { 1449 if (m->act_count >= ACT_INIT) 1450 --m->act_count; 1451 lwkt_reltoken(&vm_token); 1452 crit_exit(); 1453 return; 1454 } 1455 1456 if (m->dirty == 0) 1457 vm_page_test_dirty(m); 1458 1459 if (m->dirty || (dnw & 0x0070) == 0) { 1460 /* 1461 * Deactivate the page 3 times out of 32. 1462 */ 1463 head = 0; 1464 } else { 1465 /* 1466 * Cache the page 28 times out of every 32. Note that 1467 * the page is deactivated instead of cached, but placed 1468 * at the head of the queue instead of the tail. 1469 */ 1470 head = 1; 1471 } 1472 _vm_page_deactivate(m, head); 1473 lwkt_reltoken(&vm_token); 1474 crit_exit(); 1475 } 1476 1477 /* 1478 * Grab a page, blocking if it is busy and allocating a page if necessary. 1479 * A busy page is returned or NULL. 1480 * 1481 * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified. 1482 * If VM_ALLOC_RETRY is not specified 1483 * 1484 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 1485 * always returned if we had blocked. 1486 * This routine will never return NULL if VM_ALLOC_RETRY is set. 1487 * This routine may not be called from an interrupt. 1488 * The returned page may not be entirely valid. 1489 * 1490 * This routine may be called from mainline code without spl protection and 1491 * be guarenteed a busied page associated with the object at the specified 1492 * index. 1493 */ 1494 vm_page_t 1495 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1496 { 1497 vm_page_t m; 1498 int generation; 1499 1500 KKASSERT(allocflags & 1501 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1502 crit_enter(); 1503 lwkt_gettoken(&vm_token); 1504 retrylookup: 1505 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1506 if (m->busy || (m->flags & PG_BUSY)) { 1507 generation = object->generation; 1508 1509 while ((object->generation == generation) && 1510 (m->busy || (m->flags & PG_BUSY))) { 1511 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1512 tsleep(m, 0, "pgrbwt", 0); 1513 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1514 m = NULL; 1515 goto done; 1516 } 1517 } 1518 goto retrylookup; 1519 } else { 1520 vm_page_busy(m); 1521 goto done; 1522 } 1523 } 1524 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1525 if (m == NULL) { 1526 vm_wait(0); 1527 if ((allocflags & VM_ALLOC_RETRY) == 0) 1528 goto done; 1529 goto retrylookup; 1530 } 1531 done: 1532 lwkt_reltoken(&vm_token); 1533 crit_exit(); 1534 return(m); 1535 } 1536 1537 /* 1538 * Mapping function for valid bits or for dirty bits in 1539 * a page. May not block. 1540 * 1541 * Inputs are required to range within a page. 1542 */ 1543 __inline int 1544 vm_page_bits(int base, int size) 1545 { 1546 int first_bit; 1547 int last_bit; 1548 1549 KASSERT( 1550 base + size <= PAGE_SIZE, 1551 ("vm_page_bits: illegal base/size %d/%d", base, size) 1552 ); 1553 1554 if (size == 0) /* handle degenerate case */ 1555 return(0); 1556 1557 first_bit = base >> DEV_BSHIFT; 1558 last_bit = (base + size - 1) >> DEV_BSHIFT; 1559 1560 return ((2 << last_bit) - (1 << first_bit)); 1561 } 1562 1563 /* 1564 * Sets portions of a page valid and clean. The arguments are expected 1565 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1566 * of any partial chunks touched by the range. The invalid portion of 1567 * such chunks will be zero'd. 1568 * 1569 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 1570 * align base to DEV_BSIZE so as not to mark clean a partially 1571 * truncated device block. Otherwise the dirty page status might be 1572 * lost. 1573 * 1574 * This routine may not block. 1575 * 1576 * (base + size) must be less then or equal to PAGE_SIZE. 1577 */ 1578 static void 1579 _vm_page_zero_valid(vm_page_t m, int base, int size) 1580 { 1581 int frag; 1582 int endoff; 1583 1584 if (size == 0) /* handle degenerate case */ 1585 return; 1586 1587 /* 1588 * If the base is not DEV_BSIZE aligned and the valid 1589 * bit is clear, we have to zero out a portion of the 1590 * first block. 1591 */ 1592 1593 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1594 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1595 ) { 1596 pmap_zero_page_area( 1597 VM_PAGE_TO_PHYS(m), 1598 frag, 1599 base - frag 1600 ); 1601 } 1602 1603 /* 1604 * If the ending offset is not DEV_BSIZE aligned and the 1605 * valid bit is clear, we have to zero out a portion of 1606 * the last block. 1607 */ 1608 1609 endoff = base + size; 1610 1611 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1612 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1613 ) { 1614 pmap_zero_page_area( 1615 VM_PAGE_TO_PHYS(m), 1616 endoff, 1617 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1618 ); 1619 } 1620 } 1621 1622 /* 1623 * Set valid, clear dirty bits. If validating the entire 1624 * page we can safely clear the pmap modify bit. We also 1625 * use this opportunity to clear the PG_NOSYNC flag. If a process 1626 * takes a write fault on a MAP_NOSYNC memory area the flag will 1627 * be set again. 1628 * 1629 * We set valid bits inclusive of any overlap, but we can only 1630 * clear dirty bits for DEV_BSIZE chunks that are fully within 1631 * the range. 1632 */ 1633 void 1634 vm_page_set_valid(vm_page_t m, int base, int size) 1635 { 1636 _vm_page_zero_valid(m, base, size); 1637 m->valid |= vm_page_bits(base, size); 1638 } 1639 1640 1641 /* 1642 * Set valid bits and clear dirty bits. 1643 * 1644 * NOTE: This function does not clear the pmap modified bit. 1645 * Also note that e.g. NFS may use a byte-granular base 1646 * and size. 1647 */ 1648 void 1649 vm_page_set_validclean(vm_page_t m, int base, int size) 1650 { 1651 int pagebits; 1652 1653 _vm_page_zero_valid(m, base, size); 1654 pagebits = vm_page_bits(base, size); 1655 m->valid |= pagebits; 1656 m->dirty &= ~pagebits; 1657 if (base == 0 && size == PAGE_SIZE) { 1658 /*pmap_clear_modify(m);*/ 1659 vm_page_flag_clear(m, PG_NOSYNC); 1660 } 1661 } 1662 1663 /* 1664 * Set valid & dirty. Used by buwrite() 1665 */ 1666 void 1667 vm_page_set_validdirty(vm_page_t m, int base, int size) 1668 { 1669 int pagebits; 1670 1671 pagebits = vm_page_bits(base, size); 1672 m->valid |= pagebits; 1673 m->dirty |= pagebits; 1674 if (m->object) 1675 vm_object_set_writeable_dirty(m->object); 1676 } 1677 1678 /* 1679 * Clear dirty bits. 1680 * 1681 * NOTE: This function does not clear the pmap modified bit. 1682 * Also note that e.g. NFS may use a byte-granular base 1683 * and size. 1684 */ 1685 void 1686 vm_page_clear_dirty(vm_page_t m, int base, int size) 1687 { 1688 m->dirty &= ~vm_page_bits(base, size); 1689 if (base == 0 && size == PAGE_SIZE) { 1690 /*pmap_clear_modify(m);*/ 1691 vm_page_flag_clear(m, PG_NOSYNC); 1692 } 1693 } 1694 1695 /* 1696 * Make the page all-dirty. 1697 * 1698 * Also make sure the related object and vnode reflect the fact that the 1699 * object may now contain a dirty page. 1700 */ 1701 void 1702 vm_page_dirty(vm_page_t m) 1703 { 1704 #ifdef INVARIANTS 1705 int pqtype = m->queue - m->pc; 1706 #endif 1707 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 1708 ("vm_page_dirty: page in free/cache queue!")); 1709 if (m->dirty != VM_PAGE_BITS_ALL) { 1710 m->dirty = VM_PAGE_BITS_ALL; 1711 if (m->object) 1712 vm_object_set_writeable_dirty(m->object); 1713 } 1714 } 1715 1716 /* 1717 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1718 * valid and dirty bits for the effected areas are cleared. 1719 * 1720 * May not block. 1721 */ 1722 void 1723 vm_page_set_invalid(vm_page_t m, int base, int size) 1724 { 1725 int bits; 1726 1727 bits = vm_page_bits(base, size); 1728 m->valid &= ~bits; 1729 m->dirty &= ~bits; 1730 m->object->generation++; 1731 } 1732 1733 /* 1734 * The kernel assumes that the invalid portions of a page contain 1735 * garbage, but such pages can be mapped into memory by user code. 1736 * When this occurs, we must zero out the non-valid portions of the 1737 * page so user code sees what it expects. 1738 * 1739 * Pages are most often semi-valid when the end of a file is mapped 1740 * into memory and the file's size is not page aligned. 1741 */ 1742 void 1743 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1744 { 1745 int b; 1746 int i; 1747 1748 /* 1749 * Scan the valid bits looking for invalid sections that 1750 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1751 * valid bit may be set ) have already been zerod by 1752 * vm_page_set_validclean(). 1753 */ 1754 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1755 if (i == (PAGE_SIZE / DEV_BSIZE) || 1756 (m->valid & (1 << i)) 1757 ) { 1758 if (i > b) { 1759 pmap_zero_page_area( 1760 VM_PAGE_TO_PHYS(m), 1761 b << DEV_BSHIFT, 1762 (i - b) << DEV_BSHIFT 1763 ); 1764 } 1765 b = i + 1; 1766 } 1767 } 1768 1769 /* 1770 * setvalid is TRUE when we can safely set the zero'd areas 1771 * as being valid. We can do this if there are no cache consistency 1772 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1773 */ 1774 if (setvalid) 1775 m->valid = VM_PAGE_BITS_ALL; 1776 } 1777 1778 /* 1779 * Is a (partial) page valid? Note that the case where size == 0 1780 * will return FALSE in the degenerate case where the page is entirely 1781 * invalid, and TRUE otherwise. 1782 * 1783 * May not block. 1784 */ 1785 int 1786 vm_page_is_valid(vm_page_t m, int base, int size) 1787 { 1788 int bits = vm_page_bits(base, size); 1789 1790 if (m->valid && ((m->valid & bits) == bits)) 1791 return 1; 1792 else 1793 return 0; 1794 } 1795 1796 /* 1797 * update dirty bits from pmap/mmu. May not block. 1798 */ 1799 void 1800 vm_page_test_dirty(vm_page_t m) 1801 { 1802 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1803 vm_page_dirty(m); 1804 } 1805 } 1806 1807 /* 1808 * Issue an event on a VM page. Corresponding action structures are 1809 * removed from the page's list and called. 1810 */ 1811 void 1812 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 1813 { 1814 struct vm_page_action *scan, *next; 1815 1816 LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) { 1817 if (scan->event == event) { 1818 scan->event = VMEVENT_NONE; 1819 LIST_REMOVE(scan, entry); 1820 scan->func(m, scan); 1821 } 1822 } 1823 } 1824 1825 1826 #include "opt_ddb.h" 1827 #ifdef DDB 1828 #include <sys/kernel.h> 1829 1830 #include <ddb/ddb.h> 1831 1832 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1833 { 1834 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 1835 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 1836 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 1837 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 1838 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 1839 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 1840 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 1841 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 1842 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 1843 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 1844 } 1845 1846 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1847 { 1848 int i; 1849 db_printf("PQ_FREE:"); 1850 for(i=0;i<PQ_L2_SIZE;i++) { 1851 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1852 } 1853 db_printf("\n"); 1854 1855 db_printf("PQ_CACHE:"); 1856 for(i=0;i<PQ_L2_SIZE;i++) { 1857 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1858 } 1859 db_printf("\n"); 1860 1861 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1862 vm_page_queues[PQ_ACTIVE].lcnt, 1863 vm_page_queues[PQ_INACTIVE].lcnt); 1864 } 1865 #endif /* DDB */ 1866