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