1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 36 */ 37 38 /* 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 /* 65 * Resident memory management module. The module manipulates 'VM pages'. 66 * A VM page is the core building block for memory management. 67 */ 68 69 #include <sys/param.h> 70 #include <sys/systm.h> 71 #include <sys/malloc.h> 72 #include <sys/proc.h> 73 #include <sys/vmmeter.h> 74 #include <sys/vnode.h> 75 #include <sys/kernel.h> 76 #include <sys/alist.h> 77 #include <sys/sysctl.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/swap_pager.h> 91 92 #include <machine/inttypes.h> 93 #include <machine/md_var.h> 94 95 #include <vm/vm_page2.h> 96 #include <sys/spinlock2.h> 97 98 #define VMACTION_HSIZE 256 99 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 100 101 static void vm_page_queue_init(void); 102 static void vm_page_free_wakeup(void); 103 static vm_page_t vm_page_select_cache(u_short pg_color); 104 static vm_page_t _vm_page_list_find2(int basequeue, int index); 105 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 106 107 /* 108 * Array of tailq lists 109 */ 110 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 111 112 LIST_HEAD(vm_page_action_list, vm_page_action); 113 struct vm_page_action_list action_list[VMACTION_HSIZE]; 114 static volatile int vm_pages_waiting; 115 116 static struct alist vm_contig_alist; 117 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 118 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin); 119 120 static u_long vm_dma_reserved = 0; 121 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 122 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 123 "Memory reserved for DMA"); 124 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 125 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 126 127 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 128 vm_pindex_t, pindex); 129 130 static void 131 vm_page_queue_init(void) 132 { 133 int i; 134 135 for (i = 0; i < PQ_L2_SIZE; i++) 136 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 137 for (i = 0; i < PQ_L2_SIZE; i++) 138 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 139 for (i = 0; i < PQ_L2_SIZE; i++) 140 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count; 141 for (i = 0; i < PQ_L2_SIZE; i++) 142 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count; 143 for (i = 0; i < PQ_L2_SIZE; i++) 144 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count; 145 /* PQ_NONE has no queue */ 146 147 for (i = 0; i < PQ_COUNT; i++) { 148 TAILQ_INIT(&vm_page_queues[i].pl); 149 spin_init(&vm_page_queues[i].spin); 150 } 151 152 for (i = 0; i < VMACTION_HSIZE; i++) 153 LIST_INIT(&action_list[i]); 154 } 155 156 /* 157 * note: place in initialized data section? Is this necessary? 158 */ 159 long first_page = 0; 160 int vm_page_array_size = 0; 161 int vm_page_zero_count = 0; 162 vm_page_t vm_page_array = NULL; 163 vm_paddr_t vm_low_phys_reserved; 164 165 /* 166 * (low level boot) 167 * 168 * Sets the page size, perhaps based upon the memory size. 169 * Must be called before any use of page-size dependent functions. 170 */ 171 void 172 vm_set_page_size(void) 173 { 174 if (vmstats.v_page_size == 0) 175 vmstats.v_page_size = PAGE_SIZE; 176 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 177 panic("vm_set_page_size: page size not a power of two"); 178 } 179 180 /* 181 * (low level boot) 182 * 183 * Add a new page to the freelist for use by the system. New pages 184 * are added to both the head and tail of the associated free page 185 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 186 * requests pull 'recent' adds (higher physical addresses) first. 187 * 188 * Beware that the page zeroing daemon will also be running soon after 189 * boot, moving pages from the head to the tail of the PQ_FREE queues. 190 * 191 * Must be called in a critical section. 192 */ 193 static void 194 vm_add_new_page(vm_paddr_t pa) 195 { 196 struct vpgqueues *vpq; 197 vm_page_t m; 198 199 m = PHYS_TO_VM_PAGE(pa); 200 m->phys_addr = pa; 201 m->flags = 0; 202 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 203 #ifdef SMP 204 /* 205 * Twist for cpu localization in addition to page coloring, so 206 * different cpus selecting by m->queue get different page colors. 207 */ 208 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK; 209 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK; 210 #endif 211 /* 212 * Reserve a certain number of contiguous low memory pages for 213 * contigmalloc() to use. 214 */ 215 if (pa < vm_low_phys_reserved) { 216 atomic_add_int(&vmstats.v_page_count, 1); 217 atomic_add_int(&vmstats.v_dma_pages, 1); 218 m->queue = PQ_NONE; 219 m->wire_count = 1; 220 atomic_add_int(&vmstats.v_wire_count, 1); 221 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 222 return; 223 } 224 225 /* 226 * General page 227 */ 228 m->queue = m->pc + PQ_FREE; 229 KKASSERT(m->dirty == 0); 230 231 atomic_add_int(&vmstats.v_page_count, 1); 232 atomic_add_int(&vmstats.v_free_count, 1); 233 vpq = &vm_page_queues[m->queue]; 234 if ((vpq->flipflop & 15) == 0) { 235 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 236 m->flags |= PG_ZERO; 237 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 238 atomic_add_int(&vm_page_zero_count, 1); 239 } else { 240 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 241 } 242 ++vpq->flipflop; 243 ++vpq->lcnt; 244 } 245 246 /* 247 * (low level boot) 248 * 249 * Initializes the resident memory module. 250 * 251 * Preallocates memory for critical VM structures and arrays prior to 252 * kernel_map becoming available. 253 * 254 * Memory is allocated from (virtual2_start, virtual2_end) if available, 255 * otherwise memory is allocated from (virtual_start, virtual_end). 256 * 257 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 258 * large enough to hold vm_page_array & other structures for machines with 259 * large amounts of ram, so we want to use virtual2* when available. 260 */ 261 void 262 vm_page_startup(void) 263 { 264 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 265 vm_offset_t mapped; 266 vm_size_t npages; 267 vm_paddr_t page_range; 268 vm_paddr_t new_end; 269 int i; 270 vm_paddr_t pa; 271 int nblocks; 272 vm_paddr_t last_pa; 273 vm_paddr_t end; 274 vm_paddr_t biggestone, biggestsize; 275 vm_paddr_t total; 276 277 total = 0; 278 biggestsize = 0; 279 biggestone = 0; 280 nblocks = 0; 281 vaddr = round_page(vaddr); 282 283 for (i = 0; phys_avail[i + 1]; i += 2) { 284 phys_avail[i] = round_page64(phys_avail[i]); 285 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 286 } 287 288 for (i = 0; phys_avail[i + 1]; i += 2) { 289 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 290 291 if (size > biggestsize) { 292 biggestone = i; 293 biggestsize = size; 294 } 295 ++nblocks; 296 total += size; 297 } 298 299 end = phys_avail[biggestone+1]; 300 end = trunc_page(end); 301 302 /* 303 * Initialize the queue headers for the free queue, the active queue 304 * and the inactive queue. 305 */ 306 vm_page_queue_init(); 307 308 #if !defined(_KERNEL_VIRTUAL) 309 /* 310 * VKERNELs don't support minidumps and as such don't need 311 * vm_page_dump 312 * 313 * Allocate a bitmap to indicate that a random physical page 314 * needs to be included in a minidump. 315 * 316 * The amd64 port needs this to indicate which direct map pages 317 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 318 * 319 * However, i386 still needs this workspace internally within the 320 * minidump code. In theory, they are not needed on i386, but are 321 * included should the sf_buf code decide to use them. 322 */ 323 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 324 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 325 end -= vm_page_dump_size; 326 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 327 VM_PROT_READ | VM_PROT_WRITE); 328 bzero((void *)vm_page_dump, vm_page_dump_size); 329 #endif 330 /* 331 * Compute the number of pages of memory that will be available for 332 * use (taking into account the overhead of a page structure per 333 * page). 334 */ 335 first_page = phys_avail[0] / PAGE_SIZE; 336 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 337 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 338 339 #ifndef _KERNEL_VIRTUAL 340 /* 341 * (only applies to real kernels) 342 * 343 * Initialize the contiguous reserve map. We initially reserve up 344 * to 1/4 available physical memory or 65536 pages (~256MB), whichever 345 * is lower. 346 * 347 * Once device initialization is complete we return most of the 348 * reserved memory back to the normal page queues but leave some 349 * in reserve for things like usb attachments. 350 */ 351 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 352 if (vm_low_phys_reserved > total / 4) 353 vm_low_phys_reserved = total / 4; 354 if (vm_dma_reserved == 0) { 355 vm_dma_reserved = 16 * 1024 * 1024; /* 16MB */ 356 if (vm_dma_reserved > total / 16) 357 vm_dma_reserved = total / 16; 358 } 359 #endif 360 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 361 ALIST_RECORDS_65536); 362 363 /* 364 * Initialize the mem entry structures now, and put them in the free 365 * queue. 366 */ 367 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 368 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 369 vm_page_array = (vm_page_t)mapped; 370 371 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 372 /* 373 * since pmap_map on amd64 returns stuff out of a direct-map region, 374 * we have to manually add these pages to the minidump tracking so 375 * that they can be dumped, including the vm_page_array. 376 */ 377 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 378 dump_add_page(pa); 379 #endif 380 381 /* 382 * Clear all of the page structures 383 */ 384 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 385 vm_page_array_size = page_range; 386 387 /* 388 * Construct the free queue(s) in ascending order (by physical 389 * address) so that the first 16MB of physical memory is allocated 390 * last rather than first. On large-memory machines, this avoids 391 * the exhaustion of low physical memory before isa_dmainit has run. 392 */ 393 vmstats.v_page_count = 0; 394 vmstats.v_free_count = 0; 395 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 396 pa = phys_avail[i]; 397 if (i == biggestone) 398 last_pa = new_end; 399 else 400 last_pa = phys_avail[i + 1]; 401 while (pa < last_pa && npages-- > 0) { 402 vm_add_new_page(pa); 403 pa += PAGE_SIZE; 404 } 405 } 406 if (virtual2_start) 407 virtual2_start = vaddr; 408 else 409 virtual_start = vaddr; 410 } 411 412 /* 413 * We tended to reserve a ton of memory for contigmalloc(). Now that most 414 * drivers have initialized we want to return most the remaining free 415 * reserve back to the VM page queues so they can be used for normal 416 * allocations. 417 * 418 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 419 */ 420 static void 421 vm_page_startup_finish(void *dummy __unused) 422 { 423 alist_blk_t blk; 424 alist_blk_t rblk; 425 alist_blk_t count; 426 alist_blk_t xcount; 427 alist_blk_t bfree; 428 vm_page_t m; 429 430 spin_lock(&vm_contig_spin); 431 for (;;) { 432 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 433 if (bfree <= vm_dma_reserved / PAGE_SIZE) 434 break; 435 if (count == 0) 436 break; 437 438 /* 439 * Figure out how much of the initial reserve we have to 440 * free in order to reach our target. 441 */ 442 bfree -= vm_dma_reserved / PAGE_SIZE; 443 if (count > bfree) { 444 blk += count - bfree; 445 count = bfree; 446 } 447 448 /* 449 * Calculate the nearest power of 2 <= count. 450 */ 451 for (xcount = 1; xcount <= count; xcount <<= 1) 452 ; 453 xcount >>= 1; 454 blk += count - xcount; 455 count = xcount; 456 457 /* 458 * Allocate the pages from the alist, then free them to 459 * the normal VM page queues. 460 * 461 * Pages allocated from the alist are wired. We have to 462 * busy, unwire, and free them. We must also adjust 463 * vm_low_phys_reserved before freeing any pages to prevent 464 * confusion. 465 */ 466 rblk = alist_alloc(&vm_contig_alist, blk, count); 467 if (rblk != blk) { 468 kprintf("vm_page_startup_finish: Unable to return " 469 "dma space @0x%08x/%d -> 0x%08x\n", 470 blk, count, rblk); 471 break; 472 } 473 atomic_add_int(&vmstats.v_dma_pages, -count); 474 spin_unlock(&vm_contig_spin); 475 476 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 477 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 478 while (count) { 479 vm_page_busy_wait(m, FALSE, "cpgfr"); 480 vm_page_unwire(m, 0); 481 vm_page_free(m); 482 --count; 483 ++m; 484 } 485 spin_lock(&vm_contig_spin); 486 } 487 spin_unlock(&vm_contig_spin); 488 489 /* 490 * Print out how much DMA space drivers have already allocated and 491 * how much is left over. 492 */ 493 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 494 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 495 (PAGE_SIZE / 1024), 496 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 497 } 498 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 499 vm_page_startup_finish, NULL) 500 501 502 /* 503 * Scan comparison function for Red-Black tree scans. An inclusive 504 * (start,end) is expected. Other fields are not used. 505 */ 506 int 507 rb_vm_page_scancmp(struct vm_page *p, void *data) 508 { 509 struct rb_vm_page_scan_info *info = data; 510 511 if (p->pindex < info->start_pindex) 512 return(-1); 513 if (p->pindex > info->end_pindex) 514 return(1); 515 return(0); 516 } 517 518 int 519 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 520 { 521 if (p1->pindex < p2->pindex) 522 return(-1); 523 if (p1->pindex > p2->pindex) 524 return(1); 525 return(0); 526 } 527 528 /* 529 * Each page queue has its own spin lock, which is fairly optimal for 530 * allocating and freeing pages at least. 531 * 532 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 533 * queue spinlock via this function. Also note that m->queue cannot change 534 * unless both the page and queue are locked. 535 */ 536 static __inline 537 void 538 _vm_page_queue_spin_lock(vm_page_t m) 539 { 540 u_short queue; 541 542 queue = m->queue; 543 if (queue != PQ_NONE) { 544 spin_lock(&vm_page_queues[queue].spin); 545 KKASSERT(queue == m->queue); 546 } 547 } 548 549 static __inline 550 void 551 _vm_page_queue_spin_unlock(vm_page_t m) 552 { 553 u_short queue; 554 555 queue = m->queue; 556 cpu_ccfence(); 557 if (queue != PQ_NONE) 558 spin_unlock(&vm_page_queues[queue].spin); 559 } 560 561 static __inline 562 void 563 _vm_page_queues_spin_lock(u_short queue) 564 { 565 cpu_ccfence(); 566 if (queue != PQ_NONE) 567 spin_lock(&vm_page_queues[queue].spin); 568 } 569 570 571 static __inline 572 void 573 _vm_page_queues_spin_unlock(u_short queue) 574 { 575 cpu_ccfence(); 576 if (queue != PQ_NONE) 577 spin_unlock(&vm_page_queues[queue].spin); 578 } 579 580 void 581 vm_page_queue_spin_lock(vm_page_t m) 582 { 583 _vm_page_queue_spin_lock(m); 584 } 585 586 void 587 vm_page_queues_spin_lock(u_short queue) 588 { 589 _vm_page_queues_spin_lock(queue); 590 } 591 592 void 593 vm_page_queue_spin_unlock(vm_page_t m) 594 { 595 _vm_page_queue_spin_unlock(m); 596 } 597 598 void 599 vm_page_queues_spin_unlock(u_short queue) 600 { 601 _vm_page_queues_spin_unlock(queue); 602 } 603 604 /* 605 * This locks the specified vm_page and its queue in the proper order 606 * (page first, then queue). The queue may change so the caller must 607 * recheck on return. 608 */ 609 static __inline 610 void 611 _vm_page_and_queue_spin_lock(vm_page_t m) 612 { 613 vm_page_spin_lock(m); 614 _vm_page_queue_spin_lock(m); 615 } 616 617 static __inline 618 void 619 _vm_page_and_queue_spin_unlock(vm_page_t m) 620 { 621 _vm_page_queues_spin_unlock(m->queue); 622 vm_page_spin_unlock(m); 623 } 624 625 void 626 vm_page_and_queue_spin_unlock(vm_page_t m) 627 { 628 _vm_page_and_queue_spin_unlock(m); 629 } 630 631 void 632 vm_page_and_queue_spin_lock(vm_page_t m) 633 { 634 _vm_page_and_queue_spin_lock(m); 635 } 636 637 /* 638 * Helper function removes vm_page from its current queue. 639 * Returns the base queue the page used to be on. 640 * 641 * The vm_page and the queue must be spinlocked. 642 * This function will unlock the queue but leave the page spinlocked. 643 */ 644 static __inline u_short 645 _vm_page_rem_queue_spinlocked(vm_page_t m) 646 { 647 struct vpgqueues *pq; 648 u_short queue; 649 650 queue = m->queue; 651 if (queue != PQ_NONE) { 652 pq = &vm_page_queues[queue]; 653 TAILQ_REMOVE(&pq->pl, m, pageq); 654 atomic_add_int(pq->cnt, -1); 655 pq->lcnt--; 656 m->queue = PQ_NONE; 657 vm_page_queues_spin_unlock(queue); 658 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO)) 659 atomic_subtract_int(&vm_page_zero_count, 1); 660 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 661 return (queue - m->pc); 662 } 663 return queue; 664 } 665 666 /* 667 * Helper function places the vm_page on the specified queue. 668 * 669 * The vm_page must be spinlocked. 670 * This function will return with both the page and the queue locked. 671 */ 672 static __inline void 673 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 674 { 675 struct vpgqueues *pq; 676 677 KKASSERT(m->queue == PQ_NONE); 678 679 if (queue != PQ_NONE) { 680 vm_page_queues_spin_lock(queue); 681 pq = &vm_page_queues[queue]; 682 ++pq->lcnt; 683 atomic_add_int(pq->cnt, 1); 684 m->queue = queue; 685 686 /* 687 * Put zero'd pages on the end ( where we look for zero'd pages 688 * first ) and non-zerod pages at the head. 689 */ 690 if (queue - m->pc == PQ_FREE) { 691 if (m->flags & PG_ZERO) { 692 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 693 atomic_add_int(&vm_page_zero_count, 1); 694 } else { 695 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 696 } 697 } else if (athead) { 698 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 699 } else { 700 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 701 } 702 /* leave the queue spinlocked */ 703 } 704 } 705 706 /* 707 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 708 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we 709 * did not. Only one sleep call will be made before returning. 710 * 711 * This function does NOT busy the page and on return the page is not 712 * guaranteed to be available. 713 */ 714 void 715 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 716 { 717 u_int32_t flags; 718 719 for (;;) { 720 flags = m->flags; 721 cpu_ccfence(); 722 723 if ((flags & PG_BUSY) == 0 && 724 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) { 725 break; 726 } 727 tsleep_interlock(m, 0); 728 if (atomic_cmpset_int(&m->flags, flags, 729 flags | PG_WANTED | PG_REFERENCED)) { 730 tsleep(m, PINTERLOCKED, msg, 0); 731 break; 732 } 733 } 734 } 735 736 /* 737 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we 738 * also wait for m->busy to become 0 before setting PG_BUSY. 739 */ 740 void 741 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 742 int also_m_busy, const char *msg 743 VM_PAGE_DEBUG_ARGS) 744 { 745 u_int32_t flags; 746 747 for (;;) { 748 flags = m->flags; 749 cpu_ccfence(); 750 if (flags & PG_BUSY) { 751 tsleep_interlock(m, 0); 752 if (atomic_cmpset_int(&m->flags, flags, 753 flags | PG_WANTED | PG_REFERENCED)) { 754 tsleep(m, PINTERLOCKED, msg, 0); 755 } 756 } else if (also_m_busy && (flags & PG_SBUSY)) { 757 tsleep_interlock(m, 0); 758 if (atomic_cmpset_int(&m->flags, flags, 759 flags | PG_WANTED | PG_REFERENCED)) { 760 tsleep(m, PINTERLOCKED, msg, 0); 761 } 762 } else { 763 if (atomic_cmpset_int(&m->flags, flags, 764 flags | PG_BUSY)) { 765 #ifdef VM_PAGE_DEBUG 766 m->busy_func = func; 767 m->busy_line = lineno; 768 #endif 769 break; 770 } 771 } 772 } 773 } 774 775 /* 776 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy 777 * is also 0. 778 * 779 * Returns non-zero on failure. 780 */ 781 int 782 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 783 VM_PAGE_DEBUG_ARGS) 784 { 785 u_int32_t flags; 786 787 for (;;) { 788 flags = m->flags; 789 cpu_ccfence(); 790 if (flags & PG_BUSY) 791 return TRUE; 792 if (also_m_busy && (flags & PG_SBUSY)) 793 return TRUE; 794 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 795 #ifdef VM_PAGE_DEBUG 796 m->busy_func = func; 797 m->busy_line = lineno; 798 #endif 799 return FALSE; 800 } 801 } 802 } 803 804 /* 805 * Clear the PG_BUSY flag and return non-zero to indicate to the caller 806 * that a wakeup() should be performed. 807 * 808 * The vm_page must be spinlocked and will remain spinlocked on return. 809 * The related queue must NOT be spinlocked (which could deadlock us). 810 * 811 * (inline version) 812 */ 813 static __inline 814 int 815 _vm_page_wakeup(vm_page_t m) 816 { 817 u_int32_t flags; 818 819 for (;;) { 820 flags = m->flags; 821 cpu_ccfence(); 822 if (atomic_cmpset_int(&m->flags, flags, 823 flags & ~(PG_BUSY | PG_WANTED))) { 824 break; 825 } 826 } 827 return(flags & PG_WANTED); 828 } 829 830 /* 831 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This 832 * is typically the last call you make on a page before moving onto 833 * other things. 834 */ 835 void 836 vm_page_wakeup(vm_page_t m) 837 { 838 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 839 vm_page_spin_lock(m); 840 if (_vm_page_wakeup(m)) { 841 vm_page_spin_unlock(m); 842 wakeup(m); 843 } else { 844 vm_page_spin_unlock(m); 845 } 846 } 847 848 /* 849 * Holding a page keeps it from being reused. Other parts of the system 850 * can still disassociate the page from its current object and free it, or 851 * perform read or write I/O on it and/or otherwise manipulate the page, 852 * but if the page is held the VM system will leave the page and its data 853 * intact and not reuse the page for other purposes until the last hold 854 * reference is released. (see vm_page_wire() if you want to prevent the 855 * page from being disassociated from its object too). 856 * 857 * The caller must still validate the contents of the page and, if necessary, 858 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 859 * before manipulating the page. 860 * 861 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 862 */ 863 void 864 vm_page_hold(vm_page_t m) 865 { 866 vm_page_spin_lock(m); 867 atomic_add_int(&m->hold_count, 1); 868 if (m->queue - m->pc == PQ_FREE) { 869 _vm_page_queue_spin_lock(m); 870 _vm_page_rem_queue_spinlocked(m); 871 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 872 _vm_page_queue_spin_unlock(m); 873 } 874 vm_page_spin_unlock(m); 875 } 876 877 /* 878 * The opposite of vm_page_hold(). A page can be freed while being held, 879 * which places it on the PQ_HOLD queue. If we are able to busy the page 880 * after the hold count drops to zero we will move the page to the 881 * appropriate PQ_FREE queue by calling vm_page_free_toq(). 882 */ 883 void 884 vm_page_unhold(vm_page_t m) 885 { 886 vm_page_spin_lock(m); 887 atomic_add_int(&m->hold_count, -1); 888 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 889 _vm_page_queue_spin_lock(m); 890 _vm_page_rem_queue_spinlocked(m); 891 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 892 _vm_page_queue_spin_unlock(m); 893 } 894 vm_page_spin_unlock(m); 895 } 896 897 /* 898 * Inserts the given vm_page into the object and object list. 899 * 900 * The pagetables are not updated but will presumably fault the page 901 * in if necessary, or if a kernel page the caller will at some point 902 * enter the page into the kernel's pmap. We are not allowed to block 903 * here so we *can't* do this anyway. 904 * 905 * This routine may not block. 906 * This routine must be called with the vm_object held. 907 * This routine must be called with a critical section held. 908 * 909 * This routine returns TRUE if the page was inserted into the object 910 * successfully, and FALSE if the page already exists in the object. 911 */ 912 int 913 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 914 { 915 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 916 if (m->object != NULL) 917 panic("vm_page_insert: already inserted"); 918 919 object->generation++; 920 921 /* 922 * Record the object/offset pair in this page and add the 923 * pv_list_count of the page to the object. 924 * 925 * The vm_page spin lock is required for interactions with the pmap. 926 */ 927 vm_page_spin_lock(m); 928 m->object = object; 929 m->pindex = pindex; 930 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 931 m->object = NULL; 932 m->pindex = 0; 933 vm_page_spin_unlock(m); 934 return FALSE; 935 } 936 object->resident_page_count++; 937 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */ 938 vm_page_spin_unlock(m); 939 940 /* 941 * Since we are inserting a new and possibly dirty page, 942 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 943 */ 944 if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE)) 945 vm_object_set_writeable_dirty(object); 946 947 /* 948 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 949 */ 950 swap_pager_page_inserted(m); 951 return TRUE; 952 } 953 954 /* 955 * Removes the given vm_page_t from the (object,index) table 956 * 957 * The underlying pmap entry (if any) is NOT removed here. 958 * This routine may not block. 959 * 960 * The page must be BUSY and will remain BUSY on return. 961 * No other requirements. 962 * 963 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 964 * it busy. 965 */ 966 void 967 vm_page_remove(vm_page_t m) 968 { 969 vm_object_t object; 970 971 if (m->object == NULL) { 972 return; 973 } 974 975 if ((m->flags & PG_BUSY) == 0) 976 panic("vm_page_remove: page not busy"); 977 978 object = m->object; 979 980 vm_object_hold(object); 981 982 /* 983 * Remove the page from the object and update the object. 984 * 985 * The vm_page spin lock is required for interactions with the pmap. 986 */ 987 vm_page_spin_lock(m); 988 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 989 object->resident_page_count--; 990 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */ 991 m->object = NULL; 992 vm_page_spin_unlock(m); 993 994 object->generation++; 995 996 vm_object_drop(object); 997 } 998 999 /* 1000 * Locate and return the page at (object, pindex), or NULL if the 1001 * page could not be found. 1002 * 1003 * The caller must hold the vm_object token. 1004 */ 1005 vm_page_t 1006 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1007 { 1008 vm_page_t m; 1009 1010 /* 1011 * Search the hash table for this object/offset pair 1012 */ 1013 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1014 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1015 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1016 return(m); 1017 } 1018 1019 vm_page_t 1020 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1021 vm_pindex_t pindex, 1022 int also_m_busy, const char *msg 1023 VM_PAGE_DEBUG_ARGS) 1024 { 1025 u_int32_t flags; 1026 vm_page_t m; 1027 1028 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1029 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1030 while (m) { 1031 KKASSERT(m->object == object && m->pindex == pindex); 1032 flags = m->flags; 1033 cpu_ccfence(); 1034 if (flags & PG_BUSY) { 1035 tsleep_interlock(m, 0); 1036 if (atomic_cmpset_int(&m->flags, flags, 1037 flags | PG_WANTED | PG_REFERENCED)) { 1038 tsleep(m, PINTERLOCKED, msg, 0); 1039 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1040 pindex); 1041 } 1042 } else if (also_m_busy && (flags & PG_SBUSY)) { 1043 tsleep_interlock(m, 0); 1044 if (atomic_cmpset_int(&m->flags, flags, 1045 flags | PG_WANTED | PG_REFERENCED)) { 1046 tsleep(m, PINTERLOCKED, msg, 0); 1047 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1048 pindex); 1049 } 1050 } else if (atomic_cmpset_int(&m->flags, flags, 1051 flags | PG_BUSY)) { 1052 #ifdef VM_PAGE_DEBUG 1053 m->busy_func = func; 1054 m->busy_line = lineno; 1055 #endif 1056 break; 1057 } 1058 } 1059 return m; 1060 } 1061 1062 /* 1063 * Attempt to lookup and busy a page. 1064 * 1065 * Returns NULL if the page could not be found 1066 * 1067 * Returns a vm_page and error == TRUE if the page exists but could not 1068 * be busied. 1069 * 1070 * Returns a vm_page and error == FALSE on success. 1071 */ 1072 vm_page_t 1073 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1074 vm_pindex_t pindex, 1075 int also_m_busy, int *errorp 1076 VM_PAGE_DEBUG_ARGS) 1077 { 1078 u_int32_t flags; 1079 vm_page_t m; 1080 1081 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1082 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1083 *errorp = FALSE; 1084 while (m) { 1085 KKASSERT(m->object == object && m->pindex == pindex); 1086 flags = m->flags; 1087 cpu_ccfence(); 1088 if (flags & PG_BUSY) { 1089 *errorp = TRUE; 1090 break; 1091 } 1092 if (also_m_busy && (flags & PG_SBUSY)) { 1093 *errorp = TRUE; 1094 break; 1095 } 1096 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1097 #ifdef VM_PAGE_DEBUG 1098 m->busy_func = func; 1099 m->busy_line = lineno; 1100 #endif 1101 break; 1102 } 1103 } 1104 return m; 1105 } 1106 1107 /* 1108 * Caller must hold the related vm_object 1109 */ 1110 vm_page_t 1111 vm_page_next(vm_page_t m) 1112 { 1113 vm_page_t next; 1114 1115 next = vm_page_rb_tree_RB_NEXT(m); 1116 if (next && next->pindex != m->pindex + 1) 1117 next = NULL; 1118 return (next); 1119 } 1120 1121 /* 1122 * vm_page_rename() 1123 * 1124 * Move the given vm_page from its current object to the specified 1125 * target object/offset. The page must be busy and will remain so 1126 * on return. 1127 * 1128 * new_object must be held. 1129 * This routine might block. XXX ? 1130 * 1131 * NOTE: Swap associated with the page must be invalidated by the move. We 1132 * have to do this for several reasons: (1) we aren't freeing the 1133 * page, (2) we are dirtying the page, (3) the VM system is probably 1134 * moving the page from object A to B, and will then later move 1135 * the backing store from A to B and we can't have a conflict. 1136 * 1137 * NOTE: We *always* dirty the page. It is necessary both for the 1138 * fact that we moved it, and because we may be invalidating 1139 * swap. If the page is on the cache, we have to deactivate it 1140 * or vm_page_dirty() will panic. Dirty pages are not allowed 1141 * on the cache. 1142 */ 1143 void 1144 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1145 { 1146 KKASSERT(m->flags & PG_BUSY); 1147 ASSERT_LWKT_TOKEN_HELD(vm_object_token(new_object)); 1148 if (m->object) { 1149 ASSERT_LWKT_TOKEN_HELD(vm_object_token(m->object)); 1150 vm_page_remove(m); 1151 } 1152 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1153 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1154 new_object, new_pindex); 1155 } 1156 if (m->queue - m->pc == PQ_CACHE) 1157 vm_page_deactivate(m); 1158 vm_page_dirty(m); 1159 } 1160 1161 /* 1162 * vm_page_unqueue() without any wakeup. This routine is used when a page 1163 * is being moved between queues or otherwise is to remain BUSYied by the 1164 * caller. 1165 * 1166 * This routine may not block. 1167 */ 1168 void 1169 vm_page_unqueue_nowakeup(vm_page_t m) 1170 { 1171 vm_page_and_queue_spin_lock(m); 1172 (void)_vm_page_rem_queue_spinlocked(m); 1173 vm_page_spin_unlock(m); 1174 } 1175 1176 /* 1177 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1178 * if necessary. 1179 * 1180 * This routine may not block. 1181 */ 1182 void 1183 vm_page_unqueue(vm_page_t m) 1184 { 1185 u_short queue; 1186 1187 vm_page_and_queue_spin_lock(m); 1188 queue = _vm_page_rem_queue_spinlocked(m); 1189 if (queue == PQ_FREE || queue == PQ_CACHE) { 1190 vm_page_spin_unlock(m); 1191 pagedaemon_wakeup(); 1192 } else { 1193 vm_page_spin_unlock(m); 1194 } 1195 } 1196 1197 /* 1198 * vm_page_list_find() 1199 * 1200 * Find a page on the specified queue with color optimization. 1201 * 1202 * The page coloring optimization attempts to locate a page that does 1203 * not overload other nearby pages in the object in the cpu's L1 or L2 1204 * caches. We need this optimization because cpu caches tend to be 1205 * physical caches, while object spaces tend to be virtual. 1206 * 1207 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1208 * and the algorithm is adjusted to localize allocations on a per-core basis. 1209 * This is done by 'twisting' the colors. 1210 * 1211 * The page is returned spinlocked and removed from its queue (it will 1212 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller 1213 * is responsible for dealing with the busy-page case (usually by 1214 * deactivating the page and looping). 1215 * 1216 * NOTE: This routine is carefully inlined. A non-inlined version 1217 * is available for outside callers but the only critical path is 1218 * from within this source file. 1219 * 1220 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1221 * represent stable storage, allowing us to order our locks vm_page 1222 * first, then queue. 1223 */ 1224 static __inline 1225 vm_page_t 1226 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1227 { 1228 vm_page_t m; 1229 1230 for (;;) { 1231 if (prefer_zero) 1232 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 1233 else 1234 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1235 if (m == NULL) { 1236 m = _vm_page_list_find2(basequeue, index); 1237 return(m); 1238 } 1239 vm_page_and_queue_spin_lock(m); 1240 if (m->queue == basequeue + index) { 1241 _vm_page_rem_queue_spinlocked(m); 1242 /* vm_page_t spin held, no queue spin */ 1243 break; 1244 } 1245 vm_page_and_queue_spin_unlock(m); 1246 } 1247 return(m); 1248 } 1249 1250 static vm_page_t 1251 _vm_page_list_find2(int basequeue, int index) 1252 { 1253 int i; 1254 vm_page_t m = NULL; 1255 struct vpgqueues *pq; 1256 1257 pq = &vm_page_queues[basequeue]; 1258 1259 /* 1260 * Note that for the first loop, index+i and index-i wind up at the 1261 * same place. Even though this is not totally optimal, we've already 1262 * blown it by missing the cache case so we do not care. 1263 */ 1264 for (i = PQ_L2_SIZE / 2; i > 0; --i) { 1265 for (;;) { 1266 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl); 1267 if (m) { 1268 _vm_page_and_queue_spin_lock(m); 1269 if (m->queue == 1270 basequeue + ((index + i) & PQ_L2_MASK)) { 1271 _vm_page_rem_queue_spinlocked(m); 1272 return(m); 1273 } 1274 _vm_page_and_queue_spin_unlock(m); 1275 continue; 1276 } 1277 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl); 1278 if (m) { 1279 _vm_page_and_queue_spin_lock(m); 1280 if (m->queue == 1281 basequeue + ((index - i) & PQ_L2_MASK)) { 1282 _vm_page_rem_queue_spinlocked(m); 1283 return(m); 1284 } 1285 _vm_page_and_queue_spin_unlock(m); 1286 continue; 1287 } 1288 break; /* next i */ 1289 } 1290 } 1291 return(m); 1292 } 1293 1294 /* 1295 * Returns a vm_page candidate for allocation. The page is not busied so 1296 * it can move around. The caller must busy the page (and typically 1297 * deactivate it if it cannot be busied!) 1298 * 1299 * Returns a spinlocked vm_page that has been removed from its queue. 1300 */ 1301 vm_page_t 1302 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1303 { 1304 return(_vm_page_list_find(basequeue, index, prefer_zero)); 1305 } 1306 1307 /* 1308 * Find a page on the cache queue with color optimization, remove it 1309 * from the queue, and busy it. The returned page will not be spinlocked. 1310 * 1311 * A candidate failure will be deactivated. Candidates can fail due to 1312 * being busied by someone else, in which case they will be deactivated. 1313 * 1314 * This routine may not block. 1315 * 1316 */ 1317 static vm_page_t 1318 vm_page_select_cache(u_short pg_color) 1319 { 1320 vm_page_t m; 1321 1322 for (;;) { 1323 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE); 1324 if (m == NULL) 1325 break; 1326 /* 1327 * (m) has been removed from its queue and spinlocked 1328 */ 1329 if (vm_page_busy_try(m, TRUE)) { 1330 _vm_page_deactivate_locked(m, 0); 1331 vm_page_spin_unlock(m); 1332 #ifdef INVARIANTS 1333 kprintf("Warning: busy page %p found in cache\n", m); 1334 #endif 1335 } else { 1336 /* 1337 * We successfully busied the page 1338 */ 1339 if ((m->flags & PG_UNMANAGED) == 0 && 1340 m->hold_count == 0 && 1341 m->wire_count == 0) { 1342 vm_page_spin_unlock(m); 1343 pagedaemon_wakeup(); 1344 return(m); 1345 } 1346 _vm_page_deactivate_locked(m, 0); 1347 if (_vm_page_wakeup(m)) { 1348 vm_page_spin_unlock(m); 1349 wakeup(m); 1350 } else { 1351 vm_page_spin_unlock(m); 1352 } 1353 } 1354 } 1355 return (m); 1356 } 1357 1358 /* 1359 * Find a free or zero page, with specified preference. We attempt to 1360 * inline the nominal case and fall back to _vm_page_select_free() 1361 * otherwise. A busied page is removed from the queue and returned. 1362 * 1363 * This routine may not block. 1364 */ 1365 static __inline vm_page_t 1366 vm_page_select_free(u_short pg_color, boolean_t prefer_zero) 1367 { 1368 vm_page_t m; 1369 1370 for (;;) { 1371 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK, 1372 prefer_zero); 1373 if (m == NULL) 1374 break; 1375 if (vm_page_busy_try(m, TRUE)) { 1376 /* 1377 * Various mechanisms such as a pmap_collect can 1378 * result in a busy page on the free queue. We 1379 * have to move the page out of the way so we can 1380 * retry the allocation. If the other thread is not 1381 * allocating the page then m->valid will remain 0 and 1382 * the pageout daemon will free the page later on. 1383 * 1384 * Since we could not busy the page, however, we 1385 * cannot make assumptions as to whether the page 1386 * will be allocated by the other thread or not, 1387 * so all we can do is deactivate it to move it out 1388 * of the way. In particular, if the other thread 1389 * wires the page it may wind up on the inactive 1390 * queue and the pageout daemon will have to deal 1391 * with that case too. 1392 */ 1393 _vm_page_deactivate_locked(m, 0); 1394 vm_page_spin_unlock(m); 1395 #ifdef INVARIANTS 1396 kprintf("Warning: busy page %p found in cache\n", m); 1397 #endif 1398 } else { 1399 /* 1400 * Theoretically if we are able to busy the page 1401 * atomic with the queue removal (using the vm_page 1402 * lock) nobody else should be able to mess with the 1403 * page before us. 1404 */ 1405 KKASSERT((m->flags & PG_UNMANAGED) == 0); 1406 KKASSERT(m->hold_count == 0); 1407 KKASSERT(m->wire_count == 0); 1408 vm_page_spin_unlock(m); 1409 pagedaemon_wakeup(); 1410 1411 /* return busied and removed page */ 1412 return(m); 1413 } 1414 } 1415 return(m); 1416 } 1417 1418 /* 1419 * This implements a per-cpu cache of free, zero'd, ready-to-go pages. 1420 * The idea is to populate this cache prior to acquiring any locks so 1421 * we don't wind up potentially zeroing VM pages (under heavy loads) while 1422 * holding potentialy contending locks. 1423 * 1424 * Note that we allocate the page uninserted into anything and use a pindex 1425 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these 1426 * allocations should wind up being uncontended. However, we still want 1427 * to rove across PQ_L2_SIZE. 1428 */ 1429 void 1430 vm_page_pcpu_cache(void) 1431 { 1432 #if 0 1433 globaldata_t gd = mycpu; 1434 vm_page_t m; 1435 1436 if (gd->gd_vmpg_count < GD_MINVMPG) { 1437 crit_enter_gd(gd); 1438 while (gd->gd_vmpg_count < GD_MAXVMPG) { 1439 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask, 1440 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1441 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO); 1442 if (gd->gd_vmpg_count < GD_MAXVMPG) { 1443 if ((m->flags & PG_ZERO) == 0) { 1444 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1445 vm_page_flag_set(m, PG_ZERO); 1446 } 1447 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m; 1448 } else { 1449 vm_page_free(m); 1450 } 1451 } 1452 crit_exit_gd(gd); 1453 } 1454 #endif 1455 } 1456 1457 /* 1458 * vm_page_alloc() 1459 * 1460 * Allocate and return a memory cell associated with this VM object/offset 1461 * pair. If object is NULL an unassociated page will be allocated. 1462 * 1463 * The returned page will be busied and removed from its queues. This 1464 * routine can block and may return NULL if a race occurs and the page 1465 * is found to already exist at the specified (object, pindex). 1466 * 1467 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1468 * VM_ALLOC_QUICK like normal but cannot use cache 1469 * VM_ALLOC_SYSTEM greater free drain 1470 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1471 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1472 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1473 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1474 * (see vm_page_grab()) 1475 * VM_ALLOC_USE_GD ok to use per-gd cache 1476 * 1477 * The object must be held if not NULL 1478 * This routine may not block 1479 * 1480 * Additional special handling is required when called from an interrupt 1481 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1482 * in this case. 1483 */ 1484 vm_page_t 1485 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1486 { 1487 #ifdef SMP 1488 globaldata_t gd = mycpu; 1489 #endif 1490 vm_object_t obj; 1491 vm_page_t m; 1492 u_short pg_color; 1493 1494 #if 0 1495 /* 1496 * Special per-cpu free VM page cache. The pages are pre-busied 1497 * and pre-zerod for us. 1498 */ 1499 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1500 crit_enter_gd(gd); 1501 if (gd->gd_vmpg_count) { 1502 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1503 crit_exit_gd(gd); 1504 goto done; 1505 } 1506 crit_exit_gd(gd); 1507 } 1508 #endif 1509 m = NULL; 1510 1511 #ifdef SMP 1512 /* 1513 * Cpu twist - cpu localization algorithm 1514 */ 1515 if (object) { 1516 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) + 1517 (object->pg_color & ~ncpus_fit_mask); 1518 } else { 1519 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask); 1520 } 1521 #else 1522 /* 1523 * Normal page coloring algorithm 1524 */ 1525 if (object) { 1526 pg_color = object->pg_color + pindex; 1527 } else { 1528 pg_color = pindex; 1529 } 1530 #endif 1531 KKASSERT(page_req & 1532 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1533 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1534 1535 /* 1536 * Certain system threads (pageout daemon, buf_daemon's) are 1537 * allowed to eat deeper into the free page list. 1538 */ 1539 if (curthread->td_flags & TDF_SYSTHREAD) 1540 page_req |= VM_ALLOC_SYSTEM; 1541 1542 loop: 1543 if (vmstats.v_free_count > vmstats.v_free_reserved || 1544 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 1545 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 1546 vmstats.v_free_count > vmstats.v_interrupt_free_min) 1547 ) { 1548 /* 1549 * The free queue has sufficient free pages to take one out. 1550 */ 1551 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) 1552 m = vm_page_select_free(pg_color, TRUE); 1553 else 1554 m = vm_page_select_free(pg_color, FALSE); 1555 } else if (page_req & VM_ALLOC_NORMAL) { 1556 /* 1557 * Allocatable from the cache (non-interrupt only). On 1558 * success, we must free the page and try again, thus 1559 * ensuring that vmstats.v_*_free_min counters are replenished. 1560 */ 1561 #ifdef INVARIANTS 1562 if (curthread->td_preempted) { 1563 kprintf("vm_page_alloc(): warning, attempt to allocate" 1564 " cache page from preempting interrupt\n"); 1565 m = NULL; 1566 } else { 1567 m = vm_page_select_cache(pg_color); 1568 } 1569 #else 1570 m = vm_page_select_cache(pg_color); 1571 #endif 1572 /* 1573 * On success move the page into the free queue and loop. 1574 * 1575 * Only do this if we can safely acquire the vm_object lock, 1576 * because this is effectively a random page and the caller 1577 * might be holding the lock shared, we don't want to 1578 * deadlock. 1579 */ 1580 if (m != NULL) { 1581 KASSERT(m->dirty == 0, 1582 ("Found dirty cache page %p", m)); 1583 if ((obj = m->object) != NULL) { 1584 if (vm_object_hold_try(obj)) { 1585 vm_page_protect(m, VM_PROT_NONE); 1586 vm_page_free(m); 1587 /* m->object NULL here */ 1588 vm_object_drop(obj); 1589 } else { 1590 vm_page_deactivate(m); 1591 vm_page_wakeup(m); 1592 } 1593 } else { 1594 vm_page_protect(m, VM_PROT_NONE); 1595 vm_page_free(m); 1596 } 1597 goto loop; 1598 } 1599 1600 /* 1601 * On failure return NULL 1602 */ 1603 #if defined(DIAGNOSTIC) 1604 if (vmstats.v_cache_count > 0) 1605 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 1606 #endif 1607 vm_pageout_deficit++; 1608 pagedaemon_wakeup(); 1609 return (NULL); 1610 } else { 1611 /* 1612 * No pages available, wakeup the pageout daemon and give up. 1613 */ 1614 vm_pageout_deficit++; 1615 pagedaemon_wakeup(); 1616 return (NULL); 1617 } 1618 1619 /* 1620 * v_free_count can race so loop if we don't find the expected 1621 * page. 1622 */ 1623 if (m == NULL) 1624 goto loop; 1625 1626 /* 1627 * Good page found. The page has already been busied for us and 1628 * removed from its queues. 1629 */ 1630 KASSERT(m->dirty == 0, 1631 ("vm_page_alloc: free/cache page %p was dirty", m)); 1632 KKASSERT(m->queue == PQ_NONE); 1633 1634 #if 0 1635 done: 1636 #endif 1637 /* 1638 * Initialize the structure, inheriting some flags but clearing 1639 * all the rest. The page has already been busied for us. 1640 */ 1641 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY)); 1642 KKASSERT(m->wire_count == 0); 1643 KKASSERT(m->busy == 0); 1644 m->act_count = 0; 1645 m->valid = 0; 1646 1647 /* 1648 * Caller must be holding the object lock (asserted by 1649 * vm_page_insert()). 1650 * 1651 * NOTE: Inserting a page here does not insert it into any pmaps 1652 * (which could cause us to block allocating memory). 1653 * 1654 * NOTE: If no object an unassociated page is allocated, m->pindex 1655 * can be used by the caller for any purpose. 1656 */ 1657 if (object) { 1658 if (vm_page_insert(m, object, pindex) == FALSE) { 1659 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n", 1660 object, object->type, pindex); 1661 vm_page_free(m); 1662 m = NULL; 1663 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1664 panic("PAGE RACE"); 1665 } 1666 } else { 1667 m->pindex = pindex; 1668 } 1669 1670 /* 1671 * Don't wakeup too often - wakeup the pageout daemon when 1672 * we would be nearly out of memory. 1673 */ 1674 pagedaemon_wakeup(); 1675 1676 /* 1677 * A PG_BUSY page is returned. 1678 */ 1679 return (m); 1680 } 1681 1682 /* 1683 * Attempt to allocate contiguous physical memory with the specified 1684 * requirements. 1685 */ 1686 vm_page_t 1687 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1688 unsigned long alignment, unsigned long boundary, 1689 unsigned long size) 1690 { 1691 alist_blk_t blk; 1692 1693 alignment >>= PAGE_SHIFT; 1694 if (alignment == 0) 1695 alignment = 1; 1696 boundary >>= PAGE_SHIFT; 1697 if (boundary == 0) 1698 boundary = 1; 1699 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1700 1701 spin_lock(&vm_contig_spin); 1702 blk = alist_alloc(&vm_contig_alist, 0, size); 1703 if (blk == ALIST_BLOCK_NONE) { 1704 spin_unlock(&vm_contig_spin); 1705 if (bootverbose) { 1706 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1707 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1708 } 1709 return(NULL); 1710 } 1711 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1712 alist_free(&vm_contig_alist, blk, size); 1713 spin_unlock(&vm_contig_spin); 1714 if (bootverbose) { 1715 kprintf("vm_page_alloc_contig: %ldk high " 1716 "%016jx failed\n", 1717 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1718 (intmax_t)high); 1719 } 1720 return(NULL); 1721 } 1722 spin_unlock(&vm_contig_spin); 1723 if (bootverbose) { 1724 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 1725 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 1726 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1727 } 1728 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT)); 1729 } 1730 1731 /* 1732 * Free contiguously allocated pages. The pages will be wired but not busy. 1733 * When freeing to the alist we leave them wired and not busy. 1734 */ 1735 void 1736 vm_page_free_contig(vm_page_t m, unsigned long size) 1737 { 1738 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 1739 vm_pindex_t start = pa >> PAGE_SHIFT; 1740 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 1741 1742 if (bootverbose) { 1743 kprintf("vm_page_free_contig: %016jx/%ldk\n", 1744 (intmax_t)pa, size / 1024); 1745 } 1746 if (pa < vm_low_phys_reserved) { 1747 KKASSERT(pa + size <= vm_low_phys_reserved); 1748 spin_lock(&vm_contig_spin); 1749 alist_free(&vm_contig_alist, start, pages); 1750 spin_unlock(&vm_contig_spin); 1751 } else { 1752 while (pages) { 1753 vm_page_busy_wait(m, FALSE, "cpgfr"); 1754 vm_page_unwire(m, 0); 1755 vm_page_free(m); 1756 --pages; 1757 ++m; 1758 } 1759 1760 } 1761 } 1762 1763 1764 /* 1765 * Wait for sufficient free memory for nominal heavy memory use kernel 1766 * operations. 1767 */ 1768 void 1769 vm_wait_nominal(void) 1770 { 1771 while (vm_page_count_min(0)) 1772 vm_wait(0); 1773 } 1774 1775 /* 1776 * Test if vm_wait_nominal() would block. 1777 */ 1778 int 1779 vm_test_nominal(void) 1780 { 1781 if (vm_page_count_min(0)) 1782 return(1); 1783 return(0); 1784 } 1785 1786 /* 1787 * Block until free pages are available for allocation, called in various 1788 * places before memory allocations. 1789 * 1790 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 1791 * more generous then that. 1792 */ 1793 void 1794 vm_wait(int timo) 1795 { 1796 /* 1797 * never wait forever 1798 */ 1799 if (timo == 0) 1800 timo = hz; 1801 lwkt_gettoken(&vm_token); 1802 1803 if (curthread == pagethread) { 1804 /* 1805 * The pageout daemon itself needs pages, this is bad. 1806 */ 1807 if (vm_page_count_min(0)) { 1808 vm_pageout_pages_needed = 1; 1809 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 1810 } 1811 } else { 1812 /* 1813 * Wakeup the pageout daemon if necessary and wait. 1814 */ 1815 if (vm_page_count_target()) { 1816 if (vm_pages_needed == 0) { 1817 vm_pages_needed = 1; 1818 wakeup(&vm_pages_needed); 1819 } 1820 ++vm_pages_waiting; /* SMP race ok */ 1821 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 1822 } 1823 } 1824 lwkt_reltoken(&vm_token); 1825 } 1826 1827 /* 1828 * Block until free pages are available for allocation 1829 * 1830 * Called only from vm_fault so that processes page faulting can be 1831 * easily tracked. 1832 */ 1833 void 1834 vm_waitpfault(void) 1835 { 1836 /* 1837 * Wakeup the pageout daemon if necessary and wait. 1838 */ 1839 if (vm_page_count_target()) { 1840 lwkt_gettoken(&vm_token); 1841 if (vm_page_count_target()) { 1842 if (vm_pages_needed == 0) { 1843 vm_pages_needed = 1; 1844 wakeup(&vm_pages_needed); 1845 } 1846 ++vm_pages_waiting; /* SMP race ok */ 1847 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 1848 } 1849 lwkt_reltoken(&vm_token); 1850 } 1851 } 1852 1853 /* 1854 * Put the specified page on the active list (if appropriate). Ensure 1855 * that act_count is at least ACT_INIT but do not otherwise mess with it. 1856 * 1857 * The caller should be holding the page busied ? XXX 1858 * This routine may not block. 1859 */ 1860 void 1861 vm_page_activate(vm_page_t m) 1862 { 1863 u_short oqueue; 1864 1865 vm_page_spin_lock(m); 1866 if (m->queue - m->pc != PQ_ACTIVE) { 1867 _vm_page_queue_spin_lock(m); 1868 oqueue = _vm_page_rem_queue_spinlocked(m); 1869 /* page is left spinlocked, queue is unlocked */ 1870 1871 if (oqueue == PQ_CACHE) 1872 mycpu->gd_cnt.v_reactivated++; 1873 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1874 if (m->act_count < ACT_INIT) 1875 m->act_count = ACT_INIT; 1876 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 1877 } 1878 _vm_page_and_queue_spin_unlock(m); 1879 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 1880 pagedaemon_wakeup(); 1881 } else { 1882 if (m->act_count < ACT_INIT) 1883 m->act_count = ACT_INIT; 1884 vm_page_spin_unlock(m); 1885 } 1886 } 1887 1888 /* 1889 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1890 * routine is called when a page has been added to the cache or free 1891 * queues. 1892 * 1893 * This routine may not block. 1894 */ 1895 static __inline void 1896 vm_page_free_wakeup(void) 1897 { 1898 /* 1899 * If the pageout daemon itself needs pages, then tell it that 1900 * there are some free. 1901 */ 1902 if (vm_pageout_pages_needed && 1903 vmstats.v_cache_count + vmstats.v_free_count >= 1904 vmstats.v_pageout_free_min 1905 ) { 1906 wakeup(&vm_pageout_pages_needed); 1907 vm_pageout_pages_needed = 0; 1908 } 1909 1910 /* 1911 * Wakeup processes that are waiting on memory. 1912 * 1913 * NOTE: vm_paging_target() is the pageout daemon's target, while 1914 * vm_page_count_target() is somewhere inbetween. We want 1915 * to wake processes up prior to the pageout daemon reaching 1916 * its target to provide some hysteresis. 1917 */ 1918 if (vm_pages_waiting) { 1919 if (!vm_page_count_target()) { 1920 /* 1921 * Plenty of pages are free, wakeup everyone. 1922 */ 1923 vm_pages_waiting = 0; 1924 wakeup(&vmstats.v_free_count); 1925 ++mycpu->gd_cnt.v_ppwakeups; 1926 } else if (!vm_page_count_min(0)) { 1927 /* 1928 * Some pages are free, wakeup someone. 1929 */ 1930 int wcount = vm_pages_waiting; 1931 if (wcount > 0) 1932 --wcount; 1933 vm_pages_waiting = wcount; 1934 wakeup_one(&vmstats.v_free_count); 1935 ++mycpu->gd_cnt.v_ppwakeups; 1936 } 1937 } 1938 } 1939 1940 /* 1941 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 1942 * it from its VM object. 1943 * 1944 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1945 * return (the page will have been freed). 1946 */ 1947 void 1948 vm_page_free_toq(vm_page_t m) 1949 { 1950 mycpu->gd_cnt.v_tfree++; 1951 KKASSERT((m->flags & PG_MAPPED) == 0); 1952 KKASSERT(m->flags & PG_BUSY); 1953 1954 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1955 kprintf("vm_page_free: pindex(%lu), busy(%d), " 1956 "PG_BUSY(%d), hold(%d)\n", 1957 (u_long)m->pindex, m->busy, 1958 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 1959 if ((m->queue - m->pc) == PQ_FREE) 1960 panic("vm_page_free: freeing free page"); 1961 else 1962 panic("vm_page_free: freeing busy page"); 1963 } 1964 1965 /* 1966 * Remove from object, spinlock the page and its queues and 1967 * remove from any queue. No queue spinlock will be held 1968 * after this section (because the page was removed from any 1969 * queue). 1970 */ 1971 vm_page_remove(m); 1972 vm_page_and_queue_spin_lock(m); 1973 _vm_page_rem_queue_spinlocked(m); 1974 1975 /* 1976 * No further management of fictitious pages occurs beyond object 1977 * and queue removal. 1978 */ 1979 if ((m->flags & PG_FICTITIOUS) != 0) { 1980 vm_page_spin_unlock(m); 1981 vm_page_wakeup(m); 1982 return; 1983 } 1984 1985 m->valid = 0; 1986 vm_page_undirty(m); 1987 1988 if (m->wire_count != 0) { 1989 if (m->wire_count > 1) { 1990 panic( 1991 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1992 m->wire_count, (long)m->pindex); 1993 } 1994 panic("vm_page_free: freeing wired page"); 1995 } 1996 1997 /* 1998 * Clear the UNMANAGED flag when freeing an unmanaged page. 1999 */ 2000 if (m->flags & PG_UNMANAGED) { 2001 vm_page_flag_clear(m, PG_UNMANAGED); 2002 } 2003 2004 if (m->hold_count != 0) { 2005 vm_page_flag_clear(m, PG_ZERO); 2006 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2007 } else { 2008 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 2009 } 2010 2011 /* 2012 * This sequence allows us to clear PG_BUSY while still holding 2013 * its spin lock, which reduces contention vs allocators. We 2014 * must not leave the queue locked or _vm_page_wakeup() may 2015 * deadlock. 2016 */ 2017 _vm_page_queue_spin_unlock(m); 2018 if (_vm_page_wakeup(m)) { 2019 vm_page_spin_unlock(m); 2020 wakeup(m); 2021 } else { 2022 vm_page_spin_unlock(m); 2023 } 2024 vm_page_free_wakeup(); 2025 } 2026 2027 /* 2028 * vm_page_free_fromq_fast() 2029 * 2030 * Remove a non-zero page from one of the free queues; the page is removed for 2031 * zeroing, so do not issue a wakeup. 2032 */ 2033 vm_page_t 2034 vm_page_free_fromq_fast(void) 2035 { 2036 static int qi; 2037 vm_page_t m; 2038 int i; 2039 2040 for (i = 0; i < PQ_L2_SIZE; ++i) { 2041 m = vm_page_list_find(PQ_FREE, qi, FALSE); 2042 /* page is returned spinlocked and removed from its queue */ 2043 if (m) { 2044 if (vm_page_busy_try(m, TRUE)) { 2045 /* 2046 * We were unable to busy the page, deactivate 2047 * it and loop. 2048 */ 2049 _vm_page_deactivate_locked(m, 0); 2050 vm_page_spin_unlock(m); 2051 } else if (m->flags & PG_ZERO) { 2052 /* 2053 * The page is PG_ZERO, requeue it and loop 2054 */ 2055 _vm_page_add_queue_spinlocked(m, 2056 PQ_FREE + m->pc, 2057 0); 2058 vm_page_queue_spin_unlock(m); 2059 if (_vm_page_wakeup(m)) { 2060 vm_page_spin_unlock(m); 2061 wakeup(m); 2062 } else { 2063 vm_page_spin_unlock(m); 2064 } 2065 } else { 2066 /* 2067 * The page is not PG_ZERO'd so return it. 2068 */ 2069 vm_page_spin_unlock(m); 2070 KKASSERT((m->flags & PG_UNMANAGED) == 0); 2071 KKASSERT(m->hold_count == 0); 2072 KKASSERT(m->wire_count == 0); 2073 break; 2074 } 2075 m = NULL; 2076 } 2077 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 2078 } 2079 return (m); 2080 } 2081 2082 /* 2083 * vm_page_unmanage() 2084 * 2085 * Prevent PV management from being done on the page. The page is 2086 * removed from the paging queues as if it were wired, and as a 2087 * consequence of no longer being managed the pageout daemon will not 2088 * touch it (since there is no way to locate the pte mappings for the 2089 * page). madvise() calls that mess with the pmap will also no longer 2090 * operate on the page. 2091 * 2092 * Beyond that the page is still reasonably 'normal'. Freeing the page 2093 * will clear the flag. 2094 * 2095 * This routine is used by OBJT_PHYS objects - objects using unswappable 2096 * physical memory as backing store rather then swap-backed memory and 2097 * will eventually be extended to support 4MB unmanaged physical 2098 * mappings. 2099 * 2100 * Caller must be holding the page busy. 2101 */ 2102 void 2103 vm_page_unmanage(vm_page_t m) 2104 { 2105 KKASSERT(m->flags & PG_BUSY); 2106 if ((m->flags & PG_UNMANAGED) == 0) { 2107 if (m->wire_count == 0) 2108 vm_page_unqueue(m); 2109 } 2110 vm_page_flag_set(m, PG_UNMANAGED); 2111 } 2112 2113 /* 2114 * Mark this page as wired down by yet another map, removing it from 2115 * paging queues as necessary. 2116 * 2117 * Caller must be holding the page busy. 2118 */ 2119 void 2120 vm_page_wire(vm_page_t m) 2121 { 2122 /* 2123 * Only bump the wire statistics if the page is not already wired, 2124 * and only unqueue the page if it is on some queue (if it is unmanaged 2125 * it is already off the queues). Don't do anything with fictitious 2126 * pages because they are always wired. 2127 */ 2128 KKASSERT(m->flags & PG_BUSY); 2129 if ((m->flags & PG_FICTITIOUS) == 0) { 2130 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2131 if ((m->flags & PG_UNMANAGED) == 0) 2132 vm_page_unqueue(m); 2133 atomic_add_int(&vmstats.v_wire_count, 1); 2134 } 2135 KASSERT(m->wire_count != 0, 2136 ("vm_page_wire: wire_count overflow m=%p", m)); 2137 } 2138 } 2139 2140 /* 2141 * Release one wiring of this page, potentially enabling it to be paged again. 2142 * 2143 * Many pages placed on the inactive queue should actually go 2144 * into the cache, but it is difficult to figure out which. What 2145 * we do instead, if the inactive target is well met, is to put 2146 * clean pages at the head of the inactive queue instead of the tail. 2147 * This will cause them to be moved to the cache more quickly and 2148 * if not actively re-referenced, freed more quickly. If we just 2149 * stick these pages at the end of the inactive queue, heavy filesystem 2150 * meta-data accesses can cause an unnecessary paging load on memory bound 2151 * processes. This optimization causes one-time-use metadata to be 2152 * reused more quickly. 2153 * 2154 * BUT, if we are in a low-memory situation we have no choice but to 2155 * put clean pages on the cache queue. 2156 * 2157 * A number of routines use vm_page_unwire() to guarantee that the page 2158 * will go into either the inactive or active queues, and will NEVER 2159 * be placed in the cache - for example, just after dirtying a page. 2160 * dirty pages in the cache are not allowed. 2161 * 2162 * The page queues must be locked. 2163 * This routine may not block. 2164 */ 2165 void 2166 vm_page_unwire(vm_page_t m, int activate) 2167 { 2168 KKASSERT(m->flags & PG_BUSY); 2169 if (m->flags & PG_FICTITIOUS) { 2170 /* do nothing */ 2171 } else if (m->wire_count <= 0) { 2172 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2173 } else { 2174 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2175 atomic_add_int(&vmstats.v_wire_count, -1); 2176 if (m->flags & PG_UNMANAGED) { 2177 ; 2178 } else if (activate) { 2179 vm_page_spin_lock(m); 2180 _vm_page_add_queue_spinlocked(m, 2181 PQ_ACTIVE + m->pc, 0); 2182 _vm_page_and_queue_spin_unlock(m); 2183 } else { 2184 vm_page_spin_lock(m); 2185 vm_page_flag_clear(m, PG_WINATCFLS); 2186 _vm_page_add_queue_spinlocked(m, 2187 PQ_INACTIVE + m->pc, 0); 2188 ++vm_swapcache_inactive_heuristic; 2189 _vm_page_and_queue_spin_unlock(m); 2190 } 2191 } 2192 } 2193 } 2194 2195 /* 2196 * Move the specified page to the inactive queue. If the page has 2197 * any associated swap, the swap is deallocated. 2198 * 2199 * Normally athead is 0 resulting in LRU operation. athead is set 2200 * to 1 if we want this page to be 'as if it were placed in the cache', 2201 * except without unmapping it from the process address space. 2202 * 2203 * vm_page's spinlock must be held on entry and will remain held on return. 2204 * This routine may not block. 2205 */ 2206 static void 2207 _vm_page_deactivate_locked(vm_page_t m, int athead) 2208 { 2209 u_short oqueue; 2210 2211 /* 2212 * Ignore if already inactive. 2213 */ 2214 if (m->queue - m->pc == PQ_INACTIVE) 2215 return; 2216 _vm_page_queue_spin_lock(m); 2217 oqueue = _vm_page_rem_queue_spinlocked(m); 2218 2219 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2220 if (oqueue == PQ_CACHE) 2221 mycpu->gd_cnt.v_reactivated++; 2222 vm_page_flag_clear(m, PG_WINATCFLS); 2223 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2224 if (athead == 0) 2225 ++vm_swapcache_inactive_heuristic; 2226 } 2227 _vm_page_queue_spin_unlock(m); 2228 /* leaves vm_page spinlocked */ 2229 } 2230 2231 /* 2232 * Attempt to deactivate a page. 2233 * 2234 * No requirements. 2235 */ 2236 void 2237 vm_page_deactivate(vm_page_t m) 2238 { 2239 vm_page_spin_lock(m); 2240 _vm_page_deactivate_locked(m, 0); 2241 vm_page_spin_unlock(m); 2242 } 2243 2244 void 2245 vm_page_deactivate_locked(vm_page_t m) 2246 { 2247 _vm_page_deactivate_locked(m, 0); 2248 } 2249 2250 /* 2251 * Attempt to move a page to PQ_CACHE. 2252 * 2253 * Returns 0 on failure, 1 on success 2254 * 2255 * The page should NOT be busied by the caller. This function will validate 2256 * whether the page can be safely moved to the cache. 2257 */ 2258 int 2259 vm_page_try_to_cache(vm_page_t m) 2260 { 2261 vm_page_spin_lock(m); 2262 if (vm_page_busy_try(m, TRUE)) { 2263 vm_page_spin_unlock(m); 2264 return(0); 2265 } 2266 if (m->dirty || m->hold_count || m->wire_count || 2267 (m->flags & PG_UNMANAGED)) { 2268 if (_vm_page_wakeup(m)) { 2269 vm_page_spin_unlock(m); 2270 wakeup(m); 2271 } else { 2272 vm_page_spin_unlock(m); 2273 } 2274 return(0); 2275 } 2276 vm_page_spin_unlock(m); 2277 2278 /* 2279 * Page busied by us and no longer spinlocked. Dirty pages cannot 2280 * be moved to the cache. 2281 */ 2282 vm_page_test_dirty(m); 2283 if (m->dirty) { 2284 vm_page_wakeup(m); 2285 return(0); 2286 } 2287 vm_page_cache(m); 2288 return(1); 2289 } 2290 2291 /* 2292 * Attempt to free the page. If we cannot free it, we do nothing. 2293 * 1 is returned on success, 0 on failure. 2294 * 2295 * No requirements. 2296 */ 2297 int 2298 vm_page_try_to_free(vm_page_t m) 2299 { 2300 vm_page_spin_lock(m); 2301 if (vm_page_busy_try(m, TRUE)) { 2302 vm_page_spin_unlock(m); 2303 return(0); 2304 } 2305 2306 /* 2307 * The page can be in any state, including already being on the free 2308 * queue. Check to see if it really can be freed. 2309 */ 2310 if (m->dirty || /* can't free if it is dirty */ 2311 m->hold_count || /* or held (XXX may be wrong) */ 2312 m->wire_count || /* or wired */ 2313 (m->flags & PG_UNMANAGED) || /* or unmanaged */ 2314 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2315 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2316 if (_vm_page_wakeup(m)) { 2317 vm_page_spin_unlock(m); 2318 wakeup(m); 2319 } else { 2320 vm_page_spin_unlock(m); 2321 } 2322 return(0); 2323 } 2324 vm_page_spin_unlock(m); 2325 2326 /* 2327 * We can probably free the page. 2328 * 2329 * Page busied by us and no longer spinlocked. Dirty pages will 2330 * not be freed by this function. We have to re-test the 2331 * dirty bit after cleaning out the pmaps. 2332 */ 2333 vm_page_test_dirty(m); 2334 if (m->dirty) { 2335 vm_page_wakeup(m); 2336 return(0); 2337 } 2338 vm_page_protect(m, VM_PROT_NONE); 2339 if (m->dirty) { 2340 vm_page_wakeup(m); 2341 return(0); 2342 } 2343 vm_page_free(m); 2344 return(1); 2345 } 2346 2347 /* 2348 * vm_page_cache 2349 * 2350 * Put the specified page onto the page cache queue (if appropriate). 2351 * 2352 * The page must be busy, and this routine will release the busy and 2353 * possibly even free the page. 2354 */ 2355 void 2356 vm_page_cache(vm_page_t m) 2357 { 2358 if ((m->flags & PG_UNMANAGED) || m->busy || 2359 m->wire_count || m->hold_count) { 2360 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 2361 vm_page_wakeup(m); 2362 return; 2363 } 2364 2365 /* 2366 * Already in the cache (and thus not mapped) 2367 */ 2368 if ((m->queue - m->pc) == PQ_CACHE) { 2369 KKASSERT((m->flags & PG_MAPPED) == 0); 2370 vm_page_wakeup(m); 2371 return; 2372 } 2373 2374 /* 2375 * Caller is required to test m->dirty, but note that the act of 2376 * removing the page from its maps can cause it to become dirty 2377 * on an SMP system due to another cpu running in usermode. 2378 */ 2379 if (m->dirty) { 2380 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2381 (long)m->pindex); 2382 } 2383 2384 /* 2385 * Remove all pmaps and indicate that the page is not 2386 * writeable or mapped. Our vm_page_protect() call may 2387 * have blocked (especially w/ VM_PROT_NONE), so recheck 2388 * everything. 2389 */ 2390 vm_page_protect(m, VM_PROT_NONE); 2391 if ((m->flags & (PG_UNMANAGED|PG_MAPPED)) || m->busy || 2392 m->wire_count || m->hold_count) { 2393 vm_page_wakeup(m); 2394 } else if (m->dirty) { 2395 vm_page_deactivate(m); 2396 vm_page_wakeup(m); 2397 } else { 2398 _vm_page_and_queue_spin_lock(m); 2399 _vm_page_rem_queue_spinlocked(m); 2400 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2401 _vm_page_queue_spin_unlock(m); 2402 if (_vm_page_wakeup(m)) { 2403 vm_page_spin_unlock(m); 2404 wakeup(m); 2405 } else { 2406 vm_page_spin_unlock(m); 2407 } 2408 vm_page_free_wakeup(); 2409 } 2410 } 2411 2412 /* 2413 * vm_page_dontneed() 2414 * 2415 * Cache, deactivate, or do nothing as appropriate. This routine 2416 * is typically used by madvise() MADV_DONTNEED. 2417 * 2418 * Generally speaking we want to move the page into the cache so 2419 * it gets reused quickly. However, this can result in a silly syndrome 2420 * due to the page recycling too quickly. Small objects will not be 2421 * fully cached. On the otherhand, if we move the page to the inactive 2422 * queue we wind up with a problem whereby very large objects 2423 * unnecessarily blow away our inactive and cache queues. 2424 * 2425 * The solution is to move the pages based on a fixed weighting. We 2426 * either leave them alone, deactivate them, or move them to the cache, 2427 * where moving them to the cache has the highest weighting. 2428 * By forcing some pages into other queues we eventually force the 2429 * system to balance the queues, potentially recovering other unrelated 2430 * space from active. The idea is to not force this to happen too 2431 * often. 2432 * 2433 * The page must be busied. 2434 */ 2435 void 2436 vm_page_dontneed(vm_page_t m) 2437 { 2438 static int dnweight; 2439 int dnw; 2440 int head; 2441 2442 dnw = ++dnweight; 2443 2444 /* 2445 * occassionally leave the page alone 2446 */ 2447 if ((dnw & 0x01F0) == 0 || 2448 m->queue - m->pc == PQ_INACTIVE || 2449 m->queue - m->pc == PQ_CACHE 2450 ) { 2451 if (m->act_count >= ACT_INIT) 2452 --m->act_count; 2453 return; 2454 } 2455 2456 /* 2457 * If vm_page_dontneed() is inactivating a page, it must clear 2458 * the referenced flag; otherwise the pagedaemon will see references 2459 * on the page in the inactive queue and reactivate it. Until the 2460 * page can move to the cache queue, madvise's job is not done. 2461 */ 2462 vm_page_flag_clear(m, PG_REFERENCED); 2463 pmap_clear_reference(m); 2464 2465 if (m->dirty == 0) 2466 vm_page_test_dirty(m); 2467 2468 if (m->dirty || (dnw & 0x0070) == 0) { 2469 /* 2470 * Deactivate the page 3 times out of 32. 2471 */ 2472 head = 0; 2473 } else { 2474 /* 2475 * Cache the page 28 times out of every 32. Note that 2476 * the page is deactivated instead of cached, but placed 2477 * at the head of the queue instead of the tail. 2478 */ 2479 head = 1; 2480 } 2481 vm_page_spin_lock(m); 2482 _vm_page_deactivate_locked(m, head); 2483 vm_page_spin_unlock(m); 2484 } 2485 2486 /* 2487 * These routines manipulate the 'soft busy' count for a page. A soft busy 2488 * is almost like PG_BUSY except that it allows certain compatible operations 2489 * to occur on the page while it is busy. For example, a page undergoing a 2490 * write can still be mapped read-only. 2491 * 2492 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2493 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2494 * busy bit is cleared. 2495 */ 2496 void 2497 vm_page_io_start(vm_page_t m) 2498 { 2499 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2500 atomic_add_char(&m->busy, 1); 2501 vm_page_flag_set(m, PG_SBUSY); 2502 } 2503 2504 void 2505 vm_page_io_finish(vm_page_t m) 2506 { 2507 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2508 atomic_subtract_char(&m->busy, 1); 2509 if (m->busy == 0) 2510 vm_page_flag_clear(m, PG_SBUSY); 2511 } 2512 2513 /* 2514 * Grab a page, blocking if it is busy and allocating a page if necessary. 2515 * A busy page is returned or NULL. The page may or may not be valid and 2516 * might not be on a queue (the caller is responsible for the disposition of 2517 * the page). 2518 * 2519 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2520 * page will be zero'd and marked valid. 2521 * 2522 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2523 * valid even if it already exists. 2524 * 2525 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2526 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2527 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2528 * 2529 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2530 * always returned if we had blocked. 2531 * 2532 * This routine may not be called from an interrupt. 2533 * 2534 * PG_ZERO is *ALWAYS* cleared by this routine. 2535 * 2536 * No other requirements. 2537 */ 2538 vm_page_t 2539 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2540 { 2541 vm_page_t m; 2542 int error; 2543 2544 KKASSERT(allocflags & 2545 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2546 vm_object_hold(object); 2547 for (;;) { 2548 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2549 if (error) { 2550 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2551 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2552 m = NULL; 2553 break; 2554 } 2555 /* retry */ 2556 } else if (m == NULL) { 2557 if (allocflags & VM_ALLOC_RETRY) 2558 allocflags |= VM_ALLOC_NULL_OK; 2559 m = vm_page_alloc(object, pindex, 2560 allocflags & ~VM_ALLOC_RETRY); 2561 if (m) 2562 break; 2563 vm_wait(0); 2564 if ((allocflags & VM_ALLOC_RETRY) == 0) 2565 goto failed; 2566 } else { 2567 /* m found */ 2568 break; 2569 } 2570 } 2571 2572 /* 2573 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2574 * 2575 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2576 * valid even if already valid. 2577 */ 2578 if (m->valid == 0) { 2579 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2580 if ((m->flags & PG_ZERO) == 0) 2581 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2582 m->valid = VM_PAGE_BITS_ALL; 2583 } 2584 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2585 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2586 m->valid = VM_PAGE_BITS_ALL; 2587 } 2588 vm_page_flag_clear(m, PG_ZERO); 2589 failed: 2590 vm_object_drop(object); 2591 return(m); 2592 } 2593 2594 /* 2595 * Mapping function for valid bits or for dirty bits in 2596 * a page. May not block. 2597 * 2598 * Inputs are required to range within a page. 2599 * 2600 * No requirements. 2601 * Non blocking. 2602 */ 2603 int 2604 vm_page_bits(int base, int size) 2605 { 2606 int first_bit; 2607 int last_bit; 2608 2609 KASSERT( 2610 base + size <= PAGE_SIZE, 2611 ("vm_page_bits: illegal base/size %d/%d", base, size) 2612 ); 2613 2614 if (size == 0) /* handle degenerate case */ 2615 return(0); 2616 2617 first_bit = base >> DEV_BSHIFT; 2618 last_bit = (base + size - 1) >> DEV_BSHIFT; 2619 2620 return ((2 << last_bit) - (1 << first_bit)); 2621 } 2622 2623 /* 2624 * Sets portions of a page valid and clean. The arguments are expected 2625 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2626 * of any partial chunks touched by the range. The invalid portion of 2627 * such chunks will be zero'd. 2628 * 2629 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2630 * align base to DEV_BSIZE so as not to mark clean a partially 2631 * truncated device block. Otherwise the dirty page status might be 2632 * lost. 2633 * 2634 * This routine may not block. 2635 * 2636 * (base + size) must be less then or equal to PAGE_SIZE. 2637 */ 2638 static void 2639 _vm_page_zero_valid(vm_page_t m, int base, int size) 2640 { 2641 int frag; 2642 int endoff; 2643 2644 if (size == 0) /* handle degenerate case */ 2645 return; 2646 2647 /* 2648 * If the base is not DEV_BSIZE aligned and the valid 2649 * bit is clear, we have to zero out a portion of the 2650 * first block. 2651 */ 2652 2653 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2654 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2655 ) { 2656 pmap_zero_page_area( 2657 VM_PAGE_TO_PHYS(m), 2658 frag, 2659 base - frag 2660 ); 2661 } 2662 2663 /* 2664 * If the ending offset is not DEV_BSIZE aligned and the 2665 * valid bit is clear, we have to zero out a portion of 2666 * the last block. 2667 */ 2668 2669 endoff = base + size; 2670 2671 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2672 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 2673 ) { 2674 pmap_zero_page_area( 2675 VM_PAGE_TO_PHYS(m), 2676 endoff, 2677 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 2678 ); 2679 } 2680 } 2681 2682 /* 2683 * Set valid, clear dirty bits. If validating the entire 2684 * page we can safely clear the pmap modify bit. We also 2685 * use this opportunity to clear the PG_NOSYNC flag. If a process 2686 * takes a write fault on a MAP_NOSYNC memory area the flag will 2687 * be set again. 2688 * 2689 * We set valid bits inclusive of any overlap, but we can only 2690 * clear dirty bits for DEV_BSIZE chunks that are fully within 2691 * the range. 2692 * 2693 * Page must be busied? 2694 * No other requirements. 2695 */ 2696 void 2697 vm_page_set_valid(vm_page_t m, int base, int size) 2698 { 2699 _vm_page_zero_valid(m, base, size); 2700 m->valid |= vm_page_bits(base, size); 2701 } 2702 2703 2704 /* 2705 * Set valid bits and clear dirty bits. 2706 * 2707 * NOTE: This function does not clear the pmap modified bit. 2708 * Also note that e.g. NFS may use a byte-granular base 2709 * and size. 2710 * 2711 * WARNING: Page must be busied? But vfs_clean_one_page() will call 2712 * this without necessarily busying the page (via bdwrite()). 2713 * So for now vm_token must also be held. 2714 * 2715 * No other requirements. 2716 */ 2717 void 2718 vm_page_set_validclean(vm_page_t m, int base, int size) 2719 { 2720 int pagebits; 2721 2722 _vm_page_zero_valid(m, base, size); 2723 pagebits = vm_page_bits(base, size); 2724 m->valid |= pagebits; 2725 m->dirty &= ~pagebits; 2726 if (base == 0 && size == PAGE_SIZE) { 2727 /*pmap_clear_modify(m);*/ 2728 vm_page_flag_clear(m, PG_NOSYNC); 2729 } 2730 } 2731 2732 /* 2733 * Set valid & dirty. Used by buwrite() 2734 * 2735 * WARNING: Page must be busied? But vfs_dirty_one_page() will 2736 * call this function in buwrite() so for now vm_token must 2737 * be held. 2738 * 2739 * No other requirements. 2740 */ 2741 void 2742 vm_page_set_validdirty(vm_page_t m, int base, int size) 2743 { 2744 int pagebits; 2745 2746 pagebits = vm_page_bits(base, size); 2747 m->valid |= pagebits; 2748 m->dirty |= pagebits; 2749 if (m->object) 2750 vm_object_set_writeable_dirty(m->object); 2751 } 2752 2753 /* 2754 * Clear dirty bits. 2755 * 2756 * NOTE: This function does not clear the pmap modified bit. 2757 * Also note that e.g. NFS may use a byte-granular base 2758 * and size. 2759 * 2760 * Page must be busied? 2761 * No other requirements. 2762 */ 2763 void 2764 vm_page_clear_dirty(vm_page_t m, int base, int size) 2765 { 2766 m->dirty &= ~vm_page_bits(base, size); 2767 if (base == 0 && size == PAGE_SIZE) { 2768 /*pmap_clear_modify(m);*/ 2769 vm_page_flag_clear(m, PG_NOSYNC); 2770 } 2771 } 2772 2773 /* 2774 * Make the page all-dirty. 2775 * 2776 * Also make sure the related object and vnode reflect the fact that the 2777 * object may now contain a dirty page. 2778 * 2779 * Page must be busied? 2780 * No other requirements. 2781 */ 2782 void 2783 vm_page_dirty(vm_page_t m) 2784 { 2785 #ifdef INVARIANTS 2786 int pqtype = m->queue - m->pc; 2787 #endif 2788 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 2789 ("vm_page_dirty: page in free/cache queue!")); 2790 if (m->dirty != VM_PAGE_BITS_ALL) { 2791 m->dirty = VM_PAGE_BITS_ALL; 2792 if (m->object) 2793 vm_object_set_writeable_dirty(m->object); 2794 } 2795 } 2796 2797 /* 2798 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2799 * valid and dirty bits for the effected areas are cleared. 2800 * 2801 * Page must be busied? 2802 * Does not block. 2803 * No other requirements. 2804 */ 2805 void 2806 vm_page_set_invalid(vm_page_t m, int base, int size) 2807 { 2808 int bits; 2809 2810 bits = vm_page_bits(base, size); 2811 m->valid &= ~bits; 2812 m->dirty &= ~bits; 2813 m->object->generation++; 2814 } 2815 2816 /* 2817 * The kernel assumes that the invalid portions of a page contain 2818 * garbage, but such pages can be mapped into memory by user code. 2819 * When this occurs, we must zero out the non-valid portions of the 2820 * page so user code sees what it expects. 2821 * 2822 * Pages are most often semi-valid when the end of a file is mapped 2823 * into memory and the file's size is not page aligned. 2824 * 2825 * Page must be busied? 2826 * No other requirements. 2827 */ 2828 void 2829 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2830 { 2831 int b; 2832 int i; 2833 2834 /* 2835 * Scan the valid bits looking for invalid sections that 2836 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2837 * valid bit may be set ) have already been zerod by 2838 * vm_page_set_validclean(). 2839 */ 2840 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2841 if (i == (PAGE_SIZE / DEV_BSIZE) || 2842 (m->valid & (1 << i)) 2843 ) { 2844 if (i > b) { 2845 pmap_zero_page_area( 2846 VM_PAGE_TO_PHYS(m), 2847 b << DEV_BSHIFT, 2848 (i - b) << DEV_BSHIFT 2849 ); 2850 } 2851 b = i + 1; 2852 } 2853 } 2854 2855 /* 2856 * setvalid is TRUE when we can safely set the zero'd areas 2857 * as being valid. We can do this if there are no cache consistency 2858 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2859 */ 2860 if (setvalid) 2861 m->valid = VM_PAGE_BITS_ALL; 2862 } 2863 2864 /* 2865 * Is a (partial) page valid? Note that the case where size == 0 2866 * will return FALSE in the degenerate case where the page is entirely 2867 * invalid, and TRUE otherwise. 2868 * 2869 * Does not block. 2870 * No other requirements. 2871 */ 2872 int 2873 vm_page_is_valid(vm_page_t m, int base, int size) 2874 { 2875 int bits = vm_page_bits(base, size); 2876 2877 if (m->valid && ((m->valid & bits) == bits)) 2878 return 1; 2879 else 2880 return 0; 2881 } 2882 2883 /* 2884 * update dirty bits from pmap/mmu. May not block. 2885 * 2886 * Caller must hold the page busy 2887 */ 2888 void 2889 vm_page_test_dirty(vm_page_t m) 2890 { 2891 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 2892 vm_page_dirty(m); 2893 } 2894 } 2895 2896 /* 2897 * Register an action, associating it with its vm_page 2898 */ 2899 void 2900 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 2901 { 2902 struct vm_page_action_list *list; 2903 int hv; 2904 2905 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 2906 list = &action_list[hv]; 2907 2908 lwkt_gettoken(&vm_token); 2909 vm_page_flag_set(action->m, PG_ACTIONLIST); 2910 action->event = event; 2911 LIST_INSERT_HEAD(list, action, entry); 2912 lwkt_reltoken(&vm_token); 2913 } 2914 2915 /* 2916 * Unregister an action, disassociating it from its related vm_page 2917 */ 2918 void 2919 vm_page_unregister_action(vm_page_action_t action) 2920 { 2921 struct vm_page_action_list *list; 2922 int hv; 2923 2924 lwkt_gettoken(&vm_token); 2925 if (action->event != VMEVENT_NONE) { 2926 action->event = VMEVENT_NONE; 2927 LIST_REMOVE(action, entry); 2928 2929 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 2930 list = &action_list[hv]; 2931 if (LIST_EMPTY(list)) 2932 vm_page_flag_clear(action->m, PG_ACTIONLIST); 2933 } 2934 lwkt_reltoken(&vm_token); 2935 } 2936 2937 /* 2938 * Issue an event on a VM page. Corresponding action structures are 2939 * removed from the page's list and called. 2940 * 2941 * If the vm_page has no more pending action events we clear its 2942 * PG_ACTIONLIST flag. 2943 */ 2944 void 2945 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 2946 { 2947 struct vm_page_action_list *list; 2948 struct vm_page_action *scan; 2949 struct vm_page_action *next; 2950 int hv; 2951 int all; 2952 2953 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 2954 list = &action_list[hv]; 2955 all = 1; 2956 2957 lwkt_gettoken(&vm_token); 2958 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 2959 if (scan->m == m) { 2960 if (scan->event == event) { 2961 scan->event = VMEVENT_NONE; 2962 LIST_REMOVE(scan, entry); 2963 scan->func(m, scan); 2964 /* XXX */ 2965 } else { 2966 all = 0; 2967 } 2968 } 2969 } 2970 if (all) 2971 vm_page_flag_clear(m, PG_ACTIONLIST); 2972 lwkt_reltoken(&vm_token); 2973 } 2974 2975 #include "opt_ddb.h" 2976 #ifdef DDB 2977 #include <sys/kernel.h> 2978 2979 #include <ddb/ddb.h> 2980 2981 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2982 { 2983 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 2984 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 2985 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 2986 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 2987 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 2988 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 2989 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 2990 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 2991 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 2992 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 2993 } 2994 2995 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2996 { 2997 int i; 2998 db_printf("PQ_FREE:"); 2999 for(i=0;i<PQ_L2_SIZE;i++) { 3000 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3001 } 3002 db_printf("\n"); 3003 3004 db_printf("PQ_CACHE:"); 3005 for(i=0;i<PQ_L2_SIZE;i++) { 3006 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3007 } 3008 db_printf("\n"); 3009 3010 db_printf("PQ_ACTIVE:"); 3011 for(i=0;i<PQ_L2_SIZE;i++) { 3012 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3013 } 3014 db_printf("\n"); 3015 3016 db_printf("PQ_INACTIVE:"); 3017 for(i=0;i<PQ_L2_SIZE;i++) { 3018 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3019 } 3020 db_printf("\n"); 3021 } 3022 #endif /* DDB */ 3023