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 | PG_NEED_COMMIT)) == 0 && 1340 m->hold_count == 0 && 1341 m->wire_count == 0 && 1342 (m->dirty & m->valid) == 0) { 1343 vm_page_spin_unlock(m); 1344 pagedaemon_wakeup(); 1345 return(m); 1346 } 1347 1348 /* 1349 * The page cannot be recycled, deactivate it. 1350 */ 1351 _vm_page_deactivate_locked(m, 0); 1352 if (_vm_page_wakeup(m)) { 1353 vm_page_spin_unlock(m); 1354 wakeup(m); 1355 } else { 1356 vm_page_spin_unlock(m); 1357 } 1358 } 1359 } 1360 return (m); 1361 } 1362 1363 /* 1364 * Find a free or zero page, with specified preference. We attempt to 1365 * inline the nominal case and fall back to _vm_page_select_free() 1366 * otherwise. A busied page is removed from the queue and returned. 1367 * 1368 * This routine may not block. 1369 */ 1370 static __inline vm_page_t 1371 vm_page_select_free(u_short pg_color, boolean_t prefer_zero) 1372 { 1373 vm_page_t m; 1374 1375 for (;;) { 1376 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK, 1377 prefer_zero); 1378 if (m == NULL) 1379 break; 1380 if (vm_page_busy_try(m, TRUE)) { 1381 /* 1382 * Various mechanisms such as a pmap_collect can 1383 * result in a busy page on the free queue. We 1384 * have to move the page out of the way so we can 1385 * retry the allocation. If the other thread is not 1386 * allocating the page then m->valid will remain 0 and 1387 * the pageout daemon will free the page later on. 1388 * 1389 * Since we could not busy the page, however, we 1390 * cannot make assumptions as to whether the page 1391 * will be allocated by the other thread or not, 1392 * so all we can do is deactivate it to move it out 1393 * of the way. In particular, if the other thread 1394 * wires the page it may wind up on the inactive 1395 * queue and the pageout daemon will have to deal 1396 * with that case too. 1397 */ 1398 _vm_page_deactivate_locked(m, 0); 1399 vm_page_spin_unlock(m); 1400 #ifdef INVARIANTS 1401 kprintf("Warning: busy page %p found in cache\n", m); 1402 #endif 1403 } else { 1404 /* 1405 * Theoretically if we are able to busy the page 1406 * atomic with the queue removal (using the vm_page 1407 * lock) nobody else should be able to mess with the 1408 * page before us. 1409 */ 1410 KKASSERT((m->flags & (PG_UNMANAGED | 1411 PG_NEED_COMMIT)) == 0); 1412 KKASSERT(m->hold_count == 0); 1413 KKASSERT(m->wire_count == 0); 1414 vm_page_spin_unlock(m); 1415 pagedaemon_wakeup(); 1416 1417 /* return busied and removed page */ 1418 return(m); 1419 } 1420 } 1421 return(m); 1422 } 1423 1424 /* 1425 * This implements a per-cpu cache of free, zero'd, ready-to-go pages. 1426 * The idea is to populate this cache prior to acquiring any locks so 1427 * we don't wind up potentially zeroing VM pages (under heavy loads) while 1428 * holding potentialy contending locks. 1429 * 1430 * Note that we allocate the page uninserted into anything and use a pindex 1431 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these 1432 * allocations should wind up being uncontended. However, we still want 1433 * to rove across PQ_L2_SIZE. 1434 */ 1435 void 1436 vm_page_pcpu_cache(void) 1437 { 1438 #if 0 1439 globaldata_t gd = mycpu; 1440 vm_page_t m; 1441 1442 if (gd->gd_vmpg_count < GD_MINVMPG) { 1443 crit_enter_gd(gd); 1444 while (gd->gd_vmpg_count < GD_MAXVMPG) { 1445 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask, 1446 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1447 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO); 1448 if (gd->gd_vmpg_count < GD_MAXVMPG) { 1449 if ((m->flags & PG_ZERO) == 0) { 1450 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1451 vm_page_flag_set(m, PG_ZERO); 1452 } 1453 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m; 1454 } else { 1455 vm_page_free(m); 1456 } 1457 } 1458 crit_exit_gd(gd); 1459 } 1460 #endif 1461 } 1462 1463 /* 1464 * vm_page_alloc() 1465 * 1466 * Allocate and return a memory cell associated with this VM object/offset 1467 * pair. If object is NULL an unassociated page will be allocated. 1468 * 1469 * The returned page will be busied and removed from its queues. This 1470 * routine can block and may return NULL if a race occurs and the page 1471 * is found to already exist at the specified (object, pindex). 1472 * 1473 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1474 * VM_ALLOC_QUICK like normal but cannot use cache 1475 * VM_ALLOC_SYSTEM greater free drain 1476 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1477 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1478 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1479 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1480 * (see vm_page_grab()) 1481 * VM_ALLOC_USE_GD ok to use per-gd cache 1482 * 1483 * The object must be held if not NULL 1484 * This routine may not block 1485 * 1486 * Additional special handling is required when called from an interrupt 1487 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1488 * in this case. 1489 */ 1490 vm_page_t 1491 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1492 { 1493 #ifdef SMP 1494 globaldata_t gd = mycpu; 1495 #endif 1496 vm_object_t obj; 1497 vm_page_t m; 1498 u_short pg_color; 1499 1500 #if 0 1501 /* 1502 * Special per-cpu free VM page cache. The pages are pre-busied 1503 * and pre-zerod for us. 1504 */ 1505 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1506 crit_enter_gd(gd); 1507 if (gd->gd_vmpg_count) { 1508 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1509 crit_exit_gd(gd); 1510 goto done; 1511 } 1512 crit_exit_gd(gd); 1513 } 1514 #endif 1515 m = NULL; 1516 1517 #ifdef SMP 1518 /* 1519 * Cpu twist - cpu localization algorithm 1520 */ 1521 if (object) { 1522 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) + 1523 (object->pg_color & ~ncpus_fit_mask); 1524 } else { 1525 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask); 1526 } 1527 #else 1528 /* 1529 * Normal page coloring algorithm 1530 */ 1531 if (object) { 1532 pg_color = object->pg_color + pindex; 1533 } else { 1534 pg_color = pindex; 1535 } 1536 #endif 1537 KKASSERT(page_req & 1538 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1539 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1540 1541 /* 1542 * Certain system threads (pageout daemon, buf_daemon's) are 1543 * allowed to eat deeper into the free page list. 1544 */ 1545 if (curthread->td_flags & TDF_SYSTHREAD) 1546 page_req |= VM_ALLOC_SYSTEM; 1547 1548 loop: 1549 if (vmstats.v_free_count > vmstats.v_free_reserved || 1550 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 1551 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 1552 vmstats.v_free_count > vmstats.v_interrupt_free_min) 1553 ) { 1554 /* 1555 * The free queue has sufficient free pages to take one out. 1556 */ 1557 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) 1558 m = vm_page_select_free(pg_color, TRUE); 1559 else 1560 m = vm_page_select_free(pg_color, FALSE); 1561 } else if (page_req & VM_ALLOC_NORMAL) { 1562 /* 1563 * Allocatable from the cache (non-interrupt only). On 1564 * success, we must free the page and try again, thus 1565 * ensuring that vmstats.v_*_free_min counters are replenished. 1566 */ 1567 #ifdef INVARIANTS 1568 if (curthread->td_preempted) { 1569 kprintf("vm_page_alloc(): warning, attempt to allocate" 1570 " cache page from preempting interrupt\n"); 1571 m = NULL; 1572 } else { 1573 m = vm_page_select_cache(pg_color); 1574 } 1575 #else 1576 m = vm_page_select_cache(pg_color); 1577 #endif 1578 /* 1579 * On success move the page into the free queue and loop. 1580 * 1581 * Only do this if we can safely acquire the vm_object lock, 1582 * because this is effectively a random page and the caller 1583 * might be holding the lock shared, we don't want to 1584 * deadlock. 1585 */ 1586 if (m != NULL) { 1587 KASSERT(m->dirty == 0, 1588 ("Found dirty cache page %p", m)); 1589 if ((obj = m->object) != NULL) { 1590 if (vm_object_hold_try(obj)) { 1591 vm_page_protect(m, VM_PROT_NONE); 1592 vm_page_free(m); 1593 /* m->object NULL here */ 1594 vm_object_drop(obj); 1595 } else { 1596 vm_page_deactivate(m); 1597 vm_page_wakeup(m); 1598 } 1599 } else { 1600 vm_page_protect(m, VM_PROT_NONE); 1601 vm_page_free(m); 1602 } 1603 goto loop; 1604 } 1605 1606 /* 1607 * On failure return NULL 1608 */ 1609 #if defined(DIAGNOSTIC) 1610 if (vmstats.v_cache_count > 0) 1611 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 1612 #endif 1613 vm_pageout_deficit++; 1614 pagedaemon_wakeup(); 1615 return (NULL); 1616 } else { 1617 /* 1618 * No pages available, wakeup the pageout daemon and give up. 1619 */ 1620 vm_pageout_deficit++; 1621 pagedaemon_wakeup(); 1622 return (NULL); 1623 } 1624 1625 /* 1626 * v_free_count can race so loop if we don't find the expected 1627 * page. 1628 */ 1629 if (m == NULL) 1630 goto loop; 1631 1632 /* 1633 * Good page found. The page has already been busied for us and 1634 * removed from its queues. 1635 */ 1636 KASSERT(m->dirty == 0, 1637 ("vm_page_alloc: free/cache page %p was dirty", m)); 1638 KKASSERT(m->queue == PQ_NONE); 1639 1640 #if 0 1641 done: 1642 #endif 1643 /* 1644 * Initialize the structure, inheriting some flags but clearing 1645 * all the rest. The page has already been busied for us. 1646 */ 1647 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY)); 1648 KKASSERT(m->wire_count == 0); 1649 KKASSERT(m->busy == 0); 1650 m->act_count = 0; 1651 m->valid = 0; 1652 1653 /* 1654 * Caller must be holding the object lock (asserted by 1655 * vm_page_insert()). 1656 * 1657 * NOTE: Inserting a page here does not insert it into any pmaps 1658 * (which could cause us to block allocating memory). 1659 * 1660 * NOTE: If no object an unassociated page is allocated, m->pindex 1661 * can be used by the caller for any purpose. 1662 */ 1663 if (object) { 1664 if (vm_page_insert(m, object, pindex) == FALSE) { 1665 kprintf("PAGE RACE (%p:%d,%"PRIu64")\n", 1666 object, object->type, pindex); 1667 vm_page_free(m); 1668 m = NULL; 1669 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1670 panic("PAGE RACE"); 1671 } 1672 } else { 1673 m->pindex = pindex; 1674 } 1675 1676 /* 1677 * Don't wakeup too often - wakeup the pageout daemon when 1678 * we would be nearly out of memory. 1679 */ 1680 pagedaemon_wakeup(); 1681 1682 /* 1683 * A PG_BUSY page is returned. 1684 */ 1685 return (m); 1686 } 1687 1688 /* 1689 * Attempt to allocate contiguous physical memory with the specified 1690 * requirements. 1691 */ 1692 vm_page_t 1693 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1694 unsigned long alignment, unsigned long boundary, 1695 unsigned long size) 1696 { 1697 alist_blk_t blk; 1698 1699 alignment >>= PAGE_SHIFT; 1700 if (alignment == 0) 1701 alignment = 1; 1702 boundary >>= PAGE_SHIFT; 1703 if (boundary == 0) 1704 boundary = 1; 1705 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1706 1707 spin_lock(&vm_contig_spin); 1708 blk = alist_alloc(&vm_contig_alist, 0, size); 1709 if (blk == ALIST_BLOCK_NONE) { 1710 spin_unlock(&vm_contig_spin); 1711 if (bootverbose) { 1712 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1713 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1714 } 1715 return(NULL); 1716 } 1717 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1718 alist_free(&vm_contig_alist, blk, size); 1719 spin_unlock(&vm_contig_spin); 1720 if (bootverbose) { 1721 kprintf("vm_page_alloc_contig: %ldk high " 1722 "%016jx failed\n", 1723 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1724 (intmax_t)high); 1725 } 1726 return(NULL); 1727 } 1728 spin_unlock(&vm_contig_spin); 1729 if (bootverbose) { 1730 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 1731 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 1732 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1733 } 1734 return (PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT)); 1735 } 1736 1737 /* 1738 * Free contiguously allocated pages. The pages will be wired but not busy. 1739 * When freeing to the alist we leave them wired and not busy. 1740 */ 1741 void 1742 vm_page_free_contig(vm_page_t m, unsigned long size) 1743 { 1744 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 1745 vm_pindex_t start = pa >> PAGE_SHIFT; 1746 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 1747 1748 if (bootverbose) { 1749 kprintf("vm_page_free_contig: %016jx/%ldk\n", 1750 (intmax_t)pa, size / 1024); 1751 } 1752 if (pa < vm_low_phys_reserved) { 1753 KKASSERT(pa + size <= vm_low_phys_reserved); 1754 spin_lock(&vm_contig_spin); 1755 alist_free(&vm_contig_alist, start, pages); 1756 spin_unlock(&vm_contig_spin); 1757 } else { 1758 while (pages) { 1759 vm_page_busy_wait(m, FALSE, "cpgfr"); 1760 vm_page_unwire(m, 0); 1761 vm_page_free(m); 1762 --pages; 1763 ++m; 1764 } 1765 1766 } 1767 } 1768 1769 1770 /* 1771 * Wait for sufficient free memory for nominal heavy memory use kernel 1772 * operations. 1773 * 1774 * WARNING! Be sure never to call this in any vm_pageout code path, which 1775 * will trivially deadlock the system. 1776 */ 1777 void 1778 vm_wait_nominal(void) 1779 { 1780 while (vm_page_count_min(0)) 1781 vm_wait(0); 1782 } 1783 1784 /* 1785 * Test if vm_wait_nominal() would block. 1786 */ 1787 int 1788 vm_test_nominal(void) 1789 { 1790 if (vm_page_count_min(0)) 1791 return(1); 1792 return(0); 1793 } 1794 1795 /* 1796 * Block until free pages are available for allocation, called in various 1797 * places before memory allocations. 1798 * 1799 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 1800 * more generous then that. 1801 */ 1802 void 1803 vm_wait(int timo) 1804 { 1805 /* 1806 * never wait forever 1807 */ 1808 if (timo == 0) 1809 timo = hz; 1810 lwkt_gettoken(&vm_token); 1811 1812 if (curthread == pagethread) { 1813 /* 1814 * The pageout daemon itself needs pages, this is bad. 1815 */ 1816 if (vm_page_count_min(0)) { 1817 vm_pageout_pages_needed = 1; 1818 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 1819 } 1820 } else { 1821 /* 1822 * Wakeup the pageout daemon if necessary and wait. 1823 */ 1824 if (vm_page_count_target()) { 1825 if (vm_pages_needed == 0) { 1826 vm_pages_needed = 1; 1827 wakeup(&vm_pages_needed); 1828 } 1829 ++vm_pages_waiting; /* SMP race ok */ 1830 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 1831 } 1832 } 1833 lwkt_reltoken(&vm_token); 1834 } 1835 1836 /* 1837 * Block until free pages are available for allocation 1838 * 1839 * Called only from vm_fault so that processes page faulting can be 1840 * easily tracked. 1841 */ 1842 void 1843 vm_waitpfault(void) 1844 { 1845 /* 1846 * Wakeup the pageout daemon if necessary and wait. 1847 */ 1848 if (vm_page_count_target()) { 1849 lwkt_gettoken(&vm_token); 1850 if (vm_page_count_target()) { 1851 if (vm_pages_needed == 0) { 1852 vm_pages_needed = 1; 1853 wakeup(&vm_pages_needed); 1854 } 1855 ++vm_pages_waiting; /* SMP race ok */ 1856 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 1857 } 1858 lwkt_reltoken(&vm_token); 1859 } 1860 } 1861 1862 /* 1863 * Put the specified page on the active list (if appropriate). Ensure 1864 * that act_count is at least ACT_INIT but do not otherwise mess with it. 1865 * 1866 * The caller should be holding the page busied ? XXX 1867 * This routine may not block. 1868 */ 1869 void 1870 vm_page_activate(vm_page_t m) 1871 { 1872 u_short oqueue; 1873 1874 vm_page_spin_lock(m); 1875 if (m->queue - m->pc != PQ_ACTIVE) { 1876 _vm_page_queue_spin_lock(m); 1877 oqueue = _vm_page_rem_queue_spinlocked(m); 1878 /* page is left spinlocked, queue is unlocked */ 1879 1880 if (oqueue == PQ_CACHE) 1881 mycpu->gd_cnt.v_reactivated++; 1882 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1883 if (m->act_count < ACT_INIT) 1884 m->act_count = ACT_INIT; 1885 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 1886 } 1887 _vm_page_and_queue_spin_unlock(m); 1888 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 1889 pagedaemon_wakeup(); 1890 } else { 1891 if (m->act_count < ACT_INIT) 1892 m->act_count = ACT_INIT; 1893 vm_page_spin_unlock(m); 1894 } 1895 } 1896 1897 /* 1898 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1899 * routine is called when a page has been added to the cache or free 1900 * queues. 1901 * 1902 * This routine may not block. 1903 */ 1904 static __inline void 1905 vm_page_free_wakeup(void) 1906 { 1907 /* 1908 * If the pageout daemon itself needs pages, then tell it that 1909 * there are some free. 1910 */ 1911 if (vm_pageout_pages_needed && 1912 vmstats.v_cache_count + vmstats.v_free_count >= 1913 vmstats.v_pageout_free_min 1914 ) { 1915 wakeup(&vm_pageout_pages_needed); 1916 vm_pageout_pages_needed = 0; 1917 } 1918 1919 /* 1920 * Wakeup processes that are waiting on memory. 1921 * 1922 * NOTE: vm_paging_target() is the pageout daemon's target, while 1923 * vm_page_count_target() is somewhere inbetween. We want 1924 * to wake processes up prior to the pageout daemon reaching 1925 * its target to provide some hysteresis. 1926 */ 1927 if (vm_pages_waiting) { 1928 if (!vm_page_count_target()) { 1929 /* 1930 * Plenty of pages are free, wakeup everyone. 1931 */ 1932 vm_pages_waiting = 0; 1933 wakeup(&vmstats.v_free_count); 1934 ++mycpu->gd_cnt.v_ppwakeups; 1935 } else if (!vm_page_count_min(0)) { 1936 /* 1937 * Some pages are free, wakeup someone. 1938 */ 1939 int wcount = vm_pages_waiting; 1940 if (wcount > 0) 1941 --wcount; 1942 vm_pages_waiting = wcount; 1943 wakeup_one(&vmstats.v_free_count); 1944 ++mycpu->gd_cnt.v_ppwakeups; 1945 } 1946 } 1947 } 1948 1949 /* 1950 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 1951 * it from its VM object. 1952 * 1953 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 1954 * return (the page will have been freed). 1955 */ 1956 void 1957 vm_page_free_toq(vm_page_t m) 1958 { 1959 mycpu->gd_cnt.v_tfree++; 1960 KKASSERT((m->flags & PG_MAPPED) == 0); 1961 KKASSERT(m->flags & PG_BUSY); 1962 1963 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1964 kprintf("vm_page_free: pindex(%lu), busy(%d), " 1965 "PG_BUSY(%d), hold(%d)\n", 1966 (u_long)m->pindex, m->busy, 1967 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 1968 if ((m->queue - m->pc) == PQ_FREE) 1969 panic("vm_page_free: freeing free page"); 1970 else 1971 panic("vm_page_free: freeing busy page"); 1972 } 1973 1974 /* 1975 * Remove from object, spinlock the page and its queues and 1976 * remove from any queue. No queue spinlock will be held 1977 * after this section (because the page was removed from any 1978 * queue). 1979 */ 1980 vm_page_remove(m); 1981 vm_page_and_queue_spin_lock(m); 1982 _vm_page_rem_queue_spinlocked(m); 1983 1984 /* 1985 * No further management of fictitious pages occurs beyond object 1986 * and queue removal. 1987 */ 1988 if ((m->flags & PG_FICTITIOUS) != 0) { 1989 vm_page_spin_unlock(m); 1990 vm_page_wakeup(m); 1991 return; 1992 } 1993 1994 m->valid = 0; 1995 vm_page_undirty(m); 1996 1997 if (m->wire_count != 0) { 1998 if (m->wire_count > 1) { 1999 panic( 2000 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2001 m->wire_count, (long)m->pindex); 2002 } 2003 panic("vm_page_free: freeing wired page"); 2004 } 2005 2006 /* 2007 * Clear the UNMANAGED flag when freeing an unmanaged page. 2008 * Clear the NEED_COMMIT flag 2009 */ 2010 if (m->flags & PG_UNMANAGED) 2011 vm_page_flag_clear(m, PG_UNMANAGED); 2012 if (m->flags & PG_NEED_COMMIT) 2013 vm_page_flag_clear(m, PG_NEED_COMMIT); 2014 2015 if (m->hold_count != 0) { 2016 vm_page_flag_clear(m, PG_ZERO); 2017 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2018 } else { 2019 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 2020 } 2021 2022 /* 2023 * This sequence allows us to clear PG_BUSY while still holding 2024 * its spin lock, which reduces contention vs allocators. We 2025 * must not leave the queue locked or _vm_page_wakeup() may 2026 * deadlock. 2027 */ 2028 _vm_page_queue_spin_unlock(m); 2029 if (_vm_page_wakeup(m)) { 2030 vm_page_spin_unlock(m); 2031 wakeup(m); 2032 } else { 2033 vm_page_spin_unlock(m); 2034 } 2035 vm_page_free_wakeup(); 2036 } 2037 2038 /* 2039 * vm_page_free_fromq_fast() 2040 * 2041 * Remove a non-zero page from one of the free queues; the page is removed for 2042 * zeroing, so do not issue a wakeup. 2043 */ 2044 vm_page_t 2045 vm_page_free_fromq_fast(void) 2046 { 2047 static int qi; 2048 vm_page_t m; 2049 int i; 2050 2051 for (i = 0; i < PQ_L2_SIZE; ++i) { 2052 m = vm_page_list_find(PQ_FREE, qi, FALSE); 2053 /* page is returned spinlocked and removed from its queue */ 2054 if (m) { 2055 if (vm_page_busy_try(m, TRUE)) { 2056 /* 2057 * We were unable to busy the page, deactivate 2058 * it and loop. 2059 */ 2060 _vm_page_deactivate_locked(m, 0); 2061 vm_page_spin_unlock(m); 2062 } else if (m->flags & PG_ZERO) { 2063 /* 2064 * The page is PG_ZERO, requeue it and loop 2065 */ 2066 _vm_page_add_queue_spinlocked(m, 2067 PQ_FREE + m->pc, 2068 0); 2069 vm_page_queue_spin_unlock(m); 2070 if (_vm_page_wakeup(m)) { 2071 vm_page_spin_unlock(m); 2072 wakeup(m); 2073 } else { 2074 vm_page_spin_unlock(m); 2075 } 2076 } else { 2077 /* 2078 * The page is not PG_ZERO'd so return it. 2079 */ 2080 vm_page_spin_unlock(m); 2081 KKASSERT((m->flags & (PG_UNMANAGED | 2082 PG_NEED_COMMIT)) == 0); 2083 KKASSERT(m->hold_count == 0); 2084 KKASSERT(m->wire_count == 0); 2085 break; 2086 } 2087 m = NULL; 2088 } 2089 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 2090 } 2091 return (m); 2092 } 2093 2094 /* 2095 * vm_page_unmanage() 2096 * 2097 * Prevent PV management from being done on the page. The page is 2098 * removed from the paging queues as if it were wired, and as a 2099 * consequence of no longer being managed the pageout daemon will not 2100 * touch it (since there is no way to locate the pte mappings for the 2101 * page). madvise() calls that mess with the pmap will also no longer 2102 * operate on the page. 2103 * 2104 * Beyond that the page is still reasonably 'normal'. Freeing the page 2105 * will clear the flag. 2106 * 2107 * This routine is used by OBJT_PHYS objects - objects using unswappable 2108 * physical memory as backing store rather then swap-backed memory and 2109 * will eventually be extended to support 4MB unmanaged physical 2110 * mappings. 2111 * 2112 * Caller must be holding the page busy. 2113 */ 2114 void 2115 vm_page_unmanage(vm_page_t m) 2116 { 2117 KKASSERT(m->flags & PG_BUSY); 2118 if ((m->flags & PG_UNMANAGED) == 0) { 2119 if (m->wire_count == 0) 2120 vm_page_unqueue(m); 2121 } 2122 vm_page_flag_set(m, PG_UNMANAGED); 2123 } 2124 2125 /* 2126 * Mark this page as wired down by yet another map, removing it from 2127 * paging queues as necessary. 2128 * 2129 * Caller must be holding the page busy. 2130 */ 2131 void 2132 vm_page_wire(vm_page_t m) 2133 { 2134 /* 2135 * Only bump the wire statistics if the page is not already wired, 2136 * and only unqueue the page if it is on some queue (if it is unmanaged 2137 * it is already off the queues). Don't do anything with fictitious 2138 * pages because they are always wired. 2139 */ 2140 KKASSERT(m->flags & PG_BUSY); 2141 if ((m->flags & PG_FICTITIOUS) == 0) { 2142 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2143 if ((m->flags & PG_UNMANAGED) == 0) 2144 vm_page_unqueue(m); 2145 atomic_add_int(&vmstats.v_wire_count, 1); 2146 } 2147 KASSERT(m->wire_count != 0, 2148 ("vm_page_wire: wire_count overflow m=%p", m)); 2149 } 2150 } 2151 2152 /* 2153 * Release one wiring of this page, potentially enabling it to be paged again. 2154 * 2155 * Many pages placed on the inactive queue should actually go 2156 * into the cache, but it is difficult to figure out which. What 2157 * we do instead, if the inactive target is well met, is to put 2158 * clean pages at the head of the inactive queue instead of the tail. 2159 * This will cause them to be moved to the cache more quickly and 2160 * if not actively re-referenced, freed more quickly. If we just 2161 * stick these pages at the end of the inactive queue, heavy filesystem 2162 * meta-data accesses can cause an unnecessary paging load on memory bound 2163 * processes. This optimization causes one-time-use metadata to be 2164 * reused more quickly. 2165 * 2166 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2167 * the inactive queue. This helps the pageout daemon determine memory 2168 * pressure and act on out-of-memory situations more quickly. 2169 * 2170 * BUT, if we are in a low-memory situation we have no choice but to 2171 * put clean pages on the cache queue. 2172 * 2173 * A number of routines use vm_page_unwire() to guarantee that the page 2174 * will go into either the inactive or active queues, and will NEVER 2175 * be placed in the cache - for example, just after dirtying a page. 2176 * dirty pages in the cache are not allowed. 2177 * 2178 * The page queues must be locked. 2179 * This routine may not block. 2180 */ 2181 void 2182 vm_page_unwire(vm_page_t m, int activate) 2183 { 2184 KKASSERT(m->flags & PG_BUSY); 2185 if (m->flags & PG_FICTITIOUS) { 2186 /* do nothing */ 2187 } else if (m->wire_count <= 0) { 2188 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2189 } else { 2190 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2191 atomic_add_int(&vmstats.v_wire_count, -1); 2192 if (m->flags & PG_UNMANAGED) { 2193 ; 2194 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2195 vm_page_spin_lock(m); 2196 _vm_page_add_queue_spinlocked(m, 2197 PQ_ACTIVE + m->pc, 0); 2198 _vm_page_and_queue_spin_unlock(m); 2199 } else { 2200 vm_page_spin_lock(m); 2201 vm_page_flag_clear(m, PG_WINATCFLS); 2202 _vm_page_add_queue_spinlocked(m, 2203 PQ_INACTIVE + m->pc, 0); 2204 ++vm_swapcache_inactive_heuristic; 2205 _vm_page_and_queue_spin_unlock(m); 2206 } 2207 } 2208 } 2209 } 2210 2211 /* 2212 * Move the specified page to the inactive queue. If the page has 2213 * any associated swap, the swap is deallocated. 2214 * 2215 * Normally athead is 0 resulting in LRU operation. athead is set 2216 * to 1 if we want this page to be 'as if it were placed in the cache', 2217 * except without unmapping it from the process address space. 2218 * 2219 * vm_page's spinlock must be held on entry and will remain held on return. 2220 * This routine may not block. 2221 */ 2222 static void 2223 _vm_page_deactivate_locked(vm_page_t m, int athead) 2224 { 2225 u_short oqueue; 2226 2227 /* 2228 * Ignore if already inactive. 2229 */ 2230 if (m->queue - m->pc == PQ_INACTIVE) 2231 return; 2232 _vm_page_queue_spin_lock(m); 2233 oqueue = _vm_page_rem_queue_spinlocked(m); 2234 2235 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2236 if (oqueue == PQ_CACHE) 2237 mycpu->gd_cnt.v_reactivated++; 2238 vm_page_flag_clear(m, PG_WINATCFLS); 2239 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2240 if (athead == 0) 2241 ++vm_swapcache_inactive_heuristic; 2242 } 2243 _vm_page_queue_spin_unlock(m); 2244 /* leaves vm_page spinlocked */ 2245 } 2246 2247 /* 2248 * Attempt to deactivate a page. 2249 * 2250 * No requirements. 2251 */ 2252 void 2253 vm_page_deactivate(vm_page_t m) 2254 { 2255 vm_page_spin_lock(m); 2256 _vm_page_deactivate_locked(m, 0); 2257 vm_page_spin_unlock(m); 2258 } 2259 2260 void 2261 vm_page_deactivate_locked(vm_page_t m) 2262 { 2263 _vm_page_deactivate_locked(m, 0); 2264 } 2265 2266 /* 2267 * Attempt to move a page to PQ_CACHE. 2268 * 2269 * Returns 0 on failure, 1 on success 2270 * 2271 * The page should NOT be busied by the caller. This function will validate 2272 * whether the page can be safely moved to the cache. 2273 */ 2274 int 2275 vm_page_try_to_cache(vm_page_t m) 2276 { 2277 vm_page_spin_lock(m); 2278 if (vm_page_busy_try(m, TRUE)) { 2279 vm_page_spin_unlock(m); 2280 return(0); 2281 } 2282 if (m->dirty || m->hold_count || m->wire_count || 2283 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2284 if (_vm_page_wakeup(m)) { 2285 vm_page_spin_unlock(m); 2286 wakeup(m); 2287 } else { 2288 vm_page_spin_unlock(m); 2289 } 2290 return(0); 2291 } 2292 vm_page_spin_unlock(m); 2293 2294 /* 2295 * Page busied by us and no longer spinlocked. Dirty pages cannot 2296 * be moved to the cache. 2297 */ 2298 vm_page_test_dirty(m); 2299 if (m->dirty) { 2300 vm_page_wakeup(m); 2301 return(0); 2302 } 2303 vm_page_cache(m); 2304 return(1); 2305 } 2306 2307 /* 2308 * Attempt to free the page. If we cannot free it, we do nothing. 2309 * 1 is returned on success, 0 on failure. 2310 * 2311 * No requirements. 2312 */ 2313 int 2314 vm_page_try_to_free(vm_page_t m) 2315 { 2316 vm_page_spin_lock(m); 2317 if (vm_page_busy_try(m, TRUE)) { 2318 vm_page_spin_unlock(m); 2319 return(0); 2320 } 2321 2322 /* 2323 * The page can be in any state, including already being on the free 2324 * queue. Check to see if it really can be freed. 2325 */ 2326 if (m->dirty || /* can't free if it is dirty */ 2327 m->hold_count || /* or held (XXX may be wrong) */ 2328 m->wire_count || /* or wired */ 2329 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2330 PG_NEED_COMMIT)) || /* or needs a commit */ 2331 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2332 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2333 if (_vm_page_wakeup(m)) { 2334 vm_page_spin_unlock(m); 2335 wakeup(m); 2336 } else { 2337 vm_page_spin_unlock(m); 2338 } 2339 return(0); 2340 } 2341 vm_page_spin_unlock(m); 2342 2343 /* 2344 * We can probably free the page. 2345 * 2346 * Page busied by us and no longer spinlocked. Dirty pages will 2347 * not be freed by this function. We have to re-test the 2348 * dirty bit after cleaning out the pmaps. 2349 */ 2350 vm_page_test_dirty(m); 2351 if (m->dirty) { 2352 vm_page_wakeup(m); 2353 return(0); 2354 } 2355 vm_page_protect(m, VM_PROT_NONE); 2356 if (m->dirty) { 2357 vm_page_wakeup(m); 2358 return(0); 2359 } 2360 vm_page_free(m); 2361 return(1); 2362 } 2363 2364 /* 2365 * vm_page_cache 2366 * 2367 * Put the specified page onto the page cache queue (if appropriate). 2368 * 2369 * The page must be busy, and this routine will release the busy and 2370 * possibly even free the page. 2371 */ 2372 void 2373 vm_page_cache(vm_page_t m) 2374 { 2375 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2376 m->busy || m->wire_count || m->hold_count) { 2377 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 2378 vm_page_wakeup(m); 2379 return; 2380 } 2381 2382 /* 2383 * Already in the cache (and thus not mapped) 2384 */ 2385 if ((m->queue - m->pc) == PQ_CACHE) { 2386 KKASSERT((m->flags & PG_MAPPED) == 0); 2387 vm_page_wakeup(m); 2388 return; 2389 } 2390 2391 /* 2392 * Caller is required to test m->dirty, but note that the act of 2393 * removing the page from its maps can cause it to become dirty 2394 * on an SMP system due to another cpu running in usermode. 2395 */ 2396 if (m->dirty) { 2397 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2398 (long)m->pindex); 2399 } 2400 2401 /* 2402 * Remove all pmaps and indicate that the page is not 2403 * writeable or mapped. Our vm_page_protect() call may 2404 * have blocked (especially w/ VM_PROT_NONE), so recheck 2405 * everything. 2406 */ 2407 vm_page_protect(m, VM_PROT_NONE); 2408 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2409 m->busy || m->wire_count || m->hold_count) { 2410 vm_page_wakeup(m); 2411 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2412 vm_page_deactivate(m); 2413 vm_page_wakeup(m); 2414 } else { 2415 _vm_page_and_queue_spin_lock(m); 2416 _vm_page_rem_queue_spinlocked(m); 2417 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2418 _vm_page_queue_spin_unlock(m); 2419 if (_vm_page_wakeup(m)) { 2420 vm_page_spin_unlock(m); 2421 wakeup(m); 2422 } else { 2423 vm_page_spin_unlock(m); 2424 } 2425 vm_page_free_wakeup(); 2426 } 2427 } 2428 2429 /* 2430 * vm_page_dontneed() 2431 * 2432 * Cache, deactivate, or do nothing as appropriate. This routine 2433 * is typically used by madvise() MADV_DONTNEED. 2434 * 2435 * Generally speaking we want to move the page into the cache so 2436 * it gets reused quickly. However, this can result in a silly syndrome 2437 * due to the page recycling too quickly. Small objects will not be 2438 * fully cached. On the otherhand, if we move the page to the inactive 2439 * queue we wind up with a problem whereby very large objects 2440 * unnecessarily blow away our inactive and cache queues. 2441 * 2442 * The solution is to move the pages based on a fixed weighting. We 2443 * either leave them alone, deactivate them, or move them to the cache, 2444 * where moving them to the cache has the highest weighting. 2445 * By forcing some pages into other queues we eventually force the 2446 * system to balance the queues, potentially recovering other unrelated 2447 * space from active. The idea is to not force this to happen too 2448 * often. 2449 * 2450 * The page must be busied. 2451 */ 2452 void 2453 vm_page_dontneed(vm_page_t m) 2454 { 2455 static int dnweight; 2456 int dnw; 2457 int head; 2458 2459 dnw = ++dnweight; 2460 2461 /* 2462 * occassionally leave the page alone 2463 */ 2464 if ((dnw & 0x01F0) == 0 || 2465 m->queue - m->pc == PQ_INACTIVE || 2466 m->queue - m->pc == PQ_CACHE 2467 ) { 2468 if (m->act_count >= ACT_INIT) 2469 --m->act_count; 2470 return; 2471 } 2472 2473 /* 2474 * If vm_page_dontneed() is inactivating a page, it must clear 2475 * the referenced flag; otherwise the pagedaemon will see references 2476 * on the page in the inactive queue and reactivate it. Until the 2477 * page can move to the cache queue, madvise's job is not done. 2478 */ 2479 vm_page_flag_clear(m, PG_REFERENCED); 2480 pmap_clear_reference(m); 2481 2482 if (m->dirty == 0) 2483 vm_page_test_dirty(m); 2484 2485 if (m->dirty || (dnw & 0x0070) == 0) { 2486 /* 2487 * Deactivate the page 3 times out of 32. 2488 */ 2489 head = 0; 2490 } else { 2491 /* 2492 * Cache the page 28 times out of every 32. Note that 2493 * the page is deactivated instead of cached, but placed 2494 * at the head of the queue instead of the tail. 2495 */ 2496 head = 1; 2497 } 2498 vm_page_spin_lock(m); 2499 _vm_page_deactivate_locked(m, head); 2500 vm_page_spin_unlock(m); 2501 } 2502 2503 /* 2504 * These routines manipulate the 'soft busy' count for a page. A soft busy 2505 * is almost like PG_BUSY except that it allows certain compatible operations 2506 * to occur on the page while it is busy. For example, a page undergoing a 2507 * write can still be mapped read-only. 2508 * 2509 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2510 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2511 * busy bit is cleared. 2512 */ 2513 void 2514 vm_page_io_start(vm_page_t m) 2515 { 2516 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2517 atomic_add_char(&m->busy, 1); 2518 vm_page_flag_set(m, PG_SBUSY); 2519 } 2520 2521 void 2522 vm_page_io_finish(vm_page_t m) 2523 { 2524 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2525 atomic_subtract_char(&m->busy, 1); 2526 if (m->busy == 0) 2527 vm_page_flag_clear(m, PG_SBUSY); 2528 } 2529 2530 /* 2531 * Indicate that a clean VM page requires a filesystem commit and cannot 2532 * be reused. Used by tmpfs. 2533 */ 2534 void 2535 vm_page_need_commit(vm_page_t m) 2536 { 2537 vm_page_flag_set(m, PG_NEED_COMMIT); 2538 } 2539 2540 void 2541 vm_page_clear_commit(vm_page_t m) 2542 { 2543 vm_page_flag_clear(m, PG_NEED_COMMIT); 2544 } 2545 2546 /* 2547 * Grab a page, blocking if it is busy and allocating a page if necessary. 2548 * A busy page is returned or NULL. The page may or may not be valid and 2549 * might not be on a queue (the caller is responsible for the disposition of 2550 * the page). 2551 * 2552 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2553 * page will be zero'd and marked valid. 2554 * 2555 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2556 * valid even if it already exists. 2557 * 2558 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2559 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2560 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2561 * 2562 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2563 * always returned if we had blocked. 2564 * 2565 * This routine may not be called from an interrupt. 2566 * 2567 * PG_ZERO is *ALWAYS* cleared by this routine. 2568 * 2569 * No other requirements. 2570 */ 2571 vm_page_t 2572 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2573 { 2574 vm_page_t m; 2575 int error; 2576 2577 KKASSERT(allocflags & 2578 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2579 vm_object_hold(object); 2580 for (;;) { 2581 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2582 if (error) { 2583 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2584 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2585 m = NULL; 2586 break; 2587 } 2588 /* retry */ 2589 } else if (m == NULL) { 2590 if (allocflags & VM_ALLOC_RETRY) 2591 allocflags |= VM_ALLOC_NULL_OK; 2592 m = vm_page_alloc(object, pindex, 2593 allocflags & ~VM_ALLOC_RETRY); 2594 if (m) 2595 break; 2596 vm_wait(0); 2597 if ((allocflags & VM_ALLOC_RETRY) == 0) 2598 goto failed; 2599 } else { 2600 /* m found */ 2601 break; 2602 } 2603 } 2604 2605 /* 2606 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2607 * 2608 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2609 * valid even if already valid. 2610 */ 2611 if (m->valid == 0) { 2612 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2613 if ((m->flags & PG_ZERO) == 0) 2614 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2615 m->valid = VM_PAGE_BITS_ALL; 2616 } 2617 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2618 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2619 m->valid = VM_PAGE_BITS_ALL; 2620 } 2621 vm_page_flag_clear(m, PG_ZERO); 2622 failed: 2623 vm_object_drop(object); 2624 return(m); 2625 } 2626 2627 /* 2628 * Mapping function for valid bits or for dirty bits in 2629 * a page. May not block. 2630 * 2631 * Inputs are required to range within a page. 2632 * 2633 * No requirements. 2634 * Non blocking. 2635 */ 2636 int 2637 vm_page_bits(int base, int size) 2638 { 2639 int first_bit; 2640 int last_bit; 2641 2642 KASSERT( 2643 base + size <= PAGE_SIZE, 2644 ("vm_page_bits: illegal base/size %d/%d", base, size) 2645 ); 2646 2647 if (size == 0) /* handle degenerate case */ 2648 return(0); 2649 2650 first_bit = base >> DEV_BSHIFT; 2651 last_bit = (base + size - 1) >> DEV_BSHIFT; 2652 2653 return ((2 << last_bit) - (1 << first_bit)); 2654 } 2655 2656 /* 2657 * Sets portions of a page valid and clean. The arguments are expected 2658 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2659 * of any partial chunks touched by the range. The invalid portion of 2660 * such chunks will be zero'd. 2661 * 2662 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2663 * align base to DEV_BSIZE so as not to mark clean a partially 2664 * truncated device block. Otherwise the dirty page status might be 2665 * lost. 2666 * 2667 * This routine may not block. 2668 * 2669 * (base + size) must be less then or equal to PAGE_SIZE. 2670 */ 2671 static void 2672 _vm_page_zero_valid(vm_page_t m, int base, int size) 2673 { 2674 int frag; 2675 int endoff; 2676 2677 if (size == 0) /* handle degenerate case */ 2678 return; 2679 2680 /* 2681 * If the base is not DEV_BSIZE aligned and the valid 2682 * bit is clear, we have to zero out a portion of the 2683 * first block. 2684 */ 2685 2686 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2687 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2688 ) { 2689 pmap_zero_page_area( 2690 VM_PAGE_TO_PHYS(m), 2691 frag, 2692 base - frag 2693 ); 2694 } 2695 2696 /* 2697 * If the ending offset is not DEV_BSIZE aligned and the 2698 * valid bit is clear, we have to zero out a portion of 2699 * the last block. 2700 */ 2701 2702 endoff = base + size; 2703 2704 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2705 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 2706 ) { 2707 pmap_zero_page_area( 2708 VM_PAGE_TO_PHYS(m), 2709 endoff, 2710 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 2711 ); 2712 } 2713 } 2714 2715 /* 2716 * Set valid, clear dirty bits. If validating the entire 2717 * page we can safely clear the pmap modify bit. We also 2718 * use this opportunity to clear the PG_NOSYNC flag. If a process 2719 * takes a write fault on a MAP_NOSYNC memory area the flag will 2720 * be set again. 2721 * 2722 * We set valid bits inclusive of any overlap, but we can only 2723 * clear dirty bits for DEV_BSIZE chunks that are fully within 2724 * the range. 2725 * 2726 * Page must be busied? 2727 * No other requirements. 2728 */ 2729 void 2730 vm_page_set_valid(vm_page_t m, int base, int size) 2731 { 2732 _vm_page_zero_valid(m, base, size); 2733 m->valid |= vm_page_bits(base, size); 2734 } 2735 2736 2737 /* 2738 * Set valid bits and clear dirty bits. 2739 * 2740 * NOTE: This function does not clear the pmap modified bit. 2741 * Also note that e.g. NFS may use a byte-granular base 2742 * and size. 2743 * 2744 * WARNING: Page must be busied? But vfs_clean_one_page() will call 2745 * this without necessarily busying the page (via bdwrite()). 2746 * So for now vm_token must also be held. 2747 * 2748 * No other requirements. 2749 */ 2750 void 2751 vm_page_set_validclean(vm_page_t m, int base, int size) 2752 { 2753 int pagebits; 2754 2755 _vm_page_zero_valid(m, base, size); 2756 pagebits = vm_page_bits(base, size); 2757 m->valid |= pagebits; 2758 m->dirty &= ~pagebits; 2759 if (base == 0 && size == PAGE_SIZE) { 2760 /*pmap_clear_modify(m);*/ 2761 vm_page_flag_clear(m, PG_NOSYNC); 2762 } 2763 } 2764 2765 /* 2766 * Set valid & dirty. Used by buwrite() 2767 * 2768 * WARNING: Page must be busied? But vfs_dirty_one_page() will 2769 * call this function in buwrite() so for now vm_token must 2770 * be held. 2771 * 2772 * No other requirements. 2773 */ 2774 void 2775 vm_page_set_validdirty(vm_page_t m, int base, int size) 2776 { 2777 int pagebits; 2778 2779 pagebits = vm_page_bits(base, size); 2780 m->valid |= pagebits; 2781 m->dirty |= pagebits; 2782 if (m->object) 2783 vm_object_set_writeable_dirty(m->object); 2784 } 2785 2786 /* 2787 * Clear dirty bits. 2788 * 2789 * NOTE: This function does not clear the pmap modified bit. 2790 * Also note that e.g. NFS may use a byte-granular base 2791 * and size. 2792 * 2793 * Page must be busied? 2794 * No other requirements. 2795 */ 2796 void 2797 vm_page_clear_dirty(vm_page_t m, int base, int size) 2798 { 2799 m->dirty &= ~vm_page_bits(base, size); 2800 if (base == 0 && size == PAGE_SIZE) { 2801 /*pmap_clear_modify(m);*/ 2802 vm_page_flag_clear(m, PG_NOSYNC); 2803 } 2804 } 2805 2806 /* 2807 * Make the page all-dirty. 2808 * 2809 * Also make sure the related object and vnode reflect the fact that the 2810 * object may now contain a dirty page. 2811 * 2812 * Page must be busied? 2813 * No other requirements. 2814 */ 2815 void 2816 vm_page_dirty(vm_page_t m) 2817 { 2818 #ifdef INVARIANTS 2819 int pqtype = m->queue - m->pc; 2820 #endif 2821 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 2822 ("vm_page_dirty: page in free/cache queue!")); 2823 if (m->dirty != VM_PAGE_BITS_ALL) { 2824 m->dirty = VM_PAGE_BITS_ALL; 2825 if (m->object) 2826 vm_object_set_writeable_dirty(m->object); 2827 } 2828 } 2829 2830 /* 2831 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2832 * valid and dirty bits for the effected areas are cleared. 2833 * 2834 * Page must be busied? 2835 * Does not block. 2836 * No other requirements. 2837 */ 2838 void 2839 vm_page_set_invalid(vm_page_t m, int base, int size) 2840 { 2841 int bits; 2842 2843 bits = vm_page_bits(base, size); 2844 m->valid &= ~bits; 2845 m->dirty &= ~bits; 2846 m->object->generation++; 2847 } 2848 2849 /* 2850 * The kernel assumes that the invalid portions of a page contain 2851 * garbage, but such pages can be mapped into memory by user code. 2852 * When this occurs, we must zero out the non-valid portions of the 2853 * page so user code sees what it expects. 2854 * 2855 * Pages are most often semi-valid when the end of a file is mapped 2856 * into memory and the file's size is not page aligned. 2857 * 2858 * Page must be busied? 2859 * No other requirements. 2860 */ 2861 void 2862 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2863 { 2864 int b; 2865 int i; 2866 2867 /* 2868 * Scan the valid bits looking for invalid sections that 2869 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2870 * valid bit may be set ) have already been zerod by 2871 * vm_page_set_validclean(). 2872 */ 2873 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2874 if (i == (PAGE_SIZE / DEV_BSIZE) || 2875 (m->valid & (1 << i)) 2876 ) { 2877 if (i > b) { 2878 pmap_zero_page_area( 2879 VM_PAGE_TO_PHYS(m), 2880 b << DEV_BSHIFT, 2881 (i - b) << DEV_BSHIFT 2882 ); 2883 } 2884 b = i + 1; 2885 } 2886 } 2887 2888 /* 2889 * setvalid is TRUE when we can safely set the zero'd areas 2890 * as being valid. We can do this if there are no cache consistency 2891 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2892 */ 2893 if (setvalid) 2894 m->valid = VM_PAGE_BITS_ALL; 2895 } 2896 2897 /* 2898 * Is a (partial) page valid? Note that the case where size == 0 2899 * will return FALSE in the degenerate case where the page is entirely 2900 * invalid, and TRUE otherwise. 2901 * 2902 * Does not block. 2903 * No other requirements. 2904 */ 2905 int 2906 vm_page_is_valid(vm_page_t m, int base, int size) 2907 { 2908 int bits = vm_page_bits(base, size); 2909 2910 if (m->valid && ((m->valid & bits) == bits)) 2911 return 1; 2912 else 2913 return 0; 2914 } 2915 2916 /* 2917 * update dirty bits from pmap/mmu. May not block. 2918 * 2919 * Caller must hold the page busy 2920 */ 2921 void 2922 vm_page_test_dirty(vm_page_t m) 2923 { 2924 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 2925 vm_page_dirty(m); 2926 } 2927 } 2928 2929 /* 2930 * Register an action, associating it with its vm_page 2931 */ 2932 void 2933 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 2934 { 2935 struct vm_page_action_list *list; 2936 int hv; 2937 2938 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 2939 list = &action_list[hv]; 2940 2941 lwkt_gettoken(&vm_token); 2942 vm_page_flag_set(action->m, PG_ACTIONLIST); 2943 action->event = event; 2944 LIST_INSERT_HEAD(list, action, entry); 2945 lwkt_reltoken(&vm_token); 2946 } 2947 2948 /* 2949 * Unregister an action, disassociating it from its related vm_page 2950 */ 2951 void 2952 vm_page_unregister_action(vm_page_action_t action) 2953 { 2954 struct vm_page_action_list *list; 2955 int hv; 2956 2957 lwkt_gettoken(&vm_token); 2958 if (action->event != VMEVENT_NONE) { 2959 action->event = VMEVENT_NONE; 2960 LIST_REMOVE(action, entry); 2961 2962 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 2963 list = &action_list[hv]; 2964 if (LIST_EMPTY(list)) 2965 vm_page_flag_clear(action->m, PG_ACTIONLIST); 2966 } 2967 lwkt_reltoken(&vm_token); 2968 } 2969 2970 /* 2971 * Issue an event on a VM page. Corresponding action structures are 2972 * removed from the page's list and called. 2973 * 2974 * If the vm_page has no more pending action events we clear its 2975 * PG_ACTIONLIST flag. 2976 */ 2977 void 2978 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 2979 { 2980 struct vm_page_action_list *list; 2981 struct vm_page_action *scan; 2982 struct vm_page_action *next; 2983 int hv; 2984 int all; 2985 2986 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 2987 list = &action_list[hv]; 2988 all = 1; 2989 2990 lwkt_gettoken(&vm_token); 2991 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 2992 if (scan->m == m) { 2993 if (scan->event == event) { 2994 scan->event = VMEVENT_NONE; 2995 LIST_REMOVE(scan, entry); 2996 scan->func(m, scan); 2997 /* XXX */ 2998 } else { 2999 all = 0; 3000 } 3001 } 3002 } 3003 if (all) 3004 vm_page_flag_clear(m, PG_ACTIONLIST); 3005 lwkt_reltoken(&vm_token); 3006 } 3007 3008 #include "opt_ddb.h" 3009 #ifdef DDB 3010 #include <sys/kernel.h> 3011 3012 #include <ddb/ddb.h> 3013 3014 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3015 { 3016 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3017 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3018 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3019 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3020 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3021 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3022 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3023 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3024 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3025 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3026 } 3027 3028 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3029 { 3030 int i; 3031 db_printf("PQ_FREE:"); 3032 for(i=0;i<PQ_L2_SIZE;i++) { 3033 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3034 } 3035 db_printf("\n"); 3036 3037 db_printf("PQ_CACHE:"); 3038 for(i=0;i<PQ_L2_SIZE;i++) { 3039 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3040 } 3041 db_printf("\n"); 3042 3043 db_printf("PQ_ACTIVE:"); 3044 for(i=0;i<PQ_L2_SIZE;i++) { 3045 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3046 } 3047 db_printf("\n"); 3048 3049 db_printf("PQ_INACTIVE:"); 3050 for(i=0;i<PQ_L2_SIZE;i++) { 3051 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3052 } 3053 db_printf("\n"); 3054 } 3055 #endif /* DDB */ 3056