1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. 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 #include <machine/specialreg.h> 95 96 #include <vm/vm_page2.h> 97 #include <sys/spinlock2.h> 98 99 #define VMACTION_HSIZE 256 100 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 101 102 static void vm_page_queue_init(void); 103 static void vm_page_free_wakeup(void); 104 static vm_page_t vm_page_select_cache(u_short pg_color); 105 static vm_page_t _vm_page_list_find2(int basequeue, int index); 106 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 107 108 /* 109 * Array of tailq lists 110 */ 111 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 112 113 LIST_HEAD(vm_page_action_list, vm_page_action); 114 struct vm_page_action_list action_list[VMACTION_HSIZE]; 115 static volatile int vm_pages_waiting; 116 117 static struct alist vm_contig_alist; 118 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 119 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin); 120 121 static u_long vm_dma_reserved = 0; 122 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 123 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 124 "Memory reserved for DMA"); 125 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 126 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 127 128 static int vm_contig_verbose = 0; 129 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 130 131 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 132 vm_pindex_t, pindex); 133 134 static void 135 vm_page_queue_init(void) 136 { 137 int i; 138 139 for (i = 0; i < PQ_L2_SIZE; i++) 140 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 141 for (i = 0; i < PQ_L2_SIZE; i++) 142 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 143 for (i = 0; i < PQ_L2_SIZE; i++) 144 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count; 145 for (i = 0; i < PQ_L2_SIZE; i++) 146 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count; 147 for (i = 0; i < PQ_L2_SIZE; i++) 148 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count; 149 /* PQ_NONE has no queue */ 150 151 for (i = 0; i < PQ_COUNT; i++) { 152 TAILQ_INIT(&vm_page_queues[i].pl); 153 spin_init(&vm_page_queues[i].spin); 154 } 155 156 for (i = 0; i < VMACTION_HSIZE; i++) 157 LIST_INIT(&action_list[i]); 158 } 159 160 /* 161 * note: place in initialized data section? Is this necessary? 162 */ 163 long first_page = 0; 164 int vm_page_array_size = 0; 165 int vm_page_zero_count = 0; 166 vm_page_t vm_page_array = NULL; 167 vm_paddr_t vm_low_phys_reserved; 168 169 /* 170 * (low level boot) 171 * 172 * Sets the page size, perhaps based upon the memory size. 173 * Must be called before any use of page-size dependent functions. 174 */ 175 void 176 vm_set_page_size(void) 177 { 178 if (vmstats.v_page_size == 0) 179 vmstats.v_page_size = PAGE_SIZE; 180 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 181 panic("vm_set_page_size: page size not a power of two"); 182 } 183 184 /* 185 * (low level boot) 186 * 187 * Add a new page to the freelist for use by the system. New pages 188 * are added to both the head and tail of the associated free page 189 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 190 * requests pull 'recent' adds (higher physical addresses) first. 191 * 192 * Beware that the page zeroing daemon will also be running soon after 193 * boot, moving pages from the head to the tail of the PQ_FREE queues. 194 * 195 * Must be called in a critical section. 196 */ 197 static void 198 vm_add_new_page(vm_paddr_t pa) 199 { 200 struct vpgqueues *vpq; 201 vm_page_t m; 202 203 m = PHYS_TO_VM_PAGE(pa); 204 m->phys_addr = pa; 205 m->flags = 0; 206 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 207 m->pat_mode = PAT_WRITE_BACK; 208 /* 209 * Twist for cpu localization in addition to page coloring, so 210 * different cpus selecting by m->queue get different page colors. 211 */ 212 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK; 213 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK; 214 /* 215 * Reserve a certain number of contiguous low memory pages for 216 * contigmalloc() to use. 217 */ 218 if (pa < vm_low_phys_reserved) { 219 atomic_add_int(&vmstats.v_page_count, 1); 220 atomic_add_int(&vmstats.v_dma_pages, 1); 221 m->queue = PQ_NONE; 222 m->wire_count = 1; 223 atomic_add_int(&vmstats.v_wire_count, 1); 224 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 225 return; 226 } 227 228 /* 229 * General page 230 */ 231 m->queue = m->pc + PQ_FREE; 232 KKASSERT(m->dirty == 0); 233 234 atomic_add_int(&vmstats.v_page_count, 1); 235 atomic_add_int(&vmstats.v_free_count, 1); 236 vpq = &vm_page_queues[m->queue]; 237 if ((vpq->flipflop & 15) == 0) { 238 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 239 m->flags |= PG_ZERO; 240 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 241 atomic_add_int(&vm_page_zero_count, 1); 242 } else { 243 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 244 } 245 ++vpq->flipflop; 246 ++vpq->lcnt; 247 } 248 249 /* 250 * (low level boot) 251 * 252 * Initializes the resident memory module. 253 * 254 * Preallocates memory for critical VM structures and arrays prior to 255 * kernel_map becoming available. 256 * 257 * Memory is allocated from (virtual2_start, virtual2_end) if available, 258 * otherwise memory is allocated from (virtual_start, virtual_end). 259 * 260 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 261 * large enough to hold vm_page_array & other structures for machines with 262 * large amounts of ram, so we want to use virtual2* when available. 263 */ 264 void 265 vm_page_startup(void) 266 { 267 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 268 vm_offset_t mapped; 269 vm_size_t npages; 270 vm_paddr_t page_range; 271 vm_paddr_t new_end; 272 int i; 273 vm_paddr_t pa; 274 int nblocks; 275 vm_paddr_t last_pa; 276 vm_paddr_t end; 277 vm_paddr_t biggestone, biggestsize; 278 vm_paddr_t total; 279 280 total = 0; 281 biggestsize = 0; 282 biggestone = 0; 283 nblocks = 0; 284 vaddr = round_page(vaddr); 285 286 for (i = 0; phys_avail[i + 1]; i += 2) { 287 phys_avail[i] = round_page64(phys_avail[i]); 288 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 289 } 290 291 for (i = 0; phys_avail[i + 1]; i += 2) { 292 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 293 294 if (size > biggestsize) { 295 biggestone = i; 296 biggestsize = size; 297 } 298 ++nblocks; 299 total += size; 300 } 301 302 end = phys_avail[biggestone+1]; 303 end = trunc_page(end); 304 305 /* 306 * Initialize the queue headers for the free queue, the active queue 307 * and the inactive queue. 308 */ 309 vm_page_queue_init(); 310 311 #if !defined(_KERNEL_VIRTUAL) 312 /* 313 * VKERNELs don't support minidumps and as such don't need 314 * vm_page_dump 315 * 316 * Allocate a bitmap to indicate that a random physical page 317 * needs to be included in a minidump. 318 * 319 * The amd64 port needs this to indicate which direct map pages 320 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 321 * 322 * However, i386 still needs this workspace internally within the 323 * minidump code. In theory, they are not needed on i386, but are 324 * included should the sf_buf code decide to use them. 325 */ 326 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 327 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 328 end -= vm_page_dump_size; 329 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 330 VM_PROT_READ | VM_PROT_WRITE); 331 bzero((void *)vm_page_dump, vm_page_dump_size); 332 #endif 333 /* 334 * Compute the number of pages of memory that will be available for 335 * use (taking into account the overhead of a page structure per 336 * page). 337 */ 338 first_page = phys_avail[0] / PAGE_SIZE; 339 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 340 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 341 342 #ifndef _KERNEL_VIRTUAL 343 /* 344 * (only applies to real kernels) 345 * 346 * Initialize the contiguous reserve map. We initially reserve up 347 * to 1/4 available physical memory or 65536 pages (~256MB), whichever 348 * is lower. 349 * 350 * Once device initialization is complete we return most of the 351 * reserved memory back to the normal page queues but leave some 352 * in reserve for things like usb attachments. 353 */ 354 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 355 if (vm_low_phys_reserved > total / 4) 356 vm_low_phys_reserved = total / 4; 357 if (vm_dma_reserved == 0) { 358 vm_dma_reserved = 16 * 1024 * 1024; /* 16MB */ 359 if (vm_dma_reserved > total / 16) 360 vm_dma_reserved = total / 16; 361 } 362 #endif 363 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 364 ALIST_RECORDS_65536); 365 366 /* 367 * Initialize the mem entry structures now, and put them in the free 368 * queue. 369 */ 370 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 371 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 372 vm_page_array = (vm_page_t)mapped; 373 374 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 375 /* 376 * since pmap_map on amd64 returns stuff out of a direct-map region, 377 * we have to manually add these pages to the minidump tracking so 378 * that they can be dumped, including the vm_page_array. 379 */ 380 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 381 dump_add_page(pa); 382 #endif 383 384 /* 385 * Clear all of the page structures 386 */ 387 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 388 vm_page_array_size = page_range; 389 390 /* 391 * Construct the free queue(s) in ascending order (by physical 392 * address) so that the first 16MB of physical memory is allocated 393 * last rather than first. On large-memory machines, this avoids 394 * the exhaustion of low physical memory before isa_dmainit has run. 395 */ 396 vmstats.v_page_count = 0; 397 vmstats.v_free_count = 0; 398 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 399 pa = phys_avail[i]; 400 if (i == biggestone) 401 last_pa = new_end; 402 else 403 last_pa = phys_avail[i + 1]; 404 while (pa < last_pa && npages-- > 0) { 405 vm_add_new_page(pa); 406 pa += PAGE_SIZE; 407 } 408 } 409 if (virtual2_start) 410 virtual2_start = vaddr; 411 else 412 virtual_start = vaddr; 413 } 414 415 /* 416 * We tended to reserve a ton of memory for contigmalloc(). Now that most 417 * drivers have initialized we want to return most the remaining free 418 * reserve back to the VM page queues so they can be used for normal 419 * allocations. 420 * 421 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 422 */ 423 static void 424 vm_page_startup_finish(void *dummy __unused) 425 { 426 alist_blk_t blk; 427 alist_blk_t rblk; 428 alist_blk_t count; 429 alist_blk_t xcount; 430 alist_blk_t bfree; 431 vm_page_t m; 432 433 spin_lock(&vm_contig_spin); 434 for (;;) { 435 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 436 if (bfree <= vm_dma_reserved / PAGE_SIZE) 437 break; 438 if (count == 0) 439 break; 440 441 /* 442 * Figure out how much of the initial reserve we have to 443 * free in order to reach our target. 444 */ 445 bfree -= vm_dma_reserved / PAGE_SIZE; 446 if (count > bfree) { 447 blk += count - bfree; 448 count = bfree; 449 } 450 451 /* 452 * Calculate the nearest power of 2 <= count. 453 */ 454 for (xcount = 1; xcount <= count; xcount <<= 1) 455 ; 456 xcount >>= 1; 457 blk += count - xcount; 458 count = xcount; 459 460 /* 461 * Allocate the pages from the alist, then free them to 462 * the normal VM page queues. 463 * 464 * Pages allocated from the alist are wired. We have to 465 * busy, unwire, and free them. We must also adjust 466 * vm_low_phys_reserved before freeing any pages to prevent 467 * confusion. 468 */ 469 rblk = alist_alloc(&vm_contig_alist, blk, count); 470 if (rblk != blk) { 471 kprintf("vm_page_startup_finish: Unable to return " 472 "dma space @0x%08x/%d -> 0x%08x\n", 473 blk, count, rblk); 474 break; 475 } 476 atomic_add_int(&vmstats.v_dma_pages, -count); 477 spin_unlock(&vm_contig_spin); 478 479 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 480 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 481 while (count) { 482 vm_page_busy_wait(m, FALSE, "cpgfr"); 483 vm_page_unwire(m, 0); 484 vm_page_free(m); 485 --count; 486 ++m; 487 } 488 spin_lock(&vm_contig_spin); 489 } 490 spin_unlock(&vm_contig_spin); 491 492 /* 493 * Print out how much DMA space drivers have already allocated and 494 * how much is left over. 495 */ 496 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 497 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 498 (PAGE_SIZE / 1024), 499 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 500 } 501 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 502 vm_page_startup_finish, NULL) 503 504 505 /* 506 * Scan comparison function for Red-Black tree scans. An inclusive 507 * (start,end) is expected. Other fields are not used. 508 */ 509 int 510 rb_vm_page_scancmp(struct vm_page *p, void *data) 511 { 512 struct rb_vm_page_scan_info *info = data; 513 514 if (p->pindex < info->start_pindex) 515 return(-1); 516 if (p->pindex > info->end_pindex) 517 return(1); 518 return(0); 519 } 520 521 int 522 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 523 { 524 if (p1->pindex < p2->pindex) 525 return(-1); 526 if (p1->pindex > p2->pindex) 527 return(1); 528 return(0); 529 } 530 531 void 532 vm_page_init(vm_page_t m) 533 { 534 /* do nothing for now. Called from pmap_page_init() */ 535 } 536 537 /* 538 * Each page queue has its own spin lock, which is fairly optimal for 539 * allocating and freeing pages at least. 540 * 541 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 542 * queue spinlock via this function. Also note that m->queue cannot change 543 * unless both the page and queue are locked. 544 */ 545 static __inline 546 void 547 _vm_page_queue_spin_lock(vm_page_t m) 548 { 549 u_short queue; 550 551 queue = m->queue; 552 if (queue != PQ_NONE) { 553 spin_lock(&vm_page_queues[queue].spin); 554 KKASSERT(queue == m->queue); 555 } 556 } 557 558 static __inline 559 void 560 _vm_page_queue_spin_unlock(vm_page_t m) 561 { 562 u_short queue; 563 564 queue = m->queue; 565 cpu_ccfence(); 566 if (queue != PQ_NONE) 567 spin_unlock(&vm_page_queues[queue].spin); 568 } 569 570 static __inline 571 void 572 _vm_page_queues_spin_lock(u_short queue) 573 { 574 cpu_ccfence(); 575 if (queue != PQ_NONE) 576 spin_lock(&vm_page_queues[queue].spin); 577 } 578 579 580 static __inline 581 void 582 _vm_page_queues_spin_unlock(u_short queue) 583 { 584 cpu_ccfence(); 585 if (queue != PQ_NONE) 586 spin_unlock(&vm_page_queues[queue].spin); 587 } 588 589 void 590 vm_page_queue_spin_lock(vm_page_t m) 591 { 592 _vm_page_queue_spin_lock(m); 593 } 594 595 void 596 vm_page_queues_spin_lock(u_short queue) 597 { 598 _vm_page_queues_spin_lock(queue); 599 } 600 601 void 602 vm_page_queue_spin_unlock(vm_page_t m) 603 { 604 _vm_page_queue_spin_unlock(m); 605 } 606 607 void 608 vm_page_queues_spin_unlock(u_short queue) 609 { 610 _vm_page_queues_spin_unlock(queue); 611 } 612 613 /* 614 * This locks the specified vm_page and its queue in the proper order 615 * (page first, then queue). The queue may change so the caller must 616 * recheck on return. 617 */ 618 static __inline 619 void 620 _vm_page_and_queue_spin_lock(vm_page_t m) 621 { 622 vm_page_spin_lock(m); 623 _vm_page_queue_spin_lock(m); 624 } 625 626 static __inline 627 void 628 _vm_page_and_queue_spin_unlock(vm_page_t m) 629 { 630 _vm_page_queues_spin_unlock(m->queue); 631 vm_page_spin_unlock(m); 632 } 633 634 void 635 vm_page_and_queue_spin_unlock(vm_page_t m) 636 { 637 _vm_page_and_queue_spin_unlock(m); 638 } 639 640 void 641 vm_page_and_queue_spin_lock(vm_page_t m) 642 { 643 _vm_page_and_queue_spin_lock(m); 644 } 645 646 /* 647 * Helper function removes vm_page from its current queue. 648 * Returns the base queue the page used to be on. 649 * 650 * The vm_page and the queue must be spinlocked. 651 * This function will unlock the queue but leave the page spinlocked. 652 */ 653 static __inline u_short 654 _vm_page_rem_queue_spinlocked(vm_page_t m) 655 { 656 struct vpgqueues *pq; 657 u_short queue; 658 659 queue = m->queue; 660 if (queue != PQ_NONE) { 661 pq = &vm_page_queues[queue]; 662 TAILQ_REMOVE(&pq->pl, m, pageq); 663 atomic_add_int(pq->cnt, -1); 664 pq->lcnt--; 665 m->queue = PQ_NONE; 666 vm_page_queues_spin_unlock(queue); 667 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO)) 668 atomic_subtract_int(&vm_page_zero_count, 1); 669 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 670 return (queue - m->pc); 671 } 672 return queue; 673 } 674 675 /* 676 * Helper function places the vm_page on the specified queue. 677 * 678 * The vm_page must be spinlocked. 679 * This function will return with both the page and the queue locked. 680 */ 681 static __inline void 682 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 683 { 684 struct vpgqueues *pq; 685 686 KKASSERT(m->queue == PQ_NONE); 687 688 if (queue != PQ_NONE) { 689 vm_page_queues_spin_lock(queue); 690 pq = &vm_page_queues[queue]; 691 ++pq->lcnt; 692 atomic_add_int(pq->cnt, 1); 693 m->queue = queue; 694 695 /* 696 * Put zero'd pages on the end ( where we look for zero'd pages 697 * first ) and non-zerod pages at the head. 698 */ 699 if (queue - m->pc == PQ_FREE) { 700 if (m->flags & PG_ZERO) { 701 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 702 atomic_add_int(&vm_page_zero_count, 1); 703 } else { 704 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 705 } 706 } else if (athead) { 707 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 708 } else { 709 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 710 } 711 /* leave the queue spinlocked */ 712 } 713 } 714 715 /* 716 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 717 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we 718 * did not. Only one sleep call will be made before returning. 719 * 720 * This function does NOT busy the page and on return the page is not 721 * guaranteed to be available. 722 */ 723 void 724 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 725 { 726 u_int32_t flags; 727 728 for (;;) { 729 flags = m->flags; 730 cpu_ccfence(); 731 732 if ((flags & PG_BUSY) == 0 && 733 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) { 734 break; 735 } 736 tsleep_interlock(m, 0); 737 if (atomic_cmpset_int(&m->flags, flags, 738 flags | PG_WANTED | PG_REFERENCED)) { 739 tsleep(m, PINTERLOCKED, msg, 0); 740 break; 741 } 742 } 743 } 744 745 /* 746 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we 747 * also wait for m->busy to become 0 before setting PG_BUSY. 748 */ 749 void 750 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 751 int also_m_busy, const char *msg 752 VM_PAGE_DEBUG_ARGS) 753 { 754 u_int32_t flags; 755 756 for (;;) { 757 flags = m->flags; 758 cpu_ccfence(); 759 if (flags & PG_BUSY) { 760 tsleep_interlock(m, 0); 761 if (atomic_cmpset_int(&m->flags, flags, 762 flags | PG_WANTED | PG_REFERENCED)) { 763 tsleep(m, PINTERLOCKED, msg, 0); 764 } 765 } else if (also_m_busy && (flags & PG_SBUSY)) { 766 tsleep_interlock(m, 0); 767 if (atomic_cmpset_int(&m->flags, flags, 768 flags | PG_WANTED | PG_REFERENCED)) { 769 tsleep(m, PINTERLOCKED, msg, 0); 770 } 771 } else { 772 if (atomic_cmpset_int(&m->flags, flags, 773 flags | PG_BUSY)) { 774 #ifdef VM_PAGE_DEBUG 775 m->busy_func = func; 776 m->busy_line = lineno; 777 #endif 778 break; 779 } 780 } 781 } 782 } 783 784 /* 785 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy 786 * is also 0. 787 * 788 * Returns non-zero on failure. 789 */ 790 int 791 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 792 VM_PAGE_DEBUG_ARGS) 793 { 794 u_int32_t flags; 795 796 for (;;) { 797 flags = m->flags; 798 cpu_ccfence(); 799 if (flags & PG_BUSY) 800 return TRUE; 801 if (also_m_busy && (flags & PG_SBUSY)) 802 return TRUE; 803 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 804 #ifdef VM_PAGE_DEBUG 805 m->busy_func = func; 806 m->busy_line = lineno; 807 #endif 808 return FALSE; 809 } 810 } 811 } 812 813 /* 814 * Clear the PG_BUSY flag and return non-zero to indicate to the caller 815 * that a wakeup() should be performed. 816 * 817 * The vm_page must be spinlocked and will remain spinlocked on return. 818 * The related queue must NOT be spinlocked (which could deadlock us). 819 * 820 * (inline version) 821 */ 822 static __inline 823 int 824 _vm_page_wakeup(vm_page_t m) 825 { 826 u_int32_t flags; 827 828 for (;;) { 829 flags = m->flags; 830 cpu_ccfence(); 831 if (atomic_cmpset_int(&m->flags, flags, 832 flags & ~(PG_BUSY | PG_WANTED))) { 833 break; 834 } 835 } 836 return(flags & PG_WANTED); 837 } 838 839 /* 840 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This 841 * is typically the last call you make on a page before moving onto 842 * other things. 843 */ 844 void 845 vm_page_wakeup(vm_page_t m) 846 { 847 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 848 vm_page_spin_lock(m); 849 if (_vm_page_wakeup(m)) { 850 vm_page_spin_unlock(m); 851 wakeup(m); 852 } else { 853 vm_page_spin_unlock(m); 854 } 855 } 856 857 /* 858 * Holding a page keeps it from being reused. Other parts of the system 859 * can still disassociate the page from its current object and free it, or 860 * perform read or write I/O on it and/or otherwise manipulate the page, 861 * but if the page is held the VM system will leave the page and its data 862 * intact and not reuse the page for other purposes until the last hold 863 * reference is released. (see vm_page_wire() if you want to prevent the 864 * page from being disassociated from its object too). 865 * 866 * The caller must still validate the contents of the page and, if necessary, 867 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 868 * before manipulating the page. 869 * 870 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 871 */ 872 void 873 vm_page_hold(vm_page_t m) 874 { 875 vm_page_spin_lock(m); 876 atomic_add_int(&m->hold_count, 1); 877 if (m->queue - m->pc == PQ_FREE) { 878 _vm_page_queue_spin_lock(m); 879 _vm_page_rem_queue_spinlocked(m); 880 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 881 _vm_page_queue_spin_unlock(m); 882 } 883 vm_page_spin_unlock(m); 884 } 885 886 /* 887 * The opposite of vm_page_hold(). A page can be freed while being held, 888 * which places it on the PQ_HOLD queue. If we are able to busy the page 889 * after the hold count drops to zero we will move the page to the 890 * appropriate PQ_FREE queue by calling vm_page_free_toq(). 891 */ 892 void 893 vm_page_unhold(vm_page_t m) 894 { 895 vm_page_spin_lock(m); 896 atomic_add_int(&m->hold_count, -1); 897 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 898 _vm_page_queue_spin_lock(m); 899 _vm_page_rem_queue_spinlocked(m); 900 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 901 _vm_page_queue_spin_unlock(m); 902 } 903 vm_page_spin_unlock(m); 904 } 905 906 /* 907 * vm_page_getfake: 908 * 909 * Create a fictitious page with the specified physical address and 910 * memory attribute. The memory attribute is the only the machine- 911 * dependent aspect of a fictitious page that must be initialized. 912 */ 913 914 void 915 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 916 { 917 918 if ((m->flags & PG_FICTITIOUS) != 0) { 919 /* 920 * The page's memattr might have changed since the 921 * previous initialization. Update the pmap to the 922 * new memattr. 923 */ 924 goto memattr; 925 } 926 m->phys_addr = paddr; 927 m->queue = PQ_NONE; 928 /* Fictitious pages don't use "segind". */ 929 /* Fictitious pages don't use "order" or "pool". */ 930 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY; 931 m->wire_count = 1; 932 pmap_page_init(m); 933 memattr: 934 pmap_page_set_memattr(m, memattr); 935 } 936 937 /* 938 * Inserts the given vm_page into the object and object list. 939 * 940 * The pagetables are not updated but will presumably fault the page 941 * in if necessary, or if a kernel page the caller will at some point 942 * enter the page into the kernel's pmap. We are not allowed to block 943 * here so we *can't* do this anyway. 944 * 945 * This routine may not block. 946 * This routine must be called with the vm_object held. 947 * This routine must be called with a critical section held. 948 * 949 * This routine returns TRUE if the page was inserted into the object 950 * successfully, and FALSE if the page already exists in the object. 951 */ 952 int 953 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 954 { 955 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 956 if (m->object != NULL) 957 panic("vm_page_insert: already inserted"); 958 959 object->generation++; 960 961 /* 962 * Record the object/offset pair in this page and add the 963 * pv_list_count of the page to the object. 964 * 965 * The vm_page spin lock is required for interactions with the pmap. 966 */ 967 vm_page_spin_lock(m); 968 m->object = object; 969 m->pindex = pindex; 970 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 971 m->object = NULL; 972 m->pindex = 0; 973 vm_page_spin_unlock(m); 974 return FALSE; 975 } 976 ++object->resident_page_count; 977 ++mycpu->gd_vmtotal.t_rm; 978 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */ 979 vm_page_spin_unlock(m); 980 981 /* 982 * Since we are inserting a new and possibly dirty page, 983 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 984 */ 985 if ((m->valid & m->dirty) || 986 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 987 vm_object_set_writeable_dirty(object); 988 989 /* 990 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 991 */ 992 swap_pager_page_inserted(m); 993 return TRUE; 994 } 995 996 /* 997 * Removes the given vm_page_t from the (object,index) table 998 * 999 * The underlying pmap entry (if any) is NOT removed here. 1000 * This routine may not block. 1001 * 1002 * The page must be BUSY and will remain BUSY on return. 1003 * No other requirements. 1004 * 1005 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1006 * it busy. 1007 */ 1008 void 1009 vm_page_remove(vm_page_t m) 1010 { 1011 vm_object_t object; 1012 1013 if (m->object == NULL) { 1014 return; 1015 } 1016 1017 if ((m->flags & PG_BUSY) == 0) 1018 panic("vm_page_remove: page not busy"); 1019 1020 object = m->object; 1021 1022 vm_object_hold(object); 1023 1024 /* 1025 * Remove the page from the object and update the object. 1026 * 1027 * The vm_page spin lock is required for interactions with the pmap. 1028 */ 1029 vm_page_spin_lock(m); 1030 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1031 --object->resident_page_count; 1032 --mycpu->gd_vmtotal.t_rm; 1033 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */ 1034 m->object = NULL; 1035 vm_page_spin_unlock(m); 1036 1037 object->generation++; 1038 1039 vm_object_drop(object); 1040 } 1041 1042 /* 1043 * Locate and return the page at (object, pindex), or NULL if the 1044 * page could not be found. 1045 * 1046 * The caller must hold the vm_object token. 1047 */ 1048 vm_page_t 1049 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1050 { 1051 vm_page_t m; 1052 1053 /* 1054 * Search the hash table for this object/offset pair 1055 */ 1056 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1057 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1058 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1059 return(m); 1060 } 1061 1062 vm_page_t 1063 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1064 vm_pindex_t pindex, 1065 int also_m_busy, const char *msg 1066 VM_PAGE_DEBUG_ARGS) 1067 { 1068 u_int32_t flags; 1069 vm_page_t m; 1070 1071 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1072 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1073 while (m) { 1074 KKASSERT(m->object == object && m->pindex == pindex); 1075 flags = m->flags; 1076 cpu_ccfence(); 1077 if (flags & PG_BUSY) { 1078 tsleep_interlock(m, 0); 1079 if (atomic_cmpset_int(&m->flags, flags, 1080 flags | PG_WANTED | PG_REFERENCED)) { 1081 tsleep(m, PINTERLOCKED, msg, 0); 1082 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1083 pindex); 1084 } 1085 } else if (also_m_busy && (flags & PG_SBUSY)) { 1086 tsleep_interlock(m, 0); 1087 if (atomic_cmpset_int(&m->flags, flags, 1088 flags | PG_WANTED | PG_REFERENCED)) { 1089 tsleep(m, PINTERLOCKED, msg, 0); 1090 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1091 pindex); 1092 } 1093 } else if (atomic_cmpset_int(&m->flags, flags, 1094 flags | PG_BUSY)) { 1095 #ifdef VM_PAGE_DEBUG 1096 m->busy_func = func; 1097 m->busy_line = lineno; 1098 #endif 1099 break; 1100 } 1101 } 1102 return m; 1103 } 1104 1105 /* 1106 * Attempt to lookup and busy a page. 1107 * 1108 * Returns NULL if the page could not be found 1109 * 1110 * Returns a vm_page and error == TRUE if the page exists but could not 1111 * be busied. 1112 * 1113 * Returns a vm_page and error == FALSE on success. 1114 */ 1115 vm_page_t 1116 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1117 vm_pindex_t pindex, 1118 int also_m_busy, int *errorp 1119 VM_PAGE_DEBUG_ARGS) 1120 { 1121 u_int32_t flags; 1122 vm_page_t m; 1123 1124 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1125 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1126 *errorp = FALSE; 1127 while (m) { 1128 KKASSERT(m->object == object && m->pindex == pindex); 1129 flags = m->flags; 1130 cpu_ccfence(); 1131 if (flags & PG_BUSY) { 1132 *errorp = TRUE; 1133 break; 1134 } 1135 if (also_m_busy && (flags & PG_SBUSY)) { 1136 *errorp = TRUE; 1137 break; 1138 } 1139 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1140 #ifdef VM_PAGE_DEBUG 1141 m->busy_func = func; 1142 m->busy_line = lineno; 1143 #endif 1144 break; 1145 } 1146 } 1147 return m; 1148 } 1149 1150 /* 1151 * Caller must hold the related vm_object 1152 */ 1153 vm_page_t 1154 vm_page_next(vm_page_t m) 1155 { 1156 vm_page_t next; 1157 1158 next = vm_page_rb_tree_RB_NEXT(m); 1159 if (next && next->pindex != m->pindex + 1) 1160 next = NULL; 1161 return (next); 1162 } 1163 1164 /* 1165 * vm_page_rename() 1166 * 1167 * Move the given vm_page from its current object to the specified 1168 * target object/offset. The page must be busy and will remain so 1169 * on return. 1170 * 1171 * new_object must be held. 1172 * This routine might block. XXX ? 1173 * 1174 * NOTE: Swap associated with the page must be invalidated by the move. We 1175 * have to do this for several reasons: (1) we aren't freeing the 1176 * page, (2) we are dirtying the page, (3) the VM system is probably 1177 * moving the page from object A to B, and will then later move 1178 * the backing store from A to B and we can't have a conflict. 1179 * 1180 * NOTE: We *always* dirty the page. It is necessary both for the 1181 * fact that we moved it, and because we may be invalidating 1182 * swap. If the page is on the cache, we have to deactivate it 1183 * or vm_page_dirty() will panic. Dirty pages are not allowed 1184 * on the cache. 1185 */ 1186 void 1187 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1188 { 1189 KKASSERT(m->flags & PG_BUSY); 1190 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1191 if (m->object) { 1192 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1193 vm_page_remove(m); 1194 } 1195 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1196 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1197 new_object, new_pindex); 1198 } 1199 if (m->queue - m->pc == PQ_CACHE) 1200 vm_page_deactivate(m); 1201 vm_page_dirty(m); 1202 } 1203 1204 /* 1205 * vm_page_unqueue() without any wakeup. This routine is used when a page 1206 * is being moved between queues or otherwise is to remain BUSYied by the 1207 * caller. 1208 * 1209 * This routine may not block. 1210 */ 1211 void 1212 vm_page_unqueue_nowakeup(vm_page_t m) 1213 { 1214 vm_page_and_queue_spin_lock(m); 1215 (void)_vm_page_rem_queue_spinlocked(m); 1216 vm_page_spin_unlock(m); 1217 } 1218 1219 /* 1220 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1221 * if necessary. 1222 * 1223 * This routine may not block. 1224 */ 1225 void 1226 vm_page_unqueue(vm_page_t m) 1227 { 1228 u_short queue; 1229 1230 vm_page_and_queue_spin_lock(m); 1231 queue = _vm_page_rem_queue_spinlocked(m); 1232 if (queue == PQ_FREE || queue == PQ_CACHE) { 1233 vm_page_spin_unlock(m); 1234 pagedaemon_wakeup(); 1235 } else { 1236 vm_page_spin_unlock(m); 1237 } 1238 } 1239 1240 /* 1241 * vm_page_list_find() 1242 * 1243 * Find a page on the specified queue with color optimization. 1244 * 1245 * The page coloring optimization attempts to locate a page that does 1246 * not overload other nearby pages in the object in the cpu's L1 or L2 1247 * caches. We need this optimization because cpu caches tend to be 1248 * physical caches, while object spaces tend to be virtual. 1249 * 1250 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1251 * and the algorithm is adjusted to localize allocations on a per-core basis. 1252 * This is done by 'twisting' the colors. 1253 * 1254 * The page is returned spinlocked and removed from its queue (it will 1255 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller 1256 * is responsible for dealing with the busy-page case (usually by 1257 * deactivating the page and looping). 1258 * 1259 * NOTE: This routine is carefully inlined. A non-inlined version 1260 * is available for outside callers but the only critical path is 1261 * from within this source file. 1262 * 1263 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1264 * represent stable storage, allowing us to order our locks vm_page 1265 * first, then queue. 1266 */ 1267 static __inline 1268 vm_page_t 1269 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1270 { 1271 vm_page_t m; 1272 1273 for (;;) { 1274 if (prefer_zero) 1275 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 1276 else 1277 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1278 if (m == NULL) { 1279 m = _vm_page_list_find2(basequeue, index); 1280 return(m); 1281 } 1282 vm_page_and_queue_spin_lock(m); 1283 if (m->queue == basequeue + index) { 1284 _vm_page_rem_queue_spinlocked(m); 1285 /* vm_page_t spin held, no queue spin */ 1286 break; 1287 } 1288 vm_page_and_queue_spin_unlock(m); 1289 } 1290 return(m); 1291 } 1292 1293 static vm_page_t 1294 _vm_page_list_find2(int basequeue, int index) 1295 { 1296 int i; 1297 vm_page_t m = NULL; 1298 struct vpgqueues *pq; 1299 1300 pq = &vm_page_queues[basequeue]; 1301 1302 /* 1303 * Note that for the first loop, index+i and index-i wind up at the 1304 * same place. Even though this is not totally optimal, we've already 1305 * blown it by missing the cache case so we do not care. 1306 */ 1307 for (i = PQ_L2_SIZE / 2; i > 0; --i) { 1308 for (;;) { 1309 m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl); 1310 if (m) { 1311 _vm_page_and_queue_spin_lock(m); 1312 if (m->queue == 1313 basequeue + ((index + i) & PQ_L2_MASK)) { 1314 _vm_page_rem_queue_spinlocked(m); 1315 return(m); 1316 } 1317 _vm_page_and_queue_spin_unlock(m); 1318 continue; 1319 } 1320 m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl); 1321 if (m) { 1322 _vm_page_and_queue_spin_lock(m); 1323 if (m->queue == 1324 basequeue + ((index - i) & PQ_L2_MASK)) { 1325 _vm_page_rem_queue_spinlocked(m); 1326 return(m); 1327 } 1328 _vm_page_and_queue_spin_unlock(m); 1329 continue; 1330 } 1331 break; /* next i */ 1332 } 1333 } 1334 return(m); 1335 } 1336 1337 /* 1338 * Returns a vm_page candidate for allocation. The page is not busied so 1339 * it can move around. The caller must busy the page (and typically 1340 * deactivate it if it cannot be busied!) 1341 * 1342 * Returns a spinlocked vm_page that has been removed from its queue. 1343 */ 1344 vm_page_t 1345 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1346 { 1347 return(_vm_page_list_find(basequeue, index, prefer_zero)); 1348 } 1349 1350 /* 1351 * Find a page on the cache queue with color optimization, remove it 1352 * from the queue, and busy it. The returned page will not be spinlocked. 1353 * 1354 * A candidate failure will be deactivated. Candidates can fail due to 1355 * being busied by someone else, in which case they will be deactivated. 1356 * 1357 * This routine may not block. 1358 * 1359 */ 1360 static vm_page_t 1361 vm_page_select_cache(u_short pg_color) 1362 { 1363 vm_page_t m; 1364 1365 for (;;) { 1366 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE); 1367 if (m == NULL) 1368 break; 1369 /* 1370 * (m) has been removed from its queue and spinlocked 1371 */ 1372 if (vm_page_busy_try(m, TRUE)) { 1373 _vm_page_deactivate_locked(m, 0); 1374 vm_page_spin_unlock(m); 1375 } else { 1376 /* 1377 * We successfully busied the page 1378 */ 1379 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1380 m->hold_count == 0 && 1381 m->wire_count == 0 && 1382 (m->dirty & m->valid) == 0) { 1383 vm_page_spin_unlock(m); 1384 pagedaemon_wakeup(); 1385 return(m); 1386 } 1387 1388 /* 1389 * The page cannot be recycled, deactivate it. 1390 */ 1391 _vm_page_deactivate_locked(m, 0); 1392 if (_vm_page_wakeup(m)) { 1393 vm_page_spin_unlock(m); 1394 wakeup(m); 1395 } else { 1396 vm_page_spin_unlock(m); 1397 } 1398 } 1399 } 1400 return (m); 1401 } 1402 1403 /* 1404 * Find a free or zero page, with specified preference. We attempt to 1405 * inline the nominal case and fall back to _vm_page_select_free() 1406 * otherwise. A busied page is removed from the queue and returned. 1407 * 1408 * This routine may not block. 1409 */ 1410 static __inline vm_page_t 1411 vm_page_select_free(u_short pg_color, boolean_t prefer_zero) 1412 { 1413 vm_page_t m; 1414 1415 for (;;) { 1416 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK, 1417 prefer_zero); 1418 if (m == NULL) 1419 break; 1420 if (vm_page_busy_try(m, TRUE)) { 1421 /* 1422 * Various mechanisms such as a pmap_collect can 1423 * result in a busy page on the free queue. We 1424 * have to move the page out of the way so we can 1425 * retry the allocation. If the other thread is not 1426 * allocating the page then m->valid will remain 0 and 1427 * the pageout daemon will free the page later on. 1428 * 1429 * Since we could not busy the page, however, we 1430 * cannot make assumptions as to whether the page 1431 * will be allocated by the other thread or not, 1432 * so all we can do is deactivate it to move it out 1433 * of the way. In particular, if the other thread 1434 * wires the page it may wind up on the inactive 1435 * queue and the pageout daemon will have to deal 1436 * with that case too. 1437 */ 1438 _vm_page_deactivate_locked(m, 0); 1439 vm_page_spin_unlock(m); 1440 } else { 1441 /* 1442 * Theoretically if we are able to busy the page 1443 * atomic with the queue removal (using the vm_page 1444 * lock) nobody else should be able to mess with the 1445 * page before us. 1446 */ 1447 KKASSERT((m->flags & (PG_UNMANAGED | 1448 PG_NEED_COMMIT)) == 0); 1449 KKASSERT(m->hold_count == 0); 1450 KKASSERT(m->wire_count == 0); 1451 vm_page_spin_unlock(m); 1452 pagedaemon_wakeup(); 1453 1454 /* return busied and removed page */ 1455 return(m); 1456 } 1457 } 1458 return(m); 1459 } 1460 1461 /* 1462 * This implements a per-cpu cache of free, zero'd, ready-to-go pages. 1463 * The idea is to populate this cache prior to acquiring any locks so 1464 * we don't wind up potentially zeroing VM pages (under heavy loads) while 1465 * holding potentialy contending locks. 1466 * 1467 * Note that we allocate the page uninserted into anything and use a pindex 1468 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these 1469 * allocations should wind up being uncontended. However, we still want 1470 * to rove across PQ_L2_SIZE. 1471 */ 1472 void 1473 vm_page_pcpu_cache(void) 1474 { 1475 #if 0 1476 globaldata_t gd = mycpu; 1477 vm_page_t m; 1478 1479 if (gd->gd_vmpg_count < GD_MINVMPG) { 1480 crit_enter_gd(gd); 1481 while (gd->gd_vmpg_count < GD_MAXVMPG) { 1482 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask, 1483 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1484 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO); 1485 if (gd->gd_vmpg_count < GD_MAXVMPG) { 1486 if ((m->flags & PG_ZERO) == 0) { 1487 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1488 vm_page_flag_set(m, PG_ZERO); 1489 } 1490 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m; 1491 } else { 1492 vm_page_free(m); 1493 } 1494 } 1495 crit_exit_gd(gd); 1496 } 1497 #endif 1498 } 1499 1500 /* 1501 * vm_page_alloc() 1502 * 1503 * Allocate and return a memory cell associated with this VM object/offset 1504 * pair. If object is NULL an unassociated page will be allocated. 1505 * 1506 * The returned page will be busied and removed from its queues. This 1507 * routine can block and may return NULL if a race occurs and the page 1508 * is found to already exist at the specified (object, pindex). 1509 * 1510 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1511 * VM_ALLOC_QUICK like normal but cannot use cache 1512 * VM_ALLOC_SYSTEM greater free drain 1513 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1514 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1515 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1516 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1517 * (see vm_page_grab()) 1518 * VM_ALLOC_USE_GD ok to use per-gd cache 1519 * 1520 * The object must be held if not NULL 1521 * This routine may not block 1522 * 1523 * Additional special handling is required when called from an interrupt 1524 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1525 * in this case. 1526 */ 1527 vm_page_t 1528 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1529 { 1530 globaldata_t gd = mycpu; 1531 vm_object_t obj; 1532 vm_page_t m; 1533 u_short pg_color; 1534 1535 #if 0 1536 /* 1537 * Special per-cpu free VM page cache. The pages are pre-busied 1538 * and pre-zerod for us. 1539 */ 1540 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1541 crit_enter_gd(gd); 1542 if (gd->gd_vmpg_count) { 1543 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1544 crit_exit_gd(gd); 1545 goto done; 1546 } 1547 crit_exit_gd(gd); 1548 } 1549 #endif 1550 m = NULL; 1551 1552 /* 1553 * Cpu twist - cpu localization algorithm 1554 */ 1555 if (object) { 1556 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) + 1557 (object->pg_color & ~ncpus_fit_mask); 1558 } else { 1559 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask); 1560 } 1561 KKASSERT(page_req & 1562 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1563 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1564 1565 /* 1566 * Certain system threads (pageout daemon, buf_daemon's) are 1567 * allowed to eat deeper into the free page list. 1568 */ 1569 if (curthread->td_flags & TDF_SYSTHREAD) 1570 page_req |= VM_ALLOC_SYSTEM; 1571 1572 loop: 1573 if (vmstats.v_free_count > vmstats.v_free_reserved || 1574 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 1575 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 1576 vmstats.v_free_count > vmstats.v_interrupt_free_min) 1577 ) { 1578 /* 1579 * The free queue has sufficient free pages to take one out. 1580 */ 1581 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) 1582 m = vm_page_select_free(pg_color, TRUE); 1583 else 1584 m = vm_page_select_free(pg_color, FALSE); 1585 } else if (page_req & VM_ALLOC_NORMAL) { 1586 /* 1587 * Allocatable from the cache (non-interrupt only). On 1588 * success, we must free the page and try again, thus 1589 * ensuring that vmstats.v_*_free_min counters are replenished. 1590 */ 1591 #ifdef INVARIANTS 1592 if (curthread->td_preempted) { 1593 kprintf("vm_page_alloc(): warning, attempt to allocate" 1594 " cache page from preempting interrupt\n"); 1595 m = NULL; 1596 } else { 1597 m = vm_page_select_cache(pg_color); 1598 } 1599 #else 1600 m = vm_page_select_cache(pg_color); 1601 #endif 1602 /* 1603 * On success move the page into the free queue and loop. 1604 * 1605 * Only do this if we can safely acquire the vm_object lock, 1606 * because this is effectively a random page and the caller 1607 * might be holding the lock shared, we don't want to 1608 * deadlock. 1609 */ 1610 if (m != NULL) { 1611 KASSERT(m->dirty == 0, 1612 ("Found dirty cache page %p", m)); 1613 if ((obj = m->object) != NULL) { 1614 if (vm_object_hold_try(obj)) { 1615 vm_page_protect(m, VM_PROT_NONE); 1616 vm_page_free(m); 1617 /* m->object NULL here */ 1618 vm_object_drop(obj); 1619 } else { 1620 vm_page_deactivate(m); 1621 vm_page_wakeup(m); 1622 } 1623 } else { 1624 vm_page_protect(m, VM_PROT_NONE); 1625 vm_page_free(m); 1626 } 1627 goto loop; 1628 } 1629 1630 /* 1631 * On failure return NULL 1632 */ 1633 #if defined(DIAGNOSTIC) 1634 if (vmstats.v_cache_count > 0) 1635 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 1636 #endif 1637 vm_pageout_deficit++; 1638 pagedaemon_wakeup(); 1639 return (NULL); 1640 } else { 1641 /* 1642 * No pages available, wakeup the pageout daemon and give up. 1643 */ 1644 vm_pageout_deficit++; 1645 pagedaemon_wakeup(); 1646 return (NULL); 1647 } 1648 1649 /* 1650 * v_free_count can race so loop if we don't find the expected 1651 * page. 1652 */ 1653 if (m == NULL) 1654 goto loop; 1655 1656 /* 1657 * Good page found. The page has already been busied for us and 1658 * removed from its queues. 1659 */ 1660 KASSERT(m->dirty == 0, 1661 ("vm_page_alloc: free/cache page %p was dirty", m)); 1662 KKASSERT(m->queue == PQ_NONE); 1663 1664 #if 0 1665 done: 1666 #endif 1667 /* 1668 * Initialize the structure, inheriting some flags but clearing 1669 * all the rest. The page has already been busied for us. 1670 */ 1671 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY)); 1672 KKASSERT(m->wire_count == 0); 1673 KKASSERT(m->busy == 0); 1674 m->act_count = 0; 1675 m->valid = 0; 1676 1677 /* 1678 * Caller must be holding the object lock (asserted by 1679 * vm_page_insert()). 1680 * 1681 * NOTE: Inserting a page here does not insert it into any pmaps 1682 * (which could cause us to block allocating memory). 1683 * 1684 * NOTE: If no object an unassociated page is allocated, m->pindex 1685 * can be used by the caller for any purpose. 1686 */ 1687 if (object) { 1688 if (vm_page_insert(m, object, pindex) == FALSE) { 1689 vm_page_free(m); 1690 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1691 panic("PAGE RACE %p[%ld]/%p", 1692 object, (long)pindex, m); 1693 m = NULL; 1694 } 1695 } else { 1696 m->pindex = pindex; 1697 } 1698 1699 /* 1700 * Don't wakeup too often - wakeup the pageout daemon when 1701 * we would be nearly out of memory. 1702 */ 1703 pagedaemon_wakeup(); 1704 1705 /* 1706 * A PG_BUSY page is returned. 1707 */ 1708 return (m); 1709 } 1710 1711 /* 1712 * Attempt to allocate contiguous physical memory with the specified 1713 * requirements. 1714 */ 1715 vm_page_t 1716 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1717 unsigned long alignment, unsigned long boundary, 1718 unsigned long size, vm_memattr_t memattr) 1719 { 1720 alist_blk_t blk; 1721 vm_page_t m; 1722 int i; 1723 1724 alignment >>= PAGE_SHIFT; 1725 if (alignment == 0) 1726 alignment = 1; 1727 boundary >>= PAGE_SHIFT; 1728 if (boundary == 0) 1729 boundary = 1; 1730 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1731 1732 spin_lock(&vm_contig_spin); 1733 blk = alist_alloc(&vm_contig_alist, 0, size); 1734 if (blk == ALIST_BLOCK_NONE) { 1735 spin_unlock(&vm_contig_spin); 1736 if (bootverbose) { 1737 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1738 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1739 } 1740 return(NULL); 1741 } 1742 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1743 alist_free(&vm_contig_alist, blk, size); 1744 spin_unlock(&vm_contig_spin); 1745 if (bootverbose) { 1746 kprintf("vm_page_alloc_contig: %ldk high " 1747 "%016jx failed\n", 1748 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1749 (intmax_t)high); 1750 } 1751 return(NULL); 1752 } 1753 spin_unlock(&vm_contig_spin); 1754 if (vm_contig_verbose) { 1755 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 1756 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 1757 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1758 } 1759 1760 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 1761 if (memattr != VM_MEMATTR_DEFAULT) 1762 for (i = 0;i < size;i++) 1763 pmap_page_set_memattr(&m[i], memattr); 1764 return m; 1765 } 1766 1767 /* 1768 * Free contiguously allocated pages. The pages will be wired but not busy. 1769 * When freeing to the alist we leave them wired and not busy. 1770 */ 1771 void 1772 vm_page_free_contig(vm_page_t m, unsigned long size) 1773 { 1774 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 1775 vm_pindex_t start = pa >> PAGE_SHIFT; 1776 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 1777 1778 if (vm_contig_verbose) { 1779 kprintf("vm_page_free_contig: %016jx/%ldk\n", 1780 (intmax_t)pa, size / 1024); 1781 } 1782 if (pa < vm_low_phys_reserved) { 1783 KKASSERT(pa + size <= vm_low_phys_reserved); 1784 spin_lock(&vm_contig_spin); 1785 alist_free(&vm_contig_alist, start, pages); 1786 spin_unlock(&vm_contig_spin); 1787 } else { 1788 while (pages) { 1789 vm_page_busy_wait(m, FALSE, "cpgfr"); 1790 vm_page_unwire(m, 0); 1791 vm_page_free(m); 1792 --pages; 1793 ++m; 1794 } 1795 1796 } 1797 } 1798 1799 1800 /* 1801 * Wait for sufficient free memory for nominal heavy memory use kernel 1802 * operations. 1803 * 1804 * WARNING! Be sure never to call this in any vm_pageout code path, which 1805 * will trivially deadlock the system. 1806 */ 1807 void 1808 vm_wait_nominal(void) 1809 { 1810 while (vm_page_count_min(0)) 1811 vm_wait(0); 1812 } 1813 1814 /* 1815 * Test if vm_wait_nominal() would block. 1816 */ 1817 int 1818 vm_test_nominal(void) 1819 { 1820 if (vm_page_count_min(0)) 1821 return(1); 1822 return(0); 1823 } 1824 1825 /* 1826 * Block until free pages are available for allocation, called in various 1827 * places before memory allocations. 1828 * 1829 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 1830 * more generous then that. 1831 */ 1832 void 1833 vm_wait(int timo) 1834 { 1835 /* 1836 * never wait forever 1837 */ 1838 if (timo == 0) 1839 timo = hz; 1840 lwkt_gettoken(&vm_token); 1841 1842 if (curthread == pagethread) { 1843 /* 1844 * The pageout daemon itself needs pages, this is bad. 1845 */ 1846 if (vm_page_count_min(0)) { 1847 vm_pageout_pages_needed = 1; 1848 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 1849 } 1850 } else { 1851 /* 1852 * Wakeup the pageout daemon if necessary and wait. 1853 * 1854 * Do not wait indefinitely for the target to be reached, 1855 * as load might prevent it from being reached any time soon. 1856 * But wait a little to try to slow down page allocations 1857 * and to give more important threads (the pagedaemon) 1858 * allocation priority. 1859 */ 1860 if (vm_page_count_target()) { 1861 if (vm_pages_needed == 0) { 1862 vm_pages_needed = 1; 1863 wakeup(&vm_pages_needed); 1864 } 1865 ++vm_pages_waiting; /* SMP race ok */ 1866 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 1867 } 1868 } 1869 lwkt_reltoken(&vm_token); 1870 } 1871 1872 /* 1873 * Block until free pages are available for allocation 1874 * 1875 * Called only from vm_fault so that processes page faulting can be 1876 * easily tracked. 1877 */ 1878 void 1879 vm_wait_pfault(void) 1880 { 1881 /* 1882 * Wakeup the pageout daemon if necessary and wait. 1883 * 1884 * Do not wait indefinitely for the target to be reached, 1885 * as load might prevent it from being reached any time soon. 1886 * But wait a little to try to slow down page allocations 1887 * and to give more important threads (the pagedaemon) 1888 * allocation priority. 1889 */ 1890 if (vm_page_count_min(0)) { 1891 lwkt_gettoken(&vm_token); 1892 while (vm_page_count_severe()) { 1893 if (vm_page_count_target()) { 1894 if (vm_pages_needed == 0) { 1895 vm_pages_needed = 1; 1896 wakeup(&vm_pages_needed); 1897 } 1898 ++vm_pages_waiting; /* SMP race ok */ 1899 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 1900 } 1901 } 1902 lwkt_reltoken(&vm_token); 1903 } 1904 } 1905 1906 /* 1907 * Put the specified page on the active list (if appropriate). Ensure 1908 * that act_count is at least ACT_INIT but do not otherwise mess with it. 1909 * 1910 * The caller should be holding the page busied ? XXX 1911 * This routine may not block. 1912 */ 1913 void 1914 vm_page_activate(vm_page_t m) 1915 { 1916 u_short oqueue; 1917 1918 vm_page_spin_lock(m); 1919 if (m->queue - m->pc != PQ_ACTIVE) { 1920 _vm_page_queue_spin_lock(m); 1921 oqueue = _vm_page_rem_queue_spinlocked(m); 1922 /* page is left spinlocked, queue is unlocked */ 1923 1924 if (oqueue == PQ_CACHE) 1925 mycpu->gd_cnt.v_reactivated++; 1926 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1927 if (m->act_count < ACT_INIT) 1928 m->act_count = ACT_INIT; 1929 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 1930 } 1931 _vm_page_and_queue_spin_unlock(m); 1932 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 1933 pagedaemon_wakeup(); 1934 } else { 1935 if (m->act_count < ACT_INIT) 1936 m->act_count = ACT_INIT; 1937 vm_page_spin_unlock(m); 1938 } 1939 } 1940 1941 /* 1942 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1943 * routine is called when a page has been added to the cache or free 1944 * queues. 1945 * 1946 * This routine may not block. 1947 */ 1948 static __inline void 1949 vm_page_free_wakeup(void) 1950 { 1951 /* 1952 * If the pageout daemon itself needs pages, then tell it that 1953 * there are some free. 1954 */ 1955 if (vm_pageout_pages_needed && 1956 vmstats.v_cache_count + vmstats.v_free_count >= 1957 vmstats.v_pageout_free_min 1958 ) { 1959 vm_pageout_pages_needed = 0; 1960 wakeup(&vm_pageout_pages_needed); 1961 } 1962 1963 /* 1964 * Wakeup processes that are waiting on memory. 1965 * 1966 * Generally speaking we want to wakeup stuck processes as soon as 1967 * possible. !vm_page_count_min(0) is the absolute minimum point 1968 * where we can do this. Wait a bit longer to reduce degenerate 1969 * re-blocking (vm_page_free_hysteresis). The target check is just 1970 * to make sure the min-check w/hysteresis does not exceed the 1971 * normal target. 1972 */ 1973 if (vm_pages_waiting) { 1974 if (!vm_page_count_min(vm_page_free_hysteresis) || 1975 !vm_page_count_target()) { 1976 vm_pages_waiting = 0; 1977 wakeup(&vmstats.v_free_count); 1978 ++mycpu->gd_cnt.v_ppwakeups; 1979 } 1980 #if 0 1981 if (!vm_page_count_target()) { 1982 /* 1983 * Plenty of pages are free, wakeup everyone. 1984 */ 1985 vm_pages_waiting = 0; 1986 wakeup(&vmstats.v_free_count); 1987 ++mycpu->gd_cnt.v_ppwakeups; 1988 } else if (!vm_page_count_min(0)) { 1989 /* 1990 * Some pages are free, wakeup someone. 1991 */ 1992 int wcount = vm_pages_waiting; 1993 if (wcount > 0) 1994 --wcount; 1995 vm_pages_waiting = wcount; 1996 wakeup_one(&vmstats.v_free_count); 1997 ++mycpu->gd_cnt.v_ppwakeups; 1998 } 1999 #endif 2000 } 2001 } 2002 2003 /* 2004 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2005 * it from its VM object. 2006 * 2007 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 2008 * return (the page will have been freed). 2009 */ 2010 void 2011 vm_page_free_toq(vm_page_t m) 2012 { 2013 mycpu->gd_cnt.v_tfree++; 2014 KKASSERT((m->flags & PG_MAPPED) == 0); 2015 KKASSERT(m->flags & PG_BUSY); 2016 2017 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 2018 kprintf("vm_page_free: pindex(%lu), busy(%d), " 2019 "PG_BUSY(%d), hold(%d)\n", 2020 (u_long)m->pindex, m->busy, 2021 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 2022 if ((m->queue - m->pc) == PQ_FREE) 2023 panic("vm_page_free: freeing free page"); 2024 else 2025 panic("vm_page_free: freeing busy page"); 2026 } 2027 2028 /* 2029 * Remove from object, spinlock the page and its queues and 2030 * remove from any queue. No queue spinlock will be held 2031 * after this section (because the page was removed from any 2032 * queue). 2033 */ 2034 vm_page_remove(m); 2035 vm_page_and_queue_spin_lock(m); 2036 _vm_page_rem_queue_spinlocked(m); 2037 2038 /* 2039 * No further management of fictitious pages occurs beyond object 2040 * and queue removal. 2041 */ 2042 if ((m->flags & PG_FICTITIOUS) != 0) { 2043 vm_page_spin_unlock(m); 2044 vm_page_wakeup(m); 2045 return; 2046 } 2047 2048 m->valid = 0; 2049 vm_page_undirty(m); 2050 2051 if (m->wire_count != 0) { 2052 if (m->wire_count > 1) { 2053 panic( 2054 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2055 m->wire_count, (long)m->pindex); 2056 } 2057 panic("vm_page_free: freeing wired page"); 2058 } 2059 2060 /* 2061 * Clear the UNMANAGED flag when freeing an unmanaged page. 2062 * Clear the NEED_COMMIT flag 2063 */ 2064 if (m->flags & PG_UNMANAGED) 2065 vm_page_flag_clear(m, PG_UNMANAGED); 2066 if (m->flags & PG_NEED_COMMIT) 2067 vm_page_flag_clear(m, PG_NEED_COMMIT); 2068 2069 if (m->hold_count != 0) { 2070 vm_page_flag_clear(m, PG_ZERO); 2071 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2072 } else { 2073 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 2074 } 2075 2076 /* 2077 * This sequence allows us to clear PG_BUSY while still holding 2078 * its spin lock, which reduces contention vs allocators. We 2079 * must not leave the queue locked or _vm_page_wakeup() may 2080 * deadlock. 2081 */ 2082 _vm_page_queue_spin_unlock(m); 2083 if (_vm_page_wakeup(m)) { 2084 vm_page_spin_unlock(m); 2085 wakeup(m); 2086 } else { 2087 vm_page_spin_unlock(m); 2088 } 2089 vm_page_free_wakeup(); 2090 } 2091 2092 /* 2093 * vm_page_free_fromq_fast() 2094 * 2095 * Remove a non-zero page from one of the free queues; the page is removed for 2096 * zeroing, so do not issue a wakeup. 2097 */ 2098 vm_page_t 2099 vm_page_free_fromq_fast(void) 2100 { 2101 static int qi; 2102 vm_page_t m; 2103 int i; 2104 2105 for (i = 0; i < PQ_L2_SIZE; ++i) { 2106 m = vm_page_list_find(PQ_FREE, qi, FALSE); 2107 /* page is returned spinlocked and removed from its queue */ 2108 if (m) { 2109 if (vm_page_busy_try(m, TRUE)) { 2110 /* 2111 * We were unable to busy the page, deactivate 2112 * it and loop. 2113 */ 2114 _vm_page_deactivate_locked(m, 0); 2115 vm_page_spin_unlock(m); 2116 } else if (m->flags & PG_ZERO) { 2117 /* 2118 * The page is PG_ZERO, requeue it and loop 2119 */ 2120 _vm_page_add_queue_spinlocked(m, 2121 PQ_FREE + m->pc, 2122 0); 2123 vm_page_queue_spin_unlock(m); 2124 if (_vm_page_wakeup(m)) { 2125 vm_page_spin_unlock(m); 2126 wakeup(m); 2127 } else { 2128 vm_page_spin_unlock(m); 2129 } 2130 } else { 2131 /* 2132 * The page is not PG_ZERO'd so return it. 2133 */ 2134 vm_page_spin_unlock(m); 2135 KKASSERT((m->flags & (PG_UNMANAGED | 2136 PG_NEED_COMMIT)) == 0); 2137 KKASSERT(m->hold_count == 0); 2138 KKASSERT(m->wire_count == 0); 2139 break; 2140 } 2141 m = NULL; 2142 } 2143 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 2144 } 2145 return (m); 2146 } 2147 2148 /* 2149 * vm_page_unmanage() 2150 * 2151 * Prevent PV management from being done on the page. The page is 2152 * removed from the paging queues as if it were wired, and as a 2153 * consequence of no longer being managed the pageout daemon will not 2154 * touch it (since there is no way to locate the pte mappings for the 2155 * page). madvise() calls that mess with the pmap will also no longer 2156 * operate on the page. 2157 * 2158 * Beyond that the page is still reasonably 'normal'. Freeing the page 2159 * will clear the flag. 2160 * 2161 * This routine is used by OBJT_PHYS objects - objects using unswappable 2162 * physical memory as backing store rather then swap-backed memory and 2163 * will eventually be extended to support 4MB unmanaged physical 2164 * mappings. 2165 * 2166 * Caller must be holding the page busy. 2167 */ 2168 void 2169 vm_page_unmanage(vm_page_t m) 2170 { 2171 KKASSERT(m->flags & PG_BUSY); 2172 if ((m->flags & PG_UNMANAGED) == 0) { 2173 if (m->wire_count == 0) 2174 vm_page_unqueue(m); 2175 } 2176 vm_page_flag_set(m, PG_UNMANAGED); 2177 } 2178 2179 /* 2180 * Mark this page as wired down by yet another map, removing it from 2181 * paging queues as necessary. 2182 * 2183 * Caller must be holding the page busy. 2184 */ 2185 void 2186 vm_page_wire(vm_page_t m) 2187 { 2188 /* 2189 * Only bump the wire statistics if the page is not already wired, 2190 * and only unqueue the page if it is on some queue (if it is unmanaged 2191 * it is already off the queues). Don't do anything with fictitious 2192 * pages because they are always wired. 2193 */ 2194 KKASSERT(m->flags & PG_BUSY); 2195 if ((m->flags & PG_FICTITIOUS) == 0) { 2196 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2197 if ((m->flags & PG_UNMANAGED) == 0) 2198 vm_page_unqueue(m); 2199 atomic_add_int(&vmstats.v_wire_count, 1); 2200 } 2201 KASSERT(m->wire_count != 0, 2202 ("vm_page_wire: wire_count overflow m=%p", m)); 2203 } 2204 } 2205 2206 /* 2207 * Release one wiring of this page, potentially enabling it to be paged again. 2208 * 2209 * Many pages placed on the inactive queue should actually go 2210 * into the cache, but it is difficult to figure out which. What 2211 * we do instead, if the inactive target is well met, is to put 2212 * clean pages at the head of the inactive queue instead of the tail. 2213 * This will cause them to be moved to the cache more quickly and 2214 * if not actively re-referenced, freed more quickly. If we just 2215 * stick these pages at the end of the inactive queue, heavy filesystem 2216 * meta-data accesses can cause an unnecessary paging load on memory bound 2217 * processes. This optimization causes one-time-use metadata to be 2218 * reused more quickly. 2219 * 2220 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2221 * the inactive queue. This helps the pageout daemon determine memory 2222 * pressure and act on out-of-memory situations more quickly. 2223 * 2224 * BUT, if we are in a low-memory situation we have no choice but to 2225 * put clean pages on the cache queue. 2226 * 2227 * A number of routines use vm_page_unwire() to guarantee that the page 2228 * will go into either the inactive or active queues, and will NEVER 2229 * be placed in the cache - for example, just after dirtying a page. 2230 * dirty pages in the cache are not allowed. 2231 * 2232 * The page queues must be locked. 2233 * This routine may not block. 2234 */ 2235 void 2236 vm_page_unwire(vm_page_t m, int activate) 2237 { 2238 KKASSERT(m->flags & PG_BUSY); 2239 if (m->flags & PG_FICTITIOUS) { 2240 /* do nothing */ 2241 } else if (m->wire_count <= 0) { 2242 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2243 } else { 2244 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2245 atomic_add_int(&vmstats.v_wire_count, -1); 2246 if (m->flags & PG_UNMANAGED) { 2247 ; 2248 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2249 vm_page_spin_lock(m); 2250 _vm_page_add_queue_spinlocked(m, 2251 PQ_ACTIVE + m->pc, 0); 2252 _vm_page_and_queue_spin_unlock(m); 2253 } else { 2254 vm_page_spin_lock(m); 2255 vm_page_flag_clear(m, PG_WINATCFLS); 2256 _vm_page_add_queue_spinlocked(m, 2257 PQ_INACTIVE + m->pc, 0); 2258 ++vm_swapcache_inactive_heuristic; 2259 _vm_page_and_queue_spin_unlock(m); 2260 } 2261 } 2262 } 2263 } 2264 2265 /* 2266 * Move the specified page to the inactive queue. If the page has 2267 * any associated swap, the swap is deallocated. 2268 * 2269 * Normally athead is 0 resulting in LRU operation. athead is set 2270 * to 1 if we want this page to be 'as if it were placed in the cache', 2271 * except without unmapping it from the process address space. 2272 * 2273 * vm_page's spinlock must be held on entry and will remain held on return. 2274 * This routine may not block. 2275 */ 2276 static void 2277 _vm_page_deactivate_locked(vm_page_t m, int athead) 2278 { 2279 u_short oqueue; 2280 2281 /* 2282 * Ignore if already inactive. 2283 */ 2284 if (m->queue - m->pc == PQ_INACTIVE) 2285 return; 2286 _vm_page_queue_spin_lock(m); 2287 oqueue = _vm_page_rem_queue_spinlocked(m); 2288 2289 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2290 if (oqueue == PQ_CACHE) 2291 mycpu->gd_cnt.v_reactivated++; 2292 vm_page_flag_clear(m, PG_WINATCFLS); 2293 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2294 if (athead == 0) 2295 ++vm_swapcache_inactive_heuristic; 2296 } 2297 _vm_page_queue_spin_unlock(m); 2298 /* leaves vm_page spinlocked */ 2299 } 2300 2301 /* 2302 * Attempt to deactivate a page. 2303 * 2304 * No requirements. 2305 */ 2306 void 2307 vm_page_deactivate(vm_page_t m) 2308 { 2309 vm_page_spin_lock(m); 2310 _vm_page_deactivate_locked(m, 0); 2311 vm_page_spin_unlock(m); 2312 } 2313 2314 void 2315 vm_page_deactivate_locked(vm_page_t m) 2316 { 2317 _vm_page_deactivate_locked(m, 0); 2318 } 2319 2320 /* 2321 * Attempt to move a page to PQ_CACHE. 2322 * 2323 * Returns 0 on failure, 1 on success 2324 * 2325 * The page should NOT be busied by the caller. This function will validate 2326 * whether the page can be safely moved to the cache. 2327 */ 2328 int 2329 vm_page_try_to_cache(vm_page_t m) 2330 { 2331 vm_page_spin_lock(m); 2332 if (vm_page_busy_try(m, TRUE)) { 2333 vm_page_spin_unlock(m); 2334 return(0); 2335 } 2336 if (m->dirty || m->hold_count || m->wire_count || 2337 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2338 if (_vm_page_wakeup(m)) { 2339 vm_page_spin_unlock(m); 2340 wakeup(m); 2341 } else { 2342 vm_page_spin_unlock(m); 2343 } 2344 return(0); 2345 } 2346 vm_page_spin_unlock(m); 2347 2348 /* 2349 * Page busied by us and no longer spinlocked. Dirty pages cannot 2350 * be moved to the cache. 2351 */ 2352 vm_page_test_dirty(m); 2353 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2354 vm_page_wakeup(m); 2355 return(0); 2356 } 2357 vm_page_cache(m); 2358 return(1); 2359 } 2360 2361 /* 2362 * Attempt to free the page. If we cannot free it, we do nothing. 2363 * 1 is returned on success, 0 on failure. 2364 * 2365 * No requirements. 2366 */ 2367 int 2368 vm_page_try_to_free(vm_page_t m) 2369 { 2370 vm_page_spin_lock(m); 2371 if (vm_page_busy_try(m, TRUE)) { 2372 vm_page_spin_unlock(m); 2373 return(0); 2374 } 2375 2376 /* 2377 * The page can be in any state, including already being on the free 2378 * queue. Check to see if it really can be freed. 2379 */ 2380 if (m->dirty || /* can't free if it is dirty */ 2381 m->hold_count || /* or held (XXX may be wrong) */ 2382 m->wire_count || /* or wired */ 2383 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2384 PG_NEED_COMMIT)) || /* or needs a commit */ 2385 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2386 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2387 if (_vm_page_wakeup(m)) { 2388 vm_page_spin_unlock(m); 2389 wakeup(m); 2390 } else { 2391 vm_page_spin_unlock(m); 2392 } 2393 return(0); 2394 } 2395 vm_page_spin_unlock(m); 2396 2397 /* 2398 * We can probably free the page. 2399 * 2400 * Page busied by us and no longer spinlocked. Dirty pages will 2401 * not be freed by this function. We have to re-test the 2402 * dirty bit after cleaning out the pmaps. 2403 */ 2404 vm_page_test_dirty(m); 2405 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2406 vm_page_wakeup(m); 2407 return(0); 2408 } 2409 vm_page_protect(m, VM_PROT_NONE); 2410 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2411 vm_page_wakeup(m); 2412 return(0); 2413 } 2414 vm_page_free(m); 2415 return(1); 2416 } 2417 2418 /* 2419 * vm_page_cache 2420 * 2421 * Put the specified page onto the page cache queue (if appropriate). 2422 * 2423 * The page must be busy, and this routine will release the busy and 2424 * possibly even free the page. 2425 */ 2426 void 2427 vm_page_cache(vm_page_t m) 2428 { 2429 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2430 m->busy || m->wire_count || m->hold_count) { 2431 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 2432 vm_page_wakeup(m); 2433 return; 2434 } 2435 2436 /* 2437 * Already in the cache (and thus not mapped) 2438 */ 2439 if ((m->queue - m->pc) == PQ_CACHE) { 2440 KKASSERT((m->flags & PG_MAPPED) == 0); 2441 vm_page_wakeup(m); 2442 return; 2443 } 2444 2445 /* 2446 * Caller is required to test m->dirty, but note that the act of 2447 * removing the page from its maps can cause it to become dirty 2448 * on an SMP system due to another cpu running in usermode. 2449 */ 2450 if (m->dirty) { 2451 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2452 (long)m->pindex); 2453 } 2454 2455 /* 2456 * Remove all pmaps and indicate that the page is not 2457 * writeable or mapped. Our vm_page_protect() call may 2458 * have blocked (especially w/ VM_PROT_NONE), so recheck 2459 * everything. 2460 */ 2461 vm_page_protect(m, VM_PROT_NONE); 2462 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2463 m->busy || m->wire_count || m->hold_count) { 2464 vm_page_wakeup(m); 2465 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2466 vm_page_deactivate(m); 2467 vm_page_wakeup(m); 2468 } else { 2469 _vm_page_and_queue_spin_lock(m); 2470 _vm_page_rem_queue_spinlocked(m); 2471 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2472 _vm_page_queue_spin_unlock(m); 2473 if (_vm_page_wakeup(m)) { 2474 vm_page_spin_unlock(m); 2475 wakeup(m); 2476 } else { 2477 vm_page_spin_unlock(m); 2478 } 2479 vm_page_free_wakeup(); 2480 } 2481 } 2482 2483 /* 2484 * vm_page_dontneed() 2485 * 2486 * Cache, deactivate, or do nothing as appropriate. This routine 2487 * is typically used by madvise() MADV_DONTNEED. 2488 * 2489 * Generally speaking we want to move the page into the cache so 2490 * it gets reused quickly. However, this can result in a silly syndrome 2491 * due to the page recycling too quickly. Small objects will not be 2492 * fully cached. On the otherhand, if we move the page to the inactive 2493 * queue we wind up with a problem whereby very large objects 2494 * unnecessarily blow away our inactive and cache queues. 2495 * 2496 * The solution is to move the pages based on a fixed weighting. We 2497 * either leave them alone, deactivate them, or move them to the cache, 2498 * where moving them to the cache has the highest weighting. 2499 * By forcing some pages into other queues we eventually force the 2500 * system to balance the queues, potentially recovering other unrelated 2501 * space from active. The idea is to not force this to happen too 2502 * often. 2503 * 2504 * The page must be busied. 2505 */ 2506 void 2507 vm_page_dontneed(vm_page_t m) 2508 { 2509 static int dnweight; 2510 int dnw; 2511 int head; 2512 2513 dnw = ++dnweight; 2514 2515 /* 2516 * occassionally leave the page alone 2517 */ 2518 if ((dnw & 0x01F0) == 0 || 2519 m->queue - m->pc == PQ_INACTIVE || 2520 m->queue - m->pc == PQ_CACHE 2521 ) { 2522 if (m->act_count >= ACT_INIT) 2523 --m->act_count; 2524 return; 2525 } 2526 2527 /* 2528 * If vm_page_dontneed() is inactivating a page, it must clear 2529 * the referenced flag; otherwise the pagedaemon will see references 2530 * on the page in the inactive queue and reactivate it. Until the 2531 * page can move to the cache queue, madvise's job is not done. 2532 */ 2533 vm_page_flag_clear(m, PG_REFERENCED); 2534 pmap_clear_reference(m); 2535 2536 if (m->dirty == 0) 2537 vm_page_test_dirty(m); 2538 2539 if (m->dirty || (dnw & 0x0070) == 0) { 2540 /* 2541 * Deactivate the page 3 times out of 32. 2542 */ 2543 head = 0; 2544 } else { 2545 /* 2546 * Cache the page 28 times out of every 32. Note that 2547 * the page is deactivated instead of cached, but placed 2548 * at the head of the queue instead of the tail. 2549 */ 2550 head = 1; 2551 } 2552 vm_page_spin_lock(m); 2553 _vm_page_deactivate_locked(m, head); 2554 vm_page_spin_unlock(m); 2555 } 2556 2557 /* 2558 * These routines manipulate the 'soft busy' count for a page. A soft busy 2559 * is almost like PG_BUSY except that it allows certain compatible operations 2560 * to occur on the page while it is busy. For example, a page undergoing a 2561 * write can still be mapped read-only. 2562 * 2563 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2564 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2565 * busy bit is cleared. 2566 */ 2567 void 2568 vm_page_io_start(vm_page_t m) 2569 { 2570 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2571 atomic_add_char(&m->busy, 1); 2572 vm_page_flag_set(m, PG_SBUSY); 2573 } 2574 2575 void 2576 vm_page_io_finish(vm_page_t m) 2577 { 2578 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2579 atomic_subtract_char(&m->busy, 1); 2580 if (m->busy == 0) 2581 vm_page_flag_clear(m, PG_SBUSY); 2582 } 2583 2584 /* 2585 * Indicate that a clean VM page requires a filesystem commit and cannot 2586 * be reused. Used by tmpfs. 2587 */ 2588 void 2589 vm_page_need_commit(vm_page_t m) 2590 { 2591 vm_page_flag_set(m, PG_NEED_COMMIT); 2592 vm_object_set_writeable_dirty(m->object); 2593 } 2594 2595 void 2596 vm_page_clear_commit(vm_page_t m) 2597 { 2598 vm_page_flag_clear(m, PG_NEED_COMMIT); 2599 } 2600 2601 /* 2602 * Grab a page, blocking if it is busy and allocating a page if necessary. 2603 * A busy page is returned or NULL. The page may or may not be valid and 2604 * might not be on a queue (the caller is responsible for the disposition of 2605 * the page). 2606 * 2607 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2608 * page will be zero'd and marked valid. 2609 * 2610 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2611 * valid even if it already exists. 2612 * 2613 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2614 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2615 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2616 * 2617 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2618 * always returned if we had blocked. 2619 * 2620 * This routine may not be called from an interrupt. 2621 * 2622 * PG_ZERO is *ALWAYS* cleared by this routine. 2623 * 2624 * No other requirements. 2625 */ 2626 vm_page_t 2627 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2628 { 2629 vm_page_t m; 2630 int error; 2631 2632 KKASSERT(allocflags & 2633 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2634 vm_object_hold(object); 2635 for (;;) { 2636 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2637 if (error) { 2638 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2639 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2640 m = NULL; 2641 break; 2642 } 2643 /* retry */ 2644 } else if (m == NULL) { 2645 if (allocflags & VM_ALLOC_RETRY) 2646 allocflags |= VM_ALLOC_NULL_OK; 2647 m = vm_page_alloc(object, pindex, 2648 allocflags & ~VM_ALLOC_RETRY); 2649 if (m) 2650 break; 2651 vm_wait(0); 2652 if ((allocflags & VM_ALLOC_RETRY) == 0) 2653 goto failed; 2654 } else { 2655 /* m found */ 2656 break; 2657 } 2658 } 2659 2660 /* 2661 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2662 * 2663 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2664 * valid even if already valid. 2665 */ 2666 if (m->valid == 0) { 2667 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2668 if ((m->flags & PG_ZERO) == 0) 2669 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2670 m->valid = VM_PAGE_BITS_ALL; 2671 } 2672 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2673 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2674 m->valid = VM_PAGE_BITS_ALL; 2675 } 2676 vm_page_flag_clear(m, PG_ZERO); 2677 failed: 2678 vm_object_drop(object); 2679 return(m); 2680 } 2681 2682 /* 2683 * Mapping function for valid bits or for dirty bits in 2684 * a page. May not block. 2685 * 2686 * Inputs are required to range within a page. 2687 * 2688 * No requirements. 2689 * Non blocking. 2690 */ 2691 int 2692 vm_page_bits(int base, int size) 2693 { 2694 int first_bit; 2695 int last_bit; 2696 2697 KASSERT( 2698 base + size <= PAGE_SIZE, 2699 ("vm_page_bits: illegal base/size %d/%d", base, size) 2700 ); 2701 2702 if (size == 0) /* handle degenerate case */ 2703 return(0); 2704 2705 first_bit = base >> DEV_BSHIFT; 2706 last_bit = (base + size - 1) >> DEV_BSHIFT; 2707 2708 return ((2 << last_bit) - (1 << first_bit)); 2709 } 2710 2711 /* 2712 * Sets portions of a page valid and clean. The arguments are expected 2713 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2714 * of any partial chunks touched by the range. The invalid portion of 2715 * such chunks will be zero'd. 2716 * 2717 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2718 * align base to DEV_BSIZE so as not to mark clean a partially 2719 * truncated device block. Otherwise the dirty page status might be 2720 * lost. 2721 * 2722 * This routine may not block. 2723 * 2724 * (base + size) must be less then or equal to PAGE_SIZE. 2725 */ 2726 static void 2727 _vm_page_zero_valid(vm_page_t m, int base, int size) 2728 { 2729 int frag; 2730 int endoff; 2731 2732 if (size == 0) /* handle degenerate case */ 2733 return; 2734 2735 /* 2736 * If the base is not DEV_BSIZE aligned and the valid 2737 * bit is clear, we have to zero out a portion of the 2738 * first block. 2739 */ 2740 2741 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2742 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2743 ) { 2744 pmap_zero_page_area( 2745 VM_PAGE_TO_PHYS(m), 2746 frag, 2747 base - frag 2748 ); 2749 } 2750 2751 /* 2752 * If the ending offset is not DEV_BSIZE aligned and the 2753 * valid bit is clear, we have to zero out a portion of 2754 * the last block. 2755 */ 2756 2757 endoff = base + size; 2758 2759 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2760 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 2761 ) { 2762 pmap_zero_page_area( 2763 VM_PAGE_TO_PHYS(m), 2764 endoff, 2765 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 2766 ); 2767 } 2768 } 2769 2770 /* 2771 * Set valid, clear dirty bits. If validating the entire 2772 * page we can safely clear the pmap modify bit. We also 2773 * use this opportunity to clear the PG_NOSYNC flag. If a process 2774 * takes a write fault on a MAP_NOSYNC memory area the flag will 2775 * be set again. 2776 * 2777 * We set valid bits inclusive of any overlap, but we can only 2778 * clear dirty bits for DEV_BSIZE chunks that are fully within 2779 * the range. 2780 * 2781 * Page must be busied? 2782 * No other requirements. 2783 */ 2784 void 2785 vm_page_set_valid(vm_page_t m, int base, int size) 2786 { 2787 _vm_page_zero_valid(m, base, size); 2788 m->valid |= vm_page_bits(base, size); 2789 } 2790 2791 2792 /* 2793 * Set valid bits and clear dirty bits. 2794 * 2795 * NOTE: This function does not clear the pmap modified bit. 2796 * Also note that e.g. NFS may use a byte-granular base 2797 * and size. 2798 * 2799 * WARNING: Page must be busied? But vfs_clean_one_page() will call 2800 * this without necessarily busying the page (via bdwrite()). 2801 * So for now vm_token must also be held. 2802 * 2803 * No other requirements. 2804 */ 2805 void 2806 vm_page_set_validclean(vm_page_t m, int base, int size) 2807 { 2808 int pagebits; 2809 2810 _vm_page_zero_valid(m, base, size); 2811 pagebits = vm_page_bits(base, size); 2812 m->valid |= pagebits; 2813 m->dirty &= ~pagebits; 2814 if (base == 0 && size == PAGE_SIZE) { 2815 /*pmap_clear_modify(m);*/ 2816 vm_page_flag_clear(m, PG_NOSYNC); 2817 } 2818 } 2819 2820 /* 2821 * Set valid & dirty. Used by buwrite() 2822 * 2823 * WARNING: Page must be busied? But vfs_dirty_one_page() will 2824 * call this function in buwrite() so for now vm_token must 2825 * be held. 2826 * 2827 * No other requirements. 2828 */ 2829 void 2830 vm_page_set_validdirty(vm_page_t m, int base, int size) 2831 { 2832 int pagebits; 2833 2834 pagebits = vm_page_bits(base, size); 2835 m->valid |= pagebits; 2836 m->dirty |= pagebits; 2837 if (m->object) 2838 vm_object_set_writeable_dirty(m->object); 2839 } 2840 2841 /* 2842 * Clear dirty bits. 2843 * 2844 * NOTE: This function does not clear the pmap modified bit. 2845 * Also note that e.g. NFS may use a byte-granular base 2846 * and size. 2847 * 2848 * Page must be busied? 2849 * No other requirements. 2850 */ 2851 void 2852 vm_page_clear_dirty(vm_page_t m, int base, int size) 2853 { 2854 m->dirty &= ~vm_page_bits(base, size); 2855 if (base == 0 && size == PAGE_SIZE) { 2856 /*pmap_clear_modify(m);*/ 2857 vm_page_flag_clear(m, PG_NOSYNC); 2858 } 2859 } 2860 2861 /* 2862 * Make the page all-dirty. 2863 * 2864 * Also make sure the related object and vnode reflect the fact that the 2865 * object may now contain a dirty page. 2866 * 2867 * Page must be busied? 2868 * No other requirements. 2869 */ 2870 void 2871 vm_page_dirty(vm_page_t m) 2872 { 2873 #ifdef INVARIANTS 2874 int pqtype = m->queue - m->pc; 2875 #endif 2876 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 2877 ("vm_page_dirty: page in free/cache queue!")); 2878 if (m->dirty != VM_PAGE_BITS_ALL) { 2879 m->dirty = VM_PAGE_BITS_ALL; 2880 if (m->object) 2881 vm_object_set_writeable_dirty(m->object); 2882 } 2883 } 2884 2885 /* 2886 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2887 * valid and dirty bits for the effected areas are cleared. 2888 * 2889 * Page must be busied? 2890 * Does not block. 2891 * No other requirements. 2892 */ 2893 void 2894 vm_page_set_invalid(vm_page_t m, int base, int size) 2895 { 2896 int bits; 2897 2898 bits = vm_page_bits(base, size); 2899 m->valid &= ~bits; 2900 m->dirty &= ~bits; 2901 m->object->generation++; 2902 } 2903 2904 /* 2905 * The kernel assumes that the invalid portions of a page contain 2906 * garbage, but such pages can be mapped into memory by user code. 2907 * When this occurs, we must zero out the non-valid portions of the 2908 * page so user code sees what it expects. 2909 * 2910 * Pages are most often semi-valid when the end of a file is mapped 2911 * into memory and the file's size is not page aligned. 2912 * 2913 * Page must be busied? 2914 * No other requirements. 2915 */ 2916 void 2917 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2918 { 2919 int b; 2920 int i; 2921 2922 /* 2923 * Scan the valid bits looking for invalid sections that 2924 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2925 * valid bit may be set ) have already been zerod by 2926 * vm_page_set_validclean(). 2927 */ 2928 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2929 if (i == (PAGE_SIZE / DEV_BSIZE) || 2930 (m->valid & (1 << i)) 2931 ) { 2932 if (i > b) { 2933 pmap_zero_page_area( 2934 VM_PAGE_TO_PHYS(m), 2935 b << DEV_BSHIFT, 2936 (i - b) << DEV_BSHIFT 2937 ); 2938 } 2939 b = i + 1; 2940 } 2941 } 2942 2943 /* 2944 * setvalid is TRUE when we can safely set the zero'd areas 2945 * as being valid. We can do this if there are no cache consistency 2946 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2947 */ 2948 if (setvalid) 2949 m->valid = VM_PAGE_BITS_ALL; 2950 } 2951 2952 /* 2953 * Is a (partial) page valid? Note that the case where size == 0 2954 * will return FALSE in the degenerate case where the page is entirely 2955 * invalid, and TRUE otherwise. 2956 * 2957 * Does not block. 2958 * No other requirements. 2959 */ 2960 int 2961 vm_page_is_valid(vm_page_t m, int base, int size) 2962 { 2963 int bits = vm_page_bits(base, size); 2964 2965 if (m->valid && ((m->valid & bits) == bits)) 2966 return 1; 2967 else 2968 return 0; 2969 } 2970 2971 /* 2972 * update dirty bits from pmap/mmu. May not block. 2973 * 2974 * Caller must hold the page busy 2975 */ 2976 void 2977 vm_page_test_dirty(vm_page_t m) 2978 { 2979 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 2980 vm_page_dirty(m); 2981 } 2982 } 2983 2984 /* 2985 * Register an action, associating it with its vm_page 2986 */ 2987 void 2988 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 2989 { 2990 struct vm_page_action_list *list; 2991 int hv; 2992 2993 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 2994 list = &action_list[hv]; 2995 2996 lwkt_gettoken(&vm_token); 2997 vm_page_flag_set(action->m, PG_ACTIONLIST); 2998 action->event = event; 2999 LIST_INSERT_HEAD(list, action, entry); 3000 lwkt_reltoken(&vm_token); 3001 } 3002 3003 /* 3004 * Unregister an action, disassociating it from its related vm_page 3005 */ 3006 void 3007 vm_page_unregister_action(vm_page_action_t action) 3008 { 3009 struct vm_page_action_list *list; 3010 int hv; 3011 3012 lwkt_gettoken(&vm_token); 3013 if (action->event != VMEVENT_NONE) { 3014 action->event = VMEVENT_NONE; 3015 LIST_REMOVE(action, entry); 3016 3017 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3018 list = &action_list[hv]; 3019 if (LIST_EMPTY(list)) 3020 vm_page_flag_clear(action->m, PG_ACTIONLIST); 3021 } 3022 lwkt_reltoken(&vm_token); 3023 } 3024 3025 /* 3026 * Issue an event on a VM page. Corresponding action structures are 3027 * removed from the page's list and called. 3028 * 3029 * If the vm_page has no more pending action events we clear its 3030 * PG_ACTIONLIST flag. 3031 */ 3032 void 3033 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 3034 { 3035 struct vm_page_action_list *list; 3036 struct vm_page_action *scan; 3037 struct vm_page_action *next; 3038 int hv; 3039 int all; 3040 3041 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 3042 list = &action_list[hv]; 3043 all = 1; 3044 3045 lwkt_gettoken(&vm_token); 3046 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 3047 if (scan->m == m) { 3048 if (scan->event == event) { 3049 scan->event = VMEVENT_NONE; 3050 LIST_REMOVE(scan, entry); 3051 scan->func(m, scan); 3052 /* XXX */ 3053 } else { 3054 all = 0; 3055 } 3056 } 3057 } 3058 if (all) 3059 vm_page_flag_clear(m, PG_ACTIONLIST); 3060 lwkt_reltoken(&vm_token); 3061 } 3062 3063 #include "opt_ddb.h" 3064 #ifdef DDB 3065 #include <sys/kernel.h> 3066 3067 #include <ddb/ddb.h> 3068 3069 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3070 { 3071 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3072 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3073 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3074 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3075 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3076 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3077 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3078 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3079 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3080 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3081 } 3082 3083 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3084 { 3085 int i; 3086 db_printf("PQ_FREE:"); 3087 for(i=0;i<PQ_L2_SIZE;i++) { 3088 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3089 } 3090 db_printf("\n"); 3091 3092 db_printf("PQ_CACHE:"); 3093 for(i=0;i<PQ_L2_SIZE;i++) { 3094 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3095 } 3096 db_printf("\n"); 3097 3098 db_printf("PQ_ACTIVE:"); 3099 for(i=0;i<PQ_L2_SIZE;i++) { 3100 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3101 } 3102 db_printf("\n"); 3103 3104 db_printf("PQ_INACTIVE:"); 3105 for(i=0;i<PQ_L2_SIZE;i++) { 3106 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3107 } 3108 db_printf("\n"); 3109 } 3110 #endif /* DDB */ 3111