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, "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, "vm_page_queue_init"); 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(). If the page is on the HOLD queue 888 * it was freed while held and must be moved back to the FREE queue. 889 */ 890 void 891 vm_page_unhold(vm_page_t m) 892 { 893 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 894 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 895 m, m->hold_count, m->queue - m->pc)); 896 vm_page_spin_lock(m); 897 atomic_add_int(&m->hold_count, -1); 898 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 899 _vm_page_queue_spin_lock(m); 900 _vm_page_rem_queue_spinlocked(m); 901 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 902 _vm_page_queue_spin_unlock(m); 903 } 904 vm_page_spin_unlock(m); 905 } 906 907 /* 908 * vm_page_getfake: 909 * 910 * Create a fictitious page with the specified physical address and 911 * memory attribute. The memory attribute is the only the machine- 912 * dependent aspect of a fictitious page that must be initialized. 913 */ 914 915 void 916 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 917 { 918 919 if ((m->flags & PG_FICTITIOUS) != 0) { 920 /* 921 * The page's memattr might have changed since the 922 * previous initialization. Update the pmap to the 923 * new memattr. 924 */ 925 goto memattr; 926 } 927 m->phys_addr = paddr; 928 m->queue = PQ_NONE; 929 /* Fictitious pages don't use "segind". */ 930 /* Fictitious pages don't use "order" or "pool". */ 931 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY; 932 m->wire_count = 1; 933 pmap_page_init(m); 934 memattr: 935 pmap_page_set_memattr(m, memattr); 936 } 937 938 /* 939 * Inserts the given vm_page into the object and object list. 940 * 941 * The pagetables are not updated but will presumably fault the page 942 * in if necessary, or if a kernel page the caller will at some point 943 * enter the page into the kernel's pmap. We are not allowed to block 944 * here so we *can't* do this anyway. 945 * 946 * This routine may not block. 947 * This routine must be called with the vm_object held. 948 * This routine must be called with a critical section held. 949 * 950 * This routine returns TRUE if the page was inserted into the object 951 * successfully, and FALSE if the page already exists in the object. 952 */ 953 int 954 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 955 { 956 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 957 if (m->object != NULL) 958 panic("vm_page_insert: already inserted"); 959 960 object->generation++; 961 962 /* 963 * Record the object/offset pair in this page and add the 964 * pv_list_count of the page to the object. 965 * 966 * The vm_page spin lock is required for interactions with the pmap. 967 */ 968 vm_page_spin_lock(m); 969 m->object = object; 970 m->pindex = pindex; 971 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 972 m->object = NULL; 973 m->pindex = 0; 974 vm_page_spin_unlock(m); 975 return FALSE; 976 } 977 ++object->resident_page_count; 978 ++mycpu->gd_vmtotal.t_rm; 979 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */ 980 vm_page_spin_unlock(m); 981 982 /* 983 * Since we are inserting a new and possibly dirty page, 984 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 985 */ 986 if ((m->valid & m->dirty) || 987 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 988 vm_object_set_writeable_dirty(object); 989 990 /* 991 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 992 */ 993 swap_pager_page_inserted(m); 994 return TRUE; 995 } 996 997 /* 998 * Removes the given vm_page_t from the (object,index) table 999 * 1000 * The underlying pmap entry (if any) is NOT removed here. 1001 * This routine may not block. 1002 * 1003 * The page must be BUSY and will remain BUSY on return. 1004 * No other requirements. 1005 * 1006 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1007 * it busy. 1008 */ 1009 void 1010 vm_page_remove(vm_page_t m) 1011 { 1012 vm_object_t object; 1013 1014 if (m->object == NULL) { 1015 return; 1016 } 1017 1018 if ((m->flags & PG_BUSY) == 0) 1019 panic("vm_page_remove: page not busy"); 1020 1021 object = m->object; 1022 1023 vm_object_hold(object); 1024 1025 /* 1026 * Remove the page from the object and update the object. 1027 * 1028 * The vm_page spin lock is required for interactions with the pmap. 1029 */ 1030 vm_page_spin_lock(m); 1031 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1032 --object->resident_page_count; 1033 --mycpu->gd_vmtotal.t_rm; 1034 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */ 1035 m->object = NULL; 1036 vm_page_spin_unlock(m); 1037 1038 object->generation++; 1039 1040 vm_object_drop(object); 1041 } 1042 1043 /* 1044 * Locate and return the page at (object, pindex), or NULL if the 1045 * page could not be found. 1046 * 1047 * The caller must hold the vm_object token. 1048 */ 1049 vm_page_t 1050 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1051 { 1052 vm_page_t m; 1053 1054 /* 1055 * Search the hash table for this object/offset pair 1056 */ 1057 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1058 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1059 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1060 return(m); 1061 } 1062 1063 vm_page_t 1064 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1065 vm_pindex_t pindex, 1066 int also_m_busy, const char *msg 1067 VM_PAGE_DEBUG_ARGS) 1068 { 1069 u_int32_t flags; 1070 vm_page_t m; 1071 1072 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1073 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1074 while (m) { 1075 KKASSERT(m->object == object && m->pindex == pindex); 1076 flags = m->flags; 1077 cpu_ccfence(); 1078 if (flags & PG_BUSY) { 1079 tsleep_interlock(m, 0); 1080 if (atomic_cmpset_int(&m->flags, flags, 1081 flags | PG_WANTED | PG_REFERENCED)) { 1082 tsleep(m, PINTERLOCKED, msg, 0); 1083 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1084 pindex); 1085 } 1086 } else if (also_m_busy && (flags & PG_SBUSY)) { 1087 tsleep_interlock(m, 0); 1088 if (atomic_cmpset_int(&m->flags, flags, 1089 flags | PG_WANTED | PG_REFERENCED)) { 1090 tsleep(m, PINTERLOCKED, msg, 0); 1091 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1092 pindex); 1093 } 1094 } else if (atomic_cmpset_int(&m->flags, flags, 1095 flags | PG_BUSY)) { 1096 #ifdef VM_PAGE_DEBUG 1097 m->busy_func = func; 1098 m->busy_line = lineno; 1099 #endif 1100 break; 1101 } 1102 } 1103 return m; 1104 } 1105 1106 /* 1107 * Attempt to lookup and busy a page. 1108 * 1109 * Returns NULL if the page could not be found 1110 * 1111 * Returns a vm_page and error == TRUE if the page exists but could not 1112 * be busied. 1113 * 1114 * Returns a vm_page and error == FALSE on success. 1115 */ 1116 vm_page_t 1117 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1118 vm_pindex_t pindex, 1119 int also_m_busy, int *errorp 1120 VM_PAGE_DEBUG_ARGS) 1121 { 1122 u_int32_t flags; 1123 vm_page_t m; 1124 1125 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1126 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1127 *errorp = FALSE; 1128 while (m) { 1129 KKASSERT(m->object == object && m->pindex == pindex); 1130 flags = m->flags; 1131 cpu_ccfence(); 1132 if (flags & PG_BUSY) { 1133 *errorp = TRUE; 1134 break; 1135 } 1136 if (also_m_busy && (flags & PG_SBUSY)) { 1137 *errorp = TRUE; 1138 break; 1139 } 1140 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1141 #ifdef VM_PAGE_DEBUG 1142 m->busy_func = func; 1143 m->busy_line = lineno; 1144 #endif 1145 break; 1146 } 1147 } 1148 return m; 1149 } 1150 1151 /* 1152 * Caller must hold the related vm_object 1153 */ 1154 vm_page_t 1155 vm_page_next(vm_page_t m) 1156 { 1157 vm_page_t next; 1158 1159 next = vm_page_rb_tree_RB_NEXT(m); 1160 if (next && next->pindex != m->pindex + 1) 1161 next = NULL; 1162 return (next); 1163 } 1164 1165 /* 1166 * vm_page_rename() 1167 * 1168 * Move the given vm_page from its current object to the specified 1169 * target object/offset. The page must be busy and will remain so 1170 * on return. 1171 * 1172 * new_object must be held. 1173 * This routine might block. XXX ? 1174 * 1175 * NOTE: Swap associated with the page must be invalidated by the move. We 1176 * have to do this for several reasons: (1) we aren't freeing the 1177 * page, (2) we are dirtying the page, (3) the VM system is probably 1178 * moving the page from object A to B, and will then later move 1179 * the backing store from A to B and we can't have a conflict. 1180 * 1181 * NOTE: We *always* dirty the page. It is necessary both for the 1182 * fact that we moved it, and because we may be invalidating 1183 * swap. If the page is on the cache, we have to deactivate it 1184 * or vm_page_dirty() will panic. Dirty pages are not allowed 1185 * on the cache. 1186 */ 1187 void 1188 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1189 { 1190 KKASSERT(m->flags & PG_BUSY); 1191 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1192 if (m->object) { 1193 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1194 vm_page_remove(m); 1195 } 1196 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1197 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1198 new_object, new_pindex); 1199 } 1200 if (m->queue - m->pc == PQ_CACHE) 1201 vm_page_deactivate(m); 1202 vm_page_dirty(m); 1203 } 1204 1205 /* 1206 * vm_page_unqueue() without any wakeup. This routine is used when a page 1207 * is to remain BUSYied by the 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 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1450 "pg %p q=%d flags=%08x hold=%d wire=%d", 1451 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1452 KKASSERT(m->wire_count == 0); 1453 vm_page_spin_unlock(m); 1454 pagedaemon_wakeup(); 1455 1456 /* return busied and removed page */ 1457 return(m); 1458 } 1459 } 1460 return(m); 1461 } 1462 1463 /* 1464 * This implements a per-cpu cache of free, zero'd, ready-to-go pages. 1465 * The idea is to populate this cache prior to acquiring any locks so 1466 * we don't wind up potentially zeroing VM pages (under heavy loads) while 1467 * holding potentialy contending locks. 1468 * 1469 * Note that we allocate the page uninserted into anything and use a pindex 1470 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these 1471 * allocations should wind up being uncontended. However, we still want 1472 * to rove across PQ_L2_SIZE. 1473 */ 1474 void 1475 vm_page_pcpu_cache(void) 1476 { 1477 #if 0 1478 globaldata_t gd = mycpu; 1479 vm_page_t m; 1480 1481 if (gd->gd_vmpg_count < GD_MINVMPG) { 1482 crit_enter_gd(gd); 1483 while (gd->gd_vmpg_count < GD_MAXVMPG) { 1484 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask, 1485 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1486 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO); 1487 if (gd->gd_vmpg_count < GD_MAXVMPG) { 1488 if ((m->flags & PG_ZERO) == 0) { 1489 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 1490 vm_page_flag_set(m, PG_ZERO); 1491 } 1492 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m; 1493 } else { 1494 vm_page_free(m); 1495 } 1496 } 1497 crit_exit_gd(gd); 1498 } 1499 #endif 1500 } 1501 1502 /* 1503 * vm_page_alloc() 1504 * 1505 * Allocate and return a memory cell associated with this VM object/offset 1506 * pair. If object is NULL an unassociated page will be allocated. 1507 * 1508 * The returned page will be busied and removed from its queues. This 1509 * routine can block and may return NULL if a race occurs and the page 1510 * is found to already exist at the specified (object, pindex). 1511 * 1512 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1513 * VM_ALLOC_QUICK like normal but cannot use cache 1514 * VM_ALLOC_SYSTEM greater free drain 1515 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1516 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1517 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1518 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1519 * (see vm_page_grab()) 1520 * VM_ALLOC_USE_GD ok to use per-gd cache 1521 * 1522 * The object must be held if not NULL 1523 * This routine may not block 1524 * 1525 * Additional special handling is required when called from an interrupt 1526 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1527 * in this case. 1528 */ 1529 vm_page_t 1530 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1531 { 1532 globaldata_t gd = mycpu; 1533 vm_object_t obj; 1534 vm_page_t m; 1535 u_short pg_color; 1536 1537 #if 0 1538 /* 1539 * Special per-cpu free VM page cache. The pages are pre-busied 1540 * and pre-zerod for us. 1541 */ 1542 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1543 crit_enter_gd(gd); 1544 if (gd->gd_vmpg_count) { 1545 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1546 crit_exit_gd(gd); 1547 goto done; 1548 } 1549 crit_exit_gd(gd); 1550 } 1551 #endif 1552 m = NULL; 1553 1554 /* 1555 * Cpu twist - cpu localization algorithm 1556 */ 1557 if (object) { 1558 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask) + 1559 (object->pg_color & ~ncpus_fit_mask); 1560 } else { 1561 pg_color = gd->gd_cpuid + (pindex & ~ncpus_fit_mask); 1562 } 1563 KKASSERT(page_req & 1564 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1565 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1566 1567 /* 1568 * Certain system threads (pageout daemon, buf_daemon's) are 1569 * allowed to eat deeper into the free page list. 1570 */ 1571 if (curthread->td_flags & TDF_SYSTHREAD) 1572 page_req |= VM_ALLOC_SYSTEM; 1573 1574 loop: 1575 if (vmstats.v_free_count > vmstats.v_free_reserved || 1576 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 1577 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 1578 vmstats.v_free_count > vmstats.v_interrupt_free_min) 1579 ) { 1580 /* 1581 * The free queue has sufficient free pages to take one out. 1582 */ 1583 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) 1584 m = vm_page_select_free(pg_color, TRUE); 1585 else 1586 m = vm_page_select_free(pg_color, FALSE); 1587 } else if (page_req & VM_ALLOC_NORMAL) { 1588 /* 1589 * Allocatable from the cache (non-interrupt only). On 1590 * success, we must free the page and try again, thus 1591 * ensuring that vmstats.v_*_free_min counters are replenished. 1592 */ 1593 #ifdef INVARIANTS 1594 if (curthread->td_preempted) { 1595 kprintf("vm_page_alloc(): warning, attempt to allocate" 1596 " cache page from preempting interrupt\n"); 1597 m = NULL; 1598 } else { 1599 m = vm_page_select_cache(pg_color); 1600 } 1601 #else 1602 m = vm_page_select_cache(pg_color); 1603 #endif 1604 /* 1605 * On success move the page into the free queue and loop. 1606 * 1607 * Only do this if we can safely acquire the vm_object lock, 1608 * because this is effectively a random page and the caller 1609 * might be holding the lock shared, we don't want to 1610 * deadlock. 1611 */ 1612 if (m != NULL) { 1613 KASSERT(m->dirty == 0, 1614 ("Found dirty cache page %p", m)); 1615 if ((obj = m->object) != NULL) { 1616 if (vm_object_hold_try(obj)) { 1617 vm_page_protect(m, VM_PROT_NONE); 1618 vm_page_free(m); 1619 /* m->object NULL here */ 1620 vm_object_drop(obj); 1621 } else { 1622 vm_page_deactivate(m); 1623 vm_page_wakeup(m); 1624 } 1625 } else { 1626 vm_page_protect(m, VM_PROT_NONE); 1627 vm_page_free(m); 1628 } 1629 goto loop; 1630 } 1631 1632 /* 1633 * On failure return NULL 1634 */ 1635 #if defined(DIAGNOSTIC) 1636 if (vmstats.v_cache_count > 0) 1637 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 1638 #endif 1639 vm_pageout_deficit++; 1640 pagedaemon_wakeup(); 1641 return (NULL); 1642 } else { 1643 /* 1644 * No pages available, wakeup the pageout daemon and give up. 1645 */ 1646 vm_pageout_deficit++; 1647 pagedaemon_wakeup(); 1648 return (NULL); 1649 } 1650 1651 /* 1652 * v_free_count can race so loop if we don't find the expected 1653 * page. 1654 */ 1655 if (m == NULL) 1656 goto loop; 1657 1658 /* 1659 * Good page found. The page has already been busied for us and 1660 * removed from its queues. 1661 */ 1662 KASSERT(m->dirty == 0, 1663 ("vm_page_alloc: free/cache page %p was dirty", m)); 1664 KKASSERT(m->queue == PQ_NONE); 1665 1666 #if 0 1667 done: 1668 #endif 1669 /* 1670 * Initialize the structure, inheriting some flags but clearing 1671 * all the rest. The page has already been busied for us. 1672 */ 1673 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY)); 1674 KKASSERT(m->wire_count == 0); 1675 KKASSERT(m->busy == 0); 1676 m->act_count = 0; 1677 m->valid = 0; 1678 1679 /* 1680 * Caller must be holding the object lock (asserted by 1681 * vm_page_insert()). 1682 * 1683 * NOTE: Inserting a page here does not insert it into any pmaps 1684 * (which could cause us to block allocating memory). 1685 * 1686 * NOTE: If no object an unassociated page is allocated, m->pindex 1687 * can be used by the caller for any purpose. 1688 */ 1689 if (object) { 1690 if (vm_page_insert(m, object, pindex) == FALSE) { 1691 vm_page_free(m); 1692 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1693 panic("PAGE RACE %p[%ld]/%p", 1694 object, (long)pindex, m); 1695 m = NULL; 1696 } 1697 } else { 1698 m->pindex = pindex; 1699 } 1700 1701 /* 1702 * Don't wakeup too often - wakeup the pageout daemon when 1703 * we would be nearly out of memory. 1704 */ 1705 pagedaemon_wakeup(); 1706 1707 /* 1708 * A PG_BUSY page is returned. 1709 */ 1710 return (m); 1711 } 1712 1713 /* 1714 * Returns number of pages available in our DMA memory reserve 1715 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1716 */ 1717 vm_size_t 1718 vm_contig_avail_pages(void) 1719 { 1720 alist_blk_t blk; 1721 alist_blk_t count; 1722 alist_blk_t bfree; 1723 spin_lock(&vm_contig_spin); 1724 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1725 spin_unlock(&vm_contig_spin); 1726 1727 return bfree; 1728 } 1729 1730 /* 1731 * Attempt to allocate contiguous physical memory with the specified 1732 * requirements. 1733 */ 1734 vm_page_t 1735 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1736 unsigned long alignment, unsigned long boundary, 1737 unsigned long size, vm_memattr_t memattr) 1738 { 1739 alist_blk_t blk; 1740 vm_page_t m; 1741 int i; 1742 1743 alignment >>= PAGE_SHIFT; 1744 if (alignment == 0) 1745 alignment = 1; 1746 boundary >>= PAGE_SHIFT; 1747 if (boundary == 0) 1748 boundary = 1; 1749 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1750 1751 spin_lock(&vm_contig_spin); 1752 blk = alist_alloc(&vm_contig_alist, 0, size); 1753 if (blk == ALIST_BLOCK_NONE) { 1754 spin_unlock(&vm_contig_spin); 1755 if (bootverbose) { 1756 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1757 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1758 } 1759 return(NULL); 1760 } 1761 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1762 alist_free(&vm_contig_alist, blk, size); 1763 spin_unlock(&vm_contig_spin); 1764 if (bootverbose) { 1765 kprintf("vm_page_alloc_contig: %ldk high " 1766 "%016jx failed\n", 1767 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1768 (intmax_t)high); 1769 } 1770 return(NULL); 1771 } 1772 spin_unlock(&vm_contig_spin); 1773 if (vm_contig_verbose) { 1774 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 1775 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 1776 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1777 } 1778 1779 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 1780 if (memattr != VM_MEMATTR_DEFAULT) 1781 for (i = 0;i < size;i++) 1782 pmap_page_set_memattr(&m[i], memattr); 1783 return m; 1784 } 1785 1786 /* 1787 * Free contiguously allocated pages. The pages will be wired but not busy. 1788 * When freeing to the alist we leave them wired and not busy. 1789 */ 1790 void 1791 vm_page_free_contig(vm_page_t m, unsigned long size) 1792 { 1793 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 1794 vm_pindex_t start = pa >> PAGE_SHIFT; 1795 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 1796 1797 if (vm_contig_verbose) { 1798 kprintf("vm_page_free_contig: %016jx/%ldk\n", 1799 (intmax_t)pa, size / 1024); 1800 } 1801 if (pa < vm_low_phys_reserved) { 1802 KKASSERT(pa + size <= vm_low_phys_reserved); 1803 spin_lock(&vm_contig_spin); 1804 alist_free(&vm_contig_alist, start, pages); 1805 spin_unlock(&vm_contig_spin); 1806 } else { 1807 while (pages) { 1808 vm_page_busy_wait(m, FALSE, "cpgfr"); 1809 vm_page_unwire(m, 0); 1810 vm_page_free(m); 1811 --pages; 1812 ++m; 1813 } 1814 1815 } 1816 } 1817 1818 1819 /* 1820 * Wait for sufficient free memory for nominal heavy memory use kernel 1821 * operations. 1822 * 1823 * WARNING! Be sure never to call this in any vm_pageout code path, which 1824 * will trivially deadlock the system. 1825 */ 1826 void 1827 vm_wait_nominal(void) 1828 { 1829 while (vm_page_count_min(0)) 1830 vm_wait(0); 1831 } 1832 1833 /* 1834 * Test if vm_wait_nominal() would block. 1835 */ 1836 int 1837 vm_test_nominal(void) 1838 { 1839 if (vm_page_count_min(0)) 1840 return(1); 1841 return(0); 1842 } 1843 1844 /* 1845 * Block until free pages are available for allocation, called in various 1846 * places before memory allocations. 1847 * 1848 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 1849 * more generous then that. 1850 */ 1851 void 1852 vm_wait(int timo) 1853 { 1854 /* 1855 * never wait forever 1856 */ 1857 if (timo == 0) 1858 timo = hz; 1859 lwkt_gettoken(&vm_token); 1860 1861 if (curthread == pagethread) { 1862 /* 1863 * The pageout daemon itself needs pages, this is bad. 1864 */ 1865 if (vm_page_count_min(0)) { 1866 vm_pageout_pages_needed = 1; 1867 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 1868 } 1869 } else { 1870 /* 1871 * Wakeup the pageout daemon if necessary and wait. 1872 * 1873 * Do not wait indefinitely for the target to be reached, 1874 * as load might prevent it from being reached any time soon. 1875 * But wait a little to try to slow down page allocations 1876 * and to give more important threads (the pagedaemon) 1877 * allocation priority. 1878 */ 1879 if (vm_page_count_target()) { 1880 if (vm_pages_needed == 0) { 1881 vm_pages_needed = 1; 1882 wakeup(&vm_pages_needed); 1883 } 1884 ++vm_pages_waiting; /* SMP race ok */ 1885 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 1886 } 1887 } 1888 lwkt_reltoken(&vm_token); 1889 } 1890 1891 /* 1892 * Block until free pages are available for allocation 1893 * 1894 * Called only from vm_fault so that processes page faulting can be 1895 * easily tracked. 1896 */ 1897 void 1898 vm_wait_pfault(void) 1899 { 1900 /* 1901 * Wakeup the pageout daemon if necessary and wait. 1902 * 1903 * Do not wait indefinitely for the target to be reached, 1904 * as load might prevent it from being reached any time soon. 1905 * But wait a little to try to slow down page allocations 1906 * and to give more important threads (the pagedaemon) 1907 * allocation priority. 1908 */ 1909 if (vm_page_count_min(0)) { 1910 lwkt_gettoken(&vm_token); 1911 while (vm_page_count_severe()) { 1912 if (vm_page_count_target()) { 1913 if (vm_pages_needed == 0) { 1914 vm_pages_needed = 1; 1915 wakeup(&vm_pages_needed); 1916 } 1917 ++vm_pages_waiting; /* SMP race ok */ 1918 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 1919 } 1920 } 1921 lwkt_reltoken(&vm_token); 1922 } 1923 } 1924 1925 /* 1926 * Put the specified page on the active list (if appropriate). Ensure 1927 * that act_count is at least ACT_INIT but do not otherwise mess with it. 1928 * 1929 * The caller should be holding the page busied ? XXX 1930 * This routine may not block. 1931 */ 1932 void 1933 vm_page_activate(vm_page_t m) 1934 { 1935 u_short oqueue; 1936 1937 vm_page_spin_lock(m); 1938 if (m->queue - m->pc != PQ_ACTIVE) { 1939 _vm_page_queue_spin_lock(m); 1940 oqueue = _vm_page_rem_queue_spinlocked(m); 1941 /* page is left spinlocked, queue is unlocked */ 1942 1943 if (oqueue == PQ_CACHE) 1944 mycpu->gd_cnt.v_reactivated++; 1945 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1946 if (m->act_count < ACT_INIT) 1947 m->act_count = ACT_INIT; 1948 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 1949 } 1950 _vm_page_and_queue_spin_unlock(m); 1951 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 1952 pagedaemon_wakeup(); 1953 } else { 1954 if (m->act_count < ACT_INIT) 1955 m->act_count = ACT_INIT; 1956 vm_page_spin_unlock(m); 1957 } 1958 } 1959 1960 /* 1961 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1962 * routine is called when a page has been added to the cache or free 1963 * queues. 1964 * 1965 * This routine may not block. 1966 */ 1967 static __inline void 1968 vm_page_free_wakeup(void) 1969 { 1970 /* 1971 * If the pageout daemon itself needs pages, then tell it that 1972 * there are some free. 1973 */ 1974 if (vm_pageout_pages_needed && 1975 vmstats.v_cache_count + vmstats.v_free_count >= 1976 vmstats.v_pageout_free_min 1977 ) { 1978 vm_pageout_pages_needed = 0; 1979 wakeup(&vm_pageout_pages_needed); 1980 } 1981 1982 /* 1983 * Wakeup processes that are waiting on memory. 1984 * 1985 * Generally speaking we want to wakeup stuck processes as soon as 1986 * possible. !vm_page_count_min(0) is the absolute minimum point 1987 * where we can do this. Wait a bit longer to reduce degenerate 1988 * re-blocking (vm_page_free_hysteresis). The target check is just 1989 * to make sure the min-check w/hysteresis does not exceed the 1990 * normal target. 1991 */ 1992 if (vm_pages_waiting) { 1993 if (!vm_page_count_min(vm_page_free_hysteresis) || 1994 !vm_page_count_target()) { 1995 vm_pages_waiting = 0; 1996 wakeup(&vmstats.v_free_count); 1997 ++mycpu->gd_cnt.v_ppwakeups; 1998 } 1999 #if 0 2000 if (!vm_page_count_target()) { 2001 /* 2002 * Plenty of pages are free, wakeup everyone. 2003 */ 2004 vm_pages_waiting = 0; 2005 wakeup(&vmstats.v_free_count); 2006 ++mycpu->gd_cnt.v_ppwakeups; 2007 } else if (!vm_page_count_min(0)) { 2008 /* 2009 * Some pages are free, wakeup someone. 2010 */ 2011 int wcount = vm_pages_waiting; 2012 if (wcount > 0) 2013 --wcount; 2014 vm_pages_waiting = wcount; 2015 wakeup_one(&vmstats.v_free_count); 2016 ++mycpu->gd_cnt.v_ppwakeups; 2017 } 2018 #endif 2019 } 2020 } 2021 2022 /* 2023 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2024 * it from its VM object. 2025 * 2026 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 2027 * return (the page will have been freed). 2028 */ 2029 void 2030 vm_page_free_toq(vm_page_t m) 2031 { 2032 mycpu->gd_cnt.v_tfree++; 2033 KKASSERT((m->flags & PG_MAPPED) == 0); 2034 KKASSERT(m->flags & PG_BUSY); 2035 2036 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 2037 kprintf("vm_page_free: pindex(%lu), busy(%d), " 2038 "PG_BUSY(%d), hold(%d)\n", 2039 (u_long)m->pindex, m->busy, 2040 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 2041 if ((m->queue - m->pc) == PQ_FREE) 2042 panic("vm_page_free: freeing free page"); 2043 else 2044 panic("vm_page_free: freeing busy page"); 2045 } 2046 2047 /* 2048 * Remove from object, spinlock the page and its queues and 2049 * remove from any queue. No queue spinlock will be held 2050 * after this section (because the page was removed from any 2051 * queue). 2052 */ 2053 vm_page_remove(m); 2054 vm_page_and_queue_spin_lock(m); 2055 _vm_page_rem_queue_spinlocked(m); 2056 2057 /* 2058 * No further management of fictitious pages occurs beyond object 2059 * and queue removal. 2060 */ 2061 if ((m->flags & PG_FICTITIOUS) != 0) { 2062 vm_page_spin_unlock(m); 2063 vm_page_wakeup(m); 2064 return; 2065 } 2066 2067 m->valid = 0; 2068 vm_page_undirty(m); 2069 2070 if (m->wire_count != 0) { 2071 if (m->wire_count > 1) { 2072 panic( 2073 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2074 m->wire_count, (long)m->pindex); 2075 } 2076 panic("vm_page_free: freeing wired page"); 2077 } 2078 2079 /* 2080 * Clear the UNMANAGED flag when freeing an unmanaged page. 2081 * Clear the NEED_COMMIT flag 2082 */ 2083 if (m->flags & PG_UNMANAGED) 2084 vm_page_flag_clear(m, PG_UNMANAGED); 2085 if (m->flags & PG_NEED_COMMIT) 2086 vm_page_flag_clear(m, PG_NEED_COMMIT); 2087 2088 if (m->hold_count != 0) { 2089 vm_page_flag_clear(m, PG_ZERO); 2090 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2091 } else { 2092 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 2093 } 2094 2095 /* 2096 * This sequence allows us to clear PG_BUSY while still holding 2097 * its spin lock, which reduces contention vs allocators. We 2098 * must not leave the queue locked or _vm_page_wakeup() may 2099 * deadlock. 2100 */ 2101 _vm_page_queue_spin_unlock(m); 2102 if (_vm_page_wakeup(m)) { 2103 vm_page_spin_unlock(m); 2104 wakeup(m); 2105 } else { 2106 vm_page_spin_unlock(m); 2107 } 2108 vm_page_free_wakeup(); 2109 } 2110 2111 /* 2112 * vm_page_free_fromq_fast() 2113 * 2114 * Remove a non-zero page from one of the free queues; the page is removed for 2115 * zeroing, so do not issue a wakeup. 2116 */ 2117 vm_page_t 2118 vm_page_free_fromq_fast(void) 2119 { 2120 static int qi; 2121 vm_page_t m; 2122 int i; 2123 2124 for (i = 0; i < PQ_L2_SIZE; ++i) { 2125 m = vm_page_list_find(PQ_FREE, qi, FALSE); 2126 /* page is returned spinlocked and removed from its queue */ 2127 if (m) { 2128 if (vm_page_busy_try(m, TRUE)) { 2129 /* 2130 * We were unable to busy the page, deactivate 2131 * it and loop. 2132 */ 2133 _vm_page_deactivate_locked(m, 0); 2134 vm_page_spin_unlock(m); 2135 } else if (m->flags & PG_ZERO) { 2136 /* 2137 * The page is already PG_ZERO, requeue it and loop 2138 */ 2139 _vm_page_add_queue_spinlocked(m, 2140 PQ_FREE + m->pc, 2141 0); 2142 vm_page_queue_spin_unlock(m); 2143 if (_vm_page_wakeup(m)) { 2144 vm_page_spin_unlock(m); 2145 wakeup(m); 2146 } else { 2147 vm_page_spin_unlock(m); 2148 } 2149 } else { 2150 /* 2151 * The page is not PG_ZERO'd so return it. 2152 */ 2153 KKASSERT((m->flags & (PG_UNMANAGED | 2154 PG_NEED_COMMIT)) == 0); 2155 KKASSERT(m->hold_count == 0); 2156 KKASSERT(m->wire_count == 0); 2157 vm_page_spin_unlock(m); 2158 break; 2159 } 2160 m = NULL; 2161 } 2162 qi = (qi + PQ_PRIME2) & PQ_L2_MASK; 2163 } 2164 return (m); 2165 } 2166 2167 /* 2168 * vm_page_unmanage() 2169 * 2170 * Prevent PV management from being done on the page. The page is 2171 * removed from the paging queues as if it were wired, and as a 2172 * consequence of no longer being managed the pageout daemon will not 2173 * touch it (since there is no way to locate the pte mappings for the 2174 * page). madvise() calls that mess with the pmap will also no longer 2175 * operate on the page. 2176 * 2177 * Beyond that the page is still reasonably 'normal'. Freeing the page 2178 * will clear the flag. 2179 * 2180 * This routine is used by OBJT_PHYS objects - objects using unswappable 2181 * physical memory as backing store rather then swap-backed memory and 2182 * will eventually be extended to support 4MB unmanaged physical 2183 * mappings. 2184 * 2185 * Caller must be holding the page busy. 2186 */ 2187 void 2188 vm_page_unmanage(vm_page_t m) 2189 { 2190 KKASSERT(m->flags & PG_BUSY); 2191 if ((m->flags & PG_UNMANAGED) == 0) { 2192 if (m->wire_count == 0) 2193 vm_page_unqueue(m); 2194 } 2195 vm_page_flag_set(m, PG_UNMANAGED); 2196 } 2197 2198 /* 2199 * Mark this page as wired down by yet another map, removing it from 2200 * paging queues as necessary. 2201 * 2202 * Caller must be holding the page busy. 2203 */ 2204 void 2205 vm_page_wire(vm_page_t m) 2206 { 2207 /* 2208 * Only bump the wire statistics if the page is not already wired, 2209 * and only unqueue the page if it is on some queue (if it is unmanaged 2210 * it is already off the queues). Don't do anything with fictitious 2211 * pages because they are always wired. 2212 */ 2213 KKASSERT(m->flags & PG_BUSY); 2214 if ((m->flags & PG_FICTITIOUS) == 0) { 2215 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2216 if ((m->flags & PG_UNMANAGED) == 0) 2217 vm_page_unqueue(m); 2218 atomic_add_int(&vmstats.v_wire_count, 1); 2219 } 2220 KASSERT(m->wire_count != 0, 2221 ("vm_page_wire: wire_count overflow m=%p", m)); 2222 } 2223 } 2224 2225 /* 2226 * Release one wiring of this page, potentially enabling it to be paged again. 2227 * 2228 * Many pages placed on the inactive queue should actually go 2229 * into the cache, but it is difficult to figure out which. What 2230 * we do instead, if the inactive target is well met, is to put 2231 * clean pages at the head of the inactive queue instead of the tail. 2232 * This will cause them to be moved to the cache more quickly and 2233 * if not actively re-referenced, freed more quickly. If we just 2234 * stick these pages at the end of the inactive queue, heavy filesystem 2235 * meta-data accesses can cause an unnecessary paging load on memory bound 2236 * processes. This optimization causes one-time-use metadata to be 2237 * reused more quickly. 2238 * 2239 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2240 * the inactive queue. This helps the pageout daemon determine memory 2241 * pressure and act on out-of-memory situations more quickly. 2242 * 2243 * BUT, if we are in a low-memory situation we have no choice but to 2244 * put clean pages on the cache queue. 2245 * 2246 * A number of routines use vm_page_unwire() to guarantee that the page 2247 * will go into either the inactive or active queues, and will NEVER 2248 * be placed in the cache - for example, just after dirtying a page. 2249 * dirty pages in the cache are not allowed. 2250 * 2251 * This routine may not block. 2252 */ 2253 void 2254 vm_page_unwire(vm_page_t m, int activate) 2255 { 2256 KKASSERT(m->flags & PG_BUSY); 2257 if (m->flags & PG_FICTITIOUS) { 2258 /* do nothing */ 2259 } else if (m->wire_count <= 0) { 2260 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2261 } else { 2262 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2263 atomic_add_int(&vmstats.v_wire_count, -1); 2264 if (m->flags & PG_UNMANAGED) { 2265 ; 2266 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2267 vm_page_spin_lock(m); 2268 _vm_page_add_queue_spinlocked(m, 2269 PQ_ACTIVE + m->pc, 0); 2270 _vm_page_and_queue_spin_unlock(m); 2271 } else { 2272 vm_page_spin_lock(m); 2273 vm_page_flag_clear(m, PG_WINATCFLS); 2274 _vm_page_add_queue_spinlocked(m, 2275 PQ_INACTIVE + m->pc, 0); 2276 ++vm_swapcache_inactive_heuristic; 2277 _vm_page_and_queue_spin_unlock(m); 2278 } 2279 } 2280 } 2281 } 2282 2283 /* 2284 * Move the specified page to the inactive queue. If the page has 2285 * any associated swap, the swap is deallocated. 2286 * 2287 * Normally athead is 0 resulting in LRU operation. athead is set 2288 * to 1 if we want this page to be 'as if it were placed in the cache', 2289 * except without unmapping it from the process address space. 2290 * 2291 * vm_page's spinlock must be held on entry and will remain held on return. 2292 * This routine may not block. 2293 */ 2294 static void 2295 _vm_page_deactivate_locked(vm_page_t m, int athead) 2296 { 2297 u_short oqueue; 2298 2299 /* 2300 * Ignore if already inactive. 2301 */ 2302 if (m->queue - m->pc == PQ_INACTIVE) 2303 return; 2304 _vm_page_queue_spin_lock(m); 2305 oqueue = _vm_page_rem_queue_spinlocked(m); 2306 2307 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2308 if (oqueue == PQ_CACHE) 2309 mycpu->gd_cnt.v_reactivated++; 2310 vm_page_flag_clear(m, PG_WINATCFLS); 2311 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2312 if (athead == 0) 2313 ++vm_swapcache_inactive_heuristic; 2314 } 2315 /* NOTE: PQ_NONE if condition not taken */ 2316 _vm_page_queue_spin_unlock(m); 2317 /* leaves vm_page spinlocked */ 2318 } 2319 2320 /* 2321 * Attempt to deactivate a page. 2322 * 2323 * No requirements. 2324 */ 2325 void 2326 vm_page_deactivate(vm_page_t m) 2327 { 2328 vm_page_spin_lock(m); 2329 _vm_page_deactivate_locked(m, 0); 2330 vm_page_spin_unlock(m); 2331 } 2332 2333 void 2334 vm_page_deactivate_locked(vm_page_t m) 2335 { 2336 _vm_page_deactivate_locked(m, 0); 2337 } 2338 2339 /* 2340 * Attempt to move a page to PQ_CACHE. 2341 * 2342 * Returns 0 on failure, 1 on success 2343 * 2344 * The page should NOT be busied by the caller. This function will validate 2345 * whether the page can be safely moved to the cache. 2346 */ 2347 int 2348 vm_page_try_to_cache(vm_page_t m) 2349 { 2350 vm_page_spin_lock(m); 2351 if (vm_page_busy_try(m, TRUE)) { 2352 vm_page_spin_unlock(m); 2353 return(0); 2354 } 2355 if (m->dirty || m->hold_count || m->wire_count || 2356 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2357 if (_vm_page_wakeup(m)) { 2358 vm_page_spin_unlock(m); 2359 wakeup(m); 2360 } else { 2361 vm_page_spin_unlock(m); 2362 } 2363 return(0); 2364 } 2365 vm_page_spin_unlock(m); 2366 2367 /* 2368 * Page busied by us and no longer spinlocked. Dirty pages cannot 2369 * be moved to the cache. 2370 */ 2371 vm_page_test_dirty(m); 2372 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2373 vm_page_wakeup(m); 2374 return(0); 2375 } 2376 vm_page_cache(m); 2377 return(1); 2378 } 2379 2380 /* 2381 * Attempt to free the page. If we cannot free it, we do nothing. 2382 * 1 is returned on success, 0 on failure. 2383 * 2384 * No requirements. 2385 */ 2386 int 2387 vm_page_try_to_free(vm_page_t m) 2388 { 2389 vm_page_spin_lock(m); 2390 if (vm_page_busy_try(m, TRUE)) { 2391 vm_page_spin_unlock(m); 2392 return(0); 2393 } 2394 2395 /* 2396 * The page can be in any state, including already being on the free 2397 * queue. Check to see if it really can be freed. 2398 */ 2399 if (m->dirty || /* can't free if it is dirty */ 2400 m->hold_count || /* or held (XXX may be wrong) */ 2401 m->wire_count || /* or wired */ 2402 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2403 PG_NEED_COMMIT)) || /* or needs a commit */ 2404 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2405 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2406 if (_vm_page_wakeup(m)) { 2407 vm_page_spin_unlock(m); 2408 wakeup(m); 2409 } else { 2410 vm_page_spin_unlock(m); 2411 } 2412 return(0); 2413 } 2414 vm_page_spin_unlock(m); 2415 2416 /* 2417 * We can probably free the page. 2418 * 2419 * Page busied by us and no longer spinlocked. Dirty pages will 2420 * not be freed by this function. We have to re-test the 2421 * dirty bit after cleaning out the pmaps. 2422 */ 2423 vm_page_test_dirty(m); 2424 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2425 vm_page_wakeup(m); 2426 return(0); 2427 } 2428 vm_page_protect(m, VM_PROT_NONE); 2429 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2430 vm_page_wakeup(m); 2431 return(0); 2432 } 2433 vm_page_free(m); 2434 return(1); 2435 } 2436 2437 /* 2438 * vm_page_cache 2439 * 2440 * Put the specified page onto the page cache queue (if appropriate). 2441 * 2442 * The page must be busy, and this routine will release the busy and 2443 * possibly even free the page. 2444 */ 2445 void 2446 vm_page_cache(vm_page_t m) 2447 { 2448 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2449 m->busy || m->wire_count || m->hold_count) { 2450 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 2451 vm_page_wakeup(m); 2452 return; 2453 } 2454 2455 /* 2456 * Already in the cache (and thus not mapped) 2457 */ 2458 if ((m->queue - m->pc) == PQ_CACHE) { 2459 KKASSERT((m->flags & PG_MAPPED) == 0); 2460 vm_page_wakeup(m); 2461 return; 2462 } 2463 2464 /* 2465 * Caller is required to test m->dirty, but note that the act of 2466 * removing the page from its maps can cause it to become dirty 2467 * on an SMP system due to another cpu running in usermode. 2468 */ 2469 if (m->dirty) { 2470 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2471 (long)m->pindex); 2472 } 2473 2474 /* 2475 * Remove all pmaps and indicate that the page is not 2476 * writeable or mapped. Our vm_page_protect() call may 2477 * have blocked (especially w/ VM_PROT_NONE), so recheck 2478 * everything. 2479 */ 2480 vm_page_protect(m, VM_PROT_NONE); 2481 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2482 m->busy || m->wire_count || m->hold_count) { 2483 vm_page_wakeup(m); 2484 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2485 vm_page_deactivate(m); 2486 vm_page_wakeup(m); 2487 } else { 2488 _vm_page_and_queue_spin_lock(m); 2489 _vm_page_rem_queue_spinlocked(m); 2490 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2491 _vm_page_queue_spin_unlock(m); 2492 if (_vm_page_wakeup(m)) { 2493 vm_page_spin_unlock(m); 2494 wakeup(m); 2495 } else { 2496 vm_page_spin_unlock(m); 2497 } 2498 vm_page_free_wakeup(); 2499 } 2500 } 2501 2502 /* 2503 * vm_page_dontneed() 2504 * 2505 * Cache, deactivate, or do nothing as appropriate. This routine 2506 * is typically used by madvise() MADV_DONTNEED. 2507 * 2508 * Generally speaking we want to move the page into the cache so 2509 * it gets reused quickly. However, this can result in a silly syndrome 2510 * due to the page recycling too quickly. Small objects will not be 2511 * fully cached. On the otherhand, if we move the page to the inactive 2512 * queue we wind up with a problem whereby very large objects 2513 * unnecessarily blow away our inactive and cache queues. 2514 * 2515 * The solution is to move the pages based on a fixed weighting. We 2516 * either leave them alone, deactivate them, or move them to the cache, 2517 * where moving them to the cache has the highest weighting. 2518 * By forcing some pages into other queues we eventually force the 2519 * system to balance the queues, potentially recovering other unrelated 2520 * space from active. The idea is to not force this to happen too 2521 * often. 2522 * 2523 * The page must be busied. 2524 */ 2525 void 2526 vm_page_dontneed(vm_page_t m) 2527 { 2528 static int dnweight; 2529 int dnw; 2530 int head; 2531 2532 dnw = ++dnweight; 2533 2534 /* 2535 * occassionally leave the page alone 2536 */ 2537 if ((dnw & 0x01F0) == 0 || 2538 m->queue - m->pc == PQ_INACTIVE || 2539 m->queue - m->pc == PQ_CACHE 2540 ) { 2541 if (m->act_count >= ACT_INIT) 2542 --m->act_count; 2543 return; 2544 } 2545 2546 /* 2547 * If vm_page_dontneed() is inactivating a page, it must clear 2548 * the referenced flag; otherwise the pagedaemon will see references 2549 * on the page in the inactive queue and reactivate it. Until the 2550 * page can move to the cache queue, madvise's job is not done. 2551 */ 2552 vm_page_flag_clear(m, PG_REFERENCED); 2553 pmap_clear_reference(m); 2554 2555 if (m->dirty == 0) 2556 vm_page_test_dirty(m); 2557 2558 if (m->dirty || (dnw & 0x0070) == 0) { 2559 /* 2560 * Deactivate the page 3 times out of 32. 2561 */ 2562 head = 0; 2563 } else { 2564 /* 2565 * Cache the page 28 times out of every 32. Note that 2566 * the page is deactivated instead of cached, but placed 2567 * at the head of the queue instead of the tail. 2568 */ 2569 head = 1; 2570 } 2571 vm_page_spin_lock(m); 2572 _vm_page_deactivate_locked(m, head); 2573 vm_page_spin_unlock(m); 2574 } 2575 2576 /* 2577 * These routines manipulate the 'soft busy' count for a page. A soft busy 2578 * is almost like PG_BUSY except that it allows certain compatible operations 2579 * to occur on the page while it is busy. For example, a page undergoing a 2580 * write can still be mapped read-only. 2581 * 2582 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2583 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2584 * busy bit is cleared. 2585 */ 2586 void 2587 vm_page_io_start(vm_page_t m) 2588 { 2589 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2590 atomic_add_char(&m->busy, 1); 2591 vm_page_flag_set(m, PG_SBUSY); 2592 } 2593 2594 void 2595 vm_page_io_finish(vm_page_t m) 2596 { 2597 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2598 atomic_subtract_char(&m->busy, 1); 2599 if (m->busy == 0) 2600 vm_page_flag_clear(m, PG_SBUSY); 2601 } 2602 2603 /* 2604 * Indicate that a clean VM page requires a filesystem commit and cannot 2605 * be reused. Used by tmpfs. 2606 */ 2607 void 2608 vm_page_need_commit(vm_page_t m) 2609 { 2610 vm_page_flag_set(m, PG_NEED_COMMIT); 2611 vm_object_set_writeable_dirty(m->object); 2612 } 2613 2614 void 2615 vm_page_clear_commit(vm_page_t m) 2616 { 2617 vm_page_flag_clear(m, PG_NEED_COMMIT); 2618 } 2619 2620 /* 2621 * Grab a page, blocking if it is busy and allocating a page if necessary. 2622 * A busy page is returned or NULL. The page may or may not be valid and 2623 * might not be on a queue (the caller is responsible for the disposition of 2624 * the page). 2625 * 2626 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2627 * page will be zero'd and marked valid. 2628 * 2629 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2630 * valid even if it already exists. 2631 * 2632 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2633 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2634 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2635 * 2636 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2637 * always returned if we had blocked. 2638 * 2639 * This routine may not be called from an interrupt. 2640 * 2641 * PG_ZERO is *ALWAYS* cleared by this routine. 2642 * 2643 * No other requirements. 2644 */ 2645 vm_page_t 2646 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2647 { 2648 vm_page_t m; 2649 int error; 2650 int shared = 1; 2651 2652 KKASSERT(allocflags & 2653 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2654 vm_object_hold_shared(object); 2655 for (;;) { 2656 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2657 if (error) { 2658 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2659 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2660 m = NULL; 2661 break; 2662 } 2663 /* retry */ 2664 } else if (m == NULL) { 2665 if (shared) { 2666 vm_object_upgrade(object); 2667 shared = 0; 2668 } 2669 if (allocflags & VM_ALLOC_RETRY) 2670 allocflags |= VM_ALLOC_NULL_OK; 2671 m = vm_page_alloc(object, pindex, 2672 allocflags & ~VM_ALLOC_RETRY); 2673 if (m) 2674 break; 2675 vm_wait(0); 2676 if ((allocflags & VM_ALLOC_RETRY) == 0) 2677 goto failed; 2678 } else { 2679 /* m found */ 2680 break; 2681 } 2682 } 2683 2684 /* 2685 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2686 * 2687 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2688 * valid even if already valid. 2689 */ 2690 if (m->valid == 0) { 2691 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2692 if ((m->flags & PG_ZERO) == 0) 2693 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2694 m->valid = VM_PAGE_BITS_ALL; 2695 } 2696 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2697 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2698 m->valid = VM_PAGE_BITS_ALL; 2699 } 2700 vm_page_flag_clear(m, PG_ZERO); 2701 failed: 2702 vm_object_drop(object); 2703 return(m); 2704 } 2705 2706 /* 2707 * Mapping function for valid bits or for dirty bits in 2708 * a page. May not block. 2709 * 2710 * Inputs are required to range within a page. 2711 * 2712 * No requirements. 2713 * Non blocking. 2714 */ 2715 int 2716 vm_page_bits(int base, int size) 2717 { 2718 int first_bit; 2719 int last_bit; 2720 2721 KASSERT( 2722 base + size <= PAGE_SIZE, 2723 ("vm_page_bits: illegal base/size %d/%d", base, size) 2724 ); 2725 2726 if (size == 0) /* handle degenerate case */ 2727 return(0); 2728 2729 first_bit = base >> DEV_BSHIFT; 2730 last_bit = (base + size - 1) >> DEV_BSHIFT; 2731 2732 return ((2 << last_bit) - (1 << first_bit)); 2733 } 2734 2735 /* 2736 * Sets portions of a page valid and clean. The arguments are expected 2737 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2738 * of any partial chunks touched by the range. The invalid portion of 2739 * such chunks will be zero'd. 2740 * 2741 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2742 * align base to DEV_BSIZE so as not to mark clean a partially 2743 * truncated device block. Otherwise the dirty page status might be 2744 * lost. 2745 * 2746 * This routine may not block. 2747 * 2748 * (base + size) must be less then or equal to PAGE_SIZE. 2749 */ 2750 static void 2751 _vm_page_zero_valid(vm_page_t m, int base, int size) 2752 { 2753 int frag; 2754 int endoff; 2755 2756 if (size == 0) /* handle degenerate case */ 2757 return; 2758 2759 /* 2760 * If the base is not DEV_BSIZE aligned and the valid 2761 * bit is clear, we have to zero out a portion of the 2762 * first block. 2763 */ 2764 2765 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2766 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2767 ) { 2768 pmap_zero_page_area( 2769 VM_PAGE_TO_PHYS(m), 2770 frag, 2771 base - frag 2772 ); 2773 } 2774 2775 /* 2776 * If the ending offset is not DEV_BSIZE aligned and the 2777 * valid bit is clear, we have to zero out a portion of 2778 * the last block. 2779 */ 2780 2781 endoff = base + size; 2782 2783 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2784 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 2785 ) { 2786 pmap_zero_page_area( 2787 VM_PAGE_TO_PHYS(m), 2788 endoff, 2789 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 2790 ); 2791 } 2792 } 2793 2794 /* 2795 * Set valid, clear dirty bits. If validating the entire 2796 * page we can safely clear the pmap modify bit. We also 2797 * use this opportunity to clear the PG_NOSYNC flag. If a process 2798 * takes a write fault on a MAP_NOSYNC memory area the flag will 2799 * be set again. 2800 * 2801 * We set valid bits inclusive of any overlap, but we can only 2802 * clear dirty bits for DEV_BSIZE chunks that are fully within 2803 * the range. 2804 * 2805 * Page must be busied? 2806 * No other requirements. 2807 */ 2808 void 2809 vm_page_set_valid(vm_page_t m, int base, int size) 2810 { 2811 _vm_page_zero_valid(m, base, size); 2812 m->valid |= vm_page_bits(base, size); 2813 } 2814 2815 2816 /* 2817 * Set valid bits and clear dirty bits. 2818 * 2819 * NOTE: This function does not clear the pmap modified bit. 2820 * Also note that e.g. NFS may use a byte-granular base 2821 * and size. 2822 * 2823 * WARNING: Page must be busied? But vfs_clean_one_page() will call 2824 * this without necessarily busying the page (via bdwrite()). 2825 * So for now vm_token must also be held. 2826 * 2827 * No other requirements. 2828 */ 2829 void 2830 vm_page_set_validclean(vm_page_t m, int base, int size) 2831 { 2832 int pagebits; 2833 2834 _vm_page_zero_valid(m, base, size); 2835 pagebits = vm_page_bits(base, size); 2836 m->valid |= pagebits; 2837 m->dirty &= ~pagebits; 2838 if (base == 0 && size == PAGE_SIZE) { 2839 /*pmap_clear_modify(m);*/ 2840 vm_page_flag_clear(m, PG_NOSYNC); 2841 } 2842 } 2843 2844 /* 2845 * Set valid & dirty. Used by buwrite() 2846 * 2847 * WARNING: Page must be busied? But vfs_dirty_one_page() will 2848 * call this function in buwrite() so for now vm_token must 2849 * be held. 2850 * 2851 * No other requirements. 2852 */ 2853 void 2854 vm_page_set_validdirty(vm_page_t m, int base, int size) 2855 { 2856 int pagebits; 2857 2858 pagebits = vm_page_bits(base, size); 2859 m->valid |= pagebits; 2860 m->dirty |= pagebits; 2861 if (m->object) 2862 vm_object_set_writeable_dirty(m->object); 2863 } 2864 2865 /* 2866 * Clear dirty bits. 2867 * 2868 * NOTE: This function does not clear the pmap modified bit. 2869 * Also note that e.g. NFS may use a byte-granular base 2870 * and size. 2871 * 2872 * Page must be busied? 2873 * No other requirements. 2874 */ 2875 void 2876 vm_page_clear_dirty(vm_page_t m, int base, int size) 2877 { 2878 m->dirty &= ~vm_page_bits(base, size); 2879 if (base == 0 && size == PAGE_SIZE) { 2880 /*pmap_clear_modify(m);*/ 2881 vm_page_flag_clear(m, PG_NOSYNC); 2882 } 2883 } 2884 2885 /* 2886 * Make the page all-dirty. 2887 * 2888 * Also make sure the related object and vnode reflect the fact that the 2889 * object may now contain a dirty page. 2890 * 2891 * Page must be busied? 2892 * No other requirements. 2893 */ 2894 void 2895 vm_page_dirty(vm_page_t m) 2896 { 2897 #ifdef INVARIANTS 2898 int pqtype = m->queue - m->pc; 2899 #endif 2900 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 2901 ("vm_page_dirty: page in free/cache queue!")); 2902 if (m->dirty != VM_PAGE_BITS_ALL) { 2903 m->dirty = VM_PAGE_BITS_ALL; 2904 if (m->object) 2905 vm_object_set_writeable_dirty(m->object); 2906 } 2907 } 2908 2909 /* 2910 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2911 * valid and dirty bits for the effected areas are cleared. 2912 * 2913 * Page must be busied? 2914 * Does not block. 2915 * No other requirements. 2916 */ 2917 void 2918 vm_page_set_invalid(vm_page_t m, int base, int size) 2919 { 2920 int bits; 2921 2922 bits = vm_page_bits(base, size); 2923 m->valid &= ~bits; 2924 m->dirty &= ~bits; 2925 m->object->generation++; 2926 } 2927 2928 /* 2929 * The kernel assumes that the invalid portions of a page contain 2930 * garbage, but such pages can be mapped into memory by user code. 2931 * When this occurs, we must zero out the non-valid portions of the 2932 * page so user code sees what it expects. 2933 * 2934 * Pages are most often semi-valid when the end of a file is mapped 2935 * into memory and the file's size is not page aligned. 2936 * 2937 * Page must be busied? 2938 * No other requirements. 2939 */ 2940 void 2941 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2942 { 2943 int b; 2944 int i; 2945 2946 /* 2947 * Scan the valid bits looking for invalid sections that 2948 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2949 * valid bit may be set ) have already been zerod by 2950 * vm_page_set_validclean(). 2951 */ 2952 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2953 if (i == (PAGE_SIZE / DEV_BSIZE) || 2954 (m->valid & (1 << i)) 2955 ) { 2956 if (i > b) { 2957 pmap_zero_page_area( 2958 VM_PAGE_TO_PHYS(m), 2959 b << DEV_BSHIFT, 2960 (i - b) << DEV_BSHIFT 2961 ); 2962 } 2963 b = i + 1; 2964 } 2965 } 2966 2967 /* 2968 * setvalid is TRUE when we can safely set the zero'd areas 2969 * as being valid. We can do this if there are no cache consistency 2970 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2971 */ 2972 if (setvalid) 2973 m->valid = VM_PAGE_BITS_ALL; 2974 } 2975 2976 /* 2977 * Is a (partial) page valid? Note that the case where size == 0 2978 * will return FALSE in the degenerate case where the page is entirely 2979 * invalid, and TRUE otherwise. 2980 * 2981 * Does not block. 2982 * No other requirements. 2983 */ 2984 int 2985 vm_page_is_valid(vm_page_t m, int base, int size) 2986 { 2987 int bits = vm_page_bits(base, size); 2988 2989 if (m->valid && ((m->valid & bits) == bits)) 2990 return 1; 2991 else 2992 return 0; 2993 } 2994 2995 /* 2996 * update dirty bits from pmap/mmu. May not block. 2997 * 2998 * Caller must hold the page busy 2999 */ 3000 void 3001 vm_page_test_dirty(vm_page_t m) 3002 { 3003 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3004 vm_page_dirty(m); 3005 } 3006 } 3007 3008 /* 3009 * Register an action, associating it with its vm_page 3010 */ 3011 void 3012 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 3013 { 3014 struct vm_page_action_list *list; 3015 int hv; 3016 3017 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3018 list = &action_list[hv]; 3019 3020 lwkt_gettoken(&vm_token); 3021 vm_page_flag_set(action->m, PG_ACTIONLIST); 3022 action->event = event; 3023 LIST_INSERT_HEAD(list, action, entry); 3024 lwkt_reltoken(&vm_token); 3025 } 3026 3027 /* 3028 * Unregister an action, disassociating it from its related vm_page 3029 */ 3030 void 3031 vm_page_unregister_action(vm_page_action_t action) 3032 { 3033 struct vm_page_action_list *list; 3034 int hv; 3035 3036 lwkt_gettoken(&vm_token); 3037 if (action->event != VMEVENT_NONE) { 3038 action->event = VMEVENT_NONE; 3039 LIST_REMOVE(action, entry); 3040 3041 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3042 list = &action_list[hv]; 3043 if (LIST_EMPTY(list)) 3044 vm_page_flag_clear(action->m, PG_ACTIONLIST); 3045 } 3046 lwkt_reltoken(&vm_token); 3047 } 3048 3049 /* 3050 * Issue an event on a VM page. Corresponding action structures are 3051 * removed from the page's list and called. 3052 * 3053 * If the vm_page has no more pending action events we clear its 3054 * PG_ACTIONLIST flag. 3055 */ 3056 void 3057 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 3058 { 3059 struct vm_page_action_list *list; 3060 struct vm_page_action *scan; 3061 struct vm_page_action *next; 3062 int hv; 3063 int all; 3064 3065 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 3066 list = &action_list[hv]; 3067 all = 1; 3068 3069 lwkt_gettoken(&vm_token); 3070 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 3071 if (scan->m == m) { 3072 if (scan->event == event) { 3073 scan->event = VMEVENT_NONE; 3074 LIST_REMOVE(scan, entry); 3075 scan->func(m, scan); 3076 /* XXX */ 3077 } else { 3078 all = 0; 3079 } 3080 } 3081 } 3082 if (all) 3083 vm_page_flag_clear(m, PG_ACTIONLIST); 3084 lwkt_reltoken(&vm_token); 3085 } 3086 3087 #include "opt_ddb.h" 3088 #ifdef DDB 3089 #include <sys/kernel.h> 3090 3091 #include <ddb/ddb.h> 3092 3093 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3094 { 3095 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3096 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3097 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3098 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3099 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3100 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3101 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3102 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3103 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3104 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3105 } 3106 3107 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3108 { 3109 int i; 3110 db_printf("PQ_FREE:"); 3111 for(i=0;i<PQ_L2_SIZE;i++) { 3112 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3113 } 3114 db_printf("\n"); 3115 3116 db_printf("PQ_CACHE:"); 3117 for(i=0;i<PQ_L2_SIZE;i++) { 3118 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3119 } 3120 db_printf("\n"); 3121 3122 db_printf("PQ_ACTIVE:"); 3123 for(i=0;i<PQ_L2_SIZE;i++) { 3124 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3125 } 3126 db_printf("\n"); 3127 3128 db_printf("PQ_INACTIVE:"); 3129 for(i=0;i<PQ_L2_SIZE;i++) { 3130 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3131 } 3132 db_printf("\n"); 3133 } 3134 #endif /* DDB */ 3135