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 #include <sys/cpu_topology.h> 79 80 #include <vm/vm.h> 81 #include <vm/vm_param.h> 82 #include <sys/lock.h> 83 #include <vm/vm_kern.h> 84 #include <vm/pmap.h> 85 #include <vm/vm_map.h> 86 #include <vm/vm_object.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_pageout.h> 89 #include <vm/vm_pager.h> 90 #include <vm/vm_extern.h> 91 #include <vm/swap_pager.h> 92 93 #include <machine/inttypes.h> 94 #include <machine/md_var.h> 95 #include <machine/specialreg.h> 96 97 #include <vm/vm_page2.h> 98 #include <sys/spinlock2.h> 99 100 /* 101 * Action hash for user umtx support. 102 */ 103 #define VMACTION_HSIZE 256 104 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 105 106 /* 107 * SET - Minimum required set associative size, must be a power of 2. We 108 * want this to match or exceed the set-associativeness of the cpu. 109 * 110 * GRP - A larger set that allows bleed-over into the domains of other 111 * nearby cpus. Also must be a power of 2. Used by the page zeroing 112 * code to smooth things out a bit. 113 */ 114 #define PQ_SET_ASSOC 16 115 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1) 116 117 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2) 118 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1) 119 120 static void vm_page_queue_init(void); 121 static void vm_page_free_wakeup(void); 122 static vm_page_t vm_page_select_cache(u_short pg_color); 123 static vm_page_t _vm_page_list_find2(int basequeue, int index); 124 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 125 126 /* 127 * Array of tailq lists 128 */ 129 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 130 131 LIST_HEAD(vm_page_action_list, vm_page_action); 132 struct vm_page_action_list action_list[VMACTION_HSIZE]; 133 static volatile int vm_pages_waiting; 134 135 static struct alist vm_contig_alist; 136 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 137 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 138 139 static u_long vm_dma_reserved = 0; 140 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 141 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 142 "Memory reserved for DMA"); 143 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 144 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 145 146 static int vm_contig_verbose = 0; 147 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 148 149 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 150 vm_pindex_t, pindex); 151 152 static void 153 vm_page_queue_init(void) 154 { 155 int i; 156 157 for (i = 0; i < PQ_L2_SIZE; i++) 158 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count; 159 for (i = 0; i < PQ_L2_SIZE; i++) 160 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count; 161 for (i = 0; i < PQ_L2_SIZE; i++) 162 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count; 163 for (i = 0; i < PQ_L2_SIZE; i++) 164 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count; 165 for (i = 0; i < PQ_L2_SIZE; i++) 166 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count; 167 /* PQ_NONE has no queue */ 168 169 for (i = 0; i < PQ_COUNT; i++) { 170 TAILQ_INIT(&vm_page_queues[i].pl); 171 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 172 } 173 174 for (i = 0; i < VMACTION_HSIZE; i++) 175 LIST_INIT(&action_list[i]); 176 } 177 178 /* 179 * note: place in initialized data section? Is this necessary? 180 */ 181 long first_page = 0; 182 int vm_page_array_size = 0; 183 vm_page_t vm_page_array = NULL; 184 vm_paddr_t vm_low_phys_reserved; 185 186 /* 187 * (low level boot) 188 * 189 * Sets the page size, perhaps based upon the memory size. 190 * Must be called before any use of page-size dependent functions. 191 */ 192 void 193 vm_set_page_size(void) 194 { 195 if (vmstats.v_page_size == 0) 196 vmstats.v_page_size = PAGE_SIZE; 197 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 198 panic("vm_set_page_size: page size not a power of two"); 199 } 200 201 /* 202 * (low level boot) 203 * 204 * Add a new page to the freelist for use by the system. New pages 205 * are added to both the head and tail of the associated free page 206 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 207 * requests pull 'recent' adds (higher physical addresses) first. 208 * 209 * Beware that the page zeroing daemon will also be running soon after 210 * boot, moving pages from the head to the tail of the PQ_FREE queues. 211 * 212 * Must be called in a critical section. 213 */ 214 static void 215 vm_add_new_page(vm_paddr_t pa) 216 { 217 struct vpgqueues *vpq; 218 vm_page_t m; 219 220 m = PHYS_TO_VM_PAGE(pa); 221 m->phys_addr = pa; 222 m->flags = 0; 223 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 224 m->pat_mode = PAT_WRITE_BACK; 225 /* 226 * Twist for cpu localization in addition to page coloring, so 227 * different cpus selecting by m->queue get different page colors. 228 */ 229 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK; 230 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK; 231 /* 232 * Reserve a certain number of contiguous low memory pages for 233 * contigmalloc() to use. 234 */ 235 if (pa < vm_low_phys_reserved) { 236 atomic_add_int(&vmstats.v_page_count, 1); 237 atomic_add_int(&vmstats.v_dma_pages, 1); 238 m->queue = PQ_NONE; 239 m->wire_count = 1; 240 atomic_add_int(&vmstats.v_wire_count, 1); 241 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 242 return; 243 } 244 245 /* 246 * General page 247 */ 248 m->queue = m->pc + PQ_FREE; 249 KKASSERT(m->dirty == 0); 250 251 atomic_add_int(&vmstats.v_page_count, 1); 252 atomic_add_int(&vmstats.v_free_count, 1); 253 vpq = &vm_page_queues[m->queue]; 254 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 255 ++vpq->lcnt; 256 } 257 258 /* 259 * (low level boot) 260 * 261 * Initializes the resident memory module. 262 * 263 * Preallocates memory for critical VM structures and arrays prior to 264 * kernel_map becoming available. 265 * 266 * Memory is allocated from (virtual2_start, virtual2_end) if available, 267 * otherwise memory is allocated from (virtual_start, virtual_end). 268 * 269 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 270 * large enough to hold vm_page_array & other structures for machines with 271 * large amounts of ram, so we want to use virtual2* when available. 272 */ 273 void 274 vm_page_startup(void) 275 { 276 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 277 vm_offset_t mapped; 278 vm_size_t npages; 279 vm_paddr_t page_range; 280 vm_paddr_t new_end; 281 int i; 282 vm_paddr_t pa; 283 int nblocks; 284 vm_paddr_t last_pa; 285 vm_paddr_t end; 286 vm_paddr_t biggestone, biggestsize; 287 vm_paddr_t total; 288 289 total = 0; 290 biggestsize = 0; 291 biggestone = 0; 292 nblocks = 0; 293 vaddr = round_page(vaddr); 294 295 for (i = 0; phys_avail[i + 1]; i += 2) { 296 phys_avail[i] = round_page64(phys_avail[i]); 297 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]); 298 } 299 300 for (i = 0; phys_avail[i + 1]; i += 2) { 301 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 302 303 if (size > biggestsize) { 304 biggestone = i; 305 biggestsize = size; 306 } 307 ++nblocks; 308 total += size; 309 } 310 311 end = phys_avail[biggestone+1]; 312 end = trunc_page(end); 313 314 /* 315 * Initialize the queue headers for the free queue, the active queue 316 * and the inactive queue. 317 */ 318 vm_page_queue_init(); 319 320 #if !defined(_KERNEL_VIRTUAL) 321 /* 322 * VKERNELs don't support minidumps and as such don't need 323 * vm_page_dump 324 * 325 * Allocate a bitmap to indicate that a random physical page 326 * needs to be included in a minidump. 327 * 328 * The amd64 port needs this to indicate which direct map pages 329 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 330 * 331 * However, i386 still needs this workspace internally within the 332 * minidump code. In theory, they are not needed on i386, but are 333 * included should the sf_buf code decide to use them. 334 */ 335 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 336 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 337 end -= vm_page_dump_size; 338 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 339 VM_PROT_READ | VM_PROT_WRITE); 340 bzero((void *)vm_page_dump, vm_page_dump_size); 341 #endif 342 /* 343 * Compute the number of pages of memory that will be available for 344 * use (taking into account the overhead of a page structure per 345 * page). 346 */ 347 first_page = phys_avail[0] / PAGE_SIZE; 348 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 349 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 350 351 #ifndef _KERNEL_VIRTUAL 352 /* 353 * (only applies to real kernels) 354 * 355 * Reserve a large amount of low memory for potential 32-bit DMA 356 * space allocations. Once device initialization is complete we 357 * release most of it, but keep (vm_dma_reserved) memory reserved 358 * for later use. Typically for X / graphics. Through trial and 359 * error we find that GPUs usually requires ~60-100MB or so. 360 * 361 * By default, 128M is left in reserve on machines with 2G+ of ram. 362 */ 363 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 364 if (vm_low_phys_reserved > total / 4) 365 vm_low_phys_reserved = total / 4; 366 if (vm_dma_reserved == 0) { 367 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 368 if (vm_dma_reserved > total / 16) 369 vm_dma_reserved = total / 16; 370 } 371 #endif 372 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 373 ALIST_RECORDS_65536); 374 375 /* 376 * Initialize the mem entry structures now, and put them in the free 377 * queue. 378 */ 379 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 380 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 381 vm_page_array = (vm_page_t)mapped; 382 383 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 384 /* 385 * since pmap_map on amd64 returns stuff out of a direct-map region, 386 * we have to manually add these pages to the minidump tracking so 387 * that they can be dumped, including the vm_page_array. 388 */ 389 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 390 dump_add_page(pa); 391 #endif 392 393 /* 394 * Clear all of the page structures 395 */ 396 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 397 vm_page_array_size = page_range; 398 399 /* 400 * Construct the free queue(s) in ascending order (by physical 401 * address) so that the first 16MB of physical memory is allocated 402 * last rather than first. On large-memory machines, this avoids 403 * the exhaustion of low physical memory before isa_dmainit has run. 404 */ 405 vmstats.v_page_count = 0; 406 vmstats.v_free_count = 0; 407 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 408 pa = phys_avail[i]; 409 if (i == biggestone) 410 last_pa = new_end; 411 else 412 last_pa = phys_avail[i + 1]; 413 while (pa < last_pa && npages-- > 0) { 414 vm_add_new_page(pa); 415 pa += PAGE_SIZE; 416 } 417 } 418 if (virtual2_start) 419 virtual2_start = vaddr; 420 else 421 virtual_start = vaddr; 422 } 423 424 /* 425 * We tended to reserve a ton of memory for contigmalloc(). Now that most 426 * drivers have initialized we want to return most the remaining free 427 * reserve back to the VM page queues so they can be used for normal 428 * allocations. 429 * 430 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 431 */ 432 static void 433 vm_page_startup_finish(void *dummy __unused) 434 { 435 alist_blk_t blk; 436 alist_blk_t rblk; 437 alist_blk_t count; 438 alist_blk_t xcount; 439 alist_blk_t bfree; 440 vm_page_t m; 441 442 spin_lock(&vm_contig_spin); 443 for (;;) { 444 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 445 if (bfree <= vm_dma_reserved / PAGE_SIZE) 446 break; 447 if (count == 0) 448 break; 449 450 /* 451 * Figure out how much of the initial reserve we have to 452 * free in order to reach our target. 453 */ 454 bfree -= vm_dma_reserved / PAGE_SIZE; 455 if (count > bfree) { 456 blk += count - bfree; 457 count = bfree; 458 } 459 460 /* 461 * Calculate the nearest power of 2 <= count. 462 */ 463 for (xcount = 1; xcount <= count; xcount <<= 1) 464 ; 465 xcount >>= 1; 466 blk += count - xcount; 467 count = xcount; 468 469 /* 470 * Allocate the pages from the alist, then free them to 471 * the normal VM page queues. 472 * 473 * Pages allocated from the alist are wired. We have to 474 * busy, unwire, and free them. We must also adjust 475 * vm_low_phys_reserved before freeing any pages to prevent 476 * confusion. 477 */ 478 rblk = alist_alloc(&vm_contig_alist, blk, count); 479 if (rblk != blk) { 480 kprintf("vm_page_startup_finish: Unable to return " 481 "dma space @0x%08x/%d -> 0x%08x\n", 482 blk, count, rblk); 483 break; 484 } 485 atomic_add_int(&vmstats.v_dma_pages, -count); 486 spin_unlock(&vm_contig_spin); 487 488 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 489 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 490 while (count) { 491 vm_page_busy_wait(m, FALSE, "cpgfr"); 492 vm_page_unwire(m, 0); 493 vm_page_free(m); 494 --count; 495 ++m; 496 } 497 spin_lock(&vm_contig_spin); 498 } 499 spin_unlock(&vm_contig_spin); 500 501 /* 502 * Print out how much DMA space drivers have already allocated and 503 * how much is left over. 504 */ 505 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 506 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 507 (PAGE_SIZE / 1024), 508 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 509 } 510 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 511 vm_page_startup_finish, NULL); 512 513 514 /* 515 * Scan comparison function for Red-Black tree scans. An inclusive 516 * (start,end) is expected. Other fields are not used. 517 */ 518 int 519 rb_vm_page_scancmp(struct vm_page *p, void *data) 520 { 521 struct rb_vm_page_scan_info *info = data; 522 523 if (p->pindex < info->start_pindex) 524 return(-1); 525 if (p->pindex > info->end_pindex) 526 return(1); 527 return(0); 528 } 529 530 int 531 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 532 { 533 if (p1->pindex < p2->pindex) 534 return(-1); 535 if (p1->pindex > p2->pindex) 536 return(1); 537 return(0); 538 } 539 540 void 541 vm_page_init(vm_page_t m) 542 { 543 /* do nothing for now. Called from pmap_page_init() */ 544 } 545 546 /* 547 * Each page queue has its own spin lock, which is fairly optimal for 548 * allocating and freeing pages at least. 549 * 550 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 551 * queue spinlock via this function. Also note that m->queue cannot change 552 * unless both the page and queue are locked. 553 */ 554 static __inline 555 void 556 _vm_page_queue_spin_lock(vm_page_t m) 557 { 558 u_short queue; 559 560 queue = m->queue; 561 if (queue != PQ_NONE) { 562 spin_lock(&vm_page_queues[queue].spin); 563 KKASSERT(queue == m->queue); 564 } 565 } 566 567 static __inline 568 void 569 _vm_page_queue_spin_unlock(vm_page_t m) 570 { 571 u_short queue; 572 573 queue = m->queue; 574 cpu_ccfence(); 575 if (queue != PQ_NONE) 576 spin_unlock(&vm_page_queues[queue].spin); 577 } 578 579 static __inline 580 void 581 _vm_page_queues_spin_lock(u_short queue) 582 { 583 cpu_ccfence(); 584 if (queue != PQ_NONE) 585 spin_lock(&vm_page_queues[queue].spin); 586 } 587 588 589 static __inline 590 void 591 _vm_page_queues_spin_unlock(u_short queue) 592 { 593 cpu_ccfence(); 594 if (queue != PQ_NONE) 595 spin_unlock(&vm_page_queues[queue].spin); 596 } 597 598 void 599 vm_page_queue_spin_lock(vm_page_t m) 600 { 601 _vm_page_queue_spin_lock(m); 602 } 603 604 void 605 vm_page_queues_spin_lock(u_short queue) 606 { 607 _vm_page_queues_spin_lock(queue); 608 } 609 610 void 611 vm_page_queue_spin_unlock(vm_page_t m) 612 { 613 _vm_page_queue_spin_unlock(m); 614 } 615 616 void 617 vm_page_queues_spin_unlock(u_short queue) 618 { 619 _vm_page_queues_spin_unlock(queue); 620 } 621 622 /* 623 * This locks the specified vm_page and its queue in the proper order 624 * (page first, then queue). The queue may change so the caller must 625 * recheck on return. 626 */ 627 static __inline 628 void 629 _vm_page_and_queue_spin_lock(vm_page_t m) 630 { 631 vm_page_spin_lock(m); 632 _vm_page_queue_spin_lock(m); 633 } 634 635 static __inline 636 void 637 _vm_page_and_queue_spin_unlock(vm_page_t m) 638 { 639 _vm_page_queues_spin_unlock(m->queue); 640 vm_page_spin_unlock(m); 641 } 642 643 void 644 vm_page_and_queue_spin_unlock(vm_page_t m) 645 { 646 _vm_page_and_queue_spin_unlock(m); 647 } 648 649 void 650 vm_page_and_queue_spin_lock(vm_page_t m) 651 { 652 _vm_page_and_queue_spin_lock(m); 653 } 654 655 /* 656 * Helper function removes vm_page from its current queue. 657 * Returns the base queue the page used to be on. 658 * 659 * The vm_page and the queue must be spinlocked. 660 * This function will unlock the queue but leave the page spinlocked. 661 */ 662 static __inline u_short 663 _vm_page_rem_queue_spinlocked(vm_page_t m) 664 { 665 struct vpgqueues *pq; 666 u_short queue; 667 u_short oqueue; 668 669 queue = m->queue; 670 if (queue != PQ_NONE) { 671 pq = &vm_page_queues[queue]; 672 TAILQ_REMOVE(&pq->pl, m, pageq); 673 atomic_add_int(pq->cnt, -1); 674 pq->lcnt--; 675 m->queue = PQ_NONE; 676 oqueue = queue; 677 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE) 678 queue -= m->pc; 679 vm_page_queues_spin_unlock(oqueue); /* intended */ 680 } 681 return queue; 682 } 683 684 /* 685 * Helper function places the vm_page on the specified queue. 686 * 687 * The vm_page must be spinlocked. 688 * This function will return with both the page and the queue locked. 689 */ 690 static __inline void 691 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 692 { 693 struct vpgqueues *pq; 694 695 KKASSERT(m->queue == PQ_NONE); 696 697 if (queue != PQ_NONE) { 698 vm_page_queues_spin_lock(queue); 699 pq = &vm_page_queues[queue]; 700 ++pq->lcnt; 701 atomic_add_int(pq->cnt, 1); 702 m->queue = queue; 703 704 /* 705 * PQ_FREE is always handled LIFO style to try to provide 706 * cache-hot pages to programs. 707 */ 708 if (queue - m->pc == PQ_FREE) { 709 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 710 } else if (athead) { 711 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 712 } else { 713 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 714 } 715 /* leave the queue spinlocked */ 716 } 717 } 718 719 /* 720 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 721 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we 722 * did not. Only one sleep call will be made before returning. 723 * 724 * This function does NOT busy the page and on return the page is not 725 * guaranteed to be available. 726 */ 727 void 728 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 729 { 730 u_int32_t flags; 731 732 for (;;) { 733 flags = m->flags; 734 cpu_ccfence(); 735 736 if ((flags & PG_BUSY) == 0 && 737 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) { 738 break; 739 } 740 tsleep_interlock(m, 0); 741 if (atomic_cmpset_int(&m->flags, flags, 742 flags | PG_WANTED | PG_REFERENCED)) { 743 tsleep(m, PINTERLOCKED, msg, 0); 744 break; 745 } 746 } 747 } 748 749 /* 750 * This calculates and returns a page color given an optional VM object and 751 * either a pindex or an iterator. We attempt to return a cpu-localized 752 * pg_color that is still roughly 16-way set-associative. The CPU topology 753 * is used if it was probed. 754 * 755 * The caller may use the returned value to index into e.g. PQ_FREE when 756 * allocating a page in order to nominally obtain pages that are hopefully 757 * already localized to the requesting cpu. This function is not able to 758 * provide any sort of guarantee of this, but does its best to improve 759 * hardware cache management performance. 760 * 761 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 762 */ 763 u_short 764 vm_get_pg_color(globaldata_t gd, vm_object_t object, vm_pindex_t pindex) 765 { 766 u_short pg_color; 767 int phys_id; 768 int core_id; 769 int object_pg_color; 770 771 phys_id = get_cpu_phys_id(gd->gd_cpuid); 772 core_id = get_cpu_core_id(gd->gd_cpuid); 773 object_pg_color = object ? object->pg_color : 0; 774 775 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 776 int grpsize = PQ_L2_SIZE / cpu_topology_phys_ids; 777 778 if (grpsize / cpu_topology_core_ids >= PQ_SET_ASSOC) { 779 /* 780 * Enough space for a full break-down. 781 */ 782 pg_color = phys_id * grpsize; 783 pg_color += core_id * grpsize / cpu_topology_core_ids; 784 pg_color += (pindex + object_pg_color) % 785 (grpsize / cpu_topology_core_ids); 786 } else { 787 /* 788 * Not enough space, split up by physical package, 789 * then split up by core id but only down to a 790 * 16-set. If all else fails, force a 16-set. 791 */ 792 pg_color = phys_id * grpsize; 793 if (grpsize > 16) { 794 pg_color += 16 * (core_id % (grpsize / 16)); 795 grpsize = 16; 796 } else { 797 grpsize = 16; 798 } 799 pg_color += (pindex + object_pg_color) % 800 grpsize; 801 } 802 } else { 803 /* 804 * Unknown topology, distribute things evenly. 805 */ 806 pg_color = gd->gd_cpuid * PQ_L2_SIZE / ncpus; 807 pg_color += pindex + object_pg_color; 808 } 809 return pg_color; 810 } 811 812 /* 813 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we 814 * also wait for m->busy to become 0 before setting PG_BUSY. 815 */ 816 void 817 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 818 int also_m_busy, const char *msg 819 VM_PAGE_DEBUG_ARGS) 820 { 821 u_int32_t flags; 822 823 for (;;) { 824 flags = m->flags; 825 cpu_ccfence(); 826 if (flags & PG_BUSY) { 827 tsleep_interlock(m, 0); 828 if (atomic_cmpset_int(&m->flags, flags, 829 flags | PG_WANTED | PG_REFERENCED)) { 830 tsleep(m, PINTERLOCKED, msg, 0); 831 } 832 } else if (also_m_busy && (flags & PG_SBUSY)) { 833 tsleep_interlock(m, 0); 834 if (atomic_cmpset_int(&m->flags, flags, 835 flags | PG_WANTED | PG_REFERENCED)) { 836 tsleep(m, PINTERLOCKED, msg, 0); 837 } 838 } else { 839 if (atomic_cmpset_int(&m->flags, flags, 840 flags | PG_BUSY)) { 841 #ifdef VM_PAGE_DEBUG 842 m->busy_func = func; 843 m->busy_line = lineno; 844 #endif 845 break; 846 } 847 } 848 } 849 } 850 851 /* 852 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy 853 * is also 0. 854 * 855 * Returns non-zero on failure. 856 */ 857 int 858 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 859 VM_PAGE_DEBUG_ARGS) 860 { 861 u_int32_t flags; 862 863 for (;;) { 864 flags = m->flags; 865 cpu_ccfence(); 866 if (flags & PG_BUSY) 867 return TRUE; 868 if (also_m_busy && (flags & PG_SBUSY)) 869 return TRUE; 870 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 871 #ifdef VM_PAGE_DEBUG 872 m->busy_func = func; 873 m->busy_line = lineno; 874 #endif 875 return FALSE; 876 } 877 } 878 } 879 880 /* 881 * Clear the PG_BUSY flag and return non-zero to indicate to the caller 882 * that a wakeup() should be performed. 883 * 884 * The vm_page must be spinlocked and will remain spinlocked on return. 885 * The related queue must NOT be spinlocked (which could deadlock us). 886 * 887 * (inline version) 888 */ 889 static __inline 890 int 891 _vm_page_wakeup(vm_page_t m) 892 { 893 u_int32_t flags; 894 895 for (;;) { 896 flags = m->flags; 897 cpu_ccfence(); 898 if (atomic_cmpset_int(&m->flags, flags, 899 flags & ~(PG_BUSY | PG_WANTED))) { 900 break; 901 } 902 } 903 return(flags & PG_WANTED); 904 } 905 906 /* 907 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This 908 * is typically the last call you make on a page before moving onto 909 * other things. 910 */ 911 void 912 vm_page_wakeup(vm_page_t m) 913 { 914 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 915 vm_page_spin_lock(m); 916 if (_vm_page_wakeup(m)) { 917 vm_page_spin_unlock(m); 918 wakeup(m); 919 } else { 920 vm_page_spin_unlock(m); 921 } 922 } 923 924 /* 925 * Holding a page keeps it from being reused. Other parts of the system 926 * can still disassociate the page from its current object and free it, or 927 * perform read or write I/O on it and/or otherwise manipulate the page, 928 * but if the page is held the VM system will leave the page and its data 929 * intact and not reuse the page for other purposes until the last hold 930 * reference is released. (see vm_page_wire() if you want to prevent the 931 * page from being disassociated from its object too). 932 * 933 * The caller must still validate the contents of the page and, if necessary, 934 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 935 * before manipulating the page. 936 * 937 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 938 */ 939 void 940 vm_page_hold(vm_page_t m) 941 { 942 vm_page_spin_lock(m); 943 atomic_add_int(&m->hold_count, 1); 944 if (m->queue - m->pc == PQ_FREE) { 945 _vm_page_queue_spin_lock(m); 946 _vm_page_rem_queue_spinlocked(m); 947 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 948 _vm_page_queue_spin_unlock(m); 949 } 950 vm_page_spin_unlock(m); 951 } 952 953 /* 954 * The opposite of vm_page_hold(). If the page is on the HOLD queue 955 * it was freed while held and must be moved back to the FREE queue. 956 */ 957 void 958 vm_page_unhold(vm_page_t m) 959 { 960 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 961 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 962 m, m->hold_count, m->queue - m->pc)); 963 vm_page_spin_lock(m); 964 atomic_add_int(&m->hold_count, -1); 965 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 966 _vm_page_queue_spin_lock(m); 967 _vm_page_rem_queue_spinlocked(m); 968 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 969 _vm_page_queue_spin_unlock(m); 970 } 971 vm_page_spin_unlock(m); 972 } 973 974 /* 975 * vm_page_getfake: 976 * 977 * Create a fictitious page with the specified physical address and 978 * memory attribute. The memory attribute is the only the machine- 979 * dependent aspect of a fictitious page that must be initialized. 980 */ 981 982 void 983 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 984 { 985 986 if ((m->flags & PG_FICTITIOUS) != 0) { 987 /* 988 * The page's memattr might have changed since the 989 * previous initialization. Update the pmap to the 990 * new memattr. 991 */ 992 goto memattr; 993 } 994 m->phys_addr = paddr; 995 m->queue = PQ_NONE; 996 /* Fictitious pages don't use "segind". */ 997 /* Fictitious pages don't use "order" or "pool". */ 998 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY; 999 m->wire_count = 1; 1000 pmap_page_init(m); 1001 memattr: 1002 pmap_page_set_memattr(m, memattr); 1003 } 1004 1005 /* 1006 * Inserts the given vm_page into the object and object list. 1007 * 1008 * The pagetables are not updated but will presumably fault the page 1009 * in if necessary, or if a kernel page the caller will at some point 1010 * enter the page into the kernel's pmap. We are not allowed to block 1011 * here so we *can't* do this anyway. 1012 * 1013 * This routine may not block. 1014 * This routine must be called with the vm_object held. 1015 * This routine must be called with a critical section held. 1016 * 1017 * This routine returns TRUE if the page was inserted into the object 1018 * successfully, and FALSE if the page already exists in the object. 1019 */ 1020 int 1021 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1022 { 1023 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1024 if (m->object != NULL) 1025 panic("vm_page_insert: already inserted"); 1026 1027 object->generation++; 1028 1029 /* 1030 * Record the object/offset pair in this page and add the 1031 * pv_list_count of the page to the object. 1032 * 1033 * The vm_page spin lock is required for interactions with the pmap. 1034 */ 1035 vm_page_spin_lock(m); 1036 m->object = object; 1037 m->pindex = pindex; 1038 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1039 m->object = NULL; 1040 m->pindex = 0; 1041 vm_page_spin_unlock(m); 1042 return FALSE; 1043 } 1044 ++object->resident_page_count; 1045 ++mycpu->gd_vmtotal.t_rm; 1046 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */ 1047 vm_page_spin_unlock(m); 1048 1049 /* 1050 * Since we are inserting a new and possibly dirty page, 1051 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1052 */ 1053 if ((m->valid & m->dirty) || 1054 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1055 vm_object_set_writeable_dirty(object); 1056 1057 /* 1058 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1059 */ 1060 swap_pager_page_inserted(m); 1061 return TRUE; 1062 } 1063 1064 /* 1065 * Removes the given vm_page_t from the (object,index) table 1066 * 1067 * The underlying pmap entry (if any) is NOT removed here. 1068 * This routine may not block. 1069 * 1070 * The page must be BUSY and will remain BUSY on return. 1071 * No other requirements. 1072 * 1073 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1074 * it busy. 1075 */ 1076 void 1077 vm_page_remove(vm_page_t m) 1078 { 1079 vm_object_t object; 1080 1081 if (m->object == NULL) { 1082 return; 1083 } 1084 1085 if ((m->flags & PG_BUSY) == 0) 1086 panic("vm_page_remove: page not busy"); 1087 1088 object = m->object; 1089 1090 vm_object_hold(object); 1091 1092 /* 1093 * Remove the page from the object and update the object. 1094 * 1095 * The vm_page spin lock is required for interactions with the pmap. 1096 */ 1097 vm_page_spin_lock(m); 1098 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1099 --object->resident_page_count; 1100 --mycpu->gd_vmtotal.t_rm; 1101 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */ 1102 m->object = NULL; 1103 vm_page_spin_unlock(m); 1104 1105 object->generation++; 1106 1107 vm_object_drop(object); 1108 } 1109 1110 /* 1111 * Locate and return the page at (object, pindex), or NULL if the 1112 * page could not be found. 1113 * 1114 * The caller must hold the vm_object token. 1115 */ 1116 vm_page_t 1117 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1118 { 1119 vm_page_t m; 1120 1121 /* 1122 * Search the hash table for this object/offset pair 1123 */ 1124 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1125 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1126 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1127 return(m); 1128 } 1129 1130 vm_page_t 1131 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1132 vm_pindex_t pindex, 1133 int also_m_busy, const char *msg 1134 VM_PAGE_DEBUG_ARGS) 1135 { 1136 u_int32_t flags; 1137 vm_page_t m; 1138 1139 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1140 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1141 while (m) { 1142 KKASSERT(m->object == object && m->pindex == pindex); 1143 flags = m->flags; 1144 cpu_ccfence(); 1145 if (flags & PG_BUSY) { 1146 tsleep_interlock(m, 0); 1147 if (atomic_cmpset_int(&m->flags, flags, 1148 flags | PG_WANTED | PG_REFERENCED)) { 1149 tsleep(m, PINTERLOCKED, msg, 0); 1150 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1151 pindex); 1152 } 1153 } else if (also_m_busy && (flags & PG_SBUSY)) { 1154 tsleep_interlock(m, 0); 1155 if (atomic_cmpset_int(&m->flags, flags, 1156 flags | PG_WANTED | PG_REFERENCED)) { 1157 tsleep(m, PINTERLOCKED, msg, 0); 1158 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1159 pindex); 1160 } 1161 } else if (atomic_cmpset_int(&m->flags, flags, 1162 flags | PG_BUSY)) { 1163 #ifdef VM_PAGE_DEBUG 1164 m->busy_func = func; 1165 m->busy_line = lineno; 1166 #endif 1167 break; 1168 } 1169 } 1170 return m; 1171 } 1172 1173 /* 1174 * Attempt to lookup and busy a page. 1175 * 1176 * Returns NULL if the page could not be found 1177 * 1178 * Returns a vm_page and error == TRUE if the page exists but could not 1179 * be busied. 1180 * 1181 * Returns a vm_page and error == FALSE on success. 1182 */ 1183 vm_page_t 1184 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1185 vm_pindex_t pindex, 1186 int also_m_busy, int *errorp 1187 VM_PAGE_DEBUG_ARGS) 1188 { 1189 u_int32_t flags; 1190 vm_page_t m; 1191 1192 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1193 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1194 *errorp = FALSE; 1195 while (m) { 1196 KKASSERT(m->object == object && m->pindex == pindex); 1197 flags = m->flags; 1198 cpu_ccfence(); 1199 if (flags & PG_BUSY) { 1200 *errorp = TRUE; 1201 break; 1202 } 1203 if (also_m_busy && (flags & PG_SBUSY)) { 1204 *errorp = TRUE; 1205 break; 1206 } 1207 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1208 #ifdef VM_PAGE_DEBUG 1209 m->busy_func = func; 1210 m->busy_line = lineno; 1211 #endif 1212 break; 1213 } 1214 } 1215 return m; 1216 } 1217 1218 /* 1219 * Attempt to repurpose the passed-in page. If the passed-in page cannot 1220 * be repurposed it will be released, *must_reenter will be set to 1, and 1221 * this function will fall-through to vm_page_lookup_busy_try(). 1222 * 1223 * The passed-in page must be wired and not busy. The returned page will 1224 * be busied and not wired. 1225 * 1226 * A different page may be returned. The returned page will be busied and 1227 * not wired. 1228 * 1229 * NULL can be returned. If so, the required page could not be busied. 1230 * The passed-in page will be unwired. 1231 */ 1232 vm_page_t 1233 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex, 1234 int also_m_busy, int *errorp, vm_page_t m, 1235 int *must_reenter, int *iswired) 1236 { 1237 if (m) { 1238 /* 1239 * Do not mess with pages in a complex state, such as pages 1240 * which are mapped, as repurposing such pages can be more 1241 * expensive than simply allocatin a new one. 1242 * 1243 * NOTE: Soft-busying can deadlock against putpages or I/O 1244 * so we only allow hard-busying here. 1245 */ 1246 KKASSERT(also_m_busy == FALSE); 1247 vm_page_busy_wait(m, also_m_busy, "biodep"); 1248 1249 if ((m->flags & (PG_UNMANAGED | PG_MAPPED | 1250 PG_FICTITIOUS | PG_SBUSY)) || 1251 m->busy || m->wire_count != 1 || m->hold_count) { 1252 vm_page_unwire(m, 0); 1253 vm_page_wakeup(m); 1254 /* fall through to normal lookup */ 1255 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 1256 vm_page_unwire(m, 0); 1257 vm_page_deactivate(m); 1258 vm_page_wakeup(m); 1259 /* fall through to normal lookup */ 1260 } else { 1261 /* 1262 * We can safely repurpose the page. It should 1263 * already be unqueued. 1264 */ 1265 KKASSERT(m->queue == PQ_NONE && m->dirty == 0); 1266 vm_page_remove(m); 1267 m->valid = 0; 1268 m->act_count = 0; 1269 if (vm_page_insert(m, object, pindex)) { 1270 *errorp = 0; 1271 *iswired = 1; 1272 1273 return m; 1274 } 1275 vm_page_unwire(m, 0); 1276 vm_page_free(m); 1277 /* fall through to normal lookup */ 1278 } 1279 } 1280 1281 /* 1282 * Cannot repurpose page, attempt to locate the desired page. May 1283 * return NULL. 1284 */ 1285 *must_reenter = 1; 1286 *iswired = 0; 1287 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp); 1288 1289 return m; 1290 } 1291 1292 /* 1293 * Caller must hold the related vm_object 1294 */ 1295 vm_page_t 1296 vm_page_next(vm_page_t m) 1297 { 1298 vm_page_t next; 1299 1300 next = vm_page_rb_tree_RB_NEXT(m); 1301 if (next && next->pindex != m->pindex + 1) 1302 next = NULL; 1303 return (next); 1304 } 1305 1306 /* 1307 * vm_page_rename() 1308 * 1309 * Move the given vm_page from its current object to the specified 1310 * target object/offset. The page must be busy and will remain so 1311 * on return. 1312 * 1313 * new_object must be held. 1314 * This routine might block. XXX ? 1315 * 1316 * NOTE: Swap associated with the page must be invalidated by the move. We 1317 * have to do this for several reasons: (1) we aren't freeing the 1318 * page, (2) we are dirtying the page, (3) the VM system is probably 1319 * moving the page from object A to B, and will then later move 1320 * the backing store from A to B and we can't have a conflict. 1321 * 1322 * NOTE: We *always* dirty the page. It is necessary both for the 1323 * fact that we moved it, and because we may be invalidating 1324 * swap. If the page is on the cache, we have to deactivate it 1325 * or vm_page_dirty() will panic. Dirty pages are not allowed 1326 * on the cache. 1327 */ 1328 void 1329 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1330 { 1331 KKASSERT(m->flags & PG_BUSY); 1332 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1333 if (m->object) { 1334 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1335 vm_page_remove(m); 1336 } 1337 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1338 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1339 new_object, new_pindex); 1340 } 1341 if (m->queue - m->pc == PQ_CACHE) 1342 vm_page_deactivate(m); 1343 vm_page_dirty(m); 1344 } 1345 1346 /* 1347 * vm_page_unqueue() without any wakeup. This routine is used when a page 1348 * is to remain BUSYied by the caller. 1349 * 1350 * This routine may not block. 1351 */ 1352 void 1353 vm_page_unqueue_nowakeup(vm_page_t m) 1354 { 1355 vm_page_and_queue_spin_lock(m); 1356 (void)_vm_page_rem_queue_spinlocked(m); 1357 vm_page_spin_unlock(m); 1358 } 1359 1360 /* 1361 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1362 * if necessary. 1363 * 1364 * This routine may not block. 1365 */ 1366 void 1367 vm_page_unqueue(vm_page_t m) 1368 { 1369 u_short queue; 1370 1371 vm_page_and_queue_spin_lock(m); 1372 queue = _vm_page_rem_queue_spinlocked(m); 1373 if (queue == PQ_FREE || queue == PQ_CACHE) { 1374 vm_page_spin_unlock(m); 1375 pagedaemon_wakeup(); 1376 } else { 1377 vm_page_spin_unlock(m); 1378 } 1379 } 1380 1381 /* 1382 * vm_page_list_find() 1383 * 1384 * Find a page on the specified queue with color optimization. 1385 * 1386 * The page coloring optimization attempts to locate a page that does 1387 * not overload other nearby pages in the object in the cpu's L1 or L2 1388 * caches. We need this optimization because cpu caches tend to be 1389 * physical caches, while object spaces tend to be virtual. 1390 * 1391 * The page coloring optimization also, very importantly, tries to localize 1392 * memory to cpus and physical sockets. 1393 * 1394 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1395 * and the algorithm is adjusted to localize allocations on a per-core basis. 1396 * This is done by 'twisting' the colors. 1397 * 1398 * The page is returned spinlocked and removed from its queue (it will 1399 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller 1400 * is responsible for dealing with the busy-page case (usually by 1401 * deactivating the page and looping). 1402 * 1403 * NOTE: This routine is carefully inlined. A non-inlined version 1404 * is available for outside callers but the only critical path is 1405 * from within this source file. 1406 * 1407 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1408 * represent stable storage, allowing us to order our locks vm_page 1409 * first, then queue. 1410 */ 1411 static __inline 1412 vm_page_t 1413 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1414 { 1415 vm_page_t m; 1416 1417 for (;;) { 1418 if (prefer_zero) { 1419 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, 1420 pglist); 1421 } else { 1422 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1423 } 1424 if (m == NULL) { 1425 m = _vm_page_list_find2(basequeue, index); 1426 return(m); 1427 } 1428 vm_page_and_queue_spin_lock(m); 1429 if (m->queue == basequeue + index) { 1430 _vm_page_rem_queue_spinlocked(m); 1431 /* vm_page_t spin held, no queue spin */ 1432 break; 1433 } 1434 vm_page_and_queue_spin_unlock(m); 1435 } 1436 return(m); 1437 } 1438 1439 /* 1440 * If we could not find the page in the desired queue try to find it in 1441 * a nearby queue. 1442 */ 1443 static vm_page_t 1444 _vm_page_list_find2(int basequeue, int index) 1445 { 1446 struct vpgqueues *pq; 1447 vm_page_t m = NULL; 1448 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1449 int pqi; 1450 int i; 1451 1452 index &= PQ_L2_MASK; 1453 pq = &vm_page_queues[basequeue]; 1454 1455 /* 1456 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1457 * else fails (PQ_L2_MASK which is 255). 1458 */ 1459 do { 1460 pqmask = (pqmask << 1) | 1; 1461 for (i = 0; i <= pqmask; ++i) { 1462 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1463 m = TAILQ_FIRST(&pq[pqi].pl); 1464 if (m) { 1465 _vm_page_and_queue_spin_lock(m); 1466 if (m->queue == basequeue + pqi) { 1467 _vm_page_rem_queue_spinlocked(m); 1468 return(m); 1469 } 1470 _vm_page_and_queue_spin_unlock(m); 1471 --i; 1472 continue; 1473 } 1474 } 1475 } while (pqmask != PQ_L2_MASK); 1476 1477 return(m); 1478 } 1479 1480 /* 1481 * Returns a vm_page candidate for allocation. The page is not busied so 1482 * it can move around. The caller must busy the page (and typically 1483 * deactivate it if it cannot be busied!) 1484 * 1485 * Returns a spinlocked vm_page that has been removed from its queue. 1486 */ 1487 vm_page_t 1488 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 1489 { 1490 return(_vm_page_list_find(basequeue, index, prefer_zero)); 1491 } 1492 1493 /* 1494 * Find a page on the cache queue with color optimization, remove it 1495 * from the queue, and busy it. The returned page will not be spinlocked. 1496 * 1497 * A candidate failure will be deactivated. Candidates can fail due to 1498 * being busied by someone else, in which case they will be deactivated. 1499 * 1500 * This routine may not block. 1501 * 1502 */ 1503 static vm_page_t 1504 vm_page_select_cache(u_short pg_color) 1505 { 1506 vm_page_t m; 1507 1508 for (;;) { 1509 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE); 1510 if (m == NULL) 1511 break; 1512 /* 1513 * (m) has been removed from its queue and spinlocked 1514 */ 1515 if (vm_page_busy_try(m, TRUE)) { 1516 _vm_page_deactivate_locked(m, 0); 1517 vm_page_spin_unlock(m); 1518 } else { 1519 /* 1520 * We successfully busied the page 1521 */ 1522 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1523 m->hold_count == 0 && 1524 m->wire_count == 0 && 1525 (m->dirty & m->valid) == 0) { 1526 vm_page_spin_unlock(m); 1527 pagedaemon_wakeup(); 1528 return(m); 1529 } 1530 1531 /* 1532 * The page cannot be recycled, deactivate it. 1533 */ 1534 _vm_page_deactivate_locked(m, 0); 1535 if (_vm_page_wakeup(m)) { 1536 vm_page_spin_unlock(m); 1537 wakeup(m); 1538 } else { 1539 vm_page_spin_unlock(m); 1540 } 1541 } 1542 } 1543 return (m); 1544 } 1545 1546 /* 1547 * Find a free or zero page, with specified preference. We attempt to 1548 * inline the nominal case and fall back to _vm_page_select_free() 1549 * otherwise. A busied page is removed from the queue and returned. 1550 * 1551 * This routine may not block. 1552 */ 1553 static __inline vm_page_t 1554 vm_page_select_free(u_short pg_color, boolean_t prefer_zero) 1555 { 1556 vm_page_t m; 1557 1558 for (;;) { 1559 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK, 1560 prefer_zero); 1561 if (m == NULL) 1562 break; 1563 if (vm_page_busy_try(m, TRUE)) { 1564 /* 1565 * Various mechanisms such as a pmap_collect can 1566 * result in a busy page on the free queue. We 1567 * have to move the page out of the way so we can 1568 * retry the allocation. If the other thread is not 1569 * allocating the page then m->valid will remain 0 and 1570 * the pageout daemon will free the page later on. 1571 * 1572 * Since we could not busy the page, however, we 1573 * cannot make assumptions as to whether the page 1574 * will be allocated by the other thread or not, 1575 * so all we can do is deactivate it to move it out 1576 * of the way. In particular, if the other thread 1577 * wires the page it may wind up on the inactive 1578 * queue and the pageout daemon will have to deal 1579 * with that case too. 1580 */ 1581 _vm_page_deactivate_locked(m, 0); 1582 vm_page_spin_unlock(m); 1583 } else { 1584 /* 1585 * Theoretically if we are able to busy the page 1586 * atomic with the queue removal (using the vm_page 1587 * lock) nobody else should be able to mess with the 1588 * page before us. 1589 */ 1590 KKASSERT((m->flags & (PG_UNMANAGED | 1591 PG_NEED_COMMIT)) == 0); 1592 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1593 "pg %p q=%d flags=%08x hold=%d wire=%d", 1594 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1595 KKASSERT(m->wire_count == 0); 1596 vm_page_spin_unlock(m); 1597 pagedaemon_wakeup(); 1598 1599 /* return busied and removed page */ 1600 return(m); 1601 } 1602 } 1603 return(m); 1604 } 1605 1606 /* 1607 * vm_page_alloc() 1608 * 1609 * Allocate and return a memory cell associated with this VM object/offset 1610 * pair. If object is NULL an unassociated page will be allocated. 1611 * 1612 * The returned page will be busied and removed from its queues. This 1613 * routine can block and may return NULL if a race occurs and the page 1614 * is found to already exist at the specified (object, pindex). 1615 * 1616 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1617 * VM_ALLOC_QUICK like normal but cannot use cache 1618 * VM_ALLOC_SYSTEM greater free drain 1619 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1620 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1621 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1622 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1623 * (see vm_page_grab()) 1624 * VM_ALLOC_USE_GD ok to use per-gd cache 1625 * 1626 * The object must be held if not NULL 1627 * This routine may not block 1628 * 1629 * Additional special handling is required when called from an interrupt 1630 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1631 * in this case. 1632 */ 1633 vm_page_t 1634 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1635 { 1636 globaldata_t gd = mycpu; 1637 vm_object_t obj; 1638 vm_page_t m; 1639 u_short pg_color; 1640 1641 #if 0 1642 /* 1643 * Special per-cpu free VM page cache. The pages are pre-busied 1644 * and pre-zerod for us. 1645 */ 1646 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1647 crit_enter_gd(gd); 1648 if (gd->gd_vmpg_count) { 1649 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1650 crit_exit_gd(gd); 1651 goto done; 1652 } 1653 crit_exit_gd(gd); 1654 } 1655 #endif 1656 m = NULL; 1657 1658 /* 1659 * CPU LOCALIZATION 1660 * 1661 * CPU localization algorithm. Break the page queues up by physical 1662 * id and core id (note that two cpu threads will have the same core 1663 * id, and core_id != gd_cpuid). 1664 * 1665 * This is nowhere near perfect, for example the last pindex in a 1666 * subgroup will overflow into the next cpu or package. But this 1667 * should get us good page reuse locality in heavy mixed loads. 1668 */ 1669 pg_color = vm_get_pg_color(gd, object, pindex); 1670 1671 KKASSERT(page_req & 1672 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1673 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1674 1675 /* 1676 * Certain system threads (pageout daemon, buf_daemon's) are 1677 * allowed to eat deeper into the free page list. 1678 */ 1679 if (curthread->td_flags & TDF_SYSTHREAD) 1680 page_req |= VM_ALLOC_SYSTEM; 1681 1682 /* 1683 * Impose various limitations. Note that the v_free_reserved test 1684 * must match the opposite of vm_page_count_target() to avoid 1685 * livelocks, be careful. 1686 */ 1687 loop: 1688 if (vmstats.v_free_count >= vmstats.v_free_reserved || 1689 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) || 1690 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 && 1691 vmstats.v_free_count > vmstats.v_interrupt_free_min) 1692 ) { 1693 /* 1694 * The free queue has sufficient free pages to take one out. 1695 */ 1696 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) 1697 m = vm_page_select_free(pg_color, TRUE); 1698 else 1699 m = vm_page_select_free(pg_color, FALSE); 1700 } else if (page_req & VM_ALLOC_NORMAL) { 1701 /* 1702 * Allocatable from the cache (non-interrupt only). On 1703 * success, we must free the page and try again, thus 1704 * ensuring that vmstats.v_*_free_min counters are replenished. 1705 */ 1706 #ifdef INVARIANTS 1707 if (curthread->td_preempted) { 1708 kprintf("vm_page_alloc(): warning, attempt to allocate" 1709 " cache page from preempting interrupt\n"); 1710 m = NULL; 1711 } else { 1712 m = vm_page_select_cache(pg_color); 1713 } 1714 #else 1715 m = vm_page_select_cache(pg_color); 1716 #endif 1717 /* 1718 * On success move the page into the free queue and loop. 1719 * 1720 * Only do this if we can safely acquire the vm_object lock, 1721 * because this is effectively a random page and the caller 1722 * might be holding the lock shared, we don't want to 1723 * deadlock. 1724 */ 1725 if (m != NULL) { 1726 KASSERT(m->dirty == 0, 1727 ("Found dirty cache page %p", m)); 1728 if ((obj = m->object) != NULL) { 1729 if (vm_object_hold_try(obj)) { 1730 vm_page_protect(m, VM_PROT_NONE); 1731 vm_page_free(m); 1732 /* m->object NULL here */ 1733 vm_object_drop(obj); 1734 } else { 1735 vm_page_deactivate(m); 1736 vm_page_wakeup(m); 1737 } 1738 } else { 1739 vm_page_protect(m, VM_PROT_NONE); 1740 vm_page_free(m); 1741 } 1742 goto loop; 1743 } 1744 1745 /* 1746 * On failure return NULL 1747 */ 1748 #if defined(DIAGNOSTIC) 1749 if (vmstats.v_cache_count > 0) 1750 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count); 1751 #endif 1752 atomic_add_int(&vm_pageout_deficit, 1); 1753 pagedaemon_wakeup(); 1754 return (NULL); 1755 } else { 1756 /* 1757 * No pages available, wakeup the pageout daemon and give up. 1758 */ 1759 atomic_add_int(&vm_pageout_deficit, 1); 1760 pagedaemon_wakeup(); 1761 return (NULL); 1762 } 1763 1764 /* 1765 * v_free_count can race so loop if we don't find the expected 1766 * page. 1767 */ 1768 if (m == NULL) 1769 goto loop; 1770 1771 /* 1772 * Good page found. The page has already been busied for us and 1773 * removed from its queues. 1774 */ 1775 KASSERT(m->dirty == 0, 1776 ("vm_page_alloc: free/cache page %p was dirty", m)); 1777 KKASSERT(m->queue == PQ_NONE); 1778 1779 #if 0 1780 done: 1781 #endif 1782 /* 1783 * Initialize the structure, inheriting some flags but clearing 1784 * all the rest. The page has already been busied for us. 1785 */ 1786 vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY)); 1787 KKASSERT(m->wire_count == 0); 1788 KKASSERT(m->busy == 0); 1789 m->act_count = 0; 1790 m->valid = 0; 1791 1792 /* 1793 * Caller must be holding the object lock (asserted by 1794 * vm_page_insert()). 1795 * 1796 * NOTE: Inserting a page here does not insert it into any pmaps 1797 * (which could cause us to block allocating memory). 1798 * 1799 * NOTE: If no object an unassociated page is allocated, m->pindex 1800 * can be used by the caller for any purpose. 1801 */ 1802 if (object) { 1803 if (vm_page_insert(m, object, pindex) == FALSE) { 1804 vm_page_free(m); 1805 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1806 panic("PAGE RACE %p[%ld]/%p", 1807 object, (long)pindex, m); 1808 m = NULL; 1809 } 1810 } else { 1811 m->pindex = pindex; 1812 } 1813 1814 /* 1815 * Don't wakeup too often - wakeup the pageout daemon when 1816 * we would be nearly out of memory. 1817 */ 1818 pagedaemon_wakeup(); 1819 1820 /* 1821 * A PG_BUSY page is returned. 1822 */ 1823 return (m); 1824 } 1825 1826 /* 1827 * Returns number of pages available in our DMA memory reserve 1828 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1829 */ 1830 vm_size_t 1831 vm_contig_avail_pages(void) 1832 { 1833 alist_blk_t blk; 1834 alist_blk_t count; 1835 alist_blk_t bfree; 1836 spin_lock(&vm_contig_spin); 1837 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1838 spin_unlock(&vm_contig_spin); 1839 1840 return bfree; 1841 } 1842 1843 /* 1844 * Attempt to allocate contiguous physical memory with the specified 1845 * requirements. 1846 */ 1847 vm_page_t 1848 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1849 unsigned long alignment, unsigned long boundary, 1850 unsigned long size, vm_memattr_t memattr) 1851 { 1852 alist_blk_t blk; 1853 vm_page_t m; 1854 int i; 1855 1856 alignment >>= PAGE_SHIFT; 1857 if (alignment == 0) 1858 alignment = 1; 1859 boundary >>= PAGE_SHIFT; 1860 if (boundary == 0) 1861 boundary = 1; 1862 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1863 1864 spin_lock(&vm_contig_spin); 1865 blk = alist_alloc(&vm_contig_alist, 0, size); 1866 if (blk == ALIST_BLOCK_NONE) { 1867 spin_unlock(&vm_contig_spin); 1868 if (bootverbose) { 1869 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1870 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1871 } 1872 return(NULL); 1873 } 1874 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1875 alist_free(&vm_contig_alist, blk, size); 1876 spin_unlock(&vm_contig_spin); 1877 if (bootverbose) { 1878 kprintf("vm_page_alloc_contig: %ldk high " 1879 "%016jx failed\n", 1880 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1881 (intmax_t)high); 1882 } 1883 return(NULL); 1884 } 1885 spin_unlock(&vm_contig_spin); 1886 if (vm_contig_verbose) { 1887 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 1888 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 1889 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1890 } 1891 1892 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 1893 if (memattr != VM_MEMATTR_DEFAULT) 1894 for (i = 0;i < size;i++) 1895 pmap_page_set_memattr(&m[i], memattr); 1896 return m; 1897 } 1898 1899 /* 1900 * Free contiguously allocated pages. The pages will be wired but not busy. 1901 * When freeing to the alist we leave them wired and not busy. 1902 */ 1903 void 1904 vm_page_free_contig(vm_page_t m, unsigned long size) 1905 { 1906 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 1907 vm_pindex_t start = pa >> PAGE_SHIFT; 1908 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 1909 1910 if (vm_contig_verbose) { 1911 kprintf("vm_page_free_contig: %016jx/%ldk\n", 1912 (intmax_t)pa, size / 1024); 1913 } 1914 if (pa < vm_low_phys_reserved) { 1915 KKASSERT(pa + size <= vm_low_phys_reserved); 1916 spin_lock(&vm_contig_spin); 1917 alist_free(&vm_contig_alist, start, pages); 1918 spin_unlock(&vm_contig_spin); 1919 } else { 1920 while (pages) { 1921 vm_page_busy_wait(m, FALSE, "cpgfr"); 1922 vm_page_unwire(m, 0); 1923 vm_page_free(m); 1924 --pages; 1925 ++m; 1926 } 1927 1928 } 1929 } 1930 1931 1932 /* 1933 * Wait for sufficient free memory for nominal heavy memory use kernel 1934 * operations. 1935 * 1936 * WARNING! Be sure never to call this in any vm_pageout code path, which 1937 * will trivially deadlock the system. 1938 */ 1939 void 1940 vm_wait_nominal(void) 1941 { 1942 while (vm_page_count_min(0)) 1943 vm_wait(0); 1944 } 1945 1946 /* 1947 * Test if vm_wait_nominal() would block. 1948 */ 1949 int 1950 vm_test_nominal(void) 1951 { 1952 if (vm_page_count_min(0)) 1953 return(1); 1954 return(0); 1955 } 1956 1957 /* 1958 * Block until free pages are available for allocation, called in various 1959 * places before memory allocations. 1960 * 1961 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 1962 * more generous then that. 1963 */ 1964 void 1965 vm_wait(int timo) 1966 { 1967 /* 1968 * never wait forever 1969 */ 1970 if (timo == 0) 1971 timo = hz; 1972 lwkt_gettoken(&vm_token); 1973 1974 if (curthread == pagethread) { 1975 /* 1976 * The pageout daemon itself needs pages, this is bad. 1977 */ 1978 if (vm_page_count_min(0)) { 1979 vm_pageout_pages_needed = 1; 1980 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 1981 } 1982 } else { 1983 /* 1984 * Wakeup the pageout daemon if necessary and wait. 1985 * 1986 * Do not wait indefinitely for the target to be reached, 1987 * as load might prevent it from being reached any time soon. 1988 * But wait a little to try to slow down page allocations 1989 * and to give more important threads (the pagedaemon) 1990 * allocation priority. 1991 */ 1992 if (vm_page_count_target()) { 1993 if (vm_pages_needed == 0) { 1994 vm_pages_needed = 1; 1995 wakeup(&vm_pages_needed); 1996 } 1997 ++vm_pages_waiting; /* SMP race ok */ 1998 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 1999 } 2000 } 2001 lwkt_reltoken(&vm_token); 2002 } 2003 2004 /* 2005 * Block until free pages are available for allocation 2006 * 2007 * Called only from vm_fault so that processes page faulting can be 2008 * easily tracked. 2009 */ 2010 void 2011 vm_wait_pfault(void) 2012 { 2013 /* 2014 * Wakeup the pageout daemon if necessary and wait. 2015 * 2016 * Do not wait indefinitely for the target to be reached, 2017 * as load might prevent it from being reached any time soon. 2018 * But wait a little to try to slow down page allocations 2019 * and to give more important threads (the pagedaemon) 2020 * allocation priority. 2021 */ 2022 if (vm_page_count_min(0)) { 2023 lwkt_gettoken(&vm_token); 2024 while (vm_page_count_severe()) { 2025 if (vm_page_count_target()) { 2026 thread_t td; 2027 2028 if (vm_pages_needed == 0) { 2029 vm_pages_needed = 1; 2030 wakeup(&vm_pages_needed); 2031 } 2032 ++vm_pages_waiting; /* SMP race ok */ 2033 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2034 2035 /* 2036 * Do not stay stuck in the loop if the system is trying 2037 * to kill the process. 2038 */ 2039 td = curthread; 2040 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2041 break; 2042 } 2043 } 2044 lwkt_reltoken(&vm_token); 2045 } 2046 } 2047 2048 /* 2049 * Put the specified page on the active list (if appropriate). Ensure 2050 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2051 * 2052 * The caller should be holding the page busied ? XXX 2053 * This routine may not block. 2054 */ 2055 void 2056 vm_page_activate(vm_page_t m) 2057 { 2058 u_short oqueue; 2059 2060 vm_page_spin_lock(m); 2061 if (m->queue - m->pc != PQ_ACTIVE) { 2062 _vm_page_queue_spin_lock(m); 2063 oqueue = _vm_page_rem_queue_spinlocked(m); 2064 /* page is left spinlocked, queue is unlocked */ 2065 2066 if (oqueue == PQ_CACHE) 2067 mycpu->gd_cnt.v_reactivated++; 2068 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2069 if (m->act_count < ACT_INIT) 2070 m->act_count = ACT_INIT; 2071 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2072 } 2073 _vm_page_and_queue_spin_unlock(m); 2074 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2075 pagedaemon_wakeup(); 2076 } else { 2077 if (m->act_count < ACT_INIT) 2078 m->act_count = ACT_INIT; 2079 vm_page_spin_unlock(m); 2080 } 2081 } 2082 2083 /* 2084 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2085 * routine is called when a page has been added to the cache or free 2086 * queues. 2087 * 2088 * This routine may not block. 2089 */ 2090 static __inline void 2091 vm_page_free_wakeup(void) 2092 { 2093 /* 2094 * If the pageout daemon itself needs pages, then tell it that 2095 * there are some free. 2096 */ 2097 if (vm_pageout_pages_needed && 2098 vmstats.v_cache_count + vmstats.v_free_count >= 2099 vmstats.v_pageout_free_min 2100 ) { 2101 vm_pageout_pages_needed = 0; 2102 wakeup(&vm_pageout_pages_needed); 2103 } 2104 2105 /* 2106 * Wakeup processes that are waiting on memory. 2107 * 2108 * Generally speaking we want to wakeup stuck processes as soon as 2109 * possible. !vm_page_count_min(0) is the absolute minimum point 2110 * where we can do this. Wait a bit longer to reduce degenerate 2111 * re-blocking (vm_page_free_hysteresis). The target check is just 2112 * to make sure the min-check w/hysteresis does not exceed the 2113 * normal target. 2114 */ 2115 if (vm_pages_waiting) { 2116 if (!vm_page_count_min(vm_page_free_hysteresis) || 2117 !vm_page_count_target()) { 2118 vm_pages_waiting = 0; 2119 wakeup(&vmstats.v_free_count); 2120 ++mycpu->gd_cnt.v_ppwakeups; 2121 } 2122 #if 0 2123 if (!vm_page_count_target()) { 2124 /* 2125 * Plenty of pages are free, wakeup everyone. 2126 */ 2127 vm_pages_waiting = 0; 2128 wakeup(&vmstats.v_free_count); 2129 ++mycpu->gd_cnt.v_ppwakeups; 2130 } else if (!vm_page_count_min(0)) { 2131 /* 2132 * Some pages are free, wakeup someone. 2133 */ 2134 int wcount = vm_pages_waiting; 2135 if (wcount > 0) 2136 --wcount; 2137 vm_pages_waiting = wcount; 2138 wakeup_one(&vmstats.v_free_count); 2139 ++mycpu->gd_cnt.v_ppwakeups; 2140 } 2141 #endif 2142 } 2143 } 2144 2145 /* 2146 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2147 * it from its VM object. 2148 * 2149 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 2150 * return (the page will have been freed). 2151 */ 2152 void 2153 vm_page_free_toq(vm_page_t m) 2154 { 2155 mycpu->gd_cnt.v_tfree++; 2156 KKASSERT((m->flags & PG_MAPPED) == 0); 2157 KKASSERT(m->flags & PG_BUSY); 2158 2159 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 2160 kprintf("vm_page_free: pindex(%lu), busy(%d), " 2161 "PG_BUSY(%d), hold(%d)\n", 2162 (u_long)m->pindex, m->busy, 2163 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 2164 if ((m->queue - m->pc) == PQ_FREE) 2165 panic("vm_page_free: freeing free page"); 2166 else 2167 panic("vm_page_free: freeing busy page"); 2168 } 2169 2170 /* 2171 * Remove from object, spinlock the page and its queues and 2172 * remove from any queue. No queue spinlock will be held 2173 * after this section (because the page was removed from any 2174 * queue). 2175 */ 2176 vm_page_remove(m); 2177 vm_page_and_queue_spin_lock(m); 2178 _vm_page_rem_queue_spinlocked(m); 2179 2180 /* 2181 * No further management of fictitious pages occurs beyond object 2182 * and queue removal. 2183 */ 2184 if ((m->flags & PG_FICTITIOUS) != 0) { 2185 vm_page_spin_unlock(m); 2186 vm_page_wakeup(m); 2187 return; 2188 } 2189 2190 m->valid = 0; 2191 vm_page_undirty(m); 2192 2193 if (m->wire_count != 0) { 2194 if (m->wire_count > 1) { 2195 panic( 2196 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2197 m->wire_count, (long)m->pindex); 2198 } 2199 panic("vm_page_free: freeing wired page"); 2200 } 2201 2202 /* 2203 * Clear the UNMANAGED flag when freeing an unmanaged page. 2204 * Clear the NEED_COMMIT flag 2205 */ 2206 if (m->flags & PG_UNMANAGED) 2207 vm_page_flag_clear(m, PG_UNMANAGED); 2208 if (m->flags & PG_NEED_COMMIT) 2209 vm_page_flag_clear(m, PG_NEED_COMMIT); 2210 2211 if (m->hold_count != 0) { 2212 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2213 } else { 2214 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0); 2215 } 2216 2217 /* 2218 * This sequence allows us to clear PG_BUSY while still holding 2219 * its spin lock, which reduces contention vs allocators. We 2220 * must not leave the queue locked or _vm_page_wakeup() may 2221 * deadlock. 2222 */ 2223 _vm_page_queue_spin_unlock(m); 2224 if (_vm_page_wakeup(m)) { 2225 vm_page_spin_unlock(m); 2226 wakeup(m); 2227 } else { 2228 vm_page_spin_unlock(m); 2229 } 2230 vm_page_free_wakeup(); 2231 } 2232 2233 /* 2234 * vm_page_unmanage() 2235 * 2236 * Prevent PV management from being done on the page. The page is 2237 * removed from the paging queues as if it were wired, and as a 2238 * consequence of no longer being managed the pageout daemon will not 2239 * touch it (since there is no way to locate the pte mappings for the 2240 * page). madvise() calls that mess with the pmap will also no longer 2241 * operate on the page. 2242 * 2243 * Beyond that the page is still reasonably 'normal'. Freeing the page 2244 * will clear the flag. 2245 * 2246 * This routine is used by OBJT_PHYS objects - objects using unswappable 2247 * physical memory as backing store rather then swap-backed memory and 2248 * will eventually be extended to support 4MB unmanaged physical 2249 * mappings. 2250 * 2251 * Caller must be holding the page busy. 2252 */ 2253 void 2254 vm_page_unmanage(vm_page_t m) 2255 { 2256 KKASSERT(m->flags & PG_BUSY); 2257 if ((m->flags & PG_UNMANAGED) == 0) { 2258 if (m->wire_count == 0) 2259 vm_page_unqueue(m); 2260 } 2261 vm_page_flag_set(m, PG_UNMANAGED); 2262 } 2263 2264 /* 2265 * Mark this page as wired down by yet another map, removing it from 2266 * paging queues as necessary. 2267 * 2268 * Caller must be holding the page busy. 2269 */ 2270 void 2271 vm_page_wire(vm_page_t m) 2272 { 2273 /* 2274 * Only bump the wire statistics if the page is not already wired, 2275 * and only unqueue the page if it is on some queue (if it is unmanaged 2276 * it is already off the queues). Don't do anything with fictitious 2277 * pages because they are always wired. 2278 */ 2279 KKASSERT(m->flags & PG_BUSY); 2280 if ((m->flags & PG_FICTITIOUS) == 0) { 2281 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2282 if ((m->flags & PG_UNMANAGED) == 0) 2283 vm_page_unqueue(m); 2284 atomic_add_int(&vmstats.v_wire_count, 1); 2285 } 2286 KASSERT(m->wire_count != 0, 2287 ("vm_page_wire: wire_count overflow m=%p", m)); 2288 } 2289 } 2290 2291 /* 2292 * Release one wiring of this page, potentially enabling it to be paged again. 2293 * 2294 * Many pages placed on the inactive queue should actually go 2295 * into the cache, but it is difficult to figure out which. What 2296 * we do instead, if the inactive target is well met, is to put 2297 * clean pages at the head of the inactive queue instead of the tail. 2298 * This will cause them to be moved to the cache more quickly and 2299 * if not actively re-referenced, freed more quickly. If we just 2300 * stick these pages at the end of the inactive queue, heavy filesystem 2301 * meta-data accesses can cause an unnecessary paging load on memory bound 2302 * processes. This optimization causes one-time-use metadata to be 2303 * reused more quickly. 2304 * 2305 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2306 * the inactive queue. This helps the pageout daemon determine memory 2307 * pressure and act on out-of-memory situations more quickly. 2308 * 2309 * BUT, if we are in a low-memory situation we have no choice but to 2310 * put clean pages on the cache queue. 2311 * 2312 * A number of routines use vm_page_unwire() to guarantee that the page 2313 * will go into either the inactive or active queues, and will NEVER 2314 * be placed in the cache - for example, just after dirtying a page. 2315 * dirty pages in the cache are not allowed. 2316 * 2317 * This routine may not block. 2318 */ 2319 void 2320 vm_page_unwire(vm_page_t m, int activate) 2321 { 2322 KKASSERT(m->flags & PG_BUSY); 2323 if (m->flags & PG_FICTITIOUS) { 2324 /* do nothing */ 2325 } else if (m->wire_count <= 0) { 2326 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2327 } else { 2328 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2329 atomic_add_int(&vmstats.v_wire_count, -1); 2330 if (m->flags & PG_UNMANAGED) { 2331 ; 2332 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2333 vm_page_spin_lock(m); 2334 _vm_page_add_queue_spinlocked(m, 2335 PQ_ACTIVE + m->pc, 0); 2336 _vm_page_and_queue_spin_unlock(m); 2337 } else { 2338 vm_page_spin_lock(m); 2339 vm_page_flag_clear(m, PG_WINATCFLS); 2340 _vm_page_add_queue_spinlocked(m, 2341 PQ_INACTIVE + m->pc, 0); 2342 ++vm_swapcache_inactive_heuristic; 2343 _vm_page_and_queue_spin_unlock(m); 2344 } 2345 } 2346 } 2347 } 2348 2349 /* 2350 * Move the specified page to the inactive queue. If the page has 2351 * any associated swap, the swap is deallocated. 2352 * 2353 * Normally athead is 0 resulting in LRU operation. athead is set 2354 * to 1 if we want this page to be 'as if it were placed in the cache', 2355 * except without unmapping it from the process address space. 2356 * 2357 * vm_page's spinlock must be held on entry and will remain held on return. 2358 * This routine may not block. 2359 */ 2360 static void 2361 _vm_page_deactivate_locked(vm_page_t m, int athead) 2362 { 2363 u_short oqueue; 2364 2365 /* 2366 * Ignore if already inactive. 2367 */ 2368 if (m->queue - m->pc == PQ_INACTIVE) 2369 return; 2370 _vm_page_queue_spin_lock(m); 2371 oqueue = _vm_page_rem_queue_spinlocked(m); 2372 2373 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2374 if (oqueue == PQ_CACHE) 2375 mycpu->gd_cnt.v_reactivated++; 2376 vm_page_flag_clear(m, PG_WINATCFLS); 2377 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2378 if (athead == 0) 2379 ++vm_swapcache_inactive_heuristic; 2380 } 2381 /* NOTE: PQ_NONE if condition not taken */ 2382 _vm_page_queue_spin_unlock(m); 2383 /* leaves vm_page spinlocked */ 2384 } 2385 2386 /* 2387 * Attempt to deactivate a page. 2388 * 2389 * No requirements. 2390 */ 2391 void 2392 vm_page_deactivate(vm_page_t m) 2393 { 2394 vm_page_spin_lock(m); 2395 _vm_page_deactivate_locked(m, 0); 2396 vm_page_spin_unlock(m); 2397 } 2398 2399 void 2400 vm_page_deactivate_locked(vm_page_t m) 2401 { 2402 _vm_page_deactivate_locked(m, 0); 2403 } 2404 2405 /* 2406 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2407 * 2408 * This function returns non-zero if it successfully moved the page to 2409 * PQ_CACHE. 2410 * 2411 * This function unconditionally unbusies the page on return. 2412 */ 2413 int 2414 vm_page_try_to_cache(vm_page_t m) 2415 { 2416 vm_page_spin_lock(m); 2417 if (m->dirty || m->hold_count || m->wire_count || 2418 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2419 if (_vm_page_wakeup(m)) { 2420 vm_page_spin_unlock(m); 2421 wakeup(m); 2422 } else { 2423 vm_page_spin_unlock(m); 2424 } 2425 return(0); 2426 } 2427 vm_page_spin_unlock(m); 2428 2429 /* 2430 * Page busied by us and no longer spinlocked. Dirty pages cannot 2431 * be moved to the cache. 2432 */ 2433 vm_page_test_dirty(m); 2434 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2435 vm_page_wakeup(m); 2436 return(0); 2437 } 2438 vm_page_cache(m); 2439 return(1); 2440 } 2441 2442 /* 2443 * Attempt to free the page. If we cannot free it, we do nothing. 2444 * 1 is returned on success, 0 on failure. 2445 * 2446 * No requirements. 2447 */ 2448 int 2449 vm_page_try_to_free(vm_page_t m) 2450 { 2451 vm_page_spin_lock(m); 2452 if (vm_page_busy_try(m, TRUE)) { 2453 vm_page_spin_unlock(m); 2454 return(0); 2455 } 2456 2457 /* 2458 * The page can be in any state, including already being on the free 2459 * queue. Check to see if it really can be freed. 2460 */ 2461 if (m->dirty || /* can't free if it is dirty */ 2462 m->hold_count || /* or held (XXX may be wrong) */ 2463 m->wire_count || /* or wired */ 2464 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2465 PG_NEED_COMMIT)) || /* or needs a commit */ 2466 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2467 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2468 if (_vm_page_wakeup(m)) { 2469 vm_page_spin_unlock(m); 2470 wakeup(m); 2471 } else { 2472 vm_page_spin_unlock(m); 2473 } 2474 return(0); 2475 } 2476 vm_page_spin_unlock(m); 2477 2478 /* 2479 * We can probably free the page. 2480 * 2481 * Page busied by us and no longer spinlocked. Dirty pages will 2482 * not be freed by this function. We have to re-test the 2483 * dirty bit after cleaning out the pmaps. 2484 */ 2485 vm_page_test_dirty(m); 2486 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2487 vm_page_wakeup(m); 2488 return(0); 2489 } 2490 vm_page_protect(m, VM_PROT_NONE); 2491 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2492 vm_page_wakeup(m); 2493 return(0); 2494 } 2495 vm_page_free(m); 2496 return(1); 2497 } 2498 2499 /* 2500 * vm_page_cache 2501 * 2502 * Put the specified page onto the page cache queue (if appropriate). 2503 * 2504 * The page must be busy, and this routine will release the busy and 2505 * possibly even free the page. 2506 */ 2507 void 2508 vm_page_cache(vm_page_t m) 2509 { 2510 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2511 m->busy || m->wire_count || m->hold_count) { 2512 kprintf("vm_page_cache: attempting to cache busy/held page\n"); 2513 vm_page_wakeup(m); 2514 return; 2515 } 2516 2517 /* 2518 * Already in the cache (and thus not mapped) 2519 */ 2520 if ((m->queue - m->pc) == PQ_CACHE) { 2521 KKASSERT((m->flags & PG_MAPPED) == 0); 2522 vm_page_wakeup(m); 2523 return; 2524 } 2525 2526 /* 2527 * Caller is required to test m->dirty, but note that the act of 2528 * removing the page from its maps can cause it to become dirty 2529 * on an SMP system due to another cpu running in usermode. 2530 */ 2531 if (m->dirty) { 2532 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2533 (long)m->pindex); 2534 } 2535 2536 /* 2537 * Remove all pmaps and indicate that the page is not 2538 * writeable or mapped. Our vm_page_protect() call may 2539 * have blocked (especially w/ VM_PROT_NONE), so recheck 2540 * everything. 2541 */ 2542 vm_page_protect(m, VM_PROT_NONE); 2543 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2544 m->busy || m->wire_count || m->hold_count) { 2545 vm_page_wakeup(m); 2546 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2547 vm_page_deactivate(m); 2548 vm_page_wakeup(m); 2549 } else { 2550 _vm_page_and_queue_spin_lock(m); 2551 _vm_page_rem_queue_spinlocked(m); 2552 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2553 _vm_page_queue_spin_unlock(m); 2554 if (_vm_page_wakeup(m)) { 2555 vm_page_spin_unlock(m); 2556 wakeup(m); 2557 } else { 2558 vm_page_spin_unlock(m); 2559 } 2560 vm_page_free_wakeup(); 2561 } 2562 } 2563 2564 /* 2565 * vm_page_dontneed() 2566 * 2567 * Cache, deactivate, or do nothing as appropriate. This routine 2568 * is typically used by madvise() MADV_DONTNEED. 2569 * 2570 * Generally speaking we want to move the page into the cache so 2571 * it gets reused quickly. However, this can result in a silly syndrome 2572 * due to the page recycling too quickly. Small objects will not be 2573 * fully cached. On the otherhand, if we move the page to the inactive 2574 * queue we wind up with a problem whereby very large objects 2575 * unnecessarily blow away our inactive and cache queues. 2576 * 2577 * The solution is to move the pages based on a fixed weighting. We 2578 * either leave them alone, deactivate them, or move them to the cache, 2579 * where moving them to the cache has the highest weighting. 2580 * By forcing some pages into other queues we eventually force the 2581 * system to balance the queues, potentially recovering other unrelated 2582 * space from active. The idea is to not force this to happen too 2583 * often. 2584 * 2585 * The page must be busied. 2586 */ 2587 void 2588 vm_page_dontneed(vm_page_t m) 2589 { 2590 static int dnweight; 2591 int dnw; 2592 int head; 2593 2594 dnw = ++dnweight; 2595 2596 /* 2597 * occassionally leave the page alone 2598 */ 2599 if ((dnw & 0x01F0) == 0 || 2600 m->queue - m->pc == PQ_INACTIVE || 2601 m->queue - m->pc == PQ_CACHE 2602 ) { 2603 if (m->act_count >= ACT_INIT) 2604 --m->act_count; 2605 return; 2606 } 2607 2608 /* 2609 * If vm_page_dontneed() is inactivating a page, it must clear 2610 * the referenced flag; otherwise the pagedaemon will see references 2611 * on the page in the inactive queue and reactivate it. Until the 2612 * page can move to the cache queue, madvise's job is not done. 2613 */ 2614 vm_page_flag_clear(m, PG_REFERENCED); 2615 pmap_clear_reference(m); 2616 2617 if (m->dirty == 0) 2618 vm_page_test_dirty(m); 2619 2620 if (m->dirty || (dnw & 0x0070) == 0) { 2621 /* 2622 * Deactivate the page 3 times out of 32. 2623 */ 2624 head = 0; 2625 } else { 2626 /* 2627 * Cache the page 28 times out of every 32. Note that 2628 * the page is deactivated instead of cached, but placed 2629 * at the head of the queue instead of the tail. 2630 */ 2631 head = 1; 2632 } 2633 vm_page_spin_lock(m); 2634 _vm_page_deactivate_locked(m, head); 2635 vm_page_spin_unlock(m); 2636 } 2637 2638 /* 2639 * These routines manipulate the 'soft busy' count for a page. A soft busy 2640 * is almost like PG_BUSY except that it allows certain compatible operations 2641 * to occur on the page while it is busy. For example, a page undergoing a 2642 * write can still be mapped read-only. 2643 * 2644 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2645 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2646 * busy bit is cleared. 2647 */ 2648 void 2649 vm_page_io_start(vm_page_t m) 2650 { 2651 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2652 atomic_add_char(&m->busy, 1); 2653 vm_page_flag_set(m, PG_SBUSY); 2654 } 2655 2656 void 2657 vm_page_io_finish(vm_page_t m) 2658 { 2659 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2660 atomic_subtract_char(&m->busy, 1); 2661 if (m->busy == 0) 2662 vm_page_flag_clear(m, PG_SBUSY); 2663 } 2664 2665 /* 2666 * Indicate that a clean VM page requires a filesystem commit and cannot 2667 * be reused. Used by tmpfs. 2668 */ 2669 void 2670 vm_page_need_commit(vm_page_t m) 2671 { 2672 vm_page_flag_set(m, PG_NEED_COMMIT); 2673 vm_object_set_writeable_dirty(m->object); 2674 } 2675 2676 void 2677 vm_page_clear_commit(vm_page_t m) 2678 { 2679 vm_page_flag_clear(m, PG_NEED_COMMIT); 2680 } 2681 2682 /* 2683 * Grab a page, blocking if it is busy and allocating a page if necessary. 2684 * A busy page is returned or NULL. The page may or may not be valid and 2685 * might not be on a queue (the caller is responsible for the disposition of 2686 * the page). 2687 * 2688 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2689 * page will be zero'd and marked valid. 2690 * 2691 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2692 * valid even if it already exists. 2693 * 2694 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2695 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2696 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2697 * 2698 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2699 * always returned if we had blocked. 2700 * 2701 * This routine may not be called from an interrupt. 2702 * 2703 * No other requirements. 2704 */ 2705 vm_page_t 2706 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2707 { 2708 vm_page_t m; 2709 int error; 2710 int shared = 1; 2711 2712 KKASSERT(allocflags & 2713 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2714 vm_object_hold_shared(object); 2715 for (;;) { 2716 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2717 if (error) { 2718 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2719 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2720 m = NULL; 2721 break; 2722 } 2723 /* retry */ 2724 } else if (m == NULL) { 2725 if (shared) { 2726 vm_object_upgrade(object); 2727 shared = 0; 2728 } 2729 if (allocflags & VM_ALLOC_RETRY) 2730 allocflags |= VM_ALLOC_NULL_OK; 2731 m = vm_page_alloc(object, pindex, 2732 allocflags & ~VM_ALLOC_RETRY); 2733 if (m) 2734 break; 2735 vm_wait(0); 2736 if ((allocflags & VM_ALLOC_RETRY) == 0) 2737 goto failed; 2738 } else { 2739 /* m found */ 2740 break; 2741 } 2742 } 2743 2744 /* 2745 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2746 * 2747 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2748 * valid even if already valid. 2749 * 2750 * NOTE! We have removed all of the PG_ZERO optimizations and also 2751 * removed the idle zeroing code. These optimizations actually 2752 * slow things down on modern cpus because the zerod area is 2753 * likely uncached, placing a memory-access burden on the 2754 * accesors taking the fault. 2755 * 2756 * By always zeroing the page in-line with the fault, no 2757 * dynamic ram reads are needed and the caches are hot, ready 2758 * for userland to access the memory. 2759 */ 2760 if (m->valid == 0) { 2761 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2762 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2763 m->valid = VM_PAGE_BITS_ALL; 2764 } 2765 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2766 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2767 m->valid = VM_PAGE_BITS_ALL; 2768 } 2769 failed: 2770 vm_object_drop(object); 2771 return(m); 2772 } 2773 2774 /* 2775 * Mapping function for valid bits or for dirty bits in 2776 * a page. May not block. 2777 * 2778 * Inputs are required to range within a page. 2779 * 2780 * No requirements. 2781 * Non blocking. 2782 */ 2783 int 2784 vm_page_bits(int base, int size) 2785 { 2786 int first_bit; 2787 int last_bit; 2788 2789 KASSERT( 2790 base + size <= PAGE_SIZE, 2791 ("vm_page_bits: illegal base/size %d/%d", base, size) 2792 ); 2793 2794 if (size == 0) /* handle degenerate case */ 2795 return(0); 2796 2797 first_bit = base >> DEV_BSHIFT; 2798 last_bit = (base + size - 1) >> DEV_BSHIFT; 2799 2800 return ((2 << last_bit) - (1 << first_bit)); 2801 } 2802 2803 /* 2804 * Sets portions of a page valid and clean. The arguments are expected 2805 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2806 * of any partial chunks touched by the range. The invalid portion of 2807 * such chunks will be zero'd. 2808 * 2809 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2810 * align base to DEV_BSIZE so as not to mark clean a partially 2811 * truncated device block. Otherwise the dirty page status might be 2812 * lost. 2813 * 2814 * This routine may not block. 2815 * 2816 * (base + size) must be less then or equal to PAGE_SIZE. 2817 */ 2818 static void 2819 _vm_page_zero_valid(vm_page_t m, int base, int size) 2820 { 2821 int frag; 2822 int endoff; 2823 2824 if (size == 0) /* handle degenerate case */ 2825 return; 2826 2827 /* 2828 * If the base is not DEV_BSIZE aligned and the valid 2829 * bit is clear, we have to zero out a portion of the 2830 * first block. 2831 */ 2832 2833 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2834 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2835 ) { 2836 pmap_zero_page_area( 2837 VM_PAGE_TO_PHYS(m), 2838 frag, 2839 base - frag 2840 ); 2841 } 2842 2843 /* 2844 * If the ending offset is not DEV_BSIZE aligned and the 2845 * valid bit is clear, we have to zero out a portion of 2846 * the last block. 2847 */ 2848 2849 endoff = base + size; 2850 2851 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2852 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 2853 ) { 2854 pmap_zero_page_area( 2855 VM_PAGE_TO_PHYS(m), 2856 endoff, 2857 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 2858 ); 2859 } 2860 } 2861 2862 /* 2863 * Set valid, clear dirty bits. If validating the entire 2864 * page we can safely clear the pmap modify bit. We also 2865 * use this opportunity to clear the PG_NOSYNC flag. If a process 2866 * takes a write fault on a MAP_NOSYNC memory area the flag will 2867 * be set again. 2868 * 2869 * We set valid bits inclusive of any overlap, but we can only 2870 * clear dirty bits for DEV_BSIZE chunks that are fully within 2871 * the range. 2872 * 2873 * Page must be busied? 2874 * No other requirements. 2875 */ 2876 void 2877 vm_page_set_valid(vm_page_t m, int base, int size) 2878 { 2879 _vm_page_zero_valid(m, base, size); 2880 m->valid |= vm_page_bits(base, size); 2881 } 2882 2883 2884 /* 2885 * Set valid bits and clear dirty bits. 2886 * 2887 * NOTE: This function does not clear the pmap modified bit. 2888 * Also note that e.g. NFS may use a byte-granular base 2889 * and size. 2890 * 2891 * WARNING: Page must be busied? But vfs_clean_one_page() will call 2892 * this without necessarily busying the page (via bdwrite()). 2893 * So for now vm_token must also be held. 2894 * 2895 * No other requirements. 2896 */ 2897 void 2898 vm_page_set_validclean(vm_page_t m, int base, int size) 2899 { 2900 int pagebits; 2901 2902 _vm_page_zero_valid(m, base, size); 2903 pagebits = vm_page_bits(base, size); 2904 m->valid |= pagebits; 2905 m->dirty &= ~pagebits; 2906 if (base == 0 && size == PAGE_SIZE) { 2907 /*pmap_clear_modify(m);*/ 2908 vm_page_flag_clear(m, PG_NOSYNC); 2909 } 2910 } 2911 2912 /* 2913 * Set valid & dirty. Used by buwrite() 2914 * 2915 * WARNING: Page must be busied? But vfs_dirty_one_page() will 2916 * call this function in buwrite() so for now vm_token must 2917 * be held. 2918 * 2919 * No other requirements. 2920 */ 2921 void 2922 vm_page_set_validdirty(vm_page_t m, int base, int size) 2923 { 2924 int pagebits; 2925 2926 pagebits = vm_page_bits(base, size); 2927 m->valid |= pagebits; 2928 m->dirty |= pagebits; 2929 if (m->object) 2930 vm_object_set_writeable_dirty(m->object); 2931 } 2932 2933 /* 2934 * Clear dirty bits. 2935 * 2936 * NOTE: This function does not clear the pmap modified bit. 2937 * Also note that e.g. NFS may use a byte-granular base 2938 * and size. 2939 * 2940 * Page must be busied? 2941 * No other requirements. 2942 */ 2943 void 2944 vm_page_clear_dirty(vm_page_t m, int base, int size) 2945 { 2946 m->dirty &= ~vm_page_bits(base, size); 2947 if (base == 0 && size == PAGE_SIZE) { 2948 /*pmap_clear_modify(m);*/ 2949 vm_page_flag_clear(m, PG_NOSYNC); 2950 } 2951 } 2952 2953 /* 2954 * Make the page all-dirty. 2955 * 2956 * Also make sure the related object and vnode reflect the fact that the 2957 * object may now contain a dirty page. 2958 * 2959 * Page must be busied? 2960 * No other requirements. 2961 */ 2962 void 2963 vm_page_dirty(vm_page_t m) 2964 { 2965 #ifdef INVARIANTS 2966 int pqtype = m->queue - m->pc; 2967 #endif 2968 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 2969 ("vm_page_dirty: page in free/cache queue!")); 2970 if (m->dirty != VM_PAGE_BITS_ALL) { 2971 m->dirty = VM_PAGE_BITS_ALL; 2972 if (m->object) 2973 vm_object_set_writeable_dirty(m->object); 2974 } 2975 } 2976 2977 /* 2978 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2979 * valid and dirty bits for the effected areas are cleared. 2980 * 2981 * Page must be busied? 2982 * Does not block. 2983 * No other requirements. 2984 */ 2985 void 2986 vm_page_set_invalid(vm_page_t m, int base, int size) 2987 { 2988 int bits; 2989 2990 bits = vm_page_bits(base, size); 2991 m->valid &= ~bits; 2992 m->dirty &= ~bits; 2993 m->object->generation++; 2994 } 2995 2996 /* 2997 * The kernel assumes that the invalid portions of a page contain 2998 * garbage, but such pages can be mapped into memory by user code. 2999 * When this occurs, we must zero out the non-valid portions of the 3000 * page so user code sees what it expects. 3001 * 3002 * Pages are most often semi-valid when the end of a file is mapped 3003 * into memory and the file's size is not page aligned. 3004 * 3005 * Page must be busied? 3006 * No other requirements. 3007 */ 3008 void 3009 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3010 { 3011 int b; 3012 int i; 3013 3014 /* 3015 * Scan the valid bits looking for invalid sections that 3016 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3017 * valid bit may be set ) have already been zerod by 3018 * vm_page_set_validclean(). 3019 */ 3020 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3021 if (i == (PAGE_SIZE / DEV_BSIZE) || 3022 (m->valid & (1 << i)) 3023 ) { 3024 if (i > b) { 3025 pmap_zero_page_area( 3026 VM_PAGE_TO_PHYS(m), 3027 b << DEV_BSHIFT, 3028 (i - b) << DEV_BSHIFT 3029 ); 3030 } 3031 b = i + 1; 3032 } 3033 } 3034 3035 /* 3036 * setvalid is TRUE when we can safely set the zero'd areas 3037 * as being valid. We can do this if there are no cache consistency 3038 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3039 */ 3040 if (setvalid) 3041 m->valid = VM_PAGE_BITS_ALL; 3042 } 3043 3044 /* 3045 * Is a (partial) page valid? Note that the case where size == 0 3046 * will return FALSE in the degenerate case where the page is entirely 3047 * invalid, and TRUE otherwise. 3048 * 3049 * Does not block. 3050 * No other requirements. 3051 */ 3052 int 3053 vm_page_is_valid(vm_page_t m, int base, int size) 3054 { 3055 int bits = vm_page_bits(base, size); 3056 3057 if (m->valid && ((m->valid & bits) == bits)) 3058 return 1; 3059 else 3060 return 0; 3061 } 3062 3063 /* 3064 * update dirty bits from pmap/mmu. May not block. 3065 * 3066 * Caller must hold the page busy 3067 */ 3068 void 3069 vm_page_test_dirty(vm_page_t m) 3070 { 3071 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3072 vm_page_dirty(m); 3073 } 3074 } 3075 3076 /* 3077 * Register an action, associating it with its vm_page 3078 */ 3079 void 3080 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 3081 { 3082 struct vm_page_action_list *list; 3083 int hv; 3084 3085 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3086 list = &action_list[hv]; 3087 3088 lwkt_gettoken(&vm_token); 3089 vm_page_flag_set(action->m, PG_ACTIONLIST); 3090 action->event = event; 3091 LIST_INSERT_HEAD(list, action, entry); 3092 lwkt_reltoken(&vm_token); 3093 } 3094 3095 /* 3096 * Unregister an action, disassociating it from its related vm_page 3097 */ 3098 void 3099 vm_page_unregister_action(vm_page_action_t action) 3100 { 3101 struct vm_page_action_list *list; 3102 int hv; 3103 3104 lwkt_gettoken(&vm_token); 3105 if (action->event != VMEVENT_NONE) { 3106 action->event = VMEVENT_NONE; 3107 LIST_REMOVE(action, entry); 3108 3109 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3110 list = &action_list[hv]; 3111 if (LIST_EMPTY(list)) 3112 vm_page_flag_clear(action->m, PG_ACTIONLIST); 3113 } 3114 lwkt_reltoken(&vm_token); 3115 } 3116 3117 /* 3118 * Issue an event on a VM page. Corresponding action structures are 3119 * removed from the page's list and called. 3120 * 3121 * If the vm_page has no more pending action events we clear its 3122 * PG_ACTIONLIST flag. 3123 */ 3124 void 3125 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 3126 { 3127 struct vm_page_action_list *list; 3128 struct vm_page_action *scan; 3129 struct vm_page_action *next; 3130 int hv; 3131 int all; 3132 3133 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 3134 list = &action_list[hv]; 3135 all = 1; 3136 3137 lwkt_gettoken(&vm_token); 3138 LIST_FOREACH_MUTABLE(scan, list, entry, next) { 3139 if (scan->m == m) { 3140 if (scan->event == event) { 3141 scan->event = VMEVENT_NONE; 3142 LIST_REMOVE(scan, entry); 3143 scan->func(m, scan); 3144 /* XXX */ 3145 } else { 3146 all = 0; 3147 } 3148 } 3149 } 3150 if (all) 3151 vm_page_flag_clear(m, PG_ACTIONLIST); 3152 lwkt_reltoken(&vm_token); 3153 } 3154 3155 #include "opt_ddb.h" 3156 #ifdef DDB 3157 #include <sys/kernel.h> 3158 3159 #include <ddb/ddb.h> 3160 3161 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3162 { 3163 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3164 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3165 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3166 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3167 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3168 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3169 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3170 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3171 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3172 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3173 } 3174 3175 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3176 { 3177 int i; 3178 db_printf("PQ_FREE:"); 3179 for(i=0;i<PQ_L2_SIZE;i++) { 3180 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3181 } 3182 db_printf("\n"); 3183 3184 db_printf("PQ_CACHE:"); 3185 for(i=0;i<PQ_L2_SIZE;i++) { 3186 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3187 } 3188 db_printf("\n"); 3189 3190 db_printf("PQ_ACTIVE:"); 3191 for(i=0;i<PQ_L2_SIZE;i++) { 3192 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3193 } 3194 db_printf("\n"); 3195 3196 db_printf("PQ_INACTIVE:"); 3197 for(i=0;i<PQ_L2_SIZE;i++) { 3198 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3199 } 3200 db_printf("\n"); 3201 } 3202 #endif /* DDB */ 3203