1 /* 2 * Copyright (c) 2003-2019 The DragonFly Project. All rights reserved. 3 * Copyright (c) 1991 Regents of the University of California. 4 * All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * This code is derived from software contributed to The DragonFly Project 10 * by Matthew Dillon <dillon@backplane.com> 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 38 */ 39 40 /* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 /* 67 * Resident memory management module. The module manipulates 'VM pages'. 68 * A VM page is the core building block for memory management. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/malloc.h> 74 #include <sys/proc.h> 75 #include <sys/vmmeter.h> 76 #include <sys/vnode.h> 77 #include <sys/kernel.h> 78 #include <sys/alist.h> 79 #include <sys/sysctl.h> 80 #include <sys/cpu_topology.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <sys/lock.h> 85 #include <vm/vm_kern.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pageout.h> 91 #include <vm/vm_pager.h> 92 #include <vm/vm_extern.h> 93 #include <vm/swap_pager.h> 94 95 #include <machine/inttypes.h> 96 #include <machine/md_var.h> 97 #include <machine/specialreg.h> 98 #include <machine/bus_dma.h> 99 100 #include <vm/vm_page2.h> 101 #include <sys/spinlock2.h> 102 103 /* 104 * Cache necessary elements in the hash table itself to avoid indirecting 105 * through random vm_page's when doing a lookup. The hash table is 106 * heuristical and it is ok for races to mess up any or all fields. 107 */ 108 struct vm_page_hash_elm { 109 vm_page_t m; 110 vm_object_t object; /* heuristical */ 111 vm_pindex_t pindex; /* heuristical */ 112 int ticks; 113 int unused; 114 }; 115 116 #define VM_PAGE_HASH_SET 4 /* power of 2, set-assoc */ 117 #define VM_PAGE_HASH_MAX (8 * 1024 * 1024) /* power of 2, max size */ 118 119 /* 120 * SET - Minimum required set associative size, must be a power of 2. We 121 * want this to match or exceed the set-associativeness of the cpu, 122 * up to a reasonable limit (we will use 16). 123 */ 124 __read_mostly static int set_assoc_mask = 16 - 1; 125 126 static void vm_page_queue_init(void); 127 static void vm_page_free_wakeup(void); 128 static vm_page_t vm_page_select_cache(u_short pg_color); 129 static vm_page_t _vm_page_list_find_wide(int basequeue, int index, int *lastp); 130 static vm_page_t _vm_page_list_find2_wide(int bq1, int bq2, int index, 131 int *lastp1, int *lastp); 132 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 133 static void vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes); 134 135 /* 136 * Array of tailq lists 137 */ 138 struct vpgqueues vm_page_queues[PQ_COUNT]; 139 140 static volatile int vm_pages_waiting; 141 static struct alist vm_contig_alist; 142 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 143 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 144 145 __read_mostly static int vm_page_hash_vnode_only; 146 __read_mostly static int vm_page_hash_size; 147 __read_mostly static struct vm_page_hash_elm *vm_page_hash; 148 149 static u_long vm_dma_reserved = 0; 150 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 151 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 152 "Memory reserved for DMA"); 153 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 154 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 155 156 SYSCTL_INT(_vm, OID_AUTO, page_hash_vnode_only, CTLFLAG_RW, 157 &vm_page_hash_vnode_only, 0, "Only hash vnode pages"); 158 #if 0 159 static int vm_page_hash_debug; 160 SYSCTL_INT(_vm, OID_AUTO, page_hash_debug, CTLFLAG_RW, 161 &vm_page_hash_debug, 0, "Only hash vnode pages"); 162 #endif 163 164 static int vm_contig_verbose = 0; 165 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 166 167 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 168 vm_pindex_t, pindex); 169 170 static void 171 vm_page_queue_init(void) 172 { 173 int i; 174 175 for (i = 0; i < PQ_L2_SIZE; i++) 176 vm_page_queues[PQ_FREE+i].cnt_offset = 177 offsetof(struct vmstats, v_free_count); 178 for (i = 0; i < PQ_L2_SIZE; i++) 179 vm_page_queues[PQ_CACHE+i].cnt_offset = 180 offsetof(struct vmstats, v_cache_count); 181 for (i = 0; i < PQ_L2_SIZE; i++) 182 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 183 offsetof(struct vmstats, v_inactive_count); 184 for (i = 0; i < PQ_L2_SIZE; i++) 185 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 186 offsetof(struct vmstats, v_active_count); 187 for (i = 0; i < PQ_L2_SIZE; i++) 188 vm_page_queues[PQ_HOLD+i].cnt_offset = 189 offsetof(struct vmstats, v_active_count); 190 /* PQ_NONE has no queue */ 191 192 for (i = 0; i < PQ_COUNT; i++) { 193 vm_page_queues[i].lastq = -1; 194 TAILQ_INIT(&vm_page_queues[i].pl); 195 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 196 } 197 } 198 199 /* 200 * note: place in initialized data section? Is this necessary? 201 */ 202 vm_pindex_t first_page = 0; 203 vm_pindex_t vm_page_array_size = 0; 204 vm_page_t vm_page_array = NULL; 205 vm_paddr_t vm_low_phys_reserved; 206 207 /* 208 * (low level boot) 209 * 210 * Sets the page size, perhaps based upon the memory size. 211 * Must be called before any use of page-size dependent functions. 212 */ 213 void 214 vm_set_page_size(void) 215 { 216 if (vmstats.v_page_size == 0) 217 vmstats.v_page_size = PAGE_SIZE; 218 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 219 panic("vm_set_page_size: page size not a power of two"); 220 } 221 222 /* 223 * (low level boot) 224 * 225 * Add a new page to the freelist for use by the system. New pages 226 * are added to both the head and tail of the associated free page 227 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 228 * requests pull 'recent' adds (higher physical addresses) first. 229 * 230 * Beware that the page zeroing daemon will also be running soon after 231 * boot, moving pages from the head to the tail of the PQ_FREE queues. 232 * 233 * Must be called in a critical section. 234 */ 235 static void 236 vm_add_new_page(vm_paddr_t pa, int *badcountp) 237 { 238 struct vpgqueues *vpq; 239 vm_page_t m; 240 241 m = PHYS_TO_VM_PAGE(pa); 242 243 /* 244 * Make sure it isn't a duplicate (due to BIOS page range overlaps, 245 * which we consider bugs... but don't crash). Note that m->phys_addr 246 * is pre-initialized, so use m->queue as a check. 247 */ 248 if (m->queue) { 249 if (*badcountp < 10) { 250 kprintf("vm_add_new_page: duplicate pa %016jx\n", 251 (intmax_t)pa); 252 ++*badcountp; 253 } else if (*badcountp == 10) { 254 kprintf("vm_add_new_page: duplicate pa (many more)\n"); 255 ++*badcountp; 256 } 257 return; 258 } 259 260 m->phys_addr = pa; 261 m->flags = 0; 262 m->pat_mode = PAT_WRITE_BACK; 263 m->pc = (pa >> PAGE_SHIFT); 264 265 /* 266 * Twist for cpu localization in addition to page coloring, so 267 * different cpus selecting by m->queue get different page colors. 268 */ 269 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 270 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 271 m->pc &= PQ_L2_MASK; 272 273 /* 274 * Reserve a certain number of contiguous low memory pages for 275 * contigmalloc() to use. 276 * 277 * Even though these pages represent real ram and can be 278 * reverse-mapped, we set PG_FICTITIOUS and PG_UNQUEUED 279 * because their use is special-cased. 280 * 281 * WARNING! Once PG_FICTITIOUS is set, vm_page_wire*() 282 * and vm_page_unwire*() calls have no effect. 283 */ 284 if (pa < vm_low_phys_reserved) { 285 atomic_add_long(&vmstats.v_page_count, 1); 286 atomic_add_long(&vmstats.v_dma_pages, 1); 287 m->flags |= PG_FICTITIOUS | PG_UNQUEUED; 288 m->queue = PQ_NONE; 289 m->wire_count = 1; 290 atomic_add_long(&vmstats.v_wire_count, 1); 291 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 292 return; 293 } 294 295 /* 296 * General page 297 */ 298 m->queue = m->pc + PQ_FREE; 299 KKASSERT(m->dirty == 0); 300 301 atomic_add_long(&vmstats.v_page_count, 1); 302 atomic_add_long(&vmstats.v_free_count, 1); 303 vpq = &vm_page_queues[m->queue]; 304 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 305 ++vpq->lcnt; 306 } 307 308 /* 309 * (low level boot) 310 * 311 * Initializes the resident memory module. 312 * 313 * Preallocates memory for critical VM structures and arrays prior to 314 * kernel_map becoming available. 315 * 316 * Memory is allocated from (virtual2_start, virtual2_end) if available, 317 * otherwise memory is allocated from (virtual_start, virtual_end). 318 * 319 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 320 * large enough to hold vm_page_array & other structures for machines with 321 * large amounts of ram, so we want to use virtual2* when available. 322 */ 323 void 324 vm_page_startup(void) 325 { 326 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 327 vm_offset_t mapped; 328 vm_pindex_t npages; 329 vm_paddr_t page_range; 330 vm_paddr_t new_end; 331 int i; 332 vm_paddr_t pa; 333 vm_paddr_t last_pa; 334 vm_paddr_t end; 335 vm_paddr_t biggestone, biggestsize; 336 vm_paddr_t total; 337 vm_page_t m; 338 int badcount; 339 340 total = 0; 341 badcount = 0; 342 biggestsize = 0; 343 biggestone = 0; 344 vaddr = round_page(vaddr); 345 346 /* 347 * Make sure ranges are page-aligned. 348 */ 349 for (i = 0; phys_avail[i].phys_end; ++i) { 350 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 351 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 352 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 353 phys_avail[i].phys_end = phys_avail[i].phys_beg; 354 } 355 356 /* 357 * Locate largest block 358 */ 359 for (i = 0; phys_avail[i].phys_end; ++i) { 360 vm_paddr_t size = phys_avail[i].phys_end - 361 phys_avail[i].phys_beg; 362 363 if (size > biggestsize) { 364 biggestone = i; 365 biggestsize = size; 366 } 367 total += size; 368 } 369 --i; /* adjust to last entry for use down below */ 370 371 end = phys_avail[biggestone].phys_end; 372 end = trunc_page(end); 373 374 /* 375 * Initialize the queue headers for the free queue, the active queue 376 * and the inactive queue. 377 */ 378 vm_page_queue_init(); 379 380 #if !defined(_KERNEL_VIRTUAL) 381 /* 382 * VKERNELs don't support minidumps and as such don't need 383 * vm_page_dump 384 * 385 * Allocate a bitmap to indicate that a random physical page 386 * needs to be included in a minidump. 387 * 388 * The amd64 port needs this to indicate which direct map pages 389 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 390 * 391 * However, x86 still needs this workspace internally within the 392 * minidump code. In theory, they are not needed on x86, but are 393 * included should the sf_buf code decide to use them. 394 */ 395 page_range = phys_avail[i].phys_end / PAGE_SIZE; 396 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 397 end -= vm_page_dump_size; 398 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 399 VM_PROT_READ | VM_PROT_WRITE); 400 bzero((void *)vm_page_dump, vm_page_dump_size); 401 #endif 402 /* 403 * Compute the number of pages of memory that will be available for 404 * use (taking into account the overhead of a page structure per 405 * page). 406 */ 407 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 408 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 409 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 410 411 #ifndef _KERNEL_VIRTUAL 412 /* 413 * (only applies to real kernels) 414 * 415 * Reserve a large amount of low memory for potential 32-bit DMA 416 * space allocations. Once device initialization is complete we 417 * release most of it, but keep (vm_dma_reserved) memory reserved 418 * for later use. Typically for X / graphics. Through trial and 419 * error we find that GPUs usually requires ~60-100MB or so. 420 * 421 * By default, 128M is left in reserve on machines with 2G+ of ram. 422 */ 423 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 424 if (vm_low_phys_reserved > total / 4) 425 vm_low_phys_reserved = total / 4; 426 if (vm_dma_reserved == 0) { 427 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 428 if (vm_dma_reserved > total / 16) 429 vm_dma_reserved = total / 16; 430 } 431 #endif 432 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 433 ALIST_RECORDS_65536); 434 435 /* 436 * Initialize the mem entry structures now, and put them in the free 437 * queue. 438 */ 439 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 440 kprintf("initializing vm_page_array "); 441 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 442 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 443 vm_page_array = (vm_page_t)mapped; 444 445 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 446 /* 447 * since pmap_map on amd64 returns stuff out of a direct-map region, 448 * we have to manually add these pages to the minidump tracking so 449 * that they can be dumped, including the vm_page_array. 450 */ 451 for (pa = new_end; 452 pa < phys_avail[biggestone].phys_end; 453 pa += PAGE_SIZE) { 454 dump_add_page(pa); 455 } 456 #endif 457 458 /* 459 * Clear all of the page structures, run basic initialization so 460 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 461 * map. 462 */ 463 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 464 vm_page_array_size = page_range; 465 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 466 kprintf("size = 0x%zx\n", vm_page_array_size); 467 468 m = &vm_page_array[0]; 469 pa = ptoa(first_page); 470 for (i = 0; i < page_range; ++i) { 471 spin_init(&m->spin, "vm_page"); 472 m->phys_addr = pa; 473 pa += PAGE_SIZE; 474 ++m; 475 } 476 477 /* 478 * Construct the free queue(s) in ascending order (by physical 479 * address) so that the first 16MB of physical memory is allocated 480 * last rather than first. On large-memory machines, this avoids 481 * the exhaustion of low physical memory before isa_dma_init has run. 482 */ 483 vmstats.v_page_count = 0; 484 vmstats.v_free_count = 0; 485 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 486 pa = phys_avail[i].phys_beg; 487 if (i == biggestone) 488 last_pa = new_end; 489 else 490 last_pa = phys_avail[i].phys_end; 491 while (pa < last_pa && npages-- > 0) { 492 vm_add_new_page(pa, &badcount); 493 pa += PAGE_SIZE; 494 } 495 } 496 if (virtual2_start) 497 virtual2_start = vaddr; 498 else 499 virtual_start = vaddr; 500 mycpu->gd_vmstats = vmstats; 501 } 502 503 /* 504 * (called from early boot only) 505 * 506 * Reorganize VM pages based on numa data. May be called as many times as 507 * necessary. Will reorganize the vm_page_t page color and related queue(s) 508 * to allow vm_page_alloc() to choose pages based on socket affinity. 509 * 510 * NOTE: This function is only called while we are still in UP mode, so 511 * we only need a critical section to protect the queues (which 512 * saves a lot of time, there are likely a ton of pages). 513 */ 514 void 515 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 516 { 517 vm_paddr_t scan_beg; 518 vm_paddr_t scan_end; 519 vm_paddr_t ran_end; 520 struct vpgqueues *vpq; 521 vm_page_t m; 522 vm_page_t mend; 523 int socket_mod; 524 int socket_value; 525 int i; 526 527 /* 528 * Check if no physical information, or there was only one socket 529 * (so don't waste time doing nothing!). 530 */ 531 if (cpu_topology_phys_ids <= 1 || 532 cpu_topology_core_ids == 0) { 533 return; 534 } 535 536 /* 537 * Setup for our iteration. Note that ACPI may iterate CPU 538 * sockets starting at 0 or 1 or some other number. The 539 * cpu_topology code mod's it against the socket count. 540 */ 541 ran_end = ran_beg + bytes; 542 543 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 544 socket_value = (physid % cpu_topology_phys_ids) * socket_mod; 545 mend = &vm_page_array[vm_page_array_size]; 546 547 crit_enter(); 548 549 /* 550 * Adjust cpu_topology's phys_mem parameter 551 */ 552 if (root_cpu_node) 553 vm_numa_add_topology_mem(root_cpu_node, physid, (long)bytes); 554 555 /* 556 * Adjust vm_page->pc and requeue all affected pages. The 557 * allocator will then be able to localize memory allocations 558 * to some degree. 559 */ 560 for (i = 0; phys_avail[i].phys_end; ++i) { 561 scan_beg = phys_avail[i].phys_beg; 562 scan_end = phys_avail[i].phys_end; 563 if (scan_end <= ran_beg) 564 continue; 565 if (scan_beg >= ran_end) 566 continue; 567 if (scan_beg < ran_beg) 568 scan_beg = ran_beg; 569 if (scan_end > ran_end) 570 scan_end = ran_end; 571 if (atop(scan_end) > first_page + vm_page_array_size) 572 scan_end = ptoa(first_page + vm_page_array_size); 573 574 m = PHYS_TO_VM_PAGE(scan_beg); 575 while (scan_beg < scan_end) { 576 KKASSERT(m < mend); 577 if (m->queue != PQ_NONE) { 578 vpq = &vm_page_queues[m->queue]; 579 TAILQ_REMOVE(&vpq->pl, m, pageq); 580 --vpq->lcnt; 581 /* queue doesn't change, no need to adj cnt */ 582 m->queue -= m->pc; 583 m->pc %= socket_mod; 584 m->pc += socket_value; 585 m->pc &= PQ_L2_MASK; 586 m->queue += m->pc; 587 vpq = &vm_page_queues[m->queue]; 588 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 589 ++vpq->lcnt; 590 /* queue doesn't change, no need to adj cnt */ 591 } else { 592 m->pc %= socket_mod; 593 m->pc += socket_value; 594 m->pc &= PQ_L2_MASK; 595 } 596 scan_beg += PAGE_SIZE; 597 ++m; 598 } 599 } 600 601 crit_exit(); 602 } 603 604 /* 605 * (called from early boot only) 606 * 607 * Don't allow the NUMA organization to leave vm_page_queues[] nodes 608 * completely empty for a logical cpu. Doing so would force allocations 609 * on that cpu to always borrow from a nearby cpu, create unnecessary 610 * contention, and cause vm_page_alloc() to iterate more queues and run more 611 * slowly. 612 * 613 * This situation can occur when memory sticks are not entirely populated, 614 * populated at different densities, or in naturally assymetric systems 615 * such as the 2990WX. There could very well be many vm_page_queues[] 616 * entries with *NO* pages assigned to them. 617 * 618 * Fixing this up ensures that each logical CPU has roughly the same 619 * sized memory pool, and more importantly ensures that logical CPUs 620 * do not wind up with an empty memory pool. 621 * 622 * At them moment we just iterate the other queues and borrow pages, 623 * moving them into the queues for cpus with severe deficits even though 624 * the memory might not be local to those cpus. I am not doing this in 625 * a 'smart' way, its effectively UMA style (sorta, since its page-by-page 626 * whereas real UMA typically exchanges address bits 8-10 with high address 627 * bits). But it works extremely well and gives us fairly good deterministic 628 * results on the cpu cores associated with these secondary nodes. 629 */ 630 void 631 vm_numa_organize_finalize(void) 632 { 633 struct vpgqueues *vpq; 634 vm_page_t m; 635 long lcnt_lo; 636 long lcnt_hi; 637 int iter; 638 int i; 639 int scale_lim; 640 641 crit_enter(); 642 643 /* 644 * Machines might not use an exact power of 2 for phys_ids, 645 * core_ids, ht_ids, etc. This can slightly reduce the actual 646 * range of indices in vm_page_queues[] that are nominally used. 647 */ 648 if (cpu_topology_ht_ids) { 649 scale_lim = PQ_L2_SIZE / cpu_topology_phys_ids; 650 scale_lim = scale_lim / cpu_topology_core_ids; 651 scale_lim = scale_lim / cpu_topology_ht_ids; 652 scale_lim = scale_lim * cpu_topology_ht_ids; 653 scale_lim = scale_lim * cpu_topology_core_ids; 654 scale_lim = scale_lim * cpu_topology_phys_ids; 655 } else { 656 scale_lim = PQ_L2_SIZE; 657 } 658 659 /* 660 * Calculate an average, set hysteresis for balancing from 661 * 10% below the average to the average. 662 */ 663 lcnt_hi = 0; 664 for (i = 0; i < scale_lim; ++i) { 665 lcnt_hi += vm_page_queues[i].lcnt; 666 } 667 lcnt_hi /= scale_lim; 668 lcnt_lo = lcnt_hi - lcnt_hi / 10; 669 670 kprintf("vm_page: avg %ld pages per queue, %d queues\n", 671 lcnt_hi, scale_lim); 672 673 iter = 0; 674 for (i = 0; i < scale_lim; ++i) { 675 vpq = &vm_page_queues[PQ_FREE + i]; 676 while (vpq->lcnt < lcnt_lo) { 677 struct vpgqueues *vptmp; 678 679 iter = (iter + 1) & PQ_L2_MASK; 680 vptmp = &vm_page_queues[PQ_FREE + iter]; 681 if (vptmp->lcnt < lcnt_hi) 682 continue; 683 m = TAILQ_FIRST(&vptmp->pl); 684 KKASSERT(m->queue == PQ_FREE + iter); 685 TAILQ_REMOVE(&vptmp->pl, m, pageq); 686 --vptmp->lcnt; 687 /* queue doesn't change, no need to adj cnt */ 688 m->queue -= m->pc; 689 m->pc = i; 690 m->queue += m->pc; 691 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 692 ++vpq->lcnt; 693 } 694 } 695 crit_exit(); 696 } 697 698 static 699 void 700 vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes) 701 { 702 int cpuid; 703 int i; 704 705 switch(cpup->type) { 706 case PACKAGE_LEVEL: 707 cpup->phys_mem += bytes; 708 break; 709 case CHIP_LEVEL: 710 /* 711 * All members should have the same chipid, so we only need 712 * to pull out one member. 713 */ 714 if (CPUMASK_TESTNZERO(cpup->members)) { 715 cpuid = BSFCPUMASK(cpup->members); 716 if (physid == 717 get_chip_ID_from_APICID(CPUID_TO_APICID(cpuid))) { 718 cpup->phys_mem += bytes; 719 } 720 } 721 break; 722 case CORE_LEVEL: 723 case THREAD_LEVEL: 724 /* 725 * Just inherit from the parent node 726 */ 727 cpup->phys_mem = cpup->parent_node->phys_mem; 728 break; 729 } 730 for (i = 0; i < MAXCPU && cpup->child_node[i]; ++i) 731 vm_numa_add_topology_mem(cpup->child_node[i], physid, bytes); 732 } 733 734 /* 735 * We tended to reserve a ton of memory for contigmalloc(). Now that most 736 * drivers have initialized we want to return most the remaining free 737 * reserve back to the VM page queues so they can be used for normal 738 * allocations. 739 * 740 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 741 */ 742 static void 743 vm_page_startup_finish(void *dummy __unused) 744 { 745 alist_blk_t blk; 746 alist_blk_t rblk; 747 alist_blk_t count; 748 alist_blk_t xcount; 749 alist_blk_t bfree; 750 vm_page_t m; 751 struct vm_page_hash_elm *mp; 752 int mask; 753 754 /* 755 * Set the set_assoc_mask based on the fitted number of CPUs. 756 * This is a mask, so we subject 1. 757 * 758 * w/PQ_L2_SIZE = 1024, Don't let the associativity drop below 8. 759 * So if we have 256 CPUs, two hyper-threads will wind up sharing. 760 * 761 * The maximum is PQ_L2_SIZE. However, we limit the starting 762 * maximum to 16 (mask = 15) in order to improve the cache locality 763 * of related kernel data structures. 764 */ 765 mask = PQ_L2_SIZE / ncpus_fit - 1; 766 if (mask < 7) /* minimum is 8-way w/256 CPU threads */ 767 mask = 7; 768 if (mask < 15) 769 mask = 15; 770 cpu_ccfence(); 771 set_assoc_mask = mask; 772 773 /* 774 * Return part of the initial reserve back to the system 775 */ 776 spin_lock(&vm_contig_spin); 777 for (;;) { 778 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 779 if (bfree <= vm_dma_reserved / PAGE_SIZE) 780 break; 781 if (count == 0) 782 break; 783 784 /* 785 * Figure out how much of the initial reserve we have to 786 * free in order to reach our target. 787 */ 788 bfree -= vm_dma_reserved / PAGE_SIZE; 789 if (count > bfree) { 790 blk += count - bfree; 791 count = bfree; 792 } 793 794 /* 795 * Calculate the nearest power of 2 <= count. 796 */ 797 for (xcount = 1; xcount <= count; xcount <<= 1) 798 ; 799 xcount >>= 1; 800 blk += count - xcount; 801 count = xcount; 802 803 /* 804 * Allocate the pages from the alist, then free them to 805 * the normal VM page queues. 806 * 807 * Pages allocated from the alist are wired. We have to 808 * busy, unwire, and free them. We must also adjust 809 * vm_low_phys_reserved before freeing any pages to prevent 810 * confusion. 811 */ 812 rblk = alist_alloc(&vm_contig_alist, blk, count); 813 if (rblk != blk) { 814 kprintf("vm_page_startup_finish: Unable to return " 815 "dma space @0x%08x/%d -> 0x%08x\n", 816 blk, count, rblk); 817 break; 818 } 819 atomic_add_long(&vmstats.v_dma_pages, -(long)count); 820 spin_unlock(&vm_contig_spin); 821 822 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 823 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 824 while (count) { 825 vm_page_flag_clear(m, PG_FICTITIOUS | PG_UNQUEUED); 826 vm_page_busy_wait(m, FALSE, "cpgfr"); 827 vm_page_unwire(m, 0); 828 vm_page_free(m); 829 --count; 830 ++m; 831 } 832 spin_lock(&vm_contig_spin); 833 } 834 spin_unlock(&vm_contig_spin); 835 836 /* 837 * Print out how much DMA space drivers have already allocated and 838 * how much is left over. 839 */ 840 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 841 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 842 (PAGE_SIZE / 1024), 843 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 844 845 /* 846 * Power of 2 847 */ 848 vm_page_hash_size = 4096; 849 while (vm_page_hash_size < (vm_page_array_size / 16)) 850 vm_page_hash_size <<= 1; 851 if (vm_page_hash_size > VM_PAGE_HASH_MAX) 852 vm_page_hash_size = VM_PAGE_HASH_MAX; 853 854 /* 855 * hash table for vm_page_lookup_quick() 856 */ 857 mp = (void *)kmem_alloc3(&kernel_map, 858 (vm_page_hash_size + VM_PAGE_HASH_SET) * 859 sizeof(*vm_page_hash), 860 VM_SUBSYS_VMPGHASH, KM_CPU(0)); 861 bzero(mp, (vm_page_hash_size + VM_PAGE_HASH_SET) * sizeof(*mp)); 862 cpu_sfence(); 863 vm_page_hash = mp; 864 } 865 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 866 vm_page_startup_finish, NULL); 867 868 869 /* 870 * Scan comparison function for Red-Black tree scans. An inclusive 871 * (start,end) is expected. Other fields are not used. 872 */ 873 int 874 rb_vm_page_scancmp(struct vm_page *p, void *data) 875 { 876 struct rb_vm_page_scan_info *info = data; 877 878 if (p->pindex < info->start_pindex) 879 return(-1); 880 if (p->pindex > info->end_pindex) 881 return(1); 882 return(0); 883 } 884 885 int 886 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 887 { 888 if (p1->pindex < p2->pindex) 889 return(-1); 890 if (p1->pindex > p2->pindex) 891 return(1); 892 return(0); 893 } 894 895 void 896 vm_page_init(vm_page_t m) 897 { 898 /* do nothing for now. Called from pmap_page_init() */ 899 } 900 901 /* 902 * Each page queue has its own spin lock, which is fairly optimal for 903 * allocating and freeing pages at least. 904 * 905 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 906 * queue spinlock via this function. Also note that m->queue cannot change 907 * unless both the page and queue are locked. 908 */ 909 static __inline 910 void 911 _vm_page_queue_spin_lock(vm_page_t m) 912 { 913 u_short queue; 914 915 queue = m->queue; 916 if (queue != PQ_NONE) { 917 spin_lock(&vm_page_queues[queue].spin); 918 KKASSERT(queue == m->queue); 919 } 920 } 921 922 static __inline 923 void 924 _vm_page_queue_spin_unlock(vm_page_t m) 925 { 926 u_short queue; 927 928 queue = m->queue; 929 cpu_ccfence(); 930 if (queue != PQ_NONE) 931 spin_unlock(&vm_page_queues[queue].spin); 932 } 933 934 static __inline 935 void 936 _vm_page_queues_spin_lock(u_short queue) 937 { 938 cpu_ccfence(); 939 if (queue != PQ_NONE) 940 spin_lock(&vm_page_queues[queue].spin); 941 } 942 943 944 static __inline 945 void 946 _vm_page_queues_spin_unlock(u_short queue) 947 { 948 cpu_ccfence(); 949 if (queue != PQ_NONE) 950 spin_unlock(&vm_page_queues[queue].spin); 951 } 952 953 void 954 vm_page_queue_spin_lock(vm_page_t m) 955 { 956 _vm_page_queue_spin_lock(m); 957 } 958 959 void 960 vm_page_queues_spin_lock(u_short queue) 961 { 962 _vm_page_queues_spin_lock(queue); 963 } 964 965 void 966 vm_page_queue_spin_unlock(vm_page_t m) 967 { 968 _vm_page_queue_spin_unlock(m); 969 } 970 971 void 972 vm_page_queues_spin_unlock(u_short queue) 973 { 974 _vm_page_queues_spin_unlock(queue); 975 } 976 977 /* 978 * This locks the specified vm_page and its queue in the proper order 979 * (page first, then queue). The queue may change so the caller must 980 * recheck on return. 981 */ 982 static __inline 983 void 984 _vm_page_and_queue_spin_lock(vm_page_t m) 985 { 986 vm_page_spin_lock(m); 987 _vm_page_queue_spin_lock(m); 988 } 989 990 static __inline 991 void 992 _vm_page_and_queue_spin_unlock(vm_page_t m) 993 { 994 _vm_page_queues_spin_unlock(m->queue); 995 vm_page_spin_unlock(m); 996 } 997 998 void 999 vm_page_and_queue_spin_unlock(vm_page_t m) 1000 { 1001 _vm_page_and_queue_spin_unlock(m); 1002 } 1003 1004 void 1005 vm_page_and_queue_spin_lock(vm_page_t m) 1006 { 1007 _vm_page_and_queue_spin_lock(m); 1008 } 1009 1010 /* 1011 * Helper function removes vm_page from its current queue. 1012 * Returns the base queue the page used to be on. 1013 * 1014 * The vm_page and the queue must be spinlocked. 1015 * This function will unlock the queue but leave the page spinlocked. 1016 */ 1017 static __inline u_short 1018 _vm_page_rem_queue_spinlocked(vm_page_t m) 1019 { 1020 struct vpgqueues *pq; 1021 u_short queue; 1022 u_short oqueue; 1023 long *cnt_adj; 1024 long *cnt_gd; 1025 1026 queue = m->queue; 1027 if (queue != PQ_NONE) { 1028 pq = &vm_page_queues[queue]; 1029 TAILQ_REMOVE(&pq->pl, m, pageq); 1030 1031 /* 1032 * Primarily adjust our pcpu stats for rollup, which is 1033 * (mycpu->gd_vmstats_adj + offset). This is normally 1034 * synchronized on every hardclock(). 1035 * 1036 * However, in order for the nominal low-memory algorithms 1037 * to work properly if the unsynchronized adjustment gets 1038 * too negative and might trigger the pageout daemon, we 1039 * immediately synchronize with the global structure. 1040 * 1041 * The idea here is to reduce unnecessary SMP cache mastership 1042 * changes in the global vmstats, which can be particularly 1043 * bad in multi-socket systems. 1044 * 1045 * WARNING! In systems with low amounts of memory the 1046 * vm_paging_needed(-1024 * ncpus) test could 1047 * wind up testing a value above the paging target, 1048 * meaning it would almost always return TRUE. In 1049 * that situation we synchronize every time the 1050 * cumulative adjustment falls below -1024. 1051 */ 1052 cnt_adj = (long *)((char *)&mycpu->gd_vmstats_adj + 1053 pq->cnt_offset); 1054 cnt_gd = (long *)((char *)&mycpu->gd_vmstats + 1055 pq->cnt_offset); 1056 atomic_add_long(cnt_adj, -1); 1057 atomic_add_long(cnt_gd, -1); 1058 1059 if (*cnt_adj < -1024 && vm_paging_needed(-1024 * ncpus)) { 1060 u_long copy = atomic_swap_long(cnt_adj, 0); 1061 cnt_adj = (long *)((char *)&vmstats + pq->cnt_offset); 1062 atomic_add_long(cnt_adj, copy); 1063 } 1064 pq->lcnt--; 1065 m->queue = PQ_NONE; 1066 oqueue = queue; 1067 queue -= m->pc; 1068 vm_page_queues_spin_unlock(oqueue); /* intended */ 1069 } 1070 return queue; 1071 } 1072 1073 /* 1074 * Helper function places the vm_page on the specified queue. Generally 1075 * speaking only PQ_FREE pages are placed at the head, to allow them to 1076 * be allocated sooner rather than later on the assumption that they 1077 * are cache-hot. 1078 * 1079 * The vm_page must be spinlocked. 1080 * The vm_page must NOT be FICTITIOUS (that would be a disaster) 1081 * This function will return with both the page and the queue locked. 1082 */ 1083 static __inline void 1084 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 1085 { 1086 struct vpgqueues *pq; 1087 u_long *cnt_adj; 1088 u_long *cnt_gd; 1089 1090 KKASSERT(m->queue == PQ_NONE && 1091 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0); 1092 1093 if (queue != PQ_NONE) { 1094 vm_page_queues_spin_lock(queue); 1095 pq = &vm_page_queues[queue]; 1096 ++pq->lcnt; 1097 1098 /* 1099 * Adjust our pcpu stats. If a system entity really needs 1100 * to incorporate the count it will call vmstats_rollup() 1101 * to roll it all up into the global vmstats strufture. 1102 */ 1103 cnt_adj = (long *)((char *)&mycpu->gd_vmstats_adj + 1104 pq->cnt_offset); 1105 cnt_gd = (long *)((char *)&mycpu->gd_vmstats + 1106 pq->cnt_offset); 1107 atomic_add_long(cnt_adj, 1); 1108 atomic_add_long(cnt_gd, 1); 1109 1110 /* 1111 * PQ_FREE is always handled LIFO style to try to provide 1112 * cache-hot pages to programs. 1113 */ 1114 m->queue = queue; 1115 if (queue - m->pc == PQ_FREE) { 1116 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1117 } else if (athead) { 1118 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1119 } else { 1120 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1121 } 1122 /* leave the queue spinlocked */ 1123 } 1124 } 1125 1126 /* 1127 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 1128 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 1129 * 1130 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 1131 * call will be made before returning. 1132 * 1133 * This function does NOT busy the page and on return the page is not 1134 * guaranteed to be available. 1135 */ 1136 void 1137 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 1138 { 1139 u_int32_t busy_count; 1140 1141 for (;;) { 1142 busy_count = m->busy_count; 1143 cpu_ccfence(); 1144 1145 if ((busy_count & PBUSY_LOCKED) == 0 && 1146 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 1147 break; 1148 } 1149 tsleep_interlock(m, 0); 1150 if (atomic_cmpset_int(&m->busy_count, busy_count, 1151 busy_count | PBUSY_WANTED)) { 1152 atomic_set_int(&m->flags, PG_REFERENCED); 1153 tsleep(m, PINTERLOCKED, msg, 0); 1154 break; 1155 } 1156 } 1157 } 1158 1159 /* 1160 * This calculates and returns a page color given an optional VM object and 1161 * either a pindex or an iterator. We attempt to return a cpu-localized 1162 * pg_color that is still roughly 16-way set-associative. The CPU topology 1163 * is used if it was probed. 1164 * 1165 * The caller may use the returned value to index into e.g. PQ_FREE when 1166 * allocating a page in order to nominally obtain pages that are hopefully 1167 * already localized to the requesting cpu. This function is not able to 1168 * provide any sort of guarantee of this, but does its best to improve 1169 * hardware cache management performance. 1170 * 1171 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 1172 */ 1173 u_short 1174 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 1175 { 1176 u_short pg_color; 1177 int object_pg_color; 1178 1179 /* 1180 * WARNING! cpu_topology_core_ids might not be a power of two. 1181 * We also shouldn't make assumptions about 1182 * cpu_topology_phys_ids either. 1183 * 1184 * WARNING! ncpus might not be known at this time (during early 1185 * boot), and might be set to 1. 1186 * 1187 * General format: [phys_id][core_id][cpuid][set-associativity] 1188 * (but uses modulo, so not necessarily precise bit masks) 1189 */ 1190 object_pg_color = object ? object->pg_color : 0; 1191 1192 if (cpu_topology_ht_ids) { 1193 int phys_id; 1194 int core_id; 1195 int ht_id; 1196 int physcale; 1197 int grpscale; 1198 int cpuscale; 1199 1200 /* 1201 * Translate cpuid to socket, core, and hyperthread id. 1202 */ 1203 phys_id = get_cpu_phys_id(cpuid); 1204 core_id = get_cpu_core_id(cpuid); 1205 ht_id = get_cpu_ht_id(cpuid); 1206 1207 /* 1208 * Calculate pg_color for our array index. 1209 * 1210 * physcale - socket multiplier. 1211 * grpscale - core multiplier (cores per socket) 1212 * cpu* - cpus per core 1213 * 1214 * WARNING! In early boot, ncpus has not yet been 1215 * initialized and may be set to (1). 1216 * 1217 * WARNING! physcale must match the organization that 1218 * vm_numa_organize() creates to ensure that 1219 * we properly localize allocations to the 1220 * requested cpuid. 1221 */ 1222 physcale = PQ_L2_SIZE / cpu_topology_phys_ids; 1223 grpscale = physcale / cpu_topology_core_ids; 1224 cpuscale = grpscale / cpu_topology_ht_ids; 1225 1226 pg_color = phys_id * physcale; 1227 pg_color += core_id * grpscale; 1228 pg_color += ht_id * cpuscale; 1229 pg_color += (pindex + object_pg_color) % cpuscale; 1230 1231 #if 0 1232 if (grpsize >= 8) { 1233 pg_color += (pindex + object_pg_color) % grpsize; 1234 } else { 1235 if (grpsize <= 2) { 1236 grpsize = 8; 1237 } else { 1238 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 1239 grpsize += grpsize; 1240 if (grpsize < 8) 1241 grpsize += grpsize; 1242 } 1243 pg_color += (pindex + object_pg_color) % grpsize; 1244 } 1245 #endif 1246 } else { 1247 /* 1248 * Unknown topology, distribute things evenly. 1249 * 1250 * WARNING! In early boot, ncpus has not yet been 1251 * initialized and may be set to (1). 1252 */ 1253 int cpuscale; 1254 1255 cpuscale = PQ_L2_SIZE / ncpus; 1256 1257 pg_color = cpuid * cpuscale; 1258 pg_color += (pindex + object_pg_color) % cpuscale; 1259 } 1260 return (pg_color & PQ_L2_MASK); 1261 } 1262 1263 /* 1264 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 1265 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 1266 */ 1267 void 1268 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 1269 int also_m_busy, const char *msg 1270 VM_PAGE_DEBUG_ARGS) 1271 { 1272 u_int32_t busy_count; 1273 1274 for (;;) { 1275 busy_count = m->busy_count; 1276 cpu_ccfence(); 1277 if (busy_count & PBUSY_LOCKED) { 1278 tsleep_interlock(m, 0); 1279 if (atomic_cmpset_int(&m->busy_count, busy_count, 1280 busy_count | PBUSY_WANTED)) { 1281 atomic_set_int(&m->flags, PG_REFERENCED); 1282 tsleep(m, PINTERLOCKED, msg, 0); 1283 } 1284 } else if (also_m_busy && busy_count) { 1285 tsleep_interlock(m, 0); 1286 if (atomic_cmpset_int(&m->busy_count, busy_count, 1287 busy_count | PBUSY_WANTED)) { 1288 atomic_set_int(&m->flags, PG_REFERENCED); 1289 tsleep(m, PINTERLOCKED, msg, 0); 1290 } 1291 } else { 1292 if (atomic_cmpset_int(&m->busy_count, busy_count, 1293 busy_count | PBUSY_LOCKED)) { 1294 #ifdef VM_PAGE_DEBUG 1295 m->busy_func = func; 1296 m->busy_line = lineno; 1297 #endif 1298 break; 1299 } 1300 } 1301 } 1302 } 1303 1304 /* 1305 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1306 * m->busy_count is also 0. 1307 * 1308 * Returns non-zero on failure. 1309 */ 1310 int 1311 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1312 VM_PAGE_DEBUG_ARGS) 1313 { 1314 u_int32_t busy_count; 1315 1316 for (;;) { 1317 busy_count = m->busy_count; 1318 cpu_ccfence(); 1319 if (busy_count & PBUSY_LOCKED) 1320 return TRUE; 1321 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1322 return TRUE; 1323 if (atomic_cmpset_int(&m->busy_count, busy_count, 1324 busy_count | PBUSY_LOCKED)) { 1325 #ifdef VM_PAGE_DEBUG 1326 m->busy_func = func; 1327 m->busy_line = lineno; 1328 #endif 1329 return FALSE; 1330 } 1331 } 1332 } 1333 1334 /* 1335 * Clear the BUSY flag and return non-zero to indicate to the caller 1336 * that a wakeup() should be performed. 1337 * 1338 * (inline version) 1339 */ 1340 static __inline 1341 int 1342 _vm_page_wakeup(vm_page_t m) 1343 { 1344 u_int32_t busy_count; 1345 1346 busy_count = m->busy_count; 1347 cpu_ccfence(); 1348 for (;;) { 1349 if (atomic_fcmpset_int(&m->busy_count, &busy_count, 1350 busy_count & 1351 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1352 return((int)(busy_count & PBUSY_WANTED)); 1353 } 1354 } 1355 /* not reached */ 1356 } 1357 1358 /* 1359 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1360 * is typically the last call you make on a page before moving onto 1361 * other things. 1362 */ 1363 void 1364 vm_page_wakeup(vm_page_t m) 1365 { 1366 KASSERT(m->busy_count & PBUSY_LOCKED, 1367 ("vm_page_wakeup: page not busy!!!")); 1368 if (_vm_page_wakeup(m)) 1369 wakeup(m); 1370 } 1371 1372 /* 1373 * Hold a page, preventing reuse. This is typically only called on pages 1374 * in a known state (either held busy, special, or interlocked in some 1375 * manner). Holding a page does not ensure that it remains valid, it only 1376 * prevents reuse. The page must not already be on the FREE queue or in 1377 * any danger of being moved to the FREE queue concurrent with this call. 1378 * 1379 * Other parts of the system can still disassociate the page from its object 1380 * and attempt to free it, or perform read or write I/O on it and/or otherwise 1381 * manipulate the page, but if the page is held the VM system will leave the 1382 * page and its data intact and not cycle it through the FREE queue until 1383 * the last hold has been released. 1384 * 1385 * (see vm_page_wire() if you want to prevent the page from being 1386 * disassociated from its object too). 1387 */ 1388 void 1389 vm_page_hold(vm_page_t m) 1390 { 1391 atomic_add_int(&m->hold_count, 1); 1392 KKASSERT(m->queue - m->pc != PQ_FREE); 1393 } 1394 1395 /* 1396 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1397 * it was freed while held and must be moved back to the FREE queue. 1398 * 1399 * To avoid racing against vm_page_free*() we must re-test conditions 1400 * after obtaining the spin-lock. The initial test can also race a 1401 * vm_page_free*() that is in the middle of moving a page to PQ_HOLD, 1402 * leaving the page on PQ_HOLD with hold_count == 0. Rather than 1403 * throw a spin-lock in the critical path, we rely on the pageout 1404 * daemon to clean-up these loose ends. 1405 * 1406 * More critically, the 'easy movement' between queues without busying 1407 * a vm_page is only allowed for PQ_FREE<->PQ_HOLD. 1408 */ 1409 void 1410 vm_page_unhold(vm_page_t m) 1411 { 1412 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1413 ("vm_page_unhold: pg %p illegal hold_count (%d) or " 1414 "on FREE queue (%d)", 1415 m, m->hold_count, m->queue - m->pc)); 1416 1417 if (atomic_fetchadd_int(&m->hold_count, -1) == 1 && 1418 m->queue - m->pc == PQ_HOLD) { 1419 vm_page_spin_lock(m); 1420 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1421 _vm_page_queue_spin_lock(m); 1422 _vm_page_rem_queue_spinlocked(m); 1423 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1424 _vm_page_queue_spin_unlock(m); 1425 } 1426 vm_page_spin_unlock(m); 1427 } 1428 } 1429 1430 /* 1431 * Create a fictitious page with the specified physical address and 1432 * memory attribute. The memory attribute is the only the machine- 1433 * dependent aspect of a fictitious page that must be initialized. 1434 */ 1435 void 1436 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1437 { 1438 /* 1439 * The page's memattr might have changed since the 1440 * previous initialization. Update the pmap to the 1441 * new memattr. 1442 */ 1443 if ((m->flags & PG_FICTITIOUS) != 0) 1444 goto memattr; 1445 m->phys_addr = paddr; 1446 m->queue = PQ_NONE; 1447 /* Fictitious pages don't use "segind". */ 1448 /* Fictitious pages don't use "order" or "pool". */ 1449 m->flags = PG_FICTITIOUS | PG_UNQUEUED; 1450 m->busy_count = PBUSY_LOCKED; 1451 m->wire_count = 1; 1452 spin_init(&m->spin, "fake_page"); 1453 pmap_page_init(m); 1454 memattr: 1455 pmap_page_set_memattr(m, memattr); 1456 } 1457 1458 /* 1459 * Inserts the given vm_page into the object and object list. 1460 * 1461 * The pagetables are not updated but will presumably fault the page 1462 * in if necessary, or if a kernel page the caller will at some point 1463 * enter the page into the kernel's pmap. We are not allowed to block 1464 * here so we *can't* do this anyway. 1465 * 1466 * This routine may not block. 1467 * This routine must be called with the vm_object held. 1468 * This routine must be called with a critical section held. 1469 * 1470 * This routine returns TRUE if the page was inserted into the object 1471 * successfully, and FALSE if the page already exists in the object. 1472 */ 1473 int 1474 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1475 { 1476 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1477 if (m->object != NULL) 1478 panic("vm_page_insert: already inserted"); 1479 1480 atomic_add_int(&object->generation, 1); 1481 1482 /* 1483 * Associate the VM page with an (object, offset). 1484 * 1485 * The vm_page spin lock is required for interactions with the pmap. 1486 * XXX vm_page_spin_lock() might not be needed for this any more. 1487 */ 1488 vm_page_spin_lock(m); 1489 m->object = object; 1490 m->pindex = pindex; 1491 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1492 m->object = NULL; 1493 m->pindex = 0; 1494 vm_page_spin_unlock(m); 1495 return FALSE; 1496 } 1497 ++object->resident_page_count; 1498 ++mycpu->gd_vmtotal.t_rm; 1499 vm_page_spin_unlock(m); 1500 1501 /* 1502 * Since we are inserting a new and possibly dirty page, 1503 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1504 */ 1505 if ((m->valid & m->dirty) || 1506 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1507 vm_object_set_writeable_dirty(object); 1508 1509 /* 1510 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1511 */ 1512 swap_pager_page_inserted(m); 1513 return TRUE; 1514 } 1515 1516 /* 1517 * Removes the given vm_page_t from the (object,index) table 1518 * 1519 * The page must be BUSY and will remain BUSY on return. 1520 * No other requirements. 1521 * 1522 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1523 * it busy. 1524 * 1525 * NOTE: Caller is responsible for any pmap disposition prior to the 1526 * rename (as the pmap code will not be able to find the entries 1527 * once the object has been disassociated). The caller may choose 1528 * to leave the pmap association intact if this routine is being 1529 * called as part of a rename between shadowed objects. 1530 * 1531 * This routine may not block. 1532 */ 1533 void 1534 vm_page_remove(vm_page_t m) 1535 { 1536 vm_object_t object; 1537 1538 if (m->object == NULL) { 1539 return; 1540 } 1541 1542 if ((m->busy_count & PBUSY_LOCKED) == 0) 1543 panic("vm_page_remove: page not busy"); 1544 1545 object = m->object; 1546 1547 vm_object_hold(object); 1548 1549 /* 1550 * Remove the page from the object and update the object. 1551 * 1552 * The vm_page spin lock is required for interactions with the pmap. 1553 * XXX vm_page_spin_lock() might not be needed for this any more. 1554 */ 1555 vm_page_spin_lock(m); 1556 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1557 --object->resident_page_count; 1558 --mycpu->gd_vmtotal.t_rm; 1559 m->object = NULL; 1560 atomic_add_int(&object->generation, 1); 1561 vm_page_spin_unlock(m); 1562 1563 vm_object_drop(object); 1564 } 1565 1566 /* 1567 * Calculate the hash position for the vm_page hash heuristic. Generally 1568 * speaking we want to localize sequential lookups to reduce memory stalls. 1569 * 1570 * Mask by ~3 to offer 4-way set-assoc 1571 */ 1572 static __inline 1573 struct vm_page_hash_elm * 1574 vm_page_hash_hash(vm_object_t object, vm_pindex_t pindex) 1575 { 1576 size_t hi; 1577 1578 hi = iscsi_crc32(&object, sizeof(object)) << 2; 1579 hi ^= hi >> (23 - 2); 1580 hi += pindex * VM_PAGE_HASH_SET; 1581 #if 0 1582 /* mix it up */ 1583 hi = (intptr_t)object ^ object->pg_color ^ pindex; 1584 hi += object->pg_color * pindex; 1585 hi = hi ^ (hi >> 20); 1586 #endif 1587 hi &= vm_page_hash_size - 1; /* bounds */ 1588 1589 return (&vm_page_hash[hi]); 1590 } 1591 1592 /* 1593 * Heuristical page lookup that does not require any locks. Returns 1594 * a soft-busied page on success, NULL on failure. 1595 * 1596 * Caller must lookup the page the slow way if NULL is returned. 1597 */ 1598 vm_page_t 1599 vm_page_hash_get(vm_object_t object, vm_pindex_t pindex) 1600 { 1601 struct vm_page_hash_elm *mp; 1602 vm_page_t m; 1603 int i; 1604 1605 if (__predict_false(vm_page_hash == NULL)) 1606 return NULL; 1607 mp = vm_page_hash_hash(object, pindex); 1608 for (i = 0; i < VM_PAGE_HASH_SET; ++i, ++mp) { 1609 if (mp->object != object || 1610 mp->pindex != pindex) { 1611 continue; 1612 } 1613 m = mp->m; 1614 cpu_ccfence(); 1615 if (m == NULL) 1616 continue; 1617 if (m->object != object || m->pindex != pindex) 1618 continue; 1619 if (vm_page_sbusy_try(m)) 1620 continue; 1621 if (m->object == object && m->pindex == pindex) { 1622 /* 1623 * On-match optimization - do not update ticks 1624 * unless we have to (reduce cache coherency traffic) 1625 */ 1626 if (mp->ticks != ticks) 1627 mp->ticks = ticks; 1628 return m; 1629 } 1630 vm_page_sbusy_drop(m); 1631 } 1632 return NULL; 1633 } 1634 1635 /* 1636 * Enter page onto vm_page_hash[]. This is a heuristic, SMP collisions 1637 * are allowed. 1638 */ 1639 static __inline 1640 void 1641 vm_page_hash_enter(vm_page_t m) 1642 { 1643 struct vm_page_hash_elm *mp; 1644 struct vm_page_hash_elm *best; 1645 vm_object_t object; 1646 vm_pindex_t pindex; 1647 int best_delta; 1648 int delta; 1649 int i; 1650 1651 /* 1652 * Only enter type-stable vm_pages with well-shared objects. 1653 */ 1654 if ((m->flags & PG_MAPPEDMULTI) == 0) 1655 return; 1656 if (__predict_false(vm_page_hash == NULL || 1657 m < &vm_page_array[0] || 1658 m >= &vm_page_array[vm_page_array_size])) { 1659 return; 1660 } 1661 if (__predict_false(m->object == NULL)) 1662 return; 1663 #if 0 1664 /* 1665 * Disabled at the moment, there are some degenerate conditions 1666 * with often-exec'd programs that get ignored. In particular, 1667 * the kernel's elf loader does a vn_rdwr() on the first page of 1668 * a binary. 1669 */ 1670 if (m->object->ref_count <= 2 || (m->object->flags & OBJ_ONEMAPPING)) 1671 return; 1672 #endif 1673 if (vm_page_hash_vnode_only && m->object->type != OBJT_VNODE) 1674 return; 1675 1676 /* 1677 * Find best entry 1678 */ 1679 object = m->object; 1680 pindex = m->pindex; 1681 1682 mp = vm_page_hash_hash(object, pindex); 1683 best = mp; 1684 best_delta = ticks - best->ticks; 1685 1686 for (i = 0; i < VM_PAGE_HASH_SET; ++i, ++mp) { 1687 if (mp->m == m && 1688 mp->object == object && 1689 mp->pindex == pindex) { 1690 /* 1691 * On-match optimization - do not update ticks 1692 * unless we have to (reduce cache coherency traffic) 1693 */ 1694 if (mp->ticks != ticks) 1695 mp->ticks = ticks; 1696 return; 1697 } 1698 1699 /* 1700 * The best choice is the oldest entry. 1701 * 1702 * Also check for a field overflow, using -1 instead of 0 1703 * to deal with SMP races on accessing the 'ticks' global. 1704 */ 1705 delta = ticks - mp->ticks; 1706 if (delta < -1) 1707 best = mp; 1708 if (best_delta < delta) 1709 best = mp; 1710 } 1711 1712 /* 1713 * Load the entry. Copy a few elements to the hash entry itself 1714 * to reduce memory stalls due to memory indirects on lookups. 1715 */ 1716 best->m = m; 1717 best->object = object; 1718 best->pindex = pindex; 1719 best->ticks = ticks; 1720 } 1721 1722 /* 1723 * Locate and return the page at (object, pindex), or NULL if the 1724 * page could not be found. 1725 * 1726 * The caller must hold the vm_object token. 1727 */ 1728 vm_page_t 1729 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1730 { 1731 vm_page_t m; 1732 1733 /* 1734 * Search the hash table for this object/offset pair 1735 */ 1736 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1737 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1738 if (m) { 1739 KKASSERT(m->object == object && m->pindex == pindex); 1740 vm_page_hash_enter(m); 1741 } 1742 return(m); 1743 } 1744 1745 vm_page_t 1746 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1747 vm_pindex_t pindex, 1748 int also_m_busy, const char *msg 1749 VM_PAGE_DEBUG_ARGS) 1750 { 1751 u_int32_t busy_count; 1752 vm_page_t m; 1753 1754 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1755 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1756 while (m) { 1757 KKASSERT(m->object == object && m->pindex == pindex); 1758 busy_count = m->busy_count; 1759 cpu_ccfence(); 1760 if (busy_count & PBUSY_LOCKED) { 1761 tsleep_interlock(m, 0); 1762 if (atomic_cmpset_int(&m->busy_count, busy_count, 1763 busy_count | PBUSY_WANTED)) { 1764 atomic_set_int(&m->flags, PG_REFERENCED); 1765 tsleep(m, PINTERLOCKED, msg, 0); 1766 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1767 pindex); 1768 } 1769 } else if (also_m_busy && busy_count) { 1770 tsleep_interlock(m, 0); 1771 if (atomic_cmpset_int(&m->busy_count, busy_count, 1772 busy_count | PBUSY_WANTED)) { 1773 atomic_set_int(&m->flags, PG_REFERENCED); 1774 tsleep(m, PINTERLOCKED, msg, 0); 1775 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1776 pindex); 1777 } 1778 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1779 busy_count | PBUSY_LOCKED)) { 1780 #ifdef VM_PAGE_DEBUG 1781 m->busy_func = func; 1782 m->busy_line = lineno; 1783 #endif 1784 vm_page_hash_enter(m); 1785 break; 1786 } 1787 } 1788 return m; 1789 } 1790 1791 /* 1792 * Attempt to lookup and busy a page. 1793 * 1794 * Returns NULL if the page could not be found 1795 * 1796 * Returns a vm_page and error == TRUE if the page exists but could not 1797 * be busied. 1798 * 1799 * Returns a vm_page and error == FALSE on success. 1800 */ 1801 vm_page_t 1802 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1803 vm_pindex_t pindex, 1804 int also_m_busy, int *errorp 1805 VM_PAGE_DEBUG_ARGS) 1806 { 1807 u_int32_t busy_count; 1808 vm_page_t m; 1809 1810 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1811 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1812 *errorp = FALSE; 1813 while (m) { 1814 KKASSERT(m->object == object && m->pindex == pindex); 1815 busy_count = m->busy_count; 1816 cpu_ccfence(); 1817 if (busy_count & PBUSY_LOCKED) { 1818 *errorp = TRUE; 1819 break; 1820 } 1821 if (also_m_busy && busy_count) { 1822 *errorp = TRUE; 1823 break; 1824 } 1825 if (atomic_cmpset_int(&m->busy_count, busy_count, 1826 busy_count | PBUSY_LOCKED)) { 1827 #ifdef VM_PAGE_DEBUG 1828 m->busy_func = func; 1829 m->busy_line = lineno; 1830 #endif 1831 vm_page_hash_enter(m); 1832 break; 1833 } 1834 } 1835 return m; 1836 } 1837 1838 /* 1839 * Returns a page that is only soft-busied for use by the caller in 1840 * a read-only fashion. Returns NULL if the page could not be found, 1841 * the soft busy could not be obtained, or the page data is invalid. 1842 * 1843 * XXX Doesn't handle PG_FICTITIOUS pages at the moment, but there is 1844 * no reason why we couldn't. 1845 */ 1846 vm_page_t 1847 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1848 int pgoff, int pgbytes) 1849 { 1850 vm_page_t m; 1851 1852 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1853 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1854 if (m) { 1855 if ((m->valid != VM_PAGE_BITS_ALL && 1856 !vm_page_is_valid(m, pgoff, pgbytes)) || 1857 (m->flags & PG_FICTITIOUS)) { 1858 m = NULL; 1859 } else if (vm_page_sbusy_try(m)) { 1860 m = NULL; 1861 } else if ((m->valid != VM_PAGE_BITS_ALL && 1862 !vm_page_is_valid(m, pgoff, pgbytes)) || 1863 (m->flags & PG_FICTITIOUS)) { 1864 vm_page_sbusy_drop(m); 1865 m = NULL; 1866 } else { 1867 vm_page_hash_enter(m); 1868 } 1869 } 1870 return m; 1871 } 1872 1873 /* 1874 * Caller must hold the related vm_object 1875 */ 1876 vm_page_t 1877 vm_page_next(vm_page_t m) 1878 { 1879 vm_page_t next; 1880 1881 next = vm_page_rb_tree_RB_NEXT(m); 1882 if (next && next->pindex != m->pindex + 1) 1883 next = NULL; 1884 return (next); 1885 } 1886 1887 /* 1888 * vm_page_rename() 1889 * 1890 * Move the given vm_page from its current object to the specified 1891 * target object/offset. The page must be busy and will remain so 1892 * on return. 1893 * 1894 * new_object must be held. 1895 * This routine might block. XXX ? 1896 * 1897 * NOTE: Swap associated with the page must be invalidated by the move. We 1898 * have to do this for several reasons: (1) we aren't freeing the 1899 * page, (2) we are dirtying the page, (3) the VM system is probably 1900 * moving the page from object A to B, and will then later move 1901 * the backing store from A to B and we can't have a conflict. 1902 * 1903 * NOTE: We *always* dirty the page. It is necessary both for the 1904 * fact that we moved it, and because we may be invalidating 1905 * swap. If the page is on the cache, we have to deactivate it 1906 * or vm_page_dirty() will panic. Dirty pages are not allowed 1907 * on the cache. 1908 * 1909 * NOTE: Caller is responsible for any pmap disposition prior to the 1910 * rename (as the pmap code will not be able to find the entries 1911 * once the object has been disassociated or changed). Nominally 1912 * the caller is moving a page between shadowed objects and so the 1913 * pmap association is retained without having to remove the page 1914 * from it. 1915 */ 1916 void 1917 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1918 { 1919 KKASSERT(m->busy_count & PBUSY_LOCKED); 1920 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1921 if (m->object) { 1922 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1923 vm_page_remove(m); 1924 } 1925 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1926 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1927 new_object, new_pindex); 1928 } 1929 if (m->queue - m->pc == PQ_CACHE) 1930 vm_page_deactivate(m); 1931 vm_page_dirty(m); 1932 } 1933 1934 /* 1935 * vm_page_unqueue() without any wakeup. This routine is used when a page 1936 * is to remain BUSYied by the caller. 1937 * 1938 * This routine may not block. 1939 */ 1940 void 1941 vm_page_unqueue_nowakeup(vm_page_t m) 1942 { 1943 vm_page_and_queue_spin_lock(m); 1944 (void)_vm_page_rem_queue_spinlocked(m); 1945 vm_page_spin_unlock(m); 1946 } 1947 1948 /* 1949 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1950 * if necessary. 1951 * 1952 * This routine may not block. 1953 */ 1954 void 1955 vm_page_unqueue(vm_page_t m) 1956 { 1957 u_short queue; 1958 1959 vm_page_and_queue_spin_lock(m); 1960 queue = _vm_page_rem_queue_spinlocked(m); 1961 if (queue == PQ_FREE || queue == PQ_CACHE) { 1962 vm_page_spin_unlock(m); 1963 pagedaemon_wakeup(); 1964 } else { 1965 vm_page_spin_unlock(m); 1966 } 1967 } 1968 1969 /* 1970 * vm_page_list_find() 1971 * 1972 * Find a page on the specified queue with color optimization. 1973 * 1974 * The page coloring optimization attempts to locate a page that does 1975 * not overload other nearby pages in the object in the cpu's L1 or L2 1976 * caches. We need this optimization because cpu caches tend to be 1977 * physical caches, while object spaces tend to be virtual. 1978 * 1979 * The page coloring optimization also, very importantly, tries to localize 1980 * memory to cpus and physical sockets. 1981 * 1982 * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the 1983 * algorithm is adjusted to localize allocations on a per-core basis. 1984 * This is done by 'twisting' the colors. 1985 * 1986 * The page is returned spinlocked and removed from its queue (it will 1987 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1988 * is responsible for dealing with the busy-page case (usually by 1989 * deactivating the page and looping). 1990 * 1991 * NOTE: This routine is carefully inlined. A non-inlined version 1992 * is available for outside callers but the only critical path is 1993 * from within this source file. 1994 * 1995 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1996 * represent stable storage, allowing us to order our locks vm_page 1997 * first, then queue. 1998 */ 1999 static __inline 2000 vm_page_t 2001 _vm_page_list_find(int basequeue, int index) 2002 { 2003 struct vpgqueues *pq; 2004 vm_page_t m; 2005 2006 index &= PQ_L2_MASK; 2007 pq = &vm_page_queues[basequeue + index]; 2008 2009 /* 2010 * Try this cpu's colored queue first. Test for a page unlocked, 2011 * then lock the queue and locate a page. Note that the lock order 2012 * is reversed, but we do not want to dwadle on the page spinlock 2013 * anyway as it is held significantly longer than the queue spinlock. 2014 */ 2015 if (TAILQ_FIRST(&pq->pl)) { 2016 spin_lock(&pq->spin); 2017 TAILQ_FOREACH(m, &pq->pl, pageq) { 2018 if (spin_trylock(&m->spin) == 0) 2019 continue; 2020 KKASSERT(m->queue == basequeue + index); 2021 pq->lastq = -1; 2022 return(m); 2023 } 2024 spin_unlock(&pq->spin); 2025 } 2026 2027 m = _vm_page_list_find_wide(basequeue, index, &pq->lastq); 2028 2029 return(m); 2030 } 2031 2032 /* 2033 * If we could not find the page in the desired queue try to find it in 2034 * a nearby (NUMA-aware) queue, spreading out as we go. 2035 */ 2036 static vm_page_t 2037 _vm_page_list_find_wide(int basequeue, int index, int *lastp) 2038 { 2039 struct vpgqueues *pq; 2040 vm_page_t m = NULL; 2041 int pqmask = set_assoc_mask >> 1; 2042 int pqi; 2043 int range; 2044 int skip_start; 2045 int skip_next; 2046 int count; 2047 2048 /* 2049 * Avoid re-searching empty queues over and over again skip to 2050 * pq->last if appropriate. 2051 */ 2052 if (*lastp >= 0) 2053 index = *lastp; 2054 2055 index &= PQ_L2_MASK; 2056 pq = &vm_page_queues[basequeue]; 2057 count = 0; 2058 skip_start = -1; 2059 skip_next = -1; 2060 2061 /* 2062 * Run local sets of 16, 32, 64, 128, up to the entire queue if all 2063 * else fails (PQ_L2_MASK). 2064 * 2065 * pqmask is a mask, 15, 31, 63, etc. 2066 * 2067 * Test each queue unlocked first, then lock the queue and locate 2068 * a page. Note that the lock order is reversed, but we do not want 2069 * to dwadle on the page spinlock anyway as it is held significantly 2070 * longer than the queue spinlock. 2071 */ 2072 do { 2073 pqmask = (pqmask << 1) | 1; 2074 2075 pqi = index; 2076 range = pqmask + 1; 2077 2078 while (range > 0) { 2079 if (pqi >= skip_start && pqi < skip_next) { 2080 range -= skip_next - pqi; 2081 pqi = (pqi & ~pqmask) | (skip_next & pqmask); 2082 } 2083 if (range > 0 && TAILQ_FIRST(&pq[pqi].pl)) { 2084 spin_lock(&pq[pqi].spin); 2085 TAILQ_FOREACH(m, &pq[pqi].pl, pageq) { 2086 if (spin_trylock(&m->spin) == 0) 2087 continue; 2088 KKASSERT(m->queue == basequeue + pqi); 2089 2090 /* 2091 * If we had to wander too far, set 2092 * *lastp to skip past empty queues. 2093 */ 2094 if (count >= 8) 2095 *lastp = pqi & PQ_L2_MASK; 2096 return(m); 2097 } 2098 spin_unlock(&pq[pqi].spin); 2099 } 2100 --range; 2101 ++count; 2102 pqi = (pqi & ~pqmask) | ((pqi + 1) & pqmask); 2103 } 2104 skip_start = pqi & ~pqmask; 2105 skip_next = (pqi | pqmask) + 1; 2106 } while (pqmask != PQ_L2_MASK); 2107 2108 return(m); 2109 } 2110 2111 static __inline 2112 vm_page_t 2113 _vm_page_list_find2(int bq1, int bq2, int index) 2114 { 2115 struct vpgqueues *pq1; 2116 struct vpgqueues *pq2; 2117 vm_page_t m; 2118 2119 index &= PQ_L2_MASK; 2120 pq1 = &vm_page_queues[bq1 + index]; 2121 pq2 = &vm_page_queues[bq2 + index]; 2122 2123 /* 2124 * Try this cpu's colored queue first. Test for a page unlocked, 2125 * then lock the queue and locate a page. Note that the lock order 2126 * is reversed, but we do not want to dwadle on the page spinlock 2127 * anyway as it is held significantly longer than the queue spinlock. 2128 */ 2129 if (TAILQ_FIRST(&pq1->pl)) { 2130 spin_lock(&pq1->spin); 2131 TAILQ_FOREACH(m, &pq1->pl, pageq) { 2132 if (spin_trylock(&m->spin) == 0) 2133 continue; 2134 KKASSERT(m->queue == bq1 + index); 2135 pq1->lastq = -1; 2136 pq2->lastq = -1; 2137 return(m); 2138 } 2139 spin_unlock(&pq1->spin); 2140 } 2141 2142 m = _vm_page_list_find2_wide(bq1, bq2, index, &pq1->lastq, &pq2->lastq); 2143 2144 return(m); 2145 } 2146 2147 2148 /* 2149 * This version checks two queues at the same time, widening its search 2150 * as we progress. prefering basequeue1 2151 * and starting on basequeue2 after exhausting the first set. The idea 2152 * is to try to stay localized to the cpu. 2153 */ 2154 static vm_page_t 2155 _vm_page_list_find2_wide(int basequeue1, int basequeue2, int index, 2156 int *lastp1, int *lastp2) 2157 { 2158 struct vpgqueues *pq1; 2159 struct vpgqueues *pq2; 2160 vm_page_t m = NULL; 2161 int pqmask1, pqmask2; 2162 int pqi; 2163 int range; 2164 int skip_start1, skip_start2; 2165 int skip_next1, skip_next2; 2166 int count1, count2; 2167 2168 /* 2169 * Avoid re-searching empty queues over and over again skip to 2170 * pq->last if appropriate. 2171 */ 2172 if (*lastp1 >= 0) 2173 index = *lastp1; 2174 2175 index &= PQ_L2_MASK; 2176 2177 pqmask1 = set_assoc_mask >> 1; 2178 pq1 = &vm_page_queues[basequeue1]; 2179 count1 = 0; 2180 skip_start1 = -1; 2181 skip_next1 = -1; 2182 2183 pqmask2 = set_assoc_mask >> 1; 2184 pq2 = &vm_page_queues[basequeue2]; 2185 count2 = 0; 2186 skip_start2 = -1; 2187 skip_next2 = -1; 2188 2189 /* 2190 * Run local sets of 16, 32, 64, 128, up to the entire queue if all 2191 * else fails (PQ_L2_MASK). 2192 * 2193 * pqmask is a mask, 15, 31, 63, etc. 2194 * 2195 * Test each queue unlocked first, then lock the queue and locate 2196 * a page. Note that the lock order is reversed, but we do not want 2197 * to dwadle on the page spinlock anyway as it is held significantly 2198 * longer than the queue spinlock. 2199 */ 2200 do { 2201 if (pqmask1 == PQ_L2_MASK) 2202 goto skip2; 2203 2204 pqmask1 = (pqmask1 << 1) | 1; 2205 pqi = index; 2206 range = pqmask1 + 1; 2207 2208 while (range > 0) { 2209 if (pqi >= skip_start1 && pqi < skip_next1) { 2210 range -= skip_next1 - pqi; 2211 pqi = (pqi & ~pqmask1) | (skip_next1 & pqmask1); 2212 } 2213 if (range > 0 && TAILQ_FIRST(&pq1[pqi].pl)) { 2214 spin_lock(&pq1[pqi].spin); 2215 TAILQ_FOREACH(m, &pq1[pqi].pl, pageq) { 2216 if (spin_trylock(&m->spin) == 0) 2217 continue; 2218 KKASSERT(m->queue == basequeue1 + pqi); 2219 2220 /* 2221 * If we had to wander too far, set 2222 * *lastp to skip past empty queues. 2223 */ 2224 if (count1 >= 8) 2225 *lastp1 = pqi & PQ_L2_MASK; 2226 return(m); 2227 } 2228 spin_unlock(&pq1[pqi].spin); 2229 } 2230 --range; 2231 ++count1; 2232 pqi = (pqi & ~pqmask1) | ((pqi + 1) & pqmask1); 2233 } 2234 skip_start1 = pqi & ~pqmask1; 2235 skip_next1 = (pqi | pqmask1) + 1; 2236 skip2: 2237 if (pqmask1 < ((set_assoc_mask << 1) | 1)) 2238 continue; 2239 2240 pqmask2 = (pqmask2 << 1) | 1; 2241 pqi = index; 2242 range = pqmask2 + 1; 2243 2244 while (range > 0) { 2245 if (pqi >= skip_start2 && pqi < skip_next2) { 2246 range -= skip_next2 - pqi; 2247 pqi = (pqi & ~pqmask2) | (skip_next2 & pqmask2); 2248 } 2249 if (range > 0 && TAILQ_FIRST(&pq2[pqi].pl)) { 2250 spin_lock(&pq2[pqi].spin); 2251 TAILQ_FOREACH(m, &pq2[pqi].pl, pageq) { 2252 if (spin_trylock(&m->spin) == 0) 2253 continue; 2254 KKASSERT(m->queue == basequeue2 + pqi); 2255 2256 /* 2257 * If we had to wander too far, set 2258 * *lastp to skip past empty queues. 2259 */ 2260 if (count2 >= 8) 2261 *lastp2 = pqi & PQ_L2_MASK; 2262 return(m); 2263 } 2264 spin_unlock(&pq2[pqi].spin); 2265 } 2266 --range; 2267 ++count2; 2268 pqi = (pqi & ~pqmask2) | ((pqi + 1) & pqmask2); 2269 } 2270 skip_start2 = pqi & ~pqmask2; 2271 skip_next2 = (pqi | pqmask2) + 1; 2272 } while (pqmask1 != PQ_L2_MASK && pqmask2 != PQ_L2_MASK); 2273 2274 return(m); 2275 } 2276 2277 /* 2278 * Returns a vm_page candidate for allocation. The page is not busied so 2279 * it can move around. The caller must busy the page (and typically 2280 * deactivate it if it cannot be busied!) 2281 * 2282 * Returns a spinlocked vm_page that has been removed from its queue. 2283 * (note that _vm_page_list_find() does not remove the page from its 2284 * queue). 2285 */ 2286 vm_page_t 2287 vm_page_list_find(int basequeue, int index) 2288 { 2289 vm_page_t m; 2290 2291 m = _vm_page_list_find(basequeue, index); 2292 if (m) 2293 _vm_page_rem_queue_spinlocked(m); 2294 return m; 2295 } 2296 2297 /* 2298 * Find a page on the cache queue with color optimization, remove it 2299 * from the queue, and busy it. The returned page will not be spinlocked. 2300 * 2301 * A candidate failure will be deactivated. Candidates can fail due to 2302 * being busied by someone else, in which case they will be deactivated. 2303 * 2304 * This routine may not block. 2305 * 2306 */ 2307 static vm_page_t 2308 vm_page_select_cache(u_short pg_color) 2309 { 2310 vm_page_t m; 2311 2312 for (;;) { 2313 m = _vm_page_list_find(PQ_CACHE, pg_color); 2314 if (m == NULL) 2315 break; 2316 /* 2317 * (m) has been spinlocked 2318 */ 2319 _vm_page_rem_queue_spinlocked(m); 2320 if (vm_page_busy_try(m, TRUE)) { 2321 _vm_page_deactivate_locked(m, 0); 2322 vm_page_spin_unlock(m); 2323 } else { 2324 /* 2325 * We successfully busied the page 2326 */ 2327 if ((m->flags & PG_NEED_COMMIT) == 0 && 2328 m->hold_count == 0 && 2329 m->wire_count == 0 && 2330 (m->dirty & m->valid) == 0) { 2331 vm_page_spin_unlock(m); 2332 KKASSERT((m->flags & PG_UNQUEUED) == 0); 2333 pagedaemon_wakeup(); 2334 return(m); 2335 } 2336 2337 /* 2338 * The page cannot be recycled, deactivate it. 2339 */ 2340 _vm_page_deactivate_locked(m, 0); 2341 if (_vm_page_wakeup(m)) { 2342 vm_page_spin_unlock(m); 2343 wakeup(m); 2344 } else { 2345 vm_page_spin_unlock(m); 2346 } 2347 } 2348 } 2349 return (m); 2350 } 2351 2352 /* 2353 * Find a free page. We attempt to inline the nominal case and fall back 2354 * to _vm_page_select_free() otherwise. A busied page is removed from 2355 * the queue and returned. 2356 * 2357 * This routine may not block. 2358 */ 2359 static __inline vm_page_t 2360 vm_page_select_free(u_short pg_color) 2361 { 2362 vm_page_t m; 2363 2364 for (;;) { 2365 m = _vm_page_list_find(PQ_FREE, pg_color); 2366 if (m == NULL) 2367 break; 2368 _vm_page_rem_queue_spinlocked(m); 2369 if (vm_page_busy_try(m, TRUE)) { 2370 /* 2371 * Various mechanisms such as a pmap_collect can 2372 * result in a busy page on the free queue. We 2373 * have to move the page out of the way so we can 2374 * retry the allocation. If the other thread is not 2375 * allocating the page then m->valid will remain 0 and 2376 * the pageout daemon will free the page later on. 2377 * 2378 * Since we could not busy the page, however, we 2379 * cannot make assumptions as to whether the page 2380 * will be allocated by the other thread or not, 2381 * so all we can do is deactivate it to move it out 2382 * of the way. In particular, if the other thread 2383 * wires the page it may wind up on the inactive 2384 * queue and the pageout daemon will have to deal 2385 * with that case too. 2386 */ 2387 _vm_page_deactivate_locked(m, 0); 2388 vm_page_spin_unlock(m); 2389 } else { 2390 /* 2391 * Theoretically if we are able to busy the page 2392 * atomic with the queue removal (using the vm_page 2393 * lock) nobody else should have been able to mess 2394 * with the page before us. 2395 * 2396 * Assert the page state. Note that even though 2397 * wiring doesn't adjust queues, a page on the free 2398 * queue should never be wired at this point. 2399 */ 2400 KKASSERT((m->flags & (PG_UNQUEUED | 2401 PG_NEED_COMMIT)) == 0); 2402 KASSERT(m->hold_count == 0, 2403 ("m->hold_count is not zero " 2404 "pg %p q=%d flags=%08x hold=%d wire=%d", 2405 m, m->queue, m->flags, 2406 m->hold_count, m->wire_count)); 2407 KKASSERT(m->wire_count == 0); 2408 vm_page_spin_unlock(m); 2409 pagedaemon_wakeup(); 2410 2411 /* return busied and removed page */ 2412 return(m); 2413 } 2414 } 2415 return(m); 2416 } 2417 2418 static __inline vm_page_t 2419 vm_page_select_free_or_cache(u_short pg_color, int *fromcachep) 2420 { 2421 vm_page_t m; 2422 2423 *fromcachep = 0; 2424 for (;;) { 2425 m = _vm_page_list_find2(PQ_FREE, PQ_CACHE, pg_color); 2426 if (m == NULL) 2427 break; 2428 if (vm_page_busy_try(m, TRUE)) { 2429 _vm_page_rem_queue_spinlocked(m); 2430 _vm_page_deactivate_locked(m, 0); 2431 vm_page_spin_unlock(m); 2432 } else if (m->queue - m->pc == PQ_FREE) { 2433 /* 2434 * We successfully busied the page, PQ_FREE case 2435 */ 2436 _vm_page_rem_queue_spinlocked(m); 2437 KKASSERT((m->flags & (PG_UNQUEUED | 2438 PG_NEED_COMMIT)) == 0); 2439 KASSERT(m->hold_count == 0, 2440 ("m->hold_count is not zero " 2441 "pg %p q=%d flags=%08x hold=%d wire=%d", 2442 m, m->queue, m->flags, 2443 m->hold_count, m->wire_count)); 2444 KKASSERT(m->wire_count == 0); 2445 vm_page_spin_unlock(m); 2446 pagedaemon_wakeup(); 2447 2448 /* return busied and removed page */ 2449 return(m); 2450 } else { 2451 /* 2452 * We successfully busied the page, PQ_CACHE case 2453 */ 2454 _vm_page_rem_queue_spinlocked(m); 2455 if ((m->flags & PG_NEED_COMMIT) == 0 && 2456 m->hold_count == 0 && 2457 m->wire_count == 0 && 2458 (m->dirty & m->valid) == 0) { 2459 vm_page_spin_unlock(m); 2460 KKASSERT((m->flags & PG_UNQUEUED) == 0); 2461 pagedaemon_wakeup(); 2462 *fromcachep = 1; 2463 return(m); 2464 } 2465 2466 /* 2467 * The page cannot be recycled, deactivate it. 2468 */ 2469 _vm_page_deactivate_locked(m, 0); 2470 if (_vm_page_wakeup(m)) { 2471 vm_page_spin_unlock(m); 2472 wakeup(m); 2473 } else { 2474 vm_page_spin_unlock(m); 2475 } 2476 } 2477 } 2478 return(m); 2479 } 2480 2481 /* 2482 * vm_page_alloc() 2483 * 2484 * Allocate and return a memory cell associated with this VM object/offset 2485 * pair. If object is NULL an unassociated page will be allocated. 2486 * 2487 * The returned page will be busied and removed from its queues. This 2488 * routine can block and may return NULL if a race occurs and the page 2489 * is found to already exist at the specified (object, pindex). 2490 * 2491 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 2492 * VM_ALLOC_QUICK like normal but cannot use cache 2493 * VM_ALLOC_SYSTEM greater free drain 2494 * VM_ALLOC_INTERRUPT allow free list to be completely drained 2495 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 2496 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 2497 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 2498 * (see vm_page_grab()) 2499 * VM_ALLOC_USE_GD ok to use per-gd cache 2500 * 2501 * VM_ALLOC_CPU(n) allocate using specified cpu localization 2502 * 2503 * The object must be held if not NULL 2504 * This routine may not block 2505 * 2506 * Additional special handling is required when called from an interrupt 2507 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 2508 * in this case. 2509 */ 2510 vm_page_t 2511 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 2512 { 2513 globaldata_t gd; 2514 vm_object_t obj; 2515 vm_page_t m; 2516 u_short pg_color; 2517 int cpuid_local; 2518 int fromcache; 2519 2520 #if 0 2521 /* 2522 * Special per-cpu free VM page cache. The pages are pre-busied 2523 * and pre-zerod for us. 2524 */ 2525 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 2526 crit_enter_gd(gd); 2527 if (gd->gd_vmpg_count) { 2528 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 2529 crit_exit_gd(gd); 2530 goto done; 2531 } 2532 crit_exit_gd(gd); 2533 } 2534 #endif 2535 m = NULL; 2536 2537 /* 2538 * CPU LOCALIZATION 2539 * 2540 * CPU localization algorithm. Break the page queues up by physical 2541 * id and core id (note that two cpu threads will have the same core 2542 * id, and core_id != gd_cpuid). 2543 * 2544 * This is nowhere near perfect, for example the last pindex in a 2545 * subgroup will overflow into the next cpu or package. But this 2546 * should get us good page reuse locality in heavy mixed loads. 2547 * 2548 * (may be executed before the APs are started, so other GDs might 2549 * not exist!) 2550 */ 2551 if (page_req & VM_ALLOC_CPU_SPEC) 2552 cpuid_local = VM_ALLOC_GETCPU(page_req); 2553 else 2554 cpuid_local = mycpu->gd_cpuid; 2555 2556 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 2557 2558 KKASSERT(page_req & 2559 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 2560 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2561 2562 /* 2563 * Certain system threads (pageout daemon, buf_daemon's) are 2564 * allowed to eat deeper into the free page list. 2565 */ 2566 if (curthread->td_flags & TDF_SYSTHREAD) 2567 page_req |= VM_ALLOC_SYSTEM; 2568 2569 /* 2570 * Impose various limitations. Note that the v_free_reserved test 2571 * must match the opposite of vm_page_count_target() to avoid 2572 * livelocks, be careful. 2573 */ 2574 loop: 2575 gd = mycpu; 2576 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 2577 ((page_req & VM_ALLOC_INTERRUPT) && 2578 gd->gd_vmstats.v_free_count > 0) || 2579 ((page_req & VM_ALLOC_SYSTEM) && 2580 gd->gd_vmstats.v_cache_count == 0 && 2581 gd->gd_vmstats.v_free_count > 2582 gd->gd_vmstats.v_interrupt_free_min) 2583 ) { 2584 /* 2585 * The free queue has sufficient free pages to take one out. 2586 * 2587 * However, if the free queue is strained the scan may widen 2588 * to the entire queue and cause a great deal of SMP 2589 * contention, so we use a double-queue-scan if we can 2590 * to avoid this. 2591 */ 2592 if (page_req & VM_ALLOC_NORMAL) { 2593 m = vm_page_select_free_or_cache(pg_color, &fromcache); 2594 if (m && fromcache) 2595 goto found_cache; 2596 } else { 2597 m = vm_page_select_free(pg_color); 2598 } 2599 } else if (page_req & VM_ALLOC_NORMAL) { 2600 /* 2601 * Allocatable from the cache (non-interrupt only). On 2602 * success, we must free the page and try again, thus 2603 * ensuring that vmstats.v_*_free_min counters are replenished. 2604 */ 2605 #ifdef INVARIANTS 2606 if (curthread->td_preempted) { 2607 kprintf("vm_page_alloc(): warning, attempt to allocate" 2608 " cache page from preempting interrupt\n"); 2609 m = NULL; 2610 } else { 2611 m = vm_page_select_cache(pg_color); 2612 } 2613 #else 2614 m = vm_page_select_cache(pg_color); 2615 #endif 2616 /* 2617 * On success move the page into the free queue and loop. 2618 * 2619 * Only do this if we can safely acquire the vm_object lock, 2620 * because this is effectively a random page and the caller 2621 * might be holding the lock shared, we don't want to 2622 * deadlock. 2623 */ 2624 if (m != NULL) { 2625 found_cache: 2626 KASSERT(m->dirty == 0, 2627 ("Found dirty cache page %p", m)); 2628 if (__predict_true((m->flags & 2629 (PG_MAPPED|PG_WRITEABLE)) == 0)) { 2630 vm_page_free(m); 2631 } else if ((obj = m->object) != NULL) { 2632 if (vm_object_hold_try(obj)) { 2633 vm_page_protect(m, VM_PROT_NONE); 2634 vm_page_free(m); 2635 /* m->object NULL here */ 2636 vm_object_drop(obj); 2637 } else { 2638 vm_page_deactivate(m); 2639 vm_page_wakeup(m); 2640 } 2641 } else { 2642 vm_page_protect(m, VM_PROT_NONE); 2643 vm_page_free(m); 2644 } 2645 goto loop; 2646 } 2647 2648 /* 2649 * On failure return NULL 2650 */ 2651 atomic_add_int(&vm_pageout_deficit, 1); 2652 pagedaemon_wakeup(); 2653 return (NULL); 2654 } else { 2655 /* 2656 * No pages available, wakeup the pageout daemon and give up. 2657 */ 2658 atomic_add_int(&vm_pageout_deficit, 1); 2659 pagedaemon_wakeup(); 2660 return (NULL); 2661 } 2662 2663 /* 2664 * v_free_count can race so loop if we don't find the expected 2665 * page. 2666 */ 2667 if (m == NULL) { 2668 vmstats_rollup(); 2669 goto loop; 2670 } 2671 2672 /* 2673 * Good page found. The page has already been busied for us and 2674 * removed from its queues. 2675 */ 2676 KASSERT(m->dirty == 0, 2677 ("vm_page_alloc: free/cache page %p was dirty", m)); 2678 KKASSERT(m->queue == PQ_NONE); 2679 2680 #if 0 2681 done: 2682 #endif 2683 /* 2684 * Initialize the structure, inheriting some flags but clearing 2685 * all the rest. The page has already been busied for us. 2686 */ 2687 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 2688 2689 KKASSERT(m->wire_count == 0); 2690 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 2691 m->act_count = 0; 2692 m->valid = 0; 2693 2694 /* 2695 * Caller must be holding the object lock (asserted by 2696 * vm_page_insert()). 2697 * 2698 * NOTE: Inserting a page here does not insert it into any pmaps 2699 * (which could cause us to block allocating memory). 2700 * 2701 * NOTE: If no object an unassociated page is allocated, m->pindex 2702 * can be used by the caller for any purpose. 2703 */ 2704 if (object) { 2705 if (vm_page_insert(m, object, pindex) == FALSE) { 2706 vm_page_free(m); 2707 if ((page_req & VM_ALLOC_NULL_OK) == 0) 2708 panic("PAGE RACE %p[%ld]/%p", 2709 object, (long)pindex, m); 2710 m = NULL; 2711 } 2712 } else { 2713 m->pindex = pindex; 2714 } 2715 2716 /* 2717 * Don't wakeup too often - wakeup the pageout daemon when 2718 * we would be nearly out of memory. 2719 */ 2720 pagedaemon_wakeup(); 2721 2722 /* 2723 * A BUSY page is returned. 2724 */ 2725 return (m); 2726 } 2727 2728 /* 2729 * Returns number of pages available in our DMA memory reserve 2730 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 2731 */ 2732 vm_size_t 2733 vm_contig_avail_pages(void) 2734 { 2735 alist_blk_t blk; 2736 alist_blk_t count; 2737 alist_blk_t bfree; 2738 spin_lock(&vm_contig_spin); 2739 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 2740 spin_unlock(&vm_contig_spin); 2741 2742 return bfree; 2743 } 2744 2745 /* 2746 * Attempt to allocate contiguous physical memory with the specified 2747 * requirements. 2748 */ 2749 vm_page_t 2750 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 2751 unsigned long alignment, unsigned long boundary, 2752 unsigned long size, vm_memattr_t memattr) 2753 { 2754 alist_blk_t blk; 2755 vm_page_t m; 2756 vm_pindex_t i; 2757 #if 0 2758 static vm_pindex_t contig_rover; 2759 #endif 2760 2761 alignment >>= PAGE_SHIFT; 2762 if (alignment == 0) 2763 alignment = 1; 2764 boundary >>= PAGE_SHIFT; 2765 if (boundary == 0) 2766 boundary = 1; 2767 size = (size + PAGE_MASK) >> PAGE_SHIFT; 2768 2769 #if 0 2770 /* 2771 * Disabled temporarily until we find a solution for DRM (a flag 2772 * to always use the free space reserve, for performance). 2773 */ 2774 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 2775 boundary <= PAGE_SIZE && size == 1 && 2776 memattr == VM_MEMATTR_DEFAULT) { 2777 /* 2778 * Any page will work, use vm_page_alloc() 2779 * (e.g. when used from kmem_alloc_attr()) 2780 */ 2781 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 2782 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2783 VM_ALLOC_INTERRUPT); 2784 m->valid = VM_PAGE_BITS_ALL; 2785 vm_page_wire(m); 2786 vm_page_wakeup(m); 2787 } else 2788 #endif 2789 { 2790 /* 2791 * Use the low-memory dma reserve 2792 */ 2793 spin_lock(&vm_contig_spin); 2794 blk = alist_alloc(&vm_contig_alist, 0, size); 2795 if (blk == ALIST_BLOCK_NONE) { 2796 spin_unlock(&vm_contig_spin); 2797 if (bootverbose) { 2798 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2799 (size << PAGE_SHIFT) / 1024); 2800 print_backtrace(5); 2801 } 2802 return(NULL); 2803 } 2804 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2805 alist_free(&vm_contig_alist, blk, size); 2806 spin_unlock(&vm_contig_spin); 2807 if (bootverbose) { 2808 kprintf("vm_page_alloc_contig: %ldk high " 2809 "%016jx failed\n", 2810 (size << PAGE_SHIFT) / 1024, 2811 (intmax_t)high); 2812 } 2813 return(NULL); 2814 } 2815 spin_unlock(&vm_contig_spin); 2816 2817 /* 2818 * Base vm_page_t of range 2819 */ 2820 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2821 } 2822 if (vm_contig_verbose) { 2823 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2824 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2825 (intmax_t)m->phys_addr, 2826 (size << PAGE_SHIFT) / 1024, 2827 low, high, alignment, boundary, size, memattr); 2828 } 2829 if (memattr != VM_MEMATTR_DEFAULT) { 2830 for (i = 0; i < size; ++i) { 2831 KKASSERT(m[i].flags & PG_FICTITIOUS); 2832 pmap_page_set_memattr(&m[i], memattr); 2833 } 2834 } 2835 return m; 2836 } 2837 2838 /* 2839 * Free contiguously allocated pages. The pages will be wired but not busy. 2840 * When freeing to the alist we leave them wired and not busy. 2841 */ 2842 void 2843 vm_page_free_contig(vm_page_t m, unsigned long size) 2844 { 2845 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2846 vm_pindex_t start = pa >> PAGE_SHIFT; 2847 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2848 2849 if (vm_contig_verbose) { 2850 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2851 (intmax_t)pa, size / 1024); 2852 } 2853 if (pa < vm_low_phys_reserved) { 2854 /* 2855 * Just assert check the first page for convenience. 2856 */ 2857 KKASSERT(m->wire_count == 1); 2858 KKASSERT(m->flags & PG_FICTITIOUS); 2859 KKASSERT(pa + size <= vm_low_phys_reserved); 2860 spin_lock(&vm_contig_spin); 2861 alist_free(&vm_contig_alist, start, pages); 2862 spin_unlock(&vm_contig_spin); 2863 } else { 2864 while (pages) { 2865 /* XXX FUTURE, maybe (pair with vm_pg_contig_alloc()) */ 2866 /*vm_page_flag_clear(m, PG_FICTITIOUS | PG_UNQUEUED);*/ 2867 vm_page_busy_wait(m, FALSE, "cpgfr"); 2868 vm_page_unwire(m, 0); 2869 vm_page_free(m); 2870 --pages; 2871 ++m; 2872 } 2873 2874 } 2875 } 2876 2877 2878 /* 2879 * Wait for sufficient free memory for nominal heavy memory use kernel 2880 * operations. 2881 * 2882 * WARNING! Be sure never to call this in any vm_pageout code path, which 2883 * will trivially deadlock the system. 2884 */ 2885 void 2886 vm_wait_nominal(void) 2887 { 2888 while (vm_page_count_min(0)) 2889 vm_wait(0); 2890 } 2891 2892 /* 2893 * Test if vm_wait_nominal() would block. 2894 */ 2895 int 2896 vm_test_nominal(void) 2897 { 2898 if (vm_page_count_min(0)) 2899 return(1); 2900 return(0); 2901 } 2902 2903 /* 2904 * Block until free pages are available for allocation, called in various 2905 * places before memory allocations. 2906 * 2907 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2908 * more generous then that. 2909 */ 2910 void 2911 vm_wait(int timo) 2912 { 2913 /* 2914 * never wait forever 2915 */ 2916 if (timo == 0) 2917 timo = hz; 2918 lwkt_gettoken(&vm_token); 2919 2920 if (curthread == pagethread || 2921 curthread == emergpager) { 2922 /* 2923 * The pageout daemon itself needs pages, this is bad. 2924 */ 2925 if (vm_page_count_min(0)) { 2926 vm_pageout_pages_needed = 1; 2927 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2928 } 2929 } else { 2930 /* 2931 * Wakeup the pageout daemon if necessary and wait. 2932 * 2933 * Do not wait indefinitely for the target to be reached, 2934 * as load might prevent it from being reached any time soon. 2935 * But wait a little to try to slow down page allocations 2936 * and to give more important threads (the pagedaemon) 2937 * allocation priority. 2938 */ 2939 if (vm_page_count_target()) { 2940 if (vm_pages_needed <= 1) { 2941 ++vm_pages_needed; 2942 wakeup(&vm_pages_needed); 2943 } 2944 ++vm_pages_waiting; /* SMP race ok */ 2945 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2946 } 2947 } 2948 lwkt_reltoken(&vm_token); 2949 } 2950 2951 /* 2952 * Block until free pages are available for allocation 2953 * 2954 * Called only from vm_fault so that processes page faulting can be 2955 * easily tracked. 2956 */ 2957 void 2958 vm_wait_pfault(void) 2959 { 2960 /* 2961 * Wakeup the pageout daemon if necessary and wait. 2962 * 2963 * Do not wait indefinitely for the target to be reached, 2964 * as load might prevent it from being reached any time soon. 2965 * But wait a little to try to slow down page allocations 2966 * and to give more important threads (the pagedaemon) 2967 * allocation priority. 2968 */ 2969 if (vm_page_count_min(0)) { 2970 lwkt_gettoken(&vm_token); 2971 while (vm_page_count_severe()) { 2972 if (vm_page_count_target()) { 2973 thread_t td; 2974 2975 if (vm_pages_needed <= 1) { 2976 ++vm_pages_needed; 2977 wakeup(&vm_pages_needed); 2978 } 2979 ++vm_pages_waiting; /* SMP race ok */ 2980 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2981 2982 /* 2983 * Do not stay stuck in the loop if the system is trying 2984 * to kill the process. 2985 */ 2986 td = curthread; 2987 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2988 break; 2989 } 2990 } 2991 lwkt_reltoken(&vm_token); 2992 } 2993 } 2994 2995 /* 2996 * Put the specified page on the active list (if appropriate). Ensure 2997 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2998 * 2999 * The caller should be holding the page busied ? XXX 3000 * This routine may not block. 3001 * 3002 * It is ok if the page is wired (so buffer cache operations don't have 3003 * to mess with the page queues). 3004 */ 3005 void 3006 vm_page_activate(vm_page_t m) 3007 { 3008 u_short oqueue; 3009 3010 /* 3011 * If already active or inappropriate, just set act_count and 3012 * return. We don't have to spin-lock the page. 3013 */ 3014 if (m->queue - m->pc == PQ_ACTIVE || 3015 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 3016 if (m->act_count < ACT_INIT) 3017 m->act_count = ACT_INIT; 3018 return; 3019 } 3020 3021 vm_page_spin_lock(m); 3022 if (m->queue - m->pc != PQ_ACTIVE && 3023 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 3024 _vm_page_queue_spin_lock(m); 3025 oqueue = _vm_page_rem_queue_spinlocked(m); 3026 /* page is left spinlocked, queue is unlocked */ 3027 3028 if (oqueue == PQ_CACHE) 3029 mycpu->gd_cnt.v_reactivated++; 3030 if (m->act_count < ACT_INIT) 3031 m->act_count = ACT_INIT; 3032 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 3033 _vm_page_and_queue_spin_unlock(m); 3034 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 3035 pagedaemon_wakeup(); 3036 } else { 3037 if (m->act_count < ACT_INIT) 3038 m->act_count = ACT_INIT; 3039 vm_page_spin_unlock(m); 3040 } 3041 } 3042 3043 void 3044 vm_page_soft_activate(vm_page_t m) 3045 { 3046 if (m->queue - m->pc == PQ_ACTIVE || 3047 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 3048 if (m->act_count < ACT_INIT) 3049 m->act_count = ACT_INIT; 3050 } else { 3051 vm_page_activate(m); 3052 } 3053 } 3054 3055 /* 3056 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 3057 * routine is called when a page has been added to the cache or free 3058 * queues. 3059 * 3060 * This routine may not block. 3061 */ 3062 static __inline void 3063 vm_page_free_wakeup(void) 3064 { 3065 globaldata_t gd = mycpu; 3066 3067 /* 3068 * If the pageout daemon itself needs pages, then tell it that 3069 * there are some free. 3070 */ 3071 if (vm_pageout_pages_needed && 3072 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 3073 gd->gd_vmstats.v_pageout_free_min 3074 ) { 3075 vm_pageout_pages_needed = 0; 3076 wakeup(&vm_pageout_pages_needed); 3077 } 3078 3079 /* 3080 * Wakeup processes that are waiting on memory. 3081 * 3082 * Generally speaking we want to wakeup stuck processes as soon as 3083 * possible. !vm_page_count_min(0) is the absolute minimum point 3084 * where we can do this. Wait a bit longer to reduce degenerate 3085 * re-blocking (vm_page_free_hysteresis). The target check is just 3086 * to make sure the min-check w/hysteresis does not exceed the 3087 * normal target. 3088 */ 3089 if (vm_pages_waiting) { 3090 if (!vm_page_count_min(vm_page_free_hysteresis) || 3091 !vm_page_count_target()) { 3092 vm_pages_waiting = 0; 3093 wakeup(&vmstats.v_free_count); 3094 ++mycpu->gd_cnt.v_ppwakeups; 3095 } 3096 #if 0 3097 if (!vm_page_count_target()) { 3098 /* 3099 * Plenty of pages are free, wakeup everyone. 3100 */ 3101 vm_pages_waiting = 0; 3102 wakeup(&vmstats.v_free_count); 3103 ++mycpu->gd_cnt.v_ppwakeups; 3104 } else if (!vm_page_count_min(0)) { 3105 /* 3106 * Some pages are free, wakeup someone. 3107 */ 3108 int wcount = vm_pages_waiting; 3109 if (wcount > 0) 3110 --wcount; 3111 vm_pages_waiting = wcount; 3112 wakeup_one(&vmstats.v_free_count); 3113 ++mycpu->gd_cnt.v_ppwakeups; 3114 } 3115 #endif 3116 } 3117 } 3118 3119 /* 3120 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 3121 * it from its VM object. 3122 * 3123 * The vm_page must be BUSY on entry. BUSY will be released on 3124 * return (the page will have been freed). 3125 */ 3126 void 3127 vm_page_free_toq(vm_page_t m) 3128 { 3129 /* 3130 * The page must not be mapped when freed, but we may have to call 3131 * pmap_mapped_sync() to validate this. 3132 */ 3133 mycpu->gd_cnt.v_tfree++; 3134 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) 3135 pmap_mapped_sync(m); 3136 KKASSERT((m->flags & PG_MAPPED) == 0); 3137 KKASSERT(m->busy_count & PBUSY_LOCKED); 3138 3139 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 3140 kprintf("vm_page_free: pindex(%lu), busy %08x, " 3141 "hold(%d)\n", 3142 (u_long)m->pindex, m->busy_count, m->hold_count); 3143 if ((m->queue - m->pc) == PQ_FREE) 3144 panic("vm_page_free: freeing free page"); 3145 else 3146 panic("vm_page_free: freeing busy page"); 3147 } 3148 3149 /* 3150 * Remove from object, spinlock the page and its queues and 3151 * remove from any queue. No queue spinlock will be held 3152 * after this section (because the page was removed from any 3153 * queue). 3154 */ 3155 vm_page_remove(m); 3156 3157 /* 3158 * No further management of fictitious pages occurs beyond object 3159 * and queue removal. 3160 */ 3161 if ((m->flags & PG_FICTITIOUS) != 0) { 3162 KKASSERT(m->queue == PQ_NONE); 3163 vm_page_wakeup(m); 3164 return; 3165 } 3166 vm_page_and_queue_spin_lock(m); 3167 _vm_page_rem_queue_spinlocked(m); 3168 3169 m->valid = 0; 3170 vm_page_undirty(m); 3171 3172 if (m->wire_count != 0) { 3173 if (m->wire_count > 1) { 3174 panic( 3175 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 3176 m->wire_count, (long)m->pindex); 3177 } 3178 panic("vm_page_free: freeing wired page"); 3179 } 3180 3181 if (!MD_PAGE_FREEABLE(m)) 3182 panic("vm_page_free: page %p is still mapped!", m); 3183 3184 /* 3185 * Clear the PG_NEED_COMMIT and the PG_UNQUEUED flags. The 3186 * page returns to normal operation and will be placed in 3187 * the PQ_HOLD or PQ_FREE queue. 3188 */ 3189 vm_page_flag_clear(m, PG_NEED_COMMIT | PG_UNQUEUED); 3190 3191 if (m->hold_count != 0) { 3192 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 3193 } else { 3194 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 3195 } 3196 3197 /* 3198 * This sequence allows us to clear BUSY while still holding 3199 * its spin lock, which reduces contention vs allocators. We 3200 * must not leave the queue locked or _vm_page_wakeup() may 3201 * deadlock. 3202 */ 3203 _vm_page_queue_spin_unlock(m); 3204 if (_vm_page_wakeup(m)) { 3205 vm_page_spin_unlock(m); 3206 wakeup(m); 3207 } else { 3208 vm_page_spin_unlock(m); 3209 } 3210 vm_page_free_wakeup(); 3211 } 3212 3213 /* 3214 * Mark this page as wired down by yet another map. We do not adjust the 3215 * queue the page is on, it will be checked for wiring as-needed. 3216 * 3217 * This function has no effect on fictitious pages. 3218 * 3219 * Caller must be holding the page busy. 3220 */ 3221 void 3222 vm_page_wire(vm_page_t m) 3223 { 3224 KKASSERT(m->busy_count & PBUSY_LOCKED); 3225 if ((m->flags & PG_FICTITIOUS) == 0) { 3226 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 3227 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 3228 } 3229 KASSERT(m->wire_count != 0, 3230 ("vm_page_wire: wire_count overflow m=%p", m)); 3231 } 3232 } 3233 3234 /* 3235 * Release one wiring of this page, potentially enabling it to be paged again. 3236 * 3237 * Note that wired pages are no longer unconditionally removed from the 3238 * paging queues, so the page may already be on a queue. Move the page 3239 * to the desired queue if necessary. 3240 * 3241 * Many pages placed on the inactive queue should actually go 3242 * into the cache, but it is difficult to figure out which. What 3243 * we do instead, if the inactive target is well met, is to put 3244 * clean pages at the head of the inactive queue instead of the tail. 3245 * This will cause them to be moved to the cache more quickly and 3246 * if not actively re-referenced, freed more quickly. If we just 3247 * stick these pages at the end of the inactive queue, heavy filesystem 3248 * meta-data accesses can cause an unnecessary paging load on memory bound 3249 * processes. This optimization causes one-time-use metadata to be 3250 * reused more quickly. 3251 * 3252 * Pages marked PG_NEED_COMMIT are always activated and never placed on 3253 * the inactive queue. This helps the pageout daemon determine memory 3254 * pressure and act on out-of-memory situations more quickly. 3255 * 3256 * BUT, if we are in a low-memory situation we have no choice but to 3257 * put clean pages on the cache queue. 3258 * 3259 * A number of routines use vm_page_unwire() to guarantee that the page 3260 * will go into either the inactive or active queues, and will NEVER 3261 * be placed in the cache - for example, just after dirtying a page. 3262 * dirty pages in the cache are not allowed. 3263 * 3264 * PG_FICTITIOUS or PG_UNQUEUED pages are never moved to any queue, and 3265 * the wire_count will not be adjusted in any way for a PG_FICTITIOUS 3266 * page. 3267 * 3268 * This routine may not block. 3269 */ 3270 void 3271 vm_page_unwire(vm_page_t m, int activate) 3272 { 3273 KKASSERT(m->busy_count & PBUSY_LOCKED); 3274 if (m->flags & PG_FICTITIOUS) { 3275 /* do nothing */ 3276 } else if ((int)m->wire_count <= 0) { 3277 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 3278 } else { 3279 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 3280 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 3281 if (m->flags & PG_UNQUEUED) { 3282 ; 3283 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 3284 vm_page_activate(m); 3285 } else { 3286 vm_page_deactivate(m); 3287 } 3288 } 3289 } 3290 } 3291 3292 /* 3293 * Move the specified page to the inactive queue. 3294 * 3295 * Normally athead is 0 resulting in LRU operation. athead is set 3296 * to 1 if we want this page to be 'as if it were placed in the cache', 3297 * except without unmapping it from the process address space. 3298 * 3299 * vm_page's spinlock must be held on entry and will remain held on return. 3300 * This routine may not block. The caller does not have to hold the page 3301 * busied but should have some sort of interlock on its validity. 3302 * 3303 * It is ok if the page is wired (so buffer cache operations don't have 3304 * to mess with the page queues). 3305 */ 3306 static void 3307 _vm_page_deactivate_locked(vm_page_t m, int athead) 3308 { 3309 u_short oqueue; 3310 3311 /* 3312 * Ignore if already inactive. 3313 */ 3314 if (m->queue - m->pc == PQ_INACTIVE || 3315 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 3316 return; 3317 } 3318 3319 _vm_page_queue_spin_lock(m); 3320 oqueue = _vm_page_rem_queue_spinlocked(m); 3321 3322 if ((m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 3323 if (oqueue == PQ_CACHE) 3324 mycpu->gd_cnt.v_reactivated++; 3325 vm_page_flag_clear(m, PG_WINATCFLS); 3326 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 3327 if (athead == 0) { 3328 atomic_add_long( 3329 &vm_page_queues[PQ_INACTIVE + m->pc].adds, 1); 3330 } 3331 } 3332 /* NOTE: PQ_NONE if condition not taken */ 3333 _vm_page_queue_spin_unlock(m); 3334 /* leaves vm_page spinlocked */ 3335 } 3336 3337 /* 3338 * Attempt to deactivate a page. 3339 * 3340 * No requirements. We can pre-filter before getting the spinlock. 3341 * 3342 * It is ok if the page is wired (so buffer cache operations don't have 3343 * to mess with the page queues). 3344 */ 3345 void 3346 vm_page_deactivate(vm_page_t m) 3347 { 3348 if (m->queue - m->pc != PQ_INACTIVE && 3349 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 3350 vm_page_spin_lock(m); 3351 _vm_page_deactivate_locked(m, 0); 3352 vm_page_spin_unlock(m); 3353 } 3354 } 3355 3356 void 3357 vm_page_deactivate_locked(vm_page_t m) 3358 { 3359 _vm_page_deactivate_locked(m, 0); 3360 } 3361 3362 /* 3363 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 3364 * 3365 * This function returns non-zero if it successfully moved the page to 3366 * PQ_CACHE. 3367 * 3368 * This function unconditionally unbusies the page on return. 3369 */ 3370 int 3371 vm_page_try_to_cache(vm_page_t m) 3372 { 3373 /* 3374 * Shortcut if we obviously cannot move the page, or if the 3375 * page is already on the cache queue, or it is ficitious. 3376 * 3377 * Never allow a wired page into the cache. 3378 */ 3379 if (m->dirty || m->hold_count || m->wire_count || 3380 m->queue - m->pc == PQ_CACHE || 3381 (m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS))) { 3382 vm_page_wakeup(m); 3383 return(0); 3384 } 3385 3386 /* 3387 * Page busied by us and no longer spinlocked. Dirty pages cannot 3388 * be moved to the cache, but can be deactivated. However, users 3389 * of this function want to move pages closer to the cache so we 3390 * only deactivate it if it is in PQ_ACTIVE. We do not re-deactivate. 3391 */ 3392 vm_page_test_dirty(m); 3393 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3394 if (m->queue - m->pc == PQ_ACTIVE) 3395 vm_page_deactivate(m); 3396 vm_page_wakeup(m); 3397 return(0); 3398 } 3399 vm_page_cache(m); 3400 return(1); 3401 } 3402 3403 /* 3404 * Attempt to free the page. If we cannot free it, we do nothing. 3405 * 1 is returned on success, 0 on failure. 3406 * 3407 * The page can be in any state, including already being on the free 3408 * queue. Check to see if it really can be freed. Note that we disallow 3409 * this ad-hoc operation if the page is flagged PG_UNQUEUED. 3410 * 3411 * Caller provides an unlocked/non-busied page. 3412 * No requirements. 3413 */ 3414 int 3415 vm_page_try_to_free(vm_page_t m) 3416 { 3417 if (vm_page_busy_try(m, TRUE)) 3418 return(0); 3419 3420 if (m->dirty || /* can't free if it is dirty */ 3421 m->hold_count || /* or held (XXX may be wrong) */ 3422 m->wire_count || /* or wired */ 3423 (m->flags & (PG_UNQUEUED | /* or unqueued */ 3424 PG_NEED_COMMIT | /* or needs a commit */ 3425 PG_FICTITIOUS)) || /* or is fictitious */ 3426 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 3427 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 3428 vm_page_wakeup(m); 3429 return(0); 3430 } 3431 3432 /* 3433 * We can probably free the page. 3434 * 3435 * Page busied by us and no longer spinlocked. Dirty pages will 3436 * not be freed by this function. We have to re-test the 3437 * dirty bit after cleaning out the pmaps. 3438 */ 3439 vm_page_test_dirty(m); 3440 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3441 vm_page_wakeup(m); 3442 return(0); 3443 } 3444 vm_page_protect(m, VM_PROT_NONE); 3445 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3446 vm_page_wakeup(m); 3447 return(0); 3448 } 3449 vm_page_free(m); 3450 return(1); 3451 } 3452 3453 /* 3454 * vm_page_cache 3455 * 3456 * Put the specified page onto the page cache queue (if appropriate). 3457 * 3458 * The page must be busy, and this routine will release the busy and 3459 * possibly even free the page. 3460 */ 3461 void 3462 vm_page_cache(vm_page_t m) 3463 { 3464 /* 3465 * Not suitable for the cache 3466 */ 3467 if ((m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS)) || 3468 (m->busy_count & PBUSY_MASK) || 3469 m->wire_count || m->hold_count) { 3470 vm_page_wakeup(m); 3471 return; 3472 } 3473 3474 /* 3475 * Already in the cache (and thus not mapped) 3476 */ 3477 if ((m->queue - m->pc) == PQ_CACHE) { 3478 KKASSERT((m->flags & PG_MAPPED) == 0); 3479 vm_page_wakeup(m); 3480 return; 3481 } 3482 3483 #if 0 3484 /* 3485 * REMOVED - it is possible for dirty to get set at any time as 3486 * long as the page is still mapped and writeable. 3487 * 3488 * Caller is required to test m->dirty, but note that the act of 3489 * removing the page from its maps can cause it to become dirty 3490 * on an SMP system due to another cpu running in usermode. 3491 */ 3492 if (m->dirty) { 3493 panic("vm_page_cache: caching a dirty page, pindex: %ld", 3494 (long)m->pindex); 3495 } 3496 #endif 3497 3498 /* 3499 * Remove all pmaps and indicate that the page is not 3500 * writeable or mapped. Our vm_page_protect() call may 3501 * have blocked (especially w/ VM_PROT_NONE), so recheck 3502 * everything. 3503 */ 3504 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) { 3505 vm_page_protect(m, VM_PROT_NONE); 3506 pmap_mapped_sync(m); 3507 } 3508 if ((m->flags & (PG_UNQUEUED | PG_MAPPED)) || 3509 (m->busy_count & PBUSY_MASK) || 3510 m->wire_count || m->hold_count) { 3511 vm_page_wakeup(m); 3512 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3513 vm_page_deactivate(m); 3514 vm_page_wakeup(m); 3515 } else { 3516 _vm_page_and_queue_spin_lock(m); 3517 _vm_page_rem_queue_spinlocked(m); 3518 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 3519 _vm_page_and_queue_spin_unlock(m); 3520 vm_page_wakeup(m); 3521 vm_page_free_wakeup(); 3522 } 3523 } 3524 3525 /* 3526 * vm_page_dontneed() 3527 * 3528 * Cache, deactivate, or do nothing as appropriate. This routine 3529 * is typically used by madvise() MADV_DONTNEED. 3530 * 3531 * Generally speaking we want to move the page into the cache so 3532 * it gets reused quickly. However, this can result in a silly syndrome 3533 * due to the page recycling too quickly. Small objects will not be 3534 * fully cached. On the otherhand, if we move the page to the inactive 3535 * queue we wind up with a problem whereby very large objects 3536 * unnecessarily blow away our inactive and cache queues. 3537 * 3538 * The solution is to move the pages based on a fixed weighting. We 3539 * either leave them alone, deactivate them, or move them to the cache, 3540 * where moving them to the cache has the highest weighting. 3541 * By forcing some pages into other queues we eventually force the 3542 * system to balance the queues, potentially recovering other unrelated 3543 * space from active. The idea is to not force this to happen too 3544 * often. 3545 * 3546 * The page must be busied. 3547 */ 3548 void 3549 vm_page_dontneed(vm_page_t m) 3550 { 3551 static int dnweight; 3552 int dnw; 3553 int head; 3554 3555 dnw = ++dnweight; 3556 3557 /* 3558 * occassionally leave the page alone 3559 */ 3560 if ((dnw & 0x01F0) == 0 || 3561 m->queue - m->pc == PQ_INACTIVE || 3562 m->queue - m->pc == PQ_CACHE 3563 ) { 3564 if (m->act_count >= ACT_INIT) 3565 --m->act_count; 3566 return; 3567 } 3568 3569 /* 3570 * If vm_page_dontneed() is inactivating a page, it must clear 3571 * the referenced flag; otherwise the pagedaemon will see references 3572 * on the page in the inactive queue and reactivate it. Until the 3573 * page can move to the cache queue, madvise's job is not done. 3574 */ 3575 vm_page_flag_clear(m, PG_REFERENCED); 3576 pmap_clear_reference(m); 3577 3578 if (m->dirty == 0) 3579 vm_page_test_dirty(m); 3580 3581 if (m->dirty || (dnw & 0x0070) == 0) { 3582 /* 3583 * Deactivate the page 3 times out of 32. 3584 */ 3585 head = 0; 3586 } else { 3587 /* 3588 * Cache the page 28 times out of every 32. Note that 3589 * the page is deactivated instead of cached, but placed 3590 * at the head of the queue instead of the tail. 3591 */ 3592 head = 1; 3593 } 3594 vm_page_spin_lock(m); 3595 _vm_page_deactivate_locked(m, head); 3596 vm_page_spin_unlock(m); 3597 } 3598 3599 /* 3600 * These routines manipulate the 'soft busy' count for a page. A soft busy 3601 * is almost like a hard BUSY except that it allows certain compatible 3602 * operations to occur on the page while it is busy. For example, a page 3603 * undergoing a write can still be mapped read-only. 3604 * 3605 * We also use soft-busy to quickly pmap_enter shared read-only pages 3606 * without having to hold the page locked. 3607 * 3608 * The soft-busy count can be > 1 in situations where multiple threads 3609 * are pmap_enter()ing the same page simultaneously, or when two buffer 3610 * cache buffers overlap the same page. 3611 * 3612 * The caller must hold the page BUSY when making these two calls. 3613 */ 3614 void 3615 vm_page_io_start(vm_page_t m) 3616 { 3617 uint32_t ocount; 3618 3619 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3620 KKASSERT(ocount & PBUSY_LOCKED); 3621 } 3622 3623 void 3624 vm_page_io_finish(vm_page_t m) 3625 { 3626 uint32_t ocount; 3627 3628 ocount = atomic_fetchadd_int(&m->busy_count, -1); 3629 KKASSERT(ocount & PBUSY_MASK); 3630 #if 0 3631 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 3632 wakeup(m); 3633 #endif 3634 } 3635 3636 /* 3637 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 3638 * 3639 * We can't use fetchadd here because we might race a hard-busy and the 3640 * page freeing code asserts on a non-zero soft-busy count (even if only 3641 * temporary). 3642 * 3643 * Returns 0 on success, non-zero on failure. 3644 */ 3645 int 3646 vm_page_sbusy_try(vm_page_t m) 3647 { 3648 uint32_t ocount; 3649 3650 for (;;) { 3651 ocount = m->busy_count; 3652 cpu_ccfence(); 3653 if (ocount & PBUSY_LOCKED) 3654 return 1; 3655 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1)) 3656 break; 3657 } 3658 return 0; 3659 #if 0 3660 if (m->busy_count & PBUSY_LOCKED) 3661 return 1; 3662 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3663 if (ocount & PBUSY_LOCKED) { 3664 vm_page_sbusy_drop(m); 3665 return 1; 3666 } 3667 return 0; 3668 #endif 3669 } 3670 3671 /* 3672 * Indicate that a clean VM page requires a filesystem commit and cannot 3673 * be reused. Used by tmpfs. 3674 */ 3675 void 3676 vm_page_need_commit(vm_page_t m) 3677 { 3678 vm_page_flag_set(m, PG_NEED_COMMIT); 3679 vm_object_set_writeable_dirty(m->object); 3680 } 3681 3682 void 3683 vm_page_clear_commit(vm_page_t m) 3684 { 3685 vm_page_flag_clear(m, PG_NEED_COMMIT); 3686 } 3687 3688 /* 3689 * Grab a page, blocking if it is busy and allocating a page if necessary. 3690 * A busy page is returned or NULL. The page may or may not be valid and 3691 * might not be on a queue (the caller is responsible for the disposition of 3692 * the page). 3693 * 3694 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 3695 * page will be zero'd and marked valid. 3696 * 3697 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 3698 * valid even if it already exists. 3699 * 3700 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 3701 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 3702 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 3703 * 3704 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 3705 * always returned if we had blocked. 3706 * 3707 * This routine may not be called from an interrupt. 3708 * 3709 * No other requirements. 3710 */ 3711 vm_page_t 3712 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3713 { 3714 vm_page_t m; 3715 int error; 3716 int shared = 1; 3717 3718 KKASSERT(allocflags & 3719 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 3720 vm_object_hold_shared(object); 3721 for (;;) { 3722 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3723 if (error) { 3724 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 3725 if ((allocflags & VM_ALLOC_RETRY) == 0) { 3726 m = NULL; 3727 break; 3728 } 3729 /* retry */ 3730 } else if (m == NULL) { 3731 if (shared) { 3732 vm_object_upgrade(object); 3733 shared = 0; 3734 } 3735 if (allocflags & VM_ALLOC_RETRY) 3736 allocflags |= VM_ALLOC_NULL_OK; 3737 m = vm_page_alloc(object, pindex, 3738 allocflags & ~VM_ALLOC_RETRY); 3739 if (m) 3740 break; 3741 vm_wait(0); 3742 if ((allocflags & VM_ALLOC_RETRY) == 0) 3743 goto failed; 3744 } else { 3745 /* m found */ 3746 break; 3747 } 3748 } 3749 3750 /* 3751 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 3752 * 3753 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 3754 * valid even if already valid. 3755 * 3756 * NOTE! We have removed all of the PG_ZERO optimizations and also 3757 * removed the idle zeroing code. These optimizations actually 3758 * slow things down on modern cpus because the zerod area is 3759 * likely uncached, placing a memory-access burden on the 3760 * accesors taking the fault. 3761 * 3762 * By always zeroing the page in-line with the fault, no 3763 * dynamic ram reads are needed and the caches are hot, ready 3764 * for userland to access the memory. 3765 */ 3766 if (m->valid == 0) { 3767 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 3768 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3769 m->valid = VM_PAGE_BITS_ALL; 3770 } 3771 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 3772 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3773 m->valid = VM_PAGE_BITS_ALL; 3774 } 3775 failed: 3776 vm_object_drop(object); 3777 return(m); 3778 } 3779 3780 /* 3781 * Mapping function for valid bits or for dirty bits in 3782 * a page. May not block. 3783 * 3784 * Inputs are required to range within a page. 3785 * 3786 * No requirements. 3787 * Non blocking. 3788 */ 3789 int 3790 vm_page_bits(int base, int size) 3791 { 3792 int first_bit; 3793 int last_bit; 3794 3795 KASSERT( 3796 base + size <= PAGE_SIZE, 3797 ("vm_page_bits: illegal base/size %d/%d", base, size) 3798 ); 3799 3800 if (size == 0) /* handle degenerate case */ 3801 return(0); 3802 3803 first_bit = base >> DEV_BSHIFT; 3804 last_bit = (base + size - 1) >> DEV_BSHIFT; 3805 3806 return ((2 << last_bit) - (1 << first_bit)); 3807 } 3808 3809 /* 3810 * Sets portions of a page valid and clean. The arguments are expected 3811 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3812 * of any partial chunks touched by the range. The invalid portion of 3813 * such chunks will be zero'd. 3814 * 3815 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 3816 * align base to DEV_BSIZE so as not to mark clean a partially 3817 * truncated device block. Otherwise the dirty page status might be 3818 * lost. 3819 * 3820 * This routine may not block. 3821 * 3822 * (base + size) must be less then or equal to PAGE_SIZE. 3823 */ 3824 static void 3825 _vm_page_zero_valid(vm_page_t m, int base, int size) 3826 { 3827 int frag; 3828 int endoff; 3829 3830 if (size == 0) /* handle degenerate case */ 3831 return; 3832 3833 /* 3834 * If the base is not DEV_BSIZE aligned and the valid 3835 * bit is clear, we have to zero out a portion of the 3836 * first block. 3837 */ 3838 3839 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3840 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3841 ) { 3842 pmap_zero_page_area( 3843 VM_PAGE_TO_PHYS(m), 3844 frag, 3845 base - frag 3846 ); 3847 } 3848 3849 /* 3850 * If the ending offset is not DEV_BSIZE aligned and the 3851 * valid bit is clear, we have to zero out a portion of 3852 * the last block. 3853 */ 3854 3855 endoff = base + size; 3856 3857 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3858 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3859 ) { 3860 pmap_zero_page_area( 3861 VM_PAGE_TO_PHYS(m), 3862 endoff, 3863 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3864 ); 3865 } 3866 } 3867 3868 /* 3869 * Set valid, clear dirty bits. If validating the entire 3870 * page we can safely clear the pmap modify bit. We also 3871 * use this opportunity to clear the PG_NOSYNC flag. If a process 3872 * takes a write fault on a MAP_NOSYNC memory area the flag will 3873 * be set again. 3874 * 3875 * We set valid bits inclusive of any overlap, but we can only 3876 * clear dirty bits for DEV_BSIZE chunks that are fully within 3877 * the range. 3878 * 3879 * Page must be busied? 3880 * No other requirements. 3881 */ 3882 void 3883 vm_page_set_valid(vm_page_t m, int base, int size) 3884 { 3885 _vm_page_zero_valid(m, base, size); 3886 m->valid |= vm_page_bits(base, size); 3887 } 3888 3889 3890 /* 3891 * Set valid bits and clear dirty bits. 3892 * 3893 * Page must be busied by caller. 3894 * 3895 * NOTE: This function does not clear the pmap modified bit. 3896 * Also note that e.g. NFS may use a byte-granular base 3897 * and size. 3898 * 3899 * No other requirements. 3900 */ 3901 void 3902 vm_page_set_validclean(vm_page_t m, int base, int size) 3903 { 3904 int pagebits; 3905 3906 _vm_page_zero_valid(m, base, size); 3907 pagebits = vm_page_bits(base, size); 3908 m->valid |= pagebits; 3909 m->dirty &= ~pagebits; 3910 if (base == 0 && size == PAGE_SIZE) { 3911 /*pmap_clear_modify(m);*/ 3912 vm_page_flag_clear(m, PG_NOSYNC); 3913 } 3914 } 3915 3916 /* 3917 * Set valid & dirty. Used by buwrite() 3918 * 3919 * Page must be busied by caller. 3920 */ 3921 void 3922 vm_page_set_validdirty(vm_page_t m, int base, int size) 3923 { 3924 int pagebits; 3925 3926 pagebits = vm_page_bits(base, size); 3927 m->valid |= pagebits; 3928 m->dirty |= pagebits; 3929 if (m->object) 3930 vm_object_set_writeable_dirty(m->object); 3931 } 3932 3933 /* 3934 * Clear dirty bits. 3935 * 3936 * NOTE: This function does not clear the pmap modified bit. 3937 * Also note that e.g. NFS may use a byte-granular base 3938 * and size. 3939 * 3940 * Page must be busied? 3941 * No other requirements. 3942 */ 3943 void 3944 vm_page_clear_dirty(vm_page_t m, int base, int size) 3945 { 3946 m->dirty &= ~vm_page_bits(base, size); 3947 if (base == 0 && size == PAGE_SIZE) { 3948 /*pmap_clear_modify(m);*/ 3949 vm_page_flag_clear(m, PG_NOSYNC); 3950 } 3951 } 3952 3953 /* 3954 * Make the page all-dirty. 3955 * 3956 * Also make sure the related object and vnode reflect the fact that the 3957 * object may now contain a dirty page. 3958 * 3959 * Page must be busied? 3960 * No other requirements. 3961 */ 3962 void 3963 vm_page_dirty(vm_page_t m) 3964 { 3965 #ifdef INVARIANTS 3966 int pqtype = m->queue - m->pc; 3967 #endif 3968 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3969 ("vm_page_dirty: page in free/cache queue!")); 3970 if (m->dirty != VM_PAGE_BITS_ALL) { 3971 m->dirty = VM_PAGE_BITS_ALL; 3972 if (m->object) 3973 vm_object_set_writeable_dirty(m->object); 3974 } 3975 } 3976 3977 /* 3978 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3979 * valid and dirty bits for the effected areas are cleared. 3980 * 3981 * Page must be busied? 3982 * Does not block. 3983 * No other requirements. 3984 */ 3985 void 3986 vm_page_set_invalid(vm_page_t m, int base, int size) 3987 { 3988 int bits; 3989 3990 bits = vm_page_bits(base, size); 3991 m->valid &= ~bits; 3992 m->dirty &= ~bits; 3993 atomic_add_int(&m->object->generation, 1); 3994 } 3995 3996 /* 3997 * The kernel assumes that the invalid portions of a page contain 3998 * garbage, but such pages can be mapped into memory by user code. 3999 * When this occurs, we must zero out the non-valid portions of the 4000 * page so user code sees what it expects. 4001 * 4002 * Pages are most often semi-valid when the end of a file is mapped 4003 * into memory and the file's size is not page aligned. 4004 * 4005 * Page must be busied? 4006 * No other requirements. 4007 */ 4008 void 4009 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4010 { 4011 int b; 4012 int i; 4013 4014 /* 4015 * Scan the valid bits looking for invalid sections that 4016 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 4017 * valid bit may be set ) have already been zerod by 4018 * vm_page_set_validclean(). 4019 */ 4020 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4021 if (i == (PAGE_SIZE / DEV_BSIZE) || 4022 (m->valid & (1 << i)) 4023 ) { 4024 if (i > b) { 4025 pmap_zero_page_area( 4026 VM_PAGE_TO_PHYS(m), 4027 b << DEV_BSHIFT, 4028 (i - b) << DEV_BSHIFT 4029 ); 4030 } 4031 b = i + 1; 4032 } 4033 } 4034 4035 /* 4036 * setvalid is TRUE when we can safely set the zero'd areas 4037 * as being valid. We can do this if there are no cache consistency 4038 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4039 */ 4040 if (setvalid) 4041 m->valid = VM_PAGE_BITS_ALL; 4042 } 4043 4044 /* 4045 * Is a (partial) page valid? Note that the case where size == 0 4046 * will return FALSE in the degenerate case where the page is entirely 4047 * invalid, and TRUE otherwise. 4048 * 4049 * Does not block. 4050 * No other requirements. 4051 */ 4052 int 4053 vm_page_is_valid(vm_page_t m, int base, int size) 4054 { 4055 int bits = vm_page_bits(base, size); 4056 4057 if (m->valid && ((m->valid & bits) == bits)) 4058 return 1; 4059 else 4060 return 0; 4061 } 4062 4063 /* 4064 * Update dirty bits from pmap/mmu. May not block. 4065 * 4066 * Caller must hold the page busy 4067 * 4068 * WARNING! Unless the page has been unmapped, this function only 4069 * provides a likely dirty status. 4070 */ 4071 void 4072 vm_page_test_dirty(vm_page_t m) 4073 { 4074 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) { 4075 vm_page_dirty(m); 4076 } 4077 } 4078 4079 #include "opt_ddb.h" 4080 #ifdef DDB 4081 #include <ddb/ddb.h> 4082 4083 DB_SHOW_COMMAND(page, vm_page_print_page_info) 4084 { 4085 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 4086 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 4087 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 4088 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 4089 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 4090 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 4091 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 4092 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 4093 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 4094 db_printf("vmstats.v_inactive_target: %ld\n", 4095 vmstats.v_inactive_target); 4096 } 4097 4098 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 4099 { 4100 int i; 4101 db_printf("PQ_FREE:"); 4102 for (i = 0; i < PQ_L2_SIZE; i++) { 4103 db_printf(" %ld", vm_page_queues[PQ_FREE + i].lcnt); 4104 } 4105 db_printf("\n"); 4106 4107 db_printf("PQ_CACHE:"); 4108 for(i = 0; i < PQ_L2_SIZE; i++) { 4109 db_printf(" %ld", vm_page_queues[PQ_CACHE + i].lcnt); 4110 } 4111 db_printf("\n"); 4112 4113 db_printf("PQ_ACTIVE:"); 4114 for(i = 0; i < PQ_L2_SIZE; i++) { 4115 db_printf(" %ld", vm_page_queues[PQ_ACTIVE + i].lcnt); 4116 } 4117 db_printf("\n"); 4118 4119 db_printf("PQ_INACTIVE:"); 4120 for(i = 0; i < PQ_L2_SIZE; i++) { 4121 db_printf(" %ld", vm_page_queues[PQ_INACTIVE + i].lcnt); 4122 } 4123 db_printf("\n"); 4124 } 4125 #endif /* DDB */ 4126