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 struct vm_page_hash_elm { 104 vm_page_t m; 105 int ticks; 106 int unused01; 107 }; 108 109 #define VM_PAGE_HASH_SET 4 /* power of 2, set-assoc */ 110 #define VM_PAGE_HASH_MAX (1024 * 1024) /* power of 2, max size */ 111 112 /* 113 * SET - Minimum required set associative size, must be a power of 2. We 114 * want this to match or exceed the set-associativeness of the cpu, 115 * up to a reasonable limit (we will use 16). 116 */ 117 __read_mostly static int set_assoc_mask = 16 - 1; 118 119 static void vm_page_queue_init(void); 120 static void vm_page_free_wakeup(void); 121 static vm_page_t vm_page_select_cache(u_short pg_color); 122 static vm_page_t _vm_page_list_find2(int basequeue, int index, int *lastp); 123 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 124 static void vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes); 125 126 /* 127 * Array of tailq lists 128 */ 129 struct vpgqueues vm_page_queues[PQ_COUNT]; 130 __read_mostly long vmmeter_neg_slop_cnt = -VMMETER_SLOP_COUNT; 131 132 static volatile int vm_pages_waiting; 133 static struct alist vm_contig_alist; 134 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 135 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 136 137 __read_mostly static int vm_page_hash_vnode_only; 138 __read_mostly static int vm_page_hash_size; 139 __read_mostly static struct vm_page_hash_elm *vm_page_hash; 140 141 static u_long vm_dma_reserved = 0; 142 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 143 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 144 "Memory reserved for DMA"); 145 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 146 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 147 148 SYSCTL_INT(_vm, OID_AUTO, page_hash_vnode_only, CTLFLAG_RW, 149 &vm_page_hash_vnode_only, 0, "Only hash vnode pages"); 150 #if 0 151 static int vm_page_hash_debug; 152 SYSCTL_INT(_vm, OID_AUTO, page_hash_debug, CTLFLAG_RW, 153 &vm_page_hash_debug, 0, "Only hash vnode pages"); 154 #endif 155 156 static int vm_contig_verbose = 0; 157 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 158 159 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 160 vm_pindex_t, pindex); 161 162 static void 163 vm_page_queue_init(void) 164 { 165 int i; 166 167 for (i = 0; i < PQ_L2_SIZE; i++) 168 vm_page_queues[PQ_FREE+i].cnt_offset = 169 offsetof(struct vmstats, v_free_count); 170 for (i = 0; i < PQ_L2_SIZE; i++) 171 vm_page_queues[PQ_CACHE+i].cnt_offset = 172 offsetof(struct vmstats, v_cache_count); 173 for (i = 0; i < PQ_L2_SIZE; i++) 174 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 175 offsetof(struct vmstats, v_inactive_count); 176 for (i = 0; i < PQ_L2_SIZE; i++) 177 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 178 offsetof(struct vmstats, v_active_count); 179 for (i = 0; i < PQ_L2_SIZE; i++) 180 vm_page_queues[PQ_HOLD+i].cnt_offset = 181 offsetof(struct vmstats, v_active_count); 182 /* PQ_NONE has no queue */ 183 184 for (i = 0; i < PQ_COUNT; i++) { 185 vm_page_queues[i].lastq = -1; 186 TAILQ_INIT(&vm_page_queues[i].pl); 187 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 188 } 189 } 190 191 /* 192 * note: place in initialized data section? Is this necessary? 193 */ 194 vm_pindex_t first_page = 0; 195 vm_pindex_t vm_page_array_size = 0; 196 vm_page_t vm_page_array = NULL; 197 vm_paddr_t vm_low_phys_reserved; 198 199 /* 200 * (low level boot) 201 * 202 * Sets the page size, perhaps based upon the memory size. 203 * Must be called before any use of page-size dependent functions. 204 */ 205 void 206 vm_set_page_size(void) 207 { 208 if (vmstats.v_page_size == 0) 209 vmstats.v_page_size = PAGE_SIZE; 210 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 211 panic("vm_set_page_size: page size not a power of two"); 212 } 213 214 /* 215 * (low level boot) 216 * 217 * Add a new page to the freelist for use by the system. New pages 218 * are added to both the head and tail of the associated free page 219 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 220 * requests pull 'recent' adds (higher physical addresses) first. 221 * 222 * Beware that the page zeroing daemon will also be running soon after 223 * boot, moving pages from the head to the tail of the PQ_FREE queues. 224 * 225 * Must be called in a critical section. 226 */ 227 static void 228 vm_add_new_page(vm_paddr_t pa) 229 { 230 struct vpgqueues *vpq; 231 vm_page_t m; 232 233 m = PHYS_TO_VM_PAGE(pa); 234 m->phys_addr = pa; 235 m->flags = 0; 236 m->pat_mode = PAT_WRITE_BACK; 237 m->pc = (pa >> PAGE_SHIFT); 238 239 /* 240 * Twist for cpu localization in addition to page coloring, so 241 * different cpus selecting by m->queue get different page colors. 242 */ 243 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 244 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 245 m->pc &= PQ_L2_MASK; 246 247 /* 248 * Reserve a certain number of contiguous low memory pages for 249 * contigmalloc() to use. 250 * 251 * Even though these pages represent real ram and can be 252 * reverse-mapped, we set PG_FICTITIOUS and PG_UNQUEUED 253 * because their use is special-cased. 254 * 255 * WARNING! Once PG_FICTITIOUS is set, vm_page_wire*() 256 * and vm_page_unwire*() calls have no effect. 257 */ 258 if (pa < vm_low_phys_reserved) { 259 atomic_add_long(&vmstats.v_page_count, 1); 260 atomic_add_long(&vmstats.v_dma_pages, 1); 261 m->flags |= PG_FICTITIOUS | PG_UNQUEUED; 262 m->queue = PQ_NONE; 263 m->wire_count = 1; 264 atomic_add_long(&vmstats.v_wire_count, 1); 265 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 266 return; 267 } 268 269 /* 270 * General page 271 */ 272 m->queue = m->pc + PQ_FREE; 273 KKASSERT(m->dirty == 0); 274 275 atomic_add_long(&vmstats.v_page_count, 1); 276 atomic_add_long(&vmstats.v_free_count, 1); 277 vpq = &vm_page_queues[m->queue]; 278 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 279 ++vpq->lcnt; 280 } 281 282 /* 283 * (low level boot) 284 * 285 * Initializes the resident memory module. 286 * 287 * Preallocates memory for critical VM structures and arrays prior to 288 * kernel_map becoming available. 289 * 290 * Memory is allocated from (virtual2_start, virtual2_end) if available, 291 * otherwise memory is allocated from (virtual_start, virtual_end). 292 * 293 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 294 * large enough to hold vm_page_array & other structures for machines with 295 * large amounts of ram, so we want to use virtual2* when available. 296 */ 297 void 298 vm_page_startup(void) 299 { 300 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 301 vm_offset_t mapped; 302 vm_pindex_t npages; 303 vm_paddr_t page_range; 304 vm_paddr_t new_end; 305 int i; 306 vm_paddr_t pa; 307 vm_paddr_t last_pa; 308 vm_paddr_t end; 309 vm_paddr_t biggestone, biggestsize; 310 vm_paddr_t total; 311 vm_page_t m; 312 313 total = 0; 314 biggestsize = 0; 315 biggestone = 0; 316 vaddr = round_page(vaddr); 317 318 /* 319 * Make sure ranges are page-aligned. 320 */ 321 for (i = 0; phys_avail[i].phys_end; ++i) { 322 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 323 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 324 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 325 phys_avail[i].phys_end = phys_avail[i].phys_beg; 326 } 327 328 /* 329 * Locate largest block 330 */ 331 for (i = 0; phys_avail[i].phys_end; ++i) { 332 vm_paddr_t size = phys_avail[i].phys_end - 333 phys_avail[i].phys_beg; 334 335 if (size > biggestsize) { 336 biggestone = i; 337 biggestsize = size; 338 } 339 total += size; 340 } 341 --i; /* adjust to last entry for use down below */ 342 343 end = phys_avail[biggestone].phys_end; 344 end = trunc_page(end); 345 346 /* 347 * Initialize the queue headers for the free queue, the active queue 348 * and the inactive queue. 349 */ 350 vm_page_queue_init(); 351 352 #if !defined(_KERNEL_VIRTUAL) 353 /* 354 * VKERNELs don't support minidumps and as such don't need 355 * vm_page_dump 356 * 357 * Allocate a bitmap to indicate that a random physical page 358 * needs to be included in a minidump. 359 * 360 * The amd64 port needs this to indicate which direct map pages 361 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 362 * 363 * However, x86 still needs this workspace internally within the 364 * minidump code. In theory, they are not needed on x86, but are 365 * included should the sf_buf code decide to use them. 366 */ 367 page_range = phys_avail[i].phys_end / PAGE_SIZE; 368 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 369 end -= vm_page_dump_size; 370 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 371 VM_PROT_READ | VM_PROT_WRITE); 372 bzero((void *)vm_page_dump, vm_page_dump_size); 373 #endif 374 /* 375 * Compute the number of pages of memory that will be available for 376 * use (taking into account the overhead of a page structure per 377 * page). 378 */ 379 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 380 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 381 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 382 383 #ifndef _KERNEL_VIRTUAL 384 /* 385 * (only applies to real kernels) 386 * 387 * Reserve a large amount of low memory for potential 32-bit DMA 388 * space allocations. Once device initialization is complete we 389 * release most of it, but keep (vm_dma_reserved) memory reserved 390 * for later use. Typically for X / graphics. Through trial and 391 * error we find that GPUs usually requires ~60-100MB or so. 392 * 393 * By default, 128M is left in reserve on machines with 2G+ of ram. 394 */ 395 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 396 if (vm_low_phys_reserved > total / 4) 397 vm_low_phys_reserved = total / 4; 398 if (vm_dma_reserved == 0) { 399 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 400 if (vm_dma_reserved > total / 16) 401 vm_dma_reserved = total / 16; 402 } 403 #endif 404 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 405 ALIST_RECORDS_65536); 406 407 /* 408 * Initialize the mem entry structures now, and put them in the free 409 * queue. 410 */ 411 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 412 kprintf("initializing vm_page_array "); 413 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 414 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 415 vm_page_array = (vm_page_t)mapped; 416 417 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 418 /* 419 * since pmap_map on amd64 returns stuff out of a direct-map region, 420 * we have to manually add these pages to the minidump tracking so 421 * that they can be dumped, including the vm_page_array. 422 */ 423 for (pa = new_end; 424 pa < phys_avail[biggestone].phys_end; 425 pa += PAGE_SIZE) { 426 dump_add_page(pa); 427 } 428 #endif 429 430 /* 431 * Clear all of the page structures, run basic initialization so 432 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 433 * map. 434 */ 435 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 436 vm_page_array_size = page_range; 437 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024) 438 kprintf("size = 0x%zx\n", vm_page_array_size); 439 440 m = &vm_page_array[0]; 441 pa = ptoa(first_page); 442 for (i = 0; i < page_range; ++i) { 443 spin_init(&m->spin, "vm_page"); 444 m->phys_addr = pa; 445 pa += PAGE_SIZE; 446 ++m; 447 } 448 449 /* 450 * Construct the free queue(s) in ascending order (by physical 451 * address) so that the first 16MB of physical memory is allocated 452 * last rather than first. On large-memory machines, this avoids 453 * the exhaustion of low physical memory before isa_dma_init has run. 454 */ 455 vmstats.v_page_count = 0; 456 vmstats.v_free_count = 0; 457 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 458 pa = phys_avail[i].phys_beg; 459 if (i == biggestone) 460 last_pa = new_end; 461 else 462 last_pa = phys_avail[i].phys_end; 463 while (pa < last_pa && npages-- > 0) { 464 vm_add_new_page(pa); 465 pa += PAGE_SIZE; 466 } 467 } 468 if (virtual2_start) 469 virtual2_start = vaddr; 470 else 471 virtual_start = vaddr; 472 mycpu->gd_vmstats = vmstats; 473 } 474 475 /* 476 * (called from early boot only) 477 * 478 * Reorganize VM pages based on numa data. May be called as many times as 479 * necessary. Will reorganize the vm_page_t page color and related queue(s) 480 * to allow vm_page_alloc() to choose pages based on socket affinity. 481 * 482 * NOTE: This function is only called while we are still in UP mode, so 483 * we only need a critical section to protect the queues (which 484 * saves a lot of time, there are likely a ton of pages). 485 */ 486 void 487 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 488 { 489 vm_paddr_t scan_beg; 490 vm_paddr_t scan_end; 491 vm_paddr_t ran_end; 492 struct vpgqueues *vpq; 493 vm_page_t m; 494 vm_page_t mend; 495 int socket_mod; 496 int socket_value; 497 int i; 498 499 /* 500 * Check if no physical information, or there was only one socket 501 * (so don't waste time doing nothing!). 502 */ 503 if (cpu_topology_phys_ids <= 1 || 504 cpu_topology_core_ids == 0) { 505 return; 506 } 507 508 /* 509 * Setup for our iteration. Note that ACPI may iterate CPU 510 * sockets starting at 0 or 1 or some other number. The 511 * cpu_topology code mod's it against the socket count. 512 */ 513 ran_end = ran_beg + bytes; 514 515 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 516 socket_value = (physid % cpu_topology_phys_ids) * socket_mod; 517 mend = &vm_page_array[vm_page_array_size]; 518 519 crit_enter(); 520 521 /* 522 * Adjust cpu_topology's phys_mem parameter 523 */ 524 if (root_cpu_node) 525 vm_numa_add_topology_mem(root_cpu_node, physid, (long)bytes); 526 527 /* 528 * Adjust vm_page->pc and requeue all affected pages. The 529 * allocator will then be able to localize memory allocations 530 * to some degree. 531 */ 532 for (i = 0; phys_avail[i].phys_end; ++i) { 533 scan_beg = phys_avail[i].phys_beg; 534 scan_end = phys_avail[i].phys_end; 535 if (scan_end <= ran_beg) 536 continue; 537 if (scan_beg >= ran_end) 538 continue; 539 if (scan_beg < ran_beg) 540 scan_beg = ran_beg; 541 if (scan_end > ran_end) 542 scan_end = ran_end; 543 if (atop(scan_end) > first_page + vm_page_array_size) 544 scan_end = ptoa(first_page + vm_page_array_size); 545 546 m = PHYS_TO_VM_PAGE(scan_beg); 547 while (scan_beg < scan_end) { 548 KKASSERT(m < mend); 549 if (m->queue != PQ_NONE) { 550 vpq = &vm_page_queues[m->queue]; 551 TAILQ_REMOVE(&vpq->pl, m, pageq); 552 --vpq->lcnt; 553 /* queue doesn't change, no need to adj cnt */ 554 m->queue -= m->pc; 555 m->pc %= socket_mod; 556 m->pc += socket_value; 557 m->pc &= PQ_L2_MASK; 558 m->queue += m->pc; 559 vpq = &vm_page_queues[m->queue]; 560 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 561 ++vpq->lcnt; 562 /* queue doesn't change, no need to adj cnt */ 563 } else { 564 m->pc %= socket_mod; 565 m->pc += socket_value; 566 m->pc &= PQ_L2_MASK; 567 } 568 scan_beg += PAGE_SIZE; 569 ++m; 570 } 571 } 572 573 crit_exit(); 574 } 575 576 /* 577 * (called from early boot only) 578 * 579 * Don't allow the NUMA organization to leave vm_page_queues[] nodes 580 * completely empty for a logical cpu. Doing so would force allocations 581 * on that cpu to always borrow from a nearby cpu, create unnecessary 582 * contention, and cause vm_page_alloc() to iterate more queues and run more 583 * slowly. 584 * 585 * This situation can occur when memory sticks are not entirely populated, 586 * populated at different densities, or in naturally assymetric systems 587 * such as the 2990WX. There could very well be many vm_page_queues[] 588 * entries with *NO* pages assigned to them. 589 * 590 * Fixing this up ensures that each logical CPU has roughly the same 591 * sized memory pool, and more importantly ensures that logical CPUs 592 * do not wind up with an empty memory pool. 593 * 594 * At them moment we just iterate the other queues and borrow pages, 595 * moving them into the queues for cpus with severe deficits even though 596 * the memory might not be local to those cpus. I am not doing this in 597 * a 'smart' way, its effectively UMA style (sorta, since its page-by-page 598 * whereas real UMA typically exchanges address bits 8-10 with high address 599 * bits). But it works extremely well and gives us fairly good deterministic 600 * results on the cpu cores associated with these secondary nodes. 601 */ 602 void 603 vm_numa_organize_finalize(void) 604 { 605 struct vpgqueues *vpq; 606 vm_page_t m; 607 long lcnt_lo; 608 long lcnt_hi; 609 int iter; 610 int i; 611 int scale_lim; 612 613 crit_enter(); 614 615 /* 616 * Machines might not use an exact power of 2 for phys_ids, 617 * core_ids, ht_ids, etc. This can slightly reduce the actual 618 * range of indices in vm_page_queues[] that are nominally used. 619 */ 620 if (cpu_topology_ht_ids) { 621 scale_lim = PQ_L2_SIZE / cpu_topology_phys_ids; 622 scale_lim = scale_lim / cpu_topology_core_ids; 623 scale_lim = scale_lim / cpu_topology_ht_ids; 624 scale_lim = scale_lim * cpu_topology_ht_ids; 625 scale_lim = scale_lim * cpu_topology_core_ids; 626 scale_lim = scale_lim * cpu_topology_phys_ids; 627 } else { 628 scale_lim = PQ_L2_SIZE; 629 } 630 631 /* 632 * Calculate an average, set hysteresis for balancing from 633 * 10% below the average to the average. 634 */ 635 lcnt_hi = 0; 636 for (i = 0; i < scale_lim; ++i) { 637 lcnt_hi += vm_page_queues[i].lcnt; 638 } 639 lcnt_hi /= scale_lim; 640 lcnt_lo = lcnt_hi - lcnt_hi / 10; 641 642 kprintf("vm_page: avg %ld pages per queue, %d queues\n", 643 lcnt_hi, scale_lim); 644 645 iter = 0; 646 for (i = 0; i < scale_lim; ++i) { 647 vpq = &vm_page_queues[PQ_FREE + i]; 648 while (vpq->lcnt < lcnt_lo) { 649 struct vpgqueues *vptmp; 650 651 iter = (iter + 1) & PQ_L2_MASK; 652 vptmp = &vm_page_queues[PQ_FREE + iter]; 653 if (vptmp->lcnt < lcnt_hi) 654 continue; 655 m = TAILQ_FIRST(&vptmp->pl); 656 KKASSERT(m->queue == PQ_FREE + iter); 657 TAILQ_REMOVE(&vptmp->pl, m, pageq); 658 --vptmp->lcnt; 659 /* queue doesn't change, no need to adj cnt */ 660 m->queue -= m->pc; 661 m->pc = i; 662 m->queue += m->pc; 663 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 664 ++vpq->lcnt; 665 } 666 } 667 crit_exit(); 668 } 669 670 static 671 void 672 vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes) 673 { 674 int cpuid; 675 int i; 676 677 switch(cpup->type) { 678 case PACKAGE_LEVEL: 679 cpup->phys_mem += bytes; 680 break; 681 case CHIP_LEVEL: 682 /* 683 * All members should have the same chipid, so we only need 684 * to pull out one member. 685 */ 686 if (CPUMASK_TESTNZERO(cpup->members)) { 687 cpuid = BSFCPUMASK(cpup->members); 688 if (physid == 689 get_chip_ID_from_APICID(CPUID_TO_APICID(cpuid))) { 690 cpup->phys_mem += bytes; 691 } 692 } 693 break; 694 case CORE_LEVEL: 695 case THREAD_LEVEL: 696 /* 697 * Just inherit from the parent node 698 */ 699 cpup->phys_mem = cpup->parent_node->phys_mem; 700 break; 701 } 702 for (i = 0; i < MAXCPU && cpup->child_node[i]; ++i) 703 vm_numa_add_topology_mem(cpup->child_node[i], physid, bytes); 704 } 705 706 /* 707 * We tended to reserve a ton of memory for contigmalloc(). Now that most 708 * drivers have initialized we want to return most the remaining free 709 * reserve back to the VM page queues so they can be used for normal 710 * allocations. 711 * 712 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 713 */ 714 static void 715 vm_page_startup_finish(void *dummy __unused) 716 { 717 alist_blk_t blk; 718 alist_blk_t rblk; 719 alist_blk_t count; 720 alist_blk_t xcount; 721 alist_blk_t bfree; 722 vm_page_t m; 723 struct vm_page_hash_elm *mp; 724 int mask; 725 726 /* 727 * Set the set_assoc_mask based on the fitted number of CPUs. 728 * This is a mask, so we subject 1. 729 * 730 * w/PQ_L2_SIZE = 1024, Don't let the associativity drop below 8. 731 * So if we have 256 CPUs, two hyper-threads will wind up sharing. 732 * 733 * The maximum is PQ_L2_SIZE. However, we limit the starting 734 * maximum to 16 (mask = 15) in order to improve the cache locality 735 * of related kernel data structures. 736 */ 737 mask = PQ_L2_SIZE / ncpus_fit - 1; 738 if (mask < 7) /* minimum is 8-way w/256 CPU threads */ 739 mask = 7; 740 if (mask < 15) 741 mask = 15; 742 cpu_ccfence(); 743 set_assoc_mask = mask; 744 745 /* 746 * Return part of the initial reserve back to the system 747 */ 748 spin_lock(&vm_contig_spin); 749 for (;;) { 750 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 751 if (bfree <= vm_dma_reserved / PAGE_SIZE) 752 break; 753 if (count == 0) 754 break; 755 756 /* 757 * Figure out how much of the initial reserve we have to 758 * free in order to reach our target. 759 */ 760 bfree -= vm_dma_reserved / PAGE_SIZE; 761 if (count > bfree) { 762 blk += count - bfree; 763 count = bfree; 764 } 765 766 /* 767 * Calculate the nearest power of 2 <= count. 768 */ 769 for (xcount = 1; xcount <= count; xcount <<= 1) 770 ; 771 xcount >>= 1; 772 blk += count - xcount; 773 count = xcount; 774 775 /* 776 * Allocate the pages from the alist, then free them to 777 * the normal VM page queues. 778 * 779 * Pages allocated from the alist are wired. We have to 780 * busy, unwire, and free them. We must also adjust 781 * vm_low_phys_reserved before freeing any pages to prevent 782 * confusion. 783 */ 784 rblk = alist_alloc(&vm_contig_alist, blk, count); 785 if (rblk != blk) { 786 kprintf("vm_page_startup_finish: Unable to return " 787 "dma space @0x%08x/%d -> 0x%08x\n", 788 blk, count, rblk); 789 break; 790 } 791 atomic_add_long(&vmstats.v_dma_pages, -(long)count); 792 spin_unlock(&vm_contig_spin); 793 794 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 795 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 796 while (count) { 797 vm_page_flag_clear(m, PG_FICTITIOUS | PG_UNQUEUED); 798 vm_page_busy_wait(m, FALSE, "cpgfr"); 799 vm_page_unwire(m, 0); 800 vm_page_free(m); 801 --count; 802 ++m; 803 } 804 spin_lock(&vm_contig_spin); 805 } 806 spin_unlock(&vm_contig_spin); 807 808 /* 809 * Print out how much DMA space drivers have already allocated and 810 * how much is left over. 811 */ 812 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 813 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 814 (PAGE_SIZE / 1024), 815 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 816 817 /* 818 * Power of 2 819 */ 820 vm_page_hash_size = 4096; 821 while (vm_page_hash_size < (vm_page_array_size / 16)) 822 vm_page_hash_size <<= 1; 823 if (vm_page_hash_size > VM_PAGE_HASH_MAX) 824 vm_page_hash_size = VM_PAGE_HASH_MAX; 825 826 /* 827 * hash table for vm_page_lookup_quick() 828 */ 829 mp = (void *)kmem_alloc3(&kernel_map, 830 vm_page_hash_size * sizeof(*vm_page_hash), 831 VM_SUBSYS_VMPGHASH, KM_CPU(0)); 832 bzero(mp, vm_page_hash_size * sizeof(*mp)); 833 cpu_sfence(); 834 vm_page_hash = mp; 835 } 836 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 837 vm_page_startup_finish, NULL); 838 839 840 /* 841 * Scan comparison function for Red-Black tree scans. An inclusive 842 * (start,end) is expected. Other fields are not used. 843 */ 844 int 845 rb_vm_page_scancmp(struct vm_page *p, void *data) 846 { 847 struct rb_vm_page_scan_info *info = data; 848 849 if (p->pindex < info->start_pindex) 850 return(-1); 851 if (p->pindex > info->end_pindex) 852 return(1); 853 return(0); 854 } 855 856 int 857 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 858 { 859 if (p1->pindex < p2->pindex) 860 return(-1); 861 if (p1->pindex > p2->pindex) 862 return(1); 863 return(0); 864 } 865 866 void 867 vm_page_init(vm_page_t m) 868 { 869 /* do nothing for now. Called from pmap_page_init() */ 870 } 871 872 /* 873 * Each page queue has its own spin lock, which is fairly optimal for 874 * allocating and freeing pages at least. 875 * 876 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 877 * queue spinlock via this function. Also note that m->queue cannot change 878 * unless both the page and queue are locked. 879 */ 880 static __inline 881 void 882 _vm_page_queue_spin_lock(vm_page_t m) 883 { 884 u_short queue; 885 886 queue = m->queue; 887 if (queue != PQ_NONE) { 888 spin_lock(&vm_page_queues[queue].spin); 889 KKASSERT(queue == m->queue); 890 } 891 } 892 893 static __inline 894 void 895 _vm_page_queue_spin_unlock(vm_page_t m) 896 { 897 u_short queue; 898 899 queue = m->queue; 900 cpu_ccfence(); 901 if (queue != PQ_NONE) 902 spin_unlock(&vm_page_queues[queue].spin); 903 } 904 905 static __inline 906 void 907 _vm_page_queues_spin_lock(u_short queue) 908 { 909 cpu_ccfence(); 910 if (queue != PQ_NONE) 911 spin_lock(&vm_page_queues[queue].spin); 912 } 913 914 915 static __inline 916 void 917 _vm_page_queues_spin_unlock(u_short queue) 918 { 919 cpu_ccfence(); 920 if (queue != PQ_NONE) 921 spin_unlock(&vm_page_queues[queue].spin); 922 } 923 924 void 925 vm_page_queue_spin_lock(vm_page_t m) 926 { 927 _vm_page_queue_spin_lock(m); 928 } 929 930 void 931 vm_page_queues_spin_lock(u_short queue) 932 { 933 _vm_page_queues_spin_lock(queue); 934 } 935 936 void 937 vm_page_queue_spin_unlock(vm_page_t m) 938 { 939 _vm_page_queue_spin_unlock(m); 940 } 941 942 void 943 vm_page_queues_spin_unlock(u_short queue) 944 { 945 _vm_page_queues_spin_unlock(queue); 946 } 947 948 /* 949 * This locks the specified vm_page and its queue in the proper order 950 * (page first, then queue). The queue may change so the caller must 951 * recheck on return. 952 */ 953 static __inline 954 void 955 _vm_page_and_queue_spin_lock(vm_page_t m) 956 { 957 vm_page_spin_lock(m); 958 _vm_page_queue_spin_lock(m); 959 } 960 961 static __inline 962 void 963 _vm_page_and_queue_spin_unlock(vm_page_t m) 964 { 965 _vm_page_queues_spin_unlock(m->queue); 966 vm_page_spin_unlock(m); 967 } 968 969 void 970 vm_page_and_queue_spin_unlock(vm_page_t m) 971 { 972 _vm_page_and_queue_spin_unlock(m); 973 } 974 975 void 976 vm_page_and_queue_spin_lock(vm_page_t m) 977 { 978 _vm_page_and_queue_spin_lock(m); 979 } 980 981 /* 982 * Helper function removes vm_page from its current queue. 983 * Returns the base queue the page used to be on. 984 * 985 * The vm_page and the queue must be spinlocked. 986 * This function will unlock the queue but leave the page spinlocked. 987 */ 988 static __inline u_short 989 _vm_page_rem_queue_spinlocked(vm_page_t m) 990 { 991 struct vpgqueues *pq; 992 u_short queue; 993 u_short oqueue; 994 long *cnt; 995 996 queue = m->queue; 997 if (queue != PQ_NONE) { 998 pq = &vm_page_queues[queue]; 999 TAILQ_REMOVE(&pq->pl, m, pageq); 1000 1001 /* 1002 * Adjust our pcpu stats. In order for the nominal low-memory 1003 * algorithms to work properly we don't let any pcpu stat get 1004 * too negative before we force it to be rolled-up into the 1005 * global stats. Otherwise our pageout and vm_wait tests 1006 * will fail badly. 1007 * 1008 * The idea here is to reduce unnecessary SMP cache 1009 * mastership changes in the global vmstats, which can be 1010 * particularly bad in multi-socket systems. 1011 * 1012 * NOTE: The double *cnt test tries to avoid a global memory 1013 * read. vmmeter_neg_slop_cnt is more generous than 1014 * the baseline define, we want to try to avoid atomic 1015 * ops on the global 'vmstats' structure as much as 1016 * possible. 1017 */ 1018 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 1019 atomic_add_long(cnt, -1); 1020 if (*cnt < -VMMETER_SLOP_COUNT && *cnt < vmmeter_neg_slop_cnt) { 1021 u_long copy = atomic_swap_long(cnt, 0); 1022 cnt = (long *)((char *)&vmstats + pq->cnt_offset); 1023 atomic_add_long(cnt, copy); 1024 cnt = (long *)((char *)&mycpu->gd_vmstats + 1025 pq->cnt_offset); 1026 atomic_add_long(cnt, copy); 1027 } 1028 pq->lcnt--; 1029 m->queue = PQ_NONE; 1030 oqueue = queue; 1031 queue -= m->pc; 1032 vm_page_queues_spin_unlock(oqueue); /* intended */ 1033 } 1034 return queue; 1035 } 1036 1037 /* 1038 * Helper function places the vm_page on the specified queue. Generally 1039 * speaking only PQ_FREE pages are placed at the head, to allow them to 1040 * be allocated sooner rather than later on the assumption that they 1041 * are cache-hot. 1042 * 1043 * The vm_page must be spinlocked. 1044 * The vm_page must NOT be FICTITIOUS (that would be a disaster) 1045 * This function will return with both the page and the queue locked. 1046 */ 1047 static __inline void 1048 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 1049 { 1050 struct vpgqueues *pq; 1051 u_long *cnt; 1052 1053 KKASSERT(m->queue == PQ_NONE && 1054 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0); 1055 1056 if (queue != PQ_NONE) { 1057 vm_page_queues_spin_lock(queue); 1058 pq = &vm_page_queues[queue]; 1059 ++pq->lcnt; 1060 1061 /* 1062 * Adjust our pcpu stats. If a system entity really needs 1063 * to incorporate the count it will call vmstats_rollup() 1064 * to roll it all up into the global vmstats strufture. 1065 */ 1066 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 1067 atomic_add_long(cnt, 1); 1068 1069 /* 1070 * PQ_FREE is always handled LIFO style to try to provide 1071 * cache-hot pages to programs. 1072 */ 1073 m->queue = queue; 1074 if (queue - m->pc == PQ_FREE) { 1075 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1076 } else if (athead) { 1077 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1078 } else { 1079 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1080 } 1081 /* leave the queue spinlocked */ 1082 } 1083 } 1084 1085 /* 1086 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 1087 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 1088 * 1089 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 1090 * call will be made before returning. 1091 * 1092 * This function does NOT busy the page and on return the page is not 1093 * guaranteed to be available. 1094 */ 1095 void 1096 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 1097 { 1098 u_int32_t busy_count; 1099 1100 for (;;) { 1101 busy_count = m->busy_count; 1102 cpu_ccfence(); 1103 1104 if ((busy_count & PBUSY_LOCKED) == 0 && 1105 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 1106 break; 1107 } 1108 tsleep_interlock(m, 0); 1109 if (atomic_cmpset_int(&m->busy_count, busy_count, 1110 busy_count | PBUSY_WANTED)) { 1111 atomic_set_int(&m->flags, PG_REFERENCED); 1112 tsleep(m, PINTERLOCKED, msg, 0); 1113 break; 1114 } 1115 } 1116 } 1117 1118 /* 1119 * This calculates and returns a page color given an optional VM object and 1120 * either a pindex or an iterator. We attempt to return a cpu-localized 1121 * pg_color that is still roughly 16-way set-associative. The CPU topology 1122 * is used if it was probed. 1123 * 1124 * The caller may use the returned value to index into e.g. PQ_FREE when 1125 * allocating a page in order to nominally obtain pages that are hopefully 1126 * already localized to the requesting cpu. This function is not able to 1127 * provide any sort of guarantee of this, but does its best to improve 1128 * hardware cache management performance. 1129 * 1130 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 1131 */ 1132 u_short 1133 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 1134 { 1135 u_short pg_color; 1136 int object_pg_color; 1137 1138 /* 1139 * WARNING! cpu_topology_core_ids might not be a power of two. 1140 * We also shouldn't make assumptions about 1141 * cpu_topology_phys_ids either. 1142 * 1143 * WARNING! ncpus might not be known at this time (during early 1144 * boot), and might be set to 1. 1145 * 1146 * General format: [phys_id][core_id][cpuid][set-associativity] 1147 * (but uses modulo, so not necessarily precise bit masks) 1148 */ 1149 object_pg_color = object ? object->pg_color : 0; 1150 1151 if (cpu_topology_ht_ids) { 1152 int phys_id; 1153 int core_id; 1154 int ht_id; 1155 int physcale; 1156 int grpscale; 1157 int cpuscale; 1158 1159 /* 1160 * Translate cpuid to socket, core, and hyperthread id. 1161 */ 1162 phys_id = get_cpu_phys_id(cpuid); 1163 core_id = get_cpu_core_id(cpuid); 1164 ht_id = get_cpu_ht_id(cpuid); 1165 1166 /* 1167 * Calculate pg_color for our array index. 1168 * 1169 * physcale - socket multiplier. 1170 * grpscale - core multiplier (cores per socket) 1171 * cpu* - cpus per core 1172 * 1173 * WARNING! In early boot, ncpus has not yet been 1174 * initialized and may be set to (1). 1175 * 1176 * WARNING! physcale must match the organization that 1177 * vm_numa_organize() creates to ensure that 1178 * we properly localize allocations to the 1179 * requested cpuid. 1180 */ 1181 physcale = PQ_L2_SIZE / cpu_topology_phys_ids; 1182 grpscale = physcale / cpu_topology_core_ids; 1183 cpuscale = grpscale / cpu_topology_ht_ids; 1184 1185 pg_color = phys_id * physcale; 1186 pg_color += core_id * grpscale; 1187 pg_color += ht_id * cpuscale; 1188 pg_color += (pindex + object_pg_color) % cpuscale; 1189 1190 #if 0 1191 if (grpsize >= 8) { 1192 pg_color += (pindex + object_pg_color) % grpsize; 1193 } else { 1194 if (grpsize <= 2) { 1195 grpsize = 8; 1196 } else { 1197 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 1198 grpsize += grpsize; 1199 if (grpsize < 8) 1200 grpsize += grpsize; 1201 } 1202 pg_color += (pindex + object_pg_color) % grpsize; 1203 } 1204 #endif 1205 } else { 1206 /* 1207 * Unknown topology, distribute things evenly. 1208 * 1209 * WARNING! In early boot, ncpus has not yet been 1210 * initialized and may be set to (1). 1211 */ 1212 int cpuscale; 1213 1214 cpuscale = PQ_L2_SIZE / ncpus; 1215 1216 pg_color = cpuid * cpuscale; 1217 pg_color += (pindex + object_pg_color) % cpuscale; 1218 } 1219 return (pg_color & PQ_L2_MASK); 1220 } 1221 1222 /* 1223 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 1224 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 1225 */ 1226 void 1227 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 1228 int also_m_busy, const char *msg 1229 VM_PAGE_DEBUG_ARGS) 1230 { 1231 u_int32_t busy_count; 1232 1233 for (;;) { 1234 busy_count = m->busy_count; 1235 cpu_ccfence(); 1236 if (busy_count & PBUSY_LOCKED) { 1237 tsleep_interlock(m, 0); 1238 if (atomic_cmpset_int(&m->busy_count, busy_count, 1239 busy_count | PBUSY_WANTED)) { 1240 atomic_set_int(&m->flags, PG_REFERENCED); 1241 tsleep(m, PINTERLOCKED, msg, 0); 1242 } 1243 } else if (also_m_busy && busy_count) { 1244 tsleep_interlock(m, 0); 1245 if (atomic_cmpset_int(&m->busy_count, busy_count, 1246 busy_count | PBUSY_WANTED)) { 1247 atomic_set_int(&m->flags, PG_REFERENCED); 1248 tsleep(m, PINTERLOCKED, msg, 0); 1249 } 1250 } else { 1251 if (atomic_cmpset_int(&m->busy_count, busy_count, 1252 busy_count | PBUSY_LOCKED)) { 1253 #ifdef VM_PAGE_DEBUG 1254 m->busy_func = func; 1255 m->busy_line = lineno; 1256 #endif 1257 break; 1258 } 1259 } 1260 } 1261 } 1262 1263 /* 1264 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1265 * m->busy_count is also 0. 1266 * 1267 * Returns non-zero on failure. 1268 */ 1269 int 1270 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1271 VM_PAGE_DEBUG_ARGS) 1272 { 1273 u_int32_t busy_count; 1274 1275 for (;;) { 1276 busy_count = m->busy_count; 1277 cpu_ccfence(); 1278 if (busy_count & PBUSY_LOCKED) 1279 return TRUE; 1280 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1281 return TRUE; 1282 if (atomic_cmpset_int(&m->busy_count, busy_count, 1283 busy_count | PBUSY_LOCKED)) { 1284 #ifdef VM_PAGE_DEBUG 1285 m->busy_func = func; 1286 m->busy_line = lineno; 1287 #endif 1288 return FALSE; 1289 } 1290 } 1291 } 1292 1293 /* 1294 * Clear the BUSY flag and return non-zero to indicate to the caller 1295 * that a wakeup() should be performed. 1296 * 1297 * (inline version) 1298 */ 1299 static __inline 1300 int 1301 _vm_page_wakeup(vm_page_t m) 1302 { 1303 u_int32_t busy_count; 1304 1305 busy_count = m->busy_count; 1306 cpu_ccfence(); 1307 for (;;) { 1308 if (atomic_fcmpset_int(&m->busy_count, &busy_count, 1309 busy_count & 1310 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1311 return((int)(busy_count & PBUSY_WANTED)); 1312 } 1313 } 1314 /* not reached */ 1315 } 1316 1317 /* 1318 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1319 * is typically the last call you make on a page before moving onto 1320 * other things. 1321 */ 1322 void 1323 vm_page_wakeup(vm_page_t m) 1324 { 1325 KASSERT(m->busy_count & PBUSY_LOCKED, 1326 ("vm_page_wakeup: page not busy!!!")); 1327 if (_vm_page_wakeup(m)) 1328 wakeup(m); 1329 } 1330 1331 /* 1332 * Hold a page, preventing reuse. This is typically only called on pages 1333 * in a known state (either held busy, special, or interlocked in some 1334 * manner). Holding a page does not ensure that it remains valid, it only 1335 * prevents reuse. The page must not already be on the FREE queue or in 1336 * any danger of being moved to the FREE queue concurrent with this call. 1337 * 1338 * Other parts of the system can still disassociate the page from its object 1339 * and attempt to free it, or perform read or write I/O on it and/or otherwise 1340 * manipulate the page, but if the page is held the VM system will leave the 1341 * page and its data intact and not cycle it through the FREE queue until 1342 * the last hold has been released. 1343 * 1344 * (see vm_page_wire() if you want to prevent the page from being 1345 * disassociated from its object too). 1346 */ 1347 void 1348 vm_page_hold(vm_page_t m) 1349 { 1350 atomic_add_int(&m->hold_count, 1); 1351 KKASSERT(m->queue - m->pc != PQ_FREE); 1352 } 1353 1354 /* 1355 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1356 * it was freed while held and must be moved back to the FREE queue. 1357 * 1358 * To avoid racing against vm_page_free*() we must re-test conditions 1359 * after obtaining the spin-lock. The initial test can also race a 1360 * vm_page_free*() that is in the middle of moving a page to PQ_HOLD, 1361 * leaving the page on PQ_HOLD with hold_count == 0. Rather than 1362 * throw a spin-lock in the critical path, we rely on the pageout 1363 * daemon to clean-up these loose ends. 1364 * 1365 * More critically, the 'easy movement' between queues without busying 1366 * a vm_page is only allowed for PQ_FREE<->PQ_HOLD. 1367 */ 1368 void 1369 vm_page_unhold(vm_page_t m) 1370 { 1371 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1372 ("vm_page_unhold: pg %p illegal hold_count (%d) or " 1373 "on FREE queue (%d)", 1374 m, m->hold_count, m->queue - m->pc)); 1375 1376 if (atomic_fetchadd_int(&m->hold_count, -1) == 1 && 1377 m->queue - m->pc == PQ_HOLD) { 1378 vm_page_spin_lock(m); 1379 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1380 _vm_page_queue_spin_lock(m); 1381 _vm_page_rem_queue_spinlocked(m); 1382 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1383 _vm_page_queue_spin_unlock(m); 1384 } 1385 vm_page_spin_unlock(m); 1386 } 1387 } 1388 1389 /* 1390 * Create a fictitious page with the specified physical address and 1391 * memory attribute. The memory attribute is the only the machine- 1392 * dependent aspect of a fictitious page that must be initialized. 1393 */ 1394 void 1395 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1396 { 1397 /* 1398 * The page's memattr might have changed since the 1399 * previous initialization. Update the pmap to the 1400 * new memattr. 1401 */ 1402 if ((m->flags & PG_FICTITIOUS) != 0) 1403 goto memattr; 1404 m->phys_addr = paddr; 1405 m->queue = PQ_NONE; 1406 /* Fictitious pages don't use "segind". */ 1407 /* Fictitious pages don't use "order" or "pool". */ 1408 m->flags = PG_FICTITIOUS | PG_UNQUEUED; 1409 m->busy_count = PBUSY_LOCKED; 1410 m->wire_count = 1; 1411 spin_init(&m->spin, "fake_page"); 1412 pmap_page_init(m); 1413 memattr: 1414 pmap_page_set_memattr(m, memattr); 1415 } 1416 1417 /* 1418 * Inserts the given vm_page into the object and object list. 1419 * 1420 * The pagetables are not updated but will presumably fault the page 1421 * in if necessary, or if a kernel page the caller will at some point 1422 * enter the page into the kernel's pmap. We are not allowed to block 1423 * here so we *can't* do this anyway. 1424 * 1425 * This routine may not block. 1426 * This routine must be called with the vm_object held. 1427 * This routine must be called with a critical section held. 1428 * 1429 * This routine returns TRUE if the page was inserted into the object 1430 * successfully, and FALSE if the page already exists in the object. 1431 */ 1432 int 1433 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1434 { 1435 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1436 if (m->object != NULL) 1437 panic("vm_page_insert: already inserted"); 1438 1439 atomic_add_int(&object->generation, 1); 1440 1441 /* 1442 * Associate the VM page with an (object, offset). 1443 * 1444 * The vm_page spin lock is required for interactions with the pmap. 1445 * XXX vm_page_spin_lock() might not be needed for this any more. 1446 */ 1447 vm_page_spin_lock(m); 1448 m->object = object; 1449 m->pindex = pindex; 1450 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1451 m->object = NULL; 1452 m->pindex = 0; 1453 vm_page_spin_unlock(m); 1454 return FALSE; 1455 } 1456 ++object->resident_page_count; 1457 ++mycpu->gd_vmtotal.t_rm; 1458 vm_page_spin_unlock(m); 1459 1460 /* 1461 * Since we are inserting a new and possibly dirty page, 1462 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1463 */ 1464 if ((m->valid & m->dirty) || 1465 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1466 vm_object_set_writeable_dirty(object); 1467 1468 /* 1469 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1470 */ 1471 swap_pager_page_inserted(m); 1472 return TRUE; 1473 } 1474 1475 /* 1476 * Removes the given vm_page_t from the (object,index) table 1477 * 1478 * The page must be BUSY and will remain BUSY on return. 1479 * No other requirements. 1480 * 1481 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1482 * it busy. 1483 * 1484 * NOTE: Caller is responsible for any pmap disposition prior to the 1485 * rename (as the pmap code will not be able to find the entries 1486 * once the object has been disassociated). The caller may choose 1487 * to leave the pmap association intact if this routine is being 1488 * called as part of a rename between shadowed objects. 1489 * 1490 * This routine may not block. 1491 */ 1492 void 1493 vm_page_remove(vm_page_t m) 1494 { 1495 vm_object_t object; 1496 1497 if (m->object == NULL) { 1498 return; 1499 } 1500 1501 if ((m->busy_count & PBUSY_LOCKED) == 0) 1502 panic("vm_page_remove: page not busy"); 1503 1504 object = m->object; 1505 1506 vm_object_hold(object); 1507 1508 /* 1509 * Remove the page from the object and update the object. 1510 * 1511 * The vm_page spin lock is required for interactions with the pmap. 1512 * XXX vm_page_spin_lock() might not be needed for this any more. 1513 */ 1514 vm_page_spin_lock(m); 1515 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1516 --object->resident_page_count; 1517 --mycpu->gd_vmtotal.t_rm; 1518 m->object = NULL; 1519 atomic_add_int(&object->generation, 1); 1520 vm_page_spin_unlock(m); 1521 1522 vm_object_drop(object); 1523 } 1524 1525 /* 1526 * Calculate the hash position for the vm_page hash heuristic. 1527 * 1528 * Mask by ~3 to offer 4-way set-assoc 1529 */ 1530 static __inline 1531 struct vm_page_hash_elm * 1532 vm_page_hash_hash(vm_object_t object, vm_pindex_t pindex) 1533 { 1534 size_t hi; 1535 1536 /* mix it up */ 1537 hi = (intptr_t)object ^ object->pg_color ^ pindex; 1538 hi += object->pg_color * pindex; 1539 hi = hi ^ (hi >> 20); 1540 hi &= vm_page_hash_size - 1; /* bounds */ 1541 hi &= ~(VM_PAGE_HASH_SET - 1); /* set-assoc */ 1542 return (&vm_page_hash[hi]); 1543 } 1544 1545 /* 1546 * Heuristical page lookup that does not require any locks. Returns 1547 * a soft-busied page on success, NULL on failure. 1548 * 1549 * Caller must lookup the page the slow way if NULL is returned. 1550 */ 1551 vm_page_t 1552 vm_page_hash_get(vm_object_t object, vm_pindex_t pindex) 1553 { 1554 struct vm_page_hash_elm *mp; 1555 vm_page_t m; 1556 int i; 1557 1558 if (vm_page_hash == NULL) 1559 return NULL; 1560 mp = vm_page_hash_hash(object, pindex); 1561 for (i = 0; i < VM_PAGE_HASH_SET; ++i) { 1562 m = mp[i].m; 1563 cpu_ccfence(); 1564 if (m == NULL) 1565 continue; 1566 if (m->object != object || m->pindex != pindex) 1567 continue; 1568 if (vm_page_sbusy_try(m)) 1569 continue; 1570 if (m->object == object && m->pindex == pindex) { 1571 mp[i].ticks = ticks; 1572 return m; 1573 } 1574 vm_page_sbusy_drop(m); 1575 } 1576 return NULL; 1577 } 1578 1579 /* 1580 * Enter page onto vm_page_hash[]. This is a heuristic, SMP collisions 1581 * are allowed. 1582 */ 1583 static __inline 1584 void 1585 vm_page_hash_enter(vm_page_t m) 1586 { 1587 struct vm_page_hash_elm *mp; 1588 struct vm_page_hash_elm *best; 1589 int i; 1590 1591 /* 1592 * Only enter type-stable vm_pages with well-shared objects. 1593 */ 1594 if (vm_page_hash == NULL || 1595 m < &vm_page_array[0] || 1596 m >= &vm_page_array[vm_page_array_size]) 1597 return; 1598 if (m->object == NULL) 1599 return; 1600 #if 0 1601 /* 1602 * Disabled at the moment, there are some degenerate conditions 1603 * with often-exec'd programs that get ignored. In particular, 1604 * the kernel's elf loader does a vn_rdwr() on the first page of 1605 * a binary. 1606 */ 1607 if (m->object->ref_count <= 2 || (m->object->flags & OBJ_ONEMAPPING)) 1608 return; 1609 #endif 1610 if (vm_page_hash_vnode_only && m->object->type != OBJT_VNODE) 1611 return; 1612 1613 /* 1614 * 1615 */ 1616 mp = vm_page_hash_hash(m->object, m->pindex); 1617 best = mp; 1618 for (i = 0; i < VM_PAGE_HASH_SET; ++i) { 1619 if (mp[i].m == m) { 1620 mp[i].ticks = ticks; 1621 return; 1622 } 1623 1624 /* 1625 * The best choice is the oldest entry 1626 */ 1627 if ((ticks - best->ticks) < (ticks - mp[i].ticks) || 1628 (int)(ticks - mp[i].ticks) < 0) { 1629 best = &mp[i]; 1630 } 1631 } 1632 best->m = m; 1633 best->ticks = ticks; 1634 } 1635 1636 /* 1637 * Locate and return the page at (object, pindex), or NULL if the 1638 * page could not be found. 1639 * 1640 * The caller must hold the vm_object token. 1641 */ 1642 vm_page_t 1643 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1644 { 1645 vm_page_t m; 1646 1647 /* 1648 * Search the hash table for this object/offset pair 1649 */ 1650 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1651 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1652 if (m) { 1653 KKASSERT(m->object == object && m->pindex == pindex); 1654 vm_page_hash_enter(m); 1655 } 1656 return(m); 1657 } 1658 1659 vm_page_t 1660 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1661 vm_pindex_t pindex, 1662 int also_m_busy, const char *msg 1663 VM_PAGE_DEBUG_ARGS) 1664 { 1665 u_int32_t busy_count; 1666 vm_page_t m; 1667 1668 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1669 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1670 while (m) { 1671 KKASSERT(m->object == object && m->pindex == pindex); 1672 busy_count = m->busy_count; 1673 cpu_ccfence(); 1674 if (busy_count & PBUSY_LOCKED) { 1675 tsleep_interlock(m, 0); 1676 if (atomic_cmpset_int(&m->busy_count, busy_count, 1677 busy_count | PBUSY_WANTED)) { 1678 atomic_set_int(&m->flags, PG_REFERENCED); 1679 tsleep(m, PINTERLOCKED, msg, 0); 1680 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1681 pindex); 1682 } 1683 } else if (also_m_busy && busy_count) { 1684 tsleep_interlock(m, 0); 1685 if (atomic_cmpset_int(&m->busy_count, busy_count, 1686 busy_count | PBUSY_WANTED)) { 1687 atomic_set_int(&m->flags, PG_REFERENCED); 1688 tsleep(m, PINTERLOCKED, msg, 0); 1689 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1690 pindex); 1691 } 1692 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1693 busy_count | PBUSY_LOCKED)) { 1694 #ifdef VM_PAGE_DEBUG 1695 m->busy_func = func; 1696 m->busy_line = lineno; 1697 #endif 1698 vm_page_hash_enter(m); 1699 break; 1700 } 1701 } 1702 return m; 1703 } 1704 1705 /* 1706 * Attempt to lookup and busy a page. 1707 * 1708 * Returns NULL if the page could not be found 1709 * 1710 * Returns a vm_page and error == TRUE if the page exists but could not 1711 * be busied. 1712 * 1713 * Returns a vm_page and error == FALSE on success. 1714 */ 1715 vm_page_t 1716 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1717 vm_pindex_t pindex, 1718 int also_m_busy, int *errorp 1719 VM_PAGE_DEBUG_ARGS) 1720 { 1721 u_int32_t busy_count; 1722 vm_page_t m; 1723 1724 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1725 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1726 *errorp = FALSE; 1727 while (m) { 1728 KKASSERT(m->object == object && m->pindex == pindex); 1729 busy_count = m->busy_count; 1730 cpu_ccfence(); 1731 if (busy_count & PBUSY_LOCKED) { 1732 *errorp = TRUE; 1733 break; 1734 } 1735 if (also_m_busy && busy_count) { 1736 *errorp = TRUE; 1737 break; 1738 } 1739 if (atomic_cmpset_int(&m->busy_count, busy_count, 1740 busy_count | PBUSY_LOCKED)) { 1741 #ifdef VM_PAGE_DEBUG 1742 m->busy_func = func; 1743 m->busy_line = lineno; 1744 #endif 1745 vm_page_hash_enter(m); 1746 break; 1747 } 1748 } 1749 return m; 1750 } 1751 1752 /* 1753 * Returns a page that is only soft-busied for use by the caller in 1754 * a read-only fashion. Returns NULL if the page could not be found, 1755 * the soft busy could not be obtained, or the page data is invalid. 1756 * 1757 * XXX Doesn't handle PG_FICTITIOUS pages at the moment, but there is 1758 * no reason why we couldn't. 1759 */ 1760 vm_page_t 1761 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1762 int pgoff, int pgbytes) 1763 { 1764 vm_page_t m; 1765 1766 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1767 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1768 if (m) { 1769 if ((m->valid != VM_PAGE_BITS_ALL && 1770 !vm_page_is_valid(m, pgoff, pgbytes)) || 1771 (m->flags & PG_FICTITIOUS)) { 1772 m = NULL; 1773 } else if (vm_page_sbusy_try(m)) { 1774 m = NULL; 1775 } else if ((m->valid != VM_PAGE_BITS_ALL && 1776 !vm_page_is_valid(m, pgoff, pgbytes)) || 1777 (m->flags & PG_FICTITIOUS)) { 1778 vm_page_sbusy_drop(m); 1779 m = NULL; 1780 } else { 1781 vm_page_hash_enter(m); 1782 } 1783 } 1784 return m; 1785 } 1786 1787 /* 1788 * Caller must hold the related vm_object 1789 */ 1790 vm_page_t 1791 vm_page_next(vm_page_t m) 1792 { 1793 vm_page_t next; 1794 1795 next = vm_page_rb_tree_RB_NEXT(m); 1796 if (next && next->pindex != m->pindex + 1) 1797 next = NULL; 1798 return (next); 1799 } 1800 1801 /* 1802 * vm_page_rename() 1803 * 1804 * Move the given vm_page from its current object to the specified 1805 * target object/offset. The page must be busy and will remain so 1806 * on return. 1807 * 1808 * new_object must be held. 1809 * This routine might block. XXX ? 1810 * 1811 * NOTE: Swap associated with the page must be invalidated by the move. We 1812 * have to do this for several reasons: (1) we aren't freeing the 1813 * page, (2) we are dirtying the page, (3) the VM system is probably 1814 * moving the page from object A to B, and will then later move 1815 * the backing store from A to B and we can't have a conflict. 1816 * 1817 * NOTE: We *always* dirty the page. It is necessary both for the 1818 * fact that we moved it, and because we may be invalidating 1819 * swap. If the page is on the cache, we have to deactivate it 1820 * or vm_page_dirty() will panic. Dirty pages are not allowed 1821 * on the cache. 1822 * 1823 * NOTE: Caller is responsible for any pmap disposition prior to the 1824 * rename (as the pmap code will not be able to find the entries 1825 * once the object has been disassociated or changed). Nominally 1826 * the caller is moving a page between shadowed objects and so the 1827 * pmap association is retained without having to remove the page 1828 * from it. 1829 */ 1830 void 1831 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1832 { 1833 KKASSERT(m->busy_count & PBUSY_LOCKED); 1834 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1835 if (m->object) { 1836 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1837 vm_page_remove(m); 1838 } 1839 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1840 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1841 new_object, new_pindex); 1842 } 1843 if (m->queue - m->pc == PQ_CACHE) 1844 vm_page_deactivate(m); 1845 vm_page_dirty(m); 1846 } 1847 1848 /* 1849 * vm_page_unqueue() without any wakeup. This routine is used when a page 1850 * is to remain BUSYied by the caller. 1851 * 1852 * This routine may not block. 1853 */ 1854 void 1855 vm_page_unqueue_nowakeup(vm_page_t m) 1856 { 1857 vm_page_and_queue_spin_lock(m); 1858 (void)_vm_page_rem_queue_spinlocked(m); 1859 vm_page_spin_unlock(m); 1860 } 1861 1862 /* 1863 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1864 * if necessary. 1865 * 1866 * This routine may not block. 1867 */ 1868 void 1869 vm_page_unqueue(vm_page_t m) 1870 { 1871 u_short queue; 1872 1873 vm_page_and_queue_spin_lock(m); 1874 queue = _vm_page_rem_queue_spinlocked(m); 1875 if (queue == PQ_FREE || queue == PQ_CACHE) { 1876 vm_page_spin_unlock(m); 1877 pagedaemon_wakeup(); 1878 } else { 1879 vm_page_spin_unlock(m); 1880 } 1881 } 1882 1883 /* 1884 * vm_page_list_find() 1885 * 1886 * Find a page on the specified queue with color optimization. 1887 * 1888 * The page coloring optimization attempts to locate a page that does 1889 * not overload other nearby pages in the object in the cpu's L1 or L2 1890 * caches. We need this optimization because cpu caches tend to be 1891 * physical caches, while object spaces tend to be virtual. 1892 * 1893 * The page coloring optimization also, very importantly, tries to localize 1894 * memory to cpus and physical sockets. 1895 * 1896 * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the 1897 * algorithm is adjusted to localize allocations on a per-core basis. 1898 * This is done by 'twisting' the colors. 1899 * 1900 * The page is returned spinlocked and removed from its queue (it will 1901 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1902 * is responsible for dealing with the busy-page case (usually by 1903 * deactivating the page and looping). 1904 * 1905 * NOTE: This routine is carefully inlined. A non-inlined version 1906 * is available for outside callers but the only critical path is 1907 * from within this source file. 1908 * 1909 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1910 * represent stable storage, allowing us to order our locks vm_page 1911 * first, then queue. 1912 */ 1913 static __inline 1914 vm_page_t 1915 _vm_page_list_find(int basequeue, int index) 1916 { 1917 struct vpgqueues *pq; 1918 vm_page_t m; 1919 1920 index &= PQ_L2_MASK; 1921 pq = &vm_page_queues[basequeue + index]; 1922 1923 /* 1924 * Try this cpu's colored queue first. Test for a page unlocked, 1925 * then lock the queue and locate a page. Note that the lock order 1926 * is reversed, but we do not want to dwadle on the page spinlock 1927 * anyway as it is held significantly longer than the queue spinlock. 1928 */ 1929 if (TAILQ_FIRST(&pq->pl)) { 1930 spin_lock(&pq->spin); 1931 TAILQ_FOREACH(m, &pq->pl, pageq) { 1932 if (spin_trylock(&m->spin) == 0) 1933 continue; 1934 KKASSERT(m->queue == basequeue + index); 1935 _vm_page_rem_queue_spinlocked(m); 1936 pq->lastq = -1; 1937 return(m); 1938 } 1939 spin_unlock(&pq->spin); 1940 } 1941 1942 /* 1943 * If we are unable to get a page, do a more involved NUMA-aware 1944 * search. However, to avoid re-searching empty queues over and 1945 * over again skip to pq->last if appropriate. 1946 */ 1947 if (pq->lastq >= 0) 1948 index = pq->lastq; 1949 1950 m = _vm_page_list_find2(basequeue, index, &pq->lastq); 1951 1952 return(m); 1953 } 1954 1955 /* 1956 * If we could not find the page in the desired queue try to find it in 1957 * a nearby (NUMA-aware) queue. 1958 */ 1959 static vm_page_t 1960 _vm_page_list_find2(int basequeue, int index, int *lastp) 1961 { 1962 struct vpgqueues *pq; 1963 vm_page_t m = NULL; 1964 int pqmask = set_assoc_mask >> 1; 1965 int pqi; 1966 int range; 1967 int skip_start; 1968 int skip_next; 1969 int count; 1970 1971 index &= PQ_L2_MASK; 1972 pq = &vm_page_queues[basequeue]; 1973 count = 0; 1974 skip_start = -1; 1975 skip_next = -1; 1976 1977 /* 1978 * Run local sets of 16, 32, 64, 128, up to the entire queue if all 1979 * else fails (PQ_L2_MASK). 1980 * 1981 * pqmask is a mask, 15, 31, 63, etc. 1982 * 1983 * Test each queue unlocked first, then lock the queue and locate 1984 * a page. Note that the lock order is reversed, but we do not want 1985 * to dwadle on the page spinlock anyway as it is held significantly 1986 * longer than the queue spinlock. 1987 */ 1988 do { 1989 pqmask = (pqmask << 1) | 1; 1990 1991 pqi = index; 1992 range = pqmask + 1; 1993 1994 while (range > 0) { 1995 if (pqi >= skip_start && pqi < skip_next) { 1996 range -= skip_next - pqi; 1997 pqi = (pqi & ~pqmask) | (skip_next & pqmask); 1998 } 1999 if (range > 0 && TAILQ_FIRST(&pq[pqi].pl)) { 2000 spin_lock(&pq[pqi].spin); 2001 TAILQ_FOREACH(m, &pq[pqi].pl, pageq) { 2002 if (spin_trylock(&m->spin) == 0) 2003 continue; 2004 KKASSERT(m->queue == basequeue + pqi); 2005 _vm_page_rem_queue_spinlocked(m); 2006 2007 /* 2008 * If we had to wander too far, set 2009 * *lastp to skip past empty queues. 2010 */ 2011 if (count >= 8) 2012 *lastp = pqi & PQ_L2_MASK; 2013 return(m); 2014 } 2015 spin_unlock(&pq[pqi].spin); 2016 } 2017 --range; 2018 ++count; 2019 pqi = (pqi & ~pqmask) | ((pqi + 1) & pqmask); 2020 } 2021 skip_start = pqi & ~pqmask; 2022 skip_next = (pqi | pqmask) + 1; 2023 } while (pqmask != PQ_L2_MASK); 2024 2025 return(m); 2026 } 2027 2028 /* 2029 * Returns a vm_page candidate for allocation. The page is not busied so 2030 * it can move around. The caller must busy the page (and typically 2031 * deactivate it if it cannot be busied!) 2032 * 2033 * Returns a spinlocked vm_page that has been removed from its queue. 2034 */ 2035 vm_page_t 2036 vm_page_list_find(int basequeue, int index) 2037 { 2038 return(_vm_page_list_find(basequeue, index)); 2039 } 2040 2041 /* 2042 * Find a page on the cache queue with color optimization, remove it 2043 * from the queue, and busy it. The returned page will not be spinlocked. 2044 * 2045 * A candidate failure will be deactivated. Candidates can fail due to 2046 * being busied by someone else, in which case they will be deactivated. 2047 * 2048 * This routine may not block. 2049 * 2050 */ 2051 static vm_page_t 2052 vm_page_select_cache(u_short pg_color) 2053 { 2054 vm_page_t m; 2055 2056 for (;;) { 2057 m = _vm_page_list_find(PQ_CACHE, pg_color); 2058 if (m == NULL) 2059 break; 2060 /* 2061 * (m) has been removed from its queue and spinlocked 2062 */ 2063 if (vm_page_busy_try(m, TRUE)) { 2064 _vm_page_deactivate_locked(m, 0); 2065 vm_page_spin_unlock(m); 2066 } else { 2067 /* 2068 * We successfully busied the page 2069 */ 2070 if ((m->flags & PG_NEED_COMMIT) == 0 && 2071 m->hold_count == 0 && 2072 m->wire_count == 0 && 2073 (m->dirty & m->valid) == 0) { 2074 vm_page_spin_unlock(m); 2075 KKASSERT((m->flags & PG_UNQUEUED) == 0); 2076 pagedaemon_wakeup(); 2077 return(m); 2078 } 2079 2080 /* 2081 * The page cannot be recycled, deactivate it. 2082 */ 2083 _vm_page_deactivate_locked(m, 0); 2084 if (_vm_page_wakeup(m)) { 2085 vm_page_spin_unlock(m); 2086 wakeup(m); 2087 } else { 2088 vm_page_spin_unlock(m); 2089 } 2090 } 2091 } 2092 return (m); 2093 } 2094 2095 /* 2096 * Find a free page. We attempt to inline the nominal case and fall back 2097 * to _vm_page_select_free() otherwise. A busied page is removed from 2098 * the queue and returned. 2099 * 2100 * This routine may not block. 2101 */ 2102 static __inline vm_page_t 2103 vm_page_select_free(u_short pg_color) 2104 { 2105 vm_page_t m; 2106 2107 for (;;) { 2108 m = _vm_page_list_find(PQ_FREE, pg_color); 2109 if (m == NULL) 2110 break; 2111 if (vm_page_busy_try(m, TRUE)) { 2112 /* 2113 * Various mechanisms such as a pmap_collect can 2114 * result in a busy page on the free queue. We 2115 * have to move the page out of the way so we can 2116 * retry the allocation. If the other thread is not 2117 * allocating the page then m->valid will remain 0 and 2118 * the pageout daemon will free the page later on. 2119 * 2120 * Since we could not busy the page, however, we 2121 * cannot make assumptions as to whether the page 2122 * will be allocated by the other thread or not, 2123 * so all we can do is deactivate it to move it out 2124 * of the way. In particular, if the other thread 2125 * wires the page it may wind up on the inactive 2126 * queue and the pageout daemon will have to deal 2127 * with that case too. 2128 */ 2129 _vm_page_deactivate_locked(m, 0); 2130 vm_page_spin_unlock(m); 2131 } else { 2132 /* 2133 * Theoretically if we are able to busy the page 2134 * atomic with the queue removal (using the vm_page 2135 * lock) nobody else should have been able to mess 2136 * with the page before us. 2137 * 2138 * Assert the page state. Note that even though 2139 * wiring doesn't adjust queues, a page on the free 2140 * queue should never be wired at this point. 2141 */ 2142 KKASSERT((m->flags & (PG_UNQUEUED | 2143 PG_NEED_COMMIT)) == 0); 2144 KASSERT(m->hold_count == 0, 2145 ("m->hold_count is not zero " 2146 "pg %p q=%d flags=%08x hold=%d wire=%d", 2147 m, m->queue, m->flags, 2148 m->hold_count, m->wire_count)); 2149 KKASSERT(m->wire_count == 0); 2150 vm_page_spin_unlock(m); 2151 pagedaemon_wakeup(); 2152 2153 /* return busied and removed page */ 2154 return(m); 2155 } 2156 } 2157 return(m); 2158 } 2159 2160 /* 2161 * vm_page_alloc() 2162 * 2163 * Allocate and return a memory cell associated with this VM object/offset 2164 * pair. If object is NULL an unassociated page will be allocated. 2165 * 2166 * The returned page will be busied and removed from its queues. This 2167 * routine can block and may return NULL if a race occurs and the page 2168 * is found to already exist at the specified (object, pindex). 2169 * 2170 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 2171 * VM_ALLOC_QUICK like normal but cannot use cache 2172 * VM_ALLOC_SYSTEM greater free drain 2173 * VM_ALLOC_INTERRUPT allow free list to be completely drained 2174 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 2175 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 2176 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 2177 * (see vm_page_grab()) 2178 * VM_ALLOC_USE_GD ok to use per-gd cache 2179 * 2180 * VM_ALLOC_CPU(n) allocate using specified cpu localization 2181 * 2182 * The object must be held if not NULL 2183 * This routine may not block 2184 * 2185 * Additional special handling is required when called from an interrupt 2186 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 2187 * in this case. 2188 */ 2189 vm_page_t 2190 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 2191 { 2192 globaldata_t gd; 2193 vm_object_t obj; 2194 vm_page_t m; 2195 u_short pg_color; 2196 int cpuid_local; 2197 2198 #if 0 2199 /* 2200 * Special per-cpu free VM page cache. The pages are pre-busied 2201 * and pre-zerod for us. 2202 */ 2203 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 2204 crit_enter_gd(gd); 2205 if (gd->gd_vmpg_count) { 2206 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 2207 crit_exit_gd(gd); 2208 goto done; 2209 } 2210 crit_exit_gd(gd); 2211 } 2212 #endif 2213 m = NULL; 2214 2215 /* 2216 * CPU LOCALIZATION 2217 * 2218 * CPU localization algorithm. Break the page queues up by physical 2219 * id and core id (note that two cpu threads will have the same core 2220 * id, and core_id != gd_cpuid). 2221 * 2222 * This is nowhere near perfect, for example the last pindex in a 2223 * subgroup will overflow into the next cpu or package. But this 2224 * should get us good page reuse locality in heavy mixed loads. 2225 * 2226 * (may be executed before the APs are started, so other GDs might 2227 * not exist!) 2228 */ 2229 if (page_req & VM_ALLOC_CPU_SPEC) 2230 cpuid_local = VM_ALLOC_GETCPU(page_req); 2231 else 2232 cpuid_local = mycpu->gd_cpuid; 2233 2234 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 2235 2236 KKASSERT(page_req & 2237 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 2238 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2239 2240 /* 2241 * Certain system threads (pageout daemon, buf_daemon's) are 2242 * allowed to eat deeper into the free page list. 2243 */ 2244 if (curthread->td_flags & TDF_SYSTHREAD) 2245 page_req |= VM_ALLOC_SYSTEM; 2246 2247 /* 2248 * Impose various limitations. Note that the v_free_reserved test 2249 * must match the opposite of vm_page_count_target() to avoid 2250 * livelocks, be careful. 2251 */ 2252 loop: 2253 gd = mycpu; 2254 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 2255 ((page_req & VM_ALLOC_INTERRUPT) && 2256 gd->gd_vmstats.v_free_count > 0) || 2257 ((page_req & VM_ALLOC_SYSTEM) && 2258 gd->gd_vmstats.v_cache_count == 0 && 2259 gd->gd_vmstats.v_free_count > 2260 gd->gd_vmstats.v_interrupt_free_min) 2261 ) { 2262 /* 2263 * The free queue has sufficient free pages to take one out. 2264 */ 2265 m = vm_page_select_free(pg_color); 2266 } else if (page_req & VM_ALLOC_NORMAL) { 2267 /* 2268 * Allocatable from the cache (non-interrupt only). On 2269 * success, we must free the page and try again, thus 2270 * ensuring that vmstats.v_*_free_min counters are replenished. 2271 */ 2272 #ifdef INVARIANTS 2273 if (curthread->td_preempted) { 2274 kprintf("vm_page_alloc(): warning, attempt to allocate" 2275 " cache page from preempting interrupt\n"); 2276 m = NULL; 2277 } else { 2278 m = vm_page_select_cache(pg_color); 2279 } 2280 #else 2281 m = vm_page_select_cache(pg_color); 2282 #endif 2283 /* 2284 * On success move the page into the free queue and loop. 2285 * 2286 * Only do this if we can safely acquire the vm_object lock, 2287 * because this is effectively a random page and the caller 2288 * might be holding the lock shared, we don't want to 2289 * deadlock. 2290 */ 2291 if (m != NULL) { 2292 KASSERT(m->dirty == 0, 2293 ("Found dirty cache page %p", m)); 2294 if ((obj = m->object) != NULL) { 2295 if (vm_object_hold_try(obj)) { 2296 vm_page_protect(m, VM_PROT_NONE); 2297 vm_page_free(m); 2298 /* m->object NULL here */ 2299 vm_object_drop(obj); 2300 } else { 2301 vm_page_deactivate(m); 2302 vm_page_wakeup(m); 2303 } 2304 } else { 2305 vm_page_protect(m, VM_PROT_NONE); 2306 vm_page_free(m); 2307 } 2308 goto loop; 2309 } 2310 2311 /* 2312 * On failure return NULL 2313 */ 2314 atomic_add_int(&vm_pageout_deficit, 1); 2315 pagedaemon_wakeup(); 2316 return (NULL); 2317 } else { 2318 /* 2319 * No pages available, wakeup the pageout daemon and give up. 2320 */ 2321 atomic_add_int(&vm_pageout_deficit, 1); 2322 pagedaemon_wakeup(); 2323 return (NULL); 2324 } 2325 2326 /* 2327 * v_free_count can race so loop if we don't find the expected 2328 * page. 2329 */ 2330 if (m == NULL) { 2331 vmstats_rollup(); 2332 goto loop; 2333 } 2334 2335 /* 2336 * Good page found. The page has already been busied for us and 2337 * removed from its queues. 2338 */ 2339 KASSERT(m->dirty == 0, 2340 ("vm_page_alloc: free/cache page %p was dirty", m)); 2341 KKASSERT(m->queue == PQ_NONE); 2342 2343 #if 0 2344 done: 2345 #endif 2346 /* 2347 * Initialize the structure, inheriting some flags but clearing 2348 * all the rest. The page has already been busied for us. 2349 */ 2350 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 2351 2352 KKASSERT(m->wire_count == 0); 2353 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 2354 m->act_count = 0; 2355 m->valid = 0; 2356 2357 /* 2358 * Caller must be holding the object lock (asserted by 2359 * vm_page_insert()). 2360 * 2361 * NOTE: Inserting a page here does not insert it into any pmaps 2362 * (which could cause us to block allocating memory). 2363 * 2364 * NOTE: If no object an unassociated page is allocated, m->pindex 2365 * can be used by the caller for any purpose. 2366 */ 2367 if (object) { 2368 if (vm_page_insert(m, object, pindex) == FALSE) { 2369 vm_page_free(m); 2370 if ((page_req & VM_ALLOC_NULL_OK) == 0) 2371 panic("PAGE RACE %p[%ld]/%p", 2372 object, (long)pindex, m); 2373 m = NULL; 2374 } 2375 } else { 2376 m->pindex = pindex; 2377 } 2378 2379 /* 2380 * Don't wakeup too often - wakeup the pageout daemon when 2381 * we would be nearly out of memory. 2382 */ 2383 pagedaemon_wakeup(); 2384 2385 /* 2386 * A BUSY page is returned. 2387 */ 2388 return (m); 2389 } 2390 2391 /* 2392 * Returns number of pages available in our DMA memory reserve 2393 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 2394 */ 2395 vm_size_t 2396 vm_contig_avail_pages(void) 2397 { 2398 alist_blk_t blk; 2399 alist_blk_t count; 2400 alist_blk_t bfree; 2401 spin_lock(&vm_contig_spin); 2402 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 2403 spin_unlock(&vm_contig_spin); 2404 2405 return bfree; 2406 } 2407 2408 /* 2409 * Attempt to allocate contiguous physical memory with the specified 2410 * requirements. 2411 */ 2412 vm_page_t 2413 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 2414 unsigned long alignment, unsigned long boundary, 2415 unsigned long size, vm_memattr_t memattr) 2416 { 2417 alist_blk_t blk; 2418 vm_page_t m; 2419 vm_pindex_t i; 2420 #if 0 2421 static vm_pindex_t contig_rover; 2422 #endif 2423 2424 alignment >>= PAGE_SHIFT; 2425 if (alignment == 0) 2426 alignment = 1; 2427 boundary >>= PAGE_SHIFT; 2428 if (boundary == 0) 2429 boundary = 1; 2430 size = (size + PAGE_MASK) >> PAGE_SHIFT; 2431 2432 #if 0 2433 /* 2434 * Disabled temporarily until we find a solution for DRM (a flag 2435 * to always use the free space reserve, for performance). 2436 */ 2437 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 2438 boundary <= PAGE_SIZE && size == 1 && 2439 memattr == VM_MEMATTR_DEFAULT) { 2440 /* 2441 * Any page will work, use vm_page_alloc() 2442 * (e.g. when used from kmem_alloc_attr()) 2443 */ 2444 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 2445 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2446 VM_ALLOC_INTERRUPT); 2447 m->valid = VM_PAGE_BITS_ALL; 2448 vm_page_wire(m); 2449 vm_page_wakeup(m); 2450 } else 2451 #endif 2452 { 2453 /* 2454 * Use the low-memory dma reserve 2455 */ 2456 spin_lock(&vm_contig_spin); 2457 blk = alist_alloc(&vm_contig_alist, 0, size); 2458 if (blk == ALIST_BLOCK_NONE) { 2459 spin_unlock(&vm_contig_spin); 2460 if (bootverbose) { 2461 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2462 (size << PAGE_SHIFT) / 1024); 2463 print_backtrace(5); 2464 } 2465 return(NULL); 2466 } 2467 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2468 alist_free(&vm_contig_alist, blk, size); 2469 spin_unlock(&vm_contig_spin); 2470 if (bootverbose) { 2471 kprintf("vm_page_alloc_contig: %ldk high " 2472 "%016jx failed\n", 2473 (size << PAGE_SHIFT) / 1024, 2474 (intmax_t)high); 2475 } 2476 return(NULL); 2477 } 2478 spin_unlock(&vm_contig_spin); 2479 2480 /* 2481 * Base vm_page_t of range 2482 */ 2483 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2484 } 2485 if (vm_contig_verbose) { 2486 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2487 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2488 (intmax_t)m->phys_addr, 2489 (size << PAGE_SHIFT) / 1024, 2490 low, high, alignment, boundary, size, memattr); 2491 } 2492 if (memattr != VM_MEMATTR_DEFAULT) { 2493 for (i = 0; i < size; ++i) { 2494 KKASSERT(m[i].flags & PG_FICTITIOUS); 2495 pmap_page_set_memattr(&m[i], memattr); 2496 } 2497 } 2498 return m; 2499 } 2500 2501 /* 2502 * Free contiguously allocated pages. The pages will be wired but not busy. 2503 * When freeing to the alist we leave them wired and not busy. 2504 */ 2505 void 2506 vm_page_free_contig(vm_page_t m, unsigned long size) 2507 { 2508 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2509 vm_pindex_t start = pa >> PAGE_SHIFT; 2510 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2511 2512 if (vm_contig_verbose) { 2513 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2514 (intmax_t)pa, size / 1024); 2515 } 2516 if (pa < vm_low_phys_reserved) { 2517 /* 2518 * Just assert check the first page for convenience. 2519 */ 2520 KKASSERT(m->wire_count == 1); 2521 KKASSERT(m->flags & PG_FICTITIOUS); 2522 KKASSERT(pa + size <= vm_low_phys_reserved); 2523 spin_lock(&vm_contig_spin); 2524 alist_free(&vm_contig_alist, start, pages); 2525 spin_unlock(&vm_contig_spin); 2526 } else { 2527 while (pages) { 2528 /* XXX FUTURE, maybe (pair with vm_pg_contig_alloc()) */ 2529 /*vm_page_flag_clear(m, PG_FICTITIOUS | PG_UNQUEUED);*/ 2530 vm_page_busy_wait(m, FALSE, "cpgfr"); 2531 vm_page_unwire(m, 0); 2532 vm_page_free(m); 2533 --pages; 2534 ++m; 2535 } 2536 2537 } 2538 } 2539 2540 2541 /* 2542 * Wait for sufficient free memory for nominal heavy memory use kernel 2543 * operations. 2544 * 2545 * WARNING! Be sure never to call this in any vm_pageout code path, which 2546 * will trivially deadlock the system. 2547 */ 2548 void 2549 vm_wait_nominal(void) 2550 { 2551 while (vm_page_count_min(0)) 2552 vm_wait(0); 2553 } 2554 2555 /* 2556 * Test if vm_wait_nominal() would block. 2557 */ 2558 int 2559 vm_test_nominal(void) 2560 { 2561 if (vm_page_count_min(0)) 2562 return(1); 2563 return(0); 2564 } 2565 2566 /* 2567 * Block until free pages are available for allocation, called in various 2568 * places before memory allocations. 2569 * 2570 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2571 * more generous then that. 2572 */ 2573 void 2574 vm_wait(int timo) 2575 { 2576 /* 2577 * never wait forever 2578 */ 2579 if (timo == 0) 2580 timo = hz; 2581 lwkt_gettoken(&vm_token); 2582 2583 if (curthread == pagethread || 2584 curthread == emergpager) { 2585 /* 2586 * The pageout daemon itself needs pages, this is bad. 2587 */ 2588 if (vm_page_count_min(0)) { 2589 vm_pageout_pages_needed = 1; 2590 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2591 } 2592 } else { 2593 /* 2594 * Wakeup the pageout daemon if necessary and wait. 2595 * 2596 * Do not wait indefinitely for the target to be reached, 2597 * as load might prevent it from being reached any time soon. 2598 * But wait a little to try to slow down page allocations 2599 * and to give more important threads (the pagedaemon) 2600 * allocation priority. 2601 */ 2602 if (vm_page_count_target()) { 2603 if (vm_pages_needed == 0) { 2604 vm_pages_needed = 1; 2605 wakeup(&vm_pages_needed); 2606 } 2607 ++vm_pages_waiting; /* SMP race ok */ 2608 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2609 } 2610 } 2611 lwkt_reltoken(&vm_token); 2612 } 2613 2614 /* 2615 * Block until free pages are available for allocation 2616 * 2617 * Called only from vm_fault so that processes page faulting can be 2618 * easily tracked. 2619 */ 2620 void 2621 vm_wait_pfault(void) 2622 { 2623 /* 2624 * Wakeup the pageout daemon if necessary and wait. 2625 * 2626 * Do not wait indefinitely for the target to be reached, 2627 * as load might prevent it from being reached any time soon. 2628 * But wait a little to try to slow down page allocations 2629 * and to give more important threads (the pagedaemon) 2630 * allocation priority. 2631 */ 2632 if (vm_page_count_min(0)) { 2633 lwkt_gettoken(&vm_token); 2634 while (vm_page_count_severe()) { 2635 if (vm_page_count_target()) { 2636 thread_t td; 2637 2638 if (vm_pages_needed == 0) { 2639 vm_pages_needed = 1; 2640 wakeup(&vm_pages_needed); 2641 } 2642 ++vm_pages_waiting; /* SMP race ok */ 2643 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2644 2645 /* 2646 * Do not stay stuck in the loop if the system is trying 2647 * to kill the process. 2648 */ 2649 td = curthread; 2650 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2651 break; 2652 } 2653 } 2654 lwkt_reltoken(&vm_token); 2655 } 2656 } 2657 2658 /* 2659 * Put the specified page on the active list (if appropriate). Ensure 2660 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2661 * 2662 * The caller should be holding the page busied ? XXX 2663 * This routine may not block. 2664 * 2665 * It is ok if the page is wired (so buffer cache operations don't have 2666 * to mess with the page queues). 2667 */ 2668 void 2669 vm_page_activate(vm_page_t m) 2670 { 2671 u_short oqueue; 2672 2673 /* 2674 * If already active or inappropriate, just set act_count and 2675 * return. We don't have to spin-lock the page. 2676 */ 2677 if (m->queue - m->pc == PQ_ACTIVE || 2678 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2679 if (m->act_count < ACT_INIT) 2680 m->act_count = ACT_INIT; 2681 return; 2682 } 2683 2684 vm_page_spin_lock(m); 2685 if (m->queue - m->pc != PQ_ACTIVE && 2686 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 2687 _vm_page_queue_spin_lock(m); 2688 oqueue = _vm_page_rem_queue_spinlocked(m); 2689 /* page is left spinlocked, queue is unlocked */ 2690 2691 if (oqueue == PQ_CACHE) 2692 mycpu->gd_cnt.v_reactivated++; 2693 if (m->act_count < ACT_INIT) 2694 m->act_count = ACT_INIT; 2695 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2696 _vm_page_and_queue_spin_unlock(m); 2697 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2698 pagedaemon_wakeup(); 2699 } else { 2700 if (m->act_count < ACT_INIT) 2701 m->act_count = ACT_INIT; 2702 vm_page_spin_unlock(m); 2703 } 2704 } 2705 2706 void 2707 vm_page_soft_activate(vm_page_t m) 2708 { 2709 if (m->queue - m->pc == PQ_ACTIVE || 2710 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2711 if (m->act_count < ACT_INIT) 2712 m->act_count = ACT_INIT; 2713 } else { 2714 vm_page_activate(m); 2715 } 2716 } 2717 2718 /* 2719 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2720 * routine is called when a page has been added to the cache or free 2721 * queues. 2722 * 2723 * This routine may not block. 2724 */ 2725 static __inline void 2726 vm_page_free_wakeup(void) 2727 { 2728 globaldata_t gd = mycpu; 2729 2730 /* 2731 * If the pageout daemon itself needs pages, then tell it that 2732 * there are some free. 2733 */ 2734 if (vm_pageout_pages_needed && 2735 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2736 gd->gd_vmstats.v_pageout_free_min 2737 ) { 2738 vm_pageout_pages_needed = 0; 2739 wakeup(&vm_pageout_pages_needed); 2740 } 2741 2742 /* 2743 * Wakeup processes that are waiting on memory. 2744 * 2745 * Generally speaking we want to wakeup stuck processes as soon as 2746 * possible. !vm_page_count_min(0) is the absolute minimum point 2747 * where we can do this. Wait a bit longer to reduce degenerate 2748 * re-blocking (vm_page_free_hysteresis). The target check is just 2749 * to make sure the min-check w/hysteresis does not exceed the 2750 * normal target. 2751 */ 2752 if (vm_pages_waiting) { 2753 if (!vm_page_count_min(vm_page_free_hysteresis) || 2754 !vm_page_count_target()) { 2755 vm_pages_waiting = 0; 2756 wakeup(&vmstats.v_free_count); 2757 ++mycpu->gd_cnt.v_ppwakeups; 2758 } 2759 #if 0 2760 if (!vm_page_count_target()) { 2761 /* 2762 * Plenty of pages are free, wakeup everyone. 2763 */ 2764 vm_pages_waiting = 0; 2765 wakeup(&vmstats.v_free_count); 2766 ++mycpu->gd_cnt.v_ppwakeups; 2767 } else if (!vm_page_count_min(0)) { 2768 /* 2769 * Some pages are free, wakeup someone. 2770 */ 2771 int wcount = vm_pages_waiting; 2772 if (wcount > 0) 2773 --wcount; 2774 vm_pages_waiting = wcount; 2775 wakeup_one(&vmstats.v_free_count); 2776 ++mycpu->gd_cnt.v_ppwakeups; 2777 } 2778 #endif 2779 } 2780 } 2781 2782 /* 2783 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2784 * it from its VM object. 2785 * 2786 * The vm_page must be BUSY on entry. BUSY will be released on 2787 * return (the page will have been freed). 2788 */ 2789 void 2790 vm_page_free_toq(vm_page_t m) 2791 { 2792 mycpu->gd_cnt.v_tfree++; 2793 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) 2794 pmap_mapped_sync(m); 2795 KKASSERT((m->flags & PG_MAPPED) == 0); 2796 KKASSERT(m->busy_count & PBUSY_LOCKED); 2797 2798 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2799 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2800 "hold(%d)\n", 2801 (u_long)m->pindex, m->busy_count, m->hold_count); 2802 if ((m->queue - m->pc) == PQ_FREE) 2803 panic("vm_page_free: freeing free page"); 2804 else 2805 panic("vm_page_free: freeing busy page"); 2806 } 2807 2808 /* 2809 * Remove from object, spinlock the page and its queues and 2810 * remove from any queue. No queue spinlock will be held 2811 * after this section (because the page was removed from any 2812 * queue). 2813 */ 2814 vm_page_remove(m); 2815 2816 /* 2817 * No further management of fictitious pages occurs beyond object 2818 * and queue removal. 2819 */ 2820 if ((m->flags & PG_FICTITIOUS) != 0) { 2821 KKASSERT(m->queue == PQ_NONE); 2822 vm_page_wakeup(m); 2823 return; 2824 } 2825 vm_page_and_queue_spin_lock(m); 2826 _vm_page_rem_queue_spinlocked(m); 2827 2828 m->valid = 0; 2829 vm_page_undirty(m); 2830 2831 if (m->wire_count != 0) { 2832 if (m->wire_count > 1) { 2833 panic( 2834 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2835 m->wire_count, (long)m->pindex); 2836 } 2837 panic("vm_page_free: freeing wired page"); 2838 } 2839 2840 /* 2841 * Clear the PG_NEED_COMMIT and the PG_UNQUEUED flags. The 2842 * page returns to normal operation and will be placed in 2843 * the PQ_HOLD or PQ_FREE queue. 2844 */ 2845 vm_page_flag_clear(m, PG_NEED_COMMIT | PG_UNQUEUED); 2846 2847 if (m->hold_count != 0) { 2848 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2849 } else { 2850 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2851 } 2852 2853 /* 2854 * This sequence allows us to clear BUSY while still holding 2855 * its spin lock, which reduces contention vs allocators. We 2856 * must not leave the queue locked or _vm_page_wakeup() may 2857 * deadlock. 2858 */ 2859 _vm_page_queue_spin_unlock(m); 2860 if (_vm_page_wakeup(m)) { 2861 vm_page_spin_unlock(m); 2862 wakeup(m); 2863 } else { 2864 vm_page_spin_unlock(m); 2865 } 2866 vm_page_free_wakeup(); 2867 } 2868 2869 /* 2870 * Mark this page as wired down by yet another map. We do not adjust the 2871 * queue the page is on, it will be checked for wiring as-needed. 2872 * 2873 * This function has no effect on fictitious pages. 2874 * 2875 * Caller must be holding the page busy. 2876 */ 2877 void 2878 vm_page_wire(vm_page_t m) 2879 { 2880 KKASSERT(m->busy_count & PBUSY_LOCKED); 2881 if ((m->flags & PG_FICTITIOUS) == 0) { 2882 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2883 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2884 } 2885 KASSERT(m->wire_count != 0, 2886 ("vm_page_wire: wire_count overflow m=%p", m)); 2887 } 2888 } 2889 2890 /* 2891 * Release one wiring of this page, potentially enabling it to be paged again. 2892 * 2893 * Note that wired pages are no longer unconditionally removed from the 2894 * paging queues, so the page may already be on a queue. Move the page 2895 * to the desired queue if necessary. 2896 * 2897 * Many pages placed on the inactive queue should actually go 2898 * into the cache, but it is difficult to figure out which. What 2899 * we do instead, if the inactive target is well met, is to put 2900 * clean pages at the head of the inactive queue instead of the tail. 2901 * This will cause them to be moved to the cache more quickly and 2902 * if not actively re-referenced, freed more quickly. If we just 2903 * stick these pages at the end of the inactive queue, heavy filesystem 2904 * meta-data accesses can cause an unnecessary paging load on memory bound 2905 * processes. This optimization causes one-time-use metadata to be 2906 * reused more quickly. 2907 * 2908 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2909 * the inactive queue. This helps the pageout daemon determine memory 2910 * pressure and act on out-of-memory situations more quickly. 2911 * 2912 * BUT, if we are in a low-memory situation we have no choice but to 2913 * put clean pages on the cache queue. 2914 * 2915 * A number of routines use vm_page_unwire() to guarantee that the page 2916 * will go into either the inactive or active queues, and will NEVER 2917 * be placed in the cache - for example, just after dirtying a page. 2918 * dirty pages in the cache are not allowed. 2919 * 2920 * PG_FICTITIOUS or PG_UNQUEUED pages are never moved to any queue, and 2921 * the wire_count will not be adjusted in any way for a PG_FICTITIOUS 2922 * page. 2923 * 2924 * This routine may not block. 2925 */ 2926 void 2927 vm_page_unwire(vm_page_t m, int activate) 2928 { 2929 KKASSERT(m->busy_count & PBUSY_LOCKED); 2930 if (m->flags & PG_FICTITIOUS) { 2931 /* do nothing */ 2932 } else if ((int)m->wire_count <= 0) { 2933 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2934 } else { 2935 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2936 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2937 if (m->flags & PG_UNQUEUED) { 2938 ; 2939 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2940 vm_page_activate(m); 2941 } else { 2942 vm_page_deactivate(m); 2943 } 2944 } 2945 } 2946 } 2947 2948 /* 2949 * Move the specified page to the inactive queue. 2950 * 2951 * Normally athead is 0 resulting in LRU operation. athead is set 2952 * to 1 if we want this page to be 'as if it were placed in the cache', 2953 * except without unmapping it from the process address space. 2954 * 2955 * vm_page's spinlock must be held on entry and will remain held on return. 2956 * This routine may not block. The caller does not have to hold the page 2957 * busied but should have some sort of interlock on its validity. 2958 * 2959 * It is ok if the page is wired (so buffer cache operations don't have 2960 * to mess with the page queues). 2961 */ 2962 static void 2963 _vm_page_deactivate_locked(vm_page_t m, int athead) 2964 { 2965 u_short oqueue; 2966 2967 /* 2968 * Ignore if already inactive. 2969 */ 2970 if (m->queue - m->pc == PQ_INACTIVE || 2971 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2972 return; 2973 } 2974 2975 _vm_page_queue_spin_lock(m); 2976 oqueue = _vm_page_rem_queue_spinlocked(m); 2977 2978 if ((m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 2979 if (oqueue == PQ_CACHE) 2980 mycpu->gd_cnt.v_reactivated++; 2981 vm_page_flag_clear(m, PG_WINATCFLS); 2982 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2983 if (athead == 0) { 2984 atomic_add_long( 2985 &vm_page_queues[PQ_INACTIVE + m->pc].adds, 1); 2986 } 2987 } 2988 /* NOTE: PQ_NONE if condition not taken */ 2989 _vm_page_queue_spin_unlock(m); 2990 /* leaves vm_page spinlocked */ 2991 } 2992 2993 /* 2994 * Attempt to deactivate a page. 2995 * 2996 * No requirements. We can pre-filter before getting the spinlock. 2997 * 2998 * It is ok if the page is wired (so buffer cache operations don't have 2999 * to mess with the page queues). 3000 */ 3001 void 3002 vm_page_deactivate(vm_page_t m) 3003 { 3004 if (m->queue - m->pc != PQ_INACTIVE && 3005 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 3006 vm_page_spin_lock(m); 3007 _vm_page_deactivate_locked(m, 0); 3008 vm_page_spin_unlock(m); 3009 } 3010 } 3011 3012 void 3013 vm_page_deactivate_locked(vm_page_t m) 3014 { 3015 _vm_page_deactivate_locked(m, 0); 3016 } 3017 3018 /* 3019 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 3020 * 3021 * This function returns non-zero if it successfully moved the page to 3022 * PQ_CACHE. 3023 * 3024 * This function unconditionally unbusies the page on return. 3025 */ 3026 int 3027 vm_page_try_to_cache(vm_page_t m) 3028 { 3029 /* 3030 * Shortcut if we obviously cannot move the page, or if the 3031 * page is already on the cache queue, or it is ficitious. 3032 * 3033 * Never allow a wired page into the cache. 3034 */ 3035 if (m->dirty || m->hold_count || m->wire_count || 3036 m->queue - m->pc == PQ_CACHE || 3037 (m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS))) { 3038 vm_page_wakeup(m); 3039 return(0); 3040 } 3041 3042 /* 3043 * Page busied by us and no longer spinlocked. Dirty pages cannot 3044 * be moved to the cache. 3045 */ 3046 vm_page_test_dirty(m); 3047 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3048 vm_page_wakeup(m); 3049 return(0); 3050 } 3051 vm_page_cache(m); 3052 return(1); 3053 } 3054 3055 /* 3056 * Attempt to free the page. If we cannot free it, we do nothing. 3057 * 1 is returned on success, 0 on failure. 3058 * 3059 * The page can be in any state, including already being on the free 3060 * queue. Check to see if it really can be freed. Note that we disallow 3061 * this ad-hoc operation if the page is flagged PG_UNQUEUED. 3062 * 3063 * Caller provides an unlocked/non-busied page. 3064 * No requirements. 3065 */ 3066 int 3067 vm_page_try_to_free(vm_page_t m) 3068 { 3069 if (vm_page_busy_try(m, TRUE)) 3070 return(0); 3071 3072 if (m->dirty || /* can't free if it is dirty */ 3073 m->hold_count || /* or held (XXX may be wrong) */ 3074 m->wire_count || /* or wired */ 3075 (m->flags & (PG_UNQUEUED | /* or unqueued */ 3076 PG_NEED_COMMIT | /* or needs a commit */ 3077 PG_FICTITIOUS)) || /* or is fictitious */ 3078 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 3079 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 3080 vm_page_wakeup(m); 3081 return(0); 3082 } 3083 3084 /* 3085 * We can probably free the page. 3086 * 3087 * Page busied by us and no longer spinlocked. Dirty pages will 3088 * not be freed by this function. We have to re-test the 3089 * dirty bit after cleaning out the pmaps. 3090 */ 3091 vm_page_test_dirty(m); 3092 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3093 vm_page_wakeup(m); 3094 return(0); 3095 } 3096 vm_page_protect(m, VM_PROT_NONE); 3097 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3098 vm_page_wakeup(m); 3099 return(0); 3100 } 3101 vm_page_free(m); 3102 return(1); 3103 } 3104 3105 /* 3106 * vm_page_cache 3107 * 3108 * Put the specified page onto the page cache queue (if appropriate). 3109 * 3110 * The page must be busy, and this routine will release the busy and 3111 * possibly even free the page. 3112 */ 3113 void 3114 vm_page_cache(vm_page_t m) 3115 { 3116 /* 3117 * Not suitable for the cache 3118 */ 3119 if ((m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS)) || 3120 (m->busy_count & PBUSY_MASK) || 3121 m->wire_count || m->hold_count) { 3122 vm_page_wakeup(m); 3123 return; 3124 } 3125 3126 /* 3127 * Already in the cache (and thus not mapped) 3128 */ 3129 if ((m->queue - m->pc) == PQ_CACHE) { 3130 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) 3131 pmap_mapped_sync(m); 3132 KKASSERT((m->flags & PG_MAPPED) == 0); 3133 vm_page_wakeup(m); 3134 return; 3135 } 3136 3137 #if 0 3138 /* 3139 * REMOVED - it is possible for dirty to get set at any time as 3140 * long as the page is still mapped and writeable. 3141 * 3142 * Caller is required to test m->dirty, but note that the act of 3143 * removing the page from its maps can cause it to become dirty 3144 * on an SMP system due to another cpu running in usermode. 3145 */ 3146 if (m->dirty) { 3147 panic("vm_page_cache: caching a dirty page, pindex: %ld", 3148 (long)m->pindex); 3149 } 3150 #endif 3151 3152 /* 3153 * Remove all pmaps and indicate that the page is not 3154 * writeable or mapped. Our vm_page_protect() call may 3155 * have blocked (especially w/ VM_PROT_NONE), so recheck 3156 * everything. 3157 */ 3158 vm_page_protect(m, VM_PROT_NONE); 3159 pmap_mapped_sync(m); 3160 if ((m->flags & (PG_UNQUEUED | PG_MAPPED)) || 3161 (m->busy_count & PBUSY_MASK) || 3162 m->wire_count || m->hold_count) { 3163 vm_page_wakeup(m); 3164 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3165 vm_page_deactivate(m); 3166 vm_page_wakeup(m); 3167 } else { 3168 _vm_page_and_queue_spin_lock(m); 3169 _vm_page_rem_queue_spinlocked(m); 3170 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 3171 _vm_page_and_queue_spin_unlock(m); 3172 vm_page_wakeup(m); 3173 vm_page_free_wakeup(); 3174 } 3175 } 3176 3177 /* 3178 * vm_page_dontneed() 3179 * 3180 * Cache, deactivate, or do nothing as appropriate. This routine 3181 * is typically used by madvise() MADV_DONTNEED. 3182 * 3183 * Generally speaking we want to move the page into the cache so 3184 * it gets reused quickly. However, this can result in a silly syndrome 3185 * due to the page recycling too quickly. Small objects will not be 3186 * fully cached. On the otherhand, if we move the page to the inactive 3187 * queue we wind up with a problem whereby very large objects 3188 * unnecessarily blow away our inactive and cache queues. 3189 * 3190 * The solution is to move the pages based on a fixed weighting. We 3191 * either leave them alone, deactivate them, or move them to the cache, 3192 * where moving them to the cache has the highest weighting. 3193 * By forcing some pages into other queues we eventually force the 3194 * system to balance the queues, potentially recovering other unrelated 3195 * space from active. The idea is to not force this to happen too 3196 * often. 3197 * 3198 * The page must be busied. 3199 */ 3200 void 3201 vm_page_dontneed(vm_page_t m) 3202 { 3203 static int dnweight; 3204 int dnw; 3205 int head; 3206 3207 dnw = ++dnweight; 3208 3209 /* 3210 * occassionally leave the page alone 3211 */ 3212 if ((dnw & 0x01F0) == 0 || 3213 m->queue - m->pc == PQ_INACTIVE || 3214 m->queue - m->pc == PQ_CACHE 3215 ) { 3216 if (m->act_count >= ACT_INIT) 3217 --m->act_count; 3218 return; 3219 } 3220 3221 /* 3222 * If vm_page_dontneed() is inactivating a page, it must clear 3223 * the referenced flag; otherwise the pagedaemon will see references 3224 * on the page in the inactive queue and reactivate it. Until the 3225 * page can move to the cache queue, madvise's job is not done. 3226 */ 3227 vm_page_flag_clear(m, PG_REFERENCED); 3228 pmap_clear_reference(m); 3229 3230 if (m->dirty == 0) 3231 vm_page_test_dirty(m); 3232 3233 if (m->dirty || (dnw & 0x0070) == 0) { 3234 /* 3235 * Deactivate the page 3 times out of 32. 3236 */ 3237 head = 0; 3238 } else { 3239 /* 3240 * Cache the page 28 times out of every 32. Note that 3241 * the page is deactivated instead of cached, but placed 3242 * at the head of the queue instead of the tail. 3243 */ 3244 head = 1; 3245 } 3246 vm_page_spin_lock(m); 3247 _vm_page_deactivate_locked(m, head); 3248 vm_page_spin_unlock(m); 3249 } 3250 3251 /* 3252 * These routines manipulate the 'soft busy' count for a page. A soft busy 3253 * is almost like a hard BUSY except that it allows certain compatible 3254 * operations to occur on the page while it is busy. For example, a page 3255 * undergoing a write can still be mapped read-only. 3256 * 3257 * We also use soft-busy to quickly pmap_enter shared read-only pages 3258 * without having to hold the page locked. 3259 * 3260 * The soft-busy count can be > 1 in situations where multiple threads 3261 * are pmap_enter()ing the same page simultaneously, or when two buffer 3262 * cache buffers overlap the same page. 3263 * 3264 * The caller must hold the page BUSY when making these two calls. 3265 */ 3266 void 3267 vm_page_io_start(vm_page_t m) 3268 { 3269 uint32_t ocount; 3270 3271 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3272 KKASSERT(ocount & PBUSY_LOCKED); 3273 } 3274 3275 void 3276 vm_page_io_finish(vm_page_t m) 3277 { 3278 uint32_t ocount; 3279 3280 ocount = atomic_fetchadd_int(&m->busy_count, -1); 3281 KKASSERT(ocount & PBUSY_MASK); 3282 #if 0 3283 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 3284 wakeup(m); 3285 #endif 3286 } 3287 3288 /* 3289 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 3290 * 3291 * We can't use fetchadd here because we might race a hard-busy and the 3292 * page freeing code asserts on a non-zero soft-busy count (even if only 3293 * temporary). 3294 * 3295 * Returns 0 on success, non-zero on failure. 3296 */ 3297 int 3298 vm_page_sbusy_try(vm_page_t m) 3299 { 3300 uint32_t ocount; 3301 3302 for (;;) { 3303 ocount = m->busy_count; 3304 cpu_ccfence(); 3305 if (ocount & PBUSY_LOCKED) 3306 return 1; 3307 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1)) 3308 break; 3309 } 3310 return 0; 3311 #if 0 3312 if (m->busy_count & PBUSY_LOCKED) 3313 return 1; 3314 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3315 if (ocount & PBUSY_LOCKED) { 3316 vm_page_sbusy_drop(m); 3317 return 1; 3318 } 3319 return 0; 3320 #endif 3321 } 3322 3323 /* 3324 * Indicate that a clean VM page requires a filesystem commit and cannot 3325 * be reused. Used by tmpfs. 3326 */ 3327 void 3328 vm_page_need_commit(vm_page_t m) 3329 { 3330 vm_page_flag_set(m, PG_NEED_COMMIT); 3331 vm_object_set_writeable_dirty(m->object); 3332 } 3333 3334 void 3335 vm_page_clear_commit(vm_page_t m) 3336 { 3337 vm_page_flag_clear(m, PG_NEED_COMMIT); 3338 } 3339 3340 /* 3341 * Grab a page, blocking if it is busy and allocating a page if necessary. 3342 * A busy page is returned or NULL. The page may or may not be valid and 3343 * might not be on a queue (the caller is responsible for the disposition of 3344 * the page). 3345 * 3346 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 3347 * page will be zero'd and marked valid. 3348 * 3349 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 3350 * valid even if it already exists. 3351 * 3352 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 3353 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 3354 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 3355 * 3356 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 3357 * always returned if we had blocked. 3358 * 3359 * This routine may not be called from an interrupt. 3360 * 3361 * No other requirements. 3362 */ 3363 vm_page_t 3364 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3365 { 3366 vm_page_t m; 3367 int error; 3368 int shared = 1; 3369 3370 KKASSERT(allocflags & 3371 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 3372 vm_object_hold_shared(object); 3373 for (;;) { 3374 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3375 if (error) { 3376 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 3377 if ((allocflags & VM_ALLOC_RETRY) == 0) { 3378 m = NULL; 3379 break; 3380 } 3381 /* retry */ 3382 } else if (m == NULL) { 3383 if (shared) { 3384 vm_object_upgrade(object); 3385 shared = 0; 3386 } 3387 if (allocflags & VM_ALLOC_RETRY) 3388 allocflags |= VM_ALLOC_NULL_OK; 3389 m = vm_page_alloc(object, pindex, 3390 allocflags & ~VM_ALLOC_RETRY); 3391 if (m) 3392 break; 3393 vm_wait(0); 3394 if ((allocflags & VM_ALLOC_RETRY) == 0) 3395 goto failed; 3396 } else { 3397 /* m found */ 3398 break; 3399 } 3400 } 3401 3402 /* 3403 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 3404 * 3405 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 3406 * valid even if already valid. 3407 * 3408 * NOTE! We have removed all of the PG_ZERO optimizations and also 3409 * removed the idle zeroing code. These optimizations actually 3410 * slow things down on modern cpus because the zerod area is 3411 * likely uncached, placing a memory-access burden on the 3412 * accesors taking the fault. 3413 * 3414 * By always zeroing the page in-line with the fault, no 3415 * dynamic ram reads are needed and the caches are hot, ready 3416 * for userland to access the memory. 3417 */ 3418 if (m->valid == 0) { 3419 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 3420 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3421 m->valid = VM_PAGE_BITS_ALL; 3422 } 3423 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 3424 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3425 m->valid = VM_PAGE_BITS_ALL; 3426 } 3427 failed: 3428 vm_object_drop(object); 3429 return(m); 3430 } 3431 3432 /* 3433 * Mapping function for valid bits or for dirty bits in 3434 * a page. May not block. 3435 * 3436 * Inputs are required to range within a page. 3437 * 3438 * No requirements. 3439 * Non blocking. 3440 */ 3441 int 3442 vm_page_bits(int base, int size) 3443 { 3444 int first_bit; 3445 int last_bit; 3446 3447 KASSERT( 3448 base + size <= PAGE_SIZE, 3449 ("vm_page_bits: illegal base/size %d/%d", base, size) 3450 ); 3451 3452 if (size == 0) /* handle degenerate case */ 3453 return(0); 3454 3455 first_bit = base >> DEV_BSHIFT; 3456 last_bit = (base + size - 1) >> DEV_BSHIFT; 3457 3458 return ((2 << last_bit) - (1 << first_bit)); 3459 } 3460 3461 /* 3462 * Sets portions of a page valid and clean. The arguments are expected 3463 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3464 * of any partial chunks touched by the range. The invalid portion of 3465 * such chunks will be zero'd. 3466 * 3467 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 3468 * align base to DEV_BSIZE so as not to mark clean a partially 3469 * truncated device block. Otherwise the dirty page status might be 3470 * lost. 3471 * 3472 * This routine may not block. 3473 * 3474 * (base + size) must be less then or equal to PAGE_SIZE. 3475 */ 3476 static void 3477 _vm_page_zero_valid(vm_page_t m, int base, int size) 3478 { 3479 int frag; 3480 int endoff; 3481 3482 if (size == 0) /* handle degenerate case */ 3483 return; 3484 3485 /* 3486 * If the base is not DEV_BSIZE aligned and the valid 3487 * bit is clear, we have to zero out a portion of the 3488 * first block. 3489 */ 3490 3491 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3492 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3493 ) { 3494 pmap_zero_page_area( 3495 VM_PAGE_TO_PHYS(m), 3496 frag, 3497 base - frag 3498 ); 3499 } 3500 3501 /* 3502 * If the ending offset is not DEV_BSIZE aligned and the 3503 * valid bit is clear, we have to zero out a portion of 3504 * the last block. 3505 */ 3506 3507 endoff = base + size; 3508 3509 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3510 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3511 ) { 3512 pmap_zero_page_area( 3513 VM_PAGE_TO_PHYS(m), 3514 endoff, 3515 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3516 ); 3517 } 3518 } 3519 3520 /* 3521 * Set valid, clear dirty bits. If validating the entire 3522 * page we can safely clear the pmap modify bit. We also 3523 * use this opportunity to clear the PG_NOSYNC flag. If a process 3524 * takes a write fault on a MAP_NOSYNC memory area the flag will 3525 * be set again. 3526 * 3527 * We set valid bits inclusive of any overlap, but we can only 3528 * clear dirty bits for DEV_BSIZE chunks that are fully within 3529 * the range. 3530 * 3531 * Page must be busied? 3532 * No other requirements. 3533 */ 3534 void 3535 vm_page_set_valid(vm_page_t m, int base, int size) 3536 { 3537 _vm_page_zero_valid(m, base, size); 3538 m->valid |= vm_page_bits(base, size); 3539 } 3540 3541 3542 /* 3543 * Set valid bits and clear dirty bits. 3544 * 3545 * Page must be busied by caller. 3546 * 3547 * NOTE: This function does not clear the pmap modified bit. 3548 * Also note that e.g. NFS may use a byte-granular base 3549 * and size. 3550 * 3551 * No other requirements. 3552 */ 3553 void 3554 vm_page_set_validclean(vm_page_t m, int base, int size) 3555 { 3556 int pagebits; 3557 3558 _vm_page_zero_valid(m, base, size); 3559 pagebits = vm_page_bits(base, size); 3560 m->valid |= pagebits; 3561 m->dirty &= ~pagebits; 3562 if (base == 0 && size == PAGE_SIZE) { 3563 /*pmap_clear_modify(m);*/ 3564 vm_page_flag_clear(m, PG_NOSYNC); 3565 } 3566 } 3567 3568 /* 3569 * Set valid & dirty. Used by buwrite() 3570 * 3571 * Page must be busied by caller. 3572 */ 3573 void 3574 vm_page_set_validdirty(vm_page_t m, int base, int size) 3575 { 3576 int pagebits; 3577 3578 pagebits = vm_page_bits(base, size); 3579 m->valid |= pagebits; 3580 m->dirty |= pagebits; 3581 if (m->object) 3582 vm_object_set_writeable_dirty(m->object); 3583 } 3584 3585 /* 3586 * Clear dirty bits. 3587 * 3588 * NOTE: This function does not clear the pmap modified bit. 3589 * Also note that e.g. NFS may use a byte-granular base 3590 * and size. 3591 * 3592 * Page must be busied? 3593 * No other requirements. 3594 */ 3595 void 3596 vm_page_clear_dirty(vm_page_t m, int base, int size) 3597 { 3598 m->dirty &= ~vm_page_bits(base, size); 3599 if (base == 0 && size == PAGE_SIZE) { 3600 /*pmap_clear_modify(m);*/ 3601 vm_page_flag_clear(m, PG_NOSYNC); 3602 } 3603 } 3604 3605 /* 3606 * Make the page all-dirty. 3607 * 3608 * Also make sure the related object and vnode reflect the fact that the 3609 * object may now contain a dirty page. 3610 * 3611 * Page must be busied? 3612 * No other requirements. 3613 */ 3614 void 3615 vm_page_dirty(vm_page_t m) 3616 { 3617 #ifdef INVARIANTS 3618 int pqtype = m->queue - m->pc; 3619 #endif 3620 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3621 ("vm_page_dirty: page in free/cache queue!")); 3622 if (m->dirty != VM_PAGE_BITS_ALL) { 3623 m->dirty = VM_PAGE_BITS_ALL; 3624 if (m->object) 3625 vm_object_set_writeable_dirty(m->object); 3626 } 3627 } 3628 3629 /* 3630 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3631 * valid and dirty bits for the effected areas are cleared. 3632 * 3633 * Page must be busied? 3634 * Does not block. 3635 * No other requirements. 3636 */ 3637 void 3638 vm_page_set_invalid(vm_page_t m, int base, int size) 3639 { 3640 int bits; 3641 3642 bits = vm_page_bits(base, size); 3643 m->valid &= ~bits; 3644 m->dirty &= ~bits; 3645 atomic_add_int(&m->object->generation, 1); 3646 } 3647 3648 /* 3649 * The kernel assumes that the invalid portions of a page contain 3650 * garbage, but such pages can be mapped into memory by user code. 3651 * When this occurs, we must zero out the non-valid portions of the 3652 * page so user code sees what it expects. 3653 * 3654 * Pages are most often semi-valid when the end of a file is mapped 3655 * into memory and the file's size is not page aligned. 3656 * 3657 * Page must be busied? 3658 * No other requirements. 3659 */ 3660 void 3661 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3662 { 3663 int b; 3664 int i; 3665 3666 /* 3667 * Scan the valid bits looking for invalid sections that 3668 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3669 * valid bit may be set ) have already been zerod by 3670 * vm_page_set_validclean(). 3671 */ 3672 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3673 if (i == (PAGE_SIZE / DEV_BSIZE) || 3674 (m->valid & (1 << i)) 3675 ) { 3676 if (i > b) { 3677 pmap_zero_page_area( 3678 VM_PAGE_TO_PHYS(m), 3679 b << DEV_BSHIFT, 3680 (i - b) << DEV_BSHIFT 3681 ); 3682 } 3683 b = i + 1; 3684 } 3685 } 3686 3687 /* 3688 * setvalid is TRUE when we can safely set the zero'd areas 3689 * as being valid. We can do this if there are no cache consistency 3690 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3691 */ 3692 if (setvalid) 3693 m->valid = VM_PAGE_BITS_ALL; 3694 } 3695 3696 /* 3697 * Is a (partial) page valid? Note that the case where size == 0 3698 * will return FALSE in the degenerate case where the page is entirely 3699 * invalid, and TRUE otherwise. 3700 * 3701 * Does not block. 3702 * No other requirements. 3703 */ 3704 int 3705 vm_page_is_valid(vm_page_t m, int base, int size) 3706 { 3707 int bits = vm_page_bits(base, size); 3708 3709 if (m->valid && ((m->valid & bits) == bits)) 3710 return 1; 3711 else 3712 return 0; 3713 } 3714 3715 /* 3716 * Update dirty bits from pmap/mmu. May not block. 3717 * 3718 * Caller must hold the page busy 3719 * 3720 * WARNING! Unless the page has been unmapped, this function only 3721 * provides a likely dirty status. 3722 */ 3723 void 3724 vm_page_test_dirty(vm_page_t m) 3725 { 3726 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) { 3727 vm_page_dirty(m); 3728 } 3729 } 3730 3731 #include "opt_ddb.h" 3732 #ifdef DDB 3733 #include <ddb/ddb.h> 3734 3735 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3736 { 3737 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3738 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3739 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3740 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3741 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3742 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3743 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3744 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3745 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3746 db_printf("vmstats.v_inactive_target: %ld\n", 3747 vmstats.v_inactive_target); 3748 } 3749 3750 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3751 { 3752 int i; 3753 db_printf("PQ_FREE:"); 3754 for (i = 0; i < PQ_L2_SIZE; i++) { 3755 db_printf(" %ld", vm_page_queues[PQ_FREE + i].lcnt); 3756 } 3757 db_printf("\n"); 3758 3759 db_printf("PQ_CACHE:"); 3760 for(i = 0; i < PQ_L2_SIZE; i++) { 3761 db_printf(" %ld", vm_page_queues[PQ_CACHE + i].lcnt); 3762 } 3763 db_printf("\n"); 3764 3765 db_printf("PQ_ACTIVE:"); 3766 for(i = 0; i < PQ_L2_SIZE; i++) { 3767 db_printf(" %ld", vm_page_queues[PQ_ACTIVE + i].lcnt); 3768 } 3769 db_printf("\n"); 3770 3771 db_printf("PQ_INACTIVE:"); 3772 for(i = 0; i < PQ_L2_SIZE; i++) { 3773 db_printf(" %ld", vm_page_queues[PQ_INACTIVE + i].lcnt); 3774 } 3775 db_printf("\n"); 3776 } 3777 #endif /* DDB */ 3778