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