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_adj; 994 long *cnt_gd; 995 996 queue = m->queue; 997 if (queue != PQ_NONE) { 998 pq = &vm_page_queues[queue]; 999 TAILQ_REMOVE(&pq->pl, m, pageq); 1000 1001 /* 1002 * Primarily adjust our pcpu stats for rollup, which is 1003 * (mycpu->gd_vmstats_adj + offset). This is normally 1004 * synchronized on every hardclock(). 1005 * 1006 * However, in order for the nominal low-memory algorithms 1007 * to work properly if the unsynchronized adjustment gets 1008 * too negative and might trigger the pageout daemon, we 1009 * immediately synchronize with the global structure. 1010 * 1011 * The idea here is to reduce unnecessary SMP cache mastership 1012 * changes in the global vmstats, which can be particularly 1013 * bad in multi-socket systems. 1014 * 1015 * WARNING! In systems with low amounts of memory the 1016 * vm_paging_needed(-1024 * ncpus) test could 1017 * wind up testing a value above the paging target, 1018 * meaning it would almost always return TRUE. In 1019 * that situation we synchronize every time the 1020 * cumulative adjustment falls below -1024. 1021 */ 1022 cnt_adj = (long *)((char *)&mycpu->gd_vmstats_adj + 1023 pq->cnt_offset); 1024 cnt_gd = (long *)((char *)&mycpu->gd_vmstats + 1025 pq->cnt_offset); 1026 atomic_add_long(cnt_adj, -1); 1027 atomic_add_long(cnt_gd, -1); 1028 1029 if (*cnt_adj < -1024 && vm_paging_needed(-1024 * ncpus)) { 1030 u_long copy = atomic_swap_long(cnt_adj, 0); 1031 cnt_adj = (long *)((char *)&vmstats + pq->cnt_offset); 1032 atomic_add_long(cnt_adj, copy); 1033 } 1034 pq->lcnt--; 1035 m->queue = PQ_NONE; 1036 oqueue = queue; 1037 queue -= m->pc; 1038 vm_page_queues_spin_unlock(oqueue); /* intended */ 1039 } 1040 return queue; 1041 } 1042 1043 /* 1044 * Helper function places the vm_page on the specified queue. Generally 1045 * speaking only PQ_FREE pages are placed at the head, to allow them to 1046 * be allocated sooner rather than later on the assumption that they 1047 * are cache-hot. 1048 * 1049 * The vm_page must be spinlocked. 1050 * The vm_page must NOT be FICTITIOUS (that would be a disaster) 1051 * This function will return with both the page and the queue locked. 1052 */ 1053 static __inline void 1054 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 1055 { 1056 struct vpgqueues *pq; 1057 u_long *cnt_adj; 1058 u_long *cnt_gd; 1059 1060 KKASSERT(m->queue == PQ_NONE && 1061 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0); 1062 1063 if (queue != PQ_NONE) { 1064 vm_page_queues_spin_lock(queue); 1065 pq = &vm_page_queues[queue]; 1066 ++pq->lcnt; 1067 1068 /* 1069 * Adjust our pcpu stats. If a system entity really needs 1070 * to incorporate the count it will call vmstats_rollup() 1071 * to roll it all up into the global vmstats strufture. 1072 */ 1073 cnt_adj = (long *)((char *)&mycpu->gd_vmstats_adj + 1074 pq->cnt_offset); 1075 cnt_gd = (long *)((char *)&mycpu->gd_vmstats + 1076 pq->cnt_offset); 1077 atomic_add_long(cnt_adj, 1); 1078 atomic_add_long(cnt_gd, 1); 1079 1080 /* 1081 * PQ_FREE is always handled LIFO style to try to provide 1082 * cache-hot pages to programs. 1083 */ 1084 m->queue = queue; 1085 if (queue - m->pc == PQ_FREE) { 1086 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1087 } else if (athead) { 1088 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1089 } else { 1090 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1091 } 1092 /* leave the queue spinlocked */ 1093 } 1094 } 1095 1096 /* 1097 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 1098 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 1099 * 1100 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 1101 * call will be made before returning. 1102 * 1103 * This function does NOT busy the page and on return the page is not 1104 * guaranteed to be available. 1105 */ 1106 void 1107 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 1108 { 1109 u_int32_t busy_count; 1110 1111 for (;;) { 1112 busy_count = m->busy_count; 1113 cpu_ccfence(); 1114 1115 if ((busy_count & PBUSY_LOCKED) == 0 && 1116 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 1117 break; 1118 } 1119 tsleep_interlock(m, 0); 1120 if (atomic_cmpset_int(&m->busy_count, busy_count, 1121 busy_count | PBUSY_WANTED)) { 1122 atomic_set_int(&m->flags, PG_REFERENCED); 1123 tsleep(m, PINTERLOCKED, msg, 0); 1124 break; 1125 } 1126 } 1127 } 1128 1129 /* 1130 * This calculates and returns a page color given an optional VM object and 1131 * either a pindex or an iterator. We attempt to return a cpu-localized 1132 * pg_color that is still roughly 16-way set-associative. The CPU topology 1133 * is used if it was probed. 1134 * 1135 * The caller may use the returned value to index into e.g. PQ_FREE when 1136 * allocating a page in order to nominally obtain pages that are hopefully 1137 * already localized to the requesting cpu. This function is not able to 1138 * provide any sort of guarantee of this, but does its best to improve 1139 * hardware cache management performance. 1140 * 1141 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 1142 */ 1143 u_short 1144 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 1145 { 1146 u_short pg_color; 1147 int object_pg_color; 1148 1149 /* 1150 * WARNING! cpu_topology_core_ids might not be a power of two. 1151 * We also shouldn't make assumptions about 1152 * cpu_topology_phys_ids either. 1153 * 1154 * WARNING! ncpus might not be known at this time (during early 1155 * boot), and might be set to 1. 1156 * 1157 * General format: [phys_id][core_id][cpuid][set-associativity] 1158 * (but uses modulo, so not necessarily precise bit masks) 1159 */ 1160 object_pg_color = object ? object->pg_color : 0; 1161 1162 if (cpu_topology_ht_ids) { 1163 int phys_id; 1164 int core_id; 1165 int ht_id; 1166 int physcale; 1167 int grpscale; 1168 int cpuscale; 1169 1170 /* 1171 * Translate cpuid to socket, core, and hyperthread id. 1172 */ 1173 phys_id = get_cpu_phys_id(cpuid); 1174 core_id = get_cpu_core_id(cpuid); 1175 ht_id = get_cpu_ht_id(cpuid); 1176 1177 /* 1178 * Calculate pg_color for our array index. 1179 * 1180 * physcale - socket multiplier. 1181 * grpscale - core multiplier (cores per socket) 1182 * cpu* - cpus per core 1183 * 1184 * WARNING! In early boot, ncpus has not yet been 1185 * initialized and may be set to (1). 1186 * 1187 * WARNING! physcale must match the organization that 1188 * vm_numa_organize() creates to ensure that 1189 * we properly localize allocations to the 1190 * requested cpuid. 1191 */ 1192 physcale = PQ_L2_SIZE / cpu_topology_phys_ids; 1193 grpscale = physcale / cpu_topology_core_ids; 1194 cpuscale = grpscale / cpu_topology_ht_ids; 1195 1196 pg_color = phys_id * physcale; 1197 pg_color += core_id * grpscale; 1198 pg_color += ht_id * cpuscale; 1199 pg_color += (pindex + object_pg_color) % cpuscale; 1200 1201 #if 0 1202 if (grpsize >= 8) { 1203 pg_color += (pindex + object_pg_color) % grpsize; 1204 } else { 1205 if (grpsize <= 2) { 1206 grpsize = 8; 1207 } else { 1208 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 1209 grpsize += grpsize; 1210 if (grpsize < 8) 1211 grpsize += grpsize; 1212 } 1213 pg_color += (pindex + object_pg_color) % grpsize; 1214 } 1215 #endif 1216 } else { 1217 /* 1218 * Unknown topology, distribute things evenly. 1219 * 1220 * WARNING! In early boot, ncpus has not yet been 1221 * initialized and may be set to (1). 1222 */ 1223 int cpuscale; 1224 1225 cpuscale = PQ_L2_SIZE / ncpus; 1226 1227 pg_color = cpuid * cpuscale; 1228 pg_color += (pindex + object_pg_color) % cpuscale; 1229 } 1230 return (pg_color & PQ_L2_MASK); 1231 } 1232 1233 /* 1234 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 1235 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 1236 */ 1237 void 1238 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 1239 int also_m_busy, const char *msg 1240 VM_PAGE_DEBUG_ARGS) 1241 { 1242 u_int32_t busy_count; 1243 1244 for (;;) { 1245 busy_count = m->busy_count; 1246 cpu_ccfence(); 1247 if (busy_count & PBUSY_LOCKED) { 1248 tsleep_interlock(m, 0); 1249 if (atomic_cmpset_int(&m->busy_count, busy_count, 1250 busy_count | PBUSY_WANTED)) { 1251 atomic_set_int(&m->flags, PG_REFERENCED); 1252 tsleep(m, PINTERLOCKED, msg, 0); 1253 } 1254 } else if (also_m_busy && busy_count) { 1255 tsleep_interlock(m, 0); 1256 if (atomic_cmpset_int(&m->busy_count, busy_count, 1257 busy_count | PBUSY_WANTED)) { 1258 atomic_set_int(&m->flags, PG_REFERENCED); 1259 tsleep(m, PINTERLOCKED, msg, 0); 1260 } 1261 } else { 1262 if (atomic_cmpset_int(&m->busy_count, busy_count, 1263 busy_count | PBUSY_LOCKED)) { 1264 #ifdef VM_PAGE_DEBUG 1265 m->busy_func = func; 1266 m->busy_line = lineno; 1267 #endif 1268 break; 1269 } 1270 } 1271 } 1272 } 1273 1274 /* 1275 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1276 * m->busy_count is also 0. 1277 * 1278 * Returns non-zero on failure. 1279 */ 1280 int 1281 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1282 VM_PAGE_DEBUG_ARGS) 1283 { 1284 u_int32_t busy_count; 1285 1286 for (;;) { 1287 busy_count = m->busy_count; 1288 cpu_ccfence(); 1289 if (busy_count & PBUSY_LOCKED) 1290 return TRUE; 1291 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1292 return TRUE; 1293 if (atomic_cmpset_int(&m->busy_count, busy_count, 1294 busy_count | PBUSY_LOCKED)) { 1295 #ifdef VM_PAGE_DEBUG 1296 m->busy_func = func; 1297 m->busy_line = lineno; 1298 #endif 1299 return FALSE; 1300 } 1301 } 1302 } 1303 1304 /* 1305 * Clear the BUSY flag and return non-zero to indicate to the caller 1306 * that a wakeup() should be performed. 1307 * 1308 * (inline version) 1309 */ 1310 static __inline 1311 int 1312 _vm_page_wakeup(vm_page_t m) 1313 { 1314 u_int32_t busy_count; 1315 1316 busy_count = m->busy_count; 1317 cpu_ccfence(); 1318 for (;;) { 1319 if (atomic_fcmpset_int(&m->busy_count, &busy_count, 1320 busy_count & 1321 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1322 return((int)(busy_count & PBUSY_WANTED)); 1323 } 1324 } 1325 /* not reached */ 1326 } 1327 1328 /* 1329 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1330 * is typically the last call you make on a page before moving onto 1331 * other things. 1332 */ 1333 void 1334 vm_page_wakeup(vm_page_t m) 1335 { 1336 KASSERT(m->busy_count & PBUSY_LOCKED, 1337 ("vm_page_wakeup: page not busy!!!")); 1338 if (_vm_page_wakeup(m)) 1339 wakeup(m); 1340 } 1341 1342 /* 1343 * Hold a page, preventing reuse. This is typically only called on pages 1344 * in a known state (either held busy, special, or interlocked in some 1345 * manner). Holding a page does not ensure that it remains valid, it only 1346 * prevents reuse. The page must not already be on the FREE queue or in 1347 * any danger of being moved to the FREE queue concurrent with this call. 1348 * 1349 * Other parts of the system can still disassociate the page from its object 1350 * and attempt to free it, or perform read or write I/O on it and/or otherwise 1351 * manipulate the page, but if the page is held the VM system will leave the 1352 * page and its data intact and not cycle it through the FREE queue until 1353 * the last hold has been released. 1354 * 1355 * (see vm_page_wire() if you want to prevent the page from being 1356 * disassociated from its object too). 1357 */ 1358 void 1359 vm_page_hold(vm_page_t m) 1360 { 1361 atomic_add_int(&m->hold_count, 1); 1362 KKASSERT(m->queue - m->pc != PQ_FREE); 1363 } 1364 1365 /* 1366 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1367 * it was freed while held and must be moved back to the FREE queue. 1368 * 1369 * To avoid racing against vm_page_free*() we must re-test conditions 1370 * after obtaining the spin-lock. The initial test can also race a 1371 * vm_page_free*() that is in the middle of moving a page to PQ_HOLD, 1372 * leaving the page on PQ_HOLD with hold_count == 0. Rather than 1373 * throw a spin-lock in the critical path, we rely on the pageout 1374 * daemon to clean-up these loose ends. 1375 * 1376 * More critically, the 'easy movement' between queues without busying 1377 * a vm_page is only allowed for PQ_FREE<->PQ_HOLD. 1378 */ 1379 void 1380 vm_page_unhold(vm_page_t m) 1381 { 1382 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1383 ("vm_page_unhold: pg %p illegal hold_count (%d) or " 1384 "on FREE queue (%d)", 1385 m, m->hold_count, m->queue - m->pc)); 1386 1387 if (atomic_fetchadd_int(&m->hold_count, -1) == 1 && 1388 m->queue - m->pc == PQ_HOLD) { 1389 vm_page_spin_lock(m); 1390 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1391 _vm_page_queue_spin_lock(m); 1392 _vm_page_rem_queue_spinlocked(m); 1393 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1394 _vm_page_queue_spin_unlock(m); 1395 } 1396 vm_page_spin_unlock(m); 1397 } 1398 } 1399 1400 /* 1401 * Create a fictitious page with the specified physical address and 1402 * memory attribute. The memory attribute is the only the machine- 1403 * dependent aspect of a fictitious page that must be initialized. 1404 */ 1405 void 1406 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1407 { 1408 /* 1409 * The page's memattr might have changed since the 1410 * previous initialization. Update the pmap to the 1411 * new memattr. 1412 */ 1413 if ((m->flags & PG_FICTITIOUS) != 0) 1414 goto memattr; 1415 m->phys_addr = paddr; 1416 m->queue = PQ_NONE; 1417 /* Fictitious pages don't use "segind". */ 1418 /* Fictitious pages don't use "order" or "pool". */ 1419 m->flags = PG_FICTITIOUS | PG_UNQUEUED; 1420 m->busy_count = PBUSY_LOCKED; 1421 m->wire_count = 1; 1422 spin_init(&m->spin, "fake_page"); 1423 pmap_page_init(m); 1424 memattr: 1425 pmap_page_set_memattr(m, memattr); 1426 } 1427 1428 /* 1429 * Inserts the given vm_page into the object and object list. 1430 * 1431 * The pagetables are not updated but will presumably fault the page 1432 * in if necessary, or if a kernel page the caller will at some point 1433 * enter the page into the kernel's pmap. We are not allowed to block 1434 * here so we *can't* do this anyway. 1435 * 1436 * This routine may not block. 1437 * This routine must be called with the vm_object held. 1438 * This routine must be called with a critical section held. 1439 * 1440 * This routine returns TRUE if the page was inserted into the object 1441 * successfully, and FALSE if the page already exists in the object. 1442 */ 1443 int 1444 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1445 { 1446 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1447 if (m->object != NULL) 1448 panic("vm_page_insert: already inserted"); 1449 1450 atomic_add_int(&object->generation, 1); 1451 1452 /* 1453 * Associate the VM page with an (object, offset). 1454 * 1455 * The vm_page spin lock is required for interactions with the pmap. 1456 * XXX vm_page_spin_lock() might not be needed for this any more. 1457 */ 1458 vm_page_spin_lock(m); 1459 m->object = object; 1460 m->pindex = pindex; 1461 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1462 m->object = NULL; 1463 m->pindex = 0; 1464 vm_page_spin_unlock(m); 1465 return FALSE; 1466 } 1467 ++object->resident_page_count; 1468 ++mycpu->gd_vmtotal.t_rm; 1469 vm_page_spin_unlock(m); 1470 1471 /* 1472 * Since we are inserting a new and possibly dirty page, 1473 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1474 */ 1475 if ((m->valid & m->dirty) || 1476 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1477 vm_object_set_writeable_dirty(object); 1478 1479 /* 1480 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1481 */ 1482 swap_pager_page_inserted(m); 1483 return TRUE; 1484 } 1485 1486 /* 1487 * Removes the given vm_page_t from the (object,index) table 1488 * 1489 * The page must be BUSY and will remain BUSY on return. 1490 * No other requirements. 1491 * 1492 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1493 * it busy. 1494 * 1495 * NOTE: Caller is responsible for any pmap disposition prior to the 1496 * rename (as the pmap code will not be able to find the entries 1497 * once the object has been disassociated). The caller may choose 1498 * to leave the pmap association intact if this routine is being 1499 * called as part of a rename between shadowed objects. 1500 * 1501 * This routine may not block. 1502 */ 1503 void 1504 vm_page_remove(vm_page_t m) 1505 { 1506 vm_object_t object; 1507 1508 if (m->object == NULL) { 1509 return; 1510 } 1511 1512 if ((m->busy_count & PBUSY_LOCKED) == 0) 1513 panic("vm_page_remove: page not busy"); 1514 1515 object = m->object; 1516 1517 vm_object_hold(object); 1518 1519 /* 1520 * Remove the page from the object and update the object. 1521 * 1522 * The vm_page spin lock is required for interactions with the pmap. 1523 * XXX vm_page_spin_lock() might not be needed for this any more. 1524 */ 1525 vm_page_spin_lock(m); 1526 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1527 --object->resident_page_count; 1528 --mycpu->gd_vmtotal.t_rm; 1529 m->object = NULL; 1530 atomic_add_int(&object->generation, 1); 1531 vm_page_spin_unlock(m); 1532 1533 vm_object_drop(object); 1534 } 1535 1536 /* 1537 * Calculate the hash position for the vm_page hash heuristic. 1538 * 1539 * Mask by ~3 to offer 4-way set-assoc 1540 */ 1541 static __inline 1542 struct vm_page_hash_elm * 1543 vm_page_hash_hash(vm_object_t object, vm_pindex_t pindex) 1544 { 1545 size_t hi; 1546 1547 /* mix it up */ 1548 hi = (intptr_t)object ^ object->pg_color ^ pindex; 1549 hi += object->pg_color * pindex; 1550 hi = hi ^ (hi >> 20); 1551 hi &= vm_page_hash_size - 1; /* bounds */ 1552 hi &= ~(VM_PAGE_HASH_SET - 1); /* set-assoc */ 1553 return (&vm_page_hash[hi]); 1554 } 1555 1556 /* 1557 * Heuristical page lookup that does not require any locks. Returns 1558 * a soft-busied page on success, NULL on failure. 1559 * 1560 * Caller must lookup the page the slow way if NULL is returned. 1561 */ 1562 vm_page_t 1563 vm_page_hash_get(vm_object_t object, vm_pindex_t pindex) 1564 { 1565 struct vm_page_hash_elm *mp; 1566 vm_page_t m; 1567 int i; 1568 1569 if (vm_page_hash == NULL) 1570 return NULL; 1571 mp = vm_page_hash_hash(object, pindex); 1572 for (i = 0; i < VM_PAGE_HASH_SET; ++i) { 1573 m = mp[i].m; 1574 cpu_ccfence(); 1575 if (m == NULL) 1576 continue; 1577 if (m->object != object || m->pindex != pindex) 1578 continue; 1579 if (vm_page_sbusy_try(m)) 1580 continue; 1581 if (m->object == object && m->pindex == pindex) { 1582 mp[i].ticks = ticks; 1583 return m; 1584 } 1585 vm_page_sbusy_drop(m); 1586 } 1587 return NULL; 1588 } 1589 1590 /* 1591 * Enter page onto vm_page_hash[]. This is a heuristic, SMP collisions 1592 * are allowed. 1593 */ 1594 static __inline 1595 void 1596 vm_page_hash_enter(vm_page_t m) 1597 { 1598 struct vm_page_hash_elm *mp; 1599 struct vm_page_hash_elm *best; 1600 int i; 1601 1602 /* 1603 * Only enter type-stable vm_pages with well-shared objects. 1604 */ 1605 if (vm_page_hash == NULL || 1606 m < &vm_page_array[0] || 1607 m >= &vm_page_array[vm_page_array_size]) 1608 return; 1609 if (m->object == NULL) 1610 return; 1611 #if 0 1612 /* 1613 * Disabled at the moment, there are some degenerate conditions 1614 * with often-exec'd programs that get ignored. In particular, 1615 * the kernel's elf loader does a vn_rdwr() on the first page of 1616 * a binary. 1617 */ 1618 if (m->object->ref_count <= 2 || (m->object->flags & OBJ_ONEMAPPING)) 1619 return; 1620 #endif 1621 if (vm_page_hash_vnode_only && m->object->type != OBJT_VNODE) 1622 return; 1623 1624 /* 1625 * 1626 */ 1627 mp = vm_page_hash_hash(m->object, m->pindex); 1628 best = mp; 1629 for (i = 0; i < VM_PAGE_HASH_SET; ++i) { 1630 if (mp[i].m == m) { 1631 mp[i].ticks = ticks; 1632 return; 1633 } 1634 1635 /* 1636 * The best choice is the oldest entry 1637 */ 1638 if ((ticks - best->ticks) < (ticks - mp[i].ticks) || 1639 (int)(ticks - mp[i].ticks) < 0) { 1640 best = &mp[i]; 1641 } 1642 } 1643 best->m = m; 1644 best->ticks = ticks; 1645 } 1646 1647 /* 1648 * Locate and return the page at (object, pindex), or NULL if the 1649 * page could not be found. 1650 * 1651 * The caller must hold the vm_object token. 1652 */ 1653 vm_page_t 1654 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1655 { 1656 vm_page_t m; 1657 1658 /* 1659 * Search the hash table for this object/offset pair 1660 */ 1661 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1662 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1663 if (m) { 1664 KKASSERT(m->object == object && m->pindex == pindex); 1665 vm_page_hash_enter(m); 1666 } 1667 return(m); 1668 } 1669 1670 vm_page_t 1671 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1672 vm_pindex_t pindex, 1673 int also_m_busy, const char *msg 1674 VM_PAGE_DEBUG_ARGS) 1675 { 1676 u_int32_t busy_count; 1677 vm_page_t m; 1678 1679 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1680 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1681 while (m) { 1682 KKASSERT(m->object == object && m->pindex == pindex); 1683 busy_count = m->busy_count; 1684 cpu_ccfence(); 1685 if (busy_count & PBUSY_LOCKED) { 1686 tsleep_interlock(m, 0); 1687 if (atomic_cmpset_int(&m->busy_count, busy_count, 1688 busy_count | PBUSY_WANTED)) { 1689 atomic_set_int(&m->flags, PG_REFERENCED); 1690 tsleep(m, PINTERLOCKED, msg, 0); 1691 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1692 pindex); 1693 } 1694 } else if (also_m_busy && busy_count) { 1695 tsleep_interlock(m, 0); 1696 if (atomic_cmpset_int(&m->busy_count, busy_count, 1697 busy_count | PBUSY_WANTED)) { 1698 atomic_set_int(&m->flags, PG_REFERENCED); 1699 tsleep(m, PINTERLOCKED, msg, 0); 1700 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1701 pindex); 1702 } 1703 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1704 busy_count | PBUSY_LOCKED)) { 1705 #ifdef VM_PAGE_DEBUG 1706 m->busy_func = func; 1707 m->busy_line = lineno; 1708 #endif 1709 vm_page_hash_enter(m); 1710 break; 1711 } 1712 } 1713 return m; 1714 } 1715 1716 /* 1717 * Attempt to lookup and busy a page. 1718 * 1719 * Returns NULL if the page could not be found 1720 * 1721 * Returns a vm_page and error == TRUE if the page exists but could not 1722 * be busied. 1723 * 1724 * Returns a vm_page and error == FALSE on success. 1725 */ 1726 vm_page_t 1727 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1728 vm_pindex_t pindex, 1729 int also_m_busy, int *errorp 1730 VM_PAGE_DEBUG_ARGS) 1731 { 1732 u_int32_t busy_count; 1733 vm_page_t m; 1734 1735 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1736 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1737 *errorp = FALSE; 1738 while (m) { 1739 KKASSERT(m->object == object && m->pindex == pindex); 1740 busy_count = m->busy_count; 1741 cpu_ccfence(); 1742 if (busy_count & PBUSY_LOCKED) { 1743 *errorp = TRUE; 1744 break; 1745 } 1746 if (also_m_busy && busy_count) { 1747 *errorp = TRUE; 1748 break; 1749 } 1750 if (atomic_cmpset_int(&m->busy_count, busy_count, 1751 busy_count | PBUSY_LOCKED)) { 1752 #ifdef VM_PAGE_DEBUG 1753 m->busy_func = func; 1754 m->busy_line = lineno; 1755 #endif 1756 vm_page_hash_enter(m); 1757 break; 1758 } 1759 } 1760 return m; 1761 } 1762 1763 /* 1764 * Returns a page that is only soft-busied for use by the caller in 1765 * a read-only fashion. Returns NULL if the page could not be found, 1766 * the soft busy could not be obtained, or the page data is invalid. 1767 * 1768 * XXX Doesn't handle PG_FICTITIOUS pages at the moment, but there is 1769 * no reason why we couldn't. 1770 */ 1771 vm_page_t 1772 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1773 int pgoff, int pgbytes) 1774 { 1775 vm_page_t m; 1776 1777 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1778 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1779 if (m) { 1780 if ((m->valid != VM_PAGE_BITS_ALL && 1781 !vm_page_is_valid(m, pgoff, pgbytes)) || 1782 (m->flags & PG_FICTITIOUS)) { 1783 m = NULL; 1784 } else if (vm_page_sbusy_try(m)) { 1785 m = NULL; 1786 } else if ((m->valid != VM_PAGE_BITS_ALL && 1787 !vm_page_is_valid(m, pgoff, pgbytes)) || 1788 (m->flags & PG_FICTITIOUS)) { 1789 vm_page_sbusy_drop(m); 1790 m = NULL; 1791 } else { 1792 vm_page_hash_enter(m); 1793 } 1794 } 1795 return m; 1796 } 1797 1798 /* 1799 * Caller must hold the related vm_object 1800 */ 1801 vm_page_t 1802 vm_page_next(vm_page_t m) 1803 { 1804 vm_page_t next; 1805 1806 next = vm_page_rb_tree_RB_NEXT(m); 1807 if (next && next->pindex != m->pindex + 1) 1808 next = NULL; 1809 return (next); 1810 } 1811 1812 /* 1813 * vm_page_rename() 1814 * 1815 * Move the given vm_page from its current object to the specified 1816 * target object/offset. The page must be busy and will remain so 1817 * on return. 1818 * 1819 * new_object must be held. 1820 * This routine might block. XXX ? 1821 * 1822 * NOTE: Swap associated with the page must be invalidated by the move. We 1823 * have to do this for several reasons: (1) we aren't freeing the 1824 * page, (2) we are dirtying the page, (3) the VM system is probably 1825 * moving the page from object A to B, and will then later move 1826 * the backing store from A to B and we can't have a conflict. 1827 * 1828 * NOTE: We *always* dirty the page. It is necessary both for the 1829 * fact that we moved it, and because we may be invalidating 1830 * swap. If the page is on the cache, we have to deactivate it 1831 * or vm_page_dirty() will panic. Dirty pages are not allowed 1832 * on the cache. 1833 * 1834 * NOTE: Caller is responsible for any pmap disposition prior to the 1835 * rename (as the pmap code will not be able to find the entries 1836 * once the object has been disassociated or changed). Nominally 1837 * the caller is moving a page between shadowed objects and so the 1838 * pmap association is retained without having to remove the page 1839 * from it. 1840 */ 1841 void 1842 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1843 { 1844 KKASSERT(m->busy_count & PBUSY_LOCKED); 1845 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1846 if (m->object) { 1847 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1848 vm_page_remove(m); 1849 } 1850 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1851 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1852 new_object, new_pindex); 1853 } 1854 if (m->queue - m->pc == PQ_CACHE) 1855 vm_page_deactivate(m); 1856 vm_page_dirty(m); 1857 } 1858 1859 /* 1860 * vm_page_unqueue() without any wakeup. This routine is used when a page 1861 * is to remain BUSYied by the caller. 1862 * 1863 * This routine may not block. 1864 */ 1865 void 1866 vm_page_unqueue_nowakeup(vm_page_t m) 1867 { 1868 vm_page_and_queue_spin_lock(m); 1869 (void)_vm_page_rem_queue_spinlocked(m); 1870 vm_page_spin_unlock(m); 1871 } 1872 1873 /* 1874 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1875 * if necessary. 1876 * 1877 * This routine may not block. 1878 */ 1879 void 1880 vm_page_unqueue(vm_page_t m) 1881 { 1882 u_short queue; 1883 1884 vm_page_and_queue_spin_lock(m); 1885 queue = _vm_page_rem_queue_spinlocked(m); 1886 if (queue == PQ_FREE || queue == PQ_CACHE) { 1887 vm_page_spin_unlock(m); 1888 pagedaemon_wakeup(); 1889 } else { 1890 vm_page_spin_unlock(m); 1891 } 1892 } 1893 1894 /* 1895 * vm_page_list_find() 1896 * 1897 * Find a page on the specified queue with color optimization. 1898 * 1899 * The page coloring optimization attempts to locate a page that does 1900 * not overload other nearby pages in the object in the cpu's L1 or L2 1901 * caches. We need this optimization because cpu caches tend to be 1902 * physical caches, while object spaces tend to be virtual. 1903 * 1904 * The page coloring optimization also, very importantly, tries to localize 1905 * memory to cpus and physical sockets. 1906 * 1907 * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the 1908 * algorithm is adjusted to localize allocations on a per-core basis. 1909 * This is done by 'twisting' the colors. 1910 * 1911 * The page is returned spinlocked and removed from its queue (it will 1912 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1913 * is responsible for dealing with the busy-page case (usually by 1914 * deactivating the page and looping). 1915 * 1916 * NOTE: This routine is carefully inlined. A non-inlined version 1917 * is available for outside callers but the only critical path is 1918 * from within this source file. 1919 * 1920 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1921 * represent stable storage, allowing us to order our locks vm_page 1922 * first, then queue. 1923 */ 1924 static __inline 1925 vm_page_t 1926 _vm_page_list_find(int basequeue, int index) 1927 { 1928 struct vpgqueues *pq; 1929 vm_page_t m; 1930 1931 index &= PQ_L2_MASK; 1932 pq = &vm_page_queues[basequeue + index]; 1933 1934 /* 1935 * Try this cpu's colored queue first. Test for a page unlocked, 1936 * then lock the queue and locate a page. Note that the lock order 1937 * is reversed, but we do not want to dwadle on the page spinlock 1938 * anyway as it is held significantly longer than the queue spinlock. 1939 */ 1940 if (TAILQ_FIRST(&pq->pl)) { 1941 spin_lock(&pq->spin); 1942 TAILQ_FOREACH(m, &pq->pl, pageq) { 1943 if (spin_trylock(&m->spin) == 0) 1944 continue; 1945 KKASSERT(m->queue == basequeue + index); 1946 _vm_page_rem_queue_spinlocked(m); 1947 pq->lastq = -1; 1948 return(m); 1949 } 1950 spin_unlock(&pq->spin); 1951 } 1952 1953 /* 1954 * If we are unable to get a page, do a more involved NUMA-aware 1955 * search. However, to avoid re-searching empty queues over and 1956 * over again skip to pq->last if appropriate. 1957 */ 1958 if (pq->lastq >= 0) 1959 index = pq->lastq; 1960 1961 m = _vm_page_list_find2(basequeue, index, &pq->lastq); 1962 1963 return(m); 1964 } 1965 1966 /* 1967 * If we could not find the page in the desired queue try to find it in 1968 * a nearby (NUMA-aware) queue. 1969 */ 1970 static vm_page_t 1971 _vm_page_list_find2(int basequeue, int index, int *lastp) 1972 { 1973 struct vpgqueues *pq; 1974 vm_page_t m = NULL; 1975 int pqmask = set_assoc_mask >> 1; 1976 int pqi; 1977 int range; 1978 int skip_start; 1979 int skip_next; 1980 int count; 1981 1982 index &= PQ_L2_MASK; 1983 pq = &vm_page_queues[basequeue]; 1984 count = 0; 1985 skip_start = -1; 1986 skip_next = -1; 1987 1988 /* 1989 * Run local sets of 16, 32, 64, 128, up to the entire queue if all 1990 * else fails (PQ_L2_MASK). 1991 * 1992 * pqmask is a mask, 15, 31, 63, etc. 1993 * 1994 * Test each queue unlocked first, then lock the queue and locate 1995 * a page. Note that the lock order is reversed, but we do not want 1996 * to dwadle on the page spinlock anyway as it is held significantly 1997 * longer than the queue spinlock. 1998 */ 1999 do { 2000 pqmask = (pqmask << 1) | 1; 2001 2002 pqi = index; 2003 range = pqmask + 1; 2004 2005 while (range > 0) { 2006 if (pqi >= skip_start && pqi < skip_next) { 2007 range -= skip_next - pqi; 2008 pqi = (pqi & ~pqmask) | (skip_next & pqmask); 2009 } 2010 if (range > 0 && TAILQ_FIRST(&pq[pqi].pl)) { 2011 spin_lock(&pq[pqi].spin); 2012 TAILQ_FOREACH(m, &pq[pqi].pl, pageq) { 2013 if (spin_trylock(&m->spin) == 0) 2014 continue; 2015 KKASSERT(m->queue == basequeue + pqi); 2016 _vm_page_rem_queue_spinlocked(m); 2017 2018 /* 2019 * If we had to wander too far, set 2020 * *lastp to skip past empty queues. 2021 */ 2022 if (count >= 8) 2023 *lastp = pqi & PQ_L2_MASK; 2024 return(m); 2025 } 2026 spin_unlock(&pq[pqi].spin); 2027 } 2028 --range; 2029 ++count; 2030 pqi = (pqi & ~pqmask) | ((pqi + 1) & pqmask); 2031 } 2032 skip_start = pqi & ~pqmask; 2033 skip_next = (pqi | pqmask) + 1; 2034 } while (pqmask != PQ_L2_MASK); 2035 2036 return(m); 2037 } 2038 2039 /* 2040 * Returns a vm_page candidate for allocation. The page is not busied so 2041 * it can move around. The caller must busy the page (and typically 2042 * deactivate it if it cannot be busied!) 2043 * 2044 * Returns a spinlocked vm_page that has been removed from its queue. 2045 */ 2046 vm_page_t 2047 vm_page_list_find(int basequeue, int index) 2048 { 2049 return(_vm_page_list_find(basequeue, index)); 2050 } 2051 2052 /* 2053 * Find a page on the cache queue with color optimization, remove it 2054 * from the queue, and busy it. The returned page will not be spinlocked. 2055 * 2056 * A candidate failure will be deactivated. Candidates can fail due to 2057 * being busied by someone else, in which case they will be deactivated. 2058 * 2059 * This routine may not block. 2060 * 2061 */ 2062 static vm_page_t 2063 vm_page_select_cache(u_short pg_color) 2064 { 2065 vm_page_t m; 2066 2067 for (;;) { 2068 m = _vm_page_list_find(PQ_CACHE, pg_color); 2069 if (m == NULL) 2070 break; 2071 /* 2072 * (m) has been removed from its queue and spinlocked 2073 */ 2074 if (vm_page_busy_try(m, TRUE)) { 2075 _vm_page_deactivate_locked(m, 0); 2076 vm_page_spin_unlock(m); 2077 } else { 2078 /* 2079 * We successfully busied the page 2080 */ 2081 if ((m->flags & PG_NEED_COMMIT) == 0 && 2082 m->hold_count == 0 && 2083 m->wire_count == 0 && 2084 (m->dirty & m->valid) == 0) { 2085 vm_page_spin_unlock(m); 2086 KKASSERT((m->flags & PG_UNQUEUED) == 0); 2087 pagedaemon_wakeup(); 2088 return(m); 2089 } 2090 2091 /* 2092 * The page cannot be recycled, deactivate it. 2093 */ 2094 _vm_page_deactivate_locked(m, 0); 2095 if (_vm_page_wakeup(m)) { 2096 vm_page_spin_unlock(m); 2097 wakeup(m); 2098 } else { 2099 vm_page_spin_unlock(m); 2100 } 2101 } 2102 } 2103 return (m); 2104 } 2105 2106 /* 2107 * Find a free page. We attempt to inline the nominal case and fall back 2108 * to _vm_page_select_free() otherwise. A busied page is removed from 2109 * the queue and returned. 2110 * 2111 * This routine may not block. 2112 */ 2113 static __inline vm_page_t 2114 vm_page_select_free(u_short pg_color) 2115 { 2116 vm_page_t m; 2117 2118 for (;;) { 2119 m = _vm_page_list_find(PQ_FREE, pg_color); 2120 if (m == NULL) 2121 break; 2122 if (vm_page_busy_try(m, TRUE)) { 2123 /* 2124 * Various mechanisms such as a pmap_collect can 2125 * result in a busy page on the free queue. We 2126 * have to move the page out of the way so we can 2127 * retry the allocation. If the other thread is not 2128 * allocating the page then m->valid will remain 0 and 2129 * the pageout daemon will free the page later on. 2130 * 2131 * Since we could not busy the page, however, we 2132 * cannot make assumptions as to whether the page 2133 * will be allocated by the other thread or not, 2134 * so all we can do is deactivate it to move it out 2135 * of the way. In particular, if the other thread 2136 * wires the page it may wind up on the inactive 2137 * queue and the pageout daemon will have to deal 2138 * with that case too. 2139 */ 2140 _vm_page_deactivate_locked(m, 0); 2141 vm_page_spin_unlock(m); 2142 } else { 2143 /* 2144 * Theoretically if we are able to busy the page 2145 * atomic with the queue removal (using the vm_page 2146 * lock) nobody else should have been able to mess 2147 * with the page before us. 2148 * 2149 * Assert the page state. Note that even though 2150 * wiring doesn't adjust queues, a page on the free 2151 * queue should never be wired at this point. 2152 */ 2153 KKASSERT((m->flags & (PG_UNQUEUED | 2154 PG_NEED_COMMIT)) == 0); 2155 KASSERT(m->hold_count == 0, 2156 ("m->hold_count is not zero " 2157 "pg %p q=%d flags=%08x hold=%d wire=%d", 2158 m, m->queue, m->flags, 2159 m->hold_count, m->wire_count)); 2160 KKASSERT(m->wire_count == 0); 2161 vm_page_spin_unlock(m); 2162 pagedaemon_wakeup(); 2163 2164 /* return busied and removed page */ 2165 return(m); 2166 } 2167 } 2168 return(m); 2169 } 2170 2171 /* 2172 * vm_page_alloc() 2173 * 2174 * Allocate and return a memory cell associated with this VM object/offset 2175 * pair. If object is NULL an unassociated page will be allocated. 2176 * 2177 * The returned page will be busied and removed from its queues. This 2178 * routine can block and may return NULL if a race occurs and the page 2179 * is found to already exist at the specified (object, pindex). 2180 * 2181 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 2182 * VM_ALLOC_QUICK like normal but cannot use cache 2183 * VM_ALLOC_SYSTEM greater free drain 2184 * VM_ALLOC_INTERRUPT allow free list to be completely drained 2185 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 2186 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 2187 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 2188 * (see vm_page_grab()) 2189 * VM_ALLOC_USE_GD ok to use per-gd cache 2190 * 2191 * VM_ALLOC_CPU(n) allocate using specified cpu localization 2192 * 2193 * The object must be held if not NULL 2194 * This routine may not block 2195 * 2196 * Additional special handling is required when called from an interrupt 2197 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 2198 * in this case. 2199 */ 2200 vm_page_t 2201 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 2202 { 2203 globaldata_t gd; 2204 vm_object_t obj; 2205 vm_page_t m; 2206 u_short pg_color; 2207 int cpuid_local; 2208 2209 #if 0 2210 /* 2211 * Special per-cpu free VM page cache. The pages are pre-busied 2212 * and pre-zerod for us. 2213 */ 2214 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 2215 crit_enter_gd(gd); 2216 if (gd->gd_vmpg_count) { 2217 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 2218 crit_exit_gd(gd); 2219 goto done; 2220 } 2221 crit_exit_gd(gd); 2222 } 2223 #endif 2224 m = NULL; 2225 2226 /* 2227 * CPU LOCALIZATION 2228 * 2229 * CPU localization algorithm. Break the page queues up by physical 2230 * id and core id (note that two cpu threads will have the same core 2231 * id, and core_id != gd_cpuid). 2232 * 2233 * This is nowhere near perfect, for example the last pindex in a 2234 * subgroup will overflow into the next cpu or package. But this 2235 * should get us good page reuse locality in heavy mixed loads. 2236 * 2237 * (may be executed before the APs are started, so other GDs might 2238 * not exist!) 2239 */ 2240 if (page_req & VM_ALLOC_CPU_SPEC) 2241 cpuid_local = VM_ALLOC_GETCPU(page_req); 2242 else 2243 cpuid_local = mycpu->gd_cpuid; 2244 2245 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 2246 2247 KKASSERT(page_req & 2248 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 2249 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2250 2251 /* 2252 * Certain system threads (pageout daemon, buf_daemon's) are 2253 * allowed to eat deeper into the free page list. 2254 */ 2255 if (curthread->td_flags & TDF_SYSTHREAD) 2256 page_req |= VM_ALLOC_SYSTEM; 2257 2258 /* 2259 * Impose various limitations. Note that the v_free_reserved test 2260 * must match the opposite of vm_page_count_target() to avoid 2261 * livelocks, be careful. 2262 */ 2263 loop: 2264 gd = mycpu; 2265 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 2266 ((page_req & VM_ALLOC_INTERRUPT) && 2267 gd->gd_vmstats.v_free_count > 0) || 2268 ((page_req & VM_ALLOC_SYSTEM) && 2269 gd->gd_vmstats.v_cache_count == 0 && 2270 gd->gd_vmstats.v_free_count > 2271 gd->gd_vmstats.v_interrupt_free_min) 2272 ) { 2273 /* 2274 * The free queue has sufficient free pages to take one out. 2275 */ 2276 m = vm_page_select_free(pg_color); 2277 } else if (page_req & VM_ALLOC_NORMAL) { 2278 /* 2279 * Allocatable from the cache (non-interrupt only). On 2280 * success, we must free the page and try again, thus 2281 * ensuring that vmstats.v_*_free_min counters are replenished. 2282 */ 2283 #ifdef INVARIANTS 2284 if (curthread->td_preempted) { 2285 kprintf("vm_page_alloc(): warning, attempt to allocate" 2286 " cache page from preempting interrupt\n"); 2287 m = NULL; 2288 } else { 2289 m = vm_page_select_cache(pg_color); 2290 } 2291 #else 2292 m = vm_page_select_cache(pg_color); 2293 #endif 2294 /* 2295 * On success move the page into the free queue and loop. 2296 * 2297 * Only do this if we can safely acquire the vm_object lock, 2298 * because this is effectively a random page and the caller 2299 * might be holding the lock shared, we don't want to 2300 * deadlock. 2301 */ 2302 if (m != NULL) { 2303 KASSERT(m->dirty == 0, 2304 ("Found dirty cache page %p", m)); 2305 if ((obj = m->object) != NULL) { 2306 if (vm_object_hold_try(obj)) { 2307 vm_page_protect(m, VM_PROT_NONE); 2308 vm_page_free(m); 2309 /* m->object NULL here */ 2310 vm_object_drop(obj); 2311 } else { 2312 vm_page_deactivate(m); 2313 vm_page_wakeup(m); 2314 } 2315 } else { 2316 vm_page_protect(m, VM_PROT_NONE); 2317 vm_page_free(m); 2318 } 2319 goto loop; 2320 } 2321 2322 /* 2323 * On failure return NULL 2324 */ 2325 atomic_add_int(&vm_pageout_deficit, 1); 2326 pagedaemon_wakeup(); 2327 return (NULL); 2328 } else { 2329 /* 2330 * No pages available, wakeup the pageout daemon and give up. 2331 */ 2332 atomic_add_int(&vm_pageout_deficit, 1); 2333 pagedaemon_wakeup(); 2334 return (NULL); 2335 } 2336 2337 /* 2338 * v_free_count can race so loop if we don't find the expected 2339 * page. 2340 */ 2341 if (m == NULL) { 2342 vmstats_rollup(); 2343 goto loop; 2344 } 2345 2346 /* 2347 * Good page found. The page has already been busied for us and 2348 * removed from its queues. 2349 */ 2350 KASSERT(m->dirty == 0, 2351 ("vm_page_alloc: free/cache page %p was dirty", m)); 2352 KKASSERT(m->queue == PQ_NONE); 2353 2354 #if 0 2355 done: 2356 #endif 2357 /* 2358 * Initialize the structure, inheriting some flags but clearing 2359 * all the rest. The page has already been busied for us. 2360 */ 2361 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 2362 2363 KKASSERT(m->wire_count == 0); 2364 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 2365 m->act_count = 0; 2366 m->valid = 0; 2367 2368 /* 2369 * Caller must be holding the object lock (asserted by 2370 * vm_page_insert()). 2371 * 2372 * NOTE: Inserting a page here does not insert it into any pmaps 2373 * (which could cause us to block allocating memory). 2374 * 2375 * NOTE: If no object an unassociated page is allocated, m->pindex 2376 * can be used by the caller for any purpose. 2377 */ 2378 if (object) { 2379 if (vm_page_insert(m, object, pindex) == FALSE) { 2380 vm_page_free(m); 2381 if ((page_req & VM_ALLOC_NULL_OK) == 0) 2382 panic("PAGE RACE %p[%ld]/%p", 2383 object, (long)pindex, m); 2384 m = NULL; 2385 } 2386 } else { 2387 m->pindex = pindex; 2388 } 2389 2390 /* 2391 * Don't wakeup too often - wakeup the pageout daemon when 2392 * we would be nearly out of memory. 2393 */ 2394 pagedaemon_wakeup(); 2395 2396 /* 2397 * A BUSY page is returned. 2398 */ 2399 return (m); 2400 } 2401 2402 /* 2403 * Returns number of pages available in our DMA memory reserve 2404 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 2405 */ 2406 vm_size_t 2407 vm_contig_avail_pages(void) 2408 { 2409 alist_blk_t blk; 2410 alist_blk_t count; 2411 alist_blk_t bfree; 2412 spin_lock(&vm_contig_spin); 2413 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 2414 spin_unlock(&vm_contig_spin); 2415 2416 return bfree; 2417 } 2418 2419 /* 2420 * Attempt to allocate contiguous physical memory with the specified 2421 * requirements. 2422 */ 2423 vm_page_t 2424 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 2425 unsigned long alignment, unsigned long boundary, 2426 unsigned long size, vm_memattr_t memattr) 2427 { 2428 alist_blk_t blk; 2429 vm_page_t m; 2430 vm_pindex_t i; 2431 #if 0 2432 static vm_pindex_t contig_rover; 2433 #endif 2434 2435 alignment >>= PAGE_SHIFT; 2436 if (alignment == 0) 2437 alignment = 1; 2438 boundary >>= PAGE_SHIFT; 2439 if (boundary == 0) 2440 boundary = 1; 2441 size = (size + PAGE_MASK) >> PAGE_SHIFT; 2442 2443 #if 0 2444 /* 2445 * Disabled temporarily until we find a solution for DRM (a flag 2446 * to always use the free space reserve, for performance). 2447 */ 2448 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 2449 boundary <= PAGE_SIZE && size == 1 && 2450 memattr == VM_MEMATTR_DEFAULT) { 2451 /* 2452 * Any page will work, use vm_page_alloc() 2453 * (e.g. when used from kmem_alloc_attr()) 2454 */ 2455 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 2456 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2457 VM_ALLOC_INTERRUPT); 2458 m->valid = VM_PAGE_BITS_ALL; 2459 vm_page_wire(m); 2460 vm_page_wakeup(m); 2461 } else 2462 #endif 2463 { 2464 /* 2465 * Use the low-memory dma reserve 2466 */ 2467 spin_lock(&vm_contig_spin); 2468 blk = alist_alloc(&vm_contig_alist, 0, size); 2469 if (blk == ALIST_BLOCK_NONE) { 2470 spin_unlock(&vm_contig_spin); 2471 if (bootverbose) { 2472 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2473 (size << PAGE_SHIFT) / 1024); 2474 print_backtrace(5); 2475 } 2476 return(NULL); 2477 } 2478 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2479 alist_free(&vm_contig_alist, blk, size); 2480 spin_unlock(&vm_contig_spin); 2481 if (bootverbose) { 2482 kprintf("vm_page_alloc_contig: %ldk high " 2483 "%016jx failed\n", 2484 (size << PAGE_SHIFT) / 1024, 2485 (intmax_t)high); 2486 } 2487 return(NULL); 2488 } 2489 spin_unlock(&vm_contig_spin); 2490 2491 /* 2492 * Base vm_page_t of range 2493 */ 2494 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2495 } 2496 if (vm_contig_verbose) { 2497 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2498 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2499 (intmax_t)m->phys_addr, 2500 (size << PAGE_SHIFT) / 1024, 2501 low, high, alignment, boundary, size, memattr); 2502 } 2503 if (memattr != VM_MEMATTR_DEFAULT) { 2504 for (i = 0; i < size; ++i) { 2505 KKASSERT(m[i].flags & PG_FICTITIOUS); 2506 pmap_page_set_memattr(&m[i], memattr); 2507 } 2508 } 2509 return m; 2510 } 2511 2512 /* 2513 * Free contiguously allocated pages. The pages will be wired but not busy. 2514 * When freeing to the alist we leave them wired and not busy. 2515 */ 2516 void 2517 vm_page_free_contig(vm_page_t m, unsigned long size) 2518 { 2519 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2520 vm_pindex_t start = pa >> PAGE_SHIFT; 2521 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2522 2523 if (vm_contig_verbose) { 2524 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2525 (intmax_t)pa, size / 1024); 2526 } 2527 if (pa < vm_low_phys_reserved) { 2528 /* 2529 * Just assert check the first page for convenience. 2530 */ 2531 KKASSERT(m->wire_count == 1); 2532 KKASSERT(m->flags & PG_FICTITIOUS); 2533 KKASSERT(pa + size <= vm_low_phys_reserved); 2534 spin_lock(&vm_contig_spin); 2535 alist_free(&vm_contig_alist, start, pages); 2536 spin_unlock(&vm_contig_spin); 2537 } else { 2538 while (pages) { 2539 /* XXX FUTURE, maybe (pair with vm_pg_contig_alloc()) */ 2540 /*vm_page_flag_clear(m, PG_FICTITIOUS | PG_UNQUEUED);*/ 2541 vm_page_busy_wait(m, FALSE, "cpgfr"); 2542 vm_page_unwire(m, 0); 2543 vm_page_free(m); 2544 --pages; 2545 ++m; 2546 } 2547 2548 } 2549 } 2550 2551 2552 /* 2553 * Wait for sufficient free memory for nominal heavy memory use kernel 2554 * operations. 2555 * 2556 * WARNING! Be sure never to call this in any vm_pageout code path, which 2557 * will trivially deadlock the system. 2558 */ 2559 void 2560 vm_wait_nominal(void) 2561 { 2562 while (vm_page_count_min(0)) 2563 vm_wait(0); 2564 } 2565 2566 /* 2567 * Test if vm_wait_nominal() would block. 2568 */ 2569 int 2570 vm_test_nominal(void) 2571 { 2572 if (vm_page_count_min(0)) 2573 return(1); 2574 return(0); 2575 } 2576 2577 /* 2578 * Block until free pages are available for allocation, called in various 2579 * places before memory allocations. 2580 * 2581 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2582 * more generous then that. 2583 */ 2584 void 2585 vm_wait(int timo) 2586 { 2587 /* 2588 * never wait forever 2589 */ 2590 if (timo == 0) 2591 timo = hz; 2592 lwkt_gettoken(&vm_token); 2593 2594 if (curthread == pagethread || 2595 curthread == emergpager) { 2596 /* 2597 * The pageout daemon itself needs pages, this is bad. 2598 */ 2599 if (vm_page_count_min(0)) { 2600 vm_pageout_pages_needed = 1; 2601 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2602 } 2603 } else { 2604 /* 2605 * Wakeup the pageout daemon if necessary and wait. 2606 * 2607 * Do not wait indefinitely for the target to be reached, 2608 * as load might prevent it from being reached any time soon. 2609 * But wait a little to try to slow down page allocations 2610 * and to give more important threads (the pagedaemon) 2611 * allocation priority. 2612 */ 2613 if (vm_page_count_target()) { 2614 if (vm_pages_needed == 0) { 2615 vm_pages_needed = 1; 2616 wakeup(&vm_pages_needed); 2617 } 2618 ++vm_pages_waiting; /* SMP race ok */ 2619 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2620 } 2621 } 2622 lwkt_reltoken(&vm_token); 2623 } 2624 2625 /* 2626 * Block until free pages are available for allocation 2627 * 2628 * Called only from vm_fault so that processes page faulting can be 2629 * easily tracked. 2630 */ 2631 void 2632 vm_wait_pfault(void) 2633 { 2634 /* 2635 * Wakeup the pageout daemon if necessary and wait. 2636 * 2637 * Do not wait indefinitely for the target to be reached, 2638 * as load might prevent it from being reached any time soon. 2639 * But wait a little to try to slow down page allocations 2640 * and to give more important threads (the pagedaemon) 2641 * allocation priority. 2642 */ 2643 if (vm_page_count_min(0)) { 2644 lwkt_gettoken(&vm_token); 2645 while (vm_page_count_severe()) { 2646 if (vm_page_count_target()) { 2647 thread_t td; 2648 2649 if (vm_pages_needed == 0) { 2650 vm_pages_needed = 1; 2651 wakeup(&vm_pages_needed); 2652 } 2653 ++vm_pages_waiting; /* SMP race ok */ 2654 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2655 2656 /* 2657 * Do not stay stuck in the loop if the system is trying 2658 * to kill the process. 2659 */ 2660 td = curthread; 2661 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2662 break; 2663 } 2664 } 2665 lwkt_reltoken(&vm_token); 2666 } 2667 } 2668 2669 /* 2670 * Put the specified page on the active list (if appropriate). Ensure 2671 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2672 * 2673 * The caller should be holding the page busied ? XXX 2674 * This routine may not block. 2675 * 2676 * It is ok if the page is wired (so buffer cache operations don't have 2677 * to mess with the page queues). 2678 */ 2679 void 2680 vm_page_activate(vm_page_t m) 2681 { 2682 u_short oqueue; 2683 2684 /* 2685 * If already active or inappropriate, just set act_count and 2686 * return. We don't have to spin-lock the page. 2687 */ 2688 if (m->queue - m->pc == PQ_ACTIVE || 2689 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2690 if (m->act_count < ACT_INIT) 2691 m->act_count = ACT_INIT; 2692 return; 2693 } 2694 2695 vm_page_spin_lock(m); 2696 if (m->queue - m->pc != PQ_ACTIVE && 2697 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 2698 _vm_page_queue_spin_lock(m); 2699 oqueue = _vm_page_rem_queue_spinlocked(m); 2700 /* page is left spinlocked, queue is unlocked */ 2701 2702 if (oqueue == PQ_CACHE) 2703 mycpu->gd_cnt.v_reactivated++; 2704 if (m->act_count < ACT_INIT) 2705 m->act_count = ACT_INIT; 2706 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2707 _vm_page_and_queue_spin_unlock(m); 2708 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2709 pagedaemon_wakeup(); 2710 } else { 2711 if (m->act_count < ACT_INIT) 2712 m->act_count = ACT_INIT; 2713 vm_page_spin_unlock(m); 2714 } 2715 } 2716 2717 void 2718 vm_page_soft_activate(vm_page_t m) 2719 { 2720 if (m->queue - m->pc == PQ_ACTIVE || 2721 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2722 if (m->act_count < ACT_INIT) 2723 m->act_count = ACT_INIT; 2724 } else { 2725 vm_page_activate(m); 2726 } 2727 } 2728 2729 /* 2730 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2731 * routine is called when a page has been added to the cache or free 2732 * queues. 2733 * 2734 * This routine may not block. 2735 */ 2736 static __inline void 2737 vm_page_free_wakeup(void) 2738 { 2739 globaldata_t gd = mycpu; 2740 2741 /* 2742 * If the pageout daemon itself needs pages, then tell it that 2743 * there are some free. 2744 */ 2745 if (vm_pageout_pages_needed && 2746 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2747 gd->gd_vmstats.v_pageout_free_min 2748 ) { 2749 vm_pageout_pages_needed = 0; 2750 wakeup(&vm_pageout_pages_needed); 2751 } 2752 2753 /* 2754 * Wakeup processes that are waiting on memory. 2755 * 2756 * Generally speaking we want to wakeup stuck processes as soon as 2757 * possible. !vm_page_count_min(0) is the absolute minimum point 2758 * where we can do this. Wait a bit longer to reduce degenerate 2759 * re-blocking (vm_page_free_hysteresis). The target check is just 2760 * to make sure the min-check w/hysteresis does not exceed the 2761 * normal target. 2762 */ 2763 if (vm_pages_waiting) { 2764 if (!vm_page_count_min(vm_page_free_hysteresis) || 2765 !vm_page_count_target()) { 2766 vm_pages_waiting = 0; 2767 wakeup(&vmstats.v_free_count); 2768 ++mycpu->gd_cnt.v_ppwakeups; 2769 } 2770 #if 0 2771 if (!vm_page_count_target()) { 2772 /* 2773 * Plenty of pages are free, wakeup everyone. 2774 */ 2775 vm_pages_waiting = 0; 2776 wakeup(&vmstats.v_free_count); 2777 ++mycpu->gd_cnt.v_ppwakeups; 2778 } else if (!vm_page_count_min(0)) { 2779 /* 2780 * Some pages are free, wakeup someone. 2781 */ 2782 int wcount = vm_pages_waiting; 2783 if (wcount > 0) 2784 --wcount; 2785 vm_pages_waiting = wcount; 2786 wakeup_one(&vmstats.v_free_count); 2787 ++mycpu->gd_cnt.v_ppwakeups; 2788 } 2789 #endif 2790 } 2791 } 2792 2793 /* 2794 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2795 * it from its VM object. 2796 * 2797 * The vm_page must be BUSY on entry. BUSY will be released on 2798 * return (the page will have been freed). 2799 */ 2800 void 2801 vm_page_free_toq(vm_page_t m) 2802 { 2803 mycpu->gd_cnt.v_tfree++; 2804 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) 2805 pmap_mapped_sync(m); 2806 KKASSERT((m->flags & PG_MAPPED) == 0); 2807 KKASSERT(m->busy_count & PBUSY_LOCKED); 2808 2809 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2810 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2811 "hold(%d)\n", 2812 (u_long)m->pindex, m->busy_count, m->hold_count); 2813 if ((m->queue - m->pc) == PQ_FREE) 2814 panic("vm_page_free: freeing free page"); 2815 else 2816 panic("vm_page_free: freeing busy page"); 2817 } 2818 2819 /* 2820 * Remove from object, spinlock the page and its queues and 2821 * remove from any queue. No queue spinlock will be held 2822 * after this section (because the page was removed from any 2823 * queue). 2824 */ 2825 vm_page_remove(m); 2826 2827 /* 2828 * No further management of fictitious pages occurs beyond object 2829 * and queue removal. 2830 */ 2831 if ((m->flags & PG_FICTITIOUS) != 0) { 2832 KKASSERT(m->queue == PQ_NONE); 2833 vm_page_wakeup(m); 2834 return; 2835 } 2836 vm_page_and_queue_spin_lock(m); 2837 _vm_page_rem_queue_spinlocked(m); 2838 2839 m->valid = 0; 2840 vm_page_undirty(m); 2841 2842 if (m->wire_count != 0) { 2843 if (m->wire_count > 1) { 2844 panic( 2845 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2846 m->wire_count, (long)m->pindex); 2847 } 2848 panic("vm_page_free: freeing wired page"); 2849 } 2850 2851 /* 2852 * Clear the PG_NEED_COMMIT and the PG_UNQUEUED flags. The 2853 * page returns to normal operation and will be placed in 2854 * the PQ_HOLD or PQ_FREE queue. 2855 */ 2856 vm_page_flag_clear(m, PG_NEED_COMMIT | PG_UNQUEUED); 2857 2858 if (m->hold_count != 0) { 2859 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2860 } else { 2861 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2862 } 2863 2864 /* 2865 * This sequence allows us to clear BUSY while still holding 2866 * its spin lock, which reduces contention vs allocators. We 2867 * must not leave the queue locked or _vm_page_wakeup() may 2868 * deadlock. 2869 */ 2870 _vm_page_queue_spin_unlock(m); 2871 if (_vm_page_wakeup(m)) { 2872 vm_page_spin_unlock(m); 2873 wakeup(m); 2874 } else { 2875 vm_page_spin_unlock(m); 2876 } 2877 vm_page_free_wakeup(); 2878 } 2879 2880 /* 2881 * Mark this page as wired down by yet another map. We do not adjust the 2882 * queue the page is on, it will be checked for wiring as-needed. 2883 * 2884 * This function has no effect on fictitious pages. 2885 * 2886 * Caller must be holding the page busy. 2887 */ 2888 void 2889 vm_page_wire(vm_page_t m) 2890 { 2891 KKASSERT(m->busy_count & PBUSY_LOCKED); 2892 if ((m->flags & PG_FICTITIOUS) == 0) { 2893 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2894 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2895 } 2896 KASSERT(m->wire_count != 0, 2897 ("vm_page_wire: wire_count overflow m=%p", m)); 2898 } 2899 } 2900 2901 /* 2902 * Release one wiring of this page, potentially enabling it to be paged again. 2903 * 2904 * Note that wired pages are no longer unconditionally removed from the 2905 * paging queues, so the page may already be on a queue. Move the page 2906 * to the desired queue if necessary. 2907 * 2908 * Many pages placed on the inactive queue should actually go 2909 * into the cache, but it is difficult to figure out which. What 2910 * we do instead, if the inactive target is well met, is to put 2911 * clean pages at the head of the inactive queue instead of the tail. 2912 * This will cause them to be moved to the cache more quickly and 2913 * if not actively re-referenced, freed more quickly. If we just 2914 * stick these pages at the end of the inactive queue, heavy filesystem 2915 * meta-data accesses can cause an unnecessary paging load on memory bound 2916 * processes. This optimization causes one-time-use metadata to be 2917 * reused more quickly. 2918 * 2919 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2920 * the inactive queue. This helps the pageout daemon determine memory 2921 * pressure and act on out-of-memory situations more quickly. 2922 * 2923 * BUT, if we are in a low-memory situation we have no choice but to 2924 * put clean pages on the cache queue. 2925 * 2926 * A number of routines use vm_page_unwire() to guarantee that the page 2927 * will go into either the inactive or active queues, and will NEVER 2928 * be placed in the cache - for example, just after dirtying a page. 2929 * dirty pages in the cache are not allowed. 2930 * 2931 * PG_FICTITIOUS or PG_UNQUEUED pages are never moved to any queue, and 2932 * the wire_count will not be adjusted in any way for a PG_FICTITIOUS 2933 * page. 2934 * 2935 * This routine may not block. 2936 */ 2937 void 2938 vm_page_unwire(vm_page_t m, int activate) 2939 { 2940 KKASSERT(m->busy_count & PBUSY_LOCKED); 2941 if (m->flags & PG_FICTITIOUS) { 2942 /* do nothing */ 2943 } else if ((int)m->wire_count <= 0) { 2944 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2945 } else { 2946 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2947 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2948 if (m->flags & PG_UNQUEUED) { 2949 ; 2950 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2951 vm_page_activate(m); 2952 } else { 2953 vm_page_deactivate(m); 2954 } 2955 } 2956 } 2957 } 2958 2959 /* 2960 * Move the specified page to the inactive queue. 2961 * 2962 * Normally athead is 0 resulting in LRU operation. athead is set 2963 * to 1 if we want this page to be 'as if it were placed in the cache', 2964 * except without unmapping it from the process address space. 2965 * 2966 * vm_page's spinlock must be held on entry and will remain held on return. 2967 * This routine may not block. The caller does not have to hold the page 2968 * busied but should have some sort of interlock on its validity. 2969 * 2970 * It is ok if the page is wired (so buffer cache operations don't have 2971 * to mess with the page queues). 2972 */ 2973 static void 2974 _vm_page_deactivate_locked(vm_page_t m, int athead) 2975 { 2976 u_short oqueue; 2977 2978 /* 2979 * Ignore if already inactive. 2980 */ 2981 if (m->queue - m->pc == PQ_INACTIVE || 2982 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED))) { 2983 return; 2984 } 2985 2986 _vm_page_queue_spin_lock(m); 2987 oqueue = _vm_page_rem_queue_spinlocked(m); 2988 2989 if ((m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 2990 if (oqueue == PQ_CACHE) 2991 mycpu->gd_cnt.v_reactivated++; 2992 vm_page_flag_clear(m, PG_WINATCFLS); 2993 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2994 if (athead == 0) { 2995 atomic_add_long( 2996 &vm_page_queues[PQ_INACTIVE + m->pc].adds, 1); 2997 } 2998 } 2999 /* NOTE: PQ_NONE if condition not taken */ 3000 _vm_page_queue_spin_unlock(m); 3001 /* leaves vm_page spinlocked */ 3002 } 3003 3004 /* 3005 * Attempt to deactivate a page. 3006 * 3007 * No requirements. We can pre-filter before getting the spinlock. 3008 * 3009 * It is ok if the page is wired (so buffer cache operations don't have 3010 * to mess with the page queues). 3011 */ 3012 void 3013 vm_page_deactivate(vm_page_t m) 3014 { 3015 if (m->queue - m->pc != PQ_INACTIVE && 3016 (m->flags & (PG_FICTITIOUS | PG_UNQUEUED)) == 0) { 3017 vm_page_spin_lock(m); 3018 _vm_page_deactivate_locked(m, 0); 3019 vm_page_spin_unlock(m); 3020 } 3021 } 3022 3023 void 3024 vm_page_deactivate_locked(vm_page_t m) 3025 { 3026 _vm_page_deactivate_locked(m, 0); 3027 } 3028 3029 /* 3030 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 3031 * 3032 * This function returns non-zero if it successfully moved the page to 3033 * PQ_CACHE. 3034 * 3035 * This function unconditionally unbusies the page on return. 3036 */ 3037 int 3038 vm_page_try_to_cache(vm_page_t m) 3039 { 3040 /* 3041 * Shortcut if we obviously cannot move the page, or if the 3042 * page is already on the cache queue, or it is ficitious. 3043 * 3044 * Never allow a wired page into the cache. 3045 */ 3046 if (m->dirty || m->hold_count || m->wire_count || 3047 m->queue - m->pc == PQ_CACHE || 3048 (m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS))) { 3049 vm_page_wakeup(m); 3050 return(0); 3051 } 3052 3053 /* 3054 * Page busied by us and no longer spinlocked. Dirty pages cannot 3055 * be moved to the cache. 3056 */ 3057 vm_page_test_dirty(m); 3058 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3059 vm_page_wakeup(m); 3060 return(0); 3061 } 3062 vm_page_cache(m); 3063 return(1); 3064 } 3065 3066 /* 3067 * Attempt to free the page. If we cannot free it, we do nothing. 3068 * 1 is returned on success, 0 on failure. 3069 * 3070 * The page can be in any state, including already being on the free 3071 * queue. Check to see if it really can be freed. Note that we disallow 3072 * this ad-hoc operation if the page is flagged PG_UNQUEUED. 3073 * 3074 * Caller provides an unlocked/non-busied page. 3075 * No requirements. 3076 */ 3077 int 3078 vm_page_try_to_free(vm_page_t m) 3079 { 3080 if (vm_page_busy_try(m, TRUE)) 3081 return(0); 3082 3083 if (m->dirty || /* can't free if it is dirty */ 3084 m->hold_count || /* or held (XXX may be wrong) */ 3085 m->wire_count || /* or wired */ 3086 (m->flags & (PG_UNQUEUED | /* or unqueued */ 3087 PG_NEED_COMMIT | /* or needs a commit */ 3088 PG_FICTITIOUS)) || /* or is fictitious */ 3089 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 3090 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 3091 vm_page_wakeup(m); 3092 return(0); 3093 } 3094 3095 /* 3096 * We can probably free the page. 3097 * 3098 * Page busied by us and no longer spinlocked. Dirty pages will 3099 * not be freed by this function. We have to re-test the 3100 * dirty bit after cleaning out the pmaps. 3101 */ 3102 vm_page_test_dirty(m); 3103 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3104 vm_page_wakeup(m); 3105 return(0); 3106 } 3107 vm_page_protect(m, VM_PROT_NONE); 3108 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3109 vm_page_wakeup(m); 3110 return(0); 3111 } 3112 vm_page_free(m); 3113 return(1); 3114 } 3115 3116 /* 3117 * vm_page_cache 3118 * 3119 * Put the specified page onto the page cache queue (if appropriate). 3120 * 3121 * The page must be busy, and this routine will release the busy and 3122 * possibly even free the page. 3123 */ 3124 void 3125 vm_page_cache(vm_page_t m) 3126 { 3127 /* 3128 * Not suitable for the cache 3129 */ 3130 if ((m->flags & (PG_UNQUEUED | PG_NEED_COMMIT | PG_FICTITIOUS)) || 3131 (m->busy_count & PBUSY_MASK) || 3132 m->wire_count || m->hold_count) { 3133 vm_page_wakeup(m); 3134 return; 3135 } 3136 3137 /* 3138 * Already in the cache (and thus not mapped) 3139 */ 3140 if ((m->queue - m->pc) == PQ_CACHE) { 3141 if (m->flags & (PG_MAPPED | PG_WRITEABLE)) 3142 pmap_mapped_sync(m); 3143 KKASSERT((m->flags & PG_MAPPED) == 0); 3144 vm_page_wakeup(m); 3145 return; 3146 } 3147 3148 #if 0 3149 /* 3150 * REMOVED - it is possible for dirty to get set at any time as 3151 * long as the page is still mapped and writeable. 3152 * 3153 * Caller is required to test m->dirty, but note that the act of 3154 * removing the page from its maps can cause it to become dirty 3155 * on an SMP system due to another cpu running in usermode. 3156 */ 3157 if (m->dirty) { 3158 panic("vm_page_cache: caching a dirty page, pindex: %ld", 3159 (long)m->pindex); 3160 } 3161 #endif 3162 3163 /* 3164 * Remove all pmaps and indicate that the page is not 3165 * writeable or mapped. Our vm_page_protect() call may 3166 * have blocked (especially w/ VM_PROT_NONE), so recheck 3167 * everything. 3168 */ 3169 vm_page_protect(m, VM_PROT_NONE); 3170 pmap_mapped_sync(m); 3171 if ((m->flags & (PG_UNQUEUED | PG_MAPPED)) || 3172 (m->busy_count & PBUSY_MASK) || 3173 m->wire_count || m->hold_count) { 3174 vm_page_wakeup(m); 3175 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3176 vm_page_deactivate(m); 3177 vm_page_wakeup(m); 3178 } else { 3179 _vm_page_and_queue_spin_lock(m); 3180 _vm_page_rem_queue_spinlocked(m); 3181 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 3182 _vm_page_and_queue_spin_unlock(m); 3183 vm_page_wakeup(m); 3184 vm_page_free_wakeup(); 3185 } 3186 } 3187 3188 /* 3189 * vm_page_dontneed() 3190 * 3191 * Cache, deactivate, or do nothing as appropriate. This routine 3192 * is typically used by madvise() MADV_DONTNEED. 3193 * 3194 * Generally speaking we want to move the page into the cache so 3195 * it gets reused quickly. However, this can result in a silly syndrome 3196 * due to the page recycling too quickly. Small objects will not be 3197 * fully cached. On the otherhand, if we move the page to the inactive 3198 * queue we wind up with a problem whereby very large objects 3199 * unnecessarily blow away our inactive and cache queues. 3200 * 3201 * The solution is to move the pages based on a fixed weighting. We 3202 * either leave them alone, deactivate them, or move them to the cache, 3203 * where moving them to the cache has the highest weighting. 3204 * By forcing some pages into other queues we eventually force the 3205 * system to balance the queues, potentially recovering other unrelated 3206 * space from active. The idea is to not force this to happen too 3207 * often. 3208 * 3209 * The page must be busied. 3210 */ 3211 void 3212 vm_page_dontneed(vm_page_t m) 3213 { 3214 static int dnweight; 3215 int dnw; 3216 int head; 3217 3218 dnw = ++dnweight; 3219 3220 /* 3221 * occassionally leave the page alone 3222 */ 3223 if ((dnw & 0x01F0) == 0 || 3224 m->queue - m->pc == PQ_INACTIVE || 3225 m->queue - m->pc == PQ_CACHE 3226 ) { 3227 if (m->act_count >= ACT_INIT) 3228 --m->act_count; 3229 return; 3230 } 3231 3232 /* 3233 * If vm_page_dontneed() is inactivating a page, it must clear 3234 * the referenced flag; otherwise the pagedaemon will see references 3235 * on the page in the inactive queue and reactivate it. Until the 3236 * page can move to the cache queue, madvise's job is not done. 3237 */ 3238 vm_page_flag_clear(m, PG_REFERENCED); 3239 pmap_clear_reference(m); 3240 3241 if (m->dirty == 0) 3242 vm_page_test_dirty(m); 3243 3244 if (m->dirty || (dnw & 0x0070) == 0) { 3245 /* 3246 * Deactivate the page 3 times out of 32. 3247 */ 3248 head = 0; 3249 } else { 3250 /* 3251 * Cache the page 28 times out of every 32. Note that 3252 * the page is deactivated instead of cached, but placed 3253 * at the head of the queue instead of the tail. 3254 */ 3255 head = 1; 3256 } 3257 vm_page_spin_lock(m); 3258 _vm_page_deactivate_locked(m, head); 3259 vm_page_spin_unlock(m); 3260 } 3261 3262 /* 3263 * These routines manipulate the 'soft busy' count for a page. A soft busy 3264 * is almost like a hard BUSY except that it allows certain compatible 3265 * operations to occur on the page while it is busy. For example, a page 3266 * undergoing a write can still be mapped read-only. 3267 * 3268 * We also use soft-busy to quickly pmap_enter shared read-only pages 3269 * without having to hold the page locked. 3270 * 3271 * The soft-busy count can be > 1 in situations where multiple threads 3272 * are pmap_enter()ing the same page simultaneously, or when two buffer 3273 * cache buffers overlap the same page. 3274 * 3275 * The caller must hold the page BUSY when making these two calls. 3276 */ 3277 void 3278 vm_page_io_start(vm_page_t m) 3279 { 3280 uint32_t ocount; 3281 3282 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3283 KKASSERT(ocount & PBUSY_LOCKED); 3284 } 3285 3286 void 3287 vm_page_io_finish(vm_page_t m) 3288 { 3289 uint32_t ocount; 3290 3291 ocount = atomic_fetchadd_int(&m->busy_count, -1); 3292 KKASSERT(ocount & PBUSY_MASK); 3293 #if 0 3294 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 3295 wakeup(m); 3296 #endif 3297 } 3298 3299 /* 3300 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 3301 * 3302 * We can't use fetchadd here because we might race a hard-busy and the 3303 * page freeing code asserts on a non-zero soft-busy count (even if only 3304 * temporary). 3305 * 3306 * Returns 0 on success, non-zero on failure. 3307 */ 3308 int 3309 vm_page_sbusy_try(vm_page_t m) 3310 { 3311 uint32_t ocount; 3312 3313 for (;;) { 3314 ocount = m->busy_count; 3315 cpu_ccfence(); 3316 if (ocount & PBUSY_LOCKED) 3317 return 1; 3318 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1)) 3319 break; 3320 } 3321 return 0; 3322 #if 0 3323 if (m->busy_count & PBUSY_LOCKED) 3324 return 1; 3325 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3326 if (ocount & PBUSY_LOCKED) { 3327 vm_page_sbusy_drop(m); 3328 return 1; 3329 } 3330 return 0; 3331 #endif 3332 } 3333 3334 /* 3335 * Indicate that a clean VM page requires a filesystem commit and cannot 3336 * be reused. Used by tmpfs. 3337 */ 3338 void 3339 vm_page_need_commit(vm_page_t m) 3340 { 3341 vm_page_flag_set(m, PG_NEED_COMMIT); 3342 vm_object_set_writeable_dirty(m->object); 3343 } 3344 3345 void 3346 vm_page_clear_commit(vm_page_t m) 3347 { 3348 vm_page_flag_clear(m, PG_NEED_COMMIT); 3349 } 3350 3351 /* 3352 * Grab a page, blocking if it is busy and allocating a page if necessary. 3353 * A busy page is returned or NULL. The page may or may not be valid and 3354 * might not be on a queue (the caller is responsible for the disposition of 3355 * the page). 3356 * 3357 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 3358 * page will be zero'd and marked valid. 3359 * 3360 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 3361 * valid even if it already exists. 3362 * 3363 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 3364 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 3365 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 3366 * 3367 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 3368 * always returned if we had blocked. 3369 * 3370 * This routine may not be called from an interrupt. 3371 * 3372 * No other requirements. 3373 */ 3374 vm_page_t 3375 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3376 { 3377 vm_page_t m; 3378 int error; 3379 int shared = 1; 3380 3381 KKASSERT(allocflags & 3382 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 3383 vm_object_hold_shared(object); 3384 for (;;) { 3385 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3386 if (error) { 3387 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 3388 if ((allocflags & VM_ALLOC_RETRY) == 0) { 3389 m = NULL; 3390 break; 3391 } 3392 /* retry */ 3393 } else if (m == NULL) { 3394 if (shared) { 3395 vm_object_upgrade(object); 3396 shared = 0; 3397 } 3398 if (allocflags & VM_ALLOC_RETRY) 3399 allocflags |= VM_ALLOC_NULL_OK; 3400 m = vm_page_alloc(object, pindex, 3401 allocflags & ~VM_ALLOC_RETRY); 3402 if (m) 3403 break; 3404 vm_wait(0); 3405 if ((allocflags & VM_ALLOC_RETRY) == 0) 3406 goto failed; 3407 } else { 3408 /* m found */ 3409 break; 3410 } 3411 } 3412 3413 /* 3414 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 3415 * 3416 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 3417 * valid even if already valid. 3418 * 3419 * NOTE! We have removed all of the PG_ZERO optimizations and also 3420 * removed the idle zeroing code. These optimizations actually 3421 * slow things down on modern cpus because the zerod area is 3422 * likely uncached, placing a memory-access burden on the 3423 * accesors taking the fault. 3424 * 3425 * By always zeroing the page in-line with the fault, no 3426 * dynamic ram reads are needed and the caches are hot, ready 3427 * for userland to access the memory. 3428 */ 3429 if (m->valid == 0) { 3430 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 3431 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3432 m->valid = VM_PAGE_BITS_ALL; 3433 } 3434 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 3435 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3436 m->valid = VM_PAGE_BITS_ALL; 3437 } 3438 failed: 3439 vm_object_drop(object); 3440 return(m); 3441 } 3442 3443 /* 3444 * Mapping function for valid bits or for dirty bits in 3445 * a page. May not block. 3446 * 3447 * Inputs are required to range within a page. 3448 * 3449 * No requirements. 3450 * Non blocking. 3451 */ 3452 int 3453 vm_page_bits(int base, int size) 3454 { 3455 int first_bit; 3456 int last_bit; 3457 3458 KASSERT( 3459 base + size <= PAGE_SIZE, 3460 ("vm_page_bits: illegal base/size %d/%d", base, size) 3461 ); 3462 3463 if (size == 0) /* handle degenerate case */ 3464 return(0); 3465 3466 first_bit = base >> DEV_BSHIFT; 3467 last_bit = (base + size - 1) >> DEV_BSHIFT; 3468 3469 return ((2 << last_bit) - (1 << first_bit)); 3470 } 3471 3472 /* 3473 * Sets portions of a page valid and clean. The arguments are expected 3474 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3475 * of any partial chunks touched by the range. The invalid portion of 3476 * such chunks will be zero'd. 3477 * 3478 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 3479 * align base to DEV_BSIZE so as not to mark clean a partially 3480 * truncated device block. Otherwise the dirty page status might be 3481 * lost. 3482 * 3483 * This routine may not block. 3484 * 3485 * (base + size) must be less then or equal to PAGE_SIZE. 3486 */ 3487 static void 3488 _vm_page_zero_valid(vm_page_t m, int base, int size) 3489 { 3490 int frag; 3491 int endoff; 3492 3493 if (size == 0) /* handle degenerate case */ 3494 return; 3495 3496 /* 3497 * If the base is not DEV_BSIZE aligned and the valid 3498 * bit is clear, we have to zero out a portion of the 3499 * first block. 3500 */ 3501 3502 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3503 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3504 ) { 3505 pmap_zero_page_area( 3506 VM_PAGE_TO_PHYS(m), 3507 frag, 3508 base - frag 3509 ); 3510 } 3511 3512 /* 3513 * If the ending offset is not DEV_BSIZE aligned and the 3514 * valid bit is clear, we have to zero out a portion of 3515 * the last block. 3516 */ 3517 3518 endoff = base + size; 3519 3520 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3521 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3522 ) { 3523 pmap_zero_page_area( 3524 VM_PAGE_TO_PHYS(m), 3525 endoff, 3526 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3527 ); 3528 } 3529 } 3530 3531 /* 3532 * Set valid, clear dirty bits. If validating the entire 3533 * page we can safely clear the pmap modify bit. We also 3534 * use this opportunity to clear the PG_NOSYNC flag. If a process 3535 * takes a write fault on a MAP_NOSYNC memory area the flag will 3536 * be set again. 3537 * 3538 * We set valid bits inclusive of any overlap, but we can only 3539 * clear dirty bits for DEV_BSIZE chunks that are fully within 3540 * the range. 3541 * 3542 * Page must be busied? 3543 * No other requirements. 3544 */ 3545 void 3546 vm_page_set_valid(vm_page_t m, int base, int size) 3547 { 3548 _vm_page_zero_valid(m, base, size); 3549 m->valid |= vm_page_bits(base, size); 3550 } 3551 3552 3553 /* 3554 * Set valid bits and clear dirty bits. 3555 * 3556 * Page must be busied by caller. 3557 * 3558 * NOTE: This function does not clear the pmap modified bit. 3559 * Also note that e.g. NFS may use a byte-granular base 3560 * and size. 3561 * 3562 * No other requirements. 3563 */ 3564 void 3565 vm_page_set_validclean(vm_page_t m, int base, int size) 3566 { 3567 int pagebits; 3568 3569 _vm_page_zero_valid(m, base, size); 3570 pagebits = vm_page_bits(base, size); 3571 m->valid |= pagebits; 3572 m->dirty &= ~pagebits; 3573 if (base == 0 && size == PAGE_SIZE) { 3574 /*pmap_clear_modify(m);*/ 3575 vm_page_flag_clear(m, PG_NOSYNC); 3576 } 3577 } 3578 3579 /* 3580 * Set valid & dirty. Used by buwrite() 3581 * 3582 * Page must be busied by caller. 3583 */ 3584 void 3585 vm_page_set_validdirty(vm_page_t m, int base, int size) 3586 { 3587 int pagebits; 3588 3589 pagebits = vm_page_bits(base, size); 3590 m->valid |= pagebits; 3591 m->dirty |= pagebits; 3592 if (m->object) 3593 vm_object_set_writeable_dirty(m->object); 3594 } 3595 3596 /* 3597 * Clear dirty bits. 3598 * 3599 * NOTE: This function does not clear the pmap modified bit. 3600 * Also note that e.g. NFS may use a byte-granular base 3601 * and size. 3602 * 3603 * Page must be busied? 3604 * No other requirements. 3605 */ 3606 void 3607 vm_page_clear_dirty(vm_page_t m, int base, int size) 3608 { 3609 m->dirty &= ~vm_page_bits(base, size); 3610 if (base == 0 && size == PAGE_SIZE) { 3611 /*pmap_clear_modify(m);*/ 3612 vm_page_flag_clear(m, PG_NOSYNC); 3613 } 3614 } 3615 3616 /* 3617 * Make the page all-dirty. 3618 * 3619 * Also make sure the related object and vnode reflect the fact that the 3620 * object may now contain a dirty page. 3621 * 3622 * Page must be busied? 3623 * No other requirements. 3624 */ 3625 void 3626 vm_page_dirty(vm_page_t m) 3627 { 3628 #ifdef INVARIANTS 3629 int pqtype = m->queue - m->pc; 3630 #endif 3631 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3632 ("vm_page_dirty: page in free/cache queue!")); 3633 if (m->dirty != VM_PAGE_BITS_ALL) { 3634 m->dirty = VM_PAGE_BITS_ALL; 3635 if (m->object) 3636 vm_object_set_writeable_dirty(m->object); 3637 } 3638 } 3639 3640 /* 3641 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3642 * valid and dirty bits for the effected areas are cleared. 3643 * 3644 * Page must be busied? 3645 * Does not block. 3646 * No other requirements. 3647 */ 3648 void 3649 vm_page_set_invalid(vm_page_t m, int base, int size) 3650 { 3651 int bits; 3652 3653 bits = vm_page_bits(base, size); 3654 m->valid &= ~bits; 3655 m->dirty &= ~bits; 3656 atomic_add_int(&m->object->generation, 1); 3657 } 3658 3659 /* 3660 * The kernel assumes that the invalid portions of a page contain 3661 * garbage, but such pages can be mapped into memory by user code. 3662 * When this occurs, we must zero out the non-valid portions of the 3663 * page so user code sees what it expects. 3664 * 3665 * Pages are most often semi-valid when the end of a file is mapped 3666 * into memory and the file's size is not page aligned. 3667 * 3668 * Page must be busied? 3669 * No other requirements. 3670 */ 3671 void 3672 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3673 { 3674 int b; 3675 int i; 3676 3677 /* 3678 * Scan the valid bits looking for invalid sections that 3679 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3680 * valid bit may be set ) have already been zerod by 3681 * vm_page_set_validclean(). 3682 */ 3683 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3684 if (i == (PAGE_SIZE / DEV_BSIZE) || 3685 (m->valid & (1 << i)) 3686 ) { 3687 if (i > b) { 3688 pmap_zero_page_area( 3689 VM_PAGE_TO_PHYS(m), 3690 b << DEV_BSHIFT, 3691 (i - b) << DEV_BSHIFT 3692 ); 3693 } 3694 b = i + 1; 3695 } 3696 } 3697 3698 /* 3699 * setvalid is TRUE when we can safely set the zero'd areas 3700 * as being valid. We can do this if there are no cache consistency 3701 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3702 */ 3703 if (setvalid) 3704 m->valid = VM_PAGE_BITS_ALL; 3705 } 3706 3707 /* 3708 * Is a (partial) page valid? Note that the case where size == 0 3709 * will return FALSE in the degenerate case where the page is entirely 3710 * invalid, and TRUE otherwise. 3711 * 3712 * Does not block. 3713 * No other requirements. 3714 */ 3715 int 3716 vm_page_is_valid(vm_page_t m, int base, int size) 3717 { 3718 int bits = vm_page_bits(base, size); 3719 3720 if (m->valid && ((m->valid & bits) == bits)) 3721 return 1; 3722 else 3723 return 0; 3724 } 3725 3726 /* 3727 * Update dirty bits from pmap/mmu. May not block. 3728 * 3729 * Caller must hold the page busy 3730 * 3731 * WARNING! Unless the page has been unmapped, this function only 3732 * provides a likely dirty status. 3733 */ 3734 void 3735 vm_page_test_dirty(vm_page_t m) 3736 { 3737 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) { 3738 vm_page_dirty(m); 3739 } 3740 } 3741 3742 #include "opt_ddb.h" 3743 #ifdef DDB 3744 #include <ddb/ddb.h> 3745 3746 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3747 { 3748 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3749 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3750 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3751 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3752 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3753 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3754 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3755 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3756 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3757 db_printf("vmstats.v_inactive_target: %ld\n", 3758 vmstats.v_inactive_target); 3759 } 3760 3761 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3762 { 3763 int i; 3764 db_printf("PQ_FREE:"); 3765 for (i = 0; i < PQ_L2_SIZE; i++) { 3766 db_printf(" %ld", vm_page_queues[PQ_FREE + i].lcnt); 3767 } 3768 db_printf("\n"); 3769 3770 db_printf("PQ_CACHE:"); 3771 for(i = 0; i < PQ_L2_SIZE; i++) { 3772 db_printf(" %ld", vm_page_queues[PQ_CACHE + i].lcnt); 3773 } 3774 db_printf("\n"); 3775 3776 db_printf("PQ_ACTIVE:"); 3777 for(i = 0; i < PQ_L2_SIZE; i++) { 3778 db_printf(" %ld", vm_page_queues[PQ_ACTIVE + i].lcnt); 3779 } 3780 db_printf("\n"); 3781 3782 db_printf("PQ_INACTIVE:"); 3783 for(i = 0; i < PQ_L2_SIZE; i++) { 3784 db_printf(" %ld", vm_page_queues[PQ_INACTIVE + i].lcnt); 3785 } 3786 db_printf("\n"); 3787 } 3788 #endif /* DDB */ 3789