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