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