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 * The vm_page must NOT be FICTITIOUS (that would be a disaster) 984 * This function will return with both the page and the queue locked. 985 */ 986 static __inline void 987 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 988 { 989 struct vpgqueues *pq; 990 u_long *cnt; 991 992 KKASSERT(m->queue == PQ_NONE && (m->flags & PG_FICTITIOUS) == 0); 993 994 if (queue != PQ_NONE) { 995 vm_page_queues_spin_lock(queue); 996 pq = &vm_page_queues[queue]; 997 ++pq->lcnt; 998 999 /* 1000 * Adjust our pcpu stats. If a system entity really needs 1001 * to incorporate the count it will call vmstats_rollup() 1002 * to roll it all up into the global vmstats strufture. 1003 */ 1004 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 1005 atomic_add_long(cnt, 1); 1006 1007 /* 1008 * PQ_FREE is always handled LIFO style to try to provide 1009 * cache-hot pages to programs. 1010 */ 1011 m->queue = queue; 1012 if (queue - m->pc == PQ_FREE) { 1013 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1014 } else if (athead) { 1015 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1016 } else { 1017 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1018 } 1019 /* leave the queue spinlocked */ 1020 } 1021 } 1022 1023 /* 1024 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 1025 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 1026 * 1027 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 1028 * call will be made before returning. 1029 * 1030 * This function does NOT busy the page and on return the page is not 1031 * guaranteed to be available. 1032 */ 1033 void 1034 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 1035 { 1036 u_int32_t busy_count; 1037 1038 for (;;) { 1039 busy_count = m->busy_count; 1040 cpu_ccfence(); 1041 1042 if ((busy_count & PBUSY_LOCKED) == 0 && 1043 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 1044 break; 1045 } 1046 tsleep_interlock(m, 0); 1047 if (atomic_cmpset_int(&m->busy_count, busy_count, 1048 busy_count | PBUSY_WANTED)) { 1049 atomic_set_int(&m->flags, PG_REFERENCED); 1050 tsleep(m, PINTERLOCKED, msg, 0); 1051 break; 1052 } 1053 } 1054 } 1055 1056 /* 1057 * This calculates and returns a page color given an optional VM object and 1058 * either a pindex or an iterator. We attempt to return a cpu-localized 1059 * pg_color that is still roughly 16-way set-associative. The CPU topology 1060 * is used if it was probed. 1061 * 1062 * The caller may use the returned value to index into e.g. PQ_FREE when 1063 * allocating a page in order to nominally obtain pages that are hopefully 1064 * already localized to the requesting cpu. This function is not able to 1065 * provide any sort of guarantee of this, but does its best to improve 1066 * hardware cache management performance. 1067 * 1068 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 1069 */ 1070 u_short 1071 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 1072 { 1073 u_short pg_color; 1074 int object_pg_color; 1075 1076 /* 1077 * WARNING! cpu_topology_core_ids might not be a power of two. 1078 * We also shouldn't make assumptions about 1079 * cpu_topology_phys_ids either. 1080 * 1081 * WARNING! ncpus might not be known at this time (during early 1082 * boot), and might be set to 1. 1083 * 1084 * General format: [phys_id][core_id][cpuid][set-associativity] 1085 * (but uses modulo, so not necessarily precise bit masks) 1086 */ 1087 object_pg_color = object ? object->pg_color : 0; 1088 1089 if (cpu_topology_ht_ids) { 1090 int phys_id; 1091 int core_id; 1092 int ht_id; 1093 int physcale; 1094 int grpscale; 1095 int cpuscale; 1096 1097 /* 1098 * Translate cpuid to socket, core, and hyperthread id. 1099 */ 1100 phys_id = get_cpu_phys_id(cpuid); 1101 core_id = get_cpu_core_id(cpuid); 1102 ht_id = get_cpu_ht_id(cpuid); 1103 1104 /* 1105 * Calculate pg_color for our array index. 1106 * 1107 * physcale - socket multiplier. 1108 * grpscale - core multiplier (cores per socket) 1109 * cpu* - cpus per core 1110 * 1111 * WARNING! In early boot, ncpus has not yet been 1112 * initialized and may be set to (1). 1113 * 1114 * WARNING! physcale must match the organization that 1115 * vm_numa_organize() creates to ensure that 1116 * we properly localize allocations to the 1117 * requested cpuid. 1118 */ 1119 physcale = PQ_L2_SIZE / cpu_topology_phys_ids; 1120 grpscale = physcale / cpu_topology_core_ids; 1121 cpuscale = grpscale / cpu_topology_ht_ids; 1122 1123 pg_color = phys_id * physcale; 1124 pg_color += core_id * grpscale; 1125 pg_color += ht_id * cpuscale; 1126 pg_color += (pindex + object_pg_color) % cpuscale; 1127 1128 #if 0 1129 if (grpsize >= 8) { 1130 pg_color += (pindex + object_pg_color) % grpsize; 1131 } else { 1132 if (grpsize <= 2) { 1133 grpsize = 8; 1134 } else { 1135 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 1136 grpsize += grpsize; 1137 if (grpsize < 8) 1138 grpsize += grpsize; 1139 } 1140 pg_color += (pindex + object_pg_color) % grpsize; 1141 } 1142 #endif 1143 } else { 1144 /* 1145 * Unknown topology, distribute things evenly. 1146 * 1147 * WARNING! In early boot, ncpus has not yet been 1148 * initialized and may be set to (1). 1149 */ 1150 int cpuscale; 1151 1152 cpuscale = PQ_L2_SIZE / ncpus; 1153 1154 pg_color = cpuid * cpuscale; 1155 pg_color += (pindex + object_pg_color) % cpuscale; 1156 } 1157 return (pg_color & PQ_L2_MASK); 1158 } 1159 1160 /* 1161 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 1162 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 1163 */ 1164 void 1165 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 1166 int also_m_busy, const char *msg 1167 VM_PAGE_DEBUG_ARGS) 1168 { 1169 u_int32_t busy_count; 1170 1171 for (;;) { 1172 busy_count = m->busy_count; 1173 cpu_ccfence(); 1174 if (busy_count & PBUSY_LOCKED) { 1175 tsleep_interlock(m, 0); 1176 if (atomic_cmpset_int(&m->busy_count, busy_count, 1177 busy_count | PBUSY_WANTED)) { 1178 atomic_set_int(&m->flags, PG_REFERENCED); 1179 tsleep(m, PINTERLOCKED, msg, 0); 1180 } 1181 } else if (also_m_busy && busy_count) { 1182 tsleep_interlock(m, 0); 1183 if (atomic_cmpset_int(&m->busy_count, busy_count, 1184 busy_count | PBUSY_WANTED)) { 1185 atomic_set_int(&m->flags, PG_REFERENCED); 1186 tsleep(m, PINTERLOCKED, msg, 0); 1187 } 1188 } else { 1189 if (atomic_cmpset_int(&m->busy_count, busy_count, 1190 busy_count | PBUSY_LOCKED)) { 1191 #ifdef VM_PAGE_DEBUG 1192 m->busy_func = func; 1193 m->busy_line = lineno; 1194 #endif 1195 break; 1196 } 1197 } 1198 } 1199 } 1200 1201 /* 1202 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1203 * m->busy_count is also 0. 1204 * 1205 * Returns non-zero on failure. 1206 */ 1207 int 1208 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1209 VM_PAGE_DEBUG_ARGS) 1210 { 1211 u_int32_t busy_count; 1212 1213 for (;;) { 1214 busy_count = m->busy_count; 1215 cpu_ccfence(); 1216 if (busy_count & PBUSY_LOCKED) 1217 return TRUE; 1218 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1219 return TRUE; 1220 if (atomic_cmpset_int(&m->busy_count, busy_count, 1221 busy_count | PBUSY_LOCKED)) { 1222 #ifdef VM_PAGE_DEBUG 1223 m->busy_func = func; 1224 m->busy_line = lineno; 1225 #endif 1226 return FALSE; 1227 } 1228 } 1229 } 1230 1231 /* 1232 * Clear the BUSY flag and return non-zero to indicate to the caller 1233 * that a wakeup() should be performed. 1234 * 1235 * (inline version) 1236 */ 1237 static __inline 1238 int 1239 _vm_page_wakeup(vm_page_t m) 1240 { 1241 u_int32_t busy_count; 1242 1243 busy_count = m->busy_count; 1244 cpu_ccfence(); 1245 for (;;) { 1246 if (atomic_fcmpset_int(&m->busy_count, &busy_count, 1247 busy_count & 1248 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1249 return((int)(busy_count & PBUSY_WANTED)); 1250 } 1251 } 1252 /* not reached */ 1253 } 1254 1255 /* 1256 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1257 * is typically the last call you make on a page before moving onto 1258 * other things. 1259 */ 1260 void 1261 vm_page_wakeup(vm_page_t m) 1262 { 1263 KASSERT(m->busy_count & PBUSY_LOCKED, 1264 ("vm_page_wakeup: page not busy!!!")); 1265 if (_vm_page_wakeup(m)) 1266 wakeup(m); 1267 } 1268 1269 /* 1270 * Hold a page, preventing reuse. This is typically only called on pages 1271 * in a known state (either held busy, special, or interlocked in some 1272 * manner). Holding a page does not ensure that it remains valid, it only 1273 * prevents reuse. The page must not already be on the FREE queue or in 1274 * any danger of being moved to the FREE queue concurrent with this call. 1275 * 1276 * Other parts of the system can still disassociate the page from its object 1277 * and attempt to free it, or perform read or write I/O on it and/or otherwise 1278 * manipulate the page, but if the page is held the VM system will leave the 1279 * page and its data intact and not cycle it through the FREE queue until 1280 * the last hold has been released. 1281 * 1282 * (see vm_page_wire() if you want to prevent the page from being 1283 * disassociated from its object too). 1284 */ 1285 void 1286 vm_page_hold(vm_page_t m) 1287 { 1288 atomic_add_int(&m->hold_count, 1); 1289 KKASSERT(m->queue - m->pc != PQ_FREE); 1290 #if 0 1291 vm_page_spin_lock(m); 1292 atomic_add_int(&m->hold_count, 1); 1293 if (m->queue - m->pc == PQ_FREE) { 1294 _vm_page_queue_spin_lock(m); 1295 _vm_page_rem_queue_spinlocked(m); 1296 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1297 _vm_page_queue_spin_unlock(m); 1298 } 1299 vm_page_spin_unlock(m); 1300 #endif 1301 } 1302 1303 /* 1304 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1305 * it was freed while held and must be moved back to the FREE queue. 1306 * 1307 * To avoid racing against vm_page_free*() we must re-test conditions 1308 * after obtaining the spin-lock. The initial test can also race a 1309 * vm_page_free*() that is in the middle of moving a page to PQ_HOLD, 1310 * leaving the page on PQ_HOLD with hold_count == 0. Rather than 1311 * throw a spin-lock in the critical path, we rely on the pageout 1312 * daemon to clean-up these loose ends. 1313 * 1314 * More critically, the 'easy movement' between queues without busying 1315 * a vm_page is only allowed for PQ_FREE<->PQ_HOLD. 1316 */ 1317 void 1318 vm_page_unhold(vm_page_t m) 1319 { 1320 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1321 ("vm_page_unhold: pg %p illegal hold_count (%d) or " 1322 "on FREE queue (%d)", 1323 m, m->hold_count, m->queue - m->pc)); 1324 1325 if (atomic_fetchadd_int(&m->hold_count, -1) == 1 && 1326 m->queue - m->pc == PQ_HOLD) { 1327 vm_page_spin_lock(m); 1328 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1329 _vm_page_queue_spin_lock(m); 1330 _vm_page_rem_queue_spinlocked(m); 1331 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1332 _vm_page_queue_spin_unlock(m); 1333 } 1334 vm_page_spin_unlock(m); 1335 } 1336 } 1337 1338 /* 1339 * Create a fictitious page with the specified physical address and 1340 * memory attribute. The memory attribute is the only the machine- 1341 * dependent aspect of a fictitious page that must be initialized. 1342 */ 1343 void 1344 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1345 { 1346 1347 if ((m->flags & PG_FICTITIOUS) != 0) { 1348 /* 1349 * The page's memattr might have changed since the 1350 * previous initialization. Update the pmap to the 1351 * new memattr. 1352 */ 1353 goto memattr; 1354 } 1355 m->phys_addr = paddr; 1356 m->queue = PQ_NONE; 1357 /* Fictitious pages don't use "segind". */ 1358 /* Fictitious pages don't use "order" or "pool". */ 1359 m->flags = PG_FICTITIOUS | PG_UNMANAGED; 1360 m->busy_count = PBUSY_LOCKED; 1361 m->wire_count = 1; 1362 spin_init(&m->spin, "fake_page"); 1363 pmap_page_init(m); 1364 memattr: 1365 pmap_page_set_memattr(m, memattr); 1366 } 1367 1368 /* 1369 * Inserts the given vm_page into the object and object list. 1370 * 1371 * The pagetables are not updated but will presumably fault the page 1372 * in if necessary, or if a kernel page the caller will at some point 1373 * enter the page into the kernel's pmap. We are not allowed to block 1374 * here so we *can't* do this anyway. 1375 * 1376 * This routine may not block. 1377 * This routine must be called with the vm_object held. 1378 * This routine must be called with a critical section held. 1379 * 1380 * This routine returns TRUE if the page was inserted into the object 1381 * successfully, and FALSE if the page already exists in the object. 1382 */ 1383 int 1384 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1385 { 1386 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1387 if (m->object != NULL) 1388 panic("vm_page_insert: already inserted"); 1389 1390 atomic_add_int(&object->generation, 1); 1391 1392 /* 1393 * Record the object/offset pair in this page and add the 1394 * pv_list_count of the page to the object. 1395 * 1396 * The vm_page spin lock is required for interactions with the pmap. 1397 */ 1398 vm_page_spin_lock(m); 1399 m->object = object; 1400 m->pindex = pindex; 1401 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1402 m->object = NULL; 1403 m->pindex = 0; 1404 vm_page_spin_unlock(m); 1405 return FALSE; 1406 } 1407 ++object->resident_page_count; 1408 ++mycpu->gd_vmtotal.t_rm; 1409 vm_page_spin_unlock(m); 1410 1411 /* 1412 * Since we are inserting a new and possibly dirty page, 1413 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1414 */ 1415 if ((m->valid & m->dirty) || 1416 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1417 vm_object_set_writeable_dirty(object); 1418 1419 /* 1420 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1421 */ 1422 swap_pager_page_inserted(m); 1423 return TRUE; 1424 } 1425 1426 /* 1427 * Removes the given vm_page_t from the (object,index) table 1428 * 1429 * The underlying pmap entry (if any) is NOT removed here. 1430 * This routine may not block. 1431 * 1432 * The page must be BUSY and will remain BUSY on return. 1433 * No other requirements. 1434 * 1435 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1436 * it busy. 1437 */ 1438 void 1439 vm_page_remove(vm_page_t m) 1440 { 1441 vm_object_t object; 1442 1443 if (m->object == NULL) { 1444 return; 1445 } 1446 1447 if ((m->busy_count & PBUSY_LOCKED) == 0) 1448 panic("vm_page_remove: page not busy"); 1449 1450 object = m->object; 1451 1452 vm_object_hold(object); 1453 1454 /* 1455 * Remove the page from the object and update the object. 1456 * 1457 * The vm_page spin lock is required for interactions with the pmap. 1458 */ 1459 vm_page_spin_lock(m); 1460 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1461 --object->resident_page_count; 1462 --mycpu->gd_vmtotal.t_rm; 1463 m->object = NULL; 1464 atomic_add_int(&object->generation, 1); 1465 vm_page_spin_unlock(m); 1466 1467 vm_object_drop(object); 1468 } 1469 1470 /* 1471 * Calculate the hash position for the vm_page hash heuristic. 1472 */ 1473 static __inline 1474 struct vm_page ** 1475 vm_page_hash_hash(vm_object_t object, vm_pindex_t pindex) 1476 { 1477 size_t hi; 1478 1479 hi = (uintptr_t)object % (uintptr_t)vm_page_array_size + pindex; 1480 hi %= vm_page_array_size; 1481 return (&vm_page_hash[hi]); 1482 } 1483 1484 /* 1485 * Heuristical page lookup that does not require any locks. Returns 1486 * a soft-busied page on success, NULL on failure. 1487 * 1488 * Caller must lookup the page the slow way if NULL is returned. 1489 */ 1490 vm_page_t 1491 vm_page_hash_get(vm_object_t object, vm_pindex_t pindex) 1492 { 1493 struct vm_page **mp; 1494 vm_page_t m; 1495 1496 if (vm_page_hash == NULL) 1497 return NULL; 1498 mp = vm_page_hash_hash(object, pindex); 1499 m = *mp; 1500 cpu_ccfence(); 1501 if (m == NULL) 1502 return NULL; 1503 if (m->object != object || m->pindex != pindex) 1504 return NULL; 1505 if (vm_page_sbusy_try(m)) 1506 return NULL; 1507 if (m->object != object || m->pindex != pindex) { 1508 vm_page_wakeup(m); 1509 return NULL; 1510 } 1511 return m; 1512 } 1513 1514 /* 1515 * Enter page onto vm_page_hash[]. This is a heuristic, SMP collisions 1516 * are allowed. 1517 */ 1518 static __inline 1519 void 1520 vm_page_hash_enter(vm_page_t m) 1521 { 1522 struct vm_page **mp; 1523 1524 if (vm_page_hash && 1525 m > &vm_page_array[0] && 1526 m < &vm_page_array[vm_page_array_size]) { 1527 mp = vm_page_hash_hash(m->object, m->pindex); 1528 if (*mp != m) 1529 *mp = m; 1530 } 1531 } 1532 1533 /* 1534 * Locate and return the page at (object, pindex), or NULL if the 1535 * page could not be found. 1536 * 1537 * The caller must hold the vm_object token. 1538 */ 1539 vm_page_t 1540 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1541 { 1542 vm_page_t m; 1543 1544 /* 1545 * Search the hash table for this object/offset pair 1546 */ 1547 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1548 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1549 if (m) { 1550 KKASSERT(m->object == object && m->pindex == pindex); 1551 vm_page_hash_enter(m); 1552 } 1553 return(m); 1554 } 1555 1556 vm_page_t 1557 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1558 vm_pindex_t pindex, 1559 int also_m_busy, const char *msg 1560 VM_PAGE_DEBUG_ARGS) 1561 { 1562 u_int32_t busy_count; 1563 vm_page_t m; 1564 1565 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1566 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1567 while (m) { 1568 KKASSERT(m->object == object && m->pindex == pindex); 1569 busy_count = m->busy_count; 1570 cpu_ccfence(); 1571 if (busy_count & PBUSY_LOCKED) { 1572 tsleep_interlock(m, 0); 1573 if (atomic_cmpset_int(&m->busy_count, busy_count, 1574 busy_count | PBUSY_WANTED)) { 1575 atomic_set_int(&m->flags, PG_REFERENCED); 1576 tsleep(m, PINTERLOCKED, msg, 0); 1577 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1578 pindex); 1579 } 1580 } else if (also_m_busy && busy_count) { 1581 tsleep_interlock(m, 0); 1582 if (atomic_cmpset_int(&m->busy_count, busy_count, 1583 busy_count | PBUSY_WANTED)) { 1584 atomic_set_int(&m->flags, PG_REFERENCED); 1585 tsleep(m, PINTERLOCKED, msg, 0); 1586 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1587 pindex); 1588 } 1589 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1590 busy_count | PBUSY_LOCKED)) { 1591 #ifdef VM_PAGE_DEBUG 1592 m->busy_func = func; 1593 m->busy_line = lineno; 1594 #endif 1595 vm_page_hash_enter(m); 1596 break; 1597 } 1598 } 1599 return m; 1600 } 1601 1602 /* 1603 * Attempt to lookup and busy a page. 1604 * 1605 * Returns NULL if the page could not be found 1606 * 1607 * Returns a vm_page and error == TRUE if the page exists but could not 1608 * be busied. 1609 * 1610 * Returns a vm_page and error == FALSE on success. 1611 */ 1612 vm_page_t 1613 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1614 vm_pindex_t pindex, 1615 int also_m_busy, int *errorp 1616 VM_PAGE_DEBUG_ARGS) 1617 { 1618 u_int32_t busy_count; 1619 vm_page_t m; 1620 1621 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1622 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1623 *errorp = FALSE; 1624 while (m) { 1625 KKASSERT(m->object == object && m->pindex == pindex); 1626 busy_count = m->busy_count; 1627 cpu_ccfence(); 1628 if (busy_count & PBUSY_LOCKED) { 1629 *errorp = TRUE; 1630 break; 1631 } 1632 if (also_m_busy && busy_count) { 1633 *errorp = TRUE; 1634 break; 1635 } 1636 if (atomic_cmpset_int(&m->busy_count, busy_count, 1637 busy_count | PBUSY_LOCKED)) { 1638 #ifdef VM_PAGE_DEBUG 1639 m->busy_func = func; 1640 m->busy_line = lineno; 1641 #endif 1642 vm_page_hash_enter(m); 1643 break; 1644 } 1645 } 1646 return m; 1647 } 1648 1649 /* 1650 * Returns a page that is only soft-busied for use by the caller in 1651 * a read-only fashion. Returns NULL if the page could not be found, 1652 * the soft busy could not be obtained, or the page data is invalid. 1653 */ 1654 vm_page_t 1655 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1656 int pgoff, int pgbytes) 1657 { 1658 vm_page_t m; 1659 1660 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1661 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1662 if (m) { 1663 if ((m->valid != VM_PAGE_BITS_ALL && 1664 !vm_page_is_valid(m, pgoff, pgbytes)) || 1665 (m->flags & PG_FICTITIOUS)) { 1666 m = NULL; 1667 } else if (vm_page_sbusy_try(m)) { 1668 m = NULL; 1669 } else if ((m->valid != VM_PAGE_BITS_ALL && 1670 !vm_page_is_valid(m, pgoff, pgbytes)) || 1671 (m->flags & PG_FICTITIOUS)) { 1672 vm_page_sbusy_drop(m); 1673 m = NULL; 1674 } else { 1675 vm_page_hash_enter(m); 1676 } 1677 } 1678 return m; 1679 } 1680 1681 /* 1682 * Caller must hold the related vm_object 1683 */ 1684 vm_page_t 1685 vm_page_next(vm_page_t m) 1686 { 1687 vm_page_t next; 1688 1689 next = vm_page_rb_tree_RB_NEXT(m); 1690 if (next && next->pindex != m->pindex + 1) 1691 next = NULL; 1692 return (next); 1693 } 1694 1695 /* 1696 * vm_page_rename() 1697 * 1698 * Move the given vm_page from its current object to the specified 1699 * target object/offset. The page must be busy and will remain so 1700 * on return. 1701 * 1702 * new_object must be held. 1703 * This routine might block. XXX ? 1704 * 1705 * NOTE: Swap associated with the page must be invalidated by the move. We 1706 * have to do this for several reasons: (1) we aren't freeing the 1707 * page, (2) we are dirtying the page, (3) the VM system is probably 1708 * moving the page from object A to B, and will then later move 1709 * the backing store from A to B and we can't have a conflict. 1710 * 1711 * NOTE: We *always* dirty the page. It is necessary both for the 1712 * fact that we moved it, and because we may be invalidating 1713 * swap. If the page is on the cache, we have to deactivate it 1714 * or vm_page_dirty() will panic. Dirty pages are not allowed 1715 * on the cache. 1716 */ 1717 void 1718 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1719 { 1720 KKASSERT(m->busy_count & PBUSY_LOCKED); 1721 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1722 if (m->object) { 1723 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1724 vm_page_remove(m); 1725 } 1726 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1727 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1728 new_object, new_pindex); 1729 } 1730 if (m->queue - m->pc == PQ_CACHE) 1731 vm_page_deactivate(m); 1732 vm_page_dirty(m); 1733 } 1734 1735 /* 1736 * vm_page_unqueue() without any wakeup. This routine is used when a page 1737 * is to remain BUSYied by the caller. 1738 * 1739 * This routine may not block. 1740 */ 1741 void 1742 vm_page_unqueue_nowakeup(vm_page_t m) 1743 { 1744 vm_page_and_queue_spin_lock(m); 1745 (void)_vm_page_rem_queue_spinlocked(m); 1746 vm_page_spin_unlock(m); 1747 } 1748 1749 /* 1750 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1751 * if necessary. 1752 * 1753 * This routine may not block. 1754 */ 1755 void 1756 vm_page_unqueue(vm_page_t m) 1757 { 1758 u_short queue; 1759 1760 vm_page_and_queue_spin_lock(m); 1761 queue = _vm_page_rem_queue_spinlocked(m); 1762 if (queue == PQ_FREE || queue == PQ_CACHE) { 1763 vm_page_spin_unlock(m); 1764 pagedaemon_wakeup(); 1765 } else { 1766 vm_page_spin_unlock(m); 1767 } 1768 } 1769 1770 /* 1771 * vm_page_list_find() 1772 * 1773 * Find a page on the specified queue with color optimization. 1774 * 1775 * The page coloring optimization attempts to locate a page that does 1776 * not overload other nearby pages in the object in the cpu's L1 or L2 1777 * caches. We need this optimization because cpu caches tend to be 1778 * physical caches, while object spaces tend to be virtual. 1779 * 1780 * The page coloring optimization also, very importantly, tries to localize 1781 * memory to cpus and physical sockets. 1782 * 1783 * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the 1784 * algorithm is adjusted to localize allocations on a per-core basis. 1785 * This is done by 'twisting' the colors. 1786 * 1787 * The page is returned spinlocked and removed from its queue (it will 1788 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1789 * is responsible for dealing with the busy-page case (usually by 1790 * deactivating the page and looping). 1791 * 1792 * NOTE: This routine is carefully inlined. A non-inlined version 1793 * is available for outside callers but the only critical path is 1794 * from within this source file. 1795 * 1796 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1797 * represent stable storage, allowing us to order our locks vm_page 1798 * first, then queue. 1799 */ 1800 static __inline 1801 vm_page_t 1802 _vm_page_list_find(int basequeue, int index) 1803 { 1804 struct vpgqueues *pq; 1805 vm_page_t m; 1806 1807 index &= PQ_L2_MASK; 1808 pq = &vm_page_queues[basequeue + index]; 1809 1810 /* 1811 * Try this cpu's colored queue first. Test for a page unlocked, 1812 * then lock the queue and locate a page. Note that the lock order 1813 * is reversed, but we do not want to dwadle on the page spinlock 1814 * anyway as it is held significantly longer than the queue spinlock. 1815 */ 1816 if (TAILQ_FIRST(&pq->pl)) { 1817 spin_lock(&pq->spin); 1818 TAILQ_FOREACH(m, &pq->pl, pageq) { 1819 if (spin_trylock(&m->spin) == 0) 1820 continue; 1821 KKASSERT(m->queue == basequeue + index); 1822 _vm_page_rem_queue_spinlocked(m); 1823 return(m); 1824 } 1825 spin_unlock(&pq->spin); 1826 } 1827 1828 /* 1829 * If we are unable to get a page, do a more involved NUMA-aware 1830 * search. 1831 */ 1832 m = _vm_page_list_find2(basequeue, index); 1833 return(m); 1834 } 1835 1836 /* 1837 * If we could not find the page in the desired queue try to find it in 1838 * a nearby (NUMA-aware) queue. 1839 */ 1840 static vm_page_t 1841 _vm_page_list_find2(int basequeue, int index) 1842 { 1843 struct vpgqueues *pq; 1844 vm_page_t m = NULL; 1845 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1846 int pqi; 1847 int i; 1848 1849 index &= PQ_L2_MASK; 1850 pq = &vm_page_queues[basequeue]; 1851 1852 /* 1853 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1854 * else fails (PQ_L2_MASK which is 255). 1855 * 1856 * Test each queue unlocked first, then lock the queue and locate 1857 * a page. Note that the lock order is reversed, but we do not want 1858 * to dwadle on the page spinlock anyway as it is held significantly 1859 * longer than the queue spinlock. 1860 */ 1861 do { 1862 pqmask = (pqmask << 1) | 1; 1863 for (i = 0; i <= pqmask; ++i) { 1864 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1865 if (TAILQ_FIRST(&pq[pqi].pl)) { 1866 spin_lock(&pq[pqi].spin); 1867 TAILQ_FOREACH(m, &pq[pqi].pl, pageq) { 1868 if (spin_trylock(&m->spin) == 0) 1869 continue; 1870 KKASSERT(m->queue == basequeue + pqi); 1871 _vm_page_rem_queue_spinlocked(m); 1872 return(m); 1873 } 1874 spin_unlock(&pq[pqi].spin); 1875 } 1876 } 1877 } while (pqmask != PQ_L2_MASK); 1878 1879 return(m); 1880 } 1881 1882 /* 1883 * Returns a vm_page candidate for allocation. The page is not busied so 1884 * it can move around. The caller must busy the page (and typically 1885 * deactivate it if it cannot be busied!) 1886 * 1887 * Returns a spinlocked vm_page that has been removed from its queue. 1888 */ 1889 vm_page_t 1890 vm_page_list_find(int basequeue, int index) 1891 { 1892 return(_vm_page_list_find(basequeue, index)); 1893 } 1894 1895 /* 1896 * Find a page on the cache queue with color optimization, remove it 1897 * from the queue, and busy it. The returned page will not be spinlocked. 1898 * 1899 * A candidate failure will be deactivated. Candidates can fail due to 1900 * being busied by someone else, in which case they will be deactivated. 1901 * 1902 * This routine may not block. 1903 * 1904 */ 1905 static vm_page_t 1906 vm_page_select_cache(u_short pg_color) 1907 { 1908 vm_page_t m; 1909 1910 for (;;) { 1911 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1912 if (m == NULL) 1913 break; 1914 /* 1915 * (m) has been removed from its queue and spinlocked 1916 */ 1917 if (vm_page_busy_try(m, TRUE)) { 1918 _vm_page_deactivate_locked(m, 0); 1919 vm_page_spin_unlock(m); 1920 } else { 1921 /* 1922 * We successfully busied the page 1923 */ 1924 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1925 m->hold_count == 0 && 1926 m->wire_count == 0 && 1927 (m->dirty & m->valid) == 0) { 1928 vm_page_spin_unlock(m); 1929 pagedaemon_wakeup(); 1930 return(m); 1931 } 1932 1933 /* 1934 * The page cannot be recycled, deactivate it. 1935 */ 1936 _vm_page_deactivate_locked(m, 0); 1937 if (_vm_page_wakeup(m)) { 1938 vm_page_spin_unlock(m); 1939 wakeup(m); 1940 } else { 1941 vm_page_spin_unlock(m); 1942 } 1943 } 1944 } 1945 return (m); 1946 } 1947 1948 /* 1949 * Find a free page. We attempt to inline the nominal case and fall back 1950 * to _vm_page_select_free() otherwise. A busied page is removed from 1951 * the queue and returned. 1952 * 1953 * This routine may not block. 1954 */ 1955 static __inline vm_page_t 1956 vm_page_select_free(u_short pg_color) 1957 { 1958 vm_page_t m; 1959 1960 for (;;) { 1961 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1962 if (m == NULL) 1963 break; 1964 if (vm_page_busy_try(m, TRUE)) { 1965 /* 1966 * Various mechanisms such as a pmap_collect can 1967 * result in a busy page on the free queue. We 1968 * have to move the page out of the way so we can 1969 * retry the allocation. If the other thread is not 1970 * allocating the page then m->valid will remain 0 and 1971 * the pageout daemon will free the page later on. 1972 * 1973 * Since we could not busy the page, however, we 1974 * cannot make assumptions as to whether the page 1975 * will be allocated by the other thread or not, 1976 * so all we can do is deactivate it to move it out 1977 * of the way. In particular, if the other thread 1978 * wires the page it may wind up on the inactive 1979 * queue and the pageout daemon will have to deal 1980 * with that case too. 1981 */ 1982 _vm_page_deactivate_locked(m, 0); 1983 vm_page_spin_unlock(m); 1984 } else { 1985 /* 1986 * Theoretically if we are able to busy the page 1987 * atomic with the queue removal (using the vm_page 1988 * lock) nobody else should have been able to mess 1989 * with the page before us. 1990 * 1991 * Assert the page state. Note that even though 1992 * wiring doesn't adjust queues, a page on the free 1993 * queue should never be wired at this point. 1994 */ 1995 KKASSERT((m->flags & (PG_UNMANAGED | 1996 PG_NEED_COMMIT)) == 0); 1997 KASSERT(m->hold_count == 0, 1998 ("m->hold_count is not zero " 1999 "pg %p q=%d flags=%08x hold=%d wire=%d", 2000 m, m->queue, m->flags, 2001 m->hold_count, m->wire_count)); 2002 KKASSERT(m->wire_count == 0); 2003 vm_page_spin_unlock(m); 2004 pagedaemon_wakeup(); 2005 2006 /* return busied and removed page */ 2007 return(m); 2008 } 2009 } 2010 return(m); 2011 } 2012 2013 /* 2014 * vm_page_alloc() 2015 * 2016 * Allocate and return a memory cell associated with this VM object/offset 2017 * pair. If object is NULL an unassociated page will be allocated. 2018 * 2019 * The returned page will be busied and removed from its queues. This 2020 * routine can block and may return NULL if a race occurs and the page 2021 * is found to already exist at the specified (object, pindex). 2022 * 2023 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 2024 * VM_ALLOC_QUICK like normal but cannot use cache 2025 * VM_ALLOC_SYSTEM greater free drain 2026 * VM_ALLOC_INTERRUPT allow free list to be completely drained 2027 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 2028 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 2029 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 2030 * (see vm_page_grab()) 2031 * VM_ALLOC_USE_GD ok to use per-gd cache 2032 * 2033 * VM_ALLOC_CPU(n) allocate using specified cpu localization 2034 * 2035 * The object must be held if not NULL 2036 * This routine may not block 2037 * 2038 * Additional special handling is required when called from an interrupt 2039 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 2040 * in this case. 2041 */ 2042 vm_page_t 2043 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 2044 { 2045 globaldata_t gd; 2046 vm_object_t obj; 2047 vm_page_t m; 2048 u_short pg_color; 2049 int cpuid_local; 2050 2051 #if 0 2052 /* 2053 * Special per-cpu free VM page cache. The pages are pre-busied 2054 * and pre-zerod for us. 2055 */ 2056 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 2057 crit_enter_gd(gd); 2058 if (gd->gd_vmpg_count) { 2059 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 2060 crit_exit_gd(gd); 2061 goto done; 2062 } 2063 crit_exit_gd(gd); 2064 } 2065 #endif 2066 m = NULL; 2067 2068 /* 2069 * CPU LOCALIZATION 2070 * 2071 * CPU localization algorithm. Break the page queues up by physical 2072 * id and core id (note that two cpu threads will have the same core 2073 * id, and core_id != gd_cpuid). 2074 * 2075 * This is nowhere near perfect, for example the last pindex in a 2076 * subgroup will overflow into the next cpu or package. But this 2077 * should get us good page reuse locality in heavy mixed loads. 2078 * 2079 * (may be executed before the APs are started, so other GDs might 2080 * not exist!) 2081 */ 2082 if (page_req & VM_ALLOC_CPU_SPEC) 2083 cpuid_local = VM_ALLOC_GETCPU(page_req); 2084 else 2085 cpuid_local = mycpu->gd_cpuid; 2086 2087 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 2088 2089 KKASSERT(page_req & 2090 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 2091 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2092 2093 /* 2094 * Certain system threads (pageout daemon, buf_daemon's) are 2095 * allowed to eat deeper into the free page list. 2096 */ 2097 if (curthread->td_flags & TDF_SYSTHREAD) 2098 page_req |= VM_ALLOC_SYSTEM; 2099 2100 /* 2101 * Impose various limitations. Note that the v_free_reserved test 2102 * must match the opposite of vm_page_count_target() to avoid 2103 * livelocks, be careful. 2104 */ 2105 loop: 2106 gd = mycpu; 2107 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 2108 ((page_req & VM_ALLOC_INTERRUPT) && 2109 gd->gd_vmstats.v_free_count > 0) || 2110 ((page_req & VM_ALLOC_SYSTEM) && 2111 gd->gd_vmstats.v_cache_count == 0 && 2112 gd->gd_vmstats.v_free_count > 2113 gd->gd_vmstats.v_interrupt_free_min) 2114 ) { 2115 /* 2116 * The free queue has sufficient free pages to take one out. 2117 */ 2118 m = vm_page_select_free(pg_color); 2119 } else if (page_req & VM_ALLOC_NORMAL) { 2120 /* 2121 * Allocatable from the cache (non-interrupt only). On 2122 * success, we must free the page and try again, thus 2123 * ensuring that vmstats.v_*_free_min counters are replenished. 2124 */ 2125 #ifdef INVARIANTS 2126 if (curthread->td_preempted) { 2127 kprintf("vm_page_alloc(): warning, attempt to allocate" 2128 " cache page from preempting interrupt\n"); 2129 m = NULL; 2130 } else { 2131 m = vm_page_select_cache(pg_color); 2132 } 2133 #else 2134 m = vm_page_select_cache(pg_color); 2135 #endif 2136 /* 2137 * On success move the page into the free queue and loop. 2138 * 2139 * Only do this if we can safely acquire the vm_object lock, 2140 * because this is effectively a random page and the caller 2141 * might be holding the lock shared, we don't want to 2142 * deadlock. 2143 */ 2144 if (m != NULL) { 2145 KASSERT(m->dirty == 0, 2146 ("Found dirty cache page %p", m)); 2147 if ((obj = m->object) != NULL) { 2148 if (vm_object_hold_try(obj)) { 2149 vm_page_protect(m, VM_PROT_NONE); 2150 vm_page_free(m); 2151 /* m->object NULL here */ 2152 vm_object_drop(obj); 2153 } else { 2154 vm_page_deactivate(m); 2155 vm_page_wakeup(m); 2156 } 2157 } else { 2158 vm_page_protect(m, VM_PROT_NONE); 2159 vm_page_free(m); 2160 } 2161 goto loop; 2162 } 2163 2164 /* 2165 * On failure return NULL 2166 */ 2167 atomic_add_int(&vm_pageout_deficit, 1); 2168 pagedaemon_wakeup(); 2169 return (NULL); 2170 } else { 2171 /* 2172 * No pages available, wakeup the pageout daemon and give up. 2173 */ 2174 atomic_add_int(&vm_pageout_deficit, 1); 2175 pagedaemon_wakeup(); 2176 return (NULL); 2177 } 2178 2179 /* 2180 * v_free_count can race so loop if we don't find the expected 2181 * page. 2182 */ 2183 if (m == NULL) { 2184 vmstats_rollup(); 2185 goto loop; 2186 } 2187 2188 /* 2189 * Good page found. The page has already been busied for us and 2190 * removed from its queues. 2191 */ 2192 KASSERT(m->dirty == 0, 2193 ("vm_page_alloc: free/cache page %p was dirty", m)); 2194 KKASSERT(m->queue == PQ_NONE); 2195 2196 #if 0 2197 done: 2198 #endif 2199 /* 2200 * Initialize the structure, inheriting some flags but clearing 2201 * all the rest. The page has already been busied for us. 2202 */ 2203 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 2204 2205 KKASSERT(m->wire_count == 0); 2206 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 2207 m->act_count = 0; 2208 m->valid = 0; 2209 2210 /* 2211 * Caller must be holding the object lock (asserted by 2212 * vm_page_insert()). 2213 * 2214 * NOTE: Inserting a page here does not insert it into any pmaps 2215 * (which could cause us to block allocating memory). 2216 * 2217 * NOTE: If no object an unassociated page is allocated, m->pindex 2218 * can be used by the caller for any purpose. 2219 */ 2220 if (object) { 2221 if (vm_page_insert(m, object, pindex) == FALSE) { 2222 vm_page_free(m); 2223 if ((page_req & VM_ALLOC_NULL_OK) == 0) 2224 panic("PAGE RACE %p[%ld]/%p", 2225 object, (long)pindex, m); 2226 m = NULL; 2227 } 2228 } else { 2229 m->pindex = pindex; 2230 } 2231 2232 /* 2233 * Don't wakeup too often - wakeup the pageout daemon when 2234 * we would be nearly out of memory. 2235 */ 2236 pagedaemon_wakeup(); 2237 2238 /* 2239 * A BUSY page is returned. 2240 */ 2241 return (m); 2242 } 2243 2244 /* 2245 * Returns number of pages available in our DMA memory reserve 2246 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 2247 */ 2248 vm_size_t 2249 vm_contig_avail_pages(void) 2250 { 2251 alist_blk_t blk; 2252 alist_blk_t count; 2253 alist_blk_t bfree; 2254 spin_lock(&vm_contig_spin); 2255 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 2256 spin_unlock(&vm_contig_spin); 2257 2258 return bfree; 2259 } 2260 2261 /* 2262 * Attempt to allocate contiguous physical memory with the specified 2263 * requirements. 2264 */ 2265 vm_page_t 2266 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 2267 unsigned long alignment, unsigned long boundary, 2268 unsigned long size, vm_memattr_t memattr) 2269 { 2270 alist_blk_t blk; 2271 vm_page_t m; 2272 vm_pindex_t i; 2273 #if 0 2274 static vm_pindex_t contig_rover; 2275 #endif 2276 2277 alignment >>= PAGE_SHIFT; 2278 if (alignment == 0) 2279 alignment = 1; 2280 boundary >>= PAGE_SHIFT; 2281 if (boundary == 0) 2282 boundary = 1; 2283 size = (size + PAGE_MASK) >> PAGE_SHIFT; 2284 2285 #if 0 2286 /* 2287 * Disabled temporarily until we find a solution for DRM (a flag 2288 * to always use the free space reserve, for performance). 2289 */ 2290 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 2291 boundary <= PAGE_SIZE && size == 1 && 2292 memattr == VM_MEMATTR_DEFAULT) { 2293 /* 2294 * Any page will work, use vm_page_alloc() 2295 * (e.g. when used from kmem_alloc_attr()) 2296 */ 2297 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 2298 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2299 VM_ALLOC_INTERRUPT); 2300 m->valid = VM_PAGE_BITS_ALL; 2301 vm_page_wire(m); 2302 vm_page_wakeup(m); 2303 } else 2304 #endif 2305 { 2306 /* 2307 * Use the low-memory dma reserve 2308 */ 2309 spin_lock(&vm_contig_spin); 2310 blk = alist_alloc(&vm_contig_alist, 0, size); 2311 if (blk == ALIST_BLOCK_NONE) { 2312 spin_unlock(&vm_contig_spin); 2313 if (bootverbose) { 2314 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2315 (size << PAGE_SHIFT) / 1024); 2316 print_backtrace(5); 2317 } 2318 return(NULL); 2319 } 2320 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2321 alist_free(&vm_contig_alist, blk, size); 2322 spin_unlock(&vm_contig_spin); 2323 if (bootverbose) { 2324 kprintf("vm_page_alloc_contig: %ldk high " 2325 "%016jx failed\n", 2326 (size << PAGE_SHIFT) / 1024, 2327 (intmax_t)high); 2328 } 2329 return(NULL); 2330 } 2331 spin_unlock(&vm_contig_spin); 2332 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2333 } 2334 if (vm_contig_verbose) { 2335 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2336 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2337 (intmax_t)m->phys_addr, 2338 (size << PAGE_SHIFT) / 1024, 2339 low, high, alignment, boundary, size, memattr); 2340 } 2341 if (memattr != VM_MEMATTR_DEFAULT) { 2342 for (i = 0;i < size; i++) 2343 pmap_page_set_memattr(&m[i], memattr); 2344 } 2345 return m; 2346 } 2347 2348 /* 2349 * Free contiguously allocated pages. The pages will be wired but not busy. 2350 * When freeing to the alist we leave them wired and not busy. 2351 */ 2352 void 2353 vm_page_free_contig(vm_page_t m, unsigned long size) 2354 { 2355 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2356 vm_pindex_t start = pa >> PAGE_SHIFT; 2357 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2358 2359 if (vm_contig_verbose) { 2360 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2361 (intmax_t)pa, size / 1024); 2362 } 2363 if (pa < vm_low_phys_reserved) { 2364 KKASSERT(pa + size <= vm_low_phys_reserved); 2365 spin_lock(&vm_contig_spin); 2366 alist_free(&vm_contig_alist, start, pages); 2367 spin_unlock(&vm_contig_spin); 2368 } else { 2369 while (pages) { 2370 vm_page_busy_wait(m, FALSE, "cpgfr"); 2371 vm_page_unwire(m, 0); 2372 vm_page_free(m); 2373 --pages; 2374 ++m; 2375 } 2376 2377 } 2378 } 2379 2380 2381 /* 2382 * Wait for sufficient free memory for nominal heavy memory use kernel 2383 * operations. 2384 * 2385 * WARNING! Be sure never to call this in any vm_pageout code path, which 2386 * will trivially deadlock the system. 2387 */ 2388 void 2389 vm_wait_nominal(void) 2390 { 2391 while (vm_page_count_min(0)) 2392 vm_wait(0); 2393 } 2394 2395 /* 2396 * Test if vm_wait_nominal() would block. 2397 */ 2398 int 2399 vm_test_nominal(void) 2400 { 2401 if (vm_page_count_min(0)) 2402 return(1); 2403 return(0); 2404 } 2405 2406 /* 2407 * Block until free pages are available for allocation, called in various 2408 * places before memory allocations. 2409 * 2410 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2411 * more generous then that. 2412 */ 2413 void 2414 vm_wait(int timo) 2415 { 2416 /* 2417 * never wait forever 2418 */ 2419 if (timo == 0) 2420 timo = hz; 2421 lwkt_gettoken(&vm_token); 2422 2423 if (curthread == pagethread || 2424 curthread == emergpager) { 2425 /* 2426 * The pageout daemon itself needs pages, this is bad. 2427 */ 2428 if (vm_page_count_min(0)) { 2429 vm_pageout_pages_needed = 1; 2430 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2431 } 2432 } else { 2433 /* 2434 * Wakeup the pageout daemon if necessary and wait. 2435 * 2436 * Do not wait indefinitely for the target to be reached, 2437 * as load might prevent it from being reached any time soon. 2438 * But wait a little to try to slow down page allocations 2439 * and to give more important threads (the pagedaemon) 2440 * allocation priority. 2441 */ 2442 if (vm_page_count_target()) { 2443 if (vm_pages_needed == 0) { 2444 vm_pages_needed = 1; 2445 wakeup(&vm_pages_needed); 2446 } 2447 ++vm_pages_waiting; /* SMP race ok */ 2448 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2449 } 2450 } 2451 lwkt_reltoken(&vm_token); 2452 } 2453 2454 /* 2455 * Block until free pages are available for allocation 2456 * 2457 * Called only from vm_fault so that processes page faulting can be 2458 * easily tracked. 2459 */ 2460 void 2461 vm_wait_pfault(void) 2462 { 2463 /* 2464 * Wakeup the pageout daemon if necessary and wait. 2465 * 2466 * Do not wait indefinitely for the target to be reached, 2467 * as load might prevent it from being reached any time soon. 2468 * But wait a little to try to slow down page allocations 2469 * and to give more important threads (the pagedaemon) 2470 * allocation priority. 2471 */ 2472 if (vm_page_count_min(0)) { 2473 lwkt_gettoken(&vm_token); 2474 while (vm_page_count_severe()) { 2475 if (vm_page_count_target()) { 2476 thread_t td; 2477 2478 if (vm_pages_needed == 0) { 2479 vm_pages_needed = 1; 2480 wakeup(&vm_pages_needed); 2481 } 2482 ++vm_pages_waiting; /* SMP race ok */ 2483 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2484 2485 /* 2486 * Do not stay stuck in the loop if the system is trying 2487 * to kill the process. 2488 */ 2489 td = curthread; 2490 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2491 break; 2492 } 2493 } 2494 lwkt_reltoken(&vm_token); 2495 } 2496 } 2497 2498 /* 2499 * Put the specified page on the active list (if appropriate). Ensure 2500 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2501 * 2502 * The caller should be holding the page busied ? XXX 2503 * This routine may not block. 2504 */ 2505 void 2506 vm_page_activate(vm_page_t m) 2507 { 2508 u_short oqueue; 2509 2510 vm_page_spin_lock(m); 2511 if (m->queue - m->pc != PQ_ACTIVE && !(m->flags & PG_FICTITIOUS)) { 2512 _vm_page_queue_spin_lock(m); 2513 oqueue = _vm_page_rem_queue_spinlocked(m); 2514 /* page is left spinlocked, queue is unlocked */ 2515 2516 if (oqueue == PQ_CACHE) 2517 mycpu->gd_cnt.v_reactivated++; 2518 if ((m->flags & PG_UNMANAGED) == 0) { 2519 if (m->act_count < ACT_INIT) 2520 m->act_count = ACT_INIT; 2521 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2522 } 2523 _vm_page_and_queue_spin_unlock(m); 2524 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2525 pagedaemon_wakeup(); 2526 } else { 2527 if (m->act_count < ACT_INIT) 2528 m->act_count = ACT_INIT; 2529 vm_page_spin_unlock(m); 2530 } 2531 } 2532 2533 /* 2534 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2535 * routine is called when a page has been added to the cache or free 2536 * queues. 2537 * 2538 * This routine may not block. 2539 */ 2540 static __inline void 2541 vm_page_free_wakeup(void) 2542 { 2543 globaldata_t gd = mycpu; 2544 2545 /* 2546 * If the pageout daemon itself needs pages, then tell it that 2547 * there are some free. 2548 */ 2549 if (vm_pageout_pages_needed && 2550 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2551 gd->gd_vmstats.v_pageout_free_min 2552 ) { 2553 vm_pageout_pages_needed = 0; 2554 wakeup(&vm_pageout_pages_needed); 2555 } 2556 2557 /* 2558 * Wakeup processes that are waiting on memory. 2559 * 2560 * Generally speaking we want to wakeup stuck processes as soon as 2561 * possible. !vm_page_count_min(0) is the absolute minimum point 2562 * where we can do this. Wait a bit longer to reduce degenerate 2563 * re-blocking (vm_page_free_hysteresis). The target check is just 2564 * to make sure the min-check w/hysteresis does not exceed the 2565 * normal target. 2566 */ 2567 if (vm_pages_waiting) { 2568 if (!vm_page_count_min(vm_page_free_hysteresis) || 2569 !vm_page_count_target()) { 2570 vm_pages_waiting = 0; 2571 wakeup(&vmstats.v_free_count); 2572 ++mycpu->gd_cnt.v_ppwakeups; 2573 } 2574 #if 0 2575 if (!vm_page_count_target()) { 2576 /* 2577 * Plenty of pages are free, wakeup everyone. 2578 */ 2579 vm_pages_waiting = 0; 2580 wakeup(&vmstats.v_free_count); 2581 ++mycpu->gd_cnt.v_ppwakeups; 2582 } else if (!vm_page_count_min(0)) { 2583 /* 2584 * Some pages are free, wakeup someone. 2585 */ 2586 int wcount = vm_pages_waiting; 2587 if (wcount > 0) 2588 --wcount; 2589 vm_pages_waiting = wcount; 2590 wakeup_one(&vmstats.v_free_count); 2591 ++mycpu->gd_cnt.v_ppwakeups; 2592 } 2593 #endif 2594 } 2595 } 2596 2597 /* 2598 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2599 * it from its VM object. 2600 * 2601 * The vm_page must be BUSY on entry. BUSY will be released on 2602 * return (the page will have been freed). 2603 */ 2604 void 2605 vm_page_free_toq(vm_page_t m) 2606 { 2607 mycpu->gd_cnt.v_tfree++; 2608 KKASSERT((m->flags & PG_MAPPED) == 0); 2609 KKASSERT(m->busy_count & PBUSY_LOCKED); 2610 2611 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2612 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2613 "hold(%d)\n", 2614 (u_long)m->pindex, m->busy_count, m->hold_count); 2615 if ((m->queue - m->pc) == PQ_FREE) 2616 panic("vm_page_free: freeing free page"); 2617 else 2618 panic("vm_page_free: freeing busy page"); 2619 } 2620 2621 /* 2622 * Remove from object, spinlock the page and its queues and 2623 * remove from any queue. No queue spinlock will be held 2624 * after this section (because the page was removed from any 2625 * queue). 2626 */ 2627 vm_page_remove(m); 2628 2629 /* 2630 * No further management of fictitious pages occurs beyond object 2631 * and queue removal. 2632 */ 2633 if ((m->flags & PG_FICTITIOUS) != 0) { 2634 KKASSERT(m->queue == PQ_NONE); 2635 vm_page_wakeup(m); 2636 return; 2637 } 2638 vm_page_and_queue_spin_lock(m); 2639 _vm_page_rem_queue_spinlocked(m); 2640 2641 m->valid = 0; 2642 vm_page_undirty(m); 2643 2644 if (m->wire_count != 0) { 2645 if (m->wire_count > 1) { 2646 panic( 2647 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2648 m->wire_count, (long)m->pindex); 2649 } 2650 panic("vm_page_free: freeing wired page"); 2651 } 2652 2653 /* 2654 * Clear the UNMANAGED flag when freeing an unmanaged page. 2655 * Clear the NEED_COMMIT flag 2656 */ 2657 if (m->flags & PG_UNMANAGED) 2658 vm_page_flag_clear(m, PG_UNMANAGED); 2659 if (m->flags & PG_NEED_COMMIT) 2660 vm_page_flag_clear(m, PG_NEED_COMMIT); 2661 2662 if (m->hold_count != 0) { 2663 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2664 } else { 2665 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2666 } 2667 2668 /* 2669 * This sequence allows us to clear BUSY while still holding 2670 * its spin lock, which reduces contention vs allocators. We 2671 * must not leave the queue locked or _vm_page_wakeup() may 2672 * deadlock. 2673 */ 2674 _vm_page_queue_spin_unlock(m); 2675 if (_vm_page_wakeup(m)) { 2676 vm_page_spin_unlock(m); 2677 wakeup(m); 2678 } else { 2679 vm_page_spin_unlock(m); 2680 } 2681 vm_page_free_wakeup(); 2682 } 2683 2684 /* 2685 * vm_page_unmanage() 2686 * 2687 * Prevent PV management from being done on the page. The page is 2688 * also removed from the paging queues, and as a consequence of no longer 2689 * being managed the pageout daemon will not touch it (since there is no 2690 * way to locate the pte mappings for the page). madvise() calls that 2691 * mess with the pmap will also no longer operate on the page. 2692 * 2693 * Beyond that the page is still reasonably 'normal'. Freeing the page 2694 * will clear the flag. 2695 * 2696 * This routine is used by OBJT_PHYS objects - objects using unswappable 2697 * physical memory as backing store rather then swap-backed memory and 2698 * will eventually be extended to support 4MB unmanaged physical 2699 * mappings. 2700 * 2701 * Caller must be holding the page busy. 2702 */ 2703 void 2704 vm_page_unmanage(vm_page_t m) 2705 { 2706 KKASSERT(m->busy_count & PBUSY_LOCKED); 2707 if ((m->flags & PG_UNMANAGED) == 0) { 2708 vm_page_unqueue(m); 2709 } 2710 vm_page_flag_set(m, PG_UNMANAGED); 2711 } 2712 2713 /* 2714 * Mark this page as wired down by yet another map. We do not adjust the 2715 * queue the page is on, it will be checked for wiring as-needed. 2716 * 2717 * Caller must be holding the page busy. 2718 */ 2719 void 2720 vm_page_wire(vm_page_t m) 2721 { 2722 /* 2723 * Only bump the wire statistics if the page is not already wired, 2724 * and only unqueue the page if it is on some queue (if it is unmanaged 2725 * it is already off the queues). Don't do anything with fictitious 2726 * pages because they are always wired. 2727 */ 2728 KKASSERT(m->busy_count & PBUSY_LOCKED); 2729 if ((m->flags & PG_FICTITIOUS) == 0) { 2730 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2731 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2732 } 2733 KASSERT(m->wire_count != 0, 2734 ("vm_page_wire: wire_count overflow m=%p", m)); 2735 } 2736 } 2737 2738 /* 2739 * Release one wiring of this page, potentially enabling it to be paged again. 2740 * 2741 * Note that wired pages are no longer unconditionally removed from the 2742 * paging queues, so the page may already be on a queue. Move the page 2743 * to the desired queue if necessary. 2744 * 2745 * Many pages placed on the inactive queue should actually go 2746 * into the cache, but it is difficult to figure out which. What 2747 * we do instead, if the inactive target is well met, is to put 2748 * clean pages at the head of the inactive queue instead of the tail. 2749 * This will cause them to be moved to the cache more quickly and 2750 * if not actively re-referenced, freed more quickly. If we just 2751 * stick these pages at the end of the inactive queue, heavy filesystem 2752 * meta-data accesses can cause an unnecessary paging load on memory bound 2753 * processes. This optimization causes one-time-use metadata to be 2754 * reused more quickly. 2755 * 2756 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2757 * the inactive queue. This helps the pageout daemon determine memory 2758 * pressure and act on out-of-memory situations more quickly. 2759 * 2760 * BUT, if we are in a low-memory situation we have no choice but to 2761 * put clean pages on the cache queue. 2762 * 2763 * A number of routines use vm_page_unwire() to guarantee that the page 2764 * will go into either the inactive or active queues, and will NEVER 2765 * be placed in the cache - for example, just after dirtying a page. 2766 * dirty pages in the cache are not allowed. 2767 * 2768 * This routine may not block. 2769 */ 2770 void 2771 vm_page_unwire(vm_page_t m, int activate) 2772 { 2773 KKASSERT(m->busy_count & PBUSY_LOCKED); 2774 if (m->flags & PG_FICTITIOUS) { 2775 /* do nothing */ 2776 } else if ((int)m->wire_count <= 0) { 2777 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2778 } else { 2779 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2780 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2781 if (m->flags & PG_UNMANAGED) { 2782 ; 2783 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2784 vm_page_activate(m); 2785 #if 0 2786 vm_page_spin_lock(m); 2787 _vm_page_add_queue_spinlocked(m, 2788 PQ_ACTIVE + m->pc, 0); 2789 _vm_page_and_queue_spin_unlock(m); 2790 #endif 2791 } else { 2792 vm_page_deactivate(m); 2793 #if 0 2794 vm_page_spin_lock(m); 2795 vm_page_flag_clear(m, PG_WINATCFLS); 2796 _vm_page_add_queue_spinlocked(m, 2797 PQ_INACTIVE + m->pc, 0); 2798 _vm_page_and_queue_spin_unlock(m); 2799 #endif 2800 } 2801 } 2802 } 2803 } 2804 2805 /* 2806 * Move the specified page to the inactive queue. 2807 * 2808 * Normally athead is 0 resulting in LRU operation. athead is set 2809 * to 1 if we want this page to be 'as if it were placed in the cache', 2810 * except without unmapping it from the process address space. 2811 * 2812 * vm_page's spinlock must be held on entry and will remain held on return. 2813 * This routine may not block. The caller does not have to hold the page 2814 * busied but should have some sort of interlock on its validity. 2815 */ 2816 static void 2817 _vm_page_deactivate_locked(vm_page_t m, int athead) 2818 { 2819 u_short oqueue; 2820 2821 /* 2822 * Ignore if already inactive. 2823 */ 2824 if (m->queue - m->pc == PQ_INACTIVE || (m->flags & PG_FICTITIOUS)) 2825 return; 2826 _vm_page_queue_spin_lock(m); 2827 oqueue = _vm_page_rem_queue_spinlocked(m); 2828 2829 if ((m->flags & PG_UNMANAGED) == 0) { 2830 if (oqueue == PQ_CACHE) 2831 mycpu->gd_cnt.v_reactivated++; 2832 vm_page_flag_clear(m, PG_WINATCFLS); 2833 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2834 if (athead == 0) { 2835 atomic_add_long( 2836 &vm_page_queues[PQ_INACTIVE + m->pc].adds, 1); 2837 } 2838 } 2839 /* NOTE: PQ_NONE if condition not taken */ 2840 _vm_page_queue_spin_unlock(m); 2841 /* leaves vm_page spinlocked */ 2842 } 2843 2844 /* 2845 * Attempt to deactivate a page. 2846 * 2847 * No requirements. 2848 */ 2849 void 2850 vm_page_deactivate(vm_page_t m) 2851 { 2852 vm_page_spin_lock(m); 2853 _vm_page_deactivate_locked(m, 0); 2854 vm_page_spin_unlock(m); 2855 } 2856 2857 void 2858 vm_page_deactivate_locked(vm_page_t m) 2859 { 2860 _vm_page_deactivate_locked(m, 0); 2861 } 2862 2863 /* 2864 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2865 * 2866 * This function returns non-zero if it successfully moved the page to 2867 * PQ_CACHE. 2868 * 2869 * This function unconditionally unbusies the page on return. 2870 */ 2871 int 2872 vm_page_try_to_cache(vm_page_t m) 2873 { 2874 /* 2875 * Shortcut if we obviously cannot move the page, or if the 2876 * page is already on the cache queue, or it is ficitious. 2877 */ 2878 if (m->dirty || m->hold_count || m->wire_count || 2879 m->queue - m->pc == PQ_CACHE || 2880 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT | PG_FICTITIOUS))) { 2881 vm_page_wakeup(m); 2882 return(0); 2883 } 2884 2885 /* 2886 * Page busied by us and no longer spinlocked. Dirty pages cannot 2887 * be moved to the cache. 2888 */ 2889 vm_page_test_dirty(m); 2890 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2891 vm_page_wakeup(m); 2892 return(0); 2893 } 2894 vm_page_cache(m); 2895 return(1); 2896 } 2897 2898 /* 2899 * Attempt to free the page. If we cannot free it, we do nothing. 2900 * 1 is returned on success, 0 on failure. 2901 * 2902 * Caller provides an unlocked/non-busied page. 2903 * No requirements. 2904 */ 2905 int 2906 vm_page_try_to_free(vm_page_t m) 2907 { 2908 if (vm_page_busy_try(m, TRUE)) 2909 return(0); 2910 2911 /* 2912 * The page can be in any state, including already being on the free 2913 * queue. Check to see if it really can be freed. 2914 */ 2915 if (m->dirty || /* can't free if it is dirty */ 2916 m->hold_count || /* or held (XXX may be wrong) */ 2917 m->wire_count || /* or wired */ 2918 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2919 PG_NEED_COMMIT | /* or needs a commit */ 2920 PG_FICTITIOUS)) || /* or is fictitious */ 2921 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2922 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2923 vm_page_wakeup(m); 2924 return(0); 2925 } 2926 2927 /* 2928 * We can probably free the page. 2929 * 2930 * Page busied by us and no longer spinlocked. Dirty pages will 2931 * not be freed by this function. We have to re-test the 2932 * dirty bit after cleaning out the pmaps. 2933 */ 2934 vm_page_test_dirty(m); 2935 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2936 vm_page_wakeup(m); 2937 return(0); 2938 } 2939 vm_page_protect(m, VM_PROT_NONE); 2940 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2941 vm_page_wakeup(m); 2942 return(0); 2943 } 2944 vm_page_free(m); 2945 return(1); 2946 } 2947 2948 /* 2949 * vm_page_cache 2950 * 2951 * Put the specified page onto the page cache queue (if appropriate). 2952 * 2953 * The page must be busy, and this routine will release the busy and 2954 * possibly even free the page. 2955 */ 2956 void 2957 vm_page_cache(vm_page_t m) 2958 { 2959 /* 2960 * Not suitable for the cache 2961 */ 2962 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT | PG_FICTITIOUS)) || 2963 (m->busy_count & PBUSY_MASK) || 2964 m->wire_count || m->hold_count) { 2965 vm_page_wakeup(m); 2966 return; 2967 } 2968 2969 /* 2970 * Already in the cache (and thus not mapped) 2971 */ 2972 if ((m->queue - m->pc) == PQ_CACHE) { 2973 KKASSERT((m->flags & PG_MAPPED) == 0); 2974 vm_page_wakeup(m); 2975 return; 2976 } 2977 2978 /* 2979 * Caller is required to test m->dirty, but note that the act of 2980 * removing the page from its maps can cause it to become dirty 2981 * on an SMP system due to another cpu running in usermode. 2982 */ 2983 if (m->dirty) { 2984 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2985 (long)m->pindex); 2986 } 2987 2988 /* 2989 * Remove all pmaps and indicate that the page is not 2990 * writeable or mapped. Our vm_page_protect() call may 2991 * have blocked (especially w/ VM_PROT_NONE), so recheck 2992 * everything. 2993 */ 2994 vm_page_protect(m, VM_PROT_NONE); 2995 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2996 (m->busy_count & PBUSY_MASK) || 2997 m->wire_count || m->hold_count) { 2998 vm_page_wakeup(m); 2999 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 3000 vm_page_deactivate(m); 3001 vm_page_wakeup(m); 3002 } else { 3003 _vm_page_and_queue_spin_lock(m); 3004 _vm_page_rem_queue_spinlocked(m); 3005 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 3006 _vm_page_and_queue_spin_unlock(m); 3007 vm_page_wakeup(m); 3008 vm_page_free_wakeup(); 3009 } 3010 } 3011 3012 /* 3013 * vm_page_dontneed() 3014 * 3015 * Cache, deactivate, or do nothing as appropriate. This routine 3016 * is typically used by madvise() MADV_DONTNEED. 3017 * 3018 * Generally speaking we want to move the page into the cache so 3019 * it gets reused quickly. However, this can result in a silly syndrome 3020 * due to the page recycling too quickly. Small objects will not be 3021 * fully cached. On the otherhand, if we move the page to the inactive 3022 * queue we wind up with a problem whereby very large objects 3023 * unnecessarily blow away our inactive and cache queues. 3024 * 3025 * The solution is to move the pages based on a fixed weighting. We 3026 * either leave them alone, deactivate them, or move them to the cache, 3027 * where moving them to the cache has the highest weighting. 3028 * By forcing some pages into other queues we eventually force the 3029 * system to balance the queues, potentially recovering other unrelated 3030 * space from active. The idea is to not force this to happen too 3031 * often. 3032 * 3033 * The page must be busied. 3034 */ 3035 void 3036 vm_page_dontneed(vm_page_t m) 3037 { 3038 static int dnweight; 3039 int dnw; 3040 int head; 3041 3042 dnw = ++dnweight; 3043 3044 /* 3045 * occassionally leave the page alone 3046 */ 3047 if ((dnw & 0x01F0) == 0 || 3048 m->queue - m->pc == PQ_INACTIVE || 3049 m->queue - m->pc == PQ_CACHE 3050 ) { 3051 if (m->act_count >= ACT_INIT) 3052 --m->act_count; 3053 return; 3054 } 3055 3056 /* 3057 * If vm_page_dontneed() is inactivating a page, it must clear 3058 * the referenced flag; otherwise the pagedaemon will see references 3059 * on the page in the inactive queue and reactivate it. Until the 3060 * page can move to the cache queue, madvise's job is not done. 3061 */ 3062 vm_page_flag_clear(m, PG_REFERENCED); 3063 pmap_clear_reference(m); 3064 3065 if (m->dirty == 0) 3066 vm_page_test_dirty(m); 3067 3068 if (m->dirty || (dnw & 0x0070) == 0) { 3069 /* 3070 * Deactivate the page 3 times out of 32. 3071 */ 3072 head = 0; 3073 } else { 3074 /* 3075 * Cache the page 28 times out of every 32. Note that 3076 * the page is deactivated instead of cached, but placed 3077 * at the head of the queue instead of the tail. 3078 */ 3079 head = 1; 3080 } 3081 vm_page_spin_lock(m); 3082 _vm_page_deactivate_locked(m, head); 3083 vm_page_spin_unlock(m); 3084 } 3085 3086 /* 3087 * These routines manipulate the 'soft busy' count for a page. A soft busy 3088 * is almost like a hard BUSY except that it allows certain compatible 3089 * operations to occur on the page while it is busy. For example, a page 3090 * undergoing a write can still be mapped read-only. 3091 * 3092 * We also use soft-busy to quickly pmap_enter shared read-only pages 3093 * without having to hold the page locked. 3094 * 3095 * The soft-busy count can be > 1 in situations where multiple threads 3096 * are pmap_enter()ing the same page simultaneously, or when two buffer 3097 * cache buffers overlap the same page. 3098 * 3099 * The caller must hold the page BUSY when making these two calls. 3100 */ 3101 void 3102 vm_page_io_start(vm_page_t m) 3103 { 3104 uint32_t ocount; 3105 3106 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3107 KKASSERT(ocount & PBUSY_LOCKED); 3108 } 3109 3110 void 3111 vm_page_io_finish(vm_page_t m) 3112 { 3113 uint32_t ocount; 3114 3115 ocount = atomic_fetchadd_int(&m->busy_count, -1); 3116 KKASSERT(ocount & PBUSY_MASK); 3117 #if 0 3118 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 3119 wakeup(m); 3120 #endif 3121 } 3122 3123 /* 3124 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 3125 * 3126 * We can't use fetchadd here because we might race a hard-busy and the 3127 * page freeing code asserts on a non-zero soft-busy count (even if only 3128 * temporary). 3129 * 3130 * Returns 0 on success, non-zero on failure. 3131 */ 3132 int 3133 vm_page_sbusy_try(vm_page_t m) 3134 { 3135 uint32_t ocount; 3136 3137 for (;;) { 3138 ocount = m->busy_count; 3139 cpu_ccfence(); 3140 if (ocount & PBUSY_LOCKED) 3141 return 1; 3142 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1)) 3143 break; 3144 } 3145 return 0; 3146 #if 0 3147 if (m->busy_count & PBUSY_LOCKED) 3148 return 1; 3149 ocount = atomic_fetchadd_int(&m->busy_count, 1); 3150 if (ocount & PBUSY_LOCKED) { 3151 vm_page_sbusy_drop(m); 3152 return 1; 3153 } 3154 return 0; 3155 #endif 3156 } 3157 3158 /* 3159 * Indicate that a clean VM page requires a filesystem commit and cannot 3160 * be reused. Used by tmpfs. 3161 */ 3162 void 3163 vm_page_need_commit(vm_page_t m) 3164 { 3165 vm_page_flag_set(m, PG_NEED_COMMIT); 3166 vm_object_set_writeable_dirty(m->object); 3167 } 3168 3169 void 3170 vm_page_clear_commit(vm_page_t m) 3171 { 3172 vm_page_flag_clear(m, PG_NEED_COMMIT); 3173 } 3174 3175 /* 3176 * Grab a page, blocking if it is busy and allocating a page if necessary. 3177 * A busy page is returned or NULL. The page may or may not be valid and 3178 * might not be on a queue (the caller is responsible for the disposition of 3179 * the page). 3180 * 3181 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 3182 * page will be zero'd and marked valid. 3183 * 3184 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 3185 * valid even if it already exists. 3186 * 3187 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 3188 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 3189 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 3190 * 3191 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 3192 * always returned if we had blocked. 3193 * 3194 * This routine may not be called from an interrupt. 3195 * 3196 * No other requirements. 3197 */ 3198 vm_page_t 3199 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3200 { 3201 vm_page_t m; 3202 int error; 3203 int shared = 1; 3204 3205 KKASSERT(allocflags & 3206 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 3207 vm_object_hold_shared(object); 3208 for (;;) { 3209 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3210 if (error) { 3211 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 3212 if ((allocflags & VM_ALLOC_RETRY) == 0) { 3213 m = NULL; 3214 break; 3215 } 3216 /* retry */ 3217 } else if (m == NULL) { 3218 if (shared) { 3219 vm_object_upgrade(object); 3220 shared = 0; 3221 } 3222 if (allocflags & VM_ALLOC_RETRY) 3223 allocflags |= VM_ALLOC_NULL_OK; 3224 m = vm_page_alloc(object, pindex, 3225 allocflags & ~VM_ALLOC_RETRY); 3226 if (m) 3227 break; 3228 vm_wait(0); 3229 if ((allocflags & VM_ALLOC_RETRY) == 0) 3230 goto failed; 3231 } else { 3232 /* m found */ 3233 break; 3234 } 3235 } 3236 3237 /* 3238 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 3239 * 3240 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 3241 * valid even if already valid. 3242 * 3243 * NOTE! We have removed all of the PG_ZERO optimizations and also 3244 * removed the idle zeroing code. These optimizations actually 3245 * slow things down on modern cpus because the zerod area is 3246 * likely uncached, placing a memory-access burden on the 3247 * accesors taking the fault. 3248 * 3249 * By always zeroing the page in-line with the fault, no 3250 * dynamic ram reads are needed and the caches are hot, ready 3251 * for userland to access the memory. 3252 */ 3253 if (m->valid == 0) { 3254 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 3255 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3256 m->valid = VM_PAGE_BITS_ALL; 3257 } 3258 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 3259 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 3260 m->valid = VM_PAGE_BITS_ALL; 3261 } 3262 failed: 3263 vm_object_drop(object); 3264 return(m); 3265 } 3266 3267 /* 3268 * Mapping function for valid bits or for dirty bits in 3269 * a page. May not block. 3270 * 3271 * Inputs are required to range within a page. 3272 * 3273 * No requirements. 3274 * Non blocking. 3275 */ 3276 int 3277 vm_page_bits(int base, int size) 3278 { 3279 int first_bit; 3280 int last_bit; 3281 3282 KASSERT( 3283 base + size <= PAGE_SIZE, 3284 ("vm_page_bits: illegal base/size %d/%d", base, size) 3285 ); 3286 3287 if (size == 0) /* handle degenerate case */ 3288 return(0); 3289 3290 first_bit = base >> DEV_BSHIFT; 3291 last_bit = (base + size - 1) >> DEV_BSHIFT; 3292 3293 return ((2 << last_bit) - (1 << first_bit)); 3294 } 3295 3296 /* 3297 * Sets portions of a page valid and clean. The arguments are expected 3298 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3299 * of any partial chunks touched by the range. The invalid portion of 3300 * such chunks will be zero'd. 3301 * 3302 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 3303 * align base to DEV_BSIZE so as not to mark clean a partially 3304 * truncated device block. Otherwise the dirty page status might be 3305 * lost. 3306 * 3307 * This routine may not block. 3308 * 3309 * (base + size) must be less then or equal to PAGE_SIZE. 3310 */ 3311 static void 3312 _vm_page_zero_valid(vm_page_t m, int base, int size) 3313 { 3314 int frag; 3315 int endoff; 3316 3317 if (size == 0) /* handle degenerate case */ 3318 return; 3319 3320 /* 3321 * If the base is not DEV_BSIZE aligned and the valid 3322 * bit is clear, we have to zero out a portion of the 3323 * first block. 3324 */ 3325 3326 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3327 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3328 ) { 3329 pmap_zero_page_area( 3330 VM_PAGE_TO_PHYS(m), 3331 frag, 3332 base - frag 3333 ); 3334 } 3335 3336 /* 3337 * If the ending offset is not DEV_BSIZE aligned and the 3338 * valid bit is clear, we have to zero out a portion of 3339 * the last block. 3340 */ 3341 3342 endoff = base + size; 3343 3344 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3345 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3346 ) { 3347 pmap_zero_page_area( 3348 VM_PAGE_TO_PHYS(m), 3349 endoff, 3350 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3351 ); 3352 } 3353 } 3354 3355 /* 3356 * Set valid, clear dirty bits. If validating the entire 3357 * page we can safely clear the pmap modify bit. We also 3358 * use this opportunity to clear the PG_NOSYNC flag. If a process 3359 * takes a write fault on a MAP_NOSYNC memory area the flag will 3360 * be set again. 3361 * 3362 * We set valid bits inclusive of any overlap, but we can only 3363 * clear dirty bits for DEV_BSIZE chunks that are fully within 3364 * the range. 3365 * 3366 * Page must be busied? 3367 * No other requirements. 3368 */ 3369 void 3370 vm_page_set_valid(vm_page_t m, int base, int size) 3371 { 3372 _vm_page_zero_valid(m, base, size); 3373 m->valid |= vm_page_bits(base, size); 3374 } 3375 3376 3377 /* 3378 * Set valid bits and clear dirty bits. 3379 * 3380 * Page must be busied by caller. 3381 * 3382 * NOTE: This function does not clear the pmap modified bit. 3383 * Also note that e.g. NFS may use a byte-granular base 3384 * and size. 3385 * 3386 * No other requirements. 3387 */ 3388 void 3389 vm_page_set_validclean(vm_page_t m, int base, int size) 3390 { 3391 int pagebits; 3392 3393 _vm_page_zero_valid(m, base, size); 3394 pagebits = vm_page_bits(base, size); 3395 m->valid |= pagebits; 3396 m->dirty &= ~pagebits; 3397 if (base == 0 && size == PAGE_SIZE) { 3398 /*pmap_clear_modify(m);*/ 3399 vm_page_flag_clear(m, PG_NOSYNC); 3400 } 3401 } 3402 3403 /* 3404 * Set valid & dirty. Used by buwrite() 3405 * 3406 * Page must be busied by caller. 3407 */ 3408 void 3409 vm_page_set_validdirty(vm_page_t m, int base, int size) 3410 { 3411 int pagebits; 3412 3413 pagebits = vm_page_bits(base, size); 3414 m->valid |= pagebits; 3415 m->dirty |= pagebits; 3416 if (m->object) 3417 vm_object_set_writeable_dirty(m->object); 3418 } 3419 3420 /* 3421 * Clear dirty bits. 3422 * 3423 * NOTE: This function does not clear the pmap modified bit. 3424 * Also note that e.g. NFS may use a byte-granular base 3425 * and size. 3426 * 3427 * Page must be busied? 3428 * No other requirements. 3429 */ 3430 void 3431 vm_page_clear_dirty(vm_page_t m, int base, int size) 3432 { 3433 m->dirty &= ~vm_page_bits(base, size); 3434 if (base == 0 && size == PAGE_SIZE) { 3435 /*pmap_clear_modify(m);*/ 3436 vm_page_flag_clear(m, PG_NOSYNC); 3437 } 3438 } 3439 3440 /* 3441 * Make the page all-dirty. 3442 * 3443 * Also make sure the related object and vnode reflect the fact that the 3444 * object may now contain a dirty page. 3445 * 3446 * Page must be busied? 3447 * No other requirements. 3448 */ 3449 void 3450 vm_page_dirty(vm_page_t m) 3451 { 3452 #ifdef INVARIANTS 3453 int pqtype = m->queue - m->pc; 3454 #endif 3455 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3456 ("vm_page_dirty: page in free/cache queue!")); 3457 if (m->dirty != VM_PAGE_BITS_ALL) { 3458 m->dirty = VM_PAGE_BITS_ALL; 3459 if (m->object) 3460 vm_object_set_writeable_dirty(m->object); 3461 } 3462 } 3463 3464 /* 3465 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3466 * valid and dirty bits for the effected areas are cleared. 3467 * 3468 * Page must be busied? 3469 * Does not block. 3470 * No other requirements. 3471 */ 3472 void 3473 vm_page_set_invalid(vm_page_t m, int base, int size) 3474 { 3475 int bits; 3476 3477 bits = vm_page_bits(base, size); 3478 m->valid &= ~bits; 3479 m->dirty &= ~bits; 3480 atomic_add_int(&m->object->generation, 1); 3481 } 3482 3483 /* 3484 * The kernel assumes that the invalid portions of a page contain 3485 * garbage, but such pages can be mapped into memory by user code. 3486 * When this occurs, we must zero out the non-valid portions of the 3487 * page so user code sees what it expects. 3488 * 3489 * Pages are most often semi-valid when the end of a file is mapped 3490 * into memory and the file's size is not page aligned. 3491 * 3492 * Page must be busied? 3493 * No other requirements. 3494 */ 3495 void 3496 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3497 { 3498 int b; 3499 int i; 3500 3501 /* 3502 * Scan the valid bits looking for invalid sections that 3503 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3504 * valid bit may be set ) have already been zerod by 3505 * vm_page_set_validclean(). 3506 */ 3507 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3508 if (i == (PAGE_SIZE / DEV_BSIZE) || 3509 (m->valid & (1 << i)) 3510 ) { 3511 if (i > b) { 3512 pmap_zero_page_area( 3513 VM_PAGE_TO_PHYS(m), 3514 b << DEV_BSHIFT, 3515 (i - b) << DEV_BSHIFT 3516 ); 3517 } 3518 b = i + 1; 3519 } 3520 } 3521 3522 /* 3523 * setvalid is TRUE when we can safely set the zero'd areas 3524 * as being valid. We can do this if there are no cache consistency 3525 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3526 */ 3527 if (setvalid) 3528 m->valid = VM_PAGE_BITS_ALL; 3529 } 3530 3531 /* 3532 * Is a (partial) page valid? Note that the case where size == 0 3533 * will return FALSE in the degenerate case where the page is entirely 3534 * invalid, and TRUE otherwise. 3535 * 3536 * Does not block. 3537 * No other requirements. 3538 */ 3539 int 3540 vm_page_is_valid(vm_page_t m, int base, int size) 3541 { 3542 int bits = vm_page_bits(base, size); 3543 3544 if (m->valid && ((m->valid & bits) == bits)) 3545 return 1; 3546 else 3547 return 0; 3548 } 3549 3550 /* 3551 * update dirty bits from pmap/mmu. May not block. 3552 * 3553 * Caller must hold the page busy 3554 */ 3555 void 3556 vm_page_test_dirty(vm_page_t m) 3557 { 3558 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3559 vm_page_dirty(m); 3560 } 3561 } 3562 3563 #include "opt_ddb.h" 3564 #ifdef DDB 3565 #include <ddb/ddb.h> 3566 3567 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3568 { 3569 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3570 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3571 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3572 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3573 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3574 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3575 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3576 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3577 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3578 db_printf("vmstats.v_inactive_target: %ld\n", 3579 vmstats.v_inactive_target); 3580 } 3581 3582 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3583 { 3584 int i; 3585 db_printf("PQ_FREE:"); 3586 for (i = 0; i < PQ_L2_SIZE; i++) { 3587 db_printf(" %ld", vm_page_queues[PQ_FREE + i].lcnt); 3588 } 3589 db_printf("\n"); 3590 3591 db_printf("PQ_CACHE:"); 3592 for(i = 0; i < PQ_L2_SIZE; i++) { 3593 db_printf(" %ld", vm_page_queues[PQ_CACHE + i].lcnt); 3594 } 3595 db_printf("\n"); 3596 3597 db_printf("PQ_ACTIVE:"); 3598 for(i = 0; i < PQ_L2_SIZE; i++) { 3599 db_printf(" %ld", vm_page_queues[PQ_ACTIVE + i].lcnt); 3600 } 3601 db_printf("\n"); 3602 3603 db_printf("PQ_INACTIVE:"); 3604 for(i = 0; i < PQ_L2_SIZE; i++) { 3605 db_printf(" %ld", vm_page_queues[PQ_INACTIVE + i].lcnt); 3606 } 3607 db_printf("\n"); 3608 } 3609 #endif /* DDB */ 3610