1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 33 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 34 */ 35 36 /* 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 /* 63 * Resident memory management module. The module manipulates 'VM pages'. 64 * A VM page is the core building block for memory management. 65 */ 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/malloc.h> 70 #include <sys/proc.h> 71 #include <sys/vmmeter.h> 72 #include <sys/vnode.h> 73 #include <sys/kernel.h> 74 #include <sys/alist.h> 75 #include <sys/sysctl.h> 76 #include <sys/cpu_topology.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_param.h> 80 #include <sys/lock.h> 81 #include <vm/vm_kern.h> 82 #include <vm/pmap.h> 83 #include <vm/vm_map.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_pageout.h> 87 #include <vm/vm_pager.h> 88 #include <vm/vm_extern.h> 89 #include <vm/swap_pager.h> 90 91 #include <machine/inttypes.h> 92 #include <machine/md_var.h> 93 #include <machine/specialreg.h> 94 95 #include <vm/vm_page2.h> 96 #include <sys/spinlock2.h> 97 98 /* 99 * Action hash for user umtx support. 100 */ 101 #define VMACTION_HSIZE 256 102 #define VMACTION_HMASK (VMACTION_HSIZE - 1) 103 104 /* 105 * SET - Minimum required set associative size, must be a power of 2. We 106 * want this to match or exceed the set-associativeness of the cpu. 107 * 108 * GRP - A larger set that allows bleed-over into the domains of other 109 * nearby cpus. Also must be a power of 2. Used by the page zeroing 110 * code to smooth things out a bit. 111 */ 112 #define PQ_SET_ASSOC 16 113 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1) 114 115 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2) 116 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1) 117 118 static void vm_page_queue_init(void); 119 static void vm_page_free_wakeup(void); 120 static vm_page_t vm_page_select_cache(u_short pg_color); 121 static vm_page_t _vm_page_list_find2(int basequeue, int index); 122 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 123 124 /* 125 * Array of tailq lists 126 */ 127 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 128 129 LIST_HEAD(vm_page_action_list, vm_page_action); 130 131 struct vm_page_action_hash { 132 struct vm_page_action_list list; 133 struct lock lk; 134 } __cachealign; 135 136 struct vm_page_action_hash action_hash[VMACTION_HSIZE]; 137 static volatile int vm_pages_waiting; 138 139 static struct alist vm_contig_alist; 140 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 141 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 142 143 static u_long vm_dma_reserved = 0; 144 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 145 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 146 "Memory reserved for DMA"); 147 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 148 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 149 150 static int vm_contig_verbose = 0; 151 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 152 153 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 154 vm_pindex_t, pindex); 155 156 static void 157 vm_page_queue_init(void) 158 { 159 int i; 160 161 for (i = 0; i < PQ_L2_SIZE; i++) 162 vm_page_queues[PQ_FREE+i].cnt_offset = 163 offsetof(struct vmstats, v_free_count); 164 for (i = 0; i < PQ_L2_SIZE; i++) 165 vm_page_queues[PQ_CACHE+i].cnt_offset = 166 offsetof(struct vmstats, v_cache_count); 167 for (i = 0; i < PQ_L2_SIZE; i++) 168 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 169 offsetof(struct vmstats, v_inactive_count); 170 for (i = 0; i < PQ_L2_SIZE; i++) 171 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 172 offsetof(struct vmstats, v_active_count); 173 for (i = 0; i < PQ_L2_SIZE; i++) 174 vm_page_queues[PQ_HOLD+i].cnt_offset = 175 offsetof(struct vmstats, v_active_count); 176 /* PQ_NONE has no queue */ 177 178 for (i = 0; i < PQ_COUNT; i++) { 179 TAILQ_INIT(&vm_page_queues[i].pl); 180 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 181 } 182 183 /* 184 * NOTE: Action lock might recurse due to callback, so allow 185 * recursion. 186 */ 187 for (i = 0; i < VMACTION_HSIZE; i++) { 188 LIST_INIT(&action_hash[i].list); 189 lockinit(&action_hash[i].lk, "actlk", 0, LK_CANRECURSE); 190 } 191 } 192 193 /* 194 * note: place in initialized data section? Is this necessary? 195 */ 196 long first_page = 0; 197 int vm_page_array_size = 0; 198 vm_page_t vm_page_array = NULL; 199 vm_paddr_t vm_low_phys_reserved; 200 201 /* 202 * (low level boot) 203 * 204 * Sets the page size, perhaps based upon the memory size. 205 * Must be called before any use of page-size dependent functions. 206 */ 207 void 208 vm_set_page_size(void) 209 { 210 if (vmstats.v_page_size == 0) 211 vmstats.v_page_size = PAGE_SIZE; 212 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 213 panic("vm_set_page_size: page size not a power of two"); 214 } 215 216 /* 217 * (low level boot) 218 * 219 * Add a new page to the freelist for use by the system. New pages 220 * are added to both the head and tail of the associated free page 221 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 222 * requests pull 'recent' adds (higher physical addresses) first. 223 * 224 * Beware that the page zeroing daemon will also be running soon after 225 * boot, moving pages from the head to the tail of the PQ_FREE queues. 226 * 227 * Must be called in a critical section. 228 */ 229 static void 230 vm_add_new_page(vm_paddr_t pa) 231 { 232 struct vpgqueues *vpq; 233 vm_page_t m; 234 235 m = PHYS_TO_VM_PAGE(pa); 236 m->phys_addr = pa; 237 m->flags = 0; 238 m->pat_mode = PAT_WRITE_BACK; 239 m->pc = (pa >> PAGE_SHIFT); 240 241 /* 242 * Twist for cpu localization in addition to page coloring, so 243 * different cpus selecting by m->queue get different page colors. 244 */ 245 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 246 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 247 m->pc &= PQ_L2_MASK; 248 249 /* 250 * Reserve a certain number of contiguous low memory pages for 251 * contigmalloc() to use. 252 */ 253 if (pa < vm_low_phys_reserved) { 254 atomic_add_int(&vmstats.v_page_count, 1); 255 atomic_add_int(&vmstats.v_dma_pages, 1); 256 m->queue = PQ_NONE; 257 m->wire_count = 1; 258 atomic_add_int(&vmstats.v_wire_count, 1); 259 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 260 return; 261 } 262 263 /* 264 * General page 265 */ 266 m->queue = m->pc + PQ_FREE; 267 KKASSERT(m->dirty == 0); 268 269 atomic_add_int(&vmstats.v_page_count, 1); 270 atomic_add_int(&vmstats.v_free_count, 1); 271 vpq = &vm_page_queues[m->queue]; 272 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 273 ++vpq->lcnt; 274 } 275 276 /* 277 * (low level boot) 278 * 279 * Initializes the resident memory module. 280 * 281 * Preallocates memory for critical VM structures and arrays prior to 282 * kernel_map becoming available. 283 * 284 * Memory is allocated from (virtual2_start, virtual2_end) if available, 285 * otherwise memory is allocated from (virtual_start, virtual_end). 286 * 287 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 288 * large enough to hold vm_page_array & other structures for machines with 289 * large amounts of ram, so we want to use virtual2* when available. 290 */ 291 void 292 vm_page_startup(void) 293 { 294 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 295 vm_offset_t mapped; 296 vm_size_t npages; 297 vm_paddr_t page_range; 298 vm_paddr_t new_end; 299 int i; 300 vm_paddr_t pa; 301 vm_paddr_t last_pa; 302 vm_paddr_t end; 303 vm_paddr_t biggestone, biggestsize; 304 vm_paddr_t total; 305 vm_page_t m; 306 307 total = 0; 308 biggestsize = 0; 309 biggestone = 0; 310 vaddr = round_page(vaddr); 311 312 /* 313 * Make sure ranges are page-aligned. 314 */ 315 for (i = 0; phys_avail[i].phys_end; ++i) { 316 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 317 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 318 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 319 phys_avail[i].phys_end = phys_avail[i].phys_beg; 320 } 321 322 /* 323 * Locate largest block 324 */ 325 for (i = 0; phys_avail[i].phys_end; ++i) { 326 vm_paddr_t size = phys_avail[i].phys_end - 327 phys_avail[i].phys_beg; 328 329 if (size > biggestsize) { 330 biggestone = i; 331 biggestsize = size; 332 } 333 total += size; 334 } 335 --i; /* adjust to last entry for use down below */ 336 337 end = phys_avail[biggestone].phys_end; 338 end = trunc_page(end); 339 340 /* 341 * Initialize the queue headers for the free queue, the active queue 342 * and the inactive queue. 343 */ 344 vm_page_queue_init(); 345 346 #if !defined(_KERNEL_VIRTUAL) 347 /* 348 * VKERNELs don't support minidumps and as such don't need 349 * vm_page_dump 350 * 351 * Allocate a bitmap to indicate that a random physical page 352 * needs to be included in a minidump. 353 * 354 * The amd64 port needs this to indicate which direct map pages 355 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 356 * 357 * However, i386 still needs this workspace internally within the 358 * minidump code. In theory, they are not needed on i386, but are 359 * included should the sf_buf code decide to use them. 360 */ 361 page_range = phys_avail[i].phys_end / PAGE_SIZE; 362 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 363 end -= vm_page_dump_size; 364 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 365 VM_PROT_READ | VM_PROT_WRITE); 366 bzero((void *)vm_page_dump, vm_page_dump_size); 367 #endif 368 /* 369 * Compute the number of pages of memory that will be available for 370 * use (taking into account the overhead of a page structure per 371 * page). 372 */ 373 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 374 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 375 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 376 377 #ifndef _KERNEL_VIRTUAL 378 /* 379 * (only applies to real kernels) 380 * 381 * Reserve a large amount of low memory for potential 32-bit DMA 382 * space allocations. Once device initialization is complete we 383 * release most of it, but keep (vm_dma_reserved) memory reserved 384 * for later use. Typically for X / graphics. Through trial and 385 * error we find that GPUs usually requires ~60-100MB or so. 386 * 387 * By default, 128M is left in reserve on machines with 2G+ of ram. 388 */ 389 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 390 if (vm_low_phys_reserved > total / 4) 391 vm_low_phys_reserved = total / 4; 392 if (vm_dma_reserved == 0) { 393 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 394 if (vm_dma_reserved > total / 16) 395 vm_dma_reserved = total / 16; 396 } 397 #endif 398 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 399 ALIST_RECORDS_65536); 400 401 /* 402 * Initialize the mem entry structures now, and put them in the free 403 * queue. 404 */ 405 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 406 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 407 vm_page_array = (vm_page_t)mapped; 408 409 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 410 /* 411 * since pmap_map on amd64 returns stuff out of a direct-map region, 412 * we have to manually add these pages to the minidump tracking so 413 * that they can be dumped, including the vm_page_array. 414 */ 415 for (pa = new_end; 416 pa < phys_avail[biggestone].phys_end; 417 pa += PAGE_SIZE) { 418 dump_add_page(pa); 419 } 420 #endif 421 422 /* 423 * Clear all of the page structures, run basic initialization so 424 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 425 * map. 426 */ 427 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 428 vm_page_array_size = page_range; 429 430 m = &vm_page_array[0]; 431 pa = ptoa(first_page); 432 for (i = 0; i < page_range; ++i) { 433 spin_init(&m->spin, "vm_page"); 434 m->phys_addr = pa; 435 pa += PAGE_SIZE; 436 ++m; 437 } 438 439 /* 440 * Construct the free queue(s) in ascending order (by physical 441 * address) so that the first 16MB of physical memory is allocated 442 * last rather than first. On large-memory machines, this avoids 443 * the exhaustion of low physical memory before isa_dmainit has run. 444 */ 445 vmstats.v_page_count = 0; 446 vmstats.v_free_count = 0; 447 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 448 pa = phys_avail[i].phys_beg; 449 if (i == biggestone) 450 last_pa = new_end; 451 else 452 last_pa = phys_avail[i].phys_end; 453 while (pa < last_pa && npages-- > 0) { 454 vm_add_new_page(pa); 455 pa += PAGE_SIZE; 456 } 457 } 458 if (virtual2_start) 459 virtual2_start = vaddr; 460 else 461 virtual_start = vaddr; 462 mycpu->gd_vmstats = vmstats; 463 } 464 465 /* 466 * Reorganize VM pages based on numa data. May be called as many times as 467 * necessary. Will reorganize the vm_page_t page color and related queue(s) 468 * to allow vm_page_alloc() to choose pages based on socket affinity. 469 * 470 * NOTE: This function is only called while we are still in UP mode, so 471 * we only need a critical section to protect the queues (which 472 * saves a lot of time, there are likely a ton of pages). 473 */ 474 void 475 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 476 { 477 vm_paddr_t scan_beg; 478 vm_paddr_t scan_end; 479 vm_paddr_t ran_end; 480 struct vpgqueues *vpq; 481 vm_page_t m; 482 vm_page_t mend; 483 int i; 484 int socket_mod; 485 int socket_value; 486 487 /* 488 * Check if no physical information, or there was only one socket 489 * (so don't waste time doing nothing!). 490 */ 491 if (cpu_topology_phys_ids <= 1 || 492 cpu_topology_core_ids == 0) { 493 return; 494 } 495 496 /* 497 * Setup for our iteration. Note that ACPI may iterate CPU 498 * sockets starting at 0 or 1 or some other number. The 499 * cpu_topology code mod's it against the socket count. 500 */ 501 ran_end = ran_beg + bytes; 502 physid %= cpu_topology_phys_ids; 503 504 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 505 socket_value = physid * socket_mod; 506 mend = &vm_page_array[vm_page_array_size]; 507 508 crit_enter(); 509 510 /* 511 * Adjust vm_page->pc and requeue all affected pages. The 512 * allocator will then be able to localize memory allocations 513 * to some degree. 514 */ 515 for (i = 0; phys_avail[i].phys_end; ++i) { 516 scan_beg = phys_avail[i].phys_beg; 517 scan_end = phys_avail[i].phys_end; 518 if (scan_end <= ran_beg) 519 continue; 520 if (scan_beg >= ran_end) 521 continue; 522 if (scan_beg < ran_beg) 523 scan_beg = ran_beg; 524 if (scan_end > ran_end) 525 scan_end = ran_end; 526 if (atop(scan_end) > first_page + vm_page_array_size) 527 scan_end = ptoa(first_page + vm_page_array_size); 528 529 m = PHYS_TO_VM_PAGE(scan_beg); 530 while (scan_beg < scan_end) { 531 KKASSERT(m < mend); 532 if (m->queue != PQ_NONE) { 533 vpq = &vm_page_queues[m->queue]; 534 TAILQ_REMOVE(&vpq->pl, m, pageq); 535 --vpq->lcnt; 536 /* queue doesn't change, no need to adj cnt */ 537 m->queue -= m->pc; 538 m->pc %= socket_mod; 539 m->pc += socket_value; 540 m->pc &= PQ_L2_MASK; 541 m->queue += m->pc; 542 vpq = &vm_page_queues[m->queue]; 543 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 544 ++vpq->lcnt; 545 /* queue doesn't change, no need to adj cnt */ 546 } else { 547 m->pc %= socket_mod; 548 m->pc += socket_value; 549 m->pc &= PQ_L2_MASK; 550 } 551 scan_beg += PAGE_SIZE; 552 ++m; 553 } 554 } 555 crit_exit(); 556 } 557 558 /* 559 * We tended to reserve a ton of memory for contigmalloc(). Now that most 560 * drivers have initialized we want to return most the remaining free 561 * reserve back to the VM page queues so they can be used for normal 562 * allocations. 563 * 564 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 565 */ 566 static void 567 vm_page_startup_finish(void *dummy __unused) 568 { 569 alist_blk_t blk; 570 alist_blk_t rblk; 571 alist_blk_t count; 572 alist_blk_t xcount; 573 alist_blk_t bfree; 574 vm_page_t m; 575 576 spin_lock(&vm_contig_spin); 577 for (;;) { 578 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 579 if (bfree <= vm_dma_reserved / PAGE_SIZE) 580 break; 581 if (count == 0) 582 break; 583 584 /* 585 * Figure out how much of the initial reserve we have to 586 * free in order to reach our target. 587 */ 588 bfree -= vm_dma_reserved / PAGE_SIZE; 589 if (count > bfree) { 590 blk += count - bfree; 591 count = bfree; 592 } 593 594 /* 595 * Calculate the nearest power of 2 <= count. 596 */ 597 for (xcount = 1; xcount <= count; xcount <<= 1) 598 ; 599 xcount >>= 1; 600 blk += count - xcount; 601 count = xcount; 602 603 /* 604 * Allocate the pages from the alist, then free them to 605 * the normal VM page queues. 606 * 607 * Pages allocated from the alist are wired. We have to 608 * busy, unwire, and free them. We must also adjust 609 * vm_low_phys_reserved before freeing any pages to prevent 610 * confusion. 611 */ 612 rblk = alist_alloc(&vm_contig_alist, blk, count); 613 if (rblk != blk) { 614 kprintf("vm_page_startup_finish: Unable to return " 615 "dma space @0x%08x/%d -> 0x%08x\n", 616 blk, count, rblk); 617 break; 618 } 619 atomic_add_int(&vmstats.v_dma_pages, -count); 620 spin_unlock(&vm_contig_spin); 621 622 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 623 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 624 while (count) { 625 vm_page_busy_wait(m, FALSE, "cpgfr"); 626 vm_page_unwire(m, 0); 627 vm_page_free(m); 628 --count; 629 ++m; 630 } 631 spin_lock(&vm_contig_spin); 632 } 633 spin_unlock(&vm_contig_spin); 634 635 /* 636 * Print out how much DMA space drivers have already allocated and 637 * how much is left over. 638 */ 639 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 640 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 641 (PAGE_SIZE / 1024), 642 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 643 } 644 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 645 vm_page_startup_finish, NULL); 646 647 648 /* 649 * Scan comparison function for Red-Black tree scans. An inclusive 650 * (start,end) is expected. Other fields are not used. 651 */ 652 int 653 rb_vm_page_scancmp(struct vm_page *p, void *data) 654 { 655 struct rb_vm_page_scan_info *info = data; 656 657 if (p->pindex < info->start_pindex) 658 return(-1); 659 if (p->pindex > info->end_pindex) 660 return(1); 661 return(0); 662 } 663 664 int 665 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 666 { 667 if (p1->pindex < p2->pindex) 668 return(-1); 669 if (p1->pindex > p2->pindex) 670 return(1); 671 return(0); 672 } 673 674 void 675 vm_page_init(vm_page_t m) 676 { 677 /* do nothing for now. Called from pmap_page_init() */ 678 } 679 680 /* 681 * Each page queue has its own spin lock, which is fairly optimal for 682 * allocating and freeing pages at least. 683 * 684 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 685 * queue spinlock via this function. Also note that m->queue cannot change 686 * unless both the page and queue are locked. 687 */ 688 static __inline 689 void 690 _vm_page_queue_spin_lock(vm_page_t m) 691 { 692 u_short queue; 693 694 queue = m->queue; 695 if (queue != PQ_NONE) { 696 spin_lock(&vm_page_queues[queue].spin); 697 KKASSERT(queue == m->queue); 698 } 699 } 700 701 static __inline 702 void 703 _vm_page_queue_spin_unlock(vm_page_t m) 704 { 705 u_short queue; 706 707 queue = m->queue; 708 cpu_ccfence(); 709 if (queue != PQ_NONE) 710 spin_unlock(&vm_page_queues[queue].spin); 711 } 712 713 static __inline 714 void 715 _vm_page_queues_spin_lock(u_short queue) 716 { 717 cpu_ccfence(); 718 if (queue != PQ_NONE) 719 spin_lock(&vm_page_queues[queue].spin); 720 } 721 722 723 static __inline 724 void 725 _vm_page_queues_spin_unlock(u_short queue) 726 { 727 cpu_ccfence(); 728 if (queue != PQ_NONE) 729 spin_unlock(&vm_page_queues[queue].spin); 730 } 731 732 void 733 vm_page_queue_spin_lock(vm_page_t m) 734 { 735 _vm_page_queue_spin_lock(m); 736 } 737 738 void 739 vm_page_queues_spin_lock(u_short queue) 740 { 741 _vm_page_queues_spin_lock(queue); 742 } 743 744 void 745 vm_page_queue_spin_unlock(vm_page_t m) 746 { 747 _vm_page_queue_spin_unlock(m); 748 } 749 750 void 751 vm_page_queues_spin_unlock(u_short queue) 752 { 753 _vm_page_queues_spin_unlock(queue); 754 } 755 756 /* 757 * This locks the specified vm_page and its queue in the proper order 758 * (page first, then queue). The queue may change so the caller must 759 * recheck on return. 760 */ 761 static __inline 762 void 763 _vm_page_and_queue_spin_lock(vm_page_t m) 764 { 765 vm_page_spin_lock(m); 766 _vm_page_queue_spin_lock(m); 767 } 768 769 static __inline 770 void 771 _vm_page_and_queue_spin_unlock(vm_page_t m) 772 { 773 _vm_page_queues_spin_unlock(m->queue); 774 vm_page_spin_unlock(m); 775 } 776 777 void 778 vm_page_and_queue_spin_unlock(vm_page_t m) 779 { 780 _vm_page_and_queue_spin_unlock(m); 781 } 782 783 void 784 vm_page_and_queue_spin_lock(vm_page_t m) 785 { 786 _vm_page_and_queue_spin_lock(m); 787 } 788 789 /* 790 * Helper function removes vm_page from its current queue. 791 * Returns the base queue the page used to be on. 792 * 793 * The vm_page and the queue must be spinlocked. 794 * This function will unlock the queue but leave the page spinlocked. 795 */ 796 static __inline u_short 797 _vm_page_rem_queue_spinlocked(vm_page_t m) 798 { 799 struct vpgqueues *pq; 800 u_short queue; 801 u_short oqueue; 802 int *cnt; 803 804 queue = m->queue; 805 if (queue != PQ_NONE) { 806 pq = &vm_page_queues[queue]; 807 TAILQ_REMOVE(&pq->pl, m, pageq); 808 809 /* 810 * Adjust our pcpu stats. In order for the nominal low-memory 811 * algorithms to work properly we don't let any pcpu stat get 812 * too negative before we force it to be rolled-up into the 813 * global stats. Otherwise our pageout and vm_wait tests 814 * will fail badly. 815 * 816 * The idea here is to reduce unnecessary SMP cache 817 * mastership changes in the global vmstats, which can be 818 * particularly bad in multi-socket systems. 819 */ 820 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 821 atomic_add_int(cnt, -1); 822 if (*cnt < -VMMETER_SLOP_COUNT) { 823 u_int copy = atomic_swap_int(cnt, 0); 824 cnt = (int *)((char *)&vmstats + pq->cnt_offset); 825 atomic_add_int(cnt, copy); 826 cnt = (int *)((char *)&mycpu->gd_vmstats + 827 pq->cnt_offset); 828 atomic_add_int(cnt, copy); 829 } 830 pq->lcnt--; 831 m->queue = PQ_NONE; 832 oqueue = queue; 833 queue -= m->pc; 834 vm_page_queues_spin_unlock(oqueue); /* intended */ 835 } 836 return queue; 837 } 838 839 /* 840 * Helper function places the vm_page on the specified queue. Generally 841 * speaking only PQ_FREE pages are placed at the head, to allow them to 842 * be allocated sooner rather than later on the assumption that they 843 * are cache-hot. 844 * 845 * The vm_page must be spinlocked. 846 * This function will return with both the page and the queue locked. 847 */ 848 static __inline void 849 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 850 { 851 struct vpgqueues *pq; 852 u_int *cnt; 853 854 KKASSERT(m->queue == PQ_NONE); 855 856 if (queue != PQ_NONE) { 857 vm_page_queues_spin_lock(queue); 858 pq = &vm_page_queues[queue]; 859 ++pq->lcnt; 860 861 /* 862 * Adjust our pcpu stats. If a system entity really needs 863 * to incorporate the count it will call vmstats_rollup() 864 * to roll it all up into the global vmstats strufture. 865 */ 866 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 867 atomic_add_int(cnt, 1); 868 869 /* 870 * PQ_FREE is always handled LIFO style to try to provide 871 * cache-hot pages to programs. 872 */ 873 m->queue = queue; 874 if (queue - m->pc == PQ_FREE) { 875 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 876 } else if (athead) { 877 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 878 } else { 879 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 880 } 881 /* leave the queue spinlocked */ 882 } 883 } 884 885 /* 886 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 887 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we 888 * did not. Only one sleep call will be made before returning. 889 * 890 * This function does NOT busy the page and on return the page is not 891 * guaranteed to be available. 892 */ 893 void 894 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 895 { 896 u_int32_t flags; 897 898 for (;;) { 899 flags = m->flags; 900 cpu_ccfence(); 901 902 if ((flags & PG_BUSY) == 0 && 903 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) { 904 break; 905 } 906 tsleep_interlock(m, 0); 907 if (atomic_cmpset_int(&m->flags, flags, 908 flags | PG_WANTED | PG_REFERENCED)) { 909 tsleep(m, PINTERLOCKED, msg, 0); 910 break; 911 } 912 } 913 } 914 915 /* 916 * This calculates and returns a page color given an optional VM object and 917 * either a pindex or an iterator. We attempt to return a cpu-localized 918 * pg_color that is still roughly 16-way set-associative. The CPU topology 919 * is used if it was probed. 920 * 921 * The caller may use the returned value to index into e.g. PQ_FREE when 922 * allocating a page in order to nominally obtain pages that are hopefully 923 * already localized to the requesting cpu. This function is not able to 924 * provide any sort of guarantee of this, but does its best to improve 925 * hardware cache management performance. 926 * 927 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 928 */ 929 u_short 930 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 931 { 932 u_short pg_color; 933 int phys_id; 934 int core_id; 935 int object_pg_color; 936 937 phys_id = get_cpu_phys_id(cpuid); 938 core_id = get_cpu_core_id(cpuid); 939 object_pg_color = object ? object->pg_color : 0; 940 941 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 942 int grpsize; 943 944 /* 945 * Break us down by socket and cpu 946 */ 947 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids; 948 pg_color += core_id * PQ_L2_SIZE / 949 (cpu_topology_core_ids * cpu_topology_phys_ids); 950 951 /* 952 * Calculate remaining component for object/queue color 953 */ 954 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids * 955 cpu_topology_phys_ids); 956 if (grpsize >= 8) { 957 pg_color += (pindex + object_pg_color) % grpsize; 958 } else { 959 if (grpsize <= 2) { 960 grpsize = 8; 961 } else { 962 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 963 grpsize += grpsize; 964 if (grpsize < 8) 965 grpsize += grpsize; 966 } 967 pg_color += (pindex + object_pg_color) % grpsize; 968 } 969 } else { 970 /* 971 * Unknown topology, distribute things evenly. 972 */ 973 pg_color = cpuid * PQ_L2_SIZE / ncpus; 974 pg_color += pindex + object_pg_color; 975 } 976 return (pg_color & PQ_L2_MASK); 977 } 978 979 /* 980 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we 981 * also wait for m->busy to become 0 before setting PG_BUSY. 982 */ 983 void 984 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 985 int also_m_busy, const char *msg 986 VM_PAGE_DEBUG_ARGS) 987 { 988 u_int32_t flags; 989 990 for (;;) { 991 flags = m->flags; 992 cpu_ccfence(); 993 if (flags & PG_BUSY) { 994 tsleep_interlock(m, 0); 995 if (atomic_cmpset_int(&m->flags, flags, 996 flags | PG_WANTED | PG_REFERENCED)) { 997 tsleep(m, PINTERLOCKED, msg, 0); 998 } 999 } else if (also_m_busy && (flags & PG_SBUSY)) { 1000 tsleep_interlock(m, 0); 1001 if (atomic_cmpset_int(&m->flags, flags, 1002 flags | PG_WANTED | PG_REFERENCED)) { 1003 tsleep(m, PINTERLOCKED, msg, 0); 1004 } 1005 } else { 1006 if (atomic_cmpset_int(&m->flags, flags, 1007 flags | PG_BUSY)) { 1008 #ifdef VM_PAGE_DEBUG 1009 m->busy_func = func; 1010 m->busy_line = lineno; 1011 #endif 1012 break; 1013 } 1014 } 1015 } 1016 } 1017 1018 /* 1019 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy 1020 * is also 0. 1021 * 1022 * Returns non-zero on failure. 1023 */ 1024 int 1025 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1026 VM_PAGE_DEBUG_ARGS) 1027 { 1028 u_int32_t flags; 1029 1030 for (;;) { 1031 flags = m->flags; 1032 cpu_ccfence(); 1033 if (flags & PG_BUSY) 1034 return TRUE; 1035 if (also_m_busy && (flags & PG_SBUSY)) 1036 return TRUE; 1037 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1038 #ifdef VM_PAGE_DEBUG 1039 m->busy_func = func; 1040 m->busy_line = lineno; 1041 #endif 1042 return FALSE; 1043 } 1044 } 1045 } 1046 1047 /* 1048 * Clear the PG_BUSY flag and return non-zero to indicate to the caller 1049 * that a wakeup() should be performed. 1050 * 1051 * The vm_page must be spinlocked and will remain spinlocked on return. 1052 * The related queue must NOT be spinlocked (which could deadlock us). 1053 * 1054 * (inline version) 1055 */ 1056 static __inline 1057 int 1058 _vm_page_wakeup(vm_page_t m) 1059 { 1060 u_int32_t flags; 1061 1062 for (;;) { 1063 flags = m->flags; 1064 cpu_ccfence(); 1065 if (atomic_cmpset_int(&m->flags, flags, 1066 flags & ~(PG_BUSY | PG_WANTED))) { 1067 break; 1068 } 1069 } 1070 return(flags & PG_WANTED); 1071 } 1072 1073 /* 1074 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This 1075 * is typically the last call you make on a page before moving onto 1076 * other things. 1077 */ 1078 void 1079 vm_page_wakeup(vm_page_t m) 1080 { 1081 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 1082 vm_page_spin_lock(m); 1083 if (_vm_page_wakeup(m)) { 1084 vm_page_spin_unlock(m); 1085 wakeup(m); 1086 } else { 1087 vm_page_spin_unlock(m); 1088 } 1089 } 1090 1091 /* 1092 * Holding a page keeps it from being reused. Other parts of the system 1093 * can still disassociate the page from its current object and free it, or 1094 * perform read or write I/O on it and/or otherwise manipulate the page, 1095 * but if the page is held the VM system will leave the page and its data 1096 * intact and not reuse the page for other purposes until the last hold 1097 * reference is released. (see vm_page_wire() if you want to prevent the 1098 * page from being disassociated from its object too). 1099 * 1100 * The caller must still validate the contents of the page and, if necessary, 1101 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 1102 * before manipulating the page. 1103 * 1104 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 1105 */ 1106 void 1107 vm_page_hold(vm_page_t m) 1108 { 1109 vm_page_spin_lock(m); 1110 atomic_add_int(&m->hold_count, 1); 1111 if (m->queue - m->pc == PQ_FREE) { 1112 _vm_page_queue_spin_lock(m); 1113 _vm_page_rem_queue_spinlocked(m); 1114 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1115 _vm_page_queue_spin_unlock(m); 1116 } 1117 vm_page_spin_unlock(m); 1118 } 1119 1120 /* 1121 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1122 * it was freed while held and must be moved back to the FREE queue. 1123 */ 1124 void 1125 vm_page_unhold(vm_page_t m) 1126 { 1127 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1128 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 1129 m, m->hold_count, m->queue - m->pc)); 1130 vm_page_spin_lock(m); 1131 atomic_add_int(&m->hold_count, -1); 1132 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1133 _vm_page_queue_spin_lock(m); 1134 _vm_page_rem_queue_spinlocked(m); 1135 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1136 _vm_page_queue_spin_unlock(m); 1137 } 1138 vm_page_spin_unlock(m); 1139 } 1140 1141 /* 1142 * vm_page_getfake: 1143 * 1144 * Create a fictitious page with the specified physical address and 1145 * memory attribute. The memory attribute is the only the machine- 1146 * dependent aspect of a fictitious page that must be initialized. 1147 */ 1148 1149 void 1150 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1151 { 1152 1153 if ((m->flags & PG_FICTITIOUS) != 0) { 1154 /* 1155 * The page's memattr might have changed since the 1156 * previous initialization. Update the pmap to the 1157 * new memattr. 1158 */ 1159 goto memattr; 1160 } 1161 m->phys_addr = paddr; 1162 m->queue = PQ_NONE; 1163 /* Fictitious pages don't use "segind". */ 1164 /* Fictitious pages don't use "order" or "pool". */ 1165 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY; 1166 m->wire_count = 1; 1167 spin_init(&m->spin, "fake_page"); 1168 pmap_page_init(m); 1169 memattr: 1170 pmap_page_set_memattr(m, memattr); 1171 } 1172 1173 /* 1174 * Inserts the given vm_page into the object and object list. 1175 * 1176 * The pagetables are not updated but will presumably fault the page 1177 * in if necessary, or if a kernel page the caller will at some point 1178 * enter the page into the kernel's pmap. We are not allowed to block 1179 * here so we *can't* do this anyway. 1180 * 1181 * This routine may not block. 1182 * This routine must be called with the vm_object held. 1183 * This routine must be called with a critical section held. 1184 * 1185 * This routine returns TRUE if the page was inserted into the object 1186 * successfully, and FALSE if the page already exists in the object. 1187 */ 1188 int 1189 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1190 { 1191 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1192 if (m->object != NULL) 1193 panic("vm_page_insert: already inserted"); 1194 1195 atomic_add_int(&object->generation, 1); 1196 1197 /* 1198 * Record the object/offset pair in this page and add the 1199 * pv_list_count of the page to the object. 1200 * 1201 * The vm_page spin lock is required for interactions with the pmap. 1202 */ 1203 vm_page_spin_lock(m); 1204 m->object = object; 1205 m->pindex = pindex; 1206 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1207 m->object = NULL; 1208 m->pindex = 0; 1209 vm_page_spin_unlock(m); 1210 return FALSE; 1211 } 1212 ++object->resident_page_count; 1213 ++mycpu->gd_vmtotal.t_rm; 1214 vm_page_spin_unlock(m); 1215 1216 /* 1217 * Since we are inserting a new and possibly dirty page, 1218 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1219 */ 1220 if ((m->valid & m->dirty) || 1221 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1222 vm_object_set_writeable_dirty(object); 1223 1224 /* 1225 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1226 */ 1227 swap_pager_page_inserted(m); 1228 return TRUE; 1229 } 1230 1231 /* 1232 * Removes the given vm_page_t from the (object,index) table 1233 * 1234 * The underlying pmap entry (if any) is NOT removed here. 1235 * This routine may not block. 1236 * 1237 * The page must be BUSY and will remain BUSY on return. 1238 * No other requirements. 1239 * 1240 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1241 * it busy. 1242 */ 1243 void 1244 vm_page_remove(vm_page_t m) 1245 { 1246 vm_object_t object; 1247 1248 if (m->object == NULL) { 1249 return; 1250 } 1251 1252 if ((m->flags & PG_BUSY) == 0) 1253 panic("vm_page_remove: page not busy"); 1254 1255 object = m->object; 1256 1257 vm_object_hold(object); 1258 1259 /* 1260 * Remove the page from the object and update the object. 1261 * 1262 * The vm_page spin lock is required for interactions with the pmap. 1263 */ 1264 vm_page_spin_lock(m); 1265 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1266 --object->resident_page_count; 1267 --mycpu->gd_vmtotal.t_rm; 1268 m->object = NULL; 1269 atomic_add_int(&object->generation, 1); 1270 vm_page_spin_unlock(m); 1271 1272 vm_object_drop(object); 1273 } 1274 1275 /* 1276 * Locate and return the page at (object, pindex), or NULL if the 1277 * page could not be found. 1278 * 1279 * The caller must hold the vm_object token. 1280 */ 1281 vm_page_t 1282 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1283 { 1284 vm_page_t m; 1285 1286 /* 1287 * Search the hash table for this object/offset pair 1288 */ 1289 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1290 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1291 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1292 return(m); 1293 } 1294 1295 vm_page_t 1296 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1297 vm_pindex_t pindex, 1298 int also_m_busy, const char *msg 1299 VM_PAGE_DEBUG_ARGS) 1300 { 1301 u_int32_t flags; 1302 vm_page_t m; 1303 1304 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1305 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1306 while (m) { 1307 KKASSERT(m->object == object && m->pindex == pindex); 1308 flags = m->flags; 1309 cpu_ccfence(); 1310 if (flags & PG_BUSY) { 1311 tsleep_interlock(m, 0); 1312 if (atomic_cmpset_int(&m->flags, flags, 1313 flags | PG_WANTED | PG_REFERENCED)) { 1314 tsleep(m, PINTERLOCKED, msg, 0); 1315 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1316 pindex); 1317 } 1318 } else if (also_m_busy && (flags & PG_SBUSY)) { 1319 tsleep_interlock(m, 0); 1320 if (atomic_cmpset_int(&m->flags, flags, 1321 flags | PG_WANTED | PG_REFERENCED)) { 1322 tsleep(m, PINTERLOCKED, msg, 0); 1323 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1324 pindex); 1325 } 1326 } else if (atomic_cmpset_int(&m->flags, flags, 1327 flags | PG_BUSY)) { 1328 #ifdef VM_PAGE_DEBUG 1329 m->busy_func = func; 1330 m->busy_line = lineno; 1331 #endif 1332 break; 1333 } 1334 } 1335 return m; 1336 } 1337 1338 /* 1339 * Attempt to lookup and busy a page. 1340 * 1341 * Returns NULL if the page could not be found 1342 * 1343 * Returns a vm_page and error == TRUE if the page exists but could not 1344 * be busied. 1345 * 1346 * Returns a vm_page and error == FALSE on success. 1347 */ 1348 vm_page_t 1349 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1350 vm_pindex_t pindex, 1351 int also_m_busy, int *errorp 1352 VM_PAGE_DEBUG_ARGS) 1353 { 1354 u_int32_t flags; 1355 vm_page_t m; 1356 1357 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1358 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1359 *errorp = FALSE; 1360 while (m) { 1361 KKASSERT(m->object == object && m->pindex == pindex); 1362 flags = m->flags; 1363 cpu_ccfence(); 1364 if (flags & PG_BUSY) { 1365 *errorp = TRUE; 1366 break; 1367 } 1368 if (also_m_busy && (flags & PG_SBUSY)) { 1369 *errorp = TRUE; 1370 break; 1371 } 1372 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) { 1373 #ifdef VM_PAGE_DEBUG 1374 m->busy_func = func; 1375 m->busy_line = lineno; 1376 #endif 1377 break; 1378 } 1379 } 1380 return m; 1381 } 1382 1383 /* 1384 * Attempt to repurpose the passed-in page. If the passed-in page cannot 1385 * be repurposed it will be released, *must_reenter will be set to 1, and 1386 * this function will fall-through to vm_page_lookup_busy_try(). 1387 * 1388 * The passed-in page must be wired and not busy. The returned page will 1389 * be busied and not wired. 1390 * 1391 * A different page may be returned. The returned page will be busied and 1392 * not wired. 1393 * 1394 * NULL can be returned. If so, the required page could not be busied. 1395 * The passed-in page will be unwired. 1396 */ 1397 vm_page_t 1398 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex, 1399 int also_m_busy, int *errorp, vm_page_t m, 1400 int *must_reenter, int *iswired) 1401 { 1402 if (m) { 1403 /* 1404 * Do not mess with pages in a complex state, such as pages 1405 * which are mapped, as repurposing such pages can be more 1406 * expensive than simply allocatin a new one. 1407 * 1408 * NOTE: Soft-busying can deadlock against putpages or I/O 1409 * so we only allow hard-busying here. 1410 */ 1411 KKASSERT(also_m_busy == FALSE); 1412 vm_page_busy_wait(m, also_m_busy, "biodep"); 1413 1414 if ((m->flags & (PG_UNMANAGED | PG_MAPPED | 1415 PG_FICTITIOUS | PG_SBUSY)) || 1416 m->busy || m->wire_count != 1 || m->hold_count) { 1417 vm_page_unwire(m, 0); 1418 vm_page_wakeup(m); 1419 /* fall through to normal lookup */ 1420 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 1421 vm_page_unwire(m, 0); 1422 vm_page_deactivate(m); 1423 vm_page_wakeup(m); 1424 /* fall through to normal lookup */ 1425 } else { 1426 /* 1427 * We can safely repurpose the page. It should 1428 * already be unqueued. 1429 */ 1430 KKASSERT(m->queue == PQ_NONE && m->dirty == 0); 1431 vm_page_remove(m); 1432 m->valid = 0; 1433 m->act_count = 0; 1434 if (vm_page_insert(m, object, pindex)) { 1435 *errorp = 0; 1436 *iswired = 1; 1437 1438 return m; 1439 } 1440 vm_page_unwire(m, 0); 1441 vm_page_free(m); 1442 /* fall through to normal lookup */ 1443 } 1444 } 1445 1446 /* 1447 * Cannot repurpose page, attempt to locate the desired page. May 1448 * return NULL. 1449 */ 1450 *must_reenter = 1; 1451 *iswired = 0; 1452 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp); 1453 1454 return m; 1455 } 1456 1457 /* 1458 * Caller must hold the related vm_object 1459 */ 1460 vm_page_t 1461 vm_page_next(vm_page_t m) 1462 { 1463 vm_page_t next; 1464 1465 next = vm_page_rb_tree_RB_NEXT(m); 1466 if (next && next->pindex != m->pindex + 1) 1467 next = NULL; 1468 return (next); 1469 } 1470 1471 /* 1472 * vm_page_rename() 1473 * 1474 * Move the given vm_page from its current object to the specified 1475 * target object/offset. The page must be busy and will remain so 1476 * on return. 1477 * 1478 * new_object must be held. 1479 * This routine might block. XXX ? 1480 * 1481 * NOTE: Swap associated with the page must be invalidated by the move. We 1482 * have to do this for several reasons: (1) we aren't freeing the 1483 * page, (2) we are dirtying the page, (3) the VM system is probably 1484 * moving the page from object A to B, and will then later move 1485 * the backing store from A to B and we can't have a conflict. 1486 * 1487 * NOTE: We *always* dirty the page. It is necessary both for the 1488 * fact that we moved it, and because we may be invalidating 1489 * swap. If the page is on the cache, we have to deactivate it 1490 * or vm_page_dirty() will panic. Dirty pages are not allowed 1491 * on the cache. 1492 */ 1493 void 1494 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1495 { 1496 KKASSERT(m->flags & PG_BUSY); 1497 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1498 if (m->object) { 1499 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1500 vm_page_remove(m); 1501 } 1502 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1503 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1504 new_object, new_pindex); 1505 } 1506 if (m->queue - m->pc == PQ_CACHE) 1507 vm_page_deactivate(m); 1508 vm_page_dirty(m); 1509 } 1510 1511 /* 1512 * vm_page_unqueue() without any wakeup. This routine is used when a page 1513 * is to remain BUSYied by the caller. 1514 * 1515 * This routine may not block. 1516 */ 1517 void 1518 vm_page_unqueue_nowakeup(vm_page_t m) 1519 { 1520 vm_page_and_queue_spin_lock(m); 1521 (void)_vm_page_rem_queue_spinlocked(m); 1522 vm_page_spin_unlock(m); 1523 } 1524 1525 /* 1526 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1527 * if necessary. 1528 * 1529 * This routine may not block. 1530 */ 1531 void 1532 vm_page_unqueue(vm_page_t m) 1533 { 1534 u_short queue; 1535 1536 vm_page_and_queue_spin_lock(m); 1537 queue = _vm_page_rem_queue_spinlocked(m); 1538 if (queue == PQ_FREE || queue == PQ_CACHE) { 1539 vm_page_spin_unlock(m); 1540 pagedaemon_wakeup(); 1541 } else { 1542 vm_page_spin_unlock(m); 1543 } 1544 } 1545 1546 /* 1547 * vm_page_list_find() 1548 * 1549 * Find a page on the specified queue with color optimization. 1550 * 1551 * The page coloring optimization attempts to locate a page that does 1552 * not overload other nearby pages in the object in the cpu's L1 or L2 1553 * caches. We need this optimization because cpu caches tend to be 1554 * physical caches, while object spaces tend to be virtual. 1555 * 1556 * The page coloring optimization also, very importantly, tries to localize 1557 * memory to cpus and physical sockets. 1558 * 1559 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1560 * and the algorithm is adjusted to localize allocations on a per-core basis. 1561 * This is done by 'twisting' the colors. 1562 * 1563 * The page is returned spinlocked and removed from its queue (it will 1564 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller 1565 * is responsible for dealing with the busy-page case (usually by 1566 * deactivating the page and looping). 1567 * 1568 * NOTE: This routine is carefully inlined. A non-inlined version 1569 * is available for outside callers but the only critical path is 1570 * from within this source file. 1571 * 1572 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1573 * represent stable storage, allowing us to order our locks vm_page 1574 * first, then queue. 1575 */ 1576 static __inline 1577 vm_page_t 1578 _vm_page_list_find(int basequeue, int index) 1579 { 1580 vm_page_t m; 1581 1582 for (;;) { 1583 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1584 if (m == NULL) { 1585 m = _vm_page_list_find2(basequeue, index); 1586 return(m); 1587 } 1588 vm_page_and_queue_spin_lock(m); 1589 if (m->queue == basequeue + index) { 1590 _vm_page_rem_queue_spinlocked(m); 1591 /* vm_page_t spin held, no queue spin */ 1592 break; 1593 } 1594 vm_page_and_queue_spin_unlock(m); 1595 } 1596 return(m); 1597 } 1598 1599 /* 1600 * If we could not find the page in the desired queue try to find it in 1601 * a nearby queue. 1602 */ 1603 static vm_page_t 1604 _vm_page_list_find2(int basequeue, int index) 1605 { 1606 struct vpgqueues *pq; 1607 vm_page_t m = NULL; 1608 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1609 int pqi; 1610 int i; 1611 1612 index &= PQ_L2_MASK; 1613 pq = &vm_page_queues[basequeue]; 1614 1615 /* 1616 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1617 * else fails (PQ_L2_MASK which is 255). 1618 */ 1619 do { 1620 pqmask = (pqmask << 1) | 1; 1621 for (i = 0; i <= pqmask; ++i) { 1622 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1623 m = TAILQ_FIRST(&pq[pqi].pl); 1624 if (m) { 1625 _vm_page_and_queue_spin_lock(m); 1626 if (m->queue == basequeue + pqi) { 1627 _vm_page_rem_queue_spinlocked(m); 1628 return(m); 1629 } 1630 _vm_page_and_queue_spin_unlock(m); 1631 --i; 1632 continue; 1633 } 1634 } 1635 } while (pqmask != PQ_L2_MASK); 1636 1637 return(m); 1638 } 1639 1640 /* 1641 * Returns a vm_page candidate for allocation. The page is not busied so 1642 * it can move around. The caller must busy the page (and typically 1643 * deactivate it if it cannot be busied!) 1644 * 1645 * Returns a spinlocked vm_page that has been removed from its queue. 1646 */ 1647 vm_page_t 1648 vm_page_list_find(int basequeue, int index) 1649 { 1650 return(_vm_page_list_find(basequeue, index)); 1651 } 1652 1653 /* 1654 * Find a page on the cache queue with color optimization, remove it 1655 * from the queue, and busy it. The returned page will not be spinlocked. 1656 * 1657 * A candidate failure will be deactivated. Candidates can fail due to 1658 * being busied by someone else, in which case they will be deactivated. 1659 * 1660 * This routine may not block. 1661 * 1662 */ 1663 static vm_page_t 1664 vm_page_select_cache(u_short pg_color) 1665 { 1666 vm_page_t m; 1667 1668 for (;;) { 1669 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1670 if (m == NULL) 1671 break; 1672 /* 1673 * (m) has been removed from its queue and spinlocked 1674 */ 1675 if (vm_page_busy_try(m, TRUE)) { 1676 _vm_page_deactivate_locked(m, 0); 1677 vm_page_spin_unlock(m); 1678 } else { 1679 /* 1680 * We successfully busied the page 1681 */ 1682 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1683 m->hold_count == 0 && 1684 m->wire_count == 0 && 1685 (m->dirty & m->valid) == 0) { 1686 vm_page_spin_unlock(m); 1687 pagedaemon_wakeup(); 1688 return(m); 1689 } 1690 1691 /* 1692 * The page cannot be recycled, deactivate it. 1693 */ 1694 _vm_page_deactivate_locked(m, 0); 1695 if (_vm_page_wakeup(m)) { 1696 vm_page_spin_unlock(m); 1697 wakeup(m); 1698 } else { 1699 vm_page_spin_unlock(m); 1700 } 1701 } 1702 } 1703 return (m); 1704 } 1705 1706 /* 1707 * Find a free page. We attempt to inline the nominal case and fall back 1708 * to _vm_page_select_free() otherwise. A busied page is removed from 1709 * the queue and returned. 1710 * 1711 * This routine may not block. 1712 */ 1713 static __inline vm_page_t 1714 vm_page_select_free(u_short pg_color) 1715 { 1716 vm_page_t m; 1717 1718 for (;;) { 1719 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1720 if (m == NULL) 1721 break; 1722 if (vm_page_busy_try(m, TRUE)) { 1723 /* 1724 * Various mechanisms such as a pmap_collect can 1725 * result in a busy page on the free queue. We 1726 * have to move the page out of the way so we can 1727 * retry the allocation. If the other thread is not 1728 * allocating the page then m->valid will remain 0 and 1729 * the pageout daemon will free the page later on. 1730 * 1731 * Since we could not busy the page, however, we 1732 * cannot make assumptions as to whether the page 1733 * will be allocated by the other thread or not, 1734 * so all we can do is deactivate it to move it out 1735 * of the way. In particular, if the other thread 1736 * wires the page it may wind up on the inactive 1737 * queue and the pageout daemon will have to deal 1738 * with that case too. 1739 */ 1740 _vm_page_deactivate_locked(m, 0); 1741 vm_page_spin_unlock(m); 1742 } else { 1743 /* 1744 * Theoretically if we are able to busy the page 1745 * atomic with the queue removal (using the vm_page 1746 * lock) nobody else should be able to mess with the 1747 * page before us. 1748 */ 1749 KKASSERT((m->flags & (PG_UNMANAGED | 1750 PG_NEED_COMMIT)) == 0); 1751 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1752 "pg %p q=%d flags=%08x hold=%d wire=%d", 1753 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1754 KKASSERT(m->wire_count == 0); 1755 vm_page_spin_unlock(m); 1756 pagedaemon_wakeup(); 1757 1758 /* return busied and removed page */ 1759 return(m); 1760 } 1761 } 1762 return(m); 1763 } 1764 1765 /* 1766 * vm_page_alloc() 1767 * 1768 * Allocate and return a memory cell associated with this VM object/offset 1769 * pair. If object is NULL an unassociated page will be allocated. 1770 * 1771 * The returned page will be busied and removed from its queues. This 1772 * routine can block and may return NULL if a race occurs and the page 1773 * is found to already exist at the specified (object, pindex). 1774 * 1775 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1776 * VM_ALLOC_QUICK like normal but cannot use cache 1777 * VM_ALLOC_SYSTEM greater free drain 1778 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1779 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1780 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1781 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1782 * (see vm_page_grab()) 1783 * VM_ALLOC_USE_GD ok to use per-gd cache 1784 * 1785 * VM_ALLOC_CPU(n) allocate using specified cpu localization 1786 * 1787 * The object must be held if not NULL 1788 * This routine may not block 1789 * 1790 * Additional special handling is required when called from an interrupt 1791 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1792 * in this case. 1793 */ 1794 vm_page_t 1795 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1796 { 1797 globaldata_t gd; 1798 vm_object_t obj; 1799 vm_page_t m; 1800 u_short pg_color; 1801 int cpuid_local; 1802 1803 #if 0 1804 /* 1805 * Special per-cpu free VM page cache. The pages are pre-busied 1806 * and pre-zerod for us. 1807 */ 1808 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1809 crit_enter_gd(gd); 1810 if (gd->gd_vmpg_count) { 1811 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1812 crit_exit_gd(gd); 1813 goto done; 1814 } 1815 crit_exit_gd(gd); 1816 } 1817 #endif 1818 m = NULL; 1819 1820 /* 1821 * CPU LOCALIZATION 1822 * 1823 * CPU localization algorithm. Break the page queues up by physical 1824 * id and core id (note that two cpu threads will have the same core 1825 * id, and core_id != gd_cpuid). 1826 * 1827 * This is nowhere near perfect, for example the last pindex in a 1828 * subgroup will overflow into the next cpu or package. But this 1829 * should get us good page reuse locality in heavy mixed loads. 1830 * 1831 * (may be executed before the APs are started, so other GDs might 1832 * not exist!) 1833 */ 1834 if (page_req & VM_ALLOC_CPU_SPEC) 1835 cpuid_local = VM_ALLOC_GETCPU(page_req); 1836 else 1837 cpuid_local = mycpu->gd_cpuid; 1838 1839 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 1840 1841 KKASSERT(page_req & 1842 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1843 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1844 1845 /* 1846 * Certain system threads (pageout daemon, buf_daemon's) are 1847 * allowed to eat deeper into the free page list. 1848 */ 1849 if (curthread->td_flags & TDF_SYSTHREAD) 1850 page_req |= VM_ALLOC_SYSTEM; 1851 1852 /* 1853 * Impose various limitations. Note that the v_free_reserved test 1854 * must match the opposite of vm_page_count_target() to avoid 1855 * livelocks, be careful. 1856 */ 1857 loop: 1858 gd = mycpu; 1859 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 1860 ((page_req & VM_ALLOC_INTERRUPT) && 1861 gd->gd_vmstats.v_free_count > 0) || 1862 ((page_req & VM_ALLOC_SYSTEM) && 1863 gd->gd_vmstats.v_cache_count == 0 && 1864 gd->gd_vmstats.v_free_count > 1865 gd->gd_vmstats.v_interrupt_free_min) 1866 ) { 1867 /* 1868 * The free queue has sufficient free pages to take one out. 1869 */ 1870 m = vm_page_select_free(pg_color); 1871 } else if (page_req & VM_ALLOC_NORMAL) { 1872 /* 1873 * Allocatable from the cache (non-interrupt only). On 1874 * success, we must free the page and try again, thus 1875 * ensuring that vmstats.v_*_free_min counters are replenished. 1876 */ 1877 #ifdef INVARIANTS 1878 if (curthread->td_preempted) { 1879 kprintf("vm_page_alloc(): warning, attempt to allocate" 1880 " cache page from preempting interrupt\n"); 1881 m = NULL; 1882 } else { 1883 m = vm_page_select_cache(pg_color); 1884 } 1885 #else 1886 m = vm_page_select_cache(pg_color); 1887 #endif 1888 /* 1889 * On success move the page into the free queue and loop. 1890 * 1891 * Only do this if we can safely acquire the vm_object lock, 1892 * because this is effectively a random page and the caller 1893 * might be holding the lock shared, we don't want to 1894 * deadlock. 1895 */ 1896 if (m != NULL) { 1897 KASSERT(m->dirty == 0, 1898 ("Found dirty cache page %p", m)); 1899 if ((obj = m->object) != NULL) { 1900 if (vm_object_hold_try(obj)) { 1901 vm_page_protect(m, VM_PROT_NONE); 1902 vm_page_free(m); 1903 /* m->object NULL here */ 1904 vm_object_drop(obj); 1905 } else { 1906 vm_page_deactivate(m); 1907 vm_page_wakeup(m); 1908 } 1909 } else { 1910 vm_page_protect(m, VM_PROT_NONE); 1911 vm_page_free(m); 1912 } 1913 goto loop; 1914 } 1915 1916 /* 1917 * On failure return NULL 1918 */ 1919 atomic_add_int(&vm_pageout_deficit, 1); 1920 pagedaemon_wakeup(); 1921 return (NULL); 1922 } else { 1923 /* 1924 * No pages available, wakeup the pageout daemon and give up. 1925 */ 1926 atomic_add_int(&vm_pageout_deficit, 1); 1927 pagedaemon_wakeup(); 1928 return (NULL); 1929 } 1930 1931 /* 1932 * v_free_count can race so loop if we don't find the expected 1933 * page. 1934 */ 1935 if (m == NULL) { 1936 vmstats_rollup(); 1937 goto loop; 1938 } 1939 1940 /* 1941 * Good page found. The page has already been busied for us and 1942 * removed from its queues. 1943 */ 1944 KASSERT(m->dirty == 0, 1945 ("vm_page_alloc: free/cache page %p was dirty", m)); 1946 KKASSERT(m->queue == PQ_NONE); 1947 1948 #if 0 1949 done: 1950 #endif 1951 /* 1952 * Initialize the structure, inheriting some flags but clearing 1953 * all the rest. The page has already been busied for us. 1954 */ 1955 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 1956 1957 KKASSERT(m->wire_count == 0); 1958 KKASSERT(m->busy == 0); 1959 m->act_count = 0; 1960 m->valid = 0; 1961 1962 /* 1963 * Caller must be holding the object lock (asserted by 1964 * vm_page_insert()). 1965 * 1966 * NOTE: Inserting a page here does not insert it into any pmaps 1967 * (which could cause us to block allocating memory). 1968 * 1969 * NOTE: If no object an unassociated page is allocated, m->pindex 1970 * can be used by the caller for any purpose. 1971 */ 1972 if (object) { 1973 if (vm_page_insert(m, object, pindex) == FALSE) { 1974 vm_page_free(m); 1975 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1976 panic("PAGE RACE %p[%ld]/%p", 1977 object, (long)pindex, m); 1978 m = NULL; 1979 } 1980 } else { 1981 m->pindex = pindex; 1982 } 1983 1984 /* 1985 * Don't wakeup too often - wakeup the pageout daemon when 1986 * we would be nearly out of memory. 1987 */ 1988 pagedaemon_wakeup(); 1989 1990 /* 1991 * A PG_BUSY page is returned. 1992 */ 1993 return (m); 1994 } 1995 1996 /* 1997 * Returns number of pages available in our DMA memory reserve 1998 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1999 */ 2000 vm_size_t 2001 vm_contig_avail_pages(void) 2002 { 2003 alist_blk_t blk; 2004 alist_blk_t count; 2005 alist_blk_t bfree; 2006 spin_lock(&vm_contig_spin); 2007 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 2008 spin_unlock(&vm_contig_spin); 2009 2010 return bfree; 2011 } 2012 2013 /* 2014 * Attempt to allocate contiguous physical memory with the specified 2015 * requirements. 2016 */ 2017 vm_page_t 2018 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 2019 unsigned long alignment, unsigned long boundary, 2020 unsigned long size, vm_memattr_t memattr) 2021 { 2022 alist_blk_t blk; 2023 vm_page_t m; 2024 int i; 2025 2026 alignment >>= PAGE_SHIFT; 2027 if (alignment == 0) 2028 alignment = 1; 2029 boundary >>= PAGE_SHIFT; 2030 if (boundary == 0) 2031 boundary = 1; 2032 size = (size + PAGE_MASK) >> PAGE_SHIFT; 2033 2034 spin_lock(&vm_contig_spin); 2035 blk = alist_alloc(&vm_contig_alist, 0, size); 2036 if (blk == ALIST_BLOCK_NONE) { 2037 spin_unlock(&vm_contig_spin); 2038 if (bootverbose) { 2039 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2040 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 2041 } 2042 return(NULL); 2043 } 2044 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2045 alist_free(&vm_contig_alist, blk, size); 2046 spin_unlock(&vm_contig_spin); 2047 if (bootverbose) { 2048 kprintf("vm_page_alloc_contig: %ldk high " 2049 "%016jx failed\n", 2050 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 2051 (intmax_t)high); 2052 } 2053 return(NULL); 2054 } 2055 spin_unlock(&vm_contig_spin); 2056 if (vm_contig_verbose) { 2057 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 2058 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 2059 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 2060 } 2061 2062 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2063 if (memattr != VM_MEMATTR_DEFAULT) 2064 for (i = 0;i < size;i++) 2065 pmap_page_set_memattr(&m[i], memattr); 2066 return m; 2067 } 2068 2069 /* 2070 * Free contiguously allocated pages. The pages will be wired but not busy. 2071 * When freeing to the alist we leave them wired and not busy. 2072 */ 2073 void 2074 vm_page_free_contig(vm_page_t m, unsigned long size) 2075 { 2076 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2077 vm_pindex_t start = pa >> PAGE_SHIFT; 2078 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2079 2080 if (vm_contig_verbose) { 2081 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2082 (intmax_t)pa, size / 1024); 2083 } 2084 if (pa < vm_low_phys_reserved) { 2085 KKASSERT(pa + size <= vm_low_phys_reserved); 2086 spin_lock(&vm_contig_spin); 2087 alist_free(&vm_contig_alist, start, pages); 2088 spin_unlock(&vm_contig_spin); 2089 } else { 2090 while (pages) { 2091 vm_page_busy_wait(m, FALSE, "cpgfr"); 2092 vm_page_unwire(m, 0); 2093 vm_page_free(m); 2094 --pages; 2095 ++m; 2096 } 2097 2098 } 2099 } 2100 2101 2102 /* 2103 * Wait for sufficient free memory for nominal heavy memory use kernel 2104 * operations. 2105 * 2106 * WARNING! Be sure never to call this in any vm_pageout code path, which 2107 * will trivially deadlock the system. 2108 */ 2109 void 2110 vm_wait_nominal(void) 2111 { 2112 while (vm_page_count_min(0)) 2113 vm_wait(0); 2114 } 2115 2116 /* 2117 * Test if vm_wait_nominal() would block. 2118 */ 2119 int 2120 vm_test_nominal(void) 2121 { 2122 if (vm_page_count_min(0)) 2123 return(1); 2124 return(0); 2125 } 2126 2127 /* 2128 * Block until free pages are available for allocation, called in various 2129 * places before memory allocations. 2130 * 2131 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2132 * more generous then that. 2133 */ 2134 void 2135 vm_wait(int timo) 2136 { 2137 /* 2138 * never wait forever 2139 */ 2140 if (timo == 0) 2141 timo = hz; 2142 lwkt_gettoken(&vm_token); 2143 2144 if (curthread == pagethread) { 2145 /* 2146 * The pageout daemon itself needs pages, this is bad. 2147 */ 2148 if (vm_page_count_min(0)) { 2149 vm_pageout_pages_needed = 1; 2150 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2151 } 2152 } else { 2153 /* 2154 * Wakeup the pageout daemon if necessary and wait. 2155 * 2156 * Do not wait indefinitely for the target to be reached, 2157 * as load might prevent it from being reached any time soon. 2158 * But wait a little to try to slow down page allocations 2159 * and to give more important threads (the pagedaemon) 2160 * allocation priority. 2161 */ 2162 if (vm_page_count_target()) { 2163 if (vm_pages_needed == 0) { 2164 vm_pages_needed = 1; 2165 wakeup(&vm_pages_needed); 2166 } 2167 ++vm_pages_waiting; /* SMP race ok */ 2168 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2169 } 2170 } 2171 lwkt_reltoken(&vm_token); 2172 } 2173 2174 /* 2175 * Block until free pages are available for allocation 2176 * 2177 * Called only from vm_fault so that processes page faulting can be 2178 * easily tracked. 2179 */ 2180 void 2181 vm_wait_pfault(void) 2182 { 2183 /* 2184 * Wakeup the pageout daemon if necessary and wait. 2185 * 2186 * Do not wait indefinitely for the target to be reached, 2187 * as load might prevent it from being reached any time soon. 2188 * But wait a little to try to slow down page allocations 2189 * and to give more important threads (the pagedaemon) 2190 * allocation priority. 2191 */ 2192 if (vm_page_count_min(0)) { 2193 lwkt_gettoken(&vm_token); 2194 while (vm_page_count_severe()) { 2195 if (vm_page_count_target()) { 2196 thread_t td; 2197 2198 if (vm_pages_needed == 0) { 2199 vm_pages_needed = 1; 2200 wakeup(&vm_pages_needed); 2201 } 2202 ++vm_pages_waiting; /* SMP race ok */ 2203 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2204 2205 /* 2206 * Do not stay stuck in the loop if the system is trying 2207 * to kill the process. 2208 */ 2209 td = curthread; 2210 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2211 break; 2212 } 2213 } 2214 lwkt_reltoken(&vm_token); 2215 } 2216 } 2217 2218 /* 2219 * Put the specified page on the active list (if appropriate). Ensure 2220 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2221 * 2222 * The caller should be holding the page busied ? XXX 2223 * This routine may not block. 2224 */ 2225 void 2226 vm_page_activate(vm_page_t m) 2227 { 2228 u_short oqueue; 2229 2230 vm_page_spin_lock(m); 2231 if (m->queue - m->pc != PQ_ACTIVE) { 2232 _vm_page_queue_spin_lock(m); 2233 oqueue = _vm_page_rem_queue_spinlocked(m); 2234 /* page is left spinlocked, queue is unlocked */ 2235 2236 if (oqueue == PQ_CACHE) 2237 mycpu->gd_cnt.v_reactivated++; 2238 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2239 if (m->act_count < ACT_INIT) 2240 m->act_count = ACT_INIT; 2241 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2242 } 2243 _vm_page_and_queue_spin_unlock(m); 2244 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2245 pagedaemon_wakeup(); 2246 } else { 2247 if (m->act_count < ACT_INIT) 2248 m->act_count = ACT_INIT; 2249 vm_page_spin_unlock(m); 2250 } 2251 } 2252 2253 /* 2254 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2255 * routine is called when a page has been added to the cache or free 2256 * queues. 2257 * 2258 * This routine may not block. 2259 */ 2260 static __inline void 2261 vm_page_free_wakeup(void) 2262 { 2263 globaldata_t gd = mycpu; 2264 2265 /* 2266 * If the pageout daemon itself needs pages, then tell it that 2267 * there are some free. 2268 */ 2269 if (vm_pageout_pages_needed && 2270 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2271 gd->gd_vmstats.v_pageout_free_min 2272 ) { 2273 vm_pageout_pages_needed = 0; 2274 wakeup(&vm_pageout_pages_needed); 2275 } 2276 2277 /* 2278 * Wakeup processes that are waiting on memory. 2279 * 2280 * Generally speaking we want to wakeup stuck processes as soon as 2281 * possible. !vm_page_count_min(0) is the absolute minimum point 2282 * where we can do this. Wait a bit longer to reduce degenerate 2283 * re-blocking (vm_page_free_hysteresis). The target check is just 2284 * to make sure the min-check w/hysteresis does not exceed the 2285 * normal target. 2286 */ 2287 if (vm_pages_waiting) { 2288 if (!vm_page_count_min(vm_page_free_hysteresis) || 2289 !vm_page_count_target()) { 2290 vm_pages_waiting = 0; 2291 wakeup(&vmstats.v_free_count); 2292 ++mycpu->gd_cnt.v_ppwakeups; 2293 } 2294 #if 0 2295 if (!vm_page_count_target()) { 2296 /* 2297 * Plenty of pages are free, wakeup everyone. 2298 */ 2299 vm_pages_waiting = 0; 2300 wakeup(&vmstats.v_free_count); 2301 ++mycpu->gd_cnt.v_ppwakeups; 2302 } else if (!vm_page_count_min(0)) { 2303 /* 2304 * Some pages are free, wakeup someone. 2305 */ 2306 int wcount = vm_pages_waiting; 2307 if (wcount > 0) 2308 --wcount; 2309 vm_pages_waiting = wcount; 2310 wakeup_one(&vmstats.v_free_count); 2311 ++mycpu->gd_cnt.v_ppwakeups; 2312 } 2313 #endif 2314 } 2315 } 2316 2317 /* 2318 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2319 * it from its VM object. 2320 * 2321 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on 2322 * return (the page will have been freed). 2323 */ 2324 void 2325 vm_page_free_toq(vm_page_t m) 2326 { 2327 mycpu->gd_cnt.v_tfree++; 2328 KKASSERT((m->flags & PG_MAPPED) == 0); 2329 KKASSERT(m->flags & PG_BUSY); 2330 2331 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 2332 kprintf("vm_page_free: pindex(%lu), busy(%d), " 2333 "PG_BUSY(%d), hold(%d)\n", 2334 (u_long)m->pindex, m->busy, 2335 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count); 2336 if ((m->queue - m->pc) == PQ_FREE) 2337 panic("vm_page_free: freeing free page"); 2338 else 2339 panic("vm_page_free: freeing busy page"); 2340 } 2341 2342 /* 2343 * Remove from object, spinlock the page and its queues and 2344 * remove from any queue. No queue spinlock will be held 2345 * after this section (because the page was removed from any 2346 * queue). 2347 */ 2348 vm_page_remove(m); 2349 vm_page_and_queue_spin_lock(m); 2350 _vm_page_rem_queue_spinlocked(m); 2351 2352 /* 2353 * No further management of fictitious pages occurs beyond object 2354 * and queue removal. 2355 */ 2356 if ((m->flags & PG_FICTITIOUS) != 0) { 2357 vm_page_spin_unlock(m); 2358 vm_page_wakeup(m); 2359 return; 2360 } 2361 2362 m->valid = 0; 2363 vm_page_undirty(m); 2364 2365 if (m->wire_count != 0) { 2366 if (m->wire_count > 1) { 2367 panic( 2368 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2369 m->wire_count, (long)m->pindex); 2370 } 2371 panic("vm_page_free: freeing wired page"); 2372 } 2373 2374 /* 2375 * Clear the UNMANAGED flag when freeing an unmanaged page. 2376 * Clear the NEED_COMMIT flag 2377 */ 2378 if (m->flags & PG_UNMANAGED) 2379 vm_page_flag_clear(m, PG_UNMANAGED); 2380 if (m->flags & PG_NEED_COMMIT) 2381 vm_page_flag_clear(m, PG_NEED_COMMIT); 2382 2383 if (m->hold_count != 0) { 2384 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2385 } else { 2386 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2387 } 2388 2389 /* 2390 * This sequence allows us to clear PG_BUSY while still holding 2391 * its spin lock, which reduces contention vs allocators. We 2392 * must not leave the queue locked or _vm_page_wakeup() may 2393 * deadlock. 2394 */ 2395 _vm_page_queue_spin_unlock(m); 2396 if (_vm_page_wakeup(m)) { 2397 vm_page_spin_unlock(m); 2398 wakeup(m); 2399 } else { 2400 vm_page_spin_unlock(m); 2401 } 2402 vm_page_free_wakeup(); 2403 } 2404 2405 /* 2406 * vm_page_unmanage() 2407 * 2408 * Prevent PV management from being done on the page. The page is 2409 * removed from the paging queues as if it were wired, and as a 2410 * consequence of no longer being managed the pageout daemon will not 2411 * touch it (since there is no way to locate the pte mappings for the 2412 * page). madvise() calls that mess with the pmap will also no longer 2413 * operate on the page. 2414 * 2415 * Beyond that the page is still reasonably 'normal'. Freeing the page 2416 * will clear the flag. 2417 * 2418 * This routine is used by OBJT_PHYS objects - objects using unswappable 2419 * physical memory as backing store rather then swap-backed memory and 2420 * will eventually be extended to support 4MB unmanaged physical 2421 * mappings. 2422 * 2423 * Caller must be holding the page busy. 2424 */ 2425 void 2426 vm_page_unmanage(vm_page_t m) 2427 { 2428 KKASSERT(m->flags & PG_BUSY); 2429 if ((m->flags & PG_UNMANAGED) == 0) { 2430 if (m->wire_count == 0) 2431 vm_page_unqueue(m); 2432 } 2433 vm_page_flag_set(m, PG_UNMANAGED); 2434 } 2435 2436 /* 2437 * Mark this page as wired down by yet another map, removing it from 2438 * paging queues as necessary. 2439 * 2440 * Caller must be holding the page busy. 2441 */ 2442 void 2443 vm_page_wire(vm_page_t m) 2444 { 2445 /* 2446 * Only bump the wire statistics if the page is not already wired, 2447 * and only unqueue the page if it is on some queue (if it is unmanaged 2448 * it is already off the queues). Don't do anything with fictitious 2449 * pages because they are always wired. 2450 */ 2451 KKASSERT(m->flags & PG_BUSY); 2452 if ((m->flags & PG_FICTITIOUS) == 0) { 2453 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2454 if ((m->flags & PG_UNMANAGED) == 0) 2455 vm_page_unqueue(m); 2456 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2457 } 2458 KASSERT(m->wire_count != 0, 2459 ("vm_page_wire: wire_count overflow m=%p", m)); 2460 } 2461 } 2462 2463 /* 2464 * Release one wiring of this page, potentially enabling it to be paged again. 2465 * 2466 * Many pages placed on the inactive queue should actually go 2467 * into the cache, but it is difficult to figure out which. What 2468 * we do instead, if the inactive target is well met, is to put 2469 * clean pages at the head of the inactive queue instead of the tail. 2470 * This will cause them to be moved to the cache more quickly and 2471 * if not actively re-referenced, freed more quickly. If we just 2472 * stick these pages at the end of the inactive queue, heavy filesystem 2473 * meta-data accesses can cause an unnecessary paging load on memory bound 2474 * processes. This optimization causes one-time-use metadata to be 2475 * reused more quickly. 2476 * 2477 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2478 * the inactive queue. This helps the pageout daemon determine memory 2479 * pressure and act on out-of-memory situations more quickly. 2480 * 2481 * BUT, if we are in a low-memory situation we have no choice but to 2482 * put clean pages on the cache queue. 2483 * 2484 * A number of routines use vm_page_unwire() to guarantee that the page 2485 * will go into either the inactive or active queues, and will NEVER 2486 * be placed in the cache - for example, just after dirtying a page. 2487 * dirty pages in the cache are not allowed. 2488 * 2489 * This routine may not block. 2490 */ 2491 void 2492 vm_page_unwire(vm_page_t m, int activate) 2493 { 2494 KKASSERT(m->flags & PG_BUSY); 2495 if (m->flags & PG_FICTITIOUS) { 2496 /* do nothing */ 2497 } else if (m->wire_count <= 0) { 2498 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2499 } else { 2500 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2501 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1); 2502 if (m->flags & PG_UNMANAGED) { 2503 ; 2504 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2505 vm_page_spin_lock(m); 2506 _vm_page_add_queue_spinlocked(m, 2507 PQ_ACTIVE + m->pc, 0); 2508 _vm_page_and_queue_spin_unlock(m); 2509 } else { 2510 vm_page_spin_lock(m); 2511 vm_page_flag_clear(m, PG_WINATCFLS); 2512 _vm_page_add_queue_spinlocked(m, 2513 PQ_INACTIVE + m->pc, 0); 2514 ++vm_swapcache_inactive_heuristic; 2515 _vm_page_and_queue_spin_unlock(m); 2516 } 2517 } 2518 } 2519 } 2520 2521 /* 2522 * Move the specified page to the inactive queue. If the page has 2523 * any associated swap, the swap is deallocated. 2524 * 2525 * Normally athead is 0 resulting in LRU operation. athead is set 2526 * to 1 if we want this page to be 'as if it were placed in the cache', 2527 * except without unmapping it from the process address space. 2528 * 2529 * vm_page's spinlock must be held on entry and will remain held on return. 2530 * This routine may not block. 2531 */ 2532 static void 2533 _vm_page_deactivate_locked(vm_page_t m, int athead) 2534 { 2535 u_short oqueue; 2536 2537 /* 2538 * Ignore if already inactive. 2539 */ 2540 if (m->queue - m->pc == PQ_INACTIVE) 2541 return; 2542 _vm_page_queue_spin_lock(m); 2543 oqueue = _vm_page_rem_queue_spinlocked(m); 2544 2545 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2546 if (oqueue == PQ_CACHE) 2547 mycpu->gd_cnt.v_reactivated++; 2548 vm_page_flag_clear(m, PG_WINATCFLS); 2549 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2550 if (athead == 0) 2551 ++vm_swapcache_inactive_heuristic; 2552 } 2553 /* NOTE: PQ_NONE if condition not taken */ 2554 _vm_page_queue_spin_unlock(m); 2555 /* leaves vm_page spinlocked */ 2556 } 2557 2558 /* 2559 * Attempt to deactivate a page. 2560 * 2561 * No requirements. 2562 */ 2563 void 2564 vm_page_deactivate(vm_page_t m) 2565 { 2566 vm_page_spin_lock(m); 2567 _vm_page_deactivate_locked(m, 0); 2568 vm_page_spin_unlock(m); 2569 } 2570 2571 void 2572 vm_page_deactivate_locked(vm_page_t m) 2573 { 2574 _vm_page_deactivate_locked(m, 0); 2575 } 2576 2577 /* 2578 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2579 * 2580 * This function returns non-zero if it successfully moved the page to 2581 * PQ_CACHE. 2582 * 2583 * This function unconditionally unbusies the page on return. 2584 */ 2585 int 2586 vm_page_try_to_cache(vm_page_t m) 2587 { 2588 vm_page_spin_lock(m); 2589 if (m->dirty || m->hold_count || m->wire_count || 2590 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2591 if (_vm_page_wakeup(m)) { 2592 vm_page_spin_unlock(m); 2593 wakeup(m); 2594 } else { 2595 vm_page_spin_unlock(m); 2596 } 2597 return(0); 2598 } 2599 vm_page_spin_unlock(m); 2600 2601 /* 2602 * Page busied by us and no longer spinlocked. Dirty pages cannot 2603 * be moved to the cache. 2604 */ 2605 vm_page_test_dirty(m); 2606 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2607 vm_page_wakeup(m); 2608 return(0); 2609 } 2610 vm_page_cache(m); 2611 return(1); 2612 } 2613 2614 /* 2615 * Attempt to free the page. If we cannot free it, we do nothing. 2616 * 1 is returned on success, 0 on failure. 2617 * 2618 * No requirements. 2619 */ 2620 int 2621 vm_page_try_to_free(vm_page_t m) 2622 { 2623 vm_page_spin_lock(m); 2624 if (vm_page_busy_try(m, TRUE)) { 2625 vm_page_spin_unlock(m); 2626 return(0); 2627 } 2628 2629 /* 2630 * The page can be in any state, including already being on the free 2631 * queue. Check to see if it really can be freed. 2632 */ 2633 if (m->dirty || /* can't free if it is dirty */ 2634 m->hold_count || /* or held (XXX may be wrong) */ 2635 m->wire_count || /* or wired */ 2636 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2637 PG_NEED_COMMIT)) || /* or needs a commit */ 2638 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2639 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2640 if (_vm_page_wakeup(m)) { 2641 vm_page_spin_unlock(m); 2642 wakeup(m); 2643 } else { 2644 vm_page_spin_unlock(m); 2645 } 2646 return(0); 2647 } 2648 vm_page_spin_unlock(m); 2649 2650 /* 2651 * We can probably free the page. 2652 * 2653 * Page busied by us and no longer spinlocked. Dirty pages will 2654 * not be freed by this function. We have to re-test the 2655 * dirty bit after cleaning out the pmaps. 2656 */ 2657 vm_page_test_dirty(m); 2658 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2659 vm_page_wakeup(m); 2660 return(0); 2661 } 2662 vm_page_protect(m, VM_PROT_NONE); 2663 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2664 vm_page_wakeup(m); 2665 return(0); 2666 } 2667 vm_page_free(m); 2668 return(1); 2669 } 2670 2671 /* 2672 * vm_page_cache 2673 * 2674 * Put the specified page onto the page cache queue (if appropriate). 2675 * 2676 * The page must be busy, and this routine will release the busy and 2677 * possibly even free the page. 2678 */ 2679 void 2680 vm_page_cache(vm_page_t m) 2681 { 2682 /* 2683 * Not suitable for the cache 2684 */ 2685 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2686 m->busy || m->wire_count || m->hold_count) { 2687 vm_page_wakeup(m); 2688 return; 2689 } 2690 2691 /* 2692 * Already in the cache (and thus not mapped) 2693 */ 2694 if ((m->queue - m->pc) == PQ_CACHE) { 2695 KKASSERT((m->flags & PG_MAPPED) == 0); 2696 vm_page_wakeup(m); 2697 return; 2698 } 2699 2700 /* 2701 * Caller is required to test m->dirty, but note that the act of 2702 * removing the page from its maps can cause it to become dirty 2703 * on an SMP system due to another cpu running in usermode. 2704 */ 2705 if (m->dirty) { 2706 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2707 (long)m->pindex); 2708 } 2709 2710 /* 2711 * Remove all pmaps and indicate that the page is not 2712 * writeable or mapped. Our vm_page_protect() call may 2713 * have blocked (especially w/ VM_PROT_NONE), so recheck 2714 * everything. 2715 */ 2716 vm_page_protect(m, VM_PROT_NONE); 2717 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2718 m->busy || m->wire_count || m->hold_count) { 2719 vm_page_wakeup(m); 2720 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2721 vm_page_deactivate(m); 2722 vm_page_wakeup(m); 2723 } else { 2724 _vm_page_and_queue_spin_lock(m); 2725 _vm_page_rem_queue_spinlocked(m); 2726 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2727 _vm_page_queue_spin_unlock(m); 2728 if (_vm_page_wakeup(m)) { 2729 vm_page_spin_unlock(m); 2730 wakeup(m); 2731 } else { 2732 vm_page_spin_unlock(m); 2733 } 2734 vm_page_free_wakeup(); 2735 } 2736 } 2737 2738 /* 2739 * vm_page_dontneed() 2740 * 2741 * Cache, deactivate, or do nothing as appropriate. This routine 2742 * is typically used by madvise() MADV_DONTNEED. 2743 * 2744 * Generally speaking we want to move the page into the cache so 2745 * it gets reused quickly. However, this can result in a silly syndrome 2746 * due to the page recycling too quickly. Small objects will not be 2747 * fully cached. On the otherhand, if we move the page to the inactive 2748 * queue we wind up with a problem whereby very large objects 2749 * unnecessarily blow away our inactive and cache queues. 2750 * 2751 * The solution is to move the pages based on a fixed weighting. We 2752 * either leave them alone, deactivate them, or move them to the cache, 2753 * where moving them to the cache has the highest weighting. 2754 * By forcing some pages into other queues we eventually force the 2755 * system to balance the queues, potentially recovering other unrelated 2756 * space from active. The idea is to not force this to happen too 2757 * often. 2758 * 2759 * The page must be busied. 2760 */ 2761 void 2762 vm_page_dontneed(vm_page_t m) 2763 { 2764 static int dnweight; 2765 int dnw; 2766 int head; 2767 2768 dnw = ++dnweight; 2769 2770 /* 2771 * occassionally leave the page alone 2772 */ 2773 if ((dnw & 0x01F0) == 0 || 2774 m->queue - m->pc == PQ_INACTIVE || 2775 m->queue - m->pc == PQ_CACHE 2776 ) { 2777 if (m->act_count >= ACT_INIT) 2778 --m->act_count; 2779 return; 2780 } 2781 2782 /* 2783 * If vm_page_dontneed() is inactivating a page, it must clear 2784 * the referenced flag; otherwise the pagedaemon will see references 2785 * on the page in the inactive queue and reactivate it. Until the 2786 * page can move to the cache queue, madvise's job is not done. 2787 */ 2788 vm_page_flag_clear(m, PG_REFERENCED); 2789 pmap_clear_reference(m); 2790 2791 if (m->dirty == 0) 2792 vm_page_test_dirty(m); 2793 2794 if (m->dirty || (dnw & 0x0070) == 0) { 2795 /* 2796 * Deactivate the page 3 times out of 32. 2797 */ 2798 head = 0; 2799 } else { 2800 /* 2801 * Cache the page 28 times out of every 32. Note that 2802 * the page is deactivated instead of cached, but placed 2803 * at the head of the queue instead of the tail. 2804 */ 2805 head = 1; 2806 } 2807 vm_page_spin_lock(m); 2808 _vm_page_deactivate_locked(m, head); 2809 vm_page_spin_unlock(m); 2810 } 2811 2812 /* 2813 * These routines manipulate the 'soft busy' count for a page. A soft busy 2814 * is almost like PG_BUSY except that it allows certain compatible operations 2815 * to occur on the page while it is busy. For example, a page undergoing a 2816 * write can still be mapped read-only. 2817 * 2818 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only 2819 * adjusted while the vm_page is PG_BUSY so the flash will occur when the 2820 * busy bit is cleared. 2821 * 2822 * The caller must hold the page BUSY when making these two calls. 2823 */ 2824 void 2825 vm_page_io_start(vm_page_t m) 2826 { 2827 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!")); 2828 atomic_add_char(&m->busy, 1); 2829 vm_page_flag_set(m, PG_SBUSY); 2830 } 2831 2832 void 2833 vm_page_io_finish(vm_page_t m) 2834 { 2835 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!")); 2836 atomic_subtract_char(&m->busy, 1); 2837 if (m->busy == 0) 2838 vm_page_flag_clear(m, PG_SBUSY); 2839 } 2840 2841 /* 2842 * Indicate that a clean VM page requires a filesystem commit and cannot 2843 * be reused. Used by tmpfs. 2844 */ 2845 void 2846 vm_page_need_commit(vm_page_t m) 2847 { 2848 vm_page_flag_set(m, PG_NEED_COMMIT); 2849 vm_object_set_writeable_dirty(m->object); 2850 } 2851 2852 void 2853 vm_page_clear_commit(vm_page_t m) 2854 { 2855 vm_page_flag_clear(m, PG_NEED_COMMIT); 2856 } 2857 2858 /* 2859 * Grab a page, blocking if it is busy and allocating a page if necessary. 2860 * A busy page is returned or NULL. The page may or may not be valid and 2861 * might not be on a queue (the caller is responsible for the disposition of 2862 * the page). 2863 * 2864 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2865 * page will be zero'd and marked valid. 2866 * 2867 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2868 * valid even if it already exists. 2869 * 2870 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2871 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2872 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2873 * 2874 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2875 * always returned if we had blocked. 2876 * 2877 * This routine may not be called from an interrupt. 2878 * 2879 * No other requirements. 2880 */ 2881 vm_page_t 2882 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2883 { 2884 vm_page_t m; 2885 int error; 2886 int shared = 1; 2887 2888 KKASSERT(allocflags & 2889 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2890 vm_object_hold_shared(object); 2891 for (;;) { 2892 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2893 if (error) { 2894 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2895 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2896 m = NULL; 2897 break; 2898 } 2899 /* retry */ 2900 } else if (m == NULL) { 2901 if (shared) { 2902 vm_object_upgrade(object); 2903 shared = 0; 2904 } 2905 if (allocflags & VM_ALLOC_RETRY) 2906 allocflags |= VM_ALLOC_NULL_OK; 2907 m = vm_page_alloc(object, pindex, 2908 allocflags & ~VM_ALLOC_RETRY); 2909 if (m) 2910 break; 2911 vm_wait(0); 2912 if ((allocflags & VM_ALLOC_RETRY) == 0) 2913 goto failed; 2914 } else { 2915 /* m found */ 2916 break; 2917 } 2918 } 2919 2920 /* 2921 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2922 * 2923 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2924 * valid even if already valid. 2925 * 2926 * NOTE! We have removed all of the PG_ZERO optimizations and also 2927 * removed the idle zeroing code. These optimizations actually 2928 * slow things down on modern cpus because the zerod area is 2929 * likely uncached, placing a memory-access burden on the 2930 * accesors taking the fault. 2931 * 2932 * By always zeroing the page in-line with the fault, no 2933 * dynamic ram reads are needed and the caches are hot, ready 2934 * for userland to access the memory. 2935 */ 2936 if (m->valid == 0) { 2937 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2938 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2939 m->valid = VM_PAGE_BITS_ALL; 2940 } 2941 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2942 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2943 m->valid = VM_PAGE_BITS_ALL; 2944 } 2945 failed: 2946 vm_object_drop(object); 2947 return(m); 2948 } 2949 2950 /* 2951 * Mapping function for valid bits or for dirty bits in 2952 * a page. May not block. 2953 * 2954 * Inputs are required to range within a page. 2955 * 2956 * No requirements. 2957 * Non blocking. 2958 */ 2959 int 2960 vm_page_bits(int base, int size) 2961 { 2962 int first_bit; 2963 int last_bit; 2964 2965 KASSERT( 2966 base + size <= PAGE_SIZE, 2967 ("vm_page_bits: illegal base/size %d/%d", base, size) 2968 ); 2969 2970 if (size == 0) /* handle degenerate case */ 2971 return(0); 2972 2973 first_bit = base >> DEV_BSHIFT; 2974 last_bit = (base + size - 1) >> DEV_BSHIFT; 2975 2976 return ((2 << last_bit) - (1 << first_bit)); 2977 } 2978 2979 /* 2980 * Sets portions of a page valid and clean. The arguments are expected 2981 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2982 * of any partial chunks touched by the range. The invalid portion of 2983 * such chunks will be zero'd. 2984 * 2985 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2986 * align base to DEV_BSIZE so as not to mark clean a partially 2987 * truncated device block. Otherwise the dirty page status might be 2988 * lost. 2989 * 2990 * This routine may not block. 2991 * 2992 * (base + size) must be less then or equal to PAGE_SIZE. 2993 */ 2994 static void 2995 _vm_page_zero_valid(vm_page_t m, int base, int size) 2996 { 2997 int frag; 2998 int endoff; 2999 3000 if (size == 0) /* handle degenerate case */ 3001 return; 3002 3003 /* 3004 * If the base is not DEV_BSIZE aligned and the valid 3005 * bit is clear, we have to zero out a portion of the 3006 * first block. 3007 */ 3008 3009 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3010 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3011 ) { 3012 pmap_zero_page_area( 3013 VM_PAGE_TO_PHYS(m), 3014 frag, 3015 base - frag 3016 ); 3017 } 3018 3019 /* 3020 * If the ending offset is not DEV_BSIZE aligned and the 3021 * valid bit is clear, we have to zero out a portion of 3022 * the last block. 3023 */ 3024 3025 endoff = base + size; 3026 3027 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3028 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3029 ) { 3030 pmap_zero_page_area( 3031 VM_PAGE_TO_PHYS(m), 3032 endoff, 3033 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3034 ); 3035 } 3036 } 3037 3038 /* 3039 * Set valid, clear dirty bits. If validating the entire 3040 * page we can safely clear the pmap modify bit. We also 3041 * use this opportunity to clear the PG_NOSYNC flag. If a process 3042 * takes a write fault on a MAP_NOSYNC memory area the flag will 3043 * be set again. 3044 * 3045 * We set valid bits inclusive of any overlap, but we can only 3046 * clear dirty bits for DEV_BSIZE chunks that are fully within 3047 * the range. 3048 * 3049 * Page must be busied? 3050 * No other requirements. 3051 */ 3052 void 3053 vm_page_set_valid(vm_page_t m, int base, int size) 3054 { 3055 _vm_page_zero_valid(m, base, size); 3056 m->valid |= vm_page_bits(base, size); 3057 } 3058 3059 3060 /* 3061 * Set valid bits and clear dirty bits. 3062 * 3063 * Page must be busied by caller. 3064 * 3065 * NOTE: This function does not clear the pmap modified bit. 3066 * Also note that e.g. NFS may use a byte-granular base 3067 * and size. 3068 * 3069 * No other requirements. 3070 */ 3071 void 3072 vm_page_set_validclean(vm_page_t m, int base, int size) 3073 { 3074 int pagebits; 3075 3076 _vm_page_zero_valid(m, base, size); 3077 pagebits = vm_page_bits(base, size); 3078 m->valid |= pagebits; 3079 m->dirty &= ~pagebits; 3080 if (base == 0 && size == PAGE_SIZE) { 3081 /*pmap_clear_modify(m);*/ 3082 vm_page_flag_clear(m, PG_NOSYNC); 3083 } 3084 } 3085 3086 /* 3087 * Set valid & dirty. Used by buwrite() 3088 * 3089 * Page must be busied by caller. 3090 */ 3091 void 3092 vm_page_set_validdirty(vm_page_t m, int base, int size) 3093 { 3094 int pagebits; 3095 3096 pagebits = vm_page_bits(base, size); 3097 m->valid |= pagebits; 3098 m->dirty |= pagebits; 3099 if (m->object) 3100 vm_object_set_writeable_dirty(m->object); 3101 } 3102 3103 /* 3104 * Clear dirty bits. 3105 * 3106 * NOTE: This function does not clear the pmap modified bit. 3107 * Also note that e.g. NFS may use a byte-granular base 3108 * and size. 3109 * 3110 * Page must be busied? 3111 * No other requirements. 3112 */ 3113 void 3114 vm_page_clear_dirty(vm_page_t m, int base, int size) 3115 { 3116 m->dirty &= ~vm_page_bits(base, size); 3117 if (base == 0 && size == PAGE_SIZE) { 3118 /*pmap_clear_modify(m);*/ 3119 vm_page_flag_clear(m, PG_NOSYNC); 3120 } 3121 } 3122 3123 /* 3124 * Make the page all-dirty. 3125 * 3126 * Also make sure the related object and vnode reflect the fact that the 3127 * object may now contain a dirty page. 3128 * 3129 * Page must be busied? 3130 * No other requirements. 3131 */ 3132 void 3133 vm_page_dirty(vm_page_t m) 3134 { 3135 #ifdef INVARIANTS 3136 int pqtype = m->queue - m->pc; 3137 #endif 3138 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3139 ("vm_page_dirty: page in free/cache queue!")); 3140 if (m->dirty != VM_PAGE_BITS_ALL) { 3141 m->dirty = VM_PAGE_BITS_ALL; 3142 if (m->object) 3143 vm_object_set_writeable_dirty(m->object); 3144 } 3145 } 3146 3147 /* 3148 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3149 * valid and dirty bits for the effected areas are cleared. 3150 * 3151 * Page must be busied? 3152 * Does not block. 3153 * No other requirements. 3154 */ 3155 void 3156 vm_page_set_invalid(vm_page_t m, int base, int size) 3157 { 3158 int bits; 3159 3160 bits = vm_page_bits(base, size); 3161 m->valid &= ~bits; 3162 m->dirty &= ~bits; 3163 atomic_add_int(&m->object->generation, 1); 3164 } 3165 3166 /* 3167 * The kernel assumes that the invalid portions of a page contain 3168 * garbage, but such pages can be mapped into memory by user code. 3169 * When this occurs, we must zero out the non-valid portions of the 3170 * page so user code sees what it expects. 3171 * 3172 * Pages are most often semi-valid when the end of a file is mapped 3173 * into memory and the file's size is not page aligned. 3174 * 3175 * Page must be busied? 3176 * No other requirements. 3177 */ 3178 void 3179 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3180 { 3181 int b; 3182 int i; 3183 3184 /* 3185 * Scan the valid bits looking for invalid sections that 3186 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3187 * valid bit may be set ) have already been zerod by 3188 * vm_page_set_validclean(). 3189 */ 3190 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3191 if (i == (PAGE_SIZE / DEV_BSIZE) || 3192 (m->valid & (1 << i)) 3193 ) { 3194 if (i > b) { 3195 pmap_zero_page_area( 3196 VM_PAGE_TO_PHYS(m), 3197 b << DEV_BSHIFT, 3198 (i - b) << DEV_BSHIFT 3199 ); 3200 } 3201 b = i + 1; 3202 } 3203 } 3204 3205 /* 3206 * setvalid is TRUE when we can safely set the zero'd areas 3207 * as being valid. We can do this if there are no cache consistency 3208 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3209 */ 3210 if (setvalid) 3211 m->valid = VM_PAGE_BITS_ALL; 3212 } 3213 3214 /* 3215 * Is a (partial) page valid? Note that the case where size == 0 3216 * will return FALSE in the degenerate case where the page is entirely 3217 * invalid, and TRUE otherwise. 3218 * 3219 * Does not block. 3220 * No other requirements. 3221 */ 3222 int 3223 vm_page_is_valid(vm_page_t m, int base, int size) 3224 { 3225 int bits = vm_page_bits(base, size); 3226 3227 if (m->valid && ((m->valid & bits) == bits)) 3228 return 1; 3229 else 3230 return 0; 3231 } 3232 3233 /* 3234 * update dirty bits from pmap/mmu. May not block. 3235 * 3236 * Caller must hold the page busy 3237 */ 3238 void 3239 vm_page_test_dirty(vm_page_t m) 3240 { 3241 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3242 vm_page_dirty(m); 3243 } 3244 } 3245 3246 /* 3247 * Register an action, associating it with its vm_page 3248 */ 3249 void 3250 vm_page_register_action(vm_page_action_t action, vm_page_event_t event) 3251 { 3252 struct vm_page_action_hash *hash; 3253 int hv; 3254 3255 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3256 hash = &action_hash[hv]; 3257 3258 lockmgr(&hash->lk, LK_EXCLUSIVE); 3259 vm_page_flag_set(action->m, PG_ACTIONLIST); 3260 action->event = event; 3261 LIST_INSERT_HEAD(&hash->list, action, entry); 3262 lockmgr(&hash->lk, LK_RELEASE); 3263 } 3264 3265 /* 3266 * Unregister an action, disassociating it from its related vm_page 3267 */ 3268 void 3269 vm_page_unregister_action(vm_page_action_t action) 3270 { 3271 struct vm_page_action_hash *hash; 3272 int hv; 3273 3274 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK; 3275 hash = &action_hash[hv]; 3276 lockmgr(&hash->lk, LK_EXCLUSIVE); 3277 if (action->event != VMEVENT_NONE) { 3278 action->event = VMEVENT_NONE; 3279 LIST_REMOVE(action, entry); 3280 3281 if (LIST_EMPTY(&hash->list)) 3282 vm_page_flag_clear(action->m, PG_ACTIONLIST); 3283 } 3284 lockmgr(&hash->lk, LK_RELEASE); 3285 } 3286 3287 /* 3288 * Issue an event on a VM page. Corresponding action structures are 3289 * removed from the page's list and called. 3290 * 3291 * If the vm_page has no more pending action events we clear its 3292 * PG_ACTIONLIST flag. 3293 */ 3294 void 3295 vm_page_event_internal(vm_page_t m, vm_page_event_t event) 3296 { 3297 struct vm_page_action_hash *hash; 3298 struct vm_page_action *scan; 3299 struct vm_page_action *next; 3300 int hv; 3301 int all; 3302 3303 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK; 3304 hash = &action_hash[hv]; 3305 all = 1; 3306 3307 lockmgr(&hash->lk, LK_EXCLUSIVE); 3308 LIST_FOREACH_MUTABLE(scan, &hash->list, entry, next) { 3309 if (scan->m == m) { 3310 if (scan->event == event) { 3311 scan->event = VMEVENT_NONE; 3312 LIST_REMOVE(scan, entry); 3313 scan->func(m, scan); 3314 /* XXX */ 3315 } else { 3316 all = 0; 3317 } 3318 } 3319 } 3320 if (all) 3321 vm_page_flag_clear(m, PG_ACTIONLIST); 3322 lockmgr(&hash->lk, LK_RELEASE); 3323 } 3324 3325 #include "opt_ddb.h" 3326 #ifdef DDB 3327 #include <sys/kernel.h> 3328 3329 #include <ddb/ddb.h> 3330 3331 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3332 { 3333 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3334 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3335 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3336 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3337 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3338 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3339 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3340 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3341 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3342 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3343 } 3344 3345 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3346 { 3347 int i; 3348 db_printf("PQ_FREE:"); 3349 for (i = 0; i < PQ_L2_SIZE; i++) { 3350 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3351 } 3352 db_printf("\n"); 3353 3354 db_printf("PQ_CACHE:"); 3355 for(i = 0; i < PQ_L2_SIZE; i++) { 3356 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3357 } 3358 db_printf("\n"); 3359 3360 db_printf("PQ_ACTIVE:"); 3361 for(i = 0; i < PQ_L2_SIZE; i++) { 3362 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3363 } 3364 db_printf("\n"); 3365 3366 db_printf("PQ_INACTIVE:"); 3367 for(i = 0; i < PQ_L2_SIZE; i++) { 3368 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3369 } 3370 db_printf("\n"); 3371 } 3372 #endif /* DDB */ 3373