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