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 BUSY. If also_m_busy is TRUE we wait 863 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 864 * 865 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 866 * call will be made before returning. 867 * 868 * This function does NOT busy the page and on return the page is not 869 * guaranteed to be available. 870 */ 871 void 872 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 873 { 874 u_int32_t busy_count; 875 876 for (;;) { 877 busy_count = m->busy_count; 878 cpu_ccfence(); 879 880 if ((busy_count & PBUSY_LOCKED) == 0 && 881 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 882 break; 883 } 884 tsleep_interlock(m, 0); 885 if (atomic_cmpset_int(&m->busy_count, busy_count, 886 busy_count | PBUSY_WANTED)) { 887 atomic_set_int(&m->flags, PG_REFERENCED); 888 tsleep(m, PINTERLOCKED, msg, 0); 889 break; 890 } 891 } 892 } 893 894 /* 895 * This calculates and returns a page color given an optional VM object and 896 * either a pindex or an iterator. We attempt to return a cpu-localized 897 * pg_color that is still roughly 16-way set-associative. The CPU topology 898 * is used if it was probed. 899 * 900 * The caller may use the returned value to index into e.g. PQ_FREE when 901 * allocating a page in order to nominally obtain pages that are hopefully 902 * already localized to the requesting cpu. This function is not able to 903 * provide any sort of guarantee of this, but does its best to improve 904 * hardware cache management performance. 905 * 906 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 907 */ 908 u_short 909 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 910 { 911 u_short pg_color; 912 int phys_id; 913 int core_id; 914 int object_pg_color; 915 916 phys_id = get_cpu_phys_id(cpuid); 917 core_id = get_cpu_core_id(cpuid); 918 object_pg_color = object ? object->pg_color : 0; 919 920 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 921 int grpsize; 922 923 /* 924 * Break us down by socket and cpu 925 */ 926 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids; 927 pg_color += core_id * PQ_L2_SIZE / 928 (cpu_topology_core_ids * cpu_topology_phys_ids); 929 930 /* 931 * Calculate remaining component for object/queue color 932 */ 933 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids * 934 cpu_topology_phys_ids); 935 if (grpsize >= 8) { 936 pg_color += (pindex + object_pg_color) % grpsize; 937 } else { 938 if (grpsize <= 2) { 939 grpsize = 8; 940 } else { 941 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 942 grpsize += grpsize; 943 if (grpsize < 8) 944 grpsize += grpsize; 945 } 946 pg_color += (pindex + object_pg_color) % grpsize; 947 } 948 } else { 949 /* 950 * Unknown topology, distribute things evenly. 951 */ 952 pg_color = cpuid * PQ_L2_SIZE / ncpus; 953 pg_color += pindex + object_pg_color; 954 } 955 return (pg_color & PQ_L2_MASK); 956 } 957 958 /* 959 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 960 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 961 */ 962 void 963 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 964 int also_m_busy, const char *msg 965 VM_PAGE_DEBUG_ARGS) 966 { 967 u_int32_t busy_count; 968 969 for (;;) { 970 busy_count = m->busy_count; 971 cpu_ccfence(); 972 if (busy_count & PBUSY_LOCKED) { 973 tsleep_interlock(m, 0); 974 if (atomic_cmpset_int(&m->busy_count, busy_count, 975 busy_count | PBUSY_WANTED)) { 976 atomic_set_int(&m->flags, PG_REFERENCED); 977 tsleep(m, PINTERLOCKED, msg, 0); 978 } 979 } else if (also_m_busy && busy_count) { 980 tsleep_interlock(m, 0); 981 if (atomic_cmpset_int(&m->busy_count, busy_count, 982 busy_count | PBUSY_WANTED)) { 983 atomic_set_int(&m->flags, PG_REFERENCED); 984 tsleep(m, PINTERLOCKED, msg, 0); 985 } 986 } else { 987 if (atomic_cmpset_int(&m->busy_count, busy_count, 988 busy_count | PBUSY_LOCKED)) { 989 #ifdef VM_PAGE_DEBUG 990 m->busy_func = func; 991 m->busy_line = lineno; 992 #endif 993 break; 994 } 995 } 996 } 997 } 998 999 /* 1000 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1001 * m->busy_count is also 0. 1002 * 1003 * Returns non-zero on failure. 1004 */ 1005 int 1006 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1007 VM_PAGE_DEBUG_ARGS) 1008 { 1009 u_int32_t busy_count; 1010 1011 for (;;) { 1012 busy_count = m->busy_count; 1013 cpu_ccfence(); 1014 if (busy_count & PBUSY_LOCKED) 1015 return TRUE; 1016 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1017 return TRUE; 1018 if (atomic_cmpset_int(&m->busy_count, busy_count, 1019 busy_count | PBUSY_LOCKED)) { 1020 #ifdef VM_PAGE_DEBUG 1021 m->busy_func = func; 1022 m->busy_line = lineno; 1023 #endif 1024 return FALSE; 1025 } 1026 } 1027 } 1028 1029 /* 1030 * Clear the BUSY flag and return non-zero to indicate to the caller 1031 * that a wakeup() should be performed. 1032 * 1033 * The vm_page must be spinlocked and will remain spinlocked on return. 1034 * The related queue must NOT be spinlocked (which could deadlock us). 1035 * 1036 * (inline version) 1037 */ 1038 static __inline 1039 int 1040 _vm_page_wakeup(vm_page_t m) 1041 { 1042 u_int32_t busy_count; 1043 1044 for (;;) { 1045 busy_count = m->busy_count; 1046 cpu_ccfence(); 1047 if (atomic_cmpset_int(&m->busy_count, busy_count, 1048 busy_count & 1049 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1050 break; 1051 } 1052 } 1053 return((int)(busy_count & PBUSY_WANTED)); 1054 } 1055 1056 /* 1057 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1058 * is typically the last call you make on a page before moving onto 1059 * other things. 1060 */ 1061 void 1062 vm_page_wakeup(vm_page_t m) 1063 { 1064 KASSERT(m->busy_count & PBUSY_LOCKED, 1065 ("vm_page_wakeup: page not busy!!!")); 1066 vm_page_spin_lock(m); 1067 if (_vm_page_wakeup(m)) { 1068 vm_page_spin_unlock(m); 1069 wakeup(m); 1070 } else { 1071 vm_page_spin_unlock(m); 1072 } 1073 } 1074 1075 /* 1076 * Holding a page keeps it from being reused. Other parts of the system 1077 * can still disassociate the page from its current object and free it, or 1078 * perform read or write I/O on it and/or otherwise manipulate the page, 1079 * but if the page is held the VM system will leave the page and its data 1080 * intact and not reuse the page for other purposes until the last hold 1081 * reference is released. (see vm_page_wire() if you want to prevent the 1082 * page from being disassociated from its object too). 1083 * 1084 * The caller must still validate the contents of the page and, if necessary, 1085 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 1086 * before manipulating the page. 1087 * 1088 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 1089 */ 1090 void 1091 vm_page_hold(vm_page_t m) 1092 { 1093 vm_page_spin_lock(m); 1094 atomic_add_int(&m->hold_count, 1); 1095 if (m->queue - m->pc == PQ_FREE) { 1096 _vm_page_queue_spin_lock(m); 1097 _vm_page_rem_queue_spinlocked(m); 1098 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1099 _vm_page_queue_spin_unlock(m); 1100 } 1101 vm_page_spin_unlock(m); 1102 } 1103 1104 /* 1105 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1106 * it was freed while held and must be moved back to the FREE queue. 1107 */ 1108 void 1109 vm_page_unhold(vm_page_t m) 1110 { 1111 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1112 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 1113 m, m->hold_count, m->queue - m->pc)); 1114 vm_page_spin_lock(m); 1115 atomic_add_int(&m->hold_count, -1); 1116 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1117 _vm_page_queue_spin_lock(m); 1118 _vm_page_rem_queue_spinlocked(m); 1119 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1120 _vm_page_queue_spin_unlock(m); 1121 } 1122 vm_page_spin_unlock(m); 1123 } 1124 1125 /* 1126 * vm_page_getfake: 1127 * 1128 * Create a fictitious page with the specified physical address and 1129 * memory attribute. The memory attribute is the only the machine- 1130 * dependent aspect of a fictitious page that must be initialized. 1131 */ 1132 1133 void 1134 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1135 { 1136 1137 if ((m->flags & PG_FICTITIOUS) != 0) { 1138 /* 1139 * The page's memattr might have changed since the 1140 * previous initialization. Update the pmap to the 1141 * new memattr. 1142 */ 1143 goto memattr; 1144 } 1145 m->phys_addr = paddr; 1146 m->queue = PQ_NONE; 1147 /* Fictitious pages don't use "segind". */ 1148 /* Fictitious pages don't use "order" or "pool". */ 1149 m->flags = PG_FICTITIOUS | PG_UNMANAGED; 1150 m->busy_count = PBUSY_LOCKED; 1151 m->wire_count = 1; 1152 spin_init(&m->spin, "fake_page"); 1153 pmap_page_init(m); 1154 memattr: 1155 pmap_page_set_memattr(m, memattr); 1156 } 1157 1158 /* 1159 * Inserts the given vm_page into the object and object list. 1160 * 1161 * The pagetables are not updated but will presumably fault the page 1162 * in if necessary, or if a kernel page the caller will at some point 1163 * enter the page into the kernel's pmap. We are not allowed to block 1164 * here so we *can't* do this anyway. 1165 * 1166 * This routine may not block. 1167 * This routine must be called with the vm_object held. 1168 * This routine must be called with a critical section held. 1169 * 1170 * This routine returns TRUE if the page was inserted into the object 1171 * successfully, and FALSE if the page already exists in the object. 1172 */ 1173 int 1174 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1175 { 1176 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1177 if (m->object != NULL) 1178 panic("vm_page_insert: already inserted"); 1179 1180 atomic_add_int(&object->generation, 1); 1181 1182 /* 1183 * Record the object/offset pair in this page and add the 1184 * pv_list_count of the page to the object. 1185 * 1186 * The vm_page spin lock is required for interactions with the pmap. 1187 */ 1188 vm_page_spin_lock(m); 1189 m->object = object; 1190 m->pindex = pindex; 1191 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1192 m->object = NULL; 1193 m->pindex = 0; 1194 vm_page_spin_unlock(m); 1195 return FALSE; 1196 } 1197 ++object->resident_page_count; 1198 ++mycpu->gd_vmtotal.t_rm; 1199 vm_page_spin_unlock(m); 1200 1201 /* 1202 * Since we are inserting a new and possibly dirty page, 1203 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1204 */ 1205 if ((m->valid & m->dirty) || 1206 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1207 vm_object_set_writeable_dirty(object); 1208 1209 /* 1210 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1211 */ 1212 swap_pager_page_inserted(m); 1213 return TRUE; 1214 } 1215 1216 /* 1217 * Removes the given vm_page_t from the (object,index) table 1218 * 1219 * The underlying pmap entry (if any) is NOT removed here. 1220 * This routine may not block. 1221 * 1222 * The page must be BUSY and will remain BUSY on return. 1223 * No other requirements. 1224 * 1225 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1226 * it busy. 1227 */ 1228 void 1229 vm_page_remove(vm_page_t m) 1230 { 1231 vm_object_t object; 1232 1233 if (m->object == NULL) { 1234 return; 1235 } 1236 1237 if ((m->busy_count & PBUSY_LOCKED) == 0) 1238 panic("vm_page_remove: page not busy"); 1239 1240 object = m->object; 1241 1242 vm_object_hold(object); 1243 1244 /* 1245 * Remove the page from the object and update the object. 1246 * 1247 * The vm_page spin lock is required for interactions with the pmap. 1248 */ 1249 vm_page_spin_lock(m); 1250 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1251 --object->resident_page_count; 1252 --mycpu->gd_vmtotal.t_rm; 1253 m->object = NULL; 1254 atomic_add_int(&object->generation, 1); 1255 vm_page_spin_unlock(m); 1256 1257 vm_object_drop(object); 1258 } 1259 1260 /* 1261 * Locate and return the page at (object, pindex), or NULL if the 1262 * page could not be found. 1263 * 1264 * The caller must hold the vm_object token. 1265 */ 1266 vm_page_t 1267 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1268 { 1269 vm_page_t m; 1270 1271 /* 1272 * Search the hash table for this object/offset pair 1273 */ 1274 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1275 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1276 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1277 return(m); 1278 } 1279 1280 vm_page_t 1281 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1282 vm_pindex_t pindex, 1283 int also_m_busy, const char *msg 1284 VM_PAGE_DEBUG_ARGS) 1285 { 1286 u_int32_t busy_count; 1287 vm_page_t m; 1288 1289 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1290 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1291 while (m) { 1292 KKASSERT(m->object == object && m->pindex == pindex); 1293 busy_count = m->busy_count; 1294 cpu_ccfence(); 1295 if (busy_count & PBUSY_LOCKED) { 1296 tsleep_interlock(m, 0); 1297 if (atomic_cmpset_int(&m->busy_count, busy_count, 1298 busy_count | PBUSY_WANTED)) { 1299 atomic_set_int(&m->flags, PG_REFERENCED); 1300 tsleep(m, PINTERLOCKED, msg, 0); 1301 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1302 pindex); 1303 } 1304 } else if (also_m_busy && busy_count) { 1305 tsleep_interlock(m, 0); 1306 if (atomic_cmpset_int(&m->busy_count, busy_count, 1307 busy_count | PBUSY_WANTED)) { 1308 atomic_set_int(&m->flags, PG_REFERENCED); 1309 tsleep(m, PINTERLOCKED, msg, 0); 1310 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1311 pindex); 1312 } 1313 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1314 busy_count | PBUSY_LOCKED)) { 1315 #ifdef VM_PAGE_DEBUG 1316 m->busy_func = func; 1317 m->busy_line = lineno; 1318 #endif 1319 break; 1320 } 1321 } 1322 return m; 1323 } 1324 1325 /* 1326 * Attempt to lookup and busy a page. 1327 * 1328 * Returns NULL if the page could not be found 1329 * 1330 * Returns a vm_page and error == TRUE if the page exists but could not 1331 * be busied. 1332 * 1333 * Returns a vm_page and error == FALSE on success. 1334 */ 1335 vm_page_t 1336 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1337 vm_pindex_t pindex, 1338 int also_m_busy, int *errorp 1339 VM_PAGE_DEBUG_ARGS) 1340 { 1341 u_int32_t busy_count; 1342 vm_page_t m; 1343 1344 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1345 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1346 *errorp = FALSE; 1347 while (m) { 1348 KKASSERT(m->object == object && m->pindex == pindex); 1349 busy_count = m->busy_count; 1350 cpu_ccfence(); 1351 if (busy_count & PBUSY_LOCKED) { 1352 *errorp = TRUE; 1353 break; 1354 } 1355 if (also_m_busy && busy_count) { 1356 *errorp = TRUE; 1357 break; 1358 } 1359 if (atomic_cmpset_int(&m->busy_count, busy_count, 1360 busy_count | PBUSY_LOCKED)) { 1361 #ifdef VM_PAGE_DEBUG 1362 m->busy_func = func; 1363 m->busy_line = lineno; 1364 #endif 1365 break; 1366 } 1367 } 1368 return m; 1369 } 1370 1371 /* 1372 * Returns a page that is only soft-busied for use by the caller in 1373 * a read-only fashion. Returns NULL if the page could not be found, 1374 * the soft busy could not be obtained, or the page data is invalid. 1375 */ 1376 vm_page_t 1377 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1378 int pgoff, int pgbytes) 1379 { 1380 vm_page_t m; 1381 1382 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1383 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1384 if (m) { 1385 if ((m->valid != VM_PAGE_BITS_ALL && 1386 !vm_page_is_valid(m, pgoff, pgbytes)) || 1387 (m->flags & PG_FICTITIOUS)) { 1388 m = NULL; 1389 } else if (vm_page_sbusy_try(m)) { 1390 m = NULL; 1391 } else if ((m->valid != VM_PAGE_BITS_ALL && 1392 !vm_page_is_valid(m, pgoff, pgbytes)) || 1393 (m->flags & PG_FICTITIOUS)) { 1394 vm_page_sbusy_drop(m); 1395 m = NULL; 1396 } 1397 } 1398 return m; 1399 } 1400 1401 /* 1402 * Caller must hold the related vm_object 1403 */ 1404 vm_page_t 1405 vm_page_next(vm_page_t m) 1406 { 1407 vm_page_t next; 1408 1409 next = vm_page_rb_tree_RB_NEXT(m); 1410 if (next && next->pindex != m->pindex + 1) 1411 next = NULL; 1412 return (next); 1413 } 1414 1415 /* 1416 * vm_page_rename() 1417 * 1418 * Move the given vm_page from its current object to the specified 1419 * target object/offset. The page must be busy and will remain so 1420 * on return. 1421 * 1422 * new_object must be held. 1423 * This routine might block. XXX ? 1424 * 1425 * NOTE: Swap associated with the page must be invalidated by the move. We 1426 * have to do this for several reasons: (1) we aren't freeing the 1427 * page, (2) we are dirtying the page, (3) the VM system is probably 1428 * moving the page from object A to B, and will then later move 1429 * the backing store from A to B and we can't have a conflict. 1430 * 1431 * NOTE: We *always* dirty the page. It is necessary both for the 1432 * fact that we moved it, and because we may be invalidating 1433 * swap. If the page is on the cache, we have to deactivate it 1434 * or vm_page_dirty() will panic. Dirty pages are not allowed 1435 * on the cache. 1436 */ 1437 void 1438 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1439 { 1440 KKASSERT(m->busy_count & PBUSY_LOCKED); 1441 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1442 if (m->object) { 1443 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1444 vm_page_remove(m); 1445 } 1446 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1447 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1448 new_object, new_pindex); 1449 } 1450 if (m->queue - m->pc == PQ_CACHE) 1451 vm_page_deactivate(m); 1452 vm_page_dirty(m); 1453 } 1454 1455 /* 1456 * vm_page_unqueue() without any wakeup. This routine is used when a page 1457 * is to remain BUSYied by the caller. 1458 * 1459 * This routine may not block. 1460 */ 1461 void 1462 vm_page_unqueue_nowakeup(vm_page_t m) 1463 { 1464 vm_page_and_queue_spin_lock(m); 1465 (void)_vm_page_rem_queue_spinlocked(m); 1466 vm_page_spin_unlock(m); 1467 } 1468 1469 /* 1470 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1471 * if necessary. 1472 * 1473 * This routine may not block. 1474 */ 1475 void 1476 vm_page_unqueue(vm_page_t m) 1477 { 1478 u_short queue; 1479 1480 vm_page_and_queue_spin_lock(m); 1481 queue = _vm_page_rem_queue_spinlocked(m); 1482 if (queue == PQ_FREE || queue == PQ_CACHE) { 1483 vm_page_spin_unlock(m); 1484 pagedaemon_wakeup(); 1485 } else { 1486 vm_page_spin_unlock(m); 1487 } 1488 } 1489 1490 /* 1491 * vm_page_list_find() 1492 * 1493 * Find a page on the specified queue with color optimization. 1494 * 1495 * The page coloring optimization attempts to locate a page that does 1496 * not overload other nearby pages in the object in the cpu's L1 or L2 1497 * caches. We need this optimization because cpu caches tend to be 1498 * physical caches, while object spaces tend to be virtual. 1499 * 1500 * The page coloring optimization also, very importantly, tries to localize 1501 * memory to cpus and physical sockets. 1502 * 1503 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1504 * and the algorithm is adjusted to localize allocations on a per-core basis. 1505 * This is done by 'twisting' the colors. 1506 * 1507 * The page is returned spinlocked and removed from its queue (it will 1508 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1509 * is responsible for dealing with the busy-page case (usually by 1510 * deactivating the page and looping). 1511 * 1512 * NOTE: This routine is carefully inlined. A non-inlined version 1513 * is available for outside callers but the only critical path is 1514 * from within this source file. 1515 * 1516 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1517 * represent stable storage, allowing us to order our locks vm_page 1518 * first, then queue. 1519 */ 1520 static __inline 1521 vm_page_t 1522 _vm_page_list_find(int basequeue, int index) 1523 { 1524 vm_page_t m; 1525 1526 for (;;) { 1527 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1528 if (m == NULL) { 1529 m = _vm_page_list_find2(basequeue, index); 1530 return(m); 1531 } 1532 vm_page_and_queue_spin_lock(m); 1533 if (m->queue == basequeue + index) { 1534 _vm_page_rem_queue_spinlocked(m); 1535 /* vm_page_t spin held, no queue spin */ 1536 break; 1537 } 1538 vm_page_and_queue_spin_unlock(m); 1539 } 1540 return(m); 1541 } 1542 1543 /* 1544 * If we could not find the page in the desired queue try to find it in 1545 * a nearby queue. 1546 */ 1547 static vm_page_t 1548 _vm_page_list_find2(int basequeue, int index) 1549 { 1550 struct vpgqueues *pq; 1551 vm_page_t m = NULL; 1552 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1553 int pqi; 1554 int i; 1555 1556 index &= PQ_L2_MASK; 1557 pq = &vm_page_queues[basequeue]; 1558 1559 /* 1560 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1561 * else fails (PQ_L2_MASK which is 255). 1562 */ 1563 do { 1564 pqmask = (pqmask << 1) | 1; 1565 for (i = 0; i <= pqmask; ++i) { 1566 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1567 m = TAILQ_FIRST(&pq[pqi].pl); 1568 if (m) { 1569 _vm_page_and_queue_spin_lock(m); 1570 if (m->queue == basequeue + pqi) { 1571 _vm_page_rem_queue_spinlocked(m); 1572 return(m); 1573 } 1574 _vm_page_and_queue_spin_unlock(m); 1575 --i; 1576 continue; 1577 } 1578 } 1579 } while (pqmask != PQ_L2_MASK); 1580 1581 return(m); 1582 } 1583 1584 /* 1585 * Returns a vm_page candidate for allocation. The page is not busied so 1586 * it can move around. The caller must busy the page (and typically 1587 * deactivate it if it cannot be busied!) 1588 * 1589 * Returns a spinlocked vm_page that has been removed from its queue. 1590 */ 1591 vm_page_t 1592 vm_page_list_find(int basequeue, int index) 1593 { 1594 return(_vm_page_list_find(basequeue, index)); 1595 } 1596 1597 /* 1598 * Find a page on the cache queue with color optimization, remove it 1599 * from the queue, and busy it. The returned page will not be spinlocked. 1600 * 1601 * A candidate failure will be deactivated. Candidates can fail due to 1602 * being busied by someone else, in which case they will be deactivated. 1603 * 1604 * This routine may not block. 1605 * 1606 */ 1607 static vm_page_t 1608 vm_page_select_cache(u_short pg_color) 1609 { 1610 vm_page_t m; 1611 1612 for (;;) { 1613 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1614 if (m == NULL) 1615 break; 1616 /* 1617 * (m) has been removed from its queue and spinlocked 1618 */ 1619 if (vm_page_busy_try(m, TRUE)) { 1620 _vm_page_deactivate_locked(m, 0); 1621 vm_page_spin_unlock(m); 1622 } else { 1623 /* 1624 * We successfully busied the page 1625 */ 1626 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1627 m->hold_count == 0 && 1628 m->wire_count == 0 && 1629 (m->dirty & m->valid) == 0) { 1630 vm_page_spin_unlock(m); 1631 pagedaemon_wakeup(); 1632 return(m); 1633 } 1634 1635 /* 1636 * The page cannot be recycled, deactivate it. 1637 */ 1638 _vm_page_deactivate_locked(m, 0); 1639 if (_vm_page_wakeup(m)) { 1640 vm_page_spin_unlock(m); 1641 wakeup(m); 1642 } else { 1643 vm_page_spin_unlock(m); 1644 } 1645 } 1646 } 1647 return (m); 1648 } 1649 1650 /* 1651 * Find a free page. We attempt to inline the nominal case and fall back 1652 * to _vm_page_select_free() otherwise. A busied page is removed from 1653 * the queue and returned. 1654 * 1655 * This routine may not block. 1656 */ 1657 static __inline vm_page_t 1658 vm_page_select_free(u_short pg_color) 1659 { 1660 vm_page_t m; 1661 1662 for (;;) { 1663 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1664 if (m == NULL) 1665 break; 1666 if (vm_page_busy_try(m, TRUE)) { 1667 /* 1668 * Various mechanisms such as a pmap_collect can 1669 * result in a busy page on the free queue. We 1670 * have to move the page out of the way so we can 1671 * retry the allocation. If the other thread is not 1672 * allocating the page then m->valid will remain 0 and 1673 * the pageout daemon will free the page later on. 1674 * 1675 * Since we could not busy the page, however, we 1676 * cannot make assumptions as to whether the page 1677 * will be allocated by the other thread or not, 1678 * so all we can do is deactivate it to move it out 1679 * of the way. In particular, if the other thread 1680 * wires the page it may wind up on the inactive 1681 * queue and the pageout daemon will have to deal 1682 * with that case too. 1683 */ 1684 _vm_page_deactivate_locked(m, 0); 1685 vm_page_spin_unlock(m); 1686 } else { 1687 /* 1688 * Theoretically if we are able to busy the page 1689 * atomic with the queue removal (using the vm_page 1690 * lock) nobody else should be able to mess with the 1691 * page before us. 1692 */ 1693 KKASSERT((m->flags & (PG_UNMANAGED | 1694 PG_NEED_COMMIT)) == 0); 1695 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1696 "pg %p q=%d flags=%08x hold=%d wire=%d", 1697 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1698 KKASSERT(m->wire_count == 0); 1699 vm_page_spin_unlock(m); 1700 pagedaemon_wakeup(); 1701 1702 /* return busied and removed page */ 1703 return(m); 1704 } 1705 } 1706 return(m); 1707 } 1708 1709 /* 1710 * vm_page_alloc() 1711 * 1712 * Allocate and return a memory cell associated with this VM object/offset 1713 * pair. If object is NULL an unassociated page will be allocated. 1714 * 1715 * The returned page will be busied and removed from its queues. This 1716 * routine can block and may return NULL if a race occurs and the page 1717 * is found to already exist at the specified (object, pindex). 1718 * 1719 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1720 * VM_ALLOC_QUICK like normal but cannot use cache 1721 * VM_ALLOC_SYSTEM greater free drain 1722 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1723 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1724 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1725 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1726 * (see vm_page_grab()) 1727 * VM_ALLOC_USE_GD ok to use per-gd cache 1728 * 1729 * VM_ALLOC_CPU(n) allocate using specified cpu localization 1730 * 1731 * The object must be held if not NULL 1732 * This routine may not block 1733 * 1734 * Additional special handling is required when called from an interrupt 1735 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1736 * in this case. 1737 */ 1738 vm_page_t 1739 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1740 { 1741 globaldata_t gd; 1742 vm_object_t obj; 1743 vm_page_t m; 1744 u_short pg_color; 1745 int cpuid_local; 1746 1747 #if 0 1748 /* 1749 * Special per-cpu free VM page cache. The pages are pre-busied 1750 * and pre-zerod for us. 1751 */ 1752 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1753 crit_enter_gd(gd); 1754 if (gd->gd_vmpg_count) { 1755 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1756 crit_exit_gd(gd); 1757 goto done; 1758 } 1759 crit_exit_gd(gd); 1760 } 1761 #endif 1762 m = NULL; 1763 1764 /* 1765 * CPU LOCALIZATION 1766 * 1767 * CPU localization algorithm. Break the page queues up by physical 1768 * id and core id (note that two cpu threads will have the same core 1769 * id, and core_id != gd_cpuid). 1770 * 1771 * This is nowhere near perfect, for example the last pindex in a 1772 * subgroup will overflow into the next cpu or package. But this 1773 * should get us good page reuse locality in heavy mixed loads. 1774 * 1775 * (may be executed before the APs are started, so other GDs might 1776 * not exist!) 1777 */ 1778 if (page_req & VM_ALLOC_CPU_SPEC) 1779 cpuid_local = VM_ALLOC_GETCPU(page_req); 1780 else 1781 cpuid_local = mycpu->gd_cpuid; 1782 1783 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 1784 1785 KKASSERT(page_req & 1786 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1787 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1788 1789 /* 1790 * Certain system threads (pageout daemon, buf_daemon's) are 1791 * allowed to eat deeper into the free page list. 1792 */ 1793 if (curthread->td_flags & TDF_SYSTHREAD) 1794 page_req |= VM_ALLOC_SYSTEM; 1795 1796 /* 1797 * Impose various limitations. Note that the v_free_reserved test 1798 * must match the opposite of vm_page_count_target() to avoid 1799 * livelocks, be careful. 1800 */ 1801 loop: 1802 gd = mycpu; 1803 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 1804 ((page_req & VM_ALLOC_INTERRUPT) && 1805 gd->gd_vmstats.v_free_count > 0) || 1806 ((page_req & VM_ALLOC_SYSTEM) && 1807 gd->gd_vmstats.v_cache_count == 0 && 1808 gd->gd_vmstats.v_free_count > 1809 gd->gd_vmstats.v_interrupt_free_min) 1810 ) { 1811 /* 1812 * The free queue has sufficient free pages to take one out. 1813 */ 1814 m = vm_page_select_free(pg_color); 1815 } else if (page_req & VM_ALLOC_NORMAL) { 1816 /* 1817 * Allocatable from the cache (non-interrupt only). On 1818 * success, we must free the page and try again, thus 1819 * ensuring that vmstats.v_*_free_min counters are replenished. 1820 */ 1821 #ifdef INVARIANTS 1822 if (curthread->td_preempted) { 1823 kprintf("vm_page_alloc(): warning, attempt to allocate" 1824 " cache page from preempting interrupt\n"); 1825 m = NULL; 1826 } else { 1827 m = vm_page_select_cache(pg_color); 1828 } 1829 #else 1830 m = vm_page_select_cache(pg_color); 1831 #endif 1832 /* 1833 * On success move the page into the free queue and loop. 1834 * 1835 * Only do this if we can safely acquire the vm_object lock, 1836 * because this is effectively a random page and the caller 1837 * might be holding the lock shared, we don't want to 1838 * deadlock. 1839 */ 1840 if (m != NULL) { 1841 KASSERT(m->dirty == 0, 1842 ("Found dirty cache page %p", m)); 1843 if ((obj = m->object) != NULL) { 1844 if (vm_object_hold_try(obj)) { 1845 vm_page_protect(m, VM_PROT_NONE); 1846 vm_page_free(m); 1847 /* m->object NULL here */ 1848 vm_object_drop(obj); 1849 } else { 1850 vm_page_deactivate(m); 1851 vm_page_wakeup(m); 1852 } 1853 } else { 1854 vm_page_protect(m, VM_PROT_NONE); 1855 vm_page_free(m); 1856 } 1857 goto loop; 1858 } 1859 1860 /* 1861 * On failure return NULL 1862 */ 1863 atomic_add_int(&vm_pageout_deficit, 1); 1864 pagedaemon_wakeup(); 1865 return (NULL); 1866 } else { 1867 /* 1868 * No pages available, wakeup the pageout daemon and give up. 1869 */ 1870 atomic_add_int(&vm_pageout_deficit, 1); 1871 pagedaemon_wakeup(); 1872 return (NULL); 1873 } 1874 1875 /* 1876 * v_free_count can race so loop if we don't find the expected 1877 * page. 1878 */ 1879 if (m == NULL) { 1880 vmstats_rollup(); 1881 goto loop; 1882 } 1883 1884 /* 1885 * Good page found. The page has already been busied for us and 1886 * removed from its queues. 1887 */ 1888 KASSERT(m->dirty == 0, 1889 ("vm_page_alloc: free/cache page %p was dirty", m)); 1890 KKASSERT(m->queue == PQ_NONE); 1891 1892 #if 0 1893 done: 1894 #endif 1895 /* 1896 * Initialize the structure, inheriting some flags but clearing 1897 * all the rest. The page has already been busied for us. 1898 */ 1899 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 1900 1901 KKASSERT(m->wire_count == 0); 1902 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 1903 m->act_count = 0; 1904 m->valid = 0; 1905 1906 /* 1907 * Caller must be holding the object lock (asserted by 1908 * vm_page_insert()). 1909 * 1910 * NOTE: Inserting a page here does not insert it into any pmaps 1911 * (which could cause us to block allocating memory). 1912 * 1913 * NOTE: If no object an unassociated page is allocated, m->pindex 1914 * can be used by the caller for any purpose. 1915 */ 1916 if (object) { 1917 if (vm_page_insert(m, object, pindex) == FALSE) { 1918 vm_page_free(m); 1919 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1920 panic("PAGE RACE %p[%ld]/%p", 1921 object, (long)pindex, m); 1922 m = NULL; 1923 } 1924 } else { 1925 m->pindex = pindex; 1926 } 1927 1928 /* 1929 * Don't wakeup too often - wakeup the pageout daemon when 1930 * we would be nearly out of memory. 1931 */ 1932 pagedaemon_wakeup(); 1933 1934 /* 1935 * A BUSY page is returned. 1936 */ 1937 return (m); 1938 } 1939 1940 /* 1941 * Returns number of pages available in our DMA memory reserve 1942 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1943 */ 1944 vm_size_t 1945 vm_contig_avail_pages(void) 1946 { 1947 alist_blk_t blk; 1948 alist_blk_t count; 1949 alist_blk_t bfree; 1950 spin_lock(&vm_contig_spin); 1951 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1952 spin_unlock(&vm_contig_spin); 1953 1954 return bfree; 1955 } 1956 1957 /* 1958 * Attempt to allocate contiguous physical memory with the specified 1959 * requirements. 1960 */ 1961 vm_page_t 1962 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1963 unsigned long alignment, unsigned long boundary, 1964 unsigned long size, vm_memattr_t memattr) 1965 { 1966 alist_blk_t blk; 1967 vm_page_t m; 1968 int i; 1969 1970 alignment >>= PAGE_SHIFT; 1971 if (alignment == 0) 1972 alignment = 1; 1973 boundary >>= PAGE_SHIFT; 1974 if (boundary == 0) 1975 boundary = 1; 1976 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1977 1978 spin_lock(&vm_contig_spin); 1979 blk = alist_alloc(&vm_contig_alist, 0, size); 1980 if (blk == ALIST_BLOCK_NONE) { 1981 spin_unlock(&vm_contig_spin); 1982 if (bootverbose) { 1983 kprintf("vm_page_alloc_contig: %ldk nospace\n", 1984 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 1985 } 1986 return(NULL); 1987 } 1988 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 1989 alist_free(&vm_contig_alist, blk, size); 1990 spin_unlock(&vm_contig_spin); 1991 if (bootverbose) { 1992 kprintf("vm_page_alloc_contig: %ldk high " 1993 "%016jx failed\n", 1994 (size + PAGE_MASK) * (PAGE_SIZE / 1024), 1995 (intmax_t)high); 1996 } 1997 return(NULL); 1998 } 1999 spin_unlock(&vm_contig_spin); 2000 if (vm_contig_verbose) { 2001 kprintf("vm_page_alloc_contig: %016jx/%ldk\n", 2002 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT, 2003 (size + PAGE_MASK) * (PAGE_SIZE / 1024)); 2004 } 2005 2006 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2007 if (memattr != VM_MEMATTR_DEFAULT) 2008 for (i = 0;i < size;i++) 2009 pmap_page_set_memattr(&m[i], memattr); 2010 return m; 2011 } 2012 2013 /* 2014 * Free contiguously allocated pages. The pages will be wired but not busy. 2015 * When freeing to the alist we leave them wired and not busy. 2016 */ 2017 void 2018 vm_page_free_contig(vm_page_t m, unsigned long size) 2019 { 2020 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2021 vm_pindex_t start = pa >> PAGE_SHIFT; 2022 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2023 2024 if (vm_contig_verbose) { 2025 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2026 (intmax_t)pa, size / 1024); 2027 } 2028 if (pa < vm_low_phys_reserved) { 2029 KKASSERT(pa + size <= vm_low_phys_reserved); 2030 spin_lock(&vm_contig_spin); 2031 alist_free(&vm_contig_alist, start, pages); 2032 spin_unlock(&vm_contig_spin); 2033 } else { 2034 while (pages) { 2035 vm_page_busy_wait(m, FALSE, "cpgfr"); 2036 vm_page_unwire(m, 0); 2037 vm_page_free(m); 2038 --pages; 2039 ++m; 2040 } 2041 2042 } 2043 } 2044 2045 2046 /* 2047 * Wait for sufficient free memory for nominal heavy memory use kernel 2048 * operations. 2049 * 2050 * WARNING! Be sure never to call this in any vm_pageout code path, which 2051 * will trivially deadlock the system. 2052 */ 2053 void 2054 vm_wait_nominal(void) 2055 { 2056 while (vm_page_count_min(0)) 2057 vm_wait(0); 2058 } 2059 2060 /* 2061 * Test if vm_wait_nominal() would block. 2062 */ 2063 int 2064 vm_test_nominal(void) 2065 { 2066 if (vm_page_count_min(0)) 2067 return(1); 2068 return(0); 2069 } 2070 2071 /* 2072 * Block until free pages are available for allocation, called in various 2073 * places before memory allocations. 2074 * 2075 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2076 * more generous then that. 2077 */ 2078 void 2079 vm_wait(int timo) 2080 { 2081 /* 2082 * never wait forever 2083 */ 2084 if (timo == 0) 2085 timo = hz; 2086 lwkt_gettoken(&vm_token); 2087 2088 if (curthread == pagethread || 2089 curthread == emergpager) { 2090 /* 2091 * The pageout daemon itself needs pages, this is bad. 2092 */ 2093 if (vm_page_count_min(0)) { 2094 vm_pageout_pages_needed = 1; 2095 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2096 } 2097 } else { 2098 /* 2099 * Wakeup the pageout daemon if necessary and wait. 2100 * 2101 * Do not wait indefinitely for the target to be reached, 2102 * as load might prevent it from being reached any time soon. 2103 * But wait a little to try to slow down page allocations 2104 * and to give more important threads (the pagedaemon) 2105 * allocation priority. 2106 */ 2107 if (vm_page_count_target()) { 2108 if (vm_pages_needed == 0) { 2109 vm_pages_needed = 1; 2110 wakeup(&vm_pages_needed); 2111 } 2112 ++vm_pages_waiting; /* SMP race ok */ 2113 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2114 } 2115 } 2116 lwkt_reltoken(&vm_token); 2117 } 2118 2119 /* 2120 * Block until free pages are available for allocation 2121 * 2122 * Called only from vm_fault so that processes page faulting can be 2123 * easily tracked. 2124 */ 2125 void 2126 vm_wait_pfault(void) 2127 { 2128 /* 2129 * Wakeup the pageout daemon if necessary and wait. 2130 * 2131 * Do not wait indefinitely for the target to be reached, 2132 * as load might prevent it from being reached any time soon. 2133 * But wait a little to try to slow down page allocations 2134 * and to give more important threads (the pagedaemon) 2135 * allocation priority. 2136 */ 2137 if (vm_page_count_min(0)) { 2138 lwkt_gettoken(&vm_token); 2139 while (vm_page_count_severe()) { 2140 if (vm_page_count_target()) { 2141 thread_t td; 2142 2143 if (vm_pages_needed == 0) { 2144 vm_pages_needed = 1; 2145 wakeup(&vm_pages_needed); 2146 } 2147 ++vm_pages_waiting; /* SMP race ok */ 2148 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2149 2150 /* 2151 * Do not stay stuck in the loop if the system is trying 2152 * to kill the process. 2153 */ 2154 td = curthread; 2155 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2156 break; 2157 } 2158 } 2159 lwkt_reltoken(&vm_token); 2160 } 2161 } 2162 2163 /* 2164 * Put the specified page on the active list (if appropriate). Ensure 2165 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2166 * 2167 * The caller should be holding the page busied ? XXX 2168 * This routine may not block. 2169 */ 2170 void 2171 vm_page_activate(vm_page_t m) 2172 { 2173 u_short oqueue; 2174 2175 vm_page_spin_lock(m); 2176 if (m->queue - m->pc != PQ_ACTIVE) { 2177 _vm_page_queue_spin_lock(m); 2178 oqueue = _vm_page_rem_queue_spinlocked(m); 2179 /* page is left spinlocked, queue is unlocked */ 2180 2181 if (oqueue == PQ_CACHE) 2182 mycpu->gd_cnt.v_reactivated++; 2183 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2184 if (m->act_count < ACT_INIT) 2185 m->act_count = ACT_INIT; 2186 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2187 } 2188 _vm_page_and_queue_spin_unlock(m); 2189 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2190 pagedaemon_wakeup(); 2191 } else { 2192 if (m->act_count < ACT_INIT) 2193 m->act_count = ACT_INIT; 2194 vm_page_spin_unlock(m); 2195 } 2196 } 2197 2198 /* 2199 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2200 * routine is called when a page has been added to the cache or free 2201 * queues. 2202 * 2203 * This routine may not block. 2204 */ 2205 static __inline void 2206 vm_page_free_wakeup(void) 2207 { 2208 globaldata_t gd = mycpu; 2209 2210 /* 2211 * If the pageout daemon itself needs pages, then tell it that 2212 * there are some free. 2213 */ 2214 if (vm_pageout_pages_needed && 2215 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2216 gd->gd_vmstats.v_pageout_free_min 2217 ) { 2218 vm_pageout_pages_needed = 0; 2219 wakeup(&vm_pageout_pages_needed); 2220 } 2221 2222 /* 2223 * Wakeup processes that are waiting on memory. 2224 * 2225 * Generally speaking we want to wakeup stuck processes as soon as 2226 * possible. !vm_page_count_min(0) is the absolute minimum point 2227 * where we can do this. Wait a bit longer to reduce degenerate 2228 * re-blocking (vm_page_free_hysteresis). The target check is just 2229 * to make sure the min-check w/hysteresis does not exceed the 2230 * normal target. 2231 */ 2232 if (vm_pages_waiting) { 2233 if (!vm_page_count_min(vm_page_free_hysteresis) || 2234 !vm_page_count_target()) { 2235 vm_pages_waiting = 0; 2236 wakeup(&vmstats.v_free_count); 2237 ++mycpu->gd_cnt.v_ppwakeups; 2238 } 2239 #if 0 2240 if (!vm_page_count_target()) { 2241 /* 2242 * Plenty of pages are free, wakeup everyone. 2243 */ 2244 vm_pages_waiting = 0; 2245 wakeup(&vmstats.v_free_count); 2246 ++mycpu->gd_cnt.v_ppwakeups; 2247 } else if (!vm_page_count_min(0)) { 2248 /* 2249 * Some pages are free, wakeup someone. 2250 */ 2251 int wcount = vm_pages_waiting; 2252 if (wcount > 0) 2253 --wcount; 2254 vm_pages_waiting = wcount; 2255 wakeup_one(&vmstats.v_free_count); 2256 ++mycpu->gd_cnt.v_ppwakeups; 2257 } 2258 #endif 2259 } 2260 } 2261 2262 /* 2263 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2264 * it from its VM object. 2265 * 2266 * The vm_page must be BUSY on entry. BUSY will be released on 2267 * return (the page will have been freed). 2268 */ 2269 void 2270 vm_page_free_toq(vm_page_t m) 2271 { 2272 mycpu->gd_cnt.v_tfree++; 2273 KKASSERT((m->flags & PG_MAPPED) == 0); 2274 KKASSERT(m->busy_count & PBUSY_LOCKED); 2275 2276 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2277 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2278 "hold(%d)\n", 2279 (u_long)m->pindex, m->busy_count, m->hold_count); 2280 if ((m->queue - m->pc) == PQ_FREE) 2281 panic("vm_page_free: freeing free page"); 2282 else 2283 panic("vm_page_free: freeing busy page"); 2284 } 2285 2286 /* 2287 * Remove from object, spinlock the page and its queues and 2288 * remove from any queue. No queue spinlock will be held 2289 * after this section (because the page was removed from any 2290 * queue). 2291 */ 2292 vm_page_remove(m); 2293 vm_page_and_queue_spin_lock(m); 2294 _vm_page_rem_queue_spinlocked(m); 2295 2296 /* 2297 * No further management of fictitious pages occurs beyond object 2298 * and queue removal. 2299 */ 2300 if ((m->flags & PG_FICTITIOUS) != 0) { 2301 vm_page_spin_unlock(m); 2302 vm_page_wakeup(m); 2303 return; 2304 } 2305 2306 m->valid = 0; 2307 vm_page_undirty(m); 2308 2309 if (m->wire_count != 0) { 2310 if (m->wire_count > 1) { 2311 panic( 2312 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2313 m->wire_count, (long)m->pindex); 2314 } 2315 panic("vm_page_free: freeing wired page"); 2316 } 2317 2318 /* 2319 * Clear the UNMANAGED flag when freeing an unmanaged page. 2320 * Clear the NEED_COMMIT flag 2321 */ 2322 if (m->flags & PG_UNMANAGED) 2323 vm_page_flag_clear(m, PG_UNMANAGED); 2324 if (m->flags & PG_NEED_COMMIT) 2325 vm_page_flag_clear(m, PG_NEED_COMMIT); 2326 2327 if (m->hold_count != 0) { 2328 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2329 } else { 2330 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2331 } 2332 2333 /* 2334 * This sequence allows us to clear BUSY while still holding 2335 * its spin lock, which reduces contention vs allocators. We 2336 * must not leave the queue locked or _vm_page_wakeup() may 2337 * deadlock. 2338 */ 2339 _vm_page_queue_spin_unlock(m); 2340 if (_vm_page_wakeup(m)) { 2341 vm_page_spin_unlock(m); 2342 wakeup(m); 2343 } else { 2344 vm_page_spin_unlock(m); 2345 } 2346 vm_page_free_wakeup(); 2347 } 2348 2349 /* 2350 * vm_page_unmanage() 2351 * 2352 * Prevent PV management from being done on the page. The page is 2353 * removed from the paging queues as if it were wired, and as a 2354 * consequence of no longer being managed the pageout daemon will not 2355 * touch it (since there is no way to locate the pte mappings for the 2356 * page). madvise() calls that mess with the pmap will also no longer 2357 * operate on the page. 2358 * 2359 * Beyond that the page is still reasonably 'normal'. Freeing the page 2360 * will clear the flag. 2361 * 2362 * This routine is used by OBJT_PHYS objects - objects using unswappable 2363 * physical memory as backing store rather then swap-backed memory and 2364 * will eventually be extended to support 4MB unmanaged physical 2365 * mappings. 2366 * 2367 * Caller must be holding the page busy. 2368 */ 2369 void 2370 vm_page_unmanage(vm_page_t m) 2371 { 2372 KKASSERT(m->busy_count & PBUSY_LOCKED); 2373 if ((m->flags & PG_UNMANAGED) == 0) { 2374 if (m->wire_count == 0) 2375 vm_page_unqueue(m); 2376 } 2377 vm_page_flag_set(m, PG_UNMANAGED); 2378 } 2379 2380 /* 2381 * Mark this page as wired down by yet another map, removing it from 2382 * paging queues as necessary. 2383 * 2384 * Caller must be holding the page busy. 2385 */ 2386 void 2387 vm_page_wire(vm_page_t m) 2388 { 2389 /* 2390 * Only bump the wire statistics if the page is not already wired, 2391 * and only unqueue the page if it is on some queue (if it is unmanaged 2392 * it is already off the queues). Don't do anything with fictitious 2393 * pages because they are always wired. 2394 */ 2395 KKASSERT(m->busy_count & PBUSY_LOCKED); 2396 if ((m->flags & PG_FICTITIOUS) == 0) { 2397 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2398 if ((m->flags & PG_UNMANAGED) == 0) 2399 vm_page_unqueue(m); 2400 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2401 } 2402 KASSERT(m->wire_count != 0, 2403 ("vm_page_wire: wire_count overflow m=%p", m)); 2404 } 2405 } 2406 2407 /* 2408 * Release one wiring of this page, potentially enabling it to be paged again. 2409 * 2410 * Many pages placed on the inactive queue should actually go 2411 * into the cache, but it is difficult to figure out which. What 2412 * we do instead, if the inactive target is well met, is to put 2413 * clean pages at the head of the inactive queue instead of the tail. 2414 * This will cause them to be moved to the cache more quickly and 2415 * if not actively re-referenced, freed more quickly. If we just 2416 * stick these pages at the end of the inactive queue, heavy filesystem 2417 * meta-data accesses can cause an unnecessary paging load on memory bound 2418 * processes. This optimization causes one-time-use metadata to be 2419 * reused more quickly. 2420 * 2421 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2422 * the inactive queue. This helps the pageout daemon determine memory 2423 * pressure and act on out-of-memory situations more quickly. 2424 * 2425 * BUT, if we are in a low-memory situation we have no choice but to 2426 * put clean pages on the cache queue. 2427 * 2428 * A number of routines use vm_page_unwire() to guarantee that the page 2429 * will go into either the inactive or active queues, and will NEVER 2430 * be placed in the cache - for example, just after dirtying a page. 2431 * dirty pages in the cache are not allowed. 2432 * 2433 * This routine may not block. 2434 */ 2435 void 2436 vm_page_unwire(vm_page_t m, int activate) 2437 { 2438 KKASSERT(m->busy_count & PBUSY_LOCKED); 2439 if (m->flags & PG_FICTITIOUS) { 2440 /* do nothing */ 2441 } else if (m->wire_count <= 0) { 2442 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2443 } else { 2444 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2445 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1); 2446 if (m->flags & PG_UNMANAGED) { 2447 ; 2448 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2449 vm_page_spin_lock(m); 2450 _vm_page_add_queue_spinlocked(m, 2451 PQ_ACTIVE + m->pc, 0); 2452 _vm_page_and_queue_spin_unlock(m); 2453 } else { 2454 vm_page_spin_lock(m); 2455 vm_page_flag_clear(m, PG_WINATCFLS); 2456 _vm_page_add_queue_spinlocked(m, 2457 PQ_INACTIVE + m->pc, 0); 2458 ++vm_swapcache_inactive_heuristic; 2459 _vm_page_and_queue_spin_unlock(m); 2460 } 2461 } 2462 } 2463 } 2464 2465 /* 2466 * Move the specified page to the inactive queue. If the page has 2467 * any associated swap, the swap is deallocated. 2468 * 2469 * Normally athead is 0 resulting in LRU operation. athead is set 2470 * to 1 if we want this page to be 'as if it were placed in the cache', 2471 * except without unmapping it from the process address space. 2472 * 2473 * vm_page's spinlock must be held on entry and will remain held on return. 2474 * This routine may not block. 2475 */ 2476 static void 2477 _vm_page_deactivate_locked(vm_page_t m, int athead) 2478 { 2479 u_short oqueue; 2480 2481 /* 2482 * Ignore if already inactive. 2483 */ 2484 if (m->queue - m->pc == PQ_INACTIVE) 2485 return; 2486 _vm_page_queue_spin_lock(m); 2487 oqueue = _vm_page_rem_queue_spinlocked(m); 2488 2489 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2490 if (oqueue == PQ_CACHE) 2491 mycpu->gd_cnt.v_reactivated++; 2492 vm_page_flag_clear(m, PG_WINATCFLS); 2493 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2494 if (athead == 0) 2495 ++vm_swapcache_inactive_heuristic; 2496 } 2497 /* NOTE: PQ_NONE if condition not taken */ 2498 _vm_page_queue_spin_unlock(m); 2499 /* leaves vm_page spinlocked */ 2500 } 2501 2502 /* 2503 * Attempt to deactivate a page. 2504 * 2505 * No requirements. 2506 */ 2507 void 2508 vm_page_deactivate(vm_page_t m) 2509 { 2510 vm_page_spin_lock(m); 2511 _vm_page_deactivate_locked(m, 0); 2512 vm_page_spin_unlock(m); 2513 } 2514 2515 void 2516 vm_page_deactivate_locked(vm_page_t m) 2517 { 2518 _vm_page_deactivate_locked(m, 0); 2519 } 2520 2521 /* 2522 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2523 * 2524 * This function returns non-zero if it successfully moved the page to 2525 * PQ_CACHE. 2526 * 2527 * This function unconditionally unbusies the page on return. 2528 */ 2529 int 2530 vm_page_try_to_cache(vm_page_t m) 2531 { 2532 vm_page_spin_lock(m); 2533 if (m->dirty || m->hold_count || m->wire_count || 2534 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2535 if (_vm_page_wakeup(m)) { 2536 vm_page_spin_unlock(m); 2537 wakeup(m); 2538 } else { 2539 vm_page_spin_unlock(m); 2540 } 2541 return(0); 2542 } 2543 vm_page_spin_unlock(m); 2544 2545 /* 2546 * Page busied by us and no longer spinlocked. Dirty pages cannot 2547 * be moved to the cache. 2548 */ 2549 vm_page_test_dirty(m); 2550 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2551 vm_page_wakeup(m); 2552 return(0); 2553 } 2554 vm_page_cache(m); 2555 return(1); 2556 } 2557 2558 /* 2559 * Attempt to free the page. If we cannot free it, we do nothing. 2560 * 1 is returned on success, 0 on failure. 2561 * 2562 * No requirements. 2563 */ 2564 int 2565 vm_page_try_to_free(vm_page_t m) 2566 { 2567 vm_page_spin_lock(m); 2568 if (vm_page_busy_try(m, TRUE)) { 2569 vm_page_spin_unlock(m); 2570 return(0); 2571 } 2572 2573 /* 2574 * The page can be in any state, including already being on the free 2575 * queue. Check to see if it really can be freed. 2576 */ 2577 if (m->dirty || /* can't free if it is dirty */ 2578 m->hold_count || /* or held (XXX may be wrong) */ 2579 m->wire_count || /* or wired */ 2580 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2581 PG_NEED_COMMIT)) || /* or needs a commit */ 2582 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2583 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2584 if (_vm_page_wakeup(m)) { 2585 vm_page_spin_unlock(m); 2586 wakeup(m); 2587 } else { 2588 vm_page_spin_unlock(m); 2589 } 2590 return(0); 2591 } 2592 vm_page_spin_unlock(m); 2593 2594 /* 2595 * We can probably free the page. 2596 * 2597 * Page busied by us and no longer spinlocked. Dirty pages will 2598 * not be freed by this function. We have to re-test the 2599 * dirty bit after cleaning out the pmaps. 2600 */ 2601 vm_page_test_dirty(m); 2602 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2603 vm_page_wakeup(m); 2604 return(0); 2605 } 2606 vm_page_protect(m, VM_PROT_NONE); 2607 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2608 vm_page_wakeup(m); 2609 return(0); 2610 } 2611 vm_page_free(m); 2612 return(1); 2613 } 2614 2615 /* 2616 * vm_page_cache 2617 * 2618 * Put the specified page onto the page cache queue (if appropriate). 2619 * 2620 * The page must be busy, and this routine will release the busy and 2621 * possibly even free the page. 2622 */ 2623 void 2624 vm_page_cache(vm_page_t m) 2625 { 2626 /* 2627 * Not suitable for the cache 2628 */ 2629 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2630 (m->busy_count & PBUSY_MASK) || 2631 m->wire_count || m->hold_count) { 2632 vm_page_wakeup(m); 2633 return; 2634 } 2635 2636 /* 2637 * Already in the cache (and thus not mapped) 2638 */ 2639 if ((m->queue - m->pc) == PQ_CACHE) { 2640 KKASSERT((m->flags & PG_MAPPED) == 0); 2641 vm_page_wakeup(m); 2642 return; 2643 } 2644 2645 /* 2646 * Caller is required to test m->dirty, but note that the act of 2647 * removing the page from its maps can cause it to become dirty 2648 * on an SMP system due to another cpu running in usermode. 2649 */ 2650 if (m->dirty) { 2651 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2652 (long)m->pindex); 2653 } 2654 2655 /* 2656 * Remove all pmaps and indicate that the page is not 2657 * writeable or mapped. Our vm_page_protect() call may 2658 * have blocked (especially w/ VM_PROT_NONE), so recheck 2659 * everything. 2660 */ 2661 vm_page_protect(m, VM_PROT_NONE); 2662 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2663 (m->busy_count & PBUSY_MASK) || 2664 m->wire_count || m->hold_count) { 2665 vm_page_wakeup(m); 2666 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2667 vm_page_deactivate(m); 2668 vm_page_wakeup(m); 2669 } else { 2670 _vm_page_and_queue_spin_lock(m); 2671 _vm_page_rem_queue_spinlocked(m); 2672 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2673 _vm_page_queue_spin_unlock(m); 2674 if (_vm_page_wakeup(m)) { 2675 vm_page_spin_unlock(m); 2676 wakeup(m); 2677 } else { 2678 vm_page_spin_unlock(m); 2679 } 2680 vm_page_free_wakeup(); 2681 } 2682 } 2683 2684 /* 2685 * vm_page_dontneed() 2686 * 2687 * Cache, deactivate, or do nothing as appropriate. This routine 2688 * is typically used by madvise() MADV_DONTNEED. 2689 * 2690 * Generally speaking we want to move the page into the cache so 2691 * it gets reused quickly. However, this can result in a silly syndrome 2692 * due to the page recycling too quickly. Small objects will not be 2693 * fully cached. On the otherhand, if we move the page to the inactive 2694 * queue we wind up with a problem whereby very large objects 2695 * unnecessarily blow away our inactive and cache queues. 2696 * 2697 * The solution is to move the pages based on a fixed weighting. We 2698 * either leave them alone, deactivate them, or move them to the cache, 2699 * where moving them to the cache has the highest weighting. 2700 * By forcing some pages into other queues we eventually force the 2701 * system to balance the queues, potentially recovering other unrelated 2702 * space from active. The idea is to not force this to happen too 2703 * often. 2704 * 2705 * The page must be busied. 2706 */ 2707 void 2708 vm_page_dontneed(vm_page_t m) 2709 { 2710 static int dnweight; 2711 int dnw; 2712 int head; 2713 2714 dnw = ++dnweight; 2715 2716 /* 2717 * occassionally leave the page alone 2718 */ 2719 if ((dnw & 0x01F0) == 0 || 2720 m->queue - m->pc == PQ_INACTIVE || 2721 m->queue - m->pc == PQ_CACHE 2722 ) { 2723 if (m->act_count >= ACT_INIT) 2724 --m->act_count; 2725 return; 2726 } 2727 2728 /* 2729 * If vm_page_dontneed() is inactivating a page, it must clear 2730 * the referenced flag; otherwise the pagedaemon will see references 2731 * on the page in the inactive queue and reactivate it. Until the 2732 * page can move to the cache queue, madvise's job is not done. 2733 */ 2734 vm_page_flag_clear(m, PG_REFERENCED); 2735 pmap_clear_reference(m); 2736 2737 if (m->dirty == 0) 2738 vm_page_test_dirty(m); 2739 2740 if (m->dirty || (dnw & 0x0070) == 0) { 2741 /* 2742 * Deactivate the page 3 times out of 32. 2743 */ 2744 head = 0; 2745 } else { 2746 /* 2747 * Cache the page 28 times out of every 32. Note that 2748 * the page is deactivated instead of cached, but placed 2749 * at the head of the queue instead of the tail. 2750 */ 2751 head = 1; 2752 } 2753 vm_page_spin_lock(m); 2754 _vm_page_deactivate_locked(m, head); 2755 vm_page_spin_unlock(m); 2756 } 2757 2758 /* 2759 * These routines manipulate the 'soft busy' count for a page. A soft busy 2760 * is almost like a hard BUSY except that it allows certain compatible 2761 * operations to occur on the page while it is busy. For example, a page 2762 * undergoing a write can still be mapped read-only. 2763 * 2764 * We also use soft-busy to quickly pmap_enter shared read-only pages 2765 * without having to hold the page locked. 2766 * 2767 * The soft-busy count can be > 1 in situations where multiple threads 2768 * are pmap_enter()ing the same page simultaneously, or when two buffer 2769 * cache buffers overlap the same page. 2770 * 2771 * The caller must hold the page BUSY when making these two calls. 2772 */ 2773 void 2774 vm_page_io_start(vm_page_t m) 2775 { 2776 uint32_t ocount; 2777 2778 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2779 KKASSERT(ocount & PBUSY_LOCKED); 2780 } 2781 2782 void 2783 vm_page_io_finish(vm_page_t m) 2784 { 2785 uint32_t ocount; 2786 2787 ocount = atomic_fetchadd_int(&m->busy_count, -1); 2788 KKASSERT(ocount & PBUSY_MASK); 2789 #if 0 2790 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 2791 wakeup(m); 2792 #endif 2793 } 2794 2795 /* 2796 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 2797 * 2798 * Returns 0 on success, non-zero on failure. 2799 */ 2800 int 2801 vm_page_sbusy_try(vm_page_t m) 2802 { 2803 uint32_t ocount; 2804 2805 if (m->busy_count & PBUSY_LOCKED) 2806 return 1; 2807 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2808 if (ocount & PBUSY_LOCKED) { 2809 vm_page_sbusy_drop(m); 2810 return 1; 2811 } 2812 return 0; 2813 } 2814 2815 /* 2816 * Indicate that a clean VM page requires a filesystem commit and cannot 2817 * be reused. Used by tmpfs. 2818 */ 2819 void 2820 vm_page_need_commit(vm_page_t m) 2821 { 2822 vm_page_flag_set(m, PG_NEED_COMMIT); 2823 vm_object_set_writeable_dirty(m->object); 2824 } 2825 2826 void 2827 vm_page_clear_commit(vm_page_t m) 2828 { 2829 vm_page_flag_clear(m, PG_NEED_COMMIT); 2830 } 2831 2832 /* 2833 * Grab a page, blocking if it is busy and allocating a page if necessary. 2834 * A busy page is returned or NULL. The page may or may not be valid and 2835 * might not be on a queue (the caller is responsible for the disposition of 2836 * the page). 2837 * 2838 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2839 * page will be zero'd and marked valid. 2840 * 2841 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2842 * valid even if it already exists. 2843 * 2844 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2845 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2846 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2847 * 2848 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2849 * always returned if we had blocked. 2850 * 2851 * This routine may not be called from an interrupt. 2852 * 2853 * No other requirements. 2854 */ 2855 vm_page_t 2856 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2857 { 2858 vm_page_t m; 2859 int error; 2860 int shared = 1; 2861 2862 KKASSERT(allocflags & 2863 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2864 vm_object_hold_shared(object); 2865 for (;;) { 2866 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2867 if (error) { 2868 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2869 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2870 m = NULL; 2871 break; 2872 } 2873 /* retry */ 2874 } else if (m == NULL) { 2875 if (shared) { 2876 vm_object_upgrade(object); 2877 shared = 0; 2878 } 2879 if (allocflags & VM_ALLOC_RETRY) 2880 allocflags |= VM_ALLOC_NULL_OK; 2881 m = vm_page_alloc(object, pindex, 2882 allocflags & ~VM_ALLOC_RETRY); 2883 if (m) 2884 break; 2885 vm_wait(0); 2886 if ((allocflags & VM_ALLOC_RETRY) == 0) 2887 goto failed; 2888 } else { 2889 /* m found */ 2890 break; 2891 } 2892 } 2893 2894 /* 2895 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2896 * 2897 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2898 * valid even if already valid. 2899 * 2900 * NOTE! We have removed all of the PG_ZERO optimizations and also 2901 * removed the idle zeroing code. These optimizations actually 2902 * slow things down on modern cpus because the zerod area is 2903 * likely uncached, placing a memory-access burden on the 2904 * accesors taking the fault. 2905 * 2906 * By always zeroing the page in-line with the fault, no 2907 * dynamic ram reads are needed and the caches are hot, ready 2908 * for userland to access the memory. 2909 */ 2910 if (m->valid == 0) { 2911 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2912 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2913 m->valid = VM_PAGE_BITS_ALL; 2914 } 2915 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2916 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2917 m->valid = VM_PAGE_BITS_ALL; 2918 } 2919 failed: 2920 vm_object_drop(object); 2921 return(m); 2922 } 2923 2924 /* 2925 * Mapping function for valid bits or for dirty bits in 2926 * a page. May not block. 2927 * 2928 * Inputs are required to range within a page. 2929 * 2930 * No requirements. 2931 * Non blocking. 2932 */ 2933 int 2934 vm_page_bits(int base, int size) 2935 { 2936 int first_bit; 2937 int last_bit; 2938 2939 KASSERT( 2940 base + size <= PAGE_SIZE, 2941 ("vm_page_bits: illegal base/size %d/%d", base, size) 2942 ); 2943 2944 if (size == 0) /* handle degenerate case */ 2945 return(0); 2946 2947 first_bit = base >> DEV_BSHIFT; 2948 last_bit = (base + size - 1) >> DEV_BSHIFT; 2949 2950 return ((2 << last_bit) - (1 << first_bit)); 2951 } 2952 2953 /* 2954 * Sets portions of a page valid and clean. The arguments are expected 2955 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2956 * of any partial chunks touched by the range. The invalid portion of 2957 * such chunks will be zero'd. 2958 * 2959 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2960 * align base to DEV_BSIZE so as not to mark clean a partially 2961 * truncated device block. Otherwise the dirty page status might be 2962 * lost. 2963 * 2964 * This routine may not block. 2965 * 2966 * (base + size) must be less then or equal to PAGE_SIZE. 2967 */ 2968 static void 2969 _vm_page_zero_valid(vm_page_t m, int base, int size) 2970 { 2971 int frag; 2972 int endoff; 2973 2974 if (size == 0) /* handle degenerate case */ 2975 return; 2976 2977 /* 2978 * If the base is not DEV_BSIZE aligned and the valid 2979 * bit is clear, we have to zero out a portion of the 2980 * first block. 2981 */ 2982 2983 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2984 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 2985 ) { 2986 pmap_zero_page_area( 2987 VM_PAGE_TO_PHYS(m), 2988 frag, 2989 base - frag 2990 ); 2991 } 2992 2993 /* 2994 * If the ending offset is not DEV_BSIZE aligned and the 2995 * valid bit is clear, we have to zero out a portion of 2996 * the last block. 2997 */ 2998 2999 endoff = base + size; 3000 3001 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3002 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3003 ) { 3004 pmap_zero_page_area( 3005 VM_PAGE_TO_PHYS(m), 3006 endoff, 3007 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3008 ); 3009 } 3010 } 3011 3012 /* 3013 * Set valid, clear dirty bits. If validating the entire 3014 * page we can safely clear the pmap modify bit. We also 3015 * use this opportunity to clear the PG_NOSYNC flag. If a process 3016 * takes a write fault on a MAP_NOSYNC memory area the flag will 3017 * be set again. 3018 * 3019 * We set valid bits inclusive of any overlap, but we can only 3020 * clear dirty bits for DEV_BSIZE chunks that are fully within 3021 * the range. 3022 * 3023 * Page must be busied? 3024 * No other requirements. 3025 */ 3026 void 3027 vm_page_set_valid(vm_page_t m, int base, int size) 3028 { 3029 _vm_page_zero_valid(m, base, size); 3030 m->valid |= vm_page_bits(base, size); 3031 } 3032 3033 3034 /* 3035 * Set valid bits and clear dirty bits. 3036 * 3037 * Page must be busied by caller. 3038 * 3039 * NOTE: This function does not clear the pmap modified bit. 3040 * Also note that e.g. NFS may use a byte-granular base 3041 * and size. 3042 * 3043 * No other requirements. 3044 */ 3045 void 3046 vm_page_set_validclean(vm_page_t m, int base, int size) 3047 { 3048 int pagebits; 3049 3050 _vm_page_zero_valid(m, base, size); 3051 pagebits = vm_page_bits(base, size); 3052 m->valid |= pagebits; 3053 m->dirty &= ~pagebits; 3054 if (base == 0 && size == PAGE_SIZE) { 3055 /*pmap_clear_modify(m);*/ 3056 vm_page_flag_clear(m, PG_NOSYNC); 3057 } 3058 } 3059 3060 /* 3061 * Set valid & dirty. Used by buwrite() 3062 * 3063 * Page must be busied by caller. 3064 */ 3065 void 3066 vm_page_set_validdirty(vm_page_t m, int base, int size) 3067 { 3068 int pagebits; 3069 3070 pagebits = vm_page_bits(base, size); 3071 m->valid |= pagebits; 3072 m->dirty |= pagebits; 3073 if (m->object) 3074 vm_object_set_writeable_dirty(m->object); 3075 } 3076 3077 /* 3078 * Clear dirty bits. 3079 * 3080 * NOTE: This function does not clear the pmap modified bit. 3081 * Also note that e.g. NFS may use a byte-granular base 3082 * and size. 3083 * 3084 * Page must be busied? 3085 * No other requirements. 3086 */ 3087 void 3088 vm_page_clear_dirty(vm_page_t m, int base, int size) 3089 { 3090 m->dirty &= ~vm_page_bits(base, size); 3091 if (base == 0 && size == PAGE_SIZE) { 3092 /*pmap_clear_modify(m);*/ 3093 vm_page_flag_clear(m, PG_NOSYNC); 3094 } 3095 } 3096 3097 /* 3098 * Make the page all-dirty. 3099 * 3100 * Also make sure the related object and vnode reflect the fact that the 3101 * object may now contain a dirty page. 3102 * 3103 * Page must be busied? 3104 * No other requirements. 3105 */ 3106 void 3107 vm_page_dirty(vm_page_t m) 3108 { 3109 #ifdef INVARIANTS 3110 int pqtype = m->queue - m->pc; 3111 #endif 3112 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3113 ("vm_page_dirty: page in free/cache queue!")); 3114 if (m->dirty != VM_PAGE_BITS_ALL) { 3115 m->dirty = VM_PAGE_BITS_ALL; 3116 if (m->object) 3117 vm_object_set_writeable_dirty(m->object); 3118 } 3119 } 3120 3121 /* 3122 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3123 * valid and dirty bits for the effected areas are cleared. 3124 * 3125 * Page must be busied? 3126 * Does not block. 3127 * No other requirements. 3128 */ 3129 void 3130 vm_page_set_invalid(vm_page_t m, int base, int size) 3131 { 3132 int bits; 3133 3134 bits = vm_page_bits(base, size); 3135 m->valid &= ~bits; 3136 m->dirty &= ~bits; 3137 atomic_add_int(&m->object->generation, 1); 3138 } 3139 3140 /* 3141 * The kernel assumes that the invalid portions of a page contain 3142 * garbage, but such pages can be mapped into memory by user code. 3143 * When this occurs, we must zero out the non-valid portions of the 3144 * page so user code sees what it expects. 3145 * 3146 * Pages are most often semi-valid when the end of a file is mapped 3147 * into memory and the file's size is not page aligned. 3148 * 3149 * Page must be busied? 3150 * No other requirements. 3151 */ 3152 void 3153 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3154 { 3155 int b; 3156 int i; 3157 3158 /* 3159 * Scan the valid bits looking for invalid sections that 3160 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3161 * valid bit may be set ) have already been zerod by 3162 * vm_page_set_validclean(). 3163 */ 3164 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3165 if (i == (PAGE_SIZE / DEV_BSIZE) || 3166 (m->valid & (1 << i)) 3167 ) { 3168 if (i > b) { 3169 pmap_zero_page_area( 3170 VM_PAGE_TO_PHYS(m), 3171 b << DEV_BSHIFT, 3172 (i - b) << DEV_BSHIFT 3173 ); 3174 } 3175 b = i + 1; 3176 } 3177 } 3178 3179 /* 3180 * setvalid is TRUE when we can safely set the zero'd areas 3181 * as being valid. We can do this if there are no cache consistency 3182 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3183 */ 3184 if (setvalid) 3185 m->valid = VM_PAGE_BITS_ALL; 3186 } 3187 3188 /* 3189 * Is a (partial) page valid? Note that the case where size == 0 3190 * will return FALSE in the degenerate case where the page is entirely 3191 * invalid, and TRUE otherwise. 3192 * 3193 * Does not block. 3194 * No other requirements. 3195 */ 3196 int 3197 vm_page_is_valid(vm_page_t m, int base, int size) 3198 { 3199 int bits = vm_page_bits(base, size); 3200 3201 if (m->valid && ((m->valid & bits) == bits)) 3202 return 1; 3203 else 3204 return 0; 3205 } 3206 3207 /* 3208 * update dirty bits from pmap/mmu. May not block. 3209 * 3210 * Caller must hold the page busy 3211 */ 3212 void 3213 vm_page_test_dirty(vm_page_t m) 3214 { 3215 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3216 vm_page_dirty(m); 3217 } 3218 } 3219 3220 #include "opt_ddb.h" 3221 #ifdef DDB 3222 #include <ddb/ddb.h> 3223 3224 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3225 { 3226 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count); 3227 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count); 3228 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count); 3229 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count); 3230 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count); 3231 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved); 3232 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min); 3233 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target); 3234 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min); 3235 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target); 3236 } 3237 3238 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3239 { 3240 int i; 3241 db_printf("PQ_FREE:"); 3242 for (i = 0; i < PQ_L2_SIZE; i++) { 3243 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3244 } 3245 db_printf("\n"); 3246 3247 db_printf("PQ_CACHE:"); 3248 for(i = 0; i < PQ_L2_SIZE; i++) { 3249 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3250 } 3251 db_printf("\n"); 3252 3253 db_printf("PQ_ACTIVE:"); 3254 for(i = 0; i < PQ_L2_SIZE; i++) { 3255 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3256 } 3257 db_printf("\n"); 3258 3259 db_printf("PQ_INACTIVE:"); 3260 for(i = 0; i < PQ_L2_SIZE; i++) { 3261 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3262 } 3263 db_printf("\n"); 3264 } 3265 #endif /* DDB */ 3266