1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * This code is derived from software contributed to The DragonFly Project 10 * by Matthew Dillon <dillon@backplane.com> 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 38 */ 39 40 /* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 /* 67 * Resident memory management module. The module manipulates 'VM pages'. 68 * A VM page is the core building block for memory management. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/malloc.h> 74 #include <sys/proc.h> 75 #include <sys/vmmeter.h> 76 #include <sys/vnode.h> 77 #include <sys/kernel.h> 78 #include <sys/alist.h> 79 #include <sys/sysctl.h> 80 #include <sys/cpu_topology.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <sys/lock.h> 85 #include <vm/vm_kern.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pageout.h> 91 #include <vm/vm_pager.h> 92 #include <vm/vm_extern.h> 93 #include <vm/swap_pager.h> 94 95 #include <machine/inttypes.h> 96 #include <machine/md_var.h> 97 #include <machine/specialreg.h> 98 #include <machine/bus_dma.h> 99 100 #include <vm/vm_page2.h> 101 #include <sys/spinlock2.h> 102 103 /* 104 * SET - Minimum required set associative size, must be a power of 2. We 105 * want this to match or exceed the set-associativeness of the cpu. 106 * 107 * GRP - A larger set that allows bleed-over into the domains of other 108 * nearby cpus. Also must be a power of 2. Used by the page zeroing 109 * code to smooth things out a bit. 110 */ 111 #define PQ_SET_ASSOC 16 112 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1) 113 114 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2) 115 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1) 116 117 static void vm_page_queue_init(void); 118 static void vm_page_free_wakeup(void); 119 static vm_page_t vm_page_select_cache(u_short pg_color); 120 static vm_page_t _vm_page_list_find2(int basequeue, int index); 121 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 122 123 /* 124 * Array of tailq lists 125 */ 126 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 127 128 static volatile int vm_pages_waiting; 129 static struct alist vm_contig_alist; 130 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 131 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 132 133 static u_long vm_dma_reserved = 0; 134 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 135 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 136 "Memory reserved for DMA"); 137 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 138 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 139 140 static int vm_contig_verbose = 0; 141 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 142 143 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 144 vm_pindex_t, pindex); 145 146 static void 147 vm_page_queue_init(void) 148 { 149 int i; 150 151 for (i = 0; i < PQ_L2_SIZE; i++) 152 vm_page_queues[PQ_FREE+i].cnt_offset = 153 offsetof(struct vmstats, v_free_count); 154 for (i = 0; i < PQ_L2_SIZE; i++) 155 vm_page_queues[PQ_CACHE+i].cnt_offset = 156 offsetof(struct vmstats, v_cache_count); 157 for (i = 0; i < PQ_L2_SIZE; i++) 158 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 159 offsetof(struct vmstats, v_inactive_count); 160 for (i = 0; i < PQ_L2_SIZE; i++) 161 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 162 offsetof(struct vmstats, v_active_count); 163 for (i = 0; i < PQ_L2_SIZE; i++) 164 vm_page_queues[PQ_HOLD+i].cnt_offset = 165 offsetof(struct vmstats, v_active_count); 166 /* PQ_NONE has no queue */ 167 168 for (i = 0; i < PQ_COUNT; i++) { 169 TAILQ_INIT(&vm_page_queues[i].pl); 170 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 171 } 172 } 173 174 /* 175 * note: place in initialized data section? Is this necessary? 176 */ 177 vm_pindex_t first_page = 0; 178 vm_pindex_t vm_page_array_size = 0; 179 vm_page_t vm_page_array = NULL; 180 vm_paddr_t vm_low_phys_reserved; 181 182 /* 183 * (low level boot) 184 * 185 * Sets the page size, perhaps based upon the memory size. 186 * Must be called before any use of page-size dependent functions. 187 */ 188 void 189 vm_set_page_size(void) 190 { 191 if (vmstats.v_page_size == 0) 192 vmstats.v_page_size = PAGE_SIZE; 193 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 194 panic("vm_set_page_size: page size not a power of two"); 195 } 196 197 /* 198 * (low level boot) 199 * 200 * Add a new page to the freelist for use by the system. New pages 201 * are added to both the head and tail of the associated free page 202 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 203 * requests pull 'recent' adds (higher physical addresses) first. 204 * 205 * Beware that the page zeroing daemon will also be running soon after 206 * boot, moving pages from the head to the tail of the PQ_FREE queues. 207 * 208 * Must be called in a critical section. 209 */ 210 static void 211 vm_add_new_page(vm_paddr_t pa) 212 { 213 struct vpgqueues *vpq; 214 vm_page_t m; 215 216 m = PHYS_TO_VM_PAGE(pa); 217 m->phys_addr = pa; 218 m->flags = 0; 219 m->pat_mode = PAT_WRITE_BACK; 220 m->pc = (pa >> PAGE_SHIFT); 221 222 /* 223 * Twist for cpu localization in addition to page coloring, so 224 * different cpus selecting by m->queue get different page colors. 225 */ 226 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 227 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 228 m->pc &= PQ_L2_MASK; 229 230 /* 231 * Reserve a certain number of contiguous low memory pages for 232 * contigmalloc() to use. 233 */ 234 if (pa < vm_low_phys_reserved) { 235 atomic_add_long(&vmstats.v_page_count, 1); 236 atomic_add_long(&vmstats.v_dma_pages, 1); 237 m->queue = PQ_NONE; 238 m->wire_count = 1; 239 atomic_add_long(&vmstats.v_wire_count, 1); 240 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 241 return; 242 } 243 244 /* 245 * General page 246 */ 247 m->queue = m->pc + PQ_FREE; 248 KKASSERT(m->dirty == 0); 249 250 atomic_add_long(&vmstats.v_page_count, 1); 251 atomic_add_long(&vmstats.v_free_count, 1); 252 vpq = &vm_page_queues[m->queue]; 253 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 254 ++vpq->lcnt; 255 } 256 257 /* 258 * (low level boot) 259 * 260 * Initializes the resident memory module. 261 * 262 * Preallocates memory for critical VM structures and arrays prior to 263 * kernel_map becoming available. 264 * 265 * Memory is allocated from (virtual2_start, virtual2_end) if available, 266 * otherwise memory is allocated from (virtual_start, virtual_end). 267 * 268 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 269 * large enough to hold vm_page_array & other structures for machines with 270 * large amounts of ram, so we want to use virtual2* when available. 271 */ 272 void 273 vm_page_startup(void) 274 { 275 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 276 vm_offset_t mapped; 277 vm_pindex_t npages; 278 vm_paddr_t page_range; 279 vm_paddr_t new_end; 280 int i; 281 vm_paddr_t pa; 282 vm_paddr_t last_pa; 283 vm_paddr_t end; 284 vm_paddr_t biggestone, biggestsize; 285 vm_paddr_t total; 286 vm_page_t m; 287 288 total = 0; 289 biggestsize = 0; 290 biggestone = 0; 291 vaddr = round_page(vaddr); 292 293 /* 294 * Make sure ranges are page-aligned. 295 */ 296 for (i = 0; phys_avail[i].phys_end; ++i) { 297 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 298 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 299 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 300 phys_avail[i].phys_end = phys_avail[i].phys_beg; 301 } 302 303 /* 304 * Locate largest block 305 */ 306 for (i = 0; phys_avail[i].phys_end; ++i) { 307 vm_paddr_t size = phys_avail[i].phys_end - 308 phys_avail[i].phys_beg; 309 310 if (size > biggestsize) { 311 biggestone = i; 312 biggestsize = size; 313 } 314 total += size; 315 } 316 --i; /* adjust to last entry for use down below */ 317 318 end = phys_avail[biggestone].phys_end; 319 end = trunc_page(end); 320 321 /* 322 * Initialize the queue headers for the free queue, the active queue 323 * and the inactive queue. 324 */ 325 vm_page_queue_init(); 326 327 #if !defined(_KERNEL_VIRTUAL) 328 /* 329 * VKERNELs don't support minidumps and as such don't need 330 * vm_page_dump 331 * 332 * Allocate a bitmap to indicate that a random physical page 333 * needs to be included in a minidump. 334 * 335 * The amd64 port needs this to indicate which direct map pages 336 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 337 * 338 * However, x86 still needs this workspace internally within the 339 * minidump code. In theory, they are not needed on x86, but are 340 * included should the sf_buf code decide to use them. 341 */ 342 page_range = phys_avail[i].phys_end / PAGE_SIZE; 343 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 344 end -= vm_page_dump_size; 345 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 346 VM_PROT_READ | VM_PROT_WRITE); 347 bzero((void *)vm_page_dump, vm_page_dump_size); 348 #endif 349 /* 350 * Compute the number of pages of memory that will be available for 351 * use (taking into account the overhead of a page structure per 352 * page). 353 */ 354 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 355 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 356 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 357 358 #ifndef _KERNEL_VIRTUAL 359 /* 360 * (only applies to real kernels) 361 * 362 * Reserve a large amount of low memory for potential 32-bit DMA 363 * space allocations. Once device initialization is complete we 364 * release most of it, but keep (vm_dma_reserved) memory reserved 365 * for later use. Typically for X / graphics. Through trial and 366 * error we find that GPUs usually requires ~60-100MB or so. 367 * 368 * By default, 128M is left in reserve on machines with 2G+ of ram. 369 */ 370 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 371 if (vm_low_phys_reserved > total / 4) 372 vm_low_phys_reserved = total / 4; 373 if (vm_dma_reserved == 0) { 374 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 375 if (vm_dma_reserved > total / 16) 376 vm_dma_reserved = total / 16; 377 } 378 #endif 379 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 380 ALIST_RECORDS_65536); 381 382 /* 383 * Initialize the mem entry structures now, and put them in the free 384 * queue. 385 */ 386 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 387 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 388 vm_page_array = (vm_page_t)mapped; 389 390 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 391 /* 392 * since pmap_map on amd64 returns stuff out of a direct-map region, 393 * we have to manually add these pages to the minidump tracking so 394 * that they can be dumped, including the vm_page_array. 395 */ 396 for (pa = new_end; 397 pa < phys_avail[biggestone].phys_end; 398 pa += PAGE_SIZE) { 399 dump_add_page(pa); 400 } 401 #endif 402 403 /* 404 * Clear all of the page structures, run basic initialization so 405 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 406 * map. 407 */ 408 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 409 vm_page_array_size = page_range; 410 411 m = &vm_page_array[0]; 412 pa = ptoa(first_page); 413 for (i = 0; i < page_range; ++i) { 414 spin_init(&m->spin, "vm_page"); 415 m->phys_addr = pa; 416 pa += PAGE_SIZE; 417 ++m; 418 } 419 420 /* 421 * Construct the free queue(s) in ascending order (by physical 422 * address) so that the first 16MB of physical memory is allocated 423 * last rather than first. On large-memory machines, this avoids 424 * the exhaustion of low physical memory before isa_dma_init has run. 425 */ 426 vmstats.v_page_count = 0; 427 vmstats.v_free_count = 0; 428 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 429 pa = phys_avail[i].phys_beg; 430 if (i == biggestone) 431 last_pa = new_end; 432 else 433 last_pa = phys_avail[i].phys_end; 434 while (pa < last_pa && npages-- > 0) { 435 vm_add_new_page(pa); 436 pa += PAGE_SIZE; 437 } 438 } 439 if (virtual2_start) 440 virtual2_start = vaddr; 441 else 442 virtual_start = vaddr; 443 mycpu->gd_vmstats = vmstats; 444 } 445 446 /* 447 * Reorganize VM pages based on numa data. May be called as many times as 448 * necessary. Will reorganize the vm_page_t page color and related queue(s) 449 * to allow vm_page_alloc() to choose pages based on socket affinity. 450 * 451 * NOTE: This function is only called while we are still in UP mode, so 452 * we only need a critical section to protect the queues (which 453 * saves a lot of time, there are likely a ton of pages). 454 */ 455 void 456 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 457 { 458 vm_paddr_t scan_beg; 459 vm_paddr_t scan_end; 460 vm_paddr_t ran_end; 461 struct vpgqueues *vpq; 462 vm_page_t m; 463 vm_page_t mend; 464 int i; 465 int socket_mod; 466 int socket_value; 467 468 /* 469 * Check if no physical information, or there was only one socket 470 * (so don't waste time doing nothing!). 471 */ 472 if (cpu_topology_phys_ids <= 1 || 473 cpu_topology_core_ids == 0) { 474 return; 475 } 476 477 /* 478 * Setup for our iteration. Note that ACPI may iterate CPU 479 * sockets starting at 0 or 1 or some other number. The 480 * cpu_topology code mod's it against the socket count. 481 */ 482 ran_end = ran_beg + bytes; 483 physid %= cpu_topology_phys_ids; 484 485 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 486 socket_value = physid * socket_mod; 487 mend = &vm_page_array[vm_page_array_size]; 488 489 crit_enter(); 490 491 /* 492 * Adjust vm_page->pc and requeue all affected pages. The 493 * allocator will then be able to localize memory allocations 494 * to some degree. 495 */ 496 for (i = 0; phys_avail[i].phys_end; ++i) { 497 scan_beg = phys_avail[i].phys_beg; 498 scan_end = phys_avail[i].phys_end; 499 if (scan_end <= ran_beg) 500 continue; 501 if (scan_beg >= ran_end) 502 continue; 503 if (scan_beg < ran_beg) 504 scan_beg = ran_beg; 505 if (scan_end > ran_end) 506 scan_end = ran_end; 507 if (atop(scan_end) > first_page + vm_page_array_size) 508 scan_end = ptoa(first_page + vm_page_array_size); 509 510 m = PHYS_TO_VM_PAGE(scan_beg); 511 while (scan_beg < scan_end) { 512 KKASSERT(m < mend); 513 if (m->queue != PQ_NONE) { 514 vpq = &vm_page_queues[m->queue]; 515 TAILQ_REMOVE(&vpq->pl, m, pageq); 516 --vpq->lcnt; 517 /* queue doesn't change, no need to adj cnt */ 518 m->queue -= m->pc; 519 m->pc %= socket_mod; 520 m->pc += socket_value; 521 m->pc &= PQ_L2_MASK; 522 m->queue += m->pc; 523 vpq = &vm_page_queues[m->queue]; 524 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 525 ++vpq->lcnt; 526 /* queue doesn't change, no need to adj cnt */ 527 } else { 528 m->pc %= socket_mod; 529 m->pc += socket_value; 530 m->pc &= PQ_L2_MASK; 531 } 532 scan_beg += PAGE_SIZE; 533 ++m; 534 } 535 } 536 crit_exit(); 537 } 538 539 /* 540 * We tended to reserve a ton of memory for contigmalloc(). Now that most 541 * drivers have initialized we want to return most the remaining free 542 * reserve back to the VM page queues so they can be used for normal 543 * allocations. 544 * 545 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 546 */ 547 static void 548 vm_page_startup_finish(void *dummy __unused) 549 { 550 alist_blk_t blk; 551 alist_blk_t rblk; 552 alist_blk_t count; 553 alist_blk_t xcount; 554 alist_blk_t bfree; 555 vm_page_t m; 556 557 spin_lock(&vm_contig_spin); 558 for (;;) { 559 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 560 if (bfree <= vm_dma_reserved / PAGE_SIZE) 561 break; 562 if (count == 0) 563 break; 564 565 /* 566 * Figure out how much of the initial reserve we have to 567 * free in order to reach our target. 568 */ 569 bfree -= vm_dma_reserved / PAGE_SIZE; 570 if (count > bfree) { 571 blk += count - bfree; 572 count = bfree; 573 } 574 575 /* 576 * Calculate the nearest power of 2 <= count. 577 */ 578 for (xcount = 1; xcount <= count; xcount <<= 1) 579 ; 580 xcount >>= 1; 581 blk += count - xcount; 582 count = xcount; 583 584 /* 585 * Allocate the pages from the alist, then free them to 586 * the normal VM page queues. 587 * 588 * Pages allocated from the alist are wired. We have to 589 * busy, unwire, and free them. We must also adjust 590 * vm_low_phys_reserved before freeing any pages to prevent 591 * confusion. 592 */ 593 rblk = alist_alloc(&vm_contig_alist, blk, count); 594 if (rblk != blk) { 595 kprintf("vm_page_startup_finish: Unable to return " 596 "dma space @0x%08x/%d -> 0x%08x\n", 597 blk, count, rblk); 598 break; 599 } 600 atomic_add_long(&vmstats.v_dma_pages, -(long)count); 601 spin_unlock(&vm_contig_spin); 602 603 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 604 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 605 while (count) { 606 vm_page_busy_wait(m, FALSE, "cpgfr"); 607 vm_page_unwire(m, 0); 608 vm_page_free(m); 609 --count; 610 ++m; 611 } 612 spin_lock(&vm_contig_spin); 613 } 614 spin_unlock(&vm_contig_spin); 615 616 /* 617 * Print out how much DMA space drivers have already allocated and 618 * how much is left over. 619 */ 620 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 621 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 622 (PAGE_SIZE / 1024), 623 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 624 } 625 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 626 vm_page_startup_finish, NULL); 627 628 629 /* 630 * Scan comparison function for Red-Black tree scans. An inclusive 631 * (start,end) is expected. Other fields are not used. 632 */ 633 int 634 rb_vm_page_scancmp(struct vm_page *p, void *data) 635 { 636 struct rb_vm_page_scan_info *info = data; 637 638 if (p->pindex < info->start_pindex) 639 return(-1); 640 if (p->pindex > info->end_pindex) 641 return(1); 642 return(0); 643 } 644 645 int 646 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 647 { 648 if (p1->pindex < p2->pindex) 649 return(-1); 650 if (p1->pindex > p2->pindex) 651 return(1); 652 return(0); 653 } 654 655 void 656 vm_page_init(vm_page_t m) 657 { 658 /* do nothing for now. Called from pmap_page_init() */ 659 } 660 661 /* 662 * Each page queue has its own spin lock, which is fairly optimal for 663 * allocating and freeing pages at least. 664 * 665 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 666 * queue spinlock via this function. Also note that m->queue cannot change 667 * unless both the page and queue are locked. 668 */ 669 static __inline 670 void 671 _vm_page_queue_spin_lock(vm_page_t m) 672 { 673 u_short queue; 674 675 queue = m->queue; 676 if (queue != PQ_NONE) { 677 spin_lock(&vm_page_queues[queue].spin); 678 KKASSERT(queue == m->queue); 679 } 680 } 681 682 static __inline 683 void 684 _vm_page_queue_spin_unlock(vm_page_t m) 685 { 686 u_short queue; 687 688 queue = m->queue; 689 cpu_ccfence(); 690 if (queue != PQ_NONE) 691 spin_unlock(&vm_page_queues[queue].spin); 692 } 693 694 static __inline 695 void 696 _vm_page_queues_spin_lock(u_short queue) 697 { 698 cpu_ccfence(); 699 if (queue != PQ_NONE) 700 spin_lock(&vm_page_queues[queue].spin); 701 } 702 703 704 static __inline 705 void 706 _vm_page_queues_spin_unlock(u_short queue) 707 { 708 cpu_ccfence(); 709 if (queue != PQ_NONE) 710 spin_unlock(&vm_page_queues[queue].spin); 711 } 712 713 void 714 vm_page_queue_spin_lock(vm_page_t m) 715 { 716 _vm_page_queue_spin_lock(m); 717 } 718 719 void 720 vm_page_queues_spin_lock(u_short queue) 721 { 722 _vm_page_queues_spin_lock(queue); 723 } 724 725 void 726 vm_page_queue_spin_unlock(vm_page_t m) 727 { 728 _vm_page_queue_spin_unlock(m); 729 } 730 731 void 732 vm_page_queues_spin_unlock(u_short queue) 733 { 734 _vm_page_queues_spin_unlock(queue); 735 } 736 737 /* 738 * This locks the specified vm_page and its queue in the proper order 739 * (page first, then queue). The queue may change so the caller must 740 * recheck on return. 741 */ 742 static __inline 743 void 744 _vm_page_and_queue_spin_lock(vm_page_t m) 745 { 746 vm_page_spin_lock(m); 747 _vm_page_queue_spin_lock(m); 748 } 749 750 static __inline 751 void 752 _vm_page_and_queue_spin_unlock(vm_page_t m) 753 { 754 _vm_page_queues_spin_unlock(m->queue); 755 vm_page_spin_unlock(m); 756 } 757 758 void 759 vm_page_and_queue_spin_unlock(vm_page_t m) 760 { 761 _vm_page_and_queue_spin_unlock(m); 762 } 763 764 void 765 vm_page_and_queue_spin_lock(vm_page_t m) 766 { 767 _vm_page_and_queue_spin_lock(m); 768 } 769 770 /* 771 * Helper function removes vm_page from its current queue. 772 * Returns the base queue the page used to be on. 773 * 774 * The vm_page and the queue must be spinlocked. 775 * This function will unlock the queue but leave the page spinlocked. 776 */ 777 static __inline u_short 778 _vm_page_rem_queue_spinlocked(vm_page_t m) 779 { 780 struct vpgqueues *pq; 781 u_short queue; 782 u_short oqueue; 783 long *cnt; 784 785 queue = m->queue; 786 if (queue != PQ_NONE) { 787 pq = &vm_page_queues[queue]; 788 TAILQ_REMOVE(&pq->pl, m, pageq); 789 790 /* 791 * Adjust our pcpu stats. In order for the nominal low-memory 792 * algorithms to work properly we don't let any pcpu stat get 793 * too negative before we force it to be rolled-up into the 794 * global stats. Otherwise our pageout and vm_wait tests 795 * will fail badly. 796 * 797 * The idea here is to reduce unnecessary SMP cache 798 * mastership changes in the global vmstats, which can be 799 * particularly bad in multi-socket systems. 800 */ 801 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 802 atomic_add_long(cnt, -1); 803 if (*cnt < -VMMETER_SLOP_COUNT) { 804 u_long copy = atomic_swap_long(cnt, 0); 805 cnt = (long *)((char *)&vmstats + pq->cnt_offset); 806 atomic_add_long(cnt, copy); 807 cnt = (long *)((char *)&mycpu->gd_vmstats + 808 pq->cnt_offset); 809 atomic_add_long(cnt, copy); 810 } 811 pq->lcnt--; 812 m->queue = PQ_NONE; 813 oqueue = queue; 814 queue -= m->pc; 815 vm_page_queues_spin_unlock(oqueue); /* intended */ 816 } 817 return queue; 818 } 819 820 /* 821 * Helper function places the vm_page on the specified queue. Generally 822 * speaking only PQ_FREE pages are placed at the head, to allow them to 823 * be allocated sooner rather than later on the assumption that they 824 * are cache-hot. 825 * 826 * The vm_page must be spinlocked. 827 * This function will return with both the page and the queue locked. 828 */ 829 static __inline void 830 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 831 { 832 struct vpgqueues *pq; 833 u_long *cnt; 834 835 KKASSERT(m->queue == PQ_NONE); 836 837 if (queue != PQ_NONE) { 838 vm_page_queues_spin_lock(queue); 839 pq = &vm_page_queues[queue]; 840 ++pq->lcnt; 841 842 /* 843 * Adjust our pcpu stats. If a system entity really needs 844 * to incorporate the count it will call vmstats_rollup() 845 * to roll it all up into the global vmstats strufture. 846 */ 847 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 848 atomic_add_long(cnt, 1); 849 850 /* 851 * PQ_FREE is always handled LIFO style to try to provide 852 * cache-hot pages to programs. 853 */ 854 m->queue = queue; 855 if (queue - m->pc == PQ_FREE) { 856 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 857 } else if (athead) { 858 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 859 } else { 860 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 861 } 862 /* leave the queue spinlocked */ 863 } 864 } 865 866 /* 867 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 868 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 869 * 870 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 871 * call will be made before returning. 872 * 873 * This function does NOT busy the page and on return the page is not 874 * guaranteed to be available. 875 */ 876 void 877 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 878 { 879 u_int32_t busy_count; 880 881 for (;;) { 882 busy_count = m->busy_count; 883 cpu_ccfence(); 884 885 if ((busy_count & PBUSY_LOCKED) == 0 && 886 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 887 break; 888 } 889 tsleep_interlock(m, 0); 890 if (atomic_cmpset_int(&m->busy_count, busy_count, 891 busy_count | PBUSY_WANTED)) { 892 atomic_set_int(&m->flags, PG_REFERENCED); 893 tsleep(m, PINTERLOCKED, msg, 0); 894 break; 895 } 896 } 897 } 898 899 /* 900 * This calculates and returns a page color given an optional VM object and 901 * either a pindex or an iterator. We attempt to return a cpu-localized 902 * pg_color that is still roughly 16-way set-associative. The CPU topology 903 * is used if it was probed. 904 * 905 * The caller may use the returned value to index into e.g. PQ_FREE when 906 * allocating a page in order to nominally obtain pages that are hopefully 907 * already localized to the requesting cpu. This function is not able to 908 * provide any sort of guarantee of this, but does its best to improve 909 * hardware cache management performance. 910 * 911 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 912 */ 913 u_short 914 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 915 { 916 u_short pg_color; 917 int phys_id; 918 int core_id; 919 int object_pg_color; 920 921 phys_id = get_cpu_phys_id(cpuid); 922 core_id = get_cpu_core_id(cpuid); 923 object_pg_color = object ? object->pg_color : 0; 924 925 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 926 int grpsize; 927 928 /* 929 * Break us down by socket and cpu 930 */ 931 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids; 932 pg_color += core_id * PQ_L2_SIZE / 933 (cpu_topology_core_ids * cpu_topology_phys_ids); 934 935 /* 936 * Calculate remaining component for object/queue color 937 */ 938 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids * 939 cpu_topology_phys_ids); 940 if (grpsize >= 8) { 941 pg_color += (pindex + object_pg_color) % grpsize; 942 } else { 943 if (grpsize <= 2) { 944 grpsize = 8; 945 } else { 946 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 947 grpsize += grpsize; 948 if (grpsize < 8) 949 grpsize += grpsize; 950 } 951 pg_color += (pindex + object_pg_color) % grpsize; 952 } 953 } else { 954 /* 955 * Unknown topology, distribute things evenly. 956 */ 957 pg_color = cpuid * PQ_L2_SIZE / ncpus; 958 pg_color += pindex + object_pg_color; 959 } 960 return (pg_color & PQ_L2_MASK); 961 } 962 963 /* 964 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 965 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 966 */ 967 void 968 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 969 int also_m_busy, const char *msg 970 VM_PAGE_DEBUG_ARGS) 971 { 972 u_int32_t busy_count; 973 974 for (;;) { 975 busy_count = m->busy_count; 976 cpu_ccfence(); 977 if (busy_count & PBUSY_LOCKED) { 978 tsleep_interlock(m, 0); 979 if (atomic_cmpset_int(&m->busy_count, busy_count, 980 busy_count | PBUSY_WANTED)) { 981 atomic_set_int(&m->flags, PG_REFERENCED); 982 tsleep(m, PINTERLOCKED, msg, 0); 983 } 984 } else if (also_m_busy && busy_count) { 985 tsleep_interlock(m, 0); 986 if (atomic_cmpset_int(&m->busy_count, busy_count, 987 busy_count | PBUSY_WANTED)) { 988 atomic_set_int(&m->flags, PG_REFERENCED); 989 tsleep(m, PINTERLOCKED, msg, 0); 990 } 991 } else { 992 if (atomic_cmpset_int(&m->busy_count, busy_count, 993 busy_count | PBUSY_LOCKED)) { 994 #ifdef VM_PAGE_DEBUG 995 m->busy_func = func; 996 m->busy_line = lineno; 997 #endif 998 break; 999 } 1000 } 1001 } 1002 } 1003 1004 /* 1005 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1006 * m->busy_count is also 0. 1007 * 1008 * Returns non-zero on failure. 1009 */ 1010 int 1011 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1012 VM_PAGE_DEBUG_ARGS) 1013 { 1014 u_int32_t busy_count; 1015 1016 for (;;) { 1017 busy_count = m->busy_count; 1018 cpu_ccfence(); 1019 if (busy_count & PBUSY_LOCKED) 1020 return TRUE; 1021 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1022 return TRUE; 1023 if (atomic_cmpset_int(&m->busy_count, busy_count, 1024 busy_count | PBUSY_LOCKED)) { 1025 #ifdef VM_PAGE_DEBUG 1026 m->busy_func = func; 1027 m->busy_line = lineno; 1028 #endif 1029 return FALSE; 1030 } 1031 } 1032 } 1033 1034 /* 1035 * Clear the BUSY flag and return non-zero to indicate to the caller 1036 * that a wakeup() should be performed. 1037 * 1038 * The vm_page must be spinlocked and will remain spinlocked on return. 1039 * The related queue must NOT be spinlocked (which could deadlock us). 1040 * 1041 * (inline version) 1042 */ 1043 static __inline 1044 int 1045 _vm_page_wakeup(vm_page_t m) 1046 { 1047 u_int32_t busy_count; 1048 1049 for (;;) { 1050 busy_count = m->busy_count; 1051 cpu_ccfence(); 1052 if (atomic_cmpset_int(&m->busy_count, busy_count, 1053 busy_count & 1054 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1055 break; 1056 } 1057 } 1058 return((int)(busy_count & PBUSY_WANTED)); 1059 } 1060 1061 /* 1062 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1063 * is typically the last call you make on a page before moving onto 1064 * other things. 1065 */ 1066 void 1067 vm_page_wakeup(vm_page_t m) 1068 { 1069 KASSERT(m->busy_count & PBUSY_LOCKED, 1070 ("vm_page_wakeup: page not busy!!!")); 1071 vm_page_spin_lock(m); 1072 if (_vm_page_wakeup(m)) { 1073 vm_page_spin_unlock(m); 1074 wakeup(m); 1075 } else { 1076 vm_page_spin_unlock(m); 1077 } 1078 } 1079 1080 /* 1081 * Holding a page keeps it from being reused. Other parts of the system 1082 * can still disassociate the page from its current object and free it, or 1083 * perform read or write I/O on it and/or otherwise manipulate the page, 1084 * but if the page is held the VM system will leave the page and its data 1085 * intact and not reuse the page for other purposes until the last hold 1086 * reference is released. (see vm_page_wire() if you want to prevent the 1087 * page from being disassociated from its object too). 1088 * 1089 * The caller must still validate the contents of the page and, if necessary, 1090 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 1091 * before manipulating the page. 1092 * 1093 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 1094 */ 1095 void 1096 vm_page_hold(vm_page_t m) 1097 { 1098 vm_page_spin_lock(m); 1099 atomic_add_int(&m->hold_count, 1); 1100 if (m->queue - m->pc == PQ_FREE) { 1101 _vm_page_queue_spin_lock(m); 1102 _vm_page_rem_queue_spinlocked(m); 1103 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1104 _vm_page_queue_spin_unlock(m); 1105 } 1106 vm_page_spin_unlock(m); 1107 } 1108 1109 /* 1110 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1111 * it was freed while held and must be moved back to the FREE queue. 1112 */ 1113 void 1114 vm_page_unhold(vm_page_t m) 1115 { 1116 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1117 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 1118 m, m->hold_count, m->queue - m->pc)); 1119 vm_page_spin_lock(m); 1120 atomic_add_int(&m->hold_count, -1); 1121 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1122 _vm_page_queue_spin_lock(m); 1123 _vm_page_rem_queue_spinlocked(m); 1124 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1125 _vm_page_queue_spin_unlock(m); 1126 } 1127 vm_page_spin_unlock(m); 1128 } 1129 1130 /* 1131 * vm_page_getfake: 1132 * 1133 * Create a fictitious page with the specified physical address and 1134 * memory attribute. The memory attribute is the only the machine- 1135 * dependent aspect of a fictitious page that must be initialized. 1136 */ 1137 1138 void 1139 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1140 { 1141 1142 if ((m->flags & PG_FICTITIOUS) != 0) { 1143 /* 1144 * The page's memattr might have changed since the 1145 * previous initialization. Update the pmap to the 1146 * new memattr. 1147 */ 1148 goto memattr; 1149 } 1150 m->phys_addr = paddr; 1151 m->queue = PQ_NONE; 1152 /* Fictitious pages don't use "segind". */ 1153 /* Fictitious pages don't use "order" or "pool". */ 1154 m->flags = PG_FICTITIOUS | PG_UNMANAGED; 1155 m->busy_count = PBUSY_LOCKED; 1156 m->wire_count = 1; 1157 spin_init(&m->spin, "fake_page"); 1158 pmap_page_init(m); 1159 memattr: 1160 pmap_page_set_memattr(m, memattr); 1161 } 1162 1163 /* 1164 * Inserts the given vm_page into the object and object list. 1165 * 1166 * The pagetables are not updated but will presumably fault the page 1167 * in if necessary, or if a kernel page the caller will at some point 1168 * enter the page into the kernel's pmap. We are not allowed to block 1169 * here so we *can't* do this anyway. 1170 * 1171 * This routine may not block. 1172 * This routine must be called with the vm_object held. 1173 * This routine must be called with a critical section held. 1174 * 1175 * This routine returns TRUE if the page was inserted into the object 1176 * successfully, and FALSE if the page already exists in the object. 1177 */ 1178 int 1179 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1180 { 1181 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1182 if (m->object != NULL) 1183 panic("vm_page_insert: already inserted"); 1184 1185 atomic_add_int(&object->generation, 1); 1186 1187 /* 1188 * Record the object/offset pair in this page and add the 1189 * pv_list_count of the page to the object. 1190 * 1191 * The vm_page spin lock is required for interactions with the pmap. 1192 */ 1193 vm_page_spin_lock(m); 1194 m->object = object; 1195 m->pindex = pindex; 1196 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1197 m->object = NULL; 1198 m->pindex = 0; 1199 vm_page_spin_unlock(m); 1200 return FALSE; 1201 } 1202 ++object->resident_page_count; 1203 ++mycpu->gd_vmtotal.t_rm; 1204 vm_page_spin_unlock(m); 1205 1206 /* 1207 * Since we are inserting a new and possibly dirty page, 1208 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1209 */ 1210 if ((m->valid & m->dirty) || 1211 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1212 vm_object_set_writeable_dirty(object); 1213 1214 /* 1215 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1216 */ 1217 swap_pager_page_inserted(m); 1218 return TRUE; 1219 } 1220 1221 /* 1222 * Removes the given vm_page_t from the (object,index) table 1223 * 1224 * The underlying pmap entry (if any) is NOT removed here. 1225 * This routine may not block. 1226 * 1227 * The page must be BUSY and will remain BUSY on return. 1228 * No other requirements. 1229 * 1230 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1231 * it busy. 1232 */ 1233 void 1234 vm_page_remove(vm_page_t m) 1235 { 1236 vm_object_t object; 1237 1238 if (m->object == NULL) { 1239 return; 1240 } 1241 1242 if ((m->busy_count & PBUSY_LOCKED) == 0) 1243 panic("vm_page_remove: page not busy"); 1244 1245 object = m->object; 1246 1247 vm_object_hold(object); 1248 1249 /* 1250 * Remove the page from the object and update the object. 1251 * 1252 * The vm_page spin lock is required for interactions with the pmap. 1253 */ 1254 vm_page_spin_lock(m); 1255 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1256 --object->resident_page_count; 1257 --mycpu->gd_vmtotal.t_rm; 1258 m->object = NULL; 1259 atomic_add_int(&object->generation, 1); 1260 vm_page_spin_unlock(m); 1261 1262 vm_object_drop(object); 1263 } 1264 1265 /* 1266 * Locate and return the page at (object, pindex), or NULL if the 1267 * page could not be found. 1268 * 1269 * The caller must hold the vm_object token. 1270 */ 1271 vm_page_t 1272 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1273 { 1274 vm_page_t m; 1275 1276 /* 1277 * Search the hash table for this object/offset pair 1278 */ 1279 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1280 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1281 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1282 return(m); 1283 } 1284 1285 vm_page_t 1286 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1287 vm_pindex_t pindex, 1288 int also_m_busy, const char *msg 1289 VM_PAGE_DEBUG_ARGS) 1290 { 1291 u_int32_t busy_count; 1292 vm_page_t m; 1293 1294 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1295 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1296 while (m) { 1297 KKASSERT(m->object == object && m->pindex == pindex); 1298 busy_count = m->busy_count; 1299 cpu_ccfence(); 1300 if (busy_count & PBUSY_LOCKED) { 1301 tsleep_interlock(m, 0); 1302 if (atomic_cmpset_int(&m->busy_count, busy_count, 1303 busy_count | PBUSY_WANTED)) { 1304 atomic_set_int(&m->flags, PG_REFERENCED); 1305 tsleep(m, PINTERLOCKED, msg, 0); 1306 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1307 pindex); 1308 } 1309 } else if (also_m_busy && busy_count) { 1310 tsleep_interlock(m, 0); 1311 if (atomic_cmpset_int(&m->busy_count, busy_count, 1312 busy_count | PBUSY_WANTED)) { 1313 atomic_set_int(&m->flags, PG_REFERENCED); 1314 tsleep(m, PINTERLOCKED, msg, 0); 1315 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1316 pindex); 1317 } 1318 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1319 busy_count | PBUSY_LOCKED)) { 1320 #ifdef VM_PAGE_DEBUG 1321 m->busy_func = func; 1322 m->busy_line = lineno; 1323 #endif 1324 break; 1325 } 1326 } 1327 return m; 1328 } 1329 1330 /* 1331 * Attempt to lookup and busy a page. 1332 * 1333 * Returns NULL if the page could not be found 1334 * 1335 * Returns a vm_page and error == TRUE if the page exists but could not 1336 * be busied. 1337 * 1338 * Returns a vm_page and error == FALSE on success. 1339 */ 1340 vm_page_t 1341 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1342 vm_pindex_t pindex, 1343 int also_m_busy, int *errorp 1344 VM_PAGE_DEBUG_ARGS) 1345 { 1346 u_int32_t busy_count; 1347 vm_page_t m; 1348 1349 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1350 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1351 *errorp = FALSE; 1352 while (m) { 1353 KKASSERT(m->object == object && m->pindex == pindex); 1354 busy_count = m->busy_count; 1355 cpu_ccfence(); 1356 if (busy_count & PBUSY_LOCKED) { 1357 *errorp = TRUE; 1358 break; 1359 } 1360 if (also_m_busy && busy_count) { 1361 *errorp = TRUE; 1362 break; 1363 } 1364 if (atomic_cmpset_int(&m->busy_count, busy_count, 1365 busy_count | PBUSY_LOCKED)) { 1366 #ifdef VM_PAGE_DEBUG 1367 m->busy_func = func; 1368 m->busy_line = lineno; 1369 #endif 1370 break; 1371 } 1372 } 1373 return m; 1374 } 1375 1376 /* 1377 * Returns a page that is only soft-busied for use by the caller in 1378 * a read-only fashion. Returns NULL if the page could not be found, 1379 * the soft busy could not be obtained, or the page data is invalid. 1380 */ 1381 vm_page_t 1382 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1383 int pgoff, int pgbytes) 1384 { 1385 vm_page_t m; 1386 1387 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1388 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1389 if (m) { 1390 if ((m->valid != VM_PAGE_BITS_ALL && 1391 !vm_page_is_valid(m, pgoff, pgbytes)) || 1392 (m->flags & PG_FICTITIOUS)) { 1393 m = NULL; 1394 } else if (vm_page_sbusy_try(m)) { 1395 m = NULL; 1396 } else if ((m->valid != VM_PAGE_BITS_ALL && 1397 !vm_page_is_valid(m, pgoff, pgbytes)) || 1398 (m->flags & PG_FICTITIOUS)) { 1399 vm_page_sbusy_drop(m); 1400 m = NULL; 1401 } 1402 } 1403 return m; 1404 } 1405 1406 /* 1407 * Caller must hold the related vm_object 1408 */ 1409 vm_page_t 1410 vm_page_next(vm_page_t m) 1411 { 1412 vm_page_t next; 1413 1414 next = vm_page_rb_tree_RB_NEXT(m); 1415 if (next && next->pindex != m->pindex + 1) 1416 next = NULL; 1417 return (next); 1418 } 1419 1420 /* 1421 * vm_page_rename() 1422 * 1423 * Move the given vm_page from its current object to the specified 1424 * target object/offset. The page must be busy and will remain so 1425 * on return. 1426 * 1427 * new_object must be held. 1428 * This routine might block. XXX ? 1429 * 1430 * NOTE: Swap associated with the page must be invalidated by the move. We 1431 * have to do this for several reasons: (1) we aren't freeing the 1432 * page, (2) we are dirtying the page, (3) the VM system is probably 1433 * moving the page from object A to B, and will then later move 1434 * the backing store from A to B and we can't have a conflict. 1435 * 1436 * NOTE: We *always* dirty the page. It is necessary both for the 1437 * fact that we moved it, and because we may be invalidating 1438 * swap. If the page is on the cache, we have to deactivate it 1439 * or vm_page_dirty() will panic. Dirty pages are not allowed 1440 * on the cache. 1441 */ 1442 void 1443 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1444 { 1445 KKASSERT(m->busy_count & PBUSY_LOCKED); 1446 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1447 if (m->object) { 1448 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1449 vm_page_remove(m); 1450 } 1451 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1452 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1453 new_object, new_pindex); 1454 } 1455 if (m->queue - m->pc == PQ_CACHE) 1456 vm_page_deactivate(m); 1457 vm_page_dirty(m); 1458 } 1459 1460 /* 1461 * vm_page_unqueue() without any wakeup. This routine is used when a page 1462 * is to remain BUSYied by the caller. 1463 * 1464 * This routine may not block. 1465 */ 1466 void 1467 vm_page_unqueue_nowakeup(vm_page_t m) 1468 { 1469 vm_page_and_queue_spin_lock(m); 1470 (void)_vm_page_rem_queue_spinlocked(m); 1471 vm_page_spin_unlock(m); 1472 } 1473 1474 /* 1475 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1476 * if necessary. 1477 * 1478 * This routine may not block. 1479 */ 1480 void 1481 vm_page_unqueue(vm_page_t m) 1482 { 1483 u_short queue; 1484 1485 vm_page_and_queue_spin_lock(m); 1486 queue = _vm_page_rem_queue_spinlocked(m); 1487 if (queue == PQ_FREE || queue == PQ_CACHE) { 1488 vm_page_spin_unlock(m); 1489 pagedaemon_wakeup(); 1490 } else { 1491 vm_page_spin_unlock(m); 1492 } 1493 } 1494 1495 /* 1496 * vm_page_list_find() 1497 * 1498 * Find a page on the specified queue with color optimization. 1499 * 1500 * The page coloring optimization attempts to locate a page that does 1501 * not overload other nearby pages in the object in the cpu's L1 or L2 1502 * caches. We need this optimization because cpu caches tend to be 1503 * physical caches, while object spaces tend to be virtual. 1504 * 1505 * The page coloring optimization also, very importantly, tries to localize 1506 * memory to cpus and physical sockets. 1507 * 1508 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1509 * and the algorithm is adjusted to localize allocations on a per-core basis. 1510 * This is done by 'twisting' the colors. 1511 * 1512 * The page is returned spinlocked and removed from its queue (it will 1513 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1514 * is responsible for dealing with the busy-page case (usually by 1515 * deactivating the page and looping). 1516 * 1517 * NOTE: This routine is carefully inlined. A non-inlined version 1518 * is available for outside callers but the only critical path is 1519 * from within this source file. 1520 * 1521 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1522 * represent stable storage, allowing us to order our locks vm_page 1523 * first, then queue. 1524 */ 1525 static __inline 1526 vm_page_t 1527 _vm_page_list_find(int basequeue, int index) 1528 { 1529 vm_page_t m; 1530 1531 for (;;) { 1532 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1533 if (m == NULL) { 1534 m = _vm_page_list_find2(basequeue, index); 1535 return(m); 1536 } 1537 vm_page_and_queue_spin_lock(m); 1538 if (m->queue == basequeue + index) { 1539 _vm_page_rem_queue_spinlocked(m); 1540 /* vm_page_t spin held, no queue spin */ 1541 break; 1542 } 1543 vm_page_and_queue_spin_unlock(m); 1544 } 1545 return(m); 1546 } 1547 1548 /* 1549 * If we could not find the page in the desired queue try to find it in 1550 * a nearby queue. 1551 */ 1552 static vm_page_t 1553 _vm_page_list_find2(int basequeue, int index) 1554 { 1555 struct vpgqueues *pq; 1556 vm_page_t m = NULL; 1557 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1558 int pqi; 1559 int i; 1560 1561 index &= PQ_L2_MASK; 1562 pq = &vm_page_queues[basequeue]; 1563 1564 /* 1565 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1566 * else fails (PQ_L2_MASK which is 255). 1567 */ 1568 do { 1569 pqmask = (pqmask << 1) | 1; 1570 for (i = 0; i <= pqmask; ++i) { 1571 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1572 m = TAILQ_FIRST(&pq[pqi].pl); 1573 if (m) { 1574 _vm_page_and_queue_spin_lock(m); 1575 if (m->queue == basequeue + pqi) { 1576 _vm_page_rem_queue_spinlocked(m); 1577 return(m); 1578 } 1579 _vm_page_and_queue_spin_unlock(m); 1580 --i; 1581 continue; 1582 } 1583 } 1584 } while (pqmask != PQ_L2_MASK); 1585 1586 return(m); 1587 } 1588 1589 /* 1590 * Returns a vm_page candidate for allocation. The page is not busied so 1591 * it can move around. The caller must busy the page (and typically 1592 * deactivate it if it cannot be busied!) 1593 * 1594 * Returns a spinlocked vm_page that has been removed from its queue. 1595 */ 1596 vm_page_t 1597 vm_page_list_find(int basequeue, int index) 1598 { 1599 return(_vm_page_list_find(basequeue, index)); 1600 } 1601 1602 /* 1603 * Find a page on the cache queue with color optimization, remove it 1604 * from the queue, and busy it. The returned page will not be spinlocked. 1605 * 1606 * A candidate failure will be deactivated. Candidates can fail due to 1607 * being busied by someone else, in which case they will be deactivated. 1608 * 1609 * This routine may not block. 1610 * 1611 */ 1612 static vm_page_t 1613 vm_page_select_cache(u_short pg_color) 1614 { 1615 vm_page_t m; 1616 1617 for (;;) { 1618 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1619 if (m == NULL) 1620 break; 1621 /* 1622 * (m) has been removed from its queue and spinlocked 1623 */ 1624 if (vm_page_busy_try(m, TRUE)) { 1625 _vm_page_deactivate_locked(m, 0); 1626 vm_page_spin_unlock(m); 1627 } else { 1628 /* 1629 * We successfully busied the page 1630 */ 1631 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1632 m->hold_count == 0 && 1633 m->wire_count == 0 && 1634 (m->dirty & m->valid) == 0) { 1635 vm_page_spin_unlock(m); 1636 pagedaemon_wakeup(); 1637 return(m); 1638 } 1639 1640 /* 1641 * The page cannot be recycled, deactivate it. 1642 */ 1643 _vm_page_deactivate_locked(m, 0); 1644 if (_vm_page_wakeup(m)) { 1645 vm_page_spin_unlock(m); 1646 wakeup(m); 1647 } else { 1648 vm_page_spin_unlock(m); 1649 } 1650 } 1651 } 1652 return (m); 1653 } 1654 1655 /* 1656 * Find a free page. We attempt to inline the nominal case and fall back 1657 * to _vm_page_select_free() otherwise. A busied page is removed from 1658 * the queue and returned. 1659 * 1660 * This routine may not block. 1661 */ 1662 static __inline vm_page_t 1663 vm_page_select_free(u_short pg_color) 1664 { 1665 vm_page_t m; 1666 1667 for (;;) { 1668 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1669 if (m == NULL) 1670 break; 1671 if (vm_page_busy_try(m, TRUE)) { 1672 /* 1673 * Various mechanisms such as a pmap_collect can 1674 * result in a busy page on the free queue. We 1675 * have to move the page out of the way so we can 1676 * retry the allocation. If the other thread is not 1677 * allocating the page then m->valid will remain 0 and 1678 * the pageout daemon will free the page later on. 1679 * 1680 * Since we could not busy the page, however, we 1681 * cannot make assumptions as to whether the page 1682 * will be allocated by the other thread or not, 1683 * so all we can do is deactivate it to move it out 1684 * of the way. In particular, if the other thread 1685 * wires the page it may wind up on the inactive 1686 * queue and the pageout daemon will have to deal 1687 * with that case too. 1688 */ 1689 _vm_page_deactivate_locked(m, 0); 1690 vm_page_spin_unlock(m); 1691 } else { 1692 /* 1693 * Theoretically if we are able to busy the page 1694 * atomic with the queue removal (using the vm_page 1695 * lock) nobody else should be able to mess with the 1696 * page before us. 1697 */ 1698 KKASSERT((m->flags & (PG_UNMANAGED | 1699 PG_NEED_COMMIT)) == 0); 1700 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1701 "pg %p q=%d flags=%08x hold=%d wire=%d", 1702 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1703 KKASSERT(m->wire_count == 0); 1704 vm_page_spin_unlock(m); 1705 pagedaemon_wakeup(); 1706 1707 /* return busied and removed page */ 1708 return(m); 1709 } 1710 } 1711 return(m); 1712 } 1713 1714 /* 1715 * vm_page_alloc() 1716 * 1717 * Allocate and return a memory cell associated with this VM object/offset 1718 * pair. If object is NULL an unassociated page will be allocated. 1719 * 1720 * The returned page will be busied and removed from its queues. This 1721 * routine can block and may return NULL if a race occurs and the page 1722 * is found to already exist at the specified (object, pindex). 1723 * 1724 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1725 * VM_ALLOC_QUICK like normal but cannot use cache 1726 * VM_ALLOC_SYSTEM greater free drain 1727 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1728 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1729 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1730 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1731 * (see vm_page_grab()) 1732 * VM_ALLOC_USE_GD ok to use per-gd cache 1733 * 1734 * VM_ALLOC_CPU(n) allocate using specified cpu localization 1735 * 1736 * The object must be held if not NULL 1737 * This routine may not block 1738 * 1739 * Additional special handling is required when called from an interrupt 1740 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1741 * in this case. 1742 */ 1743 vm_page_t 1744 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1745 { 1746 globaldata_t gd; 1747 vm_object_t obj; 1748 vm_page_t m; 1749 u_short pg_color; 1750 int cpuid_local; 1751 1752 #if 0 1753 /* 1754 * Special per-cpu free VM page cache. The pages are pre-busied 1755 * and pre-zerod for us. 1756 */ 1757 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1758 crit_enter_gd(gd); 1759 if (gd->gd_vmpg_count) { 1760 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1761 crit_exit_gd(gd); 1762 goto done; 1763 } 1764 crit_exit_gd(gd); 1765 } 1766 #endif 1767 m = NULL; 1768 1769 /* 1770 * CPU LOCALIZATION 1771 * 1772 * CPU localization algorithm. Break the page queues up by physical 1773 * id and core id (note that two cpu threads will have the same core 1774 * id, and core_id != gd_cpuid). 1775 * 1776 * This is nowhere near perfect, for example the last pindex in a 1777 * subgroup will overflow into the next cpu or package. But this 1778 * should get us good page reuse locality in heavy mixed loads. 1779 * 1780 * (may be executed before the APs are started, so other GDs might 1781 * not exist!) 1782 */ 1783 if (page_req & VM_ALLOC_CPU_SPEC) 1784 cpuid_local = VM_ALLOC_GETCPU(page_req); 1785 else 1786 cpuid_local = mycpu->gd_cpuid; 1787 1788 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 1789 1790 KKASSERT(page_req & 1791 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1792 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1793 1794 /* 1795 * Certain system threads (pageout daemon, buf_daemon's) are 1796 * allowed to eat deeper into the free page list. 1797 */ 1798 if (curthread->td_flags & TDF_SYSTHREAD) 1799 page_req |= VM_ALLOC_SYSTEM; 1800 1801 /* 1802 * Impose various limitations. Note that the v_free_reserved test 1803 * must match the opposite of vm_page_count_target() to avoid 1804 * livelocks, be careful. 1805 */ 1806 loop: 1807 gd = mycpu; 1808 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 1809 ((page_req & VM_ALLOC_INTERRUPT) && 1810 gd->gd_vmstats.v_free_count > 0) || 1811 ((page_req & VM_ALLOC_SYSTEM) && 1812 gd->gd_vmstats.v_cache_count == 0 && 1813 gd->gd_vmstats.v_free_count > 1814 gd->gd_vmstats.v_interrupt_free_min) 1815 ) { 1816 /* 1817 * The free queue has sufficient free pages to take one out. 1818 */ 1819 m = vm_page_select_free(pg_color); 1820 } else if (page_req & VM_ALLOC_NORMAL) { 1821 /* 1822 * Allocatable from the cache (non-interrupt only). On 1823 * success, we must free the page and try again, thus 1824 * ensuring that vmstats.v_*_free_min counters are replenished. 1825 */ 1826 #ifdef INVARIANTS 1827 if (curthread->td_preempted) { 1828 kprintf("vm_page_alloc(): warning, attempt to allocate" 1829 " cache page from preempting interrupt\n"); 1830 m = NULL; 1831 } else { 1832 m = vm_page_select_cache(pg_color); 1833 } 1834 #else 1835 m = vm_page_select_cache(pg_color); 1836 #endif 1837 /* 1838 * On success move the page into the free queue and loop. 1839 * 1840 * Only do this if we can safely acquire the vm_object lock, 1841 * because this is effectively a random page and the caller 1842 * might be holding the lock shared, we don't want to 1843 * deadlock. 1844 */ 1845 if (m != NULL) { 1846 KASSERT(m->dirty == 0, 1847 ("Found dirty cache page %p", m)); 1848 if ((obj = m->object) != NULL) { 1849 if (vm_object_hold_try(obj)) { 1850 vm_page_protect(m, VM_PROT_NONE); 1851 vm_page_free(m); 1852 /* m->object NULL here */ 1853 vm_object_drop(obj); 1854 } else { 1855 vm_page_deactivate(m); 1856 vm_page_wakeup(m); 1857 } 1858 } else { 1859 vm_page_protect(m, VM_PROT_NONE); 1860 vm_page_free(m); 1861 } 1862 goto loop; 1863 } 1864 1865 /* 1866 * On failure return NULL 1867 */ 1868 atomic_add_int(&vm_pageout_deficit, 1); 1869 pagedaemon_wakeup(); 1870 return (NULL); 1871 } else { 1872 /* 1873 * No pages available, wakeup the pageout daemon and give up. 1874 */ 1875 atomic_add_int(&vm_pageout_deficit, 1); 1876 pagedaemon_wakeup(); 1877 return (NULL); 1878 } 1879 1880 /* 1881 * v_free_count can race so loop if we don't find the expected 1882 * page. 1883 */ 1884 if (m == NULL) { 1885 vmstats_rollup(); 1886 goto loop; 1887 } 1888 1889 /* 1890 * Good page found. The page has already been busied for us and 1891 * removed from its queues. 1892 */ 1893 KASSERT(m->dirty == 0, 1894 ("vm_page_alloc: free/cache page %p was dirty", m)); 1895 KKASSERT(m->queue == PQ_NONE); 1896 1897 #if 0 1898 done: 1899 #endif 1900 /* 1901 * Initialize the structure, inheriting some flags but clearing 1902 * all the rest. The page has already been busied for us. 1903 */ 1904 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 1905 1906 KKASSERT(m->wire_count == 0); 1907 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 1908 m->act_count = 0; 1909 m->valid = 0; 1910 1911 /* 1912 * Caller must be holding the object lock (asserted by 1913 * vm_page_insert()). 1914 * 1915 * NOTE: Inserting a page here does not insert it into any pmaps 1916 * (which could cause us to block allocating memory). 1917 * 1918 * NOTE: If no object an unassociated page is allocated, m->pindex 1919 * can be used by the caller for any purpose. 1920 */ 1921 if (object) { 1922 if (vm_page_insert(m, object, pindex) == FALSE) { 1923 vm_page_free(m); 1924 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1925 panic("PAGE RACE %p[%ld]/%p", 1926 object, (long)pindex, m); 1927 m = NULL; 1928 } 1929 } else { 1930 m->pindex = pindex; 1931 } 1932 1933 /* 1934 * Don't wakeup too often - wakeup the pageout daemon when 1935 * we would be nearly out of memory. 1936 */ 1937 pagedaemon_wakeup(); 1938 1939 /* 1940 * A BUSY page is returned. 1941 */ 1942 return (m); 1943 } 1944 1945 /* 1946 * Returns number of pages available in our DMA memory reserve 1947 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1948 */ 1949 vm_size_t 1950 vm_contig_avail_pages(void) 1951 { 1952 alist_blk_t blk; 1953 alist_blk_t count; 1954 alist_blk_t bfree; 1955 spin_lock(&vm_contig_spin); 1956 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1957 spin_unlock(&vm_contig_spin); 1958 1959 return bfree; 1960 } 1961 1962 /* 1963 * Attempt to allocate contiguous physical memory with the specified 1964 * requirements. 1965 */ 1966 vm_page_t 1967 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1968 unsigned long alignment, unsigned long boundary, 1969 unsigned long size, vm_memattr_t memattr) 1970 { 1971 alist_blk_t blk; 1972 vm_page_t m; 1973 vm_pindex_t i; 1974 #if 0 1975 static vm_pindex_t contig_rover; 1976 #endif 1977 1978 alignment >>= PAGE_SHIFT; 1979 if (alignment == 0) 1980 alignment = 1; 1981 boundary >>= PAGE_SHIFT; 1982 if (boundary == 0) 1983 boundary = 1; 1984 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1985 1986 #if 0 1987 /* 1988 * Disabled temporarily until we find a solution for DRM (a flag 1989 * to always use the free space reserve, for performance). 1990 */ 1991 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 1992 boundary <= PAGE_SIZE && size == 1 && 1993 memattr == VM_MEMATTR_DEFAULT) { 1994 /* 1995 * Any page will work, use vm_page_alloc() 1996 * (e.g. when used from kmem_alloc_attr()) 1997 */ 1998 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 1999 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 2000 VM_ALLOC_INTERRUPT); 2001 m->valid = VM_PAGE_BITS_ALL; 2002 vm_page_wire(m); 2003 vm_page_wakeup(m); 2004 } else 2005 #endif 2006 { 2007 /* 2008 * Use the low-memory dma reserve 2009 */ 2010 spin_lock(&vm_contig_spin); 2011 blk = alist_alloc(&vm_contig_alist, 0, size); 2012 if (blk == ALIST_BLOCK_NONE) { 2013 spin_unlock(&vm_contig_spin); 2014 if (bootverbose) { 2015 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2016 (size << PAGE_SHIFT) / 1024); 2017 print_backtrace(5); 2018 } 2019 return(NULL); 2020 } 2021 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2022 alist_free(&vm_contig_alist, blk, size); 2023 spin_unlock(&vm_contig_spin); 2024 if (bootverbose) { 2025 kprintf("vm_page_alloc_contig: %ldk high " 2026 "%016jx failed\n", 2027 (size << PAGE_SHIFT) / 1024, 2028 (intmax_t)high); 2029 } 2030 return(NULL); 2031 } 2032 spin_unlock(&vm_contig_spin); 2033 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2034 } 2035 if (vm_contig_verbose) { 2036 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2037 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2038 (intmax_t)m->phys_addr, 2039 (size << PAGE_SHIFT) / 1024, 2040 low, high, alignment, boundary, size, memattr); 2041 } 2042 if (memattr != VM_MEMATTR_DEFAULT) { 2043 for (i = 0;i < size; i++) 2044 pmap_page_set_memattr(&m[i], memattr); 2045 } 2046 return m; 2047 } 2048 2049 /* 2050 * Free contiguously allocated pages. The pages will be wired but not busy. 2051 * When freeing to the alist we leave them wired and not busy. 2052 */ 2053 void 2054 vm_page_free_contig(vm_page_t m, unsigned long size) 2055 { 2056 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2057 vm_pindex_t start = pa >> PAGE_SHIFT; 2058 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2059 2060 if (vm_contig_verbose) { 2061 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2062 (intmax_t)pa, size / 1024); 2063 } 2064 if (pa < vm_low_phys_reserved) { 2065 KKASSERT(pa + size <= vm_low_phys_reserved); 2066 spin_lock(&vm_contig_spin); 2067 alist_free(&vm_contig_alist, start, pages); 2068 spin_unlock(&vm_contig_spin); 2069 } else { 2070 while (pages) { 2071 vm_page_busy_wait(m, FALSE, "cpgfr"); 2072 vm_page_unwire(m, 0); 2073 vm_page_free(m); 2074 --pages; 2075 ++m; 2076 } 2077 2078 } 2079 } 2080 2081 2082 /* 2083 * Wait for sufficient free memory for nominal heavy memory use kernel 2084 * operations. 2085 * 2086 * WARNING! Be sure never to call this in any vm_pageout code path, which 2087 * will trivially deadlock the system. 2088 */ 2089 void 2090 vm_wait_nominal(void) 2091 { 2092 while (vm_page_count_min(0)) 2093 vm_wait(0); 2094 } 2095 2096 /* 2097 * Test if vm_wait_nominal() would block. 2098 */ 2099 int 2100 vm_test_nominal(void) 2101 { 2102 if (vm_page_count_min(0)) 2103 return(1); 2104 return(0); 2105 } 2106 2107 /* 2108 * Block until free pages are available for allocation, called in various 2109 * places before memory allocations. 2110 * 2111 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2112 * more generous then that. 2113 */ 2114 void 2115 vm_wait(int timo) 2116 { 2117 /* 2118 * never wait forever 2119 */ 2120 if (timo == 0) 2121 timo = hz; 2122 lwkt_gettoken(&vm_token); 2123 2124 if (curthread == pagethread || 2125 curthread == emergpager) { 2126 /* 2127 * The pageout daemon itself needs pages, this is bad. 2128 */ 2129 if (vm_page_count_min(0)) { 2130 vm_pageout_pages_needed = 1; 2131 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2132 } 2133 } else { 2134 /* 2135 * Wakeup the pageout daemon if necessary and wait. 2136 * 2137 * Do not wait indefinitely for the target to be reached, 2138 * as load might prevent it from being reached any time soon. 2139 * But wait a little to try to slow down page allocations 2140 * and to give more important threads (the pagedaemon) 2141 * allocation priority. 2142 */ 2143 if (vm_page_count_target()) { 2144 if (vm_pages_needed == 0) { 2145 vm_pages_needed = 1; 2146 wakeup(&vm_pages_needed); 2147 } 2148 ++vm_pages_waiting; /* SMP race ok */ 2149 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2150 } 2151 } 2152 lwkt_reltoken(&vm_token); 2153 } 2154 2155 /* 2156 * Block until free pages are available for allocation 2157 * 2158 * Called only from vm_fault so that processes page faulting can be 2159 * easily tracked. 2160 */ 2161 void 2162 vm_wait_pfault(void) 2163 { 2164 /* 2165 * Wakeup the pageout daemon if necessary and wait. 2166 * 2167 * Do not wait indefinitely for the target to be reached, 2168 * as load might prevent it from being reached any time soon. 2169 * But wait a little to try to slow down page allocations 2170 * and to give more important threads (the pagedaemon) 2171 * allocation priority. 2172 */ 2173 if (vm_page_count_min(0)) { 2174 lwkt_gettoken(&vm_token); 2175 while (vm_page_count_severe()) { 2176 if (vm_page_count_target()) { 2177 thread_t td; 2178 2179 if (vm_pages_needed == 0) { 2180 vm_pages_needed = 1; 2181 wakeup(&vm_pages_needed); 2182 } 2183 ++vm_pages_waiting; /* SMP race ok */ 2184 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2185 2186 /* 2187 * Do not stay stuck in the loop if the system is trying 2188 * to kill the process. 2189 */ 2190 td = curthread; 2191 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2192 break; 2193 } 2194 } 2195 lwkt_reltoken(&vm_token); 2196 } 2197 } 2198 2199 /* 2200 * Put the specified page on the active list (if appropriate). Ensure 2201 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2202 * 2203 * The caller should be holding the page busied ? XXX 2204 * This routine may not block. 2205 */ 2206 void 2207 vm_page_activate(vm_page_t m) 2208 { 2209 u_short oqueue; 2210 2211 vm_page_spin_lock(m); 2212 if (m->queue - m->pc != PQ_ACTIVE) { 2213 _vm_page_queue_spin_lock(m); 2214 oqueue = _vm_page_rem_queue_spinlocked(m); 2215 /* page is left spinlocked, queue is unlocked */ 2216 2217 if (oqueue == PQ_CACHE) 2218 mycpu->gd_cnt.v_reactivated++; 2219 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2220 if (m->act_count < ACT_INIT) 2221 m->act_count = ACT_INIT; 2222 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2223 } 2224 _vm_page_and_queue_spin_unlock(m); 2225 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2226 pagedaemon_wakeup(); 2227 } else { 2228 if (m->act_count < ACT_INIT) 2229 m->act_count = ACT_INIT; 2230 vm_page_spin_unlock(m); 2231 } 2232 } 2233 2234 /* 2235 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2236 * routine is called when a page has been added to the cache or free 2237 * queues. 2238 * 2239 * This routine may not block. 2240 */ 2241 static __inline void 2242 vm_page_free_wakeup(void) 2243 { 2244 globaldata_t gd = mycpu; 2245 2246 /* 2247 * If the pageout daemon itself needs pages, then tell it that 2248 * there are some free. 2249 */ 2250 if (vm_pageout_pages_needed && 2251 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2252 gd->gd_vmstats.v_pageout_free_min 2253 ) { 2254 vm_pageout_pages_needed = 0; 2255 wakeup(&vm_pageout_pages_needed); 2256 } 2257 2258 /* 2259 * Wakeup processes that are waiting on memory. 2260 * 2261 * Generally speaking we want to wakeup stuck processes as soon as 2262 * possible. !vm_page_count_min(0) is the absolute minimum point 2263 * where we can do this. Wait a bit longer to reduce degenerate 2264 * re-blocking (vm_page_free_hysteresis). The target check is just 2265 * to make sure the min-check w/hysteresis does not exceed the 2266 * normal target. 2267 */ 2268 if (vm_pages_waiting) { 2269 if (!vm_page_count_min(vm_page_free_hysteresis) || 2270 !vm_page_count_target()) { 2271 vm_pages_waiting = 0; 2272 wakeup(&vmstats.v_free_count); 2273 ++mycpu->gd_cnt.v_ppwakeups; 2274 } 2275 #if 0 2276 if (!vm_page_count_target()) { 2277 /* 2278 * Plenty of pages are free, wakeup everyone. 2279 */ 2280 vm_pages_waiting = 0; 2281 wakeup(&vmstats.v_free_count); 2282 ++mycpu->gd_cnt.v_ppwakeups; 2283 } else if (!vm_page_count_min(0)) { 2284 /* 2285 * Some pages are free, wakeup someone. 2286 */ 2287 int wcount = vm_pages_waiting; 2288 if (wcount > 0) 2289 --wcount; 2290 vm_pages_waiting = wcount; 2291 wakeup_one(&vmstats.v_free_count); 2292 ++mycpu->gd_cnt.v_ppwakeups; 2293 } 2294 #endif 2295 } 2296 } 2297 2298 /* 2299 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2300 * it from its VM object. 2301 * 2302 * The vm_page must be BUSY on entry. BUSY will be released on 2303 * return (the page will have been freed). 2304 */ 2305 void 2306 vm_page_free_toq(vm_page_t m) 2307 { 2308 mycpu->gd_cnt.v_tfree++; 2309 KKASSERT((m->flags & PG_MAPPED) == 0); 2310 KKASSERT(m->busy_count & PBUSY_LOCKED); 2311 2312 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2313 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2314 "hold(%d)\n", 2315 (u_long)m->pindex, m->busy_count, m->hold_count); 2316 if ((m->queue - m->pc) == PQ_FREE) 2317 panic("vm_page_free: freeing free page"); 2318 else 2319 panic("vm_page_free: freeing busy page"); 2320 } 2321 2322 /* 2323 * Remove from object, spinlock the page and its queues and 2324 * remove from any queue. No queue spinlock will be held 2325 * after this section (because the page was removed from any 2326 * queue). 2327 */ 2328 vm_page_remove(m); 2329 vm_page_and_queue_spin_lock(m); 2330 _vm_page_rem_queue_spinlocked(m); 2331 2332 /* 2333 * No further management of fictitious pages occurs beyond object 2334 * and queue removal. 2335 */ 2336 if ((m->flags & PG_FICTITIOUS) != 0) { 2337 vm_page_spin_unlock(m); 2338 vm_page_wakeup(m); 2339 return; 2340 } 2341 2342 m->valid = 0; 2343 vm_page_undirty(m); 2344 2345 if (m->wire_count != 0) { 2346 if (m->wire_count > 1) { 2347 panic( 2348 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2349 m->wire_count, (long)m->pindex); 2350 } 2351 panic("vm_page_free: freeing wired page"); 2352 } 2353 2354 /* 2355 * Clear the UNMANAGED flag when freeing an unmanaged page. 2356 * Clear the NEED_COMMIT flag 2357 */ 2358 if (m->flags & PG_UNMANAGED) 2359 vm_page_flag_clear(m, PG_UNMANAGED); 2360 if (m->flags & PG_NEED_COMMIT) 2361 vm_page_flag_clear(m, PG_NEED_COMMIT); 2362 2363 if (m->hold_count != 0) { 2364 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2365 } else { 2366 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2367 } 2368 2369 /* 2370 * This sequence allows us to clear BUSY while still holding 2371 * its spin lock, which reduces contention vs allocators. We 2372 * must not leave the queue locked or _vm_page_wakeup() may 2373 * deadlock. 2374 */ 2375 _vm_page_queue_spin_unlock(m); 2376 if (_vm_page_wakeup(m)) { 2377 vm_page_spin_unlock(m); 2378 wakeup(m); 2379 } else { 2380 vm_page_spin_unlock(m); 2381 } 2382 vm_page_free_wakeup(); 2383 } 2384 2385 /* 2386 * vm_page_unmanage() 2387 * 2388 * Prevent PV management from being done on the page. The page is 2389 * removed from the paging queues as if it were wired, and as a 2390 * consequence of no longer being managed the pageout daemon will not 2391 * touch it (since there is no way to locate the pte mappings for the 2392 * page). madvise() calls that mess with the pmap will also no longer 2393 * operate on the page. 2394 * 2395 * Beyond that the page is still reasonably 'normal'. Freeing the page 2396 * will clear the flag. 2397 * 2398 * This routine is used by OBJT_PHYS objects - objects using unswappable 2399 * physical memory as backing store rather then swap-backed memory and 2400 * will eventually be extended to support 4MB unmanaged physical 2401 * mappings. 2402 * 2403 * Caller must be holding the page busy. 2404 */ 2405 void 2406 vm_page_unmanage(vm_page_t m) 2407 { 2408 KKASSERT(m->busy_count & PBUSY_LOCKED); 2409 if ((m->flags & PG_UNMANAGED) == 0) { 2410 if (m->wire_count == 0) 2411 vm_page_unqueue(m); 2412 } 2413 vm_page_flag_set(m, PG_UNMANAGED); 2414 } 2415 2416 /* 2417 * Mark this page as wired down by yet another map, removing it from 2418 * paging queues as necessary. 2419 * 2420 * Caller must be holding the page busy. 2421 */ 2422 void 2423 vm_page_wire(vm_page_t m) 2424 { 2425 /* 2426 * Only bump the wire statistics if the page is not already wired, 2427 * and only unqueue the page if it is on some queue (if it is unmanaged 2428 * it is already off the queues). Don't do anything with fictitious 2429 * pages because they are always wired. 2430 */ 2431 KKASSERT(m->busy_count & PBUSY_LOCKED); 2432 if ((m->flags & PG_FICTITIOUS) == 0) { 2433 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2434 if ((m->flags & PG_UNMANAGED) == 0) 2435 vm_page_unqueue(m); 2436 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2437 } 2438 KASSERT(m->wire_count != 0, 2439 ("vm_page_wire: wire_count overflow m=%p", m)); 2440 } 2441 } 2442 2443 /* 2444 * Release one wiring of this page, potentially enabling it to be paged again. 2445 * 2446 * Many pages placed on the inactive queue should actually go 2447 * into the cache, but it is difficult to figure out which. What 2448 * we do instead, if the inactive target is well met, is to put 2449 * clean pages at the head of the inactive queue instead of the tail. 2450 * This will cause them to be moved to the cache more quickly and 2451 * if not actively re-referenced, freed more quickly. If we just 2452 * stick these pages at the end of the inactive queue, heavy filesystem 2453 * meta-data accesses can cause an unnecessary paging load on memory bound 2454 * processes. This optimization causes one-time-use metadata to be 2455 * reused more quickly. 2456 * 2457 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2458 * the inactive queue. This helps the pageout daemon determine memory 2459 * pressure and act on out-of-memory situations more quickly. 2460 * 2461 * BUT, if we are in a low-memory situation we have no choice but to 2462 * put clean pages on the cache queue. 2463 * 2464 * A number of routines use vm_page_unwire() to guarantee that the page 2465 * will go into either the inactive or active queues, and will NEVER 2466 * be placed in the cache - for example, just after dirtying a page. 2467 * dirty pages in the cache are not allowed. 2468 * 2469 * This routine may not block. 2470 */ 2471 void 2472 vm_page_unwire(vm_page_t m, int activate) 2473 { 2474 KKASSERT(m->busy_count & PBUSY_LOCKED); 2475 if (m->flags & PG_FICTITIOUS) { 2476 /* do nothing */ 2477 } else if (m->wire_count <= 0) { 2478 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2479 } else { 2480 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2481 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2482 if (m->flags & PG_UNMANAGED) { 2483 ; 2484 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2485 vm_page_spin_lock(m); 2486 _vm_page_add_queue_spinlocked(m, 2487 PQ_ACTIVE + m->pc, 0); 2488 _vm_page_and_queue_spin_unlock(m); 2489 } else { 2490 vm_page_spin_lock(m); 2491 vm_page_flag_clear(m, PG_WINATCFLS); 2492 _vm_page_add_queue_spinlocked(m, 2493 PQ_INACTIVE + m->pc, 0); 2494 ++vm_swapcache_inactive_heuristic; 2495 _vm_page_and_queue_spin_unlock(m); 2496 } 2497 } 2498 } 2499 } 2500 2501 /* 2502 * Move the specified page to the inactive queue. If the page has 2503 * any associated swap, the swap is deallocated. 2504 * 2505 * Normally athead is 0 resulting in LRU operation. athead is set 2506 * to 1 if we want this page to be 'as if it were placed in the cache', 2507 * except without unmapping it from the process address space. 2508 * 2509 * vm_page's spinlock must be held on entry and will remain held on return. 2510 * This routine may not block. 2511 */ 2512 static void 2513 _vm_page_deactivate_locked(vm_page_t m, int athead) 2514 { 2515 u_short oqueue; 2516 2517 /* 2518 * Ignore if already inactive. 2519 */ 2520 if (m->queue - m->pc == PQ_INACTIVE) 2521 return; 2522 _vm_page_queue_spin_lock(m); 2523 oqueue = _vm_page_rem_queue_spinlocked(m); 2524 2525 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2526 if (oqueue == PQ_CACHE) 2527 mycpu->gd_cnt.v_reactivated++; 2528 vm_page_flag_clear(m, PG_WINATCFLS); 2529 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2530 if (athead == 0) 2531 ++vm_swapcache_inactive_heuristic; 2532 } 2533 /* NOTE: PQ_NONE if condition not taken */ 2534 _vm_page_queue_spin_unlock(m); 2535 /* leaves vm_page spinlocked */ 2536 } 2537 2538 /* 2539 * Attempt to deactivate a page. 2540 * 2541 * No requirements. 2542 */ 2543 void 2544 vm_page_deactivate(vm_page_t m) 2545 { 2546 vm_page_spin_lock(m); 2547 _vm_page_deactivate_locked(m, 0); 2548 vm_page_spin_unlock(m); 2549 } 2550 2551 void 2552 vm_page_deactivate_locked(vm_page_t m) 2553 { 2554 _vm_page_deactivate_locked(m, 0); 2555 } 2556 2557 /* 2558 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2559 * 2560 * This function returns non-zero if it successfully moved the page to 2561 * PQ_CACHE. 2562 * 2563 * This function unconditionally unbusies the page on return. 2564 */ 2565 int 2566 vm_page_try_to_cache(vm_page_t m) 2567 { 2568 vm_page_spin_lock(m); 2569 if (m->dirty || m->hold_count || m->wire_count || 2570 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2571 if (_vm_page_wakeup(m)) { 2572 vm_page_spin_unlock(m); 2573 wakeup(m); 2574 } else { 2575 vm_page_spin_unlock(m); 2576 } 2577 return(0); 2578 } 2579 vm_page_spin_unlock(m); 2580 2581 /* 2582 * Page busied by us and no longer spinlocked. Dirty pages cannot 2583 * be moved to the cache. 2584 */ 2585 vm_page_test_dirty(m); 2586 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2587 vm_page_wakeup(m); 2588 return(0); 2589 } 2590 vm_page_cache(m); 2591 return(1); 2592 } 2593 2594 /* 2595 * Attempt to free the page. If we cannot free it, we do nothing. 2596 * 1 is returned on success, 0 on failure. 2597 * 2598 * No requirements. 2599 */ 2600 int 2601 vm_page_try_to_free(vm_page_t m) 2602 { 2603 vm_page_spin_lock(m); 2604 if (vm_page_busy_try(m, TRUE)) { 2605 vm_page_spin_unlock(m); 2606 return(0); 2607 } 2608 2609 /* 2610 * The page can be in any state, including already being on the free 2611 * queue. Check to see if it really can be freed. 2612 */ 2613 if (m->dirty || /* can't free if it is dirty */ 2614 m->hold_count || /* or held (XXX may be wrong) */ 2615 m->wire_count || /* or wired */ 2616 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2617 PG_NEED_COMMIT)) || /* or needs a commit */ 2618 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2619 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2620 if (_vm_page_wakeup(m)) { 2621 vm_page_spin_unlock(m); 2622 wakeup(m); 2623 } else { 2624 vm_page_spin_unlock(m); 2625 } 2626 return(0); 2627 } 2628 vm_page_spin_unlock(m); 2629 2630 /* 2631 * We can probably free the page. 2632 * 2633 * Page busied by us and no longer spinlocked. Dirty pages will 2634 * not be freed by this function. We have to re-test the 2635 * dirty bit after cleaning out the pmaps. 2636 */ 2637 vm_page_test_dirty(m); 2638 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2639 vm_page_wakeup(m); 2640 return(0); 2641 } 2642 vm_page_protect(m, VM_PROT_NONE); 2643 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2644 vm_page_wakeup(m); 2645 return(0); 2646 } 2647 vm_page_free(m); 2648 return(1); 2649 } 2650 2651 /* 2652 * vm_page_cache 2653 * 2654 * Put the specified page onto the page cache queue (if appropriate). 2655 * 2656 * The page must be busy, and this routine will release the busy and 2657 * possibly even free the page. 2658 */ 2659 void 2660 vm_page_cache(vm_page_t m) 2661 { 2662 /* 2663 * Not suitable for the cache 2664 */ 2665 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2666 (m->busy_count & PBUSY_MASK) || 2667 m->wire_count || m->hold_count) { 2668 vm_page_wakeup(m); 2669 return; 2670 } 2671 2672 /* 2673 * Already in the cache (and thus not mapped) 2674 */ 2675 if ((m->queue - m->pc) == PQ_CACHE) { 2676 KKASSERT((m->flags & PG_MAPPED) == 0); 2677 vm_page_wakeup(m); 2678 return; 2679 } 2680 2681 /* 2682 * Caller is required to test m->dirty, but note that the act of 2683 * removing the page from its maps can cause it to become dirty 2684 * on an SMP system due to another cpu running in usermode. 2685 */ 2686 if (m->dirty) { 2687 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2688 (long)m->pindex); 2689 } 2690 2691 /* 2692 * Remove all pmaps and indicate that the page is not 2693 * writeable or mapped. Our vm_page_protect() call may 2694 * have blocked (especially w/ VM_PROT_NONE), so recheck 2695 * everything. 2696 */ 2697 vm_page_protect(m, VM_PROT_NONE); 2698 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2699 (m->busy_count & PBUSY_MASK) || 2700 m->wire_count || m->hold_count) { 2701 vm_page_wakeup(m); 2702 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2703 vm_page_deactivate(m); 2704 vm_page_wakeup(m); 2705 } else { 2706 _vm_page_and_queue_spin_lock(m); 2707 _vm_page_rem_queue_spinlocked(m); 2708 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2709 _vm_page_queue_spin_unlock(m); 2710 if (_vm_page_wakeup(m)) { 2711 vm_page_spin_unlock(m); 2712 wakeup(m); 2713 } else { 2714 vm_page_spin_unlock(m); 2715 } 2716 vm_page_free_wakeup(); 2717 } 2718 } 2719 2720 /* 2721 * vm_page_dontneed() 2722 * 2723 * Cache, deactivate, or do nothing as appropriate. This routine 2724 * is typically used by madvise() MADV_DONTNEED. 2725 * 2726 * Generally speaking we want to move the page into the cache so 2727 * it gets reused quickly. However, this can result in a silly syndrome 2728 * due to the page recycling too quickly. Small objects will not be 2729 * fully cached. On the otherhand, if we move the page to the inactive 2730 * queue we wind up with a problem whereby very large objects 2731 * unnecessarily blow away our inactive and cache queues. 2732 * 2733 * The solution is to move the pages based on a fixed weighting. We 2734 * either leave them alone, deactivate them, or move them to the cache, 2735 * where moving them to the cache has the highest weighting. 2736 * By forcing some pages into other queues we eventually force the 2737 * system to balance the queues, potentially recovering other unrelated 2738 * space from active. The idea is to not force this to happen too 2739 * often. 2740 * 2741 * The page must be busied. 2742 */ 2743 void 2744 vm_page_dontneed(vm_page_t m) 2745 { 2746 static int dnweight; 2747 int dnw; 2748 int head; 2749 2750 dnw = ++dnweight; 2751 2752 /* 2753 * occassionally leave the page alone 2754 */ 2755 if ((dnw & 0x01F0) == 0 || 2756 m->queue - m->pc == PQ_INACTIVE || 2757 m->queue - m->pc == PQ_CACHE 2758 ) { 2759 if (m->act_count >= ACT_INIT) 2760 --m->act_count; 2761 return; 2762 } 2763 2764 /* 2765 * If vm_page_dontneed() is inactivating a page, it must clear 2766 * the referenced flag; otherwise the pagedaemon will see references 2767 * on the page in the inactive queue and reactivate it. Until the 2768 * page can move to the cache queue, madvise's job is not done. 2769 */ 2770 vm_page_flag_clear(m, PG_REFERENCED); 2771 pmap_clear_reference(m); 2772 2773 if (m->dirty == 0) 2774 vm_page_test_dirty(m); 2775 2776 if (m->dirty || (dnw & 0x0070) == 0) { 2777 /* 2778 * Deactivate the page 3 times out of 32. 2779 */ 2780 head = 0; 2781 } else { 2782 /* 2783 * Cache the page 28 times out of every 32. Note that 2784 * the page is deactivated instead of cached, but placed 2785 * at the head of the queue instead of the tail. 2786 */ 2787 head = 1; 2788 } 2789 vm_page_spin_lock(m); 2790 _vm_page_deactivate_locked(m, head); 2791 vm_page_spin_unlock(m); 2792 } 2793 2794 /* 2795 * These routines manipulate the 'soft busy' count for a page. A soft busy 2796 * is almost like a hard BUSY except that it allows certain compatible 2797 * operations to occur on the page while it is busy. For example, a page 2798 * undergoing a write can still be mapped read-only. 2799 * 2800 * We also use soft-busy to quickly pmap_enter shared read-only pages 2801 * without having to hold the page locked. 2802 * 2803 * The soft-busy count can be > 1 in situations where multiple threads 2804 * are pmap_enter()ing the same page simultaneously, or when two buffer 2805 * cache buffers overlap the same page. 2806 * 2807 * The caller must hold the page BUSY when making these two calls. 2808 */ 2809 void 2810 vm_page_io_start(vm_page_t m) 2811 { 2812 uint32_t ocount; 2813 2814 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2815 KKASSERT(ocount & PBUSY_LOCKED); 2816 } 2817 2818 void 2819 vm_page_io_finish(vm_page_t m) 2820 { 2821 uint32_t ocount; 2822 2823 ocount = atomic_fetchadd_int(&m->busy_count, -1); 2824 KKASSERT(ocount & PBUSY_MASK); 2825 #if 0 2826 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 2827 wakeup(m); 2828 #endif 2829 } 2830 2831 /* 2832 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 2833 * 2834 * Returns 0 on success, non-zero on failure. 2835 */ 2836 int 2837 vm_page_sbusy_try(vm_page_t m) 2838 { 2839 uint32_t ocount; 2840 2841 if (m->busy_count & PBUSY_LOCKED) 2842 return 1; 2843 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2844 if (ocount & PBUSY_LOCKED) { 2845 vm_page_sbusy_drop(m); 2846 return 1; 2847 } 2848 return 0; 2849 } 2850 2851 /* 2852 * Indicate that a clean VM page requires a filesystem commit and cannot 2853 * be reused. Used by tmpfs. 2854 */ 2855 void 2856 vm_page_need_commit(vm_page_t m) 2857 { 2858 vm_page_flag_set(m, PG_NEED_COMMIT); 2859 vm_object_set_writeable_dirty(m->object); 2860 } 2861 2862 void 2863 vm_page_clear_commit(vm_page_t m) 2864 { 2865 vm_page_flag_clear(m, PG_NEED_COMMIT); 2866 } 2867 2868 /* 2869 * Grab a page, blocking if it is busy and allocating a page if necessary. 2870 * A busy page is returned or NULL. The page may or may not be valid and 2871 * might not be on a queue (the caller is responsible for the disposition of 2872 * the page). 2873 * 2874 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2875 * page will be zero'd and marked valid. 2876 * 2877 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2878 * valid even if it already exists. 2879 * 2880 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2881 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2882 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2883 * 2884 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2885 * always returned if we had blocked. 2886 * 2887 * This routine may not be called from an interrupt. 2888 * 2889 * No other requirements. 2890 */ 2891 vm_page_t 2892 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2893 { 2894 vm_page_t m; 2895 int error; 2896 int shared = 1; 2897 2898 KKASSERT(allocflags & 2899 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2900 vm_object_hold_shared(object); 2901 for (;;) { 2902 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2903 if (error) { 2904 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2905 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2906 m = NULL; 2907 break; 2908 } 2909 /* retry */ 2910 } else if (m == NULL) { 2911 if (shared) { 2912 vm_object_upgrade(object); 2913 shared = 0; 2914 } 2915 if (allocflags & VM_ALLOC_RETRY) 2916 allocflags |= VM_ALLOC_NULL_OK; 2917 m = vm_page_alloc(object, pindex, 2918 allocflags & ~VM_ALLOC_RETRY); 2919 if (m) 2920 break; 2921 vm_wait(0); 2922 if ((allocflags & VM_ALLOC_RETRY) == 0) 2923 goto failed; 2924 } else { 2925 /* m found */ 2926 break; 2927 } 2928 } 2929 2930 /* 2931 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2932 * 2933 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2934 * valid even if already valid. 2935 * 2936 * NOTE! We have removed all of the PG_ZERO optimizations and also 2937 * removed the idle zeroing code. These optimizations actually 2938 * slow things down on modern cpus because the zerod area is 2939 * likely uncached, placing a memory-access burden on the 2940 * accesors taking the fault. 2941 * 2942 * By always zeroing the page in-line with the fault, no 2943 * dynamic ram reads are needed and the caches are hot, ready 2944 * for userland to access the memory. 2945 */ 2946 if (m->valid == 0) { 2947 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2948 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2949 m->valid = VM_PAGE_BITS_ALL; 2950 } 2951 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2952 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2953 m->valid = VM_PAGE_BITS_ALL; 2954 } 2955 failed: 2956 vm_object_drop(object); 2957 return(m); 2958 } 2959 2960 /* 2961 * Mapping function for valid bits or for dirty bits in 2962 * a page. May not block. 2963 * 2964 * Inputs are required to range within a page. 2965 * 2966 * No requirements. 2967 * Non blocking. 2968 */ 2969 int 2970 vm_page_bits(int base, int size) 2971 { 2972 int first_bit; 2973 int last_bit; 2974 2975 KASSERT( 2976 base + size <= PAGE_SIZE, 2977 ("vm_page_bits: illegal base/size %d/%d", base, size) 2978 ); 2979 2980 if (size == 0) /* handle degenerate case */ 2981 return(0); 2982 2983 first_bit = base >> DEV_BSHIFT; 2984 last_bit = (base + size - 1) >> DEV_BSHIFT; 2985 2986 return ((2 << last_bit) - (1 << first_bit)); 2987 } 2988 2989 /* 2990 * Sets portions of a page valid and clean. The arguments are expected 2991 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2992 * of any partial chunks touched by the range. The invalid portion of 2993 * such chunks will be zero'd. 2994 * 2995 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2996 * align base to DEV_BSIZE so as not to mark clean a partially 2997 * truncated device block. Otherwise the dirty page status might be 2998 * lost. 2999 * 3000 * This routine may not block. 3001 * 3002 * (base + size) must be less then or equal to PAGE_SIZE. 3003 */ 3004 static void 3005 _vm_page_zero_valid(vm_page_t m, int base, int size) 3006 { 3007 int frag; 3008 int endoff; 3009 3010 if (size == 0) /* handle degenerate case */ 3011 return; 3012 3013 /* 3014 * If the base is not DEV_BSIZE aligned and the valid 3015 * bit is clear, we have to zero out a portion of the 3016 * first block. 3017 */ 3018 3019 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3020 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3021 ) { 3022 pmap_zero_page_area( 3023 VM_PAGE_TO_PHYS(m), 3024 frag, 3025 base - frag 3026 ); 3027 } 3028 3029 /* 3030 * If the ending offset is not DEV_BSIZE aligned and the 3031 * valid bit is clear, we have to zero out a portion of 3032 * the last block. 3033 */ 3034 3035 endoff = base + size; 3036 3037 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3038 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3039 ) { 3040 pmap_zero_page_area( 3041 VM_PAGE_TO_PHYS(m), 3042 endoff, 3043 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3044 ); 3045 } 3046 } 3047 3048 /* 3049 * Set valid, clear dirty bits. If validating the entire 3050 * page we can safely clear the pmap modify bit. We also 3051 * use this opportunity to clear the PG_NOSYNC flag. If a process 3052 * takes a write fault on a MAP_NOSYNC memory area the flag will 3053 * be set again. 3054 * 3055 * We set valid bits inclusive of any overlap, but we can only 3056 * clear dirty bits for DEV_BSIZE chunks that are fully within 3057 * the range. 3058 * 3059 * Page must be busied? 3060 * No other requirements. 3061 */ 3062 void 3063 vm_page_set_valid(vm_page_t m, int base, int size) 3064 { 3065 _vm_page_zero_valid(m, base, size); 3066 m->valid |= vm_page_bits(base, size); 3067 } 3068 3069 3070 /* 3071 * Set valid bits and clear dirty bits. 3072 * 3073 * Page must be busied by caller. 3074 * 3075 * NOTE: This function does not clear the pmap modified bit. 3076 * Also note that e.g. NFS may use a byte-granular base 3077 * and size. 3078 * 3079 * No other requirements. 3080 */ 3081 void 3082 vm_page_set_validclean(vm_page_t m, int base, int size) 3083 { 3084 int pagebits; 3085 3086 _vm_page_zero_valid(m, base, size); 3087 pagebits = vm_page_bits(base, size); 3088 m->valid |= pagebits; 3089 m->dirty &= ~pagebits; 3090 if (base == 0 && size == PAGE_SIZE) { 3091 /*pmap_clear_modify(m);*/ 3092 vm_page_flag_clear(m, PG_NOSYNC); 3093 } 3094 } 3095 3096 /* 3097 * Set valid & dirty. Used by buwrite() 3098 * 3099 * Page must be busied by caller. 3100 */ 3101 void 3102 vm_page_set_validdirty(vm_page_t m, int base, int size) 3103 { 3104 int pagebits; 3105 3106 pagebits = vm_page_bits(base, size); 3107 m->valid |= pagebits; 3108 m->dirty |= pagebits; 3109 if (m->object) 3110 vm_object_set_writeable_dirty(m->object); 3111 } 3112 3113 /* 3114 * Clear dirty bits. 3115 * 3116 * NOTE: This function does not clear the pmap modified bit. 3117 * Also note that e.g. NFS may use a byte-granular base 3118 * and size. 3119 * 3120 * Page must be busied? 3121 * No other requirements. 3122 */ 3123 void 3124 vm_page_clear_dirty(vm_page_t m, int base, int size) 3125 { 3126 m->dirty &= ~vm_page_bits(base, size); 3127 if (base == 0 && size == PAGE_SIZE) { 3128 /*pmap_clear_modify(m);*/ 3129 vm_page_flag_clear(m, PG_NOSYNC); 3130 } 3131 } 3132 3133 /* 3134 * Make the page all-dirty. 3135 * 3136 * Also make sure the related object and vnode reflect the fact that the 3137 * object may now contain a dirty page. 3138 * 3139 * Page must be busied? 3140 * No other requirements. 3141 */ 3142 void 3143 vm_page_dirty(vm_page_t m) 3144 { 3145 #ifdef INVARIANTS 3146 int pqtype = m->queue - m->pc; 3147 #endif 3148 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3149 ("vm_page_dirty: page in free/cache queue!")); 3150 if (m->dirty != VM_PAGE_BITS_ALL) { 3151 m->dirty = VM_PAGE_BITS_ALL; 3152 if (m->object) 3153 vm_object_set_writeable_dirty(m->object); 3154 } 3155 } 3156 3157 /* 3158 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3159 * valid and dirty bits for the effected areas are cleared. 3160 * 3161 * Page must be busied? 3162 * Does not block. 3163 * No other requirements. 3164 */ 3165 void 3166 vm_page_set_invalid(vm_page_t m, int base, int size) 3167 { 3168 int bits; 3169 3170 bits = vm_page_bits(base, size); 3171 m->valid &= ~bits; 3172 m->dirty &= ~bits; 3173 atomic_add_int(&m->object->generation, 1); 3174 } 3175 3176 /* 3177 * The kernel assumes that the invalid portions of a page contain 3178 * garbage, but such pages can be mapped into memory by user code. 3179 * When this occurs, we must zero out the non-valid portions of the 3180 * page so user code sees what it expects. 3181 * 3182 * Pages are most often semi-valid when the end of a file is mapped 3183 * into memory and the file's size is not page aligned. 3184 * 3185 * Page must be busied? 3186 * No other requirements. 3187 */ 3188 void 3189 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3190 { 3191 int b; 3192 int i; 3193 3194 /* 3195 * Scan the valid bits looking for invalid sections that 3196 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3197 * valid bit may be set ) have already been zerod by 3198 * vm_page_set_validclean(). 3199 */ 3200 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3201 if (i == (PAGE_SIZE / DEV_BSIZE) || 3202 (m->valid & (1 << i)) 3203 ) { 3204 if (i > b) { 3205 pmap_zero_page_area( 3206 VM_PAGE_TO_PHYS(m), 3207 b << DEV_BSHIFT, 3208 (i - b) << DEV_BSHIFT 3209 ); 3210 } 3211 b = i + 1; 3212 } 3213 } 3214 3215 /* 3216 * setvalid is TRUE when we can safely set the zero'd areas 3217 * as being valid. We can do this if there are no cache consistency 3218 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3219 */ 3220 if (setvalid) 3221 m->valid = VM_PAGE_BITS_ALL; 3222 } 3223 3224 /* 3225 * Is a (partial) page valid? Note that the case where size == 0 3226 * will return FALSE in the degenerate case where the page is entirely 3227 * invalid, and TRUE otherwise. 3228 * 3229 * Does not block. 3230 * No other requirements. 3231 */ 3232 int 3233 vm_page_is_valid(vm_page_t m, int base, int size) 3234 { 3235 int bits = vm_page_bits(base, size); 3236 3237 if (m->valid && ((m->valid & bits) == bits)) 3238 return 1; 3239 else 3240 return 0; 3241 } 3242 3243 /* 3244 * update dirty bits from pmap/mmu. May not block. 3245 * 3246 * Caller must hold the page busy 3247 */ 3248 void 3249 vm_page_test_dirty(vm_page_t m) 3250 { 3251 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3252 vm_page_dirty(m); 3253 } 3254 } 3255 3256 #include "opt_ddb.h" 3257 #ifdef DDB 3258 #include <ddb/ddb.h> 3259 3260 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3261 { 3262 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3263 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3264 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3265 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3266 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3267 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3268 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3269 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3270 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3271 db_printf("vmstats.v_inactive_target: %ld\n", 3272 vmstats.v_inactive_target); 3273 } 3274 3275 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3276 { 3277 int i; 3278 db_printf("PQ_FREE:"); 3279 for (i = 0; i < PQ_L2_SIZE; i++) { 3280 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3281 } 3282 db_printf("\n"); 3283 3284 db_printf("PQ_CACHE:"); 3285 for(i = 0; i < PQ_L2_SIZE; i++) { 3286 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3287 } 3288 db_printf("\n"); 3289 3290 db_printf("PQ_ACTIVE:"); 3291 for(i = 0; i < PQ_L2_SIZE; i++) { 3292 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3293 } 3294 db_printf("\n"); 3295 3296 db_printf("PQ_INACTIVE:"); 3297 for(i = 0; i < PQ_L2_SIZE; i++) { 3298 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3299 } 3300 db_printf("\n"); 3301 } 3302 #endif /* DDB */ 3303