1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 33 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $ 34 */ 35 36 /* 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 /* 63 * Resident memory management module. The module manipulates 'VM pages'. 64 * A VM page is the core building block for memory management. 65 */ 66 67 #include <sys/param.h> 68 #include <sys/systm.h> 69 #include <sys/malloc.h> 70 #include <sys/proc.h> 71 #include <sys/vmmeter.h> 72 #include <sys/vnode.h> 73 #include <sys/kernel.h> 74 #include <sys/alist.h> 75 #include <sys/sysctl.h> 76 #include <sys/cpu_topology.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_param.h> 80 #include <sys/lock.h> 81 #include <vm/vm_kern.h> 82 #include <vm/pmap.h> 83 #include <vm/vm_map.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_pageout.h> 87 #include <vm/vm_pager.h> 88 #include <vm/vm_extern.h> 89 #include <vm/swap_pager.h> 90 91 #include <machine/inttypes.h> 92 #include <machine/md_var.h> 93 #include <machine/specialreg.h> 94 #include <machine/bus_dma.h> 95 96 #include <vm/vm_page2.h> 97 #include <sys/spinlock2.h> 98 99 /* 100 * SET - Minimum required set associative size, must be a power of 2. We 101 * want this to match or exceed the set-associativeness of the cpu. 102 * 103 * GRP - A larger set that allows bleed-over into the domains of other 104 * nearby cpus. Also must be a power of 2. Used by the page zeroing 105 * code to smooth things out a bit. 106 */ 107 #define PQ_SET_ASSOC 16 108 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1) 109 110 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2) 111 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1) 112 113 static void vm_page_queue_init(void); 114 static void vm_page_free_wakeup(void); 115 static vm_page_t vm_page_select_cache(u_short pg_color); 116 static vm_page_t _vm_page_list_find2(int basequeue, int index); 117 static void _vm_page_deactivate_locked(vm_page_t m, int athead); 118 119 /* 120 * Array of tailq lists 121 */ 122 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT]; 123 124 static volatile int vm_pages_waiting; 125 static struct alist vm_contig_alist; 126 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536]; 127 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin"); 128 129 static u_long vm_dma_reserved = 0; 130 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved); 131 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0, 132 "Memory reserved for DMA"); 133 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD, 134 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA"); 135 136 static int vm_contig_verbose = 0; 137 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose); 138 139 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare, 140 vm_pindex_t, pindex); 141 142 static void 143 vm_page_queue_init(void) 144 { 145 int i; 146 147 for (i = 0; i < PQ_L2_SIZE; i++) 148 vm_page_queues[PQ_FREE+i].cnt_offset = 149 offsetof(struct vmstats, v_free_count); 150 for (i = 0; i < PQ_L2_SIZE; i++) 151 vm_page_queues[PQ_CACHE+i].cnt_offset = 152 offsetof(struct vmstats, v_cache_count); 153 for (i = 0; i < PQ_L2_SIZE; i++) 154 vm_page_queues[PQ_INACTIVE+i].cnt_offset = 155 offsetof(struct vmstats, v_inactive_count); 156 for (i = 0; i < PQ_L2_SIZE; i++) 157 vm_page_queues[PQ_ACTIVE+i].cnt_offset = 158 offsetof(struct vmstats, v_active_count); 159 for (i = 0; i < PQ_L2_SIZE; i++) 160 vm_page_queues[PQ_HOLD+i].cnt_offset = 161 offsetof(struct vmstats, v_active_count); 162 /* PQ_NONE has no queue */ 163 164 for (i = 0; i < PQ_COUNT; i++) { 165 TAILQ_INIT(&vm_page_queues[i].pl); 166 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init"); 167 } 168 } 169 170 /* 171 * note: place in initialized data section? Is this necessary? 172 */ 173 vm_pindex_t first_page = 0; 174 vm_pindex_t vm_page_array_size = 0; 175 vm_page_t vm_page_array = NULL; 176 vm_paddr_t vm_low_phys_reserved; 177 178 /* 179 * (low level boot) 180 * 181 * Sets the page size, perhaps based upon the memory size. 182 * Must be called before any use of page-size dependent functions. 183 */ 184 void 185 vm_set_page_size(void) 186 { 187 if (vmstats.v_page_size == 0) 188 vmstats.v_page_size = PAGE_SIZE; 189 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0) 190 panic("vm_set_page_size: page size not a power of two"); 191 } 192 193 /* 194 * (low level boot) 195 * 196 * Add a new page to the freelist for use by the system. New pages 197 * are added to both the head and tail of the associated free page 198 * queue in a bottom-up fashion, so both zero'd and non-zero'd page 199 * requests pull 'recent' adds (higher physical addresses) first. 200 * 201 * Beware that the page zeroing daemon will also be running soon after 202 * boot, moving pages from the head to the tail of the PQ_FREE queues. 203 * 204 * Must be called in a critical section. 205 */ 206 static void 207 vm_add_new_page(vm_paddr_t pa) 208 { 209 struct vpgqueues *vpq; 210 vm_page_t m; 211 212 m = PHYS_TO_VM_PAGE(pa); 213 m->phys_addr = pa; 214 m->flags = 0; 215 m->pat_mode = PAT_WRITE_BACK; 216 m->pc = (pa >> PAGE_SHIFT); 217 218 /* 219 * Twist for cpu localization in addition to page coloring, so 220 * different cpus selecting by m->queue get different page colors. 221 */ 222 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE); 223 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)); 224 m->pc &= PQ_L2_MASK; 225 226 /* 227 * Reserve a certain number of contiguous low memory pages for 228 * contigmalloc() to use. 229 */ 230 if (pa < vm_low_phys_reserved) { 231 atomic_add_long(&vmstats.v_page_count, 1); 232 atomic_add_long(&vmstats.v_dma_pages, 1); 233 m->queue = PQ_NONE; 234 m->wire_count = 1; 235 atomic_add_long(&vmstats.v_wire_count, 1); 236 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1); 237 return; 238 } 239 240 /* 241 * General page 242 */ 243 m->queue = m->pc + PQ_FREE; 244 KKASSERT(m->dirty == 0); 245 246 atomic_add_long(&vmstats.v_page_count, 1); 247 atomic_add_long(&vmstats.v_free_count, 1); 248 vpq = &vm_page_queues[m->queue]; 249 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 250 ++vpq->lcnt; 251 } 252 253 /* 254 * (low level boot) 255 * 256 * Initializes the resident memory module. 257 * 258 * Preallocates memory for critical VM structures and arrays prior to 259 * kernel_map becoming available. 260 * 261 * Memory is allocated from (virtual2_start, virtual2_end) if available, 262 * otherwise memory is allocated from (virtual_start, virtual_end). 263 * 264 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be 265 * large enough to hold vm_page_array & other structures for machines with 266 * large amounts of ram, so we want to use virtual2* when available. 267 */ 268 void 269 vm_page_startup(void) 270 { 271 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start; 272 vm_offset_t mapped; 273 vm_pindex_t npages; 274 vm_paddr_t page_range; 275 vm_paddr_t new_end; 276 int i; 277 vm_paddr_t pa; 278 vm_paddr_t last_pa; 279 vm_paddr_t end; 280 vm_paddr_t biggestone, biggestsize; 281 vm_paddr_t total; 282 vm_page_t m; 283 284 total = 0; 285 biggestsize = 0; 286 biggestone = 0; 287 vaddr = round_page(vaddr); 288 289 /* 290 * Make sure ranges are page-aligned. 291 */ 292 for (i = 0; phys_avail[i].phys_end; ++i) { 293 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg); 294 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end); 295 if (phys_avail[i].phys_end < phys_avail[i].phys_beg) 296 phys_avail[i].phys_end = phys_avail[i].phys_beg; 297 } 298 299 /* 300 * Locate largest block 301 */ 302 for (i = 0; phys_avail[i].phys_end; ++i) { 303 vm_paddr_t size = phys_avail[i].phys_end - 304 phys_avail[i].phys_beg; 305 306 if (size > biggestsize) { 307 biggestone = i; 308 biggestsize = size; 309 } 310 total += size; 311 } 312 --i; /* adjust to last entry for use down below */ 313 314 end = phys_avail[biggestone].phys_end; 315 end = trunc_page(end); 316 317 /* 318 * Initialize the queue headers for the free queue, the active queue 319 * and the inactive queue. 320 */ 321 vm_page_queue_init(); 322 323 #if !defined(_KERNEL_VIRTUAL) 324 /* 325 * VKERNELs don't support minidumps and as such don't need 326 * vm_page_dump 327 * 328 * Allocate a bitmap to indicate that a random physical page 329 * needs to be included in a minidump. 330 * 331 * The amd64 port needs this to indicate which direct map pages 332 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 333 * 334 * However, i386 still needs this workspace internally within the 335 * minidump code. In theory, they are not needed on i386, but are 336 * included should the sf_buf code decide to use them. 337 */ 338 page_range = phys_avail[i].phys_end / PAGE_SIZE; 339 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 340 end -= vm_page_dump_size; 341 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size, 342 VM_PROT_READ | VM_PROT_WRITE); 343 bzero((void *)vm_page_dump, vm_page_dump_size); 344 #endif 345 /* 346 * Compute the number of pages of memory that will be available for 347 * use (taking into account the overhead of a page structure per 348 * page). 349 */ 350 first_page = phys_avail[0].phys_beg / PAGE_SIZE; 351 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page; 352 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE; 353 354 #ifndef _KERNEL_VIRTUAL 355 /* 356 * (only applies to real kernels) 357 * 358 * Reserve a large amount of low memory for potential 32-bit DMA 359 * space allocations. Once device initialization is complete we 360 * release most of it, but keep (vm_dma_reserved) memory reserved 361 * for later use. Typically for X / graphics. Through trial and 362 * error we find that GPUs usually requires ~60-100MB or so. 363 * 364 * By default, 128M is left in reserve on machines with 2G+ of ram. 365 */ 366 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT; 367 if (vm_low_phys_reserved > total / 4) 368 vm_low_phys_reserved = total / 4; 369 if (vm_dma_reserved == 0) { 370 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */ 371 if (vm_dma_reserved > total / 16) 372 vm_dma_reserved = total / 16; 373 } 374 #endif 375 alist_init(&vm_contig_alist, 65536, vm_contig_ameta, 376 ALIST_RECORDS_65536); 377 378 /* 379 * Initialize the mem entry structures now, and put them in the free 380 * queue. 381 */ 382 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 383 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); 384 vm_page_array = (vm_page_t)mapped; 385 386 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL) 387 /* 388 * since pmap_map on amd64 returns stuff out of a direct-map region, 389 * we have to manually add these pages to the minidump tracking so 390 * that they can be dumped, including the vm_page_array. 391 */ 392 for (pa = new_end; 393 pa < phys_avail[biggestone].phys_end; 394 pa += PAGE_SIZE) { 395 dump_add_page(pa); 396 } 397 #endif 398 399 /* 400 * Clear all of the page structures, run basic initialization so 401 * PHYS_TO_VM_PAGE() operates properly even on pages not in the 402 * map. 403 */ 404 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 405 vm_page_array_size = page_range; 406 407 m = &vm_page_array[0]; 408 pa = ptoa(first_page); 409 for (i = 0; i < page_range; ++i) { 410 spin_init(&m->spin, "vm_page"); 411 m->phys_addr = pa; 412 pa += PAGE_SIZE; 413 ++m; 414 } 415 416 /* 417 * Construct the free queue(s) in ascending order (by physical 418 * address) so that the first 16MB of physical memory is allocated 419 * last rather than first. On large-memory machines, this avoids 420 * the exhaustion of low physical memory before isa_dma_init has run. 421 */ 422 vmstats.v_page_count = 0; 423 vmstats.v_free_count = 0; 424 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) { 425 pa = phys_avail[i].phys_beg; 426 if (i == biggestone) 427 last_pa = new_end; 428 else 429 last_pa = phys_avail[i].phys_end; 430 while (pa < last_pa && npages-- > 0) { 431 vm_add_new_page(pa); 432 pa += PAGE_SIZE; 433 } 434 } 435 if (virtual2_start) 436 virtual2_start = vaddr; 437 else 438 virtual_start = vaddr; 439 mycpu->gd_vmstats = vmstats; 440 } 441 442 /* 443 * Reorganize VM pages based on numa data. May be called as many times as 444 * necessary. Will reorganize the vm_page_t page color and related queue(s) 445 * to allow vm_page_alloc() to choose pages based on socket affinity. 446 * 447 * NOTE: This function is only called while we are still in UP mode, so 448 * we only need a critical section to protect the queues (which 449 * saves a lot of time, there are likely a ton of pages). 450 */ 451 void 452 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid) 453 { 454 vm_paddr_t scan_beg; 455 vm_paddr_t scan_end; 456 vm_paddr_t ran_end; 457 struct vpgqueues *vpq; 458 vm_page_t m; 459 vm_page_t mend; 460 int i; 461 int socket_mod; 462 int socket_value; 463 464 /* 465 * Check if no physical information, or there was only one socket 466 * (so don't waste time doing nothing!). 467 */ 468 if (cpu_topology_phys_ids <= 1 || 469 cpu_topology_core_ids == 0) { 470 return; 471 } 472 473 /* 474 * Setup for our iteration. Note that ACPI may iterate CPU 475 * sockets starting at 0 or 1 or some other number. The 476 * cpu_topology code mod's it against the socket count. 477 */ 478 ran_end = ran_beg + bytes; 479 physid %= cpu_topology_phys_ids; 480 481 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids; 482 socket_value = physid * socket_mod; 483 mend = &vm_page_array[vm_page_array_size]; 484 485 crit_enter(); 486 487 /* 488 * Adjust vm_page->pc and requeue all affected pages. The 489 * allocator will then be able to localize memory allocations 490 * to some degree. 491 */ 492 for (i = 0; phys_avail[i].phys_end; ++i) { 493 scan_beg = phys_avail[i].phys_beg; 494 scan_end = phys_avail[i].phys_end; 495 if (scan_end <= ran_beg) 496 continue; 497 if (scan_beg >= ran_end) 498 continue; 499 if (scan_beg < ran_beg) 500 scan_beg = ran_beg; 501 if (scan_end > ran_end) 502 scan_end = ran_end; 503 if (atop(scan_end) > first_page + vm_page_array_size) 504 scan_end = ptoa(first_page + vm_page_array_size); 505 506 m = PHYS_TO_VM_PAGE(scan_beg); 507 while (scan_beg < scan_end) { 508 KKASSERT(m < mend); 509 if (m->queue != PQ_NONE) { 510 vpq = &vm_page_queues[m->queue]; 511 TAILQ_REMOVE(&vpq->pl, m, pageq); 512 --vpq->lcnt; 513 /* queue doesn't change, no need to adj cnt */ 514 m->queue -= m->pc; 515 m->pc %= socket_mod; 516 m->pc += socket_value; 517 m->pc &= PQ_L2_MASK; 518 m->queue += m->pc; 519 vpq = &vm_page_queues[m->queue]; 520 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq); 521 ++vpq->lcnt; 522 /* queue doesn't change, no need to adj cnt */ 523 } else { 524 m->pc %= socket_mod; 525 m->pc += socket_value; 526 m->pc &= PQ_L2_MASK; 527 } 528 scan_beg += PAGE_SIZE; 529 ++m; 530 } 531 } 532 crit_exit(); 533 } 534 535 /* 536 * We tended to reserve a ton of memory for contigmalloc(). Now that most 537 * drivers have initialized we want to return most the remaining free 538 * reserve back to the VM page queues so they can be used for normal 539 * allocations. 540 * 541 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool. 542 */ 543 static void 544 vm_page_startup_finish(void *dummy __unused) 545 { 546 alist_blk_t blk; 547 alist_blk_t rblk; 548 alist_blk_t count; 549 alist_blk_t xcount; 550 alist_blk_t bfree; 551 vm_page_t m; 552 553 spin_lock(&vm_contig_spin); 554 for (;;) { 555 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 556 if (bfree <= vm_dma_reserved / PAGE_SIZE) 557 break; 558 if (count == 0) 559 break; 560 561 /* 562 * Figure out how much of the initial reserve we have to 563 * free in order to reach our target. 564 */ 565 bfree -= vm_dma_reserved / PAGE_SIZE; 566 if (count > bfree) { 567 blk += count - bfree; 568 count = bfree; 569 } 570 571 /* 572 * Calculate the nearest power of 2 <= count. 573 */ 574 for (xcount = 1; xcount <= count; xcount <<= 1) 575 ; 576 xcount >>= 1; 577 blk += count - xcount; 578 count = xcount; 579 580 /* 581 * Allocate the pages from the alist, then free them to 582 * the normal VM page queues. 583 * 584 * Pages allocated from the alist are wired. We have to 585 * busy, unwire, and free them. We must also adjust 586 * vm_low_phys_reserved before freeing any pages to prevent 587 * confusion. 588 */ 589 rblk = alist_alloc(&vm_contig_alist, blk, count); 590 if (rblk != blk) { 591 kprintf("vm_page_startup_finish: Unable to return " 592 "dma space @0x%08x/%d -> 0x%08x\n", 593 blk, count, rblk); 594 break; 595 } 596 atomic_add_long(&vmstats.v_dma_pages, -(long)count); 597 spin_unlock(&vm_contig_spin); 598 599 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 600 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m); 601 while (count) { 602 vm_page_busy_wait(m, FALSE, "cpgfr"); 603 vm_page_unwire(m, 0); 604 vm_page_free(m); 605 --count; 606 ++m; 607 } 608 spin_lock(&vm_contig_spin); 609 } 610 spin_unlock(&vm_contig_spin); 611 612 /* 613 * Print out how much DMA space drivers have already allocated and 614 * how much is left over. 615 */ 616 kprintf("DMA space used: %jdk, remaining available: %jdk\n", 617 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) * 618 (PAGE_SIZE / 1024), 619 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024)); 620 } 621 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY, 622 vm_page_startup_finish, NULL); 623 624 625 /* 626 * Scan comparison function for Red-Black tree scans. An inclusive 627 * (start,end) is expected. Other fields are not used. 628 */ 629 int 630 rb_vm_page_scancmp(struct vm_page *p, void *data) 631 { 632 struct rb_vm_page_scan_info *info = data; 633 634 if (p->pindex < info->start_pindex) 635 return(-1); 636 if (p->pindex > info->end_pindex) 637 return(1); 638 return(0); 639 } 640 641 int 642 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2) 643 { 644 if (p1->pindex < p2->pindex) 645 return(-1); 646 if (p1->pindex > p2->pindex) 647 return(1); 648 return(0); 649 } 650 651 void 652 vm_page_init(vm_page_t m) 653 { 654 /* do nothing for now. Called from pmap_page_init() */ 655 } 656 657 /* 658 * Each page queue has its own spin lock, which is fairly optimal for 659 * allocating and freeing pages at least. 660 * 661 * The caller must hold the vm_page_spin_lock() before locking a vm_page's 662 * queue spinlock via this function. Also note that m->queue cannot change 663 * unless both the page and queue are locked. 664 */ 665 static __inline 666 void 667 _vm_page_queue_spin_lock(vm_page_t m) 668 { 669 u_short queue; 670 671 queue = m->queue; 672 if (queue != PQ_NONE) { 673 spin_lock(&vm_page_queues[queue].spin); 674 KKASSERT(queue == m->queue); 675 } 676 } 677 678 static __inline 679 void 680 _vm_page_queue_spin_unlock(vm_page_t m) 681 { 682 u_short queue; 683 684 queue = m->queue; 685 cpu_ccfence(); 686 if (queue != PQ_NONE) 687 spin_unlock(&vm_page_queues[queue].spin); 688 } 689 690 static __inline 691 void 692 _vm_page_queues_spin_lock(u_short queue) 693 { 694 cpu_ccfence(); 695 if (queue != PQ_NONE) 696 spin_lock(&vm_page_queues[queue].spin); 697 } 698 699 700 static __inline 701 void 702 _vm_page_queues_spin_unlock(u_short queue) 703 { 704 cpu_ccfence(); 705 if (queue != PQ_NONE) 706 spin_unlock(&vm_page_queues[queue].spin); 707 } 708 709 void 710 vm_page_queue_spin_lock(vm_page_t m) 711 { 712 _vm_page_queue_spin_lock(m); 713 } 714 715 void 716 vm_page_queues_spin_lock(u_short queue) 717 { 718 _vm_page_queues_spin_lock(queue); 719 } 720 721 void 722 vm_page_queue_spin_unlock(vm_page_t m) 723 { 724 _vm_page_queue_spin_unlock(m); 725 } 726 727 void 728 vm_page_queues_spin_unlock(u_short queue) 729 { 730 _vm_page_queues_spin_unlock(queue); 731 } 732 733 /* 734 * This locks the specified vm_page and its queue in the proper order 735 * (page first, then queue). The queue may change so the caller must 736 * recheck on return. 737 */ 738 static __inline 739 void 740 _vm_page_and_queue_spin_lock(vm_page_t m) 741 { 742 vm_page_spin_lock(m); 743 _vm_page_queue_spin_lock(m); 744 } 745 746 static __inline 747 void 748 _vm_page_and_queue_spin_unlock(vm_page_t m) 749 { 750 _vm_page_queues_spin_unlock(m->queue); 751 vm_page_spin_unlock(m); 752 } 753 754 void 755 vm_page_and_queue_spin_unlock(vm_page_t m) 756 { 757 _vm_page_and_queue_spin_unlock(m); 758 } 759 760 void 761 vm_page_and_queue_spin_lock(vm_page_t m) 762 { 763 _vm_page_and_queue_spin_lock(m); 764 } 765 766 /* 767 * Helper function removes vm_page from its current queue. 768 * Returns the base queue the page used to be on. 769 * 770 * The vm_page and the queue must be spinlocked. 771 * This function will unlock the queue but leave the page spinlocked. 772 */ 773 static __inline u_short 774 _vm_page_rem_queue_spinlocked(vm_page_t m) 775 { 776 struct vpgqueues *pq; 777 u_short queue; 778 u_short oqueue; 779 long *cnt; 780 781 queue = m->queue; 782 if (queue != PQ_NONE) { 783 pq = &vm_page_queues[queue]; 784 TAILQ_REMOVE(&pq->pl, m, pageq); 785 786 /* 787 * Adjust our pcpu stats. In order for the nominal low-memory 788 * algorithms to work properly we don't let any pcpu stat get 789 * too negative before we force it to be rolled-up into the 790 * global stats. Otherwise our pageout and vm_wait tests 791 * will fail badly. 792 * 793 * The idea here is to reduce unnecessary SMP cache 794 * mastership changes in the global vmstats, which can be 795 * particularly bad in multi-socket systems. 796 */ 797 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 798 atomic_add_long(cnt, -1); 799 if (*cnt < -VMMETER_SLOP_COUNT) { 800 u_long copy = atomic_swap_long(cnt, 0); 801 cnt = (long *)((char *)&vmstats + pq->cnt_offset); 802 atomic_add_long(cnt, copy); 803 cnt = (long *)((char *)&mycpu->gd_vmstats + 804 pq->cnt_offset); 805 atomic_add_long(cnt, copy); 806 } 807 pq->lcnt--; 808 m->queue = PQ_NONE; 809 oqueue = queue; 810 queue -= m->pc; 811 vm_page_queues_spin_unlock(oqueue); /* intended */ 812 } 813 return queue; 814 } 815 816 /* 817 * Helper function places the vm_page on the specified queue. Generally 818 * speaking only PQ_FREE pages are placed at the head, to allow them to 819 * be allocated sooner rather than later on the assumption that they 820 * are cache-hot. 821 * 822 * The vm_page must be spinlocked. 823 * This function will return with both the page and the queue locked. 824 */ 825 static __inline void 826 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead) 827 { 828 struct vpgqueues *pq; 829 u_long *cnt; 830 831 KKASSERT(m->queue == PQ_NONE); 832 833 if (queue != PQ_NONE) { 834 vm_page_queues_spin_lock(queue); 835 pq = &vm_page_queues[queue]; 836 ++pq->lcnt; 837 838 /* 839 * Adjust our pcpu stats. If a system entity really needs 840 * to incorporate the count it will call vmstats_rollup() 841 * to roll it all up into the global vmstats strufture. 842 */ 843 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset); 844 atomic_add_long(cnt, 1); 845 846 /* 847 * PQ_FREE is always handled LIFO style to try to provide 848 * cache-hot pages to programs. 849 */ 850 m->queue = queue; 851 if (queue - m->pc == PQ_FREE) { 852 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 853 } else if (athead) { 854 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 855 } else { 856 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 857 } 858 /* leave the queue spinlocked */ 859 } 860 } 861 862 /* 863 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait 864 * until the page is no longer BUSY or SBUSY (busy_count field is 0). 865 * 866 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep 867 * call will be made before returning. 868 * 869 * This function does NOT busy the page and on return the page is not 870 * guaranteed to be available. 871 */ 872 void 873 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 874 { 875 u_int32_t busy_count; 876 877 for (;;) { 878 busy_count = m->busy_count; 879 cpu_ccfence(); 880 881 if ((busy_count & PBUSY_LOCKED) == 0 && 882 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) { 883 break; 884 } 885 tsleep_interlock(m, 0); 886 if (atomic_cmpset_int(&m->busy_count, busy_count, 887 busy_count | PBUSY_WANTED)) { 888 atomic_set_int(&m->flags, PG_REFERENCED); 889 tsleep(m, PINTERLOCKED, msg, 0); 890 break; 891 } 892 } 893 } 894 895 /* 896 * This calculates and returns a page color given an optional VM object and 897 * either a pindex or an iterator. We attempt to return a cpu-localized 898 * pg_color that is still roughly 16-way set-associative. The CPU topology 899 * is used if it was probed. 900 * 901 * The caller may use the returned value to index into e.g. PQ_FREE when 902 * allocating a page in order to nominally obtain pages that are hopefully 903 * already localized to the requesting cpu. This function is not able to 904 * provide any sort of guarantee of this, but does its best to improve 905 * hardware cache management performance. 906 * 907 * WARNING! The caller must mask the returned value with PQ_L2_MASK. 908 */ 909 u_short 910 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex) 911 { 912 u_short pg_color; 913 int phys_id; 914 int core_id; 915 int object_pg_color; 916 917 phys_id = get_cpu_phys_id(cpuid); 918 core_id = get_cpu_core_id(cpuid); 919 object_pg_color = object ? object->pg_color : 0; 920 921 if (cpu_topology_phys_ids && cpu_topology_core_ids) { 922 int grpsize; 923 924 /* 925 * Break us down by socket and cpu 926 */ 927 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids; 928 pg_color += core_id * PQ_L2_SIZE / 929 (cpu_topology_core_ids * cpu_topology_phys_ids); 930 931 /* 932 * Calculate remaining component for object/queue color 933 */ 934 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids * 935 cpu_topology_phys_ids); 936 if (grpsize >= 8) { 937 pg_color += (pindex + object_pg_color) % grpsize; 938 } else { 939 if (grpsize <= 2) { 940 grpsize = 8; 941 } else { 942 /* 3->9, 4->8, 5->10, 6->12, 7->14 */ 943 grpsize += grpsize; 944 if (grpsize < 8) 945 grpsize += grpsize; 946 } 947 pg_color += (pindex + object_pg_color) % grpsize; 948 } 949 } else { 950 /* 951 * Unknown topology, distribute things evenly. 952 */ 953 pg_color = cpuid * PQ_L2_SIZE / ncpus; 954 pg_color += pindex + object_pg_color; 955 } 956 return (pg_color & PQ_L2_MASK); 957 } 958 959 /* 960 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we 961 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED. 962 */ 963 void 964 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m, 965 int also_m_busy, const char *msg 966 VM_PAGE_DEBUG_ARGS) 967 { 968 u_int32_t busy_count; 969 970 for (;;) { 971 busy_count = m->busy_count; 972 cpu_ccfence(); 973 if (busy_count & PBUSY_LOCKED) { 974 tsleep_interlock(m, 0); 975 if (atomic_cmpset_int(&m->busy_count, busy_count, 976 busy_count | PBUSY_WANTED)) { 977 atomic_set_int(&m->flags, PG_REFERENCED); 978 tsleep(m, PINTERLOCKED, msg, 0); 979 } 980 } else if (also_m_busy && busy_count) { 981 tsleep_interlock(m, 0); 982 if (atomic_cmpset_int(&m->busy_count, busy_count, 983 busy_count | PBUSY_WANTED)) { 984 atomic_set_int(&m->flags, PG_REFERENCED); 985 tsleep(m, PINTERLOCKED, msg, 0); 986 } 987 } else { 988 if (atomic_cmpset_int(&m->busy_count, busy_count, 989 busy_count | PBUSY_LOCKED)) { 990 #ifdef VM_PAGE_DEBUG 991 m->busy_func = func; 992 m->busy_line = lineno; 993 #endif 994 break; 995 } 996 } 997 } 998 } 999 1000 /* 1001 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if 1002 * m->busy_count is also 0. 1003 * 1004 * Returns non-zero on failure. 1005 */ 1006 int 1007 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy 1008 VM_PAGE_DEBUG_ARGS) 1009 { 1010 u_int32_t busy_count; 1011 1012 for (;;) { 1013 busy_count = m->busy_count; 1014 cpu_ccfence(); 1015 if (busy_count & PBUSY_LOCKED) 1016 return TRUE; 1017 if (also_m_busy && (busy_count & PBUSY_MASK) != 0) 1018 return TRUE; 1019 if (atomic_cmpset_int(&m->busy_count, busy_count, 1020 busy_count | PBUSY_LOCKED)) { 1021 #ifdef VM_PAGE_DEBUG 1022 m->busy_func = func; 1023 m->busy_line = lineno; 1024 #endif 1025 return FALSE; 1026 } 1027 } 1028 } 1029 1030 /* 1031 * Clear the BUSY flag and return non-zero to indicate to the caller 1032 * that a wakeup() should be performed. 1033 * 1034 * The vm_page must be spinlocked and will remain spinlocked on return. 1035 * The related queue must NOT be spinlocked (which could deadlock us). 1036 * 1037 * (inline version) 1038 */ 1039 static __inline 1040 int 1041 _vm_page_wakeup(vm_page_t m) 1042 { 1043 u_int32_t busy_count; 1044 1045 for (;;) { 1046 busy_count = m->busy_count; 1047 cpu_ccfence(); 1048 if (atomic_cmpset_int(&m->busy_count, busy_count, 1049 busy_count & 1050 ~(PBUSY_LOCKED | PBUSY_WANTED))) { 1051 break; 1052 } 1053 } 1054 return((int)(busy_count & PBUSY_WANTED)); 1055 } 1056 1057 /* 1058 * Clear the BUSY flag and wakeup anyone waiting for the page. This 1059 * is typically the last call you make on a page before moving onto 1060 * other things. 1061 */ 1062 void 1063 vm_page_wakeup(vm_page_t m) 1064 { 1065 KASSERT(m->busy_count & PBUSY_LOCKED, 1066 ("vm_page_wakeup: page not busy!!!")); 1067 vm_page_spin_lock(m); 1068 if (_vm_page_wakeup(m)) { 1069 vm_page_spin_unlock(m); 1070 wakeup(m); 1071 } else { 1072 vm_page_spin_unlock(m); 1073 } 1074 } 1075 1076 /* 1077 * Holding a page keeps it from being reused. Other parts of the system 1078 * can still disassociate the page from its current object and free it, or 1079 * perform read or write I/O on it and/or otherwise manipulate the page, 1080 * but if the page is held the VM system will leave the page and its data 1081 * intact and not reuse the page for other purposes until the last hold 1082 * reference is released. (see vm_page_wire() if you want to prevent the 1083 * page from being disassociated from its object too). 1084 * 1085 * The caller must still validate the contents of the page and, if necessary, 1086 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete 1087 * before manipulating the page. 1088 * 1089 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary 1090 */ 1091 void 1092 vm_page_hold(vm_page_t m) 1093 { 1094 vm_page_spin_lock(m); 1095 atomic_add_int(&m->hold_count, 1); 1096 if (m->queue - m->pc == PQ_FREE) { 1097 _vm_page_queue_spin_lock(m); 1098 _vm_page_rem_queue_spinlocked(m); 1099 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 1100 _vm_page_queue_spin_unlock(m); 1101 } 1102 vm_page_spin_unlock(m); 1103 } 1104 1105 /* 1106 * The opposite of vm_page_hold(). If the page is on the HOLD queue 1107 * it was freed while held and must be moved back to the FREE queue. 1108 */ 1109 void 1110 vm_page_unhold(vm_page_t m) 1111 { 1112 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE, 1113 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)", 1114 m, m->hold_count, m->queue - m->pc)); 1115 vm_page_spin_lock(m); 1116 atomic_add_int(&m->hold_count, -1); 1117 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) { 1118 _vm_page_queue_spin_lock(m); 1119 _vm_page_rem_queue_spinlocked(m); 1120 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 1121 _vm_page_queue_spin_unlock(m); 1122 } 1123 vm_page_spin_unlock(m); 1124 } 1125 1126 /* 1127 * vm_page_getfake: 1128 * 1129 * Create a fictitious page with the specified physical address and 1130 * memory attribute. The memory attribute is the only the machine- 1131 * dependent aspect of a fictitious page that must be initialized. 1132 */ 1133 1134 void 1135 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1136 { 1137 1138 if ((m->flags & PG_FICTITIOUS) != 0) { 1139 /* 1140 * The page's memattr might have changed since the 1141 * previous initialization. Update the pmap to the 1142 * new memattr. 1143 */ 1144 goto memattr; 1145 } 1146 m->phys_addr = paddr; 1147 m->queue = PQ_NONE; 1148 /* Fictitious pages don't use "segind". */ 1149 /* Fictitious pages don't use "order" or "pool". */ 1150 m->flags = PG_FICTITIOUS | PG_UNMANAGED; 1151 m->busy_count = PBUSY_LOCKED; 1152 m->wire_count = 1; 1153 spin_init(&m->spin, "fake_page"); 1154 pmap_page_init(m); 1155 memattr: 1156 pmap_page_set_memattr(m, memattr); 1157 } 1158 1159 /* 1160 * Inserts the given vm_page into the object and object list. 1161 * 1162 * The pagetables are not updated but will presumably fault the page 1163 * in if necessary, or if a kernel page the caller will at some point 1164 * enter the page into the kernel's pmap. We are not allowed to block 1165 * here so we *can't* do this anyway. 1166 * 1167 * This routine may not block. 1168 * This routine must be called with the vm_object held. 1169 * This routine must be called with a critical section held. 1170 * 1171 * This routine returns TRUE if the page was inserted into the object 1172 * successfully, and FALSE if the page already exists in the object. 1173 */ 1174 int 1175 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1176 { 1177 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object)); 1178 if (m->object != NULL) 1179 panic("vm_page_insert: already inserted"); 1180 1181 atomic_add_int(&object->generation, 1); 1182 1183 /* 1184 * Record the object/offset pair in this page and add the 1185 * pv_list_count of the page to the object. 1186 * 1187 * The vm_page spin lock is required for interactions with the pmap. 1188 */ 1189 vm_page_spin_lock(m); 1190 m->object = object; 1191 m->pindex = pindex; 1192 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) { 1193 m->object = NULL; 1194 m->pindex = 0; 1195 vm_page_spin_unlock(m); 1196 return FALSE; 1197 } 1198 ++object->resident_page_count; 1199 ++mycpu->gd_vmtotal.t_rm; 1200 vm_page_spin_unlock(m); 1201 1202 /* 1203 * Since we are inserting a new and possibly dirty page, 1204 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 1205 */ 1206 if ((m->valid & m->dirty) || 1207 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT))) 1208 vm_object_set_writeable_dirty(object); 1209 1210 /* 1211 * Checks for a swap assignment and sets PG_SWAPPED if appropriate. 1212 */ 1213 swap_pager_page_inserted(m); 1214 return TRUE; 1215 } 1216 1217 /* 1218 * Removes the given vm_page_t from the (object,index) table 1219 * 1220 * The underlying pmap entry (if any) is NOT removed here. 1221 * This routine may not block. 1222 * 1223 * The page must be BUSY and will remain BUSY on return. 1224 * No other requirements. 1225 * 1226 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave 1227 * it busy. 1228 */ 1229 void 1230 vm_page_remove(vm_page_t m) 1231 { 1232 vm_object_t object; 1233 1234 if (m->object == NULL) { 1235 return; 1236 } 1237 1238 if ((m->busy_count & PBUSY_LOCKED) == 0) 1239 panic("vm_page_remove: page not busy"); 1240 1241 object = m->object; 1242 1243 vm_object_hold(object); 1244 1245 /* 1246 * Remove the page from the object and update the object. 1247 * 1248 * The vm_page spin lock is required for interactions with the pmap. 1249 */ 1250 vm_page_spin_lock(m); 1251 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m); 1252 --object->resident_page_count; 1253 --mycpu->gd_vmtotal.t_rm; 1254 m->object = NULL; 1255 atomic_add_int(&object->generation, 1); 1256 vm_page_spin_unlock(m); 1257 1258 vm_object_drop(object); 1259 } 1260 1261 /* 1262 * Locate and return the page at (object, pindex), or NULL if the 1263 * page could not be found. 1264 * 1265 * The caller must hold the vm_object token. 1266 */ 1267 vm_page_t 1268 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1269 { 1270 vm_page_t m; 1271 1272 /* 1273 * Search the hash table for this object/offset pair 1274 */ 1275 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1276 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1277 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex)); 1278 return(m); 1279 } 1280 1281 vm_page_t 1282 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object, 1283 vm_pindex_t pindex, 1284 int also_m_busy, const char *msg 1285 VM_PAGE_DEBUG_ARGS) 1286 { 1287 u_int32_t busy_count; 1288 vm_page_t m; 1289 1290 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1291 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1292 while (m) { 1293 KKASSERT(m->object == object && m->pindex == pindex); 1294 busy_count = m->busy_count; 1295 cpu_ccfence(); 1296 if (busy_count & PBUSY_LOCKED) { 1297 tsleep_interlock(m, 0); 1298 if (atomic_cmpset_int(&m->busy_count, busy_count, 1299 busy_count | PBUSY_WANTED)) { 1300 atomic_set_int(&m->flags, PG_REFERENCED); 1301 tsleep(m, PINTERLOCKED, msg, 0); 1302 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1303 pindex); 1304 } 1305 } else if (also_m_busy && busy_count) { 1306 tsleep_interlock(m, 0); 1307 if (atomic_cmpset_int(&m->busy_count, busy_count, 1308 busy_count | PBUSY_WANTED)) { 1309 atomic_set_int(&m->flags, PG_REFERENCED); 1310 tsleep(m, PINTERLOCKED, msg, 0); 1311 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, 1312 pindex); 1313 } 1314 } else if (atomic_cmpset_int(&m->busy_count, busy_count, 1315 busy_count | PBUSY_LOCKED)) { 1316 #ifdef VM_PAGE_DEBUG 1317 m->busy_func = func; 1318 m->busy_line = lineno; 1319 #endif 1320 break; 1321 } 1322 } 1323 return m; 1324 } 1325 1326 /* 1327 * Attempt to lookup and busy a page. 1328 * 1329 * Returns NULL if the page could not be found 1330 * 1331 * Returns a vm_page and error == TRUE if the page exists but could not 1332 * be busied. 1333 * 1334 * Returns a vm_page and error == FALSE on success. 1335 */ 1336 vm_page_t 1337 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object, 1338 vm_pindex_t pindex, 1339 int also_m_busy, int *errorp 1340 VM_PAGE_DEBUG_ARGS) 1341 { 1342 u_int32_t busy_count; 1343 vm_page_t m; 1344 1345 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1346 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1347 *errorp = FALSE; 1348 while (m) { 1349 KKASSERT(m->object == object && m->pindex == pindex); 1350 busy_count = m->busy_count; 1351 cpu_ccfence(); 1352 if (busy_count & PBUSY_LOCKED) { 1353 *errorp = TRUE; 1354 break; 1355 } 1356 if (also_m_busy && busy_count) { 1357 *errorp = TRUE; 1358 break; 1359 } 1360 if (atomic_cmpset_int(&m->busy_count, busy_count, 1361 busy_count | PBUSY_LOCKED)) { 1362 #ifdef VM_PAGE_DEBUG 1363 m->busy_func = func; 1364 m->busy_line = lineno; 1365 #endif 1366 break; 1367 } 1368 } 1369 return m; 1370 } 1371 1372 /* 1373 * Returns a page that is only soft-busied for use by the caller in 1374 * a read-only fashion. Returns NULL if the page could not be found, 1375 * the soft busy could not be obtained, or the page data is invalid. 1376 */ 1377 vm_page_t 1378 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex, 1379 int pgoff, int pgbytes) 1380 { 1381 vm_page_t m; 1382 1383 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1384 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex); 1385 if (m) { 1386 if ((m->valid != VM_PAGE_BITS_ALL && 1387 !vm_page_is_valid(m, pgoff, pgbytes)) || 1388 (m->flags & PG_FICTITIOUS)) { 1389 m = NULL; 1390 } else if (vm_page_sbusy_try(m)) { 1391 m = NULL; 1392 } else if ((m->valid != VM_PAGE_BITS_ALL && 1393 !vm_page_is_valid(m, pgoff, pgbytes)) || 1394 (m->flags & PG_FICTITIOUS)) { 1395 vm_page_sbusy_drop(m); 1396 m = NULL; 1397 } 1398 } 1399 return m; 1400 } 1401 1402 /* 1403 * Caller must hold the related vm_object 1404 */ 1405 vm_page_t 1406 vm_page_next(vm_page_t m) 1407 { 1408 vm_page_t next; 1409 1410 next = vm_page_rb_tree_RB_NEXT(m); 1411 if (next && next->pindex != m->pindex + 1) 1412 next = NULL; 1413 return (next); 1414 } 1415 1416 /* 1417 * vm_page_rename() 1418 * 1419 * Move the given vm_page from its current object to the specified 1420 * target object/offset. The page must be busy and will remain so 1421 * on return. 1422 * 1423 * new_object must be held. 1424 * This routine might block. XXX ? 1425 * 1426 * NOTE: Swap associated with the page must be invalidated by the move. We 1427 * have to do this for several reasons: (1) we aren't freeing the 1428 * page, (2) we are dirtying the page, (3) the VM system is probably 1429 * moving the page from object A to B, and will then later move 1430 * the backing store from A to B and we can't have a conflict. 1431 * 1432 * NOTE: We *always* dirty the page. It is necessary both for the 1433 * fact that we moved it, and because we may be invalidating 1434 * swap. If the page is on the cache, we have to deactivate it 1435 * or vm_page_dirty() will panic. Dirty pages are not allowed 1436 * on the cache. 1437 */ 1438 void 1439 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1440 { 1441 KKASSERT(m->busy_count & PBUSY_LOCKED); 1442 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object)); 1443 if (m->object) { 1444 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object)); 1445 vm_page_remove(m); 1446 } 1447 if (vm_page_insert(m, new_object, new_pindex) == FALSE) { 1448 panic("vm_page_rename: target exists (%p,%"PRIu64")", 1449 new_object, new_pindex); 1450 } 1451 if (m->queue - m->pc == PQ_CACHE) 1452 vm_page_deactivate(m); 1453 vm_page_dirty(m); 1454 } 1455 1456 /* 1457 * vm_page_unqueue() without any wakeup. This routine is used when a page 1458 * is to remain BUSYied by the caller. 1459 * 1460 * This routine may not block. 1461 */ 1462 void 1463 vm_page_unqueue_nowakeup(vm_page_t m) 1464 { 1465 vm_page_and_queue_spin_lock(m); 1466 (void)_vm_page_rem_queue_spinlocked(m); 1467 vm_page_spin_unlock(m); 1468 } 1469 1470 /* 1471 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon 1472 * if necessary. 1473 * 1474 * This routine may not block. 1475 */ 1476 void 1477 vm_page_unqueue(vm_page_t m) 1478 { 1479 u_short queue; 1480 1481 vm_page_and_queue_spin_lock(m); 1482 queue = _vm_page_rem_queue_spinlocked(m); 1483 if (queue == PQ_FREE || queue == PQ_CACHE) { 1484 vm_page_spin_unlock(m); 1485 pagedaemon_wakeup(); 1486 } else { 1487 vm_page_spin_unlock(m); 1488 } 1489 } 1490 1491 /* 1492 * vm_page_list_find() 1493 * 1494 * Find a page on the specified queue with color optimization. 1495 * 1496 * The page coloring optimization attempts to locate a page that does 1497 * not overload other nearby pages in the object in the cpu's L1 or L2 1498 * caches. We need this optimization because cpu caches tend to be 1499 * physical caches, while object spaces tend to be virtual. 1500 * 1501 * The page coloring optimization also, very importantly, tries to localize 1502 * memory to cpus and physical sockets. 1503 * 1504 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock 1505 * and the algorithm is adjusted to localize allocations on a per-core basis. 1506 * This is done by 'twisting' the colors. 1507 * 1508 * The page is returned spinlocked and removed from its queue (it will 1509 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller 1510 * is responsible for dealing with the busy-page case (usually by 1511 * deactivating the page and looping). 1512 * 1513 * NOTE: This routine is carefully inlined. A non-inlined version 1514 * is available for outside callers but the only critical path is 1515 * from within this source file. 1516 * 1517 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE 1518 * represent stable storage, allowing us to order our locks vm_page 1519 * first, then queue. 1520 */ 1521 static __inline 1522 vm_page_t 1523 _vm_page_list_find(int basequeue, int index) 1524 { 1525 vm_page_t m; 1526 1527 for (;;) { 1528 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 1529 if (m == NULL) { 1530 m = _vm_page_list_find2(basequeue, index); 1531 return(m); 1532 } 1533 vm_page_and_queue_spin_lock(m); 1534 if (m->queue == basequeue + index) { 1535 _vm_page_rem_queue_spinlocked(m); 1536 /* vm_page_t spin held, no queue spin */ 1537 break; 1538 } 1539 vm_page_and_queue_spin_unlock(m); 1540 } 1541 return(m); 1542 } 1543 1544 /* 1545 * If we could not find the page in the desired queue try to find it in 1546 * a nearby queue. 1547 */ 1548 static vm_page_t 1549 _vm_page_list_find2(int basequeue, int index) 1550 { 1551 struct vpgqueues *pq; 1552 vm_page_t m = NULL; 1553 int pqmask = PQ_SET_ASSOC_MASK >> 1; 1554 int pqi; 1555 int i; 1556 1557 index &= PQ_L2_MASK; 1558 pq = &vm_page_queues[basequeue]; 1559 1560 /* 1561 * Run local sets of 16, 32, 64, 128, and the whole queue if all 1562 * else fails (PQ_L2_MASK which is 255). 1563 */ 1564 do { 1565 pqmask = (pqmask << 1) | 1; 1566 for (i = 0; i <= pqmask; ++i) { 1567 pqi = (index & ~pqmask) | ((index + i) & pqmask); 1568 m = TAILQ_FIRST(&pq[pqi].pl); 1569 if (m) { 1570 _vm_page_and_queue_spin_lock(m); 1571 if (m->queue == basequeue + pqi) { 1572 _vm_page_rem_queue_spinlocked(m); 1573 return(m); 1574 } 1575 _vm_page_and_queue_spin_unlock(m); 1576 --i; 1577 continue; 1578 } 1579 } 1580 } while (pqmask != PQ_L2_MASK); 1581 1582 return(m); 1583 } 1584 1585 /* 1586 * Returns a vm_page candidate for allocation. The page is not busied so 1587 * it can move around. The caller must busy the page (and typically 1588 * deactivate it if it cannot be busied!) 1589 * 1590 * Returns a spinlocked vm_page that has been removed from its queue. 1591 */ 1592 vm_page_t 1593 vm_page_list_find(int basequeue, int index) 1594 { 1595 return(_vm_page_list_find(basequeue, index)); 1596 } 1597 1598 /* 1599 * Find a page on the cache queue with color optimization, remove it 1600 * from the queue, and busy it. The returned page will not be spinlocked. 1601 * 1602 * A candidate failure will be deactivated. Candidates can fail due to 1603 * being busied by someone else, in which case they will be deactivated. 1604 * 1605 * This routine may not block. 1606 * 1607 */ 1608 static vm_page_t 1609 vm_page_select_cache(u_short pg_color) 1610 { 1611 vm_page_t m; 1612 1613 for (;;) { 1614 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK); 1615 if (m == NULL) 1616 break; 1617 /* 1618 * (m) has been removed from its queue and spinlocked 1619 */ 1620 if (vm_page_busy_try(m, TRUE)) { 1621 _vm_page_deactivate_locked(m, 0); 1622 vm_page_spin_unlock(m); 1623 } else { 1624 /* 1625 * We successfully busied the page 1626 */ 1627 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 && 1628 m->hold_count == 0 && 1629 m->wire_count == 0 && 1630 (m->dirty & m->valid) == 0) { 1631 vm_page_spin_unlock(m); 1632 pagedaemon_wakeup(); 1633 return(m); 1634 } 1635 1636 /* 1637 * The page cannot be recycled, deactivate it. 1638 */ 1639 _vm_page_deactivate_locked(m, 0); 1640 if (_vm_page_wakeup(m)) { 1641 vm_page_spin_unlock(m); 1642 wakeup(m); 1643 } else { 1644 vm_page_spin_unlock(m); 1645 } 1646 } 1647 } 1648 return (m); 1649 } 1650 1651 /* 1652 * Find a free page. We attempt to inline the nominal case and fall back 1653 * to _vm_page_select_free() otherwise. A busied page is removed from 1654 * the queue and returned. 1655 * 1656 * This routine may not block. 1657 */ 1658 static __inline vm_page_t 1659 vm_page_select_free(u_short pg_color) 1660 { 1661 vm_page_t m; 1662 1663 for (;;) { 1664 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK); 1665 if (m == NULL) 1666 break; 1667 if (vm_page_busy_try(m, TRUE)) { 1668 /* 1669 * Various mechanisms such as a pmap_collect can 1670 * result in a busy page on the free queue. We 1671 * have to move the page out of the way so we can 1672 * retry the allocation. If the other thread is not 1673 * allocating the page then m->valid will remain 0 and 1674 * the pageout daemon will free the page later on. 1675 * 1676 * Since we could not busy the page, however, we 1677 * cannot make assumptions as to whether the page 1678 * will be allocated by the other thread or not, 1679 * so all we can do is deactivate it to move it out 1680 * of the way. In particular, if the other thread 1681 * wires the page it may wind up on the inactive 1682 * queue and the pageout daemon will have to deal 1683 * with that case too. 1684 */ 1685 _vm_page_deactivate_locked(m, 0); 1686 vm_page_spin_unlock(m); 1687 } else { 1688 /* 1689 * Theoretically if we are able to busy the page 1690 * atomic with the queue removal (using the vm_page 1691 * lock) nobody else should be able to mess with the 1692 * page before us. 1693 */ 1694 KKASSERT((m->flags & (PG_UNMANAGED | 1695 PG_NEED_COMMIT)) == 0); 1696 KASSERT(m->hold_count == 0, ("m->hold_count is not zero " 1697 "pg %p q=%d flags=%08x hold=%d wire=%d", 1698 m, m->queue, m->flags, m->hold_count, m->wire_count)); 1699 KKASSERT(m->wire_count == 0); 1700 vm_page_spin_unlock(m); 1701 pagedaemon_wakeup(); 1702 1703 /* return busied and removed page */ 1704 return(m); 1705 } 1706 } 1707 return(m); 1708 } 1709 1710 /* 1711 * vm_page_alloc() 1712 * 1713 * Allocate and return a memory cell associated with this VM object/offset 1714 * pair. If object is NULL an unassociated page will be allocated. 1715 * 1716 * The returned page will be busied and removed from its queues. This 1717 * routine can block and may return NULL if a race occurs and the page 1718 * is found to already exist at the specified (object, pindex). 1719 * 1720 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain 1721 * VM_ALLOC_QUICK like normal but cannot use cache 1722 * VM_ALLOC_SYSTEM greater free drain 1723 * VM_ALLOC_INTERRUPT allow free list to be completely drained 1724 * VM_ALLOC_ZERO advisory request for pre-zero'd page only 1725 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only 1726 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision 1727 * (see vm_page_grab()) 1728 * VM_ALLOC_USE_GD ok to use per-gd cache 1729 * 1730 * VM_ALLOC_CPU(n) allocate using specified cpu localization 1731 * 1732 * The object must be held if not NULL 1733 * This routine may not block 1734 * 1735 * Additional special handling is required when called from an interrupt 1736 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache 1737 * in this case. 1738 */ 1739 vm_page_t 1740 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 1741 { 1742 globaldata_t gd; 1743 vm_object_t obj; 1744 vm_page_t m; 1745 u_short pg_color; 1746 int cpuid_local; 1747 1748 #if 0 1749 /* 1750 * Special per-cpu free VM page cache. The pages are pre-busied 1751 * and pre-zerod for us. 1752 */ 1753 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) { 1754 crit_enter_gd(gd); 1755 if (gd->gd_vmpg_count) { 1756 m = gd->gd_vmpg_array[--gd->gd_vmpg_count]; 1757 crit_exit_gd(gd); 1758 goto done; 1759 } 1760 crit_exit_gd(gd); 1761 } 1762 #endif 1763 m = NULL; 1764 1765 /* 1766 * CPU LOCALIZATION 1767 * 1768 * CPU localization algorithm. Break the page queues up by physical 1769 * id and core id (note that two cpu threads will have the same core 1770 * id, and core_id != gd_cpuid). 1771 * 1772 * This is nowhere near perfect, for example the last pindex in a 1773 * subgroup will overflow into the next cpu or package. But this 1774 * should get us good page reuse locality in heavy mixed loads. 1775 * 1776 * (may be executed before the APs are started, so other GDs might 1777 * not exist!) 1778 */ 1779 if (page_req & VM_ALLOC_CPU_SPEC) 1780 cpuid_local = VM_ALLOC_GETCPU(page_req); 1781 else 1782 cpuid_local = mycpu->gd_cpuid; 1783 1784 pg_color = vm_get_pg_color(cpuid_local, object, pindex); 1785 1786 KKASSERT(page_req & 1787 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK| 1788 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 1789 1790 /* 1791 * Certain system threads (pageout daemon, buf_daemon's) are 1792 * allowed to eat deeper into the free page list. 1793 */ 1794 if (curthread->td_flags & TDF_SYSTHREAD) 1795 page_req |= VM_ALLOC_SYSTEM; 1796 1797 /* 1798 * Impose various limitations. Note that the v_free_reserved test 1799 * must match the opposite of vm_page_count_target() to avoid 1800 * livelocks, be careful. 1801 */ 1802 loop: 1803 gd = mycpu; 1804 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved || 1805 ((page_req & VM_ALLOC_INTERRUPT) && 1806 gd->gd_vmstats.v_free_count > 0) || 1807 ((page_req & VM_ALLOC_SYSTEM) && 1808 gd->gd_vmstats.v_cache_count == 0 && 1809 gd->gd_vmstats.v_free_count > 1810 gd->gd_vmstats.v_interrupt_free_min) 1811 ) { 1812 /* 1813 * The free queue has sufficient free pages to take one out. 1814 */ 1815 m = vm_page_select_free(pg_color); 1816 } else if (page_req & VM_ALLOC_NORMAL) { 1817 /* 1818 * Allocatable from the cache (non-interrupt only). On 1819 * success, we must free the page and try again, thus 1820 * ensuring that vmstats.v_*_free_min counters are replenished. 1821 */ 1822 #ifdef INVARIANTS 1823 if (curthread->td_preempted) { 1824 kprintf("vm_page_alloc(): warning, attempt to allocate" 1825 " cache page from preempting interrupt\n"); 1826 m = NULL; 1827 } else { 1828 m = vm_page_select_cache(pg_color); 1829 } 1830 #else 1831 m = vm_page_select_cache(pg_color); 1832 #endif 1833 /* 1834 * On success move the page into the free queue and loop. 1835 * 1836 * Only do this if we can safely acquire the vm_object lock, 1837 * because this is effectively a random page and the caller 1838 * might be holding the lock shared, we don't want to 1839 * deadlock. 1840 */ 1841 if (m != NULL) { 1842 KASSERT(m->dirty == 0, 1843 ("Found dirty cache page %p", m)); 1844 if ((obj = m->object) != NULL) { 1845 if (vm_object_hold_try(obj)) { 1846 vm_page_protect(m, VM_PROT_NONE); 1847 vm_page_free(m); 1848 /* m->object NULL here */ 1849 vm_object_drop(obj); 1850 } else { 1851 vm_page_deactivate(m); 1852 vm_page_wakeup(m); 1853 } 1854 } else { 1855 vm_page_protect(m, VM_PROT_NONE); 1856 vm_page_free(m); 1857 } 1858 goto loop; 1859 } 1860 1861 /* 1862 * On failure return NULL 1863 */ 1864 atomic_add_int(&vm_pageout_deficit, 1); 1865 pagedaemon_wakeup(); 1866 return (NULL); 1867 } else { 1868 /* 1869 * No pages available, wakeup the pageout daemon and give up. 1870 */ 1871 atomic_add_int(&vm_pageout_deficit, 1); 1872 pagedaemon_wakeup(); 1873 return (NULL); 1874 } 1875 1876 /* 1877 * v_free_count can race so loop if we don't find the expected 1878 * page. 1879 */ 1880 if (m == NULL) { 1881 vmstats_rollup(); 1882 goto loop; 1883 } 1884 1885 /* 1886 * Good page found. The page has already been busied for us and 1887 * removed from its queues. 1888 */ 1889 KASSERT(m->dirty == 0, 1890 ("vm_page_alloc: free/cache page %p was dirty", m)); 1891 KKASSERT(m->queue == PQ_NONE); 1892 1893 #if 0 1894 done: 1895 #endif 1896 /* 1897 * Initialize the structure, inheriting some flags but clearing 1898 * all the rest. The page has already been busied for us. 1899 */ 1900 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK); 1901 1902 KKASSERT(m->wire_count == 0); 1903 KKASSERT((m->busy_count & PBUSY_MASK) == 0); 1904 m->act_count = 0; 1905 m->valid = 0; 1906 1907 /* 1908 * Caller must be holding the object lock (asserted by 1909 * vm_page_insert()). 1910 * 1911 * NOTE: Inserting a page here does not insert it into any pmaps 1912 * (which could cause us to block allocating memory). 1913 * 1914 * NOTE: If no object an unassociated page is allocated, m->pindex 1915 * can be used by the caller for any purpose. 1916 */ 1917 if (object) { 1918 if (vm_page_insert(m, object, pindex) == FALSE) { 1919 vm_page_free(m); 1920 if ((page_req & VM_ALLOC_NULL_OK) == 0) 1921 panic("PAGE RACE %p[%ld]/%p", 1922 object, (long)pindex, m); 1923 m = NULL; 1924 } 1925 } else { 1926 m->pindex = pindex; 1927 } 1928 1929 /* 1930 * Don't wakeup too often - wakeup the pageout daemon when 1931 * we would be nearly out of memory. 1932 */ 1933 pagedaemon_wakeup(); 1934 1935 /* 1936 * A BUSY page is returned. 1937 */ 1938 return (m); 1939 } 1940 1941 /* 1942 * Returns number of pages available in our DMA memory reserve 1943 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf) 1944 */ 1945 vm_size_t 1946 vm_contig_avail_pages(void) 1947 { 1948 alist_blk_t blk; 1949 alist_blk_t count; 1950 alist_blk_t bfree; 1951 spin_lock(&vm_contig_spin); 1952 bfree = alist_free_info(&vm_contig_alist, &blk, &count); 1953 spin_unlock(&vm_contig_spin); 1954 1955 return bfree; 1956 } 1957 1958 /* 1959 * Attempt to allocate contiguous physical memory with the specified 1960 * requirements. 1961 */ 1962 vm_page_t 1963 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high, 1964 unsigned long alignment, unsigned long boundary, 1965 unsigned long size, vm_memattr_t memattr) 1966 { 1967 alist_blk_t blk; 1968 vm_page_t m; 1969 vm_pindex_t i; 1970 static vm_pindex_t contig_rover; 1971 1972 alignment >>= PAGE_SHIFT; 1973 if (alignment == 0) 1974 alignment = 1; 1975 boundary >>= PAGE_SHIFT; 1976 if (boundary == 0) 1977 boundary = 1; 1978 size = (size + PAGE_MASK) >> PAGE_SHIFT; 1979 1980 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE && 1981 boundary <= PAGE_SIZE && size == 1 && 1982 memattr == VM_MEMATTR_DEFAULT) { 1983 /* 1984 * Any page will work, use vm_page_alloc() 1985 * (e.g. when used from kmem_alloc_attr()) 1986 */ 1987 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF, 1988 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | 1989 VM_ALLOC_INTERRUPT); 1990 m->valid = VM_PAGE_BITS_ALL; 1991 vm_page_wire(m); 1992 vm_page_wakeup(m); 1993 } else { 1994 /* 1995 * Use the low-memory dma reserve 1996 */ 1997 spin_lock(&vm_contig_spin); 1998 blk = alist_alloc(&vm_contig_alist, 0, size); 1999 if (blk == ALIST_BLOCK_NONE) { 2000 spin_unlock(&vm_contig_spin); 2001 if (bootverbose) { 2002 kprintf("vm_page_alloc_contig: %ldk nospace\n", 2003 (size << PAGE_SHIFT) / 1024); 2004 print_backtrace(5); 2005 } 2006 return(NULL); 2007 } 2008 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) { 2009 alist_free(&vm_contig_alist, blk, size); 2010 spin_unlock(&vm_contig_spin); 2011 if (bootverbose) { 2012 kprintf("vm_page_alloc_contig: %ldk high " 2013 "%016jx failed\n", 2014 (size << PAGE_SHIFT) / 1024, 2015 (intmax_t)high); 2016 } 2017 return(NULL); 2018 } 2019 spin_unlock(&vm_contig_spin); 2020 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT); 2021 } 2022 if (vm_contig_verbose) { 2023 kprintf("vm_page_alloc_contig: %016jx/%ldk " 2024 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n", 2025 (intmax_t)m->phys_addr, 2026 (size << PAGE_SHIFT) / 1024, 2027 low, high, alignment, boundary, size, memattr); 2028 } 2029 if (memattr != VM_MEMATTR_DEFAULT) { 2030 for (i = 0;i < size; i++) 2031 pmap_page_set_memattr(&m[i], memattr); 2032 } 2033 return m; 2034 } 2035 2036 /* 2037 * Free contiguously allocated pages. The pages will be wired but not busy. 2038 * When freeing to the alist we leave them wired and not busy. 2039 */ 2040 void 2041 vm_page_free_contig(vm_page_t m, unsigned long size) 2042 { 2043 vm_paddr_t pa = VM_PAGE_TO_PHYS(m); 2044 vm_pindex_t start = pa >> PAGE_SHIFT; 2045 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT; 2046 2047 if (vm_contig_verbose) { 2048 kprintf("vm_page_free_contig: %016jx/%ldk\n", 2049 (intmax_t)pa, size / 1024); 2050 } 2051 if (pa < vm_low_phys_reserved) { 2052 KKASSERT(pa + size <= vm_low_phys_reserved); 2053 spin_lock(&vm_contig_spin); 2054 alist_free(&vm_contig_alist, start, pages); 2055 spin_unlock(&vm_contig_spin); 2056 } else { 2057 while (pages) { 2058 vm_page_busy_wait(m, FALSE, "cpgfr"); 2059 vm_page_unwire(m, 0); 2060 vm_page_free(m); 2061 --pages; 2062 ++m; 2063 } 2064 2065 } 2066 } 2067 2068 2069 /* 2070 * Wait for sufficient free memory for nominal heavy memory use kernel 2071 * operations. 2072 * 2073 * WARNING! Be sure never to call this in any vm_pageout code path, which 2074 * will trivially deadlock the system. 2075 */ 2076 void 2077 vm_wait_nominal(void) 2078 { 2079 while (vm_page_count_min(0)) 2080 vm_wait(0); 2081 } 2082 2083 /* 2084 * Test if vm_wait_nominal() would block. 2085 */ 2086 int 2087 vm_test_nominal(void) 2088 { 2089 if (vm_page_count_min(0)) 2090 return(1); 2091 return(0); 2092 } 2093 2094 /* 2095 * Block until free pages are available for allocation, called in various 2096 * places before memory allocations. 2097 * 2098 * The caller may loop if vm_page_count_min() == FALSE so we cannot be 2099 * more generous then that. 2100 */ 2101 void 2102 vm_wait(int timo) 2103 { 2104 /* 2105 * never wait forever 2106 */ 2107 if (timo == 0) 2108 timo = hz; 2109 lwkt_gettoken(&vm_token); 2110 2111 if (curthread == pagethread || 2112 curthread == emergpager) { 2113 /* 2114 * The pageout daemon itself needs pages, this is bad. 2115 */ 2116 if (vm_page_count_min(0)) { 2117 vm_pageout_pages_needed = 1; 2118 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo); 2119 } 2120 } else { 2121 /* 2122 * Wakeup the pageout daemon if necessary and wait. 2123 * 2124 * Do not wait indefinitely for the target to be reached, 2125 * as load might prevent it from being reached any time soon. 2126 * But wait a little to try to slow down page allocations 2127 * and to give more important threads (the pagedaemon) 2128 * allocation priority. 2129 */ 2130 if (vm_page_count_target()) { 2131 if (vm_pages_needed == 0) { 2132 vm_pages_needed = 1; 2133 wakeup(&vm_pages_needed); 2134 } 2135 ++vm_pages_waiting; /* SMP race ok */ 2136 tsleep(&vmstats.v_free_count, 0, "vmwait", timo); 2137 } 2138 } 2139 lwkt_reltoken(&vm_token); 2140 } 2141 2142 /* 2143 * Block until free pages are available for allocation 2144 * 2145 * Called only from vm_fault so that processes page faulting can be 2146 * easily tracked. 2147 */ 2148 void 2149 vm_wait_pfault(void) 2150 { 2151 /* 2152 * Wakeup the pageout daemon if necessary and wait. 2153 * 2154 * Do not wait indefinitely for the target to be reached, 2155 * as load might prevent it from being reached any time soon. 2156 * But wait a little to try to slow down page allocations 2157 * and to give more important threads (the pagedaemon) 2158 * allocation priority. 2159 */ 2160 if (vm_page_count_min(0)) { 2161 lwkt_gettoken(&vm_token); 2162 while (vm_page_count_severe()) { 2163 if (vm_page_count_target()) { 2164 thread_t td; 2165 2166 if (vm_pages_needed == 0) { 2167 vm_pages_needed = 1; 2168 wakeup(&vm_pages_needed); 2169 } 2170 ++vm_pages_waiting; /* SMP race ok */ 2171 tsleep(&vmstats.v_free_count, 0, "pfault", hz); 2172 2173 /* 2174 * Do not stay stuck in the loop if the system is trying 2175 * to kill the process. 2176 */ 2177 td = curthread; 2178 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2179 break; 2180 } 2181 } 2182 lwkt_reltoken(&vm_token); 2183 } 2184 } 2185 2186 /* 2187 * Put the specified page on the active list (if appropriate). Ensure 2188 * that act_count is at least ACT_INIT but do not otherwise mess with it. 2189 * 2190 * The caller should be holding the page busied ? XXX 2191 * This routine may not block. 2192 */ 2193 void 2194 vm_page_activate(vm_page_t m) 2195 { 2196 u_short oqueue; 2197 2198 vm_page_spin_lock(m); 2199 if (m->queue - m->pc != PQ_ACTIVE) { 2200 _vm_page_queue_spin_lock(m); 2201 oqueue = _vm_page_rem_queue_spinlocked(m); 2202 /* page is left spinlocked, queue is unlocked */ 2203 2204 if (oqueue == PQ_CACHE) 2205 mycpu->gd_cnt.v_reactivated++; 2206 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2207 if (m->act_count < ACT_INIT) 2208 m->act_count = ACT_INIT; 2209 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0); 2210 } 2211 _vm_page_and_queue_spin_unlock(m); 2212 if (oqueue == PQ_CACHE || oqueue == PQ_FREE) 2213 pagedaemon_wakeup(); 2214 } else { 2215 if (m->act_count < ACT_INIT) 2216 m->act_count = ACT_INIT; 2217 vm_page_spin_unlock(m); 2218 } 2219 } 2220 2221 /* 2222 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2223 * routine is called when a page has been added to the cache or free 2224 * queues. 2225 * 2226 * This routine may not block. 2227 */ 2228 static __inline void 2229 vm_page_free_wakeup(void) 2230 { 2231 globaldata_t gd = mycpu; 2232 2233 /* 2234 * If the pageout daemon itself needs pages, then tell it that 2235 * there are some free. 2236 */ 2237 if (vm_pageout_pages_needed && 2238 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >= 2239 gd->gd_vmstats.v_pageout_free_min 2240 ) { 2241 vm_pageout_pages_needed = 0; 2242 wakeup(&vm_pageout_pages_needed); 2243 } 2244 2245 /* 2246 * Wakeup processes that are waiting on memory. 2247 * 2248 * Generally speaking we want to wakeup stuck processes as soon as 2249 * possible. !vm_page_count_min(0) is the absolute minimum point 2250 * where we can do this. Wait a bit longer to reduce degenerate 2251 * re-blocking (vm_page_free_hysteresis). The target check is just 2252 * to make sure the min-check w/hysteresis does not exceed the 2253 * normal target. 2254 */ 2255 if (vm_pages_waiting) { 2256 if (!vm_page_count_min(vm_page_free_hysteresis) || 2257 !vm_page_count_target()) { 2258 vm_pages_waiting = 0; 2259 wakeup(&vmstats.v_free_count); 2260 ++mycpu->gd_cnt.v_ppwakeups; 2261 } 2262 #if 0 2263 if (!vm_page_count_target()) { 2264 /* 2265 * Plenty of pages are free, wakeup everyone. 2266 */ 2267 vm_pages_waiting = 0; 2268 wakeup(&vmstats.v_free_count); 2269 ++mycpu->gd_cnt.v_ppwakeups; 2270 } else if (!vm_page_count_min(0)) { 2271 /* 2272 * Some pages are free, wakeup someone. 2273 */ 2274 int wcount = vm_pages_waiting; 2275 if (wcount > 0) 2276 --wcount; 2277 vm_pages_waiting = wcount; 2278 wakeup_one(&vmstats.v_free_count); 2279 ++mycpu->gd_cnt.v_ppwakeups; 2280 } 2281 #endif 2282 } 2283 } 2284 2285 /* 2286 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates 2287 * it from its VM object. 2288 * 2289 * The vm_page must be BUSY on entry. BUSY will be released on 2290 * return (the page will have been freed). 2291 */ 2292 void 2293 vm_page_free_toq(vm_page_t m) 2294 { 2295 mycpu->gd_cnt.v_tfree++; 2296 KKASSERT((m->flags & PG_MAPPED) == 0); 2297 KKASSERT(m->busy_count & PBUSY_LOCKED); 2298 2299 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) { 2300 kprintf("vm_page_free: pindex(%lu), busy %08x, " 2301 "hold(%d)\n", 2302 (u_long)m->pindex, m->busy_count, m->hold_count); 2303 if ((m->queue - m->pc) == PQ_FREE) 2304 panic("vm_page_free: freeing free page"); 2305 else 2306 panic("vm_page_free: freeing busy page"); 2307 } 2308 2309 /* 2310 * Remove from object, spinlock the page and its queues and 2311 * remove from any queue. No queue spinlock will be held 2312 * after this section (because the page was removed from any 2313 * queue). 2314 */ 2315 vm_page_remove(m); 2316 vm_page_and_queue_spin_lock(m); 2317 _vm_page_rem_queue_spinlocked(m); 2318 2319 /* 2320 * No further management of fictitious pages occurs beyond object 2321 * and queue removal. 2322 */ 2323 if ((m->flags & PG_FICTITIOUS) != 0) { 2324 vm_page_spin_unlock(m); 2325 vm_page_wakeup(m); 2326 return; 2327 } 2328 2329 m->valid = 0; 2330 vm_page_undirty(m); 2331 2332 if (m->wire_count != 0) { 2333 if (m->wire_count > 1) { 2334 panic( 2335 "vm_page_free: invalid wire count (%d), pindex: 0x%lx", 2336 m->wire_count, (long)m->pindex); 2337 } 2338 panic("vm_page_free: freeing wired page"); 2339 } 2340 2341 /* 2342 * Clear the UNMANAGED flag when freeing an unmanaged page. 2343 * Clear the NEED_COMMIT flag 2344 */ 2345 if (m->flags & PG_UNMANAGED) 2346 vm_page_flag_clear(m, PG_UNMANAGED); 2347 if (m->flags & PG_NEED_COMMIT) 2348 vm_page_flag_clear(m, PG_NEED_COMMIT); 2349 2350 if (m->hold_count != 0) { 2351 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0); 2352 } else { 2353 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1); 2354 } 2355 2356 /* 2357 * This sequence allows us to clear BUSY while still holding 2358 * its spin lock, which reduces contention vs allocators. We 2359 * must not leave the queue locked or _vm_page_wakeup() may 2360 * deadlock. 2361 */ 2362 _vm_page_queue_spin_unlock(m); 2363 if (_vm_page_wakeup(m)) { 2364 vm_page_spin_unlock(m); 2365 wakeup(m); 2366 } else { 2367 vm_page_spin_unlock(m); 2368 } 2369 vm_page_free_wakeup(); 2370 } 2371 2372 /* 2373 * vm_page_unmanage() 2374 * 2375 * Prevent PV management from being done on the page. The page is 2376 * removed from the paging queues as if it were wired, and as a 2377 * consequence of no longer being managed the pageout daemon will not 2378 * touch it (since there is no way to locate the pte mappings for the 2379 * page). madvise() calls that mess with the pmap will also no longer 2380 * operate on the page. 2381 * 2382 * Beyond that the page is still reasonably 'normal'. Freeing the page 2383 * will clear the flag. 2384 * 2385 * This routine is used by OBJT_PHYS objects - objects using unswappable 2386 * physical memory as backing store rather then swap-backed memory and 2387 * will eventually be extended to support 4MB unmanaged physical 2388 * mappings. 2389 * 2390 * Caller must be holding the page busy. 2391 */ 2392 void 2393 vm_page_unmanage(vm_page_t m) 2394 { 2395 KKASSERT(m->busy_count & PBUSY_LOCKED); 2396 if ((m->flags & PG_UNMANAGED) == 0) { 2397 if (m->wire_count == 0) 2398 vm_page_unqueue(m); 2399 } 2400 vm_page_flag_set(m, PG_UNMANAGED); 2401 } 2402 2403 /* 2404 * Mark this page as wired down by yet another map, removing it from 2405 * paging queues as necessary. 2406 * 2407 * Caller must be holding the page busy. 2408 */ 2409 void 2410 vm_page_wire(vm_page_t m) 2411 { 2412 /* 2413 * Only bump the wire statistics if the page is not already wired, 2414 * and only unqueue the page if it is on some queue (if it is unmanaged 2415 * it is already off the queues). Don't do anything with fictitious 2416 * pages because they are always wired. 2417 */ 2418 KKASSERT(m->busy_count & PBUSY_LOCKED); 2419 if ((m->flags & PG_FICTITIOUS) == 0) { 2420 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) { 2421 if ((m->flags & PG_UNMANAGED) == 0) 2422 vm_page_unqueue(m); 2423 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1); 2424 } 2425 KASSERT(m->wire_count != 0, 2426 ("vm_page_wire: wire_count overflow m=%p", m)); 2427 } 2428 } 2429 2430 /* 2431 * Release one wiring of this page, potentially enabling it to be paged again. 2432 * 2433 * Many pages placed on the inactive queue should actually go 2434 * into the cache, but it is difficult to figure out which. What 2435 * we do instead, if the inactive target is well met, is to put 2436 * clean pages at the head of the inactive queue instead of the tail. 2437 * This will cause them to be moved to the cache more quickly and 2438 * if not actively re-referenced, freed more quickly. If we just 2439 * stick these pages at the end of the inactive queue, heavy filesystem 2440 * meta-data accesses can cause an unnecessary paging load on memory bound 2441 * processes. This optimization causes one-time-use metadata to be 2442 * reused more quickly. 2443 * 2444 * Pages marked PG_NEED_COMMIT are always activated and never placed on 2445 * the inactive queue. This helps the pageout daemon determine memory 2446 * pressure and act on out-of-memory situations more quickly. 2447 * 2448 * BUT, if we are in a low-memory situation we have no choice but to 2449 * put clean pages on the cache queue. 2450 * 2451 * A number of routines use vm_page_unwire() to guarantee that the page 2452 * will go into either the inactive or active queues, and will NEVER 2453 * be placed in the cache - for example, just after dirtying a page. 2454 * dirty pages in the cache are not allowed. 2455 * 2456 * This routine may not block. 2457 */ 2458 void 2459 vm_page_unwire(vm_page_t m, int activate) 2460 { 2461 KKASSERT(m->busy_count & PBUSY_LOCKED); 2462 if (m->flags & PG_FICTITIOUS) { 2463 /* do nothing */ 2464 } else if (m->wire_count <= 0) { 2465 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 2466 } else { 2467 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) { 2468 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1); 2469 if (m->flags & PG_UNMANAGED) { 2470 ; 2471 } else if (activate || (m->flags & PG_NEED_COMMIT)) { 2472 vm_page_spin_lock(m); 2473 _vm_page_add_queue_spinlocked(m, 2474 PQ_ACTIVE + m->pc, 0); 2475 _vm_page_and_queue_spin_unlock(m); 2476 } else { 2477 vm_page_spin_lock(m); 2478 vm_page_flag_clear(m, PG_WINATCFLS); 2479 _vm_page_add_queue_spinlocked(m, 2480 PQ_INACTIVE + m->pc, 0); 2481 ++vm_swapcache_inactive_heuristic; 2482 _vm_page_and_queue_spin_unlock(m); 2483 } 2484 } 2485 } 2486 } 2487 2488 /* 2489 * Move the specified page to the inactive queue. If the page has 2490 * any associated swap, the swap is deallocated. 2491 * 2492 * Normally athead is 0 resulting in LRU operation. athead is set 2493 * to 1 if we want this page to be 'as if it were placed in the cache', 2494 * except without unmapping it from the process address space. 2495 * 2496 * vm_page's spinlock must be held on entry and will remain held on return. 2497 * This routine may not block. 2498 */ 2499 static void 2500 _vm_page_deactivate_locked(vm_page_t m, int athead) 2501 { 2502 u_short oqueue; 2503 2504 /* 2505 * Ignore if already inactive. 2506 */ 2507 if (m->queue - m->pc == PQ_INACTIVE) 2508 return; 2509 _vm_page_queue_spin_lock(m); 2510 oqueue = _vm_page_rem_queue_spinlocked(m); 2511 2512 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 2513 if (oqueue == PQ_CACHE) 2514 mycpu->gd_cnt.v_reactivated++; 2515 vm_page_flag_clear(m, PG_WINATCFLS); 2516 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead); 2517 if (athead == 0) 2518 ++vm_swapcache_inactive_heuristic; 2519 } 2520 /* NOTE: PQ_NONE if condition not taken */ 2521 _vm_page_queue_spin_unlock(m); 2522 /* leaves vm_page spinlocked */ 2523 } 2524 2525 /* 2526 * Attempt to deactivate a page. 2527 * 2528 * No requirements. 2529 */ 2530 void 2531 vm_page_deactivate(vm_page_t m) 2532 { 2533 vm_page_spin_lock(m); 2534 _vm_page_deactivate_locked(m, 0); 2535 vm_page_spin_unlock(m); 2536 } 2537 2538 void 2539 vm_page_deactivate_locked(vm_page_t m) 2540 { 2541 _vm_page_deactivate_locked(m, 0); 2542 } 2543 2544 /* 2545 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it. 2546 * 2547 * This function returns non-zero if it successfully moved the page to 2548 * PQ_CACHE. 2549 * 2550 * This function unconditionally unbusies the page on return. 2551 */ 2552 int 2553 vm_page_try_to_cache(vm_page_t m) 2554 { 2555 vm_page_spin_lock(m); 2556 if (m->dirty || m->hold_count || m->wire_count || 2557 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) { 2558 if (_vm_page_wakeup(m)) { 2559 vm_page_spin_unlock(m); 2560 wakeup(m); 2561 } else { 2562 vm_page_spin_unlock(m); 2563 } 2564 return(0); 2565 } 2566 vm_page_spin_unlock(m); 2567 2568 /* 2569 * Page busied by us and no longer spinlocked. Dirty pages cannot 2570 * be moved to the cache. 2571 */ 2572 vm_page_test_dirty(m); 2573 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2574 vm_page_wakeup(m); 2575 return(0); 2576 } 2577 vm_page_cache(m); 2578 return(1); 2579 } 2580 2581 /* 2582 * Attempt to free the page. If we cannot free it, we do nothing. 2583 * 1 is returned on success, 0 on failure. 2584 * 2585 * No requirements. 2586 */ 2587 int 2588 vm_page_try_to_free(vm_page_t m) 2589 { 2590 vm_page_spin_lock(m); 2591 if (vm_page_busy_try(m, TRUE)) { 2592 vm_page_spin_unlock(m); 2593 return(0); 2594 } 2595 2596 /* 2597 * The page can be in any state, including already being on the free 2598 * queue. Check to see if it really can be freed. 2599 */ 2600 if (m->dirty || /* can't free if it is dirty */ 2601 m->hold_count || /* or held (XXX may be wrong) */ 2602 m->wire_count || /* or wired */ 2603 (m->flags & (PG_UNMANAGED | /* or unmanaged */ 2604 PG_NEED_COMMIT)) || /* or needs a commit */ 2605 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */ 2606 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */ 2607 if (_vm_page_wakeup(m)) { 2608 vm_page_spin_unlock(m); 2609 wakeup(m); 2610 } else { 2611 vm_page_spin_unlock(m); 2612 } 2613 return(0); 2614 } 2615 vm_page_spin_unlock(m); 2616 2617 /* 2618 * We can probably free the page. 2619 * 2620 * Page busied by us and no longer spinlocked. Dirty pages will 2621 * not be freed by this function. We have to re-test the 2622 * dirty bit after cleaning out the pmaps. 2623 */ 2624 vm_page_test_dirty(m); 2625 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2626 vm_page_wakeup(m); 2627 return(0); 2628 } 2629 vm_page_protect(m, VM_PROT_NONE); 2630 if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2631 vm_page_wakeup(m); 2632 return(0); 2633 } 2634 vm_page_free(m); 2635 return(1); 2636 } 2637 2638 /* 2639 * vm_page_cache 2640 * 2641 * Put the specified page onto the page cache queue (if appropriate). 2642 * 2643 * The page must be busy, and this routine will release the busy and 2644 * possibly even free the page. 2645 */ 2646 void 2647 vm_page_cache(vm_page_t m) 2648 { 2649 /* 2650 * Not suitable for the cache 2651 */ 2652 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) || 2653 (m->busy_count & PBUSY_MASK) || 2654 m->wire_count || m->hold_count) { 2655 vm_page_wakeup(m); 2656 return; 2657 } 2658 2659 /* 2660 * Already in the cache (and thus not mapped) 2661 */ 2662 if ((m->queue - m->pc) == PQ_CACHE) { 2663 KKASSERT((m->flags & PG_MAPPED) == 0); 2664 vm_page_wakeup(m); 2665 return; 2666 } 2667 2668 /* 2669 * Caller is required to test m->dirty, but note that the act of 2670 * removing the page from its maps can cause it to become dirty 2671 * on an SMP system due to another cpu running in usermode. 2672 */ 2673 if (m->dirty) { 2674 panic("vm_page_cache: caching a dirty page, pindex: %ld", 2675 (long)m->pindex); 2676 } 2677 2678 /* 2679 * Remove all pmaps and indicate that the page is not 2680 * writeable or mapped. Our vm_page_protect() call may 2681 * have blocked (especially w/ VM_PROT_NONE), so recheck 2682 * everything. 2683 */ 2684 vm_page_protect(m, VM_PROT_NONE); 2685 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) || 2686 (m->busy_count & PBUSY_MASK) || 2687 m->wire_count || m->hold_count) { 2688 vm_page_wakeup(m); 2689 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) { 2690 vm_page_deactivate(m); 2691 vm_page_wakeup(m); 2692 } else { 2693 _vm_page_and_queue_spin_lock(m); 2694 _vm_page_rem_queue_spinlocked(m); 2695 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0); 2696 _vm_page_queue_spin_unlock(m); 2697 if (_vm_page_wakeup(m)) { 2698 vm_page_spin_unlock(m); 2699 wakeup(m); 2700 } else { 2701 vm_page_spin_unlock(m); 2702 } 2703 vm_page_free_wakeup(); 2704 } 2705 } 2706 2707 /* 2708 * vm_page_dontneed() 2709 * 2710 * Cache, deactivate, or do nothing as appropriate. This routine 2711 * is typically used by madvise() MADV_DONTNEED. 2712 * 2713 * Generally speaking we want to move the page into the cache so 2714 * it gets reused quickly. However, this can result in a silly syndrome 2715 * due to the page recycling too quickly. Small objects will not be 2716 * fully cached. On the otherhand, if we move the page to the inactive 2717 * queue we wind up with a problem whereby very large objects 2718 * unnecessarily blow away our inactive and cache queues. 2719 * 2720 * The solution is to move the pages based on a fixed weighting. We 2721 * either leave them alone, deactivate them, or move them to the cache, 2722 * where moving them to the cache has the highest weighting. 2723 * By forcing some pages into other queues we eventually force the 2724 * system to balance the queues, potentially recovering other unrelated 2725 * space from active. The idea is to not force this to happen too 2726 * often. 2727 * 2728 * The page must be busied. 2729 */ 2730 void 2731 vm_page_dontneed(vm_page_t m) 2732 { 2733 static int dnweight; 2734 int dnw; 2735 int head; 2736 2737 dnw = ++dnweight; 2738 2739 /* 2740 * occassionally leave the page alone 2741 */ 2742 if ((dnw & 0x01F0) == 0 || 2743 m->queue - m->pc == PQ_INACTIVE || 2744 m->queue - m->pc == PQ_CACHE 2745 ) { 2746 if (m->act_count >= ACT_INIT) 2747 --m->act_count; 2748 return; 2749 } 2750 2751 /* 2752 * If vm_page_dontneed() is inactivating a page, it must clear 2753 * the referenced flag; otherwise the pagedaemon will see references 2754 * on the page in the inactive queue and reactivate it. Until the 2755 * page can move to the cache queue, madvise's job is not done. 2756 */ 2757 vm_page_flag_clear(m, PG_REFERENCED); 2758 pmap_clear_reference(m); 2759 2760 if (m->dirty == 0) 2761 vm_page_test_dirty(m); 2762 2763 if (m->dirty || (dnw & 0x0070) == 0) { 2764 /* 2765 * Deactivate the page 3 times out of 32. 2766 */ 2767 head = 0; 2768 } else { 2769 /* 2770 * Cache the page 28 times out of every 32. Note that 2771 * the page is deactivated instead of cached, but placed 2772 * at the head of the queue instead of the tail. 2773 */ 2774 head = 1; 2775 } 2776 vm_page_spin_lock(m); 2777 _vm_page_deactivate_locked(m, head); 2778 vm_page_spin_unlock(m); 2779 } 2780 2781 /* 2782 * These routines manipulate the 'soft busy' count for a page. A soft busy 2783 * is almost like a hard BUSY except that it allows certain compatible 2784 * operations to occur on the page while it is busy. For example, a page 2785 * undergoing a write can still be mapped read-only. 2786 * 2787 * We also use soft-busy to quickly pmap_enter shared read-only pages 2788 * without having to hold the page locked. 2789 * 2790 * The soft-busy count can be > 1 in situations where multiple threads 2791 * are pmap_enter()ing the same page simultaneously, or when two buffer 2792 * cache buffers overlap the same page. 2793 * 2794 * The caller must hold the page BUSY when making these two calls. 2795 */ 2796 void 2797 vm_page_io_start(vm_page_t m) 2798 { 2799 uint32_t ocount; 2800 2801 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2802 KKASSERT(ocount & PBUSY_LOCKED); 2803 } 2804 2805 void 2806 vm_page_io_finish(vm_page_t m) 2807 { 2808 uint32_t ocount; 2809 2810 ocount = atomic_fetchadd_int(&m->busy_count, -1); 2811 KKASSERT(ocount & PBUSY_MASK); 2812 #if 0 2813 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0) 2814 wakeup(m); 2815 #endif 2816 } 2817 2818 /* 2819 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED. 2820 * 2821 * Returns 0 on success, non-zero on failure. 2822 */ 2823 int 2824 vm_page_sbusy_try(vm_page_t m) 2825 { 2826 uint32_t ocount; 2827 2828 if (m->busy_count & PBUSY_LOCKED) 2829 return 1; 2830 ocount = atomic_fetchadd_int(&m->busy_count, 1); 2831 if (ocount & PBUSY_LOCKED) { 2832 vm_page_sbusy_drop(m); 2833 return 1; 2834 } 2835 return 0; 2836 } 2837 2838 /* 2839 * Indicate that a clean VM page requires a filesystem commit and cannot 2840 * be reused. Used by tmpfs. 2841 */ 2842 void 2843 vm_page_need_commit(vm_page_t m) 2844 { 2845 vm_page_flag_set(m, PG_NEED_COMMIT); 2846 vm_object_set_writeable_dirty(m->object); 2847 } 2848 2849 void 2850 vm_page_clear_commit(vm_page_t m) 2851 { 2852 vm_page_flag_clear(m, PG_NEED_COMMIT); 2853 } 2854 2855 /* 2856 * Grab a page, blocking if it is busy and allocating a page if necessary. 2857 * A busy page is returned or NULL. The page may or may not be valid and 2858 * might not be on a queue (the caller is responsible for the disposition of 2859 * the page). 2860 * 2861 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the 2862 * page will be zero'd and marked valid. 2863 * 2864 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked 2865 * valid even if it already exists. 2866 * 2867 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also 2868 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified. 2869 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified. 2870 * 2871 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is 2872 * always returned if we had blocked. 2873 * 2874 * This routine may not be called from an interrupt. 2875 * 2876 * No other requirements. 2877 */ 2878 vm_page_t 2879 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2880 { 2881 vm_page_t m; 2882 int error; 2883 int shared = 1; 2884 2885 KKASSERT(allocflags & 2886 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM)); 2887 vm_object_hold_shared(object); 2888 for (;;) { 2889 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2890 if (error) { 2891 vm_page_sleep_busy(m, TRUE, "pgrbwt"); 2892 if ((allocflags & VM_ALLOC_RETRY) == 0) { 2893 m = NULL; 2894 break; 2895 } 2896 /* retry */ 2897 } else if (m == NULL) { 2898 if (shared) { 2899 vm_object_upgrade(object); 2900 shared = 0; 2901 } 2902 if (allocflags & VM_ALLOC_RETRY) 2903 allocflags |= VM_ALLOC_NULL_OK; 2904 m = vm_page_alloc(object, pindex, 2905 allocflags & ~VM_ALLOC_RETRY); 2906 if (m) 2907 break; 2908 vm_wait(0); 2909 if ((allocflags & VM_ALLOC_RETRY) == 0) 2910 goto failed; 2911 } else { 2912 /* m found */ 2913 break; 2914 } 2915 } 2916 2917 /* 2918 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid. 2919 * 2920 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set 2921 * valid even if already valid. 2922 * 2923 * NOTE! We have removed all of the PG_ZERO optimizations and also 2924 * removed the idle zeroing code. These optimizations actually 2925 * slow things down on modern cpus because the zerod area is 2926 * likely uncached, placing a memory-access burden on the 2927 * accesors taking the fault. 2928 * 2929 * By always zeroing the page in-line with the fault, no 2930 * dynamic ram reads are needed and the caches are hot, ready 2931 * for userland to access the memory. 2932 */ 2933 if (m->valid == 0) { 2934 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) { 2935 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2936 m->valid = VM_PAGE_BITS_ALL; 2937 } 2938 } else if (allocflags & VM_ALLOC_FORCE_ZERO) { 2939 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 2940 m->valid = VM_PAGE_BITS_ALL; 2941 } 2942 failed: 2943 vm_object_drop(object); 2944 return(m); 2945 } 2946 2947 /* 2948 * Mapping function for valid bits or for dirty bits in 2949 * a page. May not block. 2950 * 2951 * Inputs are required to range within a page. 2952 * 2953 * No requirements. 2954 * Non blocking. 2955 */ 2956 int 2957 vm_page_bits(int base, int size) 2958 { 2959 int first_bit; 2960 int last_bit; 2961 2962 KASSERT( 2963 base + size <= PAGE_SIZE, 2964 ("vm_page_bits: illegal base/size %d/%d", base, size) 2965 ); 2966 2967 if (size == 0) /* handle degenerate case */ 2968 return(0); 2969 2970 first_bit = base >> DEV_BSHIFT; 2971 last_bit = (base + size - 1) >> DEV_BSHIFT; 2972 2973 return ((2 << last_bit) - (1 << first_bit)); 2974 } 2975 2976 /* 2977 * Sets portions of a page valid and clean. The arguments are expected 2978 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2979 * of any partial chunks touched by the range. The invalid portion of 2980 * such chunks will be zero'd. 2981 * 2982 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically 2983 * align base to DEV_BSIZE so as not to mark clean a partially 2984 * truncated device block. Otherwise the dirty page status might be 2985 * lost. 2986 * 2987 * This routine may not block. 2988 * 2989 * (base + size) must be less then or equal to PAGE_SIZE. 2990 */ 2991 static void 2992 _vm_page_zero_valid(vm_page_t m, int base, int size) 2993 { 2994 int frag; 2995 int endoff; 2996 2997 if (size == 0) /* handle degenerate case */ 2998 return; 2999 3000 /* 3001 * If the base is not DEV_BSIZE aligned and the valid 3002 * bit is clear, we have to zero out a portion of the 3003 * first block. 3004 */ 3005 3006 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 3007 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 3008 ) { 3009 pmap_zero_page_area( 3010 VM_PAGE_TO_PHYS(m), 3011 frag, 3012 base - frag 3013 ); 3014 } 3015 3016 /* 3017 * If the ending offset is not DEV_BSIZE aligned and the 3018 * valid bit is clear, we have to zero out a portion of 3019 * the last block. 3020 */ 3021 3022 endoff = base + size; 3023 3024 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 3025 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 3026 ) { 3027 pmap_zero_page_area( 3028 VM_PAGE_TO_PHYS(m), 3029 endoff, 3030 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 3031 ); 3032 } 3033 } 3034 3035 /* 3036 * Set valid, clear dirty bits. If validating the entire 3037 * page we can safely clear the pmap modify bit. We also 3038 * use this opportunity to clear the PG_NOSYNC flag. If a process 3039 * takes a write fault on a MAP_NOSYNC memory area the flag will 3040 * be set again. 3041 * 3042 * We set valid bits inclusive of any overlap, but we can only 3043 * clear dirty bits for DEV_BSIZE chunks that are fully within 3044 * the range. 3045 * 3046 * Page must be busied? 3047 * No other requirements. 3048 */ 3049 void 3050 vm_page_set_valid(vm_page_t m, int base, int size) 3051 { 3052 _vm_page_zero_valid(m, base, size); 3053 m->valid |= vm_page_bits(base, size); 3054 } 3055 3056 3057 /* 3058 * Set valid bits and clear dirty bits. 3059 * 3060 * Page must be busied by caller. 3061 * 3062 * NOTE: This function does not clear the pmap modified bit. 3063 * Also note that e.g. NFS may use a byte-granular base 3064 * and size. 3065 * 3066 * No other requirements. 3067 */ 3068 void 3069 vm_page_set_validclean(vm_page_t m, int base, int size) 3070 { 3071 int pagebits; 3072 3073 _vm_page_zero_valid(m, base, size); 3074 pagebits = vm_page_bits(base, size); 3075 m->valid |= pagebits; 3076 m->dirty &= ~pagebits; 3077 if (base == 0 && size == PAGE_SIZE) { 3078 /*pmap_clear_modify(m);*/ 3079 vm_page_flag_clear(m, PG_NOSYNC); 3080 } 3081 } 3082 3083 /* 3084 * Set valid & dirty. Used by buwrite() 3085 * 3086 * Page must be busied by caller. 3087 */ 3088 void 3089 vm_page_set_validdirty(vm_page_t m, int base, int size) 3090 { 3091 int pagebits; 3092 3093 pagebits = vm_page_bits(base, size); 3094 m->valid |= pagebits; 3095 m->dirty |= pagebits; 3096 if (m->object) 3097 vm_object_set_writeable_dirty(m->object); 3098 } 3099 3100 /* 3101 * Clear dirty bits. 3102 * 3103 * NOTE: This function does not clear the pmap modified bit. 3104 * Also note that e.g. NFS may use a byte-granular base 3105 * and size. 3106 * 3107 * Page must be busied? 3108 * No other requirements. 3109 */ 3110 void 3111 vm_page_clear_dirty(vm_page_t m, int base, int size) 3112 { 3113 m->dirty &= ~vm_page_bits(base, size); 3114 if (base == 0 && size == PAGE_SIZE) { 3115 /*pmap_clear_modify(m);*/ 3116 vm_page_flag_clear(m, PG_NOSYNC); 3117 } 3118 } 3119 3120 /* 3121 * Make the page all-dirty. 3122 * 3123 * Also make sure the related object and vnode reflect the fact that the 3124 * object may now contain a dirty page. 3125 * 3126 * Page must be busied? 3127 * No other requirements. 3128 */ 3129 void 3130 vm_page_dirty(vm_page_t m) 3131 { 3132 #ifdef INVARIANTS 3133 int pqtype = m->queue - m->pc; 3134 #endif 3135 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE, 3136 ("vm_page_dirty: page in free/cache queue!")); 3137 if (m->dirty != VM_PAGE_BITS_ALL) { 3138 m->dirty = VM_PAGE_BITS_ALL; 3139 if (m->object) 3140 vm_object_set_writeable_dirty(m->object); 3141 } 3142 } 3143 3144 /* 3145 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3146 * valid and dirty bits for the effected areas are cleared. 3147 * 3148 * Page must be busied? 3149 * Does not block. 3150 * No other requirements. 3151 */ 3152 void 3153 vm_page_set_invalid(vm_page_t m, int base, int size) 3154 { 3155 int bits; 3156 3157 bits = vm_page_bits(base, size); 3158 m->valid &= ~bits; 3159 m->dirty &= ~bits; 3160 atomic_add_int(&m->object->generation, 1); 3161 } 3162 3163 /* 3164 * The kernel assumes that the invalid portions of a page contain 3165 * garbage, but such pages can be mapped into memory by user code. 3166 * When this occurs, we must zero out the non-valid portions of the 3167 * page so user code sees what it expects. 3168 * 3169 * Pages are most often semi-valid when the end of a file is mapped 3170 * into memory and the file's size is not page aligned. 3171 * 3172 * Page must be busied? 3173 * No other requirements. 3174 */ 3175 void 3176 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3177 { 3178 int b; 3179 int i; 3180 3181 /* 3182 * Scan the valid bits looking for invalid sections that 3183 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3184 * valid bit may be set ) have already been zerod by 3185 * vm_page_set_validclean(). 3186 */ 3187 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3188 if (i == (PAGE_SIZE / DEV_BSIZE) || 3189 (m->valid & (1 << i)) 3190 ) { 3191 if (i > b) { 3192 pmap_zero_page_area( 3193 VM_PAGE_TO_PHYS(m), 3194 b << DEV_BSHIFT, 3195 (i - b) << DEV_BSHIFT 3196 ); 3197 } 3198 b = i + 1; 3199 } 3200 } 3201 3202 /* 3203 * setvalid is TRUE when we can safely set the zero'd areas 3204 * as being valid. We can do this if there are no cache consistency 3205 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3206 */ 3207 if (setvalid) 3208 m->valid = VM_PAGE_BITS_ALL; 3209 } 3210 3211 /* 3212 * Is a (partial) page valid? Note that the case where size == 0 3213 * will return FALSE in the degenerate case where the page is entirely 3214 * invalid, and TRUE otherwise. 3215 * 3216 * Does not block. 3217 * No other requirements. 3218 */ 3219 int 3220 vm_page_is_valid(vm_page_t m, int base, int size) 3221 { 3222 int bits = vm_page_bits(base, size); 3223 3224 if (m->valid && ((m->valid & bits) == bits)) 3225 return 1; 3226 else 3227 return 0; 3228 } 3229 3230 /* 3231 * update dirty bits from pmap/mmu. May not block. 3232 * 3233 * Caller must hold the page busy 3234 */ 3235 void 3236 vm_page_test_dirty(vm_page_t m) 3237 { 3238 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 3239 vm_page_dirty(m); 3240 } 3241 } 3242 3243 #include "opt_ddb.h" 3244 #ifdef DDB 3245 #include <ddb/ddb.h> 3246 3247 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3248 { 3249 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count); 3250 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count); 3251 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count); 3252 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count); 3253 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count); 3254 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved); 3255 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min); 3256 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target); 3257 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min); 3258 db_printf("vmstats.v_inactive_target: %ld\n", 3259 vmstats.v_inactive_target); 3260 } 3261 3262 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3263 { 3264 int i; 3265 db_printf("PQ_FREE:"); 3266 for (i = 0; i < PQ_L2_SIZE; i++) { 3267 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 3268 } 3269 db_printf("\n"); 3270 3271 db_printf("PQ_CACHE:"); 3272 for(i = 0; i < PQ_L2_SIZE; i++) { 3273 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 3274 } 3275 db_printf("\n"); 3276 3277 db_printf("PQ_ACTIVE:"); 3278 for(i = 0; i < PQ_L2_SIZE; i++) { 3279 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt); 3280 } 3281 db_printf("\n"); 3282 3283 db_printf("PQ_INACTIVE:"); 3284 for(i = 0; i < PQ_L2_SIZE; i++) { 3285 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt); 3286 } 3287 db_printf("\n"); 3288 } 3289 #endif /* DDB */ 3290