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