1/*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2002-2006 Rice University 5 * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu> 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Alan L. Cox, 9 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 23 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 24 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 27 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY 30 * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 31 * POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34/* 35 * Physical memory system implementation 36 * 37 * Any external functions defined by this module are only to be used by the 38 * virtual memory system. 39 */ 40 41#include <sys/cdefs.h> 42#include "opt_ddb.h" 43#include "opt_vm.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/domainset.h> 48#include <sys/lock.h> 49#include <sys/kernel.h> 50#include <sys/malloc.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/queue.h> 54#include <sys/rwlock.h> 55#include <sys/sbuf.h> 56#include <sys/sysctl.h> 57#include <sys/tree.h> 58#include <sys/vmmeter.h> 59 60#include <ddb/ddb.h> 61 62#include <vm/vm.h> 63#include <vm/vm_extern.h> 64#include <vm/vm_param.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_object.h> 67#include <vm/vm_page.h> 68#include <vm/vm_phys.h> 69#include <vm/vm_pagequeue.h> 70 71_Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX, 72 "Too many physsegs."); 73_Static_assert(sizeof(long long) >= sizeof(vm_paddr_t), 74 "vm_paddr_t too big for ffsll, flsll."); 75 76#ifdef NUMA 77struct mem_affinity __read_mostly *mem_affinity; 78int __read_mostly *mem_locality; 79 80static int numa_disabled; 81static SYSCTL_NODE(_vm, OID_AUTO, numa, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 82 "NUMA options"); 83SYSCTL_INT(_vm_numa, OID_AUTO, disabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 84 &numa_disabled, 0, "NUMA-awareness in the allocators is disabled"); 85#endif 86 87int __read_mostly vm_ndomains = 1; 88domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1); 89 90struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX]; 91int __read_mostly vm_phys_nsegs; 92static struct vm_phys_seg vm_phys_early_segs[8]; 93static int vm_phys_early_nsegs; 94 95struct vm_phys_fictitious_seg; 96static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *, 97 struct vm_phys_fictitious_seg *); 98 99RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree = 100 RB_INITIALIZER(&vm_phys_fictitious_tree); 101 102struct vm_phys_fictitious_seg { 103 RB_ENTRY(vm_phys_fictitious_seg) node; 104 /* Memory region data */ 105 vm_paddr_t start; 106 vm_paddr_t end; 107 vm_page_t first_page; 108}; 109 110RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node, 111 vm_phys_fictitious_cmp); 112 113static struct rwlock_padalign vm_phys_fictitious_reg_lock; 114MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages"); 115 116static struct vm_freelist __aligned(CACHE_LINE_SIZE) 117 vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL] 118 [VM_NFREEORDER_MAX]; 119 120static int __read_mostly vm_nfreelists; 121 122/* 123 * These "avail lists" are globals used to communicate boot-time physical 124 * memory layout to other parts of the kernel. Each physically contiguous 125 * region of memory is defined by a start address at an even index and an 126 * end address at the following odd index. Each list is terminated by a 127 * pair of zero entries. 128 * 129 * dump_avail tells the dump code what regions to include in a crash dump, and 130 * phys_avail is all of the remaining physical memory that is available for 131 * the vm system. 132 * 133 * Initially dump_avail and phys_avail are identical. Boot time memory 134 * allocations remove extents from phys_avail that may still be included 135 * in dumps. 136 */ 137vm_paddr_t phys_avail[PHYS_AVAIL_COUNT]; 138vm_paddr_t dump_avail[PHYS_AVAIL_COUNT]; 139 140/* 141 * Provides the mapping from VM_FREELIST_* to free list indices (flind). 142 */ 143static int __read_mostly vm_freelist_to_flind[VM_NFREELIST]; 144 145CTASSERT(VM_FREELIST_DEFAULT == 0); 146 147#ifdef VM_FREELIST_DMA32 148#define VM_DMA32_BOUNDARY ((vm_paddr_t)1 << 32) 149#endif 150 151/* 152 * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about 153 * the ordering of the free list boundaries. 154 */ 155#if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY) 156CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY); 157#endif 158 159static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS); 160SYSCTL_OID(_vm, OID_AUTO, phys_free, 161 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 162 sysctl_vm_phys_free, "A", 163 "Phys Free Info"); 164 165static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS); 166SYSCTL_OID(_vm, OID_AUTO, phys_segs, 167 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 168 sysctl_vm_phys_segs, "A", 169 "Phys Seg Info"); 170 171#ifdef NUMA 172static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS); 173SYSCTL_OID(_vm, OID_AUTO, phys_locality, 174 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 175 sysctl_vm_phys_locality, "A", 176 "Phys Locality Info"); 177#endif 178 179SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD, 180 &vm_ndomains, 0, "Number of physical memory domains available."); 181 182static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain); 183static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end); 184static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, 185 int order, int tail); 186 187/* 188 * Red-black tree helpers for vm fictitious range management. 189 */ 190static inline int 191vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p, 192 struct vm_phys_fictitious_seg *range) 193{ 194 195 KASSERT(range->start != 0 && range->end != 0, 196 ("Invalid range passed on search for vm_fictitious page")); 197 if (p->start >= range->end) 198 return (1); 199 if (p->start < range->start) 200 return (-1); 201 202 return (0); 203} 204 205static int 206vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1, 207 struct vm_phys_fictitious_seg *p2) 208{ 209 210 /* Check if this is a search for a page */ 211 if (p1->end == 0) 212 return (vm_phys_fictitious_in_range(p1, p2)); 213 214 KASSERT(p2->end != 0, 215 ("Invalid range passed as second parameter to vm fictitious comparison")); 216 217 /* Searching to add a new range */ 218 if (p1->end <= p2->start) 219 return (-1); 220 if (p1->start >= p2->end) 221 return (1); 222 223 panic("Trying to add overlapping vm fictitious ranges:\n" 224 "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start, 225 (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end); 226} 227 228int 229vm_phys_domain_match(int prefer __numa_used, vm_paddr_t low __numa_used, 230 vm_paddr_t high __numa_used) 231{ 232#ifdef NUMA 233 domainset_t mask; 234 int i; 235 236 if (vm_ndomains == 1 || mem_affinity == NULL) 237 return (0); 238 239 DOMAINSET_ZERO(&mask); 240 /* 241 * Check for any memory that overlaps low, high. 242 */ 243 for (i = 0; mem_affinity[i].end != 0; i++) 244 if (mem_affinity[i].start <= high && 245 mem_affinity[i].end >= low) 246 DOMAINSET_SET(mem_affinity[i].domain, &mask); 247 if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask)) 248 return (prefer); 249 if (DOMAINSET_EMPTY(&mask)) 250 panic("vm_phys_domain_match: Impossible constraint"); 251 return (DOMAINSET_FFS(&mask) - 1); 252#else 253 return (0); 254#endif 255} 256 257/* 258 * Outputs the state of the physical memory allocator, specifically, 259 * the amount of physical memory in each free list. 260 */ 261static int 262sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS) 263{ 264 struct sbuf sbuf; 265 struct vm_freelist *fl; 266 int dom, error, flind, oind, pind; 267 268 error = sysctl_wire_old_buffer(req, 0); 269 if (error != 0) 270 return (error); 271 sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req); 272 for (dom = 0; dom < vm_ndomains; dom++) { 273 sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom); 274 for (flind = 0; flind < vm_nfreelists; flind++) { 275 sbuf_printf(&sbuf, "\nFREE LIST %d:\n" 276 "\n ORDER (SIZE) | NUMBER" 277 "\n ", flind); 278 for (pind = 0; pind < VM_NFREEPOOL; pind++) 279 sbuf_printf(&sbuf, " | POOL %d", pind); 280 sbuf_printf(&sbuf, "\n-- "); 281 for (pind = 0; pind < VM_NFREEPOOL; pind++) 282 sbuf_printf(&sbuf, "-- -- "); 283 sbuf_printf(&sbuf, "--\n"); 284 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { 285 sbuf_printf(&sbuf, " %2d (%6dK)", oind, 286 1 << (PAGE_SHIFT - 10 + oind)); 287 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 288 fl = vm_phys_free_queues[dom][flind][pind]; 289 sbuf_printf(&sbuf, " | %6d", 290 fl[oind].lcnt); 291 } 292 sbuf_printf(&sbuf, "\n"); 293 } 294 } 295 } 296 error = sbuf_finish(&sbuf); 297 sbuf_delete(&sbuf); 298 return (error); 299} 300 301/* 302 * Outputs the set of physical memory segments. 303 */ 304static int 305sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS) 306{ 307 struct sbuf sbuf; 308 struct vm_phys_seg *seg; 309 int error, segind; 310 311 error = sysctl_wire_old_buffer(req, 0); 312 if (error != 0) 313 return (error); 314 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 315 for (segind = 0; segind < vm_phys_nsegs; segind++) { 316 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind); 317 seg = &vm_phys_segs[segind]; 318 sbuf_printf(&sbuf, "start: %#jx\n", 319 (uintmax_t)seg->start); 320 sbuf_printf(&sbuf, "end: %#jx\n", 321 (uintmax_t)seg->end); 322 sbuf_printf(&sbuf, "domain: %d\n", seg->domain); 323 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues); 324 } 325 error = sbuf_finish(&sbuf); 326 sbuf_delete(&sbuf); 327 return (error); 328} 329 330/* 331 * Return affinity, or -1 if there's no affinity information. 332 */ 333int 334vm_phys_mem_affinity(int f __numa_used, int t __numa_used) 335{ 336 337#ifdef NUMA 338 if (mem_locality == NULL) 339 return (-1); 340 if (f >= vm_ndomains || t >= vm_ndomains) 341 return (-1); 342 return (mem_locality[f * vm_ndomains + t]); 343#else 344 return (-1); 345#endif 346} 347 348#ifdef NUMA 349/* 350 * Outputs the VM locality table. 351 */ 352static int 353sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS) 354{ 355 struct sbuf sbuf; 356 int error, i, j; 357 358 error = sysctl_wire_old_buffer(req, 0); 359 if (error != 0) 360 return (error); 361 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 362 363 sbuf_printf(&sbuf, "\n"); 364 365 for (i = 0; i < vm_ndomains; i++) { 366 sbuf_printf(&sbuf, "%d: ", i); 367 for (j = 0; j < vm_ndomains; j++) { 368 sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j)); 369 } 370 sbuf_printf(&sbuf, "\n"); 371 } 372 error = sbuf_finish(&sbuf); 373 sbuf_delete(&sbuf); 374 return (error); 375} 376#endif 377 378static void 379vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail) 380{ 381 382 m->order = order; 383 if (tail) 384 TAILQ_INSERT_TAIL(&fl[order].pl, m, listq); 385 else 386 TAILQ_INSERT_HEAD(&fl[order].pl, m, listq); 387 fl[order].lcnt++; 388} 389 390static void 391vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order) 392{ 393 394 TAILQ_REMOVE(&fl[order].pl, m, listq); 395 fl[order].lcnt--; 396 m->order = VM_NFREEORDER; 397} 398 399/* 400 * Create a physical memory segment. 401 */ 402static void 403_vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain) 404{ 405 struct vm_phys_seg *seg; 406 407 KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX, 408 ("vm_phys_create_seg: increase VM_PHYSSEG_MAX")); 409 KASSERT(domain >= 0 && domain < vm_ndomains, 410 ("vm_phys_create_seg: invalid domain provided")); 411 seg = &vm_phys_segs[vm_phys_nsegs++]; 412 while (seg > vm_phys_segs && (seg - 1)->start >= end) { 413 *seg = *(seg - 1); 414 seg--; 415 } 416 seg->start = start; 417 seg->end = end; 418 seg->domain = domain; 419} 420 421static void 422vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end) 423{ 424#ifdef NUMA 425 int i; 426 427 if (mem_affinity == NULL) { 428 _vm_phys_create_seg(start, end, 0); 429 return; 430 } 431 432 for (i = 0;; i++) { 433 if (mem_affinity[i].end == 0) 434 panic("Reached end of affinity info"); 435 if (mem_affinity[i].end <= start) 436 continue; 437 if (mem_affinity[i].start > start) 438 panic("No affinity info for start %jx", 439 (uintmax_t)start); 440 if (mem_affinity[i].end >= end) { 441 _vm_phys_create_seg(start, end, 442 mem_affinity[i].domain); 443 break; 444 } 445 _vm_phys_create_seg(start, mem_affinity[i].end, 446 mem_affinity[i].domain); 447 start = mem_affinity[i].end; 448 } 449#else 450 _vm_phys_create_seg(start, end, 0); 451#endif 452} 453 454/* 455 * Add a physical memory segment. 456 */ 457void 458vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end) 459{ 460 vm_paddr_t paddr; 461 462 KASSERT((start & PAGE_MASK) == 0, 463 ("vm_phys_define_seg: start is not page aligned")); 464 KASSERT((end & PAGE_MASK) == 0, 465 ("vm_phys_define_seg: end is not page aligned")); 466 467 /* 468 * Split the physical memory segment if it spans two or more free 469 * list boundaries. 470 */ 471 paddr = start; 472#ifdef VM_FREELIST_LOWMEM 473 if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) { 474 vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY); 475 paddr = VM_LOWMEM_BOUNDARY; 476 } 477#endif 478#ifdef VM_FREELIST_DMA32 479 if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) { 480 vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY); 481 paddr = VM_DMA32_BOUNDARY; 482 } 483#endif 484 vm_phys_create_seg(paddr, end); 485} 486 487/* 488 * Initialize the physical memory allocator. 489 * 490 * Requires that vm_page_array is initialized! 491 */ 492void 493vm_phys_init(void) 494{ 495 struct vm_freelist *fl; 496 struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg; 497#if defined(VM_DMA32_NPAGES_THRESHOLD) || defined(VM_PHYSSEG_SPARSE) 498 u_long npages; 499#endif 500 int dom, flind, freelist, oind, pind, segind; 501 502 /* 503 * Compute the number of free lists, and generate the mapping from the 504 * manifest constants VM_FREELIST_* to the free list indices. 505 * 506 * Initially, the entries of vm_freelist_to_flind[] are set to either 507 * 0 or 1 to indicate which free lists should be created. 508 */ 509#ifdef VM_DMA32_NPAGES_THRESHOLD 510 npages = 0; 511#endif 512 for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) { 513 seg = &vm_phys_segs[segind]; 514#ifdef VM_FREELIST_LOWMEM 515 if (seg->end <= VM_LOWMEM_BOUNDARY) 516 vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1; 517 else 518#endif 519#ifdef VM_FREELIST_DMA32 520 if ( 521#ifdef VM_DMA32_NPAGES_THRESHOLD 522 /* 523 * Create the DMA32 free list only if the amount of 524 * physical memory above physical address 4G exceeds the 525 * given threshold. 526 */ 527 npages > VM_DMA32_NPAGES_THRESHOLD && 528#endif 529 seg->end <= VM_DMA32_BOUNDARY) 530 vm_freelist_to_flind[VM_FREELIST_DMA32] = 1; 531 else 532#endif 533 { 534#ifdef VM_DMA32_NPAGES_THRESHOLD 535 npages += atop(seg->end - seg->start); 536#endif 537 vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1; 538 } 539 } 540 /* Change each entry into a running total of the free lists. */ 541 for (freelist = 1; freelist < VM_NFREELIST; freelist++) { 542 vm_freelist_to_flind[freelist] += 543 vm_freelist_to_flind[freelist - 1]; 544 } 545 vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1]; 546 KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists")); 547 /* Change each entry into a free list index. */ 548 for (freelist = 0; freelist < VM_NFREELIST; freelist++) 549 vm_freelist_to_flind[freelist]--; 550 551 /* 552 * Initialize the first_page and free_queues fields of each physical 553 * memory segment. 554 */ 555#ifdef VM_PHYSSEG_SPARSE 556 npages = 0; 557#endif 558 for (segind = 0; segind < vm_phys_nsegs; segind++) { 559 seg = &vm_phys_segs[segind]; 560#ifdef VM_PHYSSEG_SPARSE 561 seg->first_page = &vm_page_array[npages]; 562 npages += atop(seg->end - seg->start); 563#else 564 seg->first_page = PHYS_TO_VM_PAGE(seg->start); 565#endif 566#ifdef VM_FREELIST_LOWMEM 567 if (seg->end <= VM_LOWMEM_BOUNDARY) { 568 flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM]; 569 KASSERT(flind >= 0, 570 ("vm_phys_init: LOWMEM flind < 0")); 571 } else 572#endif 573#ifdef VM_FREELIST_DMA32 574 if (seg->end <= VM_DMA32_BOUNDARY) { 575 flind = vm_freelist_to_flind[VM_FREELIST_DMA32]; 576 KASSERT(flind >= 0, 577 ("vm_phys_init: DMA32 flind < 0")); 578 } else 579#endif 580 { 581 flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT]; 582 KASSERT(flind >= 0, 583 ("vm_phys_init: DEFAULT flind < 0")); 584 } 585 seg->free_queues = &vm_phys_free_queues[seg->domain][flind]; 586 } 587 588 /* 589 * Coalesce physical memory segments that are contiguous and share the 590 * same per-domain free queues. 591 */ 592 prev_seg = vm_phys_segs; 593 seg = &vm_phys_segs[1]; 594 end_seg = &vm_phys_segs[vm_phys_nsegs]; 595 while (seg < end_seg) { 596 if (prev_seg->end == seg->start && 597 prev_seg->free_queues == seg->free_queues) { 598 prev_seg->end = seg->end; 599 KASSERT(prev_seg->domain == seg->domain, 600 ("vm_phys_init: free queues cannot span domains")); 601 vm_phys_nsegs--; 602 end_seg--; 603 for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++) 604 *tmp_seg = *(tmp_seg + 1); 605 } else { 606 prev_seg = seg; 607 seg++; 608 } 609 } 610 611 /* 612 * Initialize the free queues. 613 */ 614 for (dom = 0; dom < vm_ndomains; dom++) { 615 for (flind = 0; flind < vm_nfreelists; flind++) { 616 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 617 fl = vm_phys_free_queues[dom][flind][pind]; 618 for (oind = 0; oind < VM_NFREEORDER; oind++) 619 TAILQ_INIT(&fl[oind].pl); 620 } 621 } 622 } 623 624 rw_init(&vm_phys_fictitious_reg_lock, "vmfctr"); 625} 626 627/* 628 * Register info about the NUMA topology of the system. 629 * 630 * Invoked by platform-dependent code prior to vm_phys_init(). 631 */ 632void 633vm_phys_register_domains(int ndomains __numa_used, 634 struct mem_affinity *affinity __numa_used, int *locality __numa_used) 635{ 636#ifdef NUMA 637 int i; 638 639 /* 640 * For now the only override value that we support is 1, which 641 * effectively disables NUMA-awareness in the allocators. 642 */ 643 TUNABLE_INT_FETCH("vm.numa.disabled", &numa_disabled); 644 if (numa_disabled) 645 ndomains = 1; 646 647 if (ndomains > 1) { 648 vm_ndomains = ndomains; 649 mem_affinity = affinity; 650 mem_locality = locality; 651 } 652 653 for (i = 0; i < vm_ndomains; i++) 654 DOMAINSET_SET(i, &all_domains); 655#endif 656} 657 658/* 659 * Split a contiguous, power of two-sized set of physical pages. 660 * 661 * When this function is called by a page allocation function, the caller 662 * should request insertion at the head unless the order [order, oind) queues 663 * are known to be empty. The objective being to reduce the likelihood of 664 * long-term fragmentation by promoting contemporaneous allocation and 665 * (hopefully) deallocation. 666 */ 667static __inline void 668vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order, 669 int tail) 670{ 671 vm_page_t m_buddy; 672 673 while (oind > order) { 674 oind--; 675 m_buddy = &m[1 << oind]; 676 KASSERT(m_buddy->order == VM_NFREEORDER, 677 ("vm_phys_split_pages: page %p has unexpected order %d", 678 m_buddy, m_buddy->order)); 679 vm_freelist_add(fl, m_buddy, oind, tail); 680 } 681} 682 683/* 684 * Add the physical pages [m, m + npages) at the beginning of a power-of-two 685 * aligned and sized set to the specified free list. 686 * 687 * When this function is called by a page allocation function, the caller 688 * should request insertion at the head unless the lower-order queues are 689 * known to be empty. The objective being to reduce the likelihood of long- 690 * term fragmentation by promoting contemporaneous allocation and (hopefully) 691 * deallocation. 692 * 693 * The physical page m's buddy must not be free. 694 */ 695static void 696vm_phys_enq_beg(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail) 697{ 698 int order; 699 700 KASSERT(npages == 0 || 701 (VM_PAGE_TO_PHYS(m) & 702 ((PAGE_SIZE << ilog2(npages)) - 1)) == 0, 703 ("%s: page %p and npages %u are misaligned", 704 __func__, m, npages)); 705 while (npages > 0) { 706 KASSERT(m->order == VM_NFREEORDER, 707 ("%s: page %p has unexpected order %d", 708 __func__, m, m->order)); 709 order = ilog2(npages); 710 KASSERT(order < VM_NFREEORDER, 711 ("%s: order %d is out of range", __func__, order)); 712 vm_freelist_add(fl, m, order, tail); 713 m += 1 << order; 714 npages -= 1 << order; 715 } 716} 717 718/* 719 * Add the physical pages [m, m + npages) at the end of a power-of-two aligned 720 * and sized set to the specified free list. 721 * 722 * When this function is called by a page allocation function, the caller 723 * should request insertion at the head unless the lower-order queues are 724 * known to be empty. The objective being to reduce the likelihood of long- 725 * term fragmentation by promoting contemporaneous allocation and (hopefully) 726 * deallocation. 727 * 728 * If npages is zero, this function does nothing and ignores the physical page 729 * parameter m. Otherwise, the physical page m's buddy must not be free. 730 */ 731static vm_page_t 732vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail) 733{ 734 int order; 735 736 KASSERT(npages == 0 || 737 ((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) & 738 ((PAGE_SIZE << ilog2(npages)) - 1)) == 0, 739 ("vm_phys_enq_range: page %p and npages %u are misaligned", 740 m, npages)); 741 while (npages > 0) { 742 KASSERT(m->order == VM_NFREEORDER, 743 ("vm_phys_enq_range: page %p has unexpected order %d", 744 m, m->order)); 745 order = ffs(npages) - 1; 746 KASSERT(order < VM_NFREEORDER, 747 ("vm_phys_enq_range: order %d is out of range", order)); 748 vm_freelist_add(fl, m, order, tail); 749 m += 1 << order; 750 npages -= 1 << order; 751 } 752 return (m); 753} 754 755/* 756 * Set the pool for a contiguous, power of two-sized set of physical pages. 757 */ 758static void 759vm_phys_set_pool(int pool, vm_page_t m, int order) 760{ 761 vm_page_t m_tmp; 762 763 for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++) 764 m_tmp->pool = pool; 765} 766 767/* 768 * Tries to allocate the specified number of pages from the specified pool 769 * within the specified domain. Returns the actual number of allocated pages 770 * and a pointer to each page through the array ma[]. 771 * 772 * The returned pages may not be physically contiguous. However, in contrast 773 * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0), 774 * calling this function once to allocate the desired number of pages will 775 * avoid wasted time in vm_phys_split_pages(). 776 * 777 * The free page queues for the specified domain must be locked. 778 */ 779int 780vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[]) 781{ 782 struct vm_freelist *alt, *fl; 783 vm_page_t m; 784 int avail, end, flind, freelist, i, oind, pind; 785 786 KASSERT(domain >= 0 && domain < vm_ndomains, 787 ("vm_phys_alloc_npages: domain %d is out of range", domain)); 788 KASSERT(pool < VM_NFREEPOOL, 789 ("vm_phys_alloc_npages: pool %d is out of range", pool)); 790 KASSERT(npages <= 1 << (VM_NFREEORDER - 1), 791 ("vm_phys_alloc_npages: npages %d is out of range", npages)); 792 vm_domain_free_assert_locked(VM_DOMAIN(domain)); 793 i = 0; 794 for (freelist = 0; freelist < VM_NFREELIST; freelist++) { 795 flind = vm_freelist_to_flind[freelist]; 796 if (flind < 0) 797 continue; 798 fl = vm_phys_free_queues[domain][flind][pool]; 799 for (oind = 0; oind < VM_NFREEORDER; oind++) { 800 while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) { 801 vm_freelist_rem(fl, m, oind); 802 avail = i + (1 << oind); 803 end = imin(npages, avail); 804 while (i < end) 805 ma[i++] = m++; 806 if (i == npages) { 807 /* 808 * Return excess pages to fl. Its order 809 * [0, oind) queues are empty. 810 */ 811 vm_phys_enq_range(m, avail - i, fl, 1); 812 return (npages); 813 } 814 } 815 } 816 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { 817 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 818 alt = vm_phys_free_queues[domain][flind][pind]; 819 while ((m = TAILQ_FIRST(&alt[oind].pl)) != 820 NULL) { 821 vm_freelist_rem(alt, m, oind); 822 vm_phys_set_pool(pool, m, oind); 823 avail = i + (1 << oind); 824 end = imin(npages, avail); 825 while (i < end) 826 ma[i++] = m++; 827 if (i == npages) { 828 /* 829 * Return excess pages to fl. 830 * Its order [0, oind) queues 831 * are empty. 832 */ 833 vm_phys_enq_range(m, avail - i, 834 fl, 1); 835 return (npages); 836 } 837 } 838 } 839 } 840 } 841 return (i); 842} 843 844/* 845 * Allocate a contiguous, power of two-sized set of physical pages 846 * from the free lists. 847 * 848 * The free page queues must be locked. 849 */ 850vm_page_t 851vm_phys_alloc_pages(int domain, int pool, int order) 852{ 853 vm_page_t m; 854 int freelist; 855 856 for (freelist = 0; freelist < VM_NFREELIST; freelist++) { 857 m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order); 858 if (m != NULL) 859 return (m); 860 } 861 return (NULL); 862} 863 864/* 865 * Allocate a contiguous, power of two-sized set of physical pages from the 866 * specified free list. The free list must be specified using one of the 867 * manifest constants VM_FREELIST_*. 868 * 869 * The free page queues must be locked. 870 */ 871vm_page_t 872vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order) 873{ 874 struct vm_freelist *alt, *fl; 875 vm_page_t m; 876 int oind, pind, flind; 877 878 KASSERT(domain >= 0 && domain < vm_ndomains, 879 ("vm_phys_alloc_freelist_pages: domain %d is out of range", 880 domain)); 881 KASSERT(freelist < VM_NFREELIST, 882 ("vm_phys_alloc_freelist_pages: freelist %d is out of range", 883 freelist)); 884 KASSERT(pool < VM_NFREEPOOL, 885 ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool)); 886 KASSERT(order < VM_NFREEORDER, 887 ("vm_phys_alloc_freelist_pages: order %d is out of range", order)); 888 889 flind = vm_freelist_to_flind[freelist]; 890 /* Check if freelist is present */ 891 if (flind < 0) 892 return (NULL); 893 894 vm_domain_free_assert_locked(VM_DOMAIN(domain)); 895 fl = &vm_phys_free_queues[domain][flind][pool][0]; 896 for (oind = order; oind < VM_NFREEORDER; oind++) { 897 m = TAILQ_FIRST(&fl[oind].pl); 898 if (m != NULL) { 899 vm_freelist_rem(fl, m, oind); 900 /* The order [order, oind) queues are empty. */ 901 vm_phys_split_pages(m, oind, fl, order, 1); 902 return (m); 903 } 904 } 905 906 /* 907 * The given pool was empty. Find the largest 908 * contiguous, power-of-two-sized set of pages in any 909 * pool. Transfer these pages to the given pool, and 910 * use them to satisfy the allocation. 911 */ 912 for (oind = VM_NFREEORDER - 1; oind >= order; oind--) { 913 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 914 alt = &vm_phys_free_queues[domain][flind][pind][0]; 915 m = TAILQ_FIRST(&alt[oind].pl); 916 if (m != NULL) { 917 vm_freelist_rem(alt, m, oind); 918 vm_phys_set_pool(pool, m, oind); 919 /* The order [order, oind) queues are empty. */ 920 vm_phys_split_pages(m, oind, fl, order, 1); 921 return (m); 922 } 923 } 924 } 925 return (NULL); 926} 927 928/* 929 * Find the vm_page corresponding to the given physical address. 930 */ 931vm_page_t 932vm_phys_paddr_to_vm_page(vm_paddr_t pa) 933{ 934 struct vm_phys_seg *seg; 935 936 if ((seg = vm_phys_paddr_to_seg(pa)) != NULL) 937 return (&seg->first_page[atop(pa - seg->start)]); 938 return (NULL); 939} 940 941vm_page_t 942vm_phys_fictitious_to_vm_page(vm_paddr_t pa) 943{ 944 struct vm_phys_fictitious_seg tmp, *seg; 945 vm_page_t m; 946 947 m = NULL; 948 tmp.start = pa; 949 tmp.end = 0; 950 951 rw_rlock(&vm_phys_fictitious_reg_lock); 952 seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp); 953 rw_runlock(&vm_phys_fictitious_reg_lock); 954 if (seg == NULL) 955 return (NULL); 956 957 m = &seg->first_page[atop(pa - seg->start)]; 958 KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m)); 959 960 return (m); 961} 962 963static inline void 964vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start, 965 long page_count, vm_memattr_t memattr) 966{ 967 long i; 968 969 bzero(range, page_count * sizeof(*range)); 970 for (i = 0; i < page_count; i++) { 971 vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr); 972 range[i].oflags &= ~VPO_UNMANAGED; 973 range[i].busy_lock = VPB_UNBUSIED; 974 } 975} 976 977int 978vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end, 979 vm_memattr_t memattr) 980{ 981 struct vm_phys_fictitious_seg *seg; 982 vm_page_t fp; 983 long page_count; 984#ifdef VM_PHYSSEG_DENSE 985 long pi, pe; 986 long dpage_count; 987#endif 988 989 KASSERT(start < end, 990 ("Start of segment isn't less than end (start: %jx end: %jx)", 991 (uintmax_t)start, (uintmax_t)end)); 992 993 page_count = (end - start) / PAGE_SIZE; 994 995#ifdef VM_PHYSSEG_DENSE 996 pi = atop(start); 997 pe = atop(end); 998 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 999 fp = &vm_page_array[pi - first_page]; 1000 if ((pe - first_page) > vm_page_array_size) { 1001 /* 1002 * We have a segment that starts inside 1003 * of vm_page_array, but ends outside of it. 1004 * 1005 * Use vm_page_array pages for those that are 1006 * inside of the vm_page_array range, and 1007 * allocate the remaining ones. 1008 */ 1009 dpage_count = vm_page_array_size - (pi - first_page); 1010 vm_phys_fictitious_init_range(fp, start, dpage_count, 1011 memattr); 1012 page_count -= dpage_count; 1013 start += ptoa(dpage_count); 1014 goto alloc; 1015 } 1016 /* 1017 * We can allocate the full range from vm_page_array, 1018 * so there's no need to register the range in the tree. 1019 */ 1020 vm_phys_fictitious_init_range(fp, start, page_count, memattr); 1021 return (0); 1022 } else if (pe > first_page && (pe - first_page) < vm_page_array_size) { 1023 /* 1024 * We have a segment that ends inside of vm_page_array, 1025 * but starts outside of it. 1026 */ 1027 fp = &vm_page_array[0]; 1028 dpage_count = pe - first_page; 1029 vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count, 1030 memattr); 1031 end -= ptoa(dpage_count); 1032 page_count -= dpage_count; 1033 goto alloc; 1034 } else if (pi < first_page && pe > (first_page + vm_page_array_size)) { 1035 /* 1036 * Trying to register a fictitious range that expands before 1037 * and after vm_page_array. 1038 */ 1039 return (EINVAL); 1040 } else { 1041alloc: 1042#endif 1043 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES, 1044 M_WAITOK); 1045#ifdef VM_PHYSSEG_DENSE 1046 } 1047#endif 1048 vm_phys_fictitious_init_range(fp, start, page_count, memattr); 1049 1050 seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO); 1051 seg->start = start; 1052 seg->end = end; 1053 seg->first_page = fp; 1054 1055 rw_wlock(&vm_phys_fictitious_reg_lock); 1056 RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg); 1057 rw_wunlock(&vm_phys_fictitious_reg_lock); 1058 1059 return (0); 1060} 1061 1062void 1063vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end) 1064{ 1065 struct vm_phys_fictitious_seg *seg, tmp; 1066#ifdef VM_PHYSSEG_DENSE 1067 long pi, pe; 1068#endif 1069 1070 KASSERT(start < end, 1071 ("Start of segment isn't less than end (start: %jx end: %jx)", 1072 (uintmax_t)start, (uintmax_t)end)); 1073 1074#ifdef VM_PHYSSEG_DENSE 1075 pi = atop(start); 1076 pe = atop(end); 1077 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1078 if ((pe - first_page) <= vm_page_array_size) { 1079 /* 1080 * This segment was allocated using vm_page_array 1081 * only, there's nothing to do since those pages 1082 * were never added to the tree. 1083 */ 1084 return; 1085 } 1086 /* 1087 * We have a segment that starts inside 1088 * of vm_page_array, but ends outside of it. 1089 * 1090 * Calculate how many pages were added to the 1091 * tree and free them. 1092 */ 1093 start = ptoa(first_page + vm_page_array_size); 1094 } else if (pe > first_page && (pe - first_page) < vm_page_array_size) { 1095 /* 1096 * We have a segment that ends inside of vm_page_array, 1097 * but starts outside of it. 1098 */ 1099 end = ptoa(first_page); 1100 } else if (pi < first_page && pe > (first_page + vm_page_array_size)) { 1101 /* Since it's not possible to register such a range, panic. */ 1102 panic( 1103 "Unregistering not registered fictitious range [%#jx:%#jx]", 1104 (uintmax_t)start, (uintmax_t)end); 1105 } 1106#endif 1107 tmp.start = start; 1108 tmp.end = 0; 1109 1110 rw_wlock(&vm_phys_fictitious_reg_lock); 1111 seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp); 1112 if (seg->start != start || seg->end != end) { 1113 rw_wunlock(&vm_phys_fictitious_reg_lock); 1114 panic( 1115 "Unregistering not registered fictitious range [%#jx:%#jx]", 1116 (uintmax_t)start, (uintmax_t)end); 1117 } 1118 RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg); 1119 rw_wunlock(&vm_phys_fictitious_reg_lock); 1120 free(seg->first_page, M_FICT_PAGES); 1121 free(seg, M_FICT_PAGES); 1122} 1123 1124/* 1125 * Free a contiguous, power of two-sized set of physical pages. 1126 * 1127 * The free page queues must be locked. 1128 */ 1129void 1130vm_phys_free_pages(vm_page_t m, int order) 1131{ 1132 struct vm_freelist *fl; 1133 struct vm_phys_seg *seg; 1134 vm_paddr_t pa; 1135 vm_page_t m_buddy; 1136 1137 KASSERT(m->order == VM_NFREEORDER, 1138 ("vm_phys_free_pages: page %p has unexpected order %d", 1139 m, m->order)); 1140 KASSERT(m->pool < VM_NFREEPOOL, 1141 ("vm_phys_free_pages: page %p has unexpected pool %d", 1142 m, m->pool)); 1143 KASSERT(order < VM_NFREEORDER, 1144 ("vm_phys_free_pages: order %d is out of range", order)); 1145 seg = &vm_phys_segs[m->segind]; 1146 vm_domain_free_assert_locked(VM_DOMAIN(seg->domain)); 1147 if (order < VM_NFREEORDER - 1) { 1148 pa = VM_PAGE_TO_PHYS(m); 1149 do { 1150 pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order)); 1151 if (pa < seg->start || pa >= seg->end) 1152 break; 1153 m_buddy = &seg->first_page[atop(pa - seg->start)]; 1154 if (m_buddy->order != order) 1155 break; 1156 fl = (*seg->free_queues)[m_buddy->pool]; 1157 vm_freelist_rem(fl, m_buddy, order); 1158 if (m_buddy->pool != m->pool) 1159 vm_phys_set_pool(m->pool, m_buddy, order); 1160 order++; 1161 pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1); 1162 m = &seg->first_page[atop(pa - seg->start)]; 1163 } while (order < VM_NFREEORDER - 1); 1164 } 1165 fl = (*seg->free_queues)[m->pool]; 1166 vm_freelist_add(fl, m, order, 1); 1167} 1168 1169/* 1170 * Free a contiguous, arbitrarily sized set of physical pages, without 1171 * merging across set boundaries. 1172 * 1173 * The free page queues must be locked. 1174 */ 1175void 1176vm_phys_enqueue_contig(vm_page_t m, u_long npages) 1177{ 1178 struct vm_freelist *fl; 1179 struct vm_phys_seg *seg; 1180 vm_page_t m_end; 1181 vm_paddr_t diff, lo; 1182 int order; 1183 1184 /* 1185 * Avoid unnecessary coalescing by freeing the pages in the largest 1186 * possible power-of-two-sized subsets. 1187 */ 1188 vm_domain_free_assert_locked(vm_pagequeue_domain(m)); 1189 seg = &vm_phys_segs[m->segind]; 1190 fl = (*seg->free_queues)[m->pool]; 1191 m_end = m + npages; 1192 /* Free blocks of increasing size. */ 1193 lo = atop(VM_PAGE_TO_PHYS(m)); 1194 if (m < m_end && 1195 (diff = lo ^ (lo + npages - 1)) != 0) { 1196 order = min(ilog2(diff), VM_NFREEORDER - 1); 1197 m = vm_phys_enq_range(m, roundup2(lo, 1 << order) - lo, fl, 1); 1198 } 1199 1200 /* Free blocks of maximum size. */ 1201 order = VM_NFREEORDER - 1; 1202 while (m + (1 << order) <= m_end) { 1203 KASSERT(seg == &vm_phys_segs[m->segind], 1204 ("%s: page range [%p,%p) spans multiple segments", 1205 __func__, m_end - npages, m)); 1206 vm_freelist_add(fl, m, order, 1); 1207 m += 1 << order; 1208 } 1209 /* Free blocks of diminishing size. */ 1210 vm_phys_enq_beg(m, m_end - m, fl, 1); 1211} 1212 1213/* 1214 * Free a contiguous, arbitrarily sized set of physical pages. 1215 * 1216 * The free page queues must be locked. 1217 */ 1218void 1219vm_phys_free_contig(vm_page_t m, u_long npages) 1220{ 1221 vm_paddr_t lo; 1222 vm_page_t m_start, m_end; 1223 unsigned max_order, order_start, order_end; 1224 1225 vm_domain_free_assert_locked(vm_pagequeue_domain(m)); 1226 1227 lo = atop(VM_PAGE_TO_PHYS(m)); 1228 max_order = min(ilog2(lo ^ (lo + npages)), VM_NFREEORDER - 1); 1229 1230 m_start = m; 1231 order_start = ffsll(lo) - 1; 1232 if (order_start < max_order) 1233 m_start += 1 << order_start; 1234 m_end = m + npages; 1235 order_end = ffsll(lo + npages) - 1; 1236 if (order_end < max_order) 1237 m_end -= 1 << order_end; 1238 /* 1239 * Avoid unnecessary coalescing by freeing the pages at the start and 1240 * end of the range last. 1241 */ 1242 if (m_start < m_end) 1243 vm_phys_enqueue_contig(m_start, m_end - m_start); 1244 if (order_start < max_order) 1245 vm_phys_free_pages(m, order_start); 1246 if (order_end < max_order) 1247 vm_phys_free_pages(m_end, order_end); 1248} 1249 1250/* 1251 * Identify the first address range within segment segind or greater 1252 * that matches the domain, lies within the low/high range, and has 1253 * enough pages. Return -1 if there is none. 1254 */ 1255int 1256vm_phys_find_range(vm_page_t bounds[], int segind, int domain, 1257 u_long npages, vm_paddr_t low, vm_paddr_t high) 1258{ 1259 vm_paddr_t pa_end, pa_start; 1260 struct vm_phys_seg *end_seg, *seg; 1261 1262 KASSERT(npages > 0, ("npages is zero")); 1263 KASSERT(domain >= 0 && domain < vm_ndomains, ("domain out of range")); 1264 end_seg = &vm_phys_segs[vm_phys_nsegs]; 1265 for (seg = &vm_phys_segs[segind]; seg < end_seg; seg++) { 1266 if (seg->domain != domain) 1267 continue; 1268 if (seg->start >= high) 1269 return (-1); 1270 pa_start = MAX(low, seg->start); 1271 pa_end = MIN(high, seg->end); 1272 if (pa_end - pa_start < ptoa(npages)) 1273 continue; 1274 bounds[0] = &seg->first_page[atop(pa_start - seg->start)]; 1275 bounds[1] = &seg->first_page[atop(pa_end - seg->start)]; 1276 return (seg - vm_phys_segs); 1277 } 1278 return (-1); 1279} 1280 1281/* 1282 * Search for the given physical page "m" in the free lists. If the search 1283 * succeeds, remove "m" from the free lists and return true. Otherwise, return 1284 * false, indicating that "m" is not in the free lists. 1285 * 1286 * The free page queues must be locked. 1287 */ 1288bool 1289vm_phys_unfree_page(vm_page_t m) 1290{ 1291 struct vm_freelist *fl; 1292 struct vm_phys_seg *seg; 1293 vm_paddr_t pa, pa_half; 1294 vm_page_t m_set, m_tmp; 1295 int order; 1296 1297 /* 1298 * First, find the contiguous, power of two-sized set of free 1299 * physical pages containing the given physical page "m" and 1300 * assign it to "m_set". 1301 */ 1302 seg = &vm_phys_segs[m->segind]; 1303 vm_domain_free_assert_locked(VM_DOMAIN(seg->domain)); 1304 for (m_set = m, order = 0; m_set->order == VM_NFREEORDER && 1305 order < VM_NFREEORDER - 1; ) { 1306 order++; 1307 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order)); 1308 if (pa >= seg->start) 1309 m_set = &seg->first_page[atop(pa - seg->start)]; 1310 else 1311 return (false); 1312 } 1313 if (m_set->order < order) 1314 return (false); 1315 if (m_set->order == VM_NFREEORDER) 1316 return (false); 1317 KASSERT(m_set->order < VM_NFREEORDER, 1318 ("vm_phys_unfree_page: page %p has unexpected order %d", 1319 m_set, m_set->order)); 1320 1321 /* 1322 * Next, remove "m_set" from the free lists. Finally, extract 1323 * "m" from "m_set" using an iterative algorithm: While "m_set" 1324 * is larger than a page, shrink "m_set" by returning the half 1325 * of "m_set" that does not contain "m" to the free lists. 1326 */ 1327 fl = (*seg->free_queues)[m_set->pool]; 1328 order = m_set->order; 1329 vm_freelist_rem(fl, m_set, order); 1330 while (order > 0) { 1331 order--; 1332 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order)); 1333 if (m->phys_addr < pa_half) 1334 m_tmp = &seg->first_page[atop(pa_half - seg->start)]; 1335 else { 1336 m_tmp = m_set; 1337 m_set = &seg->first_page[atop(pa_half - seg->start)]; 1338 } 1339 vm_freelist_add(fl, m_tmp, order, 0); 1340 } 1341 KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency")); 1342 return (true); 1343} 1344 1345/* 1346 * Find a run of contiguous physical pages, meeting alignment requirements, from 1347 * a list of max-sized page blocks, where we need at least two consecutive 1348 * blocks to satisfy the (large) page request. 1349 */ 1350static vm_page_t 1351vm_phys_find_freelist_contig(struct vm_freelist *fl, u_long npages, 1352 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 1353{ 1354 struct vm_phys_seg *seg; 1355 vm_page_t m, m_iter, m_ret; 1356 vm_paddr_t max_size, size; 1357 int max_order; 1358 1359 max_order = VM_NFREEORDER - 1; 1360 size = npages << PAGE_SHIFT; 1361 max_size = (vm_paddr_t)1 << (PAGE_SHIFT + max_order); 1362 KASSERT(size > max_size, ("size is too small")); 1363 1364 /* 1365 * In order to avoid examining any free max-sized page block more than 1366 * twice, identify the ones that are first in a physically-contiguous 1367 * sequence of such blocks, and only for those walk the sequence to 1368 * check if there are enough free blocks starting at a properly aligned 1369 * block. Thus, no block is checked for free-ness more than twice. 1370 */ 1371 TAILQ_FOREACH(m, &fl[max_order].pl, listq) { 1372 /* 1373 * Skip m unless it is first in a sequence of free max page 1374 * blocks >= low in its segment. 1375 */ 1376 seg = &vm_phys_segs[m->segind]; 1377 if (VM_PAGE_TO_PHYS(m) < MAX(low, seg->start)) 1378 continue; 1379 if (VM_PAGE_TO_PHYS(m) >= max_size && 1380 VM_PAGE_TO_PHYS(m) - max_size >= MAX(low, seg->start) && 1381 max_order == m[-1 << max_order].order) 1382 continue; 1383 1384 /* 1385 * Advance m_ret from m to the first of the sequence, if any, 1386 * that satisfies alignment conditions and might leave enough 1387 * space. 1388 */ 1389 m_ret = m; 1390 while (!vm_addr_ok(VM_PAGE_TO_PHYS(m_ret), 1391 size, alignment, boundary) && 1392 VM_PAGE_TO_PHYS(m_ret) + size <= MIN(high, seg->end) && 1393 max_order == m_ret[1 << max_order].order) 1394 m_ret += 1 << max_order; 1395 1396 /* 1397 * Skip m unless some block m_ret in the sequence is properly 1398 * aligned, and begins a sequence of enough pages less than 1399 * high, and in the same segment. 1400 */ 1401 if (VM_PAGE_TO_PHYS(m_ret) + size > MIN(high, seg->end)) 1402 continue; 1403 1404 /* 1405 * Skip m unless the blocks to allocate starting at m_ret are 1406 * all free. 1407 */ 1408 for (m_iter = m_ret; 1409 m_iter < m_ret + npages && max_order == m_iter->order; 1410 m_iter += 1 << max_order) { 1411 } 1412 if (m_iter < m_ret + npages) 1413 continue; 1414 return (m_ret); 1415 } 1416 return (NULL); 1417} 1418 1419/* 1420 * Find a run of contiguous physical pages from the specified free list 1421 * table. 1422 */ 1423static vm_page_t 1424vm_phys_find_queues_contig( 1425 struct vm_freelist (*queues)[VM_NFREEPOOL][VM_NFREEORDER_MAX], 1426 u_long npages, vm_paddr_t low, vm_paddr_t high, 1427 u_long alignment, vm_paddr_t boundary) 1428{ 1429 struct vm_freelist *fl; 1430 vm_page_t m_ret; 1431 vm_paddr_t pa, pa_end, size; 1432 int oind, order, pind; 1433 1434 KASSERT(npages > 0, ("npages is 0")); 1435 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 1436 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 1437 /* Compute the queue that is the best fit for npages. */ 1438 order = flsl(npages - 1); 1439 /* Search for a large enough free block. */ 1440 size = npages << PAGE_SHIFT; 1441 for (oind = order; oind < VM_NFREEORDER; oind++) { 1442 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 1443 fl = (*queues)[pind]; 1444 TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) { 1445 /* 1446 * Determine if the address range starting at pa 1447 * is within the given range, satisfies the 1448 * given alignment, and does not cross the given 1449 * boundary. 1450 */ 1451 pa = VM_PAGE_TO_PHYS(m_ret); 1452 pa_end = pa + size; 1453 if (low <= pa && pa_end <= high && 1454 vm_addr_ok(pa, size, alignment, boundary)) 1455 return (m_ret); 1456 } 1457 } 1458 } 1459 if (order < VM_NFREEORDER) 1460 return (NULL); 1461 /* Search for a long-enough sequence of max-order blocks. */ 1462 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 1463 fl = (*queues)[pind]; 1464 m_ret = vm_phys_find_freelist_contig(fl, npages, 1465 low, high, alignment, boundary); 1466 if (m_ret != NULL) 1467 return (m_ret); 1468 } 1469 return (NULL); 1470} 1471 1472/* 1473 * Allocate a contiguous set of physical pages of the given size 1474 * "npages" from the free lists. All of the physical pages must be at 1475 * or above the given physical address "low" and below the given 1476 * physical address "high". The given value "alignment" determines the 1477 * alignment of the first physical page in the set. If the given value 1478 * "boundary" is non-zero, then the set of physical pages cannot cross 1479 * any physical address boundary that is a multiple of that value. Both 1480 * "alignment" and "boundary" must be a power of two. 1481 */ 1482vm_page_t 1483vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high, 1484 u_long alignment, vm_paddr_t boundary) 1485{ 1486 vm_paddr_t pa_end, pa_start; 1487 struct vm_freelist *fl; 1488 vm_page_t m, m_run; 1489 struct vm_phys_seg *seg; 1490 struct vm_freelist (*queues)[VM_NFREEPOOL][VM_NFREEORDER_MAX]; 1491 int oind, segind; 1492 1493 KASSERT(npages > 0, ("npages is 0")); 1494 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 1495 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 1496 vm_domain_free_assert_locked(VM_DOMAIN(domain)); 1497 if (low >= high) 1498 return (NULL); 1499 queues = NULL; 1500 m_run = NULL; 1501 for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) { 1502 seg = &vm_phys_segs[segind]; 1503 if (seg->start >= high || seg->domain != domain) 1504 continue; 1505 if (low >= seg->end) 1506 break; 1507 if (low <= seg->start) 1508 pa_start = seg->start; 1509 else 1510 pa_start = low; 1511 if (high < seg->end) 1512 pa_end = high; 1513 else 1514 pa_end = seg->end; 1515 if (pa_end - pa_start < ptoa(npages)) 1516 continue; 1517 /* 1518 * If a previous segment led to a search using 1519 * the same free lists as would this segment, then 1520 * we've actually already searched within this 1521 * too. So skip it. 1522 */ 1523 if (seg->free_queues == queues) 1524 continue; 1525 queues = seg->free_queues; 1526 m_run = vm_phys_find_queues_contig(queues, npages, 1527 low, high, alignment, boundary); 1528 if (m_run != NULL) 1529 break; 1530 } 1531 if (m_run == NULL) 1532 return (NULL); 1533 1534 /* Allocate pages from the page-range found. */ 1535 for (m = m_run; m < &m_run[npages]; m = &m[1 << oind]) { 1536 fl = (*queues)[m->pool]; 1537 oind = m->order; 1538 vm_freelist_rem(fl, m, oind); 1539 if (m->pool != VM_FREEPOOL_DEFAULT) 1540 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind); 1541 } 1542 /* Return excess pages to the free lists. */ 1543 fl = (*queues)[VM_FREEPOOL_DEFAULT]; 1544 vm_phys_enq_range(&m_run[npages], m - &m_run[npages], fl, 0); 1545 1546 /* Return page verified to satisfy conditions of request. */ 1547 pa_start = VM_PAGE_TO_PHYS(m_run); 1548 KASSERT(low <= pa_start, 1549 ("memory allocated below minimum requested range")); 1550 KASSERT(pa_start + ptoa(npages) <= high, 1551 ("memory allocated above maximum requested range")); 1552 seg = &vm_phys_segs[m_run->segind]; 1553 KASSERT(seg->domain == domain, 1554 ("memory not allocated from specified domain")); 1555 KASSERT(vm_addr_ok(pa_start, ptoa(npages), alignment, boundary), 1556 ("memory alignment/boundary constraints not satisfied")); 1557 return (m_run); 1558} 1559 1560/* 1561 * Return the index of the first unused slot which may be the terminating 1562 * entry. 1563 */ 1564static int 1565vm_phys_avail_count(void) 1566{ 1567 int i; 1568 1569 for (i = 0; phys_avail[i + 1]; i += 2) 1570 continue; 1571 if (i > PHYS_AVAIL_ENTRIES) 1572 panic("Improperly terminated phys_avail %d entries", i); 1573 1574 return (i); 1575} 1576 1577/* 1578 * Assert that a phys_avail entry is valid. 1579 */ 1580static void 1581vm_phys_avail_check(int i) 1582{ 1583 if (phys_avail[i] & PAGE_MASK) 1584 panic("Unaligned phys_avail[%d]: %#jx", i, 1585 (intmax_t)phys_avail[i]); 1586 if (phys_avail[i+1] & PAGE_MASK) 1587 panic("Unaligned phys_avail[%d + 1]: %#jx", i, 1588 (intmax_t)phys_avail[i]); 1589 if (phys_avail[i + 1] < phys_avail[i]) 1590 panic("phys_avail[%d] start %#jx < end %#jx", i, 1591 (intmax_t)phys_avail[i], (intmax_t)phys_avail[i+1]); 1592} 1593 1594/* 1595 * Return the index of an overlapping phys_avail entry or -1. 1596 */ 1597#ifdef NUMA 1598static int 1599vm_phys_avail_find(vm_paddr_t pa) 1600{ 1601 int i; 1602 1603 for (i = 0; phys_avail[i + 1]; i += 2) 1604 if (phys_avail[i] <= pa && phys_avail[i + 1] > pa) 1605 return (i); 1606 return (-1); 1607} 1608#endif 1609 1610/* 1611 * Return the index of the largest entry. 1612 */ 1613int 1614vm_phys_avail_largest(void) 1615{ 1616 vm_paddr_t sz, largesz; 1617 int largest; 1618 int i; 1619 1620 largest = 0; 1621 largesz = 0; 1622 for (i = 0; phys_avail[i + 1]; i += 2) { 1623 sz = vm_phys_avail_size(i); 1624 if (sz > largesz) { 1625 largesz = sz; 1626 largest = i; 1627 } 1628 } 1629 1630 return (largest); 1631} 1632 1633vm_paddr_t 1634vm_phys_avail_size(int i) 1635{ 1636 1637 return (phys_avail[i + 1] - phys_avail[i]); 1638} 1639 1640/* 1641 * Split an entry at the address 'pa'. Return zero on success or errno. 1642 */ 1643static int 1644vm_phys_avail_split(vm_paddr_t pa, int i) 1645{ 1646 int cnt; 1647 1648 vm_phys_avail_check(i); 1649 if (pa <= phys_avail[i] || pa >= phys_avail[i + 1]) 1650 panic("vm_phys_avail_split: invalid address"); 1651 cnt = vm_phys_avail_count(); 1652 if (cnt >= PHYS_AVAIL_ENTRIES) 1653 return (ENOSPC); 1654 memmove(&phys_avail[i + 2], &phys_avail[i], 1655 (cnt - i) * sizeof(phys_avail[0])); 1656 phys_avail[i + 1] = pa; 1657 phys_avail[i + 2] = pa; 1658 vm_phys_avail_check(i); 1659 vm_phys_avail_check(i+2); 1660 1661 return (0); 1662} 1663 1664/* 1665 * Check if a given physical address can be included as part of a crash dump. 1666 */ 1667bool 1668vm_phys_is_dumpable(vm_paddr_t pa) 1669{ 1670 vm_page_t m; 1671 int i; 1672 1673 if ((m = vm_phys_paddr_to_vm_page(pa)) != NULL) 1674 return ((m->flags & PG_NODUMP) == 0); 1675 1676 for (i = 0; dump_avail[i] != 0 || dump_avail[i + 1] != 0; i += 2) { 1677 if (pa >= dump_avail[i] && pa < dump_avail[i + 1]) 1678 return (true); 1679 } 1680 return (false); 1681} 1682 1683void 1684vm_phys_early_add_seg(vm_paddr_t start, vm_paddr_t end) 1685{ 1686 struct vm_phys_seg *seg; 1687 1688 if (vm_phys_early_nsegs == -1) 1689 panic("%s: called after initialization", __func__); 1690 if (vm_phys_early_nsegs == nitems(vm_phys_early_segs)) 1691 panic("%s: ran out of early segments", __func__); 1692 1693 seg = &vm_phys_early_segs[vm_phys_early_nsegs++]; 1694 seg->start = start; 1695 seg->end = end; 1696} 1697 1698/* 1699 * This routine allocates NUMA node specific memory before the page 1700 * allocator is bootstrapped. 1701 */ 1702vm_paddr_t 1703vm_phys_early_alloc(int domain, size_t alloc_size) 1704{ 1705#ifdef NUMA 1706 int mem_index; 1707#endif 1708 int i, biggestone; 1709 vm_paddr_t pa, mem_start, mem_end, size, biggestsize, align; 1710 1711 KASSERT(domain == -1 || (domain >= 0 && domain < vm_ndomains), 1712 ("%s: invalid domain index %d", __func__, domain)); 1713 1714 /* 1715 * Search the mem_affinity array for the biggest address 1716 * range in the desired domain. This is used to constrain 1717 * the phys_avail selection below. 1718 */ 1719 biggestsize = 0; 1720 mem_start = 0; 1721 mem_end = -1; 1722#ifdef NUMA 1723 mem_index = 0; 1724 if (mem_affinity != NULL) { 1725 for (i = 0;; i++) { 1726 size = mem_affinity[i].end - mem_affinity[i].start; 1727 if (size == 0) 1728 break; 1729 if (domain != -1 && mem_affinity[i].domain != domain) 1730 continue; 1731 if (size > biggestsize) { 1732 mem_index = i; 1733 biggestsize = size; 1734 } 1735 } 1736 mem_start = mem_affinity[mem_index].start; 1737 mem_end = mem_affinity[mem_index].end; 1738 } 1739#endif 1740 1741 /* 1742 * Now find biggest physical segment in within the desired 1743 * numa domain. 1744 */ 1745 biggestsize = 0; 1746 biggestone = 0; 1747 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 1748 /* skip regions that are out of range */ 1749 if (phys_avail[i+1] - alloc_size < mem_start || 1750 phys_avail[i+1] > mem_end) 1751 continue; 1752 size = vm_phys_avail_size(i); 1753 if (size > biggestsize) { 1754 biggestone = i; 1755 biggestsize = size; 1756 } 1757 } 1758 alloc_size = round_page(alloc_size); 1759 1760 /* 1761 * Grab single pages from the front to reduce fragmentation. 1762 */ 1763 if (alloc_size == PAGE_SIZE) { 1764 pa = phys_avail[biggestone]; 1765 phys_avail[biggestone] += PAGE_SIZE; 1766 vm_phys_avail_check(biggestone); 1767 return (pa); 1768 } 1769 1770 /* 1771 * Naturally align large allocations. 1772 */ 1773 align = phys_avail[biggestone + 1] & (alloc_size - 1); 1774 if (alloc_size + align > biggestsize) 1775 panic("cannot find a large enough size\n"); 1776 if (align != 0 && 1777 vm_phys_avail_split(phys_avail[biggestone + 1] - align, 1778 biggestone) != 0) 1779 /* Wasting memory. */ 1780 phys_avail[biggestone + 1] -= align; 1781 1782 phys_avail[biggestone + 1] -= alloc_size; 1783 vm_phys_avail_check(biggestone); 1784 pa = phys_avail[biggestone + 1]; 1785 return (pa); 1786} 1787 1788void 1789vm_phys_early_startup(void) 1790{ 1791 struct vm_phys_seg *seg; 1792 int i; 1793 1794 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 1795 phys_avail[i] = round_page(phys_avail[i]); 1796 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 1797 } 1798 1799 for (i = 0; i < vm_phys_early_nsegs; i++) { 1800 seg = &vm_phys_early_segs[i]; 1801 vm_phys_add_seg(seg->start, seg->end); 1802 } 1803 vm_phys_early_nsegs = -1; 1804 1805#ifdef NUMA 1806 /* Force phys_avail to be split by domain. */ 1807 if (mem_affinity != NULL) { 1808 int idx; 1809 1810 for (i = 0; mem_affinity[i].end != 0; i++) { 1811 idx = vm_phys_avail_find(mem_affinity[i].start); 1812 if (idx != -1 && 1813 phys_avail[idx] != mem_affinity[i].start) 1814 vm_phys_avail_split(mem_affinity[i].start, idx); 1815 idx = vm_phys_avail_find(mem_affinity[i].end); 1816 if (idx != -1 && 1817 phys_avail[idx] != mem_affinity[i].end) 1818 vm_phys_avail_split(mem_affinity[i].end, idx); 1819 } 1820 } 1821#endif 1822} 1823 1824#ifdef DDB 1825/* 1826 * Show the number of physical pages in each of the free lists. 1827 */ 1828DB_SHOW_COMMAND_FLAGS(freepages, db_show_freepages, DB_CMD_MEMSAFE) 1829{ 1830 struct vm_freelist *fl; 1831 int flind, oind, pind, dom; 1832 1833 for (dom = 0; dom < vm_ndomains; dom++) { 1834 db_printf("DOMAIN: %d\n", dom); 1835 for (flind = 0; flind < vm_nfreelists; flind++) { 1836 db_printf("FREE LIST %d:\n" 1837 "\n ORDER (SIZE) | NUMBER" 1838 "\n ", flind); 1839 for (pind = 0; pind < VM_NFREEPOOL; pind++) 1840 db_printf(" | POOL %d", pind); 1841 db_printf("\n-- "); 1842 for (pind = 0; pind < VM_NFREEPOOL; pind++) 1843 db_printf("-- -- "); 1844 db_printf("--\n"); 1845 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) { 1846 db_printf(" %2.2d (%6.6dK)", oind, 1847 1 << (PAGE_SHIFT - 10 + oind)); 1848 for (pind = 0; pind < VM_NFREEPOOL; pind++) { 1849 fl = vm_phys_free_queues[dom][flind][pind]; 1850 db_printf(" | %6.6d", fl[oind].lcnt); 1851 } 1852 db_printf("\n"); 1853 } 1854 db_printf("\n"); 1855 } 1856 db_printf("\n"); 1857 } 1858} 1859#endif 1860