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