1 // SPDX-License-Identifier: GPL-2.0-or-later 2 3 #include <linux/array_size.h> 4 #include <linux/sort.h> 5 #include <linux/printk.h> 6 #include <linux/memblock.h> 7 #include <linux/numa.h> 8 #include <linux/numa_memblks.h> 9 10 #include <asm/numa.h> 11 12 int numa_distance_cnt; 13 static u8 *numa_distance; 14 15 nodemask_t numa_nodes_parsed __initdata; 16 17 static struct numa_meminfo numa_meminfo __initdata_or_meminfo; 18 static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo; 19 20 /** 21 * numa_reset_distance - Reset NUMA distance table 22 * 23 * The current table is freed. The next numa_set_distance() call will 24 * create a new one. 25 */ 26 void __init numa_reset_distance(void) 27 { 28 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); 29 30 /* numa_distance could be 1LU marking allocation failure, test cnt */ 31 if (numa_distance_cnt) 32 memblock_free(numa_distance, size); 33 numa_distance_cnt = 0; 34 numa_distance = NULL; /* enable table creation */ 35 } 36 37 static int __init numa_alloc_distance(void) 38 { 39 nodemask_t nodes_parsed; 40 size_t size; 41 int i, j, cnt = 0; 42 43 /* size the new table and allocate it */ 44 nodes_parsed = numa_nodes_parsed; 45 46 for_each_node_mask(i, nodes_parsed) 47 cnt = i; 48 cnt++; 49 size = cnt * cnt * sizeof(numa_distance[0]); 50 51 numa_distance = memblock_alloc(size, PAGE_SIZE); 52 if (!numa_distance) { 53 pr_warn("Warning: can't allocate distance table!\n"); 54 /* don't retry until explicitly reset */ 55 numa_distance = (void *)1LU; 56 return -ENOMEM; 57 } 58 59 numa_distance_cnt = cnt; 60 61 /* fill with the default distances */ 62 for (i = 0; i < cnt; i++) 63 for (j = 0; j < cnt; j++) 64 numa_distance[i * cnt + j] = i == j ? 65 LOCAL_DISTANCE : REMOTE_DISTANCE; 66 pr_debug("NUMA: Initialized distance table, cnt=%d\n", cnt); 67 68 return 0; 69 } 70 71 /** 72 * numa_set_distance - Set NUMA distance from one NUMA to another 73 * @from: the 'from' node to set distance 74 * @to: the 'to' node to set distance 75 * @distance: NUMA distance 76 * 77 * Set the distance from node @from to @to to @distance. If distance table 78 * doesn't exist, one which is large enough to accommodate all the currently 79 * known nodes will be created. 80 * 81 * If such table cannot be allocated, a warning is printed and further 82 * calls are ignored until the distance table is reset with 83 * numa_reset_distance(). 84 * 85 * If @from or @to is higher than the highest known node or lower than zero 86 * at the time of table creation or @distance doesn't make sense, the call 87 * is ignored. 88 * This is to allow simplification of specific NUMA config implementations. 89 */ 90 void __init numa_set_distance(int from, int to, int distance) 91 { 92 if (!numa_distance && numa_alloc_distance() < 0) 93 return; 94 95 if (from >= numa_distance_cnt || to >= numa_distance_cnt || 96 from < 0 || to < 0) { 97 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n", 98 from, to, distance); 99 return; 100 } 101 102 if ((u8)distance != distance || 103 (from == to && distance != LOCAL_DISTANCE)) { 104 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n", 105 from, to, distance); 106 return; 107 } 108 109 numa_distance[from * numa_distance_cnt + to] = distance; 110 } 111 112 int __node_distance(int from, int to) 113 { 114 if (from >= numa_distance_cnt || to >= numa_distance_cnt) 115 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; 116 return numa_distance[from * numa_distance_cnt + to]; 117 } 118 EXPORT_SYMBOL(__node_distance); 119 120 static int __init numa_add_memblk_to(int nid, u64 start, u64 end, 121 struct numa_meminfo *mi) 122 { 123 /* whine about and ignore invalid nid */ 124 if (nid < 0 || nid >= MAX_NUMNODES) { 125 pr_warn("Warning: invalid memblk node id %d [mem %#010Lx-%#010Lx]\n", 126 nid, start, end - 1); 127 return -EINVAL; 128 } 129 130 /* ignore zero length blks */ 131 if (start == end) 132 return 0; 133 134 /* whine about and ignore invalid ranges */ 135 if (start > end) { 136 pr_warn("Warning: invalid memblk range for node %d [mem %#010Lx-%#010Lx]\n", 137 nid, start, end - 1); 138 return 0; 139 } 140 141 if (mi->nr_blks >= NR_NODE_MEMBLKS) { 142 pr_err("too many memblk ranges\n"); 143 return -EINVAL; 144 } 145 146 mi->blk[mi->nr_blks].start = start; 147 mi->blk[mi->nr_blks].end = end; 148 mi->blk[mi->nr_blks].nid = nid; 149 mi->nr_blks++; 150 return 0; 151 } 152 153 /** 154 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo 155 * @idx: Index of memblk to remove 156 * @mi: numa_meminfo to remove memblk from 157 * 158 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and 159 * decrementing @mi->nr_blks. 160 */ 161 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi) 162 { 163 mi->nr_blks--; 164 memmove(&mi->blk[idx], &mi->blk[idx + 1], 165 (mi->nr_blks - idx) * sizeof(mi->blk[0])); 166 } 167 168 /** 169 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another 170 * @dst: numa_meminfo to append block to 171 * @idx: Index of memblk to remove 172 * @src: numa_meminfo to remove memblk from 173 */ 174 static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx, 175 struct numa_meminfo *src) 176 { 177 dst->blk[dst->nr_blks++] = src->blk[idx]; 178 numa_remove_memblk_from(idx, src); 179 } 180 181 /** 182 * numa_add_memblk - Add one numa_memblk to numa_meminfo 183 * @nid: NUMA node ID of the new memblk 184 * @start: Start address of the new memblk 185 * @end: End address of the new memblk 186 * 187 * Add a new memblk to the default numa_meminfo. 188 * On success @nid is also set in numa_nodes_parsed. 189 * 190 * RETURNS: 191 * 0 on success, -errno on failure. 192 */ 193 int __init numa_add_memblk(int nid, u64 start, u64 end) 194 { 195 int ret; 196 197 ret = numa_add_memblk_to(nid, start, end, &numa_meminfo); 198 if (!ret) 199 node_set(nid, numa_nodes_parsed); 200 201 return ret; 202 } 203 204 /** 205 * numa_add_reserved_memblk - Add one numa_memblk to numa_reserved_meminfo 206 * @nid: NUMA node ID of the new memblk 207 * @start: Start address of the new memblk 208 * @end: End address of the new memblk 209 * 210 * Add a new memblk to the numa_reserved_meminfo. 211 * 212 * Usage Case: numa_cleanup_meminfo() reconciles all numa_memblk instances 213 * against memblock_type information and moves any that intersect reserved 214 * ranges to numa_reserved_meminfo. However, when that information is known 215 * ahead of time, we use numa_add_reserved_memblk() to add the numa_memblk 216 * to numa_reserved_meminfo directly. 217 * 218 * RETURNS: 219 * 0 on success, -errno on failure. 220 */ 221 int __init numa_add_reserved_memblk(int nid, u64 start, u64 end) 222 { 223 return numa_add_memblk_to(nid, start, end, &numa_reserved_meminfo); 224 } 225 226 /** 227 * numa_cleanup_meminfo - Cleanup a numa_meminfo 228 * @mi: numa_meminfo to clean up 229 * 230 * Sanitize @mi by merging and removing unnecessary memblks. Also check for 231 * conflicts and clear unused memblks. 232 * 233 * RETURNS: 234 * 0 on success, -errno on failure. 235 */ 236 int __init numa_cleanup_meminfo(struct numa_meminfo *mi) 237 { 238 const u64 low = memblock_start_of_DRAM(); 239 const u64 high = memblock_end_of_DRAM(); 240 int i, j, k; 241 242 /* first, trim all entries */ 243 for (i = 0; i < mi->nr_blks; i++) { 244 struct numa_memblk *bi = &mi->blk[i]; 245 246 /* move / save reserved memory ranges */ 247 if (!memblock_overlaps_region(&memblock.memory, 248 bi->start, bi->end - bi->start)) { 249 numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi); 250 continue; 251 } 252 253 /* make sure all non-reserved blocks are inside the limits */ 254 bi->start = max(bi->start, low); 255 256 /* preserve info for non-RAM areas above 'max_pfn': */ 257 if (bi->end > high) { 258 numa_add_reserved_memblk(bi->nid, high, bi->end); 259 bi->end = high; 260 } 261 262 /* and there's no empty block */ 263 if (bi->start >= bi->end) 264 numa_remove_memblk_from(i--, mi); 265 } 266 267 /* merge neighboring / overlapping entries */ 268 for (i = 0; i < mi->nr_blks; i++) { 269 struct numa_memblk *bi = &mi->blk[i]; 270 271 for (j = i + 1; j < mi->nr_blks; j++) { 272 struct numa_memblk *bj = &mi->blk[j]; 273 u64 start, end; 274 275 /* 276 * See whether there are overlapping blocks. Whine 277 * about but allow overlaps of the same nid. They 278 * will be merged below. 279 */ 280 if (bi->end > bj->start && bi->start < bj->end) { 281 if (bi->nid != bj->nid) { 282 pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n", 283 bi->nid, bi->start, bi->end - 1, 284 bj->nid, bj->start, bj->end - 1); 285 return -EINVAL; 286 } 287 pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n", 288 bi->nid, bi->start, bi->end - 1, 289 bj->start, bj->end - 1); 290 } 291 292 /* 293 * Join together blocks on the same node, holes 294 * between which don't overlap with memory on other 295 * nodes. 296 */ 297 if (bi->nid != bj->nid) 298 continue; 299 start = min(bi->start, bj->start); 300 end = max(bi->end, bj->end); 301 for (k = 0; k < mi->nr_blks; k++) { 302 struct numa_memblk *bk = &mi->blk[k]; 303 304 if (bi->nid == bk->nid) 305 continue; 306 if (start < bk->end && end > bk->start) 307 break; 308 } 309 if (k < mi->nr_blks) 310 continue; 311 pr_info("NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n", 312 bi->nid, bi->start, bi->end - 1, bj->start, 313 bj->end - 1, start, end - 1); 314 bi->start = start; 315 bi->end = end; 316 numa_remove_memblk_from(j--, mi); 317 } 318 } 319 320 /* clear unused ones */ 321 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) { 322 mi->blk[i].start = mi->blk[i].end = 0; 323 mi->blk[i].nid = NUMA_NO_NODE; 324 } 325 326 return 0; 327 } 328 329 /* 330 * Mark all currently memblock-reserved physical memory (which covers the 331 * kernel's own memory ranges) as hot-unswappable. 332 */ 333 static void __init numa_clear_kernel_node_hotplug(void) 334 { 335 nodemask_t reserved_nodemask = NODE_MASK_NONE; 336 struct memblock_region *mb_region; 337 int i; 338 339 /* 340 * We have to do some preprocessing of memblock regions, to 341 * make them suitable for reservation. 342 * 343 * At this time, all memory regions reserved by memblock are 344 * used by the kernel, but those regions are not split up 345 * along node boundaries yet, and don't necessarily have their 346 * node ID set yet either. 347 * 348 * So iterate over all parsed memory blocks and use those ranges to 349 * set the nid in memblock.reserved. This will split up the 350 * memblock regions along node boundaries and will set the node IDs 351 * as well. 352 */ 353 for (i = 0; i < numa_meminfo.nr_blks; i++) { 354 struct numa_memblk *mb = numa_meminfo.blk + i; 355 int ret; 356 357 ret = memblock_set_node(mb->start, mb->end - mb->start, 358 &memblock.reserved, mb->nid); 359 WARN_ON_ONCE(ret); 360 } 361 362 /* 363 * Now go over all reserved memblock regions, to construct a 364 * node mask of all kernel reserved memory areas. 365 * 366 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel, 367 * numa_meminfo might not include all memblock.reserved 368 * memory ranges, because quirks such as trim_snb_memory() 369 * reserve specific pages for Sandy Bridge graphics. ] 370 */ 371 for_each_reserved_mem_region(mb_region) { 372 int nid = memblock_get_region_node(mb_region); 373 374 if (numa_valid_node(nid)) 375 node_set(nid, reserved_nodemask); 376 } 377 378 /* 379 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory 380 * belonging to the reserved node mask. 381 * 382 * Note that this will include memory regions that reside 383 * on nodes that contain kernel memory - entire nodes 384 * become hot-unpluggable: 385 */ 386 for (i = 0; i < numa_meminfo.nr_blks; i++) { 387 struct numa_memblk *mb = numa_meminfo.blk + i; 388 389 if (!node_isset(mb->nid, reserved_nodemask)) 390 continue; 391 392 memblock_clear_hotplug(mb->start, mb->end - mb->start); 393 } 394 } 395 396 static int __init numa_register_meminfo(struct numa_meminfo *mi) 397 { 398 int i; 399 400 /* Account for nodes with cpus and no memory */ 401 node_possible_map = numa_nodes_parsed; 402 if (WARN_ON(nodes_empty(node_possible_map))) 403 return -EINVAL; 404 405 for (i = 0; i < mi->nr_blks; i++) { 406 struct numa_memblk *mb = &mi->blk[i]; 407 408 memblock_set_node(mb->start, mb->end - mb->start, 409 &memblock.memory, mb->nid); 410 } 411 412 /* 413 * At very early time, the kernel have to use some memory such as 414 * loading the kernel image. We cannot prevent this anyway. So any 415 * node the kernel resides in should be un-hotpluggable. 416 * 417 * And when we come here, alloc node data won't fail. 418 */ 419 numa_clear_kernel_node_hotplug(); 420 421 /* 422 * If sections array is gonna be used for pfn -> nid mapping, check 423 * whether its granularity is fine enough. 424 */ 425 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) { 426 unsigned long pfn_align = node_map_pfn_alignment(); 427 428 if (pfn_align && pfn_align < PAGES_PER_SECTION) { 429 unsigned long node_align_mb = PFN_PHYS(pfn_align) / SZ_1M; 430 431 unsigned long sect_align_mb = PFN_PHYS(PAGES_PER_SECTION) / SZ_1M; 432 433 pr_warn("Node alignment %luMB < min %luMB, rejecting NUMA config\n", 434 node_align_mb, sect_align_mb); 435 return -EINVAL; 436 } 437 } 438 439 return 0; 440 } 441 442 int __init numa_memblks_init(int (*init_func)(void), 443 bool memblock_force_top_down) 444 { 445 phys_addr_t max_addr = (phys_addr_t)ULLONG_MAX; 446 int ret; 447 448 nodes_clear(numa_nodes_parsed); 449 nodes_clear(node_possible_map); 450 nodes_clear(node_online_map); 451 memset(&numa_meminfo, 0, sizeof(numa_meminfo)); 452 WARN_ON(memblock_set_node(0, max_addr, &memblock.memory, NUMA_NO_NODE)); 453 WARN_ON(memblock_set_node(0, max_addr, &memblock.reserved, 454 NUMA_NO_NODE)); 455 /* In case that parsing SRAT failed. */ 456 WARN_ON(memblock_clear_hotplug(0, max_addr)); 457 numa_reset_distance(); 458 459 ret = init_func(); 460 if (ret < 0) 461 return ret; 462 463 /* 464 * We reset memblock back to the top-down direction 465 * here because if we configured ACPI_NUMA, we have 466 * parsed SRAT in init_func(). It is ok to have the 467 * reset here even if we didn't configure ACPI_NUMA 468 * or acpi numa init fails and fallbacks to dummy 469 * numa init. 470 */ 471 if (memblock_force_top_down) 472 memblock_set_bottom_up(false); 473 474 ret = numa_cleanup_meminfo(&numa_meminfo); 475 if (ret < 0) 476 return ret; 477 478 numa_emulation(&numa_meminfo, numa_distance_cnt); 479 480 return numa_register_meminfo(&numa_meminfo); 481 } 482 483 static int __init cmp_memblk(const void *a, const void *b) 484 { 485 const struct numa_memblk *ma = *(const struct numa_memblk **)a; 486 const struct numa_memblk *mb = *(const struct numa_memblk **)b; 487 488 return (ma->start > mb->start) - (ma->start < mb->start); 489 } 490 491 static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata; 492 493 /** 494 * numa_fill_memblks - Fill gaps in numa_meminfo memblks 495 * @start: address to begin fill 496 * @end: address to end fill 497 * 498 * Find and extend numa_meminfo memblks to cover the physical 499 * address range @start-@end 500 * 501 * RETURNS: 502 * 0 : Success 503 * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end 504 */ 505 506 int __init numa_fill_memblks(u64 start, u64 end) 507 { 508 struct numa_memblk **blk = &numa_memblk_list[0]; 509 struct numa_meminfo *mi = &numa_meminfo; 510 int count = 0; 511 u64 prev_end; 512 513 /* 514 * Create a list of pointers to numa_meminfo memblks that 515 * overlap start, end. The list is used to make in-place 516 * changes that fill out the numa_meminfo memblks. 517 */ 518 for (int i = 0; i < mi->nr_blks; i++) { 519 struct numa_memblk *bi = &mi->blk[i]; 520 521 if (memblock_addrs_overlap(start, end - start, bi->start, 522 bi->end - bi->start)) { 523 blk[count] = &mi->blk[i]; 524 count++; 525 } 526 } 527 if (!count) 528 return NUMA_NO_MEMBLK; 529 530 /* Sort the list of pointers in memblk->start order */ 531 sort(&blk[0], count, sizeof(blk[0]), cmp_memblk, NULL); 532 533 /* Make sure the first/last memblks include start/end */ 534 blk[0]->start = min(blk[0]->start, start); 535 blk[count - 1]->end = max(blk[count - 1]->end, end); 536 537 /* 538 * Fill any gaps by tracking the previous memblks 539 * end address and backfilling to it if needed. 540 */ 541 prev_end = blk[0]->end; 542 for (int i = 1; i < count; i++) { 543 struct numa_memblk *curr = blk[i]; 544 545 if (prev_end >= curr->start) { 546 if (prev_end < curr->end) 547 prev_end = curr->end; 548 } else { 549 curr->start = prev_end; 550 prev_end = curr->end; 551 } 552 } 553 return 0; 554 } 555 556 #ifdef CONFIG_NUMA_KEEP_MEMINFO 557 static int meminfo_to_nid(struct numa_meminfo *mi, u64 start) 558 { 559 int i; 560 561 for (i = 0; i < mi->nr_blks; i++) 562 if (mi->blk[i].start <= start && mi->blk[i].end > start) 563 return mi->blk[i].nid; 564 return NUMA_NO_NODE; 565 } 566 567 int phys_to_target_node(u64 start) 568 { 569 int nid = meminfo_to_nid(&numa_meminfo, start); 570 int reserved_nid = meminfo_to_nid(&numa_reserved_meminfo, start); 571 572 /* 573 * Prefer online nodes unless the address is also described 574 * by reserved ranges, in which case use the reserved nid. 575 */ 576 if (nid != NUMA_NO_NODE && reserved_nid == NUMA_NO_NODE) 577 return nid; 578 579 return reserved_nid; 580 } 581 EXPORT_SYMBOL_GPL(phys_to_target_node); 582 583 int memory_add_physaddr_to_nid(u64 start) 584 { 585 int nid = meminfo_to_nid(&numa_meminfo, start); 586 587 if (nid == NUMA_NO_NODE) 588 nid = numa_meminfo.blk[0].nid; 589 return nid; 590 } 591 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); 592 593 #endif /* CONFIG_NUMA_KEEP_MEMINFO */ 594