1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2000, 2003 Silicon Graphics, Inc. All rights reserved. 4 * Copyright (c) 2001 Intel Corp. 5 * Copyright (c) 2001 Tony Luck <tony.luck@intel.com> 6 * Copyright (c) 2002 NEC Corp. 7 * Copyright (c) 2002 Kimio Suganuma <k-suganuma@da.jp.nec.com> 8 * Copyright (c) 2004 Silicon Graphics, Inc 9 * Russ Anderson <rja@sgi.com> 10 * Jesse Barnes <jbarnes@sgi.com> 11 * Jack Steiner <steiner@sgi.com> 12 */ 13 14 /* 15 * Platform initialization for Discontig Memory 16 */ 17 18 #include <linux/kernel.h> 19 #include <linux/mm.h> 20 #include <linux/nmi.h> 21 #include <linux/swap.h> 22 #include <linux/memblock.h> 23 #include <linux/acpi.h> 24 #include <linux/efi.h> 25 #include <linux/nodemask.h> 26 #include <linux/slab.h> 27 #include <asm/pgalloc.h> 28 #include <asm/tlb.h> 29 #include <asm/meminit.h> 30 #include <asm/numa.h> 31 #include <asm/sections.h> 32 33 /* 34 * Track per-node information needed to setup the boot memory allocator, the 35 * per-node areas, and the real VM. 36 */ 37 struct early_node_data { 38 struct ia64_node_data *node_data; 39 unsigned long pernode_addr; 40 unsigned long pernode_size; 41 unsigned long min_pfn; 42 unsigned long max_pfn; 43 }; 44 45 static struct early_node_data mem_data[MAX_NUMNODES] __initdata; 46 static nodemask_t memory_less_mask __initdata; 47 48 pg_data_t *pgdat_list[MAX_NUMNODES]; 49 50 /* 51 * To prevent cache aliasing effects, align per-node structures so that they 52 * start at addresses that are strided by node number. 53 */ 54 #define MAX_NODE_ALIGN_OFFSET (32 * 1024 * 1024) 55 #define NODEDATA_ALIGN(addr, node) \ 56 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + \ 57 (((node)*PERCPU_PAGE_SIZE) & (MAX_NODE_ALIGN_OFFSET - 1))) 58 59 /** 60 * build_node_maps - callback to setup mem_data structs for each node 61 * @start: physical start of range 62 * @len: length of range 63 * @node: node where this range resides 64 * 65 * Detect extents of each piece of memory that we wish to 66 * treat as a virtually contiguous block (i.e. each node). Each such block 67 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down 68 * if necessary. Any non-existent pages will simply be part of the virtual 69 * memmap. 70 */ 71 static int __init build_node_maps(unsigned long start, unsigned long len, 72 int node) 73 { 74 unsigned long spfn, epfn, end = start + len; 75 76 epfn = GRANULEROUNDUP(end) >> PAGE_SHIFT; 77 spfn = GRANULEROUNDDOWN(start) >> PAGE_SHIFT; 78 79 if (!mem_data[node].min_pfn) { 80 mem_data[node].min_pfn = spfn; 81 mem_data[node].max_pfn = epfn; 82 } else { 83 mem_data[node].min_pfn = min(spfn, mem_data[node].min_pfn); 84 mem_data[node].max_pfn = max(epfn, mem_data[node].max_pfn); 85 } 86 87 return 0; 88 } 89 90 /** 91 * early_nr_cpus_node - return number of cpus on a given node 92 * @node: node to check 93 * 94 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because 95 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been 96 * called yet. Note that node 0 will also count all non-existent cpus. 97 */ 98 static int __meminit early_nr_cpus_node(int node) 99 { 100 int cpu, n = 0; 101 102 for_each_possible_early_cpu(cpu) 103 if (node == node_cpuid[cpu].nid) 104 n++; 105 106 return n; 107 } 108 109 /** 110 * compute_pernodesize - compute size of pernode data 111 * @node: the node id. 112 */ 113 static unsigned long __meminit compute_pernodesize(int node) 114 { 115 unsigned long pernodesize = 0, cpus; 116 117 cpus = early_nr_cpus_node(node); 118 pernodesize += PERCPU_PAGE_SIZE * cpus; 119 pernodesize += node * L1_CACHE_BYTES; 120 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t)); 121 pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data)); 122 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t)); 123 pernodesize = PAGE_ALIGN(pernodesize); 124 return pernodesize; 125 } 126 127 /** 128 * per_cpu_node_setup - setup per-cpu areas on each node 129 * @cpu_data: per-cpu area on this node 130 * @node: node to setup 131 * 132 * Copy the static per-cpu data into the region we just set aside and then 133 * setup __per_cpu_offset for each CPU on this node. Return a pointer to 134 * the end of the area. 135 */ 136 static void *per_cpu_node_setup(void *cpu_data, int node) 137 { 138 #ifdef CONFIG_SMP 139 int cpu; 140 141 for_each_possible_early_cpu(cpu) { 142 void *src = cpu == 0 ? __cpu0_per_cpu : __phys_per_cpu_start; 143 144 if (node != node_cpuid[cpu].nid) 145 continue; 146 147 memcpy(__va(cpu_data), src, __per_cpu_end - __per_cpu_start); 148 __per_cpu_offset[cpu] = (char *)__va(cpu_data) - 149 __per_cpu_start; 150 151 /* 152 * percpu area for cpu0 is moved from the __init area 153 * which is setup by head.S and used till this point. 154 * Update ar.k3. This move is ensures that percpu 155 * area for cpu0 is on the correct node and its 156 * virtual address isn't insanely far from other 157 * percpu areas which is important for congruent 158 * percpu allocator. 159 */ 160 if (cpu == 0) 161 ia64_set_kr(IA64_KR_PER_CPU_DATA, 162 (unsigned long)cpu_data - 163 (unsigned long)__per_cpu_start); 164 165 cpu_data += PERCPU_PAGE_SIZE; 166 } 167 #endif 168 return cpu_data; 169 } 170 171 #ifdef CONFIG_SMP 172 /** 173 * setup_per_cpu_areas - setup percpu areas 174 * 175 * Arch code has already allocated and initialized percpu areas. All 176 * this function has to do is to teach the determined layout to the 177 * dynamic percpu allocator, which happens to be more complex than 178 * creating whole new ones using helpers. 179 */ 180 void __init setup_per_cpu_areas(void) 181 { 182 struct pcpu_alloc_info *ai; 183 struct pcpu_group_info *uninitialized_var(gi); 184 unsigned int *cpu_map; 185 void *base; 186 unsigned long base_offset; 187 unsigned int cpu; 188 ssize_t static_size, reserved_size, dyn_size; 189 int node, prev_node, unit, nr_units, rc; 190 191 ai = pcpu_alloc_alloc_info(MAX_NUMNODES, nr_cpu_ids); 192 if (!ai) 193 panic("failed to allocate pcpu_alloc_info"); 194 cpu_map = ai->groups[0].cpu_map; 195 196 /* determine base */ 197 base = (void *)ULONG_MAX; 198 for_each_possible_cpu(cpu) 199 base = min(base, 200 (void *)(__per_cpu_offset[cpu] + __per_cpu_start)); 201 base_offset = (void *)__per_cpu_start - base; 202 203 /* build cpu_map, units are grouped by node */ 204 unit = 0; 205 for_each_node(node) 206 for_each_possible_cpu(cpu) 207 if (node == node_cpuid[cpu].nid) 208 cpu_map[unit++] = cpu; 209 nr_units = unit; 210 211 /* set basic parameters */ 212 static_size = __per_cpu_end - __per_cpu_start; 213 reserved_size = PERCPU_MODULE_RESERVE; 214 dyn_size = PERCPU_PAGE_SIZE - static_size - reserved_size; 215 if (dyn_size < 0) 216 panic("percpu area overflow static=%zd reserved=%zd\n", 217 static_size, reserved_size); 218 219 ai->static_size = static_size; 220 ai->reserved_size = reserved_size; 221 ai->dyn_size = dyn_size; 222 ai->unit_size = PERCPU_PAGE_SIZE; 223 ai->atom_size = PAGE_SIZE; 224 ai->alloc_size = PERCPU_PAGE_SIZE; 225 226 /* 227 * CPUs are put into groups according to node. Walk cpu_map 228 * and create new groups at node boundaries. 229 */ 230 prev_node = NUMA_NO_NODE; 231 ai->nr_groups = 0; 232 for (unit = 0; unit < nr_units; unit++) { 233 cpu = cpu_map[unit]; 234 node = node_cpuid[cpu].nid; 235 236 if (node == prev_node) { 237 gi->nr_units++; 238 continue; 239 } 240 prev_node = node; 241 242 gi = &ai->groups[ai->nr_groups++]; 243 gi->nr_units = 1; 244 gi->base_offset = __per_cpu_offset[cpu] + base_offset; 245 gi->cpu_map = &cpu_map[unit]; 246 } 247 248 rc = pcpu_setup_first_chunk(ai, base); 249 if (rc) 250 panic("failed to setup percpu area (err=%d)", rc); 251 252 pcpu_free_alloc_info(ai); 253 } 254 #endif 255 256 /** 257 * fill_pernode - initialize pernode data. 258 * @node: the node id. 259 * @pernode: physical address of pernode data 260 * @pernodesize: size of the pernode data 261 */ 262 static void __init fill_pernode(int node, unsigned long pernode, 263 unsigned long pernodesize) 264 { 265 void *cpu_data; 266 int cpus = early_nr_cpus_node(node); 267 268 mem_data[node].pernode_addr = pernode; 269 mem_data[node].pernode_size = pernodesize; 270 memset(__va(pernode), 0, pernodesize); 271 272 cpu_data = (void *)pernode; 273 pernode += PERCPU_PAGE_SIZE * cpus; 274 pernode += node * L1_CACHE_BYTES; 275 276 pgdat_list[node] = __va(pernode); 277 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t)); 278 279 mem_data[node].node_data = __va(pernode); 280 pernode += L1_CACHE_ALIGN(sizeof(struct ia64_node_data)); 281 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t)); 282 283 cpu_data = per_cpu_node_setup(cpu_data, node); 284 285 return; 286 } 287 288 /** 289 * find_pernode_space - allocate memory for memory map and per-node structures 290 * @start: physical start of range 291 * @len: length of range 292 * @node: node where this range resides 293 * 294 * This routine reserves space for the per-cpu data struct, the list of 295 * pg_data_ts and the per-node data struct. Each node will have something like 296 * the following in the first chunk of addr. space large enough to hold it. 297 * 298 * ________________________ 299 * | | 300 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first 301 * | PERCPU_PAGE_SIZE * | start and length big enough 302 * | cpus_on_this_node | Node 0 will also have entries for all non-existent cpus. 303 * |------------------------| 304 * | local pg_data_t * | 305 * |------------------------| 306 * | local ia64_node_data | 307 * |------------------------| 308 * | ??? | 309 * |________________________| 310 * 311 * Once this space has been set aside, the bootmem maps are initialized. We 312 * could probably move the allocation of the per-cpu and ia64_node_data space 313 * outside of this function and use alloc_bootmem_node(), but doing it here 314 * is straightforward and we get the alignments we want so... 315 */ 316 static int __init find_pernode_space(unsigned long start, unsigned long len, 317 int node) 318 { 319 unsigned long spfn, epfn; 320 unsigned long pernodesize = 0, pernode; 321 322 spfn = start >> PAGE_SHIFT; 323 epfn = (start + len) >> PAGE_SHIFT; 324 325 /* 326 * Make sure this memory falls within this node's usable memory 327 * since we may have thrown some away in build_maps(). 328 */ 329 if (spfn < mem_data[node].min_pfn || epfn > mem_data[node].max_pfn) 330 return 0; 331 332 /* Don't setup this node's local space twice... */ 333 if (mem_data[node].pernode_addr) 334 return 0; 335 336 /* 337 * Calculate total size needed, incl. what's necessary 338 * for good alignment and alias prevention. 339 */ 340 pernodesize = compute_pernodesize(node); 341 pernode = NODEDATA_ALIGN(start, node); 342 343 /* Is this range big enough for what we want to store here? */ 344 if (start + len > (pernode + pernodesize)) 345 fill_pernode(node, pernode, pernodesize); 346 347 return 0; 348 } 349 350 /** 351 * reserve_pernode_space - reserve memory for per-node space 352 * 353 * Reserve the space used by the bootmem maps & per-node space in the boot 354 * allocator so that when we actually create the real mem maps we don't 355 * use their memory. 356 */ 357 static void __init reserve_pernode_space(void) 358 { 359 unsigned long base, size; 360 int node; 361 362 for_each_online_node(node) { 363 if (node_isset(node, memory_less_mask)) 364 continue; 365 366 /* Now the per-node space */ 367 size = mem_data[node].pernode_size; 368 base = __pa(mem_data[node].pernode_addr); 369 memblock_reserve(base, size); 370 } 371 } 372 373 static void __meminit scatter_node_data(void) 374 { 375 pg_data_t **dst; 376 int node; 377 378 /* 379 * for_each_online_node() can't be used at here. 380 * node_online_map is not set for hot-added nodes at this time, 381 * because we are halfway through initialization of the new node's 382 * structures. If for_each_online_node() is used, a new node's 383 * pg_data_ptrs will be not initialized. Instead of using it, 384 * pgdat_list[] is checked. 385 */ 386 for_each_node(node) { 387 if (pgdat_list[node]) { 388 dst = LOCAL_DATA_ADDR(pgdat_list[node])->pg_data_ptrs; 389 memcpy(dst, pgdat_list, sizeof(pgdat_list)); 390 } 391 } 392 } 393 394 /** 395 * initialize_pernode_data - fixup per-cpu & per-node pointers 396 * 397 * Each node's per-node area has a copy of the global pg_data_t list, so 398 * we copy that to each node here, as well as setting the per-cpu pointer 399 * to the local node data structure. 400 */ 401 static void __init initialize_pernode_data(void) 402 { 403 int cpu, node; 404 405 scatter_node_data(); 406 407 #ifdef CONFIG_SMP 408 /* Set the node_data pointer for each per-cpu struct */ 409 for_each_possible_early_cpu(cpu) { 410 node = node_cpuid[cpu].nid; 411 per_cpu(ia64_cpu_info, cpu).node_data = 412 mem_data[node].node_data; 413 } 414 #else 415 { 416 struct cpuinfo_ia64 *cpu0_cpu_info; 417 cpu = 0; 418 node = node_cpuid[cpu].nid; 419 cpu0_cpu_info = (struct cpuinfo_ia64 *)(__phys_per_cpu_start + 420 ((char *)&ia64_cpu_info - __per_cpu_start)); 421 cpu0_cpu_info->node_data = mem_data[node].node_data; 422 } 423 #endif /* CONFIG_SMP */ 424 } 425 426 /** 427 * memory_less_node_alloc - * attempt to allocate memory on the best NUMA slit 428 * node but fall back to any other node when __alloc_bootmem_node fails 429 * for best. 430 * @nid: node id 431 * @pernodesize: size of this node's pernode data 432 */ 433 static void __init *memory_less_node_alloc(int nid, unsigned long pernodesize) 434 { 435 void *ptr = NULL; 436 u8 best = 0xff; 437 int bestnode = NUMA_NO_NODE, node, anynode = 0; 438 439 for_each_online_node(node) { 440 if (node_isset(node, memory_less_mask)) 441 continue; 442 else if (node_distance(nid, node) < best) { 443 best = node_distance(nid, node); 444 bestnode = node; 445 } 446 anynode = node; 447 } 448 449 if (bestnode == NUMA_NO_NODE) 450 bestnode = anynode; 451 452 ptr = memblock_alloc_try_nid(pernodesize, PERCPU_PAGE_SIZE, 453 __pa(MAX_DMA_ADDRESS), 454 MEMBLOCK_ALLOC_ACCESSIBLE, 455 bestnode); 456 if (!ptr) 457 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%lx\n", 458 __func__, pernodesize, PERCPU_PAGE_SIZE, bestnode, 459 __pa(MAX_DMA_ADDRESS)); 460 461 return ptr; 462 } 463 464 /** 465 * memory_less_nodes - allocate and initialize CPU only nodes pernode 466 * information. 467 */ 468 static void __init memory_less_nodes(void) 469 { 470 unsigned long pernodesize; 471 void *pernode; 472 int node; 473 474 for_each_node_mask(node, memory_less_mask) { 475 pernodesize = compute_pernodesize(node); 476 pernode = memory_less_node_alloc(node, pernodesize); 477 fill_pernode(node, __pa(pernode), pernodesize); 478 } 479 480 return; 481 } 482 483 /** 484 * find_memory - walk the EFI memory map and setup the bootmem allocator 485 * 486 * Called early in boot to setup the bootmem allocator, and to 487 * allocate the per-cpu and per-node structures. 488 */ 489 void __init find_memory(void) 490 { 491 int node; 492 493 reserve_memory(); 494 efi_memmap_walk(filter_memory, register_active_ranges); 495 496 if (num_online_nodes() == 0) { 497 printk(KERN_ERR "node info missing!\n"); 498 node_set_online(0); 499 } 500 501 nodes_or(memory_less_mask, memory_less_mask, node_online_map); 502 min_low_pfn = -1; 503 max_low_pfn = 0; 504 505 /* These actually end up getting called by call_pernode_memory() */ 506 efi_memmap_walk(filter_rsvd_memory, build_node_maps); 507 efi_memmap_walk(filter_rsvd_memory, find_pernode_space); 508 efi_memmap_walk(find_max_min_low_pfn, NULL); 509 510 for_each_online_node(node) 511 if (mem_data[node].min_pfn) 512 node_clear(node, memory_less_mask); 513 514 reserve_pernode_space(); 515 memory_less_nodes(); 516 initialize_pernode_data(); 517 518 max_pfn = max_low_pfn; 519 520 find_initrd(); 521 } 522 523 #ifdef CONFIG_SMP 524 /** 525 * per_cpu_init - setup per-cpu variables 526 * 527 * find_pernode_space() does most of this already, we just need to set 528 * local_per_cpu_offset 529 */ 530 void *per_cpu_init(void) 531 { 532 int cpu; 533 static int first_time = 1; 534 535 if (first_time) { 536 first_time = 0; 537 for_each_possible_early_cpu(cpu) 538 per_cpu(local_per_cpu_offset, cpu) = __per_cpu_offset[cpu]; 539 } 540 541 return __per_cpu_start + __per_cpu_offset[smp_processor_id()]; 542 } 543 #endif /* CONFIG_SMP */ 544 545 /** 546 * call_pernode_memory - use SRAT to call callback functions with node info 547 * @start: physical start of range 548 * @len: length of range 549 * @arg: function to call for each range 550 * 551 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find 552 * out to which node a block of memory belongs. Ignore memory that we cannot 553 * identify, and split blocks that run across multiple nodes. 554 * 555 * Take this opportunity to round the start address up and the end address 556 * down to page boundaries. 557 */ 558 void call_pernode_memory(unsigned long start, unsigned long len, void *arg) 559 { 560 unsigned long rs, re, end = start + len; 561 void (*func)(unsigned long, unsigned long, int); 562 int i; 563 564 start = PAGE_ALIGN(start); 565 end &= PAGE_MASK; 566 if (start >= end) 567 return; 568 569 func = arg; 570 571 if (!num_node_memblks) { 572 /* No SRAT table, so assume one node (node 0) */ 573 if (start < end) 574 (*func)(start, end - start, 0); 575 return; 576 } 577 578 for (i = 0; i < num_node_memblks; i++) { 579 rs = max(start, node_memblk[i].start_paddr); 580 re = min(end, node_memblk[i].start_paddr + 581 node_memblk[i].size); 582 583 if (rs < re) 584 (*func)(rs, re - rs, node_memblk[i].nid); 585 586 if (re == end) 587 break; 588 } 589 } 590 591 /** 592 * paging_init - setup page tables 593 * 594 * paging_init() sets up the page tables for each node of the system and frees 595 * the bootmem allocator memory for general use. 596 */ 597 void __init paging_init(void) 598 { 599 unsigned long max_dma; 600 unsigned long pfn_offset = 0; 601 unsigned long max_pfn = 0; 602 int node; 603 unsigned long max_zone_pfns[MAX_NR_ZONES]; 604 605 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT; 606 607 sparse_memory_present_with_active_regions(MAX_NUMNODES); 608 sparse_init(); 609 610 #ifdef CONFIG_VIRTUAL_MEM_MAP 611 VMALLOC_END -= PAGE_ALIGN(ALIGN(max_low_pfn, MAX_ORDER_NR_PAGES) * 612 sizeof(struct page)); 613 vmem_map = (struct page *) VMALLOC_END; 614 efi_memmap_walk(create_mem_map_page_table, NULL); 615 printk("Virtual mem_map starts at 0x%p\n", vmem_map); 616 #endif 617 618 for_each_online_node(node) { 619 pfn_offset = mem_data[node].min_pfn; 620 621 #ifdef CONFIG_VIRTUAL_MEM_MAP 622 NODE_DATA(node)->node_mem_map = vmem_map + pfn_offset; 623 #endif 624 if (mem_data[node].max_pfn > max_pfn) 625 max_pfn = mem_data[node].max_pfn; 626 } 627 628 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 629 #ifdef CONFIG_ZONE_DMA32 630 max_zone_pfns[ZONE_DMA32] = max_dma; 631 #endif 632 max_zone_pfns[ZONE_NORMAL] = max_pfn; 633 free_area_init_nodes(max_zone_pfns); 634 635 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page)); 636 } 637 638 #ifdef CONFIG_MEMORY_HOTPLUG 639 pg_data_t *arch_alloc_nodedata(int nid) 640 { 641 unsigned long size = compute_pernodesize(nid); 642 643 return kzalloc(size, GFP_KERNEL); 644 } 645 646 void arch_free_nodedata(pg_data_t *pgdat) 647 { 648 kfree(pgdat); 649 } 650 651 void arch_refresh_nodedata(int update_node, pg_data_t *update_pgdat) 652 { 653 pgdat_list[update_node] = update_pgdat; 654 scatter_node_data(); 655 } 656 #endif 657 658 #ifdef CONFIG_SPARSEMEM_VMEMMAP 659 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 660 struct vmem_altmap *altmap) 661 { 662 return vmemmap_populate_basepages(start, end, node); 663 } 664 665 void vmemmap_free(unsigned long start, unsigned long end, 666 struct vmem_altmap *altmap) 667 { 668 } 669 #endif 670