arch/ppc64/mm/numa.c

   1 /*
   2  * pSeries NUMA support
   3  *
   4  * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11 #include <linux/threads.h>
  12 #include <linux/bootmem.h>
  13 #include <linux/init.h>
  14 #include <linux/mm.h>
  15 #include <linux/mmzone.h>
  16 #include <linux/module.h>
  17 #include <asm/lmb.h>
  18 #include <asm/machdep.h>
  19 #include <asm/abs_addr.h>
  20
  21 #if 1
  22 #define dbg(args...) printk(KERN_INFO args)
  23 #else
  24 #define dbg(args...)
  25 #endif
  26
  27 #ifdef DEBUG_NUMA
  28 #define ARRAY_INITIALISER -1
  29 #else
  30 #define ARRAY_INITIALISER 0
  31 #endif
  32
  33 int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
  34         ARRAY_INITIALISER};
  35 char *numa_memory_lookup_table;
  36 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
  37 int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
  38
  39 struct pglist_data node_data[MAX_NUMNODES];
  40 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  41 static unsigned long node0_io_hole_size;
  42
  43 EXPORT_SYMBOL(node_data);
  44 EXPORT_SYMBOL(numa_cpu_lookup_table);
  45 EXPORT_SYMBOL(numa_memory_lookup_table);
  46 EXPORT_SYMBOL(numa_cpumask_lookup_table);
  47 EXPORT_SYMBOL(nr_cpus_in_node);
  48
  49 static inline void map_cpu_to_node(int cpu, int node)
  50 {
  51         dbg("cpu %d maps to domain %d\n", cpu, node);
  52         numa_cpu_lookup_table[cpu] = node;
  53         if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node]))) {
  54                 cpu_set(cpu, numa_cpumask_lookup_table[node]);
  55                 nr_cpus_in_node[node]++;
  56         }
  57 }
  58
  59 static struct device_node * __init find_cpu_node(unsigned int cpu)
  60 {
  61         unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
  62         struct device_node *cpu_node = NULL;
  63         unsigned int *interrupt_server, *reg;
  64         int len;
  65
  66         while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
  67                 /* Try interrupt server first */
  68                 interrupt_server = (unsigned int *)get_property(cpu_node,
  69                                         "ibm,ppc-interrupt-server#s", &len);
  70
  71                 if (interrupt_server && (len > 0)) {
  72                         while (len--) {
  73                                 if (interrupt_server[len-1] == hw_cpuid)
  74                                         return cpu_node;
  75                         }
  76                 } else {
  77                         reg = (unsigned int *)get_property(cpu_node,
  78                                                            "reg", &len);
  79                         if (reg && (len > 0) && (reg[0] == hw_cpuid))
  80                                 return cpu_node;
  81                 }
  82         }
  83
  84         return NULL;
  85 }
  86
  87 /* must hold reference to node during call */
  88 static int *of_get_associativity(struct device_node *dev)
  89  {
  90         unsigned int *result;
  91         int len;
  92
  93         result = (unsigned int *)get_property(dev, "ibm,associativity", &len);
  94
  95         if (len <= 0)
  96                 return NULL;
  97
  98         return result;
  99 }
 100
 101 static int of_node_numa_domain(struct device_node *device, int depth)
 102 {
 103         int numa_domain;
 104         unsigned int *tmp;
 105
 106         tmp = of_get_associativity(device);
 107         if (tmp && (tmp[0] >= depth)) {
 108                 numa_domain = tmp[depth];
 109         } else {
 110                 printk(KERN_ERR "WARNING: no NUMA information for "
 111                        "%s\n", device->full_name);
 112                 numa_domain = 0;
 113         }
 114         return numa_domain;
 115 }
 116
 117 /*
 118  * In theory, the "ibm,associativity" property may contain multiple
 119  * associativity lists because a resource may be multiply connected
 120  * into the machine.  This resource then has different associativity
 121  * characteristics relative to its multiple connections.  We ignore
 122  * this for now.  We also assume that all cpu and memory sets have
 123  * their distances represented at a common level.  This won't be
 124  * true for heirarchical NUMA.
 125  *
 126  * In any case the ibm,associativity-reference-points should give
 127  * the correct depth for a normal NUMA system.
 128  *
 129  * - Dave Hansen <haveblue@us.ibm.com>
 130  */
 131 static int find_min_common_depth(void)
 132 {
 133         int depth;
 134         unsigned int *ref_points;
 135         struct device_node *rtas_root;
 136         unsigned int len;
 137
 138         rtas_root = of_find_node_by_path("/rtas");
 139
 140         if (!rtas_root) {
 141                 printk(KERN_ERR "WARNING: %s() could not find rtas root\n",
 142                                 __FUNCTION__);
 143                 return -1;
 144         }
 145
 146         /*
 147          * this property is 2 32-bit integers, each representing a level of
 148          * depth in the associativity nodes.  The first is for an SMP
 149          * configuration (should be all 0's) and the second is for a normal
 150          * NUMA configuration.
 151          */
 152         ref_points = (unsigned int *)get_property(rtas_root,
 153                         "ibm,associativity-reference-points", &len);
 154
 155         if ((len >= 1) && ref_points) {
 156                 depth = ref_points[1];
 157         } else {
 158                 printk(KERN_ERR "WARNING: could not find NUMA "
 159                                 "associativity reference point\n");
 160                 depth = -1;
 161         }
 162         of_node_put(rtas_root);
 163
 164         return depth;
 165 }
 166
 167 static unsigned long read_cell_ul(struct device_node *device, unsigned int **buf)
 168 {
 169         int i;
 170         unsigned long result = 0;
 171
 172         i = prom_n_size_cells(device);
 173         /* bug on i>2 ?? */
 174         while (i--) {
 175                 result = (result << 32) | **buf;
 176                 (*buf)++;
 177         }
 178         return result;
 179 }
 180
 181 static int __init parse_numa_properties(void)
 182 {
 183         struct device_node *cpu = NULL;
 184         struct device_node *memory = NULL;
 185         int depth;
 186         int max_domain = 0;
 187         long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
 188         unsigned long i;
 189
 190         if (strstr(saved_command_line, "numa=off")) {
 191                 printk(KERN_WARNING "NUMA disabled by user\n");
 192                 return -1;
 193         }
 194
 195         numa_memory_lookup_table =
 196                 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
 197
 198         for (i = 0; i < entries ; i++)
 199                 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
 200
 201         depth = find_min_common_depth();
 202
 203         printk(KERN_INFO "NUMA associativity depth for CPU/Memory: %d\n", depth);
 204         if (depth < 0)
 205                 return depth;
 206
 207         for_each_cpu(i) {
 208                 int numa_domain;
 209
 210                 cpu = find_cpu_node(i);
 211
 212                 if (cpu) {
 213                         numa_domain = of_node_numa_domain(cpu, depth);
 214                         of_node_put(cpu);
 215
 216                         if (numa_domain >= MAX_NUMNODES) {
 217                                 /*
 218                                  * POWER4 LPAR uses 0xffff as invalid node,
 219                                  * dont warn in this case.
 220                                  */
 221                                 if (numa_domain != 0xffff)
 222                                         printk(KERN_ERR "WARNING: cpu %ld "
 223                                                "maps to invalid NUMA node %d\n",
 224                                                i, numa_domain);
 225                                 numa_domain = 0;
 226                         }
 227                 } else {
 228                         printk(KERN_ERR "WARNING: no NUMA information for "
 229                                "cpu %ld\n", i);
 230                         numa_domain = 0;
 231                 }
 232
 233                 node_set_online(numa_domain);
 234
 235                 if (max_domain < numa_domain)
 236                         max_domain = numa_domain;
 237
 238                 map_cpu_to_node(i, numa_domain);
 239         }
 240
 241         memory = NULL;
 242         while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
 243                 unsigned long start;
 244                 unsigned long size;
 245                 int numa_domain;
 246                 int ranges;
 247                 unsigned int *memcell_buf;
 248                 unsigned int len;
 249
 250                 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
 251                 if (!memcell_buf || len <= 0)
 252                         continue;
 253
 254                 ranges = memory->n_addrs;
 255 new_range:
 256                 /* these are order-sensitive, and modify the buffer pointer */
 257                 start = read_cell_ul(memory, &memcell_buf);
 258                 size = read_cell_ul(memory, &memcell_buf);
 259
 260                 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
 261                 size = _ALIGN_UP(size, MEMORY_INCREMENT);
 262
 263                 numa_domain = of_node_numa_domain(memory, depth);
 264
 265                 if (numa_domain >= MAX_NUMNODES) {
 266                         if (numa_domain != 0xffff)
 267                                 printk(KERN_ERR "WARNING: memory at %lx maps "
 268                                        "to invalid NUMA node %d\n", start,
 269                                        numa_domain);
 270                         numa_domain = 0;
 271                 }
 272
 273                 node_set_online(numa_domain);
 274
 275                 if (max_domain < numa_domain)
 276                         max_domain = numa_domain;
 277
 278                 /*
 279                  * For backwards compatibility, OF splits the first node
 280                  * into two regions (the first being 0-4GB). Check for
 281                  * this simple case and complain if there is a gap in
 282                  * memory
 283                  */
 284                 if (node_data[numa_domain].node_spanned_pages) {
 285                         unsigned long shouldstart =
 286                                 node_data[numa_domain].node_start_pfn +
 287                                 node_data[numa_domain].node_spanned_pages;
 288                         if (shouldstart != (start / PAGE_SIZE)) {
 289                                 printk(KERN_ERR "Hole in node, disabling "
 290                                                 "region start %lx length %lx\n",
 291                                                 start, size);
 292                                 continue;
 293                         }
 294                         node_data[numa_domain].node_spanned_pages +=
 295                                 size / PAGE_SIZE;
 296                 } else {
 297                         node_data[numa_domain].node_start_pfn =
 298                                 start / PAGE_SIZE;
 299                         node_data[numa_domain].node_spanned_pages =
 300                                 size / PAGE_SIZE;
 301                 }
 302
 303                 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
 304                         numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
 305                                 numa_domain;
 306
 307                 dbg("memory region %lx to %lx maps to domain %d\n",
 308                     start, start+size, numa_domain);
 309
 310                 ranges--;
 311                 if (ranges)
 312                         goto new_range;
 313         }
 314
 315         numnodes = max_domain + 1;
 316
 317         return 0;
 318 }
 319
 320 static void __init setup_nonnuma(void)
 321 {
 322         unsigned long top_of_ram = lmb_end_of_DRAM();
 323         unsigned long total_ram = lmb_phys_mem_size();
 324         unsigned long i;
 325
 326         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 327                top_of_ram, total_ram);
 328         printk(KERN_INFO "Memory hole size: %ldMB\n",
 329                (top_of_ram - total_ram) >> 20);
 330
 331         if (!numa_memory_lookup_table) {
 332                 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
 333                 numa_memory_lookup_table =
 334                         (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
 335                 for (i = 0; i < entries ; i++)
 336                         numa_memory_lookup_table[i] = ARRAY_INITIALISER;
 337         }
 338
 339         for (i = 0; i < NR_CPUS; i++)
 340                 map_cpu_to_node(i, 0);
 341
 342         node_set_online(0);
 343
 344         node_data[0].node_start_pfn = 0;
 345         node_data[0].node_spanned_pages = lmb_end_of_DRAM() / PAGE_SIZE;
 346
 347         for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
 348                 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
 349
 350         node0_io_hole_size = top_of_ram - total_ram;
 351 }
 352
 353 void __init do_init_bootmem(void)
 354 {
 355         int nid;
 356
 357         min_low_pfn = 0;
 358         max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
 359
 360         if (parse_numa_properties())
 361                 setup_nonnuma();
 362
 363         for (nid = 0; nid < numnodes; nid++) {
 364                 unsigned long start_paddr, end_paddr;
 365                 int i;
 366                 unsigned long bootmem_paddr;
 367                 unsigned long bootmap_pages;
 368
 369                 if (node_data[nid].node_spanned_pages == 0)
 370                         continue;
 371
 372                 start_paddr = node_data[nid].node_start_pfn * PAGE_SIZE;
 373                 end_paddr = start_paddr +
 374                                 (node_data[nid].node_spanned_pages * PAGE_SIZE);
 375
 376                 dbg("node %d\n", nid);
 377                 dbg("start_paddr = %lx\n", start_paddr);
 378                 dbg("end_paddr = %lx\n", end_paddr);
 379
 380                 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
 381
 382                 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
 383                 dbg("bootmap_pages = %lx\n", bootmap_pages);
 384
 385                 bootmem_paddr = lmb_alloc_base(bootmap_pages << PAGE_SHIFT,
 386                                 PAGE_SIZE, end_paddr);
 387                 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
 388
 389                 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
 390                                   start_paddr >> PAGE_SHIFT,
 391                                   end_paddr >> PAGE_SHIFT);
 392
 393                 for (i = 0; i < lmb.memory.cnt; i++) {
 394                         unsigned long physbase, size;
 395
 396                         physbase = lmb.memory.region[i].physbase;
 397                         size = lmb.memory.region[i].size;
 398
 399                         if (physbase < end_paddr &&
 400                             (physbase+size) > start_paddr) {
 401                                 /* overlaps */
 402                                 if (physbase < start_paddr) {
 403                                         size -= start_paddr - physbase;
 404                                         physbase = start_paddr;
 405                                 }
 406
 407                                 if (size > end_paddr - physbase)
 408                                         size = end_paddr - physbase;
 409
 410                                 dbg("free_bootmem %lx %lx\n", physbase, size);
 411                                 free_bootmem_node(NODE_DATA(nid), physbase,
 412                                                   size);
 413                         }
 414                 }
 415
 416                 for (i = 0; i < lmb.reserved.cnt; i++) {
 417                         unsigned long physbase = lmb.reserved.region[i].physbase;
 418                         unsigned long size = lmb.reserved.region[i].size;
 419
 420                         if (physbase < end_paddr &&
 421                             (physbase+size) > start_paddr) {
 422                                 /* overlaps */
 423                                 if (physbase < start_paddr) {
 424                                         size -= start_paddr - physbase;
 425                                         physbase = start_paddr;
 426                                 }
 427
 428                                 if (size > end_paddr - physbase)
 429                                         size = end_paddr - physbase;
 430
 431                                 dbg("reserve_bootmem %lx %lx\n", physbase,
 432                                     size);
 433                                 reserve_bootmem_node(NODE_DATA(nid), physbase,
 434                                                      size);
 435                         }
 436                 }
 437         }
 438 }
 439
 440 void __init paging_init(void)
 441 {
 442         unsigned long zones_size[MAX_NR_ZONES];
 443         unsigned long zholes_size[MAX_NR_ZONES];
 444         struct page *node_mem_map;
 445         int nid;
 446
 447         memset(zones_size, 0, sizeof(zones_size));
 448         memset(zholes_size, 0, sizeof(zholes_size));
 449
 450         for (nid = 0; nid < numnodes; nid++) {
 451                 unsigned long start_pfn;
 452                 unsigned long end_pfn;
 453
 454                 start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
 455                 end_pfn = plat_node_bdata[nid].node_low_pfn;
 456
 457                 zones_size[ZONE_DMA] = end_pfn - start_pfn;
 458                 zholes_size[ZONE_DMA] = 0;
 459                 if (nid == 0)
 460                         zholes_size[ZONE_DMA] = node0_io_hole_size;
 461
 462                 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
 463                     zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
 464
 465                 /*
 466                  * Give this empty node a dummy struct page to avoid
 467                  * us from trying to allocate a node local mem_map
 468                  * in free_area_init_node (which will fail).
 469                  */
 470                 if (!node_data[nid].node_spanned_pages)
 471                         node_mem_map = alloc_bootmem(sizeof(struct page));
 472                 else
 473                         node_mem_map = NULL;
 474
 475                 free_area_init_node(nid, NODE_DATA(nid), node_mem_map,
 476                                     zones_size, start_pfn, zholes_size);
 477         }
 478 }