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...) udbg_printf(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 int __init parse_numa_properties(void)
  60 {
  61         struct device_node *cpu = NULL;
  62         struct device_node *memory = NULL;
  63         int *cpu_associativity;
  64         int *memory_associativity;
  65         int depth;
  66         int max_domain = 0;
  67         long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
  68         long i;
  69
  70         if (strstr(saved_command_line, "numa=off")) {
  71                 printk(KERN_WARNING "NUMA disabled by user\n");
  72                 return -1;
  73         }
  74
  75         numa_memory_lookup_table =
  76                 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
  77
  78         for (i = 0; i < entries ; i++)
  79                 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
  80
  81         cpu = of_find_node_by_type(NULL, "cpu");
  82         if (!cpu)
  83                 goto err;
  84
  85         memory = of_find_node_by_type(NULL, "memory");
  86         if (!memory)
  87                 goto err;
  88
  89         cpu_associativity = (int *)get_property(cpu, "ibm,associativity", NULL);
  90         if (!cpu_associativity)
  91                 goto err;
  92
  93         memory_associativity = (int *)get_property(memory, "ibm,associativity",
  94                                                    NULL);
  95         if (!memory_associativity)
  96                 goto err;
  97
  98         /* find common depth */
  99         if (cpu_associativity[0] < memory_associativity[0])
 100                 depth = cpu_associativity[0];
 101         else
 102                 depth = memory_associativity[0];
 103
 104         for (; cpu; cpu = of_find_node_by_type(cpu, "cpu")) {
 105                 int *tmp;
 106                 int cpu_nr, numa_domain;
 107
 108                 tmp = (int *)get_property(cpu, "reg", NULL);
 109                 if (!tmp)
 110                         continue;
 111                 cpu_nr = *tmp;
 112
 113                 tmp = (int *)get_property(cpu, "ibm,associativity",
 114                                           NULL);
 115                 if (!tmp)
 116                         continue;
 117                 numa_domain = tmp[depth];
 118
 119                 /* FIXME */
 120                 if (numa_domain == 0xffff) {
 121                         dbg("cpu %d has no numa doman\n", cpu_nr);
 122                         numa_domain = 0;
 123                 }
 124
 125                 if (numa_domain >= MAX_NUMNODES)
 126                         BUG();
 127
 128                 node_set_online(numa_domain);
 129
 130                 if (max_domain < numa_domain)
 131                         max_domain = numa_domain;
 132
 133                 map_cpu_to_node(cpu_nr, numa_domain);
 134                 /* register the second thread on an SMT machine */
 135                 if (cur_cpu_spec->cpu_features & CPU_FTR_SMT)
 136                         map_cpu_to_node(cpu_nr ^ 0x1, numa_domain);
 137         }
 138
 139         for (; memory; memory = of_find_node_by_type(memory, "memory")) {
 140                 unsigned int *tmp1, *tmp2;
 141                 unsigned long i;
 142                 unsigned long start = 0;
 143                 unsigned long size = 0;
 144                 int numa_domain;
 145                 int ranges;
 146
 147                 tmp1 = (int *)get_property(memory, "reg", NULL);
 148                 if (!tmp1)
 149                         continue;
 150
 151                 ranges = memory->n_addrs;
 152 new_range:
 153
 154                 i = prom_n_size_cells(memory);
 155                 while (i--) {
 156                         start = (start << 32) | *tmp1;
 157                         tmp1++;
 158                 }
 159
 160                 i = prom_n_size_cells(memory);
 161                 while (i--) {
 162                         size = (size << 32) | *tmp1;
 163                         tmp1++;
 164                 }
 165
 166                 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
 167                 size = _ALIGN_UP(size, MEMORY_INCREMENT);
 168
 169                 if ((start + size) > MAX_MEMORY)
 170                         BUG();
 171
 172                 tmp2 = (int *)get_property(memory, "ibm,associativity",
 173                                            NULL);
 174                 if (!tmp2)
 175                         continue;
 176                 numa_domain = tmp2[depth];
 177
 178                 /* FIXME */
 179                 if (numa_domain == 0xffff) {
 180                         dbg("memory has no numa doman\n");
 181                         numa_domain = 0;
 182                 }
 183
 184                 if (numa_domain >= MAX_NUMNODES)
 185                         BUG();
 186
 187                 node_set_online(numa_domain);
 188
 189                 if (max_domain < numa_domain)
 190                         max_domain = numa_domain;
 191
 192                 /*
 193                  * For backwards compatibility, OF splits the first node
 194                  * into two regions (the first being 0-4GB). Check for
 195                  * this simple case and complain if there is a gap in
 196                  * memory
 197                  */
 198                 if (node_data[numa_domain].node_spanned_pages) {
 199                         unsigned long shouldstart =
 200                                 node_data[numa_domain].node_start_pfn +
 201                                 node_data[numa_domain].node_spanned_pages;
 202                         if (shouldstart != (start / PAGE_SIZE)) {
 203                                 printk(KERN_ERR "Hole in node, disabling "
 204                                                 "region start %lx length %lx\n",
 205                                                 start, size);
 206                                 continue;
 207                         }
 208                         node_data[numa_domain].node_spanned_pages += size / PAGE_SIZE;
 209                 } else {
 210                         node_data[numa_domain].node_start_pfn =
 211                                 start / PAGE_SIZE;
 212                         node_data[numa_domain].node_spanned_pages = size / PAGE_SIZE;
 213                 }
 214
 215                 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
 216                         numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
 217                                 numa_domain;
 218
 219                 dbg("memory region %lx to %lx maps to domain %d\n",
 220                     start, start+size, numa_domain);
 221
 222                 ranges--;
 223                 if (ranges)
 224                         goto new_range;
 225         }
 226
 227         numnodes = max_domain + 1;
 228
 229         return 0;
 230 err:
 231         of_node_put(cpu);
 232         of_node_put(memory);
 233         return -1;
 234 }
 235
 236 static void __init setup_nonnuma(void)
 237 {
 238         unsigned long top_of_ram = lmb_end_of_DRAM();
 239         unsigned long total_ram = lmb_phys_mem_size();
 240         unsigned long i;
 241
 242         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 243                top_of_ram, total_ram);
 244         printk(KERN_INFO "Memory hole size: %ldMB\n",
 245                (top_of_ram - total_ram) >> 20);
 246
 247         if (!numa_memory_lookup_table) {
 248                 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
 249                 numa_memory_lookup_table =
 250                         (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
 251                 for (i = 0; i < entries ; i++)
 252                         numa_memory_lookup_table[i] = ARRAY_INITIALISER;
 253         }
 254
 255         for (i = 0; i < NR_CPUS; i++)
 256                 map_cpu_to_node(i, 0);
 257
 258         node_set_online(0);
 259
 260         node_data[0].node_start_pfn = 0;
 261         node_data[0].node_spanned_pages = lmb_end_of_DRAM() / PAGE_SIZE;
 262
 263         for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
 264                 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
 265
 266         node0_io_hole_size = top_of_ram - total_ram;
 267 }
 268
 269 void __init do_init_bootmem(void)
 270 {
 271         int nid;
 272
 273         min_low_pfn = 0;
 274         max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
 275
 276         if (parse_numa_properties())
 277                 setup_nonnuma();
 278
 279         for (nid = 0; nid < numnodes; nid++) {
 280                 unsigned long start_paddr, end_paddr;
 281                 int i;
 282                 unsigned long bootmem_paddr;
 283                 unsigned long bootmap_pages;
 284
 285                 if (node_data[nid].node_spanned_pages == 0)
 286                         continue;
 287
 288                 start_paddr = node_data[nid].node_start_pfn * PAGE_SIZE;
 289                 end_paddr = start_paddr +
 290                                 (node_data[nid].node_spanned_pages * PAGE_SIZE);
 291
 292                 dbg("node %d\n", nid);
 293                 dbg("start_paddr = %lx\n", start_paddr);
 294                 dbg("end_paddr = %lx\n", end_paddr);
 295
 296                 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
 297
 298                 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
 299                 dbg("bootmap_pages = %lx\n", bootmap_pages);
 300
 301                 bootmem_paddr = lmb_alloc_base(bootmap_pages << PAGE_SHIFT,
 302                                 PAGE_SIZE, end_paddr);
 303                 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
 304
 305                 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
 306                                   start_paddr >> PAGE_SHIFT,
 307                                   end_paddr >> PAGE_SHIFT);
 308
 309                 for (i = 0; i < lmb.memory.cnt; i++) {
 310                         unsigned long physbase, size;
 311
 312                         physbase = lmb.memory.region[i].physbase;
 313                         size = lmb.memory.region[i].size;
 314
 315                         if (physbase < end_paddr &&
 316                             (physbase+size) > start_paddr) {
 317                                 /* overlaps */
 318                                 if (physbase < start_paddr) {
 319                                         size -= start_paddr - physbase;
 320                                         physbase = start_paddr;
 321                                 }
 322
 323                                 if (size > end_paddr - physbase)
 324                                         size = end_paddr - physbase;
 325
 326                                 dbg("free_bootmem %lx %lx\n", physbase, size);
 327                                 free_bootmem_node(NODE_DATA(nid), physbase,
 328                                                   size);
 329                         }
 330                 }
 331
 332                 for (i = 0; i < lmb.reserved.cnt; i++) {
 333                         unsigned long physbase = lmb.reserved.region[i].physbase;
 334                         unsigned long size = lmb.reserved.region[i].size;
 335
 336                         if (physbase < end_paddr &&
 337                             (physbase+size) > start_paddr) {
 338                                 /* overlaps */
 339                                 if (physbase < start_paddr) {
 340                                         size -= start_paddr - physbase;
 341                                         physbase = start_paddr;
 342                                 }
 343
 344                                 if (size > end_paddr - physbase)
 345                                         size = end_paddr - physbase;
 346
 347                                 dbg("reserve_bootmem %lx %lx\n", physbase,
 348                                     size);
 349                                 reserve_bootmem_node(NODE_DATA(nid), physbase,
 350                                                      size);
 351                         }
 352                 }
 353         }
 354 }
 355
 356 void __init paging_init(void)
 357 {
 358         unsigned long zones_size[MAX_NR_ZONES];
 359         unsigned long zholes_size[MAX_NR_ZONES];
 360         struct page *node_mem_map;
 361         int nid;
 362
 363         memset(zones_size, 0, sizeof(zones_size));
 364         memset(zholes_size, 0, sizeof(zholes_size));
 365
 366         for (nid = 0; nid < numnodes; nid++) {
 367                 unsigned long start_pfn;
 368                 unsigned long end_pfn;
 369
 370                 start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
 371                 end_pfn = plat_node_bdata[nid].node_low_pfn;
 372
 373                 zones_size[ZONE_DMA] = end_pfn - start_pfn;
 374                 zholes_size[ZONE_DMA] = 0;
 375                 if (nid == 0)
 376                         zholes_size[ZONE_DMA] = node0_io_hole_size;
 377
 378                 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
 379                     zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
 380
 381                 /*
 382                  * Give this empty node a dummy struct page to avoid
 383                  * us from trying to allocate a node local mem_map
 384                  * in free_area_init_node (which will fail).
 385                  */
 386                 if (!node_data[nid].node_spanned_pages)
 387                         node_mem_map = alloc_bootmem(sizeof(struct page));
 388                 else
 389                         node_mem_map = NULL;
 390
 391                 free_area_init_node(nid, NODE_DATA(nid), node_mem_map,
 392                                     zones_size, start_pfn, zholes_size);
 393         }
 394 }