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 static int numa_enabled = 1;
  22
  23 static int numa_debug;
  24 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
  25
  26 #ifdef DEBUG_NUMA
  27 #define ARRAY_INITIALISER -1
  28 #else
  29 #define ARRAY_INITIALISER 0
  30 #endif
  31
  32 int numa_cpu_lookup_table[NR_CPUS] = { [ 0 ... (NR_CPUS - 1)] =
  33         ARRAY_INITIALISER};
  34 char *numa_memory_lookup_table;
  35 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
  36 int nr_cpus_in_node[MAX_NUMNODES] = { [0 ... (MAX_NUMNODES -1)] = 0};
  37
  38 struct pglist_data *node_data[MAX_NUMNODES];
  39 bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
  40 static unsigned long node0_io_hole_size;
  41
  42 /*
  43  * We need somewhere to store start/span for each node until we have
  44  * allocated the real node_data structures.
  45  */
  46 static struct {
  47         unsigned long node_start_pfn;
  48         unsigned long node_spanned_pages;
  49 } init_node_data[MAX_NUMNODES] __initdata;
  50
  51 EXPORT_SYMBOL(node_data);
  52 EXPORT_SYMBOL(numa_cpu_lookup_table);
  53 EXPORT_SYMBOL(numa_memory_lookup_table);
  54 EXPORT_SYMBOL(numa_cpumask_lookup_table);
  55 EXPORT_SYMBOL(nr_cpus_in_node);
  56
  57 static inline void map_cpu_to_node(int cpu, int node)
  58 {
  59         numa_cpu_lookup_table[cpu] = node;
  60         if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node]))) {
  61                 cpu_set(cpu, numa_cpumask_lookup_table[node]);
  62                 nr_cpus_in_node[node]++;
  63         }
  64 }
  65
  66 static struct device_node * __init find_cpu_node(unsigned int cpu)
  67 {
  68         unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
  69         struct device_node *cpu_node = NULL;
  70         unsigned int *interrupt_server, *reg;
  71         int len;
  72
  73         while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
  74                 /* Try interrupt server first */
  75                 interrupt_server = (unsigned int *)get_property(cpu_node,
  76                                         "ibm,ppc-interrupt-server#s", &len);
  77
  78                 if (interrupt_server && (len > 0)) {
  79                         while (len--) {
  80                                 if (interrupt_server[len-1] == hw_cpuid)
  81                                         return cpu_node;
  82                         }
  83                 } else {
  84                         reg = (unsigned int *)get_property(cpu_node,
  85                                                            "reg", &len);
  86                         if (reg && (len > 0) && (reg[0] == hw_cpuid))
  87                                 return cpu_node;
  88                 }
  89         }
  90
  91         return NULL;
  92 }
  93
  94 /* must hold reference to node during call */
  95 static int *of_get_associativity(struct device_node *dev)
  96  {
  97         unsigned int *result;
  98         int len;
  99
 100         result = (unsigned int *)get_property(dev, "ibm,associativity", &len);
 101
 102         if (len <= 0)
 103                 return NULL;
 104
 105         return result;
 106 }
 107
 108 static int of_node_numa_domain(struct device_node *device, int depth)
 109 {
 110         int numa_domain;
 111         unsigned int *tmp;
 112
 113         tmp = of_get_associativity(device);
 114         if (tmp && (tmp[0] >= depth)) {
 115                 numa_domain = tmp[depth];
 116         } else {
 117                 dbg("WARNING: no NUMA information for %s\n",
 118                     device->full_name);
 119                 numa_domain = 0;
 120         }
 121         return numa_domain;
 122 }
 123
 124 /*
 125  * In theory, the "ibm,associativity" property may contain multiple
 126  * associativity lists because a resource may be multiply connected
 127  * into the machine.  This resource then has different associativity
 128  * characteristics relative to its multiple connections.  We ignore
 129  * this for now.  We also assume that all cpu and memory sets have
 130  * their distances represented at a common level.  This won't be
 131  * true for heirarchical NUMA.
 132  *
 133  * In any case the ibm,associativity-reference-points should give
 134  * the correct depth for a normal NUMA system.
 135  *
 136  * - Dave Hansen <haveblue@us.ibm.com>
 137  */
 138 static int find_min_common_depth(void)
 139 {
 140         int depth;
 141         unsigned int *ref_points;
 142         struct device_node *rtas_root;
 143         unsigned int len;
 144
 145         rtas_root = of_find_node_by_path("/rtas");
 146
 147         if (!rtas_root)
 148                 return -1;
 149
 150         /*
 151          * this property is 2 32-bit integers, each representing a level of
 152          * depth in the associativity nodes.  The first is for an SMP
 153          * configuration (should be all 0's) and the second is for a normal
 154          * NUMA configuration.
 155          */
 156         ref_points = (unsigned int *)get_property(rtas_root,
 157                         "ibm,associativity-reference-points", &len);
 158
 159         if ((len >= 1) && ref_points) {
 160                 depth = ref_points[1];
 161         } else {
 162                 dbg("WARNING: could not find NUMA "
 163                     "associativity reference point\n");
 164                 depth = -1;
 165         }
 166         of_node_put(rtas_root);
 167
 168         return depth;
 169 }
 170
 171 static unsigned long read_cell_ul(struct device_node *device, unsigned int **buf)
 172 {
 173         int i;
 174         unsigned long result = 0;
 175
 176         i = prom_n_size_cells(device);
 177         /* bug on i>2 ?? */
 178         while (i--) {
 179                 result = (result << 32) | **buf;
 180                 (*buf)++;
 181         }
 182         return result;
 183 }
 184
 185 static int __init parse_numa_properties(void)
 186 {
 187         struct device_node *cpu = NULL;
 188         struct device_node *memory = NULL;
 189         int depth;
 190         int max_domain = 0;
 191         long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
 192         unsigned long i;
 193
 194         if (numa_enabled == 0) {
 195                 printk(KERN_WARNING "NUMA disabled by user\n");
 196                 return -1;
 197         }
 198
 199         numa_memory_lookup_table =
 200                 (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
 201         memset(numa_memory_lookup_table, 0, entries * sizeof(char));
 202
 203         for (i = 0; i < entries ; i++)
 204                 numa_memory_lookup_table[i] = ARRAY_INITIALISER;
 205
 206         depth = find_min_common_depth();
 207
 208         dbg("NUMA associativity depth for CPU/Memory: %d\n", depth);
 209         if (depth < 0)
 210                 return depth;
 211
 212         for_each_cpu(i) {
 213                 int numa_domain;
 214
 215                 cpu = find_cpu_node(i);
 216
 217                 if (cpu) {
 218                         numa_domain = of_node_numa_domain(cpu, depth);
 219                         of_node_put(cpu);
 220
 221                         if (numa_domain >= MAX_NUMNODES) {
 222                                 /*
 223                                  * POWER4 LPAR uses 0xffff as invalid node,
 224                                  * dont warn in this case.
 225                                  */
 226                                 if (numa_domain != 0xffff)
 227                                         printk(KERN_ERR "WARNING: cpu %ld "
 228                                                "maps to invalid NUMA node %d\n",
 229                                                i, numa_domain);
 230                                 numa_domain = 0;
 231                         }
 232                 } else {
 233                         dbg("WARNING: no NUMA information for cpu %ld\n", i);
 234                         numa_domain = 0;
 235                 }
 236
 237                 node_set_online(numa_domain);
 238
 239                 if (max_domain < numa_domain)
 240                         max_domain = numa_domain;
 241
 242                 map_cpu_to_node(i, numa_domain);
 243         }
 244
 245         memory = NULL;
 246         while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
 247                 unsigned long start;
 248                 unsigned long size;
 249                 int numa_domain;
 250                 int ranges;
 251                 unsigned int *memcell_buf;
 252                 unsigned int len;
 253
 254                 memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
 255                 if (!memcell_buf || len <= 0)
 256                         continue;
 257
 258                 ranges = memory->n_addrs;
 259 new_range:
 260                 /* these are order-sensitive, and modify the buffer pointer */
 261                 start = read_cell_ul(memory, &memcell_buf);
 262                 size = read_cell_ul(memory, &memcell_buf);
 263
 264                 start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
 265                 size = _ALIGN_UP(size, MEMORY_INCREMENT);
 266
 267                 numa_domain = of_node_numa_domain(memory, depth);
 268
 269                 if (numa_domain >= MAX_NUMNODES) {
 270                         if (numa_domain != 0xffff)
 271                                 printk(KERN_ERR "WARNING: memory at %lx maps "
 272                                        "to invalid NUMA node %d\n", start,
 273                                        numa_domain);
 274                         numa_domain = 0;
 275                 }
 276
 277                 node_set_online(numa_domain);
 278
 279                 if (max_domain < numa_domain)
 280                         max_domain = numa_domain;
 281
 282                 /*
 283                  * For backwards compatibility, OF splits the first node
 284                  * into two regions (the first being 0-4GB). Check for
 285                  * this simple case and complain if there is a gap in
 286                  * memory
 287                  */
 288                 if (init_node_data[numa_domain].node_spanned_pages) {
 289                         unsigned long shouldstart =
 290                                 init_node_data[numa_domain].node_start_pfn +
 291                                 init_node_data[numa_domain].node_spanned_pages;
 292                         if (shouldstart != (start / PAGE_SIZE)) {
 293                                 printk(KERN_ERR "WARNING: Hole in node, "
 294                                                 "disabling region start %lx "
 295                                                 "length %lx\n", start, size);
 296                                 continue;
 297                         }
 298                         init_node_data[numa_domain].node_spanned_pages +=
 299                                 size / PAGE_SIZE;
 300                 } else {
 301                         init_node_data[numa_domain].node_start_pfn =
 302                                 start / PAGE_SIZE;
 303                         init_node_data[numa_domain].node_spanned_pages =
 304                                 size / PAGE_SIZE;
 305                 }
 306
 307                 for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
 308                         numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] =
 309                                 numa_domain;
 310
 311                 ranges--;
 312                 if (ranges)
 313                         goto new_range;
 314         }
 315
 316         numnodes = max_domain + 1;
 317
 318         return 0;
 319 }
 320
 321 static void __init setup_nonnuma(void)
 322 {
 323         unsigned long top_of_ram = lmb_end_of_DRAM();
 324         unsigned long total_ram = lmb_phys_mem_size();
 325         unsigned long i;
 326
 327         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 328                top_of_ram, total_ram);
 329         printk(KERN_INFO "Memory hole size: %ldMB\n",
 330                (top_of_ram - total_ram) >> 20);
 331
 332         if (!numa_memory_lookup_table) {
 333                 long entries = top_of_ram >> MEMORY_INCREMENT_SHIFT;
 334                 numa_memory_lookup_table =
 335                         (char *)abs_to_virt(lmb_alloc(entries * sizeof(char), 1));
 336                 memset(numa_memory_lookup_table, 0, entries * sizeof(char));
 337                 for (i = 0; i < entries ; i++)
 338                         numa_memory_lookup_table[i] = ARRAY_INITIALISER;
 339         }
 340
 341         for (i = 0; i < NR_CPUS; i++)
 342                 map_cpu_to_node(i, 0);
 343
 344         node_set_online(0);
 345
 346         init_node_data[0].node_start_pfn = 0;
 347         init_node_data[0].node_spanned_pages = lmb_end_of_DRAM() / PAGE_SIZE;
 348
 349         for (i = 0 ; i < top_of_ram; i += MEMORY_INCREMENT)
 350                 numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] = 0;
 351
 352         node0_io_hole_size = top_of_ram - total_ram;
 353 }
 354
 355 static void __init dump_numa_topology(void)
 356 {
 357         unsigned int node;
 358         unsigned int cpu, count;
 359
 360         for (node = 0; node < MAX_NUMNODES; node++) {
 361                 if (!node_online(node))
 362                         continue;
 363
 364                 printk(KERN_INFO "Node %d CPUs:", node);
 365
 366                 count = 0;
 367                 /*
 368                  * If we used a CPU iterator here we would miss printing
 369                  * the holes in the cpumap.
 370                  */
 371                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
 372                         if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
 373                                 if (count == 0)
 374                                         printk(" %u", cpu);
 375                                 ++count;
 376                         } else {
 377                                 if (count > 1)
 378                                         printk("-%u", cpu - 1);
 379                                 count = 0;
 380                         }
 381                 }
 382
 383                 if (count > 1)
 384                         printk("-%u", NR_CPUS - 1);
 385                 printk("\n");
 386         }
 387
 388         for (node = 0; node < MAX_NUMNODES; node++) {
 389                 unsigned long i;
 390
 391                 if (!node_online(node))
 392                         continue;
 393
 394                 printk(KERN_INFO "Node %d Memory:", node);
 395
 396                 count = 0;
 397
 398                 for (i = 0; i < lmb_end_of_DRAM(); i += MEMORY_INCREMENT) {
 399                         if (numa_memory_lookup_table[i >> MEMORY_INCREMENT_SHIFT] == node) {
 400                                 if (count == 0)
 401                                         printk(" 0x%lx", i);
 402                                 ++count;
 403                         } else {
 404                                 if (count > 0)
 405                                         printk("-0x%lx", i);
 406                                 count = 0;
 407                         }
 408                 }
 409
 410                 if (count > 0)
 411                         printk("-0x%lx", i);
 412                 printk("\n");
 413         }
 414 }
 415
 416 /*
 417  * Allocate some memory, satisfying the lmb or bootmem allocator where
 418  * required. nid is the preferred node and end is the physical address of
 419  * the highest address in the node.
 420  *
 421  * Returns the physical address of the memory.
 422  */
 423 static unsigned long careful_allocation(int nid, unsigned long size,
 424                                         unsigned long align, unsigned long end)
 425 {
 426         unsigned long ret = lmb_alloc_base(size, align, end);
 427
 428         /* retry over all memory */
 429         if (!ret)
 430                 ret = lmb_alloc_base(size, align, lmb_end_of_DRAM());
 431
 432         if (!ret)
 433                 panic("numa.c: cannot allocate %lu bytes on node %d",
 434                       size, nid);
 435
 436         /*
 437          * If the memory came from a previously allocated node, we must
 438          * retry with the bootmem allocator.
 439          */
 440         if (pa_to_nid(ret) < nid) {
 441                 nid = pa_to_nid(ret);
 442                 ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(nid),
 443                                 size, align, 0);
 444
 445                 if (!ret)
 446                         panic("numa.c: cannot allocate %lu bytes on node %d",
 447                               size, nid);
 448
 449                 ret = virt_to_abs(ret);
 450
 451                 dbg("alloc_bootmem %lx %lx\n", ret, size);
 452         }
 453
 454         return ret;
 455 }
 456
 457 void __init do_init_bootmem(void)
 458 {
 459         int nid;
 460
 461         min_low_pfn = 0;
 462         max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
 463         max_pfn = max_low_pfn;
 464
 465         if (parse_numa_properties())
 466                 setup_nonnuma();
 467         else
 468                 dump_numa_topology();
 469
 470         for (nid = 0; nid < numnodes; nid++) {
 471                 unsigned long start_paddr, end_paddr;
 472                 int i;
 473                 unsigned long bootmem_paddr;
 474                 unsigned long bootmap_pages;
 475
 476                 start_paddr = init_node_data[nid].node_start_pfn * PAGE_SIZE;
 477                 end_paddr = start_paddr + (init_node_data[nid].node_spanned_pages * PAGE_SIZE);
 478
 479                 /* Allocate the node structure node local if possible */
 480                 NODE_DATA(nid) = (struct pglist_data *)careful_allocation(nid,
 481                                         sizeof(struct pglist_data),
 482                                         SMP_CACHE_BYTES, end_paddr);
 483                 NODE_DATA(nid) = abs_to_virt(NODE_DATA(nid));
 484                 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 485
 486                 dbg("node %d\n", nid);
 487                 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
 488
 489                 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
 490                 NODE_DATA(nid)->node_start_pfn =
 491                         init_node_data[nid].node_start_pfn;
 492                 NODE_DATA(nid)->node_spanned_pages =
 493                         init_node_data[nid].node_spanned_pages;
 494
 495                 if (init_node_data[nid].node_spanned_pages == 0)
 496                         continue;
 497
 498                 dbg("start_paddr = %lx\n", start_paddr);
 499                 dbg("end_paddr = %lx\n", end_paddr);
 500
 501                 bootmap_pages = bootmem_bootmap_pages((end_paddr - start_paddr) >> PAGE_SHIFT);
 502
 503                 bootmem_paddr = careful_allocation(nid,
 504                                 bootmap_pages << PAGE_SHIFT,
 505                                 PAGE_SIZE, end_paddr);
 506                 memset(abs_to_virt(bootmem_paddr), 0,
 507                        bootmap_pages << PAGE_SHIFT);
 508                 dbg("bootmap_paddr = %lx\n", bootmem_paddr);
 509
 510                 init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
 511                                   start_paddr >> PAGE_SHIFT,
 512                                   end_paddr >> PAGE_SHIFT);
 513
 514                 for (i = 0; i < lmb.memory.cnt; i++) {
 515                         unsigned long physbase, size;
 516
 517                         physbase = lmb.memory.region[i].physbase;
 518                         size = lmb.memory.region[i].size;
 519
 520                         if (physbase < end_paddr &&
 521                             (physbase+size) > start_paddr) {
 522                                 /* overlaps */
 523                                 if (physbase < start_paddr) {
 524                                         size -= start_paddr - physbase;
 525                                         physbase = start_paddr;
 526                                 }
 527
 528                                 if (size > end_paddr - physbase)
 529                                         size = end_paddr - physbase;
 530
 531                                 dbg("free_bootmem %lx %lx\n", physbase, size);
 532                                 free_bootmem_node(NODE_DATA(nid), physbase,
 533                                                   size);
 534                         }
 535                 }
 536
 537                 for (i = 0; i < lmb.reserved.cnt; i++) {
 538                         unsigned long physbase = lmb.reserved.region[i].physbase;
 539                         unsigned long size = lmb.reserved.region[i].size;
 540
 541                         if (physbase < end_paddr &&
 542                             (physbase+size) > start_paddr) {
 543                                 /* overlaps */
 544                                 if (physbase < start_paddr) {
 545                                         size -= start_paddr - physbase;
 546                                         physbase = start_paddr;
 547                                 }
 548
 549                                 if (size > end_paddr - physbase)
 550                                         size = end_paddr - physbase;
 551
 552                                 dbg("reserve_bootmem %lx %lx\n", physbase,
 553                                     size);
 554                                 reserve_bootmem_node(NODE_DATA(nid), physbase,
 555                                                      size);
 556                         }
 557                 }
 558         }
 559 }
 560
 561 void __init paging_init(void)
 562 {
 563         unsigned long zones_size[MAX_NR_ZONES];
 564         unsigned long zholes_size[MAX_NR_ZONES];
 565         int nid;
 566
 567         memset(zones_size, 0, sizeof(zones_size));
 568         memset(zholes_size, 0, sizeof(zholes_size));
 569
 570         for (nid = 0; nid < numnodes; nid++) {
 571                 unsigned long start_pfn;
 572                 unsigned long end_pfn;
 573
 574                 start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
 575                 end_pfn = plat_node_bdata[nid].node_low_pfn;
 576
 577                 zones_size[ZONE_DMA] = end_pfn - start_pfn;
 578                 zholes_size[ZONE_DMA] = 0;
 579                 if (nid == 0)
 580                         zholes_size[ZONE_DMA] = node0_io_hole_size >> PAGE_SHIFT;
 581
 582                 dbg("free_area_init node %d %lx %lx (hole: %lx)\n", nid,
 583                     zones_size[ZONE_DMA], start_pfn, zholes_size[ZONE_DMA]);
 584
 585                 free_area_init_node(nid, NODE_DATA(nid), zones_size,
 586                                                         start_pfn, zholes_size);
 587         }
 588 }
 589
 590 static int __init early_numa(char *p)
 591 {
 592         if (!p)
 593                 return 0;
 594
 595         if (strstr(p, "off"))
 596                 numa_enabled = 0;
 597
 598         if (strstr(p, "debug"))
 599                 numa_debug = 1;
 600
 601         return 0;
 602 }
 603 early_param("numa", early_numa);