arch/x86_64/mm/numa.c

   1 /*
   2  * Generic VM initialization for x86-64 NUMA setups.
   3  * Copyright 2002,2003 Andi Kleen, SuSE Labs.
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/mm.h>
   7 #include <linux/string.h>
   8 #include <linux/init.h>
   9 #include <linux/bootmem.h>
  10 #include <linux/mmzone.h>
  11 #include <linux/ctype.h>
  12 #include <linux/module.h>
  13 #include <linux/nodemask.h>
  14
  15 #include <asm/e820.h>
  16 #include <asm/proto.h>
  17 #include <asm/dma.h>
  18 #include <asm/numa.h>
  19 #include <asm/acpi.h>
  20
  21 #ifndef Dprintk
  22 #define Dprintk(x...)
  23 #endif
  24
  25 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  26 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  27
  28 struct memnode memnode;
  29
  30 unsigned char cpu_to_node[NR_CPUS] __read_mostly = {
  31         [0 ... NR_CPUS-1] = NUMA_NO_NODE
  32 };
  33 unsigned char apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  34         [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  35 };
  36 cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
  37
  38 int numa_off __initdata;
  39 unsigned long __initdata nodemap_addr;
  40 unsigned long __initdata nodemap_size;
  41
  42
  43 /*
  44  * Given a shift value, try to populate memnodemap[]
  45  * Returns :
  46  * 1 if OK
  47  * 0 if memnodmap[] too small (of shift too small)
  48  * -1 if node overlap or lost ram (shift too big)
  49  */
  50 static int __init
  51 populate_memnodemap(const struct bootnode *nodes, int numnodes, int shift)
  52 {
  53         int i;
  54         int res = -1;
  55         unsigned long addr, end;
  56
  57         memset(memnodemap, 0xff, memnodemapsize);
  58         for (i = 0; i < numnodes; i++) {
  59                 addr = nodes[i].start;
  60                 end = nodes[i].end;
  61                 if (addr >= end)
  62                         continue;
  63                 if ((end >> shift) >= memnodemapsize)
  64                         return 0;
  65                 do {
  66                         if (memnodemap[addr >> shift] != 0xff)
  67                                 return -1;
  68                         memnodemap[addr >> shift] = i;
  69                         addr += (1UL << shift);
  70                 } while (addr < end);
  71                 res = 1;
  72         }
  73         return res;
  74 }
  75
  76 static int __init allocate_cachealigned_memnodemap(void)
  77 {
  78         unsigned long pad, pad_addr;
  79
  80         memnodemap = memnode.embedded_map;
  81         if (memnodemapsize <= 48)
  82                 return 0;
  83
  84         pad = L1_CACHE_BYTES - 1;
  85         pad_addr = 0x8000;
  86         nodemap_size = pad + memnodemapsize;
  87         nodemap_addr = find_e820_area(pad_addr, end_pfn<<PAGE_SHIFT,
  88                                       nodemap_size);
  89         if (nodemap_addr == -1UL) {
  90                 printk(KERN_ERR
  91                        "NUMA: Unable to allocate Memory to Node hash map\n");
  92                 nodemap_addr = nodemap_size = 0;
  93                 return -1;
  94         }
  95         pad_addr = (nodemap_addr + pad) & ~pad;
  96         memnodemap = phys_to_virt(pad_addr);
  97
  98         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
  99                nodemap_addr, nodemap_addr + nodemap_size);
 100         return 0;
 101 }
 102
 103 /*
 104  * The LSB of all start and end addresses in the node map is the value of the
 105  * maximum possible shift.
 106  */
 107 static int __init
 108 extract_lsb_from_nodes (const struct bootnode *nodes, int numnodes)
 109 {
 110         int i, nodes_used = 0;
 111         unsigned long start, end;
 112         unsigned long bitfield = 0, memtop = 0;
 113
 114         for (i = 0; i < numnodes; i++) {
 115                 start = nodes[i].start;
 116                 end = nodes[i].end;
 117                 if (start >= end)
 118                         continue;
 119                 bitfield |= start;
 120                 nodes_used++;
 121                 if (end > memtop)
 122                         memtop = end;
 123         }
 124         if (nodes_used <= 1)
 125                 i = 63;
 126         else
 127                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 128         memnodemapsize = (memtop >> i)+1;
 129         return i;
 130 }
 131
 132 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
 133 {
 134         int shift;
 135
 136         shift = extract_lsb_from_nodes(nodes, numnodes);
 137         if (allocate_cachealigned_memnodemap())
 138                 return -1;
 139         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 140                 shift);
 141
 142         if (populate_memnodemap(nodes, numnodes, shift) != 1) {
 143                 printk(KERN_INFO
 144         "Your memory is not aligned you need to rebuild your kernel "
 145         "with a bigger NODEMAPSIZE shift=%d\n",
 146                         shift);
 147                 return -1;
 148         }
 149         return shift;
 150 }
 151
 152 #ifdef CONFIG_SPARSEMEM
 153 int early_pfn_to_nid(unsigned long pfn)
 154 {
 155         return phys_to_nid(pfn << PAGE_SHIFT);
 156 }
 157 #endif
 158
 159 static void * __init
 160 early_node_mem(int nodeid, unsigned long start, unsigned long end,
 161               unsigned long size)
 162 {
 163         unsigned long mem = find_e820_area(start, end, size);
 164         void *ptr;
 165         if (mem != -1L)
 166                 return __va(mem);
 167         ptr = __alloc_bootmem_nopanic(size,
 168                                 SMP_CACHE_BYTES, __pa(MAX_DMA_ADDRESS));
 169         if (ptr == 0) {
 170                 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 171                         size, nodeid);
 172                 return NULL;
 173         }
 174         return ptr;
 175 }
 176
 177 /* Initialize bootmem allocator for a node */
 178 void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
 179 {
 180         unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size, bootmap_start;
 181         unsigned long nodedata_phys;
 182         void *bootmap;
 183         const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
 184
 185         start = round_up(start, ZONE_ALIGN);
 186
 187         printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid, start, end);
 188
 189         start_pfn = start >> PAGE_SHIFT;
 190         end_pfn = end >> PAGE_SHIFT;
 191
 192         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size);
 193         if (node_data[nodeid] == NULL)
 194                 return;
 195         nodedata_phys = __pa(node_data[nodeid]);
 196
 197         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 198         NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
 199         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 200         NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
 201
 202         /* Find a place for the bootmem map */
 203         bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
 204         bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
 205         bootmap = early_node_mem(nodeid, bootmap_start, end,
 206                                         bootmap_pages<<PAGE_SHIFT);
 207         if (bootmap == NULL)  {
 208                 if (nodedata_phys < start || nodedata_phys >= end)
 209                         free_bootmem((unsigned long)node_data[nodeid],pgdat_size);
 210                 node_data[nodeid] = NULL;
 211                 return;
 212         }
 213         bootmap_start = __pa(bootmap);
 214         Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
 215
 216         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
 217                                          bootmap_start >> PAGE_SHIFT,
 218                                          start_pfn, end_pfn);
 219
 220         free_bootmem_with_active_regions(nodeid, end);
 221
 222         reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
 223         reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, bootmap_pages<<PAGE_SHIFT);
 224 #ifdef CONFIG_ACPI_NUMA
 225         srat_reserve_add_area(nodeid);
 226 #endif
 227         node_set_online(nodeid);
 228 }
 229
 230 /* Initialize final allocator for a zone */
 231 void __init setup_node_zones(int nodeid)
 232 {
 233         unsigned long start_pfn, end_pfn, memmapsize, limit;
 234
 235         start_pfn = node_start_pfn(nodeid);
 236         end_pfn = node_end_pfn(nodeid);
 237
 238         Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
 239                 nodeid, start_pfn, end_pfn);
 240
 241         /* Try to allocate mem_map at end to not fill up precious <4GB
 242            memory. */
 243         memmapsize = sizeof(struct page) * (end_pfn-start_pfn);
 244         limit = end_pfn << PAGE_SHIFT;
 245 #ifdef CONFIG_FLAT_NODE_MEM_MAP
 246         NODE_DATA(nodeid)->node_mem_map =
 247                 __alloc_bootmem_core(NODE_DATA(nodeid)->bdata,
 248                                 memmapsize, SMP_CACHE_BYTES,
 249                                 round_down(limit - memmapsize, PAGE_SIZE),
 250                                 limit);
 251 #endif
 252 }
 253
 254 void __init numa_init_array(void)
 255 {
 256         int rr, i;
 257         /* There are unfortunately some poorly designed mainboards around
 258            that only connect memory to a single CPU. This breaks the 1:1 cpu->node
 259            mapping. To avoid this fill in the mapping for all possible
 260            CPUs, as the number of CPUs is not known yet.
 261            We round robin the existing nodes. */
 262         rr = first_node(node_online_map);
 263         for (i = 0; i < NR_CPUS; i++) {
 264                 if (cpu_to_node[i] != NUMA_NO_NODE)
 265                         continue;
 266                 numa_set_node(i, rr);
 267                 rr = next_node(rr, node_online_map);
 268                 if (rr == MAX_NUMNODES)
 269                         rr = first_node(node_online_map);
 270         }
 271
 272 }
 273
 274 #ifdef CONFIG_NUMA_EMU
 275 /* Numa emulation */
 276 #define E820_ADDR_HOLE_SIZE(start, end)                                 \
 277         (e820_hole_size((start) >> PAGE_SHIFT, (end) >> PAGE_SHIFT) <<  \
 278         PAGE_SHIFT)
 279 char *cmdline __initdata;
 280
 281 /*
 282  * Setups up nid to range from addr to addr + size.  If the end boundary is
 283  * greater than max_addr, then max_addr is used instead.  The return value is 0
 284  * if there is additional memory left for allocation past addr and -1 otherwise.
 285  * addr is adjusted to be at the end of the node.
 286  */
 287 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
 288                                    u64 size, u64 max_addr)
 289 {
 290         int ret = 0;
 291         nodes[nid].start = *addr;
 292         *addr += size;
 293         if (*addr >= max_addr) {
 294                 *addr = max_addr;
 295                 ret = -1;
 296         }
 297         nodes[nid].end = *addr;
 298         node_set_online(nid);
 299         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 300                nodes[nid].start, nodes[nid].end,
 301                (nodes[nid].end - nodes[nid].start) >> 20);
 302         return ret;
 303 }
 304
 305 /*
 306  * Splits num_nodes nodes up equally starting at node_start.  The return value
 307  * is the number of nodes split up and addr is adjusted to be at the end of the
 308  * last node allocated.
 309  */
 310 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
 311                                       u64 max_addr, int node_start,
 312                                       int num_nodes)
 313 {
 314         unsigned int big;
 315         u64 size;
 316         int i;
 317
 318         if (num_nodes <= 0)
 319                 return -1;
 320         if (num_nodes > MAX_NUMNODES)
 321                 num_nodes = MAX_NUMNODES;
 322         size = (max_addr - *addr - E820_ADDR_HOLE_SIZE(*addr, max_addr)) /
 323                num_nodes;
 324         /*
 325          * Calculate the number of big nodes that can be allocated as a result
 326          * of consolidating the leftovers.
 327          */
 328         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
 329               FAKE_NODE_MIN_SIZE;
 330
 331         /* Round down to nearest FAKE_NODE_MIN_SIZE. */
 332         size &= FAKE_NODE_MIN_HASH_MASK;
 333         if (!size) {
 334                 printk(KERN_ERR "Not enough memory for each node.  "
 335                        "NUMA emulation disabled.\n");
 336                 return -1;
 337         }
 338
 339         for (i = node_start; i < num_nodes + node_start; i++) {
 340                 u64 end = *addr + size;
 341                 if (i < big)
 342                         end += FAKE_NODE_MIN_SIZE;
 343                 /*
 344                  * The final node can have the remaining system RAM.  Other
 345                  * nodes receive roughly the same amount of available pages.
 346                  */
 347                 if (i == num_nodes + node_start - 1)
 348                         end = max_addr;
 349                 else
 350                         while (end - *addr - E820_ADDR_HOLE_SIZE(*addr, end) <
 351                                size) {
 352                                 end += FAKE_NODE_MIN_SIZE;
 353                                 if (end > max_addr) {
 354                                         end = max_addr;
 355                                         break;
 356                                 }
 357                         }
 358                 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
 359                         break;
 360         }
 361         return i - node_start + 1;
 362 }
 363
 364 /*
 365  * Sets up the system RAM area from start_pfn to end_pfn according to the
 366  * numa=fake command-line option.
 367  */
 368 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
 369 {
 370         struct bootnode nodes[MAX_NUMNODES];
 371         u64 addr = start_pfn << PAGE_SHIFT;
 372         u64 max_addr = end_pfn << PAGE_SHIFT;
 373         unsigned int coeff;
 374         unsigned int num = 0;
 375         int num_nodes = 0;
 376         u64 size;
 377         int i;
 378
 379         memset(&nodes, 0, sizeof(nodes));
 380         /*
 381          * If the numa=fake command-line is just a single number N, split the
 382          * system RAM into N fake nodes.
 383          */
 384         if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
 385                 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0,
 386                                                 simple_strtol(cmdline, NULL, 0));
 387                 if (num_nodes < 0)
 388                         return num_nodes;
 389                 goto out;
 390         }
 391
 392         /* Parse the command line. */
 393         for (coeff = 1; ; cmdline++) {
 394                 if (*cmdline && isdigit(*cmdline)) {
 395                         num = num * 10 + *cmdline - '0';
 396                         continue;
 397                 }
 398                 if (*cmdline == '*')
 399                         coeff = num;
 400                 if (!*cmdline || *cmdline == ',') {
 401                         /*
 402                          * Round down to the nearest FAKE_NODE_MIN_SIZE.
 403                          * Command-line coefficients are in megabytes.
 404                          */
 405                         size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
 406                         if (size) {
 407                                 for (i = 0; i < coeff; i++, num_nodes++)
 408                                         if (setup_node_range(num_nodes, nodes,
 409                                                 &addr, size, max_addr) < 0)
 410                                                 goto done;
 411                                 coeff = 1;
 412                         }
 413                 }
 414                 if (!*cmdline)
 415                         break;
 416                 num = 0;
 417         }
 418 done:
 419         if (!num_nodes)
 420                 return -1;
 421         /* Fill remainder of system RAM with a final node, if appropriate. */
 422         if (addr < max_addr) {
 423                 setup_node_range(num_nodes, nodes, &addr, max_addr - addr,
 424                                  max_addr);
 425                 num_nodes++;
 426         }
 427 out:
 428         memnode_shift = compute_hash_shift(nodes, num_nodes);
 429         if (memnode_shift < 0) {
 430                 memnode_shift = 0;
 431                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 432                        "disabled.\n");
 433                 return -1;
 434         }
 435
 436         /*
 437          * We need to vacate all active ranges that may have been registered by
 438          * SRAT.
 439          */
 440         remove_all_active_ranges();
 441         for_each_online_node(i) {
 442                 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 443                                                 nodes[i].end >> PAGE_SHIFT);
 444                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 445         }
 446         numa_init_array();
 447         return 0;
 448 }
 449 #undef E820_ADDR_HOLE_SIZE
 450 #endif /* CONFIG_NUMA_EMU */
 451
 452 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 453 {
 454         int i;
 455
 456 #ifdef CONFIG_NUMA_EMU
 457         if (cmdline && !numa_emulation(start_pfn, end_pfn))
 458                 return;
 459 #endif
 460
 461 #ifdef CONFIG_ACPI_NUMA
 462         if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
 463                                           end_pfn << PAGE_SHIFT))
 464                 return;
 465 #endif
 466
 467 #ifdef CONFIG_K8_NUMA
 468         if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
 469                 return;
 470 #endif
 471         printk(KERN_INFO "%s\n",
 472                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 473
 474         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 475                start_pfn << PAGE_SHIFT,
 476                end_pfn << PAGE_SHIFT);
 477                 /* setup dummy node covering all memory */
 478         memnode_shift = 63;
 479         memnodemap = memnode.embedded_map;
 480         memnodemap[0] = 0;
 481         nodes_clear(node_online_map);
 482         node_set_online(0);
 483         for (i = 0; i < NR_CPUS; i++)
 484                 numa_set_node(i, 0);
 485         node_to_cpumask[0] = cpumask_of_cpu(0);
 486         e820_register_active_regions(0, start_pfn, end_pfn);
 487         setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
 488 }
 489
 490 __cpuinit void numa_add_cpu(int cpu)
 491 {
 492         set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
 493 }
 494
 495 void __cpuinit numa_set_node(int cpu, int node)
 496 {
 497         cpu_pda(cpu)->nodenumber = node;
 498         cpu_to_node[cpu] = node;
 499 }
 500
 501 unsigned long __init numa_free_all_bootmem(void)
 502 {
 503         int i;
 504         unsigned long pages = 0;
 505         for_each_online_node(i) {
 506                 pages += free_all_bootmem_node(NODE_DATA(i));
 507         }
 508         return pages;
 509 }
 510
 511 void __init paging_init(void)
 512 {
 513         int i;
 514         unsigned long max_zone_pfns[MAX_NR_ZONES];
 515         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 516         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 517         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 518         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 519
 520         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 521         sparse_init();
 522
 523         for_each_online_node(i) {
 524                 setup_node_zones(i);
 525         }
 526
 527         free_area_init_nodes(max_zone_pfns);
 528 }
 529
 530 static __init int numa_setup(char *opt)
 531 {
 532         if (!opt)
 533                 return -EINVAL;
 534         if (!strncmp(opt,"off",3))
 535                 numa_off = 1;
 536 #ifdef CONFIG_NUMA_EMU
 537         if (!strncmp(opt, "fake=", 5))
 538                 cmdline = opt + 5;
 539 #endif
 540 #ifdef CONFIG_ACPI_NUMA
 541         if (!strncmp(opt,"noacpi",6))
 542                 acpi_numa = -1;
 543         if (!strncmp(opt,"hotadd=", 7))
 544                 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
 545 #endif
 546         return 0;
 547 }
 548
 549 early_param("numa", numa_setup);
 550
 551 /*
 552  * Setup early cpu_to_node.
 553  *
 554  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
 555  * and apicid_to_node[] tables have valid entries for a CPU.
 556  * This means we skip cpu_to_node[] initialisation for NUMA
 557  * emulation and faking node case (when running a kernel compiled
 558  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
 559  * is already initialized in a round robin manner at numa_init_array,
 560  * prior to this call, and this initialization is good enough
 561  * for the fake NUMA cases.
 562  */
 563 void __init init_cpu_to_node(void)
 564 {
 565         int i;
 566         for (i = 0; i < NR_CPUS; i++) {
 567                 u8 apicid = x86_cpu_to_apicid[i];
 568                 if (apicid == BAD_APICID)
 569                         continue;
 570                 if (apicid_to_node[apicid] == NUMA_NO_NODE)
 571                         continue;
 572                 numa_set_node(i,apicid_to_node[apicid]);
 573         }
 574 }
 575
 576 EXPORT_SYMBOL(cpu_to_node);
 577 EXPORT_SYMBOL(node_to_cpumask);
 578 EXPORT_SYMBOL(memnode);
 579 EXPORT_SYMBOL(node_data);
 580
 581 #ifdef CONFIG_DISCONTIGMEM
 582 /*
 583  * Functions to convert PFNs from/to per node page addresses.
 584  * These are out of line because they are quite big.
 585  * They could be all tuned by pre caching more state.
 586  * Should do that.
 587  */
 588
 589 int pfn_valid(unsigned long pfn)
 590 {
 591         unsigned nid;
 592         if (pfn >= num_physpages)
 593                 return 0;
 594         nid = pfn_to_nid(pfn);
 595         if (nid == 0xff)
 596                 return 0;
 597         return pfn >= node_start_pfn(nid) && (pfn) < node_end_pfn(nid);
 598 }
 599 EXPORT_SYMBOL(pfn_valid);
 600 #endif