arch/x86/mm/numa_64.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/memblock.h>
  11 #include <linux/mmzone.h>
  12 #include <linux/ctype.h>
  13 #include <linux/module.h>
  14 #include <linux/nodemask.h>
  15 #include <linux/sched.h>
  16 #include <linux/acpi.h>
  17
  18 #include <asm/e820.h>
  19 #include <asm/proto.h>
  20 #include <asm/dma.h>
  21 #include <asm/acpi.h>
  22 #include <asm/amd_nb.h>
  23
  24 #include "numa_internal.h"
  25
  26 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  27 EXPORT_SYMBOL(node_data);
  28
  29 nodemask_t numa_nodes_parsed __initdata;
  30
  31 static struct numa_meminfo numa_meminfo __initdata;
  32 static int numa_distance_cnt;
  33 static u8 *numa_distance;
  34
  35 static void * __init early_node_mem(int nodeid, unsigned long start,
  36                                     unsigned long end, unsigned long size,
  37                                     unsigned long align)
  38 {
  39         unsigned long mem;
  40
  41         /*
  42          * put it on high as possible
  43          * something will go with NODE_DATA
  44          */
  45         if (start < (MAX_DMA_PFN<<PAGE_SHIFT))
  46                 start = MAX_DMA_PFN<<PAGE_SHIFT;
  47         if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) &&
  48             end > (MAX_DMA32_PFN<<PAGE_SHIFT))
  49                 start = MAX_DMA32_PFN<<PAGE_SHIFT;
  50         mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
  51         if (mem != MEMBLOCK_ERROR)
  52                 return __va(mem);
  53
  54         /* extend the search scope */
  55         end = max_pfn_mapped << PAGE_SHIFT;
  56         start = MAX_DMA_PFN << PAGE_SHIFT;
  57         mem = memblock_find_in_range(start, end, size, align);
  58         if (mem != MEMBLOCK_ERROR)
  59                 return __va(mem);
  60
  61         printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
  62                        size, nodeid);
  63
  64         return NULL;
  65 }
  66
  67 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
  68                                      struct numa_meminfo *mi)
  69 {
  70         /* ignore zero length blks */
  71         if (start == end)
  72                 return 0;
  73
  74         /* whine about and ignore invalid blks */
  75         if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
  76                 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
  77                            nid, start, end);
  78                 return 0;
  79         }
  80
  81         if (mi->nr_blks >= NR_NODE_MEMBLKS) {
  82                 pr_err("NUMA: too many memblk ranges\n");
  83                 return -EINVAL;
  84         }
  85
  86         mi->blk[mi->nr_blks].start = start;
  87         mi->blk[mi->nr_blks].end = end;
  88         mi->blk[mi->nr_blks].nid = nid;
  89         mi->nr_blks++;
  90         return 0;
  91 }
  92
  93 /**
  94  * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
  95  * @idx: Index of memblk to remove
  96  * @mi: numa_meminfo to remove memblk from
  97  *
  98  * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
  99  * decrementing @mi->nr_blks.
 100  */
 101 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
 102 {
 103         mi->nr_blks--;
 104         memmove(&mi->blk[idx], &mi->blk[idx + 1],
 105                 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
 106 }
 107
 108 /**
 109  * numa_add_memblk - Add one numa_memblk to numa_meminfo
 110  * @nid: NUMA node ID of the new memblk
 111  * @start: Start address of the new memblk
 112  * @end: End address of the new memblk
 113  *
 114  * Add a new memblk to the default numa_meminfo.
 115  *
 116  * RETURNS:
 117  * 0 on success, -errno on failure.
 118  */
 119 int __init numa_add_memblk(int nid, u64 start, u64 end)
 120 {
 121         return numa_add_memblk_to(nid, start, end, &numa_meminfo);
 122 }
 123
 124 /* Initialize bootmem allocator for a node */
 125 void __init
 126 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
 127 {
 128         unsigned long start_pfn, last_pfn, nodedata_phys;
 129         const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
 130         int nid;
 131
 132         if (!end)
 133                 return;
 134
 135         /*
 136          * Don't confuse VM with a node that doesn't have the
 137          * minimum amount of memory:
 138          */
 139         if (end && (end - start) < NODE_MIN_SIZE)
 140                 return;
 141
 142         start = roundup(start, ZONE_ALIGN);
 143
 144         printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid,
 145                start, end);
 146
 147         start_pfn = start >> PAGE_SHIFT;
 148         last_pfn = end >> PAGE_SHIFT;
 149
 150         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
 151                                            SMP_CACHE_BYTES);
 152         if (node_data[nodeid] == NULL)
 153                 return;
 154         nodedata_phys = __pa(node_data[nodeid]);
 155         memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
 156         printk(KERN_INFO "  NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
 157                 nodedata_phys + pgdat_size - 1);
 158         nid = early_pfn_to_nid(nodedata_phys >> PAGE_SHIFT);
 159         if (nid != nodeid)
 160                 printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nodeid, nid);
 161
 162         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 163         NODE_DATA(nodeid)->node_id = nodeid;
 164         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 165         NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
 166
 167         node_set_online(nodeid);
 168 }
 169
 170 /**
 171  * numa_cleanup_meminfo - Cleanup a numa_meminfo
 172  * @mi: numa_meminfo to clean up
 173  *
 174  * Sanitize @mi by merging and removing unncessary memblks.  Also check for
 175  * conflicts and clear unused memblks.
 176  *
 177  * RETURNS:
 178  * 0 on success, -errno on failure.
 179  */
 180 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
 181 {
 182         const u64 low = 0;
 183         const u64 high = (u64)max_pfn << PAGE_SHIFT;
 184         int i, j, k;
 185
 186         for (i = 0; i < mi->nr_blks; i++) {
 187                 struct numa_memblk *bi = &mi->blk[i];
 188
 189                 /* make sure all blocks are inside the limits */
 190                 bi->start = max(bi->start, low);
 191                 bi->end = min(bi->end, high);
 192
 193                 /* and there's no empty block */
 194                 if (bi->start == bi->end) {
 195                         numa_remove_memblk_from(i--, mi);
 196                         continue;
 197                 }
 198
 199                 for (j = i + 1; j < mi->nr_blks; j++) {
 200                         struct numa_memblk *bj = &mi->blk[j];
 201                         unsigned long start, end;
 202
 203                         /*
 204                          * See whether there are overlapping blocks.  Whine
 205                          * about but allow overlaps of the same nid.  They
 206                          * will be merged below.
 207                          */
 208                         if (bi->end > bj->start && bi->start < bj->end) {
 209                                 if (bi->nid != bj->nid) {
 210                                         pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
 211                                                bi->nid, bi->start, bi->end,
 212                                                bj->nid, bj->start, bj->end);
 213                                         return -EINVAL;
 214                                 }
 215                                 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
 216                                            bi->nid, bi->start, bi->end,
 217                                            bj->start, bj->end);
 218                         }
 219
 220                         /*
 221                          * Join together blocks on the same node, holes
 222                          * between which don't overlap with memory on other
 223                          * nodes.
 224                          */
 225                         if (bi->nid != bj->nid)
 226                                 continue;
 227                         start = max(min(bi->start, bj->start), low);
 228                         end = min(max(bi->end, bj->end), high);
 229                         for (k = 0; k < mi->nr_blks; k++) {
 230                                 struct numa_memblk *bk = &mi->blk[k];
 231
 232                                 if (bi->nid == bk->nid)
 233                                         continue;
 234                                 if (start < bk->end && end > bk->start)
 235                                         break;
 236                         }
 237                         if (k < mi->nr_blks)
 238                                 continue;
 239                         printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n",
 240                                bi->nid, bi->start, bi->end, bj->start, bj->end,
 241                                start, end);
 242                         bi->start = start;
 243                         bi->end = end;
 244                         numa_remove_memblk_from(j--, mi);
 245                 }
 246         }
 247
 248         for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
 249                 mi->blk[i].start = mi->blk[i].end = 0;
 250                 mi->blk[i].nid = NUMA_NO_NODE;
 251         }
 252
 253         return 0;
 254 }
 255
 256 /*
 257  * Set nodes, which have memory in @mi, in *@nodemask.
 258  */
 259 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
 260                                               const struct numa_meminfo *mi)
 261 {
 262         int i;
 263
 264         for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
 265                 if (mi->blk[i].start != mi->blk[i].end &&
 266                     mi->blk[i].nid != NUMA_NO_NODE)
 267                         node_set(mi->blk[i].nid, *nodemask);
 268 }
 269
 270 /**
 271  * numa_reset_distance - Reset NUMA distance table
 272  *
 273  * The current table is freed.  The next numa_set_distance() call will
 274  * create a new one.
 275  */
 276 void __init numa_reset_distance(void)
 277 {
 278         size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
 279
 280         /* numa_distance could be 1LU marking allocation failure, test cnt */
 281         if (numa_distance_cnt)
 282                 memblock_x86_free_range(__pa(numa_distance),
 283                                         __pa(numa_distance) + size);
 284         numa_distance_cnt = 0;
 285         numa_distance = NULL;   /* enable table creation */
 286 }
 287
 288 static int __init numa_alloc_distance(void)
 289 {
 290         nodemask_t nodes_parsed;
 291         size_t size;
 292         int i, j, cnt = 0;
 293         u64 phys;
 294
 295         /* size the new table and allocate it */
 296         nodes_parsed = numa_nodes_parsed;
 297         numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
 298
 299         for_each_node_mask(i, nodes_parsed)
 300                 cnt = i;
 301         cnt++;
 302         size = cnt * cnt * sizeof(numa_distance[0]);
 303
 304         phys = memblock_find_in_range(0, (u64)max_pfn_mapped << PAGE_SHIFT,
 305                                       size, PAGE_SIZE);
 306         if (phys == MEMBLOCK_ERROR) {
 307                 pr_warning("NUMA: Warning: can't allocate distance table!\n");
 308                 /* don't retry until explicitly reset */
 309                 numa_distance = (void *)1LU;
 310                 return -ENOMEM;
 311         }
 312         memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
 313
 314         numa_distance = __va(phys);
 315         numa_distance_cnt = cnt;
 316
 317         /* fill with the default distances */
 318         for (i = 0; i < cnt; i++)
 319                 for (j = 0; j < cnt; j++)
 320                         numa_distance[i * cnt + j] = i == j ?
 321                                 LOCAL_DISTANCE : REMOTE_DISTANCE;
 322         printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
 323
 324         return 0;
 325 }
 326
 327 /**
 328  * numa_set_distance - Set NUMA distance from one NUMA to another
 329  * @from: the 'from' node to set distance
 330  * @to: the 'to'  node to set distance
 331  * @distance: NUMA distance
 332  *
 333  * Set the distance from node @from to @to to @distance.  If distance table
 334  * doesn't exist, one which is large enough to accommodate all the currently
 335  * known nodes will be created.
 336  *
 337  * If such table cannot be allocated, a warning is printed and further
 338  * calls are ignored until the distance table is reset with
 339  * numa_reset_distance().
 340  *
 341  * If @from or @to is higher than the highest known node at the time of
 342  * table creation or @distance doesn't make sense, the call is ignored.
 343  * This is to allow simplification of specific NUMA config implementations.
 344  */
 345 void __init numa_set_distance(int from, int to, int distance)
 346 {
 347         if (!numa_distance && numa_alloc_distance() < 0)
 348                 return;
 349
 350         if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
 351                 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
 352                             from, to, distance);
 353                 return;
 354         }
 355
 356         if ((u8)distance != distance ||
 357             (from == to && distance != LOCAL_DISTANCE)) {
 358                 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
 359                              from, to, distance);
 360                 return;
 361         }
 362
 363         numa_distance[from * numa_distance_cnt + to] = distance;
 364 }
 365
 366 int __node_distance(int from, int to)
 367 {
 368         if (from >= numa_distance_cnt || to >= numa_distance_cnt)
 369                 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
 370         return numa_distance[from * numa_distance_cnt + to];
 371 }
 372 EXPORT_SYMBOL(__node_distance);
 373
 374 /*
 375  * Sanity check to catch more bad NUMA configurations (they are amazingly
 376  * common).  Make sure the nodes cover all memory.
 377  */
 378 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
 379 {
 380         unsigned long numaram, e820ram;
 381         int i;
 382
 383         numaram = 0;
 384         for (i = 0; i < mi->nr_blks; i++) {
 385                 unsigned long s = mi->blk[i].start >> PAGE_SHIFT;
 386                 unsigned long e = mi->blk[i].end >> PAGE_SHIFT;
 387                 numaram += e - s;
 388                 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
 389                 if ((long)numaram < 0)
 390                         numaram = 0;
 391         }
 392
 393         e820ram = max_pfn - (memblock_x86_hole_size(0,
 394                                         max_pfn << PAGE_SHIFT) >> PAGE_SHIFT);
 395         /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
 396         if ((long)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
 397                 printk(KERN_ERR "NUMA: nodes only cover %luMB of your %luMB e820 RAM. Not used.\n",
 398                        (numaram << PAGE_SHIFT) >> 20,
 399                        (e820ram << PAGE_SHIFT) >> 20);
 400                 return false;
 401         }
 402         return true;
 403 }
 404
 405 static int __init numa_register_memblks(struct numa_meminfo *mi)
 406 {
 407         int i, nid;
 408
 409         /* Account for nodes with cpus and no memory */
 410         node_possible_map = numa_nodes_parsed;
 411         numa_nodemask_from_meminfo(&node_possible_map, mi);
 412         if (WARN_ON(nodes_empty(node_possible_map)))
 413                 return -EINVAL;
 414
 415         for (i = 0; i < mi->nr_blks; i++)
 416                 memblock_x86_register_active_regions(mi->blk[i].nid,
 417                                         mi->blk[i].start >> PAGE_SHIFT,
 418                                         mi->blk[i].end >> PAGE_SHIFT);
 419
 420         /* for out of order entries */
 421         sort_node_map();
 422         if (!numa_meminfo_cover_memory(mi))
 423                 return -EINVAL;
 424
 425         /* Finally register nodes. */
 426         for_each_node_mask(nid, node_possible_map) {
 427                 u64 start = (u64)max_pfn << PAGE_SHIFT;
 428                 u64 end = 0;
 429
 430                 for (i = 0; i < mi->nr_blks; i++) {
 431                         if (nid != mi->blk[i].nid)
 432                                 continue;
 433                         start = min(mi->blk[i].start, start);
 434                         end = max(mi->blk[i].end, end);
 435                 }
 436
 437                 if (start < end)
 438                         setup_node_bootmem(nid, start, end);
 439         }
 440
 441         return 0;
 442 }
 443
 444 /**
 445  * dummy_numma_init - Fallback dummy NUMA init
 446  *
 447  * Used if there's no underlying NUMA architecture, NUMA initialization
 448  * fails, or NUMA is disabled on the command line.
 449  *
 450  * Must online at least one node and add memory blocks that cover all
 451  * allowed memory.  This function must not fail.
 452  */
 453 static int __init dummy_numa_init(void)
 454 {
 455         printk(KERN_INFO "%s\n",
 456                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 457         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 458                0LU, max_pfn << PAGE_SHIFT);
 459
 460         node_set(0, numa_nodes_parsed);
 461         numa_add_memblk(0, 0, (u64)max_pfn << PAGE_SHIFT);
 462
 463         return 0;
 464 }
 465
 466 static int __init numa_init(int (*init_func)(void))
 467 {
 468         int i;
 469         int ret;
 470
 471         for (i = 0; i < MAX_LOCAL_APIC; i++)
 472                 set_apicid_to_node(i, NUMA_NO_NODE);
 473
 474         nodes_clear(numa_nodes_parsed);
 475         nodes_clear(node_possible_map);
 476         nodes_clear(node_online_map);
 477         memset(&numa_meminfo, 0, sizeof(numa_meminfo));
 478         remove_all_active_ranges();
 479         numa_reset_distance();
 480
 481         ret = init_func();
 482         if (ret < 0)
 483                 return ret;
 484         ret = numa_cleanup_meminfo(&numa_meminfo);
 485         if (ret < 0)
 486                 return ret;
 487
 488         numa_emulation(&numa_meminfo, numa_distance_cnt);
 489
 490         ret = numa_register_memblks(&numa_meminfo);
 491         if (ret < 0)
 492                 return ret;
 493
 494         for (i = 0; i < nr_cpu_ids; i++) {
 495                 int nid = early_cpu_to_node(i);
 496
 497                 if (nid == NUMA_NO_NODE)
 498                         continue;
 499                 if (!node_online(nid))
 500                         numa_clear_node(i);
 501         }
 502         numa_init_array();
 503         return 0;
 504 }
 505
 506 void __init initmem_init(void)
 507 {
 508         int ret;
 509
 510         if (!numa_off) {
 511 #ifdef CONFIG_ACPI_NUMA
 512                 ret = numa_init(x86_acpi_numa_init);
 513                 if (!ret)
 514                         return;
 515 #endif
 516 #ifdef CONFIG_AMD_NUMA
 517                 ret = numa_init(amd_numa_init);
 518                 if (!ret)
 519                         return;
 520 #endif
 521         }
 522
 523         numa_init(dummy_numa_init);
 524 }
 525
 526 unsigned long __init numa_free_all_bootmem(void)
 527 {
 528         unsigned long pages = 0;
 529         int i;
 530
 531         for_each_online_node(i)
 532                 pages += free_all_bootmem_node(NODE_DATA(i));
 533
 534         pages += free_all_memory_core_early(MAX_NUMNODES);
 535
 536         return pages;
 537 }
 538
 539 int __cpuinit numa_cpu_node(int cpu)
 540 {
 541         int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
 542
 543         if (apicid != BAD_APICID)
 544                 return __apicid_to_node[apicid];
 545         return NUMA_NO_NODE;
 546 }