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/numa.h>
  22 #include <asm/acpi.h>
  23 #include <asm/amd_nb.h>
  24
  25 struct numa_memblk {
  26         u64                     start;
  27         u64                     end;
  28         int                     nid;
  29 };
  30
  31 struct numa_meminfo {
  32         int                     nr_blks;
  33         struct numa_memblk      blk[NR_NODE_MEMBLKS];
  34 };
  35
  36 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  37 EXPORT_SYMBOL(node_data);
  38
  39 nodemask_t numa_nodes_parsed __initdata;
  40 nodemask_t mem_nodes_parsed __initdata;
  41
  42 struct memnode memnode;
  43
  44 static unsigned long __initdata nodemap_addr;
  45 static unsigned long __initdata nodemap_size;
  46
  47 static struct numa_meminfo numa_meminfo __initdata;
  48
  49 /*
  50  * Given a shift value, try to populate memnodemap[]
  51  * Returns :
  52  * 1 if OK
  53  * 0 if memnodmap[] too small (of shift too small)
  54  * -1 if node overlap or lost ram (shift too big)
  55  */
  56 static int __init populate_memnodemap(const struct numa_meminfo *mi, int shift)
  57 {
  58         unsigned long addr, end;
  59         int i, res = -1;
  60
  61         memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
  62         for (i = 0; i < mi->nr_blks; i++) {
  63                 addr = mi->blk[i].start;
  64                 end = mi->blk[i].end;
  65                 if (addr >= end)
  66                         continue;
  67                 if ((end >> shift) >= memnodemapsize)
  68                         return 0;
  69                 do {
  70                         if (memnodemap[addr >> shift] != NUMA_NO_NODE)
  71                                 return -1;
  72                         memnodemap[addr >> shift] = mi->blk[i].nid;
  73                         addr += (1UL << shift);
  74                 } while (addr < end);
  75                 res = 1;
  76         }
  77         return res;
  78 }
  79
  80 static int __init allocate_cachealigned_memnodemap(void)
  81 {
  82         unsigned long addr;
  83
  84         memnodemap = memnode.embedded_map;
  85         if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
  86                 return 0;
  87
  88         addr = 0x8000;
  89         nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
  90         nodemap_addr = memblock_find_in_range(addr, get_max_mapped(),
  91                                       nodemap_size, L1_CACHE_BYTES);
  92         if (nodemap_addr == MEMBLOCK_ERROR) {
  93                 printk(KERN_ERR
  94                        "NUMA: Unable to allocate Memory to Node hash map\n");
  95                 nodemap_addr = nodemap_size = 0;
  96                 return -1;
  97         }
  98         memnodemap = phys_to_virt(nodemap_addr);
  99         memblock_x86_reserve_range(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
 100
 101         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
 102                nodemap_addr, nodemap_addr + nodemap_size);
 103         return 0;
 104 }
 105
 106 /*
 107  * The LSB of all start and end addresses in the node map is the value of the
 108  * maximum possible shift.
 109  */
 110 static int __init extract_lsb_from_nodes(const struct numa_meminfo *mi)
 111 {
 112         int i, nodes_used = 0;
 113         unsigned long start, end;
 114         unsigned long bitfield = 0, memtop = 0;
 115
 116         for (i = 0; i < mi->nr_blks; i++) {
 117                 start = mi->blk[i].start;
 118                 end = mi->blk[i].end;
 119                 if (start >= end)
 120                         continue;
 121                 bitfield |= start;
 122                 nodes_used++;
 123                 if (end > memtop)
 124                         memtop = end;
 125         }
 126         if (nodes_used <= 1)
 127                 i = 63;
 128         else
 129                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 130         memnodemapsize = (memtop >> i)+1;
 131         return i;
 132 }
 133
 134 static int __init compute_hash_shift(const struct numa_meminfo *mi)
 135 {
 136         int shift;
 137
 138         shift = extract_lsb_from_nodes(mi);
 139         if (allocate_cachealigned_memnodemap())
 140                 return -1;
 141         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 142                 shift);
 143
 144         if (populate_memnodemap(mi, shift) != 1) {
 145                 printk(KERN_INFO "Your memory is not aligned you need to "
 146                        "rebuild your kernel with a bigger NODEMAPSIZE "
 147                        "shift=%d\n", shift);
 148                 return -1;
 149         }
 150         return shift;
 151 }
 152
 153 int __meminit  __early_pfn_to_nid(unsigned long pfn)
 154 {
 155         return phys_to_nid(pfn << PAGE_SHIFT);
 156 }
 157
 158 static void * __init early_node_mem(int nodeid, unsigned long start,
 159                                     unsigned long end, unsigned long size,
 160                                     unsigned long align)
 161 {
 162         unsigned long mem;
 163
 164         /*
 165          * put it on high as possible
 166          * something will go with NODE_DATA
 167          */
 168         if (start < (MAX_DMA_PFN<<PAGE_SHIFT))
 169                 start = MAX_DMA_PFN<<PAGE_SHIFT;
 170         if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) &&
 171             end > (MAX_DMA32_PFN<<PAGE_SHIFT))
 172                 start = MAX_DMA32_PFN<<PAGE_SHIFT;
 173         mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
 174         if (mem != MEMBLOCK_ERROR)
 175                 return __va(mem);
 176
 177         /* extend the search scope */
 178         end = max_pfn_mapped << PAGE_SHIFT;
 179         start = MAX_DMA_PFN << PAGE_SHIFT;
 180         mem = memblock_find_in_range(start, end, size, align);
 181         if (mem != MEMBLOCK_ERROR)
 182                 return __va(mem);
 183
 184         printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 185                        size, nodeid);
 186
 187         return NULL;
 188 }
 189
 190 int __init numa_add_memblk(int nid, u64 start, u64 end)
 191 {
 192         struct numa_meminfo *mi = &numa_meminfo;
 193
 194         /* ignore zero length blks */
 195         if (start == end)
 196                 return 0;
 197
 198         /* whine about and ignore invalid blks */
 199         if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
 200                 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
 201                            nid, start, end);
 202                 return 0;
 203         }
 204
 205         if (mi->nr_blks >= NR_NODE_MEMBLKS) {
 206                 pr_err("NUMA: too many memblk ranges\n");
 207                 return -EINVAL;
 208         }
 209
 210         mi->blk[mi->nr_blks].start = start;
 211         mi->blk[mi->nr_blks].end = end;
 212         mi->blk[mi->nr_blks].nid = nid;
 213         mi->nr_blks++;
 214         return 0;
 215 }
 216
 217 static void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
 218 {
 219         mi->nr_blks--;
 220         memmove(&mi->blk[idx], &mi->blk[idx + 1],
 221                 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
 222 }
 223
 224 /* Initialize bootmem allocator for a node */
 225 void __init
 226 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
 227 {
 228         unsigned long start_pfn, last_pfn, nodedata_phys;
 229         const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
 230         int nid;
 231
 232         if (!end)
 233                 return;
 234
 235         /*
 236          * Don't confuse VM with a node that doesn't have the
 237          * minimum amount of memory:
 238          */
 239         if (end && (end - start) < NODE_MIN_SIZE)
 240                 return;
 241
 242         start = roundup(start, ZONE_ALIGN);
 243
 244         printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid,
 245                start, end);
 246
 247         start_pfn = start >> PAGE_SHIFT;
 248         last_pfn = end >> PAGE_SHIFT;
 249
 250         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
 251                                            SMP_CACHE_BYTES);
 252         if (node_data[nodeid] == NULL)
 253                 return;
 254         nodedata_phys = __pa(node_data[nodeid]);
 255         memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
 256         printk(KERN_INFO "  NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
 257                 nodedata_phys + pgdat_size - 1);
 258         nid = phys_to_nid(nodedata_phys);
 259         if (nid != nodeid)
 260                 printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nodeid, nid);
 261
 262         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 263         NODE_DATA(nodeid)->node_id = nodeid;
 264         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 265         NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
 266
 267         node_set_online(nodeid);
 268 }
 269
 270 static int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
 271 {
 272         const u64 low = 0;
 273         const u64 high = (u64)max_pfn << PAGE_SHIFT;
 274         int i, j, k;
 275
 276         for (i = 0; i < mi->nr_blks; i++) {
 277                 struct numa_memblk *bi = &mi->blk[i];
 278
 279                 /* make sure all blocks are inside the limits */
 280                 bi->start = max(bi->start, low);
 281                 bi->end = min(bi->end, high);
 282
 283                 /* and there's no empty block */
 284                 if (bi->start == bi->end) {
 285                         numa_remove_memblk_from(i--, mi);
 286                         continue;
 287                 }
 288
 289                 for (j = i + 1; j < mi->nr_blks; j++) {
 290                         struct numa_memblk *bj = &mi->blk[j];
 291                         unsigned long start, end;
 292
 293                         /*
 294                          * See whether there are overlapping blocks.  Whine
 295                          * about but allow overlaps of the same nid.  They
 296                          * will be merged below.
 297                          */
 298                         if (bi->end > bj->start && bi->start < bj->end) {
 299                                 if (bi->nid != bj->nid) {
 300                                         pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
 301                                                bi->nid, bi->start, bi->end,
 302                                                bj->nid, bj->start, bj->end);
 303                                         return -EINVAL;
 304                                 }
 305                                 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
 306                                            bi->nid, bi->start, bi->end,
 307                                            bj->start, bj->end);
 308                         }
 309
 310                         /*
 311                          * Join together blocks on the same node, holes
 312                          * between which don't overlap with memory on other
 313                          * nodes.
 314                          */
 315                         if (bi->nid != bj->nid)
 316                                 continue;
 317                         start = max(min(bi->start, bj->start), low);
 318                         end = min(max(bi->end, bj->end), high);
 319                         for (k = 0; k < mi->nr_blks; k++) {
 320                                 struct numa_memblk *bk = &mi->blk[k];
 321
 322                                 if (bi->nid == bk->nid)
 323                                         continue;
 324                                 if (start < bk->end && end > bk->start)
 325                                         break;
 326                         }
 327                         if (k < mi->nr_blks)
 328                                 continue;
 329                         printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n",
 330                                bi->nid, bi->start, bi->end, bj->start, bj->end,
 331                                start, end);
 332                         bi->start = start;
 333                         bi->end = end;
 334                         numa_remove_memblk_from(j--, mi);
 335                 }
 336         }
 337
 338         for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
 339                 mi->blk[i].start = mi->blk[i].end = 0;
 340                 mi->blk[i].nid = NUMA_NO_NODE;
 341         }
 342
 343         return 0;
 344 }
 345
 346 /*
 347  * Sanity check to catch more bad NUMA configurations (they are amazingly
 348  * common).  Make sure the nodes cover all memory.
 349  */
 350 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
 351 {
 352         unsigned long numaram, e820ram;
 353         int i;
 354
 355         numaram = 0;
 356         for (i = 0; i < mi->nr_blks; i++) {
 357                 unsigned long s = mi->blk[i].start >> PAGE_SHIFT;
 358                 unsigned long e = mi->blk[i].end >> PAGE_SHIFT;
 359                 numaram += e - s;
 360                 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
 361                 if ((long)numaram < 0)
 362                         numaram = 0;
 363         }
 364
 365         e820ram = max_pfn - (memblock_x86_hole_size(0,
 366                                         max_pfn << PAGE_SHIFT) >> PAGE_SHIFT);
 367         /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
 368         if ((long)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
 369                 printk(KERN_ERR "NUMA: nodes only cover %luMB of your %luMB e820 RAM. Not used.\n",
 370                        (numaram << PAGE_SHIFT) >> 20,
 371                        (e820ram << PAGE_SHIFT) >> 20);
 372                 return false;
 373         }
 374         return true;
 375 }
 376
 377 static int __init numa_register_memblks(struct numa_meminfo *mi)
 378 {
 379         int i, j, nid;
 380
 381         /* Account for nodes with cpus and no memory */
 382         nodes_or(node_possible_map, mem_nodes_parsed, numa_nodes_parsed);
 383         if (WARN_ON(nodes_empty(node_possible_map)))
 384                 return -EINVAL;
 385
 386         memnode_shift = compute_hash_shift(mi);
 387         if (memnode_shift < 0) {
 388                 printk(KERN_ERR "NUMA: No NUMA node hash function found. Contact maintainer\n");
 389                 return -EINVAL;
 390         }
 391
 392         for (i = 0; i < mi->nr_blks; i++)
 393                 memblock_x86_register_active_regions(mi->blk[i].nid,
 394                                         mi->blk[i].start >> PAGE_SHIFT,
 395                                         mi->blk[i].end >> PAGE_SHIFT);
 396
 397         /* for out of order entries */
 398         sort_node_map();
 399         if (!numa_meminfo_cover_memory(mi))
 400                 return -EINVAL;
 401
 402         init_memory_mapping_high();
 403
 404         /*
 405          * Finally register nodes.  Do it twice in case setup_node_bootmem
 406          * missed one due to missing bootmem.
 407          */
 408         for (i = 0; i < 2; i++) {
 409                 for_each_node_mask(nid, node_possible_map) {
 410                         u64 start = (u64)max_pfn << PAGE_SHIFT;
 411                         u64 end = 0;
 412
 413                         if (node_online(nid))
 414                                 continue;
 415
 416                         for (j = 0; j < mi->nr_blks; j++) {
 417                                 if (nid != mi->blk[j].nid)
 418                                         continue;
 419                                 start = min(mi->blk[j].start, start);
 420                                 end = max(mi->blk[j].end, end);
 421                         }
 422
 423                         if (start < end)
 424                                 setup_node_bootmem(nid, start, end);
 425                 }
 426         }
 427
 428         return 0;
 429 }
 430
 431 #ifdef CONFIG_NUMA_EMU
 432 /* Numa emulation */
 433 static struct bootnode nodes[MAX_NUMNODES] __initdata;
 434 static struct bootnode physnodes[MAX_NUMNODES] __cpuinitdata;
 435 static char *cmdline __initdata;
 436
 437 void __init numa_emu_cmdline(char *str)
 438 {
 439         cmdline = str;
 440 }
 441
 442 int __init find_node_by_addr(unsigned long addr)
 443 {
 444         const struct numa_meminfo *mi = &numa_meminfo;
 445         int i;
 446
 447         for (i = 0; i < mi->nr_blks; i++) {
 448                 /*
 449                  * Find the real node that this emulated node appears on.  For
 450                  * the sake of simplicity, we only use a real node's starting
 451                  * address to determine which emulated node it appears on.
 452                  */
 453                 if (addr >= mi->blk[i].start && addr < mi->blk[i].end)
 454                         return mi->blk[i].nid;
 455         }
 456         return NUMA_NO_NODE;
 457 }
 458
 459 static int __init setup_physnodes(unsigned long start, unsigned long end)
 460 {
 461         const struct numa_meminfo *mi = &numa_meminfo;
 462         int ret = 0;
 463         int i;
 464
 465         memset(physnodes, 0, sizeof(physnodes));
 466
 467         for (i = 0; i < mi->nr_blks; i++) {
 468                 int nid = mi->blk[i].nid;
 469
 470                 if (physnodes[nid].start == physnodes[nid].end) {
 471                         physnodes[nid].start = mi->blk[i].start;
 472                         physnodes[nid].end = mi->blk[i].end;
 473                 } else {
 474                         physnodes[nid].start = min(physnodes[nid].start,
 475                                                    mi->blk[i].start);
 476                         physnodes[nid].end = max(physnodes[nid].end,
 477                                                  mi->blk[i].end);
 478                 }
 479         }
 480
 481         /*
 482          * Basic sanity checking on the physical node map: there may be errors
 483          * if the SRAT or AMD code incorrectly reported the topology or the mem=
 484          * kernel parameter is used.
 485          */
 486         for (i = 0; i < MAX_NUMNODES; i++) {
 487                 if (physnodes[i].start == physnodes[i].end)
 488                         continue;
 489                 if (physnodes[i].start > end) {
 490                         physnodes[i].end = physnodes[i].start;
 491                         continue;
 492                 }
 493                 if (physnodes[i].end < start) {
 494                         physnodes[i].start = physnodes[i].end;
 495                         continue;
 496                 }
 497                 if (physnodes[i].start < start)
 498                         physnodes[i].start = start;
 499                 if (physnodes[i].end > end)
 500                         physnodes[i].end = end;
 501                 ret++;
 502         }
 503
 504         /*
 505          * If no physical topology was detected, a single node is faked to cover
 506          * the entire address space.
 507          */
 508         if (!ret) {
 509                 physnodes[ret].start = start;
 510                 physnodes[ret].end = end;
 511                 ret = 1;
 512         }
 513         return ret;
 514 }
 515
 516 static void __init fake_physnodes(int acpi, int amd, int nr_nodes)
 517 {
 518         int i;
 519
 520         BUG_ON(acpi && amd);
 521 #ifdef CONFIG_ACPI_NUMA
 522         if (acpi)
 523                 acpi_fake_nodes(nodes, nr_nodes);
 524 #endif
 525 #ifdef CONFIG_AMD_NUMA
 526         if (amd)
 527                 amd_fake_nodes(nodes, nr_nodes);
 528 #endif
 529         if (!acpi && !amd)
 530                 for (i = 0; i < nr_cpu_ids; i++)
 531                         numa_set_node(i, 0);
 532 }
 533
 534 /*
 535  * Setups up nid to range from addr to addr + size.  If the end
 536  * boundary is greater than max_addr, then max_addr is used instead.
 537  * The return value is 0 if there is additional memory left for
 538  * allocation past addr and -1 otherwise.  addr is adjusted to be at
 539  * the end of the node.
 540  */
 541 static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr)
 542 {
 543         int ret = 0;
 544         nodes[nid].start = *addr;
 545         *addr += size;
 546         if (*addr >= max_addr) {
 547                 *addr = max_addr;
 548                 ret = -1;
 549         }
 550         nodes[nid].end = *addr;
 551         node_set(nid, node_possible_map);
 552         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 553                nodes[nid].start, nodes[nid].end,
 554                (nodes[nid].end - nodes[nid].start) >> 20);
 555         return ret;
 556 }
 557
 558 /*
 559  * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
 560  * to max_addr.  The return value is the number of nodes allocated.
 561  */
 562 static int __init split_nodes_interleave(u64 addr, u64 max_addr, int nr_nodes)
 563 {
 564         nodemask_t physnode_mask = NODE_MASK_NONE;
 565         u64 size;
 566         int big;
 567         int ret = 0;
 568         int i;
 569
 570         if (nr_nodes <= 0)
 571                 return -1;
 572         if (nr_nodes > MAX_NUMNODES) {
 573                 pr_info("numa=fake=%d too large, reducing to %d\n",
 574                         nr_nodes, MAX_NUMNODES);
 575                 nr_nodes = MAX_NUMNODES;
 576         }
 577
 578         size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / nr_nodes;
 579         /*
 580          * Calculate the number of big nodes that can be allocated as a result
 581          * of consolidating the remainder.
 582          */
 583         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
 584                 FAKE_NODE_MIN_SIZE;
 585
 586         size &= FAKE_NODE_MIN_HASH_MASK;
 587         if (!size) {
 588                 pr_err("Not enough memory for each node.  "
 589                         "NUMA emulation disabled.\n");
 590                 return -1;
 591         }
 592
 593         for (i = 0; i < MAX_NUMNODES; i++)
 594                 if (physnodes[i].start != physnodes[i].end)
 595                         node_set(i, physnode_mask);
 596
 597         /*
 598          * Continue to fill physical nodes with fake nodes until there is no
 599          * memory left on any of them.
 600          */
 601         while (nodes_weight(physnode_mask)) {
 602                 for_each_node_mask(i, physnode_mask) {
 603                         u64 end = physnodes[i].start + size;
 604                         u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
 605
 606                         if (ret < big)
 607                                 end += FAKE_NODE_MIN_SIZE;
 608
 609                         /*
 610                          * Continue to add memory to this fake node if its
 611                          * non-reserved memory is less than the per-node size.
 612                          */
 613                         while (end - physnodes[i].start -
 614                                 memblock_x86_hole_size(physnodes[i].start, end) < size) {
 615                                 end += FAKE_NODE_MIN_SIZE;
 616                                 if (end > physnodes[i].end) {
 617                                         end = physnodes[i].end;
 618                                         break;
 619                                 }
 620                         }
 621
 622                         /*
 623                          * If there won't be at least FAKE_NODE_MIN_SIZE of
 624                          * non-reserved memory in ZONE_DMA32 for the next node,
 625                          * this one must extend to the boundary.
 626                          */
 627                         if (end < dma32_end && dma32_end - end -
 628                             memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 629                                 end = dma32_end;
 630
 631                         /*
 632                          * If there won't be enough non-reserved memory for the
 633                          * next node, this one must extend to the end of the
 634                          * physical node.
 635                          */
 636                         if (physnodes[i].end - end -
 637                             memblock_x86_hole_size(end, physnodes[i].end) < size)
 638                                 end = physnodes[i].end;
 639
 640                         /*
 641                          * Avoid allocating more nodes than requested, which can
 642                          * happen as a result of rounding down each node's size
 643                          * to FAKE_NODE_MIN_SIZE.
 644                          */
 645                         if (nodes_weight(physnode_mask) + ret >= nr_nodes)
 646                                 end = physnodes[i].end;
 647
 648                         if (setup_node_range(ret++, &physnodes[i].start,
 649                                                 end - physnodes[i].start,
 650                                                 physnodes[i].end) < 0)
 651                                 node_clear(i, physnode_mask);
 652                 }
 653         }
 654         return ret;
 655 }
 656
 657 /*
 658  * Returns the end address of a node so that there is at least `size' amount of
 659  * non-reserved memory or `max_addr' is reached.
 660  */
 661 static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
 662 {
 663         u64 end = start + size;
 664
 665         while (end - start - memblock_x86_hole_size(start, end) < size) {
 666                 end += FAKE_NODE_MIN_SIZE;
 667                 if (end > max_addr) {
 668                         end = max_addr;
 669                         break;
 670                 }
 671         }
 672         return end;
 673 }
 674
 675 /*
 676  * Sets up fake nodes of `size' interleaved over physical nodes ranging from
 677  * `addr' to `max_addr'.  The return value is the number of nodes allocated.
 678  */
 679 static int __init split_nodes_size_interleave(u64 addr, u64 max_addr, u64 size)
 680 {
 681         nodemask_t physnode_mask = NODE_MASK_NONE;
 682         u64 min_size;
 683         int ret = 0;
 684         int i;
 685
 686         if (!size)
 687                 return -1;
 688         /*
 689          * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
 690          * increased accordingly if the requested size is too small.  This
 691          * creates a uniform distribution of node sizes across the entire
 692          * machine (but not necessarily over physical nodes).
 693          */
 694         min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) /
 695                                                 MAX_NUMNODES;
 696         min_size = max(min_size, FAKE_NODE_MIN_SIZE);
 697         if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
 698                 min_size = (min_size + FAKE_NODE_MIN_SIZE) &
 699                                                 FAKE_NODE_MIN_HASH_MASK;
 700         if (size < min_size) {
 701                 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
 702                         size >> 20, min_size >> 20);
 703                 size = min_size;
 704         }
 705         size &= FAKE_NODE_MIN_HASH_MASK;
 706
 707         for (i = 0; i < MAX_NUMNODES; i++)
 708                 if (physnodes[i].start != physnodes[i].end)
 709                         node_set(i, physnode_mask);
 710         /*
 711          * Fill physical nodes with fake nodes of size until there is no memory
 712          * left on any of them.
 713          */
 714         while (nodes_weight(physnode_mask)) {
 715                 for_each_node_mask(i, physnode_mask) {
 716                         u64 dma32_end = MAX_DMA32_PFN << PAGE_SHIFT;
 717                         u64 end;
 718
 719                         end = find_end_of_node(physnodes[i].start,
 720                                                 physnodes[i].end, size);
 721                         /*
 722                          * If there won't be at least FAKE_NODE_MIN_SIZE of
 723                          * non-reserved memory in ZONE_DMA32 for the next node,
 724                          * this one must extend to the boundary.
 725                          */
 726                         if (end < dma32_end && dma32_end - end -
 727                             memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 728                                 end = dma32_end;
 729
 730                         /*
 731                          * If there won't be enough non-reserved memory for the
 732                          * next node, this one must extend to the end of the
 733                          * physical node.
 734                          */
 735                         if (physnodes[i].end - end -
 736                             memblock_x86_hole_size(end, physnodes[i].end) < size)
 737                                 end = physnodes[i].end;
 738
 739                         /*
 740                          * Setup the fake node that will be allocated as bootmem
 741                          * later.  If setup_node_range() returns non-zero, there
 742                          * is no more memory available on this physical node.
 743                          */
 744                         if (setup_node_range(ret++, &physnodes[i].start,
 745                                                 end - physnodes[i].start,
 746                                                 physnodes[i].end) < 0)
 747                                 node_clear(i, physnode_mask);
 748                 }
 749         }
 750         return ret;
 751 }
 752
 753 /*
 754  * Sets up the system RAM area from start_pfn to last_pfn according to the
 755  * numa=fake command-line option.
 756  */
 757 static int __init numa_emulation(unsigned long start_pfn,
 758                         unsigned long last_pfn, int acpi, int amd)
 759 {
 760         static struct numa_meminfo ei __initdata;
 761         u64 addr = start_pfn << PAGE_SHIFT;
 762         u64 max_addr = last_pfn << PAGE_SHIFT;
 763         int num_nodes;
 764         int i;
 765
 766         /*
 767          * If the numa=fake command-line contains a 'M' or 'G', it represents
 768          * the fixed node size.  Otherwise, if it is just a single number N,
 769          * split the system RAM into N fake nodes.
 770          */
 771         if (strchr(cmdline, 'M') || strchr(cmdline, 'G')) {
 772                 u64 size;
 773
 774                 size = memparse(cmdline, &cmdline);
 775                 num_nodes = split_nodes_size_interleave(addr, max_addr, size);
 776         } else {
 777                 unsigned long n;
 778
 779                 n = simple_strtoul(cmdline, NULL, 0);
 780                 num_nodes = split_nodes_interleave(addr, max_addr, n);
 781         }
 782
 783         if (num_nodes < 0)
 784                 return num_nodes;
 785
 786         ei.nr_blks = num_nodes;
 787         for (i = 0; i < ei.nr_blks; i++) {
 788                 ei.blk[i].start = nodes[i].start;
 789                 ei.blk[i].end = nodes[i].end;
 790                 ei.blk[i].nid = i;
 791         }
 792
 793         memnode_shift = compute_hash_shift(&ei);
 794         if (memnode_shift < 0) {
 795                 memnode_shift = 0;
 796                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 797                        "disabled.\n");
 798                 return -1;
 799         }
 800
 801         /*
 802          * We need to vacate all active ranges that may have been registered for
 803          * the e820 memory map.
 804          */
 805         remove_all_active_ranges();
 806         for_each_node_mask(i, node_possible_map)
 807                 memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 808                                                 nodes[i].end >> PAGE_SHIFT);
 809         init_memory_mapping_high();
 810         for_each_node_mask(i, node_possible_map)
 811                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 812         setup_physnodes(addr, max_addr);
 813         fake_physnodes(acpi, amd, num_nodes);
 814         numa_init_array();
 815         return 0;
 816 }
 817 #endif /* CONFIG_NUMA_EMU */
 818
 819 static int dummy_numa_init(void)
 820 {
 821         printk(KERN_INFO "%s\n",
 822                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 823         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 824                0LU, max_pfn << PAGE_SHIFT);
 825
 826         node_set(0, numa_nodes_parsed);
 827         node_set(0, mem_nodes_parsed);
 828         numa_add_memblk(0, 0, (u64)max_pfn << PAGE_SHIFT);
 829
 830         return 0;
 831 }
 832
 833 void __init initmem_init(void)
 834 {
 835         int (*numa_init[])(void) = { [2] = dummy_numa_init };
 836         int i, j;
 837
 838         if (!numa_off) {
 839 #ifdef CONFIG_ACPI_NUMA
 840                 numa_init[0] = x86_acpi_numa_init;
 841 #endif
 842 #ifdef CONFIG_AMD_NUMA
 843                 numa_init[1] = amd_numa_init;
 844 #endif
 845         }
 846
 847         for (i = 0; i < ARRAY_SIZE(numa_init); i++) {
 848                 if (!numa_init[i])
 849                         continue;
 850
 851                 for (j = 0; j < MAX_LOCAL_APIC; j++)
 852                         set_apicid_to_node(j, NUMA_NO_NODE);
 853
 854                 nodes_clear(numa_nodes_parsed);
 855                 nodes_clear(mem_nodes_parsed);
 856                 nodes_clear(node_possible_map);
 857                 nodes_clear(node_online_map);
 858                 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
 859                 remove_all_active_ranges();
 860
 861                 if (numa_init[i]() < 0)
 862                         continue;
 863
 864                 if (numa_cleanup_meminfo(&numa_meminfo) < 0)
 865                         continue;
 866 #ifdef CONFIG_NUMA_EMU
 867                 setup_physnodes(0, max_pfn << PAGE_SHIFT);
 868                 if (cmdline && !numa_emulation(0, max_pfn, i == 0, i == 1))
 869                         return;
 870                 setup_physnodes(0, max_pfn << PAGE_SHIFT);
 871                 nodes_clear(node_possible_map);
 872                 nodes_clear(node_online_map);
 873 #endif
 874                 if (numa_register_memblks(&numa_meminfo) < 0)
 875                         continue;
 876
 877                 for (j = 0; j < nr_cpu_ids; j++) {
 878                         int nid = early_cpu_to_node(j);
 879
 880                         if (nid == NUMA_NO_NODE)
 881                                 continue;
 882                         if (!node_online(nid))
 883                                 numa_clear_node(j);
 884                 }
 885                 numa_init_array();
 886                 return;
 887         }
 888         BUG();
 889 }
 890
 891 unsigned long __init numa_free_all_bootmem(void)
 892 {
 893         unsigned long pages = 0;
 894         int i;
 895
 896         for_each_online_node(i)
 897                 pages += free_all_bootmem_node(NODE_DATA(i));
 898
 899         pages += free_all_memory_core_early(MAX_NUMNODES);
 900
 901         return pages;
 902 }
 903
 904 int __cpuinit numa_cpu_node(int cpu)
 905 {
 906         int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
 907
 908         if (apicid != BAD_APICID)
 909                 return __apicid_to_node[apicid];
 910         return NUMA_NO_NODE;
 911 }
 912
 913 /*
 914  * UGLINESS AHEAD: Currently, CONFIG_NUMA_EMU is 64bit only and makes use
 915  * of 64bit specific data structures.  The distinction is artificial and
 916  * should be removed.  numa_{add|remove}_cpu() are implemented in numa.c
 917  * for both 32 and 64bit when CONFIG_NUMA_EMU is disabled but here when
 918  * enabled.
 919  *
 920  * NUMA emulation is planned to be made generic and the following and other
 921  * related code should be moved to numa.c.
 922  */
 923 #ifdef CONFIG_NUMA_EMU
 924 # ifndef CONFIG_DEBUG_PER_CPU_MAPS
 925 void __cpuinit numa_add_cpu(int cpu)
 926 {
 927         unsigned long addr;
 928         int physnid, nid;
 929
 930         nid = numa_cpu_node(cpu);
 931         if (nid == NUMA_NO_NODE)
 932                 nid = early_cpu_to_node(cpu);
 933         BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
 934
 935         /*
 936          * Use the starting address of the emulated node to find which physical
 937          * node it is allocated on.
 938          */
 939         addr = node_start_pfn(nid) << PAGE_SHIFT;
 940         for (physnid = 0; physnid < MAX_NUMNODES; physnid++)
 941                 if (addr >= physnodes[physnid].start &&
 942                     addr < physnodes[physnid].end)
 943                         break;
 944
 945         /*
 946          * Map the cpu to each emulated node that is allocated on the physical
 947          * node of the cpu's apic id.
 948          */
 949         for_each_online_node(nid) {
 950                 addr = node_start_pfn(nid) << PAGE_SHIFT;
 951                 if (addr >= physnodes[physnid].start &&
 952                     addr < physnodes[physnid].end)
 953                         cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
 954         }
 955 }
 956
 957 void __cpuinit numa_remove_cpu(int cpu)
 958 {
 959         int i;
 960
 961         for_each_online_node(i)
 962                 cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
 963 }
 964 # else  /* !CONFIG_DEBUG_PER_CPU_MAPS */
 965 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 966 {
 967         int node = early_cpu_to_node(cpu);
 968         struct cpumask *mask;
 969         int i;
 970
 971         if (node == NUMA_NO_NODE) {
 972                 /* early_cpu_to_node() already emits a warning and trace */
 973                 return;
 974         }
 975         for_each_online_node(i) {
 976                 unsigned long addr;
 977
 978                 addr = node_start_pfn(i) << PAGE_SHIFT;
 979                 if (addr < physnodes[node].start ||
 980                                         addr >= physnodes[node].end)
 981                         continue;
 982                 mask = debug_cpumask_set_cpu(cpu, enable);
 983                 if (!mask)
 984                         return;
 985
 986                 if (enable)
 987                         cpumask_set_cpu(cpu, mask);
 988                 else
 989                         cpumask_clear_cpu(cpu, mask);
 990         }
 991 }
 992
 993 void __cpuinit numa_add_cpu(int cpu)
 994 {
 995         numa_set_cpumask(cpu, 1);
 996 }
 997
 998 void __cpuinit numa_remove_cpu(int cpu)
 999 {
1000         numa_set_cpumask(cpu, 0);
1001 }
1002 # endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
1003 #endif  /* CONFIG_NUMA_EMU */