arch/x86/mm/numa_32.c

   1 /*
   2  * Written by: Patricia Gaughen <gone@us.ibm.com>, IBM Corporation
   3  * August 2002: added remote node KVA remap - Martin J. Bligh
   4  *
   5  * Copyright (C) 2002, IBM Corp.
   6  *
   7  * All rights reserved.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful, but
  15  * WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  17  * NON INFRINGEMENT.  See the GNU General Public License for more
  18  * details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/mm.h>
  26 #include <linux/bootmem.h>
  27 #include <linux/memblock.h>
  28 #include <linux/mmzone.h>
  29 #include <linux/highmem.h>
  30 #include <linux/initrd.h>
  31 #include <linux/nodemask.h>
  32 #include <linux/module.h>
  33 #include <linux/kexec.h>
  34 #include <linux/pfn.h>
  35 #include <linux/swap.h>
  36 #include <linux/acpi.h>
  37
  38 #include <asm/e820.h>
  39 #include <asm/setup.h>
  40 #include <asm/mmzone.h>
  41 #include <asm/bios_ebda.h>
  42 #include <asm/proto.h>
  43
  44 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  45 EXPORT_SYMBOL(node_data);
  46
  47 /*
  48  * numa interface - we expect the numa architecture specific code to have
  49  *                  populated the following initialisation.
  50  *
  51  * 1) node_online_map  - the map of all nodes configured (online) in the system
  52  * 2) node_start_pfn   - the starting page frame number for a node
  53  * 3) node_end_pfn     - the ending page fram number for a node
  54  */
  55 unsigned long node_start_pfn[MAX_NUMNODES] __read_mostly;
  56 unsigned long node_end_pfn[MAX_NUMNODES] __read_mostly;
  57
  58
  59 #ifdef CONFIG_DISCONTIGMEM
  60 /*
  61  * 4) physnode_map     - the mapping between a pfn and owning node
  62  * physnode_map keeps track of the physical memory layout of a generic
  63  * numa node on a 64Mb break (each element of the array will
  64  * represent 64Mb of memory and will be marked by the node id.  so,
  65  * if the first gig is on node 0, and the second gig is on node 1
  66  * physnode_map will contain:
  67  *
  68  *     physnode_map[0-15] = 0;
  69  *     physnode_map[16-31] = 1;
  70  *     physnode_map[32- ] = -1;
  71  */
  72 s8 physnode_map[MAX_ELEMENTS] __read_mostly = { [0 ... (MAX_ELEMENTS - 1)] = -1};
  73 EXPORT_SYMBOL(physnode_map);
  74
  75 void memory_present(int nid, unsigned long start, unsigned long end)
  76 {
  77         unsigned long pfn;
  78
  79         printk(KERN_INFO "Node: %d, start_pfn: %lx, end_pfn: %lx\n",
  80                         nid, start, end);
  81         printk(KERN_DEBUG "  Setting physnode_map array to node %d for pfns:\n", nid);
  82         printk(KERN_DEBUG "  ");
  83         for (pfn = start; pfn < end; pfn += PAGES_PER_ELEMENT) {
  84                 physnode_map[pfn / PAGES_PER_ELEMENT] = nid;
  85                 printk(KERN_CONT "%lx ", pfn);
  86         }
  87         printk(KERN_CONT "\n");
  88 }
  89
  90 unsigned long node_memmap_size_bytes(int nid, unsigned long start_pfn,
  91                                               unsigned long end_pfn)
  92 {
  93         unsigned long nr_pages = end_pfn - start_pfn;
  94
  95         if (!nr_pages)
  96                 return 0;
  97
  98         return (nr_pages + 1) * sizeof(struct page);
  99 }
 100 #endif
 101
 102 extern unsigned long find_max_low_pfn(void);
 103 extern unsigned long highend_pfn, highstart_pfn;
 104
 105 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
 106
 107 unsigned long node_remap_size[MAX_NUMNODES];
 108 static void *node_remap_start_vaddr[MAX_NUMNODES];
 109 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
 110
 111 static unsigned long kva_start_pfn;
 112 static unsigned long kva_pages;
 113
 114 int __cpuinit numa_cpu_node(int cpu)
 115 {
 116         return apic->x86_32_numa_cpu_node(cpu);
 117 }
 118
 119 /*
 120  * FLAT - support for basic PC memory model with discontig enabled, essentially
 121  *        a single node with all available processors in it with a flat
 122  *        memory map.
 123  */
 124 int __init get_memcfg_numa_flat(void)
 125 {
 126         printk(KERN_DEBUG "NUMA - single node, flat memory mode\n");
 127
 128         node_start_pfn[0] = 0;
 129         node_end_pfn[0] = max_pfn;
 130         memblock_x86_register_active_regions(0, 0, max_pfn);
 131         memory_present(0, 0, max_pfn);
 132         node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
 133
 134         /* Indicate there is one node available. */
 135         nodes_clear(node_online_map);
 136         node_set_online(0);
 137         return 1;
 138 }
 139
 140 /*
 141  * Find the highest page frame number we have available for the node
 142  */
 143 static void __init propagate_e820_map_node(int nid)
 144 {
 145         if (node_end_pfn[nid] > max_pfn)
 146                 node_end_pfn[nid] = max_pfn;
 147         /*
 148          * if a user has given mem=XXXX, then we need to make sure
 149          * that the node _starts_ before that, too, not just ends
 150          */
 151         if (node_start_pfn[nid] > max_pfn)
 152                 node_start_pfn[nid] = max_pfn;
 153         BUG_ON(node_start_pfn[nid] > node_end_pfn[nid]);
 154 }
 155
 156 /*
 157  * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
 158  * method.  For node zero take this from the bottom of memory, for
 159  * subsequent nodes place them at node_remap_start_vaddr which contains
 160  * node local data in physically node local memory.  See setup_memory()
 161  * for details.
 162  */
 163 static void __init allocate_pgdat(int nid)
 164 {
 165         char buf[16];
 166
 167         if (node_has_online_mem(nid) && node_remap_start_vaddr[nid])
 168                 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
 169         else {
 170                 unsigned long pgdat_phys;
 171                 pgdat_phys = memblock_find_in_range(min_low_pfn<<PAGE_SHIFT,
 172                                  max_pfn_mapped<<PAGE_SHIFT,
 173                                  sizeof(pg_data_t),
 174                                  PAGE_SIZE);
 175                 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
 176                 memset(buf, 0, sizeof(buf));
 177                 sprintf(buf, "NODE_DATA %d",  nid);
 178                 memblock_x86_reserve_range(pgdat_phys, pgdat_phys + sizeof(pg_data_t), buf);
 179         }
 180         printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
 181                 nid, (unsigned long)NODE_DATA(nid));
 182 }
 183
 184 /*
 185  * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
 186  * virtual address space (KVA) is reserved and portions of nodes are mapped
 187  * using it. This is to allow node-local memory to be allocated for
 188  * structures that would normally require ZONE_NORMAL. The memory is
 189  * allocated with alloc_remap() and callers should be prepared to allocate
 190  * from the bootmem allocator instead.
 191  */
 192 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
 193 static void *node_remap_end_vaddr[MAX_NUMNODES];
 194 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
 195 static unsigned long node_remap_offset[MAX_NUMNODES];
 196
 197 void *alloc_remap(int nid, unsigned long size)
 198 {
 199         void *allocation = node_remap_alloc_vaddr[nid];
 200
 201         size = ALIGN(size, L1_CACHE_BYTES);
 202
 203         if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
 204                 return NULL;
 205
 206         node_remap_alloc_vaddr[nid] += size;
 207         memset(allocation, 0, size);
 208
 209         return allocation;
 210 }
 211
 212 static void __init remap_numa_kva(void)
 213 {
 214         void *vaddr;
 215         unsigned long pfn;
 216         int node;
 217
 218         for_each_online_node(node) {
 219                 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
 220                 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
 221                         vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
 222                         printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
 223                                 (unsigned long)vaddr,
 224                                 node_remap_start_pfn[node] + pfn);
 225                         set_pmd_pfn((ulong) vaddr,
 226                                 node_remap_start_pfn[node] + pfn,
 227                                 PAGE_KERNEL_LARGE);
 228                 }
 229         }
 230 }
 231
 232 #ifdef CONFIG_HIBERNATION
 233 /**
 234  * resume_map_numa_kva - add KVA mapping to the temporary page tables created
 235  *                       during resume from hibernation
 236  * @pgd_base - temporary resume page directory
 237  */
 238 void resume_map_numa_kva(pgd_t *pgd_base)
 239 {
 240         int node;
 241
 242         for_each_online_node(node) {
 243                 unsigned long start_va, start_pfn, size, pfn;
 244
 245                 start_va = (unsigned long)node_remap_start_vaddr[node];
 246                 start_pfn = node_remap_start_pfn[node];
 247                 size = node_remap_size[node];
 248
 249                 printk(KERN_DEBUG "%s: node %d\n", __func__, node);
 250
 251                 for (pfn = 0; pfn < size; pfn += PTRS_PER_PTE) {
 252                         unsigned long vaddr = start_va + (pfn << PAGE_SHIFT);
 253                         pgd_t *pgd = pgd_base + pgd_index(vaddr);
 254                         pud_t *pud = pud_offset(pgd, vaddr);
 255                         pmd_t *pmd = pmd_offset(pud, vaddr);
 256
 257                         set_pmd(pmd, pfn_pmd(start_pfn + pfn,
 258                                                 PAGE_KERNEL_LARGE_EXEC));
 259
 260                         printk(KERN_DEBUG "%s: %08lx -> pfn %08lx\n",
 261                                 __func__, vaddr, start_pfn + pfn);
 262                 }
 263         }
 264 }
 265 #endif
 266
 267 static __init unsigned long calculate_numa_remap_pages(void)
 268 {
 269         int nid;
 270         unsigned long size, reserve_pages = 0;
 271
 272         for_each_online_node(nid) {
 273                 u64 node_kva_target;
 274                 u64 node_kva_final;
 275
 276                 /*
 277                  * The acpi/srat node info can show hot-add memroy zones
 278                  * where memory could be added but not currently present.
 279                  */
 280                 printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
 281                         nid, node_start_pfn[nid], node_end_pfn[nid]);
 282                 if (node_start_pfn[nid] > max_pfn)
 283                         continue;
 284                 if (!node_end_pfn[nid])
 285                         continue;
 286                 if (node_end_pfn[nid] > max_pfn)
 287                         node_end_pfn[nid] = max_pfn;
 288
 289                 /* ensure the remap includes space for the pgdat. */
 290                 size = node_remap_size[nid] + sizeof(pg_data_t);
 291
 292                 /* convert size to large (pmd size) pages, rounding up */
 293                 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
 294                 /* now the roundup is correct, convert to PAGE_SIZE pages */
 295                 size = size * PTRS_PER_PTE;
 296
 297                 node_kva_target = round_down(node_end_pfn[nid] - size,
 298                                                  PTRS_PER_PTE);
 299                 node_kva_target <<= PAGE_SHIFT;
 300                 do {
 301                         node_kva_final = memblock_find_in_range(node_kva_target,
 302                                         ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
 303                                                 ((u64)size)<<PAGE_SHIFT,
 304                                                 LARGE_PAGE_BYTES);
 305                         node_kva_target -= LARGE_PAGE_BYTES;
 306                 } while (node_kva_final == MEMBLOCK_ERROR &&
 307                          (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
 308
 309                 if (node_kva_final == MEMBLOCK_ERROR)
 310                         panic("Can not get kva ram\n");
 311
 312                 node_remap_size[nid] = size;
 313                 node_remap_offset[nid] = reserve_pages;
 314                 reserve_pages += size;
 315                 printk(KERN_DEBUG "Reserving %ld pages of KVA for lmem_map of"
 316                                   " node %d at %llx\n",
 317                                 size, nid, node_kva_final>>PAGE_SHIFT);
 318
 319                 /*
 320                  *  prevent kva address below max_low_pfn want it on system
 321                  *  with less memory later.
 322                  *  layout will be: KVA address , KVA RAM
 323                  *
 324                  *  we are supposed to only record the one less then max_low_pfn
 325                  *  but we could have some hole in high memory, and it will only
 326                  *  check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
 327                  *  to use it as free.
 328                  *  So memblock_x86_reserve_range here, hope we don't run out of that array
 329                  */
 330                 memblock_x86_reserve_range(node_kva_final,
 331                               node_kva_final+(((u64)size)<<PAGE_SHIFT),
 332                               "KVA RAM");
 333
 334                 node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
 335         }
 336         printk(KERN_INFO "Reserving total of %lx pages for numa KVA remap\n",
 337                         reserve_pages);
 338         return reserve_pages;
 339 }
 340
 341 static void init_remap_allocator(int nid)
 342 {
 343         node_remap_start_vaddr[nid] = pfn_to_kaddr(
 344                         kva_start_pfn + node_remap_offset[nid]);
 345         node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
 346                 (node_remap_size[nid] * PAGE_SIZE);
 347         node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
 348                 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
 349
 350         printk(KERN_DEBUG "node %d will remap to vaddr %08lx - %08lx\n", nid,
 351                 (ulong) node_remap_start_vaddr[nid],
 352                 (ulong) node_remap_end_vaddr[nid]);
 353 }
 354
 355 void __init initmem_init(void)
 356 {
 357         int nid;
 358         long kva_target_pfn;
 359
 360         /*
 361          * When mapping a NUMA machine we allocate the node_mem_map arrays
 362          * from node local memory.  They are then mapped directly into KVA
 363          * between zone normal and vmalloc space.  Calculate the size of
 364          * this space and use it to adjust the boundary between ZONE_NORMAL
 365          * and ZONE_HIGHMEM.
 366          */
 367
 368         get_memcfg_numa();
 369         numa_init_array();
 370
 371         kva_pages = roundup(calculate_numa_remap_pages(), PTRS_PER_PTE);
 372
 373         kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
 374         do {
 375                 kva_start_pfn = memblock_find_in_range(kva_target_pfn<<PAGE_SHIFT,
 376                                         max_low_pfn<<PAGE_SHIFT,
 377                                         kva_pages<<PAGE_SHIFT,
 378                                         PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
 379                 kva_target_pfn -= PTRS_PER_PTE;
 380         } while (kva_start_pfn == MEMBLOCK_ERROR && kva_target_pfn > min_low_pfn);
 381
 382         if (kva_start_pfn == MEMBLOCK_ERROR)
 383                 panic("Can not get kva space\n");
 384
 385         printk(KERN_INFO "kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
 386                 kva_start_pfn, max_low_pfn);
 387         printk(KERN_INFO "max_pfn = %lx\n", max_pfn);
 388
 389         /* avoid clash with initrd */
 390         memblock_x86_reserve_range(kva_start_pfn<<PAGE_SHIFT,
 391                       (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
 392                      "KVA PG");
 393 #ifdef CONFIG_HIGHMEM
 394         highstart_pfn = highend_pfn = max_pfn;
 395         if (max_pfn > max_low_pfn)
 396                 highstart_pfn = max_low_pfn;
 397         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 398                pages_to_mb(highend_pfn - highstart_pfn));
 399         num_physpages = highend_pfn;
 400         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 401 #else
 402         num_physpages = max_low_pfn;
 403         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 404 #endif
 405         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 406                         pages_to_mb(max_low_pfn));
 407         printk(KERN_DEBUG "max_low_pfn = %lx, highstart_pfn = %lx\n",
 408                         max_low_pfn, highstart_pfn);
 409
 410         printk(KERN_DEBUG "Low memory ends at vaddr %08lx\n",
 411                         (ulong) pfn_to_kaddr(max_low_pfn));
 412         for_each_online_node(nid) {
 413                 init_remap_allocator(nid);
 414
 415                 allocate_pgdat(nid);
 416         }
 417         remap_numa_kva();
 418
 419         printk(KERN_DEBUG "High memory starts at vaddr %08lx\n",
 420                         (ulong) pfn_to_kaddr(highstart_pfn));
 421         for_each_online_node(nid)
 422                 propagate_e820_map_node(nid);
 423
 424         for_each_online_node(nid) {
 425                 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 426                 NODE_DATA(nid)->node_id = nid;
 427         }
 428
 429         setup_bootmem_allocator();
 430 }
 431
 432 #ifdef CONFIG_MEMORY_HOTPLUG
 433 static int paddr_to_nid(u64 addr)
 434 {
 435         int nid;
 436         unsigned long pfn = PFN_DOWN(addr);
 437
 438         for_each_node(nid)
 439                 if (node_start_pfn[nid] <= pfn &&
 440                     pfn < node_end_pfn[nid])
 441                         return nid;
 442
 443         return -1;
 444 }
 445
 446 /*
 447  * This function is used to ask node id BEFORE memmap and mem_section's
 448  * initialization (pfn_to_nid() can't be used yet).
 449  * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
 450  */
 451 int memory_add_physaddr_to_nid(u64 addr)
 452 {
 453         int nid = paddr_to_nid(addr);
 454         return (nid >= 0) ? nid : 0;
 455 }
 456
 457 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 458 #endif
 459