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  * FLAT - support for basic PC memory model with discontig enabled, essentially
 115  *        a single node with all available processors in it with a flat
 116  *        memory map.
 117  */
 118 int __init get_memcfg_numa_flat(void)
 119 {
 120         printk(KERN_DEBUG "NUMA - single node, flat memory mode\n");
 121
 122         node_start_pfn[0] = 0;
 123         node_end_pfn[0] = max_pfn;
 124         memblock_x86_register_active_regions(0, 0, max_pfn);
 125         memory_present(0, 0, max_pfn);
 126         node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
 127
 128         /* Indicate there is one node available. */
 129         nodes_clear(node_online_map);
 130         node_set_online(0);
 131         return 1;
 132 }
 133
 134 /*
 135  * Find the highest page frame number we have available for the node
 136  */
 137 static void __init propagate_e820_map_node(int nid)
 138 {
 139         if (node_end_pfn[nid] > max_pfn)
 140                 node_end_pfn[nid] = max_pfn;
 141         /*
 142          * if a user has given mem=XXXX, then we need to make sure
 143          * that the node _starts_ before that, too, not just ends
 144          */
 145         if (node_start_pfn[nid] > max_pfn)
 146                 node_start_pfn[nid] = max_pfn;
 147         BUG_ON(node_start_pfn[nid] > node_end_pfn[nid]);
 148 }
 149
 150 /*
 151  * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
 152  * method.  For node zero take this from the bottom of memory, for
 153  * subsequent nodes place them at node_remap_start_vaddr which contains
 154  * node local data in physically node local memory.  See setup_memory()
 155  * for details.
 156  */
 157 static void __init allocate_pgdat(int nid)
 158 {
 159         char buf[16];
 160
 161         if (node_has_online_mem(nid) && node_remap_start_vaddr[nid])
 162                 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
 163         else {
 164                 unsigned long pgdat_phys;
 165                 pgdat_phys = memblock_find_in_range(min_low_pfn<<PAGE_SHIFT,
 166                                  max_pfn_mapped<<PAGE_SHIFT,
 167                                  sizeof(pg_data_t),
 168                                  PAGE_SIZE);
 169                 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
 170                 memset(buf, 0, sizeof(buf));
 171                 sprintf(buf, "NODE_DATA %d",  nid);
 172                 memblock_x86_reserve_range(pgdat_phys, pgdat_phys + sizeof(pg_data_t), buf);
 173         }
 174         printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
 175                 nid, (unsigned long)NODE_DATA(nid));
 176 }
 177
 178 /*
 179  * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
 180  * virtual address space (KVA) is reserved and portions of nodes are mapped
 181  * using it. This is to allow node-local memory to be allocated for
 182  * structures that would normally require ZONE_NORMAL. The memory is
 183  * allocated with alloc_remap() and callers should be prepared to allocate
 184  * from the bootmem allocator instead.
 185  */
 186 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
 187 static void *node_remap_end_vaddr[MAX_NUMNODES];
 188 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
 189 static unsigned long node_remap_offset[MAX_NUMNODES];
 190
 191 void *alloc_remap(int nid, unsigned long size)
 192 {
 193         void *allocation = node_remap_alloc_vaddr[nid];
 194
 195         size = ALIGN(size, L1_CACHE_BYTES);
 196
 197         if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
 198                 return NULL;
 199
 200         node_remap_alloc_vaddr[nid] += size;
 201         memset(allocation, 0, size);
 202
 203         return allocation;
 204 }
 205
 206 static void __init remap_numa_kva(void)
 207 {
 208         void *vaddr;
 209         unsigned long pfn;
 210         int node;
 211
 212         for_each_online_node(node) {
 213                 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
 214                 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
 215                         vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
 216                         printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
 217                                 (unsigned long)vaddr,
 218                                 node_remap_start_pfn[node] + pfn);
 219                         set_pmd_pfn((ulong) vaddr,
 220                                 node_remap_start_pfn[node] + pfn,
 221                                 PAGE_KERNEL_LARGE);
 222                 }
 223         }
 224 }
 225
 226 #ifdef CONFIG_HIBERNATION
 227 /**
 228  * resume_map_numa_kva - add KVA mapping to the temporary page tables created
 229  *                       during resume from hibernation
 230  * @pgd_base - temporary resume page directory
 231  */
 232 void resume_map_numa_kva(pgd_t *pgd_base)
 233 {
 234         int node;
 235
 236         for_each_online_node(node) {
 237                 unsigned long start_va, start_pfn, size, pfn;
 238
 239                 start_va = (unsigned long)node_remap_start_vaddr[node];
 240                 start_pfn = node_remap_start_pfn[node];
 241                 size = node_remap_size[node];
 242
 243                 printk(KERN_DEBUG "%s: node %d\n", __func__, node);
 244
 245                 for (pfn = 0; pfn < size; pfn += PTRS_PER_PTE) {
 246                         unsigned long vaddr = start_va + (pfn << PAGE_SHIFT);
 247                         pgd_t *pgd = pgd_base + pgd_index(vaddr);
 248                         pud_t *pud = pud_offset(pgd, vaddr);
 249                         pmd_t *pmd = pmd_offset(pud, vaddr);
 250
 251                         set_pmd(pmd, pfn_pmd(start_pfn + pfn,
 252                                                 PAGE_KERNEL_LARGE_EXEC));
 253
 254                         printk(KERN_DEBUG "%s: %08lx -> pfn %08lx\n",
 255                                 __func__, vaddr, start_pfn + pfn);
 256                 }
 257         }
 258 }
 259 #endif
 260
 261 static __init unsigned long calculate_numa_remap_pages(void)
 262 {
 263         int nid;
 264         unsigned long size, reserve_pages = 0;
 265
 266         for_each_online_node(nid) {
 267                 u64 node_kva_target;
 268                 u64 node_kva_final;
 269
 270                 /*
 271                  * The acpi/srat node info can show hot-add memroy zones
 272                  * where memory could be added but not currently present.
 273                  */
 274                 printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
 275                         nid, node_start_pfn[nid], node_end_pfn[nid]);
 276                 if (node_start_pfn[nid] > max_pfn)
 277                         continue;
 278                 if (!node_end_pfn[nid])
 279                         continue;
 280                 if (node_end_pfn[nid] > max_pfn)
 281                         node_end_pfn[nid] = max_pfn;
 282
 283                 /* ensure the remap includes space for the pgdat. */
 284                 size = node_remap_size[nid] + sizeof(pg_data_t);
 285
 286                 /* convert size to large (pmd size) pages, rounding up */
 287                 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
 288                 /* now the roundup is correct, convert to PAGE_SIZE pages */
 289                 size = size * PTRS_PER_PTE;
 290
 291                 node_kva_target = round_down(node_end_pfn[nid] - size,
 292                                                  PTRS_PER_PTE);
 293                 node_kva_target <<= PAGE_SHIFT;
 294                 do {
 295                         node_kva_final = memblock_find_in_range(node_kva_target,
 296                                         ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
 297                                                 ((u64)size)<<PAGE_SHIFT,
 298                                                 LARGE_PAGE_BYTES);
 299                         node_kva_target -= LARGE_PAGE_BYTES;
 300                 } while (node_kva_final == MEMBLOCK_ERROR &&
 301                          (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
 302
 303                 if (node_kva_final == MEMBLOCK_ERROR)
 304                         panic("Can not get kva ram\n");
 305
 306                 node_remap_size[nid] = size;
 307                 node_remap_offset[nid] = reserve_pages;
 308                 reserve_pages += size;
 309                 printk(KERN_DEBUG "Reserving %ld pages of KVA for lmem_map of"
 310                                   " node %d at %llx\n",
 311                                 size, nid, node_kva_final>>PAGE_SHIFT);
 312
 313                 /*
 314                  *  prevent kva address below max_low_pfn want it on system
 315                  *  with less memory later.
 316                  *  layout will be: KVA address , KVA RAM
 317                  *
 318                  *  we are supposed to only record the one less then max_low_pfn
 319                  *  but we could have some hole in high memory, and it will only
 320                  *  check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
 321                  *  to use it as free.
 322                  *  So memblock_x86_reserve_range here, hope we don't run out of that array
 323                  */
 324                 memblock_x86_reserve_range(node_kva_final,
 325                               node_kva_final+(((u64)size)<<PAGE_SHIFT),
 326                               "KVA RAM");
 327
 328                 node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
 329         }
 330         printk(KERN_INFO "Reserving total of %lx pages for numa KVA remap\n",
 331                         reserve_pages);
 332         return reserve_pages;
 333 }
 334
 335 static void init_remap_allocator(int nid)
 336 {
 337         node_remap_start_vaddr[nid] = pfn_to_kaddr(
 338                         kva_start_pfn + node_remap_offset[nid]);
 339         node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
 340                 (node_remap_size[nid] * PAGE_SIZE);
 341         node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
 342                 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
 343
 344         printk(KERN_DEBUG "node %d will remap to vaddr %08lx - %08lx\n", nid,
 345                 (ulong) node_remap_start_vaddr[nid],
 346                 (ulong) node_remap_end_vaddr[nid]);
 347 }
 348
 349 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn,
 350                                 int acpi, int k8)
 351 {
 352         int nid;
 353         long kva_target_pfn;
 354
 355         /*
 356          * When mapping a NUMA machine we allocate the node_mem_map arrays
 357          * from node local memory.  They are then mapped directly into KVA
 358          * between zone normal and vmalloc space.  Calculate the size of
 359          * this space and use it to adjust the boundary between ZONE_NORMAL
 360          * and ZONE_HIGHMEM.
 361          */
 362
 363         get_memcfg_numa();
 364
 365         kva_pages = roundup(calculate_numa_remap_pages(), PTRS_PER_PTE);
 366
 367         kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
 368         do {
 369                 kva_start_pfn = memblock_find_in_range(kva_target_pfn<<PAGE_SHIFT,
 370                                         max_low_pfn<<PAGE_SHIFT,
 371                                         kva_pages<<PAGE_SHIFT,
 372                                         PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
 373                 kva_target_pfn -= PTRS_PER_PTE;
 374         } while (kva_start_pfn == MEMBLOCK_ERROR && kva_target_pfn > min_low_pfn);
 375
 376         if (kva_start_pfn == MEMBLOCK_ERROR)
 377                 panic("Can not get kva space\n");
 378
 379         printk(KERN_INFO "kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
 380                 kva_start_pfn, max_low_pfn);
 381         printk(KERN_INFO "max_pfn = %lx\n", max_pfn);
 382
 383         /* avoid clash with initrd */
 384         memblock_x86_reserve_range(kva_start_pfn<<PAGE_SHIFT,
 385                       (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
 386                      "KVA PG");
 387 #ifdef CONFIG_HIGHMEM
 388         highstart_pfn = highend_pfn = max_pfn;
 389         if (max_pfn > max_low_pfn)
 390                 highstart_pfn = max_low_pfn;
 391         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 392                pages_to_mb(highend_pfn - highstart_pfn));
 393         num_physpages = highend_pfn;
 394         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 395 #else
 396         num_physpages = max_low_pfn;
 397         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 398 #endif
 399         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 400                         pages_to_mb(max_low_pfn));
 401         printk(KERN_DEBUG "max_low_pfn = %lx, highstart_pfn = %lx\n",
 402                         max_low_pfn, highstart_pfn);
 403
 404         printk(KERN_DEBUG "Low memory ends at vaddr %08lx\n",
 405                         (ulong) pfn_to_kaddr(max_low_pfn));
 406         for_each_online_node(nid) {
 407                 init_remap_allocator(nid);
 408
 409                 allocate_pgdat(nid);
 410         }
 411         remap_numa_kva();
 412
 413         printk(KERN_DEBUG "High memory starts at vaddr %08lx\n",
 414                         (ulong) pfn_to_kaddr(highstart_pfn));
 415         for_each_online_node(nid)
 416                 propagate_e820_map_node(nid);
 417
 418         for_each_online_node(nid) {
 419                 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 420                 NODE_DATA(nid)->node_id = nid;
 421         }
 422
 423         setup_bootmem_allocator();
 424 }
 425
 426 #ifdef CONFIG_MEMORY_HOTPLUG
 427 static int paddr_to_nid(u64 addr)
 428 {
 429         int nid;
 430         unsigned long pfn = PFN_DOWN(addr);
 431
 432         for_each_node(nid)
 433                 if (node_start_pfn[nid] <= pfn &&
 434                     pfn < node_end_pfn[nid])
 435                         return nid;
 436
 437         return -1;
 438 }
 439
 440 /*
 441  * This function is used to ask node id BEFORE memmap and mem_section's
 442  * initialization (pfn_to_nid() can't be used yet).
 443  * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
 444  */
 445 int memory_add_physaddr_to_nid(u64 addr)
 446 {
 447         int nid = paddr_to_nid(addr);
 448         return (nid >= 0) ? nid : 0;
 449 }
 450
 451 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 452 #endif
 453