mm/sparse.c

   1 /*
   2  * sparse memory mappings.
   3  */
   4 #include <linux/mm.h>
   5 #include <linux/mmzone.h>
   6 #include <linux/bootmem.h>
   7 #include <linux/highmem.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/vmalloc.h>
  11 #include <asm/dma.h>
  12
  13 /*
  14  * Permanent SPARSEMEM data:
  15  *
  16  * 1) mem_section       - memory sections, mem_map's for valid memory
  17  */
  18 #ifdef CONFIG_SPARSEMEM_EXTREME
  19 struct mem_section *mem_section[NR_SECTION_ROOTS]
  20         ____cacheline_internodealigned_in_smp;
  21 #else
  22 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  23         ____cacheline_internodealigned_in_smp;
  24 #endif
  25 EXPORT_SYMBOL(mem_section);
  26
  27 #ifdef NODE_NOT_IN_PAGE_FLAGS
  28 /*
  29  * If we did not store the node number in the page then we have to
  30  * do a lookup in the section_to_node_table in order to find which
  31  * node the page belongs to.
  32  */
  33 #if MAX_NUMNODES <= 256
  34 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  35 #else
  36 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  37 #endif
  38
  39 int page_to_nid(struct page *page)
  40 {
  41         return section_to_node_table[page_to_section(page)];
  42 }
  43 EXPORT_SYMBOL(page_to_nid);
  44 #endif
  45
  46 #ifdef CONFIG_SPARSEMEM_EXTREME
  47 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  48 {
  49         struct mem_section *section = NULL;
  50         unsigned long array_size = SECTIONS_PER_ROOT *
  51                                    sizeof(struct mem_section);
  52
  53         if (slab_is_available())
  54                 section = kmalloc_node(array_size, GFP_KERNEL, nid);
  55         else
  56                 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  57
  58         if (section)
  59                 memset(section, 0, array_size);
  60
  61         return section;
  62 }
  63
  64 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  65 {
  66         static DEFINE_SPINLOCK(index_init_lock);
  67         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  68         struct mem_section *section;
  69         int ret = 0;
  70
  71 #ifdef NODE_NOT_IN_PAGE_FLAGS
  72         section_to_node_table[section_nr] = nid;
  73 #endif
  74
  75         if (mem_section[root])
  76                 return -EEXIST;
  77
  78         section = sparse_index_alloc(nid);
  79         /*
  80          * This lock keeps two different sections from
  81          * reallocating for the same index
  82          */
  83         spin_lock(&index_init_lock);
  84
  85         if (mem_section[root]) {
  86                 ret = -EEXIST;
  87                 goto out;
  88         }
  89
  90         mem_section[root] = section;
  91 out:
  92         spin_unlock(&index_init_lock);
  93         return ret;
  94 }
  95 #else /* !SPARSEMEM_EXTREME */
  96 static inline int sparse_index_init(unsigned long section_nr, int nid)
  97 {
  98         return 0;
  99 }
 100 #endif
 101
 102 /*
 103  * Although written for the SPARSEMEM_EXTREME case, this happens
 104  * to also work for the flat array case becase
 105  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 106  */
 107 int __section_nr(struct mem_section* ms)
 108 {
 109         unsigned long root_nr;
 110         struct mem_section* root;
 111
 112         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 113                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 114                 if (!root)
 115                         continue;
 116
 117                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 118                      break;
 119         }
 120
 121         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 122 }
 123
 124 /*
 125  * During early boot, before section_mem_map is used for an actual
 126  * mem_map, we use section_mem_map to store the section's NUMA
 127  * node.  This keeps us from having to use another data structure.  The
 128  * node information is cleared just before we store the real mem_map.
 129  */
 130 static inline unsigned long sparse_encode_early_nid(int nid)
 131 {
 132         return (nid << SECTION_NID_SHIFT);
 133 }
 134
 135 static inline int sparse_early_nid(struct mem_section *section)
 136 {
 137         return (section->section_mem_map >> SECTION_NID_SHIFT);
 138 }
 139
 140 /* Record a memory area against a node. */
 141 void __init memory_present(int nid, unsigned long start, unsigned long end)
 142 {
 143         unsigned long pfn;
 144
 145         start &= PAGE_SECTION_MASK;
 146         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 147                 unsigned long section = pfn_to_section_nr(pfn);
 148                 struct mem_section *ms;
 149
 150                 sparse_index_init(section, nid);
 151
 152                 ms = __nr_to_section(section);
 153                 if (!ms->section_mem_map)
 154                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 155                                                         SECTION_MARKED_PRESENT;
 156         }
 157 }
 158
 159 /*
 160  * Only used by the i386 NUMA architecures, but relatively
 161  * generic code.
 162  */
 163 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 164                                                      unsigned long end_pfn)
 165 {
 166         unsigned long pfn;
 167         unsigned long nr_pages = 0;
 168
 169         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 170                 if (nid != early_pfn_to_nid(pfn))
 171                         continue;
 172
 173                 if (pfn_valid(pfn))
 174                         nr_pages += PAGES_PER_SECTION;
 175         }
 176
 177         return nr_pages * sizeof(struct page);
 178 }
 179
 180 /*
 181  * Subtle, we encode the real pfn into the mem_map such that
 182  * the identity pfn - section_mem_map will return the actual
 183  * physical page frame number.
 184  */
 185 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 186 {
 187         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 188 }
 189
 190 /*
 191  * We need this if we ever free the mem_maps.  While not implemented yet,
 192  * this function is included for parity with its sibling.
 193  */
 194 static __attribute((unused))
 195 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 196 {
 197         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 198 }
 199
 200 static int __meminit sparse_init_one_section(struct mem_section *ms,
 201                 unsigned long pnum, struct page *mem_map)
 202 {
 203         if (!valid_section(ms))
 204                 return -EINVAL;
 205
 206         ms->section_mem_map &= ~SECTION_MAP_MASK;
 207         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
 208
 209         return 1;
 210 }
 211
 212 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 213 {
 214         struct page *map;
 215         struct mem_section *ms = __nr_to_section(pnum);
 216         int nid = sparse_early_nid(ms);
 217
 218         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 219         if (map)
 220                 return map;
 221
 222         map = alloc_bootmem_node(NODE_DATA(nid),
 223                         sizeof(struct page) * PAGES_PER_SECTION);
 224         if (map)
 225                 return map;
 226
 227         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 228         ms->section_mem_map = 0;
 229         return NULL;
 230 }
 231
 232 /*
 233  * Allocate the accumulated non-linear sections, allocate a mem_map
 234  * for each and record the physical to section mapping.
 235  */
 236 void __init sparse_init(void)
 237 {
 238         unsigned long pnum;
 239         struct page *map;
 240
 241         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 242                 if (!valid_section_nr(pnum))
 243                         continue;
 244
 245                 map = sparse_early_mem_map_alloc(pnum);
 246                 if (!map)
 247                         continue;
 248                 sparse_init_one_section(__nr_to_section(pnum), pnum, map);
 249         }
 250 }
 251
 252 #ifdef CONFIG_MEMORY_HOTPLUG
 253 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 254 {
 255         struct page *page, *ret;
 256         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 257
 258         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 259         if (page)
 260                 goto got_map_page;
 261
 262         ret = vmalloc(memmap_size);
 263         if (ret)
 264                 goto got_map_ptr;
 265
 266         return NULL;
 267 got_map_page:
 268         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 269 got_map_ptr:
 270         memset(ret, 0, memmap_size);
 271
 272         return ret;
 273 }
 274
 275 static int vaddr_in_vmalloc_area(void *addr)
 276 {
 277         if (addr >= (void *)VMALLOC_START &&
 278             addr < (void *)VMALLOC_END)
 279                 return 1;
 280         return 0;
 281 }
 282
 283 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 284 {
 285         if (vaddr_in_vmalloc_area(memmap))
 286                 vfree(memmap);
 287         else
 288                 free_pages((unsigned long)memmap,
 289                            get_order(sizeof(struct page) * nr_pages));
 290 }
 291
 292 /*
 293  * returns the number of sections whose mem_maps were properly
 294  * set.  If this is <=0, then that means that the passed-in
 295  * map was not consumed and must be freed.
 296  */
 297 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 298                            int nr_pages)
 299 {
 300         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 301         struct pglist_data *pgdat = zone->zone_pgdat;
 302         struct mem_section *ms;
 303         struct page *memmap;
 304         unsigned long flags;
 305         int ret;
 306
 307         /*
 308          * no locking for this, because it does its own
 309          * plus, it does a kmalloc
 310          */
 311         sparse_index_init(section_nr, pgdat->node_id);
 312         memmap = __kmalloc_section_memmap(nr_pages);
 313
 314         pgdat_resize_lock(pgdat, &flags);
 315
 316         ms = __pfn_to_section(start_pfn);
 317         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 318                 ret = -EEXIST;
 319                 goto out;
 320         }
 321         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 322
 323         ret = sparse_init_one_section(ms, section_nr, memmap);
 324
 325 out:
 326         pgdat_resize_unlock(pgdat, &flags);
 327         if (ret <= 0)
 328                 __kfree_section_memmap(memmap, nr_pages);
 329         return ret;
 330 }
 331 #endif