mm/sparse.c

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