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_internodealigned_in_smp;
  22 #else
  23 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  24         ____cacheline_internodealigned_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         if (slab_is_available())
  36                 section = kmalloc_node(array_size, GFP_KERNEL, nid);
  37         else
  38                 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  39
  40         if (section)
  41                 memset(section, 0, array_size);
  42
  43         return section;
  44 }
  45
  46 static int sparse_index_init(unsigned long section_nr, int nid)
  47 {
  48         static spinlock_t index_init_lock = SPIN_LOCK_UNLOCKED;
  49         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  50         struct mem_section *section;
  51         int ret = 0;
  52
  53         if (mem_section[root])
  54                 return -EEXIST;
  55
  56         section = sparse_index_alloc(nid);
  57         /*
  58          * This lock keeps two different sections from
  59          * reallocating for the same index
  60          */
  61         spin_lock(&index_init_lock);
  62
  63         if (mem_section[root]) {
  64                 ret = -EEXIST;
  65                 goto out;
  66         }
  67
  68         mem_section[root] = section;
  69 out:
  70         spin_unlock(&index_init_lock);
  71         return ret;
  72 }
  73 #else /* !SPARSEMEM_EXTREME */
  74 static inline int sparse_index_init(unsigned long section_nr, int nid)
  75 {
  76         return 0;
  77 }
  78 #endif
  79
  80 /*
  81  * Although written for the SPARSEMEM_EXTREME case, this happens
  82  * to also work for the flat array case becase
  83  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
  84  */
  85 int __section_nr(struct mem_section* ms)
  86 {
  87         unsigned long root_nr;
  88         struct mem_section* root;
  89
  90         for (root_nr = 0;
  91              root_nr < NR_MEM_SECTIONS;
  92              root_nr += SECTIONS_PER_ROOT) {
  93                 root = __nr_to_section(root_nr);
  94
  95                 if (!root)
  96                         continue;
  97
  98                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
  99                      break;
 100         }
 101
 102         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 103 }
 104
 105 /* Record a memory area against a node. */
 106 void memory_present(int nid, unsigned long start, unsigned long end)
 107 {
 108         unsigned long pfn;
 109
 110         start &= PAGE_SECTION_MASK;
 111         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 112                 unsigned long section = pfn_to_section_nr(pfn);
 113                 struct mem_section *ms;
 114
 115                 sparse_index_init(section, nid);
 116
 117                 ms = __nr_to_section(section);
 118                 if (!ms->section_mem_map)
 119                         ms->section_mem_map = SECTION_MARKED_PRESENT;
 120         }
 121 }
 122
 123 /*
 124  * Only used by the i386 NUMA architecures, but relatively
 125  * generic code.
 126  */
 127 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 128                                                      unsigned long end_pfn)
 129 {
 130         unsigned long pfn;
 131         unsigned long nr_pages = 0;
 132
 133         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 134                 if (nid != early_pfn_to_nid(pfn))
 135                         continue;
 136
 137                 if (pfn_valid(pfn))
 138                         nr_pages += PAGES_PER_SECTION;
 139         }
 140
 141         return nr_pages * sizeof(struct page);
 142 }
 143
 144 /*
 145  * Subtle, we encode the real pfn into the mem_map such that
 146  * the identity pfn - section_mem_map will return the actual
 147  * physical page frame number.
 148  */
 149 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 150 {
 151         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 152 }
 153
 154 /*
 155  * We need this if we ever free the mem_maps.  While not implemented yet,
 156  * this function is included for parity with its sibling.
 157  */
 158 static __attribute((unused))
 159 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 160 {
 161         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 162 }
 163
 164 static int sparse_init_one_section(struct mem_section *ms,
 165                 unsigned long pnum, struct page *mem_map)
 166 {
 167         if (!valid_section(ms))
 168                 return -EINVAL;
 169
 170         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
 171
 172         return 1;
 173 }
 174
 175 static struct page *sparse_early_mem_map_alloc(unsigned long pnum)
 176 {
 177         struct page *map;
 178         int nid = early_pfn_to_nid(section_nr_to_pfn(pnum));
 179         struct mem_section *ms = __nr_to_section(pnum);
 180
 181         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 182         if (map)
 183                 return map;
 184
 185         map = alloc_bootmem_node(NODE_DATA(nid),
 186                         sizeof(struct page) * PAGES_PER_SECTION);
 187         if (map)
 188                 return map;
 189
 190         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 191         ms->section_mem_map = 0;
 192         return NULL;
 193 }
 194
 195 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 196 {
 197         struct page *page, *ret;
 198         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 199
 200         page = alloc_pages(GFP_KERNEL, get_order(memmap_size));
 201         if (page)
 202                 goto got_map_page;
 203
 204         ret = vmalloc(memmap_size);
 205         if (ret)
 206                 goto got_map_ptr;
 207
 208         return NULL;
 209 got_map_page:
 210         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 211 got_map_ptr:
 212         memset(ret, 0, memmap_size);
 213
 214         return ret;
 215 }
 216
 217 static int vaddr_in_vmalloc_area(void *addr)
 218 {
 219         if (addr >= (void *)VMALLOC_START &&
 220             addr < (void *)VMALLOC_END)
 221                 return 1;
 222         return 0;
 223 }
 224
 225 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 226 {
 227         if (vaddr_in_vmalloc_area(memmap))
 228                 vfree(memmap);
 229         else
 230                 free_pages((unsigned long)memmap,
 231                            get_order(sizeof(struct page) * nr_pages));
 232 }
 233
 234 /*
 235  * Allocate the accumulated non-linear sections, allocate a mem_map
 236  * for each and record the physical to section mapping.
 237  */
 238 void sparse_init(void)
 239 {
 240         unsigned long pnum;
 241         struct page *map;
 242
 243         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 244                 if (!valid_section_nr(pnum))
 245                         continue;
 246
 247                 map = sparse_early_mem_map_alloc(pnum);
 248                 if (!map)
 249                         continue;
 250                 sparse_init_one_section(__nr_to_section(pnum), pnum, map);
 251         }
 252 }
 253
 254 /*
 255  * returns the number of sections whose mem_maps were properly
 256  * set.  If this is <=0, then that means that the passed-in
 257  * map was not consumed and must be freed.
 258  */
 259 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 260                            int nr_pages)
 261 {
 262         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 263         struct pglist_data *pgdat = zone->zone_pgdat;
 264         struct mem_section *ms;
 265         struct page *memmap;
 266         unsigned long flags;
 267         int ret;
 268
 269         /*
 270          * no locking for this, because it does its own
 271          * plus, it does a kmalloc
 272          */
 273         sparse_index_init(section_nr, pgdat->node_id);
 274         memmap = __kmalloc_section_memmap(nr_pages);
 275
 276         pgdat_resize_lock(pgdat, &flags);
 277
 278         ms = __pfn_to_section(start_pfn);
 279         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 280                 ret = -EEXIST;
 281                 goto out;
 282         }
 283         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 284
 285         ret = sparse_init_one_section(ms, section_nr, memmap);
 286
 287 out:
 288         pgdat_resize_unlock(pgdat, &flags);
 289         if (ret <= 0)
 290                 __kfree_section_memmap(memmap, nr_pages);
 291         return ret;
 292 }