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/module.h>
   9 #include <linux/spinlock.h>
  10 #include <asm/dma.h>
  11
  12 /*
  13  * Permanent SPARSEMEM data:
  14  *
  15  * 1) mem_section       - memory sections, mem_map's for valid memory
  16  */
  17 #ifdef CONFIG_SPARSEMEM_EXTREME
  18 struct mem_section *mem_section[NR_SECTION_ROOTS]
  19         ____cacheline_maxaligned_in_smp;
  20 #else
  21 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  22         ____cacheline_maxaligned_in_smp;
  23 #endif
  24 EXPORT_SYMBOL(mem_section);
  25
  26 #ifdef CONFIG_SPARSEMEM_EXTREME
  27 static struct mem_section *sparse_index_alloc(int nid)
  28 {
  29         struct mem_section *section = NULL;
  30         unsigned long array_size = SECTIONS_PER_ROOT *
  31                                    sizeof(struct mem_section);
  32
  33         section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  34
  35         if (section)
  36                 memset(section, 0, array_size);
  37
  38         return section;
  39 }
  40
  41 static int sparse_index_init(unsigned long section_nr, int nid)
  42 {
  43         static spinlock_t index_init_lock = SPIN_LOCK_UNLOCKED;
  44         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  45         struct mem_section *section;
  46         int ret = 0;
  47
  48         if (mem_section[root])
  49                 return -EEXIST;
  50
  51         section = sparse_index_alloc(nid);
  52         /*
  53          * This lock keeps two different sections from
  54          * reallocating for the same index
  55          */
  56         spin_lock(&index_init_lock);
  57
  58         if (mem_section[root]) {
  59                 ret = -EEXIST;
  60                 goto out;
  61         }
  62
  63         mem_section[root] = section;
  64 out:
  65         spin_unlock(&index_init_lock);
  66         return ret;
  67 }
  68 #else /* !SPARSEMEM_EXTREME */
  69 static inline int sparse_index_init(unsigned long section_nr, int nid)
  70 {
  71         return 0;
  72 }
  73 #endif
  74
  75 /* Record a memory area against a node. */
  76 void memory_present(int nid, unsigned long start, unsigned long end)
  77 {
  78         unsigned long pfn;
  79
  80         start &= PAGE_SECTION_MASK;
  81         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
  82                 unsigned long section = pfn_to_section_nr(pfn);
  83                 struct mem_section *ms;
  84
  85                 sparse_index_init(section, nid);
  86
  87                 ms = __nr_to_section(section);
  88                 if (!ms->section_mem_map)
  89                         ms->section_mem_map = SECTION_MARKED_PRESENT;
  90         }
  91 }
  92
  93 /*
  94  * Only used by the i386 NUMA architecures, but relatively
  95  * generic code.
  96  */
  97 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
  98                                                      unsigned long end_pfn)
  99 {
 100         unsigned long pfn;
 101         unsigned long nr_pages = 0;
 102
 103         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 104                 if (nid != early_pfn_to_nid(pfn))
 105                         continue;
 106
 107                 if (pfn_valid(pfn))
 108                         nr_pages += PAGES_PER_SECTION;
 109         }
 110
 111         return nr_pages * sizeof(struct page);
 112 }
 113
 114 /*
 115  * Subtle, we encode the real pfn into the mem_map such that
 116  * the identity pfn - section_mem_map will return the actual
 117  * physical page frame number.
 118  */
 119 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 120 {
 121         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 122 }
 123
 124 /*
 125  * We need this if we ever free the mem_maps.  While not implemented yet,
 126  * this function is included for parity with its sibling.
 127  */
 128 static __attribute((unused))
 129 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 130 {
 131         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 132 }
 133
 134 static int sparse_init_one_section(struct mem_section *ms,
 135                 unsigned long pnum, struct page *mem_map)
 136 {
 137         if (!valid_section(ms))
 138                 return -EINVAL;
 139
 140         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
 141
 142         return 1;
 143 }
 144
 145 static struct page *sparse_early_mem_map_alloc(unsigned long pnum)
 146 {
 147         struct page *map;
 148         int nid = early_pfn_to_nid(section_nr_to_pfn(pnum));
 149         struct mem_section *ms = __nr_to_section(pnum);
 150
 151         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 152         if (map)
 153                 return map;
 154
 155         map = alloc_bootmem_node(NODE_DATA(nid),
 156                         sizeof(struct page) * PAGES_PER_SECTION);
 157         if (map)
 158                 return map;
 159
 160         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 161         ms->section_mem_map = 0;
 162         return NULL;
 163 }
 164
 165 /*
 166  * Allocate the accumulated non-linear sections, allocate a mem_map
 167  * for each and record the physical to section mapping.
 168  */
 169 void sparse_init(void)
 170 {
 171         unsigned long pnum;
 172         struct page *map;
 173
 174         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 175                 if (!valid_section_nr(pnum))
 176                         continue;
 177
 178                 map = sparse_early_mem_map_alloc(pnum);
 179                 if (!map)
 180                         continue;
 181                 sparse_init_one_section(__nr_to_section(pnum), pnum, map);
 182         }
 183 }
 184
 185 /*
 186  * returns the number of sections whose mem_maps were properly
 187  * set.  If this is <=0, then that means that the passed-in
 188  * map was not consumed and must be freed.
 189  */
 190 int sparse_add_one_section(unsigned long start_pfn, int nr_pages, struct page *map)
 191 {
 192         struct mem_section *ms = __pfn_to_section(start_pfn);
 193
 194         if (ms->section_mem_map & SECTION_MARKED_PRESENT)
 195                 return -EEXIST;
 196
 197         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 198
 199         return sparse_init_one_section(ms, pfn_to_section_nr(start_pfn), map);
 200 }