mm/sparse-vmemmap.c

   1 /*
   2  * Virtual Memory Map support
   3  *
   4  * (C) 2007 sgi. Christoph Lameter.
   5  *
   6  * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
   7  * virt_to_page, page_address() to be implemented as a base offset
   8  * calculation without memory access.
   9  *
  10  * However, virtual mappings need a page table and TLBs. Many Linux
  11  * architectures already map their physical space using 1-1 mappings
  12  * via TLBs. For those arches the virtual memory map is essentially
  13  * for free if we use the same page size as the 1-1 mappings. In that
  14  * case the overhead consists of a few additional pages that are
  15  * allocated to create a view of memory for vmemmap.
  16  *
  17  * The architecture is expected to provide a vmemmap_populate() function
  18  * to instantiate the mapping.
  19  */
  20 #include <linux/mm.h>
  21 #include <linux/mmzone.h>
  22 #include <linux/bootmem.h>
  23 #include <linux/memremap.h>
  24 #include <linux/highmem.h>
  25 #include <linux/slab.h>
  26 #include <linux/spinlock.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/sched.h>
  29 #include <asm/dma.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/pgtable.h>
  32
  33 /*
  34  * Allocate a block of memory to be used to back the virtual memory map
  35  * or to back the page tables that are used to create the mapping.
  36  * Uses the main allocators if they are available, else bootmem.
  37  */
  38
  39 static void * __ref __earlyonly_bootmem_alloc(int node,
  40                                 unsigned long size,
  41                                 unsigned long align,
  42                                 unsigned long goal)
  43 {
  44         return memblock_virt_alloc_try_nid(size, align, goal,
  45                                             BOOTMEM_ALLOC_ACCESSIBLE, node);
  46 }
  47
  48 static void *vmemmap_buf;
  49 static void *vmemmap_buf_end;
  50
  51 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
  52 {
  53         /* If the main allocator is up use that, fallback to bootmem. */
  54         if (slab_is_available()) {
  55                 struct page *page;
  56
  57                 if (node_state(node, N_HIGH_MEMORY))
  58                         page = alloc_pages_node(
  59                                 node, GFP_KERNEL | __GFP_ZERO | __GFP_REPEAT,
  60                                 get_order(size));
  61                 else
  62                         page = alloc_pages(
  63                                 GFP_KERNEL | __GFP_ZERO | __GFP_REPEAT,
  64                                 get_order(size));
  65                 if (page)
  66                         return page_address(page);
  67                 return NULL;
  68         } else
  69                 return __earlyonly_bootmem_alloc(node, size, size,
  70                                 __pa(MAX_DMA_ADDRESS));
  71 }
  72
  73 /* need to make sure size is all the same during early stage */
  74 static void * __meminit alloc_block_buf(unsigned long size, int node)
  75 {
  76         void *ptr;
  77
  78         if (!vmemmap_buf)
  79                 return vmemmap_alloc_block(size, node);
  80
  81         /* take the from buf */
  82         ptr = (void *)ALIGN((unsigned long)vmemmap_buf, size);
  83         if (ptr + size > vmemmap_buf_end)
  84                 return vmemmap_alloc_block(size, node);
  85
  86         vmemmap_buf = ptr + size;
  87
  88         return ptr;
  89 }
  90
  91 static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
  92 {
  93         return altmap->base_pfn + altmap->reserve + altmap->alloc
  94                 + altmap->align;
  95 }
  96
  97 static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
  98 {
  99         unsigned long allocated = altmap->alloc + altmap->align;
 100
 101         if (altmap->free > allocated)
 102                 return altmap->free - allocated;
 103         return 0;
 104 }
 105
 106 /**
 107  * vmem_altmap_alloc - allocate pages from the vmem_altmap reservation
 108  * @altmap - reserved page pool for the allocation
 109  * @nr_pfns - size (in pages) of the allocation
 110  *
 111  * Allocations are aligned to the size of the request
 112  */
 113 static unsigned long __meminit vmem_altmap_alloc(struct vmem_altmap *altmap,
 114                 unsigned long nr_pfns)
 115 {
 116         unsigned long pfn = vmem_altmap_next_pfn(altmap);
 117         unsigned long nr_align;
 118
 119         nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
 120         nr_align = ALIGN(pfn, nr_align) - pfn;
 121
 122         if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
 123                 return ULONG_MAX;
 124         altmap->alloc += nr_pfns;
 125         altmap->align += nr_align;
 126         return pfn + nr_align;
 127 }
 128
 129 static void * __meminit altmap_alloc_block_buf(unsigned long size,
 130                 struct vmem_altmap *altmap)
 131 {
 132         unsigned long pfn, nr_pfns;
 133         void *ptr;
 134
 135         if (size & ~PAGE_MASK) {
 136                 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
 137                                 __func__, size);
 138                 return NULL;
 139         }
 140
 141         nr_pfns = size >> PAGE_SHIFT;
 142         pfn = vmem_altmap_alloc(altmap, nr_pfns);
 143         if (pfn < ULONG_MAX)
 144                 ptr = __va(__pfn_to_phys(pfn));
 145         else
 146                 ptr = NULL;
 147         pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
 148                         __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
 149
 150         return ptr;
 151 }
 152
 153 /* need to make sure size is all the same during early stage */
 154 void * __meminit __vmemmap_alloc_block_buf(unsigned long size, int node,
 155                 struct vmem_altmap *altmap)
 156 {
 157         if (altmap)
 158                 return altmap_alloc_block_buf(size, altmap);
 159         return alloc_block_buf(size, node);
 160 }
 161
 162 void __meminit vmemmap_verify(pte_t *pte, int node,
 163                                 unsigned long start, unsigned long end)
 164 {
 165         unsigned long pfn = pte_pfn(*pte);
 166         int actual_node = early_pfn_to_nid(pfn);
 167
 168         if (node_distance(actual_node, node) > LOCAL_DISTANCE)
 169                 pr_warn("[%lx-%lx] potential offnode page_structs\n",
 170                         start, end - 1);
 171 }
 172
 173 pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node)
 174 {
 175         pte_t *pte = pte_offset_kernel(pmd, addr);
 176         if (pte_none(*pte)) {
 177                 pte_t entry;
 178                 void *p = alloc_block_buf(PAGE_SIZE, node);
 179                 if (!p)
 180                         return NULL;
 181                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
 182                 set_pte_at(&init_mm, addr, pte, entry);
 183         }
 184         return pte;
 185 }
 186
 187 pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
 188 {
 189         pmd_t *pmd = pmd_offset(pud, addr);
 190         if (pmd_none(*pmd)) {
 191                 void *p = vmemmap_alloc_block(PAGE_SIZE, node);
 192                 if (!p)
 193                         return NULL;
 194                 pmd_populate_kernel(&init_mm, pmd, p);
 195         }
 196         return pmd;
 197 }
 198
 199 pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
 200 {
 201         pud_t *pud = pud_offset(p4d, addr);
 202         if (pud_none(*pud)) {
 203                 void *p = vmemmap_alloc_block(PAGE_SIZE, node);
 204                 if (!p)
 205                         return NULL;
 206                 pud_populate(&init_mm, pud, p);
 207         }
 208         return pud;
 209 }
 210
 211 p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
 212 {
 213         p4d_t *p4d = p4d_offset(pgd, addr);
 214         if (p4d_none(*p4d)) {
 215                 void *p = vmemmap_alloc_block(PAGE_SIZE, node);
 216                 if (!p)
 217                         return NULL;
 218                 p4d_populate(&init_mm, p4d, p);
 219         }
 220         return p4d;
 221 }
 222
 223 pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
 224 {
 225         pgd_t *pgd = pgd_offset_k(addr);
 226         if (pgd_none(*pgd)) {
 227                 void *p = vmemmap_alloc_block(PAGE_SIZE, node);
 228                 if (!p)
 229                         return NULL;
 230                 pgd_populate(&init_mm, pgd, p);
 231         }
 232         return pgd;
 233 }
 234
 235 int __meminit vmemmap_populate_basepages(unsigned long start,
 236                                          unsigned long end, int node)
 237 {
 238         unsigned long addr = start;
 239         pgd_t *pgd;
 240         p4d_t *p4d;
 241         pud_t *pud;
 242         pmd_t *pmd;
 243         pte_t *pte;
 244
 245         for (; addr < end; addr += PAGE_SIZE) {
 246                 pgd = vmemmap_pgd_populate(addr, node);
 247                 if (!pgd)
 248                         return -ENOMEM;
 249                 p4d = vmemmap_p4d_populate(pgd, addr, node);
 250                 if (!p4d)
 251                         return -ENOMEM;
 252                 pud = vmemmap_pud_populate(p4d, addr, node);
 253                 if (!pud)
 254                         return -ENOMEM;
 255                 pmd = vmemmap_pmd_populate(pud, addr, node);
 256                 if (!pmd)
 257                         return -ENOMEM;
 258                 pte = vmemmap_pte_populate(pmd, addr, node);
 259                 if (!pte)
 260                         return -ENOMEM;
 261                 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
 262         }
 263
 264         return 0;
 265 }
 266
 267 struct page * __meminit sparse_mem_map_populate(unsigned long pnum, int nid)
 268 {
 269         unsigned long start;
 270         unsigned long end;
 271         struct page *map;
 272
 273         map = pfn_to_page(pnum * PAGES_PER_SECTION);
 274         start = (unsigned long)map;
 275         end = (unsigned long)(map + PAGES_PER_SECTION);
 276
 277         if (vmemmap_populate(start, end, nid))
 278                 return NULL;
 279
 280         return map;
 281 }
 282
 283 void __init sparse_mem_maps_populate_node(struct page **map_map,
 284                                           unsigned long pnum_begin,
 285                                           unsigned long pnum_end,
 286                                           unsigned long map_count, int nodeid)
 287 {
 288         unsigned long pnum;
 289         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 290         void *vmemmap_buf_start;
 291
 292         size = ALIGN(size, PMD_SIZE);
 293         vmemmap_buf_start = __earlyonly_bootmem_alloc(nodeid, size * map_count,
 294                          PMD_SIZE, __pa(MAX_DMA_ADDRESS));
 295
 296         if (vmemmap_buf_start) {
 297                 vmemmap_buf = vmemmap_buf_start;
 298                 vmemmap_buf_end = vmemmap_buf_start + size * map_count;
 299         }
 300
 301         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 302                 struct mem_section *ms;
 303
 304                 if (!present_section_nr(pnum))
 305                         continue;
 306
 307                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 308                 if (map_map[pnum])
 309                         continue;
 310                 ms = __nr_to_section(pnum);
 311                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 312                        __func__);
 313                 ms->section_mem_map = 0;
 314         }
 315
 316         if (vmemmap_buf_start) {
 317                 /* need to free left buf */
 318                 memblock_free_early(__pa(vmemmap_buf),
 319                                     vmemmap_buf_end - vmemmap_buf);
 320                 vmemmap_buf = NULL;
 321                 vmemmap_buf_end = NULL;
 322         }
 323 }