Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 1 | /* |
| 2 | * sparse memory mappings. |
| 3 | */ |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 4 | #include <linux/mm.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 5 | #include <linux/slab.h> |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 6 | #include <linux/mmzone.h> |
| 7 | #include <linux/bootmem.h> |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 8 | #include <linux/highmem.h> |
Paul Gortmaker | b95f1b31 | 2011-10-16 02:01:52 -0400 | [diff] [blame] | 9 | #include <linux/export.h> |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 10 | #include <linux/spinlock.h> |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 11 | #include <linux/vmalloc.h> |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 12 | #include "internal.h" |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 13 | #include <asm/dma.h> |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 14 | #include <asm/pgalloc.h> |
| 15 | #include <asm/pgtable.h> |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 16 | |
| 17 | /* |
| 18 | * Permanent SPARSEMEM data: |
| 19 | * |
| 20 | * 1) mem_section - memory sections, mem_map's for valid memory |
| 21 | */ |
Bob Picco | 3e34726 | 2005-09-03 15:54:28 -0700 | [diff] [blame] | 22 | #ifdef CONFIG_SPARSEMEM_EXTREME |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 23 | struct mem_section *mem_section[NR_SECTION_ROOTS] |
Ravikiran G Thirumalai | 22fc6ec | 2006-01-08 01:01:27 -0800 | [diff] [blame] | 24 | ____cacheline_internodealigned_in_smp; |
Bob Picco | 3e34726 | 2005-09-03 15:54:28 -0700 | [diff] [blame] | 25 | #else |
| 26 | struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT] |
Ravikiran G Thirumalai | 22fc6ec | 2006-01-08 01:01:27 -0800 | [diff] [blame] | 27 | ____cacheline_internodealigned_in_smp; |
Bob Picco | 3e34726 | 2005-09-03 15:54:28 -0700 | [diff] [blame] | 28 | #endif |
| 29 | EXPORT_SYMBOL(mem_section); |
| 30 | |
Christoph Lameter | 89689ae | 2006-12-06 20:31:45 -0800 | [diff] [blame] | 31 | #ifdef NODE_NOT_IN_PAGE_FLAGS |
| 32 | /* |
| 33 | * If we did not store the node number in the page then we have to |
| 34 | * do a lookup in the section_to_node_table in order to find which |
| 35 | * node the page belongs to. |
| 36 | */ |
| 37 | #if MAX_NUMNODES <= 256 |
| 38 | static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; |
| 39 | #else |
| 40 | static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; |
| 41 | #endif |
| 42 | |
Ian Campbell | 33dd4e0 | 2011-07-25 17:11:51 -0700 | [diff] [blame] | 43 | int page_to_nid(const struct page *page) |
Christoph Lameter | 89689ae | 2006-12-06 20:31:45 -0800 | [diff] [blame] | 44 | { |
| 45 | return section_to_node_table[page_to_section(page)]; |
| 46 | } |
| 47 | EXPORT_SYMBOL(page_to_nid); |
Andy Whitcroft | 85770ff | 2007-08-22 14:01:03 -0700 | [diff] [blame] | 48 | |
| 49 | static void set_section_nid(unsigned long section_nr, int nid) |
| 50 | { |
| 51 | section_to_node_table[section_nr] = nid; |
| 52 | } |
| 53 | #else /* !NODE_NOT_IN_PAGE_FLAGS */ |
| 54 | static inline void set_section_nid(unsigned long section_nr, int nid) |
| 55 | { |
| 56 | } |
Christoph Lameter | 89689ae | 2006-12-06 20:31:45 -0800 | [diff] [blame] | 57 | #endif |
| 58 | |
Bob Picco | 3e34726 | 2005-09-03 15:54:28 -0700 | [diff] [blame] | 59 | #ifdef CONFIG_SPARSEMEM_EXTREME |
Sam Ravnborg | 577a32f | 2007-05-17 23:29:25 +0200 | [diff] [blame] | 60 | static struct mem_section noinline __init_refok *sparse_index_alloc(int nid) |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 61 | { |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 62 | struct mem_section *section = NULL; |
| 63 | unsigned long array_size = SECTIONS_PER_ROOT * |
| 64 | sizeof(struct mem_section); |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 65 | |
Shaohua Li | f52407c | 2009-09-21 17:01:19 -0700 | [diff] [blame] | 66 | if (slab_is_available()) { |
| 67 | if (node_state(nid, N_HIGH_MEMORY)) |
Gavin Shan | 5b760e6 | 2012-07-31 16:46:02 -0700 | [diff] [blame] | 68 | section = kzalloc_node(array_size, GFP_KERNEL, nid); |
Shaohua Li | f52407c | 2009-09-21 17:01:19 -0700 | [diff] [blame] | 69 | else |
Gavin Shan | 5b760e6 | 2012-07-31 16:46:02 -0700 | [diff] [blame] | 70 | section = kzalloc(array_size, GFP_KERNEL); |
| 71 | } else { |
Mike Kravetz | 46a66ee | 2006-05-01 12:16:09 -0700 | [diff] [blame] | 72 | section = alloc_bootmem_node(NODE_DATA(nid), array_size); |
Gavin Shan | 5b760e6 | 2012-07-31 16:46:02 -0700 | [diff] [blame] | 73 | } |
Bob Picco | 3e34726 | 2005-09-03 15:54:28 -0700 | [diff] [blame] | 74 | |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 75 | return section; |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 76 | } |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 77 | |
Yasunori Goto | a3142c8 | 2007-05-08 00:23:07 -0700 | [diff] [blame] | 78 | static int __meminit sparse_index_init(unsigned long section_nr, int nid) |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 79 | { |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 80 | unsigned long root = SECTION_NR_TO_ROOT(section_nr); |
| 81 | struct mem_section *section; |
| 82 | int ret = 0; |
| 83 | |
| 84 | if (mem_section[root]) |
| 85 | return -EEXIST; |
| 86 | |
| 87 | section = sparse_index_alloc(nid); |
WANG Cong | af0cd5a | 2007-12-17 16:19:58 -0800 | [diff] [blame] | 88 | if (!section) |
| 89 | return -ENOMEM; |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 90 | |
| 91 | mem_section[root] = section; |
Gavin Shan | c1c9518 | 2012-07-31 16:46:06 -0700 | [diff] [blame] | 92 | |
Dave Hansen | 28ae55c | 2005-09-03 15:54:29 -0700 | [diff] [blame] | 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 | |
Dave Hansen | 4ca644d | 2005-10-29 18:16:51 -0700 | [diff] [blame] | 102 | /* |
| 103 | * Although written for the SPARSEMEM_EXTREME case, this happens |
Andy Whitcroft | cd881a6 | 2007-10-16 01:24:10 -0700 | [diff] [blame] | 104 | * to also work for the flat array case because |
Dave Hansen | 4ca644d | 2005-10-29 18:16:51 -0700 | [diff] [blame] | 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 | |
Mike Kravetz | 12783b0 | 2006-05-20 15:00:05 -0700 | [diff] [blame] | 112 | for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) { |
| 113 | root = __nr_to_section(root_nr * SECTIONS_PER_ROOT); |
Dave Hansen | 4ca644d | 2005-10-29 18:16:51 -0700 | [diff] [blame] | 114 | if (!root) |
| 115 | continue; |
| 116 | |
| 117 | if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT))) |
| 118 | break; |
| 119 | } |
| 120 | |
Gavin Shan | db36a46 | 2012-07-31 16:46:04 -0700 | [diff] [blame] | 121 | VM_BUG_ON(root_nr == NR_SECTION_ROOTS); |
| 122 | |
Dave Hansen | 4ca644d | 2005-10-29 18:16:51 -0700 | [diff] [blame] | 123 | return (root_nr * SECTIONS_PER_ROOT) + (ms - root); |
| 124 | } |
| 125 | |
Andy Whitcroft | 30c253e | 2006-06-23 02:03:41 -0700 | [diff] [blame] | 126 | /* |
| 127 | * During early boot, before section_mem_map is used for an actual |
| 128 | * mem_map, we use section_mem_map to store the section's NUMA |
| 129 | * node. This keeps us from having to use another data structure. The |
| 130 | * node information is cleared just before we store the real mem_map. |
| 131 | */ |
| 132 | static inline unsigned long sparse_encode_early_nid(int nid) |
| 133 | { |
| 134 | return (nid << SECTION_NID_SHIFT); |
| 135 | } |
| 136 | |
| 137 | static inline int sparse_early_nid(struct mem_section *section) |
| 138 | { |
| 139 | return (section->section_mem_map >> SECTION_NID_SHIFT); |
| 140 | } |
| 141 | |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 142 | /* Validate the physical addressing limitations of the model */ |
| 143 | void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn, |
| 144 | unsigned long *end_pfn) |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 145 | { |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 146 | unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT); |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 147 | |
Ingo Molnar | bead9a3 | 2008-04-16 01:40:00 +0200 | [diff] [blame] | 148 | /* |
| 149 | * Sanity checks - do not allow an architecture to pass |
| 150 | * in larger pfns than the maximum scope of sparsemem: |
| 151 | */ |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 152 | if (*start_pfn > max_sparsemem_pfn) { |
| 153 | mminit_dprintk(MMINIT_WARNING, "pfnvalidation", |
| 154 | "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n", |
| 155 | *start_pfn, *end_pfn, max_sparsemem_pfn); |
| 156 | WARN_ON_ONCE(1); |
| 157 | *start_pfn = max_sparsemem_pfn; |
| 158 | *end_pfn = max_sparsemem_pfn; |
Cyrill Gorcunov | ef161a9 | 2009-03-31 15:19:25 -0700 | [diff] [blame] | 159 | } else if (*end_pfn > max_sparsemem_pfn) { |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 160 | mminit_dprintk(MMINIT_WARNING, "pfnvalidation", |
| 161 | "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n", |
| 162 | *start_pfn, *end_pfn, max_sparsemem_pfn); |
| 163 | WARN_ON_ONCE(1); |
| 164 | *end_pfn = max_sparsemem_pfn; |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | /* Record a memory area against a node. */ |
| 169 | void __init memory_present(int nid, unsigned long start, unsigned long end) |
| 170 | { |
| 171 | unsigned long pfn; |
Ingo Molnar | bead9a3 | 2008-04-16 01:40:00 +0200 | [diff] [blame] | 172 | |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 173 | start &= PAGE_SECTION_MASK; |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 174 | mminit_validate_memmodel_limits(&start, &end); |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 175 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) { |
| 176 | unsigned long section = pfn_to_section_nr(pfn); |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 177 | struct mem_section *ms; |
| 178 | |
| 179 | sparse_index_init(section, nid); |
Andy Whitcroft | 85770ff | 2007-08-22 14:01:03 -0700 | [diff] [blame] | 180 | set_section_nid(section, nid); |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 181 | |
| 182 | ms = __nr_to_section(section); |
| 183 | if (!ms->section_mem_map) |
Andy Whitcroft | 30c253e | 2006-06-23 02:03:41 -0700 | [diff] [blame] | 184 | ms->section_mem_map = sparse_encode_early_nid(nid) | |
| 185 | SECTION_MARKED_PRESENT; |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 186 | } |
| 187 | } |
| 188 | |
| 189 | /* |
| 190 | * Only used by the i386 NUMA architecures, but relatively |
| 191 | * generic code. |
| 192 | */ |
| 193 | unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn, |
| 194 | unsigned long end_pfn) |
| 195 | { |
| 196 | unsigned long pfn; |
| 197 | unsigned long nr_pages = 0; |
| 198 | |
Mel Gorman | 2dbb51c | 2008-07-23 21:26:52 -0700 | [diff] [blame] | 199 | mminit_validate_memmodel_limits(&start_pfn, &end_pfn); |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 200 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { |
| 201 | if (nid != early_pfn_to_nid(pfn)) |
| 202 | continue; |
| 203 | |
Andy Whitcroft | 540557b | 2007-10-16 01:24:11 -0700 | [diff] [blame] | 204 | if (pfn_present(pfn)) |
Andy Whitcroft | d41dee3 | 2005-06-23 00:07:54 -0700 | [diff] [blame] | 205 | nr_pages += PAGES_PER_SECTION; |
| 206 | } |
| 207 | |
| 208 | return nr_pages * sizeof(struct page); |
| 209 | } |
| 210 | |
| 211 | /* |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 212 | * Subtle, we encode the real pfn into the mem_map such that |
| 213 | * the identity pfn - section_mem_map will return the actual |
| 214 | * physical page frame number. |
| 215 | */ |
| 216 | static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum) |
| 217 | { |
| 218 | return (unsigned long)(mem_map - (section_nr_to_pfn(pnum))); |
| 219 | } |
| 220 | |
| 221 | /* |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 222 | * Decode mem_map from the coded memmap |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 223 | */ |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 224 | struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum) |
| 225 | { |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 226 | /* mask off the extra low bits of information */ |
| 227 | coded_mem_map &= SECTION_MAP_MASK; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 228 | return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum); |
| 229 | } |
| 230 | |
Yasunori Goto | a3142c8 | 2007-05-08 00:23:07 -0700 | [diff] [blame] | 231 | static int __meminit sparse_init_one_section(struct mem_section *ms, |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 232 | unsigned long pnum, struct page *mem_map, |
| 233 | unsigned long *pageblock_bitmap) |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 234 | { |
Andy Whitcroft | 540557b | 2007-10-16 01:24:11 -0700 | [diff] [blame] | 235 | if (!present_section(ms)) |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 236 | return -EINVAL; |
| 237 | |
Andy Whitcroft | 30c253e | 2006-06-23 02:03:41 -0700 | [diff] [blame] | 238 | ms->section_mem_map &= ~SECTION_MAP_MASK; |
Andy Whitcroft | 540557b | 2007-10-16 01:24:11 -0700 | [diff] [blame] | 239 | ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) | |
| 240 | SECTION_HAS_MEM_MAP; |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 241 | ms->pageblock_flags = pageblock_bitmap; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 242 | |
| 243 | return 1; |
| 244 | } |
| 245 | |
Yasunori Goto | 0475327 | 2008-04-28 02:13:31 -0700 | [diff] [blame] | 246 | unsigned long usemap_size(void) |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 247 | { |
| 248 | unsigned long size_bytes; |
| 249 | size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8; |
| 250 | size_bytes = roundup(size_bytes, sizeof(unsigned long)); |
| 251 | return size_bytes; |
| 252 | } |
| 253 | |
| 254 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 255 | static unsigned long *__kmalloc_section_usemap(void) |
| 256 | { |
| 257 | return kmalloc(usemap_size(), GFP_KERNEL); |
| 258 | } |
| 259 | #endif /* CONFIG_MEMORY_HOTPLUG */ |
| 260 | |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 261 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 262 | static unsigned long * __init |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 263 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, |
Johannes Weiner | 238305b | 2012-05-29 15:06:36 -0700 | [diff] [blame] | 264 | unsigned long size) |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 265 | { |
Yinghai Lu | 99ab7b1 | 2012-07-11 14:02:53 -0700 | [diff] [blame] | 266 | unsigned long goal, limit; |
| 267 | unsigned long *p; |
| 268 | int nid; |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 269 | /* |
| 270 | * A page may contain usemaps for other sections preventing the |
| 271 | * page being freed and making a section unremovable while |
| 272 | * other sections referencing the usemap retmain active. Similarly, |
| 273 | * a pgdat can prevent a section being removed. If section A |
| 274 | * contains a pgdat and section B contains the usemap, both |
| 275 | * sections become inter-dependent. This allocates usemaps |
| 276 | * from the same section as the pgdat where possible to avoid |
| 277 | * this problem. |
| 278 | */ |
Yinghai Lu | 07b4e2b | 2012-07-11 14:02:51 -0700 | [diff] [blame] | 279 | goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT); |
Yinghai Lu | 99ab7b1 | 2012-07-11 14:02:53 -0700 | [diff] [blame] | 280 | limit = goal + (1UL << PA_SECTION_SHIFT); |
| 281 | nid = early_pfn_to_nid(goal >> PAGE_SHIFT); |
| 282 | again: |
| 283 | p = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size, |
| 284 | SMP_CACHE_BYTES, goal, limit); |
| 285 | if (!p && limit) { |
| 286 | limit = 0; |
| 287 | goto again; |
| 288 | } |
| 289 | return p; |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 290 | } |
| 291 | |
| 292 | static void __init check_usemap_section_nr(int nid, unsigned long *usemap) |
| 293 | { |
| 294 | unsigned long usemap_snr, pgdat_snr; |
| 295 | static unsigned long old_usemap_snr = NR_MEM_SECTIONS; |
| 296 | static unsigned long old_pgdat_snr = NR_MEM_SECTIONS; |
| 297 | struct pglist_data *pgdat = NODE_DATA(nid); |
| 298 | int usemap_nid; |
| 299 | |
| 300 | usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT); |
| 301 | pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT); |
| 302 | if (usemap_snr == pgdat_snr) |
| 303 | return; |
| 304 | |
| 305 | if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr) |
| 306 | /* skip redundant message */ |
| 307 | return; |
| 308 | |
| 309 | old_usemap_snr = usemap_snr; |
| 310 | old_pgdat_snr = pgdat_snr; |
| 311 | |
| 312 | usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr)); |
| 313 | if (usemap_nid != nid) { |
| 314 | printk(KERN_INFO |
| 315 | "node %d must be removed before remove section %ld\n", |
| 316 | nid, usemap_snr); |
| 317 | return; |
| 318 | } |
| 319 | /* |
| 320 | * There is a circular dependency. |
| 321 | * Some platforms allow un-removable section because they will just |
| 322 | * gather other removable sections for dynamic partitioning. |
| 323 | * Just notify un-removable section's number here. |
| 324 | */ |
| 325 | printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr, |
| 326 | pgdat_snr, nid); |
| 327 | printk(KERN_CONT |
| 328 | " have a circular dependency on usemap and pgdat allocations\n"); |
| 329 | } |
| 330 | #else |
| 331 | static unsigned long * __init |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 332 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, |
Johannes Weiner | 238305b | 2012-05-29 15:06:36 -0700 | [diff] [blame] | 333 | unsigned long size) |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 334 | { |
Johannes Weiner | 238305b | 2012-05-29 15:06:36 -0700 | [diff] [blame] | 335 | return alloc_bootmem_node_nopanic(pgdat, size); |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 336 | } |
| 337 | |
| 338 | static void __init check_usemap_section_nr(int nid, unsigned long *usemap) |
| 339 | { |
| 340 | } |
| 341 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
| 342 | |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 343 | static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map, |
| 344 | unsigned long pnum_begin, |
| 345 | unsigned long pnum_end, |
| 346 | unsigned long usemap_count, int nodeid) |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 347 | { |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 348 | void *usemap; |
| 349 | unsigned long pnum; |
| 350 | int size = usemap_size(); |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 351 | |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 352 | usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid), |
Johannes Weiner | 238305b | 2012-05-29 15:06:36 -0700 | [diff] [blame] | 353 | size * usemap_count); |
Nishanth Aravamudan | f5bf18f | 2012-03-21 16:34:07 -0700 | [diff] [blame] | 354 | if (!usemap) { |
Johannes Weiner | 238305b | 2012-05-29 15:06:36 -0700 | [diff] [blame] | 355 | printk(KERN_WARNING "%s: allocation failed\n", __func__); |
| 356 | return; |
Yasunori Goto | 48c9068 | 2008-07-23 21:28:15 -0700 | [diff] [blame] | 357 | } |
| 358 | |
Nishanth Aravamudan | f5bf18f | 2012-03-21 16:34:07 -0700 | [diff] [blame] | 359 | for (pnum = pnum_begin; pnum < pnum_end; pnum++) { |
| 360 | if (!present_section_nr(pnum)) |
| 361 | continue; |
| 362 | usemap_map[pnum] = usemap; |
| 363 | usemap += size; |
| 364 | check_usemap_section_nr(nodeid, usemap_map[pnum]); |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 365 | } |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 366 | } |
| 367 | |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 368 | #ifndef CONFIG_SPARSEMEM_VMEMMAP |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 369 | struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid) |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 370 | { |
| 371 | struct page *map; |
Yinghai Lu | e48e67e | 2010-05-24 14:31:57 -0700 | [diff] [blame] | 372 | unsigned long size; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 373 | |
| 374 | map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION); |
| 375 | if (map) |
| 376 | return map; |
| 377 | |
Yinghai Lu | e48e67e | 2010-05-24 14:31:57 -0700 | [diff] [blame] | 378 | size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION); |
| 379 | map = __alloc_bootmem_node_high(NODE_DATA(nid), size, |
| 380 | PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 381 | return map; |
| 382 | } |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 383 | void __init sparse_mem_maps_populate_node(struct page **map_map, |
| 384 | unsigned long pnum_begin, |
| 385 | unsigned long pnum_end, |
| 386 | unsigned long map_count, int nodeid) |
| 387 | { |
| 388 | void *map; |
| 389 | unsigned long pnum; |
| 390 | unsigned long size = sizeof(struct page) * PAGES_PER_SECTION; |
| 391 | |
| 392 | map = alloc_remap(nodeid, size * map_count); |
| 393 | if (map) { |
| 394 | for (pnum = pnum_begin; pnum < pnum_end; pnum++) { |
| 395 | if (!present_section_nr(pnum)) |
| 396 | continue; |
| 397 | map_map[pnum] = map; |
| 398 | map += size; |
| 399 | } |
| 400 | return; |
| 401 | } |
| 402 | |
| 403 | size = PAGE_ALIGN(size); |
Yinghai Lu | e48e67e | 2010-05-24 14:31:57 -0700 | [diff] [blame] | 404 | map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count, |
| 405 | PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 406 | if (map) { |
| 407 | for (pnum = pnum_begin; pnum < pnum_end; pnum++) { |
| 408 | if (!present_section_nr(pnum)) |
| 409 | continue; |
| 410 | map_map[pnum] = map; |
| 411 | map += size; |
| 412 | } |
| 413 | return; |
| 414 | } |
| 415 | |
| 416 | /* fallback */ |
| 417 | for (pnum = pnum_begin; pnum < pnum_end; pnum++) { |
| 418 | struct mem_section *ms; |
| 419 | |
| 420 | if (!present_section_nr(pnum)) |
| 421 | continue; |
| 422 | map_map[pnum] = sparse_mem_map_populate(pnum, nodeid); |
| 423 | if (map_map[pnum]) |
| 424 | continue; |
| 425 | ms = __nr_to_section(pnum); |
| 426 | printk(KERN_ERR "%s: sparsemem memory map backing failed " |
| 427 | "some memory will not be available.\n", __func__); |
| 428 | ms->section_mem_map = 0; |
| 429 | } |
| 430 | } |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 431 | #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ |
| 432 | |
Yinghai Lu | 81d0d95 | 2010-02-27 09:29:38 -0800 | [diff] [blame] | 433 | #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 434 | static void __init sparse_early_mem_maps_alloc_node(struct page **map_map, |
| 435 | unsigned long pnum_begin, |
| 436 | unsigned long pnum_end, |
| 437 | unsigned long map_count, int nodeid) |
| 438 | { |
| 439 | sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end, |
| 440 | map_count, nodeid); |
| 441 | } |
Yinghai Lu | 81d0d95 | 2010-02-27 09:29:38 -0800 | [diff] [blame] | 442 | #else |
Adrian Bunk | 9e5c6da | 2008-07-25 19:46:22 -0700 | [diff] [blame] | 443 | static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum) |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 444 | { |
| 445 | struct page *map; |
| 446 | struct mem_section *ms = __nr_to_section(pnum); |
| 447 | int nid = sparse_early_nid(ms); |
| 448 | |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 449 | map = sparse_mem_map_populate(pnum, nid); |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 450 | if (map) |
| 451 | return map; |
| 452 | |
Christoph Lameter | 8f6aac4 | 2007-10-16 01:24:13 -0700 | [diff] [blame] | 453 | printk(KERN_ERR "%s: sparsemem memory map backing failed " |
Harvey Harrison | d40cee2 | 2008-04-30 00:55:07 -0700 | [diff] [blame] | 454 | "some memory will not be available.\n", __func__); |
Bob Picco | 802f192 | 2005-09-03 15:54:26 -0700 | [diff] [blame] | 455 | ms->section_mem_map = 0; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 456 | return NULL; |
| 457 | } |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 458 | #endif |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 459 | |
Yinghai Lu | c2b91e2 | 2008-04-12 01:19:24 -0700 | [diff] [blame] | 460 | void __attribute__((weak)) __meminit vmemmap_populate_print_last(void) |
| 461 | { |
| 462 | } |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 463 | |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 464 | /* |
| 465 | * Allocate the accumulated non-linear sections, allocate a mem_map |
| 466 | * for each and record the physical to section mapping. |
| 467 | */ |
| 468 | void __init sparse_init(void) |
| 469 | { |
| 470 | unsigned long pnum; |
| 471 | struct page *map; |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 472 | unsigned long *usemap; |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 473 | unsigned long **usemap_map; |
Yinghai Lu | 81d0d95 | 2010-02-27 09:29:38 -0800 | [diff] [blame] | 474 | int size; |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 475 | int nodeid_begin = 0; |
| 476 | unsigned long pnum_begin = 0; |
| 477 | unsigned long usemap_count; |
Yinghai Lu | 81d0d95 | 2010-02-27 09:29:38 -0800 | [diff] [blame] | 478 | #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 479 | unsigned long map_count; |
Yinghai Lu | 81d0d95 | 2010-02-27 09:29:38 -0800 | [diff] [blame] | 480 | int size2; |
| 481 | struct page **map_map; |
| 482 | #endif |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 483 | |
Xishi Qiu | ca57df7 | 2012-07-31 16:43:19 -0700 | [diff] [blame] | 484 | /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */ |
| 485 | set_pageblock_order(); |
| 486 | |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 487 | /* |
| 488 | * map is using big page (aka 2M in x86 64 bit) |
| 489 | * usemap is less one page (aka 24 bytes) |
| 490 | * so alloc 2M (with 2M align) and 24 bytes in turn will |
| 491 | * make next 2M slip to one more 2M later. |
| 492 | * then in big system, the memory will have a lot of holes... |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 493 | * here try to allocate 2M pages continuously. |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 494 | * |
| 495 | * powerpc need to call sparse_init_one_section right after each |
| 496 | * sparse_early_mem_map_alloc, so allocate usemap_map at first. |
| 497 | */ |
| 498 | size = sizeof(unsigned long *) * NR_MEM_SECTIONS; |
| 499 | usemap_map = alloc_bootmem(size); |
| 500 | if (!usemap_map) |
| 501 | panic("can not allocate usemap_map\n"); |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 502 | |
| 503 | for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 504 | struct mem_section *ms; |
| 505 | |
Andy Whitcroft | 540557b | 2007-10-16 01:24:11 -0700 | [diff] [blame] | 506 | if (!present_section_nr(pnum)) |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 507 | continue; |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 508 | ms = __nr_to_section(pnum); |
| 509 | nodeid_begin = sparse_early_nid(ms); |
| 510 | pnum_begin = pnum; |
| 511 | break; |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 512 | } |
Yinghai Lu | a4322e1 | 2010-02-10 01:20:21 -0800 | [diff] [blame] | 513 | usemap_count = 1; |
| 514 | for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) { |
| 515 | struct mem_section *ms; |
| 516 | int nodeid; |
| 517 | |
| 518 | if (!present_section_nr(pnum)) |
| 519 | continue; |
| 520 | ms = __nr_to_section(pnum); |
| 521 | nodeid = sparse_early_nid(ms); |
| 522 | if (nodeid == nodeid_begin) { |
| 523 | usemap_count++; |
| 524 | continue; |
| 525 | } |
| 526 | /* ok, we need to take cake of from pnum_begin to pnum - 1*/ |
| 527 | sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum, |
| 528 | usemap_count, nodeid_begin); |
| 529 | /* new start, update count etc*/ |
| 530 | nodeid_begin = nodeid; |
| 531 | pnum_begin = pnum; |
| 532 | usemap_count = 1; |
| 533 | } |
| 534 | /* ok, last chunk */ |
| 535 | sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS, |
| 536 | usemap_count, nodeid_begin); |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 537 | |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 538 | #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER |
| 539 | size2 = sizeof(struct page *) * NR_MEM_SECTIONS; |
| 540 | map_map = alloc_bootmem(size2); |
| 541 | if (!map_map) |
| 542 | panic("can not allocate map_map\n"); |
| 543 | |
| 544 | for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { |
| 545 | struct mem_section *ms; |
| 546 | |
| 547 | if (!present_section_nr(pnum)) |
| 548 | continue; |
| 549 | ms = __nr_to_section(pnum); |
| 550 | nodeid_begin = sparse_early_nid(ms); |
| 551 | pnum_begin = pnum; |
| 552 | break; |
| 553 | } |
| 554 | map_count = 1; |
| 555 | for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) { |
| 556 | struct mem_section *ms; |
| 557 | int nodeid; |
| 558 | |
| 559 | if (!present_section_nr(pnum)) |
| 560 | continue; |
| 561 | ms = __nr_to_section(pnum); |
| 562 | nodeid = sparse_early_nid(ms); |
| 563 | if (nodeid == nodeid_begin) { |
| 564 | map_count++; |
| 565 | continue; |
| 566 | } |
| 567 | /* ok, we need to take cake of from pnum_begin to pnum - 1*/ |
| 568 | sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum, |
| 569 | map_count, nodeid_begin); |
| 570 | /* new start, update count etc*/ |
| 571 | nodeid_begin = nodeid; |
| 572 | pnum_begin = pnum; |
| 573 | map_count = 1; |
| 574 | } |
| 575 | /* ok, last chunk */ |
| 576 | sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS, |
| 577 | map_count, nodeid_begin); |
| 578 | #endif |
| 579 | |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 580 | for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { |
| 581 | if (!present_section_nr(pnum)) |
| 582 | continue; |
| 583 | |
| 584 | usemap = usemap_map[pnum]; |
| 585 | if (!usemap) |
| 586 | continue; |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 587 | |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 588 | #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER |
| 589 | map = map_map[pnum]; |
| 590 | #else |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 591 | map = sparse_early_mem_map_alloc(pnum); |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 592 | #endif |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 593 | if (!map) |
| 594 | continue; |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 595 | |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 596 | sparse_init_one_section(__nr_to_section(pnum), pnum, map, |
| 597 | usemap); |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 598 | } |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 599 | |
Yinghai Lu | c2b91e2 | 2008-04-12 01:19:24 -0700 | [diff] [blame] | 600 | vmemmap_populate_print_last(); |
| 601 | |
Yinghai Lu | 9bdac91 | 2010-02-10 01:20:22 -0800 | [diff] [blame] | 602 | #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER |
| 603 | free_bootmem(__pa(map_map), size2); |
| 604 | #endif |
Yinghai Lu | e123dd3 | 2008-04-13 11:51:06 -0700 | [diff] [blame] | 605 | free_bootmem(__pa(usemap_map), size); |
Stephen Rothwell | 193faea | 2007-06-08 13:46:51 -0700 | [diff] [blame] | 606 | } |
| 607 | |
| 608 | #ifdef CONFIG_MEMORY_HOTPLUG |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 609 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
| 610 | static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid, |
| 611 | unsigned long nr_pages) |
| 612 | { |
| 613 | /* This will make the necessary allocations eventually. */ |
| 614 | return sparse_mem_map_populate(pnum, nid); |
| 615 | } |
| 616 | static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages) |
| 617 | { |
| 618 | return; /* XXX: Not implemented yet */ |
| 619 | } |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 620 | static void free_map_bootmem(struct page *page, unsigned long nr_pages) |
| 621 | { |
| 622 | } |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 623 | #else |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 624 | static struct page *__kmalloc_section_memmap(unsigned long nr_pages) |
| 625 | { |
| 626 | struct page *page, *ret; |
| 627 | unsigned long memmap_size = sizeof(struct page) * nr_pages; |
| 628 | |
Yasunori Goto | f2d0aa5 | 2006-10-28 10:38:32 -0700 | [diff] [blame] | 629 | page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size)); |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 630 | if (page) |
| 631 | goto got_map_page; |
| 632 | |
| 633 | ret = vmalloc(memmap_size); |
| 634 | if (ret) |
| 635 | goto got_map_ptr; |
| 636 | |
| 637 | return NULL; |
| 638 | got_map_page: |
| 639 | ret = (struct page *)pfn_to_kaddr(page_to_pfn(page)); |
| 640 | got_map_ptr: |
| 641 | memset(ret, 0, memmap_size); |
| 642 | |
| 643 | return ret; |
| 644 | } |
| 645 | |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 646 | static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid, |
| 647 | unsigned long nr_pages) |
| 648 | { |
| 649 | return __kmalloc_section_memmap(nr_pages); |
| 650 | } |
| 651 | |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 652 | static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages) |
| 653 | { |
Christoph Lameter | 9e2779f | 2008-02-04 22:28:34 -0800 | [diff] [blame] | 654 | if (is_vmalloc_addr(memmap)) |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 655 | vfree(memmap); |
| 656 | else |
| 657 | free_pages((unsigned long)memmap, |
| 658 | get_order(sizeof(struct page) * nr_pages)); |
| 659 | } |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 660 | |
| 661 | static void free_map_bootmem(struct page *page, unsigned long nr_pages) |
| 662 | { |
| 663 | unsigned long maps_section_nr, removing_section_nr, i; |
Andrea Arcangeli | 5f24ce5 | 2011-01-13 15:47:00 -0800 | [diff] [blame] | 664 | unsigned long magic; |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 665 | |
| 666 | for (i = 0; i < nr_pages; i++, page++) { |
Andrea Arcangeli | 5f24ce5 | 2011-01-13 15:47:00 -0800 | [diff] [blame] | 667 | magic = (unsigned long) page->lru.next; |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 668 | |
| 669 | BUG_ON(magic == NODE_INFO); |
| 670 | |
| 671 | maps_section_nr = pfn_to_section_nr(page_to_pfn(page)); |
| 672 | removing_section_nr = page->private; |
| 673 | |
| 674 | /* |
| 675 | * When this function is called, the removing section is |
| 676 | * logical offlined state. This means all pages are isolated |
| 677 | * from page allocator. If removing section's memmap is placed |
| 678 | * on the same section, it must not be freed. |
| 679 | * If it is freed, page allocator may allocate it which will |
| 680 | * be removed physically soon. |
| 681 | */ |
| 682 | if (maps_section_nr != removing_section_nr) |
| 683 | put_page_bootmem(page); |
| 684 | } |
| 685 | } |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 686 | #endif /* CONFIG_SPARSEMEM_VMEMMAP */ |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 687 | |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 688 | static void free_section_usemap(struct page *memmap, unsigned long *usemap) |
| 689 | { |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 690 | struct page *usemap_page; |
| 691 | unsigned long nr_pages; |
| 692 | |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 693 | if (!usemap) |
| 694 | return; |
| 695 | |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 696 | usemap_page = virt_to_page(usemap); |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 697 | /* |
| 698 | * Check to see if allocation came from hot-plug-add |
| 699 | */ |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 700 | if (PageSlab(usemap_page)) { |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 701 | kfree(usemap); |
| 702 | if (memmap) |
| 703 | __kfree_section_memmap(memmap, PAGES_PER_SECTION); |
| 704 | return; |
| 705 | } |
| 706 | |
| 707 | /* |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 708 | * The usemap came from bootmem. This is packed with other usemaps |
| 709 | * on the section which has pgdat at boot time. Just keep it as is now. |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 710 | */ |
Yasunori Goto | 0c0a4a5 | 2008-04-28 02:13:34 -0700 | [diff] [blame] | 711 | |
| 712 | if (memmap) { |
| 713 | struct page *memmap_page; |
| 714 | memmap_page = virt_to_page(memmap); |
| 715 | |
| 716 | nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page)) |
| 717 | >> PAGE_SHIFT; |
| 718 | |
| 719 | free_map_bootmem(memmap_page, nr_pages); |
| 720 | } |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 721 | } |
| 722 | |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 723 | /* |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 724 | * returns the number of sections whose mem_maps were properly |
| 725 | * set. If this is <=0, then that means that the passed-in |
| 726 | * map was not consumed and must be freed. |
| 727 | */ |
Al Viro | 3116848 | 2008-11-22 17:33:24 +0000 | [diff] [blame] | 728 | int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn, |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 729 | int nr_pages) |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 730 | { |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 731 | unsigned long section_nr = pfn_to_section_nr(start_pfn); |
| 732 | struct pglist_data *pgdat = zone->zone_pgdat; |
| 733 | struct mem_section *ms; |
| 734 | struct page *memmap; |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 735 | unsigned long *usemap; |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 736 | unsigned long flags; |
| 737 | int ret; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 738 | |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 739 | /* |
| 740 | * no locking for this, because it does its own |
| 741 | * plus, it does a kmalloc |
| 742 | */ |
WANG Cong | bbd0682 | 2007-12-17 16:19:59 -0800 | [diff] [blame] | 743 | ret = sparse_index_init(section_nr, pgdat->node_id); |
| 744 | if (ret < 0 && ret != -EEXIST) |
| 745 | return ret; |
Yasunori Goto | 98f3cfc | 2007-10-16 01:26:14 -0700 | [diff] [blame] | 746 | memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages); |
WANG Cong | bbd0682 | 2007-12-17 16:19:59 -0800 | [diff] [blame] | 747 | if (!memmap) |
| 748 | return -ENOMEM; |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 749 | usemap = __kmalloc_section_usemap(); |
WANG Cong | bbd0682 | 2007-12-17 16:19:59 -0800 | [diff] [blame] | 750 | if (!usemap) { |
| 751 | __kfree_section_memmap(memmap, nr_pages); |
| 752 | return -ENOMEM; |
| 753 | } |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 754 | |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 755 | pgdat_resize_lock(pgdat, &flags); |
| 756 | |
| 757 | ms = __pfn_to_section(start_pfn); |
| 758 | if (ms->section_mem_map & SECTION_MARKED_PRESENT) { |
| 759 | ret = -EEXIST; |
| 760 | goto out; |
| 761 | } |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 762 | |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 763 | ms->section_mem_map |= SECTION_MARKED_PRESENT; |
| 764 | |
Mel Gorman | 5c0e306 | 2007-10-16 01:25:56 -0700 | [diff] [blame] | 765 | ret = sparse_init_one_section(ms, section_nr, memmap, usemap); |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 766 | |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 767 | out: |
| 768 | pgdat_resize_unlock(pgdat, &flags); |
WANG Cong | bbd0682 | 2007-12-17 16:19:59 -0800 | [diff] [blame] | 769 | if (ret <= 0) { |
| 770 | kfree(usemap); |
Mike Kravetz | 46a66ee | 2006-05-01 12:16:09 -0700 | [diff] [blame] | 771 | __kfree_section_memmap(memmap, nr_pages); |
WANG Cong | bbd0682 | 2007-12-17 16:19:59 -0800 | [diff] [blame] | 772 | } |
Dave Hansen | 0b0acbe | 2005-10-29 18:16:55 -0700 | [diff] [blame] | 773 | return ret; |
Andy Whitcroft | 29751f6 | 2005-06-23 00:08:00 -0700 | [diff] [blame] | 774 | } |
Badari Pulavarty | ea01ea9 | 2008-04-28 02:12:01 -0700 | [diff] [blame] | 775 | |
| 776 | void sparse_remove_one_section(struct zone *zone, struct mem_section *ms) |
| 777 | { |
| 778 | struct page *memmap = NULL; |
| 779 | unsigned long *usemap = NULL; |
| 780 | |
| 781 | if (ms->section_mem_map) { |
| 782 | usemap = ms->pageblock_flags; |
| 783 | memmap = sparse_decode_mem_map(ms->section_mem_map, |
| 784 | __section_nr(ms)); |
| 785 | ms->section_mem_map = 0; |
| 786 | ms->pageblock_flags = NULL; |
| 787 | } |
| 788 | |
| 789 | free_section_usemap(memmap, usemap); |
| 790 | } |
Yasunori Goto | a3142c8 | 2007-05-08 00:23:07 -0700 | [diff] [blame] | 791 | #endif |