blob: e06f514fe04faa748c5e50e3f4cb25789d977238 [file] [log] [blame]
Andy Whitcroftd41dee32005-06-23 00:07:54 -07001/*
2 * sparse memory mappings.
3 */
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004#include <linux/mm.h>
5#include <linux/mmzone.h>
6#include <linux/bootmem.h>
Dave Hansen0b0acbe2005-10-29 18:16:55 -07007#include <linux/highmem.h>
Andy Whitcroftd41dee32005-06-23 00:07:54 -07008#include <linux/module.h>
Dave Hansen28ae55c2005-09-03 15:54:29 -07009#include <linux/spinlock.h>
Dave Hansen0b0acbe2005-10-29 18:16:55 -070010#include <linux/vmalloc.h>
Andy Whitcroftd41dee32005-06-23 00:07:54 -070011#include <asm/dma.h>
Christoph Lameter8f6aac42007-10-16 01:24:13 -070012#include <asm/pgalloc.h>
13#include <asm/pgtable.h>
Andy Whitcroftd41dee32005-06-23 00:07:54 -070014
15/*
16 * Permanent SPARSEMEM data:
17 *
18 * 1) mem_section - memory sections, mem_map's for valid memory
19 */
Bob Picco3e347262005-09-03 15:54:28 -070020#ifdef CONFIG_SPARSEMEM_EXTREME
Bob Picco802f1922005-09-03 15:54:26 -070021struct mem_section *mem_section[NR_SECTION_ROOTS]
Ravikiran G Thirumalai22fc6ec2006-01-08 01:01:27 -080022 ____cacheline_internodealigned_in_smp;
Bob Picco3e347262005-09-03 15:54:28 -070023#else
24struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
Ravikiran G Thirumalai22fc6ec2006-01-08 01:01:27 -080025 ____cacheline_internodealigned_in_smp;
Bob Picco3e347262005-09-03 15:54:28 -070026#endif
27EXPORT_SYMBOL(mem_section);
28
Christoph Lameter89689ae2006-12-06 20:31:45 -080029#ifdef NODE_NOT_IN_PAGE_FLAGS
30/*
31 * If we did not store the node number in the page then we have to
32 * do a lookup in the section_to_node_table in order to find which
33 * node the page belongs to.
34 */
35#if MAX_NUMNODES <= 256
36static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
37#else
38static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39#endif
40
Andy Whitcroft25ba77c2006-12-06 20:33:03 -080041int page_to_nid(struct page *page)
Christoph Lameter89689ae2006-12-06 20:31:45 -080042{
43 return section_to_node_table[page_to_section(page)];
44}
45EXPORT_SYMBOL(page_to_nid);
Andy Whitcroft85770ff2007-08-22 14:01:03 -070046
47static void set_section_nid(unsigned long section_nr, int nid)
48{
49 section_to_node_table[section_nr] = nid;
50}
51#else /* !NODE_NOT_IN_PAGE_FLAGS */
52static inline void set_section_nid(unsigned long section_nr, int nid)
53{
54}
Christoph Lameter89689ae2006-12-06 20:31:45 -080055#endif
56
Bob Picco3e347262005-09-03 15:54:28 -070057#ifdef CONFIG_SPARSEMEM_EXTREME
Sam Ravnborg577a32f2007-05-17 23:29:25 +020058static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
Bob Picco802f1922005-09-03 15:54:26 -070059{
Dave Hansen28ae55c2005-09-03 15:54:29 -070060 struct mem_section *section = NULL;
61 unsigned long array_size = SECTIONS_PER_ROOT *
62 sizeof(struct mem_section);
Bob Picco802f1922005-09-03 15:54:26 -070063
Mike Kravetz39d24e62006-05-15 09:44:13 -070064 if (slab_is_available())
Mike Kravetz46a66ee2006-05-01 12:16:09 -070065 section = kmalloc_node(array_size, GFP_KERNEL, nid);
66 else
67 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
Bob Picco3e347262005-09-03 15:54:28 -070068
Dave Hansen28ae55c2005-09-03 15:54:29 -070069 if (section)
70 memset(section, 0, array_size);
Bob Picco3e347262005-09-03 15:54:28 -070071
Dave Hansen28ae55c2005-09-03 15:54:29 -070072 return section;
Bob Picco802f1922005-09-03 15:54:26 -070073}
Dave Hansen28ae55c2005-09-03 15:54:29 -070074
Yasunori Gotoa3142c82007-05-08 00:23:07 -070075static int __meminit sparse_index_init(unsigned long section_nr, int nid)
Dave Hansen28ae55c2005-09-03 15:54:29 -070076{
Ingo Molnar34af9462006-06-27 02:53:55 -070077 static DEFINE_SPINLOCK(index_init_lock);
Dave Hansen28ae55c2005-09-03 15:54:29 -070078 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
79 struct mem_section *section;
80 int ret = 0;
81
82 if (mem_section[root])
83 return -EEXIST;
84
85 section = sparse_index_alloc(nid);
86 /*
87 * This lock keeps two different sections from
88 * reallocating for the same index
89 */
90 spin_lock(&index_init_lock);
91
92 if (mem_section[root]) {
93 ret = -EEXIST;
94 goto out;
95 }
96
97 mem_section[root] = section;
98out:
99 spin_unlock(&index_init_lock);
100 return ret;
101}
102#else /* !SPARSEMEM_EXTREME */
103static inline int sparse_index_init(unsigned long section_nr, int nid)
104{
105 return 0;
106}
107#endif
108
Dave Hansen4ca644d2005-10-29 18:16:51 -0700109/*
110 * Although written for the SPARSEMEM_EXTREME case, this happens
Andy Whitcroftcd881a62007-10-16 01:24:10 -0700111 * to also work for the flat array case because
Dave Hansen4ca644d2005-10-29 18:16:51 -0700112 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
113 */
114int __section_nr(struct mem_section* ms)
115{
116 unsigned long root_nr;
117 struct mem_section* root;
118
Mike Kravetz12783b02006-05-20 15:00:05 -0700119 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
120 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
Dave Hansen4ca644d2005-10-29 18:16:51 -0700121 if (!root)
122 continue;
123
124 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
125 break;
126 }
127
128 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
129}
130
Andy Whitcroft30c253e2006-06-23 02:03:41 -0700131/*
132 * During early boot, before section_mem_map is used for an actual
133 * mem_map, we use section_mem_map to store the section's NUMA
134 * node. This keeps us from having to use another data structure. The
135 * node information is cleared just before we store the real mem_map.
136 */
137static inline unsigned long sparse_encode_early_nid(int nid)
138{
139 return (nid << SECTION_NID_SHIFT);
140}
141
142static inline int sparse_early_nid(struct mem_section *section)
143{
144 return (section->section_mem_map >> SECTION_NID_SHIFT);
145}
146
Andy Whitcroftd41dee32005-06-23 00:07:54 -0700147/* Record a memory area against a node. */
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700148void __init memory_present(int nid, unsigned long start, unsigned long end)
Andy Whitcroftd41dee32005-06-23 00:07:54 -0700149{
150 unsigned long pfn;
151
152 start &= PAGE_SECTION_MASK;
153 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
154 unsigned long section = pfn_to_section_nr(pfn);
Bob Picco802f1922005-09-03 15:54:26 -0700155 struct mem_section *ms;
156
157 sparse_index_init(section, nid);
Andy Whitcroft85770ff2007-08-22 14:01:03 -0700158 set_section_nid(section, nid);
Bob Picco802f1922005-09-03 15:54:26 -0700159
160 ms = __nr_to_section(section);
161 if (!ms->section_mem_map)
Andy Whitcroft30c253e2006-06-23 02:03:41 -0700162 ms->section_mem_map = sparse_encode_early_nid(nid) |
163 SECTION_MARKED_PRESENT;
Andy Whitcroftd41dee32005-06-23 00:07:54 -0700164 }
165}
166
167/*
168 * Only used by the i386 NUMA architecures, but relatively
169 * generic code.
170 */
171unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
172 unsigned long end_pfn)
173{
174 unsigned long pfn;
175 unsigned long nr_pages = 0;
176
177 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
178 if (nid != early_pfn_to_nid(pfn))
179 continue;
180
Andy Whitcroft540557b2007-10-16 01:24:11 -0700181 if (pfn_present(pfn))
Andy Whitcroftd41dee32005-06-23 00:07:54 -0700182 nr_pages += PAGES_PER_SECTION;
183 }
184
185 return nr_pages * sizeof(struct page);
186}
187
188/*
Andy Whitcroft29751f62005-06-23 00:08:00 -0700189 * Subtle, we encode the real pfn into the mem_map such that
190 * the identity pfn - section_mem_map will return the actual
191 * physical page frame number.
192 */
193static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
194{
195 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
196}
197
198/*
199 * We need this if we ever free the mem_maps. While not implemented yet,
200 * this function is included for parity with its sibling.
201 */
202static __attribute((unused))
203struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
204{
205 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
206}
207
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700208static int __meminit sparse_init_one_section(struct mem_section *ms,
Mel Gorman5c0e3062007-10-16 01:25:56 -0700209 unsigned long pnum, struct page *mem_map,
210 unsigned long *pageblock_bitmap)
Andy Whitcroft29751f62005-06-23 00:08:00 -0700211{
Andy Whitcroft540557b2007-10-16 01:24:11 -0700212 if (!present_section(ms))
Andy Whitcroft29751f62005-06-23 00:08:00 -0700213 return -EINVAL;
214
Andy Whitcroft30c253e2006-06-23 02:03:41 -0700215 ms->section_mem_map &= ~SECTION_MAP_MASK;
Andy Whitcroft540557b2007-10-16 01:24:11 -0700216 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
217 SECTION_HAS_MEM_MAP;
Mel Gorman5c0e3062007-10-16 01:25:56 -0700218 ms->pageblock_flags = pageblock_bitmap;
Andy Whitcroft29751f62005-06-23 00:08:00 -0700219
220 return 1;
221}
222
Mel Gorman5c0e3062007-10-16 01:25:56 -0700223static unsigned long usemap_size(void)
224{
225 unsigned long size_bytes;
226 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
227 size_bytes = roundup(size_bytes, sizeof(unsigned long));
228 return size_bytes;
229}
230
231#ifdef CONFIG_MEMORY_HOTPLUG
232static unsigned long *__kmalloc_section_usemap(void)
233{
234 return kmalloc(usemap_size(), GFP_KERNEL);
235}
236#endif /* CONFIG_MEMORY_HOTPLUG */
237
238static unsigned long *sparse_early_usemap_alloc(unsigned long pnum)
239{
240 unsigned long *usemap;
241 struct mem_section *ms = __nr_to_section(pnum);
242 int nid = sparse_early_nid(ms);
243
244 usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
245 if (usemap)
246 return usemap;
247
248 /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
249 nid = 0;
250
251 printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
252 return NULL;
253}
254
Christoph Lameter8f6aac42007-10-16 01:24:13 -0700255#ifndef CONFIG_SPARSEMEM_VMEMMAP
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700256struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
Andy Whitcroft29751f62005-06-23 00:08:00 -0700257{
258 struct page *map;
Andy Whitcroft29751f62005-06-23 00:08:00 -0700259
260 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
261 if (map)
262 return map;
263
264 map = alloc_bootmem_node(NODE_DATA(nid),
265 sizeof(struct page) * PAGES_PER_SECTION);
Christoph Lameter8f6aac42007-10-16 01:24:13 -0700266 return map;
267}
268#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
269
270struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
271{
272 struct page *map;
273 struct mem_section *ms = __nr_to_section(pnum);
274 int nid = sparse_early_nid(ms);
275
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700276 map = sparse_mem_map_populate(pnum, nid);
Andy Whitcroft29751f62005-06-23 00:08:00 -0700277 if (map)
278 return map;
279
Christoph Lameter8f6aac42007-10-16 01:24:13 -0700280 printk(KERN_ERR "%s: sparsemem memory map backing failed "
281 "some memory will not be available.\n", __FUNCTION__);
Bob Picco802f1922005-09-03 15:54:26 -0700282 ms->section_mem_map = 0;
Andy Whitcroft29751f62005-06-23 00:08:00 -0700283 return NULL;
284}
285
Stephen Rothwell193faea2007-06-08 13:46:51 -0700286/*
287 * Allocate the accumulated non-linear sections, allocate a mem_map
288 * for each and record the physical to section mapping.
289 */
290void __init sparse_init(void)
291{
292 unsigned long pnum;
293 struct page *map;
Mel Gorman5c0e3062007-10-16 01:25:56 -0700294 unsigned long *usemap;
Stephen Rothwell193faea2007-06-08 13:46:51 -0700295
296 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
Andy Whitcroft540557b2007-10-16 01:24:11 -0700297 if (!present_section_nr(pnum))
Stephen Rothwell193faea2007-06-08 13:46:51 -0700298 continue;
299
300 map = sparse_early_mem_map_alloc(pnum);
301 if (!map)
302 continue;
Mel Gorman5c0e3062007-10-16 01:25:56 -0700303
304 usemap = sparse_early_usemap_alloc(pnum);
305 if (!usemap)
306 continue;
307
308 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
309 usemap);
Stephen Rothwell193faea2007-06-08 13:46:51 -0700310 }
311}
312
313#ifdef CONFIG_MEMORY_HOTPLUG
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700314#ifdef CONFIG_SPARSEMEM_VMEMMAP
315static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
316 unsigned long nr_pages)
317{
318 /* This will make the necessary allocations eventually. */
319 return sparse_mem_map_populate(pnum, nid);
320}
321static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
322{
323 return; /* XXX: Not implemented yet */
324}
325#else
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700326static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
327{
328 struct page *page, *ret;
329 unsigned long memmap_size = sizeof(struct page) * nr_pages;
330
Yasunori Gotof2d0aa52006-10-28 10:38:32 -0700331 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700332 if (page)
333 goto got_map_page;
334
335 ret = vmalloc(memmap_size);
336 if (ret)
337 goto got_map_ptr;
338
339 return NULL;
340got_map_page:
341 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
342got_map_ptr:
343 memset(ret, 0, memmap_size);
344
345 return ret;
346}
347
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700348static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
349 unsigned long nr_pages)
350{
351 return __kmalloc_section_memmap(nr_pages);
352}
353
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700354static int vaddr_in_vmalloc_area(void *addr)
355{
356 if (addr >= (void *)VMALLOC_START &&
357 addr < (void *)VMALLOC_END)
358 return 1;
359 return 0;
360}
361
362static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
363{
364 if (vaddr_in_vmalloc_area(memmap))
365 vfree(memmap);
366 else
367 free_pages((unsigned long)memmap,
368 get_order(sizeof(struct page) * nr_pages));
369}
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700370#endif /* CONFIG_SPARSEMEM_VMEMMAP */
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700371
Andy Whitcroft29751f62005-06-23 00:08:00 -0700372/*
Andy Whitcroft29751f62005-06-23 00:08:00 -0700373 * returns the number of sections whose mem_maps were properly
374 * set. If this is <=0, then that means that the passed-in
375 * map was not consumed and must be freed.
376 */
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700377int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
378 int nr_pages)
Andy Whitcroft29751f62005-06-23 00:08:00 -0700379{
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700380 unsigned long section_nr = pfn_to_section_nr(start_pfn);
381 struct pglist_data *pgdat = zone->zone_pgdat;
382 struct mem_section *ms;
383 struct page *memmap;
Mel Gorman5c0e3062007-10-16 01:25:56 -0700384 unsigned long *usemap;
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700385 unsigned long flags;
386 int ret;
Andy Whitcroft29751f62005-06-23 00:08:00 -0700387
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700388 /*
389 * no locking for this, because it does its own
390 * plus, it does a kmalloc
391 */
392 sparse_index_init(section_nr, pgdat->node_id);
Yasunori Goto98f3cfc2007-10-16 01:26:14 -0700393 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
Mel Gorman5c0e3062007-10-16 01:25:56 -0700394 usemap = __kmalloc_section_usemap();
Andy Whitcroft29751f62005-06-23 00:08:00 -0700395
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700396 pgdat_resize_lock(pgdat, &flags);
397
398 ms = __pfn_to_section(start_pfn);
399 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
400 ret = -EEXIST;
401 goto out;
402 }
Mel Gorman5c0e3062007-10-16 01:25:56 -0700403
404 if (!usemap) {
405 ret = -ENOMEM;
406 goto out;
407 }
Andy Whitcroft29751f62005-06-23 00:08:00 -0700408 ms->section_mem_map |= SECTION_MARKED_PRESENT;
409
Mel Gorman5c0e3062007-10-16 01:25:56 -0700410 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700411
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700412out:
413 pgdat_resize_unlock(pgdat, &flags);
Mike Kravetz46a66ee2006-05-01 12:16:09 -0700414 if (ret <= 0)
415 __kfree_section_memmap(memmap, nr_pages);
Dave Hansen0b0acbe2005-10-29 18:16:55 -0700416 return ret;
Andy Whitcroft29751f62005-06-23 00:08:00 -0700417}
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700418#endif