sparse.c revision 193faea9280a809cc30e81d7e503e01b1d7b7042 
1/*
2 * sparse memory mappings.
3 */
4#include <linux/mm.h>
5#include <linux/mmzone.h>
6#include <linux/bootmem.h>
7#include <linux/highmem.h>
8#include <linux/module.h>
9#include <linux/spinlock.h>
10#include <linux/vmalloc.h>
11#include <asm/dma.h>
12
13/*
14 * Permanent SPARSEMEM data:
15 *
16 * 1) mem_section	- memory sections, mem_map's for valid memory
17 */
18#ifdef CONFIG_SPARSEMEM_EXTREME
19struct mem_section *mem_section[NR_SECTION_ROOTS]
20	____cacheline_internodealigned_in_smp;
21#else
22struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
23	____cacheline_internodealigned_in_smp;
24#endif
25EXPORT_SYMBOL(mem_section);
26
27#ifdef NODE_NOT_IN_PAGE_FLAGS
28/*
29 * If we did not store the node number in the page then we have to
30 * do a lookup in the section_to_node_table in order to find which
31 * node the page belongs to.
32 */
33#if MAX_NUMNODES <= 256
34static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
35#else
36static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
37#endif
38
39int page_to_nid(struct page *page)
40{
41	return section_to_node_table[page_to_section(page)];
42}
43EXPORT_SYMBOL(page_to_nid);
44#endif
45
46#ifdef CONFIG_SPARSEMEM_EXTREME
47static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
48{
49	struct mem_section *section = NULL;
50	unsigned long array_size = SECTIONS_PER_ROOT *
51				   sizeof(struct mem_section);
52
53	if (slab_is_available())
54		section = kmalloc_node(array_size, GFP_KERNEL, nid);
55	else
56		section = alloc_bootmem_node(NODE_DATA(nid), array_size);
57
58	if (section)
59		memset(section, 0, array_size);
60
61	return section;
62}
63
64static int __meminit sparse_index_init(unsigned long section_nr, int nid)
65{
66	static DEFINE_SPINLOCK(index_init_lock);
67	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
68	struct mem_section *section;
69	int ret = 0;
70
71#ifdef NODE_NOT_IN_PAGE_FLAGS
72	section_to_node_table[section_nr] = nid;
73#endif
74
75	if (mem_section[root])
76		return -EEXIST;
77
78	section = sparse_index_alloc(nid);
79	/*
80	 * This lock keeps two different sections from
81	 * reallocating for the same index
82	 */
83	spin_lock(&index_init_lock);
84
85	if (mem_section[root]) {
86		ret = -EEXIST;
87		goto out;
88	}
89
90	mem_section[root] = section;
91out:
92	spin_unlock(&index_init_lock);
93	return ret;
94}
95#else /* !SPARSEMEM_EXTREME */
96static inline int sparse_index_init(unsigned long section_nr, int nid)
97{
98	return 0;
99}
100#endif
101
102/*
103 * Although written for the SPARSEMEM_EXTREME case, this happens
104 * to also work for the flat array case becase
105 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
106 */
107int __section_nr(struct mem_section* ms)
108{
109	unsigned long root_nr;
110	struct mem_section* root;
111
112	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
113		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
114		if (!root)
115			continue;
116
117		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
118		     break;
119	}
120
121	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
122}
123
124/*
125 * During early boot, before section_mem_map is used for an actual
126 * mem_map, we use section_mem_map to store the section's NUMA
127 * node.  This keeps us from having to use another data structure.  The
128 * node information is cleared just before we store the real mem_map.
129 */
130static inline unsigned long sparse_encode_early_nid(int nid)
131{
132	return (nid << SECTION_NID_SHIFT);
133}
134
135static inline int sparse_early_nid(struct mem_section *section)
136{
137	return (section->section_mem_map >> SECTION_NID_SHIFT);
138}
139
140/* Record a memory area against a node. */
141void __init memory_present(int nid, unsigned long start, unsigned long end)
142{
143	unsigned long pfn;
144
145	start &= PAGE_SECTION_MASK;
146	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
147		unsigned long section = pfn_to_section_nr(pfn);
148		struct mem_section *ms;
149
150		sparse_index_init(section, nid);
151
152		ms = __nr_to_section(section);
153		if (!ms->section_mem_map)
154			ms->section_mem_map = sparse_encode_early_nid(nid) |
155							SECTION_MARKED_PRESENT;
156	}
157}
158
159/*
160 * Only used by the i386 NUMA architecures, but relatively
161 * generic code.
162 */
163unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
164						     unsigned long end_pfn)
165{
166	unsigned long pfn;
167	unsigned long nr_pages = 0;
168
169	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
170		if (nid != early_pfn_to_nid(pfn))
171			continue;
172
173		if (pfn_valid(pfn))
174			nr_pages += PAGES_PER_SECTION;
175	}
176
177	return nr_pages * sizeof(struct page);
178}
179
180/*
181 * Subtle, we encode the real pfn into the mem_map such that
182 * the identity pfn - section_mem_map will return the actual
183 * physical page frame number.
184 */
185static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
186{
187	return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
188}
189
190/*
191 * We need this if we ever free the mem_maps.  While not implemented yet,
192 * this function is included for parity with its sibling.
193 */
194static __attribute((unused))
195struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
196{
197	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
198}
199
200static int __meminit sparse_init_one_section(struct mem_section *ms,
201		unsigned long pnum, struct page *mem_map)
202{
203	if (!valid_section(ms))
204		return -EINVAL;
205
206	ms->section_mem_map &= ~SECTION_MAP_MASK;
207	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
208
209	return 1;
210}
211
212__attribute__((weak))
213void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size)
214{
215	return NULL;
216}
217
218static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
219{
220	struct page *map;
221	struct mem_section *ms = __nr_to_section(pnum);
222	int nid = sparse_early_nid(ms);
223
224	map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
225	if (map)
226		return map;
227
228  	map = alloc_bootmem_high_node(NODE_DATA(nid),
229                       sizeof(struct page) * PAGES_PER_SECTION);
230	if (map)
231		return map;
232
233	map = alloc_bootmem_node(NODE_DATA(nid),
234			sizeof(struct page) * PAGES_PER_SECTION);
235	if (map)
236		return map;
237
238	printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
239	ms->section_mem_map = 0;
240	return NULL;
241}
242
243/*
244 * Allocate the accumulated non-linear sections, allocate a mem_map
245 * for each and record the physical to section mapping.
246 */
247void __init sparse_init(void)
248{
249	unsigned long pnum;
250	struct page *map;
251
252	for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
253		if (!valid_section_nr(pnum))
254			continue;
255
256		map = sparse_early_mem_map_alloc(pnum);
257		if (!map)
258			continue;
259		sparse_init_one_section(__nr_to_section(pnum), pnum, map);
260	}
261}
262
263#ifdef CONFIG_MEMORY_HOTPLUG
264static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
265{
266	struct page *page, *ret;
267	unsigned long memmap_size = sizeof(struct page) * nr_pages;
268
269	page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
270	if (page)
271		goto got_map_page;
272
273	ret = vmalloc(memmap_size);
274	if (ret)
275		goto got_map_ptr;
276
277	return NULL;
278got_map_page:
279	ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
280got_map_ptr:
281	memset(ret, 0, memmap_size);
282
283	return ret;
284}
285
286static int vaddr_in_vmalloc_area(void *addr)
287{
288	if (addr >= (void *)VMALLOC_START &&
289	    addr < (void *)VMALLOC_END)
290		return 1;
291	return 0;
292}
293
294static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
295{
296	if (vaddr_in_vmalloc_area(memmap))
297		vfree(memmap);
298	else
299		free_pages((unsigned long)memmap,
300			   get_order(sizeof(struct page) * nr_pages));
301}
302
303/*
304 * returns the number of sections whose mem_maps were properly
305 * set.  If this is <=0, then that means that the passed-in
306 * map was not consumed and must be freed.
307 */
308int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
309			   int nr_pages)
310{
311	unsigned long section_nr = pfn_to_section_nr(start_pfn);
312	struct pglist_data *pgdat = zone->zone_pgdat;
313	struct mem_section *ms;
314	struct page *memmap;
315	unsigned long flags;
316	int ret;
317
318	/*
319	 * no locking for this, because it does its own
320	 * plus, it does a kmalloc
321	 */
322	sparse_index_init(section_nr, pgdat->node_id);
323	memmap = __kmalloc_section_memmap(nr_pages);
324
325	pgdat_resize_lock(pgdat, &flags);
326
327	ms = __pfn_to_section(start_pfn);
328	if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
329		ret = -EEXIST;
330		goto out;
331	}
332	ms->section_mem_map |= SECTION_MARKED_PRESENT;
333
334	ret = sparse_init_one_section(ms, section_nr, memmap);
335
336out:
337	pgdat_resize_unlock(pgdat, &flags);
338	if (ret <= 0)
339		__kfree_section_memmap(memmap, nr_pages);
340	return ret;
341}
342#endif
343