xref: /linux/mm/sparse-vmemmap.c (revision d524dac9279b6a41ffdf7ff7958c577f2e387db6)
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/highmem.h>
24 #include <linux/module.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 * __init_refok __earlyonly_bootmem_alloc(int node,
40 				unsigned long size,
41 				unsigned long align,
42 				unsigned long goal)
43 {
44 	return __alloc_bootmem_node_high(NODE_DATA(node), size, align, goal);
45 }
46 
47 static void *vmemmap_buf;
48 static void *vmemmap_buf_end;
49 
50 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
51 {
52 	/* If the main allocator is up use that, fallback to bootmem. */
53 	if (slab_is_available()) {
54 		struct page *page;
55 
56 		if (node_state(node, N_HIGH_MEMORY))
57 			page = alloc_pages_node(node,
58 				GFP_KERNEL | __GFP_ZERO, get_order(size));
59 		else
60 			page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
61 				get_order(size));
62 		if (page)
63 			return page_address(page);
64 		return NULL;
65 	} else
66 		return __earlyonly_bootmem_alloc(node, size, size,
67 				__pa(MAX_DMA_ADDRESS));
68 }
69 
70 /* need to make sure size is all the same during early stage */
71 void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node)
72 {
73 	void *ptr;
74 
75 	if (!vmemmap_buf)
76 		return vmemmap_alloc_block(size, node);
77 
78 	/* take the from buf */
79 	ptr = (void *)ALIGN((unsigned long)vmemmap_buf, size);
80 	if (ptr + size > vmemmap_buf_end)
81 		return vmemmap_alloc_block(size, node);
82 
83 	vmemmap_buf = ptr + size;
84 
85 	return ptr;
86 }
87 
88 void __meminit vmemmap_verify(pte_t *pte, int node,
89 				unsigned long start, unsigned long end)
90 {
91 	unsigned long pfn = pte_pfn(*pte);
92 	int actual_node = early_pfn_to_nid(pfn);
93 
94 	if (node_distance(actual_node, node) > LOCAL_DISTANCE)
95 		printk(KERN_WARNING "[%lx-%lx] potential offnode "
96 			"page_structs\n", start, end - 1);
97 }
98 
99 pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node)
100 {
101 	pte_t *pte = pte_offset_kernel(pmd, addr);
102 	if (pte_none(*pte)) {
103 		pte_t entry;
104 		void *p = vmemmap_alloc_block_buf(PAGE_SIZE, node);
105 		if (!p)
106 			return NULL;
107 		entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
108 		set_pte_at(&init_mm, addr, pte, entry);
109 	}
110 	return pte;
111 }
112 
113 pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
114 {
115 	pmd_t *pmd = pmd_offset(pud, addr);
116 	if (pmd_none(*pmd)) {
117 		void *p = vmemmap_alloc_block(PAGE_SIZE, node);
118 		if (!p)
119 			return NULL;
120 		pmd_populate_kernel(&init_mm, pmd, p);
121 	}
122 	return pmd;
123 }
124 
125 pud_t * __meminit vmemmap_pud_populate(pgd_t *pgd, unsigned long addr, int node)
126 {
127 	pud_t *pud = pud_offset(pgd, addr);
128 	if (pud_none(*pud)) {
129 		void *p = vmemmap_alloc_block(PAGE_SIZE, node);
130 		if (!p)
131 			return NULL;
132 		pud_populate(&init_mm, pud, p);
133 	}
134 	return pud;
135 }
136 
137 pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
138 {
139 	pgd_t *pgd = pgd_offset_k(addr);
140 	if (pgd_none(*pgd)) {
141 		void *p = vmemmap_alloc_block(PAGE_SIZE, node);
142 		if (!p)
143 			return NULL;
144 		pgd_populate(&init_mm, pgd, p);
145 	}
146 	return pgd;
147 }
148 
149 int __meminit vmemmap_populate_basepages(struct page *start_page,
150 						unsigned long size, int node)
151 {
152 	unsigned long addr = (unsigned long)start_page;
153 	unsigned long end = (unsigned long)(start_page + size);
154 	pgd_t *pgd;
155 	pud_t *pud;
156 	pmd_t *pmd;
157 	pte_t *pte;
158 
159 	for (; addr < end; addr += PAGE_SIZE) {
160 		pgd = vmemmap_pgd_populate(addr, node);
161 		if (!pgd)
162 			return -ENOMEM;
163 		pud = vmemmap_pud_populate(pgd, addr, node);
164 		if (!pud)
165 			return -ENOMEM;
166 		pmd = vmemmap_pmd_populate(pud, addr, node);
167 		if (!pmd)
168 			return -ENOMEM;
169 		pte = vmemmap_pte_populate(pmd, addr, node);
170 		if (!pte)
171 			return -ENOMEM;
172 		vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
173 	}
174 
175 	return 0;
176 }
177 
178 struct page * __meminit sparse_mem_map_populate(unsigned long pnum, int nid)
179 {
180 	struct page *map = pfn_to_page(pnum * PAGES_PER_SECTION);
181 	int error = vmemmap_populate(map, PAGES_PER_SECTION, nid);
182 	if (error)
183 		return NULL;
184 
185 	return map;
186 }
187 
188 void __init sparse_mem_maps_populate_node(struct page **map_map,
189 					  unsigned long pnum_begin,
190 					  unsigned long pnum_end,
191 					  unsigned long map_count, int nodeid)
192 {
193 	unsigned long pnum;
194 	unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
195 	void *vmemmap_buf_start;
196 
197 	size = ALIGN(size, PMD_SIZE);
198 	vmemmap_buf_start = __earlyonly_bootmem_alloc(nodeid, size * map_count,
199 			 PMD_SIZE, __pa(MAX_DMA_ADDRESS));
200 
201 	if (vmemmap_buf_start) {
202 		vmemmap_buf = vmemmap_buf_start;
203 		vmemmap_buf_end = vmemmap_buf_start + size * map_count;
204 	}
205 
206 	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
207 		struct mem_section *ms;
208 
209 		if (!present_section_nr(pnum))
210 			continue;
211 
212 		map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
213 		if (map_map[pnum])
214 			continue;
215 		ms = __nr_to_section(pnum);
216 		printk(KERN_ERR "%s: sparsemem memory map backing failed "
217 			"some memory will not be available.\n", __func__);
218 		ms->section_mem_map = 0;
219 	}
220 
221 	if (vmemmap_buf_start) {
222 		/* need to free left buf */
223 		free_bootmem(__pa(vmemmap_buf), vmemmap_buf_end - vmemmap_buf);
224 		vmemmap_buf = NULL;
225 		vmemmap_buf_end = NULL;
226 	}
227 }
228