]>
Commit | Line | Data |
---|---|---|
2c1b284e | 1 | #include <linux/initrd.h> |
540aca06 | 2 | #include <linux/ioport.h> |
e5b2bb55 | 3 | #include <linux/swap.h> |
540aca06 | 4 | |
e5b2bb55 | 5 | #include <asm/cacheflush.h> |
f765090a | 6 | #include <asm/e820.h> |
4fcb2083 | 7 | #include <asm/init.h> |
e5b2bb55 | 8 | #include <asm/page.h> |
540aca06 | 9 | #include <asm/page_types.h> |
e5b2bb55 | 10 | #include <asm/sections.h> |
49834396 | 11 | #include <asm/setup.h> |
e5b2bb55 | 12 | #include <asm/system.h> |
f765090a | 13 | #include <asm/tlbflush.h> |
9518e0e4 | 14 | #include <asm/tlb.h> |
76c06927 | 15 | #include <asm/proto.h> |
9518e0e4 PE |
16 | |
17 | DEFINE_PER_CPU(struct mmu_gather, mmu_gathers); | |
f765090a | 18 | |
298af9d8 PE |
19 | unsigned long __initdata e820_table_start; |
20 | unsigned long __meminitdata e820_table_end; | |
21 | unsigned long __meminitdata e820_table_top; | |
f765090a PE |
22 | |
23 | int after_bootmem; | |
24 | ||
25 | int direct_gbpages | |
26 | #ifdef CONFIG_DIRECT_GBPAGES | |
27 | = 1 | |
28 | #endif | |
29 | ; | |
30 | ||
31 | static void __init find_early_table_space(unsigned long end, int use_pse, | |
32 | int use_gbpages) | |
33 | { | |
34 | unsigned long puds, pmds, ptes, tables, start; | |
35 | ||
36 | puds = (end + PUD_SIZE - 1) >> PUD_SHIFT; | |
37 | tables = roundup(puds * sizeof(pud_t), PAGE_SIZE); | |
38 | ||
39 | if (use_gbpages) { | |
40 | unsigned long extra; | |
41 | ||
42 | extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT); | |
43 | pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT; | |
44 | } else | |
45 | pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT; | |
46 | ||
47 | tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE); | |
48 | ||
49 | if (use_pse) { | |
50 | unsigned long extra; | |
51 | ||
52 | extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT); | |
53 | #ifdef CONFIG_X86_32 | |
54 | extra += PMD_SIZE; | |
55 | #endif | |
56 | ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
57 | } else | |
58 | ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
59 | ||
60 | tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE); | |
61 | ||
62 | #ifdef CONFIG_X86_32 | |
63 | /* for fixmap */ | |
64 | tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE); | |
65 | #endif | |
66 | ||
67 | /* | |
68 | * RED-PEN putting page tables only on node 0 could | |
69 | * cause a hotspot and fill up ZONE_DMA. The page tables | |
70 | * need roughly 0.5KB per GB. | |
71 | */ | |
72 | #ifdef CONFIG_X86_32 | |
73 | start = 0x7000; | |
80989ce0 | 74 | #else |
f765090a | 75 | start = 0x8000; |
f765090a | 76 | #endif |
80989ce0 YL |
77 | e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT, |
78 | tables, PAGE_SIZE); | |
298af9d8 | 79 | if (e820_table_start == -1UL) |
f765090a PE |
80 | panic("Cannot find space for the kernel page tables"); |
81 | ||
298af9d8 PE |
82 | e820_table_start >>= PAGE_SHIFT; |
83 | e820_table_end = e820_table_start; | |
84 | e820_table_top = e820_table_start + (tables >> PAGE_SHIFT); | |
f765090a PE |
85 | |
86 | printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n", | |
298af9d8 | 87 | end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT); |
f765090a PE |
88 | } |
89 | ||
90 | struct map_range { | |
91 | unsigned long start; | |
92 | unsigned long end; | |
93 | unsigned page_size_mask; | |
94 | }; | |
95 | ||
96 | #ifdef CONFIG_X86_32 | |
97 | #define NR_RANGE_MR 3 | |
98 | #else /* CONFIG_X86_64 */ | |
99 | #define NR_RANGE_MR 5 | |
100 | #endif | |
101 | ||
dc9dd5cc JB |
102 | static int __meminit save_mr(struct map_range *mr, int nr_range, |
103 | unsigned long start_pfn, unsigned long end_pfn, | |
104 | unsigned long page_size_mask) | |
f765090a PE |
105 | { |
106 | if (start_pfn < end_pfn) { | |
107 | if (nr_range >= NR_RANGE_MR) | |
108 | panic("run out of range for init_memory_mapping\n"); | |
109 | mr[nr_range].start = start_pfn<<PAGE_SHIFT; | |
110 | mr[nr_range].end = end_pfn<<PAGE_SHIFT; | |
111 | mr[nr_range].page_size_mask = page_size_mask; | |
112 | nr_range++; | |
113 | } | |
114 | ||
115 | return nr_range; | |
116 | } | |
117 | ||
f765090a PE |
118 | /* |
119 | * Setup the direct mapping of the physical memory at PAGE_OFFSET. | |
120 | * This runs before bootmem is initialized and gets pages directly from | |
121 | * the physical memory. To access them they are temporarily mapped. | |
122 | */ | |
123 | unsigned long __init_refok init_memory_mapping(unsigned long start, | |
124 | unsigned long end) | |
125 | { | |
126 | unsigned long page_size_mask = 0; | |
127 | unsigned long start_pfn, end_pfn; | |
c77a3b59 | 128 | unsigned long ret = 0; |
f765090a | 129 | unsigned long pos; |
f765090a PE |
130 | |
131 | struct map_range mr[NR_RANGE_MR]; | |
132 | int nr_range, i; | |
133 | int use_pse, use_gbpages; | |
134 | ||
135 | printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end); | |
136 | ||
f8561296 | 137 | #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK) |
f765090a PE |
138 | /* |
139 | * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages. | |
140 | * This will simplify cpa(), which otherwise needs to support splitting | |
141 | * large pages into small in interrupt context, etc. | |
142 | */ | |
143 | use_pse = use_gbpages = 0; | |
144 | #else | |
145 | use_pse = cpu_has_pse; | |
146 | use_gbpages = direct_gbpages; | |
147 | #endif | |
148 | ||
f765090a PE |
149 | /* Enable PSE if available */ |
150 | if (cpu_has_pse) | |
151 | set_in_cr4(X86_CR4_PSE); | |
152 | ||
153 | /* Enable PGE if available */ | |
154 | if (cpu_has_pge) { | |
155 | set_in_cr4(X86_CR4_PGE); | |
156 | __supported_pte_mask |= _PAGE_GLOBAL; | |
157 | } | |
f765090a PE |
158 | |
159 | if (use_gbpages) | |
160 | page_size_mask |= 1 << PG_LEVEL_1G; | |
161 | if (use_pse) | |
162 | page_size_mask |= 1 << PG_LEVEL_2M; | |
163 | ||
164 | memset(mr, 0, sizeof(mr)); | |
165 | nr_range = 0; | |
166 | ||
167 | /* head if not big page alignment ? */ | |
168 | start_pfn = start >> PAGE_SHIFT; | |
169 | pos = start_pfn << PAGE_SHIFT; | |
170 | #ifdef CONFIG_X86_32 | |
171 | /* | |
172 | * Don't use a large page for the first 2/4MB of memory | |
173 | * because there are often fixed size MTRRs in there | |
174 | * and overlapping MTRRs into large pages can cause | |
175 | * slowdowns. | |
176 | */ | |
177 | if (pos == 0) | |
178 | end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT); | |
179 | else | |
180 | end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT) | |
181 | << (PMD_SHIFT - PAGE_SHIFT); | |
182 | #else /* CONFIG_X86_64 */ | |
183 | end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT) | |
184 | << (PMD_SHIFT - PAGE_SHIFT); | |
185 | #endif | |
186 | if (end_pfn > (end >> PAGE_SHIFT)) | |
187 | end_pfn = end >> PAGE_SHIFT; | |
188 | if (start_pfn < end_pfn) { | |
189 | nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); | |
190 | pos = end_pfn << PAGE_SHIFT; | |
191 | } | |
192 | ||
193 | /* big page (2M) range */ | |
194 | start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT) | |
195 | << (PMD_SHIFT - PAGE_SHIFT); | |
196 | #ifdef CONFIG_X86_32 | |
197 | end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT); | |
198 | #else /* CONFIG_X86_64 */ | |
199 | end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT) | |
200 | << (PUD_SHIFT - PAGE_SHIFT); | |
201 | if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT))) | |
202 | end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)); | |
203 | #endif | |
204 | ||
205 | if (start_pfn < end_pfn) { | |
206 | nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, | |
207 | page_size_mask & (1<<PG_LEVEL_2M)); | |
208 | pos = end_pfn << PAGE_SHIFT; | |
209 | } | |
210 | ||
211 | #ifdef CONFIG_X86_64 | |
212 | /* big page (1G) range */ | |
213 | start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT) | |
214 | << (PUD_SHIFT - PAGE_SHIFT); | |
215 | end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT); | |
216 | if (start_pfn < end_pfn) { | |
217 | nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, | |
218 | page_size_mask & | |
219 | ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); | |
220 | pos = end_pfn << PAGE_SHIFT; | |
221 | } | |
222 | ||
223 | /* tail is not big page (1G) alignment */ | |
224 | start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT) | |
225 | << (PMD_SHIFT - PAGE_SHIFT); | |
226 | end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT); | |
227 | if (start_pfn < end_pfn) { | |
228 | nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, | |
229 | page_size_mask & (1<<PG_LEVEL_2M)); | |
230 | pos = end_pfn << PAGE_SHIFT; | |
231 | } | |
232 | #endif | |
233 | ||
234 | /* tail is not big page (2M) alignment */ | |
235 | start_pfn = pos>>PAGE_SHIFT; | |
236 | end_pfn = end>>PAGE_SHIFT; | |
237 | nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); | |
238 | ||
239 | /* try to merge same page size and continuous */ | |
240 | for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { | |
241 | unsigned long old_start; | |
242 | if (mr[i].end != mr[i+1].start || | |
243 | mr[i].page_size_mask != mr[i+1].page_size_mask) | |
244 | continue; | |
245 | /* move it */ | |
246 | old_start = mr[i].start; | |
247 | memmove(&mr[i], &mr[i+1], | |
248 | (nr_range - 1 - i) * sizeof(struct map_range)); | |
249 | mr[i--].start = old_start; | |
250 | nr_range--; | |
251 | } | |
252 | ||
253 | for (i = 0; i < nr_range; i++) | |
254 | printk(KERN_DEBUG " %010lx - %010lx page %s\n", | |
255 | mr[i].start, mr[i].end, | |
256 | (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":( | |
257 | (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k")); | |
258 | ||
259 | /* | |
260 | * Find space for the kernel direct mapping tables. | |
261 | * | |
262 | * Later we should allocate these tables in the local node of the | |
263 | * memory mapped. Unfortunately this is done currently before the | |
264 | * nodes are discovered. | |
265 | */ | |
266 | if (!after_bootmem) | |
267 | find_early_table_space(end, use_pse, use_gbpages); | |
268 | ||
f765090a PE |
269 | for (i = 0; i < nr_range; i++) |
270 | ret = kernel_physical_mapping_init(mr[i].start, mr[i].end, | |
271 | mr[i].page_size_mask); | |
f765090a PE |
272 | |
273 | #ifdef CONFIG_X86_32 | |
274 | early_ioremap_page_table_range_init(); | |
275 | ||
276 | load_cr3(swapper_pg_dir); | |
277 | #endif | |
278 | ||
279 | #ifdef CONFIG_X86_64 | |
49834396 JB |
280 | if (!after_bootmem && !start) { |
281 | pud_t *pud; | |
282 | pmd_t *pmd; | |
283 | ||
f765090a | 284 | mmu_cr4_features = read_cr4(); |
49834396 JB |
285 | |
286 | /* | |
287 | * _brk_end cannot change anymore, but it and _end may be | |
288 | * located on different 2M pages. cleanup_highmap(), however, | |
289 | * can only consider _end when it runs, so destroy any | |
290 | * mappings beyond _brk_end here. | |
291 | */ | |
292 | pud = pud_offset(pgd_offset_k(_brk_end), _brk_end); | |
293 | pmd = pmd_offset(pud, _brk_end - 1); | |
294 | while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1)) | |
295 | pmd_clear(pmd); | |
296 | } | |
f765090a PE |
297 | #endif |
298 | __flush_tlb_all(); | |
299 | ||
298af9d8 PE |
300 | if (!after_bootmem && e820_table_end > e820_table_start) |
301 | reserve_early(e820_table_start << PAGE_SHIFT, | |
302 | e820_table_end << PAGE_SHIFT, "PGTABLE"); | |
f765090a PE |
303 | |
304 | if (!after_bootmem) | |
305 | early_memtest(start, end); | |
306 | ||
307 | return ret >> PAGE_SHIFT; | |
308 | } | |
309 | ||
e5b2bb55 | 310 | |
540aca06 PE |
311 | /* |
312 | * devmem_is_allowed() checks to see if /dev/mem access to a certain address | |
313 | * is valid. The argument is a physical page number. | |
314 | * | |
315 | * | |
316 | * On x86, access has to be given to the first megabyte of ram because that area | |
317 | * contains bios code and data regions used by X and dosemu and similar apps. | |
318 | * Access has to be given to non-kernel-ram areas as well, these contain the PCI | |
319 | * mmio resources as well as potential bios/acpi data regions. | |
320 | */ | |
321 | int devmem_is_allowed(unsigned long pagenr) | |
322 | { | |
323 | if (pagenr <= 256) | |
324 | return 1; | |
325 | if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) | |
326 | return 0; | |
327 | if (!page_is_ram(pagenr)) | |
328 | return 1; | |
329 | return 0; | |
330 | } | |
331 | ||
e5b2bb55 PE |
332 | void free_init_pages(char *what, unsigned long begin, unsigned long end) |
333 | { | |
334 | unsigned long addr = begin; | |
335 | ||
336 | if (addr >= end) | |
337 | return; | |
338 | ||
339 | /* | |
340 | * If debugging page accesses then do not free this memory but | |
341 | * mark them not present - any buggy init-section access will | |
342 | * create a kernel page fault: | |
343 | */ | |
344 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
345 | printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n", | |
346 | begin, PAGE_ALIGN(end)); | |
347 | set_memory_np(begin, (end - begin) >> PAGE_SHIFT); | |
348 | #else | |
349 | /* | |
350 | * We just marked the kernel text read only above, now that | |
351 | * we are going to free part of that, we need to make that | |
352 | * writeable first. | |
353 | */ | |
354 | set_memory_rw(begin, (end - begin) >> PAGE_SHIFT); | |
355 | ||
356 | printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10); | |
357 | ||
358 | for (; addr < end; addr += PAGE_SIZE) { | |
359 | ClearPageReserved(virt_to_page(addr)); | |
360 | init_page_count(virt_to_page(addr)); | |
361 | memset((void *)(addr & ~(PAGE_SIZE-1)), | |
362 | POISON_FREE_INITMEM, PAGE_SIZE); | |
363 | free_page(addr); | |
364 | totalram_pages++; | |
365 | } | |
366 | #endif | |
367 | } | |
368 | ||
369 | void free_initmem(void) | |
370 | { | |
371 | free_init_pages("unused kernel memory", | |
372 | (unsigned long)(&__init_begin), | |
373 | (unsigned long)(&__init_end)); | |
374 | } | |
731ddea6 PE |
375 | |
376 | #ifdef CONFIG_BLK_DEV_INITRD | |
377 | void free_initrd_mem(unsigned long start, unsigned long end) | |
378 | { | |
379 | free_init_pages("initrd memory", start, end); | |
380 | } | |
381 | #endif |