xref: /linux/arch/x86/mm/init.c (revision 0dd9ac63ce26ec87b080ca9c3e6efed33c23ace6)
1 #include <linux/gfp.h>
2 #include <linux/initrd.h>
3 #include <linux/ioport.h>
4 #include <linux/swap.h>
5 
6 #include <asm/cacheflush.h>
7 #include <asm/e820.h>
8 #include <asm/init.h>
9 #include <asm/page.h>
10 #include <asm/page_types.h>
11 #include <asm/sections.h>
12 #include <asm/setup.h>
13 #include <asm/system.h>
14 #include <asm/tlbflush.h>
15 #include <asm/tlb.h>
16 #include <asm/proto.h>
17 
18 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
19 
20 unsigned long __initdata e820_table_start;
21 unsigned long __meminitdata e820_table_end;
22 unsigned long __meminitdata e820_table_top;
23 
24 int after_bootmem;
25 
26 int direct_gbpages
27 #ifdef CONFIG_DIRECT_GBPAGES
28 				= 1
29 #endif
30 ;
31 
32 static void __init find_early_table_space(unsigned long end, int use_pse,
33 					  int use_gbpages)
34 {
35 	unsigned long puds, pmds, ptes, tables, start;
36 
37 	puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
38 	tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
39 
40 	if (use_gbpages) {
41 		unsigned long extra;
42 
43 		extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
44 		pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
45 	} else
46 		pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
47 
48 	tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
49 
50 	if (use_pse) {
51 		unsigned long extra;
52 
53 		extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
54 #ifdef CONFIG_X86_32
55 		extra += PMD_SIZE;
56 #endif
57 		ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
58 	} else
59 		ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
60 
61 	tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
62 
63 #ifdef CONFIG_X86_32
64 	/* for fixmap */
65 	tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
66 #endif
67 
68 	/*
69 	 * RED-PEN putting page tables only on node 0 could
70 	 * cause a hotspot and fill up ZONE_DMA. The page tables
71 	 * need roughly 0.5KB per GB.
72 	 */
73 #ifdef CONFIG_X86_32
74 	start = 0x7000;
75 #else
76 	start = 0x8000;
77 #endif
78 	e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
79 					tables, PAGE_SIZE);
80 	if (e820_table_start == -1UL)
81 		panic("Cannot find space for the kernel page tables");
82 
83 	e820_table_start >>= PAGE_SHIFT;
84 	e820_table_end = e820_table_start;
85 	e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
86 
87 	printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
88 		end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
89 }
90 
91 struct map_range {
92 	unsigned long start;
93 	unsigned long end;
94 	unsigned page_size_mask;
95 };
96 
97 #ifdef CONFIG_X86_32
98 #define NR_RANGE_MR 3
99 #else /* CONFIG_X86_64 */
100 #define NR_RANGE_MR 5
101 #endif
102 
103 static int __meminit save_mr(struct map_range *mr, int nr_range,
104 			     unsigned long start_pfn, unsigned long end_pfn,
105 			     unsigned long page_size_mask)
106 {
107 	if (start_pfn < end_pfn) {
108 		if (nr_range >= NR_RANGE_MR)
109 			panic("run out of range for init_memory_mapping\n");
110 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
111 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
112 		mr[nr_range].page_size_mask = page_size_mask;
113 		nr_range++;
114 	}
115 
116 	return nr_range;
117 }
118 
119 /*
120  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
121  * This runs before bootmem is initialized and gets pages directly from
122  * the physical memory. To access them they are temporarily mapped.
123  */
124 unsigned long __init_refok init_memory_mapping(unsigned long start,
125 					       unsigned long end)
126 {
127 	unsigned long page_size_mask = 0;
128 	unsigned long start_pfn, end_pfn;
129 	unsigned long ret = 0;
130 	unsigned long pos;
131 
132 	struct map_range mr[NR_RANGE_MR];
133 	int nr_range, i;
134 	int use_pse, use_gbpages;
135 
136 	printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
137 
138 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
139 	/*
140 	 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
141 	 * This will simplify cpa(), which otherwise needs to support splitting
142 	 * large pages into small in interrupt context, etc.
143 	 */
144 	use_pse = use_gbpages = 0;
145 #else
146 	use_pse = cpu_has_pse;
147 	use_gbpages = direct_gbpages;
148 #endif
149 
150 	/* Enable PSE if available */
151 	if (cpu_has_pse)
152 		set_in_cr4(X86_CR4_PSE);
153 
154 	/* Enable PGE if available */
155 	if (cpu_has_pge) {
156 		set_in_cr4(X86_CR4_PGE);
157 		__supported_pte_mask |= _PAGE_GLOBAL;
158 	}
159 
160 	if (use_gbpages)
161 		page_size_mask |= 1 << PG_LEVEL_1G;
162 	if (use_pse)
163 		page_size_mask |= 1 << PG_LEVEL_2M;
164 
165 	memset(mr, 0, sizeof(mr));
166 	nr_range = 0;
167 
168 	/* head if not big page alignment ? */
169 	start_pfn = start >> PAGE_SHIFT;
170 	pos = start_pfn << PAGE_SHIFT;
171 #ifdef CONFIG_X86_32
172 	/*
173 	 * Don't use a large page for the first 2/4MB of memory
174 	 * because there are often fixed size MTRRs in there
175 	 * and overlapping MTRRs into large pages can cause
176 	 * slowdowns.
177 	 */
178 	if (pos == 0)
179 		end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
180 	else
181 		end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
182 				 << (PMD_SHIFT - PAGE_SHIFT);
183 #else /* CONFIG_X86_64 */
184 	end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
185 			<< (PMD_SHIFT - PAGE_SHIFT);
186 #endif
187 	if (end_pfn > (end >> PAGE_SHIFT))
188 		end_pfn = end >> PAGE_SHIFT;
189 	if (start_pfn < end_pfn) {
190 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
191 		pos = end_pfn << PAGE_SHIFT;
192 	}
193 
194 	/* big page (2M) range */
195 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
196 			 << (PMD_SHIFT - PAGE_SHIFT);
197 #ifdef CONFIG_X86_32
198 	end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
199 #else /* CONFIG_X86_64 */
200 	end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
201 			 << (PUD_SHIFT - PAGE_SHIFT);
202 	if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
203 		end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
204 #endif
205 
206 	if (start_pfn < end_pfn) {
207 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
208 				page_size_mask & (1<<PG_LEVEL_2M));
209 		pos = end_pfn << PAGE_SHIFT;
210 	}
211 
212 #ifdef CONFIG_X86_64
213 	/* big page (1G) range */
214 	start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
215 			 << (PUD_SHIFT - PAGE_SHIFT);
216 	end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
217 	if (start_pfn < end_pfn) {
218 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
219 				page_size_mask &
220 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
221 		pos = end_pfn << PAGE_SHIFT;
222 	}
223 
224 	/* tail is not big page (1G) alignment */
225 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
226 			 << (PMD_SHIFT - PAGE_SHIFT);
227 	end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
228 	if (start_pfn < end_pfn) {
229 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
230 				page_size_mask & (1<<PG_LEVEL_2M));
231 		pos = end_pfn << PAGE_SHIFT;
232 	}
233 #endif
234 
235 	/* tail is not big page (2M) alignment */
236 	start_pfn = pos>>PAGE_SHIFT;
237 	end_pfn = end>>PAGE_SHIFT;
238 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
239 
240 	/* try to merge same page size and continuous */
241 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
242 		unsigned long old_start;
243 		if (mr[i].end != mr[i+1].start ||
244 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
245 			continue;
246 		/* move it */
247 		old_start = mr[i].start;
248 		memmove(&mr[i], &mr[i+1],
249 			(nr_range - 1 - i) * sizeof(struct map_range));
250 		mr[i--].start = old_start;
251 		nr_range--;
252 	}
253 
254 	for (i = 0; i < nr_range; i++)
255 		printk(KERN_DEBUG " %010lx - %010lx page %s\n",
256 				mr[i].start, mr[i].end,
257 			(mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
258 			 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
259 
260 	/*
261 	 * Find space for the kernel direct mapping tables.
262 	 *
263 	 * Later we should allocate these tables in the local node of the
264 	 * memory mapped. Unfortunately this is done currently before the
265 	 * nodes are discovered.
266 	 */
267 	if (!after_bootmem)
268 		find_early_table_space(end, use_pse, use_gbpages);
269 
270 	for (i = 0; i < nr_range; i++)
271 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
272 						   mr[i].page_size_mask);
273 
274 #ifdef CONFIG_X86_32
275 	early_ioremap_page_table_range_init();
276 
277 	load_cr3(swapper_pg_dir);
278 #endif
279 
280 #ifdef CONFIG_X86_64
281 	if (!after_bootmem && !start) {
282 		pud_t *pud;
283 		pmd_t *pmd;
284 
285 		mmu_cr4_features = read_cr4();
286 
287 		/*
288 		 * _brk_end cannot change anymore, but it and _end may be
289 		 * located on different 2M pages. cleanup_highmap(), however,
290 		 * can only consider _end when it runs, so destroy any
291 		 * mappings beyond _brk_end here.
292 		 */
293 		pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
294 		pmd = pmd_offset(pud, _brk_end - 1);
295 		while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
296 			pmd_clear(pmd);
297 	}
298 #endif
299 	__flush_tlb_all();
300 
301 	if (!after_bootmem && e820_table_end > e820_table_start)
302 		reserve_early(e820_table_start << PAGE_SHIFT,
303 				 e820_table_end << PAGE_SHIFT, "PGTABLE");
304 
305 	if (!after_bootmem)
306 		early_memtest(start, end);
307 
308 	return ret >> PAGE_SHIFT;
309 }
310 
311 
312 /*
313  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
314  * is valid. The argument is a physical page number.
315  *
316  *
317  * On x86, access has to be given to the first megabyte of ram because that area
318  * contains bios code and data regions used by X and dosemu and similar apps.
319  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
320  * mmio resources as well as potential bios/acpi data regions.
321  */
322 int devmem_is_allowed(unsigned long pagenr)
323 {
324 	if (pagenr <= 256)
325 		return 1;
326 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
327 		return 0;
328 	if (!page_is_ram(pagenr))
329 		return 1;
330 	return 0;
331 }
332 
333 void free_init_pages(char *what, unsigned long begin, unsigned long end)
334 {
335 	unsigned long addr;
336 	unsigned long begin_aligned, end_aligned;
337 
338 	/* Make sure boundaries are page aligned */
339 	begin_aligned = PAGE_ALIGN(begin);
340 	end_aligned   = end & PAGE_MASK;
341 
342 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
343 		begin = begin_aligned;
344 		end   = end_aligned;
345 	}
346 
347 	if (begin >= end)
348 		return;
349 
350 	addr = begin;
351 
352 	/*
353 	 * If debugging page accesses then do not free this memory but
354 	 * mark them not present - any buggy init-section access will
355 	 * create a kernel page fault:
356 	 */
357 #ifdef CONFIG_DEBUG_PAGEALLOC
358 	printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
359 		begin, end);
360 	set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
361 #else
362 	/*
363 	 * We just marked the kernel text read only above, now that
364 	 * we are going to free part of that, we need to make that
365 	 * writeable first.
366 	 */
367 	set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
368 
369 	printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
370 
371 	for (; addr < end; addr += PAGE_SIZE) {
372 		ClearPageReserved(virt_to_page(addr));
373 		init_page_count(virt_to_page(addr));
374 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
375 		free_page(addr);
376 		totalram_pages++;
377 	}
378 #endif
379 }
380 
381 void free_initmem(void)
382 {
383 	free_init_pages("unused kernel memory",
384 			(unsigned long)(&__init_begin),
385 			(unsigned long)(&__init_end));
386 }
387 
388 #ifdef CONFIG_BLK_DEV_INITRD
389 void free_initrd_mem(unsigned long start, unsigned long end)
390 {
391 	/*
392 	 * end could be not aligned, and We can not align that,
393 	 * decompresser could be confused by aligned initrd_end
394 	 * We already reserve the end partial page before in
395 	 *   - i386_start_kernel()
396 	 *   - x86_64_start_kernel()
397 	 *   - relocate_initrd()
398 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
399 	 */
400 	free_init_pages("initrd memory", start, PAGE_ALIGN(end));
401 }
402 #endif
403