xref: /openbmc/linux/arch/x86/mm/init.c (revision 565d76cb)
1 #include <linux/gfp.h>
2 #include <linux/initrd.h>
3 #include <linux/ioport.h>
4 #include <linux/swap.h>
5 #include <linux/memblock.h>
6 
7 #include <asm/cacheflush.h>
8 #include <asm/e820.h>
9 #include <asm/init.h>
10 #include <asm/page.h>
11 #include <asm/page_types.h>
12 #include <asm/sections.h>
13 #include <asm/setup.h>
14 #include <asm/system.h>
15 #include <asm/tlbflush.h>
16 #include <asm/tlb.h>
17 #include <asm/proto.h>
18 
19 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
20 
21 unsigned long __initdata pgt_buf_start;
22 unsigned long __meminitdata pgt_buf_end;
23 unsigned long __meminitdata pgt_buf_top;
24 
25 int after_bootmem;
26 
27 int direct_gbpages
28 #ifdef CONFIG_DIRECT_GBPAGES
29 				= 1
30 #endif
31 ;
32 
33 static void __init find_early_table_space(unsigned long end, int use_pse,
34 					  int use_gbpages)
35 {
36 	unsigned long puds, pmds, ptes, tables, start = 0, good_end = end;
37 	phys_addr_t base;
38 
39 	puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
40 	tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
41 
42 	if (use_gbpages) {
43 		unsigned long extra;
44 
45 		extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
46 		pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
47 	} else
48 		pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
49 
50 	tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
51 
52 	if (use_pse) {
53 		unsigned long extra;
54 
55 		extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
56 #ifdef CONFIG_X86_32
57 		extra += PMD_SIZE;
58 #endif
59 		ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
60 	} else
61 		ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
62 
63 	tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
64 
65 #ifdef CONFIG_X86_32
66 	/* for fixmap */
67 	tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
68 
69 	good_end = max_pfn_mapped << PAGE_SHIFT;
70 #endif
71 
72 	base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
73 	if (base == MEMBLOCK_ERROR)
74 		panic("Cannot find space for the kernel page tables");
75 
76 	pgt_buf_start = base >> PAGE_SHIFT;
77 	pgt_buf_end = pgt_buf_start;
78 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
79 
80 	printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
81 		end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
82 }
83 
84 struct map_range {
85 	unsigned long start;
86 	unsigned long end;
87 	unsigned page_size_mask;
88 };
89 
90 #ifdef CONFIG_X86_32
91 #define NR_RANGE_MR 3
92 #else /* CONFIG_X86_64 */
93 #define NR_RANGE_MR 5
94 #endif
95 
96 static int __meminit save_mr(struct map_range *mr, int nr_range,
97 			     unsigned long start_pfn, unsigned long end_pfn,
98 			     unsigned long page_size_mask)
99 {
100 	if (start_pfn < end_pfn) {
101 		if (nr_range >= NR_RANGE_MR)
102 			panic("run out of range for init_memory_mapping\n");
103 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
104 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
105 		mr[nr_range].page_size_mask = page_size_mask;
106 		nr_range++;
107 	}
108 
109 	return nr_range;
110 }
111 
112 /*
113  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
114  * This runs before bootmem is initialized and gets pages directly from
115  * the physical memory. To access them they are temporarily mapped.
116  */
117 unsigned long __init_refok init_memory_mapping(unsigned long start,
118 					       unsigned long end)
119 {
120 	unsigned long page_size_mask = 0;
121 	unsigned long start_pfn, end_pfn;
122 	unsigned long ret = 0;
123 	unsigned long pos;
124 
125 	struct map_range mr[NR_RANGE_MR];
126 	int nr_range, i;
127 	int use_pse, use_gbpages;
128 
129 	printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
130 
131 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
132 	/*
133 	 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
134 	 * This will simplify cpa(), which otherwise needs to support splitting
135 	 * large pages into small in interrupt context, etc.
136 	 */
137 	use_pse = use_gbpages = 0;
138 #else
139 	use_pse = cpu_has_pse;
140 	use_gbpages = direct_gbpages;
141 #endif
142 
143 	/* Enable PSE if available */
144 	if (cpu_has_pse)
145 		set_in_cr4(X86_CR4_PSE);
146 
147 	/* Enable PGE if available */
148 	if (cpu_has_pge) {
149 		set_in_cr4(X86_CR4_PGE);
150 		__supported_pte_mask |= _PAGE_GLOBAL;
151 	}
152 
153 	if (use_gbpages)
154 		page_size_mask |= 1 << PG_LEVEL_1G;
155 	if (use_pse)
156 		page_size_mask |= 1 << PG_LEVEL_2M;
157 
158 	memset(mr, 0, sizeof(mr));
159 	nr_range = 0;
160 
161 	/* head if not big page alignment ? */
162 	start_pfn = start >> PAGE_SHIFT;
163 	pos = start_pfn << PAGE_SHIFT;
164 #ifdef CONFIG_X86_32
165 	/*
166 	 * Don't use a large page for the first 2/4MB of memory
167 	 * because there are often fixed size MTRRs in there
168 	 * and overlapping MTRRs into large pages can cause
169 	 * slowdowns.
170 	 */
171 	if (pos == 0)
172 		end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
173 	else
174 		end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
175 				 << (PMD_SHIFT - PAGE_SHIFT);
176 #else /* CONFIG_X86_64 */
177 	end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
178 			<< (PMD_SHIFT - PAGE_SHIFT);
179 #endif
180 	if (end_pfn > (end >> PAGE_SHIFT))
181 		end_pfn = end >> PAGE_SHIFT;
182 	if (start_pfn < end_pfn) {
183 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
184 		pos = end_pfn << PAGE_SHIFT;
185 	}
186 
187 	/* big page (2M) range */
188 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
189 			 << (PMD_SHIFT - PAGE_SHIFT);
190 #ifdef CONFIG_X86_32
191 	end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
192 #else /* CONFIG_X86_64 */
193 	end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
194 			 << (PUD_SHIFT - PAGE_SHIFT);
195 	if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
196 		end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
197 #endif
198 
199 	if (start_pfn < end_pfn) {
200 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
201 				page_size_mask & (1<<PG_LEVEL_2M));
202 		pos = end_pfn << PAGE_SHIFT;
203 	}
204 
205 #ifdef CONFIG_X86_64
206 	/* big page (1G) range */
207 	start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
208 			 << (PUD_SHIFT - PAGE_SHIFT);
209 	end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
210 	if (start_pfn < end_pfn) {
211 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
212 				page_size_mask &
213 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
214 		pos = end_pfn << PAGE_SHIFT;
215 	}
216 
217 	/* tail is not big page (1G) alignment */
218 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
219 			 << (PMD_SHIFT - PAGE_SHIFT);
220 	end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
221 	if (start_pfn < end_pfn) {
222 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
223 				page_size_mask & (1<<PG_LEVEL_2M));
224 		pos = end_pfn << PAGE_SHIFT;
225 	}
226 #endif
227 
228 	/* tail is not big page (2M) alignment */
229 	start_pfn = pos>>PAGE_SHIFT;
230 	end_pfn = end>>PAGE_SHIFT;
231 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
232 
233 	/* try to merge same page size and continuous */
234 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
235 		unsigned long old_start;
236 		if (mr[i].end != mr[i+1].start ||
237 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
238 			continue;
239 		/* move it */
240 		old_start = mr[i].start;
241 		memmove(&mr[i], &mr[i+1],
242 			(nr_range - 1 - i) * sizeof(struct map_range));
243 		mr[i--].start = old_start;
244 		nr_range--;
245 	}
246 
247 	for (i = 0; i < nr_range; i++)
248 		printk(KERN_DEBUG " %010lx - %010lx page %s\n",
249 				mr[i].start, mr[i].end,
250 			(mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
251 			 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
252 
253 	/*
254 	 * Find space for the kernel direct mapping tables.
255 	 *
256 	 * Later we should allocate these tables in the local node of the
257 	 * memory mapped. Unfortunately this is done currently before the
258 	 * nodes are discovered.
259 	 */
260 	if (!after_bootmem)
261 		find_early_table_space(end, use_pse, use_gbpages);
262 
263 	for (i = 0; i < nr_range; i++)
264 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
265 						   mr[i].page_size_mask);
266 
267 #ifdef CONFIG_X86_32
268 	early_ioremap_page_table_range_init();
269 
270 	load_cr3(swapper_pg_dir);
271 #endif
272 
273 	__flush_tlb_all();
274 
275 	if (!after_bootmem && pgt_buf_end > pgt_buf_start)
276 		memblock_x86_reserve_range(pgt_buf_start << PAGE_SHIFT,
277 				 pgt_buf_end << PAGE_SHIFT, "PGTABLE");
278 
279 	if (!after_bootmem)
280 		early_memtest(start, end);
281 
282 	return ret >> PAGE_SHIFT;
283 }
284 
285 
286 /*
287  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
288  * is valid. The argument is a physical page number.
289  *
290  *
291  * On x86, access has to be given to the first megabyte of ram because that area
292  * contains bios code and data regions used by X and dosemu and similar apps.
293  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
294  * mmio resources as well as potential bios/acpi data regions.
295  */
296 int devmem_is_allowed(unsigned long pagenr)
297 {
298 	if (pagenr <= 256)
299 		return 1;
300 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
301 		return 0;
302 	if (!page_is_ram(pagenr))
303 		return 1;
304 	return 0;
305 }
306 
307 void free_init_pages(char *what, unsigned long begin, unsigned long end)
308 {
309 	unsigned long addr;
310 	unsigned long begin_aligned, end_aligned;
311 
312 	/* Make sure boundaries are page aligned */
313 	begin_aligned = PAGE_ALIGN(begin);
314 	end_aligned   = end & PAGE_MASK;
315 
316 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
317 		begin = begin_aligned;
318 		end   = end_aligned;
319 	}
320 
321 	if (begin >= end)
322 		return;
323 
324 	addr = begin;
325 
326 	/*
327 	 * If debugging page accesses then do not free this memory but
328 	 * mark them not present - any buggy init-section access will
329 	 * create a kernel page fault:
330 	 */
331 #ifdef CONFIG_DEBUG_PAGEALLOC
332 	printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
333 		begin, end);
334 	set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
335 #else
336 	/*
337 	 * We just marked the kernel text read only above, now that
338 	 * we are going to free part of that, we need to make that
339 	 * writeable and non-executable first.
340 	 */
341 	set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
342 	set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
343 
344 	printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
345 
346 	for (; addr < end; addr += PAGE_SIZE) {
347 		ClearPageReserved(virt_to_page(addr));
348 		init_page_count(virt_to_page(addr));
349 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
350 		free_page(addr);
351 		totalram_pages++;
352 	}
353 #endif
354 }
355 
356 void free_initmem(void)
357 {
358 	free_init_pages("unused kernel memory",
359 			(unsigned long)(&__init_begin),
360 			(unsigned long)(&__init_end));
361 }
362 
363 #ifdef CONFIG_BLK_DEV_INITRD
364 void free_initrd_mem(unsigned long start, unsigned long end)
365 {
366 	/*
367 	 * end could be not aligned, and We can not align that,
368 	 * decompresser could be confused by aligned initrd_end
369 	 * We already reserve the end partial page before in
370 	 *   - i386_start_kernel()
371 	 *   - x86_64_start_kernel()
372 	 *   - relocate_initrd()
373 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
374 	 */
375 	free_init_pages("initrd memory", start, PAGE_ALIGN(end));
376 }
377 #endif
378