xref: /openbmc/linux/arch/x86/mm/init.c (revision d3ea9fa0)
1 #include <linux/initrd.h>
2 #include <linux/ioport.h>
3 #include <linux/swap.h>
4 
5 #include <asm/cacheflush.h>
6 #include <asm/e820.h>
7 #include <asm/init.h>
8 #include <asm/page.h>
9 #include <asm/page_types.h>
10 #include <asm/sections.h>
11 #include <asm/setup.h>
12 #include <asm/system.h>
13 #include <asm/tlbflush.h>
14 #include <asm/tlb.h>
15 #include <asm/proto.h>
16 
17 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
18 
19 unsigned long __initdata e820_table_start;
20 unsigned long __meminitdata e820_table_end;
21 unsigned long __meminitdata e820_table_top;
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;
74 #else
75 	start = 0x8000;
76 #endif
77 	e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
78 					tables, PAGE_SIZE);
79 	if (e820_table_start == -1UL)
80 		panic("Cannot find space for the kernel page tables");
81 
82 	e820_table_start >>= PAGE_SHIFT;
83 	e820_table_end = e820_table_start;
84 	e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
85 
86 	printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
87 		end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
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 
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)
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 
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;
128 	unsigned long ret = 0;
129 	unsigned long pos;
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 
137 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
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 
149 	set_nx();
150 	if (nx_enabled)
151 		printk(KERN_INFO "NX (Execute Disable) protection: active\n");
152 
153 	/* Enable PSE if available */
154 	if (cpu_has_pse)
155 		set_in_cr4(X86_CR4_PSE);
156 
157 	/* Enable PGE if available */
158 	if (cpu_has_pge) {
159 		set_in_cr4(X86_CR4_PGE);
160 		__supported_pte_mask |= _PAGE_GLOBAL;
161 	}
162 
163 	if (use_gbpages)
164 		page_size_mask |= 1 << PG_LEVEL_1G;
165 	if (use_pse)
166 		page_size_mask |= 1 << PG_LEVEL_2M;
167 
168 	memset(mr, 0, sizeof(mr));
169 	nr_range = 0;
170 
171 	/* head if not big page alignment ? */
172 	start_pfn = start >> PAGE_SHIFT;
173 	pos = start_pfn << PAGE_SHIFT;
174 #ifdef CONFIG_X86_32
175 	/*
176 	 * Don't use a large page for the first 2/4MB of memory
177 	 * because there are often fixed size MTRRs in there
178 	 * and overlapping MTRRs into large pages can cause
179 	 * slowdowns.
180 	 */
181 	if (pos == 0)
182 		end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
183 	else
184 		end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
185 				 << (PMD_SHIFT - PAGE_SHIFT);
186 #else /* CONFIG_X86_64 */
187 	end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
188 			<< (PMD_SHIFT - PAGE_SHIFT);
189 #endif
190 	if (end_pfn > (end >> PAGE_SHIFT))
191 		end_pfn = end >> PAGE_SHIFT;
192 	if (start_pfn < end_pfn) {
193 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
194 		pos = end_pfn << PAGE_SHIFT;
195 	}
196 
197 	/* big page (2M) range */
198 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
199 			 << (PMD_SHIFT - PAGE_SHIFT);
200 #ifdef CONFIG_X86_32
201 	end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
202 #else /* CONFIG_X86_64 */
203 	end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
204 			 << (PUD_SHIFT - PAGE_SHIFT);
205 	if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
206 		end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
207 #endif
208 
209 	if (start_pfn < end_pfn) {
210 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
211 				page_size_mask & (1<<PG_LEVEL_2M));
212 		pos = end_pfn << PAGE_SHIFT;
213 	}
214 
215 #ifdef CONFIG_X86_64
216 	/* big page (1G) range */
217 	start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
218 			 << (PUD_SHIFT - PAGE_SHIFT);
219 	end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
220 	if (start_pfn < end_pfn) {
221 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
222 				page_size_mask &
223 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
224 		pos = end_pfn << PAGE_SHIFT;
225 	}
226 
227 	/* tail is not big page (1G) alignment */
228 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
229 			 << (PMD_SHIFT - PAGE_SHIFT);
230 	end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
231 	if (start_pfn < end_pfn) {
232 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
233 				page_size_mask & (1<<PG_LEVEL_2M));
234 		pos = end_pfn << PAGE_SHIFT;
235 	}
236 #endif
237 
238 	/* tail is not big page (2M) alignment */
239 	start_pfn = pos>>PAGE_SHIFT;
240 	end_pfn = end>>PAGE_SHIFT;
241 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
242 
243 	/* try to merge same page size and continuous */
244 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
245 		unsigned long old_start;
246 		if (mr[i].end != mr[i+1].start ||
247 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
248 			continue;
249 		/* move it */
250 		old_start = mr[i].start;
251 		memmove(&mr[i], &mr[i+1],
252 			(nr_range - 1 - i) * sizeof(struct map_range));
253 		mr[i--].start = old_start;
254 		nr_range--;
255 	}
256 
257 	for (i = 0; i < nr_range; i++)
258 		printk(KERN_DEBUG " %010lx - %010lx page %s\n",
259 				mr[i].start, mr[i].end,
260 			(mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
261 			 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
262 
263 	/*
264 	 * Find space for the kernel direct mapping tables.
265 	 *
266 	 * Later we should allocate these tables in the local node of the
267 	 * memory mapped. Unfortunately this is done currently before the
268 	 * nodes are discovered.
269 	 */
270 	if (!after_bootmem)
271 		find_early_table_space(end, use_pse, use_gbpages);
272 
273 #ifdef CONFIG_X86_32
274 	for (i = 0; i < nr_range; i++)
275 		kernel_physical_mapping_init(mr[i].start, mr[i].end,
276 					     mr[i].page_size_mask);
277 	ret = end;
278 #else /* CONFIG_X86_64 */
279 	for (i = 0; i < nr_range; i++)
280 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
281 						   mr[i].page_size_mask);
282 #endif
283 
284 #ifdef CONFIG_X86_32
285 	early_ioremap_page_table_range_init();
286 
287 	load_cr3(swapper_pg_dir);
288 #endif
289 
290 #ifdef CONFIG_X86_64
291 	if (!after_bootmem && !start) {
292 		pud_t *pud;
293 		pmd_t *pmd;
294 
295 		mmu_cr4_features = read_cr4();
296 
297 		/*
298 		 * _brk_end cannot change anymore, but it and _end may be
299 		 * located on different 2M pages. cleanup_highmap(), however,
300 		 * can only consider _end when it runs, so destroy any
301 		 * mappings beyond _brk_end here.
302 		 */
303 		pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
304 		pmd = pmd_offset(pud, _brk_end - 1);
305 		while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
306 			pmd_clear(pmd);
307 	}
308 #endif
309 	__flush_tlb_all();
310 
311 	if (!after_bootmem && e820_table_end > e820_table_start)
312 		reserve_early(e820_table_start << PAGE_SHIFT,
313 				 e820_table_end << PAGE_SHIFT, "PGTABLE");
314 
315 	if (!after_bootmem)
316 		early_memtest(start, end);
317 
318 	return ret >> PAGE_SHIFT;
319 }
320 
321 
322 /*
323  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
324  * is valid. The argument is a physical page number.
325  *
326  *
327  * On x86, access has to be given to the first megabyte of ram because that area
328  * contains bios code and data regions used by X and dosemu and similar apps.
329  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
330  * mmio resources as well as potential bios/acpi data regions.
331  */
332 int devmem_is_allowed(unsigned long pagenr)
333 {
334 	if (pagenr <= 256)
335 		return 1;
336 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
337 		return 0;
338 	if (!page_is_ram(pagenr))
339 		return 1;
340 	return 0;
341 }
342 
343 void free_init_pages(char *what, unsigned long begin, unsigned long end)
344 {
345 	unsigned long addr = begin;
346 
347 	if (addr >= end)
348 		return;
349 
350 	/*
351 	 * If debugging page accesses then do not free this memory but
352 	 * mark them not present - any buggy init-section access will
353 	 * create a kernel page fault:
354 	 */
355 #ifdef CONFIG_DEBUG_PAGEALLOC
356 	printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
357 		begin, PAGE_ALIGN(end));
358 	set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
359 #else
360 	/*
361 	 * We just marked the kernel text read only above, now that
362 	 * we are going to free part of that, we need to make that
363 	 * writeable first.
364 	 */
365 	set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
366 
367 	printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
368 
369 	for (; addr < end; addr += PAGE_SIZE) {
370 		ClearPageReserved(virt_to_page(addr));
371 		init_page_count(virt_to_page(addr));
372 		memset((void *)(addr & ~(PAGE_SIZE-1)),
373 			POISON_FREE_INITMEM, PAGE_SIZE);
374 		free_page(addr);
375 		totalram_pages++;
376 	}
377 #endif
378 }
379 
380 void free_initmem(void)
381 {
382 	free_init_pages("unused kernel memory",
383 			(unsigned long)(&__init_begin),
384 			(unsigned long)(&__init_end));
385 }
386 
387 #ifdef CONFIG_BLK_DEV_INITRD
388 void free_initrd_mem(unsigned long start, unsigned long end)
389 {
390 	free_init_pages("initrd memory", start, end);
391 }
392 #endif
393