xref: /openbmc/linux/arch/x86/mm/init.c (revision 78c99ba1)
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 
16 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
17 
18 unsigned long __initdata e820_table_start;
19 unsigned long __meminitdata e820_table_end;
20 unsigned long __meminitdata e820_table_top;
21 
22 int after_bootmem;
23 
24 int direct_gbpages
25 #ifdef CONFIG_DIRECT_GBPAGES
26 				= 1
27 #endif
28 ;
29 
30 int nx_enabled;
31 
32 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
33 static int disable_nx __cpuinitdata;
34 
35 /*
36  * noexec = on|off
37  *
38  * Control non-executable mappings for processes.
39  *
40  * on      Enable
41  * off     Disable
42  */
43 static int __init noexec_setup(char *str)
44 {
45 	if (!str)
46 		return -EINVAL;
47 	if (!strncmp(str, "on", 2)) {
48 		__supported_pte_mask |= _PAGE_NX;
49 		disable_nx = 0;
50 	} else if (!strncmp(str, "off", 3)) {
51 		disable_nx = 1;
52 		__supported_pte_mask &= ~_PAGE_NX;
53 	}
54 	return 0;
55 }
56 early_param("noexec", noexec_setup);
57 #endif
58 
59 #ifdef CONFIG_X86_PAE
60 static void __init set_nx(void)
61 {
62 	unsigned int v[4], l, h;
63 
64 	if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
65 		cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
66 
67 		if ((v[3] & (1 << 20)) && !disable_nx) {
68 			rdmsr(MSR_EFER, l, h);
69 			l |= EFER_NX;
70 			wrmsr(MSR_EFER, l, h);
71 			nx_enabled = 1;
72 			__supported_pte_mask |= _PAGE_NX;
73 		}
74 	}
75 }
76 #else
77 static inline void set_nx(void)
78 {
79 }
80 #endif
81 
82 #ifdef CONFIG_X86_64
83 void __cpuinit check_efer(void)
84 {
85 	unsigned long efer;
86 
87 	rdmsrl(MSR_EFER, efer);
88 	if (!(efer & EFER_NX) || disable_nx)
89 		__supported_pte_mask &= ~_PAGE_NX;
90 }
91 #endif
92 
93 static void __init find_early_table_space(unsigned long end, int use_pse,
94 					  int use_gbpages)
95 {
96 	unsigned long puds, pmds, ptes, tables, start;
97 
98 	puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
99 	tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
100 
101 	if (use_gbpages) {
102 		unsigned long extra;
103 
104 		extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
105 		pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
106 	} else
107 		pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
108 
109 	tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
110 
111 	if (use_pse) {
112 		unsigned long extra;
113 
114 		extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
115 #ifdef CONFIG_X86_32
116 		extra += PMD_SIZE;
117 #endif
118 		ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
119 	} else
120 		ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
121 
122 	tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
123 
124 #ifdef CONFIG_X86_32
125 	/* for fixmap */
126 	tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
127 #endif
128 
129 	/*
130 	 * RED-PEN putting page tables only on node 0 could
131 	 * cause a hotspot and fill up ZONE_DMA. The page tables
132 	 * need roughly 0.5KB per GB.
133 	 */
134 #ifdef CONFIG_X86_32
135 	start = 0x7000;
136 #else
137 	start = 0x8000;
138 #endif
139 	e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
140 					tables, PAGE_SIZE);
141 	if (e820_table_start == -1UL)
142 		panic("Cannot find space for the kernel page tables");
143 
144 	e820_table_start >>= PAGE_SHIFT;
145 	e820_table_end = e820_table_start;
146 	e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
147 
148 	printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
149 		end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
150 }
151 
152 struct map_range {
153 	unsigned long start;
154 	unsigned long end;
155 	unsigned page_size_mask;
156 };
157 
158 #ifdef CONFIG_X86_32
159 #define NR_RANGE_MR 3
160 #else /* CONFIG_X86_64 */
161 #define NR_RANGE_MR 5
162 #endif
163 
164 static int __meminit save_mr(struct map_range *mr, int nr_range,
165 			     unsigned long start_pfn, unsigned long end_pfn,
166 			     unsigned long page_size_mask)
167 {
168 	if (start_pfn < end_pfn) {
169 		if (nr_range >= NR_RANGE_MR)
170 			panic("run out of range for init_memory_mapping\n");
171 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
172 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
173 		mr[nr_range].page_size_mask = page_size_mask;
174 		nr_range++;
175 	}
176 
177 	return nr_range;
178 }
179 
180 #ifdef CONFIG_X86_64
181 static void __init init_gbpages(void)
182 {
183 	if (direct_gbpages && cpu_has_gbpages)
184 		printk(KERN_INFO "Using GB pages for direct mapping\n");
185 	else
186 		direct_gbpages = 0;
187 }
188 #else
189 static inline void init_gbpages(void)
190 {
191 }
192 #endif
193 
194 /*
195  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
196  * This runs before bootmem is initialized and gets pages directly from
197  * the physical memory. To access them they are temporarily mapped.
198  */
199 unsigned long __init_refok init_memory_mapping(unsigned long start,
200 					       unsigned long end)
201 {
202 	unsigned long page_size_mask = 0;
203 	unsigned long start_pfn, end_pfn;
204 	unsigned long ret = 0;
205 	unsigned long pos;
206 
207 	struct map_range mr[NR_RANGE_MR];
208 	int nr_range, i;
209 	int use_pse, use_gbpages;
210 
211 	printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
212 
213 	if (!after_bootmem)
214 		init_gbpages();
215 
216 #ifdef CONFIG_DEBUG_PAGEALLOC
217 	/*
218 	 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
219 	 * This will simplify cpa(), which otherwise needs to support splitting
220 	 * large pages into small in interrupt context, etc.
221 	 */
222 	use_pse = use_gbpages = 0;
223 #else
224 	use_pse = cpu_has_pse;
225 	use_gbpages = direct_gbpages;
226 #endif
227 
228 	set_nx();
229 	if (nx_enabled)
230 		printk(KERN_INFO "NX (Execute Disable) protection: active\n");
231 
232 	/* Enable PSE if available */
233 	if (cpu_has_pse)
234 		set_in_cr4(X86_CR4_PSE);
235 
236 	/* Enable PGE if available */
237 	if (cpu_has_pge) {
238 		set_in_cr4(X86_CR4_PGE);
239 		__supported_pte_mask |= _PAGE_GLOBAL;
240 	}
241 
242 	if (use_gbpages)
243 		page_size_mask |= 1 << PG_LEVEL_1G;
244 	if (use_pse)
245 		page_size_mask |= 1 << PG_LEVEL_2M;
246 
247 	memset(mr, 0, sizeof(mr));
248 	nr_range = 0;
249 
250 	/* head if not big page alignment ? */
251 	start_pfn = start >> PAGE_SHIFT;
252 	pos = start_pfn << PAGE_SHIFT;
253 #ifdef CONFIG_X86_32
254 	/*
255 	 * Don't use a large page for the first 2/4MB of memory
256 	 * because there are often fixed size MTRRs in there
257 	 * and overlapping MTRRs into large pages can cause
258 	 * slowdowns.
259 	 */
260 	if (pos == 0)
261 		end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
262 	else
263 		end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
264 				 << (PMD_SHIFT - PAGE_SHIFT);
265 #else /* CONFIG_X86_64 */
266 	end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
267 			<< (PMD_SHIFT - PAGE_SHIFT);
268 #endif
269 	if (end_pfn > (end >> PAGE_SHIFT))
270 		end_pfn = end >> PAGE_SHIFT;
271 	if (start_pfn < end_pfn) {
272 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
273 		pos = end_pfn << PAGE_SHIFT;
274 	}
275 
276 	/* big page (2M) range */
277 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
278 			 << (PMD_SHIFT - PAGE_SHIFT);
279 #ifdef CONFIG_X86_32
280 	end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
281 #else /* CONFIG_X86_64 */
282 	end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
283 			 << (PUD_SHIFT - PAGE_SHIFT);
284 	if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
285 		end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
286 #endif
287 
288 	if (start_pfn < end_pfn) {
289 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
290 				page_size_mask & (1<<PG_LEVEL_2M));
291 		pos = end_pfn << PAGE_SHIFT;
292 	}
293 
294 #ifdef CONFIG_X86_64
295 	/* big page (1G) range */
296 	start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
297 			 << (PUD_SHIFT - PAGE_SHIFT);
298 	end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
299 	if (start_pfn < end_pfn) {
300 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
301 				page_size_mask &
302 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
303 		pos = end_pfn << PAGE_SHIFT;
304 	}
305 
306 	/* tail is not big page (1G) alignment */
307 	start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
308 			 << (PMD_SHIFT - PAGE_SHIFT);
309 	end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
310 	if (start_pfn < end_pfn) {
311 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
312 				page_size_mask & (1<<PG_LEVEL_2M));
313 		pos = end_pfn << PAGE_SHIFT;
314 	}
315 #endif
316 
317 	/* tail is not big page (2M) alignment */
318 	start_pfn = pos>>PAGE_SHIFT;
319 	end_pfn = end>>PAGE_SHIFT;
320 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
321 
322 	/* try to merge same page size and continuous */
323 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
324 		unsigned long old_start;
325 		if (mr[i].end != mr[i+1].start ||
326 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
327 			continue;
328 		/* move it */
329 		old_start = mr[i].start;
330 		memmove(&mr[i], &mr[i+1],
331 			(nr_range - 1 - i) * sizeof(struct map_range));
332 		mr[i--].start = old_start;
333 		nr_range--;
334 	}
335 
336 	for (i = 0; i < nr_range; i++)
337 		printk(KERN_DEBUG " %010lx - %010lx page %s\n",
338 				mr[i].start, mr[i].end,
339 			(mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
340 			 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
341 
342 	/*
343 	 * Find space for the kernel direct mapping tables.
344 	 *
345 	 * Later we should allocate these tables in the local node of the
346 	 * memory mapped. Unfortunately this is done currently before the
347 	 * nodes are discovered.
348 	 */
349 	if (!after_bootmem)
350 		find_early_table_space(end, use_pse, use_gbpages);
351 
352 #ifdef CONFIG_X86_32
353 	for (i = 0; i < nr_range; i++)
354 		kernel_physical_mapping_init(mr[i].start, mr[i].end,
355 					     mr[i].page_size_mask);
356 	ret = end;
357 #else /* CONFIG_X86_64 */
358 	for (i = 0; i < nr_range; i++)
359 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
360 						   mr[i].page_size_mask);
361 #endif
362 
363 #ifdef CONFIG_X86_32
364 	early_ioremap_page_table_range_init();
365 
366 	load_cr3(swapper_pg_dir);
367 #endif
368 
369 #ifdef CONFIG_X86_64
370 	if (!after_bootmem && !start) {
371 		pud_t *pud;
372 		pmd_t *pmd;
373 
374 		mmu_cr4_features = read_cr4();
375 
376 		/*
377 		 * _brk_end cannot change anymore, but it and _end may be
378 		 * located on different 2M pages. cleanup_highmap(), however,
379 		 * can only consider _end when it runs, so destroy any
380 		 * mappings beyond _brk_end here.
381 		 */
382 		pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
383 		pmd = pmd_offset(pud, _brk_end - 1);
384 		while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
385 			pmd_clear(pmd);
386 	}
387 #endif
388 	__flush_tlb_all();
389 
390 	if (!after_bootmem && e820_table_end > e820_table_start)
391 		reserve_early(e820_table_start << PAGE_SHIFT,
392 				 e820_table_end << PAGE_SHIFT, "PGTABLE");
393 
394 	if (!after_bootmem)
395 		early_memtest(start, end);
396 
397 	return ret >> PAGE_SHIFT;
398 }
399 
400 
401 /*
402  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
403  * is valid. The argument is a physical page number.
404  *
405  *
406  * On x86, access has to be given to the first megabyte of ram because that area
407  * contains bios code and data regions used by X and dosemu and similar apps.
408  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
409  * mmio resources as well as potential bios/acpi data regions.
410  */
411 int devmem_is_allowed(unsigned long pagenr)
412 {
413 	if (pagenr <= 256)
414 		return 1;
415 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
416 		return 0;
417 	if (!page_is_ram(pagenr))
418 		return 1;
419 	return 0;
420 }
421 
422 void free_init_pages(char *what, unsigned long begin, unsigned long end)
423 {
424 	unsigned long addr = begin;
425 
426 	if (addr >= end)
427 		return;
428 
429 	/*
430 	 * If debugging page accesses then do not free this memory but
431 	 * mark them not present - any buggy init-section access will
432 	 * create a kernel page fault:
433 	 */
434 #ifdef CONFIG_DEBUG_PAGEALLOC
435 	printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
436 		begin, PAGE_ALIGN(end));
437 	set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
438 #else
439 	/*
440 	 * We just marked the kernel text read only above, now that
441 	 * we are going to free part of that, we need to make that
442 	 * writeable first.
443 	 */
444 	set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
445 
446 	printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
447 
448 	for (; addr < end; addr += PAGE_SIZE) {
449 		ClearPageReserved(virt_to_page(addr));
450 		init_page_count(virt_to_page(addr));
451 		memset((void *)(addr & ~(PAGE_SIZE-1)),
452 			POISON_FREE_INITMEM, PAGE_SIZE);
453 		free_page(addr);
454 		totalram_pages++;
455 	}
456 #endif
457 }
458 
459 void free_initmem(void)
460 {
461 	free_init_pages("unused kernel memory",
462 			(unsigned long)(&__init_begin),
463 			(unsigned long)(&__init_end));
464 }
465 
466 #ifdef CONFIG_BLK_DEV_INITRD
467 void free_initrd_mem(unsigned long start, unsigned long end)
468 {
469 	free_init_pages("initrd memory", start, end);
470 }
471 #endif
472