1 #include <linux/ioport.h>
2 #include <linux/swap.h>
4 #include <asm/cacheflush.h>
7 #include <asm/page_types.h>
8 #include <asm/sections.h>
9 #include <asm/system.h>
10 #include <asm/tlbflush.h>
13 extern void __init early_ioremap_page_table_range_init(void);
14 extern void __init kernel_physical_mapping_init(unsigned long start_pfn,
15 unsigned long end_pfn,
20 extern unsigned long __meminit
21 kernel_physical_mapping_init(unsigned long start,
23 unsigned long page_size_mask);
26 unsigned long __initdata e820_table_start;
27 unsigned long __meminitdata e820_table_end;
28 unsigned long __meminitdata e820_table_top;
33 #ifdef CONFIG_DIRECT_GBPAGES
38 static void __init find_early_table_space(unsigned long end, int use_pse,
41 unsigned long puds, pmds, ptes, tables, start;
43 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
44 tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
49 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
50 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
52 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
54 tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
59 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
63 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
65 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
67 tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
71 tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
75 * RED-PEN putting page tables only on node 0 could
76 * cause a hotspot and fill up ZONE_DMA. The page tables
77 * need roughly 0.5KB per GB.
81 e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
83 #else /* CONFIG_X86_64 */
85 e820_table_start = find_e820_area(start, end, tables, PAGE_SIZE);
87 if (e820_table_start == -1UL)
88 panic("Cannot find space for the kernel page tables");
90 e820_table_start >>= PAGE_SHIFT;
91 e820_table_end = e820_table_start;
92 e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
94 printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
95 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
101 unsigned page_size_mask;
105 #define NR_RANGE_MR 3
106 #else /* CONFIG_X86_64 */
107 #define NR_RANGE_MR 5
110 static int save_mr(struct map_range *mr, int nr_range,
111 unsigned long start_pfn, unsigned long end_pfn,
112 unsigned long page_size_mask)
114 if (start_pfn < end_pfn) {
115 if (nr_range >= NR_RANGE_MR)
116 panic("run out of range for init_memory_mapping\n");
117 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
118 mr[nr_range].end = end_pfn<<PAGE_SHIFT;
119 mr[nr_range].page_size_mask = page_size_mask;
127 static void __init init_gbpages(void)
129 if (direct_gbpages && cpu_has_gbpages)
130 printk(KERN_INFO "Using GB pages for direct mapping\n");
135 static inline void init_gbpages(void)
141 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
142 * This runs before bootmem is initialized and gets pages directly from
143 * the physical memory. To access them they are temporarily mapped.
145 unsigned long __init_refok init_memory_mapping(unsigned long start,
148 unsigned long page_size_mask = 0;
149 unsigned long start_pfn, end_pfn;
153 struct map_range mr[NR_RANGE_MR];
155 int use_pse, use_gbpages;
157 printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
162 #ifdef CONFIG_DEBUG_PAGEALLOC
164 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
165 * This will simplify cpa(), which otherwise needs to support splitting
166 * large pages into small in interrupt context, etc.
168 use_pse = use_gbpages = 0;
170 use_pse = cpu_has_pse;
171 use_gbpages = direct_gbpages;
175 #ifdef CONFIG_X86_PAE
178 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
181 /* Enable PSE if available */
183 set_in_cr4(X86_CR4_PSE);
185 /* Enable PGE if available */
187 set_in_cr4(X86_CR4_PGE);
188 __supported_pte_mask |= _PAGE_GLOBAL;
193 page_size_mask |= 1 << PG_LEVEL_1G;
195 page_size_mask |= 1 << PG_LEVEL_2M;
197 memset(mr, 0, sizeof(mr));
200 /* head if not big page alignment ? */
201 start_pfn = start >> PAGE_SHIFT;
202 pos = start_pfn << PAGE_SHIFT;
205 * Don't use a large page for the first 2/4MB of memory
206 * because there are often fixed size MTRRs in there
207 * and overlapping MTRRs into large pages can cause
211 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
213 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
214 << (PMD_SHIFT - PAGE_SHIFT);
215 #else /* CONFIG_X86_64 */
216 end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
217 << (PMD_SHIFT - PAGE_SHIFT);
219 if (end_pfn > (end >> PAGE_SHIFT))
220 end_pfn = end >> PAGE_SHIFT;
221 if (start_pfn < end_pfn) {
222 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
223 pos = end_pfn << PAGE_SHIFT;
226 /* big page (2M) range */
227 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
228 << (PMD_SHIFT - PAGE_SHIFT);
230 end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
231 #else /* CONFIG_X86_64 */
232 end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
233 << (PUD_SHIFT - PAGE_SHIFT);
234 if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
235 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
238 if (start_pfn < end_pfn) {
239 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
240 page_size_mask & (1<<PG_LEVEL_2M));
241 pos = end_pfn << PAGE_SHIFT;
245 /* big page (1G) range */
246 start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
247 << (PUD_SHIFT - PAGE_SHIFT);
248 end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
249 if (start_pfn < end_pfn) {
250 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
252 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
253 pos = end_pfn << PAGE_SHIFT;
256 /* tail is not big page (1G) alignment */
257 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
258 << (PMD_SHIFT - PAGE_SHIFT);
259 end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
260 if (start_pfn < end_pfn) {
261 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
262 page_size_mask & (1<<PG_LEVEL_2M));
263 pos = end_pfn << PAGE_SHIFT;
267 /* tail is not big page (2M) alignment */
268 start_pfn = pos>>PAGE_SHIFT;
269 end_pfn = end>>PAGE_SHIFT;
270 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
272 /* try to merge same page size and continuous */
273 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
274 unsigned long old_start;
275 if (mr[i].end != mr[i+1].start ||
276 mr[i].page_size_mask != mr[i+1].page_size_mask)
279 old_start = mr[i].start;
280 memmove(&mr[i], &mr[i+1],
281 (nr_range - 1 - i) * sizeof(struct map_range));
282 mr[i--].start = old_start;
286 for (i = 0; i < nr_range; i++)
287 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
288 mr[i].start, mr[i].end,
289 (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
290 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
293 * Find space for the kernel direct mapping tables.
295 * Later we should allocate these tables in the local node of the
296 * memory mapped. Unfortunately this is done currently before the
297 * nodes are discovered.
300 find_early_table_space(end, use_pse, use_gbpages);
303 for (i = 0; i < nr_range; i++)
304 kernel_physical_mapping_init(
305 mr[i].start >> PAGE_SHIFT,
306 mr[i].end >> PAGE_SHIFT,
307 mr[i].page_size_mask == (1<<PG_LEVEL_2M));
309 #else /* CONFIG_X86_64 */
310 for (i = 0; i < nr_range; i++)
311 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
312 mr[i].page_size_mask);
316 early_ioremap_page_table_range_init();
318 load_cr3(swapper_pg_dir);
323 mmu_cr4_features = read_cr4();
327 if (!after_bootmem && e820_table_end > e820_table_start)
328 reserve_early(e820_table_start << PAGE_SHIFT,
329 e820_table_end << PAGE_SHIFT, "PGTABLE");
332 early_memtest(start, end);
334 return ret >> PAGE_SHIFT;
339 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
340 * is valid. The argument is a physical page number.
343 * On x86, access has to be given to the first megabyte of ram because that area
344 * contains bios code and data regions used by X and dosemu and similar apps.
345 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
346 * mmio resources as well as potential bios/acpi data regions.
348 int devmem_is_allowed(unsigned long pagenr)
352 if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
354 if (!page_is_ram(pagenr))
359 void free_init_pages(char *what, unsigned long begin, unsigned long end)
361 unsigned long addr = begin;
367 * If debugging page accesses then do not free this memory but
368 * mark them not present - any buggy init-section access will
369 * create a kernel page fault:
371 #ifdef CONFIG_DEBUG_PAGEALLOC
372 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
373 begin, PAGE_ALIGN(end));
374 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
377 * We just marked the kernel text read only above, now that
378 * we are going to free part of that, we need to make that
381 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
383 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
385 for (; addr < end; addr += PAGE_SIZE) {
386 ClearPageReserved(virt_to_page(addr));
387 init_page_count(virt_to_page(addr));
388 memset((void *)(addr & ~(PAGE_SIZE-1)),
389 POISON_FREE_INITMEM, PAGE_SIZE);
396 void free_initmem(void)
398 free_init_pages("unused kernel memory",
399 (unsigned long)(&__init_begin),
400 (unsigned long)(&__init_end));
403 #ifdef CONFIG_BLK_DEV_INITRD
404 void free_initrd_mem(unsigned long start, unsigned long end)
406 free_init_pages("initrd memory", start, end);