2 * linux/arch/arm/mm/mmu.c
4 * Copyright (C) 1995-2005 Russell King
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
18 #include <linux/vmalloc.h>
19 #include <linux/sizes.h>
22 #include <asm/cputype.h>
23 #include <asm/sections.h>
24 #include <asm/cachetype.h>
25 #include <asm/setup.h>
26 #include <asm/smp_plat.h>
28 #include <asm/highmem.h>
29 #include <asm/system_info.h>
30 #include <asm/traps.h>
32 #include <asm/mach/arch.h>
33 #include <asm/mach/map.h>
38 * empty_zero_page is a special page that is used for
39 * zero-initialized data and COW.
41 struct page *empty_zero_page;
42 EXPORT_SYMBOL(empty_zero_page);
45 * The pmd table for the upper-most set of pages.
49 #define CPOLICY_UNCACHED 0
50 #define CPOLICY_BUFFERED 1
51 #define CPOLICY_WRITETHROUGH 2
52 #define CPOLICY_WRITEBACK 3
53 #define CPOLICY_WRITEALLOC 4
55 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
56 static unsigned int ecc_mask __initdata = 0;
58 pgprot_t pgprot_kernel;
60 EXPORT_SYMBOL(pgprot_user);
61 EXPORT_SYMBOL(pgprot_kernel);
64 const char policy[16];
70 static struct cachepolicy cache_policies[] __initdata = {
74 .pmd = PMD_SECT_UNCACHED,
75 .pte = L_PTE_MT_UNCACHED,
79 .pmd = PMD_SECT_BUFFERED,
80 .pte = L_PTE_MT_BUFFERABLE,
82 .policy = "writethrough",
85 .pte = L_PTE_MT_WRITETHROUGH,
87 .policy = "writeback",
90 .pte = L_PTE_MT_WRITEBACK,
92 .policy = "writealloc",
95 .pte = L_PTE_MT_WRITEALLOC,
100 * These are useful for identifying cache coherency
101 * problems by allowing the cache or the cache and
102 * writebuffer to be turned off. (Note: the write
103 * buffer should not be on and the cache off).
105 static int __init early_cachepolicy(char *p)
109 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
110 int len = strlen(cache_policies[i].policy);
112 if (memcmp(p, cache_policies[i].policy, len) == 0) {
114 cr_alignment &= ~cache_policies[i].cr_mask;
115 cr_no_alignment &= ~cache_policies[i].cr_mask;
119 if (i == ARRAY_SIZE(cache_policies))
120 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
122 * This restriction is partly to do with the way we boot; it is
123 * unpredictable to have memory mapped using two different sets of
124 * memory attributes (shared, type, and cache attribs). We can not
125 * change these attributes once the initial assembly has setup the
128 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
129 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
130 cachepolicy = CPOLICY_WRITEBACK;
133 set_cr(cr_alignment);
136 early_param("cachepolicy", early_cachepolicy);
138 static int __init early_nocache(char *__unused)
140 char *p = "buffered";
141 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
142 early_cachepolicy(p);
145 early_param("nocache", early_nocache);
147 static int __init early_nowrite(char *__unused)
149 char *p = "uncached";
150 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
151 early_cachepolicy(p);
154 early_param("nowb", early_nowrite);
156 #ifndef CONFIG_ARM_LPAE
157 static int __init early_ecc(char *p)
159 if (memcmp(p, "on", 2) == 0)
160 ecc_mask = PMD_PROTECTION;
161 else if (memcmp(p, "off", 3) == 0)
165 early_param("ecc", early_ecc);
168 static int __init noalign_setup(char *__unused)
170 cr_alignment &= ~CR_A;
171 cr_no_alignment &= ~CR_A;
172 set_cr(cr_alignment);
175 __setup("noalign", noalign_setup);
178 void adjust_cr(unsigned long mask, unsigned long set)
186 local_irq_save(flags);
188 cr_no_alignment = (cr_no_alignment & ~mask) | set;
189 cr_alignment = (cr_alignment & ~mask) | set;
191 set_cr((get_cr() & ~mask) | set);
193 local_irq_restore(flags);
197 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
198 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
200 static struct mem_type mem_types[] = {
201 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
202 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
204 .prot_l1 = PMD_TYPE_TABLE,
205 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
208 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
209 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
210 .prot_l1 = PMD_TYPE_TABLE,
211 .prot_sect = PROT_SECT_DEVICE,
214 [MT_DEVICE_CACHED] = { /* ioremap_cached */
215 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
216 .prot_l1 = PMD_TYPE_TABLE,
217 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
220 [MT_DEVICE_WC] = { /* ioremap_wc */
221 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
222 .prot_l1 = PMD_TYPE_TABLE,
223 .prot_sect = PROT_SECT_DEVICE,
227 .prot_pte = PROT_PTE_DEVICE,
228 .prot_l1 = PMD_TYPE_TABLE,
229 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
233 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
234 .domain = DOMAIN_KERNEL,
236 #ifndef CONFIG_ARM_LPAE
238 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
239 .domain = DOMAIN_KERNEL,
243 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
245 .prot_l1 = PMD_TYPE_TABLE,
246 .domain = DOMAIN_USER,
248 [MT_HIGH_VECTORS] = {
249 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
250 L_PTE_USER | L_PTE_RDONLY,
251 .prot_l1 = PMD_TYPE_TABLE,
252 .domain = DOMAIN_USER,
255 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
256 .prot_l1 = PMD_TYPE_TABLE,
257 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
258 .domain = DOMAIN_KERNEL,
261 .prot_sect = PMD_TYPE_SECT,
262 .domain = DOMAIN_KERNEL,
264 [MT_MEMORY_NONCACHED] = {
265 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
267 .prot_l1 = PMD_TYPE_TABLE,
268 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
269 .domain = DOMAIN_KERNEL,
272 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
274 .prot_l1 = PMD_TYPE_TABLE,
275 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
276 .domain = DOMAIN_KERNEL,
279 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
280 .prot_l1 = PMD_TYPE_TABLE,
281 .domain = DOMAIN_KERNEL,
284 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
286 .prot_l1 = PMD_TYPE_TABLE,
287 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
288 PMD_SECT_UNCACHED | PMD_SECT_XN,
289 .domain = DOMAIN_KERNEL,
291 [MT_MEMORY_DMA_READY] = {
292 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
293 .prot_l1 = PMD_TYPE_TABLE,
294 .domain = DOMAIN_KERNEL,
298 const struct mem_type *get_mem_type(unsigned int type)
300 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
302 EXPORT_SYMBOL(get_mem_type);
305 * Adjust the PMD section entries according to the CPU in use.
307 static void __init build_mem_type_table(void)
309 struct cachepolicy *cp;
310 unsigned int cr = get_cr();
311 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
312 int cpu_arch = cpu_architecture();
315 if (cpu_arch < CPU_ARCH_ARMv6) {
316 #if defined(CONFIG_CPU_DCACHE_DISABLE)
317 if (cachepolicy > CPOLICY_BUFFERED)
318 cachepolicy = CPOLICY_BUFFERED;
319 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
320 if (cachepolicy > CPOLICY_WRITETHROUGH)
321 cachepolicy = CPOLICY_WRITETHROUGH;
324 if (cpu_arch < CPU_ARCH_ARMv5) {
325 if (cachepolicy >= CPOLICY_WRITEALLOC)
326 cachepolicy = CPOLICY_WRITEBACK;
330 cachepolicy = CPOLICY_WRITEALLOC;
333 * Strip out features not present on earlier architectures.
334 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
335 * without extended page tables don't have the 'Shared' bit.
337 if (cpu_arch < CPU_ARCH_ARMv5)
338 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
339 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
340 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
341 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
342 mem_types[i].prot_sect &= ~PMD_SECT_S;
345 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
346 * "update-able on write" bit on ARM610). However, Xscale and
347 * Xscale3 require this bit to be cleared.
349 if (cpu_is_xscale() || cpu_is_xsc3()) {
350 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
351 mem_types[i].prot_sect &= ~PMD_BIT4;
352 mem_types[i].prot_l1 &= ~PMD_BIT4;
354 } else if (cpu_arch < CPU_ARCH_ARMv6) {
355 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
356 if (mem_types[i].prot_l1)
357 mem_types[i].prot_l1 |= PMD_BIT4;
358 if (mem_types[i].prot_sect)
359 mem_types[i].prot_sect |= PMD_BIT4;
364 * Mark the device areas according to the CPU/architecture.
366 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
367 if (!cpu_is_xsc3()) {
369 * Mark device regions on ARMv6+ as execute-never
370 * to prevent speculative instruction fetches.
372 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
373 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
374 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
375 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
377 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
379 * For ARMv7 with TEX remapping,
380 * - shared device is SXCB=1100
381 * - nonshared device is SXCB=0100
382 * - write combine device mem is SXCB=0001
383 * (Uncached Normal memory)
385 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
386 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
387 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
388 } else if (cpu_is_xsc3()) {
391 * - shared device is TEXCB=00101
392 * - nonshared device is TEXCB=01000
393 * - write combine device mem is TEXCB=00100
394 * (Inner/Outer Uncacheable in xsc3 parlance)
396 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
397 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
398 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
401 * For ARMv6 and ARMv7 without TEX remapping,
402 * - shared device is TEXCB=00001
403 * - nonshared device is TEXCB=01000
404 * - write combine device mem is TEXCB=00100
405 * (Uncached Normal in ARMv6 parlance).
407 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
408 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
409 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
413 * On others, write combining is "Uncached/Buffered"
415 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
419 * Now deal with the memory-type mappings
421 cp = &cache_policies[cachepolicy];
422 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
425 * Enable CPU-specific coherency if supported.
426 * (Only available on XSC3 at the moment.)
428 if (arch_is_coherent() && cpu_is_xsc3()) {
429 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
430 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
431 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
432 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
433 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
436 * ARMv6 and above have extended page tables.
438 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
439 #ifndef CONFIG_ARM_LPAE
441 * Mark cache clean areas and XIP ROM read only
442 * from SVC mode and no access from userspace.
444 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
445 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
446 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
451 * Mark memory with the "shared" attribute
454 user_pgprot |= L_PTE_SHARED;
455 kern_pgprot |= L_PTE_SHARED;
456 vecs_pgprot |= L_PTE_SHARED;
457 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
458 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
459 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
460 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
461 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
462 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
463 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
464 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
465 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
470 * Non-cacheable Normal - intended for memory areas that must
471 * not cause dirty cache line writebacks when used
473 if (cpu_arch >= CPU_ARCH_ARMv6) {
474 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
475 /* Non-cacheable Normal is XCB = 001 */
476 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
479 /* For both ARMv6 and non-TEX-remapping ARMv7 */
480 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
484 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
487 #ifdef CONFIG_ARM_LPAE
489 * Do not generate access flag faults for the kernel mappings.
491 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
492 mem_types[i].prot_pte |= PTE_EXT_AF;
493 if (mem_types[i].prot_sect)
494 mem_types[i].prot_sect |= PMD_SECT_AF;
496 kern_pgprot |= PTE_EXT_AF;
497 vecs_pgprot |= PTE_EXT_AF;
500 for (i = 0; i < 16; i++) {
501 unsigned long v = pgprot_val(protection_map[i]);
502 protection_map[i] = __pgprot(v | user_pgprot);
505 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
506 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
508 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
509 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
510 L_PTE_DIRTY | kern_pgprot);
512 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
513 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
514 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
515 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
516 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
517 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
518 mem_types[MT_ROM].prot_sect |= cp->pmd;
522 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
526 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
529 printk("Memory policy: ECC %sabled, Data cache %s\n",
530 ecc_mask ? "en" : "dis", cp->policy);
532 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
533 struct mem_type *t = &mem_types[i];
535 t->prot_l1 |= PMD_DOMAIN(t->domain);
537 t->prot_sect |= PMD_DOMAIN(t->domain);
541 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
542 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
543 unsigned long size, pgprot_t vma_prot)
546 return pgprot_noncached(vma_prot);
547 else if (file->f_flags & O_SYNC)
548 return pgprot_writecombine(vma_prot);
551 EXPORT_SYMBOL(phys_mem_access_prot);
554 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
556 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
558 void *ptr = __va(memblock_alloc(sz, align));
563 static void __init *early_alloc(unsigned long sz)
565 return early_alloc_aligned(sz, sz);
568 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
570 if (pmd_none(*pmd)) {
571 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
572 __pmd_populate(pmd, __pa(pte), prot);
574 BUG_ON(pmd_bad(*pmd));
575 return pte_offset_kernel(pmd, addr);
578 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
579 unsigned long end, unsigned long pfn,
580 const struct mem_type *type)
582 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
584 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
586 } while (pte++, addr += PAGE_SIZE, addr != end);
589 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
590 unsigned long end, phys_addr_t phys,
591 const struct mem_type *type)
593 pmd_t *pmd = pmd_offset(pud, addr);
596 * Try a section mapping - end, addr and phys must all be aligned
597 * to a section boundary. Note that PMDs refer to the individual
598 * L1 entries, whereas PGDs refer to a group of L1 entries making
599 * up one logical pointer to an L2 table.
601 if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
604 #ifndef CONFIG_ARM_LPAE
605 if (addr & SECTION_SIZE)
610 *pmd = __pmd(phys | type->prot_sect);
611 phys += SECTION_SIZE;
612 } while (pmd++, addr += SECTION_SIZE, addr != end);
617 * No need to loop; pte's aren't interested in the
618 * individual L1 entries.
620 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
624 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
625 unsigned long end, unsigned long phys, const struct mem_type *type)
627 pud_t *pud = pud_offset(pgd, addr);
631 next = pud_addr_end(addr, end);
632 alloc_init_section(pud, addr, next, phys, type);
634 } while (pud++, addr = next, addr != end);
637 #ifndef CONFIG_ARM_LPAE
638 static void __init create_36bit_mapping(struct map_desc *md,
639 const struct mem_type *type)
641 unsigned long addr, length, end;
646 phys = __pfn_to_phys(md->pfn);
647 length = PAGE_ALIGN(md->length);
649 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
650 printk(KERN_ERR "MM: CPU does not support supersection "
651 "mapping for 0x%08llx at 0x%08lx\n",
652 (long long)__pfn_to_phys((u64)md->pfn), addr);
656 /* N.B. ARMv6 supersections are only defined to work with domain 0.
657 * Since domain assignments can in fact be arbitrary, the
658 * 'domain == 0' check below is required to insure that ARMv6
659 * supersections are only allocated for domain 0 regardless
660 * of the actual domain assignments in use.
663 printk(KERN_ERR "MM: invalid domain in supersection "
664 "mapping for 0x%08llx at 0x%08lx\n",
665 (long long)__pfn_to_phys((u64)md->pfn), addr);
669 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
670 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
671 " at 0x%08lx invalid alignment\n",
672 (long long)__pfn_to_phys((u64)md->pfn), addr);
677 * Shift bits [35:32] of address into bits [23:20] of PMD
680 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
682 pgd = pgd_offset_k(addr);
685 pud_t *pud = pud_offset(pgd, addr);
686 pmd_t *pmd = pmd_offset(pud, addr);
689 for (i = 0; i < 16; i++)
690 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
692 addr += SUPERSECTION_SIZE;
693 phys += SUPERSECTION_SIZE;
694 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
695 } while (addr != end);
697 #endif /* !CONFIG_ARM_LPAE */
700 * Create the page directory entries and any necessary
701 * page tables for the mapping specified by `md'. We
702 * are able to cope here with varying sizes and address
703 * offsets, and we take full advantage of sections and
706 static void __init create_mapping(struct map_desc *md)
708 unsigned long addr, length, end;
710 const struct mem_type *type;
713 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
714 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
715 " at 0x%08lx in user region\n",
716 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
720 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
721 md->virtual >= PAGE_OFFSET &&
722 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
723 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
724 " at 0x%08lx out of vmalloc space\n",
725 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
728 type = &mem_types[md->type];
730 #ifndef CONFIG_ARM_LPAE
732 * Catch 36-bit addresses
734 if (md->pfn >= 0x100000) {
735 create_36bit_mapping(md, type);
740 addr = md->virtual & PAGE_MASK;
741 phys = __pfn_to_phys(md->pfn);
742 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
744 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
745 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
746 "be mapped using pages, ignoring.\n",
747 (long long)__pfn_to_phys(md->pfn), addr);
751 pgd = pgd_offset_k(addr);
754 unsigned long next = pgd_addr_end(addr, end);
756 alloc_init_pud(pgd, addr, next, phys, type);
760 } while (pgd++, addr != end);
764 * Create the architecture specific mappings
766 void __init iotable_init(struct map_desc *io_desc, int nr)
769 struct vm_struct *vm;
774 vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
776 for (md = io_desc; nr; md++, nr--) {
778 vm->addr = (void *)(md->virtual & PAGE_MASK);
779 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
780 vm->phys_addr = __pfn_to_phys(md->pfn);
781 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
782 vm->flags |= VM_ARM_MTYPE(md->type);
783 vm->caller = iotable_init;
784 vm_area_add_early(vm++);
788 #ifndef CONFIG_ARM_LPAE
791 * The Linux PMD is made of two consecutive section entries covering 2MB
792 * (see definition in include/asm/pgtable-2level.h). However a call to
793 * create_mapping() may optimize static mappings by using individual
794 * 1MB section mappings. This leaves the actual PMD potentially half
795 * initialized if the top or bottom section entry isn't used, leaving it
796 * open to problems if a subsequent ioremap() or vmalloc() tries to use
797 * the virtual space left free by that unused section entry.
799 * Let's avoid the issue by inserting dummy vm entries covering the unused
800 * PMD halves once the static mappings are in place.
803 static void __init pmd_empty_section_gap(unsigned long addr)
805 struct vm_struct *vm;
807 vm = early_alloc_aligned(sizeof(*vm), __alignof__(*vm));
808 vm->addr = (void *)addr;
809 vm->size = SECTION_SIZE;
810 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
811 vm->caller = pmd_empty_section_gap;
812 vm_area_add_early(vm);
815 static void __init fill_pmd_gaps(void)
817 struct vm_struct *vm;
818 unsigned long addr, next = 0;
821 /* we're still single threaded hence no lock needed here */
822 for (vm = vmlist; vm; vm = vm->next) {
823 if (!(vm->flags & VM_ARM_STATIC_MAPPING))
825 addr = (unsigned long)vm->addr;
830 * Check if this vm starts on an odd section boundary.
831 * If so and the first section entry for this PMD is free
832 * then we block the corresponding virtual address.
834 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
835 pmd = pmd_off_k(addr);
837 pmd_empty_section_gap(addr & PMD_MASK);
841 * Then check if this vm ends on an odd section boundary.
842 * If so and the second section entry for this PMD is empty
843 * then we block the corresponding virtual address.
846 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
847 pmd = pmd_off_k(addr) + 1;
849 pmd_empty_section_gap(addr);
852 /* no need to look at any vm entry until we hit the next PMD */
853 next = (addr + PMD_SIZE - 1) & PMD_MASK;
858 #define fill_pmd_gaps() do { } while (0)
861 static void * __initdata vmalloc_min =
862 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
865 * vmalloc=size forces the vmalloc area to be exactly 'size'
866 * bytes. This can be used to increase (or decrease) the vmalloc
867 * area - the default is 240m.
869 static int __init early_vmalloc(char *arg)
871 unsigned long vmalloc_reserve = memparse(arg, NULL);
873 if (vmalloc_reserve < SZ_16M) {
874 vmalloc_reserve = SZ_16M;
876 "vmalloc area too small, limiting to %luMB\n",
877 vmalloc_reserve >> 20);
880 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
881 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
883 "vmalloc area is too big, limiting to %luMB\n",
884 vmalloc_reserve >> 20);
887 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
890 early_param("vmalloc", early_vmalloc);
892 phys_addr_t arm_lowmem_limit __initdata = 0;
894 void __init sanity_check_meminfo(void)
896 int i, j, highmem = 0;
898 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
899 struct membank *bank = &meminfo.bank[j];
900 *bank = meminfo.bank[i];
902 if (bank->start > ULONG_MAX)
905 #ifdef CONFIG_HIGHMEM
906 if (__va(bank->start) >= vmalloc_min ||
907 __va(bank->start) < (void *)PAGE_OFFSET)
910 bank->highmem = highmem;
913 * Split those memory banks which are partially overlapping
914 * the vmalloc area greatly simplifying things later.
916 if (!highmem && __va(bank->start) < vmalloc_min &&
917 bank->size > vmalloc_min - __va(bank->start)) {
918 if (meminfo.nr_banks >= NR_BANKS) {
919 printk(KERN_CRIT "NR_BANKS too low, "
920 "ignoring high memory\n");
922 memmove(bank + 1, bank,
923 (meminfo.nr_banks - i) * sizeof(*bank));
926 bank[1].size -= vmalloc_min - __va(bank->start);
927 bank[1].start = __pa(vmalloc_min - 1) + 1;
928 bank[1].highmem = highmem = 1;
931 bank->size = vmalloc_min - __va(bank->start);
934 bank->highmem = highmem;
937 * Highmem banks not allowed with !CONFIG_HIGHMEM.
940 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
941 "(!CONFIG_HIGHMEM).\n",
942 (unsigned long long)bank->start,
943 (unsigned long long)bank->start + bank->size - 1);
948 * Check whether this memory bank would entirely overlap
951 if (__va(bank->start) >= vmalloc_min ||
952 __va(bank->start) < (void *)PAGE_OFFSET) {
953 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
954 "(vmalloc region overlap).\n",
955 (unsigned long long)bank->start,
956 (unsigned long long)bank->start + bank->size - 1);
961 * Check whether this memory bank would partially overlap
964 if (__va(bank->start + bank->size) > vmalloc_min ||
965 __va(bank->start + bank->size) < __va(bank->start)) {
966 unsigned long newsize = vmalloc_min - __va(bank->start);
967 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
968 "to -%.8llx (vmalloc region overlap).\n",
969 (unsigned long long)bank->start,
970 (unsigned long long)bank->start + bank->size - 1,
971 (unsigned long long)bank->start + newsize - 1);
972 bank->size = newsize;
975 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
976 arm_lowmem_limit = bank->start + bank->size;
980 #ifdef CONFIG_HIGHMEM
982 const char *reason = NULL;
984 if (cache_is_vipt_aliasing()) {
986 * Interactions between kmap and other mappings
987 * make highmem support with aliasing VIPT caches
990 reason = "with VIPT aliasing cache";
993 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
995 while (j > 0 && meminfo.bank[j - 1].highmem)
1000 meminfo.nr_banks = j;
1001 high_memory = __va(arm_lowmem_limit - 1) + 1;
1002 memblock_set_current_limit(arm_lowmem_limit);
1005 static inline void prepare_page_table(void)
1011 * Clear out all the mappings below the kernel image.
1013 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1014 pmd_clear(pmd_off_k(addr));
1016 #ifdef CONFIG_XIP_KERNEL
1017 /* The XIP kernel is mapped in the module area -- skip over it */
1018 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1020 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1021 pmd_clear(pmd_off_k(addr));
1024 * Find the end of the first block of lowmem.
1026 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1027 if (end >= arm_lowmem_limit)
1028 end = arm_lowmem_limit;
1031 * Clear out all the kernel space mappings, except for the first
1032 * memory bank, up to the vmalloc region.
1034 for (addr = __phys_to_virt(end);
1035 addr < VMALLOC_START; addr += PMD_SIZE)
1036 pmd_clear(pmd_off_k(addr));
1039 #ifdef CONFIG_ARM_LPAE
1040 /* the first page is reserved for pgd */
1041 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1042 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1044 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1048 * Reserve the special regions of memory
1050 void __init arm_mm_memblock_reserve(void)
1053 * Reserve the page tables. These are already in use,
1054 * and can only be in node 0.
1056 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1058 #ifdef CONFIG_SA1111
1060 * Because of the SA1111 DMA bug, we want to preserve our
1061 * precious DMA-able memory...
1063 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1068 * Set up the device mappings. Since we clear out the page tables for all
1069 * mappings above VMALLOC_START, we will remove any debug device mappings.
1070 * This means you have to be careful how you debug this function, or any
1071 * called function. This means you can't use any function or debugging
1072 * method which may touch any device, otherwise the kernel _will_ crash.
1074 static void __init devicemaps_init(struct machine_desc *mdesc)
1076 struct map_desc map;
1081 * Allocate the vector page early.
1083 vectors = early_alloc(PAGE_SIZE);
1085 early_trap_init(vectors);
1087 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1088 pmd_clear(pmd_off_k(addr));
1091 * Map the kernel if it is XIP.
1092 * It is always first in the modulearea.
1094 #ifdef CONFIG_XIP_KERNEL
1095 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1096 map.virtual = MODULES_VADDR;
1097 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1099 create_mapping(&map);
1103 * Map the cache flushing regions.
1106 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1107 map.virtual = FLUSH_BASE;
1109 map.type = MT_CACHECLEAN;
1110 create_mapping(&map);
1112 #ifdef FLUSH_BASE_MINICACHE
1113 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1114 map.virtual = FLUSH_BASE_MINICACHE;
1116 map.type = MT_MINICLEAN;
1117 create_mapping(&map);
1121 * Create a mapping for the machine vectors at the high-vectors
1122 * location (0xffff0000). If we aren't using high-vectors, also
1123 * create a mapping at the low-vectors virtual address.
1125 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1126 map.virtual = 0xffff0000;
1127 map.length = PAGE_SIZE;
1128 map.type = MT_HIGH_VECTORS;
1129 create_mapping(&map);
1131 if (!vectors_high()) {
1133 map.type = MT_LOW_VECTORS;
1134 create_mapping(&map);
1138 * Ask the machine support to map in the statically mapped devices.
1145 * Finally flush the caches and tlb to ensure that we're in a
1146 * consistent state wrt the writebuffer. This also ensures that
1147 * any write-allocated cache lines in the vector page are written
1148 * back. After this point, we can start to touch devices again.
1150 local_flush_tlb_all();
1154 static void __init kmap_init(void)
1156 #ifdef CONFIG_HIGHMEM
1157 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1158 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1162 static void __init map_lowmem(void)
1164 struct memblock_region *reg;
1166 /* Map all the lowmem memory banks. */
1167 for_each_memblock(memory, reg) {
1168 phys_addr_t start = reg->base;
1169 phys_addr_t end = start + reg->size;
1170 struct map_desc map;
1172 if (end > arm_lowmem_limit)
1173 end = arm_lowmem_limit;
1177 map.pfn = __phys_to_pfn(start);
1178 map.virtual = __phys_to_virt(start);
1179 map.length = end - start;
1180 map.type = MT_MEMORY;
1182 create_mapping(&map);
1187 * paging_init() sets up the page tables, initialises the zone memory
1188 * maps, and sets up the zero page, bad page and bad page tables.
1190 void __init paging_init(struct machine_desc *mdesc)
1194 memblock_set_current_limit(arm_lowmem_limit);
1196 build_mem_type_table();
1197 prepare_page_table();
1199 dma_contiguous_remap();
1200 devicemaps_init(mdesc);
1203 top_pmd = pmd_off_k(0xffff0000);
1205 /* allocate the zero page. */
1206 zero_page = early_alloc(PAGE_SIZE);
1210 empty_zero_page = virt_to_page(zero_page);
1211 __flush_dcache_page(NULL, empty_zero_page);