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>
19 #include <asm/cputype.h>
20 #include <asm/sections.h>
21 #include <asm/cachetype.h>
22 #include <asm/setup.h>
23 #include <asm/sizes.h>
24 #include <asm/smp_plat.h>
26 #include <asm/highmem.h>
27 #include <asm/traps.h>
29 #include <asm/mach/arch.h>
30 #include <asm/mach/map.h>
35 * empty_zero_page is a special page that is used for
36 * zero-initialized data and COW.
38 struct page *empty_zero_page;
39 EXPORT_SYMBOL(empty_zero_page);
42 * The pmd table for the upper-most set of pages.
46 #define CPOLICY_UNCACHED 0
47 #define CPOLICY_BUFFERED 1
48 #define CPOLICY_WRITETHROUGH 2
49 #define CPOLICY_WRITEBACK 3
50 #define CPOLICY_WRITEALLOC 4
52 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
53 static unsigned int ecc_mask __initdata = 0;
55 pgprot_t pgprot_kernel;
57 EXPORT_SYMBOL(pgprot_user);
58 EXPORT_SYMBOL(pgprot_kernel);
61 const char policy[16];
67 static struct cachepolicy cache_policies[] __initdata = {
71 .pmd = PMD_SECT_UNCACHED,
72 .pte = L_PTE_MT_UNCACHED,
76 .pmd = PMD_SECT_BUFFERED,
77 .pte = L_PTE_MT_BUFFERABLE,
79 .policy = "writethrough",
82 .pte = L_PTE_MT_WRITETHROUGH,
84 .policy = "writeback",
87 .pte = L_PTE_MT_WRITEBACK,
89 .policy = "writealloc",
92 .pte = L_PTE_MT_WRITEALLOC,
97 * These are useful for identifying cache coherency
98 * problems by allowing the cache or the cache and
99 * writebuffer to be turned off. (Note: the write
100 * buffer should not be on and the cache off).
102 static int __init early_cachepolicy(char *p)
106 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
107 int len = strlen(cache_policies[i].policy);
109 if (memcmp(p, cache_policies[i].policy, len) == 0) {
111 cr_alignment &= ~cache_policies[i].cr_mask;
112 cr_no_alignment &= ~cache_policies[i].cr_mask;
116 if (i == ARRAY_SIZE(cache_policies))
117 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
119 * This restriction is partly to do with the way we boot; it is
120 * unpredictable to have memory mapped using two different sets of
121 * memory attributes (shared, type, and cache attribs). We can not
122 * change these attributes once the initial assembly has setup the
125 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
126 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
127 cachepolicy = CPOLICY_WRITEBACK;
130 set_cr(cr_alignment);
133 early_param("cachepolicy", early_cachepolicy);
135 static int __init early_nocache(char *__unused)
137 char *p = "buffered";
138 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
139 early_cachepolicy(p);
142 early_param("nocache", early_nocache);
144 static int __init early_nowrite(char *__unused)
146 char *p = "uncached";
147 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
148 early_cachepolicy(p);
151 early_param("nowb", early_nowrite);
153 #ifndef CONFIG_ARM_LPAE
154 static int __init early_ecc(char *p)
156 if (memcmp(p, "on", 2) == 0)
157 ecc_mask = PMD_PROTECTION;
158 else if (memcmp(p, "off", 3) == 0)
162 early_param("ecc", early_ecc);
165 static int __init noalign_setup(char *__unused)
167 cr_alignment &= ~CR_A;
168 cr_no_alignment &= ~CR_A;
169 set_cr(cr_alignment);
172 __setup("noalign", noalign_setup);
175 void adjust_cr(unsigned long mask, unsigned long set)
183 local_irq_save(flags);
185 cr_no_alignment = (cr_no_alignment & ~mask) | set;
186 cr_alignment = (cr_alignment & ~mask) | set;
188 set_cr((get_cr() & ~mask) | set);
190 local_irq_restore(flags);
194 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
195 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
197 static struct mem_type mem_types[] = {
198 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
199 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
201 .prot_l1 = PMD_TYPE_TABLE,
202 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
205 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
206 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
207 .prot_l1 = PMD_TYPE_TABLE,
208 .prot_sect = PROT_SECT_DEVICE,
211 [MT_DEVICE_CACHED] = { /* ioremap_cached */
212 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
213 .prot_l1 = PMD_TYPE_TABLE,
214 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
217 [MT_DEVICE_WC] = { /* ioremap_wc */
218 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
219 .prot_l1 = PMD_TYPE_TABLE,
220 .prot_sect = PROT_SECT_DEVICE,
224 .prot_pte = PROT_PTE_DEVICE,
225 .prot_l1 = PMD_TYPE_TABLE,
226 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
230 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
231 .domain = DOMAIN_KERNEL,
233 #ifndef CONFIG_ARM_LPAE
235 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
236 .domain = DOMAIN_KERNEL,
240 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
242 .prot_l1 = PMD_TYPE_TABLE,
243 .domain = DOMAIN_USER,
245 [MT_HIGH_VECTORS] = {
246 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
247 L_PTE_USER | L_PTE_RDONLY,
248 .prot_l1 = PMD_TYPE_TABLE,
249 .domain = DOMAIN_USER,
252 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
253 .prot_l1 = PMD_TYPE_TABLE,
254 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
255 .domain = DOMAIN_KERNEL,
258 .prot_sect = PMD_TYPE_SECT,
259 .domain = DOMAIN_KERNEL,
261 [MT_MEMORY_NONCACHED] = {
262 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
264 .prot_l1 = PMD_TYPE_TABLE,
265 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
266 .domain = DOMAIN_KERNEL,
269 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
271 .prot_l1 = PMD_TYPE_TABLE,
272 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
273 .domain = DOMAIN_KERNEL,
276 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
277 .prot_l1 = PMD_TYPE_TABLE,
278 .domain = DOMAIN_KERNEL,
281 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
283 .prot_l1 = PMD_TYPE_TABLE,
284 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
285 PMD_SECT_UNCACHED | PMD_SECT_XN,
286 .domain = DOMAIN_KERNEL,
288 [MT_MEMORY_DMA_READY] = {
289 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
290 .prot_l1 = PMD_TYPE_TABLE,
291 .domain = DOMAIN_KERNEL,
295 const struct mem_type *get_mem_type(unsigned int type)
297 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
299 EXPORT_SYMBOL(get_mem_type);
302 * If the system supports huge pages and we are running with short descriptors,
303 * then compute the pmd and linux pte prot values for a huge page.
305 * These values are used by both the HugeTLB and THP code.
307 #if defined(CONFIG_SYS_SUPPORTS_HUGETLBFS) && !defined(CONFIG_ARM_LPAE)
308 pmdval_t arm_hugepmdprotval;
309 EXPORT_SYMBOL(arm_hugepmdprotval);
311 pteval_t arm_hugepteprotval;
312 EXPORT_SYMBOL(arm_hugepteprotval);
317 * Adjust the PMD section entries according to the CPU in use.
319 static void __init build_mem_type_table(void)
321 struct cachepolicy *cp;
322 unsigned int cr = get_cr();
323 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
324 int cpu_arch = cpu_architecture();
327 if (cpu_arch < CPU_ARCH_ARMv6) {
328 #if defined(CONFIG_CPU_DCACHE_DISABLE)
329 if (cachepolicy > CPOLICY_BUFFERED)
330 cachepolicy = CPOLICY_BUFFERED;
331 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
332 if (cachepolicy > CPOLICY_WRITETHROUGH)
333 cachepolicy = CPOLICY_WRITETHROUGH;
336 if (cpu_arch < CPU_ARCH_ARMv5) {
337 if (cachepolicy >= CPOLICY_WRITEALLOC)
338 cachepolicy = CPOLICY_WRITEBACK;
342 cachepolicy = CPOLICY_WRITEALLOC;
345 * Strip out features not present on earlier architectures.
346 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
347 * without extended page tables don't have the 'Shared' bit.
349 if (cpu_arch < CPU_ARCH_ARMv5)
350 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
351 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
352 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
353 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
354 mem_types[i].prot_sect &= ~PMD_SECT_S;
357 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
358 * "update-able on write" bit on ARM610). However, Xscale and
359 * Xscale3 require this bit to be cleared.
361 if (cpu_is_xscale() || cpu_is_xsc3()) {
362 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
363 mem_types[i].prot_sect &= ~PMD_BIT4;
364 mem_types[i].prot_l1 &= ~PMD_BIT4;
366 } else if (cpu_arch < CPU_ARCH_ARMv6) {
367 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
368 if (mem_types[i].prot_l1)
369 mem_types[i].prot_l1 |= PMD_BIT4;
370 if (mem_types[i].prot_sect)
371 mem_types[i].prot_sect |= PMD_BIT4;
376 * Mark the device areas according to the CPU/architecture.
378 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
379 if (!cpu_is_xsc3()) {
381 * Mark device regions on ARMv6+ as execute-never
382 * to prevent speculative instruction fetches.
384 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
385 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
386 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
387 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
389 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
391 * For ARMv7 with TEX remapping,
392 * - shared device is SXCB=1100
393 * - nonshared device is SXCB=0100
394 * - write combine device mem is SXCB=0001
395 * (Uncached Normal memory)
397 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
398 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
399 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
400 } else if (cpu_is_xsc3()) {
403 * - shared device is TEXCB=00101
404 * - nonshared device is TEXCB=01000
405 * - write combine device mem is TEXCB=00100
406 * (Inner/Outer Uncacheable in xsc3 parlance)
408 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
409 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
410 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
413 * For ARMv6 and ARMv7 without TEX remapping,
414 * - shared device is TEXCB=00001
415 * - nonshared device is TEXCB=01000
416 * - write combine device mem is TEXCB=00100
417 * (Uncached Normal in ARMv6 parlance).
419 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
420 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
421 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
425 * On others, write combining is "Uncached/Buffered"
427 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
431 * Now deal with the memory-type mappings
433 cp = &cache_policies[cachepolicy];
434 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
437 * Only use write-through for non-SMP systems
439 if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
440 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
443 * Enable CPU-specific coherency if supported.
444 * (Only available on XSC3 at the moment.)
446 if (arch_is_coherent() && cpu_is_xsc3()) {
447 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
448 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
449 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
450 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
451 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
454 * ARMv6 and above have extended page tables.
456 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
457 #ifndef CONFIG_ARM_LPAE
459 * Mark cache clean areas and XIP ROM read only
460 * from SVC mode and no access from userspace.
462 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
463 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
464 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
469 * Mark memory with the "shared" attribute
472 user_pgprot |= L_PTE_SHARED;
473 kern_pgprot |= L_PTE_SHARED;
474 vecs_pgprot |= L_PTE_SHARED;
475 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
476 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
477 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
478 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
479 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
480 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
481 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
482 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
483 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
488 * Non-cacheable Normal - intended for memory areas that must
489 * not cause dirty cache line writebacks when used
491 if (cpu_arch >= CPU_ARCH_ARMv6) {
492 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
493 /* Non-cacheable Normal is XCB = 001 */
494 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
497 /* For both ARMv6 and non-TEX-remapping ARMv7 */
498 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
502 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
505 #ifdef CONFIG_ARM_LPAE
507 * Do not generate access flag faults for the kernel mappings.
509 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
510 mem_types[i].prot_pte |= PTE_EXT_AF;
511 mem_types[i].prot_sect |= PMD_SECT_AF;
513 kern_pgprot |= PTE_EXT_AF;
514 vecs_pgprot |= PTE_EXT_AF;
517 for (i = 0; i < 16; i++) {
518 pteval_t v = pgprot_val(protection_map[i]);
519 protection_map[i] = __pgprot(v | user_pgprot);
522 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
523 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
525 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
526 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
527 L_PTE_DIRTY | kern_pgprot);
529 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
530 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
531 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
532 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
533 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
534 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
535 mem_types[MT_ROM].prot_sect |= cp->pmd;
539 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
543 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
546 printk("Memory policy: ECC %sabled, Data cache %s\n",
547 ecc_mask ? "en" : "dis", cp->policy);
549 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
550 struct mem_type *t = &mem_types[i];
552 t->prot_l1 |= PMD_DOMAIN(t->domain);
554 t->prot_sect |= PMD_DOMAIN(t->domain);
557 #if defined(CONFIG_SYS_SUPPORTS_HUGETLBFS) && !defined(CONFIG_ARM_LPAE)
559 * we assume all huge pages are user pages and that hardware access
560 * flag updates are disabled (i.e. SCTLR.AFE == 0b).
562 arm_hugepteprotval = mem_types[MT_MEMORY].prot_pte | L_PTE_USER | L_PTE_VALID;
564 arm_hugepmdprotval = mem_types[MT_MEMORY].prot_sect | PMD_SECT_AP_READ
570 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
571 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
572 unsigned long size, pgprot_t vma_prot)
575 return pgprot_noncached(vma_prot);
576 else if (file->f_flags & O_SYNC)
577 return pgprot_writecombine(vma_prot);
580 EXPORT_SYMBOL(phys_mem_access_prot);
583 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
585 static void __init *early_alloc(unsigned long sz)
587 void *ptr = __va(memblock_alloc(sz, sz));
592 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
594 if (pmd_none(*pmd)) {
595 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
596 __pmd_populate(pmd, __pa(pte), prot);
598 BUG_ON(pmd_bad(*pmd));
599 return pte_offset_kernel(pmd, addr);
602 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
603 unsigned long end, unsigned long pfn,
604 const struct mem_type *type)
606 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
608 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
610 } while (pte++, addr += PAGE_SIZE, addr != end);
613 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
614 unsigned long end, phys_addr_t phys,
615 const struct mem_type *type)
617 pmd_t *pmd = pmd_offset(pud, addr);
620 * Try a section mapping - end, addr and phys must all be aligned
621 * to a section boundary. Note that PMDs refer to the individual
622 * L1 entries, whereas PGDs refer to a group of L1 entries making
623 * up one logical pointer to an L2 table.
625 if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
628 #ifndef CONFIG_ARM_LPAE
629 if (addr & SECTION_SIZE)
634 *pmd = __pmd(phys | type->prot_sect);
635 phys += SECTION_SIZE;
636 } while (pmd++, addr += SECTION_SIZE, addr != end);
641 * No need to loop; pte's aren't interested in the
642 * individual L1 entries.
644 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
648 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
649 unsigned long phys, const struct mem_type *type)
651 pud_t *pud = pud_offset(pgd, addr);
655 next = pud_addr_end(addr, end);
656 alloc_init_section(pud, addr, next, phys, type);
658 } while (pud++, addr = next, addr != end);
661 #ifndef CONFIG_ARM_LPAE
662 static void __init create_36bit_mapping(struct map_desc *md,
663 const struct mem_type *type)
665 unsigned long addr, length, end;
670 phys = __pfn_to_phys(md->pfn);
671 length = PAGE_ALIGN(md->length);
673 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
674 printk(KERN_ERR "MM: CPU does not support supersection "
675 "mapping for 0x%08llx at 0x%08lx\n",
676 (long long)__pfn_to_phys((u64)md->pfn), addr);
680 /* N.B. ARMv6 supersections are only defined to work with domain 0.
681 * Since domain assignments can in fact be arbitrary, the
682 * 'domain == 0' check below is required to insure that ARMv6
683 * supersections are only allocated for domain 0 regardless
684 * of the actual domain assignments in use.
687 printk(KERN_ERR "MM: invalid domain in supersection "
688 "mapping for 0x%08llx at 0x%08lx\n",
689 (long long)__pfn_to_phys((u64)md->pfn), addr);
693 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
694 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
695 " at 0x%08lx invalid alignment\n",
696 (long long)__pfn_to_phys((u64)md->pfn), addr);
701 * Shift bits [35:32] of address into bits [23:20] of PMD
704 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
706 pgd = pgd_offset_k(addr);
709 pud_t *pud = pud_offset(pgd, addr);
710 pmd_t *pmd = pmd_offset(pud, addr);
713 for (i = 0; i < 16; i++)
714 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
716 addr += SUPERSECTION_SIZE;
717 phys += SUPERSECTION_SIZE;
718 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
719 } while (addr != end);
721 #endif /* !CONFIG_ARM_LPAE */
724 * Create the page directory entries and any necessary
725 * page tables for the mapping specified by `md'. We
726 * are able to cope here with varying sizes and address
727 * offsets, and we take full advantage of sections and
730 static void __init create_mapping(struct map_desc *md)
732 unsigned long addr, length, end;
734 const struct mem_type *type;
737 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
738 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
739 " at 0x%08lx in user region\n",
740 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
744 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
745 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
746 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
747 " at 0x%08lx overlaps vmalloc space\n",
748 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
751 type = &mem_types[md->type];
753 #ifndef CONFIG_ARM_LPAE
755 * Catch 36-bit addresses
757 if (md->pfn >= 0x100000) {
758 create_36bit_mapping(md, type);
763 addr = md->virtual & PAGE_MASK;
764 phys = __pfn_to_phys(md->pfn);
765 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
767 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
768 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
769 "be mapped using pages, ignoring.\n",
770 (long long)__pfn_to_phys(md->pfn), addr);
774 pgd = pgd_offset_k(addr);
777 unsigned long next = pgd_addr_end(addr, end);
779 alloc_init_pud(pgd, addr, next, phys, type);
783 } while (pgd++, addr != end);
787 * Create the architecture specific mappings
789 void __init iotable_init(struct map_desc *io_desc, int nr)
793 for (i = 0; i < nr; i++)
794 create_mapping(io_desc + i);
797 static void * __initdata vmalloc_min = (void *)(VMALLOC_END - SZ_128M);
800 * vmalloc=size forces the vmalloc area to be exactly 'size'
801 * bytes. This can be used to increase (or decrease) the vmalloc
802 * area - the default is 128m.
804 static int __init early_vmalloc(char *arg)
806 unsigned long vmalloc_reserve = memparse(arg, NULL);
808 if (vmalloc_reserve < SZ_16M) {
809 vmalloc_reserve = SZ_16M;
811 "vmalloc area too small, limiting to %luMB\n",
812 vmalloc_reserve >> 20);
815 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
816 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
818 "vmalloc area is too big, limiting to %luMB\n",
819 vmalloc_reserve >> 20);
822 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
825 early_param("vmalloc", early_vmalloc);
827 phys_addr_t arm_lowmem_limit __initdata = 0;
829 void __init sanity_check_meminfo(void)
831 int i, j, highmem = 0;
833 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
834 struct membank *bank = &meminfo.bank[j];
835 *bank = meminfo.bank[i];
837 if (bank->start > ULONG_MAX)
840 #ifdef CONFIG_HIGHMEM
841 if (__va(bank->start) >= vmalloc_min ||
842 __va(bank->start) < (void *)PAGE_OFFSET)
845 bank->highmem = highmem;
848 * Split those memory banks which are partially overlapping
849 * the vmalloc area greatly simplifying things later.
851 if (!highmem && __va(bank->start) < vmalloc_min &&
852 bank->size > vmalloc_min - __va(bank->start)) {
853 if (meminfo.nr_banks >= NR_BANKS) {
854 printk(KERN_CRIT "NR_BANKS too low, "
855 "ignoring high memory\n");
857 memmove(bank + 1, bank,
858 (meminfo.nr_banks - i) * sizeof(*bank));
861 bank[1].size -= vmalloc_min - __va(bank->start);
862 bank[1].start = __pa(vmalloc_min - 1) + 1;
863 bank[1].highmem = highmem = 1;
866 bank->size = vmalloc_min - __va(bank->start);
869 bank->highmem = highmem;
872 * Highmem banks not allowed with !CONFIG_HIGHMEM.
875 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
876 "(!CONFIG_HIGHMEM).\n",
877 (unsigned long long)bank->start,
878 (unsigned long long)bank->start + bank->size - 1);
883 * Check whether this memory bank would entirely overlap
886 if (__va(bank->start) >= vmalloc_min ||
887 __va(bank->start) < (void *)PAGE_OFFSET) {
888 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
889 "(vmalloc region overlap).\n",
890 (unsigned long long)bank->start,
891 (unsigned long long)bank->start + bank->size - 1);
896 * Check whether this memory bank would partially overlap
899 if (__va(bank->start + bank->size) > vmalloc_min ||
900 __va(bank->start + bank->size) < __va(bank->start)) {
901 unsigned long newsize = vmalloc_min - __va(bank->start);
902 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
903 "to -%.8llx (vmalloc region overlap).\n",
904 (unsigned long long)bank->start,
905 (unsigned long long)bank->start + bank->size - 1,
906 (unsigned long long)bank->start + newsize - 1);
907 bank->size = newsize;
910 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
911 arm_lowmem_limit = bank->start + bank->size;
915 #ifdef CONFIG_HIGHMEM
917 const char *reason = NULL;
919 if (cache_is_vipt_aliasing()) {
921 * Interactions between kmap and other mappings
922 * make highmem support with aliasing VIPT caches
925 reason = "with VIPT aliasing cache";
928 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
930 while (j > 0 && meminfo.bank[j - 1].highmem)
935 meminfo.nr_banks = j;
936 high_memory = __va(arm_lowmem_limit - 1) + 1;
937 memblock_set_current_limit(arm_lowmem_limit);
940 static inline void prepare_page_table(void)
946 * Clear out all the mappings below the kernel image.
948 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
949 pmd_clear(pmd_off_k(addr));
951 #ifdef CONFIG_XIP_KERNEL
952 /* The XIP kernel is mapped in the module area -- skip over it */
953 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
955 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
956 pmd_clear(pmd_off_k(addr));
959 * Find the end of the first block of lowmem.
961 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
962 if (end >= arm_lowmem_limit)
963 end = arm_lowmem_limit;
966 * Clear out all the kernel space mappings, except for the first
967 * memory bank, up to the end of the vmalloc region.
969 for (addr = __phys_to_virt(end);
970 addr < VMALLOC_END; addr += PMD_SIZE)
971 pmd_clear(pmd_off_k(addr));
974 #ifdef CONFIG_ARM_LPAE
975 /* the first page is reserved for pgd */
976 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
977 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
979 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
983 * Reserve the special regions of memory
985 void __init arm_mm_memblock_reserve(void)
988 * Reserve the page tables. These are already in use,
989 * and can only be in node 0.
991 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
995 * Because of the SA1111 DMA bug, we want to preserve our
996 * precious DMA-able memory...
998 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1003 * Set up device the mappings. Since we clear out the page tables for all
1004 * mappings above VMALLOC_END, we will remove any debug device mappings.
1005 * This means you have to be careful how you debug this function, or any
1006 * called function. This means you can't use any function or debugging
1007 * method which may touch any device, otherwise the kernel _will_ crash.
1009 static void __init devicemaps_init(struct machine_desc *mdesc)
1011 struct map_desc map;
1015 * Allocate the vector page early.
1017 vectors_page = early_alloc(PAGE_SIZE);
1019 for (addr = VMALLOC_END; addr; addr += PMD_SIZE)
1020 pmd_clear(pmd_off_k(addr));
1023 * Map the kernel if it is XIP.
1024 * It is always first in the modulearea.
1026 #ifdef CONFIG_XIP_KERNEL
1027 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1028 map.virtual = MODULES_VADDR;
1029 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1031 create_mapping(&map);
1035 * Map the cache flushing regions.
1038 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1039 map.virtual = FLUSH_BASE;
1041 map.type = MT_CACHECLEAN;
1042 create_mapping(&map);
1044 #ifdef FLUSH_BASE_MINICACHE
1045 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1046 map.virtual = FLUSH_BASE_MINICACHE;
1048 map.type = MT_MINICLEAN;
1049 create_mapping(&map);
1053 * Create a mapping for the machine vectors at the high-vectors
1054 * location (0xffff0000). If we aren't using high-vectors, also
1055 * create a mapping at the low-vectors virtual address.
1057 map.pfn = __phys_to_pfn(virt_to_phys(vectors_page));
1058 map.virtual = 0xffff0000;
1059 map.length = PAGE_SIZE;
1060 map.type = MT_HIGH_VECTORS;
1061 create_mapping(&map);
1063 if (!vectors_high()) {
1065 map.type = MT_LOW_VECTORS;
1066 create_mapping(&map);
1070 * Ask the machine support to map in the statically mapped devices.
1076 * Finally flush the caches and tlb to ensure that we're in a
1077 * consistent state wrt the writebuffer. This also ensures that
1078 * any write-allocated cache lines in the vector page are written
1079 * back. After this point, we can start to touch devices again.
1081 local_flush_tlb_all();
1085 static void __init kmap_init(void)
1087 #ifdef CONFIG_HIGHMEM
1088 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1089 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1093 static void __init map_lowmem(void)
1095 struct memblock_region *reg;
1097 /* Map all the lowmem memory banks. */
1098 for_each_memblock(memory, reg) {
1099 phys_addr_t start = reg->base;
1100 phys_addr_t end = start + reg->size;
1101 struct map_desc map;
1103 if (end > arm_lowmem_limit)
1104 end = arm_lowmem_limit;
1108 map.pfn = __phys_to_pfn(start);
1109 map.virtual = __phys_to_virt(start);
1110 map.length = end - start;
1111 map.type = MT_MEMORY;
1113 create_mapping(&map);
1118 * paging_init() sets up the page tables, initialises the zone memory
1119 * maps, and sets up the zero page, bad page and bad page tables.
1121 void __init paging_init(struct machine_desc *mdesc)
1125 memblock_set_current_limit(arm_lowmem_limit);
1127 build_mem_type_table();
1128 prepare_page_table();
1130 dma_contiguous_remap();
1131 devicemaps_init(mdesc);
1134 top_pmd = pmd_off_k(0xffff0000);
1136 /* allocate the zero page. */
1137 zero_page = early_alloc(PAGE_SIZE);
1141 empty_zero_page = virt_to_page(zero_page);
1142 __flush_dcache_page(NULL, empty_zero_page);