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>
21 #include <asm/cputype.h>
22 #include <asm/sections.h>
23 #include <asm/cachetype.h>
24 #include <asm/setup.h>
25 #include <asm/sizes.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 * Only use write-through for non-SMP systems
427 if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
428 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
431 * Enable CPU-specific coherency if supported.
432 * (Only available on XSC3 at the moment.)
434 if (arch_is_coherent() && cpu_is_xsc3()) {
435 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
436 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
437 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
438 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
439 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
442 * ARMv6 and above have extended page tables.
444 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
445 #ifndef CONFIG_ARM_LPAE
447 * Mark cache clean areas and XIP ROM read only
448 * from SVC mode and no access from userspace.
450 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
451 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
452 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
457 * Mark memory with the "shared" attribute
460 user_pgprot |= L_PTE_SHARED;
461 kern_pgprot |= L_PTE_SHARED;
462 vecs_pgprot |= L_PTE_SHARED;
463 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
464 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
465 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
466 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
467 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
468 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
469 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
470 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
471 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
476 * Non-cacheable Normal - intended for memory areas that must
477 * not cause dirty cache line writebacks when used
479 if (cpu_arch >= CPU_ARCH_ARMv6) {
480 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
481 /* Non-cacheable Normal is XCB = 001 */
482 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
485 /* For both ARMv6 and non-TEX-remapping ARMv7 */
486 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
490 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
493 #ifdef CONFIG_ARM_LPAE
495 * Do not generate access flag faults for the kernel mappings.
497 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
498 mem_types[i].prot_pte |= PTE_EXT_AF;
499 if (mem_types[i].prot_sect)
500 mem_types[i].prot_sect |= PMD_SECT_AF;
502 kern_pgprot |= PTE_EXT_AF;
503 vecs_pgprot |= PTE_EXT_AF;
506 for (i = 0; i < 16; i++) {
507 unsigned long v = pgprot_val(protection_map[i]);
508 protection_map[i] = __pgprot(v | user_pgprot);
511 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
512 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
514 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
515 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
516 L_PTE_DIRTY | kern_pgprot);
518 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
519 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
520 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
521 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
522 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
523 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
524 mem_types[MT_ROM].prot_sect |= cp->pmd;
528 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
532 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
535 printk("Memory policy: ECC %sabled, Data cache %s\n",
536 ecc_mask ? "en" : "dis", cp->policy);
538 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
539 struct mem_type *t = &mem_types[i];
541 t->prot_l1 |= PMD_DOMAIN(t->domain);
543 t->prot_sect |= PMD_DOMAIN(t->domain);
547 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
548 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
549 unsigned long size, pgprot_t vma_prot)
552 return pgprot_noncached(vma_prot);
553 else if (file->f_flags & O_SYNC)
554 return pgprot_writecombine(vma_prot);
557 EXPORT_SYMBOL(phys_mem_access_prot);
560 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
562 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
564 void *ptr = __va(memblock_alloc(sz, align));
569 static void __init *early_alloc(unsigned long sz)
571 return early_alloc_aligned(sz, sz);
574 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
576 if (pmd_none(*pmd)) {
577 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
578 __pmd_populate(pmd, __pa(pte), prot);
580 BUG_ON(pmd_bad(*pmd));
581 return pte_offset_kernel(pmd, addr);
584 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
585 unsigned long end, unsigned long pfn,
586 const struct mem_type *type)
588 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
590 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
592 } while (pte++, addr += PAGE_SIZE, addr != end);
595 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
596 unsigned long end, phys_addr_t phys,
597 const struct mem_type *type)
599 pmd_t *pmd = pmd_offset(pud, addr);
602 * Try a section mapping - end, addr and phys must all be aligned
603 * to a section boundary. Note that PMDs refer to the individual
604 * L1 entries, whereas PGDs refer to a group of L1 entries making
605 * up one logical pointer to an L2 table.
607 if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
610 #ifndef CONFIG_ARM_LPAE
611 if (addr & SECTION_SIZE)
616 *pmd = __pmd(phys | type->prot_sect);
617 phys += SECTION_SIZE;
618 } while (pmd++, addr += SECTION_SIZE, addr != end);
623 * No need to loop; pte's aren't interested in the
624 * individual L1 entries.
626 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
630 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
631 unsigned long end, unsigned long phys, const struct mem_type *type)
633 pud_t *pud = pud_offset(pgd, addr);
637 next = pud_addr_end(addr, end);
638 alloc_init_section(pud, addr, next, phys, type);
640 } while (pud++, addr = next, addr != end);
643 #ifndef CONFIG_ARM_LPAE
644 static void __init create_36bit_mapping(struct map_desc *md,
645 const struct mem_type *type)
647 unsigned long addr, length, end;
652 phys = __pfn_to_phys(md->pfn);
653 length = PAGE_ALIGN(md->length);
655 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
656 printk(KERN_ERR "MM: CPU does not support supersection "
657 "mapping for 0x%08llx at 0x%08lx\n",
658 (long long)__pfn_to_phys((u64)md->pfn), addr);
662 /* N.B. ARMv6 supersections are only defined to work with domain 0.
663 * Since domain assignments can in fact be arbitrary, the
664 * 'domain == 0' check below is required to insure that ARMv6
665 * supersections are only allocated for domain 0 regardless
666 * of the actual domain assignments in use.
669 printk(KERN_ERR "MM: invalid domain in supersection "
670 "mapping for 0x%08llx at 0x%08lx\n",
671 (long long)__pfn_to_phys((u64)md->pfn), addr);
675 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
676 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
677 " at 0x%08lx invalid alignment\n",
678 (long long)__pfn_to_phys((u64)md->pfn), addr);
683 * Shift bits [35:32] of address into bits [23:20] of PMD
686 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
688 pgd = pgd_offset_k(addr);
691 pud_t *pud = pud_offset(pgd, addr);
692 pmd_t *pmd = pmd_offset(pud, addr);
695 for (i = 0; i < 16; i++)
696 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
698 addr += SUPERSECTION_SIZE;
699 phys += SUPERSECTION_SIZE;
700 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
701 } while (addr != end);
703 #endif /* !CONFIG_ARM_LPAE */
706 * Create the page directory entries and any necessary
707 * page tables for the mapping specified by `md'. We
708 * are able to cope here with varying sizes and address
709 * offsets, and we take full advantage of sections and
712 static void __init create_mapping(struct map_desc *md)
714 unsigned long addr, length, end;
716 const struct mem_type *type;
719 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
720 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
721 " at 0x%08lx in user region\n",
722 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
726 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
727 md->virtual >= PAGE_OFFSET &&
728 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
729 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
730 " at 0x%08lx out of vmalloc space\n",
731 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
734 type = &mem_types[md->type];
736 #ifndef CONFIG_ARM_LPAE
738 * Catch 36-bit addresses
740 if (md->pfn >= 0x100000) {
741 create_36bit_mapping(md, type);
746 addr = md->virtual & PAGE_MASK;
747 phys = __pfn_to_phys(md->pfn);
748 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
750 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
751 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
752 "be mapped using pages, ignoring.\n",
753 (long long)__pfn_to_phys(md->pfn), addr);
757 pgd = pgd_offset_k(addr);
760 unsigned long next = pgd_addr_end(addr, end);
762 alloc_init_pud(pgd, addr, next, phys, type);
766 } while (pgd++, addr != end);
770 * Create the architecture specific mappings
772 void __init iotable_init(struct map_desc *io_desc, int nr)
775 struct vm_struct *vm;
780 vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
782 for (md = io_desc; nr; md++, nr--) {
784 vm->addr = (void *)(md->virtual & PAGE_MASK);
785 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
786 vm->phys_addr = __pfn_to_phys(md->pfn);
787 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
788 vm->flags |= VM_ARM_MTYPE(md->type);
789 vm->caller = iotable_init;
790 vm_area_add_early(vm++);
794 #ifndef CONFIG_ARM_LPAE
797 * The Linux PMD is made of two consecutive section entries covering 2MB
798 * (see definition in include/asm/pgtable-2level.h). However a call to
799 * create_mapping() may optimize static mappings by using individual
800 * 1MB section mappings. This leaves the actual PMD potentially half
801 * initialized if the top or bottom section entry isn't used, leaving it
802 * open to problems if a subsequent ioremap() or vmalloc() tries to use
803 * the virtual space left free by that unused section entry.
805 * Let's avoid the issue by inserting dummy vm entries covering the unused
806 * PMD halves once the static mappings are in place.
809 static void __init pmd_empty_section_gap(unsigned long addr)
811 struct vm_struct *vm;
813 vm = early_alloc_aligned(sizeof(*vm), __alignof__(*vm));
814 vm->addr = (void *)addr;
815 vm->size = SECTION_SIZE;
816 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
817 vm->caller = pmd_empty_section_gap;
818 vm_area_add_early(vm);
821 static void __init fill_pmd_gaps(void)
823 struct vm_struct *vm;
824 unsigned long addr, next = 0;
827 /* we're still single threaded hence no lock needed here */
828 for (vm = vmlist; vm; vm = vm->next) {
829 if (!(vm->flags & VM_ARM_STATIC_MAPPING))
831 addr = (unsigned long)vm->addr;
836 * Check if this vm starts on an odd section boundary.
837 * If so and the first section entry for this PMD is free
838 * then we block the corresponding virtual address.
840 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
841 pmd = pmd_off_k(addr);
843 pmd_empty_section_gap(addr & PMD_MASK);
847 * Then check if this vm ends on an odd section boundary.
848 * If so and the second section entry for this PMD is empty
849 * then we block the corresponding virtual address.
852 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
853 pmd = pmd_off_k(addr) + 1;
855 pmd_empty_section_gap(addr);
858 /* no need to look at any vm entry until we hit the next PMD */
859 next = (addr + PMD_SIZE - 1) & PMD_MASK;
864 #define fill_pmd_gaps() do { } while (0)
867 static void * __initdata vmalloc_min =
868 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
871 * vmalloc=size forces the vmalloc area to be exactly 'size'
872 * bytes. This can be used to increase (or decrease) the vmalloc
873 * area - the default is 240m.
875 static int __init early_vmalloc(char *arg)
877 unsigned long vmalloc_reserve = memparse(arg, NULL);
879 if (vmalloc_reserve < SZ_16M) {
880 vmalloc_reserve = SZ_16M;
882 "vmalloc area too small, limiting to %luMB\n",
883 vmalloc_reserve >> 20);
886 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
887 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
889 "vmalloc area is too big, limiting to %luMB\n",
890 vmalloc_reserve >> 20);
893 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
896 early_param("vmalloc", early_vmalloc);
898 phys_addr_t arm_lowmem_limit __initdata = 0;
900 void __init sanity_check_meminfo(void)
902 int i, j, highmem = 0;
904 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
905 struct membank *bank = &meminfo.bank[j];
906 *bank = meminfo.bank[i];
908 if (bank->start > ULONG_MAX)
911 #ifdef CONFIG_HIGHMEM
912 if (__va(bank->start) >= vmalloc_min ||
913 __va(bank->start) < (void *)PAGE_OFFSET)
916 bank->highmem = highmem;
919 * Split those memory banks which are partially overlapping
920 * the vmalloc area greatly simplifying things later.
922 if (!highmem && __va(bank->start) < vmalloc_min &&
923 bank->size > vmalloc_min - __va(bank->start)) {
924 if (meminfo.nr_banks >= NR_BANKS) {
925 printk(KERN_CRIT "NR_BANKS too low, "
926 "ignoring high memory\n");
928 memmove(bank + 1, bank,
929 (meminfo.nr_banks - i) * sizeof(*bank));
932 bank[1].size -= vmalloc_min - __va(bank->start);
933 bank[1].start = __pa(vmalloc_min - 1) + 1;
934 bank[1].highmem = highmem = 1;
937 bank->size = vmalloc_min - __va(bank->start);
940 bank->highmem = highmem;
943 * Highmem banks not allowed with !CONFIG_HIGHMEM.
946 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
947 "(!CONFIG_HIGHMEM).\n",
948 (unsigned long long)bank->start,
949 (unsigned long long)bank->start + bank->size - 1);
954 * Check whether this memory bank would entirely overlap
957 if (__va(bank->start) >= vmalloc_min ||
958 __va(bank->start) < (void *)PAGE_OFFSET) {
959 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
960 "(vmalloc region overlap).\n",
961 (unsigned long long)bank->start,
962 (unsigned long long)bank->start + bank->size - 1);
967 * Check whether this memory bank would partially overlap
970 if (__va(bank->start + bank->size) > vmalloc_min ||
971 __va(bank->start + bank->size) < __va(bank->start)) {
972 unsigned long newsize = vmalloc_min - __va(bank->start);
973 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
974 "to -%.8llx (vmalloc region overlap).\n",
975 (unsigned long long)bank->start,
976 (unsigned long long)bank->start + bank->size - 1,
977 (unsigned long long)bank->start + newsize - 1);
978 bank->size = newsize;
981 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
982 arm_lowmem_limit = bank->start + bank->size;
986 #ifdef CONFIG_HIGHMEM
988 const char *reason = NULL;
990 if (cache_is_vipt_aliasing()) {
992 * Interactions between kmap and other mappings
993 * make highmem support with aliasing VIPT caches
996 reason = "with VIPT aliasing cache";
999 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
1001 while (j > 0 && meminfo.bank[j - 1].highmem)
1006 meminfo.nr_banks = j;
1007 high_memory = __va(arm_lowmem_limit - 1) + 1;
1008 memblock_set_current_limit(arm_lowmem_limit);
1011 static inline void prepare_page_table(void)
1017 * Clear out all the mappings below the kernel image.
1019 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1020 pmd_clear(pmd_off_k(addr));
1022 #ifdef CONFIG_XIP_KERNEL
1023 /* The XIP kernel is mapped in the module area -- skip over it */
1024 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1026 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1027 pmd_clear(pmd_off_k(addr));
1030 * Find the end of the first block of lowmem.
1032 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1033 if (end >= arm_lowmem_limit)
1034 end = arm_lowmem_limit;
1037 * Clear out all the kernel space mappings, except for the first
1038 * memory bank, up to the vmalloc region.
1040 for (addr = __phys_to_virt(end);
1041 addr < VMALLOC_START; addr += PMD_SIZE)
1042 pmd_clear(pmd_off_k(addr));
1045 #ifdef CONFIG_ARM_LPAE
1046 /* the first page is reserved for pgd */
1047 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1048 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1050 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1054 * Reserve the special regions of memory
1056 void __init arm_mm_memblock_reserve(void)
1059 * Reserve the page tables. These are already in use,
1060 * and can only be in node 0.
1062 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1064 #ifdef CONFIG_SA1111
1066 * Because of the SA1111 DMA bug, we want to preserve our
1067 * precious DMA-able memory...
1069 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1074 * Set up the device mappings. Since we clear out the page tables for all
1075 * mappings above VMALLOC_START, we will remove any debug device mappings.
1076 * This means you have to be careful how you debug this function, or any
1077 * called function. This means you can't use any function or debugging
1078 * method which may touch any device, otherwise the kernel _will_ crash.
1080 static void __init devicemaps_init(struct machine_desc *mdesc)
1082 struct map_desc map;
1087 * Allocate the vector page early.
1089 vectors = early_alloc(PAGE_SIZE);
1091 early_trap_init(vectors);
1093 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1094 pmd_clear(pmd_off_k(addr));
1097 * Map the kernel if it is XIP.
1098 * It is always first in the modulearea.
1100 #ifdef CONFIG_XIP_KERNEL
1101 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1102 map.virtual = MODULES_VADDR;
1103 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1105 create_mapping(&map);
1109 * Map the cache flushing regions.
1112 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1113 map.virtual = FLUSH_BASE;
1115 map.type = MT_CACHECLEAN;
1116 create_mapping(&map);
1118 #ifdef FLUSH_BASE_MINICACHE
1119 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1120 map.virtual = FLUSH_BASE_MINICACHE;
1122 map.type = MT_MINICLEAN;
1123 create_mapping(&map);
1127 * Create a mapping for the machine vectors at the high-vectors
1128 * location (0xffff0000). If we aren't using high-vectors, also
1129 * create a mapping at the low-vectors virtual address.
1131 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1132 map.virtual = 0xffff0000;
1133 map.length = PAGE_SIZE;
1134 map.type = MT_HIGH_VECTORS;
1135 create_mapping(&map);
1137 if (!vectors_high()) {
1139 map.type = MT_LOW_VECTORS;
1140 create_mapping(&map);
1144 * Ask the machine support to map in the statically mapped devices.
1151 * Finally flush the caches and tlb to ensure that we're in a
1152 * consistent state wrt the writebuffer. This also ensures that
1153 * any write-allocated cache lines in the vector page are written
1154 * back. After this point, we can start to touch devices again.
1156 local_flush_tlb_all();
1160 static void __init kmap_init(void)
1162 #ifdef CONFIG_HIGHMEM
1163 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1164 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1168 static void __init map_lowmem(void)
1170 struct memblock_region *reg;
1172 /* Map all the lowmem memory banks. */
1173 for_each_memblock(memory, reg) {
1174 phys_addr_t start = reg->base;
1175 phys_addr_t end = start + reg->size;
1176 struct map_desc map;
1178 if (end > arm_lowmem_limit)
1179 end = arm_lowmem_limit;
1183 map.pfn = __phys_to_pfn(start);
1184 map.virtual = __phys_to_virt(start);
1185 map.length = end - start;
1186 map.type = MT_MEMORY;
1188 create_mapping(&map);
1193 * paging_init() sets up the page tables, initialises the zone memory
1194 * maps, and sets up the zero page, bad page and bad page tables.
1196 void __init paging_init(struct machine_desc *mdesc)
1200 memblock_set_current_limit(arm_lowmem_limit);
1202 build_mem_type_table();
1203 prepare_page_table();
1205 dma_contiguous_remap();
1206 devicemaps_init(mdesc);
1209 top_pmd = pmd_off_k(0xffff0000);
1211 /* allocate the zero page. */
1212 zero_page = early_alloc(PAGE_SIZE);
1216 empty_zero_page = virt_to_page(zero_page);
1217 __flush_dcache_page(NULL, empty_zero_page);