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>
34 #include <asm/mach/pci.h>
39 * empty_zero_page is a special page that is used for
40 * zero-initialized data and COW.
42 struct page *empty_zero_page;
43 EXPORT_SYMBOL(empty_zero_page);
46 * The pmd table for the upper-most set of pages.
50 #define CPOLICY_UNCACHED 0
51 #define CPOLICY_BUFFERED 1
52 #define CPOLICY_WRITETHROUGH 2
53 #define CPOLICY_WRITEBACK 3
54 #define CPOLICY_WRITEALLOC 4
56 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
57 static unsigned int ecc_mask __initdata = 0;
59 pgprot_t pgprot_kernel;
61 EXPORT_SYMBOL(pgprot_user);
62 EXPORT_SYMBOL(pgprot_kernel);
65 const char policy[16];
71 static struct cachepolicy cache_policies[] __initdata = {
75 .pmd = PMD_SECT_UNCACHED,
76 .pte = L_PTE_MT_UNCACHED,
80 .pmd = PMD_SECT_BUFFERED,
81 .pte = L_PTE_MT_BUFFERABLE,
83 .policy = "writethrough",
86 .pte = L_PTE_MT_WRITETHROUGH,
88 .policy = "writeback",
91 .pte = L_PTE_MT_WRITEBACK,
93 .policy = "writealloc",
96 .pte = L_PTE_MT_WRITEALLOC,
101 * These are useful for identifying cache coherency
102 * problems by allowing the cache or the cache and
103 * writebuffer to be turned off. (Note: the write
104 * buffer should not be on and the cache off).
106 static int __init early_cachepolicy(char *p)
110 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
111 int len = strlen(cache_policies[i].policy);
113 if (memcmp(p, cache_policies[i].policy, len) == 0) {
115 cr_alignment &= ~cache_policies[i].cr_mask;
116 cr_no_alignment &= ~cache_policies[i].cr_mask;
120 if (i == ARRAY_SIZE(cache_policies))
121 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
123 * This restriction is partly to do with the way we boot; it is
124 * unpredictable to have memory mapped using two different sets of
125 * memory attributes (shared, type, and cache attribs). We can not
126 * change these attributes once the initial assembly has setup the
129 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
130 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
131 cachepolicy = CPOLICY_WRITEBACK;
134 set_cr(cr_alignment);
137 early_param("cachepolicy", early_cachepolicy);
139 static int __init early_nocache(char *__unused)
141 char *p = "buffered";
142 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
143 early_cachepolicy(p);
146 early_param("nocache", early_nocache);
148 static int __init early_nowrite(char *__unused)
150 char *p = "uncached";
151 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
152 early_cachepolicy(p);
155 early_param("nowb", early_nowrite);
157 #ifndef CONFIG_ARM_LPAE
158 static int __init early_ecc(char *p)
160 if (memcmp(p, "on", 2) == 0)
161 ecc_mask = PMD_PROTECTION;
162 else if (memcmp(p, "off", 3) == 0)
166 early_param("ecc", early_ecc);
169 static int __init noalign_setup(char *__unused)
171 cr_alignment &= ~CR_A;
172 cr_no_alignment &= ~CR_A;
173 set_cr(cr_alignment);
176 __setup("noalign", noalign_setup);
179 void adjust_cr(unsigned long mask, unsigned long set)
187 local_irq_save(flags);
189 cr_no_alignment = (cr_no_alignment & ~mask) | set;
190 cr_alignment = (cr_alignment & ~mask) | set;
192 set_cr((get_cr() & ~mask) | set);
194 local_irq_restore(flags);
198 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
199 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
201 static struct mem_type mem_types[] = {
202 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
203 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
205 .prot_l1 = PMD_TYPE_TABLE,
206 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
209 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
210 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
211 .prot_l1 = PMD_TYPE_TABLE,
212 .prot_sect = PROT_SECT_DEVICE,
215 [MT_DEVICE_CACHED] = { /* ioremap_cached */
216 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
217 .prot_l1 = PMD_TYPE_TABLE,
218 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
221 [MT_DEVICE_WC] = { /* ioremap_wc */
222 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
223 .prot_l1 = PMD_TYPE_TABLE,
224 .prot_sect = PROT_SECT_DEVICE,
228 .prot_pte = PROT_PTE_DEVICE,
229 .prot_l1 = PMD_TYPE_TABLE,
230 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
234 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
235 .domain = DOMAIN_KERNEL,
237 #ifndef CONFIG_ARM_LPAE
239 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
240 .domain = DOMAIN_KERNEL,
244 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
246 .prot_l1 = PMD_TYPE_TABLE,
247 .domain = DOMAIN_USER,
249 [MT_HIGH_VECTORS] = {
250 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
251 L_PTE_USER | L_PTE_RDONLY,
252 .prot_l1 = PMD_TYPE_TABLE,
253 .domain = DOMAIN_USER,
256 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
257 .prot_l1 = PMD_TYPE_TABLE,
258 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
259 .domain = DOMAIN_KERNEL,
262 .prot_sect = PMD_TYPE_SECT,
263 .domain = DOMAIN_KERNEL,
265 [MT_MEMORY_NONCACHED] = {
266 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
268 .prot_l1 = PMD_TYPE_TABLE,
269 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
270 .domain = DOMAIN_KERNEL,
273 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
275 .prot_l1 = PMD_TYPE_TABLE,
276 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
277 .domain = DOMAIN_KERNEL,
280 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
281 .prot_l1 = PMD_TYPE_TABLE,
282 .domain = DOMAIN_KERNEL,
285 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
287 .prot_l1 = PMD_TYPE_TABLE,
288 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
289 PMD_SECT_UNCACHED | PMD_SECT_XN,
290 .domain = DOMAIN_KERNEL,
292 [MT_MEMORY_DMA_READY] = {
293 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
294 .prot_l1 = PMD_TYPE_TABLE,
295 .domain = DOMAIN_KERNEL,
299 const struct mem_type *get_mem_type(unsigned int type)
301 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
303 EXPORT_SYMBOL(get_mem_type);
306 * Adjust the PMD section entries according to the CPU in use.
308 static void __init build_mem_type_table(void)
310 struct cachepolicy *cp;
311 unsigned int cr = get_cr();
312 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
313 int cpu_arch = cpu_architecture();
316 if (cpu_arch < CPU_ARCH_ARMv6) {
317 #if defined(CONFIG_CPU_DCACHE_DISABLE)
318 if (cachepolicy > CPOLICY_BUFFERED)
319 cachepolicy = CPOLICY_BUFFERED;
320 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
321 if (cachepolicy > CPOLICY_WRITETHROUGH)
322 cachepolicy = CPOLICY_WRITETHROUGH;
325 if (cpu_arch < CPU_ARCH_ARMv5) {
326 if (cachepolicy >= CPOLICY_WRITEALLOC)
327 cachepolicy = CPOLICY_WRITEBACK;
331 cachepolicy = CPOLICY_WRITEALLOC;
334 * Strip out features not present on earlier architectures.
335 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
336 * without extended page tables don't have the 'Shared' bit.
338 if (cpu_arch < CPU_ARCH_ARMv5)
339 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
340 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
341 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
342 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
343 mem_types[i].prot_sect &= ~PMD_SECT_S;
346 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
347 * "update-able on write" bit on ARM610). However, Xscale and
348 * Xscale3 require this bit to be cleared.
350 if (cpu_is_xscale() || cpu_is_xsc3()) {
351 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
352 mem_types[i].prot_sect &= ~PMD_BIT4;
353 mem_types[i].prot_l1 &= ~PMD_BIT4;
355 } else if (cpu_arch < CPU_ARCH_ARMv6) {
356 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
357 if (mem_types[i].prot_l1)
358 mem_types[i].prot_l1 |= PMD_BIT4;
359 if (mem_types[i].prot_sect)
360 mem_types[i].prot_sect |= PMD_BIT4;
365 * Mark the device areas according to the CPU/architecture.
367 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
368 if (!cpu_is_xsc3()) {
370 * Mark device regions on ARMv6+ as execute-never
371 * to prevent speculative instruction fetches.
373 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
374 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
375 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
376 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
378 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
380 * For ARMv7 with TEX remapping,
381 * - shared device is SXCB=1100
382 * - nonshared device is SXCB=0100
383 * - write combine device mem is SXCB=0001
384 * (Uncached Normal memory)
386 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
387 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
388 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
389 } else if (cpu_is_xsc3()) {
392 * - shared device is TEXCB=00101
393 * - nonshared device is TEXCB=01000
394 * - write combine device mem is TEXCB=00100
395 * (Inner/Outer Uncacheable in xsc3 parlance)
397 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
398 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
399 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
402 * For ARMv6 and ARMv7 without TEX remapping,
403 * - shared device is TEXCB=00001
404 * - nonshared device is TEXCB=01000
405 * - write combine device mem is TEXCB=00100
406 * (Uncached Normal in ARMv6 parlance).
408 mem_types[MT_DEVICE].prot_sect |= 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);
414 * On others, write combining is "Uncached/Buffered"
416 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
420 * Now deal with the memory-type mappings
422 cp = &cache_policies[cachepolicy];
423 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
426 * Only use write-through for non-SMP systems
428 if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
429 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].pte;
432 * Enable CPU-specific coherency if supported.
433 * (Only available on XSC3 at the moment.)
435 if (arch_is_coherent() && cpu_is_xsc3()) {
436 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
437 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
438 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
439 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
440 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
443 * ARMv6 and above have extended page tables.
445 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
446 #ifndef CONFIG_ARM_LPAE
448 * Mark cache clean areas and XIP ROM read only
449 * from SVC mode and no access from userspace.
451 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
452 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
453 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
458 * Mark memory with the "shared" attribute
461 user_pgprot |= L_PTE_SHARED;
462 kern_pgprot |= L_PTE_SHARED;
463 vecs_pgprot |= L_PTE_SHARED;
464 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
465 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
466 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
467 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
468 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
469 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
470 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
471 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
472 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
477 * Non-cacheable Normal - intended for memory areas that must
478 * not cause dirty cache line writebacks when used
480 if (cpu_arch >= CPU_ARCH_ARMv6) {
481 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
482 /* Non-cacheable Normal is XCB = 001 */
483 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
486 /* For both ARMv6 and non-TEX-remapping ARMv7 */
487 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
491 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
494 #ifdef CONFIG_ARM_LPAE
496 * Do not generate access flag faults for the kernel mappings.
498 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
499 mem_types[i].prot_pte |= PTE_EXT_AF;
500 if (mem_types[i].prot_sect)
501 mem_types[i].prot_sect |= PMD_SECT_AF;
503 kern_pgprot |= PTE_EXT_AF;
504 vecs_pgprot |= PTE_EXT_AF;
507 for (i = 0; i < 16; i++) {
508 unsigned long v = pgprot_val(protection_map[i]);
509 protection_map[i] = __pgprot(v | user_pgprot);
512 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
513 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
515 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
516 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
517 L_PTE_DIRTY | kern_pgprot);
519 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
520 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
521 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
522 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
523 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
524 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
525 mem_types[MT_ROM].prot_sect |= cp->pmd;
529 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
533 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
536 printk("Memory policy: ECC %sabled, Data cache %s\n",
537 ecc_mask ? "en" : "dis", cp->policy);
539 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
540 struct mem_type *t = &mem_types[i];
542 t->prot_l1 |= PMD_DOMAIN(t->domain);
544 t->prot_sect |= PMD_DOMAIN(t->domain);
548 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
549 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
550 unsigned long size, pgprot_t vma_prot)
553 return pgprot_noncached(vma_prot);
554 else if (file->f_flags & O_SYNC)
555 return pgprot_writecombine(vma_prot);
558 EXPORT_SYMBOL(phys_mem_access_prot);
561 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
563 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
565 void *ptr = __va(memblock_alloc(sz, align));
570 static void __init *early_alloc(unsigned long sz)
572 return early_alloc_aligned(sz, sz);
575 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
577 if (pmd_none(*pmd)) {
578 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
579 __pmd_populate(pmd, __pa(pte), prot);
581 BUG_ON(pmd_bad(*pmd));
582 return pte_offset_kernel(pmd, addr);
585 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
586 unsigned long end, unsigned long pfn,
587 const struct mem_type *type)
589 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
591 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
593 } while (pte++, addr += PAGE_SIZE, addr != end);
596 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
597 unsigned long end, phys_addr_t phys,
598 const struct mem_type *type)
600 pmd_t *pmd = pmd_offset(pud, addr);
603 * Try a section mapping - end, addr and phys must all be aligned
604 * to a section boundary. Note that PMDs refer to the individual
605 * L1 entries, whereas PGDs refer to a group of L1 entries making
606 * up one logical pointer to an L2 table.
608 if (type->prot_sect && ((addr | end | phys) & ~SECTION_MASK) == 0) {
611 #ifndef CONFIG_ARM_LPAE
612 if (addr & SECTION_SIZE)
617 *pmd = __pmd(phys | type->prot_sect);
618 phys += SECTION_SIZE;
619 } while (pmd++, addr += SECTION_SIZE, addr != end);
624 * No need to loop; pte's aren't interested in the
625 * individual L1 entries.
627 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
631 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
632 unsigned long end, unsigned long phys, const struct mem_type *type)
634 pud_t *pud = pud_offset(pgd, addr);
638 next = pud_addr_end(addr, end);
639 alloc_init_section(pud, addr, next, phys, type);
641 } while (pud++, addr = next, addr != end);
644 #ifndef CONFIG_ARM_LPAE
645 static void __init create_36bit_mapping(struct map_desc *md,
646 const struct mem_type *type)
648 unsigned long addr, length, end;
653 phys = __pfn_to_phys(md->pfn);
654 length = PAGE_ALIGN(md->length);
656 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
657 printk(KERN_ERR "MM: CPU does not support supersection "
658 "mapping for 0x%08llx at 0x%08lx\n",
659 (long long)__pfn_to_phys((u64)md->pfn), addr);
663 /* N.B. ARMv6 supersections are only defined to work with domain 0.
664 * Since domain assignments can in fact be arbitrary, the
665 * 'domain == 0' check below is required to insure that ARMv6
666 * supersections are only allocated for domain 0 regardless
667 * of the actual domain assignments in use.
670 printk(KERN_ERR "MM: invalid domain in supersection "
671 "mapping for 0x%08llx at 0x%08lx\n",
672 (long long)__pfn_to_phys((u64)md->pfn), addr);
676 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
677 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
678 " at 0x%08lx invalid alignment\n",
679 (long long)__pfn_to_phys((u64)md->pfn), addr);
684 * Shift bits [35:32] of address into bits [23:20] of PMD
687 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
689 pgd = pgd_offset_k(addr);
692 pud_t *pud = pud_offset(pgd, addr);
693 pmd_t *pmd = pmd_offset(pud, addr);
696 for (i = 0; i < 16; i++)
697 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
699 addr += SUPERSECTION_SIZE;
700 phys += SUPERSECTION_SIZE;
701 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
702 } while (addr != end);
704 #endif /* !CONFIG_ARM_LPAE */
707 * Create the page directory entries and any necessary
708 * page tables for the mapping specified by `md'. We
709 * are able to cope here with varying sizes and address
710 * offsets, and we take full advantage of sections and
713 static void __init create_mapping(struct map_desc *md)
715 unsigned long addr, length, end;
717 const struct mem_type *type;
720 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
721 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
722 " at 0x%08lx in user region\n",
723 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
727 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
728 md->virtual >= PAGE_OFFSET &&
729 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
730 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
731 " at 0x%08lx out of vmalloc space\n",
732 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
735 type = &mem_types[md->type];
737 #ifndef CONFIG_ARM_LPAE
739 * Catch 36-bit addresses
741 if (md->pfn >= 0x100000) {
742 create_36bit_mapping(md, type);
747 addr = md->virtual & PAGE_MASK;
748 phys = __pfn_to_phys(md->pfn);
749 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
751 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
752 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
753 "be mapped using pages, ignoring.\n",
754 (long long)__pfn_to_phys(md->pfn), addr);
758 pgd = pgd_offset_k(addr);
761 unsigned long next = pgd_addr_end(addr, end);
763 alloc_init_pud(pgd, addr, next, phys, type);
767 } while (pgd++, addr != end);
771 * Create the architecture specific mappings
773 void __init iotable_init(struct map_desc *io_desc, int nr)
776 struct vm_struct *vm;
781 vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
783 for (md = io_desc; nr; md++, nr--) {
785 vm->addr = (void *)(md->virtual & PAGE_MASK);
786 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
787 vm->phys_addr = __pfn_to_phys(md->pfn);
788 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
789 vm->flags |= VM_ARM_MTYPE(md->type);
790 vm->caller = iotable_init;
791 vm_area_add_early(vm++);
795 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
798 struct vm_struct *vm;
800 vm = early_alloc_aligned(sizeof(*vm), __alignof__(*vm));
801 vm->addr = (void *)addr;
803 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
805 vm_area_add_early(vm);
808 #ifndef CONFIG_ARM_LPAE
811 * The Linux PMD is made of two consecutive section entries covering 2MB
812 * (see definition in include/asm/pgtable-2level.h). However a call to
813 * create_mapping() may optimize static mappings by using individual
814 * 1MB section mappings. This leaves the actual PMD potentially half
815 * initialized if the top or bottom section entry isn't used, leaving it
816 * open to problems if a subsequent ioremap() or vmalloc() tries to use
817 * the virtual space left free by that unused section entry.
819 * Let's avoid the issue by inserting dummy vm entries covering the unused
820 * PMD halves once the static mappings are in place.
823 static void __init pmd_empty_section_gap(unsigned long addr)
825 vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
828 static void __init fill_pmd_gaps(void)
830 struct vm_struct *vm;
831 unsigned long addr, next = 0;
834 /* we're still single threaded hence no lock needed here */
835 for (vm = vmlist; vm; vm = vm->next) {
836 if (!(vm->flags & VM_ARM_STATIC_MAPPING))
838 addr = (unsigned long)vm->addr;
843 * Check if this vm starts on an odd section boundary.
844 * If so and the first section entry for this PMD is free
845 * then we block the corresponding virtual address.
847 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
848 pmd = pmd_off_k(addr);
850 pmd_empty_section_gap(addr & PMD_MASK);
854 * Then check if this vm ends on an odd section boundary.
855 * If so and the second section entry for this PMD is empty
856 * then we block the corresponding virtual address.
859 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
860 pmd = pmd_off_k(addr) + 1;
862 pmd_empty_section_gap(addr);
865 /* no need to look at any vm entry until we hit the next PMD */
866 next = (addr + PMD_SIZE - 1) & PMD_MASK;
871 #define fill_pmd_gaps() do { } while (0)
874 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
875 static void __init pci_reserve_io(void)
877 struct vm_struct *vm;
880 /* we're still single threaded hence no lock needed here */
881 for (vm = vmlist; vm; vm = vm->next) {
882 if (!(vm->flags & VM_ARM_STATIC_MAPPING))
884 addr = (unsigned long)vm->addr;
885 addr &= ~(SZ_2M - 1);
886 if (addr == PCI_IO_VIRT_BASE)
890 vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
893 #define pci_reserve_io() do { } while (0)
896 static void * __initdata vmalloc_min =
897 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
900 * vmalloc=size forces the vmalloc area to be exactly 'size'
901 * bytes. This can be used to increase (or decrease) the vmalloc
902 * area - the default is 240m.
904 static int __init early_vmalloc(char *arg)
906 unsigned long vmalloc_reserve = memparse(arg, NULL);
908 if (vmalloc_reserve < SZ_16M) {
909 vmalloc_reserve = SZ_16M;
911 "vmalloc area too small, limiting to %luMB\n",
912 vmalloc_reserve >> 20);
915 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
916 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
918 "vmalloc area is too big, limiting to %luMB\n",
919 vmalloc_reserve >> 20);
922 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
925 early_param("vmalloc", early_vmalloc);
927 phys_addr_t arm_lowmem_limit __initdata = 0;
929 void __init sanity_check_meminfo(void)
931 int i, j, highmem = 0;
933 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
934 struct membank *bank = &meminfo.bank[j];
935 *bank = meminfo.bank[i];
937 if (bank->start > ULONG_MAX)
940 #ifdef CONFIG_HIGHMEM
941 if (__va(bank->start) >= vmalloc_min ||
942 __va(bank->start) < (void *)PAGE_OFFSET)
945 bank->highmem = highmem;
948 * Split those memory banks which are partially overlapping
949 * the vmalloc area greatly simplifying things later.
951 if (!highmem && __va(bank->start) < vmalloc_min &&
952 bank->size > vmalloc_min - __va(bank->start)) {
953 if (meminfo.nr_banks >= NR_BANKS) {
954 printk(KERN_CRIT "NR_BANKS too low, "
955 "ignoring high memory\n");
957 memmove(bank + 1, bank,
958 (meminfo.nr_banks - i) * sizeof(*bank));
961 bank[1].size -= vmalloc_min - __va(bank->start);
962 bank[1].start = __pa(vmalloc_min - 1) + 1;
963 bank[1].highmem = highmem = 1;
966 bank->size = vmalloc_min - __va(bank->start);
969 bank->highmem = highmem;
972 * Highmem banks not allowed with !CONFIG_HIGHMEM.
975 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
976 "(!CONFIG_HIGHMEM).\n",
977 (unsigned long long)bank->start,
978 (unsigned long long)bank->start + bank->size - 1);
983 * Check whether this memory bank would entirely overlap
986 if (__va(bank->start) >= vmalloc_min ||
987 __va(bank->start) < (void *)PAGE_OFFSET) {
988 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
989 "(vmalloc region overlap).\n",
990 (unsigned long long)bank->start,
991 (unsigned long long)bank->start + bank->size - 1);
996 * Check whether this memory bank would partially overlap
999 if (__va(bank->start + bank->size) > vmalloc_min ||
1000 __va(bank->start + bank->size) < __va(bank->start)) {
1001 unsigned long newsize = vmalloc_min - __va(bank->start);
1002 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
1003 "to -%.8llx (vmalloc region overlap).\n",
1004 (unsigned long long)bank->start,
1005 (unsigned long long)bank->start + bank->size - 1,
1006 (unsigned long long)bank->start + newsize - 1);
1007 bank->size = newsize;
1010 if (!bank->highmem && bank->start + bank->size > arm_lowmem_limit)
1011 arm_lowmem_limit = bank->start + bank->size;
1015 #ifdef CONFIG_HIGHMEM
1017 const char *reason = NULL;
1019 if (cache_is_vipt_aliasing()) {
1021 * Interactions between kmap and other mappings
1022 * make highmem support with aliasing VIPT caches
1025 reason = "with VIPT aliasing cache";
1028 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
1030 while (j > 0 && meminfo.bank[j - 1].highmem)
1035 meminfo.nr_banks = j;
1036 high_memory = __va(arm_lowmem_limit - 1) + 1;
1037 memblock_set_current_limit(arm_lowmem_limit);
1040 static inline void prepare_page_table(void)
1046 * Clear out all the mappings below the kernel image.
1048 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1049 pmd_clear(pmd_off_k(addr));
1051 #ifdef CONFIG_XIP_KERNEL
1052 /* The XIP kernel is mapped in the module area -- skip over it */
1053 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1055 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1056 pmd_clear(pmd_off_k(addr));
1059 * Find the end of the first block of lowmem.
1061 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1062 if (end >= arm_lowmem_limit)
1063 end = arm_lowmem_limit;
1066 * Clear out all the kernel space mappings, except for the first
1067 * memory bank, up to the vmalloc region.
1069 for (addr = __phys_to_virt(end);
1070 addr < VMALLOC_START; addr += PMD_SIZE)
1071 pmd_clear(pmd_off_k(addr));
1074 #ifdef CONFIG_ARM_LPAE
1075 /* the first page is reserved for pgd */
1076 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1077 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1079 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1083 * Reserve the special regions of memory
1085 void __init arm_mm_memblock_reserve(void)
1088 * Reserve the page tables. These are already in use,
1089 * and can only be in node 0.
1091 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1093 #ifdef CONFIG_SA1111
1095 * Because of the SA1111 DMA bug, we want to preserve our
1096 * precious DMA-able memory...
1098 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1103 * Set up the device mappings. Since we clear out the page tables for all
1104 * mappings above VMALLOC_START, we will remove any debug device mappings.
1105 * This means you have to be careful how you debug this function, or any
1106 * called function. This means you can't use any function or debugging
1107 * method which may touch any device, otherwise the kernel _will_ crash.
1109 static void __init devicemaps_init(struct machine_desc *mdesc)
1111 struct map_desc map;
1116 * Allocate the vector page early.
1118 vectors = early_alloc(PAGE_SIZE);
1120 early_trap_init(vectors);
1122 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1123 pmd_clear(pmd_off_k(addr));
1126 * Map the kernel if it is XIP.
1127 * It is always first in the modulearea.
1129 #ifdef CONFIG_XIP_KERNEL
1130 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1131 map.virtual = MODULES_VADDR;
1132 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1134 create_mapping(&map);
1138 * Map the cache flushing regions.
1141 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1142 map.virtual = FLUSH_BASE;
1144 map.type = MT_CACHECLEAN;
1145 create_mapping(&map);
1147 #ifdef FLUSH_BASE_MINICACHE
1148 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1149 map.virtual = FLUSH_BASE_MINICACHE;
1151 map.type = MT_MINICLEAN;
1152 create_mapping(&map);
1156 * Create a mapping for the machine vectors at the high-vectors
1157 * location (0xffff0000). If we aren't using high-vectors, also
1158 * create a mapping at the low-vectors virtual address.
1160 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1161 map.virtual = 0xffff0000;
1162 map.length = PAGE_SIZE;
1163 map.type = MT_HIGH_VECTORS;
1164 create_mapping(&map);
1166 if (!vectors_high()) {
1168 map.type = MT_LOW_VECTORS;
1169 create_mapping(&map);
1173 * Ask the machine support to map in the statically mapped devices.
1179 /* Reserve fixed i/o space in VMALLOC region */
1183 * Finally flush the caches and tlb to ensure that we're in a
1184 * consistent state wrt the writebuffer. This also ensures that
1185 * any write-allocated cache lines in the vector page are written
1186 * back. After this point, we can start to touch devices again.
1188 local_flush_tlb_all();
1192 static void __init kmap_init(void)
1194 #ifdef CONFIG_HIGHMEM
1195 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1196 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1200 static void __init map_lowmem(void)
1202 struct memblock_region *reg;
1204 /* Map all the lowmem memory banks. */
1205 for_each_memblock(memory, reg) {
1206 phys_addr_t start = reg->base;
1207 phys_addr_t end = start + reg->size;
1208 struct map_desc map;
1210 if (end > arm_lowmem_limit)
1211 end = arm_lowmem_limit;
1215 map.pfn = __phys_to_pfn(start);
1216 map.virtual = __phys_to_virt(start);
1217 map.length = end - start;
1218 map.type = MT_MEMORY;
1220 create_mapping(&map);
1225 * paging_init() sets up the page tables, initialises the zone memory
1226 * maps, and sets up the zero page, bad page and bad page tables.
1228 void __init paging_init(struct machine_desc *mdesc)
1232 memblock_set_current_limit(arm_lowmem_limit);
1234 build_mem_type_table();
1235 prepare_page_table();
1237 dma_contiguous_remap();
1238 devicemaps_init(mdesc);
1241 top_pmd = pmd_off_k(0xffff0000);
1243 /* allocate the zero page. */
1244 zero_page = early_alloc(PAGE_SIZE);
1248 empty_zero_page = virt_to_page(zero_page);
1249 __flush_dcache_page(NULL, empty_zero_page);