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/traps.h>
31 #include <asm/mach/arch.h>
32 #include <asm/mach/map.h>
37 * empty_zero_page is a special page that is used for
38 * zero-initialized data and COW.
40 struct page *empty_zero_page;
41 EXPORT_SYMBOL(empty_zero_page);
44 * The pmd table for the upper-most set of pages.
48 #define CPOLICY_UNCACHED 0
49 #define CPOLICY_BUFFERED 1
50 #define CPOLICY_WRITETHROUGH 2
51 #define CPOLICY_WRITEBACK 3
52 #define CPOLICY_WRITEALLOC 4
54 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
55 static unsigned int ecc_mask __initdata = 0;
57 pgprot_t pgprot_kernel;
59 EXPORT_SYMBOL(pgprot_user);
60 EXPORT_SYMBOL(pgprot_kernel);
63 const char policy[16];
69 static struct cachepolicy cache_policies[] __initdata = {
73 .pmd = PMD_SECT_UNCACHED,
74 .pte = L_PTE_MT_UNCACHED,
78 .pmd = PMD_SECT_BUFFERED,
79 .pte = L_PTE_MT_BUFFERABLE,
81 .policy = "writethrough",
84 .pte = L_PTE_MT_WRITETHROUGH,
86 .policy = "writeback",
89 .pte = L_PTE_MT_WRITEBACK,
91 .policy = "writealloc",
94 .pte = L_PTE_MT_WRITEALLOC,
99 * These are useful for identifying cache coherency
100 * problems by allowing the cache or the cache and
101 * writebuffer to be turned off. (Note: the write
102 * buffer should not be on and the cache off).
104 static int __init early_cachepolicy(char *p)
108 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
109 int len = strlen(cache_policies[i].policy);
111 if (memcmp(p, cache_policies[i].policy, len) == 0) {
113 cr_alignment &= ~cache_policies[i].cr_mask;
114 cr_no_alignment &= ~cache_policies[i].cr_mask;
118 if (i == ARRAY_SIZE(cache_policies))
119 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
121 * This restriction is partly to do with the way we boot; it is
122 * unpredictable to have memory mapped using two different sets of
123 * memory attributes (shared, type, and cache attribs). We can not
124 * change these attributes once the initial assembly has setup the
127 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
128 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
129 cachepolicy = CPOLICY_WRITEBACK;
132 set_cr(cr_alignment);
135 early_param("cachepolicy", early_cachepolicy);
137 static int __init early_nocache(char *__unused)
139 char *p = "buffered";
140 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
141 early_cachepolicy(p);
144 early_param("nocache", early_nocache);
146 static int __init early_nowrite(char *__unused)
148 char *p = "uncached";
149 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
150 early_cachepolicy(p);
153 early_param("nowb", early_nowrite);
155 #ifndef CONFIG_ARM_LPAE
156 static int __init early_ecc(char *p)
158 if (memcmp(p, "on", 2) == 0)
159 ecc_mask = PMD_PROTECTION;
160 else if (memcmp(p, "off", 3) == 0)
164 early_param("ecc", early_ecc);
167 static int __init noalign_setup(char *__unused)
169 cr_alignment &= ~CR_A;
170 cr_no_alignment &= ~CR_A;
171 set_cr(cr_alignment);
174 __setup("noalign", noalign_setup);
177 void adjust_cr(unsigned long mask, unsigned long set)
185 local_irq_save(flags);
187 cr_no_alignment = (cr_no_alignment & ~mask) | set;
188 cr_alignment = (cr_alignment & ~mask) | set;
190 set_cr((get_cr() & ~mask) | set);
192 local_irq_restore(flags);
196 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
197 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
199 static struct mem_type mem_types[] = {
200 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
201 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
203 .prot_l1 = PMD_TYPE_TABLE,
204 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
207 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
208 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
209 .prot_l1 = PMD_TYPE_TABLE,
210 .prot_sect = PROT_SECT_DEVICE,
213 [MT_DEVICE_CACHED] = { /* ioremap_cached */
214 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
215 .prot_l1 = PMD_TYPE_TABLE,
216 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
219 [MT_DEVICE_WC] = { /* ioremap_wc */
220 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
221 .prot_l1 = PMD_TYPE_TABLE,
222 .prot_sect = PROT_SECT_DEVICE,
226 .prot_pte = PROT_PTE_DEVICE,
227 .prot_l1 = PMD_TYPE_TABLE,
228 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
232 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
233 .domain = DOMAIN_KERNEL,
235 #ifndef CONFIG_ARM_LPAE
237 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
238 .domain = DOMAIN_KERNEL,
242 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
244 .prot_l1 = PMD_TYPE_TABLE,
245 .domain = DOMAIN_USER,
247 [MT_HIGH_VECTORS] = {
248 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
249 L_PTE_USER | L_PTE_RDONLY,
250 .prot_l1 = PMD_TYPE_TABLE,
251 .domain = DOMAIN_USER,
254 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
255 .prot_l1 = PMD_TYPE_TABLE,
256 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
257 .domain = DOMAIN_KERNEL,
260 .prot_sect = PMD_TYPE_SECT,
261 .domain = DOMAIN_KERNEL,
263 [MT_MEMORY_NONCACHED] = {
264 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
266 .prot_l1 = PMD_TYPE_TABLE,
267 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
268 .domain = DOMAIN_KERNEL,
271 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
273 .prot_l1 = PMD_TYPE_TABLE,
274 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
275 .domain = DOMAIN_KERNEL,
278 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
279 .prot_l1 = PMD_TYPE_TABLE,
280 .domain = DOMAIN_KERNEL,
283 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
285 .prot_l1 = PMD_TYPE_TABLE,
286 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
287 PMD_SECT_UNCACHED | PMD_SECT_XN,
288 .domain = DOMAIN_KERNEL,
292 const struct mem_type *get_mem_type(unsigned int type)
294 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
296 EXPORT_SYMBOL(get_mem_type);
299 * Adjust the PMD section entries according to the CPU in use.
301 static void __init build_mem_type_table(void)
303 struct cachepolicy *cp;
304 unsigned int cr = get_cr();
305 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
306 int cpu_arch = cpu_architecture();
309 if (cpu_arch < CPU_ARCH_ARMv6) {
310 #if defined(CONFIG_CPU_DCACHE_DISABLE)
311 if (cachepolicy > CPOLICY_BUFFERED)
312 cachepolicy = CPOLICY_BUFFERED;
313 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
314 if (cachepolicy > CPOLICY_WRITETHROUGH)
315 cachepolicy = CPOLICY_WRITETHROUGH;
318 if (cpu_arch < CPU_ARCH_ARMv5) {
319 if (cachepolicy >= CPOLICY_WRITEALLOC)
320 cachepolicy = CPOLICY_WRITEBACK;
324 cachepolicy = CPOLICY_WRITEALLOC;
327 * Strip out features not present on earlier architectures.
328 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
329 * without extended page tables don't have the 'Shared' bit.
331 if (cpu_arch < CPU_ARCH_ARMv5)
332 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
333 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
334 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
335 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
336 mem_types[i].prot_sect &= ~PMD_SECT_S;
339 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
340 * "update-able on write" bit on ARM610). However, Xscale and
341 * Xscale3 require this bit to be cleared.
343 if (cpu_is_xscale() || cpu_is_xsc3()) {
344 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
345 mem_types[i].prot_sect &= ~PMD_BIT4;
346 mem_types[i].prot_l1 &= ~PMD_BIT4;
348 } else if (cpu_arch < CPU_ARCH_ARMv6) {
349 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
350 if (mem_types[i].prot_l1)
351 mem_types[i].prot_l1 |= PMD_BIT4;
352 if (mem_types[i].prot_sect)
353 mem_types[i].prot_sect |= PMD_BIT4;
358 * Mark the device areas according to the CPU/architecture.
360 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
361 if (!cpu_is_xsc3()) {
363 * Mark device regions on ARMv6+ as execute-never
364 * to prevent speculative instruction fetches.
366 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
367 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
368 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
369 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
371 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
373 * For ARMv7 with TEX remapping,
374 * - shared device is SXCB=1100
375 * - nonshared device is SXCB=0100
376 * - write combine device mem is SXCB=0001
377 * (Uncached Normal memory)
379 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
380 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
381 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
382 } else if (cpu_is_xsc3()) {
385 * - shared device is TEXCB=00101
386 * - nonshared device is TEXCB=01000
387 * - write combine device mem is TEXCB=00100
388 * (Inner/Outer Uncacheable in xsc3 parlance)
390 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
391 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
392 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
395 * For ARMv6 and ARMv7 without TEX remapping,
396 * - shared device is TEXCB=00001
397 * - nonshared device is TEXCB=01000
398 * - write combine device mem is TEXCB=00100
399 * (Uncached Normal in ARMv6 parlance).
401 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
402 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
403 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
407 * On others, write combining is "Uncached/Buffered"
409 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
413 * Now deal with the memory-type mappings
415 cp = &cache_policies[cachepolicy];
416 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
419 * Only use write-through for non-SMP systems
421 if (!is_smp() && cpu_arch >= CPU_ARCH_ARMv5 && cachepolicy > CPOLICY_WRITETHROUGH)
422 vecs_pgprot = cache_policies[CPOLICY_WRITETHROUGH].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_NONCACHED].prot_sect |= PMD_SECT_S;
432 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
435 * ARMv6 and above have extended page tables.
437 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
438 #ifndef CONFIG_ARM_LPAE
440 * Mark cache clean areas and XIP ROM read only
441 * from SVC mode and no access from userspace.
443 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
444 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
445 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
450 * Mark memory with the "shared" attribute
453 user_pgprot |= L_PTE_SHARED;
454 kern_pgprot |= L_PTE_SHARED;
455 vecs_pgprot |= L_PTE_SHARED;
456 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
457 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
458 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
459 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
460 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
461 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
462 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
463 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
468 * Non-cacheable Normal - intended for memory areas that must
469 * not cause dirty cache line writebacks when used
471 if (cpu_arch >= CPU_ARCH_ARMv6) {
472 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
473 /* Non-cacheable Normal is XCB = 001 */
474 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
477 /* For both ARMv6 and non-TEX-remapping ARMv7 */
478 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
482 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
485 #ifdef CONFIG_ARM_LPAE
487 * Do not generate access flag faults for the kernel mappings.
489 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
490 mem_types[i].prot_pte |= PTE_EXT_AF;
491 mem_types[i].prot_sect |= PMD_SECT_AF;
493 kern_pgprot |= PTE_EXT_AF;
494 vecs_pgprot |= PTE_EXT_AF;
497 for (i = 0; i < 16; i++) {
498 unsigned long v = pgprot_val(protection_map[i]);
499 protection_map[i] = __pgprot(v | user_pgprot);
502 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
503 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
505 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
506 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
507 L_PTE_DIRTY | kern_pgprot);
509 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
510 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
511 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
512 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
513 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
514 mem_types[MT_ROM].prot_sect |= cp->pmd;
518 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
522 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
525 printk("Memory policy: ECC %sabled, Data cache %s\n",
526 ecc_mask ? "en" : "dis", cp->policy);
528 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
529 struct mem_type *t = &mem_types[i];
531 t->prot_l1 |= PMD_DOMAIN(t->domain);
533 t->prot_sect |= PMD_DOMAIN(t->domain);
537 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
538 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
539 unsigned long size, pgprot_t vma_prot)
542 return pgprot_noncached(vma_prot);
543 else if (file->f_flags & O_SYNC)
544 return pgprot_writecombine(vma_prot);
547 EXPORT_SYMBOL(phys_mem_access_prot);
550 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
552 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
554 void *ptr = __va(memblock_alloc(sz, align));
559 static void __init *early_alloc(unsigned long sz)
561 return early_alloc_aligned(sz, sz);
564 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
566 if (pmd_none(*pmd)) {
567 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
568 __pmd_populate(pmd, __pa(pte), prot);
570 BUG_ON(pmd_bad(*pmd));
571 return pte_offset_kernel(pmd, addr);
574 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
575 unsigned long end, unsigned long pfn,
576 const struct mem_type *type)
578 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
580 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
582 } while (pte++, addr += PAGE_SIZE, addr != end);
585 static void __init alloc_init_section(pud_t *pud, unsigned long addr,
586 unsigned long end, phys_addr_t phys,
587 const struct mem_type *type)
589 pmd_t *pmd = pmd_offset(pud, addr);
592 * Try a section mapping - end, addr and phys must all be aligned
593 * to a section boundary. Note that PMDs refer to the individual
594 * L1 entries, whereas PGDs refer to a group of L1 entries making
595 * up one logical pointer to an L2 table.
597 if (((addr | end | phys) & ~SECTION_MASK) == 0) {
600 #ifndef CONFIG_ARM_LPAE
601 if (addr & SECTION_SIZE)
606 *pmd = __pmd(phys | type->prot_sect);
607 phys += SECTION_SIZE;
608 } while (pmd++, addr += SECTION_SIZE, addr != end);
613 * No need to loop; pte's aren't interested in the
614 * individual L1 entries.
616 alloc_init_pte(pmd, addr, end, __phys_to_pfn(phys), type);
620 static void alloc_init_pud(pgd_t *pgd, unsigned long addr, unsigned long end,
621 unsigned long phys, const struct mem_type *type)
623 pud_t *pud = pud_offset(pgd, addr);
627 next = pud_addr_end(addr, end);
628 alloc_init_section(pud, addr, next, phys, type);
630 } while (pud++, addr = next, addr != end);
633 #ifndef CONFIG_ARM_LPAE
634 static void __init create_36bit_mapping(struct map_desc *md,
635 const struct mem_type *type)
637 unsigned long addr, length, end;
642 phys = __pfn_to_phys(md->pfn);
643 length = PAGE_ALIGN(md->length);
645 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
646 printk(KERN_ERR "MM: CPU does not support supersection "
647 "mapping for 0x%08llx at 0x%08lx\n",
648 (long long)__pfn_to_phys((u64)md->pfn), addr);
652 /* N.B. ARMv6 supersections are only defined to work with domain 0.
653 * Since domain assignments can in fact be arbitrary, the
654 * 'domain == 0' check below is required to insure that ARMv6
655 * supersections are only allocated for domain 0 regardless
656 * of the actual domain assignments in use.
659 printk(KERN_ERR "MM: invalid domain in supersection "
660 "mapping for 0x%08llx at 0x%08lx\n",
661 (long long)__pfn_to_phys((u64)md->pfn), addr);
665 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
666 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
667 " at 0x%08lx invalid alignment\n",
668 (long long)__pfn_to_phys((u64)md->pfn), addr);
673 * Shift bits [35:32] of address into bits [23:20] of PMD
676 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
678 pgd = pgd_offset_k(addr);
681 pud_t *pud = pud_offset(pgd, addr);
682 pmd_t *pmd = pmd_offset(pud, addr);
685 for (i = 0; i < 16; i++)
686 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
688 addr += SUPERSECTION_SIZE;
689 phys += SUPERSECTION_SIZE;
690 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
691 } while (addr != end);
693 #endif /* !CONFIG_ARM_LPAE */
696 * Create the page directory entries and any necessary
697 * page tables for the mapping specified by `md'. We
698 * are able to cope here with varying sizes and address
699 * offsets, and we take full advantage of sections and
702 static void __init create_mapping(struct map_desc *md)
704 unsigned long addr, length, end;
706 const struct mem_type *type;
709 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
710 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
711 " at 0x%08lx in user region\n",
712 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
716 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
717 md->virtual >= PAGE_OFFSET &&
718 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
719 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
720 " at 0x%08lx out of vmalloc space\n",
721 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
724 type = &mem_types[md->type];
726 #ifndef CONFIG_ARM_LPAE
728 * Catch 36-bit addresses
730 if (md->pfn >= 0x100000) {
731 create_36bit_mapping(md, type);
736 addr = md->virtual & PAGE_MASK;
737 phys = __pfn_to_phys(md->pfn);
738 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
740 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
741 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
742 "be mapped using pages, ignoring.\n",
743 (long long)__pfn_to_phys(md->pfn), addr);
747 pgd = pgd_offset_k(addr);
750 unsigned long next = pgd_addr_end(addr, end);
752 alloc_init_pud(pgd, addr, next, phys, type);
756 } while (pgd++, addr != end);
760 * Create the architecture specific mappings
762 void __init iotable_init(struct map_desc *io_desc, int nr)
765 struct vm_struct *vm;
770 vm = early_alloc_aligned(sizeof(*vm) * nr, __alignof__(*vm));
772 for (md = io_desc; nr; md++, nr--) {
774 vm->addr = (void *)(md->virtual & PAGE_MASK);
775 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
776 vm->phys_addr = __pfn_to_phys(md->pfn);
777 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
778 vm->flags |= VM_ARM_MTYPE(md->type);
779 vm->caller = iotable_init;
780 vm_area_add_early(vm++);
784 static void * __initdata vmalloc_min =
785 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
788 * vmalloc=size forces the vmalloc area to be exactly 'size'
789 * bytes. This can be used to increase (or decrease) the vmalloc
790 * area - the default is 240m.
792 static int __init early_vmalloc(char *arg)
794 unsigned long vmalloc_reserve = memparse(arg, NULL);
796 if (vmalloc_reserve < SZ_16M) {
797 vmalloc_reserve = SZ_16M;
799 "vmalloc area too small, limiting to %luMB\n",
800 vmalloc_reserve >> 20);
803 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
804 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
806 "vmalloc area is too big, limiting to %luMB\n",
807 vmalloc_reserve >> 20);
810 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
813 early_param("vmalloc", early_vmalloc);
815 static phys_addr_t lowmem_limit __initdata = 0;
817 void __init sanity_check_meminfo(void)
819 int i, j, highmem = 0;
821 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
822 struct membank *bank = &meminfo.bank[j];
823 *bank = meminfo.bank[i];
825 if (bank->start > ULONG_MAX)
828 #ifdef CONFIG_HIGHMEM
829 if (__va(bank->start) >= vmalloc_min ||
830 __va(bank->start) < (void *)PAGE_OFFSET)
833 bank->highmem = highmem;
836 * Split those memory banks which are partially overlapping
837 * the vmalloc area greatly simplifying things later.
839 if (!highmem && __va(bank->start) < vmalloc_min &&
840 bank->size > vmalloc_min - __va(bank->start)) {
841 if (meminfo.nr_banks >= NR_BANKS) {
842 printk(KERN_CRIT "NR_BANKS too low, "
843 "ignoring high memory\n");
845 memmove(bank + 1, bank,
846 (meminfo.nr_banks - i) * sizeof(*bank));
849 bank[1].size -= vmalloc_min - __va(bank->start);
850 bank[1].start = __pa(vmalloc_min - 1) + 1;
851 bank[1].highmem = highmem = 1;
854 bank->size = vmalloc_min - __va(bank->start);
857 bank->highmem = highmem;
860 * Highmem banks not allowed with !CONFIG_HIGHMEM.
863 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
864 "(!CONFIG_HIGHMEM).\n",
865 (unsigned long long)bank->start,
866 (unsigned long long)bank->start + bank->size - 1);
871 * Check whether this memory bank would entirely overlap
874 if (__va(bank->start) >= vmalloc_min ||
875 __va(bank->start) < (void *)PAGE_OFFSET) {
876 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
877 "(vmalloc region overlap).\n",
878 (unsigned long long)bank->start,
879 (unsigned long long)bank->start + bank->size - 1);
884 * Check whether this memory bank would partially overlap
887 if (__va(bank->start + bank->size) > vmalloc_min ||
888 __va(bank->start + bank->size) < __va(bank->start)) {
889 unsigned long newsize = vmalloc_min - __va(bank->start);
890 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
891 "to -%.8llx (vmalloc region overlap).\n",
892 (unsigned long long)bank->start,
893 (unsigned long long)bank->start + bank->size - 1,
894 (unsigned long long)bank->start + newsize - 1);
895 bank->size = newsize;
898 if (!bank->highmem && bank->start + bank->size > lowmem_limit)
899 lowmem_limit = bank->start + bank->size;
903 #ifdef CONFIG_HIGHMEM
905 const char *reason = NULL;
907 if (cache_is_vipt_aliasing()) {
909 * Interactions between kmap and other mappings
910 * make highmem support with aliasing VIPT caches
913 reason = "with VIPT aliasing cache";
916 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
918 while (j > 0 && meminfo.bank[j - 1].highmem)
923 meminfo.nr_banks = j;
924 high_memory = __va(lowmem_limit - 1) + 1;
925 memblock_set_current_limit(lowmem_limit);
928 static inline void prepare_page_table(void)
934 * Clear out all the mappings below the kernel image.
936 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
937 pmd_clear(pmd_off_k(addr));
939 #ifdef CONFIG_XIP_KERNEL
940 /* The XIP kernel is mapped in the module area -- skip over it */
941 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
943 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
944 pmd_clear(pmd_off_k(addr));
947 * Find the end of the first block of lowmem.
949 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
950 if (end >= lowmem_limit)
954 * Clear out all the kernel space mappings, except for the first
955 * memory bank, up to the vmalloc region.
957 for (addr = __phys_to_virt(end);
958 addr < VMALLOC_START; addr += PMD_SIZE)
959 pmd_clear(pmd_off_k(addr));
962 #ifdef CONFIG_ARM_LPAE
963 /* the first page is reserved for pgd */
964 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
965 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
967 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
971 * Reserve the special regions of memory
973 void __init arm_mm_memblock_reserve(void)
976 * Reserve the page tables. These are already in use,
977 * and can only be in node 0.
979 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
983 * Because of the SA1111 DMA bug, we want to preserve our
984 * precious DMA-able memory...
986 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
991 * Set up the device mappings. Since we clear out the page tables for all
992 * mappings above VMALLOC_START, we will remove any debug device mappings.
993 * This means you have to be careful how you debug this function, or any
994 * called function. This means you can't use any function or debugging
995 * method which may touch any device, otherwise the kernel _will_ crash.
997 static void __init devicemaps_init(struct machine_desc *mdesc)
1004 * Allocate the vector page early.
1006 vectors = early_alloc(PAGE_SIZE);
1008 early_trap_init(vectors);
1010 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1011 pmd_clear(pmd_off_k(addr));
1014 * Map the kernel if it is XIP.
1015 * It is always first in the modulearea.
1017 #ifdef CONFIG_XIP_KERNEL
1018 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1019 map.virtual = MODULES_VADDR;
1020 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1022 create_mapping(&map);
1026 * Map the cache flushing regions.
1029 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1030 map.virtual = FLUSH_BASE;
1032 map.type = MT_CACHECLEAN;
1033 create_mapping(&map);
1035 #ifdef FLUSH_BASE_MINICACHE
1036 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1037 map.virtual = FLUSH_BASE_MINICACHE;
1039 map.type = MT_MINICLEAN;
1040 create_mapping(&map);
1044 * Create a mapping for the machine vectors at the high-vectors
1045 * location (0xffff0000). If we aren't using high-vectors, also
1046 * create a mapping at the low-vectors virtual address.
1048 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1049 map.virtual = 0xffff0000;
1050 map.length = PAGE_SIZE;
1051 map.type = MT_HIGH_VECTORS;
1052 create_mapping(&map);
1054 if (!vectors_high()) {
1056 map.type = MT_LOW_VECTORS;
1057 create_mapping(&map);
1061 * Ask the machine support to map in the statically mapped devices.
1067 * Finally flush the caches and tlb to ensure that we're in a
1068 * consistent state wrt the writebuffer. This also ensures that
1069 * any write-allocated cache lines in the vector page are written
1070 * back. After this point, we can start to touch devices again.
1072 local_flush_tlb_all();
1076 static void __init kmap_init(void)
1078 #ifdef CONFIG_HIGHMEM
1079 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1080 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1084 static void __init map_lowmem(void)
1086 struct memblock_region *reg;
1088 /* Map all the lowmem memory banks. */
1089 for_each_memblock(memory, reg) {
1090 phys_addr_t start = reg->base;
1091 phys_addr_t end = start + reg->size;
1092 struct map_desc map;
1094 if (end > lowmem_limit)
1099 map.pfn = __phys_to_pfn(start);
1100 map.virtual = __phys_to_virt(start);
1101 map.length = end - start;
1102 map.type = MT_MEMORY;
1104 create_mapping(&map);
1109 * paging_init() sets up the page tables, initialises the zone memory
1110 * maps, and sets up the zero page, bad page and bad page tables.
1112 void __init paging_init(struct machine_desc *mdesc)
1116 memblock_set_current_limit(lowmem_limit);
1118 build_mem_type_table();
1119 prepare_page_table();
1121 devicemaps_init(mdesc);
1124 top_pmd = pmd_off_k(0xffff0000);
1126 /* allocate the zero page. */
1127 zero_page = early_alloc(PAGE_SIZE);
1131 empty_zero_page = virt_to_page(zero_page);
1132 __flush_dcache_page(NULL, empty_zero_page);