2 * linux/arch/arm/mm/mmu.c
4 * Copyright (C) 1995-2005 Russell King
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
18 #include <linux/vmalloc.h>
19 #include <linux/sizes.h>
22 #include <asm/cputype.h>
23 #include <asm/sections.h>
24 #include <asm/cachetype.h>
25 #include <asm/setup.h>
26 #include <asm/smp_plat.h>
28 #include <asm/highmem.h>
29 #include <asm/system_info.h>
30 #include <asm/traps.h>
32 #include <asm/mach/arch.h>
33 #include <asm/mach/map.h>
34 #include <asm/mach/pci.h>
40 * empty_zero_page is a special page that is used for
41 * zero-initialized data and COW.
43 struct page *empty_zero_page;
44 EXPORT_SYMBOL(empty_zero_page);
47 * The pmd table for the upper-most set of pages.
51 #define CPOLICY_UNCACHED 0
52 #define CPOLICY_BUFFERED 1
53 #define CPOLICY_WRITETHROUGH 2
54 #define CPOLICY_WRITEBACK 3
55 #define CPOLICY_WRITEALLOC 4
57 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
58 static unsigned int ecc_mask __initdata = 0;
60 pgprot_t pgprot_kernel;
61 pgprot_t pgprot_hyp_device;
63 pgprot_t pgprot_s2_device;
65 EXPORT_SYMBOL(pgprot_user);
66 EXPORT_SYMBOL(pgprot_kernel);
69 const char policy[16];
76 #ifdef CONFIG_ARM_LPAE
77 #define s2_policy(policy) policy
79 #define s2_policy(policy) 0
82 static struct cachepolicy cache_policies[] __initdata = {
86 .pmd = PMD_SECT_UNCACHED,
87 .pte = L_PTE_MT_UNCACHED,
88 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED),
92 .pmd = PMD_SECT_BUFFERED,
93 .pte = L_PTE_MT_BUFFERABLE,
94 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED),
96 .policy = "writethrough",
99 .pte = L_PTE_MT_WRITETHROUGH,
100 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITETHROUGH),
102 .policy = "writeback",
105 .pte = L_PTE_MT_WRITEBACK,
106 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK),
108 .policy = "writealloc",
110 .pmd = PMD_SECT_WBWA,
111 .pte = L_PTE_MT_WRITEALLOC,
112 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK),
116 #ifdef CONFIG_CPU_CP15
118 * These are useful for identifying cache coherency
119 * problems by allowing the cache or the cache and
120 * writebuffer to be turned off. (Note: the write
121 * buffer should not be on and the cache off).
123 static int __init early_cachepolicy(char *p)
127 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
128 int len = strlen(cache_policies[i].policy);
130 if (memcmp(p, cache_policies[i].policy, len) == 0) {
132 cr_alignment &= ~cache_policies[i].cr_mask;
133 cr_no_alignment &= ~cache_policies[i].cr_mask;
137 if (i == ARRAY_SIZE(cache_policies))
138 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
140 * This restriction is partly to do with the way we boot; it is
141 * unpredictable to have memory mapped using two different sets of
142 * memory attributes (shared, type, and cache attribs). We can not
143 * change these attributes once the initial assembly has setup the
146 if (cpu_architecture() >= CPU_ARCH_ARMv6) {
147 printk(KERN_WARNING "Only cachepolicy=writeback supported on ARMv6 and later\n");
148 cachepolicy = CPOLICY_WRITEBACK;
151 set_cr(cr_alignment);
154 early_param("cachepolicy", early_cachepolicy);
156 static int __init early_nocache(char *__unused)
158 char *p = "buffered";
159 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
160 early_cachepolicy(p);
163 early_param("nocache", early_nocache);
165 static int __init early_nowrite(char *__unused)
167 char *p = "uncached";
168 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
169 early_cachepolicy(p);
172 early_param("nowb", early_nowrite);
174 #ifndef CONFIG_ARM_LPAE
175 static int __init early_ecc(char *p)
177 if (memcmp(p, "on", 2) == 0)
178 ecc_mask = PMD_PROTECTION;
179 else if (memcmp(p, "off", 3) == 0)
183 early_param("ecc", early_ecc);
186 static int __init noalign_setup(char *__unused)
188 cr_alignment &= ~CR_A;
189 cr_no_alignment &= ~CR_A;
190 set_cr(cr_alignment);
193 __setup("noalign", noalign_setup);
196 void adjust_cr(unsigned long mask, unsigned long set)
204 local_irq_save(flags);
206 cr_no_alignment = (cr_no_alignment & ~mask) | set;
207 cr_alignment = (cr_alignment & ~mask) | set;
209 set_cr((get_cr() & ~mask) | set);
211 local_irq_restore(flags);
215 #else /* ifdef CONFIG_CPU_CP15 */
217 static int __init early_cachepolicy(char *p)
219 pr_warning("cachepolicy kernel parameter not supported without cp15\n");
221 early_param("cachepolicy", early_cachepolicy);
223 static int __init noalign_setup(char *__unused)
225 pr_warning("noalign kernel parameter not supported without cp15\n");
227 __setup("noalign", noalign_setup);
229 #endif /* ifdef CONFIG_CPU_CP15 / else */
231 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
232 #define PROT_PTE_S2_DEVICE PROT_PTE_DEVICE
233 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
235 static struct mem_type mem_types[] = {
236 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
237 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
239 .prot_pte_s2 = s2_policy(PROT_PTE_S2_DEVICE) |
240 s2_policy(L_PTE_S2_MT_DEV_SHARED) |
242 .prot_l1 = PMD_TYPE_TABLE,
243 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
246 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
247 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
248 .prot_l1 = PMD_TYPE_TABLE,
249 .prot_sect = PROT_SECT_DEVICE,
252 [MT_DEVICE_CACHED] = { /* ioremap_cached */
253 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
254 .prot_l1 = PMD_TYPE_TABLE,
255 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
258 [MT_DEVICE_WC] = { /* ioremap_wc */
259 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
260 .prot_l1 = PMD_TYPE_TABLE,
261 .prot_sect = PROT_SECT_DEVICE,
265 .prot_pte = PROT_PTE_DEVICE,
266 .prot_l1 = PMD_TYPE_TABLE,
267 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
271 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
272 .domain = DOMAIN_KERNEL,
274 #ifndef CONFIG_ARM_LPAE
276 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
277 .domain = DOMAIN_KERNEL,
281 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
283 .prot_l1 = PMD_TYPE_TABLE,
284 .domain = DOMAIN_USER,
286 [MT_HIGH_VECTORS] = {
287 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
288 L_PTE_USER | L_PTE_RDONLY,
289 .prot_l1 = PMD_TYPE_TABLE,
290 .domain = DOMAIN_USER,
293 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
294 .prot_l1 = PMD_TYPE_TABLE,
295 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
296 .domain = DOMAIN_KERNEL,
299 .prot_sect = PMD_TYPE_SECT,
300 .domain = DOMAIN_KERNEL,
302 [MT_MEMORY_NONCACHED] = {
303 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
305 .prot_l1 = PMD_TYPE_TABLE,
306 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
307 .domain = DOMAIN_KERNEL,
310 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
312 .prot_l1 = PMD_TYPE_TABLE,
313 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
314 .domain = DOMAIN_KERNEL,
317 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
318 .prot_l1 = PMD_TYPE_TABLE,
319 .domain = DOMAIN_KERNEL,
322 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
323 L_PTE_MT_UNCACHED | L_PTE_XN,
324 .prot_l1 = PMD_TYPE_TABLE,
325 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
326 PMD_SECT_UNCACHED | PMD_SECT_XN,
327 .domain = DOMAIN_KERNEL,
329 [MT_MEMORY_DMA_READY] = {
330 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
331 .prot_l1 = PMD_TYPE_TABLE,
332 .domain = DOMAIN_KERNEL,
336 const struct mem_type *get_mem_type(unsigned int type)
338 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
340 EXPORT_SYMBOL(get_mem_type);
343 * Adjust the PMD section entries according to the CPU in use.
345 static void __init build_mem_type_table(void)
347 struct cachepolicy *cp;
348 unsigned int cr = get_cr();
349 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
350 pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
351 int cpu_arch = cpu_architecture();
354 if (cpu_arch < CPU_ARCH_ARMv6) {
355 #if defined(CONFIG_CPU_DCACHE_DISABLE)
356 if (cachepolicy > CPOLICY_BUFFERED)
357 cachepolicy = CPOLICY_BUFFERED;
358 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
359 if (cachepolicy > CPOLICY_WRITETHROUGH)
360 cachepolicy = CPOLICY_WRITETHROUGH;
363 if (cpu_arch < CPU_ARCH_ARMv5) {
364 if (cachepolicy >= CPOLICY_WRITEALLOC)
365 cachepolicy = CPOLICY_WRITEBACK;
369 cachepolicy = CPOLICY_WRITEALLOC;
372 * Strip out features not present on earlier architectures.
373 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
374 * without extended page tables don't have the 'Shared' bit.
376 if (cpu_arch < CPU_ARCH_ARMv5)
377 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
378 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
379 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
380 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
381 mem_types[i].prot_sect &= ~PMD_SECT_S;
384 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
385 * "update-able on write" bit on ARM610). However, Xscale and
386 * Xscale3 require this bit to be cleared.
388 if (cpu_is_xscale() || cpu_is_xsc3()) {
389 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
390 mem_types[i].prot_sect &= ~PMD_BIT4;
391 mem_types[i].prot_l1 &= ~PMD_BIT4;
393 } else if (cpu_arch < CPU_ARCH_ARMv6) {
394 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
395 if (mem_types[i].prot_l1)
396 mem_types[i].prot_l1 |= PMD_BIT4;
397 if (mem_types[i].prot_sect)
398 mem_types[i].prot_sect |= PMD_BIT4;
403 * Mark the device areas according to the CPU/architecture.
405 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
406 if (!cpu_is_xsc3()) {
408 * Mark device regions on ARMv6+ as execute-never
409 * to prevent speculative instruction fetches.
411 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
412 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
413 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
414 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
416 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
418 * For ARMv7 with TEX remapping,
419 * - shared device is SXCB=1100
420 * - nonshared device is SXCB=0100
421 * - write combine device mem is SXCB=0001
422 * (Uncached Normal memory)
424 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
425 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
426 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
427 } else if (cpu_is_xsc3()) {
430 * - shared device is TEXCB=00101
431 * - nonshared device is TEXCB=01000
432 * - write combine device mem is TEXCB=00100
433 * (Inner/Outer Uncacheable in xsc3 parlance)
435 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
436 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
437 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
440 * For ARMv6 and ARMv7 without TEX remapping,
441 * - shared device is TEXCB=00001
442 * - nonshared device is TEXCB=01000
443 * - write combine device mem is TEXCB=00100
444 * (Uncached Normal in ARMv6 parlance).
446 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
447 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
448 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
452 * On others, write combining is "Uncached/Buffered"
454 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
458 * Now deal with the memory-type mappings
460 cp = &cache_policies[cachepolicy];
461 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
462 s2_pgprot = cp->pte_s2;
463 hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte;
464 s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2;
467 * ARMv6 and above have extended page tables.
469 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
470 #ifndef CONFIG_ARM_LPAE
472 * Mark cache clean areas and XIP ROM read only
473 * from SVC mode and no access from userspace.
475 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
476 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
477 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
482 * Mark memory with the "shared" attribute
485 user_pgprot |= L_PTE_SHARED;
486 kern_pgprot |= L_PTE_SHARED;
487 vecs_pgprot |= L_PTE_SHARED;
488 s2_pgprot |= L_PTE_SHARED;
489 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
490 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
491 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
492 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
493 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
494 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
495 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
496 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_S;
497 mem_types[MT_MEMORY_NONCACHED].prot_pte |= L_PTE_SHARED;
502 * Non-cacheable Normal - intended for memory areas that must
503 * not cause dirty cache line writebacks when used
505 if (cpu_arch >= CPU_ARCH_ARMv6) {
506 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
507 /* Non-cacheable Normal is XCB = 001 */
508 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
511 /* For both ARMv6 and non-TEX-remapping ARMv7 */
512 mem_types[MT_MEMORY_NONCACHED].prot_sect |=
516 mem_types[MT_MEMORY_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
519 #ifdef CONFIG_ARM_LPAE
521 * Do not generate access flag faults for the kernel mappings.
523 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
524 mem_types[i].prot_pte |= PTE_EXT_AF;
525 if (mem_types[i].prot_sect)
526 mem_types[i].prot_sect |= PMD_SECT_AF;
528 kern_pgprot |= PTE_EXT_AF;
529 vecs_pgprot |= PTE_EXT_AF;
532 for (i = 0; i < 16; i++) {
533 pteval_t v = pgprot_val(protection_map[i]);
534 protection_map[i] = __pgprot(v | user_pgprot);
537 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
538 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
540 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
541 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
542 L_PTE_DIRTY | kern_pgprot);
543 pgprot_s2 = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
544 pgprot_s2_device = __pgprot(s2_device_pgprot);
545 pgprot_hyp_device = __pgprot(hyp_device_pgprot);
547 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
548 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
549 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
550 mem_types[MT_MEMORY].prot_pte |= kern_pgprot;
551 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
552 mem_types[MT_MEMORY_NONCACHED].prot_sect |= ecc_mask;
553 mem_types[MT_ROM].prot_sect |= cp->pmd;
557 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
561 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
564 printk("Memory policy: ECC %sabled, Data cache %s\n",
565 ecc_mask ? "en" : "dis", cp->policy);
567 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
568 struct mem_type *t = &mem_types[i];
570 t->prot_l1 |= PMD_DOMAIN(t->domain);
572 t->prot_sect |= PMD_DOMAIN(t->domain);
576 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
577 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
578 unsigned long size, pgprot_t vma_prot)
581 return pgprot_noncached(vma_prot);
582 else if (file->f_flags & O_SYNC)
583 return pgprot_writecombine(vma_prot);
586 EXPORT_SYMBOL(phys_mem_access_prot);
589 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
591 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
593 void *ptr = __va(memblock_alloc(sz, align));
598 static void __init *early_alloc(unsigned long sz)
600 return early_alloc_aligned(sz, sz);
603 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
605 if (pmd_none(*pmd)) {
606 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
607 __pmd_populate(pmd, __pa(pte), prot);
609 BUG_ON(pmd_bad(*pmd));
610 return pte_offset_kernel(pmd, addr);
613 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
614 unsigned long end, unsigned long pfn,
615 const struct mem_type *type)
617 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
619 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
621 } while (pte++, addr += PAGE_SIZE, addr != end);
624 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
625 unsigned long end, phys_addr_t phys,
626 const struct mem_type *type)
630 #ifndef CONFIG_ARM_LPAE
632 * In classic MMU format, puds and pmds are folded in to
633 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
634 * group of L1 entries making up one logical pointer to
635 * an L2 table (2MB), where as PMDs refer to the individual
636 * L1 entries (1MB). Hence increment to get the correct
637 * offset for odd 1MB sections.
638 * (See arch/arm/include/asm/pgtable-2level.h)
640 if (addr & SECTION_SIZE)
644 *pmd = __pmd(phys | type->prot_sect);
645 phys += SECTION_SIZE;
646 } while (pmd++, addr += SECTION_SIZE, addr != end);
651 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
652 unsigned long end, phys_addr_t phys,
653 const struct mem_type *type)
655 pmd_t *pmd = pmd_offset(pud, addr);
660 * With LPAE, we must loop over to map
661 * all the pmds for the given range.
663 next = pmd_addr_end(addr, end);
666 * Try a section mapping - addr, next and phys must all be
667 * aligned to a section boundary.
669 if (type->prot_sect &&
670 ((addr | next | phys) & ~SECTION_MASK) == 0) {
671 __map_init_section(pmd, addr, next, phys, type);
673 alloc_init_pte(pmd, addr, next,
674 __phys_to_pfn(phys), type);
679 } while (pmd++, addr = next, addr != end);
682 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
683 unsigned long end, phys_addr_t phys,
684 const struct mem_type *type)
686 pud_t *pud = pud_offset(pgd, addr);
690 next = pud_addr_end(addr, end);
691 alloc_init_pmd(pud, addr, next, phys, type);
693 } while (pud++, addr = next, addr != end);
696 #ifndef CONFIG_ARM_LPAE
697 static void __init create_36bit_mapping(struct map_desc *md,
698 const struct mem_type *type)
700 unsigned long addr, length, end;
705 phys = __pfn_to_phys(md->pfn);
706 length = PAGE_ALIGN(md->length);
708 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
709 printk(KERN_ERR "MM: CPU does not support supersection "
710 "mapping for 0x%08llx at 0x%08lx\n",
711 (long long)__pfn_to_phys((u64)md->pfn), addr);
715 /* N.B. ARMv6 supersections are only defined to work with domain 0.
716 * Since domain assignments can in fact be arbitrary, the
717 * 'domain == 0' check below is required to insure that ARMv6
718 * supersections are only allocated for domain 0 regardless
719 * of the actual domain assignments in use.
722 printk(KERN_ERR "MM: invalid domain in supersection "
723 "mapping for 0x%08llx at 0x%08lx\n",
724 (long long)__pfn_to_phys((u64)md->pfn), addr);
728 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
729 printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
730 " at 0x%08lx invalid alignment\n",
731 (long long)__pfn_to_phys((u64)md->pfn), addr);
736 * Shift bits [35:32] of address into bits [23:20] of PMD
739 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
741 pgd = pgd_offset_k(addr);
744 pud_t *pud = pud_offset(pgd, addr);
745 pmd_t *pmd = pmd_offset(pud, addr);
748 for (i = 0; i < 16; i++)
749 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
751 addr += SUPERSECTION_SIZE;
752 phys += SUPERSECTION_SIZE;
753 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
754 } while (addr != end);
756 #endif /* !CONFIG_ARM_LPAE */
759 * Create the page directory entries and any necessary
760 * page tables for the mapping specified by `md'. We
761 * are able to cope here with varying sizes and address
762 * offsets, and we take full advantage of sections and
765 static void __init create_mapping(struct map_desc *md)
767 unsigned long addr, length, end;
769 const struct mem_type *type;
772 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
773 printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
774 " at 0x%08lx in user region\n",
775 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
779 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
780 md->virtual >= PAGE_OFFSET &&
781 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
782 printk(KERN_WARNING "BUG: mapping for 0x%08llx"
783 " at 0x%08lx out of vmalloc space\n",
784 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
787 type = &mem_types[md->type];
789 #ifndef CONFIG_ARM_LPAE
791 * Catch 36-bit addresses
793 if (md->pfn >= 0x100000) {
794 create_36bit_mapping(md, type);
799 addr = md->virtual & PAGE_MASK;
800 phys = __pfn_to_phys(md->pfn);
801 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
803 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
804 printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
805 "be mapped using pages, ignoring.\n",
806 (long long)__pfn_to_phys(md->pfn), addr);
810 pgd = pgd_offset_k(addr);
813 unsigned long next = pgd_addr_end(addr, end);
815 alloc_init_pud(pgd, addr, next, phys, type);
819 } while (pgd++, addr != end);
823 * Create the architecture specific mappings
825 void __init iotable_init(struct map_desc *io_desc, int nr)
828 struct vm_struct *vm;
829 struct static_vm *svm;
834 svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm));
836 for (md = io_desc; nr; md++, nr--) {
840 vm->addr = (void *)(md->virtual & PAGE_MASK);
841 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
842 vm->phys_addr = __pfn_to_phys(md->pfn);
843 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
844 vm->flags |= VM_ARM_MTYPE(md->type);
845 vm->caller = iotable_init;
846 add_static_vm_early(svm++);
850 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
853 struct vm_struct *vm;
854 struct static_vm *svm;
856 svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm));
859 vm->addr = (void *)addr;
861 vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
863 add_static_vm_early(svm);
866 #ifndef CONFIG_ARM_LPAE
869 * The Linux PMD is made of two consecutive section entries covering 2MB
870 * (see definition in include/asm/pgtable-2level.h). However a call to
871 * create_mapping() may optimize static mappings by using individual
872 * 1MB section mappings. This leaves the actual PMD potentially half
873 * initialized if the top or bottom section entry isn't used, leaving it
874 * open to problems if a subsequent ioremap() or vmalloc() tries to use
875 * the virtual space left free by that unused section entry.
877 * Let's avoid the issue by inserting dummy vm entries covering the unused
878 * PMD halves once the static mappings are in place.
881 static void __init pmd_empty_section_gap(unsigned long addr)
883 vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
886 static void __init fill_pmd_gaps(void)
888 struct static_vm *svm;
889 struct vm_struct *vm;
890 unsigned long addr, next = 0;
893 list_for_each_entry(svm, &static_vmlist, list) {
895 addr = (unsigned long)vm->addr;
900 * Check if this vm starts on an odd section boundary.
901 * If so and the first section entry for this PMD is free
902 * then we block the corresponding virtual address.
904 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
905 pmd = pmd_off_k(addr);
907 pmd_empty_section_gap(addr & PMD_MASK);
911 * Then check if this vm ends on an odd section boundary.
912 * If so and the second section entry for this PMD is empty
913 * then we block the corresponding virtual address.
916 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
917 pmd = pmd_off_k(addr) + 1;
919 pmd_empty_section_gap(addr);
922 /* no need to look at any vm entry until we hit the next PMD */
923 next = (addr + PMD_SIZE - 1) & PMD_MASK;
928 #define fill_pmd_gaps() do { } while (0)
931 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
932 static void __init pci_reserve_io(void)
934 struct static_vm *svm;
936 svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
940 vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
943 #define pci_reserve_io() do { } while (0)
946 #ifdef CONFIG_DEBUG_LL
947 void __init debug_ll_io_init(void)
951 debug_ll_addr(&map.pfn, &map.virtual);
952 if (!map.pfn || !map.virtual)
954 map.pfn = __phys_to_pfn(map.pfn);
955 map.virtual &= PAGE_MASK;
956 map.length = PAGE_SIZE;
957 map.type = MT_DEVICE;
958 iotable_init(&map, 1);
962 static void * __initdata vmalloc_min =
963 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
966 * vmalloc=size forces the vmalloc area to be exactly 'size'
967 * bytes. This can be used to increase (or decrease) the vmalloc
968 * area - the default is 240m.
970 static int __init early_vmalloc(char *arg)
972 unsigned long vmalloc_reserve = memparse(arg, NULL);
974 if (vmalloc_reserve < SZ_16M) {
975 vmalloc_reserve = SZ_16M;
977 "vmalloc area too small, limiting to %luMB\n",
978 vmalloc_reserve >> 20);
981 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
982 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
984 "vmalloc area is too big, limiting to %luMB\n",
985 vmalloc_reserve >> 20);
988 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
991 early_param("vmalloc", early_vmalloc);
993 phys_addr_t arm_lowmem_limit __initdata = 0;
995 void __init sanity_check_meminfo(void)
997 phys_addr_t memblock_limit = 0;
998 int i, j, highmem = 0;
999 phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1;
1001 for (i = 0, j = 0; i < meminfo.nr_banks; i++) {
1002 struct membank *bank = &meminfo.bank[j];
1003 phys_addr_t size_limit;
1005 *bank = meminfo.bank[i];
1006 size_limit = bank->size;
1008 if (bank->start >= vmalloc_limit)
1011 size_limit = vmalloc_limit - bank->start;
1013 bank->highmem = highmem;
1015 #ifdef CONFIG_HIGHMEM
1017 * Split those memory banks which are partially overlapping
1018 * the vmalloc area greatly simplifying things later.
1020 if (!highmem && bank->size > size_limit) {
1021 if (meminfo.nr_banks >= NR_BANKS) {
1022 printk(KERN_CRIT "NR_BANKS too low, "
1023 "ignoring high memory\n");
1025 memmove(bank + 1, bank,
1026 (meminfo.nr_banks - i) * sizeof(*bank));
1029 bank[1].size -= size_limit;
1030 bank[1].start = vmalloc_limit;
1031 bank[1].highmem = highmem = 1;
1034 bank->size = size_limit;
1038 * Highmem banks not allowed with !CONFIG_HIGHMEM.
1041 printk(KERN_NOTICE "Ignoring RAM at %.8llx-%.8llx "
1042 "(!CONFIG_HIGHMEM).\n",
1043 (unsigned long long)bank->start,
1044 (unsigned long long)bank->start + bank->size - 1);
1049 * Check whether this memory bank would partially overlap
1052 if (bank->size > size_limit) {
1053 printk(KERN_NOTICE "Truncating RAM at %.8llx-%.8llx "
1054 "to -%.8llx (vmalloc region overlap).\n",
1055 (unsigned long long)bank->start,
1056 (unsigned long long)bank->start + bank->size - 1,
1057 (unsigned long long)bank->start + size_limit - 1);
1058 bank->size = size_limit;
1061 if (!bank->highmem) {
1062 phys_addr_t bank_end = bank->start + bank->size;
1064 if (bank_end > arm_lowmem_limit)
1065 arm_lowmem_limit = bank_end;
1068 * Find the first non-section-aligned page, and point
1069 * memblock_limit at it. This relies on rounding the
1070 * limit down to be section-aligned, which happens at
1071 * the end of this function.
1073 * With this algorithm, the start or end of almost any
1074 * bank can be non-section-aligned. The only exception
1075 * is that the start of the bank 0 must be section-
1076 * aligned, since otherwise memory would need to be
1077 * allocated when mapping the start of bank 0, which
1078 * occurs before any free memory is mapped.
1080 if (!memblock_limit) {
1081 if (!IS_ALIGNED(bank->start, SECTION_SIZE))
1082 memblock_limit = bank->start;
1083 else if (!IS_ALIGNED(bank_end, SECTION_SIZE))
1084 memblock_limit = bank_end;
1089 #ifdef CONFIG_HIGHMEM
1091 const char *reason = NULL;
1093 if (cache_is_vipt_aliasing()) {
1095 * Interactions between kmap and other mappings
1096 * make highmem support with aliasing VIPT caches
1099 reason = "with VIPT aliasing cache";
1102 printk(KERN_CRIT "HIGHMEM is not supported %s, ignoring high memory\n",
1104 while (j > 0 && meminfo.bank[j - 1].highmem)
1109 meminfo.nr_banks = j;
1110 high_memory = __va(arm_lowmem_limit - 1) + 1;
1113 * Round the memblock limit down to a section size. This
1114 * helps to ensure that we will allocate memory from the
1115 * last full section, which should be mapped.
1118 memblock_limit = round_down(memblock_limit, SECTION_SIZE);
1119 if (!memblock_limit)
1120 memblock_limit = arm_lowmem_limit;
1122 memblock_set_current_limit(memblock_limit);
1125 static inline void prepare_page_table(void)
1131 * Clear out all the mappings below the kernel image.
1133 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1134 pmd_clear(pmd_off_k(addr));
1136 #ifdef CONFIG_XIP_KERNEL
1137 /* The XIP kernel is mapped in the module area -- skip over it */
1138 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1140 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1141 pmd_clear(pmd_off_k(addr));
1144 * Find the end of the first block of lowmem.
1146 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1147 if (end >= arm_lowmem_limit)
1148 end = arm_lowmem_limit;
1151 * Clear out all the kernel space mappings, except for the first
1152 * memory bank, up to the vmalloc region.
1154 for (addr = __phys_to_virt(end);
1155 addr < VMALLOC_START; addr += PMD_SIZE)
1156 pmd_clear(pmd_off_k(addr));
1159 #ifdef CONFIG_ARM_LPAE
1160 /* the first page is reserved for pgd */
1161 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1162 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1164 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1168 * Reserve the special regions of memory
1170 void __init arm_mm_memblock_reserve(void)
1173 * Reserve the page tables. These are already in use,
1174 * and can only be in node 0.
1176 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1178 #ifdef CONFIG_SA1111
1180 * Because of the SA1111 DMA bug, we want to preserve our
1181 * precious DMA-able memory...
1183 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1188 * Set up the device mappings. Since we clear out the page tables for all
1189 * mappings above VMALLOC_START, we will remove any debug device mappings.
1190 * This means you have to be careful how you debug this function, or any
1191 * called function. This means you can't use any function or debugging
1192 * method which may touch any device, otherwise the kernel _will_ crash.
1194 static void __init devicemaps_init(const struct machine_desc *mdesc)
1196 struct map_desc map;
1201 * Allocate the vector page early.
1203 vectors = early_alloc(PAGE_SIZE * 2);
1205 early_trap_init(vectors);
1207 for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1208 pmd_clear(pmd_off_k(addr));
1211 * Map the kernel if it is XIP.
1212 * It is always first in the modulearea.
1214 #ifdef CONFIG_XIP_KERNEL
1215 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1216 map.virtual = MODULES_VADDR;
1217 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1219 create_mapping(&map);
1223 * Map the cache flushing regions.
1226 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1227 map.virtual = FLUSH_BASE;
1229 map.type = MT_CACHECLEAN;
1230 create_mapping(&map);
1232 #ifdef FLUSH_BASE_MINICACHE
1233 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1234 map.virtual = FLUSH_BASE_MINICACHE;
1236 map.type = MT_MINICLEAN;
1237 create_mapping(&map);
1241 * Create a mapping for the machine vectors at the high-vectors
1242 * location (0xffff0000). If we aren't using high-vectors, also
1243 * create a mapping at the low-vectors virtual address.
1245 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1246 map.virtual = 0xffff0000;
1247 map.length = PAGE_SIZE;
1248 #ifdef CONFIG_KUSER_HELPERS
1249 map.type = MT_HIGH_VECTORS;
1251 map.type = MT_LOW_VECTORS;
1253 create_mapping(&map);
1255 if (!vectors_high()) {
1257 map.length = PAGE_SIZE * 2;
1258 map.type = MT_LOW_VECTORS;
1259 create_mapping(&map);
1262 /* Now create a kernel read-only mapping */
1264 map.virtual = 0xffff0000 + PAGE_SIZE;
1265 map.length = PAGE_SIZE;
1266 map.type = MT_LOW_VECTORS;
1267 create_mapping(&map);
1270 * Ask the machine support to map in the statically mapped devices.
1278 /* Reserve fixed i/o space in VMALLOC region */
1282 * Finally flush the caches and tlb to ensure that we're in a
1283 * consistent state wrt the writebuffer. This also ensures that
1284 * any write-allocated cache lines in the vector page are written
1285 * back. After this point, we can start to touch devices again.
1287 local_flush_tlb_all();
1291 static void __init kmap_init(void)
1293 #ifdef CONFIG_HIGHMEM
1294 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1295 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1299 static void __init map_lowmem(void)
1301 struct memblock_region *reg;
1303 /* Map all the lowmem memory banks. */
1304 for_each_memblock(memory, reg) {
1305 phys_addr_t start = reg->base;
1306 phys_addr_t end = start + reg->size;
1307 struct map_desc map;
1309 if (end > arm_lowmem_limit)
1310 end = arm_lowmem_limit;
1314 map.pfn = __phys_to_pfn(start);
1315 map.virtual = __phys_to_virt(start);
1316 map.length = end - start;
1317 map.type = MT_MEMORY;
1319 create_mapping(&map);
1324 * paging_init() sets up the page tables, initialises the zone memory
1325 * maps, and sets up the zero page, bad page and bad page tables.
1327 void __init paging_init(const struct machine_desc *mdesc)
1331 build_mem_type_table();
1332 prepare_page_table();
1334 dma_contiguous_remap();
1335 devicemaps_init(mdesc);
1339 top_pmd = pmd_off_k(0xffff0000);
1341 /* allocate the zero page. */
1342 zero_page = early_alloc(PAGE_SIZE);
1346 empty_zero_page = virt_to_page(zero_page);
1347 __flush_dcache_page(NULL, empty_zero_page);