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/bootmem.h>
15 #include <linux/mman.h>
16 #include <linux/nodemask.h>
18 #include <asm/mach-types.h>
19 #include <asm/setup.h>
20 #include <asm/sizes.h>
23 #include <asm/mach/arch.h>
24 #include <asm/mach/map.h>
28 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
30 extern void _stext, _etext, __data_start, _end;
31 extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
34 * empty_zero_page is a special page that is used for
35 * zero-initialized data and COW.
37 struct page *empty_zero_page;
40 * The pmd table for the upper-most set of pages.
44 #define CPOLICY_UNCACHED 0
45 #define CPOLICY_BUFFERED 1
46 #define CPOLICY_WRITETHROUGH 2
47 #define CPOLICY_WRITEBACK 3
48 #define CPOLICY_WRITEALLOC 4
50 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
51 static unsigned int ecc_mask __initdata = 0;
52 pgprot_t pgprot_kernel;
54 EXPORT_SYMBOL(pgprot_kernel);
57 const char policy[16];
63 static struct cachepolicy cache_policies[] __initdata = {
67 .pmd = PMD_SECT_UNCACHED,
72 .pmd = PMD_SECT_BUFFERED,
73 .pte = PTE_BUFFERABLE,
75 .policy = "writethrough",
80 .policy = "writeback",
83 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
85 .policy = "writealloc",
88 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
93 * These are useful for identifing cache coherency
94 * problems by allowing the cache or the cache and
95 * writebuffer to be turned off. (Note: the write
96 * buffer should not be on and the cache off).
98 static void __init early_cachepolicy(char **p)
102 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
103 int len = strlen(cache_policies[i].policy);
105 if (memcmp(*p, cache_policies[i].policy, len) == 0) {
107 cr_alignment &= ~cache_policies[i].cr_mask;
108 cr_no_alignment &= ~cache_policies[i].cr_mask;
113 if (i == ARRAY_SIZE(cache_policies))
114 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
116 set_cr(cr_alignment);
118 __early_param("cachepolicy=", early_cachepolicy);
120 static void __init early_nocache(char **__unused)
122 char *p = "buffered";
123 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
124 early_cachepolicy(&p);
126 __early_param("nocache", early_nocache);
128 static void __init early_nowrite(char **__unused)
130 char *p = "uncached";
131 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
132 early_cachepolicy(&p);
134 __early_param("nowb", early_nowrite);
136 static void __init early_ecc(char **p)
138 if (memcmp(*p, "on", 2) == 0) {
139 ecc_mask = PMD_PROTECTION;
141 } else if (memcmp(*p, "off", 3) == 0) {
146 __early_param("ecc=", early_ecc);
148 static int __init noalign_setup(char *__unused)
150 cr_alignment &= ~CR_A;
151 cr_no_alignment &= ~CR_A;
152 set_cr(cr_alignment);
155 __setup("noalign", noalign_setup);
158 unsigned int prot_pte;
159 unsigned int prot_l1;
160 unsigned int prot_sect;
164 static struct mem_types mem_types[] __initdata = {
166 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
168 .prot_l1 = PMD_TYPE_TABLE,
169 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
174 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
175 .domain = DOMAIN_KERNEL,
178 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_MINICACHE,
179 .domain = DOMAIN_KERNEL,
182 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
184 .prot_l1 = PMD_TYPE_TABLE,
185 .domain = DOMAIN_USER,
187 [MT_HIGH_VECTORS] = {
188 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
189 L_PTE_USER | L_PTE_EXEC,
190 .prot_l1 = PMD_TYPE_TABLE,
191 .domain = DOMAIN_USER,
194 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_AP_WRITE,
195 .domain = DOMAIN_KERNEL,
198 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
199 .domain = DOMAIN_KERNEL,
201 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
202 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
204 .prot_l1 = PMD_TYPE_TABLE,
205 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
206 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
210 [MT_NONSHARED_DEVICE] = {
211 .prot_l1 = PMD_TYPE_TABLE,
212 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_NONSHARED_DEV |
219 * Adjust the PMD section entries according to the CPU in use.
221 static void __init build_mem_type_table(void)
223 struct cachepolicy *cp;
224 unsigned int cr = get_cr();
225 unsigned int user_pgprot, kern_pgprot;
226 int cpu_arch = cpu_architecture();
229 #if defined(CONFIG_CPU_DCACHE_DISABLE)
230 if (cachepolicy > CPOLICY_BUFFERED)
231 cachepolicy = CPOLICY_BUFFERED;
232 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
233 if (cachepolicy > CPOLICY_WRITETHROUGH)
234 cachepolicy = CPOLICY_WRITETHROUGH;
236 if (cpu_arch < CPU_ARCH_ARMv5) {
237 if (cachepolicy >= CPOLICY_WRITEALLOC)
238 cachepolicy = CPOLICY_WRITEBACK;
243 * Xscale must not have PMD bit 4 set for section mappings.
246 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
247 mem_types[i].prot_sect &= ~PMD_BIT4;
250 * ARMv5 and lower, excluding Xscale, bit 4 must be set for
253 if (cpu_arch < CPU_ARCH_ARMv6 && !cpu_is_xscale())
254 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
255 if (mem_types[i].prot_l1)
256 mem_types[i].prot_l1 |= PMD_BIT4;
258 cp = &cache_policies[cachepolicy];
259 kern_pgprot = user_pgprot = cp->pte;
262 * Enable CPU-specific coherency if supported.
263 * (Only available on XSC3 at the moment.)
265 if (arch_is_coherent()) {
267 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
268 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
273 * ARMv6 and above have extended page tables.
275 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
277 * bit 4 becomes XN which we must clear for the
278 * kernel memory mapping.
280 mem_types[MT_MEMORY].prot_sect &= ~PMD_SECT_XN;
281 mem_types[MT_ROM].prot_sect &= ~PMD_SECT_XN;
284 * Mark cache clean areas and XIP ROM read only
285 * from SVC mode and no access from userspace.
287 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
288 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
289 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
292 * Mark the device area as "shared device"
294 mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
295 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
298 * User pages need to be mapped with the ASID
301 user_pgprot |= L_PTE_ASID;
305 * Mark memory with the "shared" attribute for SMP systems
307 user_pgprot |= L_PTE_SHARED;
308 kern_pgprot |= L_PTE_SHARED;
309 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
313 for (i = 0; i < 16; i++) {
314 unsigned long v = pgprot_val(protection_map[i]);
315 v = (v & ~(L_PTE_BUFFERABLE|L_PTE_CACHEABLE)) | user_pgprot;
316 protection_map[i] = __pgprot(v);
319 mem_types[MT_LOW_VECTORS].prot_pte |= kern_pgprot;
320 mem_types[MT_HIGH_VECTORS].prot_pte |= kern_pgprot;
322 if (cpu_arch >= CPU_ARCH_ARMv5) {
325 * Only use write-through for non-SMP systems
327 mem_types[MT_LOW_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
328 mem_types[MT_HIGH_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
331 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
334 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
335 L_PTE_DIRTY | L_PTE_WRITE |
336 L_PTE_EXEC | kern_pgprot);
338 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
339 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
340 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
341 mem_types[MT_ROM].prot_sect |= cp->pmd;
345 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
349 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
352 printk("Memory policy: ECC %sabled, Data cache %s\n",
353 ecc_mask ? "en" : "dis", cp->policy);
356 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
359 * Create a SECTION PGD between VIRT and PHYS in domain
360 * DOMAIN with protection PROT. This operates on half-
361 * pgdir entry increments.
364 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
366 pmd_t *pmdp = pmd_off_k(virt);
368 if (virt & (1 << 20))
371 *pmdp = __pmd(phys | prot);
372 flush_pmd_entry(pmdp);
376 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
379 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
383 for (i = 0; i < 16; i += 1) {
384 alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
386 virt += (PGDIR_SIZE / 2);
391 * Add a PAGE mapping between VIRT and PHYS in domain
392 * DOMAIN with protection PROT. Note that due to the
393 * way we map the PTEs, we must allocate two PTE_SIZE'd
394 * blocks - one for the Linux pte table, and one for
395 * the hardware pte table.
398 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
400 pmd_t *pmdp = pmd_off_k(virt);
403 if (pmd_none(*pmdp)) {
404 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
407 __pmd_populate(pmdp, __pa(ptep) | prot_l1);
409 ptep = pte_offset_kernel(pmdp, virt);
411 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
415 * Create the page directory entries and any necessary
416 * page tables for the mapping specified by `md'. We
417 * are able to cope here with varying sizes and address
418 * offsets, and we take full advantage of sections and
421 void __init create_mapping(struct map_desc *md)
423 unsigned long virt, length;
424 int prot_sect, prot_l1, domain;
426 unsigned long off = (u32)__pfn_to_phys(md->pfn);
428 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
429 printk(KERN_WARNING "BUG: not creating mapping for "
430 "0x%08llx at 0x%08lx in user region\n",
431 __pfn_to_phys((u64)md->pfn), md->virtual);
435 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
436 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
437 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
438 "overlaps vmalloc space\n",
439 __pfn_to_phys((u64)md->pfn), md->virtual);
442 domain = mem_types[md->type].domain;
443 prot_pte = __pgprot(mem_types[md->type].prot_pte);
444 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
445 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
448 * Catch 36-bit addresses
450 if(md->pfn >= 0x100000) {
452 printk(KERN_ERR "MM: invalid domain in supersection "
453 "mapping for 0x%08llx at 0x%08lx\n",
454 __pfn_to_phys((u64)md->pfn), md->virtual);
457 if((md->virtual | md->length | __pfn_to_phys(md->pfn))
458 & ~SUPERSECTION_MASK) {
459 printk(KERN_ERR "MM: cannot create mapping for "
460 "0x%08llx at 0x%08lx invalid alignment\n",
461 __pfn_to_phys((u64)md->pfn), md->virtual);
466 * Shift bits [35:32] of address into bits [23:20] of PMD
469 off |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
476 if (mem_types[md->type].prot_l1 == 0 &&
477 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
478 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
479 "be mapped using pages, ignoring.\n",
480 __pfn_to_phys(md->pfn), md->virtual);
484 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
485 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
491 /* N.B. ARMv6 supersections are only defined to work with domain 0.
492 * Since domain assignments can in fact be arbitrary, the
493 * 'domain == 0' check below is required to insure that ARMv6
494 * supersections are only allocated for domain 0 regardless
495 * of the actual domain assignments in use.
497 if ((cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())
500 * Align to supersection boundary if !high pages.
501 * High pages have already been checked for proper
502 * alignment above and they will fail the SUPSERSECTION_MASK
503 * check because of the way the address is encoded into
506 if (md->pfn <= 0x100000) {
507 while ((virt & ~SUPERSECTION_MASK ||
508 (virt + off) & ~SUPERSECTION_MASK) &&
509 length >= (PGDIR_SIZE / 2)) {
510 alloc_init_section(virt, virt + off, prot_sect);
512 virt += (PGDIR_SIZE / 2);
513 length -= (PGDIR_SIZE / 2);
517 while (length >= SUPERSECTION_SIZE) {
518 alloc_init_supersection(virt, virt + off, prot_sect);
520 virt += SUPERSECTION_SIZE;
521 length -= SUPERSECTION_SIZE;
526 * A section mapping covers half a "pgdir" entry.
528 while (length >= (PGDIR_SIZE / 2)) {
529 alloc_init_section(virt, virt + off, prot_sect);
531 virt += (PGDIR_SIZE / 2);
532 length -= (PGDIR_SIZE / 2);
535 while (length >= PAGE_SIZE) {
536 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
544 * Create the architecture specific mappings
546 void __init iotable_init(struct map_desc *io_desc, int nr)
550 for (i = 0; i < nr; i++)
551 create_mapping(io_desc + i);
554 static inline void prepare_page_table(struct meminfo *mi)
559 * Clear out all the mappings below the kernel image.
561 for (addr = 0; addr < MODULE_START; addr += PGDIR_SIZE)
562 pmd_clear(pmd_off_k(addr));
564 #ifdef CONFIG_XIP_KERNEL
565 /* The XIP kernel is mapped in the module area -- skip over it */
566 addr = ((unsigned long)&_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
568 for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
569 pmd_clear(pmd_off_k(addr));
572 * Clear out all the kernel space mappings, except for the first
573 * memory bank, up to the end of the vmalloc region.
575 for (addr = __phys_to_virt(mi->bank[0].start + mi->bank[0].size);
576 addr < VMALLOC_END; addr += PGDIR_SIZE)
577 pmd_clear(pmd_off_k(addr));
581 * Reserve the various regions of node 0
583 void __init reserve_node_zero(pg_data_t *pgdat)
585 unsigned long res_size = 0;
588 * Register the kernel text and data with bootmem.
589 * Note that this can only be in node 0.
591 #ifdef CONFIG_XIP_KERNEL
592 reserve_bootmem_node(pgdat, __pa(&__data_start), &_end - &__data_start);
594 reserve_bootmem_node(pgdat, __pa(&_stext), &_end - &_stext);
598 * Reserve the page tables. These are already in use,
599 * and can only be in node 0.
601 reserve_bootmem_node(pgdat, __pa(swapper_pg_dir),
602 PTRS_PER_PGD * sizeof(pgd_t));
605 * Hmm... This should go elsewhere, but we really really need to
606 * stop things allocating the low memory; ideally we need a better
607 * implementation of GFP_DMA which does not assume that DMA-able
608 * memory starts at zero.
610 if (machine_is_integrator() || machine_is_cintegrator())
611 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
614 * These should likewise go elsewhere. They pre-reserve the
615 * screen memory region at the start of main system memory.
617 if (machine_is_edb7211())
618 res_size = 0x00020000;
619 if (machine_is_p720t())
620 res_size = 0x00014000;
622 /* H1940 and RX3715 need to reserve this for suspend */
624 if (machine_is_h1940() || machine_is_rx3715()) {
625 reserve_bootmem_node(pgdat, 0x30003000, 0x1000);
626 reserve_bootmem_node(pgdat, 0x30081000, 0x1000);
631 * Because of the SA1111 DMA bug, we want to preserve our
632 * precious DMA-able memory...
634 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
637 reserve_bootmem_node(pgdat, PHYS_OFFSET, res_size);
641 * Set up device the mappings. Since we clear out the page tables for all
642 * mappings above VMALLOC_END, we will remove any debug device mappings.
643 * This means you have to be careful how you debug this function, or any
644 * called function. This means you can't use any function or debugging
645 * method which may touch any device, otherwise the kernel _will_ crash.
647 static void __init devicemaps_init(struct machine_desc *mdesc)
654 * Allocate the vector page early.
656 vectors = alloc_bootmem_low_pages(PAGE_SIZE);
659 for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
660 pmd_clear(pmd_off_k(addr));
663 * Map the kernel if it is XIP.
664 * It is always first in the modulearea.
666 #ifdef CONFIG_XIP_KERNEL
667 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
668 map.virtual = MODULE_START;
669 map.length = ((unsigned long)&_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
671 create_mapping(&map);
675 * Map the cache flushing regions.
678 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
679 map.virtual = FLUSH_BASE;
681 map.type = MT_CACHECLEAN;
682 create_mapping(&map);
684 #ifdef FLUSH_BASE_MINICACHE
685 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
686 map.virtual = FLUSH_BASE_MINICACHE;
688 map.type = MT_MINICLEAN;
689 create_mapping(&map);
693 * Create a mapping for the machine vectors at the high-vectors
694 * location (0xffff0000). If we aren't using high-vectors, also
695 * create a mapping at the low-vectors virtual address.
697 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
698 map.virtual = 0xffff0000;
699 map.length = PAGE_SIZE;
700 map.type = MT_HIGH_VECTORS;
701 create_mapping(&map);
703 if (!vectors_high()) {
705 map.type = MT_LOW_VECTORS;
706 create_mapping(&map);
710 * Ask the machine support to map in the statically mapped devices.
716 * Finally flush the caches and tlb to ensure that we're in a
717 * consistent state wrt the writebuffer. This also ensures that
718 * any write-allocated cache lines in the vector page are written
719 * back. After this point, we can start to touch devices again.
721 local_flush_tlb_all();
726 * paging_init() sets up the page tables, initialises the zone memory
727 * maps, and sets up the zero page, bad page and bad page tables.
729 void __init paging_init(struct meminfo *mi, struct machine_desc *mdesc)
733 build_mem_type_table();
734 prepare_page_table(mi);
736 devicemaps_init(mdesc);
738 top_pmd = pmd_off_k(0xffff0000);
741 * allocate the zero page. Note that we count on this going ok.
743 zero_page = alloc_bootmem_low_pages(PAGE_SIZE);
744 memzero(zero_page, PAGE_SIZE);
745 empty_zero_page = virt_to_page(zero_page);
746 flush_dcache_page(empty_zero_page);
750 * In order to soft-boot, we need to insert a 1:1 mapping in place of
751 * the user-mode pages. This will then ensure that we have predictable
752 * results when turning the mmu off
754 void setup_mm_for_reboot(char mode)
756 unsigned long base_pmdval;
760 if (current->mm && current->mm->pgd)
761 pgd = current->mm->pgd;
765 base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
766 if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
767 base_pmdval |= PMD_BIT4;
769 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
770 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
773 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
774 pmd[0] = __pmd(pmdval);
775 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
776 flush_pmd_entry(pmd);