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 void adjust_cr(unsigned long mask, unsigned long set)
166 local_irq_save(flags);
168 cr_no_alignment = (cr_no_alignment & ~mask) | set;
169 cr_alignment = (cr_alignment & ~mask) | set;
171 set_cr((get_cr() & ~mask) | set);
173 local_irq_restore(flags);
178 unsigned int prot_pte;
179 unsigned int prot_l1;
180 unsigned int prot_sect;
184 static struct mem_types mem_types[] __initdata = {
186 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
188 .prot_l1 = PMD_TYPE_TABLE,
189 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
194 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
195 .domain = DOMAIN_KERNEL,
198 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_MINICACHE,
199 .domain = DOMAIN_KERNEL,
202 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
204 .prot_l1 = PMD_TYPE_TABLE,
205 .domain = DOMAIN_USER,
207 [MT_HIGH_VECTORS] = {
208 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
209 L_PTE_USER | L_PTE_EXEC,
210 .prot_l1 = PMD_TYPE_TABLE,
211 .domain = DOMAIN_USER,
214 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_AP_WRITE,
215 .domain = DOMAIN_KERNEL,
218 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
219 .domain = DOMAIN_KERNEL,
221 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
222 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
224 .prot_l1 = PMD_TYPE_TABLE,
225 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
226 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
230 [MT_NONSHARED_DEVICE] = {
231 .prot_l1 = PMD_TYPE_TABLE,
232 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_NONSHARED_DEV |
239 * Adjust the PMD section entries according to the CPU in use.
241 static void __init build_mem_type_table(void)
243 struct cachepolicy *cp;
244 unsigned int cr = get_cr();
245 unsigned int user_pgprot, kern_pgprot;
246 int cpu_arch = cpu_architecture();
249 #if defined(CONFIG_CPU_DCACHE_DISABLE)
250 if (cachepolicy > CPOLICY_BUFFERED)
251 cachepolicy = CPOLICY_BUFFERED;
252 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
253 if (cachepolicy > CPOLICY_WRITETHROUGH)
254 cachepolicy = CPOLICY_WRITETHROUGH;
256 if (cpu_arch < CPU_ARCH_ARMv5) {
257 if (cachepolicy >= CPOLICY_WRITEALLOC)
258 cachepolicy = CPOLICY_WRITEBACK;
263 * Xscale must not have PMD bit 4 set for section mappings.
266 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
267 mem_types[i].prot_sect &= ~PMD_BIT4;
270 * ARMv5 and lower, excluding Xscale, bit 4 must be set for
273 if (cpu_arch < CPU_ARCH_ARMv6 && !cpu_is_xscale())
274 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
275 if (mem_types[i].prot_l1)
276 mem_types[i].prot_l1 |= PMD_BIT4;
278 cp = &cache_policies[cachepolicy];
279 kern_pgprot = user_pgprot = cp->pte;
282 * Enable CPU-specific coherency if supported.
283 * (Only available on XSC3 at the moment.)
285 if (arch_is_coherent()) {
287 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
288 mem_types[MT_MEMORY].prot_pte |= L_PTE_SHARED;
293 * ARMv6 and above have extended page tables.
295 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
297 * bit 4 becomes XN which we must clear for the
298 * kernel memory mapping.
300 mem_types[MT_MEMORY].prot_sect &= ~PMD_SECT_XN;
301 mem_types[MT_ROM].prot_sect &= ~PMD_SECT_XN;
304 * Mark cache clean areas and XIP ROM read only
305 * from SVC mode and no access from userspace.
307 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
308 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
309 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
312 * Mark the device area as "shared device"
314 mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
315 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
319 * Mark memory with the "shared" attribute for SMP systems
321 user_pgprot |= L_PTE_SHARED;
322 kern_pgprot |= L_PTE_SHARED;
323 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
327 for (i = 0; i < 16; i++) {
328 unsigned long v = pgprot_val(protection_map[i]);
329 v = (v & ~(L_PTE_BUFFERABLE|L_PTE_CACHEABLE)) | user_pgprot;
330 protection_map[i] = __pgprot(v);
333 mem_types[MT_LOW_VECTORS].prot_pte |= kern_pgprot;
334 mem_types[MT_HIGH_VECTORS].prot_pte |= kern_pgprot;
336 if (cpu_arch >= CPU_ARCH_ARMv5) {
339 * Only use write-through for non-SMP systems
341 mem_types[MT_LOW_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
342 mem_types[MT_HIGH_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
345 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
348 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
349 L_PTE_DIRTY | L_PTE_WRITE |
350 L_PTE_EXEC | kern_pgprot);
352 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
353 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
354 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
355 mem_types[MT_ROM].prot_sect |= cp->pmd;
359 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
363 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
366 printk("Memory policy: ECC %sabled, Data cache %s\n",
367 ecc_mask ? "en" : "dis", cp->policy);
370 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
373 * Create a SECTION PGD between VIRT and PHYS in domain
374 * DOMAIN with protection PROT. This operates on half-
375 * pgdir entry increments.
378 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
380 pmd_t *pmdp = pmd_off_k(virt);
382 if (virt & (1 << 20))
385 *pmdp = __pmd(phys | prot);
386 flush_pmd_entry(pmdp);
390 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
393 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
397 for (i = 0; i < 16; i += 1) {
398 alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
400 virt += (PGDIR_SIZE / 2);
405 * Add a PAGE mapping between VIRT and PHYS in domain
406 * DOMAIN with protection PROT. Note that due to the
407 * way we map the PTEs, we must allocate two PTE_SIZE'd
408 * blocks - one for the Linux pte table, and one for
409 * the hardware pte table.
412 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
414 pmd_t *pmdp = pmd_off_k(virt);
417 if (pmd_none(*pmdp)) {
418 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
421 __pmd_populate(pmdp, __pa(ptep) | prot_l1);
423 ptep = pte_offset_kernel(pmdp, virt);
425 set_pte_ext(ptep, pfn_pte(phys >> PAGE_SHIFT, prot), 0);
429 * Create the page directory entries and any necessary
430 * page tables for the mapping specified by `md'. We
431 * are able to cope here with varying sizes and address
432 * offsets, and we take full advantage of sections and
435 void __init create_mapping(struct map_desc *md)
437 unsigned long virt, length;
438 int prot_sect, prot_l1, domain;
440 unsigned long off = (u32)__pfn_to_phys(md->pfn);
442 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
443 printk(KERN_WARNING "BUG: not creating mapping for "
444 "0x%08llx at 0x%08lx in user region\n",
445 __pfn_to_phys((u64)md->pfn), md->virtual);
449 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
450 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
451 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
452 "overlaps vmalloc space\n",
453 __pfn_to_phys((u64)md->pfn), md->virtual);
456 domain = mem_types[md->type].domain;
457 prot_pte = __pgprot(mem_types[md->type].prot_pte);
458 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
459 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
462 * Catch 36-bit addresses
464 if(md->pfn >= 0x100000) {
466 printk(KERN_ERR "MM: invalid domain in supersection "
467 "mapping for 0x%08llx at 0x%08lx\n",
468 __pfn_to_phys((u64)md->pfn), md->virtual);
471 if((md->virtual | md->length | __pfn_to_phys(md->pfn))
472 & ~SUPERSECTION_MASK) {
473 printk(KERN_ERR "MM: cannot create mapping for "
474 "0x%08llx at 0x%08lx invalid alignment\n",
475 __pfn_to_phys((u64)md->pfn), md->virtual);
480 * Shift bits [35:32] of address into bits [23:20] of PMD
483 off |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
490 if (mem_types[md->type].prot_l1 == 0 &&
491 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
492 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
493 "be mapped using pages, ignoring.\n",
494 __pfn_to_phys(md->pfn), md->virtual);
498 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
499 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
505 /* N.B. ARMv6 supersections are only defined to work with domain 0.
506 * Since domain assignments can in fact be arbitrary, the
507 * 'domain == 0' check below is required to insure that ARMv6
508 * supersections are only allocated for domain 0 regardless
509 * of the actual domain assignments in use.
511 if ((cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())
514 * Align to supersection boundary if !high pages.
515 * High pages have already been checked for proper
516 * alignment above and they will fail the SUPSERSECTION_MASK
517 * check because of the way the address is encoded into
520 if (md->pfn <= 0x100000) {
521 while ((virt & ~SUPERSECTION_MASK ||
522 (virt + off) & ~SUPERSECTION_MASK) &&
523 length >= (PGDIR_SIZE / 2)) {
524 alloc_init_section(virt, virt + off, prot_sect);
526 virt += (PGDIR_SIZE / 2);
527 length -= (PGDIR_SIZE / 2);
531 while (length >= SUPERSECTION_SIZE) {
532 alloc_init_supersection(virt, virt + off, prot_sect);
534 virt += SUPERSECTION_SIZE;
535 length -= SUPERSECTION_SIZE;
540 * A section mapping covers half a "pgdir" entry.
542 while (length >= (PGDIR_SIZE / 2)) {
543 alloc_init_section(virt, virt + off, prot_sect);
545 virt += (PGDIR_SIZE / 2);
546 length -= (PGDIR_SIZE / 2);
549 while (length >= PAGE_SIZE) {
550 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
558 * Create the architecture specific mappings
560 void __init iotable_init(struct map_desc *io_desc, int nr)
564 for (i = 0; i < nr; i++)
565 create_mapping(io_desc + i);
568 static inline void prepare_page_table(struct meminfo *mi)
573 * Clear out all the mappings below the kernel image.
575 for (addr = 0; addr < MODULE_START; addr += PGDIR_SIZE)
576 pmd_clear(pmd_off_k(addr));
578 #ifdef CONFIG_XIP_KERNEL
579 /* The XIP kernel is mapped in the module area -- skip over it */
580 addr = ((unsigned long)&_etext + PGDIR_SIZE - 1) & PGDIR_MASK;
582 for ( ; addr < PAGE_OFFSET; addr += PGDIR_SIZE)
583 pmd_clear(pmd_off_k(addr));
586 * Clear out all the kernel space mappings, except for the first
587 * memory bank, up to the end of the vmalloc region.
589 for (addr = __phys_to_virt(mi->bank[0].start + mi->bank[0].size);
590 addr < VMALLOC_END; addr += PGDIR_SIZE)
591 pmd_clear(pmd_off_k(addr));
595 * Reserve the various regions of node 0
597 void __init reserve_node_zero(pg_data_t *pgdat)
599 unsigned long res_size = 0;
602 * Register the kernel text and data with bootmem.
603 * Note that this can only be in node 0.
605 #ifdef CONFIG_XIP_KERNEL
606 reserve_bootmem_node(pgdat, __pa(&__data_start), &_end - &__data_start);
608 reserve_bootmem_node(pgdat, __pa(&_stext), &_end - &_stext);
612 * Reserve the page tables. These are already in use,
613 * and can only be in node 0.
615 reserve_bootmem_node(pgdat, __pa(swapper_pg_dir),
616 PTRS_PER_PGD * sizeof(pgd_t));
619 * Hmm... This should go elsewhere, but we really really need to
620 * stop things allocating the low memory; ideally we need a better
621 * implementation of GFP_DMA which does not assume that DMA-able
622 * memory starts at zero.
624 if (machine_is_integrator() || machine_is_cintegrator())
625 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
628 * These should likewise go elsewhere. They pre-reserve the
629 * screen memory region at the start of main system memory.
631 if (machine_is_edb7211())
632 res_size = 0x00020000;
633 if (machine_is_p720t())
634 res_size = 0x00014000;
636 /* H1940 and RX3715 need to reserve this for suspend */
638 if (machine_is_h1940() || machine_is_rx3715()) {
639 reserve_bootmem_node(pgdat, 0x30003000, 0x1000);
640 reserve_bootmem_node(pgdat, 0x30081000, 0x1000);
645 * Because of the SA1111 DMA bug, we want to preserve our
646 * precious DMA-able memory...
648 res_size = __pa(swapper_pg_dir) - PHYS_OFFSET;
651 reserve_bootmem_node(pgdat, PHYS_OFFSET, res_size);
655 * Set up device the mappings. Since we clear out the page tables for all
656 * mappings above VMALLOC_END, we will remove any debug device mappings.
657 * This means you have to be careful how you debug this function, or any
658 * called function. This means you can't use any function or debugging
659 * method which may touch any device, otherwise the kernel _will_ crash.
661 static void __init devicemaps_init(struct machine_desc *mdesc)
668 * Allocate the vector page early.
670 vectors = alloc_bootmem_low_pages(PAGE_SIZE);
673 for (addr = VMALLOC_END; addr; addr += PGDIR_SIZE)
674 pmd_clear(pmd_off_k(addr));
677 * Map the kernel if it is XIP.
678 * It is always first in the modulearea.
680 #ifdef CONFIG_XIP_KERNEL
681 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
682 map.virtual = MODULE_START;
683 map.length = ((unsigned long)&_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
685 create_mapping(&map);
689 * Map the cache flushing regions.
692 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
693 map.virtual = FLUSH_BASE;
695 map.type = MT_CACHECLEAN;
696 create_mapping(&map);
698 #ifdef FLUSH_BASE_MINICACHE
699 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
700 map.virtual = FLUSH_BASE_MINICACHE;
702 map.type = MT_MINICLEAN;
703 create_mapping(&map);
707 * Create a mapping for the machine vectors at the high-vectors
708 * location (0xffff0000). If we aren't using high-vectors, also
709 * create a mapping at the low-vectors virtual address.
711 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
712 map.virtual = 0xffff0000;
713 map.length = PAGE_SIZE;
714 map.type = MT_HIGH_VECTORS;
715 create_mapping(&map);
717 if (!vectors_high()) {
719 map.type = MT_LOW_VECTORS;
720 create_mapping(&map);
724 * Ask the machine support to map in the statically mapped devices.
730 * Finally flush the caches and tlb to ensure that we're in a
731 * consistent state wrt the writebuffer. This also ensures that
732 * any write-allocated cache lines in the vector page are written
733 * back. After this point, we can start to touch devices again.
735 local_flush_tlb_all();
740 * paging_init() sets up the page tables, initialises the zone memory
741 * maps, and sets up the zero page, bad page and bad page tables.
743 void __init paging_init(struct meminfo *mi, struct machine_desc *mdesc)
747 build_mem_type_table();
748 prepare_page_table(mi);
750 devicemaps_init(mdesc);
752 top_pmd = pmd_off_k(0xffff0000);
755 * allocate the zero page. Note that we count on this going ok.
757 zero_page = alloc_bootmem_low_pages(PAGE_SIZE);
758 memzero(zero_page, PAGE_SIZE);
759 empty_zero_page = virt_to_page(zero_page);
760 flush_dcache_page(empty_zero_page);
764 * In order to soft-boot, we need to insert a 1:1 mapping in place of
765 * the user-mode pages. This will then ensure that we have predictable
766 * results when turning the mmu off
768 void setup_mm_for_reboot(char mode)
770 unsigned long base_pmdval;
774 if (current->mm && current->mm->pgd)
775 pgd = current->mm->pgd;
779 base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
780 if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
781 base_pmdval |= PMD_BIT4;
783 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
784 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
787 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
788 pmd[0] = __pmd(pmdval);
789 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
790 flush_pmd_entry(pmd);