2 * linux/arch/arm/mm/mm-armv.c
4 * Copyright (C) 1998-2005 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 * Page table sludge for ARM v3 and v4 processor architectures.
12 #include <linux/config.h>
13 #include <linux/module.h>
15 #include <linux/init.h>
16 #include <linux/bootmem.h>
17 #include <linux/highmem.h>
18 #include <linux/nodemask.h>
20 #include <asm/pgalloc.h>
22 #include <asm/setup.h>
23 #include <asm/tlbflush.h>
25 #include <asm/mach/map.h>
27 #define CPOLICY_UNCACHED 0
28 #define CPOLICY_BUFFERED 1
29 #define CPOLICY_WRITETHROUGH 2
30 #define CPOLICY_WRITEBACK 3
31 #define CPOLICY_WRITEALLOC 4
33 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
34 static unsigned int ecc_mask __initdata = 0;
35 pgprot_t pgprot_kernel;
37 EXPORT_SYMBOL(pgprot_kernel);
42 const char policy[16];
48 static struct cachepolicy cache_policies[] __initdata = {
52 .pmd = PMD_SECT_UNCACHED,
57 .pmd = PMD_SECT_BUFFERED,
58 .pte = PTE_BUFFERABLE,
60 .policy = "writethrough",
65 .policy = "writeback",
68 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
70 .policy = "writealloc",
73 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
78 * These are useful for identifing cache coherency
79 * problems by allowing the cache or the cache and
80 * writebuffer to be turned off. (Note: the write
81 * buffer should not be on and the cache off).
83 static void __init early_cachepolicy(char **p)
87 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
88 int len = strlen(cache_policies[i].policy);
90 if (memcmp(*p, cache_policies[i].policy, len) == 0) {
92 cr_alignment &= ~cache_policies[i].cr_mask;
93 cr_no_alignment &= ~cache_policies[i].cr_mask;
98 if (i == ARRAY_SIZE(cache_policies))
99 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
101 set_cr(cr_alignment);
104 static void __init early_nocache(char **__unused)
106 char *p = "buffered";
107 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
108 early_cachepolicy(&p);
111 static void __init early_nowrite(char **__unused)
113 char *p = "uncached";
114 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
115 early_cachepolicy(&p);
118 static void __init early_ecc(char **p)
120 if (memcmp(*p, "on", 2) == 0) {
121 ecc_mask = PMD_PROTECTION;
123 } else if (memcmp(*p, "off", 3) == 0) {
129 __early_param("nocache", early_nocache);
130 __early_param("nowb", early_nowrite);
131 __early_param("cachepolicy=", early_cachepolicy);
132 __early_param("ecc=", early_ecc);
134 static int __init noalign_setup(char *__unused)
136 cr_alignment &= ~CR_A;
137 cr_no_alignment &= ~CR_A;
138 set_cr(cr_alignment);
142 __setup("noalign", noalign_setup);
144 #define FIRST_KERNEL_PGD_NR (FIRST_USER_PGD_NR + USER_PTRS_PER_PGD)
146 static inline pmd_t *pmd_off(pgd_t *pgd, unsigned long virt)
148 return pmd_offset(pgd, virt);
151 static inline pmd_t *pmd_off_k(unsigned long virt)
153 return pmd_off(pgd_offset_k(virt), virt);
157 * need to get a 16k page for level 1
159 pgd_t *get_pgd_slow(struct mm_struct *mm)
161 pgd_t *new_pgd, *init_pgd;
162 pmd_t *new_pmd, *init_pmd;
163 pte_t *new_pte, *init_pte;
165 new_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, 2);
169 memzero(new_pgd, FIRST_KERNEL_PGD_NR * sizeof(pgd_t));
172 * Copy over the kernel and IO PGD entries
174 init_pgd = pgd_offset_k(0);
175 memcpy(new_pgd + FIRST_KERNEL_PGD_NR, init_pgd + FIRST_KERNEL_PGD_NR,
176 (PTRS_PER_PGD - FIRST_KERNEL_PGD_NR) * sizeof(pgd_t));
178 clean_dcache_area(new_pgd, PTRS_PER_PGD * sizeof(pgd_t));
180 if (!vectors_high()) {
182 * On ARM, first page must always be allocated since it
183 * contains the machine vectors.
185 new_pmd = pmd_alloc(mm, new_pgd, 0);
189 new_pte = pte_alloc_map(mm, new_pmd, 0);
193 init_pmd = pmd_offset(init_pgd, 0);
194 init_pte = pte_offset_map_nested(init_pmd, 0);
195 set_pte(new_pte, *init_pte);
196 pte_unmap_nested(init_pte);
205 free_pages((unsigned long)new_pgd, 2);
210 void free_pgd_slow(pgd_t *pgd)
218 /* pgd is always present and good */
219 pmd = pmd_off(pgd, 0);
228 pte = pmd_page(*pmd);
230 dec_page_state(nr_page_table_pages);
231 pte_lock_deinit(pte);
235 free_pages((unsigned long) pgd, 2);
239 * Create a SECTION PGD between VIRT and PHYS in domain
240 * DOMAIN with protection PROT. This operates on half-
241 * pgdir entry increments.
244 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
246 pmd_t *pmdp = pmd_off_k(virt);
248 if (virt & (1 << 20))
251 *pmdp = __pmd(phys | prot);
252 flush_pmd_entry(pmdp);
256 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
259 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
263 for (i = 0; i < 16; i += 1) {
264 alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
266 virt += (PGDIR_SIZE / 2);
271 * Add a PAGE mapping between VIRT and PHYS in domain
272 * DOMAIN with protection PROT. Note that due to the
273 * way we map the PTEs, we must allocate two PTE_SIZE'd
274 * blocks - one for the Linux pte table, and one for
275 * the hardware pte table.
278 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
280 pmd_t *pmdp = pmd_off_k(virt);
283 if (pmd_none(*pmdp)) {
284 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
287 __pmd_populate(pmdp, __pa(ptep) | prot_l1);
289 ptep = pte_offset_kernel(pmdp, virt);
291 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
295 unsigned int prot_pte;
296 unsigned int prot_l1;
297 unsigned int prot_sect;
301 static struct mem_types mem_types[] __initdata = {
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_UNCACHED |
311 .prot_sect = PMD_TYPE_SECT,
312 .domain = DOMAIN_KERNEL,
315 .prot_sect = PMD_TYPE_SECT | PMD_SECT_MINICACHE,
316 .domain = DOMAIN_KERNEL,
319 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
321 .prot_l1 = PMD_TYPE_TABLE,
322 .domain = DOMAIN_USER,
324 [MT_HIGH_VECTORS] = {
325 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
326 L_PTE_USER | L_PTE_EXEC,
327 .prot_l1 = PMD_TYPE_TABLE,
328 .domain = DOMAIN_USER,
331 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
332 .domain = DOMAIN_KERNEL,
335 .prot_sect = PMD_TYPE_SECT,
336 .domain = DOMAIN_KERNEL,
338 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
339 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
341 .prot_l1 = PMD_TYPE_TABLE,
342 .prot_sect = PMD_TYPE_SECT | PMD_SECT_UNCACHED |
343 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
347 [MT_NONSHARED_DEVICE] = {
348 .prot_l1 = PMD_TYPE_TABLE,
349 .prot_sect = PMD_TYPE_SECT | PMD_SECT_NONSHARED_DEV |
356 * Adjust the PMD section entries according to the CPU in use.
358 void __init build_mem_type_table(void)
360 struct cachepolicy *cp;
361 unsigned int cr = get_cr();
362 unsigned int user_pgprot, kern_pgprot;
363 int cpu_arch = cpu_architecture();
366 #if defined(CONFIG_CPU_DCACHE_DISABLE)
367 if (cachepolicy > CPOLICY_BUFFERED)
368 cachepolicy = CPOLICY_BUFFERED;
369 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
370 if (cachepolicy > CPOLICY_WRITETHROUGH)
371 cachepolicy = CPOLICY_WRITETHROUGH;
373 if (cpu_arch < CPU_ARCH_ARMv5) {
374 if (cachepolicy >= CPOLICY_WRITEALLOC)
375 cachepolicy = CPOLICY_WRITEBACK;
379 if (cpu_arch <= CPU_ARCH_ARMv5TEJ) {
380 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
381 if (mem_types[i].prot_l1)
382 mem_types[i].prot_l1 |= PMD_BIT4;
383 if (mem_types[i].prot_sect)
384 mem_types[i].prot_sect |= PMD_BIT4;
388 cp = &cache_policies[cachepolicy];
389 kern_pgprot = user_pgprot = cp->pte;
392 * ARMv6 and above have extended page tables.
394 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
396 * bit 4 becomes XN which we must clear for the
397 * kernel memory mapping.
399 mem_types[MT_MEMORY].prot_sect &= ~PMD_BIT4;
400 mem_types[MT_ROM].prot_sect &= ~PMD_BIT4;
403 * Mark cache clean areas and XIP ROM read only
404 * from SVC mode and no access from userspace.
406 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
407 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
408 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
411 * Mark the device area as "shared device"
413 mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
414 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
417 * User pages need to be mapped with the ASID
420 user_pgprot |= L_PTE_ASID;
424 * Mark memory with the "shared" attribute for SMP systems
426 user_pgprot |= L_PTE_SHARED;
427 kern_pgprot |= L_PTE_SHARED;
428 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
432 for (i = 0; i < 16; i++) {
433 unsigned long v = pgprot_val(protection_map[i]);
434 v = (v & ~(L_PTE_BUFFERABLE|L_PTE_CACHEABLE)) | user_pgprot;
435 protection_map[i] = __pgprot(v);
438 mem_types[MT_LOW_VECTORS].prot_pte |= kern_pgprot;
439 mem_types[MT_HIGH_VECTORS].prot_pte |= kern_pgprot;
441 if (cpu_arch >= CPU_ARCH_ARMv5) {
444 * Only use write-through for non-SMP systems
446 mem_types[MT_LOW_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
447 mem_types[MT_HIGH_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
450 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
453 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
454 L_PTE_DIRTY | L_PTE_WRITE |
455 L_PTE_EXEC | kern_pgprot);
457 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
458 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
459 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
460 mem_types[MT_ROM].prot_sect |= cp->pmd;
464 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
468 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
471 printk("Memory policy: ECC %sabled, Data cache %s\n",
472 ecc_mask ? "en" : "dis", cp->policy);
475 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
478 * Create the page directory entries and any necessary
479 * page tables for the mapping specified by `md'. We
480 * are able to cope here with varying sizes and address
481 * offsets, and we take full advantage of sections and
484 void __init create_mapping(struct map_desc *md)
486 unsigned long virt, length;
487 int prot_sect, prot_l1, domain;
489 unsigned long off = (u32)__pfn_to_phys(md->pfn);
491 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
492 printk(KERN_WARNING "BUG: not creating mapping for "
493 "0x%08llx at 0x%08lx in user region\n",
494 __pfn_to_phys((u64)md->pfn), md->virtual);
498 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
499 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
500 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
501 "overlaps vmalloc space\n",
502 __pfn_to_phys((u64)md->pfn), md->virtual);
505 domain = mem_types[md->type].domain;
506 prot_pte = __pgprot(mem_types[md->type].prot_pte);
507 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
508 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
511 * Catch 36-bit addresses
513 if(md->pfn >= 0x100000) {
515 printk(KERN_ERR "MM: invalid domain in supersection "
516 "mapping for 0x%08llx at 0x%08lx\n",
517 __pfn_to_phys((u64)md->pfn), md->virtual);
520 if((md->virtual | md->length | __pfn_to_phys(md->pfn))
521 & ~SUPERSECTION_MASK) {
522 printk(KERN_ERR "MM: cannot create mapping for "
523 "0x%08llx at 0x%08lx invalid alignment\n",
524 __pfn_to_phys((u64)md->pfn), md->virtual);
529 * Shift bits [35:32] of address into bits [23:20] of PMD
532 off |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
539 if (mem_types[md->type].prot_l1 == 0 &&
540 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
541 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
542 "be mapped using pages, ignoring.\n",
543 __pfn_to_phys(md->pfn), md->virtual);
547 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
548 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
554 /* N.B. ARMv6 supersections are only defined to work with domain 0.
555 * Since domain assignments can in fact be arbitrary, the
556 * 'domain == 0' check below is required to insure that ARMv6
557 * supersections are only allocated for domain 0 regardless
558 * of the actual domain assignments in use.
560 if (cpu_architecture() >= CPU_ARCH_ARMv6 && domain == 0) {
562 * Align to supersection boundary if !high pages.
563 * High pages have already been checked for proper
564 * alignment above and they will fail the SUPSERSECTION_MASK
565 * check because of the way the address is encoded into
568 if (md->pfn <= 0x100000) {
569 while ((virt & ~SUPERSECTION_MASK ||
570 (virt + off) & ~SUPERSECTION_MASK) &&
571 length >= (PGDIR_SIZE / 2)) {
572 alloc_init_section(virt, virt + off, prot_sect);
574 virt += (PGDIR_SIZE / 2);
575 length -= (PGDIR_SIZE / 2);
579 while (length >= SUPERSECTION_SIZE) {
580 alloc_init_supersection(virt, virt + off, prot_sect);
582 virt += SUPERSECTION_SIZE;
583 length -= SUPERSECTION_SIZE;
588 * A section mapping covers half a "pgdir" entry.
590 while (length >= (PGDIR_SIZE / 2)) {
591 alloc_init_section(virt, virt + off, prot_sect);
593 virt += (PGDIR_SIZE / 2);
594 length -= (PGDIR_SIZE / 2);
597 while (length >= PAGE_SIZE) {
598 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
606 * In order to soft-boot, we need to insert a 1:1 mapping in place of
607 * the user-mode pages. This will then ensure that we have predictable
608 * results when turning the mmu off
610 void setup_mm_for_reboot(char mode)
612 unsigned long base_pmdval;
616 if (current->mm && current->mm->pgd)
617 pgd = current->mm->pgd;
621 base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
622 if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ)
623 base_pmdval |= PMD_BIT4;
625 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
626 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
629 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
630 pmd[0] = __pmd(pmdval);
631 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
632 flush_pmd_entry(pmd);
637 * Create the architecture specific mappings
639 void __init iotable_init(struct map_desc *io_desc, int nr)
643 for (i = 0; i < nr; i++)
644 create_mapping(io_desc + i);