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/module.h>
14 #include <linux/init.h>
15 #include <linux/bootmem.h>
16 #include <linux/highmem.h>
17 #include <linux/nodemask.h>
19 #include <asm/pgalloc.h>
21 #include <asm/setup.h>
22 #include <asm/tlbflush.h>
24 #include <asm/mach/map.h>
26 #define CPOLICY_UNCACHED 0
27 #define CPOLICY_BUFFERED 1
28 #define CPOLICY_WRITETHROUGH 2
29 #define CPOLICY_WRITEBACK 3
30 #define CPOLICY_WRITEALLOC 4
32 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
33 static unsigned int ecc_mask __initdata = 0;
34 pgprot_t pgprot_kernel;
36 EXPORT_SYMBOL(pgprot_kernel);
41 const char policy[16];
47 static struct cachepolicy cache_policies[] __initdata = {
51 .pmd = PMD_SECT_UNCACHED,
56 .pmd = PMD_SECT_BUFFERED,
57 .pte = PTE_BUFFERABLE,
59 .policy = "writethrough",
64 .policy = "writeback",
67 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
69 .policy = "writealloc",
72 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
77 * These are useful for identifing cache coherency
78 * problems by allowing the cache or the cache and
79 * writebuffer to be turned off. (Note: the write
80 * buffer should not be on and the cache off).
82 static void __init early_cachepolicy(char **p)
86 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
87 int len = strlen(cache_policies[i].policy);
89 if (memcmp(*p, cache_policies[i].policy, len) == 0) {
91 cr_alignment &= ~cache_policies[i].cr_mask;
92 cr_no_alignment &= ~cache_policies[i].cr_mask;
97 if (i == ARRAY_SIZE(cache_policies))
98 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
100 set_cr(cr_alignment);
103 static void __init early_nocache(char **__unused)
105 char *p = "buffered";
106 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
107 early_cachepolicy(&p);
110 static void __init early_nowrite(char **__unused)
112 char *p = "uncached";
113 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
114 early_cachepolicy(&p);
117 static void __init early_ecc(char **p)
119 if (memcmp(*p, "on", 2) == 0) {
120 ecc_mask = PMD_PROTECTION;
122 } else if (memcmp(*p, "off", 3) == 0) {
128 __early_param("nocache", early_nocache);
129 __early_param("nowb", early_nowrite);
130 __early_param("cachepolicy=", early_cachepolicy);
131 __early_param("ecc=", early_ecc);
133 static int __init noalign_setup(char *__unused)
135 cr_alignment &= ~CR_A;
136 cr_no_alignment &= ~CR_A;
137 set_cr(cr_alignment);
141 __setup("noalign", noalign_setup);
143 #define FIRST_KERNEL_PGD_NR (FIRST_USER_PGD_NR + USER_PTRS_PER_PGD)
145 static inline pmd_t *pmd_off(pgd_t *pgd, unsigned long virt)
147 return pmd_offset(pgd, virt);
150 static inline pmd_t *pmd_off_k(unsigned long virt)
152 return pmd_off(pgd_offset_k(virt), virt);
156 * need to get a 16k page for level 1
158 pgd_t *get_pgd_slow(struct mm_struct *mm)
160 pgd_t *new_pgd, *init_pgd;
161 pmd_t *new_pmd, *init_pmd;
162 pte_t *new_pte, *init_pte;
164 new_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, 2);
168 memzero(new_pgd, FIRST_KERNEL_PGD_NR * sizeof(pgd_t));
171 * Copy over the kernel and IO PGD entries
173 init_pgd = pgd_offset_k(0);
174 memcpy(new_pgd + FIRST_KERNEL_PGD_NR, init_pgd + FIRST_KERNEL_PGD_NR,
175 (PTRS_PER_PGD - FIRST_KERNEL_PGD_NR) * sizeof(pgd_t));
177 clean_dcache_area(new_pgd, PTRS_PER_PGD * sizeof(pgd_t));
179 if (!vectors_high()) {
181 * On ARM, first page must always be allocated since it
182 * contains the machine vectors.
184 new_pmd = pmd_alloc(mm, new_pgd, 0);
188 new_pte = pte_alloc_map(mm, new_pmd, 0);
192 init_pmd = pmd_offset(init_pgd, 0);
193 init_pte = pte_offset_map_nested(init_pmd, 0);
194 set_pte(new_pte, *init_pte);
195 pte_unmap_nested(init_pte);
204 free_pages((unsigned long)new_pgd, 2);
209 void free_pgd_slow(pgd_t *pgd)
217 /* pgd is always present and good */
218 pmd = pmd_off(pgd, 0);
227 pte = pmd_page(*pmd);
229 dec_zone_page_state(virt_to_page((unsigned long *)pgd), NR_PAGETABLE);
230 pte_lock_deinit(pte);
234 free_pages((unsigned long) pgd, 2);
238 * Create a SECTION PGD between VIRT and PHYS in domain
239 * DOMAIN with protection PROT. This operates on half-
240 * pgdir entry increments.
243 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
245 pmd_t *pmdp = pmd_off_k(virt);
247 if (virt & (1 << 20))
250 *pmdp = __pmd(phys | prot);
251 flush_pmd_entry(pmdp);
255 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
258 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
262 for (i = 0; i < 16; i += 1) {
263 alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
265 virt += (PGDIR_SIZE / 2);
270 * Add a PAGE mapping between VIRT and PHYS in domain
271 * DOMAIN with protection PROT. Note that due to the
272 * way we map the PTEs, we must allocate two PTE_SIZE'd
273 * blocks - one for the Linux pte table, and one for
274 * the hardware pte table.
277 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
279 pmd_t *pmdp = pmd_off_k(virt);
282 if (pmd_none(*pmdp)) {
283 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
286 __pmd_populate(pmdp, __pa(ptep) | prot_l1);
288 ptep = pte_offset_kernel(pmdp, virt);
290 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
294 unsigned int prot_pte;
295 unsigned int prot_l1;
296 unsigned int prot_sect;
300 static struct mem_types mem_types[] __initdata = {
302 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
304 .prot_l1 = PMD_TYPE_TABLE,
305 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
310 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
311 .domain = DOMAIN_KERNEL,
314 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_MINICACHE,
315 .domain = DOMAIN_KERNEL,
318 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
320 .prot_l1 = PMD_TYPE_TABLE,
321 .domain = DOMAIN_USER,
323 [MT_HIGH_VECTORS] = {
324 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
325 L_PTE_USER | L_PTE_EXEC,
326 .prot_l1 = PMD_TYPE_TABLE,
327 .domain = DOMAIN_USER,
330 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_AP_WRITE,
331 .domain = DOMAIN_KERNEL,
334 .prot_sect = PMD_TYPE_SECT | PMD_BIT4,
335 .domain = DOMAIN_KERNEL,
337 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
338 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
340 .prot_l1 = PMD_TYPE_TABLE,
341 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_UNCACHED |
342 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
346 [MT_NONSHARED_DEVICE] = {
347 .prot_l1 = PMD_TYPE_TABLE,
348 .prot_sect = PMD_TYPE_SECT | PMD_BIT4 | PMD_SECT_NONSHARED_DEV |
355 * Adjust the PMD section entries according to the CPU in use.
357 void __init build_mem_type_table(void)
359 struct cachepolicy *cp;
360 unsigned int cr = get_cr();
361 unsigned int user_pgprot, kern_pgprot;
362 int cpu_arch = cpu_architecture();
365 #if defined(CONFIG_CPU_DCACHE_DISABLE)
366 if (cachepolicy > CPOLICY_BUFFERED)
367 cachepolicy = CPOLICY_BUFFERED;
368 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
369 if (cachepolicy > CPOLICY_WRITETHROUGH)
370 cachepolicy = CPOLICY_WRITETHROUGH;
372 if (cpu_arch < CPU_ARCH_ARMv5) {
373 if (cachepolicy >= CPOLICY_WRITEALLOC)
374 cachepolicy = CPOLICY_WRITEBACK;
379 * Xscale must not have PMD bit 4 set for section mappings.
382 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
383 mem_types[i].prot_sect &= ~PMD_BIT4;
386 * ARMv5 and lower, excluding Xscale, bit 4 must be set for
389 if (cpu_arch < CPU_ARCH_ARMv6 && !cpu_is_xscale())
390 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
391 if (mem_types[i].prot_l1)
392 mem_types[i].prot_l1 |= PMD_BIT4;
394 cp = &cache_policies[cachepolicy];
395 kern_pgprot = user_pgprot = cp->pte;
398 * Enable CPU-specific coherency if supported.
399 * (Only available on XSC3 at the moment.)
401 if (arch_is_coherent()) {
403 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
404 mem_types[MT_MEMORY].prot_pte |= L_PTE_COHERENT;
409 * ARMv6 and above have extended page tables.
411 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
413 * bit 4 becomes XN which we must clear for the
414 * kernel memory mapping.
416 mem_types[MT_MEMORY].prot_sect &= ~PMD_SECT_XN;
417 mem_types[MT_ROM].prot_sect &= ~PMD_SECT_XN;
420 * Mark cache clean areas and XIP ROM read only
421 * from SVC mode and no access from userspace.
423 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
424 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
425 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
428 * Mark the device area as "shared device"
430 mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
431 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
434 * User pages need to be mapped with the ASID
437 user_pgprot |= L_PTE_ASID;
441 * Mark memory with the "shared" attribute for SMP systems
443 user_pgprot |= L_PTE_SHARED;
444 kern_pgprot |= L_PTE_SHARED;
445 mem_types[MT_MEMORY].prot_sect |= PMD_SECT_S;
449 for (i = 0; i < 16; i++) {
450 unsigned long v = pgprot_val(protection_map[i]);
451 v = (v & ~(L_PTE_BUFFERABLE|L_PTE_CACHEABLE)) | user_pgprot;
452 protection_map[i] = __pgprot(v);
455 mem_types[MT_LOW_VECTORS].prot_pte |= kern_pgprot;
456 mem_types[MT_HIGH_VECTORS].prot_pte |= kern_pgprot;
458 if (cpu_arch >= CPU_ARCH_ARMv5) {
461 * Only use write-through for non-SMP systems
463 mem_types[MT_LOW_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
464 mem_types[MT_HIGH_VECTORS].prot_pte &= ~L_PTE_BUFFERABLE;
467 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
470 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
471 L_PTE_DIRTY | L_PTE_WRITE |
472 L_PTE_EXEC | kern_pgprot);
474 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
475 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
476 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
477 mem_types[MT_ROM].prot_sect |= cp->pmd;
481 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
485 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
488 printk("Memory policy: ECC %sabled, Data cache %s\n",
489 ecc_mask ? "en" : "dis", cp->policy);
492 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
495 * Create the page directory entries and any necessary
496 * page tables for the mapping specified by `md'. We
497 * are able to cope here with varying sizes and address
498 * offsets, and we take full advantage of sections and
501 void __init create_mapping(struct map_desc *md)
503 unsigned long virt, length;
504 int prot_sect, prot_l1, domain;
506 unsigned long off = (u32)__pfn_to_phys(md->pfn);
508 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
509 printk(KERN_WARNING "BUG: not creating mapping for "
510 "0x%08llx at 0x%08lx in user region\n",
511 __pfn_to_phys((u64)md->pfn), md->virtual);
515 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
516 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
517 printk(KERN_WARNING "BUG: mapping for 0x%08llx at 0x%08lx "
518 "overlaps vmalloc space\n",
519 __pfn_to_phys((u64)md->pfn), md->virtual);
522 domain = mem_types[md->type].domain;
523 prot_pte = __pgprot(mem_types[md->type].prot_pte);
524 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
525 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
528 * Catch 36-bit addresses
530 if(md->pfn >= 0x100000) {
532 printk(KERN_ERR "MM: invalid domain in supersection "
533 "mapping for 0x%08llx at 0x%08lx\n",
534 __pfn_to_phys((u64)md->pfn), md->virtual);
537 if((md->virtual | md->length | __pfn_to_phys(md->pfn))
538 & ~SUPERSECTION_MASK) {
539 printk(KERN_ERR "MM: cannot create mapping for "
540 "0x%08llx at 0x%08lx invalid alignment\n",
541 __pfn_to_phys((u64)md->pfn), md->virtual);
546 * Shift bits [35:32] of address into bits [23:20] of PMD
549 off |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
556 if (mem_types[md->type].prot_l1 == 0 &&
557 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
558 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
559 "be mapped using pages, ignoring.\n",
560 __pfn_to_phys(md->pfn), md->virtual);
564 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
565 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
571 /* N.B. ARMv6 supersections are only defined to work with domain 0.
572 * Since domain assignments can in fact be arbitrary, the
573 * 'domain == 0' check below is required to insure that ARMv6
574 * supersections are only allocated for domain 0 regardless
575 * of the actual domain assignments in use.
577 if ((cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())
580 * Align to supersection boundary if !high pages.
581 * High pages have already been checked for proper
582 * alignment above and they will fail the SUPSERSECTION_MASK
583 * check because of the way the address is encoded into
586 if (md->pfn <= 0x100000) {
587 while ((virt & ~SUPERSECTION_MASK ||
588 (virt + off) & ~SUPERSECTION_MASK) &&
589 length >= (PGDIR_SIZE / 2)) {
590 alloc_init_section(virt, virt + off, prot_sect);
592 virt += (PGDIR_SIZE / 2);
593 length -= (PGDIR_SIZE / 2);
597 while (length >= SUPERSECTION_SIZE) {
598 alloc_init_supersection(virt, virt + off, prot_sect);
600 virt += SUPERSECTION_SIZE;
601 length -= SUPERSECTION_SIZE;
606 * A section mapping covers half a "pgdir" entry.
608 while (length >= (PGDIR_SIZE / 2)) {
609 alloc_init_section(virt, virt + off, prot_sect);
611 virt += (PGDIR_SIZE / 2);
612 length -= (PGDIR_SIZE / 2);
615 while (length >= PAGE_SIZE) {
616 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
624 * In order to soft-boot, we need to insert a 1:1 mapping in place of
625 * the user-mode pages. This will then ensure that we have predictable
626 * results when turning the mmu off
628 void setup_mm_for_reboot(char mode)
630 unsigned long base_pmdval;
634 if (current->mm && current->mm->pgd)
635 pgd = current->mm->pgd;
639 base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
640 if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ && !cpu_is_xscale())
641 base_pmdval |= PMD_BIT4;
643 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
644 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
647 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
648 pmd[0] = __pmd(pmdval);
649 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
650 flush_pmd_entry(pmd);
655 * Create the architecture specific mappings
657 void __init iotable_init(struct map_desc *io_desc, int nr)
661 for (i = 0; i < nr; i++)
662 create_mapping(io_desc + i);