2 * linux/arch/arm/mm/mm-armv.c
4 * Copyright (C) 1998-2002 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>
23 #include <asm/setup.h>
24 #include <asm/tlbflush.h>
26 #include <asm/mach/map.h>
28 #define CPOLICY_UNCACHED 0
29 #define CPOLICY_BUFFERED 1
30 #define CPOLICY_WRITETHROUGH 2
31 #define CPOLICY_WRITEBACK 3
32 #define CPOLICY_WRITEALLOC 4
34 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
35 static unsigned int ecc_mask __initdata = 0;
36 pgprot_t pgprot_kernel;
38 EXPORT_SYMBOL(pgprot_kernel);
43 const char policy[16];
49 static struct cachepolicy cache_policies[] __initdata = {
53 .pmd = PMD_SECT_UNCACHED,
58 .pmd = PMD_SECT_BUFFERED,
59 .pte = PTE_BUFFERABLE,
61 .policy = "writethrough",
66 .policy = "writeback",
69 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
71 .policy = "writealloc",
74 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
79 * These are useful for identifing cache coherency
80 * problems by allowing the cache or the cache and
81 * writebuffer to be turned off. (Note: the write
82 * buffer should not be on and the cache off).
84 static void __init early_cachepolicy(char **p)
88 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
89 int len = strlen(cache_policies[i].policy);
91 if (memcmp(*p, cache_policies[i].policy, len) == 0) {
93 cr_alignment &= ~cache_policies[i].cr_mask;
94 cr_no_alignment &= ~cache_policies[i].cr_mask;
99 if (i == ARRAY_SIZE(cache_policies))
100 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
102 set_cr(cr_alignment);
105 static void __init early_nocache(char **__unused)
107 char *p = "buffered";
108 printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
109 early_cachepolicy(&p);
112 static void __init early_nowrite(char **__unused)
114 char *p = "uncached";
115 printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
116 early_cachepolicy(&p);
119 static void __init early_ecc(char **p)
121 if (memcmp(*p, "on", 2) == 0) {
122 ecc_mask = PMD_PROTECTION;
124 } else if (memcmp(*p, "off", 3) == 0) {
130 __early_param("nocache", early_nocache);
131 __early_param("nowb", early_nowrite);
132 __early_param("cachepolicy=", early_cachepolicy);
133 __early_param("ecc=", early_ecc);
135 static int __init noalign_setup(char *__unused)
137 cr_alignment &= ~CR_A;
138 cr_no_alignment &= ~CR_A;
139 set_cr(cr_alignment);
143 __setup("noalign", noalign_setup);
145 #define FIRST_KERNEL_PGD_NR (FIRST_USER_PGD_NR + USER_PTRS_PER_PGD)
147 static inline pmd_t *pmd_off(pgd_t *pgd, unsigned long virt)
149 return pmd_offset(pgd, virt);
152 static inline pmd_t *pmd_off_k(unsigned long virt)
154 return pmd_off(pgd_offset_k(virt), virt);
158 * need to get a 16k page for level 1
160 pgd_t *get_pgd_slow(struct mm_struct *mm)
162 pgd_t *new_pgd, *init_pgd;
163 pmd_t *new_pmd, *init_pmd;
164 pte_t *new_pte, *init_pte;
166 new_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, 2);
170 memzero(new_pgd, FIRST_KERNEL_PGD_NR * sizeof(pgd_t));
173 * Copy over the kernel and IO PGD entries
175 init_pgd = pgd_offset_k(0);
176 memcpy(new_pgd + FIRST_KERNEL_PGD_NR, init_pgd + FIRST_KERNEL_PGD_NR,
177 (PTRS_PER_PGD - FIRST_KERNEL_PGD_NR) * sizeof(pgd_t));
179 clean_dcache_area(new_pgd, PTRS_PER_PGD * sizeof(pgd_t));
181 if (!vectors_high()) {
183 * This lock is here just to satisfy pmd_alloc and pte_lock
185 spin_lock(&mm->page_table_lock);
188 * On ARM, first page must always be allocated since it
189 * contains the machine vectors.
191 new_pmd = pmd_alloc(mm, new_pgd, 0);
195 new_pte = pte_alloc_map(mm, new_pmd, 0);
199 init_pmd = pmd_offset(init_pgd, 0);
200 init_pte = pte_offset_map_nested(init_pmd, 0);
201 set_pte(new_pte, *init_pte);
202 pte_unmap_nested(init_pte);
205 spin_unlock(&mm->page_table_lock);
211 spin_unlock(&mm->page_table_lock);
213 free_pages((unsigned long)new_pgd, 2);
217 spin_unlock(&mm->page_table_lock);
218 free_pages((unsigned long)new_pgd, 2);
225 void free_pgd_slow(pgd_t *pgd)
233 /* pgd is always present and good */
234 pmd = pmd_off(pgd, 0);
243 pte = pmd_page(*pmd);
245 dec_page_state(nr_page_table_pages);
249 free_pages((unsigned long) pgd, 2);
253 * Create a SECTION PGD between VIRT and PHYS in domain
254 * DOMAIN with protection PROT. This operates on half-
255 * pgdir entry increments.
258 alloc_init_section(unsigned long virt, unsigned long phys, int prot)
260 pmd_t *pmdp = pmd_off_k(virt);
262 if (virt & (1 << 20))
265 *pmdp = __pmd(phys | prot);
266 flush_pmd_entry(pmdp);
270 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
273 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
277 for (i = 0; i < 16; i += 1) {
278 alloc_init_section(virt, phys, prot | PMD_SECT_SUPER);
280 virt += (PGDIR_SIZE / 2);
285 * Add a PAGE mapping between VIRT and PHYS in domain
286 * DOMAIN with protection PROT. Note that due to the
287 * way we map the PTEs, we must allocate two PTE_SIZE'd
288 * blocks - one for the Linux pte table, and one for
289 * the hardware pte table.
292 alloc_init_page(unsigned long virt, unsigned long phys, unsigned int prot_l1, pgprot_t prot)
294 pmd_t *pmdp = pmd_off_k(virt);
297 if (pmd_none(*pmdp)) {
298 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
301 __pmd_populate(pmdp, __pa(ptep) | prot_l1);
303 ptep = pte_offset_kernel(pmdp, virt);
305 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
309 * Clear any PGD mapping. On a two-level page table system,
310 * the clearance is done by the middle-level functions (pmd)
311 * rather than the top-level (pgd) functions.
313 static inline void clear_mapping(unsigned long virt)
315 pmd_clear(pmd_off_k(virt));
319 unsigned int prot_pte;
320 unsigned int prot_l1;
321 unsigned int prot_sect;
325 static struct mem_types mem_types[] __initdata = {
327 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
329 .prot_l1 = PMD_TYPE_TABLE,
330 .prot_sect = PMD_TYPE_SECT | PMD_SECT_UNCACHED |
335 .prot_sect = PMD_TYPE_SECT,
336 .domain = DOMAIN_KERNEL,
339 .prot_sect = PMD_TYPE_SECT | PMD_SECT_MINICACHE,
340 .domain = DOMAIN_KERNEL,
343 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
345 .prot_l1 = PMD_TYPE_TABLE,
346 .domain = DOMAIN_USER,
348 [MT_HIGH_VECTORS] = {
349 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
350 L_PTE_USER | L_PTE_EXEC,
351 .prot_l1 = PMD_TYPE_TABLE,
352 .domain = DOMAIN_USER,
355 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
356 .domain = DOMAIN_KERNEL,
359 .prot_sect = PMD_TYPE_SECT,
360 .domain = DOMAIN_KERNEL,
362 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
363 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
365 .prot_l1 = PMD_TYPE_TABLE,
366 .prot_sect = PMD_TYPE_SECT | PMD_SECT_UNCACHED |
367 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
374 * Adjust the PMD section entries according to the CPU in use.
376 static void __init build_mem_type_table(void)
378 struct cachepolicy *cp;
379 unsigned int cr = get_cr();
380 unsigned int user_pgprot;
381 int cpu_arch = cpu_architecture();
384 #if defined(CONFIG_CPU_DCACHE_DISABLE)
385 if (cachepolicy > CPOLICY_BUFFERED)
386 cachepolicy = CPOLICY_BUFFERED;
387 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
388 if (cachepolicy > CPOLICY_WRITETHROUGH)
389 cachepolicy = CPOLICY_WRITETHROUGH;
391 if (cpu_arch < CPU_ARCH_ARMv5) {
392 if (cachepolicy >= CPOLICY_WRITEALLOC)
393 cachepolicy = CPOLICY_WRITEBACK;
397 if (cpu_arch <= CPU_ARCH_ARMv5TEJ) {
398 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
399 if (mem_types[i].prot_l1)
400 mem_types[i].prot_l1 |= PMD_BIT4;
401 if (mem_types[i].prot_sect)
402 mem_types[i].prot_sect |= PMD_BIT4;
406 cp = &cache_policies[cachepolicy];
407 user_pgprot = cp->pte;
410 * ARMv6 and above have extended page tables.
412 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
414 * bit 4 becomes XN which we must clear for the
415 * kernel memory mapping.
417 mem_types[MT_MEMORY].prot_sect &= ~PMD_BIT4;
418 mem_types[MT_ROM].prot_sect &= ~PMD_BIT4;
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;
440 if (cpu_arch >= CPU_ARCH_ARMv5) {
441 mem_types[MT_LOW_VECTORS].prot_pte |= cp->pte & PTE_CACHEABLE;
442 mem_types[MT_HIGH_VECTORS].prot_pte |= cp->pte & PTE_CACHEABLE;
444 mem_types[MT_LOW_VECTORS].prot_pte |= cp->pte;
445 mem_types[MT_HIGH_VECTORS].prot_pte |= cp->pte;
446 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
449 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
450 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
451 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
452 mem_types[MT_ROM].prot_sect |= cp->pmd;
454 for (i = 0; i < 16; i++) {
455 unsigned long v = pgprot_val(protection_map[i]);
456 v = (v & ~(PTE_BUFFERABLE|PTE_CACHEABLE)) | user_pgprot;
457 protection_map[i] = __pgprot(v);
460 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
461 L_PTE_DIRTY | L_PTE_WRITE |
462 L_PTE_EXEC | cp->pte);
466 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
470 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
473 printk("Memory policy: ECC %sabled, Data cache %s\n",
474 ecc_mask ? "en" : "dis", cp->policy);
477 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
480 * Create the page directory entries and any necessary
481 * page tables for the mapping specified by `md'. We
482 * are able to cope here with varying sizes and address
483 * offsets, and we take full advantage of sections and
486 static void __init create_mapping(struct map_desc *md)
488 unsigned long virt, length;
489 int prot_sect, prot_l1, domain;
493 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
494 printk(KERN_WARNING "BUG: not creating mapping for "
495 "0x%08lx at 0x%08lx in user region\n",
496 md->physical, md->virtual);
500 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
501 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
502 printk(KERN_WARNING "BUG: mapping for 0x%08lx at 0x%08lx "
503 "overlaps vmalloc space\n",
504 md->physical, md->virtual);
507 domain = mem_types[md->type].domain;
508 prot_pte = __pgprot(mem_types[md->type].prot_pte);
509 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
510 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
513 off = md->physical - virt;
516 if (mem_types[md->type].prot_l1 == 0 &&
517 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
518 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
519 "be mapped using pages, ignoring.\n",
520 md->physical, md->virtual);
524 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
525 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
531 /* N.B. ARMv6 supersections are only defined to work with domain 0.
532 * Since domain assignments can in fact be arbitrary, the
533 * 'domain == 0' check below is required to insure that ARMv6
534 * supersections are only allocated for domain 0 regardless
535 * of the actual domain assignments in use.
537 if (cpu_architecture() >= CPU_ARCH_ARMv6 && domain == 0) {
538 /* Align to supersection boundary */
539 while ((virt & ~SUPERSECTION_MASK || (virt + off) &
540 ~SUPERSECTION_MASK) && length >= (PGDIR_SIZE / 2)) {
541 alloc_init_section(virt, virt + off, prot_sect);
543 virt += (PGDIR_SIZE / 2);
544 length -= (PGDIR_SIZE / 2);
547 while (length >= SUPERSECTION_SIZE) {
548 alloc_init_supersection(virt, virt + off, prot_sect);
550 virt += SUPERSECTION_SIZE;
551 length -= SUPERSECTION_SIZE;
556 * A section mapping covers half a "pgdir" entry.
558 while (length >= (PGDIR_SIZE / 2)) {
559 alloc_init_section(virt, virt + off, prot_sect);
561 virt += (PGDIR_SIZE / 2);
562 length -= (PGDIR_SIZE / 2);
565 while (length >= PAGE_SIZE) {
566 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
574 * In order to soft-boot, we need to insert a 1:1 mapping in place of
575 * the user-mode pages. This will then ensure that we have predictable
576 * results when turning the mmu off
578 void setup_mm_for_reboot(char mode)
580 unsigned long base_pmdval;
584 if (current->mm && current->mm->pgd)
585 pgd = current->mm->pgd;
589 base_pmdval = PMD_SECT_AP_WRITE | PMD_SECT_AP_READ | PMD_TYPE_SECT;
590 if (cpu_architecture() <= CPU_ARCH_ARMv5TEJ)
591 base_pmdval |= PMD_BIT4;
593 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++, pgd++) {
594 unsigned long pmdval = (i << PGDIR_SHIFT) | base_pmdval;
597 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
598 pmd[0] = __pmd(pmdval);
599 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
600 flush_pmd_entry(pmd);
604 extern void _stext, _etext;
607 * Setup initial mappings. We use the page we allocated for zero page to hold
608 * the mappings, which will get overwritten by the vectors in traps_init().
609 * The mappings must be in virtual address order.
611 void __init memtable_init(struct meminfo *mi)
613 struct map_desc *init_maps, *p, *q;
614 unsigned long address = 0;
617 build_mem_type_table();
619 init_maps = p = alloc_bootmem_low_pages(PAGE_SIZE);
621 #ifdef CONFIG_XIP_KERNEL
622 p->physical = CONFIG_XIP_PHYS_ADDR & PMD_MASK;
623 p->virtual = (unsigned long)&_stext & PMD_MASK;
624 p->length = ((unsigned long)&_etext - p->virtual + ~PMD_MASK) & PMD_MASK;
629 for (i = 0; i < mi->nr_banks; i++) {
630 if (mi->bank[i].size == 0)
633 p->physical = mi->bank[i].start;
634 p->virtual = __phys_to_virt(p->physical);
635 p->length = mi->bank[i].size;
641 p->physical = FLUSH_BASE_PHYS;
642 p->virtual = FLUSH_BASE;
643 p->length = PGDIR_SIZE;
644 p->type = MT_CACHECLEAN;
648 #ifdef FLUSH_BASE_MINICACHE
649 p->physical = FLUSH_BASE_PHYS + PGDIR_SIZE;
650 p->virtual = FLUSH_BASE_MINICACHE;
651 p->length = PGDIR_SIZE;
652 p->type = MT_MINICLEAN;
657 * Go through the initial mappings, but clear out any
658 * pgdir entries that are not in the description.
662 if (address < q->virtual || q == p) {
663 clear_mapping(address);
664 address += PGDIR_SIZE;
668 address = q->virtual + q->length;
669 address = (address + PGDIR_SIZE - 1) & PGDIR_MASK;
673 } while (address != 0);
676 * Create a mapping for the machine vectors at the high-vectors
677 * location (0xffff0000). If we aren't using high-vectors, also
678 * create a mapping at the low-vectors virtual address.
680 init_maps->physical = virt_to_phys(init_maps);
681 init_maps->virtual = 0xffff0000;
682 init_maps->length = PAGE_SIZE;
683 init_maps->type = MT_HIGH_VECTORS;
684 create_mapping(init_maps);
686 if (!vectors_high()) {
687 init_maps->virtual = 0;
688 init_maps->type = MT_LOW_VECTORS;
689 create_mapping(init_maps);
693 local_flush_tlb_all();
695 top_pmd = pmd_off_k(0xffff0000);
699 * Create the architecture specific mappings
701 void __init iotable_init(struct map_desc *io_desc, int nr)
705 for (i = 0; i < nr; i++)
706 create_mapping(io_desc + i);