[ARM] 2853/1: Make alloc_init_supersection() work with 36-bit mappings
[linux-2.6/kmemtrace.git] / arch / arm / mm / mm-armv.c
blob4dae00bf7a563f5c9c101cde18536b5f48f5463d
1 /*
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>
14 #include <linux/mm.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>
21 #include <asm/page.h>
22 #include <asm/io.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);
40 pmd_t *top_pmd;
42 struct cachepolicy {
43 const char policy[16];
44 unsigned int cr_mask;
45 unsigned int pmd;
46 unsigned int pte;
49 static struct cachepolicy cache_policies[] __initdata = {
51 .policy = "uncached",
52 .cr_mask = CR_W|CR_C,
53 .pmd = PMD_SECT_UNCACHED,
54 .pte = 0,
55 }, {
56 .policy = "buffered",
57 .cr_mask = CR_C,
58 .pmd = PMD_SECT_BUFFERED,
59 .pte = PTE_BUFFERABLE,
60 }, {
61 .policy = "writethrough",
62 .cr_mask = 0,
63 .pmd = PMD_SECT_WT,
64 .pte = PTE_CACHEABLE,
65 }, {
66 .policy = "writeback",
67 .cr_mask = 0,
68 .pmd = PMD_SECT_WB,
69 .pte = PTE_BUFFERABLE|PTE_CACHEABLE,
70 }, {
71 .policy = "writealloc",
72 .cr_mask = 0,
73 .pmd = PMD_SECT_WBWA,
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)
86 int i;
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) {
92 cachepolicy = i;
93 cr_alignment &= ~cache_policies[i].cr_mask;
94 cr_no_alignment &= ~cache_policies[i].cr_mask;
95 *p += len;
96 break;
99 if (i == ARRAY_SIZE(cache_policies))
100 printk(KERN_ERR "ERROR: unknown or unsupported cache policy\n");
101 flush_cache_all();
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;
123 *p += 2;
124 } else if (memcmp(*p, "off", 3) == 0) {
125 ecc_mask = 0;
126 *p += 3;
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);
140 return 1;
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);
167 if (!new_pgd)
168 goto no_pgd;
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);
192 if (!new_pmd)
193 goto no_pmd;
195 new_pte = pte_alloc_map(mm, new_pmd, 0);
196 if (!new_pte)
197 goto no_pte;
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);
203 pte_unmap(new_pte);
205 spin_unlock(&mm->page_table_lock);
208 return new_pgd;
210 no_pte:
211 spin_unlock(&mm->page_table_lock);
212 pmd_free(new_pmd);
213 free_pages((unsigned long)new_pgd, 2);
214 return NULL;
216 no_pmd:
217 spin_unlock(&mm->page_table_lock);
218 free_pages((unsigned long)new_pgd, 2);
219 return NULL;
221 no_pgd:
222 return NULL;
225 void free_pgd_slow(pgd_t *pgd)
227 pmd_t *pmd;
228 struct page *pte;
230 if (!pgd)
231 return;
233 /* pgd is always present and good */
234 pmd = pmd_off(pgd, 0);
235 if (pmd_none(*pmd))
236 goto free;
237 if (pmd_bad(*pmd)) {
238 pmd_ERROR(*pmd);
239 pmd_clear(pmd);
240 goto free;
243 pte = pmd_page(*pmd);
244 pmd_clear(pmd);
245 dec_page_state(nr_page_table_pages);
246 pte_free(pte);
247 pmd_free(pmd);
248 free:
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.
257 static inline void
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))
263 pmdp++;
265 *pmdp = __pmd(phys | prot);
266 flush_pmd_entry(pmdp);
270 * Create a SUPER SECTION PGD between VIRT and PHYS with protection PROT
272 static inline void
273 alloc_init_supersection(unsigned long virt, unsigned long phys, int prot)
275 int i;
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.
291 static inline void
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);
295 pte_t *ptep;
297 if (pmd_none(*pmdp)) {
298 unsigned long pmdval;
299 ptep = alloc_bootmem_low_pages(2 * PTRS_PER_PTE *
300 sizeof(pte_t));
302 pmdval = __pa(ptep) | prot_l1;
303 pmdp[0] = __pmd(pmdval);
304 pmdp[1] = __pmd(pmdval + 256 * sizeof(pte_t));
305 flush_pmd_entry(pmdp);
307 ptep = pte_offset_kernel(pmdp, virt);
309 set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
313 * Clear any PGD mapping. On a two-level page table system,
314 * the clearance is done by the middle-level functions (pmd)
315 * rather than the top-level (pgd) functions.
317 static inline void clear_mapping(unsigned long virt)
319 pmd_clear(pmd_off_k(virt));
322 struct mem_types {
323 unsigned int prot_pte;
324 unsigned int prot_l1;
325 unsigned int prot_sect;
326 unsigned int domain;
329 static struct mem_types mem_types[] __initdata = {
330 [MT_DEVICE] = {
331 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
332 L_PTE_WRITE,
333 .prot_l1 = PMD_TYPE_TABLE,
334 .prot_sect = PMD_TYPE_SECT | PMD_SECT_UNCACHED |
335 PMD_SECT_AP_WRITE,
336 .domain = DOMAIN_IO,
338 [MT_CACHECLEAN] = {
339 .prot_sect = PMD_TYPE_SECT,
340 .domain = DOMAIN_KERNEL,
342 [MT_MINICLEAN] = {
343 .prot_sect = PMD_TYPE_SECT | PMD_SECT_MINICACHE,
344 .domain = DOMAIN_KERNEL,
346 [MT_LOW_VECTORS] = {
347 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
348 L_PTE_EXEC,
349 .prot_l1 = PMD_TYPE_TABLE,
350 .domain = DOMAIN_USER,
352 [MT_HIGH_VECTORS] = {
353 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
354 L_PTE_USER | L_PTE_EXEC,
355 .prot_l1 = PMD_TYPE_TABLE,
356 .domain = DOMAIN_USER,
358 [MT_MEMORY] = {
359 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
360 .domain = DOMAIN_KERNEL,
362 [MT_ROM] = {
363 .prot_sect = PMD_TYPE_SECT,
364 .domain = DOMAIN_KERNEL,
366 [MT_IXP2000_DEVICE] = { /* IXP2400 requires XCB=101 for on-chip I/O */
367 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
368 L_PTE_WRITE,
369 .prot_l1 = PMD_TYPE_TABLE,
370 .prot_sect = PMD_TYPE_SECT | PMD_SECT_UNCACHED |
371 PMD_SECT_AP_WRITE | PMD_SECT_BUFFERABLE |
372 PMD_SECT_TEX(1),
373 .domain = DOMAIN_IO,
378 * Adjust the PMD section entries according to the CPU in use.
380 static void __init build_mem_type_table(void)
382 struct cachepolicy *cp;
383 unsigned int cr = get_cr();
384 unsigned int user_pgprot;
385 int cpu_arch = cpu_architecture();
386 int i;
388 #if defined(CONFIG_CPU_DCACHE_DISABLE)
389 if (cachepolicy > CPOLICY_BUFFERED)
390 cachepolicy = CPOLICY_BUFFERED;
391 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
392 if (cachepolicy > CPOLICY_WRITETHROUGH)
393 cachepolicy = CPOLICY_WRITETHROUGH;
394 #endif
395 if (cpu_arch < CPU_ARCH_ARMv5) {
396 if (cachepolicy >= CPOLICY_WRITEALLOC)
397 cachepolicy = CPOLICY_WRITEBACK;
398 ecc_mask = 0;
401 if (cpu_arch <= CPU_ARCH_ARMv5TEJ) {
402 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
403 if (mem_types[i].prot_l1)
404 mem_types[i].prot_l1 |= PMD_BIT4;
405 if (mem_types[i].prot_sect)
406 mem_types[i].prot_sect |= PMD_BIT4;
410 cp = &cache_policies[cachepolicy];
411 user_pgprot = cp->pte;
414 * ARMv6 and above have extended page tables.
416 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
418 * bit 4 becomes XN which we must clear for the
419 * kernel memory mapping.
421 mem_types[MT_MEMORY].prot_sect &= ~PMD_BIT4;
422 mem_types[MT_ROM].prot_sect &= ~PMD_BIT4;
424 * Mark cache clean areas and XIP ROM read only
425 * from SVC mode and no access from userspace.
427 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
428 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
429 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
432 * Mark the device area as "shared device"
434 mem_types[MT_DEVICE].prot_pte |= L_PTE_BUFFERABLE;
435 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
438 * User pages need to be mapped with the ASID
439 * (iow, non-global)
441 user_pgprot |= L_PTE_ASID;
444 if (cpu_arch >= CPU_ARCH_ARMv5) {
445 mem_types[MT_LOW_VECTORS].prot_pte |= cp->pte & PTE_CACHEABLE;
446 mem_types[MT_HIGH_VECTORS].prot_pte |= cp->pte & PTE_CACHEABLE;
447 } else {
448 mem_types[MT_LOW_VECTORS].prot_pte |= cp->pte;
449 mem_types[MT_HIGH_VECTORS].prot_pte |= cp->pte;
450 mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
453 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
454 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
455 mem_types[MT_MEMORY].prot_sect |= ecc_mask | cp->pmd;
456 mem_types[MT_ROM].prot_sect |= cp->pmd;
458 for (i = 0; i < 16; i++) {
459 unsigned long v = pgprot_val(protection_map[i]);
460 v &= (~(PTE_BUFFERABLE|PTE_CACHEABLE)) | user_pgprot;
461 protection_map[i] = __pgprot(v);
464 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
465 L_PTE_DIRTY | L_PTE_WRITE |
466 L_PTE_EXEC | cp->pte);
468 switch (cp->pmd) {
469 case PMD_SECT_WT:
470 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
471 break;
472 case PMD_SECT_WB:
473 case PMD_SECT_WBWA:
474 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
475 break;
477 printk("Memory policy: ECC %sabled, Data cache %s\n",
478 ecc_mask ? "en" : "dis", cp->policy);
481 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
484 * Create the page directory entries and any necessary
485 * page tables for the mapping specified by `md'. We
486 * are able to cope here with varying sizes and address
487 * offsets, and we take full advantage of sections and
488 * supersections.
490 static void __init create_mapping(struct map_desc *md)
492 unsigned long virt, length;
493 int prot_sect, prot_l1, domain;
494 pgprot_t prot_pte;
495 long off;
497 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
498 printk(KERN_WARNING "BUG: not creating mapping for "
499 "0x%08lx at 0x%08lx in user region\n",
500 md->physical, md->virtual);
501 return;
504 if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
505 md->virtual >= PAGE_OFFSET && md->virtual < VMALLOC_END) {
506 printk(KERN_WARNING "BUG: mapping for 0x%08lx at 0x%08lx "
507 "overlaps vmalloc space\n",
508 md->physical, md->virtual);
511 domain = mem_types[md->type].domain;
512 prot_pte = __pgprot(mem_types[md->type].prot_pte);
513 prot_l1 = mem_types[md->type].prot_l1 | PMD_DOMAIN(domain);
514 prot_sect = mem_types[md->type].prot_sect | PMD_DOMAIN(domain);
516 virt = md->virtual;
517 off = md->physical - virt;
518 length = md->length;
520 if (mem_types[md->type].prot_l1 == 0 &&
521 (virt & 0xfffff || (virt + off) & 0xfffff || (virt + length) & 0xfffff)) {
522 printk(KERN_WARNING "BUG: map for 0x%08lx at 0x%08lx can not "
523 "be mapped using pages, ignoring.\n",
524 md->physical, md->virtual);
525 return;
528 while ((virt & 0xfffff || (virt + off) & 0xfffff) && length >= PAGE_SIZE) {
529 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
531 virt += PAGE_SIZE;
532 length -= PAGE_SIZE;
535 /* N.B. ARMv6 supersections are only defined to work with domain 0.
536 * Since domain assignments can in fact be arbitrary, the
537 * 'domain == 0' check below is required to insure that ARMv6
538 * supersections are only allocated for domain 0 regardless
539 * of the actual domain assignments in use.
541 if (cpu_architecture() >= CPU_ARCH_ARMv6 && domain == 0) {
542 /* Align to supersection boundary */
543 while ((virt & ~SUPERSECTION_MASK || (virt + off) &
544 ~SUPERSECTION_MASK) && length >= (PGDIR_SIZE / 2)) {
545 alloc_init_section(virt, virt + off, prot_sect);
547 virt += (PGDIR_SIZE / 2);
548 length -= (PGDIR_SIZE / 2);
551 while (length >= SUPERSECTION_SIZE) {
552 alloc_init_supersection(virt, virt + off, prot_sect);
554 virt += SUPERSECTION_SIZE;
555 length -= SUPERSECTION_SIZE;
560 * A section mapping covers half a "pgdir" entry.
562 while (length >= (PGDIR_SIZE / 2)) {
563 alloc_init_section(virt, virt + off, prot_sect);
565 virt += (PGDIR_SIZE / 2);
566 length -= (PGDIR_SIZE / 2);
569 while (length >= PAGE_SIZE) {
570 alloc_init_page(virt, virt + off, prot_l1, prot_pte);
572 virt += PAGE_SIZE;
573 length -= PAGE_SIZE;
578 * In order to soft-boot, we need to insert a 1:1 mapping in place of
579 * the user-mode pages. This will then ensure that we have predictable
580 * results when turning the mmu off
582 void setup_mm_for_reboot(char mode)
584 unsigned long pmdval;
585 pgd_t *pgd;
586 pmd_t *pmd;
587 int i;
588 int cpu_arch = cpu_architecture();
590 if (current->mm && current->mm->pgd)
591 pgd = current->mm->pgd;
592 else
593 pgd = init_mm.pgd;
595 for (i = 0; i < FIRST_USER_PGD_NR + USER_PTRS_PER_PGD; i++) {
596 pmdval = (i << PGDIR_SHIFT) |
597 PMD_SECT_AP_WRITE | PMD_SECT_AP_READ |
598 PMD_TYPE_SECT;
599 if (cpu_arch <= CPU_ARCH_ARMv5TEJ)
600 pmdval |= PMD_BIT4;
601 pmd = pmd_off(pgd, i << PGDIR_SHIFT);
602 pmd[0] = __pmd(pmdval);
603 pmd[1] = __pmd(pmdval + (1 << (PGDIR_SHIFT - 1)));
604 flush_pmd_entry(pmd);
608 extern void _stext, _etext;
611 * Setup initial mappings. We use the page we allocated for zero page to hold
612 * the mappings, which will get overwritten by the vectors in traps_init().
613 * The mappings must be in virtual address order.
615 void __init memtable_init(struct meminfo *mi)
617 struct map_desc *init_maps, *p, *q;
618 unsigned long address = 0;
619 int i;
621 build_mem_type_table();
623 init_maps = p = alloc_bootmem_low_pages(PAGE_SIZE);
625 #ifdef CONFIG_XIP_KERNEL
626 p->physical = CONFIG_XIP_PHYS_ADDR & PMD_MASK;
627 p->virtual = (unsigned long)&_stext & PMD_MASK;
628 p->length = ((unsigned long)&_etext - p->virtual + ~PMD_MASK) & PMD_MASK;
629 p->type = MT_ROM;
630 p ++;
631 #endif
633 for (i = 0; i < mi->nr_banks; i++) {
634 if (mi->bank[i].size == 0)
635 continue;
637 p->physical = mi->bank[i].start;
638 p->virtual = __phys_to_virt(p->physical);
639 p->length = mi->bank[i].size;
640 p->type = MT_MEMORY;
641 p ++;
644 #ifdef FLUSH_BASE
645 p->physical = FLUSH_BASE_PHYS;
646 p->virtual = FLUSH_BASE;
647 p->length = PGDIR_SIZE;
648 p->type = MT_CACHECLEAN;
649 p ++;
650 #endif
652 #ifdef FLUSH_BASE_MINICACHE
653 p->physical = FLUSH_BASE_PHYS + PGDIR_SIZE;
654 p->virtual = FLUSH_BASE_MINICACHE;
655 p->length = PGDIR_SIZE;
656 p->type = MT_MINICLEAN;
657 p ++;
658 #endif
661 * Go through the initial mappings, but clear out any
662 * pgdir entries that are not in the description.
664 q = init_maps;
665 do {
666 if (address < q->virtual || q == p) {
667 clear_mapping(address);
668 address += PGDIR_SIZE;
669 } else {
670 create_mapping(q);
672 address = q->virtual + q->length;
673 address = (address + PGDIR_SIZE - 1) & PGDIR_MASK;
675 q ++;
677 } while (address != 0);
680 * Create a mapping for the machine vectors at the high-vectors
681 * location (0xffff0000). If we aren't using high-vectors, also
682 * create a mapping at the low-vectors virtual address.
684 init_maps->physical = virt_to_phys(init_maps);
685 init_maps->virtual = 0xffff0000;
686 init_maps->length = PAGE_SIZE;
687 init_maps->type = MT_HIGH_VECTORS;
688 create_mapping(init_maps);
690 if (!vectors_high()) {
691 init_maps->virtual = 0;
692 init_maps->type = MT_LOW_VECTORS;
693 create_mapping(init_maps);
696 flush_cache_all();
697 local_flush_tlb_all();
699 top_pmd = pmd_off_k(0xffff0000);
703 * Create the architecture specific mappings
705 void __init iotable_init(struct map_desc *io_desc, int nr)
707 int i;
709 for (i = 0; i < nr; i++)
710 create_mapping(io_desc + i);