2 * Re-map IO memory to kernel address space so that we can access it.
3 * This is needed for high PCI addresses that aren't mapped in the
4 * 640k-1MB IO memory area on PC's
6 * (C) Copyright 1995 1996 Linus Torvalds
9 #include <linux/bootmem.h>
10 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/mmiotrace.h>
17 #include <asm/cacheflush.h>
19 #include <asm/fixmap.h>
20 #include <asm/pgtable.h>
21 #include <asm/tlbflush.h>
22 #include <asm/pgalloc.h>
27 unsigned long __phys_addr(unsigned long x)
29 if (x >= __START_KERNEL_map)
30 return x - __START_KERNEL_map + phys_base;
31 return x - PAGE_OFFSET;
33 EXPORT_SYMBOL(__phys_addr);
35 static inline int phys_addr_valid(unsigned long addr)
37 return addr < (1UL << boot_cpu_data.x86_phys_bits);
42 static inline int phys_addr_valid(unsigned long addr)
49 int page_is_ram(unsigned long pagenr)
51 resource_size_t addr, end;
55 * A special case is the first 4Kb of memory;
56 * This is a BIOS owned area, not kernel ram, but generally
57 * not listed as such in the E820 table.
63 * Second special case: Some BIOSen report the PC BIOS
64 * area (640->1Mb) as ram even though it is not.
66 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
67 pagenr < (BIOS_END >> PAGE_SHIFT))
70 for (i = 0; i < e820.nr_map; i++) {
74 if (e820.map[i].type != E820_RAM)
76 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
77 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
80 if ((pagenr >= addr) && (pagenr < end))
87 * Fix up the linear direct mapping of the kernel to avoid cache attribute
90 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
91 unsigned long prot_val)
93 unsigned long nrpages = size >> PAGE_SHIFT;
99 err = _set_memory_uc(vaddr, nrpages);
102 err = _set_memory_wc(vaddr, nrpages);
105 err = _set_memory_wb(vaddr, nrpages);
113 * Remap an arbitrary physical address space into the kernel virtual
114 * address space. Needed when the kernel wants to access high addresses
117 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
118 * have to convert them into an offset in a page-aligned mapping, but the
119 * caller shouldn't need to know that small detail.
121 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
122 unsigned long size, unsigned long prot_val, void *caller)
124 unsigned long pfn, offset, vaddr;
125 resource_size_t last_addr;
126 const resource_size_t unaligned_phys_addr = phys_addr;
127 const unsigned long unaligned_size = size;
128 struct vm_struct *area;
129 unsigned long new_prot_val;
132 void __iomem *ret_addr;
134 /* Don't allow wraparound or zero size */
135 last_addr = phys_addr + size - 1;
136 if (!size || last_addr < phys_addr)
139 if (!phys_addr_valid(phys_addr)) {
140 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
141 (unsigned long long)phys_addr);
147 * Don't remap the low PCI/ISA area, it's always mapped..
149 if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
150 return (__force void __iomem *)phys_to_virt(phys_addr);
153 * Don't allow anybody to remap normal RAM that we're using..
155 for (pfn = phys_addr >> PAGE_SHIFT;
156 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
159 int is_ram = page_is_ram(pfn);
161 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
163 WARN_ON_ONCE(is_ram);
167 * Mappings have to be page-aligned
169 offset = phys_addr & ~PAGE_MASK;
170 phys_addr &= PAGE_MASK;
171 size = PAGE_ALIGN(last_addr+1) - phys_addr;
173 retval = reserve_memtype(phys_addr, phys_addr + size,
174 prot_val, &new_prot_val);
176 pr_debug("Warning: reserve_memtype returned %d\n", retval);
180 if (prot_val != new_prot_val) {
182 * Do not fallback to certain memory types with certain
184 * - request is uc-, return cannot be write-back
185 * - request is uc-, return cannot be write-combine
186 * - request is write-combine, return cannot be write-back
188 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
189 (new_prot_val == _PAGE_CACHE_WB ||
190 new_prot_val == _PAGE_CACHE_WC)) ||
191 (prot_val == _PAGE_CACHE_WC &&
192 new_prot_val == _PAGE_CACHE_WB)) {
194 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
195 (unsigned long long)phys_addr,
196 (unsigned long long)(phys_addr + size),
197 prot_val, new_prot_val);
198 free_memtype(phys_addr, phys_addr + size);
201 prot_val = new_prot_val;
207 prot = PAGE_KERNEL_NOCACHE;
209 case _PAGE_CACHE_UC_MINUS:
210 prot = PAGE_KERNEL_UC_MINUS;
213 prot = PAGE_KERNEL_WC;
223 area = get_vm_area_caller(size, VM_IOREMAP, caller);
226 area->phys_addr = phys_addr;
227 vaddr = (unsigned long) area->addr;
228 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
229 free_memtype(phys_addr, phys_addr + size);
234 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
235 free_memtype(phys_addr, phys_addr + size);
240 ret_addr = (void __iomem *) (vaddr + offset);
241 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
247 * ioremap_nocache - map bus memory into CPU space
248 * @offset: bus address of the memory
249 * @size: size of the resource to map
251 * ioremap_nocache performs a platform specific sequence of operations to
252 * make bus memory CPU accessible via the readb/readw/readl/writeb/
253 * writew/writel functions and the other mmio helpers. The returned
254 * address is not guaranteed to be usable directly as a virtual
257 * This version of ioremap ensures that the memory is marked uncachable
258 * on the CPU as well as honouring existing caching rules from things like
259 * the PCI bus. Note that there are other caches and buffers on many
260 * busses. In particular driver authors should read up on PCI writes
262 * It's useful if some control registers are in such an area and
263 * write combining or read caching is not desirable:
265 * Must be freed with iounmap.
267 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
270 * Ideally, this should be:
271 * pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
273 * Till we fix all X drivers to use ioremap_wc(), we will use
276 unsigned long val = _PAGE_CACHE_UC_MINUS;
278 return __ioremap_caller(phys_addr, size, val,
279 __builtin_return_address(0));
281 EXPORT_SYMBOL(ioremap_nocache);
284 * ioremap_wc - map memory into CPU space write combined
285 * @offset: bus address of the memory
286 * @size: size of the resource to map
288 * This version of ioremap ensures that the memory is marked write combining.
289 * Write combining allows faster writes to some hardware devices.
291 * Must be freed with iounmap.
293 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
296 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
297 __builtin_return_address(0));
299 return ioremap_nocache(phys_addr, size);
301 EXPORT_SYMBOL(ioremap_wc);
303 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
305 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
306 __builtin_return_address(0));
308 EXPORT_SYMBOL(ioremap_cache);
311 * iounmap - Free a IO remapping
312 * @addr: virtual address from ioremap_*
314 * Caller must ensure there is only one unmapping for the same pointer.
316 void iounmap(volatile void __iomem *addr)
318 struct vm_struct *p, *o;
320 if ((void __force *)addr <= high_memory)
324 * __ioremap special-cases the PCI/ISA range by not instantiating a
325 * vm_area and by simply returning an address into the kernel mapping
326 * of ISA space. So handle that here.
328 if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
329 addr < phys_to_virt(ISA_END_ADDRESS))
332 addr = (volatile void __iomem *)
333 (PAGE_MASK & (unsigned long __force)addr);
335 mmiotrace_iounmap(addr);
337 /* Use the vm area unlocked, assuming the caller
338 ensures there isn't another iounmap for the same address
339 in parallel. Reuse of the virtual address is prevented by
340 leaving it in the global lists until we're done with it.
341 cpa takes care of the direct mappings. */
342 read_lock(&vmlist_lock);
343 for (p = vmlist; p; p = p->next) {
347 read_unlock(&vmlist_lock);
350 printk(KERN_ERR "iounmap: bad address %p\n", addr);
355 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
357 /* Finally remove it */
358 o = remove_vm_area((void *)addr);
359 BUG_ON(p != o || o == NULL);
362 EXPORT_SYMBOL(iounmap);
365 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
368 void *xlate_dev_mem_ptr(unsigned long phys)
371 unsigned long start = phys & PAGE_MASK;
373 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
374 if (page_is_ram(start >> PAGE_SHIFT))
377 addr = (void *)ioremap(start, PAGE_SIZE);
379 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
384 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
386 if (page_is_ram(phys >> PAGE_SHIFT))
389 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
395 int __initdata early_ioremap_debug;
397 static int __init early_ioremap_debug_setup(char *str)
399 early_ioremap_debug = 1;
403 early_param("early_ioremap_debug", early_ioremap_debug_setup);
405 static __initdata int after_paging_init;
406 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
407 __section(.bss.page_aligned);
409 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
411 /* Don't assume we're using swapper_pg_dir at this point */
412 pgd_t *base = __va(read_cr3());
413 pgd_t *pgd = &base[pgd_index(addr)];
414 pud_t *pud = pud_offset(pgd, addr);
415 pmd_t *pmd = pmd_offset(pud, addr);
420 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
422 return &bm_pte[pte_index(addr)];
425 void __init early_ioremap_init(void)
429 if (early_ioremap_debug)
430 printk(KERN_INFO "early_ioremap_init()\n");
432 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
433 memset(bm_pte, 0, sizeof(bm_pte));
434 pmd_populate_kernel(&init_mm, pmd, bm_pte);
437 * The boot-ioremap range spans multiple pmds, for which
438 * we are not prepared:
440 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
442 printk(KERN_WARNING "pmd %p != %p\n",
443 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
444 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
445 fix_to_virt(FIX_BTMAP_BEGIN));
446 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
447 fix_to_virt(FIX_BTMAP_END));
449 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
450 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
455 void __init early_ioremap_clear(void)
459 if (early_ioremap_debug)
460 printk(KERN_INFO "early_ioremap_clear()\n");
462 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
464 paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
468 void __init early_ioremap_reset(void)
470 enum fixed_addresses idx;
471 unsigned long addr, phys;
474 after_paging_init = 1;
475 for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
476 addr = fix_to_virt(idx);
477 pte = early_ioremap_pte(addr);
478 if (pte_present(*pte)) {
479 phys = pte_val(*pte) & PAGE_MASK;
480 set_fixmap(idx, phys);
485 static void __init __early_set_fixmap(enum fixed_addresses idx,
486 unsigned long phys, pgprot_t flags)
488 unsigned long addr = __fix_to_virt(idx);
491 if (idx >= __end_of_fixed_addresses) {
495 pte = early_ioremap_pte(addr);
496 if (pgprot_val(flags))
497 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
499 pte_clear(NULL, addr, pte);
500 __flush_tlb_one(addr);
503 static inline void __init early_set_fixmap(enum fixed_addresses idx,
506 if (after_paging_init)
507 set_fixmap(idx, phys);
509 __early_set_fixmap(idx, phys, PAGE_KERNEL);
512 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
514 if (after_paging_init)
517 __early_set_fixmap(idx, 0, __pgprot(0));
521 int __initdata early_ioremap_nested;
523 static int __init check_early_ioremap_leak(void)
525 if (!early_ioremap_nested)
529 "Debug warning: early ioremap leak of %d areas detected.\n",
530 early_ioremap_nested);
532 "please boot with early_ioremap_debug and report the dmesg.\n");
537 late_initcall(check_early_ioremap_leak);
539 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
541 unsigned long offset, last_addr;
542 unsigned int nrpages, nesting;
543 enum fixed_addresses idx0, idx;
545 WARN_ON(system_state != SYSTEM_BOOTING);
547 nesting = early_ioremap_nested;
548 if (early_ioremap_debug) {
549 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
550 phys_addr, size, nesting);
554 /* Don't allow wraparound or zero size */
555 last_addr = phys_addr + size - 1;
556 if (!size || last_addr < phys_addr) {
561 if (nesting >= FIX_BTMAPS_NESTING) {
565 early_ioremap_nested++;
567 * Mappings have to be page-aligned
569 offset = phys_addr & ~PAGE_MASK;
570 phys_addr &= PAGE_MASK;
571 size = PAGE_ALIGN(last_addr) - phys_addr;
574 * Mappings have to fit in the FIX_BTMAP area.
576 nrpages = size >> PAGE_SHIFT;
577 if (nrpages > NR_FIX_BTMAPS) {
585 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
587 while (nrpages > 0) {
588 early_set_fixmap(idx, phys_addr);
589 phys_addr += PAGE_SIZE;
593 if (early_ioremap_debug)
594 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
596 return (void *) (offset + fix_to_virt(idx0));
599 void __init early_iounmap(void *addr, unsigned long size)
601 unsigned long virt_addr;
602 unsigned long offset;
603 unsigned int nrpages;
604 enum fixed_addresses idx;
607 nesting = --early_ioremap_nested;
608 if (WARN_ON(nesting < 0))
611 if (early_ioremap_debug) {
612 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
617 virt_addr = (unsigned long)addr;
618 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
622 offset = virt_addr & ~PAGE_MASK;
623 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
625 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
626 while (nrpages > 0) {
627 early_clear_fixmap(idx);
633 void __this_fixmap_does_not_exist(void)
638 #endif /* CONFIG_X86_32 */