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 struct vm_struct *area;
127 unsigned long new_prot_val;
130 void __iomem *ret_addr;
132 /* Don't allow wraparound or zero size */
133 last_addr = phys_addr + size - 1;
134 if (!size || last_addr < phys_addr)
137 if (!phys_addr_valid(phys_addr)) {
138 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
139 (unsigned long long)phys_addr);
145 * Don't remap the low PCI/ISA area, it's always mapped..
147 if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
148 return (__force void __iomem *)phys_to_virt(phys_addr);
151 * Don't allow anybody to remap normal RAM that we're using..
153 for (pfn = phys_addr >> PAGE_SHIFT;
154 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
157 int is_ram = page_is_ram(pfn);
159 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
161 WARN_ON_ONCE(is_ram);
165 * Mappings have to be page-aligned
167 offset = phys_addr & ~PAGE_MASK;
168 phys_addr &= PAGE_MASK;
169 size = PAGE_ALIGN(last_addr+1) - phys_addr;
171 retval = reserve_memtype(phys_addr, phys_addr + size,
172 prot_val, &new_prot_val);
174 pr_debug("Warning: reserve_memtype returned %d\n", retval);
178 if (prot_val != new_prot_val) {
180 * Do not fallback to certain memory types with certain
182 * - request is uc-, return cannot be write-back
183 * - request is uc-, return cannot be write-combine
184 * - request is write-combine, return cannot be write-back
186 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
187 (new_prot_val == _PAGE_CACHE_WB ||
188 new_prot_val == _PAGE_CACHE_WC)) ||
189 (prot_val == _PAGE_CACHE_WC &&
190 new_prot_val == _PAGE_CACHE_WB)) {
192 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
193 (unsigned long long)phys_addr,
194 (unsigned long long)(phys_addr + size),
195 prot_val, new_prot_val);
196 free_memtype(phys_addr, phys_addr + size);
199 prot_val = new_prot_val;
205 prot = PAGE_KERNEL_NOCACHE;
207 case _PAGE_CACHE_UC_MINUS:
208 prot = PAGE_KERNEL_UC_MINUS;
211 prot = PAGE_KERNEL_WC;
221 area = get_vm_area_caller(size, VM_IOREMAP, caller);
224 area->phys_addr = phys_addr;
225 vaddr = (unsigned long) area->addr;
226 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
227 free_memtype(phys_addr, phys_addr + size);
232 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
233 free_memtype(phys_addr, phys_addr + size);
238 ret_addr = (void __iomem *) (vaddr + offset);
239 mmiotrace_ioremap(phys_addr, size, ret_addr);
245 * ioremap_nocache - map bus memory into CPU space
246 * @offset: bus address of the memory
247 * @size: size of the resource to map
249 * ioremap_nocache performs a platform specific sequence of operations to
250 * make bus memory CPU accessible via the readb/readw/readl/writeb/
251 * writew/writel functions and the other mmio helpers. The returned
252 * address is not guaranteed to be usable directly as a virtual
255 * This version of ioremap ensures that the memory is marked uncachable
256 * on the CPU as well as honouring existing caching rules from things like
257 * the PCI bus. Note that there are other caches and buffers on many
258 * busses. In particular driver authors should read up on PCI writes
260 * It's useful if some control registers are in such an area and
261 * write combining or read caching is not desirable:
263 * Must be freed with iounmap.
265 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
268 * Ideally, this should be:
269 * pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
271 * Till we fix all X drivers to use ioremap_wc(), we will use
274 unsigned long val = _PAGE_CACHE_UC_MINUS;
276 return __ioremap_caller(phys_addr, size, val,
277 __builtin_return_address(0));
279 EXPORT_SYMBOL(ioremap_nocache);
282 * ioremap_wc - map memory into CPU space write combined
283 * @offset: bus address of the memory
284 * @size: size of the resource to map
286 * This version of ioremap ensures that the memory is marked write combining.
287 * Write combining allows faster writes to some hardware devices.
289 * Must be freed with iounmap.
291 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
294 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
295 __builtin_return_address(0));
297 return ioremap_nocache(phys_addr, size);
299 EXPORT_SYMBOL(ioremap_wc);
301 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
303 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
304 __builtin_return_address(0));
306 EXPORT_SYMBOL(ioremap_cache);
309 * iounmap - Free a IO remapping
310 * @addr: virtual address from ioremap_*
312 * Caller must ensure there is only one unmapping for the same pointer.
314 void iounmap(volatile void __iomem *addr)
316 struct vm_struct *p, *o;
318 if ((void __force *)addr <= high_memory)
322 * __ioremap special-cases the PCI/ISA range by not instantiating a
323 * vm_area and by simply returning an address into the kernel mapping
324 * of ISA space. So handle that here.
326 if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
327 addr < phys_to_virt(ISA_END_ADDRESS))
330 addr = (volatile void __iomem *)
331 (PAGE_MASK & (unsigned long __force)addr);
333 mmiotrace_iounmap(addr);
335 /* Use the vm area unlocked, assuming the caller
336 ensures there isn't another iounmap for the same address
337 in parallel. Reuse of the virtual address is prevented by
338 leaving it in the global lists until we're done with it.
339 cpa takes care of the direct mappings. */
340 read_lock(&vmlist_lock);
341 for (p = vmlist; p; p = p->next) {
345 read_unlock(&vmlist_lock);
348 printk(KERN_ERR "iounmap: bad address %p\n", addr);
353 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
355 /* Finally remove it */
356 o = remove_vm_area((void *)addr);
357 BUG_ON(p != o || o == NULL);
360 EXPORT_SYMBOL(iounmap);
363 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
366 void *xlate_dev_mem_ptr(unsigned long phys)
369 unsigned long start = phys & PAGE_MASK;
371 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
372 if (page_is_ram(start >> PAGE_SHIFT))
375 addr = (void *)ioremap(start, PAGE_SIZE);
377 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
382 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
384 if (page_is_ram(phys >> PAGE_SHIFT))
387 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
393 int __initdata early_ioremap_debug;
395 static int __init early_ioremap_debug_setup(char *str)
397 early_ioremap_debug = 1;
401 early_param("early_ioremap_debug", early_ioremap_debug_setup);
403 static __initdata int after_paging_init;
404 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
405 __section(.bss.page_aligned);
407 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
409 /* Don't assume we're using swapper_pg_dir at this point */
410 pgd_t *base = __va(read_cr3());
411 pgd_t *pgd = &base[pgd_index(addr)];
412 pud_t *pud = pud_offset(pgd, addr);
413 pmd_t *pmd = pmd_offset(pud, addr);
418 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
420 return &bm_pte[pte_index(addr)];
423 void __init early_ioremap_init(void)
427 if (early_ioremap_debug)
428 printk(KERN_INFO "early_ioremap_init()\n");
430 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
431 memset(bm_pte, 0, sizeof(bm_pte));
432 pmd_populate_kernel(&init_mm, pmd, bm_pte);
435 * The boot-ioremap range spans multiple pmds, for which
436 * we are not prepared:
438 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
440 printk(KERN_WARNING "pmd %p != %p\n",
441 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
442 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
443 fix_to_virt(FIX_BTMAP_BEGIN));
444 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
445 fix_to_virt(FIX_BTMAP_END));
447 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
448 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
453 void __init early_ioremap_clear(void)
457 if (early_ioremap_debug)
458 printk(KERN_INFO "early_ioremap_clear()\n");
460 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
462 paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
466 void __init early_ioremap_reset(void)
468 enum fixed_addresses idx;
469 unsigned long addr, phys;
472 after_paging_init = 1;
473 for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
474 addr = fix_to_virt(idx);
475 pte = early_ioremap_pte(addr);
476 if (pte_present(*pte)) {
477 phys = pte_val(*pte) & PAGE_MASK;
478 set_fixmap(idx, phys);
483 static void __init __early_set_fixmap(enum fixed_addresses idx,
484 unsigned long phys, pgprot_t flags)
486 unsigned long addr = __fix_to_virt(idx);
489 if (idx >= __end_of_fixed_addresses) {
493 pte = early_ioremap_pte(addr);
494 if (pgprot_val(flags))
495 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
497 pte_clear(NULL, addr, pte);
498 __flush_tlb_one(addr);
501 static inline void __init early_set_fixmap(enum fixed_addresses idx,
504 if (after_paging_init)
505 set_fixmap(idx, phys);
507 __early_set_fixmap(idx, phys, PAGE_KERNEL);
510 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
512 if (after_paging_init)
515 __early_set_fixmap(idx, 0, __pgprot(0));
519 int __initdata early_ioremap_nested;
521 static int __init check_early_ioremap_leak(void)
523 if (!early_ioremap_nested)
527 "Debug warning: early ioremap leak of %d areas detected.\n",
528 early_ioremap_nested);
530 "please boot with early_ioremap_debug and report the dmesg.\n");
535 late_initcall(check_early_ioremap_leak);
537 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
539 unsigned long offset, last_addr;
540 unsigned int nrpages, nesting;
541 enum fixed_addresses idx0, idx;
543 WARN_ON(system_state != SYSTEM_BOOTING);
545 nesting = early_ioremap_nested;
546 if (early_ioremap_debug) {
547 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
548 phys_addr, size, nesting);
552 /* Don't allow wraparound or zero size */
553 last_addr = phys_addr + size - 1;
554 if (!size || last_addr < phys_addr) {
559 if (nesting >= FIX_BTMAPS_NESTING) {
563 early_ioremap_nested++;
565 * Mappings have to be page-aligned
567 offset = phys_addr & ~PAGE_MASK;
568 phys_addr &= PAGE_MASK;
569 size = PAGE_ALIGN(last_addr) - phys_addr;
572 * Mappings have to fit in the FIX_BTMAP area.
574 nrpages = size >> PAGE_SHIFT;
575 if (nrpages > NR_FIX_BTMAPS) {
583 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
585 while (nrpages > 0) {
586 early_set_fixmap(idx, phys_addr);
587 phys_addr += PAGE_SIZE;
591 if (early_ioremap_debug)
592 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
594 return (void *) (offset + fix_to_virt(idx0));
597 void __init early_iounmap(void *addr, unsigned long size)
599 unsigned long virt_addr;
600 unsigned long offset;
601 unsigned int nrpages;
602 enum fixed_addresses idx;
603 unsigned int nesting;
605 nesting = --early_ioremap_nested;
606 WARN_ON(nesting < 0);
608 if (early_ioremap_debug) {
609 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
614 virt_addr = (unsigned long)addr;
615 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
619 offset = virt_addr & ~PAGE_MASK;
620 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
622 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
623 while (nrpages > 0) {
624 early_clear_fixmap(idx);
630 void __this_fixmap_does_not_exist(void)
635 #endif /* CONFIG_X86_32 */