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>
16 #include <asm/cacheflush.h>
18 #include <asm/fixmap.h>
19 #include <asm/pgtable.h>
20 #include <asm/tlbflush.h>
21 #include <asm/pgalloc.h>
26 unsigned long __phys_addr(unsigned long x)
28 if (x >= __START_KERNEL_map)
29 return x - __START_KERNEL_map + phys_base;
30 return x - PAGE_OFFSET;
32 EXPORT_SYMBOL(__phys_addr);
34 static inline int phys_addr_valid(unsigned long addr)
36 return addr < (1UL << boot_cpu_data.x86_phys_bits);
41 static inline int phys_addr_valid(unsigned long addr)
48 int page_is_ram(unsigned long pagenr)
50 resource_size_t addr, end;
54 * A special case is the first 4Kb of memory;
55 * This is a BIOS owned area, not kernel ram, but generally
56 * not listed as such in the E820 table.
62 * Second special case: Some BIOSen report the PC BIOS
63 * area (640->1Mb) as ram even though it is not.
65 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
66 pagenr < (BIOS_END >> PAGE_SHIFT))
69 for (i = 0; i < e820.nr_map; i++) {
73 if (e820.map[i].type != E820_RAM)
75 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
76 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
79 if ((pagenr >= addr) && (pagenr < end))
86 * Fix up the linear direct mapping of the kernel to avoid cache attribute
89 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
90 unsigned long prot_val)
92 unsigned long nrpages = size >> PAGE_SHIFT;
98 err = _set_memory_uc(vaddr, nrpages);
101 err = _set_memory_wc(vaddr, nrpages);
104 err = _set_memory_wb(vaddr, nrpages);
112 * Remap an arbitrary physical address space into the kernel virtual
113 * address space. Needed when the kernel wants to access high addresses
116 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
117 * have to convert them into an offset in a page-aligned mapping, but the
118 * caller shouldn't need to know that small detail.
120 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
121 unsigned long size, unsigned long prot_val, void *caller)
123 unsigned long pfn, offset, vaddr;
124 resource_size_t last_addr;
125 struct vm_struct *area;
126 unsigned long new_prot_val;
130 /* Don't allow wraparound or zero size */
131 last_addr = phys_addr + size - 1;
132 if (!size || last_addr < phys_addr)
135 if (!phys_addr_valid(phys_addr)) {
136 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
137 (unsigned long long)phys_addr);
143 * Don't remap the low PCI/ISA area, it's always mapped..
145 if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
146 return (__force void __iomem *)phys_to_virt(phys_addr);
149 * Don't allow anybody to remap normal RAM that we're using..
151 for (pfn = phys_addr >> PAGE_SHIFT; pfn < max_pfn_mapped &&
152 (pfn << PAGE_SHIFT) < last_addr; pfn++) {
154 int is_ram = page_is_ram(pfn);
156 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
158 WARN_ON_ONCE(is_ram);
162 * Mappings have to be page-aligned
164 offset = phys_addr & ~PAGE_MASK;
165 phys_addr &= PAGE_MASK;
166 size = PAGE_ALIGN(last_addr+1) - phys_addr;
168 retval = reserve_memtype(phys_addr, phys_addr + size,
169 prot_val, &new_prot_val);
171 pr_debug("Warning: reserve_memtype returned %d\n", retval);
175 if (prot_val != new_prot_val) {
177 * Do not fallback to certain memory types with certain
179 * - request is uc-, return cannot be write-back
180 * - request is uc-, return cannot be write-combine
181 * - request is write-combine, return cannot be write-back
183 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
184 (new_prot_val == _PAGE_CACHE_WB ||
185 new_prot_val == _PAGE_CACHE_WC)) ||
186 (prot_val == _PAGE_CACHE_WC &&
187 new_prot_val == _PAGE_CACHE_WB)) {
189 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
190 (unsigned long long)phys_addr,
191 (unsigned long long)(phys_addr + size),
192 prot_val, new_prot_val);
193 free_memtype(phys_addr, phys_addr + size);
196 prot_val = new_prot_val;
202 prot = PAGE_KERNEL_NOCACHE;
204 case _PAGE_CACHE_UC_MINUS:
205 prot = PAGE_KERNEL_UC_MINUS;
208 prot = PAGE_KERNEL_WC;
218 area = get_vm_area_caller(size, VM_IOREMAP, caller);
221 area->phys_addr = phys_addr;
222 vaddr = (unsigned long) area->addr;
223 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
224 free_memtype(phys_addr, phys_addr + size);
229 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
230 free_memtype(phys_addr, phys_addr + size);
235 return (void __iomem *) (vaddr + offset);
239 * ioremap_nocache - map bus memory into CPU space
240 * @offset: bus address of the memory
241 * @size: size of the resource to map
243 * ioremap_nocache performs a platform specific sequence of operations to
244 * make bus memory CPU accessible via the readb/readw/readl/writeb/
245 * writew/writel functions and the other mmio helpers. The returned
246 * address is not guaranteed to be usable directly as a virtual
249 * This version of ioremap ensures that the memory is marked uncachable
250 * on the CPU as well as honouring existing caching rules from things like
251 * the PCI bus. Note that there are other caches and buffers on many
252 * busses. In particular driver authors should read up on PCI writes
254 * It's useful if some control registers are in such an area and
255 * write combining or read caching is not desirable:
257 * Must be freed with iounmap.
259 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
262 * Ideally, this should be:
263 * pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
265 * Till we fix all X drivers to use ioremap_wc(), we will use
268 unsigned long val = _PAGE_CACHE_UC_MINUS;
270 return __ioremap_caller(phys_addr, size, val,
271 __builtin_return_address(0));
273 EXPORT_SYMBOL(ioremap_nocache);
276 * ioremap_wc - map memory into CPU space write combined
277 * @offset: bus address of the memory
278 * @size: size of the resource to map
280 * This version of ioremap ensures that the memory is marked write combining.
281 * Write combining allows faster writes to some hardware devices.
283 * Must be freed with iounmap.
285 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
288 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
289 __builtin_return_address(0));
291 return ioremap_nocache(phys_addr, size);
293 EXPORT_SYMBOL(ioremap_wc);
295 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
297 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
298 __builtin_return_address(0));
300 EXPORT_SYMBOL(ioremap_cache);
303 * iounmap - Free a IO remapping
304 * @addr: virtual address from ioremap_*
306 * Caller must ensure there is only one unmapping for the same pointer.
308 void iounmap(volatile void __iomem *addr)
310 struct vm_struct *p, *o;
312 if ((void __force *)addr <= high_memory)
316 * __ioremap special-cases the PCI/ISA range by not instantiating a
317 * vm_area and by simply returning an address into the kernel mapping
318 * of ISA space. So handle that here.
320 if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
321 addr < phys_to_virt(ISA_END_ADDRESS))
324 addr = (volatile void __iomem *)
325 (PAGE_MASK & (unsigned long __force)addr);
327 /* Use the vm area unlocked, assuming the caller
328 ensures there isn't another iounmap for the same address
329 in parallel. Reuse of the virtual address is prevented by
330 leaving it in the global lists until we're done with it.
331 cpa takes care of the direct mappings. */
332 read_lock(&vmlist_lock);
333 for (p = vmlist; p; p = p->next) {
337 read_unlock(&vmlist_lock);
340 printk(KERN_ERR "iounmap: bad address %p\n", addr);
345 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
347 /* Finally remove it */
348 o = remove_vm_area((void *)addr);
349 BUG_ON(p != o || o == NULL);
352 EXPORT_SYMBOL(iounmap);
355 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
358 void *xlate_dev_mem_ptr(unsigned long phys)
361 unsigned long start = phys & PAGE_MASK;
363 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
364 if (page_is_ram(start >> PAGE_SHIFT))
367 addr = (void *)ioremap(start, PAGE_SIZE);
369 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
374 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
376 if (page_is_ram(phys >> PAGE_SHIFT))
379 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
385 int __initdata early_ioremap_debug;
387 static int __init early_ioremap_debug_setup(char *str)
389 early_ioremap_debug = 1;
393 early_param("early_ioremap_debug", early_ioremap_debug_setup);
395 static __initdata int after_paging_init;
396 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
397 __section(.bss.page_aligned);
399 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
401 /* Don't assume we're using swapper_pg_dir at this point */
402 pgd_t *base = __va(read_cr3());
403 pgd_t *pgd = &base[pgd_index(addr)];
404 pud_t *pud = pud_offset(pgd, addr);
405 pmd_t *pmd = pmd_offset(pud, addr);
410 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
412 return &bm_pte[pte_index(addr)];
415 void __init early_ioremap_init(void)
419 if (early_ioremap_debug)
420 printk(KERN_INFO "early_ioremap_init()\n");
422 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
423 memset(bm_pte, 0, sizeof(bm_pte));
424 pmd_populate_kernel(&init_mm, pmd, bm_pte);
427 * The boot-ioremap range spans multiple pmds, for which
428 * we are not prepared:
430 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
432 printk(KERN_WARNING "pmd %p != %p\n",
433 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
434 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
435 fix_to_virt(FIX_BTMAP_BEGIN));
436 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
437 fix_to_virt(FIX_BTMAP_END));
439 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
440 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
445 void __init early_ioremap_clear(void)
449 if (early_ioremap_debug)
450 printk(KERN_INFO "early_ioremap_clear()\n");
452 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
454 paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
458 void __init early_ioremap_reset(void)
460 enum fixed_addresses idx;
461 unsigned long addr, phys;
464 after_paging_init = 1;
465 for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
466 addr = fix_to_virt(idx);
467 pte = early_ioremap_pte(addr);
468 if (pte_present(*pte)) {
469 phys = pte_val(*pte) & PAGE_MASK;
470 set_fixmap(idx, phys);
475 static void __init __early_set_fixmap(enum fixed_addresses idx,
476 unsigned long phys, pgprot_t flags)
478 unsigned long addr = __fix_to_virt(idx);
481 if (idx >= __end_of_fixed_addresses) {
485 pte = early_ioremap_pte(addr);
486 if (pgprot_val(flags))
487 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
489 pte_clear(NULL, addr, pte);
490 __flush_tlb_one(addr);
493 static inline void __init early_set_fixmap(enum fixed_addresses idx,
496 if (after_paging_init)
497 set_fixmap(idx, phys);
499 __early_set_fixmap(idx, phys, PAGE_KERNEL);
502 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
504 if (after_paging_init)
507 __early_set_fixmap(idx, 0, __pgprot(0));
511 int __initdata early_ioremap_nested;
513 static int __init check_early_ioremap_leak(void)
515 if (!early_ioremap_nested)
519 "Debug warning: early ioremap leak of %d areas detected.\n",
520 early_ioremap_nested);
522 "please boot with early_ioremap_debug and report the dmesg.\n");
527 late_initcall(check_early_ioremap_leak);
529 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
531 unsigned long offset, last_addr;
532 unsigned int nrpages, nesting;
533 enum fixed_addresses idx0, idx;
535 WARN_ON(system_state != SYSTEM_BOOTING);
537 nesting = early_ioremap_nested;
538 if (early_ioremap_debug) {
539 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
540 phys_addr, size, nesting);
544 /* Don't allow wraparound or zero size */
545 last_addr = phys_addr + size - 1;
546 if (!size || last_addr < phys_addr) {
551 if (nesting >= FIX_BTMAPS_NESTING) {
555 early_ioremap_nested++;
557 * Mappings have to be page-aligned
559 offset = phys_addr & ~PAGE_MASK;
560 phys_addr &= PAGE_MASK;
561 size = PAGE_ALIGN(last_addr) - phys_addr;
564 * Mappings have to fit in the FIX_BTMAP area.
566 nrpages = size >> PAGE_SHIFT;
567 if (nrpages > NR_FIX_BTMAPS) {
575 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
577 while (nrpages > 0) {
578 early_set_fixmap(idx, phys_addr);
579 phys_addr += PAGE_SIZE;
583 if (early_ioremap_debug)
584 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
586 return (void *) (offset + fix_to_virt(idx0));
589 void __init early_iounmap(void *addr, unsigned long size)
591 unsigned long virt_addr;
592 unsigned long offset;
593 unsigned int nrpages;
594 enum fixed_addresses idx;
595 unsigned int nesting;
597 nesting = --early_ioremap_nested;
598 WARN_ON(nesting < 0);
600 if (early_ioremap_debug) {
601 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
606 virt_addr = (unsigned long)addr;
607 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
611 offset = virt_addr & ~PAGE_MASK;
612 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
614 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
615 while (nrpages > 0) {
616 early_clear_fixmap(idx);
622 void __this_fixmap_does_not_exist(void)
627 #endif /* CONFIG_X86_32 */