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 static inline int phys_addr_valid(unsigned long addr)
29 return addr < (1UL << boot_cpu_data.x86_phys_bits);
32 unsigned long __phys_addr(unsigned long x)
34 if (x >= __START_KERNEL_map) {
35 x -= __START_KERNEL_map;
36 VIRTUAL_BUG_ON(x >= KERNEL_IMAGE_SIZE);
39 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
41 VIRTUAL_BUG_ON(system_state == SYSTEM_BOOTING ? x > MAXMEM :
46 EXPORT_SYMBOL(__phys_addr);
48 bool __virt_addr_valid(unsigned long x)
50 if (x >= __START_KERNEL_map) {
51 x -= __START_KERNEL_map;
52 if (x >= KERNEL_IMAGE_SIZE)
59 if (system_state == SYSTEM_BOOTING ?
60 x > MAXMEM : !phys_addr_valid(x)) {
65 return pfn_valid(x >> PAGE_SHIFT);
67 EXPORT_SYMBOL(__virt_addr_valid);
71 static inline int phys_addr_valid(unsigned long addr)
76 #ifdef CONFIG_DEBUG_VIRTUAL
77 unsigned long __phys_addr(unsigned long x)
79 /* VMALLOC_* aren't constants; not available at the boot time */
80 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
81 VIRTUAL_BUG_ON(system_state != SYSTEM_BOOTING &&
82 is_vmalloc_addr((void *) x));
83 return x - PAGE_OFFSET;
85 EXPORT_SYMBOL(__phys_addr);
88 bool __virt_addr_valid(unsigned long x)
92 if (system_state != SYSTEM_BOOTING && is_vmalloc_addr((void *) x))
94 return pfn_valid((x - PAGE_OFFSET) >> PAGE_SHIFT);
96 EXPORT_SYMBOL(__virt_addr_valid);
100 int page_is_ram(unsigned long pagenr)
102 resource_size_t addr, end;
106 * A special case is the first 4Kb of memory;
107 * This is a BIOS owned area, not kernel ram, but generally
108 * not listed as such in the E820 table.
114 * Second special case: Some BIOSen report the PC BIOS
115 * area (640->1Mb) as ram even though it is not.
117 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
118 pagenr < (BIOS_END >> PAGE_SHIFT))
121 for (i = 0; i < e820.nr_map; i++) {
125 if (e820.map[i].type != E820_RAM)
127 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
128 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
131 if ((pagenr >= addr) && (pagenr < end))
137 int pagerange_is_ram(unsigned long start, unsigned long end)
139 int ram_page = 0, not_rampage = 0;
140 unsigned long page_nr;
142 for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
144 if (page_is_ram(page_nr))
149 if (ram_page == not_rampage)
157 * Fix up the linear direct mapping of the kernel to avoid cache attribute
160 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
161 unsigned long prot_val)
163 unsigned long nrpages = size >> PAGE_SHIFT;
169 err = _set_memory_uc(vaddr, nrpages);
172 err = _set_memory_wc(vaddr, nrpages);
175 err = _set_memory_wb(vaddr, nrpages);
183 * Remap an arbitrary physical address space into the kernel virtual
184 * address space. Needed when the kernel wants to access high addresses
187 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
188 * have to convert them into an offset in a page-aligned mapping, but the
189 * caller shouldn't need to know that small detail.
191 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
192 unsigned long size, unsigned long prot_val, void *caller)
194 unsigned long pfn, offset, vaddr;
195 resource_size_t last_addr;
196 const resource_size_t unaligned_phys_addr = phys_addr;
197 const unsigned long unaligned_size = size;
198 struct vm_struct *area;
199 unsigned long new_prot_val;
202 void __iomem *ret_addr;
204 /* Don't allow wraparound or zero size */
205 last_addr = phys_addr + size - 1;
206 if (!size || last_addr < phys_addr)
209 if (!phys_addr_valid(phys_addr)) {
210 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
211 (unsigned long long)phys_addr);
217 * Don't remap the low PCI/ISA area, it's always mapped..
219 if (is_ISA_range(phys_addr, last_addr))
220 return (__force void __iomem *)phys_to_virt(phys_addr);
223 * Don't allow anybody to remap normal RAM that we're using..
225 for (pfn = phys_addr >> PAGE_SHIFT;
226 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
229 int is_ram = page_is_ram(pfn);
231 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
233 WARN_ON_ONCE(is_ram);
237 * Mappings have to be page-aligned
239 offset = phys_addr & ~PAGE_MASK;
240 phys_addr &= PAGE_MASK;
241 size = PAGE_ALIGN(last_addr+1) - phys_addr;
243 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
244 prot_val, &new_prot_val);
246 pr_debug("Warning: reserve_memtype returned %d\n", retval);
250 if (prot_val != new_prot_val) {
252 * Do not fallback to certain memory types with certain
254 * - request is uc-, return cannot be write-back
255 * - request is uc-, return cannot be write-combine
256 * - request is write-combine, return cannot be write-back
258 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
259 (new_prot_val == _PAGE_CACHE_WB ||
260 new_prot_val == _PAGE_CACHE_WC)) ||
261 (prot_val == _PAGE_CACHE_WC &&
262 new_prot_val == _PAGE_CACHE_WB)) {
264 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
265 (unsigned long long)phys_addr,
266 (unsigned long long)(phys_addr + size),
267 prot_val, new_prot_val);
268 free_memtype(phys_addr, phys_addr + size);
271 prot_val = new_prot_val;
277 prot = PAGE_KERNEL_IO_NOCACHE;
279 case _PAGE_CACHE_UC_MINUS:
280 prot = PAGE_KERNEL_IO_UC_MINUS;
283 prot = PAGE_KERNEL_IO_WC;
286 prot = PAGE_KERNEL_IO;
293 area = get_vm_area_caller(size, VM_IOREMAP, caller);
296 area->phys_addr = phys_addr;
297 vaddr = (unsigned long) area->addr;
298 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
299 free_memtype(phys_addr, phys_addr + size);
304 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
305 free_memtype(phys_addr, phys_addr + size);
310 ret_addr = (void __iomem *) (vaddr + offset);
311 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
317 * ioremap_nocache - map bus memory into CPU space
318 * @offset: bus address of the memory
319 * @size: size of the resource to map
321 * ioremap_nocache performs a platform specific sequence of operations to
322 * make bus memory CPU accessible via the readb/readw/readl/writeb/
323 * writew/writel functions and the other mmio helpers. The returned
324 * address is not guaranteed to be usable directly as a virtual
327 * This version of ioremap ensures that the memory is marked uncachable
328 * on the CPU as well as honouring existing caching rules from things like
329 * the PCI bus. Note that there are other caches and buffers on many
330 * busses. In particular driver authors should read up on PCI writes
332 * It's useful if some control registers are in such an area and
333 * write combining or read caching is not desirable:
335 * Must be freed with iounmap.
337 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
340 * Ideally, this should be:
341 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
343 * Till we fix all X drivers to use ioremap_wc(), we will use
346 unsigned long val = _PAGE_CACHE_UC_MINUS;
348 return __ioremap_caller(phys_addr, size, val,
349 __builtin_return_address(0));
351 EXPORT_SYMBOL(ioremap_nocache);
354 * ioremap_wc - map memory into CPU space write combined
355 * @offset: bus address of the memory
356 * @size: size of the resource to map
358 * This version of ioremap ensures that the memory is marked write combining.
359 * Write combining allows faster writes to some hardware devices.
361 * Must be freed with iounmap.
363 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
366 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
367 __builtin_return_address(0));
369 return ioremap_nocache(phys_addr, size);
371 EXPORT_SYMBOL(ioremap_wc);
373 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
375 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
376 __builtin_return_address(0));
378 EXPORT_SYMBOL(ioremap_cache);
380 static void __iomem *ioremap_default(resource_size_t phys_addr,
388 * - WB for WB-able memory and no other conflicting mappings
389 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
390 * - Inherit from confliting mappings otherwise
392 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
396 ret = (void *) __ioremap_caller(phys_addr, size, flags,
397 __builtin_return_address(0));
399 free_memtype(phys_addr, phys_addr + size);
400 return (void __iomem *)ret;
403 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
404 unsigned long prot_val)
406 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
407 __builtin_return_address(0));
409 EXPORT_SYMBOL(ioremap_prot);
412 * iounmap - Free a IO remapping
413 * @addr: virtual address from ioremap_*
415 * Caller must ensure there is only one unmapping for the same pointer.
417 void iounmap(volatile void __iomem *addr)
419 struct vm_struct *p, *o;
421 if ((void __force *)addr <= high_memory)
425 * __ioremap special-cases the PCI/ISA range by not instantiating a
426 * vm_area and by simply returning an address into the kernel mapping
427 * of ISA space. So handle that here.
429 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
430 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
433 addr = (volatile void __iomem *)
434 (PAGE_MASK & (unsigned long __force)addr);
436 mmiotrace_iounmap(addr);
438 /* Use the vm area unlocked, assuming the caller
439 ensures there isn't another iounmap for the same address
440 in parallel. Reuse of the virtual address is prevented by
441 leaving it in the global lists until we're done with it.
442 cpa takes care of the direct mappings. */
443 read_lock(&vmlist_lock);
444 for (p = vmlist; p; p = p->next) {
445 if (p->addr == (void __force *)addr)
448 read_unlock(&vmlist_lock);
451 printk(KERN_ERR "iounmap: bad address %p\n", addr);
456 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
458 /* Finally remove it */
459 o = remove_vm_area((void __force *)addr);
460 BUG_ON(p != o || o == NULL);
463 EXPORT_SYMBOL(iounmap);
466 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
469 void *xlate_dev_mem_ptr(unsigned long phys)
472 unsigned long start = phys & PAGE_MASK;
474 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
475 if (page_is_ram(start >> PAGE_SHIFT))
478 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
480 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
485 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
487 if (page_is_ram(phys >> PAGE_SHIFT))
490 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
494 static int __initdata early_ioremap_debug;
496 static int __init early_ioremap_debug_setup(char *str)
498 early_ioremap_debug = 1;
502 early_param("early_ioremap_debug", early_ioremap_debug_setup);
504 static __initdata int after_paging_init;
505 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
507 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
509 /* Don't assume we're using swapper_pg_dir at this point */
510 pgd_t *base = __va(read_cr3());
511 pgd_t *pgd = &base[pgd_index(addr)];
512 pud_t *pud = pud_offset(pgd, addr);
513 pmd_t *pmd = pmd_offset(pud, addr);
518 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
520 return &bm_pte[pte_index(addr)];
523 void __init early_ioremap_init(void)
527 if (early_ioremap_debug)
528 printk(KERN_INFO "early_ioremap_init()\n");
530 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
531 memset(bm_pte, 0, sizeof(bm_pte));
532 pmd_populate_kernel(&init_mm, pmd, bm_pte);
535 * The boot-ioremap range spans multiple pmds, for which
536 * we are not prepared:
538 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
540 printk(KERN_WARNING "pmd %p != %p\n",
541 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
542 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
543 fix_to_virt(FIX_BTMAP_BEGIN));
544 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
545 fix_to_virt(FIX_BTMAP_END));
547 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
548 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
553 void __init early_ioremap_clear(void)
557 if (early_ioremap_debug)
558 printk(KERN_INFO "early_ioremap_clear()\n");
560 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
562 paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
566 void __init early_ioremap_reset(void)
568 enum fixed_addresses idx;
569 unsigned long addr, phys;
572 after_paging_init = 1;
573 for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
574 addr = fix_to_virt(idx);
575 pte = early_ioremap_pte(addr);
576 if (pte_present(*pte)) {
577 phys = pte_val(*pte) & PAGE_MASK;
578 set_fixmap(idx, phys);
583 static void __init __early_set_fixmap(enum fixed_addresses idx,
584 unsigned long phys, pgprot_t flags)
586 unsigned long addr = __fix_to_virt(idx);
589 if (idx >= __end_of_fixed_addresses) {
593 pte = early_ioremap_pte(addr);
595 if (pgprot_val(flags))
596 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
598 pte_clear(&init_mm, addr, pte);
599 __flush_tlb_one(addr);
602 static inline void __init early_set_fixmap(enum fixed_addresses idx,
603 unsigned long phys, pgprot_t prot)
605 if (after_paging_init)
606 __set_fixmap(idx, phys, prot);
608 __early_set_fixmap(idx, phys, prot);
611 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
613 if (after_paging_init)
616 __early_set_fixmap(idx, 0, __pgprot(0));
620 static int __initdata early_ioremap_nested;
622 static int __init check_early_ioremap_leak(void)
624 if (!early_ioremap_nested)
627 "Debug warning: early ioremap leak of %d areas detected.\n",
628 early_ioremap_nested);
630 "please boot with early_ioremap_debug and report the dmesg.\n");
634 late_initcall(check_early_ioremap_leak);
636 static void __init *__early_ioremap(unsigned long phys_addr, unsigned long size, pgprot_t prot)
638 unsigned long offset, last_addr;
639 unsigned int nrpages, nesting;
640 enum fixed_addresses idx0, idx;
642 WARN_ON(system_state != SYSTEM_BOOTING);
644 nesting = early_ioremap_nested;
645 if (early_ioremap_debug) {
646 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
647 phys_addr, size, nesting);
651 /* Don't allow wraparound or zero size */
652 last_addr = phys_addr + size - 1;
653 if (!size || last_addr < phys_addr) {
658 if (nesting >= FIX_BTMAPS_NESTING) {
662 early_ioremap_nested++;
664 * Mappings have to be page-aligned
666 offset = phys_addr & ~PAGE_MASK;
667 phys_addr &= PAGE_MASK;
668 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
671 * Mappings have to fit in the FIX_BTMAP area.
673 nrpages = size >> PAGE_SHIFT;
674 if (nrpages > NR_FIX_BTMAPS) {
682 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
684 while (nrpages > 0) {
685 early_set_fixmap(idx, phys_addr, prot);
686 phys_addr += PAGE_SIZE;
690 if (early_ioremap_debug)
691 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
693 return (void *) (offset + fix_to_virt(idx0));
696 /* Remap an IO device */
697 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
699 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
703 void __init *early_memremap(unsigned long phys_addr, unsigned long size)
705 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
708 void __init early_iounmap(void *addr, unsigned long size)
710 unsigned long virt_addr;
711 unsigned long offset;
712 unsigned int nrpages;
713 enum fixed_addresses idx;
716 nesting = --early_ioremap_nested;
717 if (WARN_ON(nesting < 0))
720 if (early_ioremap_debug) {
721 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
726 virt_addr = (unsigned long)addr;
727 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
731 offset = virt_addr & ~PAGE_MASK;
732 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
734 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
735 while (nrpages > 0) {
736 early_clear_fixmap(idx);
742 void __this_fixmap_does_not_exist(void)