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))
138 * Fix up the linear direct mapping of the kernel to avoid cache attribute
141 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
142 unsigned long prot_val)
144 unsigned long nrpages = size >> PAGE_SHIFT;
150 err = _set_memory_uc(vaddr, nrpages);
153 err = _set_memory_wc(vaddr, nrpages);
156 err = _set_memory_wb(vaddr, nrpages);
164 * Remap an arbitrary physical address space into the kernel virtual
165 * address space. Needed when the kernel wants to access high addresses
168 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
169 * have to convert them into an offset in a page-aligned mapping, but the
170 * caller shouldn't need to know that small detail.
172 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
173 unsigned long size, unsigned long prot_val, void *caller)
175 unsigned long pfn, offset, vaddr;
176 resource_size_t last_addr;
177 const resource_size_t unaligned_phys_addr = phys_addr;
178 const unsigned long unaligned_size = size;
179 struct vm_struct *area;
180 unsigned long new_prot_val;
183 void __iomem *ret_addr;
185 /* Don't allow wraparound or zero size */
186 last_addr = phys_addr + size - 1;
187 if (!size || last_addr < phys_addr)
190 if (!phys_addr_valid(phys_addr)) {
191 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
192 (unsigned long long)phys_addr);
198 * Don't remap the low PCI/ISA area, it's always mapped..
200 if (is_ISA_range(phys_addr, last_addr))
201 return (__force void __iomem *)phys_to_virt(phys_addr);
204 * Check if the request spans more than any BAR in the iomem resource
207 WARN_ONCE(iomem_map_sanity_check(phys_addr, size),
208 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
211 * Don't allow anybody to remap normal RAM that we're using..
213 for (pfn = phys_addr >> PAGE_SHIFT;
214 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
217 int is_ram = page_is_ram(pfn);
219 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
221 WARN_ON_ONCE(is_ram);
225 * Mappings have to be page-aligned
227 offset = phys_addr & ~PAGE_MASK;
228 phys_addr &= PAGE_MASK;
229 size = PAGE_ALIGN(last_addr+1) - phys_addr;
231 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
232 prot_val, &new_prot_val);
234 pr_debug("Warning: reserve_memtype returned %d\n", retval);
238 if (prot_val != new_prot_val) {
240 * Do not fallback to certain memory types with certain
242 * - request is uc-, return cannot be write-back
243 * - request is uc-, return cannot be write-combine
244 * - request is write-combine, return cannot be write-back
246 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
247 (new_prot_val == _PAGE_CACHE_WB ||
248 new_prot_val == _PAGE_CACHE_WC)) ||
249 (prot_val == _PAGE_CACHE_WC &&
250 new_prot_val == _PAGE_CACHE_WB)) {
252 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
253 (unsigned long long)phys_addr,
254 (unsigned long long)(phys_addr + size),
255 prot_val, new_prot_val);
256 free_memtype(phys_addr, phys_addr + size);
259 prot_val = new_prot_val;
265 prot = PAGE_KERNEL_IO_NOCACHE;
267 case _PAGE_CACHE_UC_MINUS:
268 prot = PAGE_KERNEL_IO_UC_MINUS;
271 prot = PAGE_KERNEL_IO_WC;
274 prot = PAGE_KERNEL_IO;
281 area = get_vm_area_caller(size, VM_IOREMAP, caller);
284 area->phys_addr = phys_addr;
285 vaddr = (unsigned long) area->addr;
286 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
287 free_memtype(phys_addr, phys_addr + size);
292 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
293 free_memtype(phys_addr, phys_addr + size);
298 ret_addr = (void __iomem *) (vaddr + offset);
299 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
305 * ioremap_nocache - map bus memory into CPU space
306 * @offset: bus address of the memory
307 * @size: size of the resource to map
309 * ioremap_nocache performs a platform specific sequence of operations to
310 * make bus memory CPU accessible via the readb/readw/readl/writeb/
311 * writew/writel functions and the other mmio helpers. The returned
312 * address is not guaranteed to be usable directly as a virtual
315 * This version of ioremap ensures that the memory is marked uncachable
316 * on the CPU as well as honouring existing caching rules from things like
317 * the PCI bus. Note that there are other caches and buffers on many
318 * busses. In particular driver authors should read up on PCI writes
320 * It's useful if some control registers are in such an area and
321 * write combining or read caching is not desirable:
323 * Must be freed with iounmap.
325 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
328 * Ideally, this should be:
329 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
331 * Till we fix all X drivers to use ioremap_wc(), we will use
334 unsigned long val = _PAGE_CACHE_UC_MINUS;
336 return __ioremap_caller(phys_addr, size, val,
337 __builtin_return_address(0));
339 EXPORT_SYMBOL(ioremap_nocache);
342 * ioremap_wc - map memory into CPU space write combined
343 * @offset: bus address of the memory
344 * @size: size of the resource to map
346 * This version of ioremap ensures that the memory is marked write combining.
347 * Write combining allows faster writes to some hardware devices.
349 * Must be freed with iounmap.
351 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
354 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
355 __builtin_return_address(0));
357 return ioremap_nocache(phys_addr, size);
359 EXPORT_SYMBOL(ioremap_wc);
361 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
363 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
364 __builtin_return_address(0));
366 EXPORT_SYMBOL(ioremap_cache);
368 static void __iomem *ioremap_default(resource_size_t phys_addr,
376 * - WB for WB-able memory and no other conflicting mappings
377 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
378 * - Inherit from confliting mappings otherwise
380 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
384 ret = __ioremap_caller(phys_addr, size, flags,
385 __builtin_return_address(0));
387 free_memtype(phys_addr, phys_addr + size);
391 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
392 unsigned long prot_val)
394 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
395 __builtin_return_address(0));
397 EXPORT_SYMBOL(ioremap_prot);
400 * iounmap - Free a IO remapping
401 * @addr: virtual address from ioremap_*
403 * Caller must ensure there is only one unmapping for the same pointer.
405 void iounmap(volatile void __iomem *addr)
407 struct vm_struct *p, *o;
409 if ((void __force *)addr <= high_memory)
413 * __ioremap special-cases the PCI/ISA range by not instantiating a
414 * vm_area and by simply returning an address into the kernel mapping
415 * of ISA space. So handle that here.
417 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
418 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
421 addr = (volatile void __iomem *)
422 (PAGE_MASK & (unsigned long __force)addr);
424 mmiotrace_iounmap(addr);
426 /* Use the vm area unlocked, assuming the caller
427 ensures there isn't another iounmap for the same address
428 in parallel. Reuse of the virtual address is prevented by
429 leaving it in the global lists until we're done with it.
430 cpa takes care of the direct mappings. */
431 read_lock(&vmlist_lock);
432 for (p = vmlist; p; p = p->next) {
433 if (p->addr == (void __force *)addr)
436 read_unlock(&vmlist_lock);
439 printk(KERN_ERR "iounmap: bad address %p\n", addr);
444 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
446 /* Finally remove it */
447 o = remove_vm_area((void __force *)addr);
448 BUG_ON(p != o || o == NULL);
451 EXPORT_SYMBOL(iounmap);
454 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
457 void *xlate_dev_mem_ptr(unsigned long phys)
460 unsigned long start = phys & PAGE_MASK;
462 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
463 if (page_is_ram(start >> PAGE_SHIFT))
466 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
468 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
473 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
475 if (page_is_ram(phys >> PAGE_SHIFT))
478 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
482 static int __initdata early_ioremap_debug;
484 static int __init early_ioremap_debug_setup(char *str)
486 early_ioremap_debug = 1;
490 early_param("early_ioremap_debug", early_ioremap_debug_setup);
492 static __initdata int after_paging_init;
493 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
495 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
497 /* Don't assume we're using swapper_pg_dir at this point */
498 pgd_t *base = __va(read_cr3());
499 pgd_t *pgd = &base[pgd_index(addr)];
500 pud_t *pud = pud_offset(pgd, addr);
501 pmd_t *pmd = pmd_offset(pud, addr);
506 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
508 return &bm_pte[pte_index(addr)];
511 void __init early_ioremap_init(void)
515 if (early_ioremap_debug)
516 printk(KERN_INFO "early_ioremap_init()\n");
518 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
519 memset(bm_pte, 0, sizeof(bm_pte));
520 pmd_populate_kernel(&init_mm, pmd, bm_pte);
523 * The boot-ioremap range spans multiple pmds, for which
524 * we are not prepared:
526 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
528 printk(KERN_WARNING "pmd %p != %p\n",
529 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
530 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
531 fix_to_virt(FIX_BTMAP_BEGIN));
532 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
533 fix_to_virt(FIX_BTMAP_END));
535 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
536 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
541 void __init early_ioremap_reset(void)
543 after_paging_init = 1;
546 static void __init __early_set_fixmap(enum fixed_addresses idx,
547 unsigned long phys, pgprot_t flags)
549 unsigned long addr = __fix_to_virt(idx);
552 if (idx >= __end_of_fixed_addresses) {
556 pte = early_ioremap_pte(addr);
558 if (pgprot_val(flags))
559 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
561 pte_clear(&init_mm, addr, pte);
562 __flush_tlb_one(addr);
565 static inline void __init early_set_fixmap(enum fixed_addresses idx,
566 unsigned long phys, pgprot_t prot)
568 if (after_paging_init)
569 __set_fixmap(idx, phys, prot);
571 __early_set_fixmap(idx, phys, prot);
574 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
576 if (after_paging_init)
579 __early_set_fixmap(idx, 0, __pgprot(0));
582 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
583 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
584 static int __init check_early_ioremap_leak(void)
589 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
596 "Debug warning: early ioremap leak of %d areas detected.\n",
599 "please boot with early_ioremap_debug and report the dmesg.\n");
603 late_initcall(check_early_ioremap_leak);
605 static void __init __iomem *__early_ioremap(unsigned long phys_addr, unsigned long size, pgprot_t prot)
607 unsigned long offset, last_addr;
608 unsigned int nrpages;
609 enum fixed_addresses idx0, idx;
612 WARN_ON(system_state != SYSTEM_BOOTING);
615 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
623 printk(KERN_INFO "early_iomap(%08lx, %08lx) not found slot\n",
629 if (early_ioremap_debug) {
630 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
631 phys_addr, size, slot);
635 /* Don't allow wraparound or zero size */
636 last_addr = phys_addr + size - 1;
637 if (!size || last_addr < phys_addr) {
642 prev_size[slot] = size;
644 * Mappings have to be page-aligned
646 offset = phys_addr & ~PAGE_MASK;
647 phys_addr &= PAGE_MASK;
648 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
651 * Mappings have to fit in the FIX_BTMAP area.
653 nrpages = size >> PAGE_SHIFT;
654 if (nrpages > NR_FIX_BTMAPS) {
662 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
664 while (nrpages > 0) {
665 early_set_fixmap(idx, phys_addr, prot);
666 phys_addr += PAGE_SIZE;
670 if (early_ioremap_debug)
671 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
673 prev_map[slot] = (void __iomem *)(offset + fix_to_virt(idx0));
674 return prev_map[slot];
677 /* Remap an IO device */
678 void __init __iomem *early_ioremap(unsigned long phys_addr, unsigned long size)
680 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
684 void __init __iomem *early_memremap(unsigned long phys_addr, unsigned long size)
686 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
689 void __init early_iounmap(void __iomem *addr, unsigned long size)
691 unsigned long virt_addr;
692 unsigned long offset;
693 unsigned int nrpages;
694 enum fixed_addresses idx;
698 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
699 if (prev_map[i] == addr) {
706 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
712 if (prev_size[slot] != size) {
713 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
714 addr, size, slot, prev_size[slot]);
719 if (early_ioremap_debug) {
720 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
725 virt_addr = (unsigned long)addr;
726 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
730 offset = virt_addr & ~PAGE_MASK;
731 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
733 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
734 while (nrpages > 0) {
735 early_clear_fixmap(idx);
739 prev_map[slot] = NULL;
742 void __this_fixmap_does_not_exist(void)