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(!phys_addr_valid(x));
45 EXPORT_SYMBOL(__phys_addr);
47 bool __virt_addr_valid(unsigned long x)
49 if (x >= __START_KERNEL_map) {
50 x -= __START_KERNEL_map;
51 if (x >= KERNEL_IMAGE_SIZE)
58 if (!phys_addr_valid(x))
62 return pfn_valid(x >> PAGE_SHIFT);
64 EXPORT_SYMBOL(__virt_addr_valid);
68 static inline int phys_addr_valid(unsigned long addr)
73 #ifdef CONFIG_DEBUG_VIRTUAL
74 unsigned long __phys_addr(unsigned long x)
76 /* VMALLOC_* aren't constants */
77 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
78 VIRTUAL_BUG_ON(__vmalloc_start_set && is_vmalloc_addr((void *) x));
79 return x - PAGE_OFFSET;
81 EXPORT_SYMBOL(__phys_addr);
84 bool __virt_addr_valid(unsigned long x)
88 if (__vmalloc_start_set && is_vmalloc_addr((void *) x))
90 return pfn_valid((x - PAGE_OFFSET) >> PAGE_SHIFT);
92 EXPORT_SYMBOL(__virt_addr_valid);
96 int page_is_ram(unsigned long pagenr)
98 resource_size_t addr, end;
102 * A special case is the first 4Kb of memory;
103 * This is a BIOS owned area, not kernel ram, but generally
104 * not listed as such in the E820 table.
110 * Second special case: Some BIOSen report the PC BIOS
111 * area (640->1Mb) as ram even though it is not.
113 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
114 pagenr < (BIOS_END >> PAGE_SHIFT))
117 for (i = 0; i < e820.nr_map; i++) {
121 if (e820.map[i].type != E820_RAM)
123 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
124 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
127 if ((pagenr >= addr) && (pagenr < end))
134 * Fix up the linear direct mapping of the kernel to avoid cache attribute
137 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
138 unsigned long prot_val)
140 unsigned long nrpages = size >> PAGE_SHIFT;
146 err = _set_memory_uc(vaddr, nrpages);
149 err = _set_memory_wc(vaddr, nrpages);
152 err = _set_memory_wb(vaddr, nrpages);
160 * Remap an arbitrary physical address space into the kernel virtual
161 * address space. Needed when the kernel wants to access high addresses
164 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
165 * have to convert them into an offset in a page-aligned mapping, but the
166 * caller shouldn't need to know that small detail.
168 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
169 unsigned long size, unsigned long prot_val, void *caller)
171 unsigned long pfn, offset, vaddr;
172 resource_size_t last_addr;
173 const resource_size_t unaligned_phys_addr = phys_addr;
174 const unsigned long unaligned_size = size;
175 struct vm_struct *area;
176 unsigned long new_prot_val;
179 void __iomem *ret_addr;
181 /* Don't allow wraparound or zero size */
182 last_addr = phys_addr + size - 1;
183 if (!size || last_addr < phys_addr)
186 if (!phys_addr_valid(phys_addr)) {
187 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
188 (unsigned long long)phys_addr);
194 * Don't remap the low PCI/ISA area, it's always mapped..
196 if (is_ISA_range(phys_addr, last_addr))
197 return (__force void __iomem *)phys_to_virt(phys_addr);
200 * Check if the request spans more than any BAR in the iomem resource
203 WARN_ONCE(iomem_map_sanity_check(phys_addr, size),
204 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
207 * Don't allow anybody to remap normal RAM that we're using..
209 for (pfn = phys_addr >> PAGE_SHIFT;
210 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
213 int is_ram = page_is_ram(pfn);
215 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
217 WARN_ON_ONCE(is_ram);
221 * Mappings have to be page-aligned
223 offset = phys_addr & ~PAGE_MASK;
224 phys_addr &= PAGE_MASK;
225 size = PAGE_ALIGN(last_addr+1) - phys_addr;
227 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
228 prot_val, &new_prot_val);
230 pr_debug("Warning: reserve_memtype returned %d\n", retval);
234 if (prot_val != new_prot_val) {
236 * Do not fallback to certain memory types with certain
238 * - request is uc-, return cannot be write-back
239 * - request is uc-, return cannot be write-combine
240 * - request is write-combine, return cannot be write-back
242 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
243 (new_prot_val == _PAGE_CACHE_WB ||
244 new_prot_val == _PAGE_CACHE_WC)) ||
245 (prot_val == _PAGE_CACHE_WC &&
246 new_prot_val == _PAGE_CACHE_WB)) {
248 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
249 (unsigned long long)phys_addr,
250 (unsigned long long)(phys_addr + size),
251 prot_val, new_prot_val);
252 free_memtype(phys_addr, phys_addr + size);
255 prot_val = new_prot_val;
261 prot = PAGE_KERNEL_IO_NOCACHE;
263 case _PAGE_CACHE_UC_MINUS:
264 prot = PAGE_KERNEL_IO_UC_MINUS;
267 prot = PAGE_KERNEL_IO_WC;
270 prot = PAGE_KERNEL_IO;
277 area = get_vm_area_caller(size, VM_IOREMAP, caller);
280 area->phys_addr = phys_addr;
281 vaddr = (unsigned long) area->addr;
282 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
283 free_memtype(phys_addr, phys_addr + size);
288 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
289 free_memtype(phys_addr, phys_addr + size);
294 ret_addr = (void __iomem *) (vaddr + offset);
295 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
301 * ioremap_nocache - map bus memory into CPU space
302 * @offset: bus address of the memory
303 * @size: size of the resource to map
305 * ioremap_nocache performs a platform specific sequence of operations to
306 * make bus memory CPU accessible via the readb/readw/readl/writeb/
307 * writew/writel functions and the other mmio helpers. The returned
308 * address is not guaranteed to be usable directly as a virtual
311 * This version of ioremap ensures that the memory is marked uncachable
312 * on the CPU as well as honouring existing caching rules from things like
313 * the PCI bus. Note that there are other caches and buffers on many
314 * busses. In particular driver authors should read up on PCI writes
316 * It's useful if some control registers are in such an area and
317 * write combining or read caching is not desirable:
319 * Must be freed with iounmap.
321 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
324 * Ideally, this should be:
325 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
327 * Till we fix all X drivers to use ioremap_wc(), we will use
330 unsigned long val = _PAGE_CACHE_UC_MINUS;
332 return __ioremap_caller(phys_addr, size, val,
333 __builtin_return_address(0));
335 EXPORT_SYMBOL(ioremap_nocache);
338 * ioremap_wc - map memory into CPU space write combined
339 * @offset: bus address of the memory
340 * @size: size of the resource to map
342 * This version of ioremap ensures that the memory is marked write combining.
343 * Write combining allows faster writes to some hardware devices.
345 * Must be freed with iounmap.
347 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
350 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
351 __builtin_return_address(0));
353 return ioremap_nocache(phys_addr, size);
355 EXPORT_SYMBOL(ioremap_wc);
357 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
359 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
360 __builtin_return_address(0));
362 EXPORT_SYMBOL(ioremap_cache);
364 static void __iomem *ioremap_default(resource_size_t phys_addr,
372 * - WB for WB-able memory and no other conflicting mappings
373 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
374 * - Inherit from confliting mappings otherwise
376 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
380 ret = __ioremap_caller(phys_addr, size, flags,
381 __builtin_return_address(0));
383 free_memtype(phys_addr, phys_addr + size);
387 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
388 unsigned long prot_val)
390 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
391 __builtin_return_address(0));
393 EXPORT_SYMBOL(ioremap_prot);
396 * iounmap - Free a IO remapping
397 * @addr: virtual address from ioremap_*
399 * Caller must ensure there is only one unmapping for the same pointer.
401 void iounmap(volatile void __iomem *addr)
403 struct vm_struct *p, *o;
405 if ((void __force *)addr <= high_memory)
409 * __ioremap special-cases the PCI/ISA range by not instantiating a
410 * vm_area and by simply returning an address into the kernel mapping
411 * of ISA space. So handle that here.
413 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
414 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
417 addr = (volatile void __iomem *)
418 (PAGE_MASK & (unsigned long __force)addr);
420 mmiotrace_iounmap(addr);
422 /* Use the vm area unlocked, assuming the caller
423 ensures there isn't another iounmap for the same address
424 in parallel. Reuse of the virtual address is prevented by
425 leaving it in the global lists until we're done with it.
426 cpa takes care of the direct mappings. */
427 read_lock(&vmlist_lock);
428 for (p = vmlist; p; p = p->next) {
429 if (p->addr == (void __force *)addr)
432 read_unlock(&vmlist_lock);
435 printk(KERN_ERR "iounmap: bad address %p\n", addr);
440 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
442 /* Finally remove it */
443 o = remove_vm_area((void __force *)addr);
444 BUG_ON(p != o || o == NULL);
447 EXPORT_SYMBOL(iounmap);
450 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
453 void *xlate_dev_mem_ptr(unsigned long phys)
456 unsigned long start = phys & PAGE_MASK;
458 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
459 if (page_is_ram(start >> PAGE_SHIFT))
462 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
464 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
469 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
471 if (page_is_ram(phys >> PAGE_SHIFT))
474 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
478 static int __initdata early_ioremap_debug;
480 static int __init early_ioremap_debug_setup(char *str)
482 early_ioremap_debug = 1;
486 early_param("early_ioremap_debug", early_ioremap_debug_setup);
488 static __initdata int after_paging_init;
489 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
491 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
493 /* Don't assume we're using swapper_pg_dir at this point */
494 pgd_t *base = __va(read_cr3());
495 pgd_t *pgd = &base[pgd_index(addr)];
496 pud_t *pud = pud_offset(pgd, addr);
497 pmd_t *pmd = pmd_offset(pud, addr);
502 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
504 return &bm_pte[pte_index(addr)];
507 void __init early_ioremap_init(void)
511 if (early_ioremap_debug)
512 printk(KERN_INFO "early_ioremap_init()\n");
514 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
515 memset(bm_pte, 0, sizeof(bm_pte));
516 pmd_populate_kernel(&init_mm, pmd, bm_pte);
519 * The boot-ioremap range spans multiple pmds, for which
520 * we are not prepared:
522 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
524 printk(KERN_WARNING "pmd %p != %p\n",
525 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
526 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
527 fix_to_virt(FIX_BTMAP_BEGIN));
528 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
529 fix_to_virt(FIX_BTMAP_END));
531 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
532 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
537 void __init early_ioremap_reset(void)
539 after_paging_init = 1;
542 static void __init __early_set_fixmap(enum fixed_addresses idx,
543 unsigned long phys, pgprot_t flags)
545 unsigned long addr = __fix_to_virt(idx);
548 if (idx >= __end_of_fixed_addresses) {
552 pte = early_ioremap_pte(addr);
554 if (pgprot_val(flags))
555 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
557 pte_clear(&init_mm, addr, pte);
558 __flush_tlb_one(addr);
561 static inline void __init early_set_fixmap(enum fixed_addresses idx,
562 unsigned long phys, pgprot_t prot)
564 if (after_paging_init)
565 __set_fixmap(idx, phys, prot);
567 __early_set_fixmap(idx, phys, prot);
570 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
572 if (after_paging_init)
575 __early_set_fixmap(idx, 0, __pgprot(0));
578 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
579 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
580 static int __init check_early_ioremap_leak(void)
585 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
592 "Debug warning: early ioremap leak of %d areas detected.\n",
595 "please boot with early_ioremap_debug and report the dmesg.\n");
599 late_initcall(check_early_ioremap_leak);
601 static void __init __iomem *__early_ioremap(unsigned long phys_addr, unsigned long size, pgprot_t prot)
603 unsigned long offset, last_addr;
604 unsigned int nrpages;
605 enum fixed_addresses idx0, idx;
608 WARN_ON(system_state != SYSTEM_BOOTING);
611 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
619 printk(KERN_INFO "early_iomap(%08lx, %08lx) not found slot\n",
625 if (early_ioremap_debug) {
626 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
627 phys_addr, size, slot);
631 /* Don't allow wraparound or zero size */
632 last_addr = phys_addr + size - 1;
633 if (!size || last_addr < phys_addr) {
638 prev_size[slot] = size;
640 * Mappings have to be page-aligned
642 offset = phys_addr & ~PAGE_MASK;
643 phys_addr &= PAGE_MASK;
644 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
647 * Mappings have to fit in the FIX_BTMAP area.
649 nrpages = size >> PAGE_SHIFT;
650 if (nrpages > NR_FIX_BTMAPS) {
658 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
660 while (nrpages > 0) {
661 early_set_fixmap(idx, phys_addr, prot);
662 phys_addr += PAGE_SIZE;
666 if (early_ioremap_debug)
667 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
669 prev_map[slot] = (void __iomem *)(offset + fix_to_virt(idx0));
670 return prev_map[slot];
673 /* Remap an IO device */
674 void __init __iomem *early_ioremap(unsigned long phys_addr, unsigned long size)
676 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
680 void __init __iomem *early_memremap(unsigned long phys_addr, unsigned long size)
682 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
685 void __init early_iounmap(void __iomem *addr, unsigned long size)
687 unsigned long virt_addr;
688 unsigned long offset;
689 unsigned int nrpages;
690 enum fixed_addresses idx;
694 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
695 if (prev_map[i] == addr) {
702 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
708 if (prev_size[slot] != size) {
709 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
710 addr, size, slot, prev_size[slot]);
715 if (early_ioremap_debug) {
716 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
721 virt_addr = (unsigned long)addr;
722 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
726 offset = virt_addr & ~PAGE_MASK;
727 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
729 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
730 while (nrpages > 0) {
731 early_clear_fixmap(idx);
735 prev_map[slot] = NULL;
738 void __this_fixmap_does_not_exist(void)