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 if (x >= FIXADDR_START)
92 return pfn_valid((x - PAGE_OFFSET) >> PAGE_SHIFT);
94 EXPORT_SYMBOL(__virt_addr_valid);
98 int page_is_ram(unsigned long pagenr)
100 resource_size_t addr, end;
104 * A special case is the first 4Kb of memory;
105 * This is a BIOS owned area, not kernel ram, but generally
106 * not listed as such in the E820 table.
112 * Second special case: Some BIOSen report the PC BIOS
113 * area (640->1Mb) as ram even though it is not.
115 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
116 pagenr < (BIOS_END >> PAGE_SHIFT))
119 for (i = 0; i < e820.nr_map; i++) {
123 if (e820.map[i].type != E820_RAM)
125 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
126 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
129 if ((pagenr >= addr) && (pagenr < end))
136 * Fix up the linear direct mapping of the kernel to avoid cache attribute
139 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
140 unsigned long prot_val)
142 unsigned long nrpages = size >> PAGE_SHIFT;
148 err = _set_memory_uc(vaddr, nrpages);
151 err = _set_memory_wc(vaddr, nrpages);
154 err = _set_memory_wb(vaddr, nrpages);
162 * Remap an arbitrary physical address space into the kernel virtual
163 * address space. Needed when the kernel wants to access high addresses
166 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
167 * have to convert them into an offset in a page-aligned mapping, but the
168 * caller shouldn't need to know that small detail.
170 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
171 unsigned long size, unsigned long prot_val, void *caller)
173 unsigned long pfn, offset, vaddr;
174 resource_size_t last_addr;
175 const resource_size_t unaligned_phys_addr = phys_addr;
176 const unsigned long unaligned_size = size;
177 struct vm_struct *area;
178 unsigned long new_prot_val;
181 void __iomem *ret_addr;
183 /* Don't allow wraparound or zero size */
184 last_addr = phys_addr + size - 1;
185 if (!size || last_addr < phys_addr)
188 if (!phys_addr_valid(phys_addr)) {
189 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
190 (unsigned long long)phys_addr);
196 * Don't remap the low PCI/ISA area, it's always mapped..
198 if (is_ISA_range(phys_addr, last_addr))
199 return (__force void __iomem *)phys_to_virt(phys_addr);
202 * Check if the request spans more than any BAR in the iomem resource
205 WARN_ONCE(iomem_map_sanity_check(phys_addr, size),
206 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
209 * Don't allow anybody to remap normal RAM that we're using..
211 for (pfn = phys_addr >> PAGE_SHIFT;
212 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
215 int is_ram = page_is_ram(pfn);
217 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
219 WARN_ON_ONCE(is_ram);
223 * Mappings have to be page-aligned
225 offset = phys_addr & ~PAGE_MASK;
226 phys_addr &= PAGE_MASK;
227 size = PAGE_ALIGN(last_addr+1) - phys_addr;
229 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
230 prot_val, &new_prot_val);
232 pr_debug("Warning: reserve_memtype returned %d\n", retval);
236 if (prot_val != new_prot_val) {
238 * Do not fallback to certain memory types with certain
240 * - request is uc-, return cannot be write-back
241 * - request is uc-, return cannot be write-combine
242 * - request is write-combine, return cannot be write-back
244 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
245 (new_prot_val == _PAGE_CACHE_WB ||
246 new_prot_val == _PAGE_CACHE_WC)) ||
247 (prot_val == _PAGE_CACHE_WC &&
248 new_prot_val == _PAGE_CACHE_WB)) {
250 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
251 (unsigned long long)phys_addr,
252 (unsigned long long)(phys_addr + size),
253 prot_val, new_prot_val);
254 free_memtype(phys_addr, phys_addr + size);
257 prot_val = new_prot_val;
263 prot = PAGE_KERNEL_IO_NOCACHE;
265 case _PAGE_CACHE_UC_MINUS:
266 prot = PAGE_KERNEL_IO_UC_MINUS;
269 prot = PAGE_KERNEL_IO_WC;
272 prot = PAGE_KERNEL_IO;
279 area = get_vm_area_caller(size, VM_IOREMAP, caller);
282 area->phys_addr = phys_addr;
283 vaddr = (unsigned long) area->addr;
284 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
285 free_memtype(phys_addr, phys_addr + size);
290 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
291 free_memtype(phys_addr, phys_addr + size);
296 ret_addr = (void __iomem *) (vaddr + offset);
297 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
303 * ioremap_nocache - map bus memory into CPU space
304 * @offset: bus address of the memory
305 * @size: size of the resource to map
307 * ioremap_nocache performs a platform specific sequence of operations to
308 * make bus memory CPU accessible via the readb/readw/readl/writeb/
309 * writew/writel functions and the other mmio helpers. The returned
310 * address is not guaranteed to be usable directly as a virtual
313 * This version of ioremap ensures that the memory is marked uncachable
314 * on the CPU as well as honouring existing caching rules from things like
315 * the PCI bus. Note that there are other caches and buffers on many
316 * busses. In particular driver authors should read up on PCI writes
318 * It's useful if some control registers are in such an area and
319 * write combining or read caching is not desirable:
321 * Must be freed with iounmap.
323 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
326 * Ideally, this should be:
327 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
329 * Till we fix all X drivers to use ioremap_wc(), we will use
332 unsigned long val = _PAGE_CACHE_UC_MINUS;
334 return __ioremap_caller(phys_addr, size, val,
335 __builtin_return_address(0));
337 EXPORT_SYMBOL(ioremap_nocache);
340 * ioremap_wc - map memory into CPU space write combined
341 * @offset: bus address of the memory
342 * @size: size of the resource to map
344 * This version of ioremap ensures that the memory is marked write combining.
345 * Write combining allows faster writes to some hardware devices.
347 * Must be freed with iounmap.
349 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
352 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
353 __builtin_return_address(0));
355 return ioremap_nocache(phys_addr, size);
357 EXPORT_SYMBOL(ioremap_wc);
359 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
361 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
362 __builtin_return_address(0));
364 EXPORT_SYMBOL(ioremap_cache);
366 static void __iomem *ioremap_default(resource_size_t phys_addr,
374 * - WB for WB-able memory and no other conflicting mappings
375 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
376 * - Inherit from confliting mappings otherwise
378 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
382 ret = __ioremap_caller(phys_addr, size, flags,
383 __builtin_return_address(0));
385 free_memtype(phys_addr, phys_addr + size);
389 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
390 unsigned long prot_val)
392 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
393 __builtin_return_address(0));
395 EXPORT_SYMBOL(ioremap_prot);
398 * iounmap - Free a IO remapping
399 * @addr: virtual address from ioremap_*
401 * Caller must ensure there is only one unmapping for the same pointer.
403 void iounmap(volatile void __iomem *addr)
405 struct vm_struct *p, *o;
407 if ((void __force *)addr <= high_memory)
411 * __ioremap special-cases the PCI/ISA range by not instantiating a
412 * vm_area and by simply returning an address into the kernel mapping
413 * of ISA space. So handle that here.
415 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
416 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
419 addr = (volatile void __iomem *)
420 (PAGE_MASK & (unsigned long __force)addr);
422 mmiotrace_iounmap(addr);
424 /* Use the vm area unlocked, assuming the caller
425 ensures there isn't another iounmap for the same address
426 in parallel. Reuse of the virtual address is prevented by
427 leaving it in the global lists until we're done with it.
428 cpa takes care of the direct mappings. */
429 read_lock(&vmlist_lock);
430 for (p = vmlist; p; p = p->next) {
431 if (p->addr == (void __force *)addr)
434 read_unlock(&vmlist_lock);
437 printk(KERN_ERR "iounmap: bad address %p\n", addr);
442 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
444 /* Finally remove it */
445 o = remove_vm_area((void __force *)addr);
446 BUG_ON(p != o || o == NULL);
449 EXPORT_SYMBOL(iounmap);
452 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
455 void *xlate_dev_mem_ptr(unsigned long phys)
458 unsigned long start = phys & PAGE_MASK;
460 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
461 if (page_is_ram(start >> PAGE_SHIFT))
464 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
466 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
471 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
473 if (page_is_ram(phys >> PAGE_SHIFT))
476 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
480 static int __initdata early_ioremap_debug;
482 static int __init early_ioremap_debug_setup(char *str)
484 early_ioremap_debug = 1;
488 early_param("early_ioremap_debug", early_ioremap_debug_setup);
490 static __initdata int after_paging_init;
491 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
493 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
495 /* Don't assume we're using swapper_pg_dir at this point */
496 pgd_t *base = __va(read_cr3());
497 pgd_t *pgd = &base[pgd_index(addr)];
498 pud_t *pud = pud_offset(pgd, addr);
499 pmd_t *pmd = pmd_offset(pud, addr);
504 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
506 return &bm_pte[pte_index(addr)];
509 void __init early_ioremap_init(void)
513 if (early_ioremap_debug)
514 printk(KERN_INFO "early_ioremap_init()\n");
516 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
517 memset(bm_pte, 0, sizeof(bm_pte));
518 pmd_populate_kernel(&init_mm, pmd, bm_pte);
521 * The boot-ioremap range spans multiple pmds, for which
522 * we are not prepared:
524 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
526 printk(KERN_WARNING "pmd %p != %p\n",
527 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
528 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
529 fix_to_virt(FIX_BTMAP_BEGIN));
530 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
531 fix_to_virt(FIX_BTMAP_END));
533 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
534 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
539 void __init early_ioremap_reset(void)
541 after_paging_init = 1;
544 static void __init __early_set_fixmap(enum fixed_addresses idx,
545 unsigned long phys, pgprot_t flags)
547 unsigned long addr = __fix_to_virt(idx);
550 if (idx >= __end_of_fixed_addresses) {
554 pte = early_ioremap_pte(addr);
556 if (pgprot_val(flags))
557 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
559 pte_clear(&init_mm, addr, pte);
560 __flush_tlb_one(addr);
563 static inline void __init early_set_fixmap(enum fixed_addresses idx,
564 unsigned long phys, pgprot_t prot)
566 if (after_paging_init)
567 __set_fixmap(idx, phys, prot);
569 __early_set_fixmap(idx, phys, prot);
572 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
574 if (after_paging_init)
577 __early_set_fixmap(idx, 0, __pgprot(0));
580 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
581 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
582 static int __init check_early_ioremap_leak(void)
587 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
594 "Debug warning: early ioremap leak of %d areas detected.\n",
597 "please boot with early_ioremap_debug and report the dmesg.\n");
601 late_initcall(check_early_ioremap_leak);
603 static void __init __iomem *__early_ioremap(unsigned long phys_addr, unsigned long size, pgprot_t prot)
605 unsigned long offset, last_addr;
606 unsigned int nrpages;
607 enum fixed_addresses idx0, idx;
610 WARN_ON(system_state != SYSTEM_BOOTING);
613 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
621 printk(KERN_INFO "early_iomap(%08lx, %08lx) not found slot\n",
627 if (early_ioremap_debug) {
628 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
629 phys_addr, size, slot);
633 /* Don't allow wraparound or zero size */
634 last_addr = phys_addr + size - 1;
635 if (!size || last_addr < phys_addr) {
640 prev_size[slot] = size;
642 * Mappings have to be page-aligned
644 offset = phys_addr & ~PAGE_MASK;
645 phys_addr &= PAGE_MASK;
646 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
649 * Mappings have to fit in the FIX_BTMAP area.
651 nrpages = size >> PAGE_SHIFT;
652 if (nrpages > NR_FIX_BTMAPS) {
660 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
662 while (nrpages > 0) {
663 early_set_fixmap(idx, phys_addr, prot);
664 phys_addr += PAGE_SIZE;
668 if (early_ioremap_debug)
669 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
671 prev_map[slot] = (void __iomem *)(offset + fix_to_virt(idx0));
672 return prev_map[slot];
675 /* Remap an IO device */
676 void __init __iomem *early_ioremap(unsigned long phys_addr, unsigned long size)
678 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
682 void __init __iomem *early_memremap(unsigned long phys_addr, unsigned long size)
684 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
687 void __init early_iounmap(void __iomem *addr, unsigned long size)
689 unsigned long virt_addr;
690 unsigned long offset;
691 unsigned int nrpages;
692 enum fixed_addresses idx;
696 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
697 if (prev_map[i] == addr) {
704 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
710 if (prev_size[slot] != size) {
711 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
712 addr, size, slot, prev_size[slot]);
717 if (early_ioremap_debug) {
718 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
723 virt_addr = (unsigned long)addr;
724 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
728 offset = virt_addr & ~PAGE_MASK;
729 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
731 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
732 while (nrpages > 0) {
733 early_clear_fixmap(idx);
737 prev_map[slot] = NULL;
740 void __this_fixmap_does_not_exist(void)