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>
25 static inline int phys_addr_valid(resource_size_t addr)
27 #ifdef CONFIG_PHYS_ADDR_T_64BIT
28 return !(addr >> boot_cpu_data.x86_phys_bits);
36 unsigned long __phys_addr(unsigned long x)
38 if (x >= __START_KERNEL_map) {
39 x -= __START_KERNEL_map;
40 VIRTUAL_BUG_ON(x >= KERNEL_IMAGE_SIZE);
43 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
45 VIRTUAL_BUG_ON(!phys_addr_valid(x));
49 EXPORT_SYMBOL(__phys_addr);
51 bool __virt_addr_valid(unsigned long x)
53 if (x >= __START_KERNEL_map) {
54 x -= __START_KERNEL_map;
55 if (x >= KERNEL_IMAGE_SIZE)
62 if (!phys_addr_valid(x))
66 return pfn_valid(x >> PAGE_SHIFT);
68 EXPORT_SYMBOL(__virt_addr_valid);
72 #ifdef CONFIG_DEBUG_VIRTUAL
73 unsigned long __phys_addr(unsigned long x)
75 /* VMALLOC_* aren't constants */
76 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
77 VIRTUAL_BUG_ON(__vmalloc_start_set && is_vmalloc_addr((void *) x));
78 return x - PAGE_OFFSET;
80 EXPORT_SYMBOL(__phys_addr);
83 bool __virt_addr_valid(unsigned long x)
87 if (__vmalloc_start_set && is_vmalloc_addr((void *) x))
89 if (x >= FIXADDR_START)
91 return pfn_valid((x - PAGE_OFFSET) >> PAGE_SHIFT);
93 EXPORT_SYMBOL(__virt_addr_valid);
97 int page_is_ram(unsigned long pagenr)
99 resource_size_t addr, end;
103 * A special case is the first 4Kb of memory;
104 * This is a BIOS owned area, not kernel ram, but generally
105 * not listed as such in the E820 table.
111 * Second special case: Some BIOSen report the PC BIOS
112 * area (640->1Mb) as ram even though it is not.
114 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
115 pagenr < (BIOS_END >> PAGE_SHIFT))
118 for (i = 0; i < e820.nr_map; i++) {
122 if (e820.map[i].type != E820_RAM)
124 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
125 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
128 if ((pagenr >= addr) && (pagenr < end))
135 * Fix up the linear direct mapping of the kernel to avoid cache attribute
138 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
139 unsigned long prot_val)
141 unsigned long nrpages = size >> PAGE_SHIFT;
147 err = _set_memory_uc(vaddr, nrpages);
150 err = _set_memory_wc(vaddr, nrpages);
153 err = _set_memory_wb(vaddr, nrpages);
161 * Remap an arbitrary physical address space into the kernel virtual
162 * address space. Needed when the kernel wants to access high addresses
165 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
166 * have to convert them into an offset in a page-aligned mapping, but the
167 * caller shouldn't need to know that small detail.
169 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
170 unsigned long size, unsigned long prot_val, void *caller)
172 unsigned long pfn, offset, vaddr;
173 resource_size_t last_addr;
174 const resource_size_t unaligned_phys_addr = phys_addr;
175 const unsigned long unaligned_size = size;
176 struct vm_struct *area;
177 unsigned long new_prot_val;
180 void __iomem *ret_addr;
182 /* Don't allow wraparound or zero size */
183 last_addr = phys_addr + size - 1;
184 if (!size || last_addr < phys_addr)
187 if (!phys_addr_valid(phys_addr)) {
188 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
189 (unsigned long long)phys_addr);
195 * Don't remap the low PCI/ISA area, it's always mapped..
197 if (is_ISA_range(phys_addr, last_addr))
198 return (__force void __iomem *)phys_to_virt(phys_addr);
201 * Check if the request spans more than any BAR in the iomem resource
204 WARN_ONCE(iomem_map_sanity_check(phys_addr, size),
205 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
208 * Don't allow anybody to remap normal RAM that we're using..
210 for (pfn = phys_addr >> PAGE_SHIFT;
211 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
214 int is_ram = page_is_ram(pfn);
216 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
218 WARN_ON_ONCE(is_ram);
222 * Mappings have to be page-aligned
224 offset = phys_addr & ~PAGE_MASK;
225 phys_addr &= PAGE_MASK;
226 size = PAGE_ALIGN(last_addr+1) - phys_addr;
228 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
229 prot_val, &new_prot_val);
231 pr_debug("Warning: reserve_memtype returned %d\n", retval);
235 if (prot_val != new_prot_val) {
237 * Do not fallback to certain memory types with certain
239 * - request is uc-, return cannot be write-back
240 * - request is uc-, return cannot be write-combine
241 * - request is write-combine, return cannot be write-back
243 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
244 (new_prot_val == _PAGE_CACHE_WB ||
245 new_prot_val == _PAGE_CACHE_WC)) ||
246 (prot_val == _PAGE_CACHE_WC &&
247 new_prot_val == _PAGE_CACHE_WB)) {
249 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
250 (unsigned long long)phys_addr,
251 (unsigned long long)(phys_addr + size),
252 prot_val, new_prot_val);
253 free_memtype(phys_addr, phys_addr + size);
256 prot_val = new_prot_val;
262 prot = PAGE_KERNEL_IO_NOCACHE;
264 case _PAGE_CACHE_UC_MINUS:
265 prot = PAGE_KERNEL_IO_UC_MINUS;
268 prot = PAGE_KERNEL_IO_WC;
271 prot = PAGE_KERNEL_IO;
278 area = get_vm_area_caller(size, VM_IOREMAP, caller);
281 area->phys_addr = phys_addr;
282 vaddr = (unsigned long) area->addr;
283 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
284 free_memtype(phys_addr, phys_addr + size);
289 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
290 free_memtype(phys_addr, phys_addr + size);
295 ret_addr = (void __iomem *) (vaddr + offset);
296 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
302 * ioremap_nocache - map bus memory into CPU space
303 * @offset: bus address of the memory
304 * @size: size of the resource to map
306 * ioremap_nocache performs a platform specific sequence of operations to
307 * make bus memory CPU accessible via the readb/readw/readl/writeb/
308 * writew/writel functions and the other mmio helpers. The returned
309 * address is not guaranteed to be usable directly as a virtual
312 * This version of ioremap ensures that the memory is marked uncachable
313 * on the CPU as well as honouring existing caching rules from things like
314 * the PCI bus. Note that there are other caches and buffers on many
315 * busses. In particular driver authors should read up on PCI writes
317 * It's useful if some control registers are in such an area and
318 * write combining or read caching is not desirable:
320 * Must be freed with iounmap.
322 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
325 * Ideally, this should be:
326 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
328 * Till we fix all X drivers to use ioremap_wc(), we will use
331 unsigned long val = _PAGE_CACHE_UC_MINUS;
333 return __ioremap_caller(phys_addr, size, val,
334 __builtin_return_address(0));
336 EXPORT_SYMBOL(ioremap_nocache);
339 * ioremap_wc - map memory into CPU space write combined
340 * @offset: bus address of the memory
341 * @size: size of the resource to map
343 * This version of ioremap ensures that the memory is marked write combining.
344 * Write combining allows faster writes to some hardware devices.
346 * Must be freed with iounmap.
348 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
351 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
352 __builtin_return_address(0));
354 return ioremap_nocache(phys_addr, size);
356 EXPORT_SYMBOL(ioremap_wc);
358 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
360 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
361 __builtin_return_address(0));
363 EXPORT_SYMBOL(ioremap_cache);
365 static void __iomem *ioremap_default(resource_size_t phys_addr,
373 * - WB for WB-able memory and no other conflicting mappings
374 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
375 * - Inherit from confliting mappings otherwise
377 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
381 ret = __ioremap_caller(phys_addr, size, flags,
382 __builtin_return_address(0));
384 free_memtype(phys_addr, phys_addr + size);
388 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
389 unsigned long prot_val)
391 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
392 __builtin_return_address(0));
394 EXPORT_SYMBOL(ioremap_prot);
397 * iounmap - Free a IO remapping
398 * @addr: virtual address from ioremap_*
400 * Caller must ensure there is only one unmapping for the same pointer.
402 void iounmap(volatile void __iomem *addr)
404 struct vm_struct *p, *o;
406 if ((void __force *)addr <= high_memory)
410 * __ioremap special-cases the PCI/ISA range by not instantiating a
411 * vm_area and by simply returning an address into the kernel mapping
412 * of ISA space. So handle that here.
414 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
415 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
418 addr = (volatile void __iomem *)
419 (PAGE_MASK & (unsigned long __force)addr);
421 mmiotrace_iounmap(addr);
423 /* Use the vm area unlocked, assuming the caller
424 ensures there isn't another iounmap for the same address
425 in parallel. Reuse of the virtual address is prevented by
426 leaving it in the global lists until we're done with it.
427 cpa takes care of the direct mappings. */
428 read_lock(&vmlist_lock);
429 for (p = vmlist; p; p = p->next) {
430 if (p->addr == (void __force *)addr)
433 read_unlock(&vmlist_lock);
436 printk(KERN_ERR "iounmap: bad address %p\n", addr);
441 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
443 /* Finally remove it */
444 o = remove_vm_area((void __force *)addr);
445 BUG_ON(p != o || o == NULL);
448 EXPORT_SYMBOL(iounmap);
451 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
454 void *xlate_dev_mem_ptr(unsigned long phys)
457 unsigned long start = phys & PAGE_MASK;
459 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
460 if (page_is_ram(start >> PAGE_SHIFT))
463 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
465 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
470 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
472 if (page_is_ram(phys >> PAGE_SHIFT))
475 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
479 static int __initdata early_ioremap_debug;
481 static int __init early_ioremap_debug_setup(char *str)
483 early_ioremap_debug = 1;
487 early_param("early_ioremap_debug", early_ioremap_debug_setup);
489 static __initdata int after_paging_init;
490 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
492 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
494 /* Don't assume we're using swapper_pg_dir at this point */
495 pgd_t *base = __va(read_cr3());
496 pgd_t *pgd = &base[pgd_index(addr)];
497 pud_t *pud = pud_offset(pgd, addr);
498 pmd_t *pmd = pmd_offset(pud, addr);
503 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
505 return &bm_pte[pte_index(addr)];
508 static unsigned long slot_virt[FIX_BTMAPS_SLOTS] __initdata;
510 void __init early_ioremap_init(void)
515 if (early_ioremap_debug)
516 printk(KERN_INFO "early_ioremap_init()\n");
518 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
519 slot_virt[i] = __fix_to_virt(FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*i);
521 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
522 memset(bm_pte, 0, sizeof(bm_pte));
523 pmd_populate_kernel(&init_mm, pmd, bm_pte);
526 * The boot-ioremap range spans multiple pmds, for which
527 * we are not prepared:
529 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
531 printk(KERN_WARNING "pmd %p != %p\n",
532 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
533 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
534 fix_to_virt(FIX_BTMAP_BEGIN));
535 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
536 fix_to_virt(FIX_BTMAP_END));
538 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
539 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
544 void __init early_ioremap_reset(void)
546 after_paging_init = 1;
549 static void __init __early_set_fixmap(enum fixed_addresses idx,
550 phys_addr_t phys, pgprot_t flags)
552 unsigned long addr = __fix_to_virt(idx);
555 if (idx >= __end_of_fixed_addresses) {
559 pte = early_ioremap_pte(addr);
561 if (pgprot_val(flags))
562 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
564 pte_clear(&init_mm, addr, pte);
565 __flush_tlb_one(addr);
568 static inline void __init early_set_fixmap(enum fixed_addresses idx,
569 phys_addr_t phys, pgprot_t prot)
571 if (after_paging_init)
572 __set_fixmap(idx, phys, prot);
574 __early_set_fixmap(idx, phys, prot);
577 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
579 if (after_paging_init)
582 __early_set_fixmap(idx, 0, __pgprot(0));
585 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
586 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
588 static int __init check_early_ioremap_leak(void)
593 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
600 "Debug warning: early ioremap leak of %d areas detected.\n",
603 "please boot with early_ioremap_debug and report the dmesg.\n");
607 late_initcall(check_early_ioremap_leak);
609 static void __init __iomem *
610 __early_ioremap(resource_size_t phys_addr, unsigned long size, pgprot_t prot)
612 unsigned long offset;
613 resource_size_t last_addr;
614 unsigned int nrpages;
615 enum fixed_addresses idx0, idx;
618 WARN_ON(system_state != SYSTEM_BOOTING);
621 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
629 printk(KERN_INFO "early_iomap(%08llx, %08lx) not found slot\n",
630 (u64)phys_addr, size);
635 if (early_ioremap_debug) {
636 printk(KERN_INFO "early_ioremap(%08llx, %08lx) [%d] => ",
637 (u64)phys_addr, size, slot);
641 /* Don't allow wraparound or zero size */
642 last_addr = phys_addr + size - 1;
643 if (!size || last_addr < phys_addr) {
648 prev_size[slot] = size;
650 * Mappings have to be page-aligned
652 offset = phys_addr & ~PAGE_MASK;
653 phys_addr &= PAGE_MASK;
654 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
657 * Mappings have to fit in the FIX_BTMAP area.
659 nrpages = size >> PAGE_SHIFT;
660 if (nrpages > NR_FIX_BTMAPS) {
668 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
670 while (nrpages > 0) {
671 early_set_fixmap(idx, phys_addr, prot);
672 phys_addr += PAGE_SIZE;
676 if (early_ioremap_debug)
677 printk(KERN_CONT "%08lx + %08lx\n", offset, slot_virt[slot]);
679 prev_map[slot] = (void __iomem *)(offset + slot_virt[slot]);
680 return prev_map[slot];
683 /* Remap an IO device */
684 void __init __iomem *
685 early_ioremap(resource_size_t phys_addr, unsigned long size)
687 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
691 void __init __iomem *
692 early_memremap(resource_size_t phys_addr, unsigned long size)
694 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
697 void __init early_iounmap(void __iomem *addr, unsigned long size)
699 unsigned long virt_addr;
700 unsigned long offset;
701 unsigned int nrpages;
702 enum fixed_addresses idx;
706 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
707 if (prev_map[i] == addr) {
714 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
720 if (prev_size[slot] != size) {
721 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
722 addr, size, slot, prev_size[slot]);
727 if (early_ioremap_debug) {
728 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
733 virt_addr = (unsigned long)addr;
734 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
738 offset = virt_addr & ~PAGE_MASK;
739 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
741 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
742 while (nrpages > 0) {
743 early_clear_fixmap(idx);
747 prev_map[slot] = NULL;