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 */
80 VIRTUAL_BUG_ON(x < PAGE_OFFSET);
81 VIRTUAL_BUG_ON(__vmalloc_start_set && is_vmalloc_addr((void *) x));
82 return x - PAGE_OFFSET;
84 EXPORT_SYMBOL(__phys_addr);
87 bool __virt_addr_valid(unsigned long x)
91 if (__vmalloc_start_set && is_vmalloc_addr((void *) x))
93 return pfn_valid((x - PAGE_OFFSET) >> PAGE_SHIFT);
95 EXPORT_SYMBOL(__virt_addr_valid);
99 int page_is_ram(unsigned long pagenr)
101 resource_size_t addr, end;
105 * A special case is the first 4Kb of memory;
106 * This is a BIOS owned area, not kernel ram, but generally
107 * not listed as such in the E820 table.
113 * Second special case: Some BIOSen report the PC BIOS
114 * area (640->1Mb) as ram even though it is not.
116 if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
117 pagenr < (BIOS_END >> PAGE_SHIFT))
120 for (i = 0; i < e820.nr_map; i++) {
124 if (e820.map[i].type != E820_RAM)
126 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
127 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
130 if ((pagenr >= addr) && (pagenr < end))
137 * Fix up the linear direct mapping of the kernel to avoid cache attribute
140 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
141 unsigned long prot_val)
143 unsigned long nrpages = size >> PAGE_SHIFT;
149 err = _set_memory_uc(vaddr, nrpages);
152 err = _set_memory_wc(vaddr, nrpages);
155 err = _set_memory_wb(vaddr, nrpages);
163 * Remap an arbitrary physical address space into the kernel virtual
164 * address space. Needed when the kernel wants to access high addresses
167 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
168 * have to convert them into an offset in a page-aligned mapping, but the
169 * caller shouldn't need to know that small detail.
171 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
172 unsigned long size, unsigned long prot_val, void *caller)
174 unsigned long pfn, offset, vaddr;
175 resource_size_t last_addr;
176 const resource_size_t unaligned_phys_addr = phys_addr;
177 const unsigned long unaligned_size = size;
178 struct vm_struct *area;
179 unsigned long new_prot_val;
182 void __iomem *ret_addr;
184 /* Don't allow wraparound or zero size */
185 last_addr = phys_addr + size - 1;
186 if (!size || last_addr < phys_addr)
189 if (!phys_addr_valid(phys_addr)) {
190 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
191 (unsigned long long)phys_addr);
197 * Don't remap the low PCI/ISA area, it's always mapped..
199 if (is_ISA_range(phys_addr, last_addr))
200 return (__force void __iomem *)phys_to_virt(phys_addr);
203 * Check if the request spans more than any BAR in the iomem resource
206 WARN_ONCE(iomem_map_sanity_check(phys_addr, size),
207 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
210 * Don't allow anybody to remap normal RAM that we're using..
212 for (pfn = phys_addr >> PAGE_SHIFT;
213 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
216 int is_ram = page_is_ram(pfn);
218 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
220 WARN_ON_ONCE(is_ram);
224 * Mappings have to be page-aligned
226 offset = phys_addr & ~PAGE_MASK;
227 phys_addr &= PAGE_MASK;
228 size = PAGE_ALIGN(last_addr+1) - phys_addr;
230 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
231 prot_val, &new_prot_val);
233 pr_debug("Warning: reserve_memtype returned %d\n", retval);
237 if (prot_val != new_prot_val) {
239 * Do not fallback to certain memory types with certain
241 * - request is uc-, return cannot be write-back
242 * - request is uc-, return cannot be write-combine
243 * - request is write-combine, return cannot be write-back
245 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
246 (new_prot_val == _PAGE_CACHE_WB ||
247 new_prot_val == _PAGE_CACHE_WC)) ||
248 (prot_val == _PAGE_CACHE_WC &&
249 new_prot_val == _PAGE_CACHE_WB)) {
251 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
252 (unsigned long long)phys_addr,
253 (unsigned long long)(phys_addr + size),
254 prot_val, new_prot_val);
255 free_memtype(phys_addr, phys_addr + size);
258 prot_val = new_prot_val;
264 prot = PAGE_KERNEL_IO_NOCACHE;
266 case _PAGE_CACHE_UC_MINUS:
267 prot = PAGE_KERNEL_IO_UC_MINUS;
270 prot = PAGE_KERNEL_IO_WC;
273 prot = PAGE_KERNEL_IO;
280 area = get_vm_area_caller(size, VM_IOREMAP, caller);
283 area->phys_addr = phys_addr;
284 vaddr = (unsigned long) area->addr;
285 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
286 free_memtype(phys_addr, phys_addr + size);
291 if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
292 free_memtype(phys_addr, phys_addr + size);
297 ret_addr = (void __iomem *) (vaddr + offset);
298 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
304 * ioremap_nocache - map bus memory into CPU space
305 * @offset: bus address of the memory
306 * @size: size of the resource to map
308 * ioremap_nocache performs a platform specific sequence of operations to
309 * make bus memory CPU accessible via the readb/readw/readl/writeb/
310 * writew/writel functions and the other mmio helpers. The returned
311 * address is not guaranteed to be usable directly as a virtual
314 * This version of ioremap ensures that the memory is marked uncachable
315 * on the CPU as well as honouring existing caching rules from things like
316 * the PCI bus. Note that there are other caches and buffers on many
317 * busses. In particular driver authors should read up on PCI writes
319 * It's useful if some control registers are in such an area and
320 * write combining or read caching is not desirable:
322 * Must be freed with iounmap.
324 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
327 * Ideally, this should be:
328 * pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
330 * Till we fix all X drivers to use ioremap_wc(), we will use
333 unsigned long val = _PAGE_CACHE_UC_MINUS;
335 return __ioremap_caller(phys_addr, size, val,
336 __builtin_return_address(0));
338 EXPORT_SYMBOL(ioremap_nocache);
341 * ioremap_wc - map memory into CPU space write combined
342 * @offset: bus address of the memory
343 * @size: size of the resource to map
345 * This version of ioremap ensures that the memory is marked write combining.
346 * Write combining allows faster writes to some hardware devices.
348 * Must be freed with iounmap.
350 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
353 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
354 __builtin_return_address(0));
356 return ioremap_nocache(phys_addr, size);
358 EXPORT_SYMBOL(ioremap_wc);
360 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
362 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
363 __builtin_return_address(0));
365 EXPORT_SYMBOL(ioremap_cache);
367 static void __iomem *ioremap_default(resource_size_t phys_addr,
375 * - WB for WB-able memory and no other conflicting mappings
376 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
377 * - Inherit from confliting mappings otherwise
379 err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
383 ret = __ioremap_caller(phys_addr, size, flags,
384 __builtin_return_address(0));
386 free_memtype(phys_addr, phys_addr + size);
390 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
391 unsigned long prot_val)
393 return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
394 __builtin_return_address(0));
396 EXPORT_SYMBOL(ioremap_prot);
399 * iounmap - Free a IO remapping
400 * @addr: virtual address from ioremap_*
402 * Caller must ensure there is only one unmapping for the same pointer.
404 void iounmap(volatile void __iomem *addr)
406 struct vm_struct *p, *o;
408 if ((void __force *)addr <= high_memory)
412 * __ioremap special-cases the PCI/ISA range by not instantiating a
413 * vm_area and by simply returning an address into the kernel mapping
414 * of ISA space. So handle that here.
416 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
417 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
420 addr = (volatile void __iomem *)
421 (PAGE_MASK & (unsigned long __force)addr);
423 mmiotrace_iounmap(addr);
425 /* Use the vm area unlocked, assuming the caller
426 ensures there isn't another iounmap for the same address
427 in parallel. Reuse of the virtual address is prevented by
428 leaving it in the global lists until we're done with it.
429 cpa takes care of the direct mappings. */
430 read_lock(&vmlist_lock);
431 for (p = vmlist; p; p = p->next) {
432 if (p->addr == (void __force *)addr)
435 read_unlock(&vmlist_lock);
438 printk(KERN_ERR "iounmap: bad address %p\n", addr);
443 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
445 /* Finally remove it */
446 o = remove_vm_area((void __force *)addr);
447 BUG_ON(p != o || o == NULL);
450 EXPORT_SYMBOL(iounmap);
453 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
456 void *xlate_dev_mem_ptr(unsigned long phys)
459 unsigned long start = phys & PAGE_MASK;
461 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
462 if (page_is_ram(start >> PAGE_SHIFT))
465 addr = (void __force *)ioremap_default(start, PAGE_SIZE);
467 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
472 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
474 if (page_is_ram(phys >> PAGE_SHIFT))
477 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
481 static int __initdata early_ioremap_debug;
483 static int __init early_ioremap_debug_setup(char *str)
485 early_ioremap_debug = 1;
489 early_param("early_ioremap_debug", early_ioremap_debug_setup);
491 static __initdata int after_paging_init;
492 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
494 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
496 /* Don't assume we're using swapper_pg_dir at this point */
497 pgd_t *base = __va(read_cr3());
498 pgd_t *pgd = &base[pgd_index(addr)];
499 pud_t *pud = pud_offset(pgd, addr);
500 pmd_t *pmd = pmd_offset(pud, addr);
505 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
507 return &bm_pte[pte_index(addr)];
510 void __init early_ioremap_init(void)
514 if (early_ioremap_debug)
515 printk(KERN_INFO "early_ioremap_init()\n");
517 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
518 memset(bm_pte, 0, sizeof(bm_pte));
519 pmd_populate_kernel(&init_mm, pmd, bm_pte);
522 * The boot-ioremap range spans multiple pmds, for which
523 * we are not prepared:
525 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
527 printk(KERN_WARNING "pmd %p != %p\n",
528 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
529 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
530 fix_to_virt(FIX_BTMAP_BEGIN));
531 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
532 fix_to_virt(FIX_BTMAP_END));
534 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
535 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
540 void __init early_ioremap_reset(void)
542 after_paging_init = 1;
545 static void __init __early_set_fixmap(enum fixed_addresses idx,
546 unsigned long phys, pgprot_t flags)
548 unsigned long addr = __fix_to_virt(idx);
551 if (idx >= __end_of_fixed_addresses) {
555 pte = early_ioremap_pte(addr);
557 if (pgprot_val(flags))
558 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
560 pte_clear(&init_mm, addr, pte);
561 __flush_tlb_one(addr);
564 static inline void __init early_set_fixmap(enum fixed_addresses idx,
565 unsigned long phys, pgprot_t prot)
567 if (after_paging_init)
568 __set_fixmap(idx, phys, prot);
570 __early_set_fixmap(idx, phys, prot);
573 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
575 if (after_paging_init)
578 __early_set_fixmap(idx, 0, __pgprot(0));
581 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
582 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
583 static int __init check_early_ioremap_leak(void)
588 for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
595 "Debug warning: early ioremap leak of %d areas detected.\n",
598 "please boot with early_ioremap_debug and report the dmesg.\n");
602 late_initcall(check_early_ioremap_leak);
604 static void __init __iomem *__early_ioremap(unsigned long phys_addr, unsigned long size, pgprot_t prot)
606 unsigned long offset, last_addr;
607 unsigned int nrpages;
608 enum fixed_addresses idx0, idx;
611 WARN_ON(system_state != SYSTEM_BOOTING);
614 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
622 printk(KERN_INFO "early_iomap(%08lx, %08lx) not found slot\n",
628 if (early_ioremap_debug) {
629 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
630 phys_addr, size, slot);
634 /* Don't allow wraparound or zero size */
635 last_addr = phys_addr + size - 1;
636 if (!size || last_addr < phys_addr) {
641 prev_size[slot] = size;
643 * Mappings have to be page-aligned
645 offset = phys_addr & ~PAGE_MASK;
646 phys_addr &= PAGE_MASK;
647 size = PAGE_ALIGN(last_addr + 1) - phys_addr;
650 * Mappings have to fit in the FIX_BTMAP area.
652 nrpages = size >> PAGE_SHIFT;
653 if (nrpages > NR_FIX_BTMAPS) {
661 idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
663 while (nrpages > 0) {
664 early_set_fixmap(idx, phys_addr, prot);
665 phys_addr += PAGE_SIZE;
669 if (early_ioremap_debug)
670 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
672 prev_map[slot] = (void __iomem *)(offset + fix_to_virt(idx0));
673 return prev_map[slot];
676 /* Remap an IO device */
677 void __init __iomem *early_ioremap(unsigned long phys_addr, unsigned long size)
679 return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
683 void __init __iomem *early_memremap(unsigned long phys_addr, unsigned long size)
685 return __early_ioremap(phys_addr, size, PAGE_KERNEL);
688 void __init early_iounmap(void __iomem *addr, unsigned long size)
690 unsigned long virt_addr;
691 unsigned long offset;
692 unsigned int nrpages;
693 enum fixed_addresses idx;
697 for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
698 if (prev_map[i] == addr) {
705 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
711 if (prev_size[slot] != size) {
712 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
713 addr, size, slot, prev_size[slot]);
718 if (early_ioremap_debug) {
719 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
724 virt_addr = (unsigned long)addr;
725 if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
729 offset = virt_addr & ~PAGE_MASK;
730 nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
732 idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
733 while (nrpages > 0) {
734 early_clear_fixmap(idx);
738 prev_map[slot] = NULL;
741 void __this_fixmap_does_not_exist(void)