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>
28 * Fix up the linear direct mapping of the kernel to avoid cache attribute
31 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
32 enum page_cache_mode pcm)
34 unsigned long nrpages = size >> PAGE_SHIFT;
38 case _PAGE_CACHE_MODE_UC:
40 err = _set_memory_uc(vaddr, nrpages);
42 case _PAGE_CACHE_MODE_WC:
43 err = _set_memory_wc(vaddr, nrpages);
45 case _PAGE_CACHE_MODE_WT:
46 err = _set_memory_wt(vaddr, nrpages);
48 case _PAGE_CACHE_MODE_WB:
49 err = _set_memory_wb(vaddr, nrpages);
56 static int __ioremap_check_ram(unsigned long start_pfn, unsigned long nr_pages,
61 for (i = 0; i < nr_pages; ++i)
62 if (pfn_valid(start_pfn + i) &&
63 !PageReserved(pfn_to_page(start_pfn + i)))
70 * Remap an arbitrary physical address space into the kernel virtual
71 * address space. It transparently creates kernel huge I/O mapping when
72 * the physical address is aligned by a huge page size (1GB or 2MB) and
73 * the requested size is at least the huge page size.
75 * NOTE: MTRRs can override PAT memory types with a 4KB granularity.
76 * Therefore, the mapping code falls back to use a smaller page toward 4KB
77 * when a mapping range is covered by non-WB type of MTRRs.
79 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
80 * have to convert them into an offset in a page-aligned mapping, but the
81 * caller shouldn't need to know that small detail.
83 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
84 unsigned long size, enum page_cache_mode pcm, void *caller)
86 unsigned long offset, vaddr;
87 resource_size_t pfn, last_pfn, last_addr;
88 const resource_size_t unaligned_phys_addr = phys_addr;
89 const unsigned long unaligned_size = size;
90 struct vm_struct *area;
91 enum page_cache_mode new_pcm;
94 void __iomem *ret_addr;
97 /* Don't allow wraparound or zero size */
98 last_addr = phys_addr + size - 1;
99 if (!size || last_addr < phys_addr)
102 if (!phys_addr_valid(phys_addr)) {
103 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
104 (unsigned long long)phys_addr);
110 * Don't remap the low PCI/ISA area, it's always mapped..
112 if (is_ISA_range(phys_addr, last_addr))
113 return (__force void __iomem *)phys_to_virt(phys_addr);
116 * Don't allow anybody to remap normal RAM that we're using..
118 /* First check if whole region can be identified as RAM or not */
119 ram_region = region_is_ram(phys_addr, size);
120 if (ram_region > 0) {
121 WARN_ONCE(1, "ioremap on RAM at 0x%lx - 0x%lx\n",
122 (unsigned long int)phys_addr,
123 (unsigned long int)last_addr);
127 /* If could not be identified(-1), check page by page */
128 if (ram_region < 0) {
129 pfn = phys_addr >> PAGE_SHIFT;
130 last_pfn = last_addr >> PAGE_SHIFT;
131 if (walk_system_ram_range(pfn, last_pfn - pfn + 1, NULL,
132 __ioremap_check_ram) == 1) {
133 WARN_ONCE(1, "ioremap on RAM at 0x%llx - 0x%llx\n",
134 phys_addr, last_addr);
139 * Mappings have to be page-aligned
141 offset = phys_addr & ~PAGE_MASK;
142 phys_addr &= PHYSICAL_PAGE_MASK;
143 size = PAGE_ALIGN(last_addr+1) - phys_addr;
145 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
148 printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
152 if (pcm != new_pcm) {
153 if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
155 "ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
156 (unsigned long long)phys_addr,
157 (unsigned long long)(phys_addr + size),
159 goto err_free_memtype;
164 prot = PAGE_KERNEL_IO;
166 case _PAGE_CACHE_MODE_UC:
168 prot = __pgprot(pgprot_val(prot) |
169 cachemode2protval(_PAGE_CACHE_MODE_UC));
171 case _PAGE_CACHE_MODE_UC_MINUS:
172 prot = __pgprot(pgprot_val(prot) |
173 cachemode2protval(_PAGE_CACHE_MODE_UC_MINUS));
175 case _PAGE_CACHE_MODE_WC:
176 prot = __pgprot(pgprot_val(prot) |
177 cachemode2protval(_PAGE_CACHE_MODE_WC));
179 case _PAGE_CACHE_MODE_WT:
180 prot = __pgprot(pgprot_val(prot) |
181 cachemode2protval(_PAGE_CACHE_MODE_WT));
183 case _PAGE_CACHE_MODE_WB:
190 area = get_vm_area_caller(size, VM_IOREMAP, caller);
192 goto err_free_memtype;
193 area->phys_addr = phys_addr;
194 vaddr = (unsigned long) area->addr;
196 if (kernel_map_sync_memtype(phys_addr, size, pcm))
199 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
202 ret_addr = (void __iomem *) (vaddr + offset);
203 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
206 * Check if the request spans more than any BAR in the iomem resource
209 WARN_ONCE(iomem_map_sanity_check(unaligned_phys_addr, unaligned_size),
210 KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
216 free_memtype(phys_addr, phys_addr + size);
221 * ioremap_nocache - map bus memory into CPU space
222 * @phys_addr: bus address of the memory
223 * @size: size of the resource to map
225 * ioremap_nocache performs a platform specific sequence of operations to
226 * make bus memory CPU accessible via the readb/readw/readl/writeb/
227 * writew/writel functions and the other mmio helpers. The returned
228 * address is not guaranteed to be usable directly as a virtual
231 * This version of ioremap ensures that the memory is marked uncachable
232 * on the CPU as well as honouring existing caching rules from things like
233 * the PCI bus. Note that there are other caches and buffers on many
234 * busses. In particular driver authors should read up on PCI writes
236 * It's useful if some control registers are in such an area and
237 * write combining or read caching is not desirable:
239 * Must be freed with iounmap.
241 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
244 * Ideally, this should be:
245 * pat_enabled() ? _PAGE_CACHE_MODE_UC : _PAGE_CACHE_MODE_UC_MINUS;
247 * Till we fix all X drivers to use ioremap_wc(), we will use
248 * UC MINUS. Drivers that are certain they need or can already
249 * be converted over to strong UC can use ioremap_uc().
251 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC_MINUS;
253 return __ioremap_caller(phys_addr, size, pcm,
254 __builtin_return_address(0));
256 EXPORT_SYMBOL(ioremap_nocache);
259 * ioremap_uc - map bus memory into CPU space as strongly uncachable
260 * @phys_addr: bus address of the memory
261 * @size: size of the resource to map
263 * ioremap_uc performs a platform specific sequence of operations to
264 * make bus memory CPU accessible via the readb/readw/readl/writeb/
265 * writew/writel functions and the other mmio helpers. The returned
266 * address is not guaranteed to be usable directly as a virtual
269 * This version of ioremap ensures that the memory is marked with a strong
270 * preference as completely uncachable on the CPU when possible. For non-PAT
271 * systems this ends up setting page-attribute flags PCD=1, PWT=1. For PAT
272 * systems this will set the PAT entry for the pages as strong UC. This call
273 * will honor existing caching rules from things like the PCI bus. Note that
274 * there are other caches and buffers on many busses. In particular driver
275 * authors should read up on PCI writes.
277 * It's useful if some control registers are in such an area and
278 * write combining or read caching is not desirable:
280 * Must be freed with iounmap.
282 void __iomem *ioremap_uc(resource_size_t phys_addr, unsigned long size)
284 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC;
286 return __ioremap_caller(phys_addr, size, pcm,
287 __builtin_return_address(0));
289 EXPORT_SYMBOL_GPL(ioremap_uc);
292 * ioremap_wc - map memory into CPU space write combined
293 * @phys_addr: bus address of the memory
294 * @size: size of the resource to map
296 * This version of ioremap ensures that the memory is marked write combining.
297 * Write combining allows faster writes to some hardware devices.
299 * Must be freed with iounmap.
301 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
303 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WC,
304 __builtin_return_address(0));
306 EXPORT_SYMBOL(ioremap_wc);
309 * ioremap_wt - map memory into CPU space write through
310 * @phys_addr: bus address of the memory
311 * @size: size of the resource to map
313 * This version of ioremap ensures that the memory is marked write through.
314 * Write through stores data into memory while keeping the cache up-to-date.
316 * Must be freed with iounmap.
318 void __iomem *ioremap_wt(resource_size_t phys_addr, unsigned long size)
320 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WT,
321 __builtin_return_address(0));
323 EXPORT_SYMBOL(ioremap_wt);
325 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
327 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
328 __builtin_return_address(0));
330 EXPORT_SYMBOL(ioremap_cache);
332 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
333 unsigned long prot_val)
335 return __ioremap_caller(phys_addr, size,
336 pgprot2cachemode(__pgprot(prot_val)),
337 __builtin_return_address(0));
339 EXPORT_SYMBOL(ioremap_prot);
342 * iounmap - Free a IO remapping
343 * @addr: virtual address from ioremap_*
345 * Caller must ensure there is only one unmapping for the same pointer.
347 void iounmap(volatile void __iomem *addr)
349 struct vm_struct *p, *o;
351 if ((void __force *)addr <= high_memory)
355 * __ioremap special-cases the PCI/ISA range by not instantiating a
356 * vm_area and by simply returning an address into the kernel mapping
357 * of ISA space. So handle that here.
359 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
360 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
363 addr = (volatile void __iomem *)
364 (PAGE_MASK & (unsigned long __force)addr);
366 mmiotrace_iounmap(addr);
368 /* Use the vm area unlocked, assuming the caller
369 ensures there isn't another iounmap for the same address
370 in parallel. Reuse of the virtual address is prevented by
371 leaving it in the global lists until we're done with it.
372 cpa takes care of the direct mappings. */
373 p = find_vm_area((void __force *)addr);
376 printk(KERN_ERR "iounmap: bad address %p\n", addr);
381 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
383 /* Finally remove it */
384 o = remove_vm_area((void __force *)addr);
385 BUG_ON(p != o || o == NULL);
388 EXPORT_SYMBOL(iounmap);
390 int __init arch_ioremap_pud_supported(void)
393 return cpu_has_gbpages;
399 int __init arch_ioremap_pmd_supported(void)
405 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
408 void *xlate_dev_mem_ptr(phys_addr_t phys)
410 unsigned long start = phys & PAGE_MASK;
411 unsigned long offset = phys & ~PAGE_MASK;
414 /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
415 if (page_is_ram(start >> PAGE_SHIFT))
418 vaddr = ioremap_cache(start, PAGE_SIZE);
419 /* Only add the offset on success and return NULL if the ioremap() failed: */
426 void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr)
428 if (page_is_ram(phys >> PAGE_SHIFT))
431 iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
434 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
436 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
438 /* Don't assume we're using swapper_pg_dir at this point */
439 pgd_t *base = __va(read_cr3());
440 pgd_t *pgd = &base[pgd_index(addr)];
441 pud_t *pud = pud_offset(pgd, addr);
442 pmd_t *pmd = pmd_offset(pud, addr);
447 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
449 return &bm_pte[pte_index(addr)];
452 bool __init is_early_ioremap_ptep(pte_t *ptep)
454 return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
457 void __init early_ioremap_init(void)
462 BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
464 WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
467 early_ioremap_setup();
469 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
470 memset(bm_pte, 0, sizeof(bm_pte));
471 pmd_populate_kernel(&init_mm, pmd, bm_pte);
474 * The boot-ioremap range spans multiple pmds, for which
475 * we are not prepared:
477 #define __FIXADDR_TOP (-PAGE_SIZE)
478 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
479 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
481 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
483 printk(KERN_WARNING "pmd %p != %p\n",
484 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
485 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
486 fix_to_virt(FIX_BTMAP_BEGIN));
487 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
488 fix_to_virt(FIX_BTMAP_END));
490 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
491 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
496 void __init __early_set_fixmap(enum fixed_addresses idx,
497 phys_addr_t phys, pgprot_t flags)
499 unsigned long addr = __fix_to_virt(idx);
502 if (idx >= __end_of_fixed_addresses) {
506 pte = early_ioremap_pte(addr);
508 if (pgprot_val(flags))
509 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
511 pte_clear(&init_mm, addr, pte);
512 __flush_tlb_one(addr);