arch/x86/mm/ioremap.c

   1 /*
   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
   5  *
   6  * (C) Copyright 1995 1996 Linus Torvalds
   7  */
   8
   9 #include <linux/bootmem.h>
  10 #include <linux/init.h>
  11 #include <linux/io.h>
  12 #include <linux/module.h>
  13 #include <linux/slab.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/mmiotrace.h>
  16
  17 #include <asm/cacheflush.h>
  18 #include <asm/e820.h>
  19 #include <asm/fixmap.h>
  20 #include <asm/pgtable.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/pgalloc.h>
  23 #include <asm/pat.h>
  24
  25 #include "physaddr.h"
  26
  27 /*
  28  * Fix up the linear direct mapping of the kernel to avoid cache attribute
  29  * conflicts.
  30  */
  31 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
  32                                unsigned long prot_val)
  33 {
  34         unsigned long nrpages = size >> PAGE_SHIFT;
  35         int err;
  36
  37         switch (prot_val) {
  38         case _PAGE_CACHE_UC:
  39         default:
  40                 err = _set_memory_uc(vaddr, nrpages);
  41                 break;
  42         case _PAGE_CACHE_WC:
  43                 err = _set_memory_wc(vaddr, nrpages);
  44                 break;
  45         case _PAGE_CACHE_WB:
  46                 err = _set_memory_wb(vaddr, nrpages);
  47                 break;
  48         }
  49
  50         return err;
  51 }
  52
  53 /*
  54  * Remap an arbitrary physical address space into the kernel virtual
  55  * address space. Needed when the kernel wants to access high addresses
  56  * directly.
  57  *
  58  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
  59  * have to convert them into an offset in a page-aligned mapping, but the
  60  * caller shouldn't need to know that small detail.
  61  */
  62 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
  63                 unsigned long size, unsigned long prot_val, void *caller)
  64 {
  65         unsigned long offset, vaddr;
  66         resource_size_t pfn, last_pfn, last_addr;
  67         const resource_size_t unaligned_phys_addr = phys_addr;
  68         const unsigned long unaligned_size = size;
  69         struct vm_struct *area;
  70         unsigned long new_prot_val;
  71         pgprot_t prot;
  72         int retval;
  73         void __iomem *ret_addr;
  74
  75         /* Don't allow wraparound or zero size */
  76         last_addr = phys_addr + size - 1;
  77         if (!size || last_addr < phys_addr)
  78                 return NULL;
  79
  80         if (!phys_addr_valid(phys_addr)) {
  81                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
  82                        (unsigned long long)phys_addr);
  83                 WARN_ON_ONCE(1);
  84                 return NULL;
  85         }
  86
  87         /*
  88          * Don't remap the low PCI/ISA area, it's always mapped..
  89          */
  90         if (is_ISA_range(phys_addr, last_addr))
  91                 return (__force void __iomem *)phys_to_virt(phys_addr);
  92
  93         /*
  94          * Don't allow anybody to remap normal RAM that we're using..
  95          */
  96         last_pfn = last_addr >> PAGE_SHIFT;
  97         for (pfn = phys_addr >> PAGE_SHIFT; pfn <= last_pfn; pfn++) {
  98                 int is_ram = page_is_ram(pfn);
  99
 100                 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
 101                         return NULL;
 102                 WARN_ON_ONCE(is_ram);
 103         }
 104
 105         /*
 106          * Mappings have to be page-aligned
 107          */
 108         offset = phys_addr & ~PAGE_MASK;
 109         phys_addr &= PHYSICAL_PAGE_MASK;
 110         size = PAGE_ALIGN(last_addr+1) - phys_addr;
 111
 112         retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
 113                                                 prot_val, &new_prot_val);
 114         if (retval) {
 115                 printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
 116                 return NULL;
 117         }
 118
 119         if (prot_val != new_prot_val) {
 120                 if (!is_new_memtype_allowed(phys_addr, size,
 121                                             prot_val, new_prot_val)) {
 122                         printk(KERN_ERR
 123                 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
 124                                 (unsigned long long)phys_addr,
 125                                 (unsigned long long)(phys_addr + size),
 126                                 prot_val, new_prot_val);
 127                         goto err_free_memtype;
 128                 }
 129                 prot_val = new_prot_val;
 130         }
 131
 132         switch (prot_val) {
 133         case _PAGE_CACHE_UC:
 134         default:
 135                 prot = PAGE_KERNEL_IO_NOCACHE;
 136                 break;
 137         case _PAGE_CACHE_UC_MINUS:
 138                 prot = PAGE_KERNEL_IO_UC_MINUS;
 139                 break;
 140         case _PAGE_CACHE_WC:
 141                 prot = PAGE_KERNEL_IO_WC;
 142                 break;
 143         case _PAGE_CACHE_WB:
 144                 prot = PAGE_KERNEL_IO;
 145                 break;
 146         }
 147
 148         /*
 149          * Ok, go for it..
 150          */
 151         area = get_vm_area_caller(size, VM_IOREMAP, caller);
 152         if (!area)
 153                 goto err_free_memtype;
 154         area->phys_addr = phys_addr;
 155         vaddr = (unsigned long) area->addr;
 156
 157         if (kernel_map_sync_memtype(phys_addr, size, prot_val))
 158                 goto err_free_area;
 159
 160         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
 161                 goto err_free_area;
 162
 163         ret_addr = (void __iomem *) (vaddr + offset);
 164         mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
 165
 166         /*
 167          * Check if the request spans more than any BAR in the iomem resource
 168          * tree.
 169          */
 170         WARN_ONCE(iomem_map_sanity_check(unaligned_phys_addr, unaligned_size),
 171                   KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
 172
 173         return ret_addr;
 174 err_free_area:
 175         free_vm_area(area);
 176 err_free_memtype:
 177         free_memtype(phys_addr, phys_addr + size);
 178         return NULL;
 179 }
 180
 181 /**
 182  * ioremap_nocache     -   map bus memory into CPU space
 183  * @phys_addr:    bus address of the memory
 184  * @size:      size of the resource to map
 185  *
 186  * ioremap_nocache performs a platform specific sequence of operations to
 187  * make bus memory CPU accessible via the readb/readw/readl/writeb/
 188  * writew/writel functions and the other mmio helpers. The returned
 189  * address is not guaranteed to be usable directly as a virtual
 190  * address.
 191  *
 192  * This version of ioremap ensures that the memory is marked uncachable
 193  * on the CPU as well as honouring existing caching rules from things like
 194  * the PCI bus. Note that there are other caches and buffers on many
 195  * busses. In particular driver authors should read up on PCI writes
 196  *
 197  * It's useful if some control registers are in such an area and
 198  * write combining or read caching is not desirable:
 199  *
 200  * Must be freed with iounmap.
 201  */
 202 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 203 {
 204         /*
 205          * Ideally, this should be:
 206          *      pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
 207          *
 208          * Till we fix all X drivers to use ioremap_wc(), we will use
 209          * UC MINUS.
 210          */
 211         unsigned long val = _PAGE_CACHE_UC_MINUS;
 212
 213         return __ioremap_caller(phys_addr, size, val,
 214                                 __builtin_return_address(0));
 215 }
 216 EXPORT_SYMBOL(ioremap_nocache);
 217
 218 /**
 219  * ioremap_wc   -       map memory into CPU space write combined
 220  * @phys_addr:  bus address of the memory
 221  * @size:       size of the resource to map
 222  *
 223  * This version of ioremap ensures that the memory is marked write combining.
 224  * Write combining allows faster writes to some hardware devices.
 225  *
 226  * Must be freed with iounmap.
 227  */
 228 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
 229 {
 230         if (pat_enabled)
 231                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
 232                                         __builtin_return_address(0));
 233         else
 234                 return ioremap_nocache(phys_addr, size);
 235 }
 236 EXPORT_SYMBOL(ioremap_wc);
 237
 238 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
 239 {
 240         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
 241                                 __builtin_return_address(0));
 242 }
 243 EXPORT_SYMBOL(ioremap_cache);
 244
 245 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
 246                                 unsigned long prot_val)
 247 {
 248         return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
 249                                 __builtin_return_address(0));
 250 }
 251 EXPORT_SYMBOL(ioremap_prot);
 252
 253 /**
 254  * iounmap - Free a IO remapping
 255  * @addr: virtual address from ioremap_*
 256  *
 257  * Caller must ensure there is only one unmapping for the same pointer.
 258  */
 259 void iounmap(volatile void __iomem *addr)
 260 {
 261         struct vm_struct *p, *o;
 262
 263         if ((void __force *)addr <= high_memory)
 264                 return;
 265
 266         /*
 267          * __ioremap special-cases the PCI/ISA range by not instantiating a
 268          * vm_area and by simply returning an address into the kernel mapping
 269          * of ISA space.   So handle that here.
 270          */
 271         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
 272             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
 273                 return;
 274
 275         addr = (volatile void __iomem *)
 276                 (PAGE_MASK & (unsigned long __force)addr);
 277
 278         mmiotrace_iounmap(addr);
 279
 280         /* Use the vm area unlocked, assuming the caller
 281            ensures there isn't another iounmap for the same address
 282            in parallel. Reuse of the virtual address is prevented by
 283            leaving it in the global lists until we're done with it.
 284            cpa takes care of the direct mappings. */
 285         p = find_vm_area((void __force *)addr);
 286
 287         if (!p) {
 288                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
 289                 dump_stack();
 290                 return;
 291         }
 292
 293         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
 294
 295         /* Finally remove it */
 296         o = remove_vm_area((void __force *)addr);
 297         BUG_ON(p != o || o == NULL);
 298         kfree(p);
 299 }
 300 EXPORT_SYMBOL(iounmap);
 301
 302 /*
 303  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
 304  * access
 305  */
 306 void *xlate_dev_mem_ptr(unsigned long phys)
 307 {
 308         void *addr;
 309         unsigned long start = phys & PAGE_MASK;
 310
 311         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
 312         if (page_is_ram(start >> PAGE_SHIFT))
 313                 return __va(phys);
 314
 315         addr = (void __force *)ioremap_cache(start, PAGE_SIZE);
 316         if (addr)
 317                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
 318
 319         return addr;
 320 }
 321
 322 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
 323 {
 324         if (page_is_ram(phys >> PAGE_SHIFT))
 325                 return;
 326
 327         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
 328         return;
 329 }
 330
 331 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
 332
 333 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
 334 {
 335         /* Don't assume we're using swapper_pg_dir at this point */
 336         pgd_t *base = __va(read_cr3());
 337         pgd_t *pgd = &base[pgd_index(addr)];
 338         pud_t *pud = pud_offset(pgd, addr);
 339         pmd_t *pmd = pmd_offset(pud, addr);
 340
 341         return pmd;
 342 }
 343
 344 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
 345 {
 346         return &bm_pte[pte_index(addr)];
 347 }
 348
 349 bool __init is_early_ioremap_ptep(pte_t *ptep)
 350 {
 351         return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
 352 }
 353
 354 void __init early_ioremap_init(void)
 355 {
 356         pmd_t *pmd;
 357
 358 #ifdef CONFIG_X86_64
 359         BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 360 #else
 361         WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 362 #endif
 363
 364         early_ioremap_setup();
 365
 366         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 367         memset(bm_pte, 0, sizeof(bm_pte));
 368         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 369
 370         /*
 371          * The boot-ioremap range spans multiple pmds, for which
 372          * we are not prepared:
 373          */
 374 #define __FIXADDR_TOP (-PAGE_SIZE)
 375         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 376                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 377 #undef __FIXADDR_TOP
 378         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 379                 WARN_ON(1);
 380                 printk(KERN_WARNING "pmd %p != %p\n",
 381                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
 382                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 383                         fix_to_virt(FIX_BTMAP_BEGIN));
 384                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 385                         fix_to_virt(FIX_BTMAP_END));
 386
 387                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 388                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
 389                        FIX_BTMAP_BEGIN);
 390         }
 391 }
 392
 393 void __init __early_set_fixmap(enum fixed_addresses idx,
 394                                phys_addr_t phys, pgprot_t flags)
 395 {
 396         unsigned long addr = __fix_to_virt(idx);
 397         pte_t *pte;
 398
 399         if (idx >= __end_of_fixed_addresses) {
 400                 BUG();
 401                 return;
 402         }
 403         pte = early_ioremap_pte(addr);
 404
 405         if (pgprot_val(flags))
 406                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 407         else
 408                 pte_clear(&init_mm, addr, pte);
 409         __flush_tlb_one(addr);
 410 }