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 static int __ioremap_check_ram(unsigned long start_pfn, unsigned long nr_pages,
  54                                void *arg)
  55 {
  56         unsigned long i;
  57
  58         for (i = 0; i < nr_pages; ++i)
  59                 if (pfn_valid(start_pfn + i) &&
  60                     !PageReserved(pfn_to_page(start_pfn + i)))
  61                         return 1;
  62
  63         WARN_ONCE(1, "ioremap on RAM pfn 0x%lx\n", start_pfn);
  64
  65         return 0;
  66 }
  67
  68 /*
  69  * Remap an arbitrary physical address space into the kernel virtual
  70  * address space. Needed when the kernel wants to access high addresses
  71  * directly.
  72  *
  73  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
  74  * have to convert them into an offset in a page-aligned mapping, but the
  75  * caller shouldn't need to know that small detail.
  76  */
  77 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
  78                 unsigned long size, unsigned long prot_val, void *caller)
  79 {
  80         unsigned long offset, vaddr;
  81         resource_size_t pfn, last_pfn, last_addr;
  82         const resource_size_t unaligned_phys_addr = phys_addr;
  83         const unsigned long unaligned_size = size;
  84         struct vm_struct *area;
  85         unsigned long new_prot_val;
  86         pgprot_t prot;
  87         int retval;
  88         void __iomem *ret_addr;
  89
  90         /* Don't allow wraparound or zero size */
  91         last_addr = phys_addr + size - 1;
  92         if (!size || last_addr < phys_addr)
  93                 return NULL;
  94
  95         if (!phys_addr_valid(phys_addr)) {
  96                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
  97                        (unsigned long long)phys_addr);
  98                 WARN_ON_ONCE(1);
  99                 return NULL;
 100         }
 101
 102         /*
 103          * Don't remap the low PCI/ISA area, it's always mapped..
 104          */
 105         if (is_ISA_range(phys_addr, last_addr))
 106                 return (__force void __iomem *)phys_to_virt(phys_addr);
 107
 108         /*
 109          * Don't allow anybody to remap normal RAM that we're using..
 110          */
 111         pfn      = phys_addr >> PAGE_SHIFT;
 112         last_pfn = last_addr >> PAGE_SHIFT;
 113         if (walk_system_ram_range(pfn, last_pfn - pfn + 1, NULL,
 114                                   __ioremap_check_ram) == 1)
 115                 return NULL;
 116
 117         /*
 118          * Mappings have to be page-aligned
 119          */
 120         offset = phys_addr & ~PAGE_MASK;
 121         phys_addr &= PHYSICAL_PAGE_MASK;
 122         size = PAGE_ALIGN(last_addr+1) - phys_addr;
 123
 124         retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
 125                                                 prot_val, &new_prot_val);
 126         if (retval) {
 127                 printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
 128                 return NULL;
 129         }
 130
 131         if (prot_val != new_prot_val) {
 132                 if (!is_new_memtype_allowed(phys_addr, size,
 133                                             prot_val, new_prot_val)) {
 134                         printk(KERN_ERR
 135                 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
 136                                 (unsigned long long)phys_addr,
 137                                 (unsigned long long)(phys_addr + size),
 138                                 prot_val, new_prot_val);
 139                         goto err_free_memtype;
 140                 }
 141                 prot_val = new_prot_val;
 142         }
 143
 144         switch (prot_val) {
 145         case _PAGE_CACHE_UC:
 146         default:
 147                 prot = PAGE_KERNEL_IO_NOCACHE;
 148                 break;
 149         case _PAGE_CACHE_UC_MINUS:
 150                 prot = PAGE_KERNEL_IO_UC_MINUS;
 151                 break;
 152         case _PAGE_CACHE_WC:
 153                 prot = PAGE_KERNEL_IO_WC;
 154                 break;
 155         case _PAGE_CACHE_WB:
 156                 prot = PAGE_KERNEL_IO;
 157                 break;
 158         }
 159
 160         /*
 161          * Ok, go for it..
 162          */
 163         area = get_vm_area_caller(size, VM_IOREMAP, caller);
 164         if (!area)
 165                 goto err_free_memtype;
 166         area->phys_addr = phys_addr;
 167         vaddr = (unsigned long) area->addr;
 168
 169         if (kernel_map_sync_memtype(phys_addr, size, prot_val))
 170                 goto err_free_area;
 171
 172         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
 173                 goto err_free_area;
 174
 175         ret_addr = (void __iomem *) (vaddr + offset);
 176         mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
 177
 178         /*
 179          * Check if the request spans more than any BAR in the iomem resource
 180          * tree.
 181          */
 182         WARN_ONCE(iomem_map_sanity_check(unaligned_phys_addr, unaligned_size),
 183                   KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
 184
 185         return ret_addr;
 186 err_free_area:
 187         free_vm_area(area);
 188 err_free_memtype:
 189         free_memtype(phys_addr, phys_addr + size);
 190         return NULL;
 191 }
 192
 193 /**
 194  * ioremap_nocache     -   map bus memory into CPU space
 195  * @phys_addr:    bus address of the memory
 196  * @size:      size of the resource to map
 197  *
 198  * ioremap_nocache performs a platform specific sequence of operations to
 199  * make bus memory CPU accessible via the readb/readw/readl/writeb/
 200  * writew/writel functions and the other mmio helpers. The returned
 201  * address is not guaranteed to be usable directly as a virtual
 202  * address.
 203  *
 204  * This version of ioremap ensures that the memory is marked uncachable
 205  * on the CPU as well as honouring existing caching rules from things like
 206  * the PCI bus. Note that there are other caches and buffers on many
 207  * busses. In particular driver authors should read up on PCI writes
 208  *
 209  * It's useful if some control registers are in such an area and
 210  * write combining or read caching is not desirable:
 211  *
 212  * Must be freed with iounmap.
 213  */
 214 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 215 {
 216         /*
 217          * Ideally, this should be:
 218          *      pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
 219          *
 220          * Till we fix all X drivers to use ioremap_wc(), we will use
 221          * UC MINUS.
 222          */
 223         unsigned long val = _PAGE_CACHE_UC_MINUS;
 224
 225         return __ioremap_caller(phys_addr, size, val,
 226                                 __builtin_return_address(0));
 227 }
 228 EXPORT_SYMBOL(ioremap_nocache);
 229
 230 /**
 231  * ioremap_wc   -       map memory into CPU space write combined
 232  * @phys_addr:  bus address of the memory
 233  * @size:       size of the resource to map
 234  *
 235  * This version of ioremap ensures that the memory is marked write combining.
 236  * Write combining allows faster writes to some hardware devices.
 237  *
 238  * Must be freed with iounmap.
 239  */
 240 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
 241 {
 242         if (pat_enabled)
 243                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
 244                                         __builtin_return_address(0));
 245         else
 246                 return ioremap_nocache(phys_addr, size);
 247 }
 248 EXPORT_SYMBOL(ioremap_wc);
 249
 250 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
 251 {
 252         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
 253                                 __builtin_return_address(0));
 254 }
 255 EXPORT_SYMBOL(ioremap_cache);
 256
 257 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
 258                                 unsigned long prot_val)
 259 {
 260         return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
 261                                 __builtin_return_address(0));
 262 }
 263 EXPORT_SYMBOL(ioremap_prot);
 264
 265 /**
 266  * iounmap - Free a IO remapping
 267  * @addr: virtual address from ioremap_*
 268  *
 269  * Caller must ensure there is only one unmapping for the same pointer.
 270  */
 271 void iounmap(volatile void __iomem *addr)
 272 {
 273         struct vm_struct *p, *o;
 274
 275         if ((void __force *)addr <= high_memory)
 276                 return;
 277
 278         /*
 279          * __ioremap special-cases the PCI/ISA range by not instantiating a
 280          * vm_area and by simply returning an address into the kernel mapping
 281          * of ISA space.   So handle that here.
 282          */
 283         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
 284             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
 285                 return;
 286
 287         addr = (volatile void __iomem *)
 288                 (PAGE_MASK & (unsigned long __force)addr);
 289
 290         mmiotrace_iounmap(addr);
 291
 292         /* Use the vm area unlocked, assuming the caller
 293            ensures there isn't another iounmap for the same address
 294            in parallel. Reuse of the virtual address is prevented by
 295            leaving it in the global lists until we're done with it.
 296            cpa takes care of the direct mappings. */
 297         p = find_vm_area((void __force *)addr);
 298
 299         if (!p) {
 300                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
 301                 dump_stack();
 302                 return;
 303         }
 304
 305         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
 306
 307         /* Finally remove it */
 308         o = remove_vm_area((void __force *)addr);
 309         BUG_ON(p != o || o == NULL);
 310         kfree(p);
 311 }
 312 EXPORT_SYMBOL(iounmap);
 313
 314 /*
 315  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
 316  * access
 317  */
 318 void *xlate_dev_mem_ptr(unsigned long phys)
 319 {
 320         void *addr;
 321         unsigned long start = phys & PAGE_MASK;
 322
 323         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
 324         if (page_is_ram(start >> PAGE_SHIFT))
 325                 return __va(phys);
 326
 327         addr = (void __force *)ioremap_cache(start, PAGE_SIZE);
 328         if (addr)
 329                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
 330
 331         return addr;
 332 }
 333
 334 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
 335 {
 336         if (page_is_ram(phys >> PAGE_SHIFT))
 337                 return;
 338
 339         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
 340         return;
 341 }
 342
 343 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
 344
 345 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
 346 {
 347         /* Don't assume we're using swapper_pg_dir at this point */
 348         pgd_t *base = __va(read_cr3());
 349         pgd_t *pgd = &base[pgd_index(addr)];
 350         pud_t *pud = pud_offset(pgd, addr);
 351         pmd_t *pmd = pmd_offset(pud, addr);
 352
 353         return pmd;
 354 }
 355
 356 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
 357 {
 358         return &bm_pte[pte_index(addr)];
 359 }
 360
 361 bool __init is_early_ioremap_ptep(pte_t *ptep)
 362 {
 363         return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
 364 }
 365
 366 void __init early_ioremap_init(void)
 367 {
 368         pmd_t *pmd;
 369
 370 #ifdef CONFIG_X86_64
 371         BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 372 #else
 373         WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 374 #endif
 375
 376         early_ioremap_setup();
 377
 378         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 379         memset(bm_pte, 0, sizeof(bm_pte));
 380         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 381
 382         /*
 383          * The boot-ioremap range spans multiple pmds, for which
 384          * we are not prepared:
 385          */
 386 #define __FIXADDR_TOP (-PAGE_SIZE)
 387         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 388                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 389 #undef __FIXADDR_TOP
 390         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 391                 WARN_ON(1);
 392                 printk(KERN_WARNING "pmd %p != %p\n",
 393                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
 394                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 395                         fix_to_virt(FIX_BTMAP_BEGIN));
 396                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 397                         fix_to_virt(FIX_BTMAP_END));
 398
 399                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 400                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
 401                        FIX_BTMAP_BEGIN);
 402         }
 403 }
 404
 405 void __init __early_set_fixmap(enum fixed_addresses idx,
 406                                phys_addr_t phys, pgprot_t flags)
 407 {
 408         unsigned long addr = __fix_to_virt(idx);
 409         pte_t *pte;
 410
 411         if (idx >= __end_of_fixed_addresses) {
 412                 BUG();
 413                 return;
 414         }
 415         pte = early_ioremap_pte(addr);
 416
 417         if (pgprot_val(flags))
 418                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 419         else
 420                 pte_clear(&init_mm, addr, pte);
 421         __flush_tlb_one(addr);
 422 }