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                         enum page_cache_mode pcm)
  33 {
  34         unsigned long nrpages = size >> PAGE_SHIFT;
  35         int err;
  36
  37         switch (pcm) {
  38         case _PAGE_CACHE_MODE_UC:
  39         default:
  40                 err = _set_memory_uc(vaddr, nrpages);
  41                 break;
  42         case _PAGE_CACHE_MODE_WC:
  43                 err = _set_memory_wc(vaddr, nrpages);
  44                 break;
  45         case _PAGE_CACHE_MODE_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, enum page_cache_mode pcm, 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         enum page_cache_mode new_pcm;
  86         pgprot_t prot;
  87         int retval;
  88         void __iomem *ret_addr;
  89         int ram_region;
  90
  91         /* Don't allow wraparound or zero size */
  92         last_addr = phys_addr + size - 1;
  93         if (!size || last_addr < phys_addr)
  94                 return NULL;
  95
  96         if (!phys_addr_valid(phys_addr)) {
  97                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
  98                        (unsigned long long)phys_addr);
  99                 WARN_ON_ONCE(1);
 100                 return NULL;
 101         }
 102
 103         /*
 104          * Don't remap the low PCI/ISA area, it's always mapped..
 105          */
 106         if (is_ISA_range(phys_addr, last_addr))
 107                 return (__force void __iomem *)phys_to_virt(phys_addr);
 108
 109         /*
 110          * Don't allow anybody to remap normal RAM that we're using..
 111          */
 112         /* First check if whole region can be identified as RAM or not */
 113         ram_region = region_is_ram(phys_addr, size);
 114         if (ram_region > 0) {
 115                 WARN_ONCE(1, "ioremap on RAM at 0x%lx - 0x%lx\n",
 116                                 (unsigned long int)phys_addr,
 117                                 (unsigned long int)last_addr);
 118                 return NULL;
 119         }
 120
 121         /* If could not be identified(-1), check page by page */
 122         if (ram_region < 0) {
 123                 pfn      = phys_addr >> PAGE_SHIFT;
 124                 last_pfn = last_addr >> PAGE_SHIFT;
 125                 if (walk_system_ram_range(pfn, last_pfn - pfn + 1, NULL,
 126                                           __ioremap_check_ram) == 1)
 127                         return NULL;
 128         }
 129         /*
 130          * Mappings have to be page-aligned
 131          */
 132         offset = phys_addr & ~PAGE_MASK;
 133         phys_addr &= PHYSICAL_PAGE_MASK;
 134         size = PAGE_ALIGN(last_addr+1) - phys_addr;
 135
 136         retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
 137                                                 pcm, &new_pcm);
 138         if (retval) {
 139                 printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
 140                 return NULL;
 141         }
 142
 143         if (pcm != new_pcm) {
 144                 if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
 145                         printk(KERN_ERR
 146                 "ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
 147                                 (unsigned long long)phys_addr,
 148                                 (unsigned long long)(phys_addr + size),
 149                                 pcm, new_pcm);
 150                         goto err_free_memtype;
 151                 }
 152                 pcm = new_pcm;
 153         }
 154
 155         prot = PAGE_KERNEL_IO;
 156         switch (pcm) {
 157         case _PAGE_CACHE_MODE_UC:
 158         default:
 159                 prot = __pgprot(pgprot_val(prot) |
 160                                 cachemode2protval(_PAGE_CACHE_MODE_UC));
 161                 break;
 162         case _PAGE_CACHE_MODE_UC_MINUS:
 163                 prot = __pgprot(pgprot_val(prot) |
 164                                 cachemode2protval(_PAGE_CACHE_MODE_UC_MINUS));
 165                 break;
 166         case _PAGE_CACHE_MODE_WC:
 167                 prot = __pgprot(pgprot_val(prot) |
 168                                 cachemode2protval(_PAGE_CACHE_MODE_WC));
 169                 break;
 170         case _PAGE_CACHE_MODE_WB:
 171                 break;
 172         }
 173
 174         /*
 175          * Ok, go for it..
 176          */
 177         area = get_vm_area_caller(size, VM_IOREMAP, caller);
 178         if (!area)
 179                 goto err_free_memtype;
 180         area->phys_addr = phys_addr;
 181         vaddr = (unsigned long) area->addr;
 182
 183         if (kernel_map_sync_memtype(phys_addr, size, pcm))
 184                 goto err_free_area;
 185
 186         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
 187                 goto err_free_area;
 188
 189         ret_addr = (void __iomem *) (vaddr + offset);
 190         mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
 191
 192         /*
 193          * Check if the request spans more than any BAR in the iomem resource
 194          * tree.
 195          */
 196         WARN_ONCE(iomem_map_sanity_check(unaligned_phys_addr, unaligned_size),
 197                   KERN_INFO "Info: mapping multiple BARs. Your kernel is fine.");
 198
 199         return ret_addr;
 200 err_free_area:
 201         free_vm_area(area);
 202 err_free_memtype:
 203         free_memtype(phys_addr, phys_addr + size);
 204         return NULL;
 205 }
 206
 207 /**
 208  * ioremap_nocache     -   map bus memory into CPU space
 209  * @phys_addr:    bus address of the memory
 210  * @size:      size of the resource to map
 211  *
 212  * ioremap_nocache performs a platform specific sequence of operations to
 213  * make bus memory CPU accessible via the readb/readw/readl/writeb/
 214  * writew/writel functions and the other mmio helpers. The returned
 215  * address is not guaranteed to be usable directly as a virtual
 216  * address.
 217  *
 218  * This version of ioremap ensures that the memory is marked uncachable
 219  * on the CPU as well as honouring existing caching rules from things like
 220  * the PCI bus. Note that there are other caches and buffers on many
 221  * busses. In particular driver authors should read up on PCI writes
 222  *
 223  * It's useful if some control registers are in such an area and
 224  * write combining or read caching is not desirable:
 225  *
 226  * Must be freed with iounmap.
 227  */
 228 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 229 {
 230         /*
 231          * Ideally, this should be:
 232          *      pat_enabled ? _PAGE_CACHE_MODE_UC : _PAGE_CACHE_MODE_UC_MINUS;
 233          *
 234          * Till we fix all X drivers to use ioremap_wc(), we will use
 235          * UC MINUS.
 236          */
 237         enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC_MINUS;
 238
 239         return __ioremap_caller(phys_addr, size, pcm,
 240                                 __builtin_return_address(0));
 241 }
 242 EXPORT_SYMBOL(ioremap_nocache);
 243
 244 /**
 245  * ioremap_wc   -       map memory into CPU space write combined
 246  * @phys_addr:  bus address of the memory
 247  * @size:       size of the resource to map
 248  *
 249  * This version of ioremap ensures that the memory is marked write combining.
 250  * Write combining allows faster writes to some hardware devices.
 251  *
 252  * Must be freed with iounmap.
 253  */
 254 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
 255 {
 256         if (pat_enabled)
 257                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WC,
 258                                         __builtin_return_address(0));
 259         else
 260                 return ioremap_nocache(phys_addr, size);
 261 }
 262 EXPORT_SYMBOL(ioremap_wc);
 263
 264 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
 265 {
 266         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
 267                                 __builtin_return_address(0));
 268 }
 269 EXPORT_SYMBOL(ioremap_cache);
 270
 271 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
 272                                 unsigned long prot_val)
 273 {
 274         return __ioremap_caller(phys_addr, size,
 275                                 pgprot2cachemode(__pgprot(prot_val)),
 276                                 __builtin_return_address(0));
 277 }
 278 EXPORT_SYMBOL(ioremap_prot);
 279
 280 /**
 281  * iounmap - Free a IO remapping
 282  * @addr: virtual address from ioremap_*
 283  *
 284  * Caller must ensure there is only one unmapping for the same pointer.
 285  */
 286 void iounmap(volatile void __iomem *addr)
 287 {
 288         struct vm_struct *p, *o;
 289
 290         if ((void __force *)addr <= high_memory)
 291                 return;
 292
 293         /*
 294          * __ioremap special-cases the PCI/ISA range by not instantiating a
 295          * vm_area and by simply returning an address into the kernel mapping
 296          * of ISA space.   So handle that here.
 297          */
 298         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
 299             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
 300                 return;
 301
 302         addr = (volatile void __iomem *)
 303                 (PAGE_MASK & (unsigned long __force)addr);
 304
 305         mmiotrace_iounmap(addr);
 306
 307         /* Use the vm area unlocked, assuming the caller
 308            ensures there isn't another iounmap for the same address
 309            in parallel. Reuse of the virtual address is prevented by
 310            leaving it in the global lists until we're done with it.
 311            cpa takes care of the direct mappings. */
 312         p = find_vm_area((void __force *)addr);
 313
 314         if (!p) {
 315                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
 316                 dump_stack();
 317                 return;
 318         }
 319
 320         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
 321
 322         /* Finally remove it */
 323         o = remove_vm_area((void __force *)addr);
 324         BUG_ON(p != o || o == NULL);
 325         kfree(p);
 326 }
 327 EXPORT_SYMBOL(iounmap);
 328
 329 /*
 330  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
 331  * access
 332  */
 333 void *xlate_dev_mem_ptr(phys_addr_t phys)
 334 {
 335         void *addr;
 336         unsigned long start = phys & PAGE_MASK;
 337
 338         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
 339         if (page_is_ram(start >> PAGE_SHIFT))
 340                 return __va(phys);
 341
 342         addr = (void __force *)ioremap_cache(start, PAGE_SIZE);
 343         if (addr)
 344                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
 345
 346         return addr;
 347 }
 348
 349 void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr)
 350 {
 351         if (page_is_ram(phys >> PAGE_SHIFT))
 352                 return;
 353
 354         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
 355         return;
 356 }
 357
 358 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
 359
 360 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
 361 {
 362         /* Don't assume we're using swapper_pg_dir at this point */
 363         pgd_t *base = __va(read_cr3());
 364         pgd_t *pgd = &base[pgd_index(addr)];
 365         pud_t *pud = pud_offset(pgd, addr);
 366         pmd_t *pmd = pmd_offset(pud, addr);
 367
 368         return pmd;
 369 }
 370
 371 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
 372 {
 373         return &bm_pte[pte_index(addr)];
 374 }
 375
 376 bool __init is_early_ioremap_ptep(pte_t *ptep)
 377 {
 378         return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
 379 }
 380
 381 void __init early_ioremap_init(void)
 382 {
 383         pmd_t *pmd;
 384
 385 #ifdef CONFIG_X86_64
 386         BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 387 #else
 388         WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
 389 #endif
 390
 391         early_ioremap_setup();
 392
 393         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 394         memset(bm_pte, 0, sizeof(bm_pte));
 395         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 396
 397         /*
 398          * The boot-ioremap range spans multiple pmds, for which
 399          * we are not prepared:
 400          */
 401 #define __FIXADDR_TOP (-PAGE_SIZE)
 402         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 403                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 404 #undef __FIXADDR_TOP
 405         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 406                 WARN_ON(1);
 407                 printk(KERN_WARNING "pmd %p != %p\n",
 408                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
 409                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 410                         fix_to_virt(FIX_BTMAP_BEGIN));
 411                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 412                         fix_to_virt(FIX_BTMAP_END));
 413
 414                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 415                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
 416                        FIX_BTMAP_BEGIN);
 417         }
 418 }
 419
 420 void __init __early_set_fixmap(enum fixed_addresses idx,
 421                                phys_addr_t phys, pgprot_t flags)
 422 {
 423         unsigned long addr = __fix_to_virt(idx);
 424         pte_t *pte;
 425
 426         if (idx >= __end_of_fixed_addresses) {
 427                 BUG();
 428                 return;
 429         }
 430         pte = early_ioremap_pte(addr);
 431
 432         if (pgprot_val(flags))
 433                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 434         else
 435                 pte_clear(&init_mm, addr, pte);
 436         __flush_tlb_one(addr);
 437 }