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 #ifdef CONFIG_X86_64
  26
  27 unsigned long __phys_addr(unsigned long x)
  28 {
  29         if (x >= __START_KERNEL_map)
  30                 return x - __START_KERNEL_map + phys_base;
  31         return x - PAGE_OFFSET;
  32 }
  33 EXPORT_SYMBOL(__phys_addr);
  34
  35 static inline int phys_addr_valid(unsigned long addr)
  36 {
  37         return addr < (1UL << boot_cpu_data.x86_phys_bits);
  38 }
  39
  40 #else
  41
  42 static inline int phys_addr_valid(unsigned long addr)
  43 {
  44         return 1;
  45 }
  46
  47 #endif
  48
  49 int page_is_ram(unsigned long pagenr)
  50 {
  51         resource_size_t addr, end;
  52         int i;
  53
  54         /*
  55          * A special case is the first 4Kb of memory;
  56          * This is a BIOS owned area, not kernel ram, but generally
  57          * not listed as such in the E820 table.
  58          */
  59         if (pagenr == 0)
  60                 return 0;
  61
  62         /*
  63          * Second special case: Some BIOSen report the PC BIOS
  64          * area (640->1Mb) as ram even though it is not.
  65          */
  66         if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
  67                     pagenr < (BIOS_END >> PAGE_SHIFT))
  68                 return 0;
  69
  70         for (i = 0; i < e820.nr_map; i++) {
  71                 /*
  72                  * Not usable memory:
  73                  */
  74                 if (e820.map[i].type != E820_RAM)
  75                         continue;
  76                 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
  77                 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
  78
  79
  80                 if ((pagenr >= addr) && (pagenr < end))
  81                         return 1;
  82         }
  83         return 0;
  84 }
  85
  86 /*
  87  * Fix up the linear direct mapping of the kernel to avoid cache attribute
  88  * conflicts.
  89  */
  90 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
  91                                unsigned long prot_val)
  92 {
  93         unsigned long nrpages = size >> PAGE_SHIFT;
  94         int err;
  95
  96         switch (prot_val) {
  97         case _PAGE_CACHE_UC:
  98         default:
  99                 err = _set_memory_uc(vaddr, nrpages);
 100                 break;
 101         case _PAGE_CACHE_WC:
 102                 err = _set_memory_wc(vaddr, nrpages);
 103                 break;
 104         case _PAGE_CACHE_WB:
 105                 err = _set_memory_wb(vaddr, nrpages);
 106                 break;
 107         }
 108
 109         return err;
 110 }
 111
 112 /*
 113  * Remap an arbitrary physical address space into the kernel virtual
 114  * address space. Needed when the kernel wants to access high addresses
 115  * directly.
 116  *
 117  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
 118  * have to convert them into an offset in a page-aligned mapping, but the
 119  * caller shouldn't need to know that small detail.
 120  */
 121 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
 122                 unsigned long size, unsigned long prot_val, void *caller)
 123 {
 124         unsigned long pfn, offset, vaddr;
 125         resource_size_t last_addr;
 126         const resource_size_t unaligned_phys_addr = phys_addr;
 127         const unsigned long unaligned_size = size;
 128         struct vm_struct *area;
 129         unsigned long new_prot_val;
 130         pgprot_t prot;
 131         int retval;
 132         void __iomem *ret_addr;
 133
 134         /* Don't allow wraparound or zero size */
 135         last_addr = phys_addr + size - 1;
 136         if (!size || last_addr < phys_addr)
 137                 return NULL;
 138
 139         if (!phys_addr_valid(phys_addr)) {
 140                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
 141                        (unsigned long long)phys_addr);
 142                 WARN_ON_ONCE(1);
 143                 return NULL;
 144         }
 145
 146         /*
 147          * Don't remap the low PCI/ISA area, it's always mapped..
 148          */
 149         if (is_ISA_range(phys_addr, last_addr))
 150                 return (__force void __iomem *)phys_to_virt(phys_addr);
 151
 152         /*
 153          * Don't allow anybody to remap normal RAM that we're using..
 154          */
 155         for (pfn = phys_addr >> PAGE_SHIFT;
 156                                 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
 157                                 pfn++) {
 158
 159                 int is_ram = page_is_ram(pfn);
 160
 161                 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
 162                         return NULL;
 163                 WARN_ON_ONCE(is_ram);
 164         }
 165
 166         /*
 167          * Mappings have to be page-aligned
 168          */
 169         offset = phys_addr & ~PAGE_MASK;
 170         phys_addr &= PAGE_MASK;
 171         size = PAGE_ALIGN(last_addr+1) - phys_addr;
 172
 173         retval = reserve_memtype(phys_addr, phys_addr + size,
 174                                                 prot_val, &new_prot_val);
 175         if (retval) {
 176                 pr_debug("Warning: reserve_memtype returned %d\n", retval);
 177                 return NULL;
 178         }
 179
 180         if (prot_val != new_prot_val) {
 181                 /*
 182                  * Do not fallback to certain memory types with certain
 183                  * requested type:
 184                  * - request is uc-, return cannot be write-back
 185                  * - request is uc-, return cannot be write-combine
 186                  * - request is write-combine, return cannot be write-back
 187                  */
 188                 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
 189                      (new_prot_val == _PAGE_CACHE_WB ||
 190                       new_prot_val == _PAGE_CACHE_WC)) ||
 191                     (prot_val == _PAGE_CACHE_WC &&
 192                      new_prot_val == _PAGE_CACHE_WB)) {
 193                         pr_debug(
 194                 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
 195                                 (unsigned long long)phys_addr,
 196                                 (unsigned long long)(phys_addr + size),
 197                                 prot_val, new_prot_val);
 198                         free_memtype(phys_addr, phys_addr + size);
 199                         return NULL;
 200                 }
 201                 prot_val = new_prot_val;
 202         }
 203
 204         switch (prot_val) {
 205         case _PAGE_CACHE_UC:
 206         default:
 207                 prot = PAGE_KERNEL_NOCACHE;
 208                 break;
 209         case _PAGE_CACHE_UC_MINUS:
 210                 prot = PAGE_KERNEL_UC_MINUS;
 211                 break;
 212         case _PAGE_CACHE_WC:
 213                 prot = PAGE_KERNEL_WC;
 214                 break;
 215         case _PAGE_CACHE_WB:
 216                 prot = PAGE_KERNEL;
 217                 break;
 218         }
 219
 220         /*
 221          * Ok, go for it..
 222          */
 223         area = get_vm_area_caller(size, VM_IOREMAP, caller);
 224         if (!area)
 225                 return NULL;
 226         area->phys_addr = phys_addr;
 227         vaddr = (unsigned long) area->addr;
 228         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
 229                 free_memtype(phys_addr, phys_addr + size);
 230                 free_vm_area(area);
 231                 return NULL;
 232         }
 233
 234         if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
 235                 free_memtype(phys_addr, phys_addr + size);
 236                 vunmap(area->addr);
 237                 return NULL;
 238         }
 239
 240         ret_addr = (void __iomem *) (vaddr + offset);
 241         mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
 242
 243         return ret_addr;
 244 }
 245
 246 /**
 247  * ioremap_nocache     -   map bus memory into CPU space
 248  * @offset:    bus address of the memory
 249  * @size:      size of the resource to map
 250  *
 251  * ioremap_nocache performs a platform specific sequence of operations to
 252  * make bus memory CPU accessible via the readb/readw/readl/writeb/
 253  * writew/writel functions and the other mmio helpers. The returned
 254  * address is not guaranteed to be usable directly as a virtual
 255  * address.
 256  *
 257  * This version of ioremap ensures that the memory is marked uncachable
 258  * on the CPU as well as honouring existing caching rules from things like
 259  * the PCI bus. Note that there are other caches and buffers on many
 260  * busses. In particular driver authors should read up on PCI writes
 261  *
 262  * It's useful if some control registers are in such an area and
 263  * write combining or read caching is not desirable:
 264  *
 265  * Must be freed with iounmap.
 266  */
 267 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
 268 {
 269         /*
 270          * Ideally, this should be:
 271          *      pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
 272          *
 273          * Till we fix all X drivers to use ioremap_wc(), we will use
 274          * UC MINUS.
 275          */
 276         unsigned long val = _PAGE_CACHE_UC_MINUS;
 277
 278         return __ioremap_caller(phys_addr, size, val,
 279                                 __builtin_return_address(0));
 280 }
 281 EXPORT_SYMBOL(ioremap_nocache);
 282
 283 /**
 284  * ioremap_wc   -       map memory into CPU space write combined
 285  * @offset:     bus address of the memory
 286  * @size:       size of the resource to map
 287  *
 288  * This version of ioremap ensures that the memory is marked write combining.
 289  * Write combining allows faster writes to some hardware devices.
 290  *
 291  * Must be freed with iounmap.
 292  */
 293 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
 294 {
 295         if (pat_enabled)
 296                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
 297                                         __builtin_return_address(0));
 298         else
 299                 return ioremap_nocache(phys_addr, size);
 300 }
 301 EXPORT_SYMBOL(ioremap_wc);
 302
 303 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
 304 {
 305         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
 306                                 __builtin_return_address(0));
 307 }
 308 EXPORT_SYMBOL(ioremap_cache);
 309
 310 static void __iomem *ioremap_default(resource_size_t phys_addr,
 311                                         unsigned long size)
 312 {
 313         unsigned long flags;
 314         void *ret;
 315         int err;
 316
 317         /*
 318          * - WB for WB-able memory and no other conflicting mappings
 319          * - UC_MINUS for non-WB-able memory with no other conflicting mappings
 320          * - Inherit from confliting mappings otherwise
 321          */
 322         err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
 323         if (err < 0)
 324                 return NULL;
 325
 326         ret = (void *) __ioremap_caller(phys_addr, size, flags,
 327                                         __builtin_return_address(0));
 328
 329         free_memtype(phys_addr, phys_addr + size);
 330         return (void __iomem *)ret;
 331 }
 332
 333 /**
 334  * iounmap - Free a IO remapping
 335  * @addr: virtual address from ioremap_*
 336  *
 337  * Caller must ensure there is only one unmapping for the same pointer.
 338  */
 339 void iounmap(volatile void __iomem *addr)
 340 {
 341         struct vm_struct *p, *o;
 342
 343         if ((void __force *)addr <= high_memory)
 344                 return;
 345
 346         /*
 347          * __ioremap special-cases the PCI/ISA range by not instantiating a
 348          * vm_area and by simply returning an address into the kernel mapping
 349          * of ISA space.   So handle that here.
 350          */
 351         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
 352             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
 353                 return;
 354
 355         addr = (volatile void __iomem *)
 356                 (PAGE_MASK & (unsigned long __force)addr);
 357
 358         mmiotrace_iounmap(addr);
 359
 360         /* Use the vm area unlocked, assuming the caller
 361            ensures there isn't another iounmap for the same address
 362            in parallel. Reuse of the virtual address is prevented by
 363            leaving it in the global lists until we're done with it.
 364            cpa takes care of the direct mappings. */
 365         read_lock(&vmlist_lock);
 366         for (p = vmlist; p; p = p->next) {
 367                 if (p->addr == (void __force *)addr)
 368                         break;
 369         }
 370         read_unlock(&vmlist_lock);
 371
 372         if (!p) {
 373                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
 374                 dump_stack();
 375                 return;
 376         }
 377
 378         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
 379
 380         /* Finally remove it */
 381         o = remove_vm_area((void __force *)addr);
 382         BUG_ON(p != o || o == NULL);
 383         kfree(p);
 384 }
 385 EXPORT_SYMBOL(iounmap);
 386
 387 /*
 388  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
 389  * access
 390  */
 391 void *xlate_dev_mem_ptr(unsigned long phys)
 392 {
 393         void *addr;
 394         unsigned long start = phys & PAGE_MASK;
 395
 396         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
 397         if (page_is_ram(start >> PAGE_SHIFT))
 398                 return __va(phys);
 399
 400         addr = (void __force *)ioremap_default(start, PAGE_SIZE);
 401         if (addr)
 402                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
 403
 404         return addr;
 405 }
 406
 407 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
 408 {
 409         if (page_is_ram(phys >> PAGE_SHIFT))
 410                 return;
 411
 412         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
 413         return;
 414 }
 415
 416 int __initdata early_ioremap_debug;
 417
 418 static int __init early_ioremap_debug_setup(char *str)
 419 {
 420         early_ioremap_debug = 1;
 421
 422         return 0;
 423 }
 424 early_param("early_ioremap_debug", early_ioremap_debug_setup);
 425
 426 static __initdata int after_paging_init;
 427 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
 428
 429 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
 430 {
 431         /* Don't assume we're using swapper_pg_dir at this point */
 432         pgd_t *base = __va(read_cr3());
 433         pgd_t *pgd = &base[pgd_index(addr)];
 434         pud_t *pud = pud_offset(pgd, addr);
 435         pmd_t *pmd = pmd_offset(pud, addr);
 436
 437         return pmd;
 438 }
 439
 440 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
 441 {
 442         return &bm_pte[pte_index(addr)];
 443 }
 444
 445 void __init early_ioremap_init(void)
 446 {
 447         pmd_t *pmd;
 448
 449         if (early_ioremap_debug)
 450                 printk(KERN_INFO "early_ioremap_init()\n");
 451
 452         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 453         memset(bm_pte, 0, sizeof(bm_pte));
 454         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 455
 456         /*
 457          * The boot-ioremap range spans multiple pmds, for which
 458          * we are not prepared:
 459          */
 460         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 461                 WARN_ON(1);
 462                 printk(KERN_WARNING "pmd %p != %p\n",
 463                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
 464                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 465                         fix_to_virt(FIX_BTMAP_BEGIN));
 466                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 467                         fix_to_virt(FIX_BTMAP_END));
 468
 469                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 470                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
 471                        FIX_BTMAP_BEGIN);
 472         }
 473 }
 474
 475 void __init early_ioremap_clear(void)
 476 {
 477         pmd_t *pmd;
 478
 479         if (early_ioremap_debug)
 480                 printk(KERN_INFO "early_ioremap_clear()\n");
 481
 482         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 483         pmd_clear(pmd);
 484         paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
 485         __flush_tlb_all();
 486 }
 487
 488 void __init early_ioremap_reset(void)
 489 {
 490         enum fixed_addresses idx;
 491         unsigned long addr, phys;
 492         pte_t *pte;
 493
 494         after_paging_init = 1;
 495         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
 496                 addr = fix_to_virt(idx);
 497                 pte = early_ioremap_pte(addr);
 498                 if (pte_present(*pte)) {
 499                         phys = pte_val(*pte) & PAGE_MASK;
 500                         set_fixmap(idx, phys);
 501                 }
 502         }
 503 }
 504
 505 static void __init __early_set_fixmap(enum fixed_addresses idx,
 506                                    unsigned long phys, pgprot_t flags)
 507 {
 508         unsigned long addr = __fix_to_virt(idx);
 509         pte_t *pte;
 510
 511         if (idx >= __end_of_fixed_addresses) {
 512                 BUG();
 513                 return;
 514         }
 515         pte = early_ioremap_pte(addr);
 516
 517         if (pgprot_val(flags))
 518                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 519         else
 520                 pte_clear(&init_mm, addr, pte);
 521         __flush_tlb_one(addr);
 522 }
 523
 524 static inline void __init early_set_fixmap(enum fixed_addresses idx,
 525                                         unsigned long phys)
 526 {
 527         if (after_paging_init)
 528                 set_fixmap(idx, phys);
 529         else
 530                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
 531 }
 532
 533 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
 534 {
 535         if (after_paging_init)
 536                 clear_fixmap(idx);
 537         else
 538                 __early_set_fixmap(idx, 0, __pgprot(0));
 539 }
 540
 541
 542 int __initdata early_ioremap_nested;
 543
 544 static int __init check_early_ioremap_leak(void)
 545 {
 546         if (!early_ioremap_nested)
 547                 return 0;
 548
 549         printk(KERN_WARNING
 550                "Debug warning: early ioremap leak of %d areas detected.\n",
 551                early_ioremap_nested);
 552         printk(KERN_WARNING
 553                "please boot with early_ioremap_debug and report the dmesg.\n");
 554         WARN_ON(1);
 555
 556         return 1;
 557 }
 558 late_initcall(check_early_ioremap_leak);
 559
 560 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
 561 {
 562         unsigned long offset, last_addr;
 563         unsigned int nrpages, nesting;
 564         enum fixed_addresses idx0, idx;
 565
 566         WARN_ON(system_state != SYSTEM_BOOTING);
 567
 568         nesting = early_ioremap_nested;
 569         if (early_ioremap_debug) {
 570                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
 571                        phys_addr, size, nesting);
 572                 dump_stack();
 573         }
 574
 575         /* Don't allow wraparound or zero size */
 576         last_addr = phys_addr + size - 1;
 577         if (!size || last_addr < phys_addr) {
 578                 WARN_ON(1);
 579                 return NULL;
 580         }
 581
 582         if (nesting >= FIX_BTMAPS_NESTING) {
 583                 WARN_ON(1);
 584                 return NULL;
 585         }
 586         early_ioremap_nested++;
 587         /*
 588          * Mappings have to be page-aligned
 589          */
 590         offset = phys_addr & ~PAGE_MASK;
 591         phys_addr &= PAGE_MASK;
 592         size = PAGE_ALIGN(last_addr) - phys_addr;
 593
 594         /*
 595          * Mappings have to fit in the FIX_BTMAP area.
 596          */
 597         nrpages = size >> PAGE_SHIFT;
 598         if (nrpages > NR_FIX_BTMAPS) {
 599                 WARN_ON(1);
 600                 return NULL;
 601         }
 602
 603         /*
 604          * Ok, go for it..
 605          */
 606         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
 607         idx = idx0;
 608         while (nrpages > 0) {
 609                 early_set_fixmap(idx, phys_addr);
 610                 phys_addr += PAGE_SIZE;
 611                 --idx;
 612                 --nrpages;
 613         }
 614         if (early_ioremap_debug)
 615                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
 616
 617         return (void *) (offset + fix_to_virt(idx0));
 618 }
 619
 620 void __init early_iounmap(void *addr, unsigned long size)
 621 {
 622         unsigned long virt_addr;
 623         unsigned long offset;
 624         unsigned int nrpages;
 625         enum fixed_addresses idx;
 626         int nesting;
 627
 628         nesting = --early_ioremap_nested;
 629         if (WARN_ON(nesting < 0))
 630                 return;
 631
 632         if (early_ioremap_debug) {
 633                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
 634                        size, nesting);
 635                 dump_stack();
 636         }
 637
 638         virt_addr = (unsigned long)addr;
 639         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
 640                 WARN_ON(1);
 641                 return;
 642         }
 643         offset = virt_addr & ~PAGE_MASK;
 644         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
 645
 646         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
 647         while (nrpages > 0) {
 648                 early_clear_fixmap(idx);
 649                 --idx;
 650                 --nrpages;
 651         }
 652 }
 653
 654 void __this_fixmap_does_not_exist(void)
 655 {
 656         WARN_ON(1);
 657 }