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  * @offset:    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  * @offset:     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         read_lock(&vmlist_lock);
 286         for (p = vmlist; p; p = p->next) {
 287                 if (p->addr == (void __force *)addr)
 288                         break;
 289         }
 290         read_unlock(&vmlist_lock);
 291
 292         if (!p) {
 293                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
 294                 dump_stack();
 295                 return;
 296         }
 297
 298         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
 299
 300         /* Finally remove it */
 301         o = remove_vm_area((void __force *)addr);
 302         BUG_ON(p != o || o == NULL);
 303         kfree(p);
 304 }
 305 EXPORT_SYMBOL(iounmap);
 306
 307 /*
 308  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
 309  * access
 310  */
 311 void *xlate_dev_mem_ptr(unsigned long phys)
 312 {
 313         void *addr;
 314         unsigned long start = phys & PAGE_MASK;
 315
 316         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
 317         if (page_is_ram(start >> PAGE_SHIFT))
 318                 return __va(phys);
 319
 320         addr = (void __force *)ioremap_cache(start, PAGE_SIZE);
 321         if (addr)
 322                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
 323
 324         return addr;
 325 }
 326
 327 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
 328 {
 329         if (page_is_ram(phys >> PAGE_SHIFT))
 330                 return;
 331
 332         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
 333         return;
 334 }
 335
 336 static int __initdata early_ioremap_debug;
 337
 338 static int __init early_ioremap_debug_setup(char *str)
 339 {
 340         early_ioremap_debug = 1;
 341
 342         return 0;
 343 }
 344 early_param("early_ioremap_debug", early_ioremap_debug_setup);
 345
 346 static __initdata int after_paging_init;
 347 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
 348
 349 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
 350 {
 351         /* Don't assume we're using swapper_pg_dir at this point */
 352         pgd_t *base = __va(read_cr3());
 353         pgd_t *pgd = &base[pgd_index(addr)];
 354         pud_t *pud = pud_offset(pgd, addr);
 355         pmd_t *pmd = pmd_offset(pud, addr);
 356
 357         return pmd;
 358 }
 359
 360 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
 361 {
 362         return &bm_pte[pte_index(addr)];
 363 }
 364
 365 bool __init is_early_ioremap_ptep(pte_t *ptep)
 366 {
 367         return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
 368 }
 369
 370 static unsigned long slot_virt[FIX_BTMAPS_SLOTS] __initdata;
 371
 372 void __init early_ioremap_init(void)
 373 {
 374         pmd_t *pmd;
 375         int i;
 376
 377         if (early_ioremap_debug)
 378                 printk(KERN_INFO "early_ioremap_init()\n");
 379
 380         for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
 381                 slot_virt[i] = __fix_to_virt(FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*i);
 382
 383         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 384         memset(bm_pte, 0, sizeof(bm_pte));
 385         pmd_populate_kernel(&init_mm, pmd, bm_pte);
 386
 387         /*
 388          * The boot-ioremap range spans multiple pmds, for which
 389          * we are not prepared:
 390          */
 391 #define __FIXADDR_TOP (-PAGE_SIZE)
 392         BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
 393                      != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
 394 #undef __FIXADDR_TOP
 395         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
 396                 WARN_ON(1);
 397                 printk(KERN_WARNING "pmd %p != %p\n",
 398                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
 399                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
 400                         fix_to_virt(FIX_BTMAP_BEGIN));
 401                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
 402                         fix_to_virt(FIX_BTMAP_END));
 403
 404                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
 405                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
 406                        FIX_BTMAP_BEGIN);
 407         }
 408 }
 409
 410 void __init early_ioremap_reset(void)
 411 {
 412         after_paging_init = 1;
 413 }
 414
 415 static void __init __early_set_fixmap(enum fixed_addresses idx,
 416                                       phys_addr_t phys, pgprot_t flags)
 417 {
 418         unsigned long addr = __fix_to_virt(idx);
 419         pte_t *pte;
 420
 421         if (idx >= __end_of_fixed_addresses) {
 422                 BUG();
 423                 return;
 424         }
 425         pte = early_ioremap_pte(addr);
 426
 427         if (pgprot_val(flags))
 428                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
 429         else
 430                 pte_clear(&init_mm, addr, pte);
 431         __flush_tlb_one(addr);
 432 }
 433
 434 static inline void __init early_set_fixmap(enum fixed_addresses idx,
 435                                            phys_addr_t phys, pgprot_t prot)
 436 {
 437         if (after_paging_init)
 438                 __set_fixmap(idx, phys, prot);
 439         else
 440                 __early_set_fixmap(idx, phys, prot);
 441 }
 442
 443 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
 444 {
 445         if (after_paging_init)
 446                 clear_fixmap(idx);
 447         else
 448                 __early_set_fixmap(idx, 0, __pgprot(0));
 449 }
 450
 451 static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
 452 static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
 453
 454 void __init fixup_early_ioremap(void)
 455 {
 456         int i;
 457
 458         for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
 459                 if (prev_map[i]) {
 460                         WARN_ON(1);
 461                         break;
 462                 }
 463         }
 464
 465         early_ioremap_init();
 466 }
 467
 468 static int __init check_early_ioremap_leak(void)
 469 {
 470         int count = 0;
 471         int i;
 472
 473         for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
 474                 if (prev_map[i])
 475                         count++;
 476
 477         if (!count)
 478                 return 0;
 479         WARN(1, KERN_WARNING
 480                "Debug warning: early ioremap leak of %d areas detected.\n",
 481                 count);
 482         printk(KERN_WARNING
 483                 "please boot with early_ioremap_debug and report the dmesg.\n");
 484
 485         return 1;
 486 }
 487 late_initcall(check_early_ioremap_leak);
 488
 489 static void __init __iomem *
 490 __early_ioremap(resource_size_t phys_addr, unsigned long size, pgprot_t prot)
 491 {
 492         unsigned long offset;
 493         resource_size_t last_addr;
 494         unsigned int nrpages;
 495         enum fixed_addresses idx0, idx;
 496         int i, slot;
 497
 498         WARN_ON(system_state != SYSTEM_BOOTING);
 499
 500         slot = -1;
 501         for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
 502                 if (!prev_map[i]) {
 503                         slot = i;
 504                         break;
 505                 }
 506         }
 507
 508         if (slot < 0) {
 509                 printk(KERN_INFO "early_iomap(%08llx, %08lx) not found slot\n",
 510                          (u64)phys_addr, size);
 511                 WARN_ON(1);
 512                 return NULL;
 513         }
 514
 515         if (early_ioremap_debug) {
 516                 printk(KERN_INFO "early_ioremap(%08llx, %08lx) [%d] => ",
 517                        (u64)phys_addr, size, slot);
 518                 dump_stack();
 519         }
 520
 521         /* Don't allow wraparound or zero size */
 522         last_addr = phys_addr + size - 1;
 523         if (!size || last_addr < phys_addr) {
 524                 WARN_ON(1);
 525                 return NULL;
 526         }
 527
 528         prev_size[slot] = size;
 529         /*
 530          * Mappings have to be page-aligned
 531          */
 532         offset = phys_addr & ~PAGE_MASK;
 533         phys_addr &= PAGE_MASK;
 534         size = PAGE_ALIGN(last_addr + 1) - phys_addr;
 535
 536         /*
 537          * Mappings have to fit in the FIX_BTMAP area.
 538          */
 539         nrpages = size >> PAGE_SHIFT;
 540         if (nrpages > NR_FIX_BTMAPS) {
 541                 WARN_ON(1);
 542                 return NULL;
 543         }
 544
 545         /*
 546          * Ok, go for it..
 547          */
 548         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
 549         idx = idx0;
 550         while (nrpages > 0) {
 551                 early_set_fixmap(idx, phys_addr, prot);
 552                 phys_addr += PAGE_SIZE;
 553                 --idx;
 554                 --nrpages;
 555         }
 556         if (early_ioremap_debug)
 557                 printk(KERN_CONT "%08lx + %08lx\n", offset, slot_virt[slot]);
 558
 559         prev_map[slot] = (void __iomem *)(offset + slot_virt[slot]);
 560         return prev_map[slot];
 561 }
 562
 563 /* Remap an IO device */
 564 void __init __iomem *
 565 early_ioremap(resource_size_t phys_addr, unsigned long size)
 566 {
 567         return __early_ioremap(phys_addr, size, PAGE_KERNEL_IO);
 568 }
 569
 570 /* Remap memory */
 571 void __init __iomem *
 572 early_memremap(resource_size_t phys_addr, unsigned long size)
 573 {
 574         return __early_ioremap(phys_addr, size, PAGE_KERNEL);
 575 }
 576
 577 void __init early_iounmap(void __iomem *addr, unsigned long size)
 578 {
 579         unsigned long virt_addr;
 580         unsigned long offset;
 581         unsigned int nrpages;
 582         enum fixed_addresses idx;
 583         int i, slot;
 584
 585         slot = -1;
 586         for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
 587                 if (prev_map[i] == addr) {
 588                         slot = i;
 589                         break;
 590                 }
 591         }
 592
 593         if (slot < 0) {
 594                 printk(KERN_INFO "early_iounmap(%p, %08lx) not found slot\n",
 595                          addr, size);
 596                 WARN_ON(1);
 597                 return;
 598         }
 599
 600         if (prev_size[slot] != size) {
 601                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d] size not consistent %08lx\n",
 602                          addr, size, slot, prev_size[slot]);
 603                 WARN_ON(1);
 604                 return;
 605         }
 606
 607         if (early_ioremap_debug) {
 608                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
 609                        size, slot);
 610                 dump_stack();
 611         }
 612
 613         virt_addr = (unsigned long)addr;
 614         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
 615                 WARN_ON(1);
 616                 return;
 617         }
 618         offset = virt_addr & ~PAGE_MASK;
 619         nrpages = PAGE_ALIGN(offset + size) >> PAGE_SHIFT;
 620
 621         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;
 622         while (nrpages > 0) {
 623                 early_clear_fixmap(idx);
 624                 --idx;
 625                 --nrpages;
 626         }
 627         prev_map[slot] = NULL;
 628 }