fedora core 6 1.2949 + vserver 2.2.0
[linux-2.6.git] / arch / i386 / mm / ioremap.c
index 7b11be0..fff08ae 100644 (file)
 #include <linux/vmalloc.h>
 #include <linux/init.h>
 #include <linux/slab.h>
-#include <asm/io.h>
+#include <linux/module.h>
+#include <linux/io.h>
 #include <asm/fixmap.h>
 #include <asm/cacheflush.h>
 #include <asm/tlbflush.h>
 #include <asm/pgtable.h>
 
-static inline void remap_area_pte(pte_t * pte, unsigned long address, unsigned long size,
-       unsigned long phys_addr, unsigned long flags)
-{
-       unsigned long end;
-       unsigned long pfn;
-
-       address &= ~PMD_MASK;
-       end = address + size;
-       if (end > PMD_SIZE)
-               end = PMD_SIZE;
-       if (address >= end)
-               BUG();
-       pfn = phys_addr >> PAGE_SHIFT;
-       do {
-               if (!pte_none(*pte)) {
-                       printk("remap_area_pte: page already exists\n");
-                       BUG();
-               }
-               set_pte(pte, pfn_pte(pfn, __pgprot(_PAGE_PRESENT | _PAGE_RW | 
-                                       _PAGE_DIRTY | _PAGE_ACCESSED | flags)));
-               address += PAGE_SIZE;
-               pfn++;
-               pte++;
-       } while (address && (address < end));
-}
-
-static inline int remap_area_pmd(pmd_t * pmd, unsigned long address, unsigned long size,
-       unsigned long phys_addr, unsigned long flags)
-{
-       unsigned long end;
-
-       address &= ~PGDIR_MASK;
-       end = address + size;
-       if (end > PGDIR_SIZE)
-               end = PGDIR_SIZE;
-       phys_addr -= address;
-       if (address >= end)
-               BUG();
-       do {
-               pte_t * pte = pte_alloc_kernel(&init_mm, pmd, address);
-               if (!pte)
-                       return -ENOMEM;
-               remap_area_pte(pte, address, end - address, address + phys_addr, flags);
-               address = (address + PMD_SIZE) & PMD_MASK;
-               pmd++;
-       } while (address && (address < end));
-       return 0;
-}
-
-static int remap_area_pages(unsigned long address, unsigned long phys_addr,
-                                unsigned long size, unsigned long flags)
-{
-       int error;
-       pgd_t * dir;
-       unsigned long end = address + size;
-
-       phys_addr -= address;
-       dir = pgd_offset(&init_mm, address);
-       flush_cache_all();
-       if (address >= end)
-               BUG();
-       spin_lock(&init_mm.page_table_lock);
-       do {
-               pmd_t *pmd;
-               pmd = pmd_alloc(&init_mm, dir, address);
-               error = -ENOMEM;
-               if (!pmd)
-                       break;
-               if (remap_area_pmd(pmd, address, end - address,
-                                        phys_addr + address, flags))
-                       break;
-               error = 0;
-               address = (address + PGDIR_SIZE) & PGDIR_MASK;
-               dir++;
-       } while (address && (address < end));
-       spin_unlock(&init_mm.page_table_lock);
-       flush_tlb_all();
-       return error;
-}
+#define ISA_START_ADDRESS      0xa0000
+#define ISA_END_ADDRESS                0x100000
 
 /*
  * Generic mapping function (not visible outside):
@@ -110,11 +34,12 @@ static int remap_area_pages(unsigned long address, unsigned long phys_addr,
  * have to convert them into an offset in a page-aligned mapping, but the
  * caller shouldn't need to know that small detail.
  */
-void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flags)
+void __iomem * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flags)
 {
-       void * addr;
+       void __iomem * addr;
        struct vm_struct * area;
        unsigned long offset, last_addr;
+       pgprot_t prot;
 
        /* Don't allow wraparound or zero size */
        last_addr = phys_addr + size - 1;
@@ -124,13 +49,13 @@ void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flag
        /*
         * Don't remap the low PCI/ISA area, it's always mapped..
         */
-       if (phys_addr >= 0xA0000 && last_addr < 0x100000)
-               return phys_to_virt(phys_addr);
+       if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
+               return (void __iomem *) phys_to_virt(phys_addr);
 
        /*
         * Don't allow anybody to remap normal RAM that we're using..
         */
-       if (phys_addr < virt_to_phys(high_memory)) {
+       if (phys_addr <= virt_to_phys(high_memory - 1)) {
                char *t_addr, *t_end;
                struct page *page;
 
@@ -142,6 +67,9 @@ void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flag
                                return NULL;
        }
 
+       prot = __pgprot(_PAGE_PRESENT | _PAGE_RW | _PAGE_DIRTY
+                       | _PAGE_ACCESSED | flags);
+
        /*
         * Mappings have to be page-aligned
         */
@@ -152,18 +80,19 @@ void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flag
        /*
         * Ok, go for it..
         */
-       area = get_vm_area(size, VM_IOREMAP);
+       area = get_vm_area(size, VM_IOREMAP | (flags << 20));
        if (!area)
                return NULL;
        area->phys_addr = phys_addr;
-       addr = area->addr;
-       if (remap_area_pages((unsigned long) addr, phys_addr, size, flags)) {
-               vunmap(addr);
+       addr = (void __iomem *) area->addr;
+       if (ioremap_page_range((unsigned long) addr,
+                       (unsigned long) addr + size, phys_addr, prot)) {
+               vunmap((void __force *) addr);
                return NULL;
        }
-       return (void *) (offset + (char *)addr);
+       return (void __iomem *) (offset + (char __iomem *)addr);
 }
-
+EXPORT_SYMBOL(__ioremap);
 
 /**
  * ioremap_nocache     -   map bus memory into CPU space
@@ -187,17 +116,17 @@ void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flag
  * Must be freed with iounmap.
  */
 
-void *ioremap_nocache (unsigned long phys_addr, unsigned long size)
+void __iomem *ioremap_nocache (unsigned long phys_addr, unsigned long size)
 {
        unsigned long last_addr;
-       void *p = __ioremap(phys_addr, size, _PAGE_PCD);
+       void __iomem *p = __ioremap(phys_addr, size, _PAGE_PCD);
        if (!p) 
                return p; 
 
        /* Guaranteed to be > phys_addr, as per __ioremap() */
        last_addr = phys_addr + size - 1;
 
-       if (last_addr < virt_to_phys(high_memory)) { 
+       if (last_addr < virt_to_phys(high_memory) - 1) {
                struct page *ppage = virt_to_page(__va(phys_addr));             
                unsigned long npages;
 
@@ -220,26 +149,64 @@ void *ioremap_nocache (unsigned long phys_addr, unsigned long size)
 
        return p;                                       
 }
+EXPORT_SYMBOL(ioremap_nocache);
 
-void iounmap(void *addr)
+/**
+ * iounmap - Free a IO remapping
+ * @addr: virtual address from ioremap_*
+ *
+ * Caller must ensure there is only one unmapping for the same pointer.
+ */
+void iounmap(volatile void __iomem *addr)
 {
-       struct vm_struct *p;
-       if (addr <= high_memory) 
-               return; 
-       p = remove_vm_area((void *) (PAGE_MASK & (unsigned long) addr));
-       if (!p) { 
-               printk("__iounmap: bad address %p\n", addr);
+       struct vm_struct *p, *o;
+
+       if ((void __force *)addr <= high_memory)
                return;
-       } 
 
-       if (p->flags && p->phys_addr < virt_to_phys(high_memory)) { 
+       /*
+        * __ioremap special-cases the PCI/ISA range by not instantiating a
+        * vm_area and by simply returning an address into the kernel mapping
+        * of ISA space.   So handle that here.
+        */
+       if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
+                       addr < phys_to_virt(ISA_END_ADDRESS))
+               return;
+
+       addr = (volatile void __iomem *)(PAGE_MASK & (unsigned long __force)addr);
+
+       /* Use the vm area unlocked, assuming the caller
+          ensures there isn't another iounmap for the same address
+          in parallel. Reuse of the virtual address is prevented by
+          leaving it in the global lists until we're done with it.
+          cpa takes care of the direct mappings. */
+       read_lock(&vmlist_lock);
+       for (p = vmlist; p; p = p->next) {
+               if (p->addr == addr)
+                       break;
+       }
+       read_unlock(&vmlist_lock);
+
+       if (!p) {
+               printk("iounmap: bad address %p\n", addr);
+               dump_stack();
+               return;
+       }
+
+       /* Reset the direct mapping. Can block */
+       if ((p->flags >> 20) && p->phys_addr < virt_to_phys(high_memory) - 1) {
                change_page_attr(virt_to_page(__va(p->phys_addr)),
                                 p->size >> PAGE_SHIFT,
-                                PAGE_KERNEL);                           
+                                PAGE_KERNEL);
                global_flush_tlb();
        } 
+
+       /* Finally remove it */
+       o = remove_vm_area((void *)addr);
+       BUG_ON(p != o || o == NULL);
        kfree(p); 
 }
+EXPORT_SYMBOL(iounmap);
 
 void __init *bt_ioremap(unsigned long phys_addr, unsigned long size)
 {
@@ -255,7 +222,7 @@ void __init *bt_ioremap(unsigned long phys_addr, unsigned long size)
        /*
         * Don't remap the low PCI/ISA area, it's always mapped..
         */
-       if (phys_addr >= 0xA0000 && last_addr < 0x100000)
+       if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
                return phys_to_virt(phys_addr);
 
        /*